323 files changed, 118084 insertions, 0 deletions

diff --git a/src/rocksdb/util/aligned_buffer.h b/src/rocksdb/util/aligned_buffer.h
new file mode 100644
index 000000000..95ee5dfe8
--- /dev/null
+++ b/src/rocksdb/util/aligned_buffer.h
@@ -0,0 +1,234 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+//
+// Copyright (c) 2011 The LevelDB Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file. See the AUTHORS file for names of contributors.
+#pragma once
+
+#include <algorithm>
+
+#include "port/port.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+// This file contains utilities to handle the alignment of pages and buffers.
+
+// Truncate to a multiple of page_size, which is also a page boundary. This
+// helps to figuring out the right alignment.
+// Example:
+//   TruncateToPageBoundary(4096, 5000)  => 4096
+//   TruncateToPageBoundary((4096, 10000) => 8192
+inline size_t TruncateToPageBoundary(size_t page_size, size_t s) {
+  s -= (s & (page_size - 1));
+  assert((s % page_size) == 0);
+  return s;
+}
+
+// Round up x to a multiple of y.
+// Example:
+//   Roundup(13, 5)   => 15
+//   Roundup(201, 16) => 208
+inline size_t Roundup(size_t x, size_t y) { return ((x + y - 1) / y) * y; }
+
+// Round down x to a multiple of y.
+// Example:
+//   Rounddown(13, 5)   => 10
+//   Rounddown(201, 16) => 192
+inline size_t Rounddown(size_t x, size_t y) { return (x / y) * y; }
+
+// AlignedBuffer manages a buffer by taking alignment into consideration, and
+// aligns the buffer start and end positions. It is mainly used for direct I/O,
+// though it can be used other purposes as well.
+// It also supports expanding the managed buffer, and copying whole or part of
+// the data from old buffer into the new expanded buffer. Such a copy especially
+// helps in cases avoiding an IO to re-fetch the data from disk.
+//
+// Example:
+//   AlignedBuffer buf;
+//   buf.Alignment(alignment);
+//   buf.AllocateNewBuffer(user_requested_buf_size);
+//   ...
+//   buf.AllocateNewBuffer(2*user_requested_buf_size, /*copy_data*/ true,
+//                         copy_offset, copy_len);
+class AlignedBuffer {
+  size_t alignment_;
+  std::unique_ptr<char[]> buf_;
+  size_t capacity_;
+  size_t cursize_;
+  char* bufstart_;
+
+ public:
+  AlignedBuffer()
+      : alignment_(), capacity_(0), cursize_(0), bufstart_(nullptr) {}
+
+  AlignedBuffer(AlignedBuffer&& o) noexcept { *this = std::move(o); }
+
+  AlignedBuffer& operator=(AlignedBuffer&& o) noexcept {
+    alignment_ = std::move(o.alignment_);
+    buf_ = std::move(o.buf_);
+    capacity_ = std::move(o.capacity_);
+    cursize_ = std::move(o.cursize_);
+    bufstart_ = std::move(o.bufstart_);
+    return *this;
+  }
+
+  AlignedBuffer(const AlignedBuffer&) = delete;
+
+  AlignedBuffer& operator=(const AlignedBuffer&) = delete;
+
+  static bool isAligned(const void* ptr, size_t alignment) {
+    return reinterpret_cast<uintptr_t>(ptr) % alignment == 0;
+  }
+
+  static bool isAligned(size_t n, size_t alignment) {
+    return n % alignment == 0;
+  }
+
+  size_t Alignment() const { return alignment_; }
+
+  size_t Capacity() const { return capacity_; }
+
+  size_t CurrentSize() const { return cursize_; }
+
+  const char* BufferStart() const { return bufstart_; }
+
+  char* BufferStart() { return bufstart_; }
+
+  void Clear() { cursize_ = 0; }
+
+  char* Release() {
+    cursize_ = 0;
+    capacity_ = 0;
+    bufstart_ = nullptr;
+    return buf_.release();
+  }
+
+  void Alignment(size_t alignment) {
+    assert(alignment > 0);
+    assert((alignment & (alignment - 1)) == 0);
+    alignment_ = alignment;
+  }
+
+  // Allocates a new buffer and sets the start position to the first aligned
+  // byte.
+  //
+  // requested_capacity: requested new buffer capacity. This capacity will be
+  //     rounded up based on alignment.
+  // copy_data: Copy data from old buffer to new buffer. If copy_offset and
+  //     copy_len are not passed in and the new requested capacity is bigger
+  //     than the existing buffer's capacity, the data in the exising buffer is
+  //     fully copied over to the new buffer.
+  // copy_offset: Copy data from this offset in old buffer.
+  // copy_len: Number of bytes to copy.
+  //
+  // The function does nothing if the new requested_capacity is smaller than
+  // the current buffer capacity and copy_data is true i.e. the old buffer is
+  // retained as is.
+  void AllocateNewBuffer(size_t requested_capacity, bool copy_data = false,
+                         uint64_t copy_offset = 0, size_t copy_len = 0) {
+    assert(alignment_ > 0);
+    assert((alignment_ & (alignment_ - 1)) == 0);
+
+    copy_len = copy_len > 0 ? copy_len : cursize_;
+    if (copy_data && requested_capacity < copy_len) {
+      // If we are downsizing to a capacity that is smaller than the current
+      // data in the buffer -- Ignore the request.
+      return;
+    }
+
+    size_t new_capacity = Roundup(requested_capacity, alignment_);
+    char* new_buf = new char[new_capacity + alignment_];
+    char* new_bufstart = reinterpret_cast<char*>(
+        (reinterpret_cast<uintptr_t>(new_buf) + (alignment_ - 1)) &
+        ~static_cast<uintptr_t>(alignment_ - 1));
+
+    if (copy_data) {
+      assert(bufstart_ + copy_offset + copy_len <= bufstart_ + cursize_);
+      memcpy(new_bufstart, bufstart_ + copy_offset, copy_len);
+      cursize_ = copy_len;
+    } else {
+      cursize_ = 0;
+    }
+
+    bufstart_ = new_bufstart;
+    capacity_ = new_capacity;
+    buf_.reset(new_buf);
+  }
+
+  // Append to the buffer.
+  //
+  // src         : source to copy the data from.
+  // append_size : number of bytes to copy from src.
+  // Returns the number of bytes appended.
+  //
+  // If append_size is more than the remaining buffer size only the
+  // remaining-size worth of bytes are copied.
+  size_t Append(const char* src, size_t append_size) {
+    size_t buffer_remaining = capacity_ - cursize_;
+    size_t to_copy = std::min(append_size, buffer_remaining);
+
+    if (to_copy > 0) {
+      memcpy(bufstart_ + cursize_, src, to_copy);
+      cursize_ += to_copy;
+    }
+    return to_copy;
+  }
+
+  // Read from the buffer.
+  //
+  // dest      : destination buffer to copy the data to.
+  // offset    : the buffer offset to start reading from.
+  // read_size : the number of bytes to copy from the buffer to dest.
+  // Returns the number of bytes read/copied to dest.
+  size_t Read(char* dest, size_t offset, size_t read_size) const {
+    assert(offset < cursize_);
+
+    size_t to_read = 0;
+    if (offset < cursize_) {
+      to_read = std::min(cursize_ - offset, read_size);
+    }
+    if (to_read > 0) {
+      memcpy(dest, bufstart_ + offset, to_read);
+    }
+    return to_read;
+  }
+
+  // Pad to the end of alignment with "padding"
+  void PadToAlignmentWith(int padding) {
+    size_t total_size = Roundup(cursize_, alignment_);
+    size_t pad_size = total_size - cursize_;
+
+    if (pad_size > 0) {
+      assert((pad_size + cursize_) <= capacity_);
+      memset(bufstart_ + cursize_, padding, pad_size);
+      cursize_ += pad_size;
+    }
+  }
+
+  void PadWith(size_t pad_size, int padding) {
+    assert((pad_size + cursize_) <= capacity_);
+    memset(bufstart_ + cursize_, padding, pad_size);
+    cursize_ += pad_size;
+  }
+
+  // After a partial flush move the tail to the beginning of the buffer.
+  void RefitTail(size_t tail_offset, size_t tail_size) {
+    if (tail_size > 0) {
+      memmove(bufstart_, bufstart_ + tail_offset, tail_size);
+    }
+    cursize_ = tail_size;
+  }
+
+  // Returns a place to start appending.
+  // WARNING: Note that it is possible to write past the end of the buffer if
+  // the buffer is modified without using the write APIs or encapsulation
+  // offered by AlignedBuffer. It is up to the user to guard against such
+  // errors.
+  char* Destination() { return bufstart_ + cursize_; }
+
+  void Size(size_t cursize) { cursize_ = cursize; }
+};
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/util/async_file_reader.cc b/src/rocksdb/util/async_file_reader.cc
new file mode 100644
index 000000000..8401a6b44
--- /dev/null
+++ b/src/rocksdb/util/async_file_reader.cc
@@ -0,0 +1,73 @@
+//  Copyright (c) Meta Platforms, Inc. and affiliates.
+//
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+//
+#if USE_COROUTINES
+#include "util/async_file_reader.h"
+
+namespace ROCKSDB_NAMESPACE {
+bool AsyncFileReader::MultiReadAsyncImpl(ReadAwaiter* awaiter) {
+  if (tail_) {
+    tail_->next_ = awaiter;
+  }
+  tail_ = awaiter;
+  if (!head_) {
+    head_ = awaiter;
+  }
+  num_reqs_ += awaiter->num_reqs_;
+  awaiter->io_handle_.resize(awaiter->num_reqs_);
+  awaiter->del_fn_.resize(awaiter->num_reqs_);
+  for (size_t i = 0; i < awaiter->num_reqs_; ++i) {
+    awaiter->file_
+        ->ReadAsync(
+            awaiter->read_reqs_[i], awaiter->opts_,
+            [](const FSReadRequest& req, void* cb_arg) {
+              FSReadRequest* read_req = static_cast<FSReadRequest*>(cb_arg);
+              read_req->status = req.status;
+              read_req->result = req.result;
+            },
+            &awaiter->read_reqs_[i], &awaiter->io_handle_[i],
+            &awaiter->del_fn_[i], /*aligned_buf=*/nullptr)
+        .PermitUncheckedError();
+  }
+  return true;
+}
+
+void AsyncFileReader::Wait() {
+  if (!head_) {
+    return;
+  }
+  ReadAwaiter* waiter;
+  std::vector<void*> io_handles;
+  io_handles.reserve(num_reqs_);
+  waiter = head_;
+  do {
+    for (size_t i = 0; i < waiter->num_reqs_; ++i) {
+      if (waiter->io_handle_[i]) {
+        io_handles.push_back(waiter->io_handle_[i]);
+      }
+    }
+  } while (waiter != tail_ && (waiter = waiter->next_));
+  if (io_handles.size() > 0) {
+    StopWatch sw(SystemClock::Default().get(), stats_, POLL_WAIT_MICROS);
+    fs_->Poll(io_handles, io_handles.size()).PermitUncheckedError();
+  }
+  do {
+    waiter = head_;
+    head_ = waiter->next_;
+
+    for (size_t i = 0; i < waiter->num_reqs_; ++i) {
+      if (waiter->io_handle_[i] && waiter->del_fn_[i]) {
+        waiter->del_fn_[i](waiter->io_handle_[i]);
+      }
+    }
+    waiter->awaiting_coro_.resume();
+  } while (waiter != tail_);
+  head_ = tail_ = nullptr;
+  RecordInHistogram(stats_, MULTIGET_IO_BATCH_SIZE, num_reqs_);
+  num_reqs_ = 0;
+}
+}  // namespace ROCKSDB_NAMESPACE
+#endif  // USE_COROUTINES
diff --git a/src/rocksdb/util/async_file_reader.h b/src/rocksdb/util/async_file_reader.h
new file mode 100644
index 000000000..df69a840e
--- /dev/null
+++ b/src/rocksdb/util/async_file_reader.h
@@ -0,0 +1,144 @@
+//  Copyright (c) Meta Platforms, Inc. and affiliates.
+//
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).#pragma once
+#pragma once
+
+#if USE_COROUTINES
+#include "file/random_access_file_reader.h"
+#include "folly/experimental/coro/ViaIfAsync.h"
+#include "port/port.h"
+#include "rocksdb/file_system.h"
+#include "rocksdb/statistics.h"
+#include "util/autovector.h"
+#include "util/stop_watch.h"
+
+namespace ROCKSDB_NAMESPACE {
+class SingleThreadExecutor;
+
+// AsyncFileReader implements the Awaitable concept, which allows calling
+// coroutines to co_await it. When the AsyncFileReader Awaitable is
+// resumed, it initiates the fie reads requested by the awaiting caller
+// by calling RandomAccessFileReader's ReadAsync. It then suspends the
+// awaiting coroutine. The suspended awaiter is later resumed by Wait().
+class AsyncFileReader {
+  class ReadAwaiter;
+  template <typename Awaiter>
+  class ReadOperation;
+
+ public:
+  AsyncFileReader(FileSystem* fs, Statistics* stats) : fs_(fs), stats_(stats) {}
+
+  ~AsyncFileReader() {}
+
+  ReadOperation<ReadAwaiter> MultiReadAsync(RandomAccessFileReader* file,
+                                            const IOOptions& opts,
+                                            FSReadRequest* read_reqs,
+                                            size_t num_reqs,
+                                            AlignedBuf* aligned_buf) noexcept {
+    return ReadOperation<ReadAwaiter>{*this,     file,     opts,
+                                      read_reqs, num_reqs, aligned_buf};
+  }
+
+ private:
+  friend SingleThreadExecutor;
+
+  // Implementation of the Awaitable concept
+  class ReadAwaiter {
+   public:
+    explicit ReadAwaiter(AsyncFileReader& reader, RandomAccessFileReader* file,
+                         const IOOptions& opts, FSReadRequest* read_reqs,
+                         size_t num_reqs, AlignedBuf* /*aligned_buf*/) noexcept
+        : reader_(reader),
+          file_(file),
+          opts_(opts),
+          read_reqs_(read_reqs),
+          num_reqs_(num_reqs),
+          next_(nullptr) {}
+
+    bool await_ready() noexcept { return false; }
+
+    // A return value of true means suspend the awaiter (calling coroutine). The
+    // awaiting_coro parameter is the handle of the awaiter. The handle can be
+    // resumed later, so we cache it here.
+    bool await_suspend(
+        folly::coro::impl::coroutine_handle<> awaiting_coro) noexcept {
+      awaiting_coro_ = awaiting_coro;
+      // MultiReadAsyncImpl always returns true, so caller will be suspended
+      return reader_.MultiReadAsyncImpl(this);
+    }
+
+    void await_resume() noexcept {}
+
+   private:
+    friend AsyncFileReader;
+
+    // The parameters passed to MultiReadAsync are cached here when the caller
+    // calls MultiReadAsync. Later, when the execution of this awaitable is
+    // started, these are used to do the actual IO
+    AsyncFileReader& reader_;
+    RandomAccessFileReader* file_;
+    const IOOptions& opts_;
+    FSReadRequest* read_reqs_;
+    size_t num_reqs_;
+    autovector<void*, 32> io_handle_;
+    autovector<IOHandleDeleter, 32> del_fn_;
+    folly::coro::impl::coroutine_handle<> awaiting_coro_;
+    // Use this to link to the next ReadAwaiter in the suspended coroutine
+    // list. The head and tail of the list are tracked by AsyncFileReader.
+    // We use this approach rather than an STL container in order to avoid
+    // extra memory allocations. The coroutine call already allocates a
+    // ReadAwaiter object.
+    ReadAwaiter* next_;
+  };
+
+  // An instance of ReadOperation is returned to the caller of MultiGetAsync.
+  // This represents an awaitable that can be started later.
+  template <typename Awaiter>
+  class ReadOperation {
+   public:
+    explicit ReadOperation(AsyncFileReader& reader,
+                           RandomAccessFileReader* file, const IOOptions& opts,
+                           FSReadRequest* read_reqs, size_t num_reqs,
+                           AlignedBuf* aligned_buf) noexcept
+        : reader_(reader),
+          file_(file),
+          opts_(opts),
+          read_reqs_(read_reqs),
+          num_reqs_(num_reqs),
+          aligned_buf_(aligned_buf) {}
+
+    auto viaIfAsync(folly::Executor::KeepAlive<> executor) const {
+      return folly::coro::co_viaIfAsync(
+          std::move(executor),
+          Awaiter{reader_, file_, opts_, read_reqs_, num_reqs_, aligned_buf_});
+    }
+
+   private:
+    AsyncFileReader& reader_;
+    RandomAccessFileReader* file_;
+    const IOOptions& opts_;
+    FSReadRequest* read_reqs_;
+    size_t num_reqs_;
+    AlignedBuf* aligned_buf_;
+  };
+
+  // This function does the actual work when this awaitable starts execution
+  bool MultiReadAsyncImpl(ReadAwaiter* awaiter);
+
+  // Called by the SingleThreadExecutor to poll for async IO completion.
+  // This also resumes the awaiting coroutines.
+  void Wait();
+
+  // Head of the queue of awaiters waiting for async IO completion
+  ReadAwaiter* head_ = nullptr;
+  // Tail of the awaiter queue
+  ReadAwaiter* tail_ = nullptr;
+  // Total number of pending async IOs
+  size_t num_reqs_ = 0;
+  FileSystem* fs_;
+  Statistics* stats_;
+};
+}  // namespace ROCKSDB_NAMESPACE
+#endif  // USE_COROUTINES
diff --git a/src/rocksdb/util/autovector.h b/src/rocksdb/util/autovector.h
new file mode 100644
index 000000000..f758473b7
--- /dev/null
+++ b/src/rocksdb/util/autovector.h
@@ -0,0 +1,406 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+#pragma once
+
+#include <algorithm>
+#include <cassert>
+#include <initializer_list>
+#include <iterator>
+#include <stdexcept>
+#include <vector>
+
+#include "port/lang.h"
+#include "rocksdb/rocksdb_namespace.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+#ifdef ROCKSDB_LITE
+template <class T, size_t kSize = 8>
+class autovector : public std::vector<T> {
+  using std::vector<T>::vector;
+
+ public:
+  autovector() {
+    // Make sure the initial vector has space for kSize elements
+    std::vector<T>::reserve(kSize);
+  }
+};
+#else
+// A vector that leverages pre-allocated stack-based array to achieve better
+// performance for array with small amount of items.
+//
+// The interface resembles that of vector, but with less features since we aim
+// to solve the problem that we have in hand, rather than implementing a
+// full-fledged generic container.
+//
+// Currently we don't support:
+//  * shrink_to_fit()
+//     If used correctly, in most cases, people should not touch the
+//     underlying vector at all.
+//  * random insert()/erase(), please only use push_back()/pop_back().
+//  * No move/swap operations. Each autovector instance has a
+//     stack-allocated array and if we want support move/swap operations, we
+//     need to copy the arrays other than just swapping the pointers. In this
+//     case we'll just explicitly forbid these operations since they may
+//     lead users to make false assumption by thinking they are inexpensive
+//     operations.
+//
+// Naming style of public methods almost follows that of the STL's.
+template <class T, size_t kSize = 8>
+class autovector {
+ public:
+  // General STL-style container member types.
+  using value_type = T;
+  using difference_type = typename std::vector<T>::difference_type;
+  using size_type = typename std::vector<T>::size_type;
+  using reference = value_type&;
+  using const_reference = const value_type&;
+  using pointer = value_type*;
+  using const_pointer = const value_type*;
+
+  // This class is the base for regular/const iterator
+  template <class TAutoVector, class TValueType>
+  class iterator_impl {
+   public:
+    // -- iterator traits
+    using self_type = iterator_impl<TAutoVector, TValueType>;
+    using value_type = TValueType;
+    using reference = TValueType&;
+    using pointer = TValueType*;
+    using difference_type = typename TAutoVector::difference_type;
+    using iterator_category = std::random_access_iterator_tag;
+
+    iterator_impl(TAutoVector* vect, size_t index)
+        : vect_(vect), index_(index){};
+    iterator_impl(const iterator_impl&) = default;
+    ~iterator_impl() {}
+    iterator_impl& operator=(const iterator_impl&) = default;
+
+    // -- Advancement
+    // ++iterator
+    self_type& operator++() {
+      ++index_;
+      return *this;
+    }
+
+    // iterator++
+    self_type operator++(int) {
+      auto old = *this;
+      ++index_;
+      return old;
+    }
+
+    // --iterator
+    self_type& operator--() {
+      --index_;
+      return *this;
+    }
+
+    // iterator--
+    self_type operator--(int) {
+      auto old = *this;
+      --index_;
+      return old;
+    }
+
+    self_type operator-(difference_type len) const {
+      return self_type(vect_, index_ - len);
+    }
+
+    difference_type operator-(const self_type& other) const {
+      assert(vect_ == other.vect_);
+      return index_ - other.index_;
+    }
+
+    self_type operator+(difference_type len) const {
+      return self_type(vect_, index_ + len);
+    }
+
+    self_type& operator+=(difference_type len) {
+      index_ += len;
+      return *this;
+    }
+
+    self_type& operator-=(difference_type len) {
+      index_ -= len;
+      return *this;
+    }
+
+    // -- Reference
+    reference operator*() const {
+      assert(vect_->size() >= index_);
+      return (*vect_)[index_];
+    }
+
+    pointer operator->() const {
+      assert(vect_->size() >= index_);
+      return &(*vect_)[index_];
+    }
+
+    reference operator[](difference_type len) const { return *(*this + len); }
+
+    // -- Logical Operators
+    bool operator==(const self_type& other) const {
+      assert(vect_ == other.vect_);
+      return index_ == other.index_;
+    }
+
+    bool operator!=(const self_type& other) const { return !(*this == other); }
+
+    bool operator>(const self_type& other) const {
+      assert(vect_ == other.vect_);
+      return index_ > other.index_;
+    }
+
+    bool operator<(const self_type& other) const {
+      assert(vect_ == other.vect_);
+      return index_ < other.index_;
+    }
+
+    bool operator>=(const self_type& other) const {
+      assert(vect_ == other.vect_);
+      return index_ >= other.index_;
+    }
+
+    bool operator<=(const self_type& other) const {
+      assert(vect_ == other.vect_);
+      return index_ <= other.index_;
+    }
+
+   private:
+    TAutoVector* vect_ = nullptr;
+    size_t index_ = 0;
+  };
+
+  using iterator = iterator_impl<autovector, value_type>;
+  using const_iterator = iterator_impl<const autovector, const value_type>;
+  using reverse_iterator = std::reverse_iterator<iterator>;
+  using const_reverse_iterator = std::reverse_iterator<const_iterator>;
+
+  autovector() : values_(reinterpret_cast<pointer>(buf_)) {}
+
+  autovector(std::initializer_list<T> init_list)
+      : values_(reinterpret_cast<pointer>(buf_)) {
+    for (const T& item : init_list) {
+      push_back(item);
+    }
+  }
+
+  ~autovector() { clear(); }
+
+  // -- Immutable operations
+  // Indicate if all data resides in in-stack data structure.
+  bool only_in_stack() const {
+    // If no element was inserted at all, the vector's capacity will be `0`.
+    return vect_.capacity() == 0;
+  }
+
+  size_type size() const { return num_stack_items_ + vect_.size(); }
+
+  // resize does not guarantee anything about the contents of the newly
+  // available elements
+  void resize(size_type n) {
+    if (n > kSize) {
+      vect_.resize(n - kSize);
+      while (num_stack_items_ < kSize) {
+        new ((void*)(&values_[num_stack_items_++])) value_type();
+      }
+      num_stack_items_ = kSize;
+    } else {
+      vect_.clear();
+      while (num_stack_items_ < n) {
+        new ((void*)(&values_[num_stack_items_++])) value_type();
+      }
+      while (num_stack_items_ > n) {
+        values_[--num_stack_items_].~value_type();
+      }
+    }
+  }
+
+  bool empty() const { return size() == 0; }
+
+  size_type capacity() const { return kSize + vect_.capacity(); }
+
+  void reserve(size_t cap) {
+    if (cap > kSize) {
+      vect_.reserve(cap - kSize);
+    }
+
+    assert(cap <= capacity());
+  }
+
+  const_reference operator[](size_type n) const {
+    assert(n < size());
+    if (n < kSize) {
+      return values_[n];
+    }
+    return vect_[n - kSize];
+  }
+
+  reference operator[](size_type n) {
+    assert(n < size());
+    if (n < kSize) {
+      return values_[n];
+    }
+    return vect_[n - kSize];
+  }
+
+  const_reference at(size_type n) const {
+    assert(n < size());
+    return (*this)[n];
+  }
+
+  reference at(size_type n) {
+    assert(n < size());
+    return (*this)[n];
+  }
+
+  reference front() {
+    assert(!empty());
+    return *begin();
+  }
+
+  const_reference front() const {
+    assert(!empty());
+    return *begin();
+  }
+
+  reference back() {
+    assert(!empty());
+    return *(end() - 1);
+  }
+
+  const_reference back() const {
+    assert(!empty());
+    return *(end() - 1);
+  }
+
+  // -- Mutable Operations
+  void push_back(T&& item) {
+    if (num_stack_items_ < kSize) {
+      new ((void*)(&values_[num_stack_items_])) value_type();
+      values_[num_stack_items_++] = std::move(item);
+    } else {
+      vect_.push_back(item);
+    }
+  }
+
+  void push_back(const T& item) {
+    if (num_stack_items_ < kSize) {
+      new ((void*)(&values_[num_stack_items_])) value_type();
+      values_[num_stack_items_++] = item;
+    } else {
+      vect_.push_back(item);
+    }
+  }
+
+  template <class... Args>
+#if _LIBCPP_STD_VER > 14
+  reference emplace_back(Args&&... args) {
+    if (num_stack_items_ < kSize) {
+      return *(new ((void*)(&values_[num_stack_items_++]))
+                   value_type(std::forward<Args>(args)...));
+    } else {
+      return vect_.emplace_back(std::forward<Args>(args)...);
+    }
+  }
+#else
+  void emplace_back(Args&&... args) {
+    if (num_stack_items_ < kSize) {
+      new ((void*)(&values_[num_stack_items_++]))
+          value_type(std::forward<Args>(args)...);
+    } else {
+      vect_.emplace_back(std::forward<Args>(args)...);
+    }
+  }
+#endif
+
+  void pop_back() {
+    assert(!empty());
+    if (!vect_.empty()) {
+      vect_.pop_back();
+    } else {
+      values_[--num_stack_items_].~value_type();
+    }
+  }
+
+  void clear() {
+    while (num_stack_items_ > 0) {
+      values_[--num_stack_items_].~value_type();
+    }
+    vect_.clear();
+  }
+
+  // -- Copy and Assignment
+  autovector& assign(const autovector& other);
+
+  autovector(const autovector& other) { assign(other); }
+
+  autovector& operator=(const autovector& other) { return assign(other); }
+
+  autovector(autovector&& other) noexcept { *this = std::move(other); }
+  autovector& operator=(autovector&& other);
+
+  // -- Iterator Operations
+  iterator begin() { return iterator(this, 0); }
+
+  const_iterator begin() const { return const_iterator(this, 0); }
+
+  iterator end() { return iterator(this, this->size()); }
+
+  const_iterator end() const { return const_iterator(this, this->size()); }
+
+  reverse_iterator rbegin() { return reverse_iterator(end()); }
+
+  const_reverse_iterator rbegin() const {
+    return const_reverse_iterator(end());
+  }
+
+  reverse_iterator rend() { return reverse_iterator(begin()); }
+
+  const_reverse_iterator rend() const {
+    return const_reverse_iterator(begin());
+  }
+
+ private:
+  size_type num_stack_items_ = 0;  // current number of items
+  alignas(alignof(
+      value_type)) char buf_[kSize *
+                             sizeof(value_type)];  // the first `kSize` items
+  pointer values_;
+  // used only if there are more than `kSize` items.
+  std::vector<T> vect_;
+};
+
+template <class T, size_t kSize>
+autovector<T, kSize>& autovector<T, kSize>::assign(
+    const autovector<T, kSize>& other) {
+  values_ = reinterpret_cast<pointer>(buf_);
+  // copy the internal vector
+  vect_.assign(other.vect_.begin(), other.vect_.end());
+
+  // copy array
+  num_stack_items_ = other.num_stack_items_;
+  std::copy(other.values_, other.values_ + num_stack_items_, values_);
+
+  return *this;
+}
+
+template <class T, size_t kSize>
+autovector<T, kSize>& autovector<T, kSize>::operator=(
+    autovector<T, kSize>&& other) {
+  values_ = reinterpret_cast<pointer>(buf_);
+  vect_ = std::move(other.vect_);
+  size_t n = other.num_stack_items_;
+  num_stack_items_ = n;
+  other.num_stack_items_ = 0;
+  for (size_t i = 0; i < n; ++i) {
+    values_[i] = std::move(other.values_[i]);
+  }
+  return *this;
+}
+
+#endif  // ROCKSDB_LITE
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/util/autovector_test.cc b/src/rocksdb/util/autovector_test.cc
new file mode 100644
index 000000000..8c7c39ce6
--- /dev/null
+++ b/src/rocksdb/util/autovector_test.cc
@@ -0,0 +1,331 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#include "util/autovector.h"
+
+#include <atomic>
+#include <iostream>
+#include <string>
+#include <utility>
+
+#include "rocksdb/env.h"
+#include "test_util/testharness.h"
+#include "test_util/testutil.h"
+#include "util/string_util.h"
+
+using std::cout;
+using std::endl;
+
+namespace ROCKSDB_NAMESPACE {
+
+class AutoVectorTest : public testing::Test {};
+const unsigned long kSize = 8;
+
+namespace {
+template <class T>
+void AssertAutoVectorOnlyInStack(autovector<T, kSize>* vec, bool result) {
+#ifndef ROCKSDB_LITE
+  ASSERT_EQ(vec->only_in_stack(), result);
+#else
+  (void)vec;
+  (void)result;
+#endif  // !ROCKSDB_LITE
+}
+}  // namespace
+
+TEST_F(AutoVectorTest, PushBackAndPopBack) {
+  autovector<size_t, kSize> vec;
+  ASSERT_TRUE(vec.empty());
+  ASSERT_EQ(0ul, vec.size());
+
+  for (size_t i = 0; i < 1000 * kSize; ++i) {
+    vec.push_back(i);
+    ASSERT_TRUE(!vec.empty());
+    if (i < kSize) {
+      AssertAutoVectorOnlyInStack(&vec, true);
+    } else {
+      AssertAutoVectorOnlyInStack(&vec, false);
+    }
+    ASSERT_EQ(i + 1, vec.size());
+    ASSERT_EQ(i, vec[i]);
+    ASSERT_EQ(i, vec.at(i));
+  }
+
+  size_t size = vec.size();
+  while (size != 0) {
+    vec.pop_back();
+    // will always be in heap
+    AssertAutoVectorOnlyInStack(&vec, false);
+    ASSERT_EQ(--size, vec.size());
+  }
+
+  ASSERT_TRUE(vec.empty());
+}
+
+TEST_F(AutoVectorTest, EmplaceBack) {
+  using ValType = std::pair<size_t, std::string>;
+  autovector<ValType, kSize> vec;
+
+  for (size_t i = 0; i < 1000 * kSize; ++i) {
+    vec.emplace_back(i, std::to_string(i + 123));
+    ASSERT_TRUE(!vec.empty());
+    if (i < kSize) {
+      AssertAutoVectorOnlyInStack(&vec, true);
+    } else {
+      AssertAutoVectorOnlyInStack(&vec, false);
+    }
+
+    ASSERT_EQ(i + 1, vec.size());
+    ASSERT_EQ(i, vec[i].first);
+    ASSERT_EQ(std::to_string(i + 123), vec[i].second);
+  }
+
+  vec.clear();
+  ASSERT_TRUE(vec.empty());
+  AssertAutoVectorOnlyInStack(&vec, false);
+}
+
+TEST_F(AutoVectorTest, Resize) {
+  autovector<size_t, kSize> vec;
+
+  vec.resize(kSize);
+  AssertAutoVectorOnlyInStack(&vec, true);
+  for (size_t i = 0; i < kSize; ++i) {
+    vec[i] = i;
+  }
+
+  vec.resize(kSize * 2);
+  AssertAutoVectorOnlyInStack(&vec, false);
+  for (size_t i = 0; i < kSize; ++i) {
+    ASSERT_EQ(vec[i], i);
+  }
+  for (size_t i = 0; i < kSize; ++i) {
+    vec[i + kSize] = i;
+  }
+
+  vec.resize(1);
+  ASSERT_EQ(1U, vec.size());
+}
+
+namespace {
+void AssertEqual(const autovector<size_t, kSize>& a,
+                 const autovector<size_t, kSize>& b) {
+  ASSERT_EQ(a.size(), b.size());
+  ASSERT_EQ(a.empty(), b.empty());
+#ifndef ROCKSDB_LITE
+  ASSERT_EQ(a.only_in_stack(), b.only_in_stack());
+#endif  // !ROCKSDB_LITE
+  for (size_t i = 0; i < a.size(); ++i) {
+    ASSERT_EQ(a[i], b[i]);
+  }
+}
+}  // namespace
+
+TEST_F(AutoVectorTest, CopyAndAssignment) {
+  // Test both heap-allocated and stack-allocated cases.
+  for (auto size : {kSize / 2, kSize * 1000}) {
+    autovector<size_t, kSize> vec;
+    for (size_t i = 0; i < size; ++i) {
+      vec.push_back(i);
+    }
+
+    {
+      autovector<size_t, kSize> other;
+      other = vec;
+      AssertEqual(other, vec);
+    }
+
+    {
+      autovector<size_t, kSize> other(vec);
+      AssertEqual(other, vec);
+    }
+  }
+}
+
+TEST_F(AutoVectorTest, Iterators) {
+  autovector<std::string, kSize> vec;
+  for (size_t i = 0; i < kSize * 1000; ++i) {
+    vec.push_back(std::to_string(i));
+  }
+
+  // basic operator test
+  ASSERT_EQ(vec.front(), *vec.begin());
+  ASSERT_EQ(vec.back(), *(vec.end() - 1));
+  ASSERT_TRUE(vec.begin() < vec.end());
+
+  // non-const iterator
+  size_t index = 0;
+  for (const auto& item : vec) {
+    ASSERT_EQ(vec[index++], item);
+  }
+
+  index = vec.size() - 1;
+  for (auto pos = vec.rbegin(); pos != vec.rend(); ++pos) {
+    ASSERT_EQ(vec[index--], *pos);
+  }
+
+  // const iterator
+  const auto& cvec = vec;
+  index = 0;
+  for (const auto& item : cvec) {
+    ASSERT_EQ(cvec[index++], item);
+  }
+
+  index = vec.size() - 1;
+  for (auto pos = cvec.rbegin(); pos != cvec.rend(); ++pos) {
+    ASSERT_EQ(cvec[index--], *pos);
+  }
+
+  // forward and backward
+  auto pos = vec.begin();
+  while (pos != vec.end()) {
+    auto old_val = *pos;
+    auto old = pos++;
+    // HACK: make sure -> works
+    ASSERT_TRUE(!old->empty());
+    ASSERT_EQ(old_val, *old);
+    ASSERT_TRUE(pos == vec.end() || old_val != *pos);
+  }
+
+  pos = vec.begin();
+  for (size_t i = 0; i < vec.size(); i += 2) {
+    // Cannot use ASSERT_EQ since that macro depends on iostream serialization
+    ASSERT_TRUE(pos + 2 - 2 == pos);
+    pos += 2;
+    ASSERT_TRUE(pos >= vec.begin());
+    ASSERT_TRUE(pos <= vec.end());
+
+    size_t diff = static_cast<size_t>(pos - vec.begin());
+    ASSERT_EQ(i + 2, diff);
+  }
+}
+
+namespace {
+std::vector<std::string> GetTestKeys(size_t size) {
+  std::vector<std::string> keys;
+  keys.resize(size);
+
+  int index = 0;
+  for (auto& key : keys) {
+    key = "item-" + std::to_string(index++);
+  }
+  return keys;
+}
+}  // namespace
+
+template <class TVector>
+void BenchmarkVectorCreationAndInsertion(
+    std::string name, size_t ops, size_t item_size,
+    const std::vector<typename TVector::value_type>& items) {
+  auto env = Env::Default();
+
+  int index = 0;
+  auto start_time = env->NowNanos();
+  auto ops_remaining = ops;
+  while (ops_remaining--) {
+    TVector v;
+    for (size_t i = 0; i < item_size; ++i) {
+      v.push_back(items[index++]);
+    }
+  }
+  auto elapsed = env->NowNanos() - start_time;
+  cout << "created " << ops << " " << name << " instances:\n\t"
+       << "each was inserted with " << item_size << " elements\n\t"
+       << "total time elapsed: " << elapsed << " (ns)" << endl;
+}
+
+template <class TVector>
+size_t BenchmarkSequenceAccess(std::string name, size_t ops, size_t elem_size) {
+  TVector v;
+  for (const auto& item : GetTestKeys(elem_size)) {
+    v.push_back(item);
+  }
+  auto env = Env::Default();
+
+  auto ops_remaining = ops;
+  auto start_time = env->NowNanos();
+  size_t total = 0;
+  while (ops_remaining--) {
+    auto end = v.end();
+    for (auto pos = v.begin(); pos != end; ++pos) {
+      total += pos->size();
+    }
+  }
+  auto elapsed = env->NowNanos() - start_time;
+  cout << "performed " << ops << " sequence access against " << name << "\n\t"
+       << "size: " << elem_size << "\n\t"
+       << "total time elapsed: " << elapsed << " (ns)" << endl;
+  // HACK avoid compiler's optimization to ignore total
+  return total;
+}
+
+// This test case only reports the performance between std::vector<std::string>
+// and autovector<std::string>. We chose string for comparison because in most
+// of our use cases we used std::vector<std::string>.
+TEST_F(AutoVectorTest, PerfBench) {
+  // We run same operations for kOps times in order to get a more fair result.
+  size_t kOps = 100000;
+
+  // Creation and insertion test
+  // Test the case when there is:
+  //  * no element inserted: internal array of std::vector may not really get
+  //    initialize.
+  //  * one element inserted: internal array of std::vector must have
+  //    initialized.
+  //  * kSize elements inserted. This shows the most time we'll spend if we
+  //    keep everything in stack.
+  //  * 2 * kSize elements inserted. The internal vector of
+  //    autovector must have been initialized.
+  cout << "=====================================================" << endl;
+  cout << "Creation and Insertion Test (value type: std::string)" << endl;
+  cout << "=====================================================" << endl;
+
+  // pre-generated unique keys
+  auto string_keys = GetTestKeys(kOps * 2 * kSize);
+  for (auto insertions : {0ul, 1ul, kSize / 2, kSize, 2 * kSize}) {
+    BenchmarkVectorCreationAndInsertion<std::vector<std::string>>(
+        "std::vector<std::string>", kOps, insertions, string_keys);
+    BenchmarkVectorCreationAndInsertion<autovector<std::string, kSize>>(
+        "autovector<std::string>", kOps, insertions, string_keys);
+    cout << "-----------------------------------" << endl;
+  }
+
+  cout << "=====================================================" << endl;
+  cout << "Creation and Insertion Test (value type: uint64_t)" << endl;
+  cout << "=====================================================" << endl;
+
+  // pre-generated unique keys
+  std::vector<uint64_t> int_keys(kOps * 2 * kSize);
+  for (size_t i = 0; i < kOps * 2 * kSize; ++i) {
+    int_keys[i] = i;
+  }
+  for (auto insertions : {0ul, 1ul, kSize / 2, kSize, 2 * kSize}) {
+    BenchmarkVectorCreationAndInsertion<std::vector<uint64_t>>(
+        "std::vector<uint64_t>", kOps, insertions, int_keys);
+    BenchmarkVectorCreationAndInsertion<autovector<uint64_t, kSize>>(
+        "autovector<uint64_t>", kOps, insertions, int_keys);
+    cout << "-----------------------------------" << endl;
+  }
+
+  // Sequence Access Test
+  cout << "=====================================================" << endl;
+  cout << "Sequence Access Test" << endl;
+  cout << "=====================================================" << endl;
+  for (auto elem_size : {kSize / 2, kSize, 2 * kSize}) {
+    BenchmarkSequenceAccess<std::vector<std::string>>("std::vector", kOps,
+                                                      elem_size);
+    BenchmarkSequenceAccess<autovector<std::string, kSize>>("autovector", kOps,
+                                                            elem_size);
+    cout << "-----------------------------------" << endl;
+  }
+}
+
+}  // namespace ROCKSDB_NAMESPACE
+
+int main(int argc, char** argv) {
+  ROCKSDB_NAMESPACE::port::InstallStackTraceHandler();
+  ::testing::InitGoogleTest(&argc, argv);
+  return RUN_ALL_TESTS();
+}
diff --git a/src/rocksdb/util/bloom_impl.h b/src/rocksdb/util/bloom_impl.h
new file mode 100644
index 000000000..fadd012d3
--- /dev/null
+++ b/src/rocksdb/util/bloom_impl.h
@@ -0,0 +1,489 @@
+//  Copyright (c) 2019-present, Facebook, Inc. All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+//
+// Implementation details of various Bloom filter implementations used in
+// RocksDB. (DynamicBloom is in a separate file for now because it
+// supports concurrent write.)
+
+#pragma once
+#include <stddef.h>
+#include <stdint.h>
+
+#include <cmath>
+
+#include "port/port.h"  // for PREFETCH
+#include "rocksdb/slice.h"
+#include "util/hash.h"
+
+#ifdef HAVE_AVX2
+#include <immintrin.h>
+#endif
+
+namespace ROCKSDB_NAMESPACE {
+
+class BloomMath {
+ public:
+  // False positive rate of a standard Bloom filter, for given ratio of
+  // filter memory bits to added keys, and number of probes per operation.
+  // (The false positive rate is effectively independent of scale, assuming
+  // the implementation scales OK.)
+  static double StandardFpRate(double bits_per_key, int num_probes) {
+    // Standard very-good-estimate formula. See
+    // https://en.wikipedia.org/wiki/Bloom_filter#Probability_of_false_positives
+    return std::pow(1.0 - std::exp(-num_probes / bits_per_key), num_probes);
+  }
+
+  // False positive rate of a "blocked"/"shareded"/"cache-local" Bloom filter,
+  // for given ratio of filter memory bits to added keys, number of probes per
+  // operation (all within the given block or cache line size), and block or
+  // cache line size.
+  static double CacheLocalFpRate(double bits_per_key, int num_probes,
+                                 int cache_line_bits) {
+    if (bits_per_key <= 0.0) {
+      // Fix a discontinuity
+      return 1.0;
+    }
+    double keys_per_cache_line = cache_line_bits / bits_per_key;
+    // A reasonable estimate is the average of the FP rates for one standard
+    // deviation above and below the mean bucket occupancy. See
+    // https://github.com/facebook/rocksdb/wiki/RocksDB-Bloom-Filter#the-math
+    double keys_stddev = std::sqrt(keys_per_cache_line);
+    double crowded_fp = StandardFpRate(
+        cache_line_bits / (keys_per_cache_line + keys_stddev), num_probes);
+    double uncrowded_fp = StandardFpRate(
+        cache_line_bits / (keys_per_cache_line - keys_stddev), num_probes);
+    return (crowded_fp + uncrowded_fp) / 2;
+  }
+
+  // False positive rate of querying a new item against `num_keys` items, all
+  // hashed to `fingerprint_bits` bits. (This assumes the fingerprint hashes
+  // themselves are stored losslessly. See Section 4 of
+  // http://www.ccs.neu.edu/home/pete/pub/bloom-filters-verification.pdf)
+  static double FingerprintFpRate(size_t num_keys, int fingerprint_bits) {
+    double inv_fingerprint_space = std::pow(0.5, fingerprint_bits);
+    // Base estimate assumes each key maps to a unique fingerprint.
+    // Could be > 1 in extreme cases.
+    double base_estimate = num_keys * inv_fingerprint_space;
+    // To account for potential overlap, we choose between two formulas
+    if (base_estimate > 0.0001) {
+      // A very good formula assuming we don't construct a floating point
+      // number extremely close to 1. Always produces a probability < 1.
+      return 1.0 - std::exp(-base_estimate);
+    } else {
+      // A very good formula when base_estimate is far below 1. (Subtract
+      // away the integral-approximated sum that some key has same hash as
+      // one coming before it in a list.)
+      return base_estimate - (base_estimate * base_estimate * 0.5);
+    }
+  }
+
+  // Returns the probably of either of two independent(-ish) events
+  // happening, given their probabilities. (This is useful for combining
+  // results from StandardFpRate or CacheLocalFpRate with FingerprintFpRate
+  // for a hash-efficient Bloom filter's FP rate. See Section 4 of
+  // http://www.ccs.neu.edu/home/pete/pub/bloom-filters-verification.pdf)
+  static double IndependentProbabilitySum(double rate1, double rate2) {
+    // Use formula that avoids floating point extremely close to 1 if
+    // rates are extremely small.
+    return rate1 + rate2 - (rate1 * rate2);
+  }
+};
+
+// A fast, flexible, and accurate cache-local Bloom implementation with
+// SIMD-optimized query performance (currently using AVX2 on Intel). Write
+// performance and non-SIMD read are very good, benefiting from FastRange32
+// used in place of % and single-cycle multiplication on recent processors.
+//
+// Most other SIMD Bloom implementations sacrifice flexibility and/or
+// accuracy by requiring num_probes to be a power of two and restricting
+// where each probe can occur in a cache line. This implementation sacrifices
+// SIMD-optimization for add (might still be possible, especially with AVX512)
+// in favor of allowing any num_probes, not crossing cache line boundary,
+// and accuracy close to theoretical best accuracy for a cache-local Bloom.
+// E.g. theoretical best for 10 bits/key, num_probes=6, and 512-bit bucket
+// (Intel cache line size) is 0.9535% FP rate. This implementation yields
+// about 0.957%. (Compare to LegacyLocalityBloomImpl<false> at 1.138%, or
+// about 0.951% for 1024-bit buckets, cache line size for some ARM CPUs.)
+//
+// This implementation can use a 32-bit hash (let h2 be h1 * 0x9e3779b9) or
+// a 64-bit hash (split into two uint32s). With many millions of keys, the
+// false positive rate associated with using a 32-bit hash can dominate the
+// false positive rate of the underlying filter. At 10 bits/key setting, the
+// inflection point is about 40 million keys, so 32-bit hash is a bad idea
+// with 10s of millions of keys or more.
+//
+// Despite accepting a 64-bit hash, this implementation uses 32-bit fastrange
+// to pick a cache line, which can be faster than 64-bit in some cases.
+// This only hurts accuracy as you get into 10s of GB for a single filter,
+// and accuracy abruptly breaks down at 256GB (2^32 cache lines). Switch to
+// 64-bit fastrange if you need filters so big. ;)
+//
+// Using only a 32-bit input hash within each cache line has negligible
+// impact for any reasonable cache line / bucket size, for arbitrary filter
+// size, and potentially saves intermediate data size in some cases vs.
+// tracking full 64 bits. (Even in an implementation using 64-bit arithmetic
+// to generate indices, I might do the same, as a single multiplication
+// suffices to generate a sufficiently mixed 64 bits from 32 bits.)
+//
+// This implementation is currently tied to Intel cache line size, 64 bytes ==
+// 512 bits. If there's sufficient demand for other cache line sizes, this is
+// a pretty good implementation to extend, but slight performance enhancements
+// are possible with an alternate implementation (probably not very compatible
+// with SIMD):
+// (1) Use rotation in addition to multiplication for remixing
+// (like murmur hash). (Using multiplication alone *slightly* hurts accuracy
+// because lower bits never depend on original upper bits.)
+// (2) Extract more than one bit index from each re-mix. (Only if rotation
+// or similar is part of remix, because otherwise you're making the
+// multiplication-only problem worse.)
+// (3) Re-mix full 64 bit hash, to get maximum number of bit indices per
+// re-mix.
+//
+class FastLocalBloomImpl {
+ public:
+  // NOTE: this has only been validated to enough accuracy for producing
+  // reasonable warnings / user feedback, not for making functional decisions.
+  static double EstimatedFpRate(size_t keys, size_t bytes, int num_probes,
+                                int hash_bits) {
+    return BloomMath::IndependentProbabilitySum(
+        BloomMath::CacheLocalFpRate(8.0 * bytes / keys, num_probes,
+                                    /*cache line bits*/ 512),
+        BloomMath::FingerprintFpRate(keys, hash_bits));
+  }
+
+  static inline int ChooseNumProbes(int millibits_per_key) {
+    // Since this implementation can (with AVX2) make up to 8 probes
+    // for the same cost, we pick the most accurate num_probes, based
+    // on actual tests of the implementation. Note that for higher
+    // bits/key, the best choice for cache-local Bloom can be notably
+    // smaller than standard bloom, e.g. 9 instead of 11 @ 16 b/k.
+    if (millibits_per_key <= 2080) {
+      return 1;
+    } else if (millibits_per_key <= 3580) {
+      return 2;
+    } else if (millibits_per_key <= 5100) {
+      return 3;
+    } else if (millibits_per_key <= 6640) {
+      return 4;
+    } else if (millibits_per_key <= 8300) {
+      return 5;
+    } else if (millibits_per_key <= 10070) {
+      return 6;
+    } else if (millibits_per_key <= 11720) {
+      return 7;
+    } else if (millibits_per_key <= 14001) {
+      // Would be something like <= 13800 but sacrificing *slightly* for
+      // more settings using <= 8 probes.
+      return 8;
+    } else if (millibits_per_key <= 16050) {
+      return 9;
+    } else if (millibits_per_key <= 18300) {
+      return 10;
+    } else if (millibits_per_key <= 22001) {
+      return 11;
+    } else if (millibits_per_key <= 25501) {
+      return 12;
+    } else if (millibits_per_key > 50000) {
+      // Top out at 24 probes (three sets of 8)
+      return 24;
+    } else {
+      // Roughly optimal choices for remaining range
+      // e.g.
+      // 28000 -> 12, 28001 -> 13
+      // 50000 -> 23, 50001 -> 24
+      return (millibits_per_key - 1) / 2000 - 1;
+    }
+  }
+
+  static inline void AddHash(uint32_t h1, uint32_t h2, uint32_t len_bytes,
+                             int num_probes, char *data) {
+    uint32_t bytes_to_cache_line = FastRange32(len_bytes >> 6, h1) << 6;
+    AddHashPrepared(h2, num_probes, data + bytes_to_cache_line);
+  }
+
+  static inline void AddHashPrepared(uint32_t h2, int num_probes,
+                                     char *data_at_cache_line) {
+    uint32_t h = h2;
+    for (int i = 0; i < num_probes; ++i, h *= uint32_t{0x9e3779b9}) {
+      // 9-bit address within 512 bit cache line
+      int bitpos = h >> (32 - 9);
+      data_at_cache_line[bitpos >> 3] |= (uint8_t{1} << (bitpos & 7));
+    }
+  }
+
+  static inline void PrepareHash(uint32_t h1, uint32_t len_bytes,
+                                 const char *data,
+                                 uint32_t /*out*/ *byte_offset) {
+    uint32_t bytes_to_cache_line = FastRange32(len_bytes >> 6, h1) << 6;
+    PREFETCH(data + bytes_to_cache_line, 0 /* rw */, 1 /* locality */);
+    PREFETCH(data + bytes_to_cache_line + 63, 0 /* rw */, 1 /* locality */);
+    *byte_offset = bytes_to_cache_line;
+  }
+
+  static inline bool HashMayMatch(uint32_t h1, uint32_t h2, uint32_t len_bytes,
+                                  int num_probes, const char *data) {
+    uint32_t bytes_to_cache_line = FastRange32(len_bytes >> 6, h1) << 6;
+    return HashMayMatchPrepared(h2, num_probes, data + bytes_to_cache_line);
+  }
+
+  static inline bool HashMayMatchPrepared(uint32_t h2, int num_probes,
+                                          const char *data_at_cache_line) {
+    uint32_t h = h2;
+#ifdef HAVE_AVX2
+    int rem_probes = num_probes;
+
+    // NOTE: For better performance for num_probes in {1, 2, 9, 10, 17, 18,
+    // etc.} one can insert specialized code for rem_probes <= 2, bypassing
+    // the SIMD code in those cases. There is a detectable but minor overhead
+    // applied to other values of num_probes (when not statically determined),
+    // but smoother performance curve vs. num_probes. But for now, when
+    // in doubt, don't add unnecessary code.
+
+    // Powers of 32-bit golden ratio, mod 2**32.
+    const __m256i multipliers =
+        _mm256_setr_epi32(0x00000001, 0x9e3779b9, 0xe35e67b1, 0x734297e9,
+                          0x35fbe861, 0xdeb7c719, 0x448b211, 0x3459b749);
+
+    for (;;) {
+      // Eight copies of hash
+      __m256i hash_vector = _mm256_set1_epi32(h);
+
+      // Same effect as repeated multiplication by 0x9e3779b9 thanks to
+      // associativity of multiplication.
+      hash_vector = _mm256_mullo_epi32(hash_vector, multipliers);
+
+      // Now the top 9 bits of each of the eight 32-bit values in
+      // hash_vector are bit addresses for probes within the cache line.
+      // While the platform-independent code uses byte addressing (6 bits
+      // to pick a byte + 3 bits to pick a bit within a byte), here we work
+      // with 32-bit words (4 bits to pick a word + 5 bits to pick a bit
+      // within a word) because that works well with AVX2 and is equivalent
+      // under little-endian.
+
+      // Shift each right by 28 bits to get 4-bit word addresses.
+      const __m256i word_addresses = _mm256_srli_epi32(hash_vector, 28);
+
+      // Gather 32-bit values spread over 512 bits by 4-bit address. In
+      // essence, we are dereferencing eight pointers within the cache
+      // line.
+      //
+      // Option 1: AVX2 gather (seems to be a little slow - understandable)
+      // const __m256i value_vector =
+      //     _mm256_i32gather_epi32(static_cast<const int
+      //     *>(data_at_cache_line),
+      //                            word_addresses,
+      //                            /*bytes / i32*/ 4);
+      // END Option 1
+      // Potentially unaligned as we're not *always* cache-aligned -> loadu
+      const __m256i *mm_data =
+          reinterpret_cast<const __m256i *>(data_at_cache_line);
+      __m256i lower = _mm256_loadu_si256(mm_data);
+      __m256i upper = _mm256_loadu_si256(mm_data + 1);
+      // Option 2: AVX512VL permute hack
+      // Only negligibly faster than Option 3, so not yet worth supporting
+      // const __m256i value_vector =
+      //    _mm256_permutex2var_epi32(lower, word_addresses, upper);
+      // END Option 2
+      // Option 3: AVX2 permute+blend hack
+      // Use lowest three bits to order probing values, as if all from same
+      // 256 bit piece.
+      lower = _mm256_permutevar8x32_epi32(lower, word_addresses);
+      upper = _mm256_permutevar8x32_epi32(upper, word_addresses);
+      // Just top 1 bit of address, to select between lower and upper.
+      const __m256i upper_lower_selector = _mm256_srai_epi32(hash_vector, 31);
+      // Finally: the next 8 probed 32-bit values, in probing sequence order.
+      const __m256i value_vector =
+          _mm256_blendv_epi8(lower, upper, upper_lower_selector);
+      // END Option 3
+
+      // We might not need to probe all 8, so build a mask for selecting only
+      // what we need. (The k_selector(s) could be pre-computed but that
+      // doesn't seem to make a noticeable performance difference.)
+      const __m256i zero_to_seven = _mm256_setr_epi32(0, 1, 2, 3, 4, 5, 6, 7);
+      // Subtract rem_probes from each of those constants
+      __m256i k_selector =
+          _mm256_sub_epi32(zero_to_seven, _mm256_set1_epi32(rem_probes));
+      // Negative after subtract -> use/select
+      // Keep only high bit (logical shift right each by 31).
+      k_selector = _mm256_srli_epi32(k_selector, 31);
+
+      // Strip off the 4 bit word address (shift left)
+      __m256i bit_addresses = _mm256_slli_epi32(hash_vector, 4);
+      // And keep only 5-bit (32 - 27) bit-within-32-bit-word addresses.
+      bit_addresses = _mm256_srli_epi32(bit_addresses, 27);
+      // Build a bit mask
+      const __m256i bit_mask = _mm256_sllv_epi32(k_selector, bit_addresses);
+
+      // Like ((~value_vector) & bit_mask) == 0)
+      bool match = _mm256_testc_si256(value_vector, bit_mask) != 0;
+
+      // This check first so that it's easy for branch predictor to optimize
+      // num_probes <= 8 case, making it free of unpredictable branches.
+      if (rem_probes <= 8) {
+        return match;
+      } else if (!match) {
+        return false;
+      }
+      // otherwise
+      // Need another iteration. 0xab25f4c1 == golden ratio to the 8th power
+      h *= 0xab25f4c1;
+      rem_probes -= 8;
+    }
+#else
+    for (int i = 0; i < num_probes; ++i, h *= uint32_t{0x9e3779b9}) {
+      // 9-bit address within 512 bit cache line
+      int bitpos = h >> (32 - 9);
+      if ((data_at_cache_line[bitpos >> 3] & (char(1) << (bitpos & 7))) == 0) {
+        return false;
+      }
+    }
+    return true;
+#endif
+  }
+};
+
+// A legacy Bloom filter implementation with no locality of probes (slow).
+// It uses double hashing to generate a sequence of hash values.
+// Asymptotic analysis is in [Kirsch,Mitzenmacher 2006], but known to have
+// subtle accuracy flaws for practical sizes [Dillinger,Manolios 2004].
+//
+// DO NOT REUSE
+//
+class LegacyNoLocalityBloomImpl {
+ public:
+  static inline int ChooseNumProbes(int bits_per_key) {
+    // We intentionally round down to reduce probing cost a little bit
+    int num_probes = static_cast<int>(bits_per_key * 0.69);  // 0.69 =~ ln(2)
+    if (num_probes < 1) num_probes = 1;
+    if (num_probes > 30) num_probes = 30;
+    return num_probes;
+  }
+
+  static inline void AddHash(uint32_t h, uint32_t total_bits, int num_probes,
+                             char *data) {
+    const uint32_t delta = (h >> 17) | (h << 15);  // Rotate right 17 bits
+    for (int i = 0; i < num_probes; i++) {
+      const uint32_t bitpos = h % total_bits;
+      data[bitpos / 8] |= (1 << (bitpos % 8));
+      h += delta;
+    }
+  }
+
+  static inline bool HashMayMatch(uint32_t h, uint32_t total_bits,
+                                  int num_probes, const char *data) {
+    const uint32_t delta = (h >> 17) | (h << 15);  // Rotate right 17 bits
+    for (int i = 0; i < num_probes; i++) {
+      const uint32_t bitpos = h % total_bits;
+      if ((data[bitpos / 8] & (1 << (bitpos % 8))) == 0) {
+        return false;
+      }
+      h += delta;
+    }
+    return true;
+  }
+};
+
+// A legacy Bloom filter implementation with probes local to a single
+// cache line (fast). Because SST files might be transported between
+// platforms, the cache line size is a parameter rather than hard coded.
+// (But if specified as a constant parameter, an optimizing compiler
+// should take advantage of that.)
+//
+// When ExtraRotates is false, this implementation is notably deficient in
+// accuracy. Specifically, it uses double hashing with a 1/512 chance of the
+// increment being zero (when cache line size is 512 bits). Thus, there's a
+// 1/512 chance of probing only one index, which we'd expect to incur about
+// a 1/2 * 1/512 or absolute 0.1% FP rate penalty. More detail at
+// https://github.com/facebook/rocksdb/issues/4120
+//
+// DO NOT REUSE
+//
+template <bool ExtraRotates>
+class LegacyLocalityBloomImpl {
+ private:
+  static inline uint32_t GetLine(uint32_t h, uint32_t num_lines) {
+    uint32_t offset_h = ExtraRotates ? (h >> 11) | (h << 21) : h;
+    return offset_h % num_lines;
+  }
+
+ public:
+  // NOTE: this has only been validated to enough accuracy for producing
+  // reasonable warnings / user feedback, not for making functional decisions.
+  static double EstimatedFpRate(size_t keys, size_t bytes, int num_probes) {
+    double bits_per_key = 8.0 * bytes / keys;
+    double filter_rate = BloomMath::CacheLocalFpRate(bits_per_key, num_probes,
+                                                     /*cache line bits*/ 512);
+    if (!ExtraRotates) {
+      // Good estimate of impact of flaw in index computation.
+      // Adds roughly 0.002 around 50 bits/key and 0.001 around 100 bits/key.
+      // The + 22 shifts it nicely to fit for lower bits/key.
+      filter_rate += 0.1 / (bits_per_key * 0.75 + 22);
+    } else {
+      // Not yet validated
+      assert(false);
+    }
+    // Always uses 32-bit hash
+    double fingerprint_rate = BloomMath::FingerprintFpRate(keys, 32);
+    return BloomMath::IndependentProbabilitySum(filter_rate, fingerprint_rate);
+  }
+
+  static inline void AddHash(uint32_t h, uint32_t num_lines, int num_probes,
+                             char *data, int log2_cache_line_bytes) {
+    const int log2_cache_line_bits = log2_cache_line_bytes + 3;
+
+    char *data_at_offset =
+        data + (GetLine(h, num_lines) << log2_cache_line_bytes);
+    const uint32_t delta = (h >> 17) | (h << 15);
+    for (int i = 0; i < num_probes; ++i) {
+      // Mask to bit-within-cache-line address
+      const uint32_t bitpos = h & ((1 << log2_cache_line_bits) - 1);
+      data_at_offset[bitpos / 8] |= (1 << (bitpos % 8));
+      if (ExtraRotates) {
+        h = (h >> log2_cache_line_bits) | (h << (32 - log2_cache_line_bits));
+      }
+      h += delta;
+    }
+  }
+
+  static inline void PrepareHashMayMatch(uint32_t h, uint32_t num_lines,
+                                         const char *data,
+                                         uint32_t /*out*/ *byte_offset,
+                                         int log2_cache_line_bytes) {
+    uint32_t b = GetLine(h, num_lines) << log2_cache_line_bytes;
+    PREFETCH(data + b, 0 /* rw */, 1 /* locality */);
+    PREFETCH(data + b + ((1 << log2_cache_line_bytes) - 1), 0 /* rw */,
+             1 /* locality */);
+    *byte_offset = b;
+  }
+
+  static inline bool HashMayMatch(uint32_t h, uint32_t num_lines,
+                                  int num_probes, const char *data,
+                                  int log2_cache_line_bytes) {
+    uint32_t b = GetLine(h, num_lines) << log2_cache_line_bytes;
+    return HashMayMatchPrepared(h, num_probes, data + b, log2_cache_line_bytes);
+  }
+
+  static inline bool HashMayMatchPrepared(uint32_t h, int num_probes,
+                                          const char *data_at_offset,
+                                          int log2_cache_line_bytes) {
+    const int log2_cache_line_bits = log2_cache_line_bytes + 3;
+
+    const uint32_t delta = (h >> 17) | (h << 15);
+    for (int i = 0; i < num_probes; ++i) {
+      // Mask to bit-within-cache-line address
+      const uint32_t bitpos = h & ((1 << log2_cache_line_bits) - 1);
+      if (((data_at_offset[bitpos / 8]) & (1 << (bitpos % 8))) == 0) {
+        return false;
+      }
+      if (ExtraRotates) {
+        h = (h >> log2_cache_line_bits) | (h << (32 - log2_cache_line_bits));
+      }
+      h += delta;
+    }
+    return true;
+  }
+};
+
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/util/bloom_test.cc b/src/rocksdb/util/bloom_test.cc
new file mode 100644
index 000000000..9d509ac3d
--- /dev/null
+++ b/src/rocksdb/util/bloom_test.cc
@@ -0,0 +1,1175 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+//
+// Copyright (c) 2012 The LevelDB Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file. See the AUTHORS file for names of contributors.
+
+#ifndef GFLAGS
+#include <cstdio>
+int main() {
+  fprintf(stderr, "Please install gflags to run this test... Skipping...\n");
+  return 0;
+}
+#else
+
+#include <array>
+#include <cmath>
+#include <vector>
+
+#include "cache/cache_entry_roles.h"
+#include "cache/cache_reservation_manager.h"
+#include "memory/arena.h"
+#include "port/jemalloc_helper.h"
+#include "rocksdb/filter_policy.h"
+#include "table/block_based/filter_policy_internal.h"
+#include "test_util/testharness.h"
+#include "test_util/testutil.h"
+#include "util/gflags_compat.h"
+#include "util/hash.h"
+
+using GFLAGS_NAMESPACE::ParseCommandLineFlags;
+
+// The test is not fully designed for bits_per_key other than 10, but with
+// this parameter you can easily explore the behavior of other bits_per_key.
+// See also filter_bench.
+DEFINE_int32(bits_per_key, 10, "");
+
+namespace ROCKSDB_NAMESPACE {
+
+namespace {
+const std::string kLegacyBloom = test::LegacyBloomFilterPolicy::kClassName();
+const std::string kFastLocalBloom =
+    test::FastLocalBloomFilterPolicy::kClassName();
+const std::string kStandard128Ribbon =
+    test::Standard128RibbonFilterPolicy::kClassName();
+}  // namespace
+
+static const int kVerbose = 1;
+
+static Slice Key(int i, char* buffer) {
+  std::string s;
+  PutFixed32(&s, static_cast<uint32_t>(i));
+  memcpy(buffer, s.c_str(), sizeof(i));
+  return Slice(buffer, sizeof(i));
+}
+
+static int NextLength(int length) {
+  if (length < 10) {
+    length += 1;
+  } else if (length < 100) {
+    length += 10;
+  } else if (length < 1000) {
+    length += 100;
+  } else {
+    length += 1000;
+  }
+  return length;
+}
+
+class FullBloomTest : public testing::TestWithParam<std::string> {
+ protected:
+  BlockBasedTableOptions table_options_;
+
+ private:
+  std::shared_ptr<const FilterPolicy>& policy_;
+  std::unique_ptr<FilterBitsBuilder> bits_builder_;
+  std::unique_ptr<FilterBitsReader> bits_reader_;
+  std::unique_ptr<const char[]> buf_;
+  size_t filter_size_;
+
+ public:
+  FullBloomTest() : policy_(table_options_.filter_policy), filter_size_(0) {
+    ResetPolicy();
+  }
+
+  BuiltinFilterBitsBuilder* GetBuiltinFilterBitsBuilder() {
+    // Throws on bad cast
+    return dynamic_cast<BuiltinFilterBitsBuilder*>(bits_builder_.get());
+  }
+
+  const BloomLikeFilterPolicy* GetBloomLikeFilterPolicy() {
+    // Throws on bad cast
+    return &dynamic_cast<const BloomLikeFilterPolicy&>(*policy_);
+  }
+
+  void Reset() {
+    bits_builder_.reset(BloomFilterPolicy::GetBuilderFromContext(
+        FilterBuildingContext(table_options_)));
+    bits_reader_.reset(nullptr);
+    buf_.reset(nullptr);
+    filter_size_ = 0;
+  }
+
+  void ResetPolicy(double bits_per_key) {
+    policy_ = BloomLikeFilterPolicy::Create(GetParam(), bits_per_key);
+    Reset();
+  }
+
+  void ResetPolicy() { ResetPolicy(FLAGS_bits_per_key); }
+
+  void Add(const Slice& s) { bits_builder_->AddKey(s); }
+
+  void OpenRaw(const Slice& s) {
+    bits_reader_.reset(policy_->GetFilterBitsReader(s));
+  }
+
+  void Build() {
+    Slice filter = bits_builder_->Finish(&buf_);
+    bits_reader_.reset(policy_->GetFilterBitsReader(filter));
+    filter_size_ = filter.size();
+  }
+
+  size_t FilterSize() const { return filter_size_; }
+
+  Slice FilterData() { return Slice(buf_.get(), filter_size_); }
+
+  int GetNumProbesFromFilterData() {
+    assert(filter_size_ >= 5);
+    int8_t raw_num_probes = static_cast<int8_t>(buf_.get()[filter_size_ - 5]);
+    if (raw_num_probes == -1) {  // New bloom filter marker
+      return static_cast<uint8_t>(buf_.get()[filter_size_ - 3]);
+    } else {
+      return raw_num_probes;
+    }
+  }
+
+  int GetRibbonSeedFromFilterData() {
+    assert(filter_size_ >= 5);
+    // Check for ribbon marker
+    assert(-2 == static_cast<int8_t>(buf_.get()[filter_size_ - 5]));
+    return static_cast<uint8_t>(buf_.get()[filter_size_ - 4]);
+  }
+
+  bool Matches(const Slice& s) {
+    if (bits_reader_ == nullptr) {
+      Build();
+    }
+    return bits_reader_->MayMatch(s);
+  }
+
+  // Provides a kind of fingerprint on the Bloom filter's
+  // behavior, for reasonbly high FP rates.
+  uint64_t PackedMatches() {
+    char buffer[sizeof(int)];
+    uint64_t result = 0;
+    for (int i = 0; i < 64; i++) {
+      if (Matches(Key(i + 12345, buffer))) {
+        result |= uint64_t{1} << i;
+      }
+    }
+    return result;
+  }
+
+  // Provides a kind of fingerprint on the Bloom filter's
+  // behavior, for lower FP rates.
+  std::string FirstFPs(int count) {
+    char buffer[sizeof(int)];
+    std::string rv;
+    int fp_count = 0;
+    for (int i = 0; i < 1000000; i++) {
+      // Pack four match booleans into each hexadecimal digit
+      if (Matches(Key(i + 1000000, buffer))) {
+        ++fp_count;
+        rv += std::to_string(i);
+        if (fp_count == count) {
+          break;
+        }
+        rv += ',';
+      }
+    }
+    return rv;
+  }
+
+  double FalsePositiveRate() {
+    char buffer[sizeof(int)];
+    int result = 0;
+    for (int i = 0; i < 10000; i++) {
+      if (Matches(Key(i + 1000000000, buffer))) {
+        result++;
+      }
+    }
+    return result / 10000.0;
+  }
+};
+
+TEST_P(FullBloomTest, FilterSize) {
+  // In addition to checking the consistency of space computation, we are
+  // checking that denoted and computed doubles are interpreted as expected
+  // as bits_per_key values.
+  bool some_computed_less_than_denoted = false;
+  // Note: to avoid unproductive configurations, bits_per_key < 0.5 is rounded
+  // down to 0 (no filter), and 0.5 <= bits_per_key < 1.0 is rounded up to 1
+  // bit per key (1000 millibits). Also, enforced maximum is 100 bits per key
+  // (100000 millibits).
+  for (auto bpk : std::vector<std::pair<double, int> >{{-HUGE_VAL, 0},
+                                                       {-INFINITY, 0},
+                                                       {0.0, 0},
+                                                       {0.499, 0},
+                                                       {0.5, 1000},
+                                                       {1.234, 1234},
+                                                       {3.456, 3456},
+                                                       {9.5, 9500},
+                                                       {10.0, 10000},
+                                                       {10.499, 10499},
+                                                       {21.345, 21345},
+                                                       {99.999, 99999},
+                                                       {1234.0, 100000},
+                                                       {HUGE_VAL, 100000},
+                                                       {INFINITY, 100000},
+                                                       {NAN, 100000}}) {
+    ResetPolicy(bpk.first);
+    auto bfp = GetBloomLikeFilterPolicy();
+    EXPECT_EQ(bpk.second, bfp->GetMillibitsPerKey());
+    EXPECT_EQ((bpk.second + 500) / 1000, bfp->GetWholeBitsPerKey());
+
+    double computed = bpk.first;
+    // This transforms e.g. 9.5 -> 9.499999999999998, which we still
+    // round to 10 for whole bits per key.
+    computed += 0.5;
+    computed /= 1234567.0;
+    computed *= 1234567.0;
+    computed -= 0.5;
+    some_computed_less_than_denoted |= (computed < bpk.first);
+    ResetPolicy(computed);
+    bfp = GetBloomLikeFilterPolicy();
+    EXPECT_EQ(bpk.second, bfp->GetMillibitsPerKey());
+    EXPECT_EQ((bpk.second + 500) / 1000, bfp->GetWholeBitsPerKey());
+
+    auto bits_builder = GetBuiltinFilterBitsBuilder();
+    if (bpk.second == 0) {
+      ASSERT_EQ(bits_builder, nullptr);
+      continue;
+    }
+
+    size_t n = 1;
+    size_t space = 0;
+    for (; n < 1000000; n += 1 + n / 1000) {
+      // Ensure consistency between CalculateSpace and ApproximateNumEntries
+      space = bits_builder->CalculateSpace(n);
+      size_t n2 = bits_builder->ApproximateNumEntries(space);
+      EXPECT_GE(n2, n);
+      size_t space2 = bits_builder->CalculateSpace(n2);
+      if (n > 12000 && GetParam() == kStandard128Ribbon) {
+        // TODO(peterd): better approximation?
+        EXPECT_GE(space2, space);
+        EXPECT_LE(space2 * 0.998, space * 1.0);
+      } else {
+        EXPECT_EQ(space2, space);
+      }
+    }
+    // Until size_t overflow
+    for (; n < (n + n / 3); n += n / 3) {
+      // Ensure space computation is not overflowing; capped is OK
+      size_t space2 = bits_builder->CalculateSpace(n);
+      EXPECT_GE(space2, space);
+      space = space2;
+    }
+  }
+  // Check that the compiler hasn't optimized our computation into nothing
+  EXPECT_TRUE(some_computed_less_than_denoted);
+  ResetPolicy();
+}
+
+TEST_P(FullBloomTest, FullEmptyFilter) {
+  // Empty filter is not match, at this level
+  ASSERT_TRUE(!Matches("hello"));
+  ASSERT_TRUE(!Matches("world"));
+}
+
+TEST_P(FullBloomTest, FullSmall) {
+  Add("hello");
+  Add("world");
+  ASSERT_TRUE(Matches("hello"));
+  ASSERT_TRUE(Matches("world"));
+  ASSERT_TRUE(!Matches("x"));
+  ASSERT_TRUE(!Matches("foo"));
+}
+
+TEST_P(FullBloomTest, FullVaryingLengths) {
+  char buffer[sizeof(int)];
+
+  // Count number of filters that significantly exceed the false positive rate
+  int mediocre_filters = 0;
+  int good_filters = 0;
+
+  for (int length = 1; length <= 10000; length = NextLength(length)) {
+    Reset();
+    for (int i = 0; i < length; i++) {
+      Add(Key(i, buffer));
+    }
+    Build();
+
+    EXPECT_LE(FilterSize(), (size_t)((length * FLAGS_bits_per_key / 8) +
+                                     CACHE_LINE_SIZE * 2 + 5));
+
+    // All added keys must match
+    for (int i = 0; i < length; i++) {
+      ASSERT_TRUE(Matches(Key(i, buffer)))
+          << "Length " << length << "; key " << i;
+    }
+
+    // Check false positive rate
+    double rate = FalsePositiveRate();
+    if (kVerbose >= 1) {
+      fprintf(stderr, "False positives: %5.2f%% @ length = %6d ; bytes = %6d\n",
+              rate * 100.0, length, static_cast<int>(FilterSize()));
+    }
+    if (FLAGS_bits_per_key == 10) {
+      EXPECT_LE(rate, 0.02);  // Must not be over 2%
+      if (rate > 0.0125) {
+        mediocre_filters++;  // Allowed, but not too often
+      } else {
+        good_filters++;
+      }
+    }
+  }
+  if (kVerbose >= 1) {
+    fprintf(stderr, "Filters: %d good, %d mediocre\n", good_filters,
+            mediocre_filters);
+  }
+  EXPECT_LE(mediocre_filters, good_filters / 5);
+}
+
+TEST_P(FullBloomTest, OptimizeForMemory) {
+  char buffer[sizeof(int)];
+  for (bool offm : {true, false}) {
+    table_options_.optimize_filters_for_memory = offm;
+    ResetPolicy();
+    Random32 rnd(12345);
+    uint64_t total_size = 0;
+    uint64_t total_mem = 0;
+    int64_t total_keys = 0;
+    double total_fp_rate = 0;
+    constexpr int nfilters = 100;
+    for (int i = 0; i < nfilters; ++i) {
+      int nkeys = static_cast<int>(rnd.Uniformish(10000)) + 100;
+      Reset();
+      for (int j = 0; j < nkeys; ++j) {
+        Add(Key(j, buffer));
+      }
+      Build();
+      size_t size = FilterData().size();
+      total_size += size;
+      // optimize_filters_for_memory currently depends on malloc_usable_size
+      // but we run the rest of the test to ensure no bad behavior without it.
+#ifdef ROCKSDB_MALLOC_USABLE_SIZE
+      size = malloc_usable_size(const_cast<char*>(FilterData().data()));
+#endif  // ROCKSDB_MALLOC_USABLE_SIZE
+      total_mem += size;
+      total_keys += nkeys;
+      total_fp_rate += FalsePositiveRate();
+    }
+    if (FLAGS_bits_per_key == 10) {
+      EXPECT_LE(total_fp_rate / double{nfilters}, 0.011);
+      EXPECT_GE(total_fp_rate / double{nfilters},
+                CACHE_LINE_SIZE >= 256 ? 0.007 : 0.008);
+    }
+
+    int64_t ex_min_total_size = int64_t{FLAGS_bits_per_key} * total_keys / 8;
+    if (GetParam() == kStandard128Ribbon) {
+      // ~ 30% savings vs. Bloom filter
+      ex_min_total_size = 7 * ex_min_total_size / 10;
+    }
+    EXPECT_GE(static_cast<int64_t>(total_size), ex_min_total_size);
+
+    int64_t blocked_bloom_overhead = nfilters * (CACHE_LINE_SIZE + 5);
+    if (GetParam() == kLegacyBloom) {
+      // this config can add extra cache line to make odd number
+      blocked_bloom_overhead += nfilters * CACHE_LINE_SIZE;
+    }
+
+    EXPECT_GE(total_mem, total_size);
+
+    // optimize_filters_for_memory not implemented with legacy Bloom
+    if (offm && GetParam() != kLegacyBloom) {
+      // This value can include a small extra penalty for kExtraPadding
+      fprintf(stderr, "Internal fragmentation (optimized): %g%%\n",
+              (total_mem - total_size) * 100.0 / total_size);
+      // Less than 1% internal fragmentation
+      EXPECT_LE(total_mem, total_size * 101 / 100);
+      // Up to 2% storage penalty
+      EXPECT_LE(static_cast<int64_t>(total_size),
+                ex_min_total_size * 102 / 100 + blocked_bloom_overhead);
+    } else {
+      fprintf(stderr, "Internal fragmentation (not optimized): %g%%\n",
+              (total_mem - total_size) * 100.0 / total_size);
+      // TODO: add control checks for more allocators?
+#ifdef ROCKSDB_JEMALLOC
+      fprintf(stderr, "Jemalloc detected? %d\n", HasJemalloc());
+      if (HasJemalloc()) {
+#ifdef ROCKSDB_MALLOC_USABLE_SIZE
+        // More than 5% internal fragmentation
+        EXPECT_GE(total_mem, total_size * 105 / 100);
+#endif  // ROCKSDB_MALLOC_USABLE_SIZE
+      }
+#endif  // ROCKSDB_JEMALLOC
+      // No storage penalty, just usual overhead
+      EXPECT_LE(static_cast<int64_t>(total_size),
+                ex_min_total_size + blocked_bloom_overhead);
+    }
+  }
+}
+
+class ChargeFilterConstructionTest : public testing::Test {};
+TEST_F(ChargeFilterConstructionTest, RibbonFilterFallBackOnLargeBanding) {
+  constexpr std::size_t kCacheCapacity =
+      8 * CacheReservationManagerImpl<
+              CacheEntryRole::kFilterConstruction>::GetDummyEntrySize();
+  constexpr std::size_t num_entries_for_cache_full = kCacheCapacity / 8;
+
+  for (CacheEntryRoleOptions::Decision charge_filter_construction_mem :
+       {CacheEntryRoleOptions::Decision::kEnabled,
+        CacheEntryRoleOptions::Decision::kDisabled}) {
+    bool will_fall_back = charge_filter_construction_mem ==
+                          CacheEntryRoleOptions::Decision::kEnabled;
+
+    BlockBasedTableOptions table_options;
+    table_options.cache_usage_options.options_overrides.insert(
+        {CacheEntryRole::kFilterConstruction,
+         {/*.charged = */ charge_filter_construction_mem}});
+    LRUCacheOptions lo;
+    lo.capacity = kCacheCapacity;
+    lo.num_shard_bits = 0;  // 2^0 shard
+    lo.strict_capacity_limit = true;
+    std::shared_ptr<Cache> cache(NewLRUCache(lo));
+    table_options.block_cache = cache;
+    table_options.filter_policy =
+        BloomLikeFilterPolicy::Create(kStandard128Ribbon, FLAGS_bits_per_key);
+    FilterBuildingContext ctx(table_options);
+    std::unique_ptr<FilterBitsBuilder> filter_bits_builder(
+        table_options.filter_policy->GetBuilderWithContext(ctx));
+
+    char key_buffer[sizeof(int)];
+    for (std::size_t i = 0; i < num_entries_for_cache_full; ++i) {
+      filter_bits_builder->AddKey(Key(static_cast<int>(i), key_buffer));
+    }
+
+    std::unique_ptr<const char[]> buf;
+    Slice filter = filter_bits_builder->Finish(&buf);
+
+    // To verify Ribbon Filter fallbacks to Bloom Filter properly
+    // based on cache charging result
+    // See BloomFilterPolicy::GetBloomBitsReader re: metadata
+    // -1 = Marker for newer Bloom implementations
+    // -2 = Marker for Standard128 Ribbon
+    if (will_fall_back) {
+      EXPECT_EQ(filter.data()[filter.size() - 5], static_cast<char>(-1));
+    } else {
+      EXPECT_EQ(filter.data()[filter.size() - 5], static_cast<char>(-2));
+    }
+
+    if (charge_filter_construction_mem ==
+        CacheEntryRoleOptions::Decision::kEnabled) {
+      const size_t dummy_entry_num = static_cast<std::size_t>(std::ceil(
+          filter.size() * 1.0 /
+          CacheReservationManagerImpl<
+              CacheEntryRole::kFilterConstruction>::GetDummyEntrySize()));
+      EXPECT_GE(
+          cache->GetPinnedUsage(),
+          dummy_entry_num *
+              CacheReservationManagerImpl<
+                  CacheEntryRole::kFilterConstruction>::GetDummyEntrySize());
+      EXPECT_LT(
+          cache->GetPinnedUsage(),
+          (dummy_entry_num + 1) *
+              CacheReservationManagerImpl<
+                  CacheEntryRole::kFilterConstruction>::GetDummyEntrySize());
+    } else {
+      EXPECT_EQ(cache->GetPinnedUsage(), 0);
+    }
+  }
+}
+
+namespace {
+inline uint32_t SelectByCacheLineSize(uint32_t for64, uint32_t for128,
+                                      uint32_t for256) {
+  (void)for64;
+  (void)for128;
+  (void)for256;
+#if CACHE_LINE_SIZE == 64
+  return for64;
+#elif CACHE_LINE_SIZE == 128
+  return for128;
+#elif CACHE_LINE_SIZE == 256
+  return for256;
+#else
+#error "CACHE_LINE_SIZE unknown or unrecognized"
+#endif
+}
+}  // namespace
+
+// Ensure the implementation doesn't accidentally change in an
+// incompatible way. This test doesn't check the reading side
+// (FirstFPs/PackedMatches) for LegacyBloom because it requires the
+// ability to read filters generated using other cache line sizes.
+// See RawSchema.
+TEST_P(FullBloomTest, Schema) {
+#define EXPECT_EQ_Bloom(a, b)               \
+  {                                         \
+    if (GetParam() != kStandard128Ribbon) { \
+      EXPECT_EQ(a, b);                      \
+    }                                       \
+  }
+#define EXPECT_EQ_Ribbon(a, b)              \
+  {                                         \
+    if (GetParam() == kStandard128Ribbon) { \
+      EXPECT_EQ(a, b);                      \
+    }                                       \
+  }
+#define EXPECT_EQ_FastBloom(a, b)        \
+  {                                      \
+    if (GetParam() == kFastLocalBloom) { \
+      EXPECT_EQ(a, b);                   \
+    }                                    \
+  }
+#define EXPECT_EQ_LegacyBloom(a, b)   \
+  {                                   \
+    if (GetParam() == kLegacyBloom) { \
+      EXPECT_EQ(a, b);                \
+    }                                 \
+  }
+#define EXPECT_EQ_NotLegacy(a, b)     \
+  {                                   \
+    if (GetParam() != kLegacyBloom) { \
+      EXPECT_EQ(a, b);                \
+    }                                 \
+  }
+
+  char buffer[sizeof(int)];
+
+  // First do a small number of keys, where Ribbon config will fall back on
+  // fast Bloom filter and generate the same data
+  ResetPolicy(5);  // num_probes = 3
+  for (int key = 0; key < 87; key++) {
+    Add(Key(key, buffer));
+  }
+  Build();
+  EXPECT_EQ(GetNumProbesFromFilterData(), 3);
+
+  EXPECT_EQ_NotLegacy(BloomHash(FilterData()), 4130687756U);
+
+  EXPECT_EQ_NotLegacy("31,38,40,43,61,83,86,112,125,131", FirstFPs(10));
+
+  // Now use enough keys so that changing bits / key by 1 is guaranteed to
+  // change number of allocated cache lines. So keys > max cache line bits.
+
+  // Note that the first attempted Ribbon seed is determined by the hash
+  // of the first key added (for pseudorandomness in practice, determinism in
+  // testing)
+
+  ResetPolicy(2);  // num_probes = 1
+  for (int key = 0; key < 2087; key++) {
+    Add(Key(key, buffer));
+  }
+  Build();
+  EXPECT_EQ_Bloom(GetNumProbesFromFilterData(), 1);
+  EXPECT_EQ_Ribbon(GetRibbonSeedFromFilterData(), 61);
+
+  EXPECT_EQ_LegacyBloom(
+      BloomHash(FilterData()),
+      SelectByCacheLineSize(1567096579, 1964771444, 2659542661U));
+  EXPECT_EQ_FastBloom(BloomHash(FilterData()), 3817481309U);
+  EXPECT_EQ_Ribbon(BloomHash(FilterData()), 1705851228U);
+
+  EXPECT_EQ_FastBloom("11,13,17,25,29,30,35,37,45,53", FirstFPs(10));
+  EXPECT_EQ_Ribbon("3,8,10,17,19,20,23,28,31,32", FirstFPs(10));
+
+  ResetPolicy(3);  // num_probes = 2
+  for (int key = 0; key < 2087; key++) {
+    Add(Key(key, buffer));
+  }
+  Build();
+  EXPECT_EQ_Bloom(GetNumProbesFromFilterData(), 2);
+  EXPECT_EQ_Ribbon(GetRibbonSeedFromFilterData(), 61);
+
+  EXPECT_EQ_LegacyBloom(
+      BloomHash(FilterData()),
+      SelectByCacheLineSize(2707206547U, 2571983456U, 218344685));
+  EXPECT_EQ_FastBloom(BloomHash(FilterData()), 2807269961U);
+  EXPECT_EQ_Ribbon(BloomHash(FilterData()), 1095342358U);
+
+  EXPECT_EQ_FastBloom("4,15,17,24,27,28,29,53,63,70", FirstFPs(10));
+  EXPECT_EQ_Ribbon("3,17,20,28,32,33,36,43,49,54", FirstFPs(10));
+
+  ResetPolicy(5);  // num_probes = 3
+  for (int key = 0; key < 2087; key++) {
+    Add(Key(key, buffer));
+  }
+  Build();
+  EXPECT_EQ_Bloom(GetNumProbesFromFilterData(), 3);
+  EXPECT_EQ_Ribbon(GetRibbonSeedFromFilterData(), 61);
+
+  EXPECT_EQ_LegacyBloom(
+      BloomHash(FilterData()),
+      SelectByCacheLineSize(515748486, 94611728, 2436112214U));
+  EXPECT_EQ_FastBloom(BloomHash(FilterData()), 204628445U);
+  EXPECT_EQ_Ribbon(BloomHash(FilterData()), 3971337699U);
+
+  EXPECT_EQ_FastBloom("15,24,29,39,53,87,89,100,103,104", FirstFPs(10));
+  EXPECT_EQ_Ribbon("3,33,36,43,67,70,76,78,84,102", FirstFPs(10));
+
+  ResetPolicy(8);  // num_probes = 5
+  for (int key = 0; key < 2087; key++) {
+    Add(Key(key, buffer));
+  }
+  Build();
+  EXPECT_EQ_Bloom(GetNumProbesFromFilterData(), 5);
+  EXPECT_EQ_Ribbon(GetRibbonSeedFromFilterData(), 61);
+
+  EXPECT_EQ_LegacyBloom(
+      BloomHash(FilterData()),
+      SelectByCacheLineSize(1302145999, 2811644657U, 756553699));
+  EXPECT_EQ_FastBloom(BloomHash(FilterData()), 355564975U);
+  EXPECT_EQ_Ribbon(BloomHash(FilterData()), 3651449053U);
+
+  EXPECT_EQ_FastBloom("16,60,66,126,220,238,244,256,265,287", FirstFPs(10));
+  EXPECT_EQ_Ribbon("33,187,203,296,300,322,411,419,547,582", FirstFPs(10));
+
+  ResetPolicy(9);  // num_probes = 6
+  for (int key = 0; key < 2087; key++) {
+    Add(Key(key, buffer));
+  }
+  Build();
+  EXPECT_EQ_Bloom(GetNumProbesFromFilterData(), 6);
+  EXPECT_EQ_Ribbon(GetRibbonSeedFromFilterData(), 61);
+
+  EXPECT_EQ_LegacyBloom(
+      BloomHash(FilterData()),
+      SelectByCacheLineSize(2092755149, 661139132, 1182970461));
+  EXPECT_EQ_FastBloom(BloomHash(FilterData()), 2137566013U);
+  EXPECT_EQ_Ribbon(BloomHash(FilterData()), 1005676675U);
+
+  EXPECT_EQ_FastBloom("156,367,791,872,945,1015,1139,1159,1265", FirstFPs(9));
+  EXPECT_EQ_Ribbon("33,187,203,296,411,419,604,612,615,619", FirstFPs(10));
+
+  ResetPolicy(11);  // num_probes = 7
+  for (int key = 0; key < 2087; key++) {
+    Add(Key(key, buffer));
+  }
+  Build();
+  EXPECT_EQ_Bloom(GetNumProbesFromFilterData(), 7);
+  EXPECT_EQ_Ribbon(GetRibbonSeedFromFilterData(), 61);
+
+  EXPECT_EQ_LegacyBloom(
+      BloomHash(FilterData()),
+      SelectByCacheLineSize(3755609649U, 1812694762, 1449142939));
+  EXPECT_EQ_FastBloom(BloomHash(FilterData()), 2561502687U);
+  EXPECT_EQ_Ribbon(BloomHash(FilterData()), 3129900846U);
+
+  EXPECT_EQ_FastBloom("34,74,130,236,643,882,962,1015,1035,1110", FirstFPs(10));
+  EXPECT_EQ_Ribbon("411,419,623,665,727,794,955,1052,1323,1330", FirstFPs(10));
+
+  // This used to be 9 probes, but 8 is a better choice for speed,
+  // especially with SIMD groups of 8 probes, with essentially no
+  // change in FP rate.
+  // FP rate @ 9 probes, old Bloom: 0.4321%
+  // FP rate @ 9 probes, new Bloom: 0.1846%
+  // FP rate @ 8 probes, new Bloom: 0.1843%
+  ResetPolicy(14);  // num_probes = 8 (new), 9 (old)
+  for (int key = 0; key < 2087; key++) {
+    Add(Key(key, buffer));
+  }
+  Build();
+  EXPECT_EQ_LegacyBloom(GetNumProbesFromFilterData(), 9);
+  EXPECT_EQ_FastBloom(GetNumProbesFromFilterData(), 8);
+  EXPECT_EQ_Ribbon(GetRibbonSeedFromFilterData(), 61);
+
+  EXPECT_EQ_LegacyBloom(
+      BloomHash(FilterData()),
+      SelectByCacheLineSize(178861123, 379087593, 2574136516U));
+  EXPECT_EQ_FastBloom(BloomHash(FilterData()), 3709876890U);
+  EXPECT_EQ_Ribbon(BloomHash(FilterData()), 1855638875U);
+
+  EXPECT_EQ_FastBloom("130,240,522,565,989,2002,2526,3147,3543", FirstFPs(9));
+  EXPECT_EQ_Ribbon("665,727,1323,1755,3866,4232,4442,4492,4736", FirstFPs(9));
+
+  // This used to be 11 probes, but 9 is a better choice for speed
+  // AND accuracy.
+  // FP rate @ 11 probes, old Bloom: 0.3571%
+  // FP rate @ 11 probes, new Bloom: 0.0884%
+  // FP rate @  9 probes, new Bloom: 0.0843%
+  ResetPolicy(16);  // num_probes = 9 (new), 11 (old)
+  for (int key = 0; key < 2087; key++) {
+    Add(Key(key, buffer));
+  }
+  Build();
+  EXPECT_EQ_LegacyBloom(GetNumProbesFromFilterData(), 11);
+  EXPECT_EQ_FastBloom(GetNumProbesFromFilterData(), 9);
+  EXPECT_EQ_Ribbon(GetRibbonSeedFromFilterData(), 61);
+
+  EXPECT_EQ_LegacyBloom(
+      BloomHash(FilterData()),
+      SelectByCacheLineSize(1129406313, 3049154394U, 1727750964));
+  EXPECT_EQ_FastBloom(BloomHash(FilterData()), 1087138490U);
+  EXPECT_EQ_Ribbon(BloomHash(FilterData()), 459379967U);
+
+  EXPECT_EQ_FastBloom("3299,3611,3916,6620,7822,8079,8482,8942", FirstFPs(8));
+  EXPECT_EQ_Ribbon("727,1323,1755,4442,4736,5386,6974,7154,8222", FirstFPs(9));
+
+  ResetPolicy(10);  // num_probes = 6, but different memory ratio vs. 9
+  for (int key = 0; key < 2087; key++) {
+    Add(Key(key, buffer));
+  }
+  Build();
+  EXPECT_EQ_Bloom(GetNumProbesFromFilterData(), 6);
+  EXPECT_EQ_Ribbon(GetRibbonSeedFromFilterData(), 61);
+
+  EXPECT_EQ_LegacyBloom(
+      BloomHash(FilterData()),
+      SelectByCacheLineSize(1478976371, 2910591341U, 1182970461));
+  EXPECT_EQ_FastBloom(BloomHash(FilterData()), 2498541272U);
+  EXPECT_EQ_Ribbon(BloomHash(FilterData()), 1273231667U);
+
+  EXPECT_EQ_FastBloom("16,126,133,422,466,472,813,1002,1035", FirstFPs(9));
+  EXPECT_EQ_Ribbon("296,411,419,612,619,623,630,665,686,727", FirstFPs(10));
+
+  ResetPolicy(10);
+  for (int key = /*CHANGED*/ 1; key < 2087; key++) {
+    Add(Key(key, buffer));
+  }
+  Build();
+  EXPECT_EQ_Bloom(GetNumProbesFromFilterData(), 6);
+  EXPECT_EQ_Ribbon(GetRibbonSeedFromFilterData(), /*CHANGED*/ 184);
+
+  EXPECT_EQ_LegacyBloom(
+      BloomHash(FilterData()),
+      SelectByCacheLineSize(4205696321U, 1132081253U, 2385981855U));
+  EXPECT_EQ_FastBloom(BloomHash(FilterData()), 2058382345U);
+  EXPECT_EQ_Ribbon(BloomHash(FilterData()), 3007790572U);
+
+  EXPECT_EQ_FastBloom("16,126,133,422,466,472,813,1002,1035", FirstFPs(9));
+  EXPECT_EQ_Ribbon("33,152,383,497,589,633,737,781,911,990", FirstFPs(10));
+
+  ResetPolicy(10);
+  for (int key = 1; key < /*CHANGED*/ 2088; key++) {
+    Add(Key(key, buffer));
+  }
+  Build();
+  EXPECT_EQ_Bloom(GetNumProbesFromFilterData(), 6);
+  EXPECT_EQ_Ribbon(GetRibbonSeedFromFilterData(), 184);
+
+  EXPECT_EQ_LegacyBloom(
+      BloomHash(FilterData()),
+      SelectByCacheLineSize(2885052954U, 769447944, 4175124908U));
+  EXPECT_EQ_FastBloom(BloomHash(FilterData()), 23699164U);
+  EXPECT_EQ_Ribbon(BloomHash(FilterData()), 1942323379U);
+
+  EXPECT_EQ_FastBloom("16,126,133,422,466,472,813,1002,1035", FirstFPs(9));
+  EXPECT_EQ_Ribbon("33,95,360,589,737,911,990,1048,1081,1414", FirstFPs(10));
+
+  // With new fractional bits_per_key, check that we are rounding to
+  // whole bits per key for old Bloom filters but fractional for
+  // new Bloom filter.
+  ResetPolicy(9.5);
+  for (int key = 1; key < 2088; key++) {
+    Add(Key(key, buffer));
+  }
+  Build();
+  EXPECT_EQ_Bloom(GetNumProbesFromFilterData(), 6);
+  EXPECT_EQ_Ribbon(GetRibbonSeedFromFilterData(), 184);
+
+  EXPECT_EQ_LegacyBloom(
+      BloomHash(FilterData()),
+      /*SAME*/ SelectByCacheLineSize(2885052954U, 769447944, 4175124908U));
+  EXPECT_EQ_FastBloom(BloomHash(FilterData()), 3166884174U);
+  EXPECT_EQ_Ribbon(BloomHash(FilterData()), 1148258663U);
+
+  EXPECT_EQ_FastBloom("126,156,367,444,458,791,813,976,1015", FirstFPs(9));
+  EXPECT_EQ_Ribbon("33,54,95,360,589,693,737,911,990,1048", FirstFPs(10));
+
+  ResetPolicy(10.499);
+  for (int key = 1; key < 2088; key++) {
+    Add(Key(key, buffer));
+  }
+  Build();
+  EXPECT_EQ_LegacyBloom(GetNumProbesFromFilterData(), 6);
+  EXPECT_EQ_FastBloom(GetNumProbesFromFilterData(), 7);
+  EXPECT_EQ_Ribbon(GetRibbonSeedFromFilterData(), 184);
+
+  EXPECT_EQ_LegacyBloom(
+      BloomHash(FilterData()),
+      /*SAME*/ SelectByCacheLineSize(2885052954U, 769447944, 4175124908U));
+  EXPECT_EQ_FastBloom(BloomHash(FilterData()), 4098502778U);
+  EXPECT_EQ_Ribbon(BloomHash(FilterData()), 792138188U);
+
+  EXPECT_EQ_FastBloom("16,236,240,472,1015,1045,1111,1409,1465", FirstFPs(9));
+  EXPECT_EQ_Ribbon("33,95,360,589,737,990,1048,1081,1414,1643", FirstFPs(10));
+
+  ResetPolicy();
+}
+
+// A helper class for testing custom or corrupt filter bits as read by
+// built-in FilterBitsReaders.
+struct RawFilterTester {
+  // Buffer, from which we always return a tail Slice, so the
+  // last five bytes are always the metadata bytes.
+  std::array<char, 3000> data_;
+  // Points five bytes from the end
+  char* metadata_ptr_;
+
+  RawFilterTester() : metadata_ptr_(&*(data_.end() - 5)) {}
+
+  Slice ResetNoFill(uint32_t len_without_metadata, uint32_t num_lines,
+                    uint32_t num_probes) {
+    metadata_ptr_[0] = static_cast<char>(num_probes);
+    EncodeFixed32(metadata_ptr_ + 1, num_lines);
+    uint32_t len = len_without_metadata + /*metadata*/ 5;
+    assert(len <= data_.size());
+    return Slice(metadata_ptr_ - len_without_metadata, len);
+  }
+
+  Slice Reset(uint32_t len_without_metadata, uint32_t num_lines,
+              uint32_t num_probes, bool fill_ones) {
+    data_.fill(fill_ones ? 0xff : 0);
+    return ResetNoFill(len_without_metadata, num_lines, num_probes);
+  }
+
+  Slice ResetWeirdFill(uint32_t len_without_metadata, uint32_t num_lines,
+                       uint32_t num_probes) {
+    for (uint32_t i = 0; i < data_.size(); ++i) {
+      data_[i] = static_cast<char>(0x7b7b >> (i % 7));
+    }
+    return ResetNoFill(len_without_metadata, num_lines, num_probes);
+  }
+};
+
+TEST_P(FullBloomTest, RawSchema) {
+  RawFilterTester cft;
+  // Legacy Bloom configurations
+  // Two probes, about 3/4 bits set: ~50% "FP" rate
+  // One 256-byte cache line.
+  OpenRaw(cft.ResetWeirdFill(256, 1, 2));
+  EXPECT_EQ(uint64_t{11384799501900898790U}, PackedMatches());
+
+  // Two 128-byte cache lines.
+  OpenRaw(cft.ResetWeirdFill(256, 2, 2));
+  EXPECT_EQ(uint64_t{10157853359773492589U}, PackedMatches());
+
+  // Four 64-byte cache lines.
+  OpenRaw(cft.ResetWeirdFill(256, 4, 2));
+  EXPECT_EQ(uint64_t{7123594913907464682U}, PackedMatches());
+
+  // Fast local Bloom configurations (marker 255 -> -1)
+  // Two probes, about 3/4 bits set: ~50% "FP" rate
+  // Four 64-byte cache lines.
+  OpenRaw(cft.ResetWeirdFill(256, 2U << 8, 255));
+  EXPECT_EQ(uint64_t{9957045189927952471U}, PackedMatches());
+
+  // Ribbon configurations (marker 254 -> -2)
+
+  // Even though the builder never builds configurations this
+  // small (preferring Bloom), we can test that the configuration
+  // can be read, for possible future-proofing.
+
+  // 256 slots, one result column = 32 bytes (2 blocks, seed 0)
+  // ~50% FP rate:
+  // 0b0101010111110101010000110000011011011111100100001110010011101010
+  OpenRaw(cft.ResetWeirdFill(32, 2U << 8, 254));
+  EXPECT_EQ(uint64_t{6193930559317665002U}, PackedMatches());
+
+  // 256 slots, three-to-four result columns = 112 bytes
+  // ~ 1 in 10 FP rate:
+  // 0b0000000000100000000000000000000001000001000000010000101000000000
+  OpenRaw(cft.ResetWeirdFill(112, 2U << 8, 254));
+  EXPECT_EQ(uint64_t{9007200345328128U}, PackedMatches());
+}
+
+TEST_P(FullBloomTest, CorruptFilters) {
+  RawFilterTester cft;
+
+  for (bool fill : {false, true}) {
+    // Legacy Bloom configurations
+    // Good filter bits - returns same as fill
+    OpenRaw(cft.Reset(CACHE_LINE_SIZE, 1, 6, fill));
+    ASSERT_EQ(fill, Matches("hello"));
+    ASSERT_EQ(fill, Matches("world"));
+
+    // Good filter bits - returns same as fill
+    OpenRaw(cft.Reset(CACHE_LINE_SIZE * 3, 3, 6, fill));
+    ASSERT_EQ(fill, Matches("hello"));
+    ASSERT_EQ(fill, Matches("world"));
+
+    // Good filter bits - returns same as fill
+    // 256 is unusual but legal cache line size
+    OpenRaw(cft.Reset(256 * 3, 3, 6, fill));
+    ASSERT_EQ(fill, Matches("hello"));
+    ASSERT_EQ(fill, Matches("world"));
+
+    // Good filter bits - returns same as fill
+    // 30 should be max num_probes
+    OpenRaw(cft.Reset(CACHE_LINE_SIZE, 1, 30, fill));
+    ASSERT_EQ(fill, Matches("hello"));
+    ASSERT_EQ(fill, Matches("world"));
+
+    // Good filter bits - returns same as fill
+    // 1 should be min num_probes
+    OpenRaw(cft.Reset(CACHE_LINE_SIZE, 1, 1, fill));
+    ASSERT_EQ(fill, Matches("hello"));
+    ASSERT_EQ(fill, Matches("world"));
+
+    // Type 1 trivial filter bits - returns true as if FP by zero probes
+    OpenRaw(cft.Reset(CACHE_LINE_SIZE, 1, 0, fill));
+    ASSERT_TRUE(Matches("hello"));
+    ASSERT_TRUE(Matches("world"));
+
+    // Type 2 trivial filter bits - returns false as if built from zero keys
+    OpenRaw(cft.Reset(0, 0, 6, fill));
+    ASSERT_FALSE(Matches("hello"));
+    ASSERT_FALSE(Matches("world"));
+
+    // Type 2 trivial filter bits - returns false as if built from zero keys
+    OpenRaw(cft.Reset(0, 37, 6, fill));
+    ASSERT_FALSE(Matches("hello"));
+    ASSERT_FALSE(Matches("world"));
+
+    // Type 2 trivial filter bits - returns false as 0 size trumps 0 probes
+    OpenRaw(cft.Reset(0, 0, 0, fill));
+    ASSERT_FALSE(Matches("hello"));
+    ASSERT_FALSE(Matches("world"));
+
+    // Bad filter bits - returns true for safety
+    // No solution to 0 * x == CACHE_LINE_SIZE
+    OpenRaw(cft.Reset(CACHE_LINE_SIZE, 0, 6, fill));
+    ASSERT_TRUE(Matches("hello"));
+    ASSERT_TRUE(Matches("world"));
+
+    // Bad filter bits - returns true for safety
+    // Can't have 3 * x == 4 for integer x
+    OpenRaw(cft.Reset(4, 3, 6, fill));
+    ASSERT_TRUE(Matches("hello"));
+    ASSERT_TRUE(Matches("world"));
+
+    // Bad filter bits - returns true for safety
+    // 97 bytes is not a power of two, so not a legal cache line size
+    OpenRaw(cft.Reset(97 * 3, 3, 6, fill));
+    ASSERT_TRUE(Matches("hello"));
+    ASSERT_TRUE(Matches("world"));
+
+    // Bad filter bits - returns true for safety
+    // 65 bytes is not a power of two, so not a legal cache line size
+    OpenRaw(cft.Reset(65 * 3, 3, 6, fill));
+    ASSERT_TRUE(Matches("hello"));
+    ASSERT_TRUE(Matches("world"));
+
+    // Bad filter bits - returns false as if built from zero keys
+    // < 5 bytes overall means missing even metadata
+    OpenRaw(cft.Reset(static_cast<uint32_t>(-1), 3, 6, fill));
+    ASSERT_FALSE(Matches("hello"));
+    ASSERT_FALSE(Matches("world"));
+
+    OpenRaw(cft.Reset(static_cast<uint32_t>(-5), 3, 6, fill));
+    ASSERT_FALSE(Matches("hello"));
+    ASSERT_FALSE(Matches("world"));
+
+    // Dubious filter bits - returns same as fill (for now)
+    // 31 is not a useful num_probes, nor generated by RocksDB unless directly
+    // using filter bits API without BloomFilterPolicy.
+    OpenRaw(cft.Reset(CACHE_LINE_SIZE, 1, 31, fill));
+    ASSERT_EQ(fill, Matches("hello"));
+    ASSERT_EQ(fill, Matches("world"));
+
+    // Dubious filter bits - returns same as fill (for now)
+    // Similar, with 127, largest positive char
+    OpenRaw(cft.Reset(CACHE_LINE_SIZE, 1, 127, fill));
+    ASSERT_EQ(fill, Matches("hello"));
+    ASSERT_EQ(fill, Matches("world"));
+
+    // Dubious filter bits - returns true (for now)
+    // num_probes set to 128 / -128, lowest negative char
+    // NB: Bug in implementation interprets this as negative and has same
+    // effect as zero probes, but effectively reserves negative char values
+    // for future use.
+    OpenRaw(cft.Reset(CACHE_LINE_SIZE, 1, 128, fill));
+    ASSERT_TRUE(Matches("hello"));
+    ASSERT_TRUE(Matches("world"));
+
+    // Dubious filter bits - returns true (for now)
+    // Similar, with 253 / -3
+    OpenRaw(cft.Reset(CACHE_LINE_SIZE, 1, 253, fill));
+    ASSERT_TRUE(Matches("hello"));
+    ASSERT_TRUE(Matches("world"));
+
+    // #########################################################
+    // Fast local Bloom configurations (marker 255 -> -1)
+    // Good config with six probes
+    OpenRaw(cft.Reset(CACHE_LINE_SIZE, 6U << 8, 255, fill));
+    ASSERT_EQ(fill, Matches("hello"));
+    ASSERT_EQ(fill, Matches("world"));
+
+    // Becomes bad/reserved config (always true) if any other byte set
+    OpenRaw(cft.Reset(CACHE_LINE_SIZE, (6U << 8) | 1U, 255, fill));
+    ASSERT_TRUE(Matches("hello"));
+    ASSERT_TRUE(Matches("world"));
+
+    OpenRaw(cft.Reset(CACHE_LINE_SIZE, (6U << 8) | (1U << 16), 255, fill));
+    ASSERT_TRUE(Matches("hello"));
+    ASSERT_TRUE(Matches("world"));
+
+    OpenRaw(cft.Reset(CACHE_LINE_SIZE, (6U << 8) | (1U << 24), 255, fill));
+    ASSERT_TRUE(Matches("hello"));
+    ASSERT_TRUE(Matches("world"));
+
+    // Good config, max 30 probes
+    OpenRaw(cft.Reset(CACHE_LINE_SIZE, 30U << 8, 255, fill));
+    ASSERT_EQ(fill, Matches("hello"));
+    ASSERT_EQ(fill, Matches("world"));
+
+    // Bad/reserved config (always true) if more than 30
+    OpenRaw(cft.Reset(CACHE_LINE_SIZE, 31U << 8, 255, fill));
+    ASSERT_TRUE(Matches("hello"));
+    ASSERT_TRUE(Matches("world"));
+
+    OpenRaw(cft.Reset(CACHE_LINE_SIZE, 33U << 8, 255, fill));
+    ASSERT_TRUE(Matches("hello"));
+    ASSERT_TRUE(Matches("world"));
+
+    OpenRaw(cft.Reset(CACHE_LINE_SIZE, 66U << 8, 255, fill));
+    ASSERT_TRUE(Matches("hello"));
+    ASSERT_TRUE(Matches("world"));
+
+    OpenRaw(cft.Reset(CACHE_LINE_SIZE, 130U << 8, 255, fill));
+    ASSERT_TRUE(Matches("hello"));
+    ASSERT_TRUE(Matches("world"));
+  }
+
+  // #########################################################
+  // Ribbon configurations (marker 254 -> -2)
+  // ("fill" doesn't work to detect good configurations, we just
+  // have to rely on TN probability)
+
+  // Good: 2 blocks * 16 bytes / segment * 4 columns = 128 bytes
+  // seed = 123
+  OpenRaw(cft.Reset(128, (2U << 8) + 123U, 254, false));
+  ASSERT_FALSE(Matches("hello"));
+  ASSERT_FALSE(Matches("world"));
+
+  // Good: 2 blocks * 16 bytes / segment * 8 columns = 256 bytes
+  OpenRaw(cft.Reset(256, (2U << 8) + 123U, 254, false));
+  ASSERT_FALSE(Matches("hello"));
+  ASSERT_FALSE(Matches("world"));
+
+  // Surprisingly OK: 5000 blocks (640,000 slots) in only 1024 bits
+  // -> average close to 0 columns
+  OpenRaw(cft.Reset(128, (5000U << 8) + 123U, 254, false));
+  // *Almost* all FPs
+  ASSERT_TRUE(Matches("hello"));
+  ASSERT_TRUE(Matches("world"));
+  // Need many queries to find a "true negative"
+  for (int i = 0; Matches(std::to_string(i)); ++i) {
+    ASSERT_LT(i, 1000);
+  }
+
+  // Bad: 1 block not allowed (for implementation detail reasons)
+  OpenRaw(cft.Reset(128, (1U << 8) + 123U, 254, false));
+  ASSERT_TRUE(Matches("hello"));
+  ASSERT_TRUE(Matches("world"));
+
+  // Bad: 0 blocks not allowed
+  OpenRaw(cft.Reset(128, (0U << 8) + 123U, 254, false));
+  ASSERT_TRUE(Matches("hello"));
+  ASSERT_TRUE(Matches("world"));
+}
+
+INSTANTIATE_TEST_CASE_P(Full, FullBloomTest,
+                        testing::Values(kLegacyBloom, kFastLocalBloom,
+                                        kStandard128Ribbon));
+
+static double GetEffectiveBitsPerKey(FilterBitsBuilder* builder) {
+  union {
+    uint64_t key_value = 0;
+    char key_bytes[8];
+  };
+
+  const unsigned kNumKeys = 1000;
+
+  Slice key_slice{key_bytes, 8};
+  for (key_value = 0; key_value < kNumKeys; ++key_value) {
+    builder->AddKey(key_slice);
+  }
+
+  std::unique_ptr<const char[]> buf;
+  auto filter = builder->Finish(&buf);
+  return filter.size() * /*bits per byte*/ 8 / (1.0 * kNumKeys);
+}
+
+static void SetTestingLevel(int levelish, FilterBuildingContext* ctx) {
+  if (levelish == -1) {
+    // Flush is treated as level -1 for this option but actually level 0
+    ctx->level_at_creation = 0;
+    ctx->reason = TableFileCreationReason::kFlush;
+  } else {
+    ctx->level_at_creation = levelish;
+    ctx->reason = TableFileCreationReason::kCompaction;
+  }
+}
+
+TEST(RibbonTest, RibbonTestLevelThreshold) {
+  BlockBasedTableOptions opts;
+  FilterBuildingContext ctx(opts);
+  // A few settings
+  for (CompactionStyle cs : {kCompactionStyleLevel, kCompactionStyleUniversal,
+                             kCompactionStyleFIFO, kCompactionStyleNone}) {
+    ctx.compaction_style = cs;
+    for (int bloom_before_level : {-1, 0, 1, 10}) {
+      std::vector<std::unique_ptr<const FilterPolicy> > policies;
+      policies.emplace_back(NewRibbonFilterPolicy(10, bloom_before_level));
+
+      if (bloom_before_level == 0) {
+        // Also test new API default
+        policies.emplace_back(NewRibbonFilterPolicy(10));
+      }
+
+      for (std::unique_ptr<const FilterPolicy>& policy : policies) {
+        // Claim to be generating filter for this level
+        SetTestingLevel(bloom_before_level, &ctx);
+
+        std::unique_ptr<FilterBitsBuilder> builder{
+            policy->GetBuilderWithContext(ctx)};
+
+        // Must be Ribbon (more space efficient than 10 bits per key)
+        ASSERT_LT(GetEffectiveBitsPerKey(builder.get()), 8);
+
+        if (bloom_before_level >= 0) {
+          // Claim to be generating filter for previous level
+          SetTestingLevel(bloom_before_level - 1, &ctx);
+
+          builder.reset(policy->GetBuilderWithContext(ctx));
+
+          if (cs == kCompactionStyleLevel || cs == kCompactionStyleUniversal) {
+            // Level is considered.
+            // Must be Bloom (~ 10 bits per key)
+            ASSERT_GT(GetEffectiveBitsPerKey(builder.get()), 9);
+          } else {
+            // Level is ignored under non-traditional compaction styles.
+            // Must be Ribbon (more space efficient than 10 bits per key)
+            ASSERT_LT(GetEffectiveBitsPerKey(builder.get()), 8);
+          }
+        }
+
+        // Like SST file writer
+        ctx.level_at_creation = -1;
+        ctx.reason = TableFileCreationReason::kMisc;
+
+        builder.reset(policy->GetBuilderWithContext(ctx));
+
+        // Must be Ribbon (more space efficient than 10 bits per key)
+        ASSERT_LT(GetEffectiveBitsPerKey(builder.get()), 8);
+      }
+    }
+  }
+}
+
+}  // namespace ROCKSDB_NAMESPACE
+
+int main(int argc, char** argv) {
+  ROCKSDB_NAMESPACE::port::InstallStackTraceHandler();
+  ::testing::InitGoogleTest(&argc, argv);
+  ParseCommandLineFlags(&argc, &argv, true);
+
+  return RUN_ALL_TESTS();
+}
+
+#endif  // GFLAGS
diff --git a/src/rocksdb/util/build_version.cc.in b/src/rocksdb/util/build_version.cc.in
new file mode 100644
index 000000000..c1706dc1f
--- /dev/null
+++ b/src/rocksdb/util/build_version.cc.in
@@ -0,0 +1,81 @@
+// Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.
+
+#include <memory>
+
+#include "rocksdb/version.h"
+#include "rocksdb/utilities/object_registry.h"
+#include "util/string_util.h"
+
+// The build script may replace these values with real values based
+// on whether or not GIT is available and the platform settings
+static const std::string rocksdb_build_git_sha  = "rocksdb_build_git_sha:@GIT_SHA@";
+static const std::string rocksdb_build_git_tag = "rocksdb_build_git_tag:@GIT_TAG@";
+#define HAS_GIT_CHANGES @GIT_MOD@
+#if HAS_GIT_CHANGES == 0
+// If HAS_GIT_CHANGES is 0, the GIT date is used.
+// Use the time the branch/tag was last modified
+static const std::string rocksdb_build_date = "rocksdb_build_date:@GIT_DATE@";
+#else
+// If HAS_GIT_CHANGES is > 0, the branch/tag has modifications.
+// Use the time the build was created.
+static const std::string rocksdb_build_date = "rocksdb_build_date:@BUILD_DATE@";
+#endif
+
+#ifndef ROCKSDB_LITE
+extern "C" {
+@ROCKSDB_PLUGIN_EXTERNS@
+} // extern "C"
+
+std::unordered_map<std::string, ROCKSDB_NAMESPACE::RegistrarFunc> ROCKSDB_NAMESPACE::ObjectRegistry::builtins_ = {
+  @ROCKSDB_PLUGIN_BUILTINS@
+};
+#endif //ROCKSDB_LITE
+
+namespace ROCKSDB_NAMESPACE {
+static void AddProperty(std::unordered_map<std::string, std::string> *props, const std::string& name) {
+  size_t colon = name.find(":");
+  if (colon != std::string::npos && colon > 0 && colon < name.length() - 1) {
+    // If we found a "@:", then this property was a build-time substitution that failed.  Skip it
+    size_t at = name.find("@", colon);
+    if (at != colon + 1) {
+      // Everything before the colon is the name, after is the value
+      (*props)[name.substr(0, colon)] = name.substr(colon + 1);
+    }
+  }
+}
+
+static std::unordered_map<std::string, std::string>* LoadPropertiesSet() {
+  auto * properties = new std::unordered_map<std::string, std::string>();
+  AddProperty(properties, rocksdb_build_git_sha);
+  AddProperty(properties, rocksdb_build_git_tag);
+  AddProperty(properties, rocksdb_build_date);
+  return properties;
+}
+
+const std::unordered_map<std::string, std::string>& GetRocksBuildProperties() {
+  static std::unique_ptr<std::unordered_map<std::string, std::string>> props(LoadPropertiesSet());
+  return *props;
+}
+
+std::string GetRocksVersionAsString(bool with_patch) {
+  std::string version = std::to_string(ROCKSDB_MAJOR) + "." + std::to_string(ROCKSDB_MINOR);
+  if (with_patch) {
+    return version + "." + std::to_string(ROCKSDB_PATCH);
+  } else {
+    return version;
+ }
+}
+
+std::string GetRocksBuildInfoAsString(const std::string& program, bool verbose) {
+  std::string info = program + " (RocksDB) " + GetRocksVersionAsString(true);
+  if (verbose) {
+    for (const auto& it : GetRocksBuildProperties()) {
+      info.append("\n    ");
+      info.append(it.first);
+      info.append(": ");
+      info.append(it.second);
+    }
+  }
+  return info;
+}
+} // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/util/cast_util.h b/src/rocksdb/util/cast_util.h
new file mode 100644
index 000000000..c91b6ff1e
--- /dev/null
+++ b/src/rocksdb/util/cast_util.h
@@ -0,0 +1,42 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#pragma once
+
+#include <type_traits>
+
+#include "rocksdb/rocksdb_namespace.h"
+
+namespace ROCKSDB_NAMESPACE {
+// The helper function to assert the move from dynamic_cast<> to
+// static_cast<> is correct. This function is to deal with legacy code.
+// It is not recommended to add new code to issue class casting. The preferred
+// solution is to implement the functionality without a need of casting.
+template <class DestClass, class SrcClass>
+inline DestClass* static_cast_with_check(SrcClass* x) {
+  DestClass* ret = static_cast<DestClass*>(x);
+#ifdef ROCKSDB_USE_RTTI
+  assert(ret == dynamic_cast<DestClass*>(x));
+#endif
+  return ret;
+}
+
+// A wrapper around static_cast for lossless conversion between integral
+// types, including enum types. For example, this can be used for converting
+// between signed/unsigned or enum type and underlying type without fear of
+// stripping away data, now or in the future.
+template <typename To, typename From>
+inline To lossless_cast(From x) {
+  using FromValue = typename std::remove_reference<From>::type;
+  static_assert(
+      std::is_integral<FromValue>::value || std::is_enum<FromValue>::value,
+      "Only works on integral types");
+  static_assert(std::is_integral<To>::value || std::is_enum<To>::value,
+                "Only works on integral types");
+  static_assert(sizeof(To) >= sizeof(FromValue), "Must be lossless");
+  return static_cast<To>(x);
+}
+
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/util/channel.h b/src/rocksdb/util/channel.h
new file mode 100644
index 000000000..19b956297
--- /dev/null
+++ b/src/rocksdb/util/channel.h
@@ -0,0 +1,69 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#pragma once
+
+#include <condition_variable>
+#include <mutex>
+#include <queue>
+#include <utility>
+
+#include "rocksdb/rocksdb_namespace.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+template <class T>
+class channel {
+ public:
+  explicit channel() : eof_(false) {}
+
+  channel(const channel&) = delete;
+  void operator=(const channel&) = delete;
+
+  void sendEof() {
+    std::lock_guard<std::mutex> lk(lock_);
+    eof_ = true;
+    cv_.notify_all();
+  }
+
+  bool eof() {
+    std::lock_guard<std::mutex> lk(lock_);
+    return buffer_.empty() && eof_;
+  }
+
+  size_t size() const {
+    std::lock_guard<std::mutex> lk(lock_);
+    return buffer_.size();
+  }
+
+  // writes elem to the queue
+  void write(T&& elem) {
+    std::unique_lock<std::mutex> lk(lock_);
+    buffer_.emplace(std::forward<T>(elem));
+    cv_.notify_one();
+  }
+
+  /// Moves a dequeued element onto elem, blocking until an element
+  /// is available.
+  // returns false if EOF
+  bool read(T& elem) {
+    std::unique_lock<std::mutex> lk(lock_);
+    cv_.wait(lk, [&] { return eof_ || !buffer_.empty(); });
+    if (eof_ && buffer_.empty()) {
+      return false;
+    }
+    elem = std::move(buffer_.front());
+    buffer_.pop();
+    cv_.notify_one();
+    return true;
+  }
+
+ private:
+  std::condition_variable cv_;
+  mutable std::mutex lock_;
+  std::queue<T> buffer_;
+  bool eof_;
+};
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/util/cleanable.cc b/src/rocksdb/util/cleanable.cc
new file mode 100644
index 000000000..89a7ab9be
--- /dev/null
+++ b/src/rocksdb/util/cleanable.cc
@@ -0,0 +1,181 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+//
+// Copyright (c) 2011 The LevelDB Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file. See the AUTHORS file for names of contributors.
+
+#include "rocksdb/cleanable.h"
+
+#include <atomic>
+#include <cassert>
+#include <utility>
+
+namespace ROCKSDB_NAMESPACE {
+
+Cleanable::Cleanable() {
+  cleanup_.function = nullptr;
+  cleanup_.next = nullptr;
+}
+
+Cleanable::~Cleanable() { DoCleanup(); }
+
+Cleanable::Cleanable(Cleanable&& other) noexcept { *this = std::move(other); }
+
+Cleanable& Cleanable::operator=(Cleanable&& other) noexcept {
+  assert(this != &other);  // https://stackoverflow.com/a/9322542/454544
+  cleanup_ = other.cleanup_;
+  other.cleanup_.function = nullptr;
+  other.cleanup_.next = nullptr;
+  return *this;
+}
+
+// If the entire linked list was on heap we could have simply add attach one
+// link list to another. However the head is an embeded object to avoid the cost
+// of creating objects for most of the use cases when the Cleanable has only one
+// Cleanup to do. We could put evernything on heap if benchmarks show no
+// negative impact on performance.
+// Also we need to iterate on the linked list since there is no pointer to the
+// tail. We can add the tail pointer but maintainin it might negatively impact
+// the perforamnce for the common case of one cleanup where tail pointer is not
+// needed. Again benchmarks could clarify that.
+// Even without a tail pointer we could iterate on the list, find the tail, and
+// have only that node updated without the need to insert the Cleanups one by
+// one. This however would be redundant when the source Cleanable has one or a
+// few Cleanups which is the case most of the time.
+// TODO(myabandeh): if the list is too long we should maintain a tail pointer
+// and have the entire list (minus the head that has to be inserted separately)
+// merged with the target linked list at once.
+void Cleanable::DelegateCleanupsTo(Cleanable* other) {
+  assert(other != nullptr);
+  if (cleanup_.function == nullptr) {
+    return;
+  }
+  Cleanup* c = &cleanup_;
+  other->RegisterCleanup(c->function, c->arg1, c->arg2);
+  c = c->next;
+  while (c != nullptr) {
+    Cleanup* next = c->next;
+    other->RegisterCleanup(c);
+    c = next;
+  }
+  cleanup_.function = nullptr;
+  cleanup_.next = nullptr;
+}
+
+void Cleanable::RegisterCleanup(Cleanable::Cleanup* c) {
+  assert(c != nullptr);
+  if (cleanup_.function == nullptr) {
+    cleanup_.function = c->function;
+    cleanup_.arg1 = c->arg1;
+    cleanup_.arg2 = c->arg2;
+    delete c;
+  } else {
+    c->next = cleanup_.next;
+    cleanup_.next = c;
+  }
+}
+
+void Cleanable::RegisterCleanup(CleanupFunction func, void* arg1, void* arg2) {
+  assert(func != nullptr);
+  Cleanup* c;
+  if (cleanup_.function == nullptr) {
+    c = &cleanup_;
+  } else {
+    c = new Cleanup;
+    c->next = cleanup_.next;
+    cleanup_.next = c;
+  }
+  c->function = func;
+  c->arg1 = arg1;
+  c->arg2 = arg2;
+}
+
+struct SharedCleanablePtr::Impl : public Cleanable {
+  std::atomic<unsigned> ref_count{1};  // Start with 1 ref
+  void Ref() { ref_count.fetch_add(1, std::memory_order_relaxed); }
+  void Unref() {
+    if (ref_count.fetch_sub(1, std::memory_order_relaxed) == 1) {
+      // Last ref
+      delete this;
+    }
+  }
+  static void UnrefWrapper(void* arg1, void* /*arg2*/) {
+    static_cast<SharedCleanablePtr::Impl*>(arg1)->Unref();
+  }
+};
+
+void SharedCleanablePtr::Reset() {
+  if (ptr_) {
+    ptr_->Unref();
+    ptr_ = nullptr;
+  }
+}
+
+void SharedCleanablePtr::Allocate() {
+  Reset();
+  ptr_ = new Impl();
+}
+
+SharedCleanablePtr::SharedCleanablePtr(const SharedCleanablePtr& from) {
+  *this = from;
+}
+
+SharedCleanablePtr::SharedCleanablePtr(SharedCleanablePtr&& from) noexcept {
+  *this = std::move(from);
+}
+
+SharedCleanablePtr& SharedCleanablePtr::operator=(
+    const SharedCleanablePtr& from) {
+  if (this != &from) {
+    Reset();
+    ptr_ = from.ptr_;
+    if (ptr_) {
+      ptr_->Ref();
+    }
+  }
+  return *this;
+}
+
+SharedCleanablePtr& SharedCleanablePtr::operator=(
+    SharedCleanablePtr&& from) noexcept {
+  assert(this != &from);  // https://stackoverflow.com/a/9322542/454544
+  Reset();
+  ptr_ = from.ptr_;
+  from.ptr_ = nullptr;
+  return *this;
+}
+
+SharedCleanablePtr::~SharedCleanablePtr() { Reset(); }
+
+Cleanable& SharedCleanablePtr::operator*() {
+  return *ptr_;  // implicit upcast
+}
+
+Cleanable* SharedCleanablePtr::operator->() {
+  return ptr_;  // implicit upcast
+}
+
+Cleanable* SharedCleanablePtr::get() {
+  return ptr_;  // implicit upcast
+}
+
+void SharedCleanablePtr::RegisterCopyWith(Cleanable* target) {
+  if (ptr_) {
+    // "Virtual" copy of the pointer
+    ptr_->Ref();
+    target->RegisterCleanup(&Impl::UnrefWrapper, ptr_, nullptr);
+  }
+}
+
+void SharedCleanablePtr::MoveAsCleanupTo(Cleanable* target) {
+  if (ptr_) {
+    // "Virtual" move of the pointer
+    target->RegisterCleanup(&Impl::UnrefWrapper, ptr_, nullptr);
+    ptr_ = nullptr;
+  }
+}
+
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/util/coding.cc b/src/rocksdb/util/coding.cc
new file mode 100644
index 000000000..3da8afaa2
--- /dev/null
+++ b/src/rocksdb/util/coding.cc
@@ -0,0 +1,90 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+//
+// Copyright (c) 2011 The LevelDB Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file. See the AUTHORS file for names of contributors.
+
+#include "util/coding.h"
+
+#include <algorithm>
+
+#include "rocksdb/slice.h"
+#include "rocksdb/slice_transform.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+// conversion' conversion from 'type1' to 'type2', possible loss of data
+#if defined(_MSC_VER)
+#pragma warning(push)
+#pragma warning(disable : 4244)
+#endif
+char* EncodeVarint32(char* dst, uint32_t v) {
+  // Operate on characters as unsigneds
+  unsigned char* ptr = reinterpret_cast<unsigned char*>(dst);
+  static const int B = 128;
+  if (v < (1 << 7)) {
+    *(ptr++) = v;
+  } else if (v < (1 << 14)) {
+    *(ptr++) = v | B;
+    *(ptr++) = v >> 7;
+  } else if (v < (1 << 21)) {
+    *(ptr++) = v | B;
+    *(ptr++) = (v >> 7) | B;
+    *(ptr++) = v >> 14;
+  } else if (v < (1 << 28)) {
+    *(ptr++) = v | B;
+    *(ptr++) = (v >> 7) | B;
+    *(ptr++) = (v >> 14) | B;
+    *(ptr++) = v >> 21;
+  } else {
+    *(ptr++) = v | B;
+    *(ptr++) = (v >> 7) | B;
+    *(ptr++) = (v >> 14) | B;
+    *(ptr++) = (v >> 21) | B;
+    *(ptr++) = v >> 28;
+  }
+  return reinterpret_cast<char*>(ptr);
+}
+#if defined(_MSC_VER)
+#pragma warning(pop)
+#endif
+
+const char* GetVarint32PtrFallback(const char* p, const char* limit,
+                                   uint32_t* value) {
+  uint32_t result = 0;
+  for (uint32_t shift = 0; shift <= 28 && p < limit; shift += 7) {
+    uint32_t byte = *(reinterpret_cast<const unsigned char*>(p));
+    p++;
+    if (byte & 128) {
+      // More bytes are present
+      result |= ((byte & 127) << shift);
+    } else {
+      result |= (byte << shift);
+      *value = result;
+      return reinterpret_cast<const char*>(p);
+    }
+  }
+  return nullptr;
+}
+
+const char* GetVarint64Ptr(const char* p, const char* limit, uint64_t* value) {
+  uint64_t result = 0;
+  for (uint32_t shift = 0; shift <= 63 && p < limit; shift += 7) {
+    uint64_t byte = *(reinterpret_cast<const unsigned char*>(p));
+    p++;
+    if (byte & 128) {
+      // More bytes are present
+      result |= ((byte & 127) << shift);
+    } else {
+      result |= (byte << shift);
+      *value = result;
+      return reinterpret_cast<const char*>(p);
+    }
+  }
+  return nullptr;
+}
+
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/util/coding.h b/src/rocksdb/util/coding.h
new file mode 100644
index 000000000..3168fd2fd
--- /dev/null
+++ b/src/rocksdb/util/coding.h
@@ -0,0 +1,389 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+//
+// Copyright (c) 2011 The LevelDB Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file. See the AUTHORS file for names of contributors.
+//
+// Encoding independent of machine byte order:
+// * Fixed-length numbers are encoded with least-significant byte first
+//   (little endian, native order on Intel and others)
+// * In addition we support variable length "varint" encoding
+// * Strings are encoded prefixed by their length in varint format
+//
+// Some related functions are provided in coding_lean.h
+
+#pragma once
+#include <algorithm>
+#include <string>
+
+#include "port/port.h"
+#include "rocksdb/slice.h"
+#include "util/coding_lean.h"
+
+// Some processors does not allow unaligned access to memory
+#if defined(__sparc)
+#define PLATFORM_UNALIGNED_ACCESS_NOT_ALLOWED
+#endif
+
+namespace ROCKSDB_NAMESPACE {
+
+// The maximum length of a varint in bytes for 64-bit.
+const uint32_t kMaxVarint64Length = 10;
+
+// Standard Put... routines append to a string
+extern void PutFixed16(std::string* dst, uint16_t value);
+extern void PutFixed32(std::string* dst, uint32_t value);
+extern void PutFixed64(std::string* dst, uint64_t value);
+extern void PutVarint32(std::string* dst, uint32_t value);
+extern void PutVarint32Varint32(std::string* dst, uint32_t value1,
+                                uint32_t value2);
+extern void PutVarint32Varint32Varint32(std::string* dst, uint32_t value1,
+                                        uint32_t value2, uint32_t value3);
+extern void PutVarint64(std::string* dst, uint64_t value);
+extern void PutVarint64Varint64(std::string* dst, uint64_t value1,
+                                uint64_t value2);
+extern void PutVarint32Varint64(std::string* dst, uint32_t value1,
+                                uint64_t value2);
+extern void PutVarint32Varint32Varint64(std::string* dst, uint32_t value1,
+                                        uint32_t value2, uint64_t value3);
+extern void PutLengthPrefixedSlice(std::string* dst, const Slice& value);
+extern void PutLengthPrefixedSliceParts(std::string* dst,
+                                        const SliceParts& slice_parts);
+extern void PutLengthPrefixedSlicePartsWithPadding(
+    std::string* dst, const SliceParts& slice_parts, size_t pad_sz);
+
+// Standard Get... routines parse a value from the beginning of a Slice
+// and advance the slice past the parsed value.
+extern bool GetFixed64(Slice* input, uint64_t* value);
+extern bool GetFixed32(Slice* input, uint32_t* value);
+extern bool GetFixed16(Slice* input, uint16_t* value);
+extern bool GetVarint32(Slice* input, uint32_t* value);
+extern bool GetVarint64(Slice* input, uint64_t* value);
+extern bool GetVarsignedint64(Slice* input, int64_t* value);
+extern bool GetLengthPrefixedSlice(Slice* input, Slice* result);
+// This function assumes data is well-formed.
+extern Slice GetLengthPrefixedSlice(const char* data);
+
+extern Slice GetSliceUntil(Slice* slice, char delimiter);
+
+// Borrowed from
+// https://github.com/facebook/fbthrift/blob/449a5f77f9f9bae72c9eb5e78093247eef185c04/thrift/lib/cpp/util/VarintUtils-inl.h#L202-L208
+constexpr inline uint64_t i64ToZigzag(const int64_t l) {
+  return (static_cast<uint64_t>(l) << 1) ^ static_cast<uint64_t>(l >> 63);
+}
+inline int64_t zigzagToI64(uint64_t n) {
+  return (n >> 1) ^ -static_cast<int64_t>(n & 1);
+}
+
+// Pointer-based variants of GetVarint...  These either store a value
+// in *v and return a pointer just past the parsed value, or return
+// nullptr on error.  These routines only look at bytes in the range
+// [p..limit-1]
+extern const char* GetVarint32Ptr(const char* p, const char* limit,
+                                  uint32_t* v);
+extern const char* GetVarint64Ptr(const char* p, const char* limit,
+                                  uint64_t* v);
+inline const char* GetVarsignedint64Ptr(const char* p, const char* limit,
+                                        int64_t* value) {
+  uint64_t u = 0;
+  const char* ret = GetVarint64Ptr(p, limit, &u);
+  *value = zigzagToI64(u);
+  return ret;
+}
+
+// Returns the length of the varint32 or varint64 encoding of "v"
+extern int VarintLength(uint64_t v);
+
+// Lower-level versions of Put... that write directly into a character buffer
+// and return a pointer just past the last byte written.
+// REQUIRES: dst has enough space for the value being written
+extern char* EncodeVarint32(char* dst, uint32_t value);
+extern char* EncodeVarint64(char* dst, uint64_t value);
+
+// Internal routine for use by fallback path of GetVarint32Ptr
+extern const char* GetVarint32PtrFallback(const char* p, const char* limit,
+                                          uint32_t* value);
+inline const char* GetVarint32Ptr(const char* p, const char* limit,
+                                  uint32_t* value) {
+  if (p < limit) {
+    uint32_t result = *(reinterpret_cast<const unsigned char*>(p));
+    if ((result & 128) == 0) {
+      *value = result;
+      return p + 1;
+    }
+  }
+  return GetVarint32PtrFallback(p, limit, value);
+}
+
+// Pull the last 8 bits and cast it to a character
+inline void PutFixed16(std::string* dst, uint16_t value) {
+  if (port::kLittleEndian) {
+    dst->append(const_cast<const char*>(reinterpret_cast<char*>(&value)),
+                sizeof(value));
+  } else {
+    char buf[sizeof(value)];
+    EncodeFixed16(buf, value);
+    dst->append(buf, sizeof(buf));
+  }
+}
+
+inline void PutFixed32(std::string* dst, uint32_t value) {
+  if (port::kLittleEndian) {
+    dst->append(const_cast<const char*>(reinterpret_cast<char*>(&value)),
+                sizeof(value));
+  } else {
+    char buf[sizeof(value)];
+    EncodeFixed32(buf, value);
+    dst->append(buf, sizeof(buf));
+  }
+}
+
+inline void PutFixed64(std::string* dst, uint64_t value) {
+  if (port::kLittleEndian) {
+    dst->append(const_cast<const char*>(reinterpret_cast<char*>(&value)),
+                sizeof(value));
+  } else {
+    char buf[sizeof(value)];
+    EncodeFixed64(buf, value);
+    dst->append(buf, sizeof(buf));
+  }
+}
+
+inline void PutVarint32(std::string* dst, uint32_t v) {
+  char buf[5];
+  char* ptr = EncodeVarint32(buf, v);
+  dst->append(buf, static_cast<size_t>(ptr - buf));
+}
+
+inline void PutVarint32Varint32(std::string* dst, uint32_t v1, uint32_t v2) {
+  char buf[10];
+  char* ptr = EncodeVarint32(buf, v1);
+  ptr = EncodeVarint32(ptr, v2);
+  dst->append(buf, static_cast<size_t>(ptr - buf));
+}
+
+inline void PutVarint32Varint32Varint32(std::string* dst, uint32_t v1,
+                                        uint32_t v2, uint32_t v3) {
+  char buf[15];
+  char* ptr = EncodeVarint32(buf, v1);
+  ptr = EncodeVarint32(ptr, v2);
+  ptr = EncodeVarint32(ptr, v3);
+  dst->append(buf, static_cast<size_t>(ptr - buf));
+}
+
+inline char* EncodeVarint64(char* dst, uint64_t v) {
+  static const unsigned int B = 128;
+  unsigned char* ptr = reinterpret_cast<unsigned char*>(dst);
+  while (v >= B) {
+    *(ptr++) = (v & (B - 1)) | B;
+    v >>= 7;
+  }
+  *(ptr++) = static_cast<unsigned char>(v);
+  return reinterpret_cast<char*>(ptr);
+}
+
+inline void PutVarint64(std::string* dst, uint64_t v) {
+  char buf[kMaxVarint64Length];
+  char* ptr = EncodeVarint64(buf, v);
+  dst->append(buf, static_cast<size_t>(ptr - buf));
+}
+
+inline void PutVarsignedint64(std::string* dst, int64_t v) {
+  char buf[kMaxVarint64Length];
+  // Using Zigzag format to convert signed to unsigned
+  char* ptr = EncodeVarint64(buf, i64ToZigzag(v));
+  dst->append(buf, static_cast<size_t>(ptr - buf));
+}
+
+inline void PutVarint64Varint64(std::string* dst, uint64_t v1, uint64_t v2) {
+  char buf[20];
+  char* ptr = EncodeVarint64(buf, v1);
+  ptr = EncodeVarint64(ptr, v2);
+  dst->append(buf, static_cast<size_t>(ptr - buf));
+}
+
+inline void PutVarint32Varint64(std::string* dst, uint32_t v1, uint64_t v2) {
+  char buf[15];
+  char* ptr = EncodeVarint32(buf, v1);
+  ptr = EncodeVarint64(ptr, v2);
+  dst->append(buf, static_cast<size_t>(ptr - buf));
+}
+
+inline void PutVarint32Varint32Varint64(std::string* dst, uint32_t v1,
+                                        uint32_t v2, uint64_t v3) {
+  char buf[20];
+  char* ptr = EncodeVarint32(buf, v1);
+  ptr = EncodeVarint32(ptr, v2);
+  ptr = EncodeVarint64(ptr, v3);
+  dst->append(buf, static_cast<size_t>(ptr - buf));
+}
+
+inline void PutLengthPrefixedSlice(std::string* dst, const Slice& value) {
+  PutVarint32(dst, static_cast<uint32_t>(value.size()));
+  dst->append(value.data(), value.size());
+}
+
+inline void PutLengthPrefixedSliceParts(std::string* dst, size_t total_bytes,
+                                        const SliceParts& slice_parts) {
+  for (int i = 0; i < slice_parts.num_parts; ++i) {
+    total_bytes += slice_parts.parts[i].size();
+  }
+  PutVarint32(dst, static_cast<uint32_t>(total_bytes));
+  for (int i = 0; i < slice_parts.num_parts; ++i) {
+    dst->append(slice_parts.parts[i].data(), slice_parts.parts[i].size());
+  }
+}
+
+inline void PutLengthPrefixedSliceParts(std::string* dst,
+                                        const SliceParts& slice_parts) {
+  PutLengthPrefixedSliceParts(dst, /*total_bytes=*/0, slice_parts);
+}
+
+inline void PutLengthPrefixedSlicePartsWithPadding(
+    std::string* dst, const SliceParts& slice_parts, size_t pad_sz) {
+  PutLengthPrefixedSliceParts(dst, /*total_bytes=*/pad_sz, slice_parts);
+  dst->append(pad_sz, '\0');
+}
+
+inline int VarintLength(uint64_t v) {
+  int len = 1;
+  while (v >= 128) {
+    v >>= 7;
+    len++;
+  }
+  return len;
+}
+
+inline bool GetFixed64(Slice* input, uint64_t* value) {
+  if (input->size() < sizeof(uint64_t)) {
+    return false;
+  }
+  *value = DecodeFixed64(input->data());
+  input->remove_prefix(sizeof(uint64_t));
+  return true;
+}
+
+inline bool GetFixed32(Slice* input, uint32_t* value) {
+  if (input->size() < sizeof(uint32_t)) {
+    return false;
+  }
+  *value = DecodeFixed32(input->data());
+  input->remove_prefix(sizeof(uint32_t));
+  return true;
+}
+
+inline bool GetFixed16(Slice* input, uint16_t* value) {
+  if (input->size() < sizeof(uint16_t)) {
+    return false;
+  }
+  *value = DecodeFixed16(input->data());
+  input->remove_prefix(sizeof(uint16_t));
+  return true;
+}
+
+inline bool GetVarint32(Slice* input, uint32_t* value) {
+  const char* p = input->data();
+  const char* limit = p + input->size();
+  const char* q = GetVarint32Ptr(p, limit, value);
+  if (q == nullptr) {
+    return false;
+  } else {
+    *input = Slice(q, static_cast<size_t>(limit - q));
+    return true;
+  }
+}
+
+inline bool GetVarint64(Slice* input, uint64_t* value) {
+  const char* p = input->data();
+  const char* limit = p + input->size();
+  const char* q = GetVarint64Ptr(p, limit, value);
+  if (q == nullptr) {
+    return false;
+  } else {
+    *input = Slice(q, static_cast<size_t>(limit - q));
+    return true;
+  }
+}
+
+inline bool GetVarsignedint64(Slice* input, int64_t* value) {
+  const char* p = input->data();
+  const char* limit = p + input->size();
+  const char* q = GetVarsignedint64Ptr(p, limit, value);
+  if (q == nullptr) {
+    return false;
+  } else {
+    *input = Slice(q, static_cast<size_t>(limit - q));
+    return true;
+  }
+}
+
+inline bool GetLengthPrefixedSlice(Slice* input, Slice* result) {
+  uint32_t len = 0;
+  if (GetVarint32(input, &len) && input->size() >= len) {
+    *result = Slice(input->data(), len);
+    input->remove_prefix(len);
+    return true;
+  } else {
+    return false;
+  }
+}
+
+inline Slice GetLengthPrefixedSlice(const char* data) {
+  uint32_t len = 0;
+  // +5: we assume "data" is not corrupted
+  // unsigned char is 7 bits, uint32_t is 32 bits, need 5 unsigned char
+  auto p = GetVarint32Ptr(data, data + 5 /* limit */, &len);
+  return Slice(p, len);
+}
+
+inline Slice GetSliceUntil(Slice* slice, char delimiter) {
+  uint32_t len = 0;
+  for (len = 0; len < slice->size() && slice->data()[len] != delimiter; ++len) {
+    // nothing
+  }
+
+  Slice ret(slice->data(), len);
+  slice->remove_prefix(len + ((len < slice->size()) ? 1 : 0));
+  return ret;
+}
+
+template <class T>
+#ifdef ROCKSDB_UBSAN_RUN
+#if defined(__clang__)
+__attribute__((__no_sanitize__("alignment")))
+#elif defined(__GNUC__)
+__attribute__((__no_sanitize_undefined__))
+#endif
+#endif
+inline void
+PutUnaligned(T* memory, const T& value) {
+#if defined(PLATFORM_UNALIGNED_ACCESS_NOT_ALLOWED)
+  char* nonAlignedMemory = reinterpret_cast<char*>(memory);
+  memcpy(nonAlignedMemory, reinterpret_cast<const char*>(&value), sizeof(T));
+#else
+  *memory = value;
+#endif
+}
+
+template <class T>
+#ifdef ROCKSDB_UBSAN_RUN
+#if defined(__clang__)
+__attribute__((__no_sanitize__("alignment")))
+#elif defined(__GNUC__)
+__attribute__((__no_sanitize_undefined__))
+#endif
+#endif
+inline void
+GetUnaligned(const T* memory, T* value) {
+#if defined(PLATFORM_UNALIGNED_ACCESS_NOT_ALLOWED)
+  char* nonAlignedMemory = reinterpret_cast<char*>(value);
+  memcpy(nonAlignedMemory, reinterpret_cast<const char*>(memory), sizeof(T));
+#else
+  *value = *memory;
+#endif
+}
+
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/util/coding_lean.h b/src/rocksdb/util/coding_lean.h
new file mode 100644
index 000000000..6966f7a66
--- /dev/null
+++ b/src/rocksdb/util/coding_lean.h
@@ -0,0 +1,101 @@
+//  Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+// Encoding independent of machine byte order:
+// * Fixed-length numbers are encoded with least-significant byte first
+//   (little endian, native order on Intel and others)
+//
+// More functions in coding.h
+
+#pragma once
+
+#include <cstdint>
+#include <cstring>
+
+#include "port/port.h"  // for port::kLittleEndian
+
+namespace ROCKSDB_NAMESPACE {
+
+// Lower-level versions of Put... that write directly into a character buffer
+// REQUIRES: dst has enough space for the value being written
+// -- Implementation of the functions declared above
+inline void EncodeFixed16(char* buf, uint16_t value) {
+  if (port::kLittleEndian) {
+    memcpy(buf, &value, sizeof(value));
+  } else {
+    buf[0] = value & 0xff;
+    buf[1] = (value >> 8) & 0xff;
+  }
+}
+
+inline void EncodeFixed32(char* buf, uint32_t value) {
+  if (port::kLittleEndian) {
+    memcpy(buf, &value, sizeof(value));
+  } else {
+    buf[0] = value & 0xff;
+    buf[1] = (value >> 8) & 0xff;
+    buf[2] = (value >> 16) & 0xff;
+    buf[3] = (value >> 24) & 0xff;
+  }
+}
+
+inline void EncodeFixed64(char* buf, uint64_t value) {
+  if (port::kLittleEndian) {
+    memcpy(buf, &value, sizeof(value));
+  } else {
+    buf[0] = value & 0xff;
+    buf[1] = (value >> 8) & 0xff;
+    buf[2] = (value >> 16) & 0xff;
+    buf[3] = (value >> 24) & 0xff;
+    buf[4] = (value >> 32) & 0xff;
+    buf[5] = (value >> 40) & 0xff;
+    buf[6] = (value >> 48) & 0xff;
+    buf[7] = (value >> 56) & 0xff;
+  }
+}
+
+// Lower-level versions of Get... that read directly from a character buffer
+// without any bounds checking.
+
+inline uint16_t DecodeFixed16(const char* ptr) {
+  if (port::kLittleEndian) {
+    // Load the raw bytes
+    uint16_t result;
+    memcpy(&result, ptr, sizeof(result));  // gcc optimizes this to a plain load
+    return result;
+  } else {
+    return ((static_cast<uint16_t>(static_cast<unsigned char>(ptr[0]))) |
+            (static_cast<uint16_t>(static_cast<unsigned char>(ptr[1])) << 8));
+  }
+}
+
+inline uint32_t DecodeFixed32(const char* ptr) {
+  if (port::kLittleEndian) {
+    // Load the raw bytes
+    uint32_t result;
+    memcpy(&result, ptr, sizeof(result));  // gcc optimizes this to a plain load
+    return result;
+  } else {
+    return ((static_cast<uint32_t>(static_cast<unsigned char>(ptr[0]))) |
+            (static_cast<uint32_t>(static_cast<unsigned char>(ptr[1])) << 8) |
+            (static_cast<uint32_t>(static_cast<unsigned char>(ptr[2])) << 16) |
+            (static_cast<uint32_t>(static_cast<unsigned char>(ptr[3])) << 24));
+  }
+}
+
+inline uint64_t DecodeFixed64(const char* ptr) {
+  if (port::kLittleEndian) {
+    // Load the raw bytes
+    uint64_t result;
+    memcpy(&result, ptr, sizeof(result));  // gcc optimizes this to a plain load
+    return result;
+  } else {
+    uint64_t lo = DecodeFixed32(ptr);
+    uint64_t hi = DecodeFixed32(ptr + 4);
+    return (hi << 32) | lo;
+  }
+}
+
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/util/coding_test.cc b/src/rocksdb/util/coding_test.cc
new file mode 100644
index 000000000..79dd7b82e
--- /dev/null
+++ b/src/rocksdb/util/coding_test.cc
@@ -0,0 +1,217 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+//
+// Copyright (c) 2011 The LevelDB Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file. See the AUTHORS file for names of contributors.
+
+#include "util/coding.h"
+
+#include "test_util/testharness.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+class Coding {};
+TEST(Coding, Fixed16) {
+  std::string s;
+  for (uint16_t v = 0; v < 0xFFFF; v++) {
+    PutFixed16(&s, v);
+  }
+
+  const char* p = s.data();
+  for (uint16_t v = 0; v < 0xFFFF; v++) {
+    uint16_t actual = DecodeFixed16(p);
+    ASSERT_EQ(v, actual);
+    p += sizeof(uint16_t);
+  }
+}
+
+TEST(Coding, Fixed32) {
+  std::string s;
+  for (uint32_t v = 0; v < 100000; v++) {
+    PutFixed32(&s, v);
+  }
+
+  const char* p = s.data();
+  for (uint32_t v = 0; v < 100000; v++) {
+    uint32_t actual = DecodeFixed32(p);
+    ASSERT_EQ(v, actual);
+    p += sizeof(uint32_t);
+  }
+}
+
+TEST(Coding, Fixed64) {
+  std::string s;
+  for (int power = 0; power <= 63; power++) {
+    uint64_t v = static_cast<uint64_t>(1) << power;
+    PutFixed64(&s, v - 1);
+    PutFixed64(&s, v + 0);
+    PutFixed64(&s, v + 1);
+  }
+
+  const char* p = s.data();
+  for (int power = 0; power <= 63; power++) {
+    uint64_t v = static_cast<uint64_t>(1) << power;
+    uint64_t actual = 0;
+    actual = DecodeFixed64(p);
+    ASSERT_EQ(v - 1, actual);
+    p += sizeof(uint64_t);
+
+    actual = DecodeFixed64(p);
+    ASSERT_EQ(v + 0, actual);
+    p += sizeof(uint64_t);
+
+    actual = DecodeFixed64(p);
+    ASSERT_EQ(v + 1, actual);
+    p += sizeof(uint64_t);
+  }
+}
+
+// Test that encoding routines generate little-endian encodings
+TEST(Coding, EncodingOutput) {
+  std::string dst;
+  PutFixed32(&dst, 0x04030201);
+  ASSERT_EQ(4U, dst.size());
+  ASSERT_EQ(0x01, static_cast<int>(dst[0]));
+  ASSERT_EQ(0x02, static_cast<int>(dst[1]));
+  ASSERT_EQ(0x03, static_cast<int>(dst[2]));
+  ASSERT_EQ(0x04, static_cast<int>(dst[3]));
+
+  dst.clear();
+  PutFixed64(&dst, 0x0807060504030201ull);
+  ASSERT_EQ(8U, dst.size());
+  ASSERT_EQ(0x01, static_cast<int>(dst[0]));
+  ASSERT_EQ(0x02, static_cast<int>(dst[1]));
+  ASSERT_EQ(0x03, static_cast<int>(dst[2]));
+  ASSERT_EQ(0x04, static_cast<int>(dst[3]));
+  ASSERT_EQ(0x05, static_cast<int>(dst[4]));
+  ASSERT_EQ(0x06, static_cast<int>(dst[5]));
+  ASSERT_EQ(0x07, static_cast<int>(dst[6]));
+  ASSERT_EQ(0x08, static_cast<int>(dst[7]));
+}
+
+TEST(Coding, Varint32) {
+  std::string s;
+  for (uint32_t i = 0; i < (32 * 32); i++) {
+    uint32_t v = (i / 32) << (i % 32);
+    PutVarint32(&s, v);
+  }
+
+  const char* p = s.data();
+  const char* limit = p + s.size();
+  for (uint32_t i = 0; i < (32 * 32); i++) {
+    uint32_t expected = (i / 32) << (i % 32);
+    uint32_t actual = 0;
+    const char* start = p;
+    p = GetVarint32Ptr(p, limit, &actual);
+    ASSERT_TRUE(p != nullptr);
+    ASSERT_EQ(expected, actual);
+    ASSERT_EQ(VarintLength(actual), p - start);
+  }
+  ASSERT_EQ(p, s.data() + s.size());
+}
+
+TEST(Coding, Varint64) {
+  // Construct the list of values to check
+  std::vector<uint64_t> values;
+  // Some special values
+  values.push_back(0);
+  values.push_back(100);
+  values.push_back(~static_cast<uint64_t>(0));
+  values.push_back(~static_cast<uint64_t>(0) - 1);
+  for (uint32_t k = 0; k < 64; k++) {
+    // Test values near powers of two
+    const uint64_t power = 1ull << k;
+    values.push_back(power);
+    values.push_back(power - 1);
+    values.push_back(power + 1);
+  };
+
+  std::string s;
+  for (unsigned int i = 0; i < values.size(); i++) {
+    PutVarint64(&s, values[i]);
+  }
+
+  const char* p = s.data();
+  const char* limit = p + s.size();
+  for (unsigned int i = 0; i < values.size(); i++) {
+    ASSERT_TRUE(p < limit);
+    uint64_t actual = 0;
+    const char* start = p;
+    p = GetVarint64Ptr(p, limit, &actual);
+    ASSERT_TRUE(p != nullptr);
+    ASSERT_EQ(values[i], actual);
+    ASSERT_EQ(VarintLength(actual), p - start);
+  }
+  ASSERT_EQ(p, limit);
+}
+
+TEST(Coding, Varint32Overflow) {
+  uint32_t result;
+  std::string input("\x81\x82\x83\x84\x85\x11");
+  ASSERT_TRUE(GetVarint32Ptr(input.data(), input.data() + input.size(),
+                             &result) == nullptr);
+}
+
+TEST(Coding, Varint32Truncation) {
+  uint32_t large_value = (1u << 31) + 100;
+  std::string s;
+  PutVarint32(&s, large_value);
+  uint32_t result;
+  for (unsigned int len = 0; len + 1 < s.size(); len++) {
+    ASSERT_TRUE(GetVarint32Ptr(s.data(), s.data() + len, &result) == nullptr);
+  }
+  ASSERT_TRUE(GetVarint32Ptr(s.data(), s.data() + s.size(), &result) !=
+              nullptr);
+  ASSERT_EQ(large_value, result);
+}
+
+TEST(Coding, Varint64Overflow) {
+  uint64_t result;
+  std::string input("\x81\x82\x83\x84\x85\x81\x82\x83\x84\x85\x11");
+  ASSERT_TRUE(GetVarint64Ptr(input.data(), input.data() + input.size(),
+                             &result) == nullptr);
+}
+
+TEST(Coding, Varint64Truncation) {
+  uint64_t large_value = (1ull << 63) + 100ull;
+  std::string s;
+  PutVarint64(&s, large_value);
+  uint64_t result;
+  for (unsigned int len = 0; len + 1 < s.size(); len++) {
+    ASSERT_TRUE(GetVarint64Ptr(s.data(), s.data() + len, &result) == nullptr);
+  }
+  ASSERT_TRUE(GetVarint64Ptr(s.data(), s.data() + s.size(), &result) !=
+              nullptr);
+  ASSERT_EQ(large_value, result);
+}
+
+TEST(Coding, Strings) {
+  std::string s;
+  PutLengthPrefixedSlice(&s, Slice(""));
+  PutLengthPrefixedSlice(&s, Slice("foo"));
+  PutLengthPrefixedSlice(&s, Slice("bar"));
+  PutLengthPrefixedSlice(&s, Slice(std::string(200, 'x')));
+
+  Slice input(s);
+  Slice v;
+  ASSERT_TRUE(GetLengthPrefixedSlice(&input, &v));
+  ASSERT_EQ("", v.ToString());
+  ASSERT_TRUE(GetLengthPrefixedSlice(&input, &v));
+  ASSERT_EQ("foo", v.ToString());
+  ASSERT_TRUE(GetLengthPrefixedSlice(&input, &v));
+  ASSERT_EQ("bar", v.ToString());
+  ASSERT_TRUE(GetLengthPrefixedSlice(&input, &v));
+  ASSERT_EQ(std::string(200, 'x'), v.ToString());
+  ASSERT_EQ("", input.ToString());
+}
+
+}  // namespace ROCKSDB_NAMESPACE
+
+int main(int argc, char** argv) {
+  ROCKSDB_NAMESPACE::port::InstallStackTraceHandler();
+  ::testing::InitGoogleTest(&argc, argv);
+  return RUN_ALL_TESTS();
+}
diff --git a/src/rocksdb/util/compaction_job_stats_impl.cc b/src/rocksdb/util/compaction_job_stats_impl.cc
new file mode 100644
index 000000000..cfab2a4fe
--- /dev/null
+++ b/src/rocksdb/util/compaction_job_stats_impl.cc
@@ -0,0 +1,100 @@
+// Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#include "rocksdb/compaction_job_stats.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+#ifndef ROCKSDB_LITE
+
+void CompactionJobStats::Reset() {
+  elapsed_micros = 0;
+  cpu_micros = 0;
+
+  num_input_records = 0;
+  num_blobs_read = 0;
+  num_input_files = 0;
+  num_input_files_at_output_level = 0;
+
+  num_output_records = 0;
+  num_output_files = 0;
+  num_output_files_blob = 0;
+
+  is_full_compaction = false;
+  is_manual_compaction = false;
+
+  total_input_bytes = 0;
+  total_blob_bytes_read = 0;
+  total_output_bytes = 0;
+  total_output_bytes_blob = 0;
+
+  num_records_replaced = 0;
+
+  total_input_raw_key_bytes = 0;
+  total_input_raw_value_bytes = 0;
+
+  num_input_deletion_records = 0;
+  num_expired_deletion_records = 0;
+
+  num_corrupt_keys = 0;
+
+  file_write_nanos = 0;
+  file_range_sync_nanos = 0;
+  file_fsync_nanos = 0;
+  file_prepare_write_nanos = 0;
+
+  smallest_output_key_prefix.clear();
+  largest_output_key_prefix.clear();
+
+  num_single_del_fallthru = 0;
+  num_single_del_mismatch = 0;
+}
+
+void CompactionJobStats::Add(const CompactionJobStats& stats) {
+  elapsed_micros += stats.elapsed_micros;
+  cpu_micros += stats.cpu_micros;
+
+  num_input_records += stats.num_input_records;
+  num_blobs_read += stats.num_blobs_read;
+  num_input_files += stats.num_input_files;
+  num_input_files_at_output_level += stats.num_input_files_at_output_level;
+
+  num_output_records += stats.num_output_records;
+  num_output_files += stats.num_output_files;
+  num_output_files_blob += stats.num_output_files_blob;
+
+  total_input_bytes += stats.total_input_bytes;
+  total_blob_bytes_read += stats.total_blob_bytes_read;
+  total_output_bytes += stats.total_output_bytes;
+  total_output_bytes_blob += stats.total_output_bytes_blob;
+
+  num_records_replaced += stats.num_records_replaced;
+
+  total_input_raw_key_bytes += stats.total_input_raw_key_bytes;
+  total_input_raw_value_bytes += stats.total_input_raw_value_bytes;
+
+  num_input_deletion_records += stats.num_input_deletion_records;
+  num_expired_deletion_records += stats.num_expired_deletion_records;
+
+  num_corrupt_keys += stats.num_corrupt_keys;
+
+  file_write_nanos += stats.file_write_nanos;
+  file_range_sync_nanos += stats.file_range_sync_nanos;
+  file_fsync_nanos += stats.file_fsync_nanos;
+  file_prepare_write_nanos += stats.file_prepare_write_nanos;
+
+  num_single_del_fallthru += stats.num_single_del_fallthru;
+  num_single_del_mismatch += stats.num_single_del_mismatch;
+}
+
+#else
+
+void CompactionJobStats::Reset() {}
+
+void CompactionJobStats::Add(const CompactionJobStats& /*stats*/) {}
+
+#endif  // !ROCKSDB_LITE
+
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/util/comparator.cc b/src/rocksdb/util/comparator.cc
new file mode 100644
index 000000000..f85ed69ee
--- /dev/null
+++ b/src/rocksdb/util/comparator.cc
@@ -0,0 +1,391 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+//
+// Copyright (c) 2011 The LevelDB Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file. See the AUTHORS file for names of contributors.
+
+#include "rocksdb/comparator.h"
+
+#include <stdint.h>
+
+#include <algorithm>
+#include <memory>
+#include <mutex>
+#include <sstream>
+
+#include "db/dbformat.h"
+#include "port/lang.h"
+#include "port/port.h"
+#include "rocksdb/convenience.h"
+#include "rocksdb/slice.h"
+#include "rocksdb/utilities/customizable_util.h"
+#include "rocksdb/utilities/object_registry.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+namespace {
+class BytewiseComparatorImpl : public Comparator {
+ public:
+  BytewiseComparatorImpl() {}
+  static const char* kClassName() { return "leveldb.BytewiseComparator"; }
+  const char* Name() const override { return kClassName(); }
+
+  int Compare(const Slice& a, const Slice& b) const override {
+    return a.compare(b);
+  }
+
+  bool Equal(const Slice& a, const Slice& b) const override { return a == b; }
+
+  void FindShortestSeparator(std::string* start,
+                             const Slice& limit) const override {
+    // Find length of common prefix
+    size_t min_length = std::min(start->size(), limit.size());
+    size_t diff_index = 0;
+    while ((diff_index < min_length) &&
+           ((*start)[diff_index] == limit[diff_index])) {
+      diff_index++;
+    }
+
+    if (diff_index >= min_length) {
+      // Do not shorten if one string is a prefix of the other
+    } else {
+      uint8_t start_byte = static_cast<uint8_t>((*start)[diff_index]);
+      uint8_t limit_byte = static_cast<uint8_t>(limit[diff_index]);
+      if (start_byte >= limit_byte) {
+        // Cannot shorten since limit is smaller than start or start is
+        // already the shortest possible.
+        return;
+      }
+      assert(start_byte < limit_byte);
+
+      if (diff_index < limit.size() - 1 || start_byte + 1 < limit_byte) {
+        (*start)[diff_index]++;
+        start->resize(diff_index + 1);
+      } else {
+        //     v
+        // A A 1 A A A
+        // A A 2
+        //
+        // Incrementing the current byte will make start bigger than limit, we
+        // will skip this byte, and find the first non 0xFF byte in start and
+        // increment it.
+        diff_index++;
+
+        while (diff_index < start->size()) {
+          // Keep moving until we find the first non 0xFF byte to
+          // increment it
+          if (static_cast<uint8_t>((*start)[diff_index]) <
+              static_cast<uint8_t>(0xff)) {
+            (*start)[diff_index]++;
+            start->resize(diff_index + 1);
+            break;
+          }
+          diff_index++;
+        }
+      }
+      assert(Compare(*start, limit) < 0);
+    }
+  }
+
+  void FindShortSuccessor(std::string* key) const override {
+    // Find first character that can be incremented
+    size_t n = key->size();
+    for (size_t i = 0; i < n; i++) {
+      const uint8_t byte = (*key)[i];
+      if (byte != static_cast<uint8_t>(0xff)) {
+        (*key)[i] = byte + 1;
+        key->resize(i + 1);
+        return;
+      }
+    }
+    // *key is a run of 0xffs.  Leave it alone.
+  }
+
+  bool IsSameLengthImmediateSuccessor(const Slice& s,
+                                      const Slice& t) const override {
+    if (s.size() != t.size() || s.size() == 0) {
+      return false;
+    }
+    size_t diff_ind = s.difference_offset(t);
+    // same slice
+    if (diff_ind >= s.size()) return false;
+    uint8_t byte_s = static_cast<uint8_t>(s[diff_ind]);
+    uint8_t byte_t = static_cast<uint8_t>(t[diff_ind]);
+    // first different byte must be consecutive, and remaining bytes must be
+    // 0xff for s and 0x00 for t
+    if (byte_s != uint8_t{0xff} && byte_s + 1 == byte_t) {
+      for (size_t i = diff_ind + 1; i < s.size(); ++i) {
+        byte_s = static_cast<uint8_t>(s[i]);
+        byte_t = static_cast<uint8_t>(t[i]);
+        if (byte_s != uint8_t{0xff} || byte_t != uint8_t{0x00}) {
+          return false;
+        }
+      }
+      return true;
+    } else {
+      return false;
+    }
+  }
+
+  bool CanKeysWithDifferentByteContentsBeEqual() const override {
+    return false;
+  }
+
+  using Comparator::CompareWithoutTimestamp;
+  int CompareWithoutTimestamp(const Slice& a, bool /*a_has_ts*/, const Slice& b,
+                              bool /*b_has_ts*/) const override {
+    return a.compare(b);
+  }
+
+  bool EqualWithoutTimestamp(const Slice& a, const Slice& b) const override {
+    return a == b;
+  }
+};
+
+class ReverseBytewiseComparatorImpl : public BytewiseComparatorImpl {
+ public:
+  ReverseBytewiseComparatorImpl() {}
+
+  static const char* kClassName() {
+    return "rocksdb.ReverseBytewiseComparator";
+  }
+  const char* Name() const override { return kClassName(); }
+
+  int Compare(const Slice& a, const Slice& b) const override {
+    return -a.compare(b);
+  }
+
+  void FindShortestSeparator(std::string* start,
+                             const Slice& limit) const override {
+    // Find length of common prefix
+    size_t min_length = std::min(start->size(), limit.size());
+    size_t diff_index = 0;
+    while ((diff_index < min_length) &&
+           ((*start)[diff_index] == limit[diff_index])) {
+      diff_index++;
+    }
+
+    assert(diff_index <= min_length);
+    if (diff_index == min_length) {
+      // Do not shorten if one string is a prefix of the other
+      //
+      // We could handle cases like:
+      //     V
+      // A A 2 X Y
+      // A A 2
+      // in a similar way as BytewiseComparator::FindShortestSeparator().
+      // We keep it simple by not implementing it. We can come back to it
+      // later when needed.
+    } else {
+      uint8_t start_byte = static_cast<uint8_t>((*start)[diff_index]);
+      uint8_t limit_byte = static_cast<uint8_t>(limit[diff_index]);
+      if (start_byte > limit_byte && diff_index < start->size() - 1) {
+        // Case like
+        //     V
+        // A A 3 A A
+        // A A 1 B B
+        //
+        // or
+        //     v
+        // A A 2 A A
+        // A A 1 B B
+        // In this case "AA2" will be good.
+#ifndef NDEBUG
+        std::string old_start = *start;
+#endif
+        start->resize(diff_index + 1);
+#ifndef NDEBUG
+        assert(old_start >= *start);
+#endif
+        assert(Slice(*start).compare(limit) > 0);
+      }
+    }
+  }
+
+  void FindShortSuccessor(std::string* /*key*/) const override {
+    // Don't do anything for simplicity.
+  }
+
+  bool IsSameLengthImmediateSuccessor(const Slice& s,
+                                      const Slice& t) const override {
+    // Always returning false to prevent surfacing design flaws in
+    // auto_prefix_mode
+    (void)s, (void)t;
+    return false;
+    // "Correct" implementation:
+    // return BytewiseComparatorImpl::IsSameLengthImmediateSuccessor(t, s);
+  }
+
+  bool CanKeysWithDifferentByteContentsBeEqual() const override {
+    return false;
+  }
+
+  using Comparator::CompareWithoutTimestamp;
+  int CompareWithoutTimestamp(const Slice& a, bool /*a_has_ts*/, const Slice& b,
+                              bool /*b_has_ts*/) const override {
+    return -a.compare(b);
+  }
+};
+
+// EXPERIMENTAL
+// Comparator with 64-bit integer timestamp.
+// We did not performance test this yet.
+template <typename TComparator>
+class ComparatorWithU64TsImpl : public Comparator {
+  static_assert(std::is_base_of<Comparator, TComparator>::value,
+                "template type must be a inherited type of comparator");
+
+ public:
+  explicit ComparatorWithU64TsImpl() : Comparator(/*ts_sz=*/sizeof(uint64_t)) {
+    assert(cmp_without_ts_.timestamp_size() == 0);
+  }
+
+  static const char* kClassName() {
+    static std::string class_name = kClassNameInternal();
+    return class_name.c_str();
+  }
+
+  const char* Name() const override { return kClassName(); }
+
+  void FindShortSuccessor(std::string*) const override {}
+  void FindShortestSeparator(std::string*, const Slice&) const override {}
+  int Compare(const Slice& a, const Slice& b) const override {
+    int ret = CompareWithoutTimestamp(a, b);
+    size_t ts_sz = timestamp_size();
+    if (ret != 0) {
+      return ret;
+    }
+    // Compare timestamp.
+    // For the same user key with different timestamps, larger (newer) timestamp
+    // comes first.
+    return -CompareTimestamp(ExtractTimestampFromUserKey(a, ts_sz),
+                             ExtractTimestampFromUserKey(b, ts_sz));
+  }
+  using Comparator::CompareWithoutTimestamp;
+  int CompareWithoutTimestamp(const Slice& a, bool a_has_ts, const Slice& b,
+                              bool b_has_ts) const override {
+    const size_t ts_sz = timestamp_size();
+    assert(!a_has_ts || a.size() >= ts_sz);
+    assert(!b_has_ts || b.size() >= ts_sz);
+    Slice lhs = a_has_ts ? StripTimestampFromUserKey(a, ts_sz) : a;
+    Slice rhs = b_has_ts ? StripTimestampFromUserKey(b, ts_sz) : b;
+    return cmp_without_ts_.Compare(lhs, rhs);
+  }
+  int CompareTimestamp(const Slice& ts1, const Slice& ts2) const override {
+    assert(ts1.size() == sizeof(uint64_t));
+    assert(ts2.size() == sizeof(uint64_t));
+    uint64_t lhs = DecodeFixed64(ts1.data());
+    uint64_t rhs = DecodeFixed64(ts2.data());
+    if (lhs < rhs) {
+      return -1;
+    } else if (lhs > rhs) {
+      return 1;
+    } else {
+      return 0;
+    }
+  }
+
+ private:
+  static std::string kClassNameInternal() {
+    std::stringstream ss;
+    ss << TComparator::kClassName() << ".u64ts";
+    return ss.str();
+  }
+
+  TComparator cmp_without_ts_;
+};
+
+}  // namespace
+
+const Comparator* BytewiseComparator() {
+  STATIC_AVOID_DESTRUCTION(BytewiseComparatorImpl, bytewise);
+  return &bytewise;
+}
+
+const Comparator* ReverseBytewiseComparator() {
+  STATIC_AVOID_DESTRUCTION(ReverseBytewiseComparatorImpl, rbytewise);
+  return &rbytewise;
+}
+
+const Comparator* BytewiseComparatorWithU64Ts() {
+  STATIC_AVOID_DESTRUCTION(ComparatorWithU64TsImpl<BytewiseComparatorImpl>,
+                           comp_with_u64_ts);
+  return &comp_with_u64_ts;
+}
+
+#ifndef ROCKSDB_LITE
+static int RegisterBuiltinComparators(ObjectLibrary& library,
+                                      const std::string& /*arg*/) {
+  library.AddFactory<const Comparator>(
+      BytewiseComparatorImpl::kClassName(),
+      [](const std::string& /*uri*/,
+         std::unique_ptr<const Comparator>* /*guard */,
+         std::string* /* errmsg */) { return BytewiseComparator(); });
+  library.AddFactory<const Comparator>(
+      ReverseBytewiseComparatorImpl::kClassName(),
+      [](const std::string& /*uri*/,
+         std::unique_ptr<const Comparator>* /*guard */,
+         std::string* /* errmsg */) { return ReverseBytewiseComparator(); });
+  library.AddFactory<const Comparator>(
+      ComparatorWithU64TsImpl<BytewiseComparatorImpl>::kClassName(),
+      [](const std::string& /*uri*/,
+         std::unique_ptr<const Comparator>* /*guard */,
+         std::string* /* errmsg */) { return BytewiseComparatorWithU64Ts(); });
+  return 3;
+}
+#endif  // ROCKSDB_LITE
+
+Status Comparator::CreateFromString(const ConfigOptions& config_options,
+                                    const std::string& value,
+                                    const Comparator** result) {
+#ifndef ROCKSDB_LITE
+  static std::once_flag once;
+  std::call_once(once, [&]() {
+    RegisterBuiltinComparators(*(ObjectLibrary::Default().get()), "");
+  });
+#endif  // ROCKSDB_LITE
+  std::string id;
+  std::unordered_map<std::string, std::string> opt_map;
+  Status status = Customizable::GetOptionsMap(config_options, *result, value,
+                                              &id, &opt_map);
+  if (!status.ok()) {  // GetOptionsMap failed
+    return status;
+  }
+  if (id == BytewiseComparatorImpl::kClassName()) {
+    *result = BytewiseComparator();
+  } else if (id == ReverseBytewiseComparatorImpl::kClassName()) {
+    *result = ReverseBytewiseComparator();
+  } else if (id ==
+             ComparatorWithU64TsImpl<BytewiseComparatorImpl>::kClassName()) {
+    *result = BytewiseComparatorWithU64Ts();
+  } else if (value.empty()) {
+    // No Id and no options.  Clear the object
+    *result = nullptr;
+    return Status::OK();
+  } else if (id.empty()) {  // We have no Id but have options.  Not good
+    return Status::NotSupported("Cannot reset object ", id);
+  } else {
+#ifndef ROCKSDB_LITE
+    status = config_options.registry->NewStaticObject(id, result);
+#else
+    status = Status::NotSupported("Cannot load object in LITE mode ", id);
+#endif  // ROCKSDB_LITE
+    if (!status.ok()) {
+      if (config_options.ignore_unsupported_options &&
+          status.IsNotSupported()) {
+        return Status::OK();
+      } else {
+        return status;
+      }
+    } else {
+      Comparator* comparator = const_cast<Comparator*>(*result);
+      status =
+          Customizable::ConfigureNewObject(config_options, comparator, opt_map);
+    }
+  }
+  return status;
+}
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/util/compression.cc b/src/rocksdb/util/compression.cc
new file mode 100644
index 000000000..8e2f01b12
--- /dev/null
+++ b/src/rocksdb/util/compression.cc
@@ -0,0 +1,122 @@
+// Copyright (c) 2022-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#include "util/compression.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+StreamingCompress* StreamingCompress::Create(CompressionType compression_type,
+                                             const CompressionOptions& opts,
+                                             uint32_t compress_format_version,
+                                             size_t max_output_len) {
+  switch (compression_type) {
+    case kZSTD: {
+      if (!ZSTD_Streaming_Supported()) {
+        return nullptr;
+      }
+      return new ZSTDStreamingCompress(opts, compress_format_version,
+                                       max_output_len);
+    }
+    default:
+      return nullptr;
+  }
+}
+
+StreamingUncompress* StreamingUncompress::Create(
+    CompressionType compression_type, uint32_t compress_format_version,
+    size_t max_output_len) {
+  switch (compression_type) {
+    case kZSTD: {
+      if (!ZSTD_Streaming_Supported()) {
+        return nullptr;
+      }
+      return new ZSTDStreamingUncompress(compress_format_version,
+                                         max_output_len);
+    }
+    default:
+      return nullptr;
+  }
+}
+
+int ZSTDStreamingCompress::Compress(const char* input, size_t input_size,
+                                    char* output, size_t* output_pos) {
+  assert(input != nullptr && output != nullptr && output_pos != nullptr);
+  *output_pos = 0;
+  // Don't need to compress an empty input
+  if (input_size == 0) {
+    return 0;
+  }
+#ifndef ZSTD_STREAMING
+  (void)input;
+  (void)input_size;
+  (void)output;
+  return -1;
+#else
+  if (input_buffer_.src == nullptr || input_buffer_.src != input) {
+    // New input
+    // Catch errors where the previous input was not fully decompressed.
+    assert(input_buffer_.pos == input_buffer_.size);
+    input_buffer_ = {input, input_size, /*pos=*/0};
+  } else if (input_buffer_.src == input) {
+    // Same input, not fully compressed.
+  }
+  ZSTD_outBuffer output_buffer = {output, max_output_len_, /*pos=*/0};
+  const size_t remaining =
+      ZSTD_compressStream2(cctx_, &output_buffer, &input_buffer_, ZSTD_e_end);
+  if (ZSTD_isError(remaining)) {
+    // Failure
+    Reset();
+    return -1;
+  }
+  // Success
+  *output_pos = output_buffer.pos;
+  return (int)remaining;
+#endif
+}
+
+void ZSTDStreamingCompress::Reset() {
+#ifdef ZSTD_STREAMING
+  ZSTD_CCtx_reset(cctx_, ZSTD_ResetDirective::ZSTD_reset_session_only);
+  input_buffer_ = {/*src=*/nullptr, /*size=*/0, /*pos=*/0};
+#endif
+}
+
+int ZSTDStreamingUncompress::Uncompress(const char* input, size_t input_size,
+                                        char* output, size_t* output_pos) {
+  assert(input != nullptr && output != nullptr && output_pos != nullptr);
+  *output_pos = 0;
+  // Don't need to uncompress an empty input
+  if (input_size == 0) {
+    return 0;
+  }
+#ifdef ZSTD_STREAMING
+  if (input_buffer_.src != input) {
+    // New input
+    input_buffer_ = {input, input_size, /*pos=*/0};
+  }
+  ZSTD_outBuffer output_buffer = {output, max_output_len_, /*pos=*/0};
+  size_t ret = ZSTD_decompressStream(dctx_, &output_buffer, &input_buffer_);
+  if (ZSTD_isError(ret)) {
+    Reset();
+    return -1;
+  }
+  *output_pos = output_buffer.pos;
+  return (int)(input_buffer_.size - input_buffer_.pos);
+#else
+  (void)input;
+  (void)input_size;
+  (void)output;
+  return -1;
+#endif
+}
+
+void ZSTDStreamingUncompress::Reset() {
+#ifdef ZSTD_STREAMING
+  ZSTD_DCtx_reset(dctx_, ZSTD_ResetDirective::ZSTD_reset_session_only);
+  input_buffer_ = {/*src=*/nullptr, /*size=*/0, /*pos=*/0};
+#endif
+}
+
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/util/compression.h b/src/rocksdb/util/compression.h
new file mode 100644
index 000000000..0d4febcfb
--- /dev/null
+++ b/src/rocksdb/util/compression.h
@@ -0,0 +1,1786 @@
+// Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+//
+// Copyright (c) 2011 The LevelDB Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file. See the AUTHORS file for names of contributors.
+//
+#pragma once
+
+#include <algorithm>
+#include <limits>
+#ifdef ROCKSDB_MALLOC_USABLE_SIZE
+#ifdef OS_FREEBSD
+#include <malloc_np.h>
+#else  // OS_FREEBSD
+#include <malloc.h>
+#endif  // OS_FREEBSD
+#endif  // ROCKSDB_MALLOC_USABLE_SIZE
+#include <string>
+
+#include "memory/memory_allocator.h"
+#include "rocksdb/options.h"
+#include "rocksdb/table.h"
+#include "test_util/sync_point.h"
+#include "util/coding.h"
+#include "util/compression_context_cache.h"
+#include "util/string_util.h"
+
+#ifdef SNAPPY
+#include <snappy.h>
+#endif
+
+#ifdef ZLIB
+#include <zlib.h>
+#endif
+
+#ifdef BZIP2
+#include <bzlib.h>
+#endif
+
+#if defined(LZ4)
+#include <lz4.h>
+#include <lz4hc.h>
+#endif
+
+#if defined(ZSTD)
+#include <zstd.h>
+// v1.1.3+
+#if ZSTD_VERSION_NUMBER >= 10103
+#include <zdict.h>
+#endif  // ZSTD_VERSION_NUMBER >= 10103
+// v1.4.0+
+#if ZSTD_VERSION_NUMBER >= 10400
+#define ZSTD_STREAMING
+#endif  // ZSTD_VERSION_NUMBER >= 10400
+namespace ROCKSDB_NAMESPACE {
+// Need this for the context allocation override
+// On windows we need to do this explicitly
+#if (ZSTD_VERSION_NUMBER >= 500)
+#if defined(ROCKSDB_JEMALLOC) && defined(OS_WIN) && \
+    defined(ZSTD_STATIC_LINKING_ONLY)
+#define ROCKSDB_ZSTD_CUSTOM_MEM
+namespace port {
+ZSTD_customMem GetJeZstdAllocationOverrides();
+}  // namespace port
+#endif  // defined(ROCKSDB_JEMALLOC) && defined(OS_WIN) &&
+        // defined(ZSTD_STATIC_LINKING_ONLY)
+
+// We require `ZSTD_sizeof_DDict` and `ZSTD_createDDict_byReference` to use
+// `ZSTD_DDict`. The former was introduced in v1.0.0 and the latter was
+// introduced in v1.1.3. But an important bug fix for `ZSTD_sizeof_DDict` came
+// in v1.1.4, so that is the version we require. As of today's latest version
+// (v1.3.8), they are both still in the experimental API, which means they are
+// only exported when the compiler flag `ZSTD_STATIC_LINKING_ONLY` is set.
+#if defined(ZSTD_STATIC_LINKING_ONLY) && ZSTD_VERSION_NUMBER >= 10104
+#define ROCKSDB_ZSTD_DDICT
+#endif  // defined(ZSTD_STATIC_LINKING_ONLY) && ZSTD_VERSION_NUMBER >= 10104
+
+// Cached data represents a portion that can be re-used
+// If, in the future we have more than one native context to
+// cache we can arrange this as a tuple
+class ZSTDUncompressCachedData {
+ public:
+  using ZSTDNativeContext = ZSTD_DCtx*;
+  ZSTDUncompressCachedData() {}
+  // Init from cache
+  ZSTDUncompressCachedData(const ZSTDUncompressCachedData& o) = delete;
+  ZSTDUncompressCachedData& operator=(const ZSTDUncompressCachedData&) = delete;
+  ZSTDUncompressCachedData(ZSTDUncompressCachedData&& o) noexcept
+      : ZSTDUncompressCachedData() {
+    *this = std::move(o);
+  }
+  ZSTDUncompressCachedData& operator=(ZSTDUncompressCachedData&& o) noexcept {
+    assert(zstd_ctx_ == nullptr);
+    std::swap(zstd_ctx_, o.zstd_ctx_);
+    std::swap(cache_idx_, o.cache_idx_);
+    return *this;
+  }
+  ZSTDNativeContext Get() const { return zstd_ctx_; }
+  int64_t GetCacheIndex() const { return cache_idx_; }
+  void CreateIfNeeded() {
+    if (zstd_ctx_ == nullptr) {
+#ifdef ROCKSDB_ZSTD_CUSTOM_MEM
+      zstd_ctx_ =
+          ZSTD_createDCtx_advanced(port::GetJeZstdAllocationOverrides());
+#else   // ROCKSDB_ZSTD_CUSTOM_MEM
+      zstd_ctx_ = ZSTD_createDCtx();
+#endif  // ROCKSDB_ZSTD_CUSTOM_MEM
+      cache_idx_ = -1;
+    }
+  }
+  void InitFromCache(const ZSTDUncompressCachedData& o, int64_t idx) {
+    zstd_ctx_ = o.zstd_ctx_;
+    cache_idx_ = idx;
+  }
+  ~ZSTDUncompressCachedData() {
+    if (zstd_ctx_ != nullptr && cache_idx_ == -1) {
+      ZSTD_freeDCtx(zstd_ctx_);
+    }
+  }
+
+ private:
+  ZSTDNativeContext zstd_ctx_ = nullptr;
+  int64_t cache_idx_ = -1;  // -1 means this instance owns the context
+};
+#endif  // (ZSTD_VERSION_NUMBER >= 500)
+}  // namespace ROCKSDB_NAMESPACE
+#endif  // ZSTD
+
+#if !(defined ZSTD) || !(ZSTD_VERSION_NUMBER >= 500)
+namespace ROCKSDB_NAMESPACE {
+class ZSTDUncompressCachedData {
+  void* padding;  // unused
+ public:
+  using ZSTDNativeContext = void*;
+  ZSTDUncompressCachedData() {}
+  ZSTDUncompressCachedData(const ZSTDUncompressCachedData&) {}
+  ZSTDUncompressCachedData& operator=(const ZSTDUncompressCachedData&) = delete;
+  ZSTDUncompressCachedData(ZSTDUncompressCachedData&&) noexcept = default;
+  ZSTDUncompressCachedData& operator=(ZSTDUncompressCachedData&&) noexcept =
+      default;
+  ZSTDNativeContext Get() const { return nullptr; }
+  int64_t GetCacheIndex() const { return -1; }
+  void CreateIfNeeded() {}
+  void InitFromCache(const ZSTDUncompressCachedData&, int64_t) {}
+
+ private:
+  void ignore_padding__() { padding = nullptr; }
+};
+}  // namespace ROCKSDB_NAMESPACE
+#endif
+
+#if defined(XPRESS)
+#include "port/xpress.h"
+#endif
+
+namespace ROCKSDB_NAMESPACE {
+
+// Holds dictionary and related data, like ZSTD's digested compression
+// dictionary.
+struct CompressionDict {
+#if ZSTD_VERSION_NUMBER >= 700
+  ZSTD_CDict* zstd_cdict_ = nullptr;
+#endif  // ZSTD_VERSION_NUMBER >= 700
+  std::string dict_;
+
+ public:
+#if ZSTD_VERSION_NUMBER >= 700
+  CompressionDict(std::string dict, CompressionType type, int level) {
+#else   // ZSTD_VERSION_NUMBER >= 700
+  CompressionDict(std::string dict, CompressionType /*type*/, int /*level*/) {
+#endif  // ZSTD_VERSION_NUMBER >= 700
+    dict_ = std::move(dict);
+#if ZSTD_VERSION_NUMBER >= 700
+    zstd_cdict_ = nullptr;
+    if (!dict_.empty() && (type == kZSTD || type == kZSTDNotFinalCompression)) {
+      if (level == CompressionOptions::kDefaultCompressionLevel) {
+        // 3 is the value of ZSTD_CLEVEL_DEFAULT (not exposed publicly), see
+        // https://github.com/facebook/zstd/issues/1148
+        level = 3;
+      }
+      // Should be safe (but slower) if below call fails as we'll use the
+      // raw dictionary to compress.
+      zstd_cdict_ = ZSTD_createCDict(dict_.data(), dict_.size(), level);
+      assert(zstd_cdict_ != nullptr);
+    }
+#endif  // ZSTD_VERSION_NUMBER >= 700
+  }
+
+  ~CompressionDict() {
+#if ZSTD_VERSION_NUMBER >= 700
+    size_t res = 0;
+    if (zstd_cdict_ != nullptr) {
+      res = ZSTD_freeCDict(zstd_cdict_);
+    }
+    assert(res == 0);  // Last I checked they can't fail
+    (void)res;         // prevent unused var warning
+#endif                 // ZSTD_VERSION_NUMBER >= 700
+  }
+
+#if ZSTD_VERSION_NUMBER >= 700
+  const ZSTD_CDict* GetDigestedZstdCDict() const { return zstd_cdict_; }
+#endif  // ZSTD_VERSION_NUMBER >= 700
+
+  Slice GetRawDict() const { return dict_; }
+
+  static const CompressionDict& GetEmptyDict() {
+    static CompressionDict empty_dict{};
+    return empty_dict;
+  }
+
+  CompressionDict() = default;
+  // Disable copy/move
+  CompressionDict(const CompressionDict&) = delete;
+  CompressionDict& operator=(const CompressionDict&) = delete;
+  CompressionDict(CompressionDict&&) = delete;
+  CompressionDict& operator=(CompressionDict&&) = delete;
+};
+
+// Holds dictionary and related data, like ZSTD's digested uncompression
+// dictionary.
+struct UncompressionDict {
+  // Block containing the data for the compression dictionary in case the
+  // constructor that takes a string parameter is used.
+  std::string dict_;
+
+  // Block containing the data for the compression dictionary in case the
+  // constructor that takes a Slice parameter is used and the passed in
+  // CacheAllocationPtr is not nullptr.
+  CacheAllocationPtr allocation_;
+
+  // Slice pointing to the compression dictionary data. Can point to
+  // dict_, allocation_, or some other memory location, depending on how
+  // the object was constructed.
+  Slice slice_;
+
+#ifdef ROCKSDB_ZSTD_DDICT
+  // Processed version of the contents of slice_ for ZSTD compression.
+  ZSTD_DDict* zstd_ddict_ = nullptr;
+#endif  // ROCKSDB_ZSTD_DDICT
+
+#ifdef ROCKSDB_ZSTD_DDICT
+  UncompressionDict(std::string dict, bool using_zstd)
+#else   // ROCKSDB_ZSTD_DDICT
+  UncompressionDict(std::string dict, bool /* using_zstd */)
+#endif  // ROCKSDB_ZSTD_DDICT
+      : dict_(std::move(dict)), slice_(dict_) {
+#ifdef ROCKSDB_ZSTD_DDICT
+    if (!slice_.empty() && using_zstd) {
+      zstd_ddict_ = ZSTD_createDDict_byReference(slice_.data(), slice_.size());
+      assert(zstd_ddict_ != nullptr);
+    }
+#endif  // ROCKSDB_ZSTD_DDICT
+  }
+
+#ifdef ROCKSDB_ZSTD_DDICT
+  UncompressionDict(Slice slice, CacheAllocationPtr&& allocation,
+                    bool using_zstd)
+#else   // ROCKSDB_ZSTD_DDICT
+  UncompressionDict(Slice slice, CacheAllocationPtr&& allocation,
+                    bool /* using_zstd */)
+#endif  // ROCKSDB_ZSTD_DDICT
+      : allocation_(std::move(allocation)), slice_(std::move(slice)) {
+#ifdef ROCKSDB_ZSTD_DDICT
+    if (!slice_.empty() && using_zstd) {
+      zstd_ddict_ = ZSTD_createDDict_byReference(slice_.data(), slice_.size());
+      assert(zstd_ddict_ != nullptr);
+    }
+#endif  // ROCKSDB_ZSTD_DDICT
+  }
+
+  UncompressionDict(UncompressionDict&& rhs)
+      : dict_(std::move(rhs.dict_)),
+        allocation_(std::move(rhs.allocation_)),
+        slice_(std::move(rhs.slice_))
+#ifdef ROCKSDB_ZSTD_DDICT
+        ,
+        zstd_ddict_(rhs.zstd_ddict_)
+#endif
+  {
+#ifdef ROCKSDB_ZSTD_DDICT
+    rhs.zstd_ddict_ = nullptr;
+#endif
+  }
+
+  ~UncompressionDict() {
+#ifdef ROCKSDB_ZSTD_DDICT
+    size_t res = 0;
+    if (zstd_ddict_ != nullptr) {
+      res = ZSTD_freeDDict(zstd_ddict_);
+    }
+    assert(res == 0);  // Last I checked they can't fail
+    (void)res;         // prevent unused var warning
+#endif                 // ROCKSDB_ZSTD_DDICT
+  }
+
+  UncompressionDict& operator=(UncompressionDict&& rhs) {
+    if (this == &rhs) {
+      return *this;
+    }
+
+    dict_ = std::move(rhs.dict_);
+    allocation_ = std::move(rhs.allocation_);
+    slice_ = std::move(rhs.slice_);
+
+#ifdef ROCKSDB_ZSTD_DDICT
+    zstd_ddict_ = rhs.zstd_ddict_;
+    rhs.zstd_ddict_ = nullptr;
+#endif
+
+    return *this;
+  }
+
+  // The object is self-contained if the string constructor is used, or the
+  // Slice constructor is invoked with a non-null allocation. Otherwise, it
+  // is the caller's responsibility to ensure that the underlying storage
+  // outlives this object.
+  bool own_bytes() const { return !dict_.empty() || allocation_; }
+
+  const Slice& GetRawDict() const { return slice_; }
+
+#ifdef ROCKSDB_ZSTD_DDICT
+  const ZSTD_DDict* GetDigestedZstdDDict() const { return zstd_ddict_; }
+#endif  // ROCKSDB_ZSTD_DDICT
+
+  static const UncompressionDict& GetEmptyDict() {
+    static UncompressionDict empty_dict{};
+    return empty_dict;
+  }
+
+  size_t ApproximateMemoryUsage() const {
+    size_t usage = sizeof(struct UncompressionDict);
+    usage += dict_.size();
+    if (allocation_) {
+      auto allocator = allocation_.get_deleter().allocator;
+      if (allocator) {
+        usage += allocator->UsableSize(allocation_.get(), slice_.size());
+      } else {
+        usage += slice_.size();
+      }
+    }
+#ifdef ROCKSDB_ZSTD_DDICT
+    usage += ZSTD_sizeof_DDict(zstd_ddict_);
+#endif  // ROCKSDB_ZSTD_DDICT
+    return usage;
+  }
+
+  UncompressionDict() = default;
+  // Disable copy
+  UncompressionDict(const CompressionDict&) = delete;
+  UncompressionDict& operator=(const CompressionDict&) = delete;
+};
+
+class CompressionContext {
+ private:
+#if defined(ZSTD) && (ZSTD_VERSION_NUMBER >= 500)
+  ZSTD_CCtx* zstd_ctx_ = nullptr;
+  void CreateNativeContext(CompressionType type) {
+    if (type == kZSTD || type == kZSTDNotFinalCompression) {
+#ifdef ROCKSDB_ZSTD_CUSTOM_MEM
+      zstd_ctx_ =
+          ZSTD_createCCtx_advanced(port::GetJeZstdAllocationOverrides());
+#else   // ROCKSDB_ZSTD_CUSTOM_MEM
+      zstd_ctx_ = ZSTD_createCCtx();
+#endif  // ROCKSDB_ZSTD_CUSTOM_MEM
+    }
+  }
+  void DestroyNativeContext() {
+    if (zstd_ctx_ != nullptr) {
+      ZSTD_freeCCtx(zstd_ctx_);
+    }
+  }
+
+ public:
+  // callable inside ZSTD_Compress
+  ZSTD_CCtx* ZSTDPreallocCtx() const {
+    assert(zstd_ctx_ != nullptr);
+    return zstd_ctx_;
+  }
+
+#else   // ZSTD && (ZSTD_VERSION_NUMBER >= 500)
+ private:
+  void CreateNativeContext(CompressionType /* type */) {}
+  void DestroyNativeContext() {}
+#endif  // ZSTD && (ZSTD_VERSION_NUMBER >= 500)
+ public:
+  explicit CompressionContext(CompressionType type) {
+    CreateNativeContext(type);
+  }
+  ~CompressionContext() { DestroyNativeContext(); }
+  CompressionContext(const CompressionContext&) = delete;
+  CompressionContext& operator=(const CompressionContext&) = delete;
+};
+
+class CompressionInfo {
+  const CompressionOptions& opts_;
+  const CompressionContext& context_;
+  const CompressionDict& dict_;
+  const CompressionType type_;
+  const uint64_t sample_for_compression_;
+
+ public:
+  CompressionInfo(const CompressionOptions& _opts,
+                  const CompressionContext& _context,
+                  const CompressionDict& _dict, CompressionType _type,
+                  uint64_t _sample_for_compression)
+      : opts_(_opts),
+        context_(_context),
+        dict_(_dict),
+        type_(_type),
+        sample_for_compression_(_sample_for_compression) {}
+
+  const CompressionOptions& options() const { return opts_; }
+  const CompressionContext& context() const { return context_; }
+  const CompressionDict& dict() const { return dict_; }
+  CompressionType type() const { return type_; }
+  uint64_t SampleForCompression() const { return sample_for_compression_; }
+};
+
+class UncompressionContext {
+ private:
+  CompressionContextCache* ctx_cache_ = nullptr;
+  ZSTDUncompressCachedData uncomp_cached_data_;
+
+ public:
+  explicit UncompressionContext(CompressionType type) {
+    if (type == kZSTD || type == kZSTDNotFinalCompression) {
+      ctx_cache_ = CompressionContextCache::Instance();
+      uncomp_cached_data_ = ctx_cache_->GetCachedZSTDUncompressData();
+    }
+  }
+  ~UncompressionContext() {
+    if (uncomp_cached_data_.GetCacheIndex() != -1) {
+      assert(ctx_cache_ != nullptr);
+      ctx_cache_->ReturnCachedZSTDUncompressData(
+          uncomp_cached_data_.GetCacheIndex());
+    }
+  }
+  UncompressionContext(const UncompressionContext&) = delete;
+  UncompressionContext& operator=(const UncompressionContext&) = delete;
+
+  ZSTDUncompressCachedData::ZSTDNativeContext GetZSTDContext() const {
+    return uncomp_cached_data_.Get();
+  }
+};
+
+class UncompressionInfo {
+  const UncompressionContext& context_;
+  const UncompressionDict& dict_;
+  const CompressionType type_;
+
+ public:
+  UncompressionInfo(const UncompressionContext& _context,
+                    const UncompressionDict& _dict, CompressionType _type)
+      : context_(_context), dict_(_dict), type_(_type) {}
+
+  const UncompressionContext& context() const { return context_; }
+  const UncompressionDict& dict() const { return dict_; }
+  CompressionType type() const { return type_; }
+};
+
+inline bool Snappy_Supported() {
+#ifdef SNAPPY
+  return true;
+#else
+  return false;
+#endif
+}
+
+inline bool Zlib_Supported() {
+#ifdef ZLIB
+  return true;
+#else
+  return false;
+#endif
+}
+
+inline bool BZip2_Supported() {
+#ifdef BZIP2
+  return true;
+#else
+  return false;
+#endif
+}
+
+inline bool LZ4_Supported() {
+#ifdef LZ4
+  return true;
+#else
+  return false;
+#endif
+}
+
+inline bool XPRESS_Supported() {
+#ifdef XPRESS
+  return true;
+#else
+  return false;
+#endif
+}
+
+inline bool ZSTD_Supported() {
+#ifdef ZSTD
+  // ZSTD format is finalized since version 0.8.0.
+  return (ZSTD_versionNumber() >= 800);
+#else
+  return false;
+#endif
+}
+
+inline bool ZSTDNotFinal_Supported() {
+#ifdef ZSTD
+  return true;
+#else
+  return false;
+#endif
+}
+
+inline bool ZSTD_Streaming_Supported() {
+#if defined(ZSTD) && defined(ZSTD_STREAMING)
+  return true;
+#else
+  return false;
+#endif
+}
+
+inline bool StreamingCompressionTypeSupported(
+    CompressionType compression_type) {
+  switch (compression_type) {
+    case kNoCompression:
+      return true;
+    case kZSTD:
+      return ZSTD_Streaming_Supported();
+    default:
+      return false;
+  }
+}
+
+inline bool CompressionTypeSupported(CompressionType compression_type) {
+  switch (compression_type) {
+    case kNoCompression:
+      return true;
+    case kSnappyCompression:
+      return Snappy_Supported();
+    case kZlibCompression:
+      return Zlib_Supported();
+    case kBZip2Compression:
+      return BZip2_Supported();
+    case kLZ4Compression:
+      return LZ4_Supported();
+    case kLZ4HCCompression:
+      return LZ4_Supported();
+    case kXpressCompression:
+      return XPRESS_Supported();
+    case kZSTDNotFinalCompression:
+      return ZSTDNotFinal_Supported();
+    case kZSTD:
+      return ZSTD_Supported();
+    default:
+      assert(false);
+      return false;
+  }
+}
+
+inline bool DictCompressionTypeSupported(CompressionType compression_type) {
+  switch (compression_type) {
+    case kNoCompression:
+      return false;
+    case kSnappyCompression:
+      return false;
+    case kZlibCompression:
+      return Zlib_Supported();
+    case kBZip2Compression:
+      return false;
+    case kLZ4Compression:
+    case kLZ4HCCompression:
+#if LZ4_VERSION_NUMBER >= 10400  // r124+
+      return LZ4_Supported();
+#else
+      return false;
+#endif
+    case kXpressCompression:
+      return false;
+    case kZSTDNotFinalCompression:
+#if ZSTD_VERSION_NUMBER >= 500  // v0.5.0+
+      return ZSTDNotFinal_Supported();
+#else
+      return false;
+#endif
+    case kZSTD:
+#if ZSTD_VERSION_NUMBER >= 500  // v0.5.0+
+      return ZSTD_Supported();
+#else
+      return false;
+#endif
+    default:
+      assert(false);
+      return false;
+  }
+}
+
+inline std::string CompressionTypeToString(CompressionType compression_type) {
+  switch (compression_type) {
+    case kNoCompression:
+      return "NoCompression";
+    case kSnappyCompression:
+      return "Snappy";
+    case kZlibCompression:
+      return "Zlib";
+    case kBZip2Compression:
+      return "BZip2";
+    case kLZ4Compression:
+      return "LZ4";
+    case kLZ4HCCompression:
+      return "LZ4HC";
+    case kXpressCompression:
+      return "Xpress";
+    case kZSTD:
+      return "ZSTD";
+    case kZSTDNotFinalCompression:
+      return "ZSTDNotFinal";
+    case kDisableCompressionOption:
+      return "DisableOption";
+    default:
+      assert(false);
+      return "";
+  }
+}
+
+inline std::string CompressionOptionsToString(
+    CompressionOptions& compression_options) {
+  std::string result;
+  result.reserve(512);
+  result.append("window_bits=")
+      .append(std::to_string(compression_options.window_bits))
+      .append("; ");
+  result.append("level=")
+      .append(std::to_string(compression_options.level))
+      .append("; ");
+  result.append("strategy=")
+      .append(std::to_string(compression_options.strategy))
+      .append("; ");
+  result.append("max_dict_bytes=")
+      .append(std::to_string(compression_options.max_dict_bytes))
+      .append("; ");
+  result.append("zstd_max_train_bytes=")
+      .append(std::to_string(compression_options.zstd_max_train_bytes))
+      .append("; ");
+  result.append("enabled=")
+      .append(std::to_string(compression_options.enabled))
+      .append("; ");
+  result.append("max_dict_buffer_bytes=")
+      .append(std::to_string(compression_options.max_dict_buffer_bytes))
+      .append("; ");
+  result.append("use_zstd_dict_trainer=")
+      .append(std::to_string(compression_options.use_zstd_dict_trainer))
+      .append("; ");
+  return result;
+}
+
+// compress_format_version can have two values:
+// 1 -- decompressed sizes for BZip2 and Zlib are not included in the compressed
+// block. Also, decompressed sizes for LZ4 are encoded in platform-dependent
+// way.
+// 2 -- Zlib, BZip2 and LZ4 encode decompressed size as Varint32 just before the
+// start of compressed block. Snappy format is the same as version 1.
+
+inline bool Snappy_Compress(const CompressionInfo& /*info*/, const char* input,
+                            size_t length, ::std::string* output) {
+#ifdef SNAPPY
+  output->resize(snappy::MaxCompressedLength(length));
+  size_t outlen;
+  snappy::RawCompress(input, length, &(*output)[0], &outlen);
+  output->resize(outlen);
+  return true;
+#else
+  (void)input;
+  (void)length;
+  (void)output;
+  return false;
+#endif
+}
+
+inline CacheAllocationPtr Snappy_Uncompress(
+    const char* input, size_t length, size_t* uncompressed_size,
+    MemoryAllocator* allocator = nullptr) {
+#ifdef SNAPPY
+  size_t uncompressed_length = 0;
+  if (!snappy::GetUncompressedLength(input, length, &uncompressed_length)) {
+    return nullptr;
+  }
+
+  CacheAllocationPtr output = AllocateBlock(uncompressed_length, allocator);
+
+  if (!snappy::RawUncompress(input, length, output.get())) {
+    return nullptr;
+  }
+
+  *uncompressed_size = uncompressed_length;
+
+  return output;
+#else
+  (void)input;
+  (void)length;
+  (void)uncompressed_size;
+  (void)allocator;
+  return nullptr;
+#endif
+}
+
+namespace compression {
+// returns size
+inline size_t PutDecompressedSizeInfo(std::string* output, uint32_t length) {
+  PutVarint32(output, length);
+  return output->size();
+}
+
+inline bool GetDecompressedSizeInfo(const char** input_data,
+                                    size_t* input_length,
+                                    uint32_t* output_len) {
+  auto new_input_data =
+      GetVarint32Ptr(*input_data, *input_data + *input_length, output_len);
+  if (new_input_data == nullptr) {
+    return false;
+  }
+  *input_length -= (new_input_data - *input_data);
+  *input_data = new_input_data;
+  return true;
+}
+}  // namespace compression
+
+// compress_format_version == 1 -- decompressed size is not included in the
+// block header
+// compress_format_version == 2 -- decompressed size is included in the block
+// header in varint32 format
+// @param compression_dict Data for presetting the compression library's
+//    dictionary.
+inline bool Zlib_Compress(const CompressionInfo& info,
+                          uint32_t compress_format_version, const char* input,
+                          size_t length, ::std::string* output) {
+#ifdef ZLIB
+  if (length > std::numeric_limits<uint32_t>::max()) {
+    // Can't compress more than 4GB
+    return false;
+  }
+
+  size_t output_header_len = 0;
+  if (compress_format_version == 2) {
+    output_header_len = compression::PutDecompressedSizeInfo(
+        output, static_cast<uint32_t>(length));
+  }
+
+  // The memLevel parameter specifies how much memory should be allocated for
+  // the internal compression state.
+  // memLevel=1 uses minimum memory but is slow and reduces compression ratio.
+  // memLevel=9 uses maximum memory for optimal speed.
+  // The default value is 8. See zconf.h for more details.
+  static const int memLevel = 8;
+  int level;
+  if (info.options().level == CompressionOptions::kDefaultCompressionLevel) {
+    level = Z_DEFAULT_COMPRESSION;
+  } else {
+    level = info.options().level;
+  }
+  z_stream _stream;
+  memset(&_stream, 0, sizeof(z_stream));
+  int st = deflateInit2(&_stream, level, Z_DEFLATED, info.options().window_bits,
+                        memLevel, info.options().strategy);
+  if (st != Z_OK) {
+    return false;
+  }
+
+  Slice compression_dict = info.dict().GetRawDict();
+  if (compression_dict.size()) {
+    // Initialize the compression library's dictionary
+    st = deflateSetDictionary(
+        &_stream, reinterpret_cast<const Bytef*>(compression_dict.data()),
+        static_cast<unsigned int>(compression_dict.size()));
+    if (st != Z_OK) {
+      deflateEnd(&_stream);
+      return false;
+    }
+  }
+
+  // Get an upper bound on the compressed size.
+  size_t upper_bound =
+      deflateBound(&_stream, static_cast<unsigned long>(length));
+  output->resize(output_header_len + upper_bound);
+
+  // Compress the input, and put compressed data in output.
+  _stream.next_in = (Bytef*)input;
+  _stream.avail_in = static_cast<unsigned int>(length);
+
+  // Initialize the output size.
+  _stream.avail_out = static_cast<unsigned int>(upper_bound);
+  _stream.next_out = reinterpret_cast<Bytef*>(&(*output)[output_header_len]);
+
+  bool compressed = false;
+  st = deflate(&_stream, Z_FINISH);
+  if (st == Z_STREAM_END) {
+    compressed = true;
+    output->resize(output->size() - _stream.avail_out);
+  }
+  // The only return value we really care about is Z_STREAM_END.
+  // Z_OK means insufficient output space. This means the compression is
+  // bigger than decompressed size. Just fail the compression in that case.
+
+  deflateEnd(&_stream);
+  return compressed;
+#else
+  (void)info;
+  (void)compress_format_version;
+  (void)input;
+  (void)length;
+  (void)output;
+  return false;
+#endif
+}
+
+// compress_format_version == 1 -- decompressed size is not included in the
+// block header
+// compress_format_version == 2 -- decompressed size is included in the block
+// header in varint32 format
+// @param compression_dict Data for presetting the compression library's
+//    dictionary.
+inline CacheAllocationPtr Zlib_Uncompress(
+    const UncompressionInfo& info, const char* input_data, size_t input_length,
+    size_t* uncompressed_size, uint32_t compress_format_version,
+    MemoryAllocator* allocator = nullptr, int windowBits = -14) {
+#ifdef ZLIB
+  uint32_t output_len = 0;
+  if (compress_format_version == 2) {
+    if (!compression::GetDecompressedSizeInfo(&input_data, &input_length,
+                                              &output_len)) {
+      return nullptr;
+    }
+  } else {
+    // Assume the decompressed data size will 5x of compressed size, but round
+    // to the page size
+    size_t proposed_output_len = ((input_length * 5) & (~(4096 - 1))) + 4096;
+    output_len = static_cast<uint32_t>(
+        std::min(proposed_output_len,
+                 static_cast<size_t>(std::numeric_limits<uint32_t>::max())));
+  }
+
+  z_stream _stream;
+  memset(&_stream, 0, sizeof(z_stream));
+
+  // For raw inflate, the windowBits should be -8..-15.
+  // If windowBits is bigger than zero, it will use either zlib
+  // header or gzip header. Adding 32 to it will do automatic detection.
+  int st =
+      inflateInit2(&_stream, windowBits > 0 ? windowBits + 32 : windowBits);
+  if (st != Z_OK) {
+    return nullptr;
+  }
+
+  const Slice& compression_dict = info.dict().GetRawDict();
+  if (compression_dict.size()) {
+    // Initialize the compression library's dictionary
+    st = inflateSetDictionary(
+        &_stream, reinterpret_cast<const Bytef*>(compression_dict.data()),
+        static_cast<unsigned int>(compression_dict.size()));
+    if (st != Z_OK) {
+      return nullptr;
+    }
+  }
+
+  _stream.next_in = (Bytef*)input_data;
+  _stream.avail_in = static_cast<unsigned int>(input_length);
+
+  auto output = AllocateBlock(output_len, allocator);
+
+  _stream.next_out = (Bytef*)output.get();
+  _stream.avail_out = static_cast<unsigned int>(output_len);
+
+  bool done = false;
+  while (!done) {
+    st = inflate(&_stream, Z_SYNC_FLUSH);
+    switch (st) {
+      case Z_STREAM_END:
+        done = true;
+        break;
+      case Z_OK: {
+        // No output space. Increase the output space by 20%.
+        // We should never run out of output space if
+        // compress_format_version == 2
+        assert(compress_format_version != 2);
+        size_t old_sz = output_len;
+        uint32_t output_len_delta = output_len / 5;
+        output_len += output_len_delta < 10 ? 10 : output_len_delta;
+        auto tmp = AllocateBlock(output_len, allocator);
+        memcpy(tmp.get(), output.get(), old_sz);
+        output = std::move(tmp);
+
+        // Set more output.
+        _stream.next_out = (Bytef*)(output.get() + old_sz);
+        _stream.avail_out = static_cast<unsigned int>(output_len - old_sz);
+        break;
+      }
+      case Z_BUF_ERROR:
+      default:
+        inflateEnd(&_stream);
+        return nullptr;
+    }
+  }
+
+  // If we encoded decompressed block size, we should have no bytes left
+  assert(compress_format_version != 2 || _stream.avail_out == 0);
+  assert(output_len >= _stream.avail_out);
+  *uncompressed_size = output_len - _stream.avail_out;
+  inflateEnd(&_stream);
+  return output;
+#else
+  (void)info;
+  (void)input_data;
+  (void)input_length;
+  (void)uncompressed_size;
+  (void)compress_format_version;
+  (void)allocator;
+  (void)windowBits;
+  return nullptr;
+#endif
+}
+
+// compress_format_version == 1 -- decompressed size is not included in the
+// block header
+// compress_format_version == 2 -- decompressed size is included in the block
+// header in varint32 format
+inline bool BZip2_Compress(const CompressionInfo& /*info*/,
+                           uint32_t compress_format_version, const char* input,
+                           size_t length, ::std::string* output) {
+#ifdef BZIP2
+  if (length > std::numeric_limits<uint32_t>::max()) {
+    // Can't compress more than 4GB
+    return false;
+  }
+  size_t output_header_len = 0;
+  if (compress_format_version == 2) {
+    output_header_len = compression::PutDecompressedSizeInfo(
+        output, static_cast<uint32_t>(length));
+  }
+  // Resize output to be the plain data length.
+  // This may not be big enough if the compression actually expands data.
+  output->resize(output_header_len + length);
+
+  bz_stream _stream;
+  memset(&_stream, 0, sizeof(bz_stream));
+
+  // Block size 1 is 100K.
+  // 0 is for silent.
+  // 30 is the default workFactor
+  int st = BZ2_bzCompressInit(&_stream, 1, 0, 30);
+  if (st != BZ_OK) {
+    return false;
+  }
+
+  // Compress the input, and put compressed data in output.
+  _stream.next_in = (char*)input;
+  _stream.avail_in = static_cast<unsigned int>(length);
+
+  // Initialize the output size.
+  _stream.avail_out = static_cast<unsigned int>(length);
+  _stream.next_out = reinterpret_cast<char*>(&(*output)[output_header_len]);
+
+  bool compressed = false;
+  st = BZ2_bzCompress(&_stream, BZ_FINISH);
+  if (st == BZ_STREAM_END) {
+    compressed = true;
+    output->resize(output->size() - _stream.avail_out);
+  }
+  // The only return value we really care about is BZ_STREAM_END.
+  // BZ_FINISH_OK means insufficient output space. This means the compression
+  // is bigger than decompressed size. Just fail the compression in that case.
+
+  BZ2_bzCompressEnd(&_stream);
+  return compressed;
+#else
+  (void)compress_format_version;
+  (void)input;
+  (void)length;
+  (void)output;
+  return false;
+#endif
+}
+
+// compress_format_version == 1 -- decompressed size is not included in the
+// block header
+// compress_format_version == 2 -- decompressed size is included in the block
+// header in varint32 format
+inline CacheAllocationPtr BZip2_Uncompress(
+    const char* input_data, size_t input_length, size_t* uncompressed_size,
+    uint32_t compress_format_version, MemoryAllocator* allocator = nullptr) {
+#ifdef BZIP2
+  uint32_t output_len = 0;
+  if (compress_format_version == 2) {
+    if (!compression::GetDecompressedSizeInfo(&input_data, &input_length,
+                                              &output_len)) {
+      return nullptr;
+    }
+  } else {
+    // Assume the decompressed data size will 5x of compressed size, but round
+    // to the next page size
+    size_t proposed_output_len = ((input_length * 5) & (~(4096 - 1))) + 4096;
+    output_len = static_cast<uint32_t>(
+        std::min(proposed_output_len,
+                 static_cast<size_t>(std::numeric_limits<uint32_t>::max())));
+  }
+
+  bz_stream _stream;
+  memset(&_stream, 0, sizeof(bz_stream));
+
+  int st = BZ2_bzDecompressInit(&_stream, 0, 0);
+  if (st != BZ_OK) {
+    return nullptr;
+  }
+
+  _stream.next_in = (char*)input_data;
+  _stream.avail_in = static_cast<unsigned int>(input_length);
+
+  auto output = AllocateBlock(output_len, allocator);
+
+  _stream.next_out = (char*)output.get();
+  _stream.avail_out = static_cast<unsigned int>(output_len);
+
+  bool done = false;
+  while (!done) {
+    st = BZ2_bzDecompress(&_stream);
+    switch (st) {
+      case BZ_STREAM_END:
+        done = true;
+        break;
+      case BZ_OK: {
+        // No output space. Increase the output space by 20%.
+        // We should never run out of output space if
+        // compress_format_version == 2
+        assert(compress_format_version != 2);
+        uint32_t old_sz = output_len;
+        output_len = output_len * 1.2;
+        auto tmp = AllocateBlock(output_len, allocator);
+        memcpy(tmp.get(), output.get(), old_sz);
+        output = std::move(tmp);
+
+        // Set more output.
+        _stream.next_out = (char*)(output.get() + old_sz);
+        _stream.avail_out = static_cast<unsigned int>(output_len - old_sz);
+        break;
+      }
+      default:
+        BZ2_bzDecompressEnd(&_stream);
+        return nullptr;
+    }
+  }
+
+  // If we encoded decompressed block size, we should have no bytes left
+  assert(compress_format_version != 2 || _stream.avail_out == 0);
+  assert(output_len >= _stream.avail_out);
+  *uncompressed_size = output_len - _stream.avail_out;
+  BZ2_bzDecompressEnd(&_stream);
+  return output;
+#else
+  (void)input_data;
+  (void)input_length;
+  (void)uncompressed_size;
+  (void)compress_format_version;
+  (void)allocator;
+  return nullptr;
+#endif
+}
+
+// compress_format_version == 1 -- decompressed size is included in the
+// block header using memcpy, which makes database non-portable)
+// compress_format_version == 2 -- decompressed size is included in the block
+// header in varint32 format
+// @param compression_dict Data for presetting the compression library's
+//    dictionary.
+inline bool LZ4_Compress(const CompressionInfo& info,
+                         uint32_t compress_format_version, const char* input,
+                         size_t length, ::std::string* output) {
+#ifdef LZ4
+  if (length > std::numeric_limits<uint32_t>::max()) {
+    // Can't compress more than 4GB
+    return false;
+  }
+
+  size_t output_header_len = 0;
+  if (compress_format_version == 2) {
+    // new encoding, using varint32 to store size information
+    output_header_len = compression::PutDecompressedSizeInfo(
+        output, static_cast<uint32_t>(length));
+  } else {
+    // legacy encoding, which is not really portable (depends on big/little
+    // endianness)
+    output_header_len = 8;
+    output->resize(output_header_len);
+    char* p = const_cast<char*>(output->c_str());
+    memcpy(p, &length, sizeof(length));
+  }
+  int compress_bound = LZ4_compressBound(static_cast<int>(length));
+  output->resize(static_cast<size_t>(output_header_len + compress_bound));
+
+  int outlen;
+#if LZ4_VERSION_NUMBER >= 10400  // r124+
+  LZ4_stream_t* stream = LZ4_createStream();
+  Slice compression_dict = info.dict().GetRawDict();
+  if (compression_dict.size()) {
+    LZ4_loadDict(stream, compression_dict.data(),
+                 static_cast<int>(compression_dict.size()));
+  }
+#if LZ4_VERSION_NUMBER >= 10700  // r129+
+  outlen =
+      LZ4_compress_fast_continue(stream, input, &(*output)[output_header_len],
+                                 static_cast<int>(length), compress_bound, 1);
+#else  // up to r128
+  outlen = LZ4_compress_limitedOutput_continue(
+      stream, input, &(*output)[output_header_len], static_cast<int>(length),
+      compress_bound);
+#endif
+  LZ4_freeStream(stream);
+#else   // up to r123
+  outlen = LZ4_compress_limitedOutput(input, &(*output)[output_header_len],
+                                      static_cast<int>(length), compress_bound);
+#endif  // LZ4_VERSION_NUMBER >= 10400
+
+  if (outlen == 0) {
+    return false;
+  }
+  output->resize(static_cast<size_t>(output_header_len + outlen));
+  return true;
+#else  // LZ4
+  (void)info;
+  (void)compress_format_version;
+  (void)input;
+  (void)length;
+  (void)output;
+  return false;
+#endif
+}
+
+// compress_format_version == 1 -- decompressed size is included in the
+// block header using memcpy, which makes database non-portable)
+// compress_format_version == 2 -- decompressed size is included in the block
+// header in varint32 format
+// @param compression_dict Data for presetting the compression library's
+//    dictionary.
+inline CacheAllocationPtr LZ4_Uncompress(const UncompressionInfo& info,
+                                         const char* input_data,
+                                         size_t input_length,
+                                         size_t* uncompressed_size,
+                                         uint32_t compress_format_version,
+                                         MemoryAllocator* allocator = nullptr) {
+#ifdef LZ4
+  uint32_t output_len = 0;
+  if (compress_format_version == 2) {
+    // new encoding, using varint32 to store size information
+    if (!compression::GetDecompressedSizeInfo(&input_data, &input_length,
+                                              &output_len)) {
+      return nullptr;
+    }
+  } else {
+    // legacy encoding, which is not really portable (depends on big/little
+    // endianness)
+    if (input_length < 8) {
+      return nullptr;
+    }
+    if (port::kLittleEndian) {
+      memcpy(&output_len, input_data, sizeof(output_len));
+    } else {
+      memcpy(&output_len, input_data + 4, sizeof(output_len));
+    }
+    input_length -= 8;
+    input_data += 8;
+  }
+
+  auto output = AllocateBlock(output_len, allocator);
+
+  int decompress_bytes = 0;
+
+#if LZ4_VERSION_NUMBER >= 10400  // r124+
+  LZ4_streamDecode_t* stream = LZ4_createStreamDecode();
+  const Slice& compression_dict = info.dict().GetRawDict();
+  if (compression_dict.size()) {
+    LZ4_setStreamDecode(stream, compression_dict.data(),
+                        static_cast<int>(compression_dict.size()));
+  }
+  decompress_bytes = LZ4_decompress_safe_continue(
+      stream, input_data, output.get(), static_cast<int>(input_length),
+      static_cast<int>(output_len));
+  LZ4_freeStreamDecode(stream);
+#else   // up to r123
+  decompress_bytes = LZ4_decompress_safe(input_data, output.get(),
+                                         static_cast<int>(input_length),
+                                         static_cast<int>(output_len));
+#endif  // LZ4_VERSION_NUMBER >= 10400
+
+  if (decompress_bytes < 0) {
+    return nullptr;
+  }
+  assert(decompress_bytes == static_cast<int>(output_len));
+  *uncompressed_size = decompress_bytes;
+  return output;
+#else  // LZ4
+  (void)info;
+  (void)input_data;
+  (void)input_length;
+  (void)uncompressed_size;
+  (void)compress_format_version;
+  (void)allocator;
+  return nullptr;
+#endif
+}
+
+// compress_format_version == 1 -- decompressed size is included in the
+// block header using memcpy, which makes database non-portable)
+// compress_format_version == 2 -- decompressed size is included in the block
+// header in varint32 format
+// @param compression_dict Data for presetting the compression library's
+//    dictionary.
+inline bool LZ4HC_Compress(const CompressionInfo& info,
+                           uint32_t compress_format_version, const char* input,
+                           size_t length, ::std::string* output) {
+#ifdef LZ4
+  if (length > std::numeric_limits<uint32_t>::max()) {
+    // Can't compress more than 4GB
+    return false;
+  }
+
+  size_t output_header_len = 0;
+  if (compress_format_version == 2) {
+    // new encoding, using varint32 to store size information
+    output_header_len = compression::PutDecompressedSizeInfo(
+        output, static_cast<uint32_t>(length));
+  } else {
+    // legacy encoding, which is not really portable (depends on big/little
+    // endianness)
+    output_header_len = 8;
+    output->resize(output_header_len);
+    char* p = const_cast<char*>(output->c_str());
+    memcpy(p, &length, sizeof(length));
+  }
+  int compress_bound = LZ4_compressBound(static_cast<int>(length));
+  output->resize(static_cast<size_t>(output_header_len + compress_bound));
+
+  int outlen;
+  int level;
+  if (info.options().level == CompressionOptions::kDefaultCompressionLevel) {
+    level = 0;  // lz4hc.h says any value < 1 will be sanitized to default
+  } else {
+    level = info.options().level;
+  }
+#if LZ4_VERSION_NUMBER >= 10400  // r124+
+  LZ4_streamHC_t* stream = LZ4_createStreamHC();
+  LZ4_resetStreamHC(stream, level);
+  Slice compression_dict = info.dict().GetRawDict();
+  const char* compression_dict_data =
+      compression_dict.size() > 0 ? compression_dict.data() : nullptr;
+  size_t compression_dict_size = compression_dict.size();
+  if (compression_dict_data != nullptr) {
+    LZ4_loadDictHC(stream, compression_dict_data,
+                   static_cast<int>(compression_dict_size));
+  }
+
+#if LZ4_VERSION_NUMBER >= 10700  // r129+
+  outlen =
+      LZ4_compress_HC_continue(stream, input, &(*output)[output_header_len],
+                               static_cast<int>(length), compress_bound);
+#else   // r124-r128
+  outlen = LZ4_compressHC_limitedOutput_continue(
+      stream, input, &(*output)[output_header_len], static_cast<int>(length),
+      compress_bound);
+#endif  // LZ4_VERSION_NUMBER >= 10700
+  LZ4_freeStreamHC(stream);
+
+#elif LZ4_VERSION_MAJOR  // r113-r123
+  outlen = LZ4_compressHC2_limitedOutput(input, &(*output)[output_header_len],
+                                         static_cast<int>(length),
+                                         compress_bound, level);
+#else                    // up to r112
+  outlen =
+      LZ4_compressHC_limitedOutput(input, &(*output)[output_header_len],
+                                   static_cast<int>(length), compress_bound);
+#endif                   // LZ4_VERSION_NUMBER >= 10400
+
+  if (outlen == 0) {
+    return false;
+  }
+  output->resize(static_cast<size_t>(output_header_len + outlen));
+  return true;
+#else  // LZ4
+  (void)info;
+  (void)compress_format_version;
+  (void)input;
+  (void)length;
+  (void)output;
+  return false;
+#endif
+}
+
+#ifdef XPRESS
+inline bool XPRESS_Compress(const char* input, size_t length,
+                            std::string* output) {
+  return port::xpress::Compress(input, length, output);
+}
+#else
+inline bool XPRESS_Compress(const char* /*input*/, size_t /*length*/,
+                            std::string* /*output*/) {
+  return false;
+}
+#endif
+
+#ifdef XPRESS
+inline char* XPRESS_Uncompress(const char* input_data, size_t input_length,
+                               size_t* uncompressed_size) {
+  return port::xpress::Decompress(input_data, input_length, uncompressed_size);
+}
+#else
+inline char* XPRESS_Uncompress(const char* /*input_data*/,
+                               size_t /*input_length*/,
+                               size_t* /*uncompressed_size*/) {
+  return nullptr;
+}
+#endif
+
+inline bool ZSTD_Compress(const CompressionInfo& info, const char* input,
+                          size_t length, ::std::string* output) {
+#ifdef ZSTD
+  if (length > std::numeric_limits<uint32_t>::max()) {
+    // Can't compress more than 4GB
+    return false;
+  }
+
+  size_t output_header_len = compression::PutDecompressedSizeInfo(
+      output, static_cast<uint32_t>(length));
+
+  size_t compressBound = ZSTD_compressBound(length);
+  output->resize(static_cast<size_t>(output_header_len + compressBound));
+  size_t outlen = 0;
+  int level;
+  if (info.options().level == CompressionOptions::kDefaultCompressionLevel) {
+    // 3 is the value of ZSTD_CLEVEL_DEFAULT (not exposed publicly), see
+    // https://github.com/facebook/zstd/issues/1148
+    level = 3;
+  } else {
+    level = info.options().level;
+  }
+#if ZSTD_VERSION_NUMBER >= 500  // v0.5.0+
+  ZSTD_CCtx* context = info.context().ZSTDPreallocCtx();
+  assert(context != nullptr);
+#if ZSTD_VERSION_NUMBER >= 700  // v0.7.0+
+  if (info.dict().GetDigestedZstdCDict() != nullptr) {
+    outlen = ZSTD_compress_usingCDict(context, &(*output)[output_header_len],
+                                      compressBound, input, length,
+                                      info.dict().GetDigestedZstdCDict());
+  }
+#endif  // ZSTD_VERSION_NUMBER >= 700
+  if (outlen == 0) {
+    outlen = ZSTD_compress_usingDict(context, &(*output)[output_header_len],
+                                     compressBound, input, length,
+                                     info.dict().GetRawDict().data(),
+                                     info.dict().GetRawDict().size(), level);
+  }
+#else   // up to v0.4.x
+  outlen = ZSTD_compress(&(*output)[output_header_len], compressBound, input,
+                         length, level);
+#endif  // ZSTD_VERSION_NUMBER >= 500
+  if (outlen == 0) {
+    return false;
+  }
+  output->resize(output_header_len + outlen);
+  return true;
+#else  // ZSTD
+  (void)info;
+  (void)input;
+  (void)length;
+  (void)output;
+  return false;
+#endif
+}
+
+// @param compression_dict Data for presetting the compression library's
+//    dictionary.
+inline CacheAllocationPtr ZSTD_Uncompress(
+    const UncompressionInfo& info, const char* input_data, size_t input_length,
+    size_t* uncompressed_size, MemoryAllocator* allocator = nullptr) {
+#ifdef ZSTD
+  uint32_t output_len = 0;
+  if (!compression::GetDecompressedSizeInfo(&input_data, &input_length,
+                                            &output_len)) {
+    return nullptr;
+  }
+
+  auto output = AllocateBlock(output_len, allocator);
+  size_t actual_output_length = 0;
+#if ZSTD_VERSION_NUMBER >= 500  // v0.5.0+
+  ZSTD_DCtx* context = info.context().GetZSTDContext();
+  assert(context != nullptr);
+#ifdef ROCKSDB_ZSTD_DDICT
+  if (info.dict().GetDigestedZstdDDict() != nullptr) {
+    actual_output_length = ZSTD_decompress_usingDDict(
+        context, output.get(), output_len, input_data, input_length,
+        info.dict().GetDigestedZstdDDict());
+  }
+#endif  // ROCKSDB_ZSTD_DDICT
+  if (actual_output_length == 0) {
+    actual_output_length = ZSTD_decompress_usingDict(
+        context, output.get(), output_len, input_data, input_length,
+        info.dict().GetRawDict().data(), info.dict().GetRawDict().size());
+  }
+#else   // up to v0.4.x
+  (void)info;
+  actual_output_length =
+      ZSTD_decompress(output.get(), output_len, input_data, input_length);
+#endif  // ZSTD_VERSION_NUMBER >= 500
+  assert(actual_output_length == output_len);
+  *uncompressed_size = actual_output_length;
+  return output;
+#else  // ZSTD
+  (void)info;
+  (void)input_data;
+  (void)input_length;
+  (void)uncompressed_size;
+  (void)allocator;
+  return nullptr;
+#endif
+}
+
+inline bool ZSTD_TrainDictionarySupported() {
+#ifdef ZSTD
+  // Dictionary trainer is available since v0.6.1 for static linking, but not
+  // available for dynamic linking until v1.1.3. For now we enable the feature
+  // in v1.1.3+ only.
+  return (ZSTD_versionNumber() >= 10103);
+#else
+  return false;
+#endif
+}
+
+inline std::string ZSTD_TrainDictionary(const std::string& samples,
+                                        const std::vector<size_t>& sample_lens,
+                                        size_t max_dict_bytes) {
+  // Dictionary trainer is available since v0.6.1 for static linking, but not
+  // available for dynamic linking until v1.1.3. For now we enable the feature
+  // in v1.1.3+ only.
+#if ZSTD_VERSION_NUMBER >= 10103  // v1.1.3+
+  assert(samples.empty() == sample_lens.empty());
+  if (samples.empty()) {
+    return "";
+  }
+  std::string dict_data(max_dict_bytes, '\0');
+  size_t dict_len = ZDICT_trainFromBuffer(
+      &dict_data[0], max_dict_bytes, &samples[0], &sample_lens[0],
+      static_cast<unsigned>(sample_lens.size()));
+  if (ZDICT_isError(dict_len)) {
+    return "";
+  }
+  assert(dict_len <= max_dict_bytes);
+  dict_data.resize(dict_len);
+  return dict_data;
+#else   // up to v1.1.2
+  assert(false);
+  (void)samples;
+  (void)sample_lens;
+  (void)max_dict_bytes;
+  return "";
+#endif  // ZSTD_VERSION_NUMBER >= 10103
+}
+
+inline std::string ZSTD_TrainDictionary(const std::string& samples,
+                                        size_t sample_len_shift,
+                                        size_t max_dict_bytes) {
+  // Dictionary trainer is available since v0.6.1, but ZSTD was marked stable
+  // only since v0.8.0. For now we enable the feature in stable versions only.
+#if ZSTD_VERSION_NUMBER >= 10103  // v1.1.3+
+  // skips potential partial sample at the end of "samples"
+  size_t num_samples = samples.size() >> sample_len_shift;
+  std::vector<size_t> sample_lens(num_samples, size_t(1) << sample_len_shift);
+  return ZSTD_TrainDictionary(samples, sample_lens, max_dict_bytes);
+#else   // up to v1.1.2
+  assert(false);
+  (void)samples;
+  (void)sample_len_shift;
+  (void)max_dict_bytes;
+  return "";
+#endif  // ZSTD_VERSION_NUMBER >= 10103
+}
+
+inline bool ZSTD_FinalizeDictionarySupported() {
+#ifdef ZSTD
+  // ZDICT_finalizeDictionary API is stable since v1.4.5
+  return (ZSTD_versionNumber() >= 10405);
+#else
+  return false;
+#endif
+}
+
+inline std::string ZSTD_FinalizeDictionary(
+    const std::string& samples, const std::vector<size_t>& sample_lens,
+    size_t max_dict_bytes, int level) {
+  // ZDICT_finalizeDictionary is stable since version v1.4.5
+#if ZSTD_VERSION_NUMBER >= 10405  // v1.4.5+
+  assert(samples.empty() == sample_lens.empty());
+  if (samples.empty()) {
+    return "";
+  }
+  if (level == CompressionOptions::kDefaultCompressionLevel) {
+    // 3 is the value of ZSTD_CLEVEL_DEFAULT (not exposed publicly), see
+    // https://github.com/facebook/zstd/issues/1148
+    level = 3;
+  }
+  std::string dict_data(max_dict_bytes, '\0');
+  size_t dict_len = ZDICT_finalizeDictionary(
+      dict_data.data(), max_dict_bytes, samples.data(),
+      std::min(static_cast<size_t>(samples.size()), max_dict_bytes),
+      samples.data(), sample_lens.data(),
+      static_cast<unsigned>(sample_lens.size()),
+      {level, 0 /* notificationLevel */, 0 /* dictID */});
+  if (ZDICT_isError(dict_len)) {
+    return "";
+  } else {
+    assert(dict_len <= max_dict_bytes);
+    dict_data.resize(dict_len);
+    return dict_data;
+  }
+#else   // up to v1.4.4
+  (void)samples;
+  (void)sample_lens;
+  (void)max_dict_bytes;
+  (void)level;
+  return "";
+#endif  // ZSTD_VERSION_NUMBER >= 10405
+}
+
+inline bool CompressData(const Slice& raw,
+                         const CompressionInfo& compression_info,
+                         uint32_t compress_format_version,
+                         std::string* compressed_output) {
+  bool ret = false;
+
+  // Will return compressed block contents if (1) the compression method is
+  // supported in this platform and (2) the compression rate is "good enough".
+  switch (compression_info.type()) {
+    case kSnappyCompression:
+      ret = Snappy_Compress(compression_info, raw.data(), raw.size(),
+                            compressed_output);
+      break;
+    case kZlibCompression:
+      ret = Zlib_Compress(compression_info, compress_format_version, raw.data(),
+                          raw.size(), compressed_output);
+      break;
+    case kBZip2Compression:
+      ret = BZip2_Compress(compression_info, compress_format_version,
+                           raw.data(), raw.size(), compressed_output);
+      break;
+    case kLZ4Compression:
+      ret = LZ4_Compress(compression_info, compress_format_version, raw.data(),
+                         raw.size(), compressed_output);
+      break;
+    case kLZ4HCCompression:
+      ret = LZ4HC_Compress(compression_info, compress_format_version,
+                           raw.data(), raw.size(), compressed_output);
+      break;
+    case kXpressCompression:
+      ret = XPRESS_Compress(raw.data(), raw.size(), compressed_output);
+      break;
+    case kZSTD:
+    case kZSTDNotFinalCompression:
+      ret = ZSTD_Compress(compression_info, raw.data(), raw.size(),
+                          compressed_output);
+      break;
+    default:
+      // Do not recognize this compression type
+      break;
+  }
+
+  TEST_SYNC_POINT_CALLBACK("CompressData:TamperWithReturnValue",
+                           static_cast<void*>(&ret));
+
+  return ret;
+}
+
+inline CacheAllocationPtr UncompressData(
+    const UncompressionInfo& uncompression_info, const char* data, size_t n,
+    size_t* uncompressed_size, uint32_t compress_format_version,
+    MemoryAllocator* allocator = nullptr) {
+  switch (uncompression_info.type()) {
+    case kSnappyCompression:
+      return Snappy_Uncompress(data, n, uncompressed_size, allocator);
+    case kZlibCompression:
+      return Zlib_Uncompress(uncompression_info, data, n, uncompressed_size,
+                             compress_format_version, allocator);
+    case kBZip2Compression:
+      return BZip2_Uncompress(data, n, uncompressed_size,
+                              compress_format_version, allocator);
+    case kLZ4Compression:
+    case kLZ4HCCompression:
+      return LZ4_Uncompress(uncompression_info, data, n, uncompressed_size,
+                            compress_format_version, allocator);
+    case kXpressCompression:
+      // XPRESS allocates memory internally, thus no support for custom
+      // allocator.
+      return CacheAllocationPtr(XPRESS_Uncompress(data, n, uncompressed_size));
+    case kZSTD:
+    case kZSTDNotFinalCompression:
+      return ZSTD_Uncompress(uncompression_info, data, n, uncompressed_size,
+                             allocator);
+    default:
+      return CacheAllocationPtr();
+  }
+}
+
+// Records the compression type for subsequent WAL records.
+class CompressionTypeRecord {
+ public:
+  explicit CompressionTypeRecord(CompressionType compression_type)
+      : compression_type_(compression_type) {}
+
+  CompressionType GetCompressionType() const { return compression_type_; }
+
+  inline void EncodeTo(std::string* dst) const {
+    assert(dst != nullptr);
+    PutFixed32(dst, compression_type_);
+  }
+
+  inline Status DecodeFrom(Slice* src) {
+    constexpr char class_name[] = "CompressionTypeRecord";
+
+    uint32_t val;
+    if (!GetFixed32(src, &val)) {
+      return Status::Corruption(class_name,
+                                "Error decoding WAL compression type");
+    }
+    CompressionType compression_type = static_cast<CompressionType>(val);
+    if (!StreamingCompressionTypeSupported(compression_type)) {
+      return Status::Corruption(class_name,
+                                "WAL compression type not supported");
+    }
+    compression_type_ = compression_type;
+    return Status::OK();
+  }
+
+  inline std::string DebugString() const {
+    return "compression_type: " + CompressionTypeToString(compression_type_);
+  }
+
+ private:
+  CompressionType compression_type_;
+};
+
+// Base class to implement compression for a stream of buffers.
+// Instantiate an implementation of the class using Create() with the
+// compression type and use Compress() repeatedly.
+// The output buffer needs to be at least max_output_len.
+// Call Reset() in between frame boundaries or in case of an error.
+// NOTE: This class is not thread safe.
+class StreamingCompress {
+ public:
+  StreamingCompress(CompressionType compression_type,
+                    const CompressionOptions& opts,
+                    uint32_t compress_format_version, size_t max_output_len)
+      : compression_type_(compression_type),
+        opts_(opts),
+        compress_format_version_(compress_format_version),
+        max_output_len_(max_output_len) {}
+  virtual ~StreamingCompress() = default;
+  // compress should be called repeatedly with the same input till the method
+  // returns 0
+  // Parameters:
+  // input - buffer to compress
+  // input_size - size of input buffer
+  // output - compressed buffer allocated by caller, should be at least
+  // max_output_len
+  // output_size - size of the output buffer
+  // Returns -1 for errors, the remaining size of the input buffer that needs to
+  // be compressed
+  virtual int Compress(const char* input, size_t input_size, char* output,
+                       size_t* output_pos) = 0;
+  // static method to create object of a class inherited from StreamingCompress
+  // based on the actual compression type.
+  static StreamingCompress* Create(CompressionType compression_type,
+                                   const CompressionOptions& opts,
+                                   uint32_t compress_format_version,
+                                   size_t max_output_len);
+  virtual void Reset() = 0;
+
+ protected:
+  const CompressionType compression_type_;
+  const CompressionOptions opts_;
+  const uint32_t compress_format_version_;
+  const size_t max_output_len_;
+};
+
+// Base class to uncompress a stream of compressed buffers.
+// Instantiate an implementation of the class using Create() with the
+// compression type and use Uncompress() repeatedly.
+// The output buffer needs to be at least max_output_len.
+// Call Reset() in between frame boundaries or in case of an error.
+// NOTE: This class is not thread safe.
+class StreamingUncompress {
+ public:
+  StreamingUncompress(CompressionType compression_type,
+                      uint32_t compress_format_version, size_t max_output_len)
+      : compression_type_(compression_type),
+        compress_format_version_(compress_format_version),
+        max_output_len_(max_output_len) {}
+  virtual ~StreamingUncompress() = default;
+  // uncompress should be called again with the same input if output_size is
+  // equal to max_output_len or with the next input fragment.
+  // Parameters:
+  // input - buffer to uncompress
+  // input_size - size of input buffer
+  // output - uncompressed buffer allocated by caller, should be at least
+  // max_output_len
+  // output_size - size of the output buffer
+  // Returns -1 for errors, remaining input to be processed otherwise.
+  virtual int Uncompress(const char* input, size_t input_size, char* output,
+                         size_t* output_pos) = 0;
+  static StreamingUncompress* Create(CompressionType compression_type,
+                                     uint32_t compress_format_version,
+                                     size_t max_output_len);
+  virtual void Reset() = 0;
+
+ protected:
+  CompressionType compression_type_;
+  uint32_t compress_format_version_;
+  size_t max_output_len_;
+};
+
+class ZSTDStreamingCompress final : public StreamingCompress {
+ public:
+  explicit ZSTDStreamingCompress(const CompressionOptions& opts,
+                                 uint32_t compress_format_version,
+                                 size_t max_output_len)
+      : StreamingCompress(kZSTD, opts, compress_format_version,
+                          max_output_len) {
+#ifdef ZSTD_STREAMING
+    cctx_ = ZSTD_createCCtx();
+    // Each compressed frame will have a checksum
+    ZSTD_CCtx_setParameter(cctx_, ZSTD_c_checksumFlag, 1);
+    assert(cctx_ != nullptr);
+    input_buffer_ = {/*src=*/nullptr, /*size=*/0, /*pos=*/0};
+#endif
+  }
+  ~ZSTDStreamingCompress() override {
+#ifdef ZSTD_STREAMING
+    ZSTD_freeCCtx(cctx_);
+#endif
+  }
+  int Compress(const char* input, size_t input_size, char* output,
+               size_t* output_pos) override;
+  void Reset() override;
+#ifdef ZSTD_STREAMING
+  ZSTD_CCtx* cctx_;
+  ZSTD_inBuffer input_buffer_;
+#endif
+};
+
+class ZSTDStreamingUncompress final : public StreamingUncompress {
+ public:
+  explicit ZSTDStreamingUncompress(uint32_t compress_format_version,
+                                   size_t max_output_len)
+      : StreamingUncompress(kZSTD, compress_format_version, max_output_len) {
+#ifdef ZSTD_STREAMING
+    dctx_ = ZSTD_createDCtx();
+    assert(dctx_ != nullptr);
+    input_buffer_ = {/*src=*/nullptr, /*size=*/0, /*pos=*/0};
+#endif
+  }
+  ~ZSTDStreamingUncompress() override {
+#ifdef ZSTD_STREAMING
+    ZSTD_freeDCtx(dctx_);
+#endif
+  }
+  int Uncompress(const char* input, size_t input_size, char* output,
+                 size_t* output_size) override;
+  void Reset() override;
+
+ private:
+#ifdef ZSTD_STREAMING
+  ZSTD_DCtx* dctx_;
+  ZSTD_inBuffer input_buffer_;
+#endif
+};
+
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/util/compression_context_cache.cc b/src/rocksdb/util/compression_context_cache.cc
new file mode 100644
index 000000000..52c3fac72
--- /dev/null
+++ b/src/rocksdb/util/compression_context_cache.cc
@@ -0,0 +1,106 @@
+// Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+// Copyright (c) 2011 The LevelDB Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file. See the AUTHORS file for names of contributors.
+//
+
+#include "util/compression_context_cache.h"
+
+#include <atomic>
+
+#include "util/compression.h"
+#include "util/core_local.h"
+
+namespace ROCKSDB_NAMESPACE {
+namespace compression_cache {
+
+void* const SentinelValue = nullptr;
+// Cache ZSTD uncompression contexts for reads
+// if needed we can add ZSTD compression context caching
+// which is currently is not done since BlockBasedTableBuilder
+// simply creates one compression context per new SST file.
+struct ZSTDCachedData {
+  // We choose to cache the below structure instead of a ptr
+  // because we want to avoid a) native types leak b) make
+  // cache use transparent for the user
+  ZSTDUncompressCachedData uncomp_cached_data_;
+  std::atomic<void*> zstd_uncomp_sentinel_;
+
+  char
+      padding[(CACHE_LINE_SIZE -
+               (sizeof(ZSTDUncompressCachedData) + sizeof(std::atomic<void*>)) %
+                   CACHE_LINE_SIZE)];  // unused padding field
+
+  ZSTDCachedData() : zstd_uncomp_sentinel_(&uncomp_cached_data_) {}
+  ZSTDCachedData(const ZSTDCachedData&) = delete;
+  ZSTDCachedData& operator=(const ZSTDCachedData&) = delete;
+
+  ZSTDUncompressCachedData GetUncompressData(int64_t idx) {
+    ZSTDUncompressCachedData result;
+    void* expected = &uncomp_cached_data_;
+    if (zstd_uncomp_sentinel_.compare_exchange_strong(expected,
+                                                      SentinelValue)) {
+      uncomp_cached_data_.CreateIfNeeded();
+      result.InitFromCache(uncomp_cached_data_, idx);
+    } else {
+      // Creates one time use data
+      result.CreateIfNeeded();
+    }
+    return result;
+  }
+  // Return the entry back into circulation
+  // This is executed only when we successfully obtained
+  // in the first place
+  void ReturnUncompressData() {
+    if (zstd_uncomp_sentinel_.exchange(&uncomp_cached_data_) != SentinelValue) {
+      // Means we are returning while not having it acquired.
+      assert(false);
+    }
+  }
+};
+static_assert(sizeof(ZSTDCachedData) % CACHE_LINE_SIZE == 0,
+              "Expected CACHE_LINE_SIZE alignment");
+}  // namespace compression_cache
+
+class CompressionContextCache::Rep {
+ public:
+  Rep() {}
+  ZSTDUncompressCachedData GetZSTDUncompressData() {
+    auto p = per_core_uncompr_.AccessElementAndIndex();
+    int64_t idx = static_cast<int64_t>(p.second);
+    return p.first->GetUncompressData(idx);
+  }
+  void ReturnZSTDUncompressData(int64_t idx) {
+    assert(idx >= 0);
+    auto* cn = per_core_uncompr_.AccessAtCore(static_cast<size_t>(idx));
+    cn->ReturnUncompressData();
+  }
+
+ private:
+  CoreLocalArray<compression_cache::ZSTDCachedData> per_core_uncompr_;
+};
+
+CompressionContextCache::CompressionContextCache() : rep_(new Rep()) {}
+
+CompressionContextCache* CompressionContextCache::Instance() {
+  static CompressionContextCache instance;
+  return &instance;
+}
+
+void CompressionContextCache::InitSingleton() { Instance(); }
+
+ZSTDUncompressCachedData
+CompressionContextCache::GetCachedZSTDUncompressData() {
+  return rep_->GetZSTDUncompressData();
+}
+
+void CompressionContextCache::ReturnCachedZSTDUncompressData(int64_t idx) {
+  rep_->ReturnZSTDUncompressData(idx);
+}
+
+CompressionContextCache::~CompressionContextCache() { delete rep_; }
+
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/util/compression_context_cache.h b/src/rocksdb/util/compression_context_cache.h
new file mode 100644
index 000000000..7b7b2d507
--- /dev/null
+++ b/src/rocksdb/util/compression_context_cache.h
@@ -0,0 +1,47 @@
+// Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+// Copyright (c) 2011 The LevelDB Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file. See the AUTHORS file for names of contributors.
+//
+
+// Compression context cache allows to cache compression/uncompression contexts
+// This helps with Random Read latencies and reduces CPU utilization
+// Caching is implemented using CoreLocal facility. Compression/Uncompression
+// instances are cached on a per core basis using CoreLocalArray. A borrowed
+// instance is atomically replaced with a sentinel value for the time of being
+// used. If it turns out that another thread is already makes use of the
+// instance we still create one on the heap which is later is destroyed.
+
+#pragma once
+
+#include <stdint.h>
+
+#include "rocksdb/rocksdb_namespace.h"
+
+namespace ROCKSDB_NAMESPACE {
+class ZSTDUncompressCachedData;
+
+class CompressionContextCache {
+ public:
+  // Singleton
+  static CompressionContextCache* Instance();
+  static void InitSingleton();
+  CompressionContextCache(const CompressionContextCache&) = delete;
+  CompressionContextCache& operator=(const CompressionContextCache&) = delete;
+
+  ZSTDUncompressCachedData GetCachedZSTDUncompressData();
+  void ReturnCachedZSTDUncompressData(int64_t idx);
+
+ private:
+  // Singleton
+  CompressionContextCache();
+  ~CompressionContextCache();
+
+  class Rep;
+  Rep* rep_;
+};
+
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/util/concurrent_task_limiter_impl.cc b/src/rocksdb/util/concurrent_task_limiter_impl.cc
new file mode 100644
index 000000000..a0fc7331f
--- /dev/null
+++ b/src/rocksdb/util/concurrent_task_limiter_impl.cc
@@ -0,0 +1,64 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+//
+// Copyright (c) 2011 The LevelDB Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file. See the AUTHORS file for names of contributors.
+
+#include "util/concurrent_task_limiter_impl.h"
+
+#include "rocksdb/concurrent_task_limiter.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+ConcurrentTaskLimiterImpl::ConcurrentTaskLimiterImpl(
+    const std::string& name, int32_t max_outstanding_task)
+    : name_(name),
+      max_outstanding_tasks_{max_outstanding_task},
+      outstanding_tasks_{0} {}
+
+ConcurrentTaskLimiterImpl::~ConcurrentTaskLimiterImpl() {
+  assert(outstanding_tasks_ == 0);
+}
+
+const std::string& ConcurrentTaskLimiterImpl::GetName() const { return name_; }
+
+void ConcurrentTaskLimiterImpl::SetMaxOutstandingTask(int32_t limit) {
+  max_outstanding_tasks_.store(limit, std::memory_order_relaxed);
+}
+
+void ConcurrentTaskLimiterImpl::ResetMaxOutstandingTask() {
+  max_outstanding_tasks_.store(-1, std::memory_order_relaxed);
+}
+
+int32_t ConcurrentTaskLimiterImpl::GetOutstandingTask() const {
+  return outstanding_tasks_.load(std::memory_order_relaxed);
+}
+
+std::unique_ptr<TaskLimiterToken> ConcurrentTaskLimiterImpl::GetToken(
+    bool force) {
+  int32_t limit = max_outstanding_tasks_.load(std::memory_order_relaxed);
+  int32_t tasks = outstanding_tasks_.load(std::memory_order_relaxed);
+  // force = true, bypass the throttle.
+  // limit < 0 means unlimited tasks.
+  while (force || limit < 0 || tasks < limit) {
+    if (outstanding_tasks_.compare_exchange_weak(tasks, tasks + 1)) {
+      return std::unique_ptr<TaskLimiterToken>(new TaskLimiterToken(this));
+    }
+  }
+  return nullptr;
+}
+
+ConcurrentTaskLimiter* NewConcurrentTaskLimiter(const std::string& name,
+                                                int32_t limit) {
+  return new ConcurrentTaskLimiterImpl(name, limit);
+}
+
+TaskLimiterToken::~TaskLimiterToken() {
+  --limiter_->outstanding_tasks_;
+  assert(limiter_->outstanding_tasks_ >= 0);
+}
+
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/util/concurrent_task_limiter_impl.h b/src/rocksdb/util/concurrent_task_limiter_impl.h
new file mode 100644
index 000000000..4952ae23a
--- /dev/null
+++ b/src/rocksdb/util/concurrent_task_limiter_impl.h
@@ -0,0 +1,67 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+//
+// Copyright (c) 2011 The LevelDB Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file. See the AUTHORS file for names of contributors.
+
+#pragma once
+#include <atomic>
+#include <memory>
+
+#include "rocksdb/concurrent_task_limiter.h"
+#include "rocksdb/env.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+class TaskLimiterToken;
+
+class ConcurrentTaskLimiterImpl : public ConcurrentTaskLimiter {
+ public:
+  explicit ConcurrentTaskLimiterImpl(const std::string& name,
+                                     int32_t max_outstanding_task);
+  // No copying allowed
+  ConcurrentTaskLimiterImpl(const ConcurrentTaskLimiterImpl&) = delete;
+  ConcurrentTaskLimiterImpl& operator=(const ConcurrentTaskLimiterImpl&) =
+      delete;
+
+  virtual ~ConcurrentTaskLimiterImpl();
+
+  virtual const std::string& GetName() const override;
+
+  virtual void SetMaxOutstandingTask(int32_t limit) override;
+
+  virtual void ResetMaxOutstandingTask() override;
+
+  virtual int32_t GetOutstandingTask() const override;
+
+  // Request token for adding a new task.
+  // If force == true, it requests a token bypassing throttle.
+  // Returns nullptr if it got throttled.
+  virtual std::unique_ptr<TaskLimiterToken> GetToken(bool force);
+
+ private:
+  friend class TaskLimiterToken;
+
+  std::string name_;
+  std::atomic<int32_t> max_outstanding_tasks_;
+  std::atomic<int32_t> outstanding_tasks_;
+};
+
+class TaskLimiterToken {
+ public:
+  explicit TaskLimiterToken(ConcurrentTaskLimiterImpl* limiter)
+      : limiter_(limiter) {}
+  ~TaskLimiterToken();
+
+ private:
+  ConcurrentTaskLimiterImpl* limiter_;
+
+  // no copying allowed
+  TaskLimiterToken(const TaskLimiterToken&) = delete;
+  void operator=(const TaskLimiterToken&) = delete;
+};
+
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/util/core_local.h b/src/rocksdb/util/core_local.h
new file mode 100644
index 000000000..b444a1152
--- /dev/null
+++ b/src/rocksdb/util/core_local.h
@@ -0,0 +1,83 @@
+//  Copyright (c) 2017-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#pragma once
+
+#include <cstddef>
+#include <thread>
+#include <utility>
+#include <vector>
+
+#include "port/likely.h"
+#include "port/port.h"
+#include "util/random.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+// An array of core-local values. Ideally the value type, T, is cache aligned to
+// prevent false sharing.
+template <typename T>
+class CoreLocalArray {
+ public:
+  CoreLocalArray();
+
+  size_t Size() const;
+  // returns pointer to the element corresponding to the core that the thread
+  // currently runs on.
+  T* Access() const;
+  // same as above, but also returns the core index, which the client can cache
+  // to reduce how often core ID needs to be retrieved. Only do this if some
+  // inaccuracy is tolerable, as the thread may migrate to a different core.
+  std::pair<T*, size_t> AccessElementAndIndex() const;
+  // returns pointer to element for the specified core index. This can be used,
+  // e.g., for aggregation, or if the client caches core index.
+  T* AccessAtCore(size_t core_idx) const;
+
+ private:
+  std::unique_ptr<T[]> data_;
+  int size_shift_;
+};
+
+template <typename T>
+CoreLocalArray<T>::CoreLocalArray() {
+  int num_cpus = static_cast<int>(std::thread::hardware_concurrency());
+  // find a power of two >= num_cpus and >= 8
+  size_shift_ = 3;
+  while (1 << size_shift_ < num_cpus) {
+    ++size_shift_;
+  }
+  data_.reset(new T[static_cast<size_t>(1) << size_shift_]);
+}
+
+template <typename T>
+size_t CoreLocalArray<T>::Size() const {
+  return static_cast<size_t>(1) << size_shift_;
+}
+
+template <typename T>
+T* CoreLocalArray<T>::Access() const {
+  return AccessElementAndIndex().first;
+}
+
+template <typename T>
+std::pair<T*, size_t> CoreLocalArray<T>::AccessElementAndIndex() const {
+  int cpuid = port::PhysicalCoreID();
+  size_t core_idx;
+  if (UNLIKELY(cpuid < 0)) {
+    // cpu id unavailable, just pick randomly
+    core_idx = Random::GetTLSInstance()->Uniform(1 << size_shift_);
+  } else {
+    core_idx = static_cast<size_t>(cpuid & ((1 << size_shift_) - 1));
+  }
+  return {AccessAtCore(core_idx), core_idx};
+}
+
+template <typename T>
+T* CoreLocalArray<T>::AccessAtCore(size_t core_idx) const {
+  assert(core_idx < static_cast<size_t>(1) << size_shift_);
+  return &data_[core_idx];
+}
+
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/util/coro_utils.h b/src/rocksdb/util/coro_utils.h
new file mode 100644
index 000000000..5b4211135
--- /dev/null
+++ b/src/rocksdb/util/coro_utils.h
@@ -0,0 +1,112 @@
+//  Copyright (c) Meta Platforms, Inc. and affiliates.
+//
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#if defined(USE_COROUTINES)
+#include "folly/experimental/coro/Coroutine.h"
+#include "folly/experimental/coro/Task.h"
+#endif
+#include "rocksdb/rocksdb_namespace.h"
+
+// This file has two sctions. The first section applies to all instances of
+// header file inclusion and has an include guard. The second section is
+// meant for multiple inclusions in the same source file, and is idempotent.
+namespace ROCKSDB_NAMESPACE {
+
+#ifndef UTIL_CORO_UTILS_H_
+#define UTIL_CORO_UTILS_H_
+
+#if defined(USE_COROUTINES)
+
+// The follwoing macros expand to regular and coroutine function
+// declarations for a given function
+#define DECLARE_SYNC_AND_ASYNC(__ret_type__, __func_name__, ...) \
+  __ret_type__ __func_name__(__VA_ARGS__);                       \
+  folly::coro::Task<__ret_type__> __func_name__##Coroutine(__VA_ARGS__);
+
+#define DECLARE_SYNC_AND_ASYNC_OVERRIDE(__ret_type__, __func_name__, ...) \
+  __ret_type__ __func_name__(__VA_ARGS__) override;                       \
+  folly::coro::Task<__ret_type__> __func_name__##Coroutine(__VA_ARGS__)   \
+      override;
+
+#define DECLARE_SYNC_AND_ASYNC_CONST(__ret_type__, __func_name__, ...) \
+  __ret_type__ __func_name__(__VA_ARGS__) const;                       \
+  folly::coro::Task<__ret_type__> __func_name__##Coroutine(__VA_ARGS__) const;
+
+constexpr bool using_coroutines() { return true; }
+#else  // !USE_COROUTINES
+
+// The follwoing macros expand to a regular function declaration for a given
+// function
+#define DECLARE_SYNC_AND_ASYNC(__ret_type__, __func_name__, ...) \
+  __ret_type__ __func_name__(__VA_ARGS__);
+
+#define DECLARE_SYNC_AND_ASYNC_OVERRIDE(__ret_type__, __func_name__, ...) \
+  __ret_type__ __func_name__(__VA_ARGS__) override;
+
+#define DECLARE_SYNC_AND_ASYNC_CONST(__ret_type__, __func_name__, ...) \
+  __ret_type__ __func_name__(__VA_ARGS__) const;
+
+constexpr bool using_coroutines() { return false; }
+#endif  // USE_COROUTINES
+#endif  // UTIL_CORO_UTILS_H_
+
+// The following section of the file is meant to be included twice in a
+// source file - once defining WITH_COROUTINES and once defining
+// WITHOUT_COROUTINES
+#undef DEFINE_SYNC_AND_ASYNC
+#undef CO_AWAIT
+#undef CO_RETURN
+
+#if defined(WITH_COROUTINES) && defined(USE_COROUTINES)
+
+// This macro should be used in the beginning of the function
+// definition. The declaration should have been done using one of the
+// DECLARE_SYNC_AND_ASYNC* macros. It expands to the return type and
+// the function name with the Coroutine suffix. For example -
+// DEFINE_SYNC_AND_ASYNC(int, foo)(bool bar) {}
+// would expand to -
+// folly::coro::Task<int> fooCoroutine(bool bar) {}
+#define DEFINE_SYNC_AND_ASYNC(__ret_type__, __func_name__) \
+  folly::coro::Task<__ret_type__> __func_name__##Coroutine
+
+// This macro should be used to call a function that might be a
+// coroutine. It expands to the correct function name and prefixes
+// the co_await operator if necessary. For example -
+// s = CO_AWAIT(foo)(true);
+// if the code is compiled WITH_COROUTINES, would expand to
+// s = co_await fooCoroutine(true);
+// if compiled WITHOUT_COROUTINES, would expand to
+// s = foo(true);
+#define CO_AWAIT(__func_name__) co_await __func_name__##Coroutine
+
+#define CO_RETURN co_return
+
+#elif defined(WITHOUT_COROUTINES)
+
+// This macro should be used in the beginning of the function
+// definition. The declaration should have been done using one of the
+// DECLARE_SYNC_AND_ASYNC* macros. It expands to the return type and
+// the function name without the Coroutine suffix. For example -
+// DEFINE_SYNC_AND_ASYNC(int, foo)(bool bar) {}
+// would expand to -
+// int foo(bool bar) {}
+#define DEFINE_SYNC_AND_ASYNC(__ret_type__, __func_name__) \
+  __ret_type__ __func_name__
+
+// This macro should be used to call a function that might be a
+// coroutine. It expands to the correct function name and prefixes
+// the co_await operator if necessary. For example -
+// s = CO_AWAIT(foo)(true);
+// if the code is compiled WITH_COROUTINES, would expand to
+// s = co_await fooCoroutine(true);
+// if compiled WITHOUT_COROUTINES, would expand to
+// s = foo(true);
+#define CO_AWAIT(__func_name__) __func_name__
+
+#define CO_RETURN return
+
+#endif  // DO_NOT_USE_COROUTINES
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/util/crc32c.cc b/src/rocksdb/util/crc32c.cc
new file mode 100644
index 000000000..d71c71c2e
--- /dev/null
+++ b/src/rocksdb/util/crc32c.cc
@@ -0,0 +1,1351 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+//
+// Copyright (c) 2011 The LevelDB Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file. See the AUTHORS file for names of contributors.
+//
+// A portable implementation of crc32c, optimized to handle
+// four bytes at a time.
+#include "util/crc32c.h"
+
+#include <stdint.h>
+
+#include <array>
+#include <utility>
+#ifdef HAVE_SSE42
+#include <nmmintrin.h>
+#include <wmmintrin.h>
+#endif
+
+#include "port/lang.h"
+#include "util/coding.h"
+#include "util/crc32c_arm64.h"
+#include "util/math.h"
+
+#ifdef __powerpc64__
+#include "util/crc32c_ppc.h"
+#include "util/crc32c_ppc_constants.h"
+
+#if __linux__
+#ifdef ROCKSDB_AUXV_GETAUXVAL_PRESENT
+#include <sys/auxv.h>
+#endif
+
+#ifndef PPC_FEATURE2_VEC_CRYPTO
+#define PPC_FEATURE2_VEC_CRYPTO 0x02000000
+#endif
+
+#ifndef AT_HWCAP2
+#define AT_HWCAP2 26
+#endif
+
+#elif __FreeBSD__
+#include <machine/cpu.h>
+#include <sys/auxv.h>
+#include <sys/elf_common.h>
+#endif /* __linux__ */
+
+#endif
+
+#if defined(HAVE_ARM64_CRC)
+bool pmull_runtime_flag = false;
+#endif
+
+namespace ROCKSDB_NAMESPACE {
+namespace crc32c {
+
+#if defined(HAVE_POWER8) && defined(HAS_ALTIVEC)
+#ifdef __powerpc64__
+static int arch_ppc_crc32 = 0;
+#endif /* __powerpc64__ */
+#endif
+
+static const uint32_t table0_[256] = {
+    0x00000000, 0xf26b8303, 0xe13b70f7, 0x1350f3f4, 0xc79a971f, 0x35f1141c,
+    0x26a1e7e8, 0xd4ca64eb, 0x8ad958cf, 0x78b2dbcc, 0x6be22838, 0x9989ab3b,
+    0x4d43cfd0, 0xbf284cd3, 0xac78bf27, 0x5e133c24, 0x105ec76f, 0xe235446c,
+    0xf165b798, 0x030e349b, 0xd7c45070, 0x25afd373, 0x36ff2087, 0xc494a384,
+    0x9a879fa0, 0x68ec1ca3, 0x7bbcef57, 0x89d76c54, 0x5d1d08bf, 0xaf768bbc,
+    0xbc267848, 0x4e4dfb4b, 0x20bd8ede, 0xd2d60ddd, 0xc186fe29, 0x33ed7d2a,
+    0xe72719c1, 0x154c9ac2, 0x061c6936, 0xf477ea35, 0xaa64d611, 0x580f5512,
+    0x4b5fa6e6, 0xb93425e5, 0x6dfe410e, 0x9f95c20d, 0x8cc531f9, 0x7eaeb2fa,
+    0x30e349b1, 0xc288cab2, 0xd1d83946, 0x23b3ba45, 0xf779deae, 0x05125dad,
+    0x1642ae59, 0xe4292d5a, 0xba3a117e, 0x4851927d, 0x5b016189, 0xa96ae28a,
+    0x7da08661, 0x8fcb0562, 0x9c9bf696, 0x6ef07595, 0x417b1dbc, 0xb3109ebf,
+    0xa0406d4b, 0x522bee48, 0x86e18aa3, 0x748a09a0, 0x67dafa54, 0x95b17957,
+    0xcba24573, 0x39c9c670, 0x2a993584, 0xd8f2b687, 0x0c38d26c, 0xfe53516f,
+    0xed03a29b, 0x1f682198, 0x5125dad3, 0xa34e59d0, 0xb01eaa24, 0x42752927,
+    0x96bf4dcc, 0x64d4cecf, 0x77843d3b, 0x85efbe38, 0xdbfc821c, 0x2997011f,
+    0x3ac7f2eb, 0xc8ac71e8, 0x1c661503, 0xee0d9600, 0xfd5d65f4, 0x0f36e6f7,
+    0x61c69362, 0x93ad1061, 0x80fde395, 0x72966096, 0xa65c047d, 0x5437877e,
+    0x4767748a, 0xb50cf789, 0xeb1fcbad, 0x197448ae, 0x0a24bb5a, 0xf84f3859,
+    0x2c855cb2, 0xdeeedfb1, 0xcdbe2c45, 0x3fd5af46, 0x7198540d, 0x83f3d70e,
+    0x90a324fa, 0x62c8a7f9, 0xb602c312, 0x44694011, 0x5739b3e5, 0xa55230e6,
+    0xfb410cc2, 0x092a8fc1, 0x1a7a7c35, 0xe811ff36, 0x3cdb9bdd, 0xceb018de,
+    0xdde0eb2a, 0x2f8b6829, 0x82f63b78, 0x709db87b, 0x63cd4b8f, 0x91a6c88c,
+    0x456cac67, 0xb7072f64, 0xa457dc90, 0x563c5f93, 0x082f63b7, 0xfa44e0b4,
+    0xe9141340, 0x1b7f9043, 0xcfb5f4a8, 0x3dde77ab, 0x2e8e845f, 0xdce5075c,
+    0x92a8fc17, 0x60c37f14, 0x73938ce0, 0x81f80fe3, 0x55326b08, 0xa759e80b,
+    0xb4091bff, 0x466298fc, 0x1871a4d8, 0xea1a27db, 0xf94ad42f, 0x0b21572c,
+    0xdfeb33c7, 0x2d80b0c4, 0x3ed04330, 0xccbbc033, 0xa24bb5a6, 0x502036a5,
+    0x4370c551, 0xb11b4652, 0x65d122b9, 0x97baa1ba, 0x84ea524e, 0x7681d14d,
+    0x2892ed69, 0xdaf96e6a, 0xc9a99d9e, 0x3bc21e9d, 0xef087a76, 0x1d63f975,
+    0x0e330a81, 0xfc588982, 0xb21572c9, 0x407ef1ca, 0x532e023e, 0xa145813d,
+    0x758fe5d6, 0x87e466d5, 0x94b49521, 0x66df1622, 0x38cc2a06, 0xcaa7a905,
+    0xd9f75af1, 0x2b9cd9f2, 0xff56bd19, 0x0d3d3e1a, 0x1e6dcdee, 0xec064eed,
+    0xc38d26c4, 0x31e6a5c7, 0x22b65633, 0xd0ddd530, 0x0417b1db, 0xf67c32d8,
+    0xe52cc12c, 0x1747422f, 0x49547e0b, 0xbb3ffd08, 0xa86f0efc, 0x5a048dff,
+    0x8ecee914, 0x7ca56a17, 0x6ff599e3, 0x9d9e1ae0, 0xd3d3e1ab, 0x21b862a8,
+    0x32e8915c, 0xc083125f, 0x144976b4, 0xe622f5b7, 0xf5720643, 0x07198540,
+    0x590ab964, 0xab613a67, 0xb831c993, 0x4a5a4a90, 0x9e902e7b, 0x6cfbad78,
+    0x7fab5e8c, 0x8dc0dd8f, 0xe330a81a, 0x115b2b19, 0x020bd8ed, 0xf0605bee,
+    0x24aa3f05, 0xd6c1bc06, 0xc5914ff2, 0x37faccf1, 0x69e9f0d5, 0x9b8273d6,
+    0x88d28022, 0x7ab90321, 0xae7367ca, 0x5c18e4c9, 0x4f48173d, 0xbd23943e,
+    0xf36e6f75, 0x0105ec76, 0x12551f82, 0xe03e9c81, 0x34f4f86a, 0xc69f7b69,
+    0xd5cf889d, 0x27a40b9e, 0x79b737ba, 0x8bdcb4b9, 0x988c474d, 0x6ae7c44e,
+    0xbe2da0a5, 0x4c4623a6, 0x5f16d052, 0xad7d5351};
+static const uint32_t table1_[256] = {
+    0x00000000, 0x13a29877, 0x274530ee, 0x34e7a899, 0x4e8a61dc, 0x5d28f9ab,
+    0x69cf5132, 0x7a6dc945, 0x9d14c3b8, 0x8eb65bcf, 0xba51f356, 0xa9f36b21,
+    0xd39ea264, 0xc03c3a13, 0xf4db928a, 0xe7790afd, 0x3fc5f181, 0x2c6769f6,
+    0x1880c16f, 0x0b225918, 0x714f905d, 0x62ed082a, 0x560aa0b3, 0x45a838c4,
+    0xa2d13239, 0xb173aa4e, 0x859402d7, 0x96369aa0, 0xec5b53e5, 0xfff9cb92,
+    0xcb1e630b, 0xd8bcfb7c, 0x7f8be302, 0x6c297b75, 0x58ced3ec, 0x4b6c4b9b,
+    0x310182de, 0x22a31aa9, 0x1644b230, 0x05e62a47, 0xe29f20ba, 0xf13db8cd,
+    0xc5da1054, 0xd6788823, 0xac154166, 0xbfb7d911, 0x8b507188, 0x98f2e9ff,
+    0x404e1283, 0x53ec8af4, 0x670b226d, 0x74a9ba1a, 0x0ec4735f, 0x1d66eb28,
+    0x298143b1, 0x3a23dbc6, 0xdd5ad13b, 0xcef8494c, 0xfa1fe1d5, 0xe9bd79a2,
+    0x93d0b0e7, 0x80722890, 0xb4958009, 0xa737187e, 0xff17c604, 0xecb55e73,
+    0xd852f6ea, 0xcbf06e9d, 0xb19da7d8, 0xa23f3faf, 0x96d89736, 0x857a0f41,
+    0x620305bc, 0x71a19dcb, 0x45463552, 0x56e4ad25, 0x2c896460, 0x3f2bfc17,
+    0x0bcc548e, 0x186eccf9, 0xc0d23785, 0xd370aff2, 0xe797076b, 0xf4359f1c,
+    0x8e585659, 0x9dface2e, 0xa91d66b7, 0xbabffec0, 0x5dc6f43d, 0x4e646c4a,
+    0x7a83c4d3, 0x69215ca4, 0x134c95e1, 0x00ee0d96, 0x3409a50f, 0x27ab3d78,
+    0x809c2506, 0x933ebd71, 0xa7d915e8, 0xb47b8d9f, 0xce1644da, 0xddb4dcad,
+    0xe9537434, 0xfaf1ec43, 0x1d88e6be, 0x0e2a7ec9, 0x3acdd650, 0x296f4e27,
+    0x53028762, 0x40a01f15, 0x7447b78c, 0x67e52ffb, 0xbf59d487, 0xacfb4cf0,
+    0x981ce469, 0x8bbe7c1e, 0xf1d3b55b, 0xe2712d2c, 0xd69685b5, 0xc5341dc2,
+    0x224d173f, 0x31ef8f48, 0x050827d1, 0x16aabfa6, 0x6cc776e3, 0x7f65ee94,
+    0x4b82460d, 0x5820de7a, 0xfbc3faf9, 0xe861628e, 0xdc86ca17, 0xcf245260,
+    0xb5499b25, 0xa6eb0352, 0x920cabcb, 0x81ae33bc, 0x66d73941, 0x7575a136,
+    0x419209af, 0x523091d8, 0x285d589d, 0x3bffc0ea, 0x0f186873, 0x1cbaf004,
+    0xc4060b78, 0xd7a4930f, 0xe3433b96, 0xf0e1a3e1, 0x8a8c6aa4, 0x992ef2d3,
+    0xadc95a4a, 0xbe6bc23d, 0x5912c8c0, 0x4ab050b7, 0x7e57f82e, 0x6df56059,
+    0x1798a91c, 0x043a316b, 0x30dd99f2, 0x237f0185, 0x844819fb, 0x97ea818c,
+    0xa30d2915, 0xb0afb162, 0xcac27827, 0xd960e050, 0xed8748c9, 0xfe25d0be,
+    0x195cda43, 0x0afe4234, 0x3e19eaad, 0x2dbb72da, 0x57d6bb9f, 0x447423e8,
+    0x70938b71, 0x63311306, 0xbb8de87a, 0xa82f700d, 0x9cc8d894, 0x8f6a40e3,
+    0xf50789a6, 0xe6a511d1, 0xd242b948, 0xc1e0213f, 0x26992bc2, 0x353bb3b5,
+    0x01dc1b2c, 0x127e835b, 0x68134a1e, 0x7bb1d269, 0x4f567af0, 0x5cf4e287,
+    0x04d43cfd, 0x1776a48a, 0x23910c13, 0x30339464, 0x4a5e5d21, 0x59fcc556,
+    0x6d1b6dcf, 0x7eb9f5b8, 0x99c0ff45, 0x8a626732, 0xbe85cfab, 0xad2757dc,
+    0xd74a9e99, 0xc4e806ee, 0xf00fae77, 0xe3ad3600, 0x3b11cd7c, 0x28b3550b,
+    0x1c54fd92, 0x0ff665e5, 0x759baca0, 0x663934d7, 0x52de9c4e, 0x417c0439,
+    0xa6050ec4, 0xb5a796b3, 0x81403e2a, 0x92e2a65d, 0xe88f6f18, 0xfb2df76f,
+    0xcfca5ff6, 0xdc68c781, 0x7b5fdfff, 0x68fd4788, 0x5c1aef11, 0x4fb87766,
+    0x35d5be23, 0x26772654, 0x12908ecd, 0x013216ba, 0xe64b1c47, 0xf5e98430,
+    0xc10e2ca9, 0xd2acb4de, 0xa8c17d9b, 0xbb63e5ec, 0x8f844d75, 0x9c26d502,
+    0x449a2e7e, 0x5738b609, 0x63df1e90, 0x707d86e7, 0x0a104fa2, 0x19b2d7d5,
+    0x2d557f4c, 0x3ef7e73b, 0xd98eedc6, 0xca2c75b1, 0xfecbdd28, 0xed69455f,
+    0x97048c1a, 0x84a6146d, 0xb041bcf4, 0xa3e32483};
+static const uint32_t table2_[256] = {
+    0x00000000, 0xa541927e, 0x4f6f520d, 0xea2ec073, 0x9edea41a, 0x3b9f3664,
+    0xd1b1f617, 0x74f06469, 0x38513ec5, 0x9d10acbb, 0x773e6cc8, 0xd27ffeb6,
+    0xa68f9adf, 0x03ce08a1, 0xe9e0c8d2, 0x4ca15aac, 0x70a27d8a, 0xd5e3eff4,
+    0x3fcd2f87, 0x9a8cbdf9, 0xee7cd990, 0x4b3d4bee, 0xa1138b9d, 0x045219e3,
+    0x48f3434f, 0xedb2d131, 0x079c1142, 0xa2dd833c, 0xd62de755, 0x736c752b,
+    0x9942b558, 0x3c032726, 0xe144fb14, 0x4405696a, 0xae2ba919, 0x0b6a3b67,
+    0x7f9a5f0e, 0xdadbcd70, 0x30f50d03, 0x95b49f7d, 0xd915c5d1, 0x7c5457af,
+    0x967a97dc, 0x333b05a2, 0x47cb61cb, 0xe28af3b5, 0x08a433c6, 0xade5a1b8,
+    0x91e6869e, 0x34a714e0, 0xde89d493, 0x7bc846ed, 0x0f382284, 0xaa79b0fa,
+    0x40577089, 0xe516e2f7, 0xa9b7b85b, 0x0cf62a25, 0xe6d8ea56, 0x43997828,
+    0x37691c41, 0x92288e3f, 0x78064e4c, 0xdd47dc32, 0xc76580d9, 0x622412a7,
+    0x880ad2d4, 0x2d4b40aa, 0x59bb24c3, 0xfcfab6bd, 0x16d476ce, 0xb395e4b0,
+    0xff34be1c, 0x5a752c62, 0xb05bec11, 0x151a7e6f, 0x61ea1a06, 0xc4ab8878,
+    0x2e85480b, 0x8bc4da75, 0xb7c7fd53, 0x12866f2d, 0xf8a8af5e, 0x5de93d20,
+    0x29195949, 0x8c58cb37, 0x66760b44, 0xc337993a, 0x8f96c396, 0x2ad751e8,
+    0xc0f9919b, 0x65b803e5, 0x1148678c, 0xb409f5f2, 0x5e273581, 0xfb66a7ff,
+    0x26217bcd, 0x8360e9b3, 0x694e29c0, 0xcc0fbbbe, 0xb8ffdfd7, 0x1dbe4da9,
+    0xf7908dda, 0x52d11fa4, 0x1e704508, 0xbb31d776, 0x511f1705, 0xf45e857b,
+    0x80aee112, 0x25ef736c, 0xcfc1b31f, 0x6a802161, 0x56830647, 0xf3c29439,
+    0x19ec544a, 0xbcadc634, 0xc85da25d, 0x6d1c3023, 0x8732f050, 0x2273622e,
+    0x6ed23882, 0xcb93aafc, 0x21bd6a8f, 0x84fcf8f1, 0xf00c9c98, 0x554d0ee6,
+    0xbf63ce95, 0x1a225ceb, 0x8b277743, 0x2e66e53d, 0xc448254e, 0x6109b730,
+    0x15f9d359, 0xb0b84127, 0x5a968154, 0xffd7132a, 0xb3764986, 0x1637dbf8,
+    0xfc191b8b, 0x595889f5, 0x2da8ed9c, 0x88e97fe2, 0x62c7bf91, 0xc7862def,
+    0xfb850ac9, 0x5ec498b7, 0xb4ea58c4, 0x11abcaba, 0x655baed3, 0xc01a3cad,
+    0x2a34fcde, 0x8f756ea0, 0xc3d4340c, 0x6695a672, 0x8cbb6601, 0x29faf47f,
+    0x5d0a9016, 0xf84b0268, 0x1265c21b, 0xb7245065, 0x6a638c57, 0xcf221e29,
+    0x250cde5a, 0x804d4c24, 0xf4bd284d, 0x51fcba33, 0xbbd27a40, 0x1e93e83e,
+    0x5232b292, 0xf77320ec, 0x1d5de09f, 0xb81c72e1, 0xccec1688, 0x69ad84f6,
+    0x83834485, 0x26c2d6fb, 0x1ac1f1dd, 0xbf8063a3, 0x55aea3d0, 0xf0ef31ae,
+    0x841f55c7, 0x215ec7b9, 0xcb7007ca, 0x6e3195b4, 0x2290cf18, 0x87d15d66,
+    0x6dff9d15, 0xc8be0f6b, 0xbc4e6b02, 0x190ff97c, 0xf321390f, 0x5660ab71,
+    0x4c42f79a, 0xe90365e4, 0x032da597, 0xa66c37e9, 0xd29c5380, 0x77ddc1fe,
+    0x9df3018d, 0x38b293f3, 0x7413c95f, 0xd1525b21, 0x3b7c9b52, 0x9e3d092c,
+    0xeacd6d45, 0x4f8cff3b, 0xa5a23f48, 0x00e3ad36, 0x3ce08a10, 0x99a1186e,
+    0x738fd81d, 0xd6ce4a63, 0xa23e2e0a, 0x077fbc74, 0xed517c07, 0x4810ee79,
+    0x04b1b4d5, 0xa1f026ab, 0x4bdee6d8, 0xee9f74a6, 0x9a6f10cf, 0x3f2e82b1,
+    0xd50042c2, 0x7041d0bc, 0xad060c8e, 0x08479ef0, 0xe2695e83, 0x4728ccfd,
+    0x33d8a894, 0x96993aea, 0x7cb7fa99, 0xd9f668e7, 0x9557324b, 0x3016a035,
+    0xda386046, 0x7f79f238, 0x0b899651, 0xaec8042f, 0x44e6c45c, 0xe1a75622,
+    0xdda47104, 0x78e5e37a, 0x92cb2309, 0x378ab177, 0x437ad51e, 0xe63b4760,
+    0x0c158713, 0xa954156d, 0xe5f54fc1, 0x40b4ddbf, 0xaa9a1dcc, 0x0fdb8fb2,
+    0x7b2bebdb, 0xde6a79a5, 0x3444b9d6, 0x91052ba8};
+static const uint32_t table3_[256] = {
+    0x00000000, 0xdd45aab8, 0xbf672381, 0x62228939, 0x7b2231f3, 0xa6679b4b,
+    0xc4451272, 0x1900b8ca, 0xf64463e6, 0x2b01c95e, 0x49234067, 0x9466eadf,
+    0x8d665215, 0x5023f8ad, 0x32017194, 0xef44db2c, 0xe964b13d, 0x34211b85,
+    0x560392bc, 0x8b463804, 0x924680ce, 0x4f032a76, 0x2d21a34f, 0xf06409f7,
+    0x1f20d2db, 0xc2657863, 0xa047f15a, 0x7d025be2, 0x6402e328, 0xb9474990,
+    0xdb65c0a9, 0x06206a11, 0xd725148b, 0x0a60be33, 0x6842370a, 0xb5079db2,
+    0xac072578, 0x71428fc0, 0x136006f9, 0xce25ac41, 0x2161776d, 0xfc24ddd5,
+    0x9e0654ec, 0x4343fe54, 0x5a43469e, 0x8706ec26, 0xe524651f, 0x3861cfa7,
+    0x3e41a5b6, 0xe3040f0e, 0x81268637, 0x5c632c8f, 0x45639445, 0x98263efd,
+    0xfa04b7c4, 0x27411d7c, 0xc805c650, 0x15406ce8, 0x7762e5d1, 0xaa274f69,
+    0xb327f7a3, 0x6e625d1b, 0x0c40d422, 0xd1057e9a, 0xaba65fe7, 0x76e3f55f,
+    0x14c17c66, 0xc984d6de, 0xd0846e14, 0x0dc1c4ac, 0x6fe34d95, 0xb2a6e72d,
+    0x5de23c01, 0x80a796b9, 0xe2851f80, 0x3fc0b538, 0x26c00df2, 0xfb85a74a,
+    0x99a72e73, 0x44e284cb, 0x42c2eeda, 0x9f874462, 0xfda5cd5b, 0x20e067e3,
+    0x39e0df29, 0xe4a57591, 0x8687fca8, 0x5bc25610, 0xb4868d3c, 0x69c32784,
+    0x0be1aebd, 0xd6a40405, 0xcfa4bccf, 0x12e11677, 0x70c39f4e, 0xad8635f6,
+    0x7c834b6c, 0xa1c6e1d4, 0xc3e468ed, 0x1ea1c255, 0x07a17a9f, 0xdae4d027,
+    0xb8c6591e, 0x6583f3a6, 0x8ac7288a, 0x57828232, 0x35a00b0b, 0xe8e5a1b3,
+    0xf1e51979, 0x2ca0b3c1, 0x4e823af8, 0x93c79040, 0x95e7fa51, 0x48a250e9,
+    0x2a80d9d0, 0xf7c57368, 0xeec5cba2, 0x3380611a, 0x51a2e823, 0x8ce7429b,
+    0x63a399b7, 0xbee6330f, 0xdcc4ba36, 0x0181108e, 0x1881a844, 0xc5c402fc,
+    0xa7e68bc5, 0x7aa3217d, 0x52a0c93f, 0x8fe56387, 0xedc7eabe, 0x30824006,
+    0x2982f8cc, 0xf4c75274, 0x96e5db4d, 0x4ba071f5, 0xa4e4aad9, 0x79a10061,
+    0x1b838958, 0xc6c623e0, 0xdfc69b2a, 0x02833192, 0x60a1b8ab, 0xbde41213,
+    0xbbc47802, 0x6681d2ba, 0x04a35b83, 0xd9e6f13b, 0xc0e649f1, 0x1da3e349,
+    0x7f816a70, 0xa2c4c0c8, 0x4d801be4, 0x90c5b15c, 0xf2e73865, 0x2fa292dd,
+    0x36a22a17, 0xebe780af, 0x89c50996, 0x5480a32e, 0x8585ddb4, 0x58c0770c,
+    0x3ae2fe35, 0xe7a7548d, 0xfea7ec47, 0x23e246ff, 0x41c0cfc6, 0x9c85657e,
+    0x73c1be52, 0xae8414ea, 0xcca69dd3, 0x11e3376b, 0x08e38fa1, 0xd5a62519,
+    0xb784ac20, 0x6ac10698, 0x6ce16c89, 0xb1a4c631, 0xd3864f08, 0x0ec3e5b0,
+    0x17c35d7a, 0xca86f7c2, 0xa8a47efb, 0x75e1d443, 0x9aa50f6f, 0x47e0a5d7,
+    0x25c22cee, 0xf8878656, 0xe1873e9c, 0x3cc29424, 0x5ee01d1d, 0x83a5b7a5,
+    0xf90696d8, 0x24433c60, 0x4661b559, 0x9b241fe1, 0x8224a72b, 0x5f610d93,
+    0x3d4384aa, 0xe0062e12, 0x0f42f53e, 0xd2075f86, 0xb025d6bf, 0x6d607c07,
+    0x7460c4cd, 0xa9256e75, 0xcb07e74c, 0x16424df4, 0x106227e5, 0xcd278d5d,
+    0xaf050464, 0x7240aedc, 0x6b401616, 0xb605bcae, 0xd4273597, 0x09629f2f,
+    0xe6264403, 0x3b63eebb, 0x59416782, 0x8404cd3a, 0x9d0475f0, 0x4041df48,
+    0x22635671, 0xff26fcc9, 0x2e238253, 0xf36628eb, 0x9144a1d2, 0x4c010b6a,
+    0x5501b3a0, 0x88441918, 0xea669021, 0x37233a99, 0xd867e1b5, 0x05224b0d,
+    0x6700c234, 0xba45688c, 0xa345d046, 0x7e007afe, 0x1c22f3c7, 0xc167597f,
+    0xc747336e, 0x1a0299d6, 0x782010ef, 0xa565ba57, 0xbc65029d, 0x6120a825,
+    0x0302211c, 0xde478ba4, 0x31035088, 0xec46fa30, 0x8e647309, 0x5321d9b1,
+    0x4a21617b, 0x9764cbc3, 0xf54642fa, 0x2803e842};
+
+// Used to fetch a naturally-aligned 32-bit word in little endian byte-order
+static inline uint32_t LE_LOAD32(const uint8_t* p) {
+  return DecodeFixed32(reinterpret_cast<const char*>(p));
+}
+
+#if defined(HAVE_SSE42) && (defined(__LP64__) || defined(_WIN64))
+static inline uint64_t LE_LOAD64(const uint8_t* p) {
+  return DecodeFixed64(reinterpret_cast<const char*>(p));
+}
+#endif
+
+static inline void Slow_CRC32(uint64_t* l, uint8_t const** p) {
+  uint32_t c = static_cast<uint32_t>(*l ^ LE_LOAD32(*p));
+  *p += 4;
+  *l = table3_[c & 0xff] ^ table2_[(c >> 8) & 0xff] ^
+       table1_[(c >> 16) & 0xff] ^ table0_[c >> 24];
+  // DO it twice.
+  c = static_cast<uint32_t>(*l ^ LE_LOAD32(*p));
+  *p += 4;
+  *l = table3_[c & 0xff] ^ table2_[(c >> 8) & 0xff] ^
+       table1_[(c >> 16) & 0xff] ^ table0_[c >> 24];
+}
+
+#if (!(defined(HAVE_POWER8) && defined(HAS_ALTIVEC))) && \
+        (!defined(HAVE_ARM64_CRC)) ||                    \
+    defined(NO_THREEWAY_CRC32C)
+static inline void Fast_CRC32(uint64_t* l, uint8_t const** p) {
+#ifndef HAVE_SSE42
+  Slow_CRC32(l, p);
+#elif defined(__LP64__) || defined(_WIN64)
+  *l = _mm_crc32_u64(*l, LE_LOAD64(*p));
+  *p += 8;
+#else
+  *l = _mm_crc32_u32(static_cast<unsigned int>(*l), LE_LOAD32(*p));
+  *p += 4;
+  *l = _mm_crc32_u32(static_cast<unsigned int>(*l), LE_LOAD32(*p));
+  *p += 4;
+#endif
+}
+#endif
+
+template <void (*CRC32)(uint64_t*, uint8_t const**)>
+uint32_t ExtendImpl(uint32_t crc, const char* buf, size_t size) {
+  const uint8_t* p = reinterpret_cast<const uint8_t*>(buf);
+  const uint8_t* e = p + size;
+  uint64_t l = crc ^ 0xffffffffu;
+
+// Align n to (1 << m) byte boundary
+#define ALIGN(n, m) ((n + ((1 << m) - 1)) & ~((1 << m) - 1))
+
+#define STEP1                  \
+  do {                         \
+    int c = (l & 0xff) ^ *p++; \
+    l = table0_[c] ^ (l >> 8); \
+  } while (0)
+
+  // Point x at first 16-byte aligned byte in string.  This might be
+  // just past the end of the string.
+  const uintptr_t pval = reinterpret_cast<uintptr_t>(p);
+  const uint8_t* x = reinterpret_cast<const uint8_t*>(ALIGN(pval, 4));
+  if (x <= e) {
+    // Process bytes until finished or p is 16-byte aligned
+    while (p != x) {
+      STEP1;
+    }
+  }
+  // Process bytes 16 at a time
+  while ((e - p) >= 16) {
+    CRC32(&l, &p);
+    CRC32(&l, &p);
+  }
+  // Process bytes 8 at a time
+  while ((e - p) >= 8) {
+    CRC32(&l, &p);
+  }
+  // Process the last few bytes
+  while (p != e) {
+    STEP1;
+  }
+#undef STEP1
+#undef ALIGN
+  return static_cast<uint32_t>(l ^ 0xffffffffu);
+}
+
+// Detect if ARM64 CRC or not.
+#ifndef HAVE_ARM64_CRC
+// Detect if SS42 or not.
+#ifndef HAVE_POWER8
+
+static bool isSSE42() {
+#ifndef HAVE_SSE42
+  return false;
+#elif defined(__GNUC__) && defined(__x86_64__) && !defined(IOS_CROSS_COMPILE)
+  uint32_t c_;
+  __asm__("cpuid" : "=c"(c_) : "a"(1) : "ebx", "edx");
+  return c_ & (1U << 20);  // copied from CpuId.h in Folly. Test SSE42
+#elif defined(_WIN64)
+  int info[4];
+  __cpuidex(info, 0x00000001, 0);
+  return (info[2] & ((int)1 << 20)) != 0;
+#else
+  return false;
+#endif
+}
+
+static bool isPCLMULQDQ() {
+#ifndef HAVE_SSE42
+  // in build_detect_platform we set this macro when both SSE42 and PCLMULQDQ
+  // are supported by compiler
+  return false;
+#elif defined(__GNUC__) && defined(__x86_64__) && !defined(IOS_CROSS_COMPILE)
+  uint32_t c_;
+  __asm__("cpuid" : "=c"(c_) : "a"(1) : "ebx", "edx");
+  return c_ & (1U << 1);  // PCLMULQDQ is in bit 1 (not bit 0)
+#elif defined(_WIN64)
+  int info[4];
+  __cpuidex(info, 0x00000001, 0);
+  return (info[2] & ((int)1 << 1)) != 0;
+#else
+  return false;
+#endif
+}
+
+#endif  // HAVE_POWER8
+#endif  // HAVE_ARM64_CRC
+
+using Function = uint32_t (*)(uint32_t, const char*, size_t);
+
+#if defined(HAVE_POWER8) && defined(HAS_ALTIVEC)
+uint32_t ExtendPPCImpl(uint32_t crc, const char* buf, size_t size) {
+  return crc32c_ppc(crc, (const unsigned char*)buf, size);
+}
+
+#if __linux__
+static int arch_ppc_probe(void) {
+  arch_ppc_crc32 = 0;
+
+#if defined(__powerpc64__) && defined(ROCKSDB_AUXV_GETAUXVAL_PRESENT)
+  if (getauxval(AT_HWCAP2) & PPC_FEATURE2_VEC_CRYPTO) arch_ppc_crc32 = 1;
+#endif /* __powerpc64__ */
+
+  return arch_ppc_crc32;
+}
+#elif __FreeBSD__
+static int arch_ppc_probe(void) {
+  unsigned long cpufeatures;
+  arch_ppc_crc32 = 0;
+
+#if defined(__powerpc64__)
+  elf_aux_info(AT_HWCAP2, &cpufeatures, sizeof(cpufeatures));
+  if (cpufeatures & PPC_FEATURE2_HAS_VEC_CRYPTO) arch_ppc_crc32 = 1;
+#endif  /* __powerpc64__ */
+
+  return arch_ppc_crc32;
+}
+#endif  // __linux__
+
+static bool isAltiVec() {
+  if (arch_ppc_probe()) {
+    return true;
+  } else {
+    return false;
+  }
+}
+#endif
+
+#if defined(HAVE_ARM64_CRC)
+uint32_t ExtendARMImpl(uint32_t crc, const char* buf, size_t size) {
+  return crc32c_arm64(crc, (const unsigned char*)buf, size);
+}
+#endif
+
+std::string IsFastCrc32Supported() {
+  bool has_fast_crc = false;
+  std::string fast_zero_msg;
+  std::string arch;
+#ifdef HAVE_POWER8
+#ifdef HAS_ALTIVEC
+  if (arch_ppc_probe()) {
+    has_fast_crc = true;
+    arch = "PPC";
+  }
+#else
+  has_fast_crc = false;
+  arch = "PPC";
+#endif
+#elif defined(HAVE_ARM64_CRC)
+  if (crc32c_runtime_check()) {
+    has_fast_crc = true;
+    arch = "Arm64";
+    pmull_runtime_flag = crc32c_pmull_runtime_check();
+  } else {
+    has_fast_crc = false;
+    arch = "Arm64";
+  }
+#else
+  has_fast_crc = isSSE42();
+  arch = "x86";
+#endif
+  if (has_fast_crc) {
+    fast_zero_msg.append("Supported on " + arch);
+  } else {
+    fast_zero_msg.append("Not supported on " + arch);
+  }
+  return fast_zero_msg;
+}
+
+/*
+ * Copyright 2016 Ferry Toth, Exalon Delft BV, The Netherlands
+ *  This software is provided 'as-is', without any express or implied
+ * warranty.  In no event will the author be held liable for any damages
+ * arising from the use of this software.
+ *  Permission is granted to anyone to use this software for any purpose,
+ * including commercial applications, and to alter it and redistribute it
+ * freely, subject to the following restrictions:
+ *  1. The origin of this software must not be misrepresented; you must not
+ *   claim that you wrote the original software. If you use this software
+ *   in a product, an acknowledgment in the product documentation would be
+ *   appreciated but is not required.
+ * 2. Altered source versions must be plainly marked as such, and must not be
+ *   misrepresented as being the original software.
+ * 3. This notice may not be removed or altered from any source distribution.
+ *  Ferry Toth
+ * ftoth@exalondelft.nl
+ *
+ * https://github.com/htot/crc32c
+ *
+ * Modified by Facebook
+ *
+ * Original intel whitepaper:
+ * "Fast CRC Computation for iSCSI Polynomial Using CRC32 Instruction"
+ * https://www.intel.com/content/dam/www/public/us/en/documents/white-papers/crc-iscsi-polynomial-crc32-instruction-paper.pdf
+ *
+ * This version is from the folly library, created by Dave Watson
+ * <davejwatson@fb.com>
+ *
+ */
+#if defined HAVE_SSE42 && defined HAVE_PCLMUL
+
+#define CRCtriplet(crc, buf, offset)                  \
+  crc##0 = _mm_crc32_u64(crc##0, *(buf##0 + offset)); \
+  crc##1 = _mm_crc32_u64(crc##1, *(buf##1 + offset)); \
+  crc##2 = _mm_crc32_u64(crc##2, *(buf##2 + offset));
+
+#define CRCduplet(crc, buf, offset)                   \
+  crc##0 = _mm_crc32_u64(crc##0, *(buf##0 + offset)); \
+  crc##1 = _mm_crc32_u64(crc##1, *(buf##1 + offset));
+
+#define CRCsinglet(crc, buf, offset) \
+  crc = _mm_crc32_u64(crc, *(uint64_t*)(buf + offset));
+
+// Numbers taken directly from intel whitepaper.
+// clang-format off
+const uint64_t clmul_constants[] = {
+    0x14cd00bd6, 0x105ec76f0, 0x0ba4fc28e, 0x14cd00bd6,
+    0x1d82c63da, 0x0f20c0dfe, 0x09e4addf8, 0x0ba4fc28e,
+    0x039d3b296, 0x1384aa63a, 0x102f9b8a2, 0x1d82c63da,
+    0x14237f5e6, 0x01c291d04, 0x00d3b6092, 0x09e4addf8,
+    0x0c96cfdc0, 0x0740eef02, 0x18266e456, 0x039d3b296,
+    0x0daece73e, 0x0083a6eec, 0x0ab7aff2a, 0x102f9b8a2,
+    0x1248ea574, 0x1c1733996, 0x083348832, 0x14237f5e6,
+    0x12c743124, 0x02ad91c30, 0x0b9e02b86, 0x00d3b6092,
+    0x018b33a4e, 0x06992cea2, 0x1b331e26a, 0x0c96cfdc0,
+    0x17d35ba46, 0x07e908048, 0x1bf2e8b8a, 0x18266e456,
+    0x1a3e0968a, 0x11ed1f9d8, 0x0ce7f39f4, 0x0daece73e,
+    0x061d82e56, 0x0f1d0f55e, 0x0d270f1a2, 0x0ab7aff2a,
+    0x1c3f5f66c, 0x0a87ab8a8, 0x12ed0daac, 0x1248ea574,
+    0x065863b64, 0x08462d800, 0x11eef4f8e, 0x083348832,
+    0x1ee54f54c, 0x071d111a8, 0x0b3e32c28, 0x12c743124,
+    0x0064f7f26, 0x0ffd852c6, 0x0dd7e3b0c, 0x0b9e02b86,
+    0x0f285651c, 0x0dcb17aa4, 0x010746f3c, 0x018b33a4e,
+    0x1c24afea4, 0x0f37c5aee, 0x0271d9844, 0x1b331e26a,
+    0x08e766a0c, 0x06051d5a2, 0x093a5f730, 0x17d35ba46,
+    0x06cb08e5c, 0x11d5ca20e, 0x06b749fb2, 0x1bf2e8b8a,
+    0x1167f94f2, 0x021f3d99c, 0x0cec3662e, 0x1a3e0968a,
+    0x19329634a, 0x08f158014, 0x0e6fc4e6a, 0x0ce7f39f4,
+    0x08227bb8a, 0x1a5e82106, 0x0b0cd4768, 0x061d82e56,
+    0x13c2b89c4, 0x188815ab2, 0x0d7a4825c, 0x0d270f1a2,
+    0x10f5ff2ba, 0x105405f3e, 0x00167d312, 0x1c3f5f66c,
+    0x0f6076544, 0x0e9adf796, 0x026f6a60a, 0x12ed0daac,
+    0x1a2adb74e, 0x096638b34, 0x19d34af3a, 0x065863b64,
+    0x049c3cc9c, 0x1e50585a0, 0x068bce87a, 0x11eef4f8e,
+    0x1524fa6c6, 0x19f1c69dc, 0x16cba8aca, 0x1ee54f54c,
+    0x042d98888, 0x12913343e, 0x1329d9f7e, 0x0b3e32c28,
+    0x1b1c69528, 0x088f25a3a, 0x02178513a, 0x0064f7f26,
+    0x0e0ac139e, 0x04e36f0b0, 0x0170076fa, 0x0dd7e3b0c,
+    0x141a1a2e2, 0x0bd6f81f8, 0x16ad828b4, 0x0f285651c,
+    0x041d17b64, 0x19425cbba, 0x1fae1cc66, 0x010746f3c,
+    0x1a75b4b00, 0x18db37e8a, 0x0f872e54c, 0x1c24afea4,
+    0x01e41e9fc, 0x04c144932, 0x086d8e4d2, 0x0271d9844,
+    0x160f7af7a, 0x052148f02, 0x05bb8f1bc, 0x08e766a0c,
+    0x0a90fd27a, 0x0a3c6f37a, 0x0b3af077a, 0x093a5f730,
+    0x04984d782, 0x1d22c238e, 0x0ca6ef3ac, 0x06cb08e5c,
+    0x0234e0b26, 0x063ded06a, 0x1d88abd4a, 0x06b749fb2,
+    0x04597456a, 0x04d56973c, 0x0e9e28eb4, 0x1167f94f2,
+    0x07b3ff57a, 0x19385bf2e, 0x0c9c8b782, 0x0cec3662e,
+    0x13a9cba9e, 0x0e417f38a, 0x093e106a4, 0x19329634a,
+    0x167001a9c, 0x14e727980, 0x1ddffc5d4, 0x0e6fc4e6a,
+    0x00df04680, 0x0d104b8fc, 0x02342001e, 0x08227bb8a,
+    0x00a2a8d7e, 0x05b397730, 0x168763fa6, 0x0b0cd4768,
+    0x1ed5a407a, 0x0e78eb416, 0x0d2c3ed1a, 0x13c2b89c4,
+    0x0995a5724, 0x1641378f0, 0x19b1afbc4, 0x0d7a4825c,
+    0x109ffedc0, 0x08d96551c, 0x0f2271e60, 0x10f5ff2ba,
+    0x00b0bf8ca, 0x00bf80dd2, 0x123888b7a, 0x00167d312,
+    0x1e888f7dc, 0x18dcddd1c, 0x002ee03b2, 0x0f6076544,
+    0x183e8d8fe, 0x06a45d2b2, 0x133d7a042, 0x026f6a60a,
+    0x116b0f50c, 0x1dd3e10e8, 0x05fabe670, 0x1a2adb74e,
+    0x130004488, 0x0de87806c, 0x000bcf5f6, 0x19d34af3a,
+    0x18f0c7078, 0x014338754, 0x017f27698, 0x049c3cc9c,
+    0x058ca5f00, 0x15e3e77ee, 0x1af900c24, 0x068bce87a,
+    0x0b5cfca28, 0x0dd07448e, 0x0ded288f8, 0x1524fa6c6,
+    0x059f229bc, 0x1d8048348, 0x06d390dec, 0x16cba8aca,
+    0x037170390, 0x0a3e3e02c, 0x06353c1cc, 0x042d98888,
+    0x0c4584f5c, 0x0d73c7bea, 0x1f16a3418, 0x1329d9f7e,
+    0x0531377e2, 0x185137662, 0x1d8d9ca7c, 0x1b1c69528,
+    0x0b25b29f2, 0x18a08b5bc, 0x19fb2a8b0, 0x02178513a,
+    0x1a08fe6ac, 0x1da758ae0, 0x045cddf4e, 0x0e0ac139e,
+    0x1a91647f2, 0x169cf9eb0, 0x1a0f717c4, 0x0170076fa,
+};
+
+// Compute the crc32c value for buffer smaller than 8
+#ifdef ROCKSDB_UBSAN_RUN
+#if defined(__clang__)
+__attribute__((__no_sanitize__("alignment")))
+#elif defined(__GNUC__)
+__attribute__((__no_sanitize_undefined__))
+#endif
+#endif
+inline void align_to_8(
+    size_t len,
+    uint64_t& crc0, // crc so far, updated on return
+    const unsigned char*& next) { // next data pointer, updated on return
+  uint32_t crc32bit = static_cast<uint32_t>(crc0);
+  if (len & 0x04) {
+    crc32bit = _mm_crc32_u32(crc32bit, *(uint32_t*)next);
+    next += sizeof(uint32_t);
+  }
+  if (len & 0x02) {
+    crc32bit = _mm_crc32_u16(crc32bit, *(uint16_t*)next);
+    next += sizeof(uint16_t);
+  }
+  if (len & 0x01) {
+    crc32bit = _mm_crc32_u8(crc32bit, *(next));
+    next++;
+  }
+  crc0 = crc32bit;
+}
+
+//
+// CombineCRC performs pclmulqdq multiplication of 2 partial CRC's and a well
+// chosen constant and xor's these with the remaining CRC.
+//
+inline uint64_t CombineCRC(
+    size_t block_size,
+    uint64_t crc0,
+    uint64_t crc1,
+    uint64_t crc2,
+    const uint64_t* next2) {
+  const auto multiplier =
+      *(reinterpret_cast<const __m128i*>(clmul_constants) + block_size - 1);
+  const auto crc0_xmm = _mm_set_epi64x(0, crc0);
+  const auto res0 = _mm_clmulepi64_si128(crc0_xmm, multiplier, 0x00);
+  const auto crc1_xmm = _mm_set_epi64x(0, crc1);
+  const auto res1 = _mm_clmulepi64_si128(crc1_xmm, multiplier, 0x10);
+  const auto res = _mm_xor_si128(res0, res1);
+  crc0 = _mm_cvtsi128_si64(res);
+  crc0 = crc0 ^ *((uint64_t*)next2 - 1);
+  crc2 = _mm_crc32_u64(crc2, crc0);
+  return crc2;
+}
+
+// Compute CRC-32C using the Intel hardware instruction.
+#ifdef ROCKSDB_UBSAN_RUN
+#if defined(__clang__)
+__attribute__((__no_sanitize__("alignment")))
+#elif defined(__GNUC__)
+__attribute__((__no_sanitize_undefined__))
+#endif
+#endif
+uint32_t crc32c_3way(uint32_t crc, const char* buf, size_t len) {
+  const unsigned char* next = (const unsigned char*)buf;
+  uint64_t count;
+  uint64_t crc0, crc1, crc2;
+  crc0 = crc ^ 0xffffffffu;
+
+
+  if (len >= 8) {
+    // if len > 216 then align and use triplets
+    if (len > 216) {
+      {
+        // Work on the bytes (< 8) before the first 8-byte alignment addr starts
+        uint64_t align_bytes = (8 - (uintptr_t)next) & 7;
+        len -= align_bytes;
+        align_to_8(align_bytes, crc0, next);
+      }
+
+      // Now work on the remaining blocks
+      count = len / 24; // number of triplets
+      len %= 24; // bytes remaining
+      uint64_t n = count >> 7; // #blocks = first block + full blocks
+      uint64_t block_size = count & 127;
+      if (block_size == 0) {
+        block_size = 128;
+      } else {
+        n++;
+      }
+      // points to the first byte of the next block
+      const uint64_t* next0 = (uint64_t*)next + block_size;
+      const uint64_t* next1 = next0 + block_size;
+      const uint64_t* next2 = next1 + block_size;
+
+      crc1 = crc2 = 0;
+      // Use Duff's device, a for() loop inside a switch()
+      // statement. This needs to execute at least once, round len
+      // down to nearest triplet multiple
+      switch (block_size) {
+        case 128:
+          do {
+            // jumps here for a full block of len 128
+            CRCtriplet(crc, next, -128);
+            FALLTHROUGH_INTENDED;
+            case 127:
+              // jumps here or below for the first block smaller
+              CRCtriplet(crc, next, -127);
+              FALLTHROUGH_INTENDED;
+            case 126:
+              CRCtriplet(crc, next, -126); // than 128
+              FALLTHROUGH_INTENDED;
+            case 125:
+              CRCtriplet(crc, next, -125);
+              FALLTHROUGH_INTENDED;
+            case 124:
+              CRCtriplet(crc, next, -124);
+              FALLTHROUGH_INTENDED;
+            case 123:
+              CRCtriplet(crc, next, -123);
+              FALLTHROUGH_INTENDED;
+            case 122:
+              CRCtriplet(crc, next, -122);
+              FALLTHROUGH_INTENDED;
+            case 121:
+              CRCtriplet(crc, next, -121);
+              FALLTHROUGH_INTENDED;
+            case 120:
+              CRCtriplet(crc, next, -120);
+              FALLTHROUGH_INTENDED;
+            case 119:
+              CRCtriplet(crc, next, -119);
+              FALLTHROUGH_INTENDED;
+            case 118:
+              CRCtriplet(crc, next, -118);
+              FALLTHROUGH_INTENDED;
+            case 117:
+              CRCtriplet(crc, next, -117);
+              FALLTHROUGH_INTENDED;
+            case 116:
+              CRCtriplet(crc, next, -116);
+              FALLTHROUGH_INTENDED;
+            case 115:
+              CRCtriplet(crc, next, -115);
+              FALLTHROUGH_INTENDED;
+            case 114:
+              CRCtriplet(crc, next, -114);
+              FALLTHROUGH_INTENDED;
+            case 113:
+              CRCtriplet(crc, next, -113);
+              FALLTHROUGH_INTENDED;
+            case 112:
+              CRCtriplet(crc, next, -112);
+              FALLTHROUGH_INTENDED;
+            case 111:
+              CRCtriplet(crc, next, -111);
+              FALLTHROUGH_INTENDED;
+            case 110:
+              CRCtriplet(crc, next, -110);
+              FALLTHROUGH_INTENDED;
+            case 109:
+              CRCtriplet(crc, next, -109);
+              FALLTHROUGH_INTENDED;
+            case 108:
+              CRCtriplet(crc, next, -108);
+              FALLTHROUGH_INTENDED;
+            case 107:
+              CRCtriplet(crc, next, -107);
+              FALLTHROUGH_INTENDED;
+            case 106:
+              CRCtriplet(crc, next, -106);
+              FALLTHROUGH_INTENDED;
+            case 105:
+              CRCtriplet(crc, next, -105);
+              FALLTHROUGH_INTENDED;
+            case 104:
+              CRCtriplet(crc, next, -104);
+              FALLTHROUGH_INTENDED;
+            case 103:
+              CRCtriplet(crc, next, -103);
+              FALLTHROUGH_INTENDED;
+            case 102:
+              CRCtriplet(crc, next, -102);
+              FALLTHROUGH_INTENDED;
+            case 101:
+              CRCtriplet(crc, next, -101);
+              FALLTHROUGH_INTENDED;
+            case 100:
+              CRCtriplet(crc, next, -100);
+              FALLTHROUGH_INTENDED;
+            case 99:
+              CRCtriplet(crc, next, -99);
+              FALLTHROUGH_INTENDED;
+            case 98:
+              CRCtriplet(crc, next, -98);
+              FALLTHROUGH_INTENDED;
+            case 97:
+              CRCtriplet(crc, next, -97);
+              FALLTHROUGH_INTENDED;
+            case 96:
+              CRCtriplet(crc, next, -96);
+              FALLTHROUGH_INTENDED;
+            case 95:
+              CRCtriplet(crc, next, -95);
+              FALLTHROUGH_INTENDED;
+            case 94:
+              CRCtriplet(crc, next, -94);
+              FALLTHROUGH_INTENDED;
+            case 93:
+              CRCtriplet(crc, next, -93);
+              FALLTHROUGH_INTENDED;
+            case 92:
+              CRCtriplet(crc, next, -92);
+              FALLTHROUGH_INTENDED;
+            case 91:
+              CRCtriplet(crc, next, -91);
+              FALLTHROUGH_INTENDED;
+            case 90:
+              CRCtriplet(crc, next, -90);
+              FALLTHROUGH_INTENDED;
+            case 89:
+              CRCtriplet(crc, next, -89);
+              FALLTHROUGH_INTENDED;
+            case 88:
+              CRCtriplet(crc, next, -88);
+              FALLTHROUGH_INTENDED;
+            case 87:
+              CRCtriplet(crc, next, -87);
+              FALLTHROUGH_INTENDED;
+            case 86:
+              CRCtriplet(crc, next, -86);
+              FALLTHROUGH_INTENDED;
+            case 85:
+              CRCtriplet(crc, next, -85);
+              FALLTHROUGH_INTENDED;
+            case 84:
+              CRCtriplet(crc, next, -84);
+              FALLTHROUGH_INTENDED;
+            case 83:
+              CRCtriplet(crc, next, -83);
+              FALLTHROUGH_INTENDED;
+            case 82:
+              CRCtriplet(crc, next, -82);
+              FALLTHROUGH_INTENDED;
+            case 81:
+              CRCtriplet(crc, next, -81);
+              FALLTHROUGH_INTENDED;
+            case 80:
+              CRCtriplet(crc, next, -80);
+              FALLTHROUGH_INTENDED;
+            case 79:
+              CRCtriplet(crc, next, -79);
+              FALLTHROUGH_INTENDED;
+            case 78:
+              CRCtriplet(crc, next, -78);
+              FALLTHROUGH_INTENDED;
+            case 77:
+              CRCtriplet(crc, next, -77);
+              FALLTHROUGH_INTENDED;
+            case 76:
+              CRCtriplet(crc, next, -76);
+              FALLTHROUGH_INTENDED;
+            case 75:
+              CRCtriplet(crc, next, -75);
+              FALLTHROUGH_INTENDED;
+            case 74:
+              CRCtriplet(crc, next, -74);
+              FALLTHROUGH_INTENDED;
+            case 73:
+              CRCtriplet(crc, next, -73);
+              FALLTHROUGH_INTENDED;
+            case 72:
+              CRCtriplet(crc, next, -72);
+              FALLTHROUGH_INTENDED;
+            case 71:
+              CRCtriplet(crc, next, -71);
+              FALLTHROUGH_INTENDED;
+            case 70:
+              CRCtriplet(crc, next, -70);
+              FALLTHROUGH_INTENDED;
+            case 69:
+              CRCtriplet(crc, next, -69);
+              FALLTHROUGH_INTENDED;
+            case 68:
+              CRCtriplet(crc, next, -68);
+              FALLTHROUGH_INTENDED;
+            case 67:
+              CRCtriplet(crc, next, -67);
+              FALLTHROUGH_INTENDED;
+            case 66:
+              CRCtriplet(crc, next, -66);
+              FALLTHROUGH_INTENDED;
+            case 65:
+              CRCtriplet(crc, next, -65);
+              FALLTHROUGH_INTENDED;
+            case 64:
+              CRCtriplet(crc, next, -64);
+              FALLTHROUGH_INTENDED;
+            case 63:
+              CRCtriplet(crc, next, -63);
+              FALLTHROUGH_INTENDED;
+            case 62:
+              CRCtriplet(crc, next, -62);
+              FALLTHROUGH_INTENDED;
+            case 61:
+              CRCtriplet(crc, next, -61);
+              FALLTHROUGH_INTENDED;
+            case 60:
+              CRCtriplet(crc, next, -60);
+              FALLTHROUGH_INTENDED;
+            case 59:
+              CRCtriplet(crc, next, -59);
+              FALLTHROUGH_INTENDED;
+            case 58:
+              CRCtriplet(crc, next, -58);
+              FALLTHROUGH_INTENDED;
+            case 57:
+              CRCtriplet(crc, next, -57);
+              FALLTHROUGH_INTENDED;
+            case 56:
+              CRCtriplet(crc, next, -56);
+              FALLTHROUGH_INTENDED;
+            case 55:
+              CRCtriplet(crc, next, -55);
+              FALLTHROUGH_INTENDED;
+            case 54:
+              CRCtriplet(crc, next, -54);
+              FALLTHROUGH_INTENDED;
+            case 53:
+              CRCtriplet(crc, next, -53);
+              FALLTHROUGH_INTENDED;
+            case 52:
+              CRCtriplet(crc, next, -52);
+              FALLTHROUGH_INTENDED;
+            case 51:
+              CRCtriplet(crc, next, -51);
+              FALLTHROUGH_INTENDED;
+            case 50:
+              CRCtriplet(crc, next, -50);
+              FALLTHROUGH_INTENDED;
+            case 49:
+              CRCtriplet(crc, next, -49);
+              FALLTHROUGH_INTENDED;
+            case 48:
+              CRCtriplet(crc, next, -48);
+              FALLTHROUGH_INTENDED;
+            case 47:
+              CRCtriplet(crc, next, -47);
+              FALLTHROUGH_INTENDED;
+            case 46:
+              CRCtriplet(crc, next, -46);
+              FALLTHROUGH_INTENDED;
+            case 45:
+              CRCtriplet(crc, next, -45);
+              FALLTHROUGH_INTENDED;
+            case 44:
+              CRCtriplet(crc, next, -44);
+              FALLTHROUGH_INTENDED;
+            case 43:
+              CRCtriplet(crc, next, -43);
+              FALLTHROUGH_INTENDED;
+            case 42:
+              CRCtriplet(crc, next, -42);
+              FALLTHROUGH_INTENDED;
+            case 41:
+              CRCtriplet(crc, next, -41);
+              FALLTHROUGH_INTENDED;
+            case 40:
+              CRCtriplet(crc, next, -40);
+              FALLTHROUGH_INTENDED;
+            case 39:
+              CRCtriplet(crc, next, -39);
+              FALLTHROUGH_INTENDED;
+            case 38:
+              CRCtriplet(crc, next, -38);
+              FALLTHROUGH_INTENDED;
+            case 37:
+              CRCtriplet(crc, next, -37);
+              FALLTHROUGH_INTENDED;
+            case 36:
+              CRCtriplet(crc, next, -36);
+              FALLTHROUGH_INTENDED;
+            case 35:
+              CRCtriplet(crc, next, -35);
+              FALLTHROUGH_INTENDED;
+            case 34:
+              CRCtriplet(crc, next, -34);
+              FALLTHROUGH_INTENDED;
+            case 33:
+              CRCtriplet(crc, next, -33);
+              FALLTHROUGH_INTENDED;
+            case 32:
+              CRCtriplet(crc, next, -32);
+              FALLTHROUGH_INTENDED;
+            case 31:
+              CRCtriplet(crc, next, -31);
+              FALLTHROUGH_INTENDED;
+            case 30:
+              CRCtriplet(crc, next, -30);
+              FALLTHROUGH_INTENDED;
+            case 29:
+              CRCtriplet(crc, next, -29);
+              FALLTHROUGH_INTENDED;
+            case 28:
+              CRCtriplet(crc, next, -28);
+              FALLTHROUGH_INTENDED;
+            case 27:
+              CRCtriplet(crc, next, -27);
+              FALLTHROUGH_INTENDED;
+            case 26:
+              CRCtriplet(crc, next, -26);
+              FALLTHROUGH_INTENDED;
+            case 25:
+              CRCtriplet(crc, next, -25);
+              FALLTHROUGH_INTENDED;
+            case 24:
+              CRCtriplet(crc, next, -24);
+              FALLTHROUGH_INTENDED;
+            case 23:
+              CRCtriplet(crc, next, -23);
+              FALLTHROUGH_INTENDED;
+            case 22:
+              CRCtriplet(crc, next, -22);
+              FALLTHROUGH_INTENDED;
+            case 21:
+              CRCtriplet(crc, next, -21);
+              FALLTHROUGH_INTENDED;
+            case 20:
+              CRCtriplet(crc, next, -20);
+              FALLTHROUGH_INTENDED;
+            case 19:
+              CRCtriplet(crc, next, -19);
+              FALLTHROUGH_INTENDED;
+            case 18:
+              CRCtriplet(crc, next, -18);
+              FALLTHROUGH_INTENDED;
+            case 17:
+              CRCtriplet(crc, next, -17);
+              FALLTHROUGH_INTENDED;
+            case 16:
+              CRCtriplet(crc, next, -16);
+              FALLTHROUGH_INTENDED;
+            case 15:
+              CRCtriplet(crc, next, -15);
+              FALLTHROUGH_INTENDED;
+            case 14:
+              CRCtriplet(crc, next, -14);
+              FALLTHROUGH_INTENDED;
+            case 13:
+              CRCtriplet(crc, next, -13);
+              FALLTHROUGH_INTENDED;
+            case 12:
+              CRCtriplet(crc, next, -12);
+              FALLTHROUGH_INTENDED;
+            case 11:
+              CRCtriplet(crc, next, -11);
+              FALLTHROUGH_INTENDED;
+            case 10:
+              CRCtriplet(crc, next, -10);
+              FALLTHROUGH_INTENDED;
+            case 9:
+              CRCtriplet(crc, next, -9);
+              FALLTHROUGH_INTENDED;
+            case 8:
+              CRCtriplet(crc, next, -8);
+              FALLTHROUGH_INTENDED;
+            case 7:
+              CRCtriplet(crc, next, -7);
+              FALLTHROUGH_INTENDED;
+            case 6:
+              CRCtriplet(crc, next, -6);
+              FALLTHROUGH_INTENDED;
+            case 5:
+              CRCtriplet(crc, next, -5);
+              FALLTHROUGH_INTENDED;
+            case 4:
+              CRCtriplet(crc, next, -4);
+              FALLTHROUGH_INTENDED;
+            case 3:
+              CRCtriplet(crc, next, -3);
+              FALLTHROUGH_INTENDED;
+            case 2:
+              CRCtriplet(crc, next, -2);
+              FALLTHROUGH_INTENDED;
+            case 1:
+              CRCduplet(crc, next, -1); // the final triplet is actually only 2
+              //{ CombineCRC(); }
+              crc0 = CombineCRC(block_size, crc0, crc1, crc2, next2);
+              if (--n > 0) {
+                crc1 = crc2 = 0;
+                block_size = 128;
+                // points to the first byte of the next block
+                next0 = next2 + 128;
+                next1 = next0 + 128; // from here on all blocks are 128 long
+                next2 = next1 + 128;
+              }
+              FALLTHROUGH_INTENDED;
+            case 0:;
+          } while (n > 0);
+      }
+      next = (const unsigned char*)next2;
+    }
+    uint64_t count2 = len >> 3; // 216 of less bytes is 27 or less singlets
+    len = len & 7;
+    next += (count2 * 8);
+    switch (count2) {
+      case 27:
+        CRCsinglet(crc0, next, -27 * 8);
+        FALLTHROUGH_INTENDED;
+      case 26:
+        CRCsinglet(crc0, next, -26 * 8);
+        FALLTHROUGH_INTENDED;
+      case 25:
+        CRCsinglet(crc0, next, -25 * 8);
+        FALLTHROUGH_INTENDED;
+      case 24:
+        CRCsinglet(crc0, next, -24 * 8);
+        FALLTHROUGH_INTENDED;
+      case 23:
+        CRCsinglet(crc0, next, -23 * 8);
+        FALLTHROUGH_INTENDED;
+      case 22:
+        CRCsinglet(crc0, next, -22 * 8);
+        FALLTHROUGH_INTENDED;
+      case 21:
+        CRCsinglet(crc0, next, -21 * 8);
+        FALLTHROUGH_INTENDED;
+      case 20:
+        CRCsinglet(crc0, next, -20 * 8);
+        FALLTHROUGH_INTENDED;
+      case 19:
+        CRCsinglet(crc0, next, -19 * 8);
+        FALLTHROUGH_INTENDED;
+      case 18:
+        CRCsinglet(crc0, next, -18 * 8);
+        FALLTHROUGH_INTENDED;
+      case 17:
+        CRCsinglet(crc0, next, -17 * 8);
+        FALLTHROUGH_INTENDED;
+      case 16:
+        CRCsinglet(crc0, next, -16 * 8);
+        FALLTHROUGH_INTENDED;
+      case 15:
+        CRCsinglet(crc0, next, -15 * 8);
+        FALLTHROUGH_INTENDED;
+      case 14:
+        CRCsinglet(crc0, next, -14 * 8);
+        FALLTHROUGH_INTENDED;
+      case 13:
+        CRCsinglet(crc0, next, -13 * 8);
+        FALLTHROUGH_INTENDED;
+      case 12:
+        CRCsinglet(crc0, next, -12 * 8);
+        FALLTHROUGH_INTENDED;
+      case 11:
+        CRCsinglet(crc0, next, -11 * 8);
+        FALLTHROUGH_INTENDED;
+      case 10:
+        CRCsinglet(crc0, next, -10 * 8);
+        FALLTHROUGH_INTENDED;
+      case 9:
+        CRCsinglet(crc0, next, -9 * 8);
+        FALLTHROUGH_INTENDED;
+      case 8:
+        CRCsinglet(crc0, next, -8 * 8);
+        FALLTHROUGH_INTENDED;
+      case 7:
+        CRCsinglet(crc0, next, -7 * 8);
+        FALLTHROUGH_INTENDED;
+      case 6:
+        CRCsinglet(crc0, next, -6 * 8);
+        FALLTHROUGH_INTENDED;
+      case 5:
+        CRCsinglet(crc0, next, -5 * 8);
+        FALLTHROUGH_INTENDED;
+      case 4:
+        CRCsinglet(crc0, next, -4 * 8);
+        FALLTHROUGH_INTENDED;
+      case 3:
+        CRCsinglet(crc0, next, -3 * 8);
+        FALLTHROUGH_INTENDED;
+      case 2:
+        CRCsinglet(crc0, next, -2 * 8);
+        FALLTHROUGH_INTENDED;
+      case 1:
+        CRCsinglet(crc0, next, -1 * 8);
+        FALLTHROUGH_INTENDED;
+      case 0:;
+    }
+  }
+  {
+    align_to_8(len, crc0, next);
+    return (uint32_t)crc0 ^ 0xffffffffu;
+  }
+}
+
+#endif //HAVE_SSE42 && HAVE_PCLMUL
+
+static inline Function Choose_Extend() {
+#ifdef HAVE_POWER8
+  return isAltiVec() ? ExtendPPCImpl : ExtendImpl<Slow_CRC32>;
+#elif defined(HAVE_ARM64_CRC)
+  if(crc32c_runtime_check()) {
+    pmull_runtime_flag = crc32c_pmull_runtime_check();
+    return ExtendARMImpl;
+  } else {
+    return ExtendImpl<Slow_CRC32>;
+  }
+#else
+  if (isSSE42()) {
+    if (isPCLMULQDQ()) {
+#if (defined HAVE_SSE42 && defined HAVE_PCLMUL) && !defined NO_THREEWAY_CRC32C
+      return crc32c_3way;
+#else
+    return ExtendImpl<Fast_CRC32>; // Fast_CRC32 will check HAVE_SSE42 itself
+#endif
+    }
+    else {  // no runtime PCLMULQDQ support but has SSE42 support
+      return ExtendImpl<Fast_CRC32>;
+    }
+  } // end of isSSE42()
+  else {
+    return ExtendImpl<Slow_CRC32>;
+  }
+#endif
+}
+
+static Function ChosenExtend = Choose_Extend();
+uint32_t Extend(uint32_t crc, const char* buf, size_t size) {
+  return ChosenExtend(crc, buf, size);
+}
+
+// The code for crc32c combine, copied with permission from folly
+
+// Standard galois-field multiply.  The only modification is that a,
+// b, m, and p are all bit-reflected.
+//
+// https://en.wikipedia.org/wiki/Finite_field_arithmetic
+static constexpr uint32_t gf_multiply_sw_1(
+    size_t i, uint32_t p, uint32_t a, uint32_t b, uint32_t m) {
+  // clang-format off
+  return i == 32 ? p : gf_multiply_sw_1(
+      /* i = */ i + 1,
+      /* p = */ p ^ ((0u-((b >> 31) & 1)) & a),
+      /* a = */ (a >> 1) ^ ((0u-(a & 1)) & m),
+      /* b = */ b << 1,
+      /* m = */ m);
+  // clang-format on
+}
+static constexpr uint32_t gf_multiply_sw(uint32_t a, uint32_t b, uint32_t m) {
+  return gf_multiply_sw_1(/* i = */ 0, /* p = */ 0, a, b, m);
+}
+
+static constexpr uint32_t gf_square_sw(uint32_t a, uint32_t m) {
+  return gf_multiply_sw(a, a, m);
+}
+
+template <size_t i, uint32_t m>
+struct gf_powers_memo {
+  static constexpr uint32_t value =
+      gf_square_sw(gf_powers_memo<i - 1, m>::value, m);
+};
+template <uint32_t m>
+struct gf_powers_memo<0, m> {
+  static constexpr uint32_t value = m;
+};
+
+template <typename T, T... Ints>
+struct integer_sequence {
+  using value_type = T;
+  static constexpr size_t size() { return sizeof...(Ints); }
+};
+
+template <typename T, std::size_t N, T... Is>
+struct make_integer_sequence : make_integer_sequence<T, N - 1, N - 1, Is...> {};
+
+template <typename T, T... Is>
+struct make_integer_sequence<T, 0, Is...> : integer_sequence<T, Is...> {};
+
+template <std::size_t N>
+using make_index_sequence = make_integer_sequence<std::size_t, N>;
+
+template <uint32_t m>
+struct gf_powers_make {
+  template <size_t... i>
+  using index_sequence = integer_sequence<size_t, i...>;
+  template <size_t... i>
+  constexpr std::array<uint32_t, sizeof...(i)> operator()(
+      index_sequence<i...>) const {
+    return std::array<uint32_t, sizeof...(i)>{{gf_powers_memo<i, m>::value...}};
+  }
+};
+
+static constexpr uint32_t crc32c_m = 0x82f63b78;
+
+static constexpr std::array<uint32_t, 62> const crc32c_powers =
+    gf_powers_make<crc32c_m>{}(make_index_sequence<62>{});
+
+// Expects a "pure" crc (see Crc32cCombine)
+static uint32_t Crc32AppendZeroes(
+    uint32_t crc, size_t len_over_4, uint32_t polynomial,
+    std::array<uint32_t, 62> const& powers_array) {
+  auto powers = powers_array.data();
+  // Append by multiplying by consecutive powers of two of the zeroes
+  // array
+  size_t len_bits = len_over_4;
+
+  while (len_bits) {
+    // Advance directly to next bit set.
+    auto r = CountTrailingZeroBits(len_bits);
+    len_bits >>= r;
+    powers += r;
+
+    crc = gf_multiply_sw(crc, *powers, polynomial);
+
+    len_bits >>= 1;
+    powers++;
+  }
+
+  return crc;
+}
+
+static inline uint32_t InvertedToPure(uint32_t crc) { return ~crc; }
+
+static inline uint32_t PureToInverted(uint32_t crc) { return ~crc; }
+
+static inline uint32_t PureExtend(uint32_t crc, const char* buf, size_t size) {
+  return InvertedToPure(Extend(PureToInverted(crc), buf, size));
+}
+
+// Background:
+// RocksDB uses two kinds of crc32c values: masked and unmasked. Neither is
+// a "pure" CRC because a pure CRC satisfies (^ for xor)
+//  crc(a ^ b) = crc(a) ^ crc(b)
+// The unmasked is closest, and this function takes unmasked crc32c values.
+// The unmasked values are impure in two ways:
+// * The initial setting at the start of CRC computation is all 1 bits
+// (like -1) instead of zero.
+// * The result has all bits invered.
+// Note that together, these result in the empty string having a crc32c of
+// zero. See
+// https://en.wikipedia.org/wiki/Computation_of_cyclic_redundancy_checks#CRC_variants
+//
+// Simplified version of strategy, using xor through pure CRCs (+ for concat):
+//
+// pure_crc(str1 + str2) = pure_crc(str1 + zeros(len(str2))) ^
+//                         pure_crc(zeros(len(str1)) + str2)
+//
+// because the xor of these two zero-padded strings is str1 + str2. For pure
+// CRC, leading zeros don't affect the result, so we only need
+//
+// pure_crc(str1 + str2) = pure_crc(str1 + zeros(len(str2))) ^
+//                         pure_crc(str2)
+//
+// Considering we aren't working with pure CRCs, what is actually in the input?
+//
+// crc1 = PureToInverted(PureExtendCrc32c(-1, zeros, crc1len) ^
+//                       PureCrc32c(str1, crc1len))
+// crc2 = PureToInverted(PureExtendCrc32c(-1, zeros, crc2len) ^
+//                       PureCrc32c(str2, crc2len))
+//
+// The result we want to compute is
+// combined = PureToInverted(PureExtendCrc32c(PureExtendCrc32c(-1, zeros,
+//                                                             crc1len) ^
+//                                            PureCrc32c(str1, crc1len),
+//                                            zeros, crc2len) ^
+//                           PureCrc32c(str2, crc2len))
+//
+// Thus, in addition to extending crc1 over the length of str2 in (virtual)
+// zeros, we need to cancel out the -1 initializer that was used in computing
+// crc2. To cancel it out, we also need to extend it over crc2len in zeros.
+// To simplify, since the end of str1 and that -1 initializer for crc2 are at
+// the same logical position, we can combine them before we extend over the
+// zeros.
+uint32_t Crc32cCombine(uint32_t crc1, uint32_t crc2, size_t crc2len) {
+  uint32_t pure_crc1_with_init = InvertedToPure(crc1);
+  uint32_t pure_crc2_with_init = InvertedToPure(crc2);
+  uint32_t pure_crc2_init = static_cast<uint32_t>(-1);
+
+  // Append up to 32 bits of zeroes in the normal way
+  char zeros[4] = {0, 0, 0, 0};
+  auto len = crc2len & 3;
+  uint32_t tmp = pure_crc1_with_init ^ pure_crc2_init;
+  if (len) {
+    tmp = PureExtend(tmp, zeros, len);
+  }
+  return PureToInverted(
+      Crc32AppendZeroes(tmp, crc2len / 4, crc32c_m, crc32c_powers) ^
+      pure_crc2_with_init);
+}
+
+}  // namespace crc32c
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/util/crc32c.h b/src/rocksdb/util/crc32c.h
new file mode 100644
index 000000000..a08ad60af
--- /dev/null
+++ b/src/rocksdb/util/crc32c.h
@@ -0,0 +1,56 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+//
+// Copyright (c) 2011 The LevelDB Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file. See the AUTHORS file for names of contributors.
+
+#pragma once
+#include <stddef.h>
+#include <stdint.h>
+
+#include <string>
+
+#include "rocksdb/rocksdb_namespace.h"
+
+namespace ROCKSDB_NAMESPACE {
+namespace crc32c {
+
+extern std::string IsFastCrc32Supported();
+
+// Return the crc32c of concat(A, data[0,n-1]) where init_crc is the
+// crc32c of some string A.  Extend() is often used to maintain the
+// crc32c of a stream of data.
+extern uint32_t Extend(uint32_t init_crc, const char* data, size_t n);
+
+// Takes two unmasked crc32c values, and the length of the string from
+// which `crc2` was computed, and computes a crc32c value for the
+// concatenation of the original two input strings. Running time is
+// ~ log(crc2len).
+extern uint32_t Crc32cCombine(uint32_t crc1, uint32_t crc2, size_t crc2len);
+
+// Return the crc32c of data[0,n-1]
+inline uint32_t Value(const char* data, size_t n) { return Extend(0, data, n); }
+
+static const uint32_t kMaskDelta = 0xa282ead8ul;
+
+// Return a masked representation of crc.
+//
+// Motivation: it is problematic to compute the CRC of a string that
+// contains embedded CRCs.  Therefore we recommend that CRCs stored
+// somewhere (e.g., in files) should be masked before being stored.
+inline uint32_t Mask(uint32_t crc) {
+  // Rotate right by 15 bits and add a constant.
+  return ((crc >> 15) | (crc << 17)) + kMaskDelta;
+}
+
+// Return the crc whose masked representation is masked_crc.
+inline uint32_t Unmask(uint32_t masked_crc) {
+  uint32_t rot = masked_crc - kMaskDelta;
+  return ((rot >> 17) | (rot << 15));
+}
+
+}  // namespace crc32c
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/util/crc32c_arm64.cc b/src/rocksdb/util/crc32c_arm64.cc
new file mode 100644
index 000000000..4885f4fe1
--- /dev/null
+++ b/src/rocksdb/util/crc32c_arm64.cc
@@ -0,0 +1,215 @@
+//  Copyright (c) 2018, Arm Limited and affiliates. All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#include "util/crc32c_arm64.h"
+
+#if defined(HAVE_ARM64_CRC)
+
+#if defined(__linux__)
+#include <asm/hwcap.h>
+#endif
+#ifdef ROCKSDB_AUXV_GETAUXVAL_PRESENT
+#include <sys/auxv.h>
+#endif
+#ifndef HWCAP_CRC32
+#define HWCAP_CRC32 (1 << 7)
+#endif
+#ifndef HWCAP_PMULL
+#define HWCAP_PMULL (1 << 4)
+#endif
+#if defined(__APPLE__)
+#include <sys/sysctl.h>
+#endif
+#if defined(__OpenBSD__)
+#include <sys/types.h>
+#include <sys/sysctl.h>
+#include <machine/cpu.h>
+#include <machine/armreg.h>
+#endif
+
+#ifdef HAVE_ARM64_CRYPTO
+/* unfolding to compute 8 * 3 = 24 bytes parallelly */
+#define CRC32C24BYTES(ITR)                                    \
+  crc1 = crc32c_u64(crc1, *(buf64 + BLK_LENGTH + (ITR)));     \
+  crc2 = crc32c_u64(crc2, *(buf64 + BLK_LENGTH * 2 + (ITR))); \
+  crc0 = crc32c_u64(crc0, *(buf64 + (ITR)));
+
+/* unfolding to compute 24 * 7 = 168 bytes parallelly */
+#define CRC32C7X24BYTES(ITR)   \
+  do {                         \
+    CRC32C24BYTES((ITR)*7 + 0) \
+    CRC32C24BYTES((ITR)*7 + 1) \
+    CRC32C24BYTES((ITR)*7 + 2) \
+    CRC32C24BYTES((ITR)*7 + 3) \
+    CRC32C24BYTES((ITR)*7 + 4) \
+    CRC32C24BYTES((ITR)*7 + 5) \
+    CRC32C24BYTES((ITR)*7 + 6) \
+  } while (0)
+#endif
+
+extern bool pmull_runtime_flag;
+
+uint32_t crc32c_runtime_check(void) {
+#if defined(ROCKSDB_AUXV_GETAUXVAL_PRESENT) || defined(__FreeBSD__)
+  uint64_t auxv = 0;
+#if defined(ROCKSDB_AUXV_GETAUXVAL_PRESENT)
+  auxv = getauxval(AT_HWCAP);
+#elif defined(__FreeBSD__)
+  elf_aux_info(AT_HWCAP, &auxv, sizeof(auxv));
+#endif
+  return (auxv & HWCAP_CRC32) != 0;
+#elif defined(__APPLE__)
+  int r;
+  size_t l = sizeof(r);
+  if (sysctlbyname("hw.optional.armv8_crc32", &r, &l, NULL, 0) == -1) return 0;
+  return r == 1;
+#elif defined(__OpenBSD__)
+  int r = 0;
+  const int isar0_mib[] = { CTL_MACHDEP, CPU_ID_AA64ISAR0 };
+  uint64_t isar0;
+  size_t len = sizeof(isar0);
+
+  if (sysctl(isar0_mib, 2, &isar0, &len, NULL, 0) != -1) {
+      if (ID_AA64ISAR0_CRC32(isar0) >= ID_AA64ISAR0_CRC32_BASE)
+        r = 1;
+  }
+  return r;
+#else
+  return 0;
+#endif
+}
+
+bool crc32c_pmull_runtime_check(void) {
+#if defined(ROCKSDB_AUXV_GETAUXVAL_PRESENT) || defined(__FreeBSD__)
+  uint64_t auxv = 0;
+#if defined(ROCKSDB_AUXV_GETAUXVAL_PRESENT)
+  auxv = getauxval(AT_HWCAP);
+#elif defined(__FreeBSD__)
+  elf_aux_info(AT_HWCAP, &auxv, sizeof(auxv));
+#endif
+  return (auxv & HWCAP_PMULL) != 0;
+#elif defined(__APPLE__)
+  return true;
+#elif defined(__OpenBSD__)
+  bool r = false;
+  const int isar0_mib[] = { CTL_MACHDEP, CPU_ID_AA64ISAR0 };
+  uint64_t isar0;
+  size_t len = sizeof(isar0);
+
+  if (sysctl(isar0_mib, 2, &isar0, &len, NULL, 0) != -1) {
+      if (ID_AA64ISAR0_AES(isar0) >= ID_AA64ISAR0_AES_PMULL)
+        r = true;
+  }
+  return r;
+#else
+  return false;
+#endif
+}
+
+#ifdef ROCKSDB_UBSAN_RUN
+#if defined(__clang__)
+__attribute__((__no_sanitize__("alignment")))
+#elif defined(__GNUC__)
+__attribute__((__no_sanitize_undefined__))
+#endif
+#endif
+uint32_t
+crc32c_arm64(uint32_t crc, unsigned char const *data, size_t len) {
+  const uint8_t *buf8;
+  const uint64_t *buf64 = (uint64_t *)data;
+  int length = (int)len;
+  crc ^= 0xffffffff;
+
+  /*
+   * Pmull runtime check here.
+   * Raspberry Pi supports crc32 but doesn't support pmull.
+   * Skip Crc32c Parallel computation if no crypto extension available.
+   */
+  if (pmull_runtime_flag) {
+/* Macro (HAVE_ARM64_CRYPTO) is used for compiling check  */
+#ifdef HAVE_ARM64_CRYPTO
+/* Crc32c Parallel computation
+ *   Algorithm comes from Intel whitepaper:
+ *   crc-iscsi-polynomial-crc32-instruction-paper
+ *
+ * Input data is divided into three equal-sized blocks
+ *   Three parallel blocks (crc0, crc1, crc2) for 1024 Bytes
+ *   One Block: 42(BLK_LENGTH) * 8(step length: crc32c_u64) bytes
+ */
+#define BLK_LENGTH 42
+    while (length >= 1024) {
+      uint64_t t0, t1;
+      uint32_t crc0 = 0, crc1 = 0, crc2 = 0;
+
+      /* Parallel Param:
+       *   k0 = CRC32(x ^ (42 * 8 * 8 * 2 - 1));
+       *   k1 = CRC32(x ^ (42 * 8 * 8 - 1));
+       */
+      uint32_t k0 = 0xe417f38a, k1 = 0x8f158014;
+
+      /* Prefetch data for following block to avoid cache miss */
+      PREF1KL1((uint8_t *)buf64, 1024);
+
+      /* First 8 byte for better pipelining */
+      crc0 = crc32c_u64(crc, *buf64++);
+
+      /* 3 blocks crc32c parallel computation
+       * Macro unfolding to compute parallelly
+       * 168 * 6 = 1008 (bytes)
+       */
+      CRC32C7X24BYTES(0);
+      CRC32C7X24BYTES(1);
+      CRC32C7X24BYTES(2);
+      CRC32C7X24BYTES(3);
+      CRC32C7X24BYTES(4);
+      CRC32C7X24BYTES(5);
+      buf64 += (BLK_LENGTH * 3);
+
+      /* Last 8 bytes */
+      crc = crc32c_u64(crc2, *buf64++);
+
+      t0 = (uint64_t)vmull_p64(crc0, k0);
+      t1 = (uint64_t)vmull_p64(crc1, k1);
+
+      /* Merge (crc0, crc1, crc2) -> crc */
+      crc1 = crc32c_u64(0, t1);
+      crc ^= crc1;
+      crc0 = crc32c_u64(0, t0);
+      crc ^= crc0;
+
+      length -= 1024;
+    }
+
+    if (length == 0) return crc ^ (0xffffffffU);
+#endif
+  }  // if Pmull runtime check here
+
+  buf8 = (const uint8_t *)buf64;
+  while (length >= 8) {
+    crc = crc32c_u64(crc, *(const uint64_t *)buf8);
+    buf8 += 8;
+    length -= 8;
+  }
+
+  /* The following is more efficient than the straight loop */
+  if (length >= 4) {
+    crc = crc32c_u32(crc, *(const uint32_t *)buf8);
+    buf8 += 4;
+    length -= 4;
+  }
+
+  if (length >= 2) {
+    crc = crc32c_u16(crc, *(const uint16_t *)buf8);
+    buf8 += 2;
+    length -= 2;
+  }
+
+  if (length >= 1) crc = crc32c_u8(crc, *buf8);
+
+  crc ^= 0xffffffff;
+  return crc;
+}
+
+#endif
diff --git a/src/rocksdb/util/crc32c_arm64.h b/src/rocksdb/util/crc32c_arm64.h
new file mode 100644
index 000000000..4b27fe871
--- /dev/null
+++ b/src/rocksdb/util/crc32c_arm64.h
@@ -0,0 +1,52 @@
+//  Copyright (c) 2018, Arm Limited and affiliates. All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#ifndef UTIL_CRC32C_ARM64_H
+#define UTIL_CRC32C_ARM64_H
+
+#include <cinttypes>
+#include <cstddef>
+
+#if defined(__aarch64__) || defined(__AARCH64__)
+
+#ifdef __ARM_FEATURE_CRC32
+#define HAVE_ARM64_CRC
+#include <arm_acle.h>
+#define crc32c_u8(crc, v) __crc32cb(crc, v)
+#define crc32c_u16(crc, v) __crc32ch(crc, v)
+#define crc32c_u32(crc, v) __crc32cw(crc, v)
+#define crc32c_u64(crc, v) __crc32cd(crc, v)
+// clang-format off
+#define PREF4X64L1(buffer, PREF_OFFSET, ITR)                \
+  __asm__("PRFM PLDL1KEEP, [%x[v],%[c]]" ::[v] "r"(buffer), \
+          [c] "I"((PREF_OFFSET) + ((ITR) + 0) * 64));       \
+  __asm__("PRFM PLDL1KEEP, [%x[v],%[c]]" ::[v] "r"(buffer), \
+          [c] "I"((PREF_OFFSET) + ((ITR) + 1) * 64));       \
+  __asm__("PRFM PLDL1KEEP, [%x[v],%[c]]" ::[v] "r"(buffer), \
+          [c] "I"((PREF_OFFSET) + ((ITR) + 2) * 64));       \
+  __asm__("PRFM PLDL1KEEP, [%x[v],%[c]]" ::[v] "r"(buffer), \
+          [c] "I"((PREF_OFFSET) + ((ITR) + 3) * 64));
+// clang-format on
+
+#define PREF1KL1(buffer, PREF_OFFSET)  \
+  PREF4X64L1(buffer, (PREF_OFFSET), 0) \
+  PREF4X64L1(buffer, (PREF_OFFSET), 4) \
+  PREF4X64L1(buffer, (PREF_OFFSET), 8) \
+  PREF4X64L1(buffer, (PREF_OFFSET), 12)
+
+extern uint32_t crc32c_arm64(uint32_t crc, unsigned char const *data,
+                             size_t len);
+extern uint32_t crc32c_runtime_check(void);
+extern bool crc32c_pmull_runtime_check(void);
+
+#ifdef __ARM_FEATURE_CRYPTO
+#define HAVE_ARM64_CRYPTO
+#include <arm_neon.h>
+#endif  // __ARM_FEATURE_CRYPTO
+#endif  // __ARM_FEATURE_CRC32
+
+#endif  // defined(__aarch64__) || defined(__AARCH64__)
+
+#endif
diff --git a/src/rocksdb/util/crc32c_ppc.c b/src/rocksdb/util/crc32c_ppc.c
new file mode 100644
index 000000000..b37dfb158
--- /dev/null
+++ b/src/rocksdb/util/crc32c_ppc.c
@@ -0,0 +1,94 @@
+//  Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.
+//  Copyright (c) 2017 International Business Machines Corp.
+//  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#define CRC_TABLE
+#include <stdint.h>
+#include <stdlib.h>
+#include <strings.h>
+#include "util/crc32c_ppc_constants.h"
+
+#define VMX_ALIGN 16
+#define VMX_ALIGN_MASK (VMX_ALIGN - 1)
+
+#ifdef REFLECT
+static unsigned int crc32_align(unsigned int crc, unsigned char const *p,
+                                unsigned long len) {
+  while (len--) crc = crc_table[(crc ^ *p++) & 0xff] ^ (crc >> 8);
+  return crc;
+}
+#endif
+
+#ifdef HAVE_POWER8
+unsigned int __crc32_vpmsum(unsigned int crc, unsigned char const *p,
+                            unsigned long len);
+
+static uint32_t crc32_vpmsum(uint32_t crc, unsigned char const *data,
+                             size_t len) {
+  unsigned int prealign;
+  unsigned int tail;
+
+#ifdef CRC_XOR
+  crc ^= 0xffffffff;
+#endif
+
+  if (len < VMX_ALIGN + VMX_ALIGN_MASK) {
+    crc = crc32_align(crc, data, (unsigned long)len);
+    goto out;
+  }
+
+  if ((unsigned long)data & VMX_ALIGN_MASK) {
+    prealign = VMX_ALIGN - ((unsigned long)data & VMX_ALIGN_MASK);
+    crc = crc32_align(crc, data, prealign);
+    len -= prealign;
+    data += prealign;
+  }
+
+  crc = __crc32_vpmsum(crc, data, (unsigned long)len & ~VMX_ALIGN_MASK);
+
+  tail = len & VMX_ALIGN_MASK;
+  if (tail) {
+    data += len & ~VMX_ALIGN_MASK;
+    crc = crc32_align(crc, data, tail);
+  }
+
+out:
+#ifdef CRC_XOR
+  crc ^= 0xffffffff;
+#endif
+
+  return crc;
+}
+
+/* This wrapper function works around the fact that crc32_vpmsum
+ * does not gracefully handle the case where the data pointer is NULL.  There
+ * may be room for performance improvement here.
+ */
+uint32_t crc32c_ppc(uint32_t crc, unsigned char const *data, size_t len) {
+  unsigned char *buf2;
+
+  if (!data) {
+    buf2 = (unsigned char *)malloc(len);
+    bzero(buf2, len);
+    crc = crc32_vpmsum(crc, buf2, len);
+    free(buf2);
+  } else {
+    crc = crc32_vpmsum(crc, data, (unsigned long)len);
+  }
+  return crc;
+}
+
+#else /* HAVE_POWER8 */
+
+/* This symbol has to exist on non-ppc architectures (and on legacy
+ * ppc systems using power7 or below) in order to compile properly
+ * there, even though it won't be called.
+ */
+uint32_t crc32c_ppc(uint32_t crc, unsigned char const *data, size_t len) {
+  return 0;
+}
+
+#endif /* HAVE_POWER8 */
diff --git a/src/rocksdb/util/crc32c_ppc.h b/src/rocksdb/util/crc32c_ppc.h
new file mode 100644
index 000000000..f0b0b66d5
--- /dev/null
+++ b/src/rocksdb/util/crc32c_ppc.h
@@ -0,0 +1,22 @@
+//  Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.
+//  Copyright (c) 2017 International Business Machines Corp.
+//  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#pragma once
+
+#include <cstddef>
+#include <cstdint>
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+extern uint32_t crc32c_ppc(uint32_t crc, unsigned char const *buffer,
+                           size_t len);
+
+#ifdef __cplusplus
+}
+#endif
diff --git a/src/rocksdb/util/crc32c_ppc_asm.S b/src/rocksdb/util/crc32c_ppc_asm.S
new file mode 100644
index 000000000..6959ba839
--- /dev/null
+++ b/src/rocksdb/util/crc32c_ppc_asm.S
@@ -0,0 +1,756 @@
+//  Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.
+//  Copyright (c) 2015 Anton Blanchard <anton@au.ibm.com>, IBM
+//  Copyright (c) 2017 International Business Machines Corp.
+//  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#if defined (__clang__)
+#include "third-party/gcc/ppc-asm.h"
+#else
+#include <ppc-asm.h>
+#endif
+#include "ppc-opcode.h"
+
+#undef toc
+
+#ifndef r1
+#define r1 1
+#endif
+
+#ifndef r2
+#define r2 2
+#endif
+
+	.section	.rodata
+.balign 16
+
+.byteswap_constant:
+	/* byte reverse permute constant */
+	.octa 0x0F0E0D0C0B0A09080706050403020100
+
+#define __ASSEMBLY__
+#include "crc32c_ppc_constants.h"
+
+	.text
+
+#if defined(__BIG_ENDIAN__) && defined(REFLECT)
+#define BYTESWAP_DATA
+#elif defined(__LITTLE_ENDIAN__) && !defined(REFLECT)
+#define BYTESWAP_DATA
+#else
+#undef BYTESWAP_DATA
+#endif
+
+#define off16		r25
+#define off32		r26
+#define off48		r27
+#define off64		r28
+#define off80		r29
+#define off96		r30
+#define off112		r31
+
+#define const1		v24
+#define const2		v25
+
+#define byteswap	v26
+#define	mask_32bit	v27
+#define	mask_64bit	v28
+#define zeroes		v29
+
+#ifdef BYTESWAP_DATA
+#define VPERM(A, B, C, D) vperm	A, B, C, D
+#else
+#define VPERM(A, B, C, D)
+#endif
+
+/* unsigned int __crc32_vpmsum(unsigned int crc, void *p, unsigned long len) */
+FUNC_START(__crc32_vpmsum)
+	std	r31,-8(r1)
+	std	r30,-16(r1)
+	std	r29,-24(r1)
+	std	r28,-32(r1)
+	std	r27,-40(r1)
+	std	r26,-48(r1)
+	std	r25,-56(r1)
+
+	li	off16,16
+	li	off32,32
+	li	off48,48
+	li	off64,64
+	li	off80,80
+	li	off96,96
+	li	off112,112
+	li	r0,0
+
+	/* Enough room for saving 10 non volatile VMX registers */
+	subi	r6,r1,56+10*16
+	subi	r7,r1,56+2*16
+
+	stvx	v20,0,r6
+	stvx	v21,off16,r6
+	stvx	v22,off32,r6
+	stvx	v23,off48,r6
+	stvx	v24,off64,r6
+	stvx	v25,off80,r6
+	stvx	v26,off96,r6
+	stvx	v27,off112,r6
+	stvx	v28,0,r7
+	stvx	v29,off16,r7
+
+	mr	r10,r3
+
+	vxor	zeroes,zeroes,zeroes
+	vspltisw v0,-1
+
+	vsldoi	mask_32bit,zeroes,v0,4
+	vsldoi	mask_64bit,zeroes,v0,8
+
+	/* Get the initial value into v8 */
+	vxor	v8,v8,v8
+	MTVRD(v8, r3)
+#ifdef REFLECT
+	vsldoi	v8,zeroes,v8,8	/* shift into bottom 32 bits */
+#else
+	vsldoi	v8,v8,zeroes,4	/* shift into top 32 bits */
+#endif
+
+#ifdef BYTESWAP_DATA
+	addis	r3,r2,.byteswap_constant@toc@ha
+	addi	r3,r3,.byteswap_constant@toc@l
+
+	lvx	byteswap,0,r3
+	addi	r3,r3,16
+#endif
+
+	cmpdi	r5,256
+	blt	.Lshort
+
+	rldicr	r6,r5,0,56
+
+	/* Checksum in blocks of MAX_SIZE */
+1:	lis	r7,MAX_SIZE@h
+	ori	r7,r7,MAX_SIZE@l
+	mr	r9,r7
+	cmpd	r6,r7
+	bgt	2f
+	mr	r7,r6
+2:	subf	r6,r7,r6
+
+	/* our main loop does 128 bytes at a time */
+	srdi	r7,r7,7
+
+	/*
+	 * Work out the offset into the constants table to start at. Each
+	 * constant is 16 bytes, and it is used against 128 bytes of input
+	 * data - 128 / 16 = 8
+	 */
+	sldi	r8,r7,4
+	srdi	r9,r9,3
+	subf	r8,r8,r9
+
+	/* We reduce our final 128 bytes in a separate step */
+	addi	r7,r7,-1
+	mtctr	r7
+
+	addis	r3,r2,.constants@toc@ha
+	addi	r3,r3,.constants@toc@l
+
+	/* Find the start of our constants */
+	add	r3,r3,r8
+
+	/* zero v0-v7 which will contain our checksums */
+	vxor	v0,v0,v0
+	vxor	v1,v1,v1
+	vxor	v2,v2,v2
+	vxor	v3,v3,v3
+	vxor	v4,v4,v4
+	vxor	v5,v5,v5
+	vxor	v6,v6,v6
+	vxor	v7,v7,v7
+
+	lvx	const1,0,r3
+
+	/*
+	 * If we are looping back to consume more data we use the values
+	 * already in v16-v23.
+	 */
+	cmpdi	r0,1
+	beq	2f
+
+	/* First warm up pass */
+	lvx	v16,0,r4
+	lvx	v17,off16,r4
+	VPERM(v16,v16,v16,byteswap)
+	VPERM(v17,v17,v17,byteswap)
+	lvx	v18,off32,r4
+	lvx	v19,off48,r4
+	VPERM(v18,v18,v18,byteswap)
+	VPERM(v19,v19,v19,byteswap)
+	lvx	v20,off64,r4
+	lvx	v21,off80,r4
+	VPERM(v20,v20,v20,byteswap)
+	VPERM(v21,v21,v21,byteswap)
+	lvx	v22,off96,r4
+	lvx	v23,off112,r4
+	VPERM(v22,v22,v22,byteswap)
+	VPERM(v23,v23,v23,byteswap)
+	addi	r4,r4,8*16
+
+	/* xor in initial value */
+	vxor	v16,v16,v8
+
+2:	bdz	.Lfirst_warm_up_done
+
+	addi	r3,r3,16
+	lvx	const2,0,r3
+
+	/* Second warm up pass */
+	VPMSUMD(v8,v16,const1)
+	lvx	v16,0,r4
+	VPERM(v16,v16,v16,byteswap)
+	ori	r2,r2,0
+
+	VPMSUMD(v9,v17,const1)
+	lvx	v17,off16,r4
+	VPERM(v17,v17,v17,byteswap)
+	ori	r2,r2,0
+
+	VPMSUMD(v10,v18,const1)
+	lvx	v18,off32,r4
+	VPERM(v18,v18,v18,byteswap)
+	ori	r2,r2,0
+
+	VPMSUMD(v11,v19,const1)
+	lvx	v19,off48,r4
+	VPERM(v19,v19,v19,byteswap)
+	ori	r2,r2,0
+
+	VPMSUMD(v12,v20,const1)
+	lvx	v20,off64,r4
+	VPERM(v20,v20,v20,byteswap)
+	ori	r2,r2,0
+
+	VPMSUMD(v13,v21,const1)
+	lvx	v21,off80,r4
+	VPERM(v21,v21,v21,byteswap)
+	ori	r2,r2,0
+
+	VPMSUMD(v14,v22,const1)
+	lvx	v22,off96,r4
+	VPERM(v22,v22,v22,byteswap)
+	ori	r2,r2,0
+
+	VPMSUMD(v15,v23,const1)
+	lvx	v23,off112,r4
+	VPERM(v23,v23,v23,byteswap)
+
+	addi	r4,r4,8*16
+
+	bdz	.Lfirst_cool_down
+
+	/*
+	 * main loop. We modulo schedule it such that it takes three iterations
+	 * to complete - first iteration load, second iteration vpmsum, third
+	 * iteration xor.
+	 */
+	.balign	16
+4:	lvx	const1,0,r3
+	addi	r3,r3,16
+	ori	r2,r2,0
+
+	vxor	v0,v0,v8
+	VPMSUMD(v8,v16,const2)
+	lvx	v16,0,r4
+	VPERM(v16,v16,v16,byteswap)
+	ori	r2,r2,0
+
+	vxor	v1,v1,v9
+	VPMSUMD(v9,v17,const2)
+	lvx	v17,off16,r4
+	VPERM(v17,v17,v17,byteswap)
+	ori	r2,r2,0
+
+	vxor	v2,v2,v10
+	VPMSUMD(v10,v18,const2)
+	lvx	v18,off32,r4
+	VPERM(v18,v18,v18,byteswap)
+	ori	r2,r2,0
+
+	vxor	v3,v3,v11
+	VPMSUMD(v11,v19,const2)
+	lvx	v19,off48,r4
+	VPERM(v19,v19,v19,byteswap)
+	lvx	const2,0,r3
+	ori	r2,r2,0
+
+	vxor	v4,v4,v12
+	VPMSUMD(v12,v20,const1)
+	lvx	v20,off64,r4
+	VPERM(v20,v20,v20,byteswap)
+	ori	r2,r2,0
+
+	vxor	v5,v5,v13
+	VPMSUMD(v13,v21,const1)
+	lvx	v21,off80,r4
+	VPERM(v21,v21,v21,byteswap)
+	ori	r2,r2,0
+
+	vxor	v6,v6,v14
+	VPMSUMD(v14,v22,const1)
+	lvx	v22,off96,r4
+	VPERM(v22,v22,v22,byteswap)
+	ori	r2,r2,0
+
+	vxor	v7,v7,v15
+	VPMSUMD(v15,v23,const1)
+	lvx	v23,off112,r4
+	VPERM(v23,v23,v23,byteswap)
+
+	addi	r4,r4,8*16
+
+	bdnz	4b
+
+.Lfirst_cool_down:
+	/* First cool down pass */
+	lvx	const1,0,r3
+	addi	r3,r3,16
+
+	vxor	v0,v0,v8
+	VPMSUMD(v8,v16,const1)
+	ori	r2,r2,0
+
+	vxor	v1,v1,v9
+	VPMSUMD(v9,v17,const1)
+	ori	r2,r2,0
+
+	vxor	v2,v2,v10
+	VPMSUMD(v10,v18,const1)
+	ori	r2,r2,0
+
+	vxor	v3,v3,v11
+	VPMSUMD(v11,v19,const1)
+	ori	r2,r2,0
+
+	vxor	v4,v4,v12
+	VPMSUMD(v12,v20,const1)
+	ori	r2,r2,0
+
+	vxor	v5,v5,v13
+	VPMSUMD(v13,v21,const1)
+	ori	r2,r2,0
+
+	vxor	v6,v6,v14
+	VPMSUMD(v14,v22,const1)
+	ori	r2,r2,0
+
+	vxor	v7,v7,v15
+	VPMSUMD(v15,v23,const1)
+	ori	r2,r2,0
+
+.Lsecond_cool_down:
+	/* Second cool down pass */
+	vxor	v0,v0,v8
+	vxor	v1,v1,v9
+	vxor	v2,v2,v10
+	vxor	v3,v3,v11
+	vxor	v4,v4,v12
+	vxor	v5,v5,v13
+	vxor	v6,v6,v14
+	vxor	v7,v7,v15
+
+#ifdef REFLECT
+	/*
+	 * vpmsumd produces a 96 bit result in the least significant bits
+	 * of the register. Since we are bit reflected we have to shift it
+	 * left 32 bits so it occupies the least significant bits in the
+	 * bit reflected domain.
+	 */
+	vsldoi	v0,v0,zeroes,4
+	vsldoi	v1,v1,zeroes,4
+	vsldoi	v2,v2,zeroes,4
+	vsldoi	v3,v3,zeroes,4
+	vsldoi	v4,v4,zeroes,4
+	vsldoi	v5,v5,zeroes,4
+	vsldoi	v6,v6,zeroes,4
+	vsldoi	v7,v7,zeroes,4
+#endif
+
+	/* xor with last 1024 bits */
+	lvx	v8,0,r4
+	lvx	v9,off16,r4
+	VPERM(v8,v8,v8,byteswap)
+	VPERM(v9,v9,v9,byteswap)
+	lvx	v10,off32,r4
+	lvx	v11,off48,r4
+	VPERM(v10,v10,v10,byteswap)
+	VPERM(v11,v11,v11,byteswap)
+	lvx	v12,off64,r4
+	lvx	v13,off80,r4
+	VPERM(v12,v12,v12,byteswap)
+	VPERM(v13,v13,v13,byteswap)
+	lvx	v14,off96,r4
+	lvx	v15,off112,r4
+	VPERM(v14,v14,v14,byteswap)
+	VPERM(v15,v15,v15,byteswap)
+
+	addi	r4,r4,8*16
+
+	vxor	v16,v0,v8
+	vxor	v17,v1,v9
+	vxor	v18,v2,v10
+	vxor	v19,v3,v11
+	vxor	v20,v4,v12
+	vxor	v21,v5,v13
+	vxor	v22,v6,v14
+	vxor	v23,v7,v15
+
+	li	r0,1
+	cmpdi	r6,0
+	addi	r6,r6,128
+	bne	1b
+
+	/* Work out how many bytes we have left */
+	andi.	r5,r5,127
+
+	/* Calculate where in the constant table we need to start */
+	subfic	r6,r5,128
+	add	r3,r3,r6
+
+	/* How many 16 byte chunks are in the tail */
+	srdi	r7,r5,4
+	mtctr	r7
+
+	/*
+	 * Reduce the previously calculated 1024 bits to 64 bits, shifting
+	 * 32 bits to include the trailing 32 bits of zeros
+	 */
+	lvx	v0,0,r3
+	lvx	v1,off16,r3
+	lvx	v2,off32,r3
+	lvx	v3,off48,r3
+	lvx	v4,off64,r3
+	lvx	v5,off80,r3
+	lvx	v6,off96,r3
+	lvx	v7,off112,r3
+	addi	r3,r3,8*16
+
+	VPMSUMW(v0,v16,v0)
+	VPMSUMW(v1,v17,v1)
+	VPMSUMW(v2,v18,v2)
+	VPMSUMW(v3,v19,v3)
+	VPMSUMW(v4,v20,v4)
+	VPMSUMW(v5,v21,v5)
+	VPMSUMW(v6,v22,v6)
+	VPMSUMW(v7,v23,v7)
+
+	/* Now reduce the tail (0 - 112 bytes) */
+	cmpdi	r7,0
+	beq	1f
+
+	lvx	v16,0,r4
+	lvx	v17,0,r3
+	VPERM(v16,v16,v16,byteswap)
+	VPMSUMW(v16,v16,v17)
+	vxor	v0,v0,v16
+	bdz	1f
+
+	lvx	v16,off16,r4
+	lvx	v17,off16,r3
+	VPERM(v16,v16,v16,byteswap)
+	VPMSUMW(v16,v16,v17)
+	vxor	v0,v0,v16
+	bdz	1f
+
+	lvx	v16,off32,r4
+	lvx	v17,off32,r3
+	VPERM(v16,v16,v16,byteswap)
+	VPMSUMW(v16,v16,v17)
+	vxor	v0,v0,v16
+	bdz	1f
+
+	lvx	v16,off48,r4
+	lvx	v17,off48,r3
+	VPERM(v16,v16,v16,byteswap)
+	VPMSUMW(v16,v16,v17)
+	vxor	v0,v0,v16
+	bdz	1f
+
+	lvx	v16,off64,r4
+	lvx	v17,off64,r3
+	VPERM(v16,v16,v16,byteswap)
+	VPMSUMW(v16,v16,v17)
+	vxor	v0,v0,v16
+	bdz	1f
+
+	lvx	v16,off80,r4
+	lvx	v17,off80,r3
+	VPERM(v16,v16,v16,byteswap)
+	VPMSUMW(v16,v16,v17)
+	vxor	v0,v0,v16
+	bdz	1f
+
+	lvx	v16,off96,r4
+	lvx	v17,off96,r3
+	VPERM(v16,v16,v16,byteswap)
+	VPMSUMW(v16,v16,v17)
+	vxor	v0,v0,v16
+
+	/* Now xor all the parallel chunks together */
+1:	vxor	v0,v0,v1
+	vxor	v2,v2,v3
+	vxor	v4,v4,v5
+	vxor	v6,v6,v7
+
+	vxor	v0,v0,v2
+	vxor	v4,v4,v6
+
+	vxor	v0,v0,v4
+
+.Lbarrett_reduction:
+	/* Barrett constants */
+	addis	r3,r2,.barrett_constants@toc@ha
+	addi	r3,r3,.barrett_constants@toc@l
+
+	lvx	const1,0,r3
+	lvx	const2,off16,r3
+
+	vsldoi	v1,v0,v0,8
+	vxor	v0,v0,v1		/* xor two 64 bit results together */
+
+#ifdef REFLECT
+	/* shift left one bit */
+	vspltisb v1,1
+	vsl	v0,v0,v1
+#endif
+
+	vand	v0,v0,mask_64bit
+
+#ifndef REFLECT
+	/*
+	 * Now for the Barrett reduction algorithm. The idea is to calculate q,
+	 * the multiple of our polynomial that we need to subtract. By
+	 * doing the computation 2x bits higher (ie 64 bits) and shifting the
+	 * result back down 2x bits, we round down to the nearest multiple.
+	 */
+	VPMSUMD(v1,v0,const1)	/* ma */
+	vsldoi	v1,zeroes,v1,8	/* q = floor(ma/(2^64)) */
+	VPMSUMD(v1,v1,const2)	/* qn */
+	vxor	v0,v0,v1	/* a - qn, subtraction is xor in GF(2) */
+
+	/*
+	 * Get the result into r3. We need to shift it left 8 bytes:
+	 * V0 [ 0 1 2 X ]
+	 * V0 [ 0 X 2 3 ]
+	 */
+	vsldoi	v0,v0,zeroes,8	/* shift result into top 64 bits */
+#else
+	/*
+	 * The reflected version of Barrett reduction. Instead of bit
+	 * reflecting our data (which is expensive to do), we bit reflect our
+	 * constants and our algorithm, which means the intermediate data in
+	 * our vector registers goes from 0-63 instead of 63-0. We can reflect
+	 * the algorithm because we don't carry in mod 2 arithmetic.
+	 */
+	vand	v1,v0,mask_32bit	/* bottom 32 bits of a */
+	VPMSUMD(v1,v1,const1)		/* ma */
+	vand	v1,v1,mask_32bit	/* bottom 32bits of ma */
+	VPMSUMD(v1,v1,const2)		/* qn */
+	vxor	v0,v0,v1		/* a - qn, subtraction is xor in GF(2) */
+
+	/*
+	 * Since we are bit reflected, the result (ie the low 32 bits) is in
+	 * the high 32 bits. We just need to shift it left 4 bytes
+	 * V0 [ 0 1 X 3 ]
+	 * V0 [ 0 X 2 3 ]
+	 */
+	vsldoi	v0,v0,zeroes,4		/* shift result into top 64 bits of */
+#endif
+
+	/* Get it into r3 */
+	MFVRD(r3, v0)
+
+.Lout:
+	subi	r6,r1,56+10*16
+	subi	r7,r1,56+2*16
+
+	lvx	v20,0,r6
+	lvx	v21,off16,r6
+	lvx	v22,off32,r6
+	lvx	v23,off48,r6
+	lvx	v24,off64,r6
+	lvx	v25,off80,r6
+	lvx	v26,off96,r6
+	lvx	v27,off112,r6
+	lvx	v28,0,r7
+	lvx	v29,off16,r7
+
+	ld	r31,-8(r1)
+	ld	r30,-16(r1)
+	ld	r29,-24(r1)
+	ld	r28,-32(r1)
+	ld	r27,-40(r1)
+	ld	r26,-48(r1)
+	ld	r25,-56(r1)
+
+	blr
+
+.Lfirst_warm_up_done:
+	lvx	const1,0,r3
+	addi	r3,r3,16
+
+	VPMSUMD(v8,v16,const1)
+	VPMSUMD(v9,v17,const1)
+	VPMSUMD(v10,v18,const1)
+	VPMSUMD(v11,v19,const1)
+	VPMSUMD(v12,v20,const1)
+	VPMSUMD(v13,v21,const1)
+	VPMSUMD(v14,v22,const1)
+	VPMSUMD(v15,v23,const1)
+
+	b	.Lsecond_cool_down
+
+.Lshort:
+	cmpdi	r5,0
+	beq	.Lzero
+
+	addis	r3,r2,.short_constants@toc@ha
+	addi	r3,r3,.short_constants@toc@l
+
+	/* Calculate where in the constant table we need to start */
+	subfic	r6,r5,256
+	add	r3,r3,r6
+
+	/* How many 16 byte chunks? */
+	srdi	r7,r5,4
+	mtctr	r7
+
+	vxor	v19,v19,v19
+	vxor	v20,v20,v20
+
+	lvx	v0,0,r4
+	lvx	v16,0,r3
+	VPERM(v0,v0,v16,byteswap)
+	vxor	v0,v0,v8	/* xor in initial value */
+	VPMSUMW(v0,v0,v16)
+	bdz	.Lv0
+
+	lvx	v1,off16,r4
+	lvx	v17,off16,r3
+	VPERM(v1,v1,v17,byteswap)
+	VPMSUMW(v1,v1,v17)
+	bdz	.Lv1
+
+	lvx	v2,off32,r4
+	lvx	v16,off32,r3
+	VPERM(v2,v2,v16,byteswap)
+	VPMSUMW(v2,v2,v16)
+	bdz	.Lv2
+
+	lvx	v3,off48,r4
+	lvx	v17,off48,r3
+	VPERM(v3,v3,v17,byteswap)
+	VPMSUMW(v3,v3,v17)
+	bdz	.Lv3
+
+	lvx	v4,off64,r4
+	lvx	v16,off64,r3
+	VPERM(v4,v4,v16,byteswap)
+	VPMSUMW(v4,v4,v16)
+	bdz	.Lv4
+
+	lvx	v5,off80,r4
+	lvx	v17,off80,r3
+	VPERM(v5,v5,v17,byteswap)
+	VPMSUMW(v5,v5,v17)
+	bdz	.Lv5
+
+	lvx	v6,off96,r4
+	lvx	v16,off96,r3
+	VPERM(v6,v6,v16,byteswap)
+	VPMSUMW(v6,v6,v16)
+	bdz	.Lv6
+
+	lvx	v7,off112,r4
+	lvx	v17,off112,r3
+	VPERM(v7,v7,v17,byteswap)
+	VPMSUMW(v7,v7,v17)
+	bdz	.Lv7
+
+	addi	r3,r3,128
+	addi	r4,r4,128
+
+	lvx	v8,0,r4
+	lvx	v16,0,r3
+	VPERM(v8,v8,v16,byteswap)
+	VPMSUMW(v8,v8,v16)
+	bdz	.Lv8
+
+	lvx	v9,off16,r4
+	lvx	v17,off16,r3
+	VPERM(v9,v9,v17,byteswap)
+	VPMSUMW(v9,v9,v17)
+	bdz	.Lv9
+
+	lvx	v10,off32,r4
+	lvx	v16,off32,r3
+	VPERM(v10,v10,v16,byteswap)
+	VPMSUMW(v10,v10,v16)
+	bdz	.Lv10
+
+	lvx	v11,off48,r4
+	lvx	v17,off48,r3
+	VPERM(v11,v11,v17,byteswap)
+	VPMSUMW(v11,v11,v17)
+	bdz	.Lv11
+
+	lvx	v12,off64,r4
+	lvx	v16,off64,r3
+	VPERM(v12,v12,v16,byteswap)
+	VPMSUMW(v12,v12,v16)
+	bdz	.Lv12
+
+	lvx	v13,off80,r4
+	lvx	v17,off80,r3
+	VPERM(v13,v13,v17,byteswap)
+	VPMSUMW(v13,v13,v17)
+	bdz	.Lv13
+
+	lvx	v14,off96,r4
+	lvx	v16,off96,r3
+	VPERM(v14,v14,v16,byteswap)
+	VPMSUMW(v14,v14,v16)
+	bdz	.Lv14
+
+	lvx	v15,off112,r4
+	lvx	v17,off112,r3
+	VPERM(v15,v15,v17,byteswap)
+	VPMSUMW(v15,v15,v17)
+
+.Lv15:	vxor	v19,v19,v15
+.Lv14:	vxor	v20,v20,v14
+.Lv13:	vxor	v19,v19,v13
+.Lv12:	vxor	v20,v20,v12
+.Lv11:	vxor	v19,v19,v11
+.Lv10:	vxor	v20,v20,v10
+.Lv9:	vxor	v19,v19,v9
+.Lv8:	vxor	v20,v20,v8
+.Lv7:	vxor	v19,v19,v7
+.Lv6:	vxor	v20,v20,v6
+.Lv5:	vxor	v19,v19,v5
+.Lv4:	vxor	v20,v20,v4
+.Lv3:	vxor	v19,v19,v3
+.Lv2:	vxor	v20,v20,v2
+.Lv1:	vxor	v19,v19,v1
+.Lv0:	vxor	v20,v20,v0
+
+	vxor	v0,v19,v20
+
+	b	.Lbarrett_reduction
+
+.Lzero:
+	mr	r3,r10
+	b	.Lout
+
+FUNC_END(__crc32_vpmsum)
diff --git a/src/rocksdb/util/crc32c_ppc_constants.h b/src/rocksdb/util/crc32c_ppc_constants.h
new file mode 100644
index 000000000..f6494cd01
--- /dev/null
+++ b/src/rocksdb/util/crc32c_ppc_constants.h
@@ -0,0 +1,900 @@
+//  Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.
+//  Copyright (C) 2015, 2017 International Business Machines Corp.
+//  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#pragma once
+
+#define CRC 0x1edc6f41
+#define REFLECT
+#define CRC_XOR
+
+#ifndef __ASSEMBLY__
+#ifdef CRC_TABLE
+static const unsigned int crc_table[] = {
+    0x00000000, 0xf26b8303, 0xe13b70f7, 0x1350f3f4, 0xc79a971f, 0x35f1141c,
+    0x26a1e7e8, 0xd4ca64eb, 0x8ad958cf, 0x78b2dbcc, 0x6be22838, 0x9989ab3b,
+    0x4d43cfd0, 0xbf284cd3, 0xac78bf27, 0x5e133c24, 0x105ec76f, 0xe235446c,
+    0xf165b798, 0x030e349b, 0xd7c45070, 0x25afd373, 0x36ff2087, 0xc494a384,
+    0x9a879fa0, 0x68ec1ca3, 0x7bbcef57, 0x89d76c54, 0x5d1d08bf, 0xaf768bbc,
+    0xbc267848, 0x4e4dfb4b, 0x20bd8ede, 0xd2d60ddd, 0xc186fe29, 0x33ed7d2a,
+    0xe72719c1, 0x154c9ac2, 0x061c6936, 0xf477ea35, 0xaa64d611, 0x580f5512,
+    0x4b5fa6e6, 0xb93425e5, 0x6dfe410e, 0x9f95c20d, 0x8cc531f9, 0x7eaeb2fa,
+    0x30e349b1, 0xc288cab2, 0xd1d83946, 0x23b3ba45, 0xf779deae, 0x05125dad,
+    0x1642ae59, 0xe4292d5a, 0xba3a117e, 0x4851927d, 0x5b016189, 0xa96ae28a,
+    0x7da08661, 0x8fcb0562, 0x9c9bf696, 0x6ef07595, 0x417b1dbc, 0xb3109ebf,
+    0xa0406d4b, 0x522bee48, 0x86e18aa3, 0x748a09a0, 0x67dafa54, 0x95b17957,
+    0xcba24573, 0x39c9c670, 0x2a993584, 0xd8f2b687, 0x0c38d26c, 0xfe53516f,
+    0xed03a29b, 0x1f682198, 0x5125dad3, 0xa34e59d0, 0xb01eaa24, 0x42752927,
+    0x96bf4dcc, 0x64d4cecf, 0x77843d3b, 0x85efbe38, 0xdbfc821c, 0x2997011f,
+    0x3ac7f2eb, 0xc8ac71e8, 0x1c661503, 0xee0d9600, 0xfd5d65f4, 0x0f36e6f7,
+    0x61c69362, 0x93ad1061, 0x80fde395, 0x72966096, 0xa65c047d, 0x5437877e,
+    0x4767748a, 0xb50cf789, 0xeb1fcbad, 0x197448ae, 0x0a24bb5a, 0xf84f3859,
+    0x2c855cb2, 0xdeeedfb1, 0xcdbe2c45, 0x3fd5af46, 0x7198540d, 0x83f3d70e,
+    0x90a324fa, 0x62c8a7f9, 0xb602c312, 0x44694011, 0x5739b3e5, 0xa55230e6,
+    0xfb410cc2, 0x092a8fc1, 0x1a7a7c35, 0xe811ff36, 0x3cdb9bdd, 0xceb018de,
+    0xdde0eb2a, 0x2f8b6829, 0x82f63b78, 0x709db87b, 0x63cd4b8f, 0x91a6c88c,
+    0x456cac67, 0xb7072f64, 0xa457dc90, 0x563c5f93, 0x082f63b7, 0xfa44e0b4,
+    0xe9141340, 0x1b7f9043, 0xcfb5f4a8, 0x3dde77ab, 0x2e8e845f, 0xdce5075c,
+    0x92a8fc17, 0x60c37f14, 0x73938ce0, 0x81f80fe3, 0x55326b08, 0xa759e80b,
+    0xb4091bff, 0x466298fc, 0x1871a4d8, 0xea1a27db, 0xf94ad42f, 0x0b21572c,
+    0xdfeb33c7, 0x2d80b0c4, 0x3ed04330, 0xccbbc033, 0xa24bb5a6, 0x502036a5,
+    0x4370c551, 0xb11b4652, 0x65d122b9, 0x97baa1ba, 0x84ea524e, 0x7681d14d,
+    0x2892ed69, 0xdaf96e6a, 0xc9a99d9e, 0x3bc21e9d, 0xef087a76, 0x1d63f975,
+    0x0e330a81, 0xfc588982, 0xb21572c9, 0x407ef1ca, 0x532e023e, 0xa145813d,
+    0x758fe5d6, 0x87e466d5, 0x94b49521, 0x66df1622, 0x38cc2a06, 0xcaa7a905,
+    0xd9f75af1, 0x2b9cd9f2, 0xff56bd19, 0x0d3d3e1a, 0x1e6dcdee, 0xec064eed,
+    0xc38d26c4, 0x31e6a5c7, 0x22b65633, 0xd0ddd530, 0x0417b1db, 0xf67c32d8,
+    0xe52cc12c, 0x1747422f, 0x49547e0b, 0xbb3ffd08, 0xa86f0efc, 0x5a048dff,
+    0x8ecee914, 0x7ca56a17, 0x6ff599e3, 0x9d9e1ae0, 0xd3d3e1ab, 0x21b862a8,
+    0x32e8915c, 0xc083125f, 0x144976b4, 0xe622f5b7, 0xf5720643, 0x07198540,
+    0x590ab964, 0xab613a67, 0xb831c993, 0x4a5a4a90, 0x9e902e7b, 0x6cfbad78,
+    0x7fab5e8c, 0x8dc0dd8f, 0xe330a81a, 0x115b2b19, 0x020bd8ed, 0xf0605bee,
+    0x24aa3f05, 0xd6c1bc06, 0xc5914ff2, 0x37faccf1, 0x69e9f0d5, 0x9b8273d6,
+    0x88d28022, 0x7ab90321, 0xae7367ca, 0x5c18e4c9, 0x4f48173d, 0xbd23943e,
+    0xf36e6f75, 0x0105ec76, 0x12551f82, 0xe03e9c81, 0x34f4f86a, 0xc69f7b69,
+    0xd5cf889d, 0x27a40b9e, 0x79b737ba, 0x8bdcb4b9, 0x988c474d, 0x6ae7c44e,
+    0xbe2da0a5, 0x4c4623a6, 0x5f16d052, 0xad7d5351,
+};
+
+#endif
+
+#else
+#define MAX_SIZE 32768
+.constants :
+
+        /* Reduce 262144 kbits to 1024 bits */
+        /* x^261120 mod p(x)` << 1, x^261184 mod p(x)` << 1 */
+        .octa 0x00000000b6ca9e20000000009c37c408
+
+        /* x^260096 mod p(x)` << 1, x^260160 mod p(x)` << 1 */
+        .octa 0x00000000350249a800000001b51df26c
+
+        /* x^259072 mod p(x)` << 1, x^259136 mod p(x)` << 1 */
+        .octa 0x00000001862dac54000000000724b9d0
+
+        /* x^258048 mod p(x)` << 1, x^258112 mod p(x)` << 1 */
+        .octa 0x00000001d87fb48c00000001c00532fe
+
+        /* x^257024 mod p(x)` << 1, x^257088 mod p(x)` << 1 */
+        .octa 0x00000001f39b699e00000000f05a9362
+
+        /* x^256000 mod p(x)` << 1, x^256064 mod p(x)` << 1 */
+        .octa 0x0000000101da11b400000001e1007970
+
+        /* x^254976 mod p(x)` << 1, x^255040 mod p(x)` << 1 */
+        .octa 0x00000001cab571e000000000a57366ee
+
+        /* x^253952 mod p(x)` << 1, x^254016 mod p(x)` << 1 */
+        .octa 0x00000000c7020cfe0000000192011284
+
+        /* x^252928 mod p(x)` << 1, x^252992 mod p(x)` << 1 */
+        .octa 0x00000000cdaed1ae0000000162716d9a
+
+        /* x^251904 mod p(x)` << 1, x^251968 mod p(x)` << 1 */
+        .octa 0x00000001e804effc00000000cd97ecde
+
+        /* x^250880 mod p(x)` << 1, x^250944 mod p(x)` << 1 */
+        .octa 0x0000000077c3ea3a0000000058812bc0
+
+        /* x^249856 mod p(x)` << 1, x^249920 mod p(x)` << 1 */
+        .octa 0x0000000068df31b40000000088b8c12e
+
+        /* x^248832 mod p(x)` << 1, x^248896 mod p(x)` << 1 */
+        .octa 0x00000000b059b6c200000001230b234c
+
+        /* x^247808 mod p(x)` << 1, x^247872 mod p(x)` << 1 */
+        .octa 0x0000000145fb8ed800000001120b416e
+
+        /* x^246784 mod p(x)` << 1, x^246848 mod p(x)` << 1 */
+        .octa 0x00000000cbc0916800000001974aecb0
+
+        /* x^245760 mod p(x)` << 1, x^245824 mod p(x)` << 1 */
+        .octa 0x000000005ceeedc2000000008ee3f226
+
+        /* x^244736 mod p(x)` << 1, x^244800 mod p(x)` << 1 */
+        .octa 0x0000000047d74e8600000001089aba9a
+
+        /* x^243712 mod p(x)` << 1, x^243776 mod p(x)` << 1 */
+        .octa 0x00000001407e9e220000000065113872
+
+        /* x^242688 mod p(x)` << 1, x^242752 mod p(x)` << 1 */
+        .octa 0x00000001da967bda000000005c07ec10
+
+        /* x^241664 mod p(x)` << 1, x^241728 mod p(x)` << 1 */
+        .octa 0x000000006c8983680000000187590924
+
+        /* x^240640 mod p(x)` << 1, x^240704 mod p(x)` << 1 */
+        .octa 0x00000000f2d14c9800000000e35da7c6
+
+        /* x^239616 mod p(x)` << 1, x^239680 mod p(x)` << 1 */
+        .octa 0x00000001993c6ad4000000000415855a
+
+        /* x^238592 mod p(x)` << 1, x^238656 mod p(x)` << 1 */
+        .octa 0x000000014683d1ac0000000073617758
+
+        /* x^237568 mod p(x)` << 1, x^237632 mod p(x)` << 1 */
+        .octa 0x00000001a7c93e6c0000000176021d28
+
+        /* x^236544 mod p(x)` << 1, x^236608 mod p(x)` << 1 */
+        .octa 0x000000010211e90a00000001c358fd0a
+
+        /* x^235520 mod p(x)` << 1, x^235584 mod p(x)` << 1 */
+        .octa 0x000000001119403e00000001ff7a2c18
+
+        /* x^234496 mod p(x)` << 1, x^234560 mod p(x)` << 1 */
+        .octa 0x000000001c3261aa00000000f2d9f7e4
+
+        /* x^233472 mod p(x)` << 1, x^233536 mod p(x)` << 1 */
+        .octa 0x000000014e37a634000000016cf1f9c8
+
+        /* x^232448 mod p(x)` << 1, x^232512 mod p(x)` << 1 */
+        .octa 0x0000000073786c0c000000010af9279a
+
+        /* x^231424 mod p(x)` << 1, x^231488 mod p(x)` << 1 */
+        .octa 0x000000011dc037f80000000004f101e8
+
+        /* x^230400 mod p(x)` << 1, x^230464 mod p(x)` << 1 */
+        .octa 0x0000000031433dfc0000000070bcf184
+
+        /* x^229376 mod p(x)` << 1, x^229440 mod p(x)` << 1 */
+        .octa 0x000000009cde8348000000000a8de642
+
+        /* x^228352 mod p(x)` << 1, x^228416 mod p(x)` << 1 */
+        .octa 0x0000000038d3c2a60000000062ea130c
+
+        /* x^227328 mod p(x)` << 1, x^227392 mod p(x)` << 1 */
+        .octa 0x000000011b25f26000000001eb31cbb2
+
+        /* x^226304 mod p(x)` << 1, x^226368 mod p(x)` << 1 */
+        .octa 0x000000001629e6f00000000170783448
+
+        /* x^225280 mod p(x)` << 1, x^225344 mod p(x)` << 1 */
+        .octa 0x0000000160838b4c00000001a684b4c6
+
+        /* x^224256 mod p(x)` << 1, x^224320 mod p(x)` << 1 */
+        .octa 0x000000007a44011c00000000253ca5b4
+
+        /* x^223232 mod p(x)` << 1, x^223296 mod p(x)` << 1 */
+        .octa 0x00000000226f417a0000000057b4b1e2
+
+        /* x^222208 mod p(x)` << 1, x^222272 mod p(x)` << 1 */
+        .octa 0x0000000045eb2eb400000000b6bd084c
+
+        /* x^221184 mod p(x)` << 1, x^221248 mod p(x)` << 1 */
+        .octa 0x000000014459d70c0000000123c2d592
+
+        /* x^220160 mod p(x)` << 1, x^220224 mod p(x)` << 1 */
+        .octa 0x00000001d406ed8200000000159dafce
+
+        /* x^219136 mod p(x)` << 1, x^219200 mod p(x)` << 1 */
+        .octa 0x0000000160c8e1a80000000127e1a64e
+
+        /* x^218112 mod p(x)` << 1, x^218176 mod p(x)` << 1 */
+        .octa 0x0000000027ba80980000000056860754
+
+        /* x^217088 mod p(x)` << 1, x^217152 mod p(x)` << 1 */
+        .octa 0x000000006d92d01800000001e661aae8
+
+        /* x^216064 mod p(x)` << 1, x^216128 mod p(x)` << 1 */
+        .octa 0x000000012ed7e3f200000000f82c6166
+
+        /* x^215040 mod p(x)` << 1, x^215104 mod p(x)` << 1 */
+        .octa 0x000000002dc8778800000000c4f9c7ae
+
+        /* x^214016 mod p(x)` << 1, x^214080 mod p(x)` << 1 */
+        .octa 0x0000000018240bb80000000074203d20
+
+        /* x^212992 mod p(x)` << 1, x^213056 mod p(x)` << 1 */
+        .octa 0x000000001ad381580000000198173052
+
+        /* x^211968 mod p(x)` << 1, x^212032 mod p(x)` << 1 */
+        .octa 0x00000001396b78f200000001ce8aba54
+
+        /* x^210944 mod p(x)` << 1, x^211008 mod p(x)` << 1 */
+        .octa 0x000000011a68133400000001850d5d94
+
+        /* x^209920 mod p(x)` << 1, x^209984 mod p(x)` << 1 */
+        .octa 0x000000012104732e00000001d609239c
+
+        /* x^208896 mod p(x)` << 1, x^208960 mod p(x)` << 1 */
+        .octa 0x00000000a140d90c000000001595f048
+
+        /* x^207872 mod p(x)` << 1, x^207936 mod p(x)` << 1 */
+        .octa 0x00000001b7215eda0000000042ccee08
+
+        /* x^206848 mod p(x)` << 1, x^206912 mod p(x)` << 1 */
+        .octa 0x00000001aaf1df3c000000010a389d74
+
+        /* x^205824 mod p(x)` << 1, x^205888 mod p(x)` << 1 */
+        .octa 0x0000000029d15b8a000000012a840da6
+
+        /* x^204800 mod p(x)` << 1, x^204864 mod p(x)` << 1 */
+        .octa 0x00000000f1a96922000000001d181c0c
+
+        /* x^203776 mod p(x)` << 1, x^203840 mod p(x)` << 1 */
+        .octa 0x00000001ac80d03c0000000068b7d1f6
+
+        /* x^202752 mod p(x)` << 1, x^202816 mod p(x)` << 1 */
+        .octa 0x000000000f11d56a000000005b0f14fc
+
+        /* x^201728 mod p(x)` << 1, x^201792 mod p(x)` << 1 */
+        .octa 0x00000001f1c022a20000000179e9e730
+
+        /* x^200704 mod p(x)` << 1, x^200768 mod p(x)` << 1 */
+        .octa 0x0000000173d00ae200000001ce1368d6
+
+        /* x^199680 mod p(x)` << 1, x^199744 mod p(x)` << 1 */
+        .octa 0x00000001d4ffe4ac0000000112c3a84c
+
+        /* x^198656 mod p(x)` << 1, x^198720 mod p(x)` << 1 */
+        .octa 0x000000016edc5ae400000000de940fee
+
+        /* x^197632 mod p(x)` << 1, x^197696 mod p(x)` << 1 */
+        .octa 0x00000001f1a0214000000000fe896b7e
+
+        /* x^196608 mod p(x)` << 1, x^196672 mod p(x)` << 1 */
+        .octa 0x00000000ca0b28a000000001f797431c
+
+        /* x^195584 mod p(x)` << 1, x^195648 mod p(x)` << 1 */
+        .octa 0x00000001928e30a20000000053e989ba
+
+        /* x^194560 mod p(x)` << 1, x^194624 mod p(x)` << 1 */
+        .octa 0x0000000097b1b002000000003920cd16
+
+        /* x^193536 mod p(x)` << 1, x^193600 mod p(x)` << 1 */
+        .octa 0x00000000b15bf90600000001e6f579b8
+
+        /* x^192512 mod p(x)` << 1, x^192576 mod p(x)` << 1 */
+        .octa 0x00000000411c5d52000000007493cb0a
+
+        /* x^191488 mod p(x)` << 1, x^191552 mod p(x)` << 1 */
+        .octa 0x00000001c36f330000000001bdd376d8
+
+        /* x^190464 mod p(x)` << 1, x^190528 mod p(x)` << 1 */
+        .octa 0x00000001119227e0000000016badfee6
+
+        /* x^189440 mod p(x)` << 1, x^189504 mod p(x)` << 1 */
+        .octa 0x00000000114d47020000000071de5c58
+
+        /* x^188416 mod p(x)` << 1, x^188480 mod p(x)` << 1 */
+        .octa 0x00000000458b5b9800000000453f317c
+
+        /* x^187392 mod p(x)` << 1, x^187456 mod p(x)` << 1 */
+        .octa 0x000000012e31fb8e0000000121675cce
+
+        /* x^186368 mod p(x)` << 1, x^186432 mod p(x)` << 1 */
+        .octa 0x000000005cf619d800000001f409ee92
+
+        /* x^185344 mod p(x)` << 1, x^185408 mod p(x)` << 1 */
+        .octa 0x0000000063f4d8b200000000f36b9c88
+
+        /* x^184320 mod p(x)` << 1, x^184384 mod p(x)` << 1 */
+        .octa 0x000000004138dc8a0000000036b398f4
+
+        /* x^183296 mod p(x)` << 1, x^183360 mod p(x)` << 1 */
+        .octa 0x00000001d29ee8e000000001748f9adc
+
+        /* x^182272 mod p(x)` << 1, x^182336 mod p(x)` << 1 */
+        .octa 0x000000006a08ace800000001be94ec00
+
+        /* x^181248 mod p(x)` << 1, x^181312 mod p(x)` << 1 */
+        .octa 0x0000000127d4201000000000b74370d6
+
+        /* x^180224 mod p(x)` << 1, x^180288 mod p(x)` << 1 */
+        .octa 0x0000000019d76b6200000001174d0b98
+
+        /* x^179200 mod p(x)` << 1, x^179264 mod p(x)` << 1 */
+        .octa 0x00000001b1471f6e00000000befc06a4
+
+        /* x^178176 mod p(x)` << 1, x^178240 mod p(x)` << 1 */
+        .octa 0x00000001f64c19cc00000001ae125288
+
+        /* x^177152 mod p(x)` << 1, x^177216 mod p(x)` << 1 */
+        .octa 0x00000000003c0ea00000000095c19b34
+
+        /* x^176128 mod p(x)` << 1, x^176192 mod p(x)` << 1 */
+        .octa 0x000000014d73abf600000001a78496f2
+
+        /* x^175104 mod p(x)` << 1, x^175168 mod p(x)` << 1 */
+        .octa 0x00000001620eb84400000001ac5390a0
+
+        /* x^174080 mod p(x)` << 1, x^174144 mod p(x)` << 1 */
+        .octa 0x0000000147655048000000002a80ed6e
+
+        /* x^173056 mod p(x)` << 1, x^173120 mod p(x)` << 1 */
+        .octa 0x0000000067b5077e00000001fa9b0128
+
+        /* x^172032 mod p(x)` << 1, x^172096 mod p(x)` << 1 */
+        .octa 0x0000000010ffe20600000001ea94929e
+
+        /* x^171008 mod p(x)` << 1, x^171072 mod p(x)` << 1 */
+        .octa 0x000000000fee8f1e0000000125f4305c
+
+        /* x^169984 mod p(x)` << 1, x^170048 mod p(x)` << 1 */
+        .octa 0x00000001da26fbae00000001471e2002
+
+        /* x^168960 mod p(x)` << 1, x^169024 mod p(x)` << 1 */
+        .octa 0x00000001b3a8bd880000000132d2253a
+
+        /* x^167936 mod p(x)` << 1, x^168000 mod p(x)` << 1 */
+        .octa 0x00000000e8f3898e00000000f26b3592
+
+        /* x^166912 mod p(x)` << 1, x^166976 mod p(x)` << 1 */
+        .octa 0x00000000b0d0d28c00000000bc8b67b0
+
+        /* x^165888 mod p(x)` << 1, x^165952 mod p(x)` << 1 */
+        .octa 0x0000000030f2a798000000013a826ef2
+
+        /* x^164864 mod p(x)` << 1, x^164928 mod p(x)` << 1 */
+        .octa 0x000000000fba10020000000081482c84
+
+        /* x^163840 mod p(x)` << 1, x^163904 mod p(x)` << 1 */
+        .octa 0x00000000bdb9bd7200000000e77307c2
+
+        /* x^162816 mod p(x)` << 1, x^162880 mod p(x)` << 1 */
+        .octa 0x0000000075d3bf5a00000000d4a07ec8
+
+        /* x^161792 mod p(x)` << 1, x^161856 mod p(x)` << 1 */
+        .octa 0x00000000ef1f98a00000000017102100
+
+        /* x^160768 mod p(x)` << 1, x^160832 mod p(x)` << 1 */
+        .octa 0x00000000689c760200000000db406486
+
+        /* x^159744 mod p(x)` << 1, x^159808 mod p(x)` << 1 */
+        .octa 0x000000016d5fa5fe0000000192db7f88
+
+        /* x^158720 mod p(x)` << 1, x^158784 mod p(x)` << 1 */
+        .octa 0x00000001d0d2b9ca000000018bf67b1e
+
+        /* x^157696 mod p(x)` << 1, x^157760 mod p(x)` << 1 */
+        .octa 0x0000000041e7b470000000007c09163e
+
+        /* x^156672 mod p(x)` << 1, x^156736 mod p(x)` << 1 */
+        .octa 0x00000001cbb6495e000000000adac060
+
+        /* x^155648 mod p(x)` << 1, x^155712 mod p(x)` << 1 */
+        .octa 0x000000010052a0b000000000bd8316ae
+
+        /* x^154624 mod p(x)` << 1, x^154688 mod p(x)` << 1 */
+        .octa 0x00000001d8effb5c000000019f09ab54
+
+        /* x^153600 mod p(x)` << 1, x^153664 mod p(x)` << 1 */
+        .octa 0x00000001d969853c0000000125155542
+
+        /* x^152576 mod p(x)` << 1, x^152640 mod p(x)` << 1 */
+        .octa 0x00000000523ccce2000000018fdb5882
+
+        /* x^151552 mod p(x)` << 1, x^151616 mod p(x)` << 1 */
+        .octa 0x000000001e2436bc00000000e794b3f4
+
+        /* x^150528 mod p(x)` << 1, x^150592 mod p(x)` << 1 */
+        .octa 0x00000000ddd1c3a2000000016f9bb022
+
+        /* x^149504 mod p(x)` << 1, x^149568 mod p(x)` << 1 */
+        .octa 0x0000000019fcfe3800000000290c9978
+
+        /* x^148480 mod p(x)` << 1, x^148544 mod p(x)` << 1 */
+        .octa 0x00000001ce95db640000000083c0f350
+
+        /* x^147456 mod p(x)` << 1, x^147520 mod p(x)` << 1 */
+        .octa 0x00000000af5828060000000173ea6628
+
+        /* x^146432 mod p(x)` << 1, x^146496 mod p(x)` << 1 */
+        .octa 0x00000001006388f600000001c8b4e00a
+
+        /* x^145408 mod p(x)` << 1, x^145472 mod p(x)` << 1 */
+        .octa 0x0000000179eca00a00000000de95d6aa
+
+        /* x^144384 mod p(x)` << 1, x^144448 mod p(x)` << 1 */
+        .octa 0x0000000122410a6a000000010b7f7248
+
+        /* x^143360 mod p(x)` << 1, x^143424 mod p(x)` << 1 */
+        .octa 0x000000004288e87c00000001326e3a06
+
+        /* x^142336 mod p(x)` << 1, x^142400 mod p(x)` << 1 */
+        .octa 0x000000016c5490da00000000bb62c2e6
+
+        /* x^141312 mod p(x)` << 1, x^141376 mod p(x)` << 1 */
+        .octa 0x00000000d1c71f6e0000000156a4b2c2
+
+        /* x^140288 mod p(x)` << 1, x^140352 mod p(x)` << 1 */
+        .octa 0x00000001b4ce08a6000000011dfe763a
+
+        /* x^139264 mod p(x)` << 1, x^139328 mod p(x)` << 1 */
+        .octa 0x00000001466ba60c000000007bcca8e2
+
+        /* x^138240 mod p(x)` << 1, x^138304 mod p(x)` << 1 */
+        .octa 0x00000001f6c488a40000000186118faa
+
+        /* x^137216 mod p(x)` << 1, x^137280 mod p(x)` << 1 */
+        .octa 0x000000013bfb06820000000111a65a88
+
+        /* x^136192 mod p(x)` << 1, x^136256 mod p(x)` << 1 */
+        .octa 0x00000000690e9e54000000003565e1c4
+
+        /* x^135168 mod p(x)` << 1, x^135232 mod p(x)` << 1 */
+        .octa 0x00000000281346b6000000012ed02a82
+
+        /* x^134144 mod p(x)` << 1, x^134208 mod p(x)` << 1 */
+        .octa 0x000000015646402400000000c486ecfc
+
+        /* x^133120 mod p(x)` << 1, x^133184 mod p(x)` << 1 */
+        .octa 0x000000016063a8dc0000000001b951b2
+
+        /* x^132096 mod p(x)` << 1, x^132160 mod p(x)` << 1 */
+        .octa 0x0000000116a663620000000048143916
+
+        /* x^131072 mod p(x)` << 1, x^131136 mod p(x)` << 1 */
+        .octa 0x000000017e8aa4d200000001dc2ae124
+
+        /* x^130048 mod p(x)` << 1, x^130112 mod p(x)` << 1 */
+        .octa 0x00000001728eb10c00000001416c58d6
+
+        /* x^129024 mod p(x)` << 1, x^129088 mod p(x)` << 1 */
+        .octa 0x00000001b08fd7fa00000000a479744a
+
+        /* x^128000 mod p(x)` << 1, x^128064 mod p(x)` << 1 */
+        .octa 0x00000001092a16e80000000096ca3a26
+
+        /* x^126976 mod p(x)` << 1, x^127040 mod p(x)` << 1 */
+        .octa 0x00000000a505637c00000000ff223d4e
+
+        /* x^125952 mod p(x)` << 1, x^126016 mod p(x)` << 1 */
+        .octa 0x00000000d94869b2000000010e84da42
+
+        /* x^124928 mod p(x)` << 1, x^124992 mod p(x)` << 1 */
+        .octa 0x00000001c8b203ae00000001b61ba3d0
+
+        /* x^123904 mod p(x)` << 1, x^123968 mod p(x)` << 1 */
+        .octa 0x000000005704aea000000000680f2de8
+
+        /* x^122880 mod p(x)` << 1, x^122944 mod p(x)` << 1 */
+        .octa 0x000000012e295fa2000000008772a9a8
+
+        /* x^121856 mod p(x)` << 1, x^121920 mod p(x)` << 1 */
+        .octa 0x000000011d0908bc0000000155f295bc
+
+        /* x^120832 mod p(x)` << 1, x^120896 mod p(x)` << 1 */
+        .octa 0x0000000193ed97ea00000000595f9282
+
+        /* x^119808 mod p(x)` << 1, x^119872 mod p(x)` << 1 */
+        .octa 0x000000013a0f1c520000000164b1c25a
+
+        /* x^118784 mod p(x)` << 1, x^118848 mod p(x)` << 1 */
+        .octa 0x000000010c2c40c000000000fbd67c50
+
+        /* x^117760 mod p(x)` << 1, x^117824 mod p(x)` << 1 */
+        .octa 0x00000000ff6fac3e0000000096076268
+
+        /* x^116736 mod p(x)` << 1, x^116800 mod p(x)` << 1 */
+        .octa 0x000000017b3609c000000001d288e4cc
+
+        /* x^115712 mod p(x)` << 1, x^115776 mod p(x)` << 1 */
+        .octa 0x0000000088c8c92200000001eaac1bdc
+
+        /* x^114688 mod p(x)` << 1, x^114752 mod p(x)` << 1 */
+        .octa 0x00000001751baae600000001f1ea39e2
+
+        /* x^113664 mod p(x)` << 1, x^113728 mod p(x)` << 1 */
+        .octa 0x000000010795297200000001eb6506fc
+
+        /* x^112640 mod p(x)` << 1, x^112704 mod p(x)` << 1 */
+        .octa 0x0000000162b00abe000000010f806ffe
+
+        /* x^111616 mod p(x)` << 1, x^111680 mod p(x)` << 1 */
+        .octa 0x000000000d7b404c000000010408481e
+
+        /* x^110592 mod p(x)` << 1, x^110656 mod p(x)` << 1 */
+        .octa 0x00000000763b13d40000000188260534
+
+        /* x^109568 mod p(x)` << 1, x^109632 mod p(x)` << 1 */
+        .octa 0x00000000f6dc22d80000000058fc73e0
+
+        /* x^108544 mod p(x)` << 1, x^108608 mod p(x)` << 1 */
+        .octa 0x000000007daae06000000000391c59b8
+
+        /* x^107520 mod p(x)` << 1, x^107584 mod p(x)` << 1 */
+        .octa 0x000000013359ab7c000000018b638400
+
+        /* x^106496 mod p(x)` << 1, x^106560 mod p(x)` << 1 */
+        .octa 0x000000008add438a000000011738f5c4
+
+        /* x^105472 mod p(x)` << 1, x^105536 mod p(x)` << 1 */
+        .octa 0x00000001edbefdea000000008cf7c6da
+
+        /* x^104448 mod p(x)` << 1, x^104512 mod p(x)` << 1 */
+        .octa 0x000000004104e0f800000001ef97fb16
+
+        /* x^103424 mod p(x)` << 1, x^103488 mod p(x)` << 1 */
+        .octa 0x00000000b48a82220000000102130e20
+
+        /* x^102400 mod p(x)` << 1, x^102464 mod p(x)` << 1 */
+        .octa 0x00000001bcb4684400000000db968898
+
+        /* x^101376 mod p(x)` << 1, x^101440 mod p(x)` << 1 */
+        .octa 0x000000013293ce0a00000000b5047b5e
+
+        /* x^100352 mod p(x)` << 1, x^100416 mod p(x)` << 1 */
+        .octa 0x00000001710d0844000000010b90fdb2
+
+        /* x^99328 mod p(x)` << 1, x^99392 mod p(x)` << 1 */
+        .octa 0x0000000117907f6e000000004834a32e
+
+        /* x^98304 mod p(x)` << 1, x^98368 mod p(x)` << 1 */
+        .octa 0x0000000087ddf93e0000000059c8f2b0
+
+        /* x^97280 mod p(x)` << 1, x^97344 mod p(x)` << 1 */
+        .octa 0x000000005970e9b00000000122cec508
+
+        /* x^96256 mod p(x)` << 1, x^96320 mod p(x)` << 1 */
+        .octa 0x0000000185b2b7d0000000000a330cda
+
+        /* x^95232 mod p(x)` << 1, x^95296 mod p(x)` << 1 */
+        .octa 0x00000001dcee0efc000000014a47148c
+
+        /* x^94208 mod p(x)` << 1, x^94272 mod p(x)` << 1 */
+        .octa 0x0000000030da27220000000042c61cb8
+
+        /* x^93184 mod p(x)` << 1, x^93248 mod p(x)` << 1 */
+        .octa 0x000000012f925a180000000012fe6960
+
+        /* x^92160 mod p(x)` << 1, x^92224 mod p(x)` << 1 */
+        .octa 0x00000000dd2e357c00000000dbda2c20
+
+        /* x^91136 mod p(x)` << 1, x^91200 mod p(x)` << 1 */
+        .octa 0x00000000071c80de000000011122410c
+
+        /* x^90112 mod p(x)` << 1, x^90176 mod p(x)` << 1 */
+        .octa 0x000000011513140a00000000977b2070
+
+        /* x^89088 mod p(x)` << 1, x^89152 mod p(x)` << 1 */
+        .octa 0x00000001df876e8e000000014050438e
+
+        /* x^88064 mod p(x)` << 1, x^88128 mod p(x)` << 1 */
+        .octa 0x000000015f81d6ce0000000147c840e8
+
+        /* x^87040 mod p(x)` << 1, x^87104 mod p(x)` << 1 */
+        .octa 0x000000019dd94dbe00000001cc7c88ce
+
+        /* x^86016 mod p(x)` << 1, x^86080 mod p(x)` << 1 */
+        .octa 0x00000001373d206e00000001476b35a4
+
+        /* x^84992 mod p(x)` << 1, x^85056 mod p(x)` << 1 */
+        .octa 0x00000000668ccade000000013d52d508
+
+        /* x^83968 mod p(x)` << 1, x^84032 mod p(x)` << 1 */
+        .octa 0x00000001b192d268000000008e4be32e
+
+        /* x^82944 mod p(x)` << 1, x^83008 mod p(x)` << 1 */
+        .octa 0x00000000e30f3a7800000000024120fe
+
+        /* x^81920 mod p(x)` << 1, x^81984 mod p(x)` << 1 */
+        .octa 0x000000010ef1f7bc00000000ddecddb4
+
+        /* x^80896 mod p(x)` << 1, x^80960 mod p(x)` << 1 */
+        .octa 0x00000001f5ac738000000000d4d403bc
+
+        /* x^79872 mod p(x)` << 1, x^79936 mod p(x)` << 1 */
+        .octa 0x000000011822ea7000000001734b89aa
+
+        /* x^78848 mod p(x)` << 1, x^78912 mod p(x)` << 1 */
+        .octa 0x00000000c3a33848000000010e7a58d6
+
+        /* x^77824 mod p(x)` << 1, x^77888 mod p(x)` << 1 */
+        .octa 0x00000001bd151c2400000001f9f04e9c
+
+        /* x^76800 mod p(x)` << 1, x^76864 mod p(x)` << 1 */
+        .octa 0x0000000056002d7600000000b692225e
+
+        /* x^75776 mod p(x)` << 1, x^75840 mod p(x)` << 1 */
+        .octa 0x000000014657c4f4000000019b8d3f3e
+
+        /* x^74752 mod p(x)` << 1, x^74816 mod p(x)` << 1 */
+        .octa 0x0000000113742d7c00000001a874f11e
+
+        /* x^73728 mod p(x)` << 1, x^73792 mod p(x)` << 1 */
+        .octa 0x000000019c5920ba000000010d5a4254
+
+        /* x^72704 mod p(x)` << 1, x^72768 mod p(x)` << 1 */
+        .octa 0x000000005216d2d600000000bbb2f5d6
+
+        /* x^71680 mod p(x)` << 1, x^71744 mod p(x)` << 1 */
+        .octa 0x0000000136f5ad8a0000000179cc0e36
+
+        /* x^70656 mod p(x)` << 1, x^70720 mod p(x)` << 1 */
+        .octa 0x000000018b07beb600000001dca1da4a
+
+        /* x^69632 mod p(x)` << 1, x^69696 mod p(x)` << 1 */
+        .octa 0x00000000db1e93b000000000feb1a192
+
+        /* x^68608 mod p(x)` << 1, x^68672 mod p(x)` << 1 */
+        .octa 0x000000000b96fa3a00000000d1eeedd6
+
+        /* x^67584 mod p(x)` << 1, x^67648 mod p(x)` << 1 */
+        .octa 0x00000001d9968af0000000008fad9bb4
+
+        /* x^66560 mod p(x)` << 1, x^66624 mod p(x)` << 1 */
+        .octa 0x000000000e4a77a200000001884938e4
+
+        /* x^65536 mod p(x)` << 1, x^65600 mod p(x)` << 1 */
+        .octa 0x00000000508c2ac800000001bc2e9bc0
+
+        /* x^64512 mod p(x)` << 1, x^64576 mod p(x)` << 1 */
+        .octa 0x0000000021572a8000000001f9658a68
+
+        /* x^63488 mod p(x)` << 1, x^63552 mod p(x)` << 1 */
+        .octa 0x00000001b859daf2000000001b9224fc
+
+        /* x^62464 mod p(x)` << 1, x^62528 mod p(x)` << 1 */
+        .octa 0x000000016f7884740000000055b2fb84
+
+        /* x^61440 mod p(x)` << 1, x^61504 mod p(x)` << 1 */
+        .octa 0x00000001b438810e000000018b090348
+
+        /* x^60416 mod p(x)` << 1, x^60480 mod p(x)` << 1 */
+        .octa 0x0000000095ddc6f2000000011ccbd5ea
+
+        /* x^59392 mod p(x)` << 1, x^59456 mod p(x)` << 1 */
+        .octa 0x00000001d977c20c0000000007ae47f8
+
+        /* x^58368 mod p(x)` << 1, x^58432 mod p(x)` << 1 */
+        .octa 0x00000000ebedb99a0000000172acbec0
+
+        /* x^57344 mod p(x)` << 1, x^57408 mod p(x)` << 1 */
+        .octa 0x00000001df9e9e9200000001c6e3ff20
+
+        /* x^56320 mod p(x)` << 1, x^56384 mod p(x)` << 1 */
+        .octa 0x00000001a4a3f95200000000e1b38744
+
+        /* x^55296 mod p(x)` << 1, x^55360 mod p(x)` << 1 */
+        .octa 0x00000000e2f5122000000000791585b2
+
+        /* x^54272 mod p(x)` << 1, x^54336 mod p(x)` << 1 */
+        .octa 0x000000004aa01f3e00000000ac53b894
+
+        /* x^53248 mod p(x)` << 1, x^53312 mod p(x)` << 1 */
+        .octa 0x00000000b3e90a5800000001ed5f2cf4
+
+        /* x^52224 mod p(x)` << 1, x^52288 mod p(x)` << 1 */
+        .octa 0x000000000c9ca2aa00000001df48b2e0
+
+        /* x^51200 mod p(x)` << 1, x^51264 mod p(x)` << 1 */
+        .octa 0x000000015168231600000000049c1c62
+
+        /* x^50176 mod p(x)` << 1, x^50240 mod p(x)` << 1 */
+        .octa 0x0000000036fce78c000000017c460c12
+
+        /* x^49152 mod p(x)` << 1, x^49216 mod p(x)` << 1 */
+        .octa 0x000000009037dc10000000015be4da7e
+
+        /* x^48128 mod p(x)` << 1, x^48192 mod p(x)` << 1 */
+        .octa 0x00000000d3298582000000010f38f668
+
+        /* x^47104 mod p(x)` << 1, x^47168 mod p(x)` << 1 */
+        .octa 0x00000001b42e8ad60000000039f40a00
+
+        /* x^46080 mod p(x)` << 1, x^46144 mod p(x)` << 1 */
+        .octa 0x00000000142a983800000000bd4c10c4
+
+        /* x^45056 mod p(x)` << 1, x^45120 mod p(x)` << 1 */
+        .octa 0x0000000109c7f1900000000042db1d98
+
+        /* x^44032 mod p(x)` << 1, x^44096 mod p(x)` << 1 */
+        .octa 0x0000000056ff931000000001c905bae6
+
+        /* x^43008 mod p(x)` << 1, x^43072 mod p(x)` << 1 */
+        .octa 0x00000001594513aa00000000069d40ea
+
+        /* x^41984 mod p(x)` << 1, x^42048 mod p(x)` << 1 */
+        .octa 0x00000001e3b5b1e8000000008e4fbad0
+
+        /* x^40960 mod p(x)` << 1, x^41024 mod p(x)` << 1 */
+        .octa 0x000000011dd5fc080000000047bedd46
+
+        /* x^39936 mod p(x)` << 1, x^40000 mod p(x)` << 1 */
+        .octa 0x00000001675f0cc20000000026396bf8
+
+        /* x^38912 mod p(x)` << 1, x^38976 mod p(x)` << 1 */
+        .octa 0x00000000d1c8dd4400000000379beb92
+
+        /* x^37888 mod p(x)` << 1, x^37952 mod p(x)` << 1 */
+        .octa 0x0000000115ebd3d8000000000abae54a
+
+        /* x^36864 mod p(x)` << 1, x^36928 mod p(x)` << 1 */
+        .octa 0x00000001ecbd0dac0000000007e6a128
+
+        /* x^35840 mod p(x)` << 1, x^35904 mod p(x)` << 1 */
+        .octa 0x00000000cdf67af2000000000ade29d2
+
+        /* x^34816 mod p(x)` << 1, x^34880 mod p(x)` << 1 */
+        .octa 0x000000004c01ff4c00000000f974c45c
+
+        /* x^33792 mod p(x)` << 1, x^33856 mod p(x)` << 1 */
+        .octa 0x00000000f2d8657e00000000e77ac60a
+
+        /* x^32768 mod p(x)` << 1, x^32832 mod p(x)` << 1 */
+        .octa 0x000000006bae74c40000000145895816
+
+        /* x^31744 mod p(x)` << 1, x^31808 mod p(x)` << 1 */
+        .octa 0x0000000152af8aa00000000038e362be
+
+        /* x^30720 mod p(x)` << 1, x^30784 mod p(x)` << 1 */
+        .octa 0x0000000004663802000000007f991a64
+
+        /* x^29696 mod p(x)` << 1, x^29760 mod p(x)` << 1 */
+        .octa 0x00000001ab2f5afc00000000fa366d3a
+
+        /* x^28672 mod p(x)` << 1, x^28736 mod p(x)` << 1 */
+        .octa 0x0000000074a4ebd400000001a2bb34f0
+
+        /* x^27648 mod p(x)` << 1, x^27712 mod p(x)` << 1 */
+        .octa 0x00000001d7ab3a4c0000000028a9981e
+
+        /* x^26624 mod p(x)` << 1, x^26688 mod p(x)` << 1 */
+        .octa 0x00000001a8da60c600000001dbc672be
+
+        /* x^25600 mod p(x)` << 1, x^25664 mod p(x)` << 1 */
+        .octa 0x000000013cf6382000000000b04d77f6
+
+        /* x^24576 mod p(x)` << 1, x^24640 mod p(x)` << 1 */
+        .octa 0x00000000bec12e1e0000000124400d96
+
+        /* x^23552 mod p(x)` << 1, x^23616 mod p(x)` << 1 */
+        .octa 0x00000001c6368010000000014ca4b414
+
+        /* x^22528 mod p(x)` << 1, x^22592 mod p(x)` << 1 */
+        .octa 0x00000001e6e78758000000012fe2c938
+
+        /* x^21504 mod p(x)` << 1, x^21568 mod p(x)` << 1 */
+        .octa 0x000000008d7f2b3c00000001faed01e6
+
+        /* x^20480 mod p(x)` << 1, x^20544 mod p(x)` << 1 */
+        .octa 0x000000016b4a156e000000007e80ecfe
+
+        /* x^19456 mod p(x)` << 1, x^19520 mod p(x)` << 1 */
+        .octa 0x00000001c63cfeb60000000098daee94
+
+        /* x^18432 mod p(x)` << 1, x^18496 mod p(x)` << 1 */
+        .octa 0x000000015f902670000000010a04edea
+
+        /* x^17408 mod p(x)` << 1, x^17472 mod p(x)` << 1 */
+        .octa 0x00000001cd5de11e00000001c00b4524
+
+        /* x^16384 mod p(x)` << 1, x^16448 mod p(x)` << 1 */
+        .octa 0x000000001acaec540000000170296550
+
+        /* x^15360 mod p(x)` << 1, x^15424 mod p(x)` << 1 */
+        .octa 0x000000002bd0ca780000000181afaa48
+
+        /* x^14336 mod p(x)` << 1, x^14400 mod p(x)` << 1 */
+        .octa 0x0000000032d63d5c0000000185a31ffa
+
+        /* x^13312 mod p(x)` << 1, x^13376 mod p(x)` << 1 */
+        .octa 0x000000001c6d4e4c000000002469f608
+
+        /* x^12288 mod p(x)` << 1, x^12352 mod p(x)` << 1 */
+        .octa 0x0000000106a60b92000000006980102a
+
+        /* x^11264 mod p(x)` << 1, x^11328 mod p(x)` << 1 */
+        .octa 0x00000000d3855e120000000111ea9ca8
+
+        /* x^10240 mod p(x)` << 1, x^10304 mod p(x)` << 1 */
+        .octa 0x00000000e312563600000001bd1d29ce
+
+        /* x^9216 mod p(x)` << 1, x^9280 mod p(x)` << 1 */
+        .octa 0x000000009e8f7ea400000001b34b9580
+
+        /* x^8192 mod p(x)` << 1, x^8256 mod p(x)` << 1 */
+        .octa 0x00000001c82e562c000000003076054e
+
+        /* x^7168 mod p(x)` << 1, x^7232 mod p(x)` << 1 */
+        .octa 0x00000000ca9f09ce000000012a608ea4
+
+        /* x^6144 mod p(x)` << 1, x^6208 mod p(x)` << 1 */
+        .octa 0x00000000c63764e600000000784d05fe
+
+        /* x^5120 mod p(x)` << 1, x^5184 mod p(x)` << 1 */
+        .octa 0x0000000168d2e49e000000016ef0d82a
+
+        /* x^4096 mod p(x)` << 1, x^4160 mod p(x)` << 1 */
+        .octa 0x00000000e986c1480000000075bda454
+
+        /* x^3072 mod p(x)` << 1, x^3136 mod p(x)` << 1 */
+        .octa 0x00000000cfb65894000000003dc0a1c4
+
+        /* x^2048 mod p(x)` << 1, x^2112 mod p(x)` << 1 */
+        .octa 0x0000000111cadee400000000e9a5d8be
+
+        /* x^1024 mod p(x)` << 1, x^1088 mod p(x)` << 1 */
+        .octa 0x0000000171fb63ce00000001609bc4b4
+
+        .short_constants :
+
+        /* Reduce final 1024-2048 bits to 64 bits, shifting 32 bits to include
+           the trailing 32 bits of zeros */
+        /* x^1952 mod p(x)`, x^1984 mod p(x)`, x^2016 mod p(x)`, x^2048 mod
+           p(x)` */
+        .octa 0x7fec2963e5bf80485cf015c388e56f72
+
+        /* x^1824 mod p(x)`, x^1856 mod p(x)`, x^1888 mod p(x)`, x^1920 mod
+           p(x)` */
+        .octa 0x38e888d4844752a9963a18920246e2e6
+
+        /* x^1696 mod p(x)`, x^1728 mod p(x)`, x^1760 mod p(x)`, x^1792 mod
+           p(x)` */
+        .octa 0x42316c00730206ad419a441956993a31
+
+        /* x^1568 mod p(x)`, x^1600 mod p(x)`, x^1632 mod p(x)`, x^1664 mod
+           p(x)` */
+        .octa 0x543d5c543e65ddf9924752ba2b830011
+
+        /* x^1440 mod p(x)`, x^1472 mod p(x)`, x^1504 mod p(x)`, x^1536 mod
+           p(x)` */
+        .octa 0x78e87aaf56767c9255bd7f9518e4a304
+
+        /* x^1312 mod p(x)`, x^1344 mod p(x)`, x^1376 mod p(x)`, x^1408 mod
+           p(x)` */
+        .octa 0x8f68fcec1903da7f6d76739fe0553f1e
+
+        /* x^1184 mod p(x)`, x^1216 mod p(x)`, x^1248 mod p(x)`, x^1280 mod
+           p(x)` */
+        .octa 0x3f4840246791d588c133722b1fe0b5c3
+
+        /* x^1056 mod p(x)`, x^1088 mod p(x)`, x^1120 mod p(x)`, x^1152 mod
+           p(x)` */
+        .octa 0x34c96751b04de25a64b67ee0e55ef1f3
+
+        /* x^928 mod p(x)`, x^960 mod p(x)`, x^992 mod p(x)`, x^1024 mod p(x)`
+         */
+        .octa 0x156c8e180b4a395b069db049b8fdb1e7
+
+        /* x^800 mod p(x)`, x^832 mod p(x)`, x^864 mod p(x)`, x^896 mod p(x)` */
+        .octa 0xe0b99ccbe661f7bea11bfaf3c9e90b9e
+
+        /* x^672 mod p(x)`, x^704 mod p(x)`, x^736 mod p(x)`, x^768 mod p(x)` */
+        .octa 0x041d37768cd75659817cdc5119b29a35
+
+        /* x^544 mod p(x)`, x^576 mod p(x)`, x^608 mod p(x)`, x^640 mod p(x)` */
+        .octa 0x3a0777818cfaa9651ce9d94b36c41f1c
+
+        /* x^416 mod p(x)`, x^448 mod p(x)`, x^480 mod p(x)`, x^512 mod p(x)` */
+        .octa 0x0e148e8252377a554f256efcb82be955
+
+        /* x^288 mod p(x)`, x^320 mod p(x)`, x^352 mod p(x)`, x^384 mod p(x)` */
+        .octa 0x9c25531d19e65ddeec1631edb2dea967
+
+        /* x^160 mod p(x)`, x^192 mod p(x)`, x^224 mod p(x)`, x^256 mod p(x)` */
+        .octa 0x790606ff9957c0a65d27e147510ac59a
+
+        /* x^32 mod p(x)`, x^64 mod p(x)`, x^96 mod p(x)`, x^128 mod p(x)` */
+        .octa 0x82f63b786ea2d55ca66805eb18b8ea18
+
+        .barrett_constants :
+        /* 33 bit reflected Barrett constant m - (4^32)/n */
+        .octa 0x000000000000000000000000dea713f1 /* x^64 div p(x)` */
+        /* 33 bit reflected Barrett constant n */
+        .octa 0x00000000000000000000000105ec76f1
+#endif
diff --git a/src/rocksdb/util/crc32c_test.cc b/src/rocksdb/util/crc32c_test.cc
new file mode 100644
index 000000000..715d63e2d
--- /dev/null
+++ b/src/rocksdb/util/crc32c_test.cc
@@ -0,0 +1,213 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+//
+// Copyright (c) 2011 The LevelDB Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file. See the AUTHORS file for names of contributors.
+#include "util/crc32c.h"
+
+#include "test_util/testharness.h"
+#include "util/coding.h"
+#include "util/random.h"
+
+namespace ROCKSDB_NAMESPACE {
+namespace crc32c {
+
+class CRC {};
+
+// Tests for 3-way crc32c algorithm. We need these tests because it uses
+// different lookup tables than the original Fast_CRC32
+const unsigned int BUFFER_SIZE = 512 * 1024 * sizeof(uint64_t);
+char buffer[BUFFER_SIZE];
+
+struct ExpectedResult {
+  size_t offset;
+  size_t length;
+  uint32_t crc32c;
+};
+
+ExpectedResult expectedResults[] = {
+    // Zero-byte input
+    {0, 0, ~0U},
+    // Small aligned inputs to test special cases in SIMD implementations
+    {8, 1, 1543413366},
+    {8, 2, 523493126},
+    {8, 3, 1560427360},
+    {8, 4, 3422504776},
+    {8, 5, 447841138},
+    {8, 6, 3910050499},
+    {8, 7, 3346241981},
+    // Small unaligned inputs
+    {9, 1, 3855826643},
+    {10, 2, 560880875},
+    {11, 3, 1479707779},
+    {12, 4, 2237687071},
+    {13, 5, 4063855784},
+    {14, 6, 2553454047},
+    {15, 7, 1349220140},
+    // Larger inputs to test leftover chunks at the end of aligned blocks
+    {8, 8, 627613930},
+    {8, 9, 2105929409},
+    {8, 10, 2447068514},
+    {8, 11, 863807079},
+    {8, 12, 292050879},
+    {8, 13, 1411837737},
+    {8, 14, 2614515001},
+    {8, 15, 3579076296},
+    {8, 16, 2897079161},
+    {8, 17, 675168386},
+    // // Much larger inputs
+    {0, BUFFER_SIZE, 2096790750},
+    {1, BUFFER_SIZE / 2, 3854797577},
+
+};
+
+TEST(CRC, StandardResults) {
+  // Original Fast_CRC32 tests.
+  // From rfc3720 section B.4.
+  char buf[32];
+
+  memset(buf, 0, sizeof(buf));
+  ASSERT_EQ(0x8a9136aaU, Value(buf, sizeof(buf)));
+
+  memset(buf, 0xff, sizeof(buf));
+  ASSERT_EQ(0x62a8ab43U, Value(buf, sizeof(buf)));
+
+  for (int i = 0; i < 32; i++) {
+    buf[i] = static_cast<char>(i);
+  }
+  ASSERT_EQ(0x46dd794eU, Value(buf, sizeof(buf)));
+
+  for (int i = 0; i < 32; i++) {
+    buf[i] = static_cast<char>(31 - i);
+  }
+  ASSERT_EQ(0x113fdb5cU, Value(buf, sizeof(buf)));
+
+  unsigned char data[48] = {
+      0x01, 0xc0, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+      0x00, 0x00, 0x00, 0x00, 0x14, 0x00, 0x00, 0x00, 0x00, 0x00, 0x04, 0x00,
+      0x00, 0x00, 0x00, 0x14, 0x00, 0x00, 0x00, 0x18, 0x28, 0x00, 0x00, 0x00,
+      0x00, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+  };
+  ASSERT_EQ(0xd9963a56, Value(reinterpret_cast<char*>(data), sizeof(data)));
+
+  // 3-Way Crc32c tests ported from folly.
+  // Test 1: single computation
+  for (auto expected : expectedResults) {
+    uint32_t result = Value(buffer + expected.offset, expected.length);
+    EXPECT_EQ(~expected.crc32c, result);
+  }
+
+  // Test 2: stitching two computations
+  for (auto expected : expectedResults) {
+    size_t partialLength = expected.length / 2;
+    uint32_t partialChecksum = Value(buffer + expected.offset, partialLength);
+    uint32_t result =
+        Extend(partialChecksum, buffer + expected.offset + partialLength,
+               expected.length - partialLength);
+    EXPECT_EQ(~expected.crc32c, result);
+  }
+}
+
+TEST(CRC, Values) { ASSERT_NE(Value("a", 1), Value("foo", 3)); }
+
+TEST(CRC, Extend) {
+  ASSERT_EQ(Value("hello world", 11), Extend(Value("hello ", 6), "world", 5));
+}
+
+TEST(CRC, Mask) {
+  uint32_t crc = Value("foo", 3);
+  ASSERT_NE(crc, Mask(crc));
+  ASSERT_NE(crc, Mask(Mask(crc)));
+  ASSERT_EQ(crc, Unmask(Mask(crc)));
+  ASSERT_EQ(crc, Unmask(Unmask(Mask(Mask(crc)))));
+}
+
+TEST(CRC, Crc32cCombineBasicTest) {
+  uint32_t crc1 = Value("hello ", 6);
+  uint32_t crc2 = Value("world", 5);
+  uint32_t crc3 = Value("hello world", 11);
+  uint32_t crc1_2_combine = Crc32cCombine(crc1, crc2, 5);
+  ASSERT_EQ(crc3, crc1_2_combine);
+}
+
+TEST(CRC, Crc32cCombineOrderMattersTest) {
+  uint32_t crc1 = Value("hello ", 6);
+  uint32_t crc2 = Value("world", 5);
+  uint32_t crc3 = Value("hello world", 11);
+  uint32_t crc2_1_combine = Crc32cCombine(crc2, crc1, 6);
+  ASSERT_NE(crc3, crc2_1_combine);
+}
+
+TEST(CRC, Crc32cCombineFullCoverTest) {
+  int scale = 4 * 1024;
+  Random rnd(test::RandomSeed());
+  int size_1 = 1024 * 1024;
+  std::string s1 = rnd.RandomBinaryString(size_1);
+  uint32_t crc1 = Value(s1.data(), size_1);
+  for (int i = 0; i < scale; i++) {
+    int size_2 = i;
+    std::string s2 = rnd.RandomBinaryString(size_2);
+    uint32_t crc2 = Value(s2.data(), s2.size());
+    uint32_t crc1_2 = Extend(crc1, s2.data(), s2.size());
+    uint32_t crc1_2_combine = Crc32cCombine(crc1, crc2, size_2);
+    ASSERT_EQ(crc1_2, crc1_2_combine);
+  }
+}
+
+TEST(CRC, Crc32cCombineBigSizeTest) {
+  Random rnd(test::RandomSeed());
+  int size_1 = 1024 * 1024;
+  std::string s1 = rnd.RandomBinaryString(size_1);
+  uint32_t crc1 = Value(s1.data(), size_1);
+  int size_2 = 16 * 1024 * 1024 - 1;
+  std::string s2 = rnd.RandomBinaryString(size_2);
+  uint32_t crc2 = Value(s2.data(), s2.size());
+  uint32_t crc1_2 = Extend(crc1, s2.data(), s2.size());
+  uint32_t crc1_2_combine = Crc32cCombine(crc1, crc2, size_2);
+  ASSERT_EQ(crc1_2, crc1_2_combine);
+}
+
+}  // namespace crc32c
+}  // namespace ROCKSDB_NAMESPACE
+
+// copied from folly
+const uint64_t FNV_64_HASH_START = 14695981039346656037ULL;
+inline uint64_t fnv64_buf(const void* buf, size_t n,
+                          uint64_t hash = FNV_64_HASH_START) {
+  // forcing signed char, since other platforms can use unsigned
+  const signed char* char_buf = reinterpret_cast<const signed char*>(buf);
+
+  for (size_t i = 0; i < n; ++i) {
+    hash += (hash << 1) + (hash << 4) + (hash << 5) + (hash << 7) +
+            (hash << 8) + (hash << 40);
+    hash ^= char_buf[i];
+  }
+  return hash;
+}
+
+int main(int argc, char** argv) {
+  ROCKSDB_NAMESPACE::port::InstallStackTraceHandler();
+  ::testing::InitGoogleTest(&argc, argv);
+
+  // Populate a buffer with a deterministic pattern
+  // on which to compute checksums
+
+  const uint8_t* src = (uint8_t*)ROCKSDB_NAMESPACE::crc32c::buffer;
+  uint64_t* dst = (uint64_t*)ROCKSDB_NAMESPACE::crc32c::buffer;
+  const uint64_t* end =
+      (const uint64_t*)(ROCKSDB_NAMESPACE::crc32c::buffer +
+                        ROCKSDB_NAMESPACE::crc32c::BUFFER_SIZE);
+  *dst++ = 0;
+  while (dst < end) {
+    ROCKSDB_NAMESPACE::EncodeFixed64(
+        reinterpret_cast<char*>(dst),
+        fnv64_buf((const char*)src, sizeof(uint64_t)));
+    dst++;
+    src += sizeof(uint64_t);
+  }
+
+  return RUN_ALL_TESTS();
+}
diff --git a/src/rocksdb/util/defer.h b/src/rocksdb/util/defer.h
new file mode 100644
index 000000000..f71e67ba9
--- /dev/null
+++ b/src/rocksdb/util/defer.h
@@ -0,0 +1,82 @@
+// Copyright (c) 2011-present, Facebook, Inc. All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#pragma once
+
+#include <functional>
+
+#include "rocksdb/rocksdb_namespace.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+// Defers the execution of the provided function until the Defer
+// object goes out of scope.
+//
+// Usage example:
+//
+// Status DeferTest() {
+//   Status s;
+//   Defer defer([&s]() {
+//     if (!s.ok()) {
+//       // do cleanups ...
+//     }
+//   });
+//   // do something ...
+//   if (!s.ok()) return;
+//   // do some other things ...
+//   return s;
+// }
+//
+// The above code ensures that cleanups will always happen on returning.
+//
+// Without the help of Defer, you can
+// 1. every time when !s.ok(), do the cleanup;
+// 2. instead of returning when !s.ok(), continue the work only when s.ok(),
+//    but sometimes, this might lead to nested blocks of "if (s.ok()) {...}".
+//
+// With the help of Defer, you can centralize the cleanup logic inside the
+// lambda passed to Defer, and you can return immediately on failure when
+// necessary.
+class Defer final {
+ public:
+  explicit Defer(std::function<void()>&& fn) : fn_(std::move(fn)) {}
+  ~Defer() { fn_(); }
+
+  // Disallow copy.
+  Defer(const Defer&) = delete;
+  Defer& operator=(const Defer&) = delete;
+
+ private:
+  std::function<void()> fn_;
+};
+
+// An RAII utility object that saves the current value of an object so that
+// it can be overwritten, and restores it to the saved value when the
+// SaveAndRestore object goes out of scope.
+template <typename T>
+class SaveAndRestore {
+ public:
+  // obj is non-null pointer to value to be saved and later restored.
+  explicit SaveAndRestore(T* obj) : obj_(obj), saved_(*obj) {}
+  // new_value is stored in *obj
+  SaveAndRestore(T* obj, const T& new_value)
+      : obj_(obj), saved_(std::move(*obj)) {
+    *obj = new_value;
+  }
+  SaveAndRestore(T* obj, T&& new_value) : obj_(obj), saved_(std::move(*obj)) {
+    *obj = std::move(new_value);
+  }
+  ~SaveAndRestore() { *obj_ = std::move(saved_); }
+
+  // No copies
+  SaveAndRestore(const SaveAndRestore&) = delete;
+  SaveAndRestore& operator=(const SaveAndRestore&) = delete;
+
+ private:
+  T* const obj_;
+  T saved_;
+};
+
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/util/defer_test.cc b/src/rocksdb/util/defer_test.cc
new file mode 100644
index 000000000..0e98f68b6
--- /dev/null
+++ b/src/rocksdb/util/defer_test.cc
@@ -0,0 +1,51 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#include "util/defer.h"
+
+#include "port/port.h"
+#include "port/stack_trace.h"
+#include "test_util/testharness.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+class DeferTest {};
+
+TEST(DeferTest, BlockScope) {
+  int v = 1;
+  {
+    Defer defer([&v]() { v *= 2; });
+  }
+  ASSERT_EQ(2, v);
+}
+
+TEST(DeferTest, FunctionScope) {
+  int v = 1;
+  auto f = [&v]() {
+    Defer defer([&v]() { v *= 2; });
+    v = 2;
+  };
+  f();
+  ASSERT_EQ(4, v);
+}
+
+TEST(SaveAndRestoreTest, BlockScope) {
+  int v = 1;
+  {
+    SaveAndRestore<int> sr(&v);
+    ASSERT_EQ(v, 1);
+    v = 2;
+    ASSERT_EQ(v, 2);
+  }
+  ASSERT_EQ(v, 1);
+}
+
+}  // namespace ROCKSDB_NAMESPACE
+
+int main(int argc, char** argv) {
+  ROCKSDB_NAMESPACE::port::InstallStackTraceHandler();
+  ::testing::InitGoogleTest(&argc, argv);
+  return RUN_ALL_TESTS();
+}
diff --git a/src/rocksdb/util/distributed_mutex.h b/src/rocksdb/util/distributed_mutex.h
new file mode 100644
index 000000000..9675a1e2d
--- /dev/null
+++ b/src/rocksdb/util/distributed_mutex.h
@@ -0,0 +1,48 @@
+//  Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#pragma once
+
+#include "rocksdb/rocksdb_namespace.h"
+
+// This file declares a wrapper around the efficient folly DistributedMutex
+// that falls back on a standard mutex when not available. See
+// https://github.com/facebook/folly/blob/main/folly/synchronization/DistributedMutex.h
+// for benefits and limitations.
+
+// At the moment, only scoped locking is supported using DMutexLock
+// RAII wrapper, because lock/unlock APIs will vary.
+
+#ifdef USE_FOLLY
+
+#include <folly/synchronization/DistributedMutex.h>
+
+namespace ROCKSDB_NAMESPACE {
+
+class DMutex : public folly::DistributedMutex {
+ public:
+  static const char* kName() { return "folly::DistributedMutex"; }
+
+  explicit DMutex(bool IGNORED_adaptive = false) { (void)IGNORED_adaptive; }
+
+  // currently no-op
+  void AssertHeld() {}
+};
+using DMutexLock = std::lock_guard<folly::DistributedMutex>;
+
+}  // namespace ROCKSDB_NAMESPACE
+
+#else
+
+#include "port/port.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+using DMutex = port::Mutex;
+using DMutexLock = std::lock_guard<DMutex>;
+
+}  // namespace ROCKSDB_NAMESPACE
+
+#endif
diff --git a/src/rocksdb/util/duplicate_detector.h b/src/rocksdb/util/duplicate_detector.h
new file mode 100644
index 000000000..d778622db
--- /dev/null
+++ b/src/rocksdb/util/duplicate_detector.h
@@ -0,0 +1,71 @@
+// Copyright (c) 2011-present, Facebook, Inc. All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#pragma once
+
+#include <cstdint>
+
+#include "db/db_impl/db_impl.h"
+#include "logging/logging.h"
+#include "util/set_comparator.h"
+
+namespace ROCKSDB_NAMESPACE {
+// During recovery if the memtable is flushed we cannot rely on its help on
+// duplicate key detection and as key insert will not be attempted. This class
+// will be used as a emulator of memtable to tell if insertion of a key/seq
+// would have resulted in duplication.
+class DuplicateDetector {
+ public:
+  explicit DuplicateDetector(DBImpl* db) : db_(db) {}
+  bool IsDuplicateKeySeq(uint32_t cf, const Slice& key, SequenceNumber seq) {
+    assert(seq >= batch_seq_);
+    if (batch_seq_ != seq) {  // it is a new batch
+      keys_.clear();
+    }
+    batch_seq_ = seq;
+    CFKeys& cf_keys = keys_[cf];
+    if (cf_keys.size() == 0) {  // just inserted
+      InitWithComp(cf);
+    }
+    auto it = cf_keys.insert(key);
+    if (it.second == false) {  // second is false if a element already existed.
+      keys_.clear();
+      InitWithComp(cf);
+      keys_[cf].insert(key);
+      return true;
+    }
+    return false;
+  }
+
+ private:
+  SequenceNumber batch_seq_ = 0;
+  DBImpl* db_;
+  using CFKeys = std::set<Slice, SetComparator>;
+  std::map<uint32_t, CFKeys> keys_;
+  void InitWithComp(const uint32_t cf) {
+    auto h = db_->GetColumnFamilyHandle(cf);
+    if (!h) {
+      // TODO(myabandeh): This is not a concern in MyRocks as drop cf is not
+      // implemented yet. When it does, we should return proper error instead
+      // of throwing exception.
+      ROCKS_LOG_FATAL(
+          db_->immutable_db_options().info_log,
+          "Recovering an entry from the dropped column family %" PRIu32
+          ". WAL must must have been emptied before dropping the column "
+          "family",
+          cf);
+#ifndef ROCKSDB_LITE
+      throw std::runtime_error(
+          "Recovering an entry from a dropped column family. "
+          "WAL must must have been flushed before dropping the column "
+          "family");
+#endif
+      return;
+    }
+    auto cmp = h->GetComparator();
+    keys_[cf] = CFKeys(SetComparator(cmp));
+  }
+};
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/util/dynamic_bloom.cc b/src/rocksdb/util/dynamic_bloom.cc
new file mode 100644
index 000000000..0ff3b4a75
--- /dev/null
+++ b/src/rocksdb/util/dynamic_bloom.cc
@@ -0,0 +1,70 @@
+// Copyright (c) 2011-present, Facebook, Inc. All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#include "dynamic_bloom.h"
+
+#include <algorithm>
+
+#include "memory/allocator.h"
+#include "port/port.h"
+#include "rocksdb/slice.h"
+#include "util/hash.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+namespace {
+
+uint32_t roundUpToPow2(uint32_t x) {
+  uint32_t rv = 1;
+  while (rv < x) {
+    rv <<= 1;
+  }
+  return rv;
+}
+}  // namespace
+
+DynamicBloom::DynamicBloom(Allocator* allocator, uint32_t total_bits,
+                           uint32_t num_probes, size_t huge_page_tlb_size,
+                           Logger* logger)
+    // Round down, except round up with 1
+    : kNumDoubleProbes((num_probes + (num_probes == 1)) / 2) {
+  assert(num_probes % 2 == 0);  // limitation of current implementation
+  assert(num_probes <= 10);     // limitation of current implementation
+  assert(kNumDoubleProbes > 0);
+
+  // Determine how much to round off + align by so that x ^ i (that's xor) is
+  // a valid u64 index if x is a valid u64 index and 0 <= i < kNumDoubleProbes.
+  uint32_t block_bytes = /*bytes/u64*/ 8 *
+                         /*u64s*/ std::max(1U, roundUpToPow2(kNumDoubleProbes));
+  uint32_t block_bits = block_bytes * 8;
+  uint32_t blocks = (total_bits + block_bits - 1) / block_bits;
+  uint32_t sz = blocks * block_bytes;
+  kLen = sz / /*bytes/u64*/ 8;
+  assert(kLen > 0);
+#ifndef NDEBUG
+  for (uint32_t i = 0; i < kNumDoubleProbes; ++i) {
+    // Ensure probes starting at last word are in range
+    assert(((kLen - 1) ^ i) < kLen);
+  }
+#endif
+
+  // Padding to correct for allocation not originally aligned on block_bytes
+  // boundary
+  sz += block_bytes - 1;
+  assert(allocator);
+
+  char* raw = allocator->AllocateAligned(sz, huge_page_tlb_size, logger);
+  memset(raw, 0, sz);
+  auto block_offset = reinterpret_cast<uintptr_t>(raw) % block_bytes;
+  if (block_offset > 0) {
+    // Align on block_bytes boundary
+    raw += block_bytes - block_offset;
+  }
+  static_assert(sizeof(std::atomic<uint64_t>) == sizeof(uint64_t),
+                "Expecting zero-space-overhead atomic");
+  data_ = reinterpret_cast<std::atomic<uint64_t>*>(raw);
+}
+
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/util/dynamic_bloom.h b/src/rocksdb/util/dynamic_bloom.h
new file mode 100644
index 000000000..40cd29404
--- /dev/null
+++ b/src/rocksdb/util/dynamic_bloom.h
@@ -0,0 +1,214 @@
+// Copyright (c) 2011-present, Facebook, Inc. All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#pragma once
+
+#include <array>
+#include <atomic>
+#include <memory>
+#include <string>
+
+#include "port/port.h"
+#include "rocksdb/slice.h"
+#include "table/multiget_context.h"
+#include "util/hash.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+class Slice;
+class Allocator;
+class Logger;
+
+// A Bloom filter intended only to be used in memory, never serialized in a way
+// that could lead to schema incompatibility. Supports opt-in lock-free
+// concurrent access.
+//
+// This implementation is also intended for applications generally preferring
+// speed vs. maximum accuracy: roughly 0.9x BF op latency for 1.1x FP rate.
+// For 1% FP rate, that means that the latency of a look-up triggered by an FP
+// should be less than roughly 100x the cost of a Bloom filter op.
+//
+// For simplicity and performance, the current implementation requires
+// num_probes to be a multiple of two and <= 10.
+//
+class DynamicBloom {
+ public:
+  // allocator: pass allocator to bloom filter, hence trace the usage of memory
+  // total_bits: fixed total bits for the bloom
+  // num_probes: number of hash probes for a single key
+  // hash_func:  customized hash function
+  // huge_page_tlb_size:  if >0, try to allocate bloom bytes from huge page TLB
+  //                      within this page size. Need to reserve huge pages for
+  //                      it to be allocated, like:
+  //                         sysctl -w vm.nr_hugepages=20
+  //                     See linux doc Documentation/vm/hugetlbpage.txt
+  explicit DynamicBloom(Allocator* allocator, uint32_t total_bits,
+                        uint32_t num_probes = 6, size_t huge_page_tlb_size = 0,
+                        Logger* logger = nullptr);
+
+  ~DynamicBloom() {}
+
+  // Assuming single threaded access to this function.
+  void Add(const Slice& key);
+
+  // Like Add, but may be called concurrent with other functions.
+  void AddConcurrently(const Slice& key);
+
+  // Assuming single threaded access to this function.
+  void AddHash(uint32_t hash);
+
+  // Like AddHash, but may be called concurrent with other functions.
+  void AddHashConcurrently(uint32_t hash);
+
+  // Multithreaded access to this function is OK
+  bool MayContain(const Slice& key) const;
+
+  void MayContain(int num_keys, Slice* keys, bool* may_match) const;
+
+  // Multithreaded access to this function is OK
+  bool MayContainHash(uint32_t hash) const;
+
+  void Prefetch(uint32_t h);
+
+ private:
+  // Length of the structure, in 64-bit words. For this structure, "word"
+  // will always refer to 64-bit words.
+  uint32_t kLen;
+  // We make the k probes in pairs, two for each 64-bit read/write. Thus,
+  // this stores k/2, the number of words to double-probe.
+  const uint32_t kNumDoubleProbes;
+
+  std::atomic<uint64_t>* data_;
+
+  // or_func(ptr, mask) should effect *ptr |= mask with the appropriate
+  // concurrency safety, working with bytes.
+  template <typename OrFunc>
+  void AddHash(uint32_t hash, const OrFunc& or_func);
+
+  bool DoubleProbe(uint32_t h32, size_t a) const;
+};
+
+inline void DynamicBloom::Add(const Slice& key) { AddHash(BloomHash(key)); }
+
+inline void DynamicBloom::AddConcurrently(const Slice& key) {
+  AddHashConcurrently(BloomHash(key));
+}
+
+inline void DynamicBloom::AddHash(uint32_t hash) {
+  AddHash(hash, [](std::atomic<uint64_t>* ptr, uint64_t mask) {
+    ptr->store(ptr->load(std::memory_order_relaxed) | mask,
+               std::memory_order_relaxed);
+  });
+}
+
+inline void DynamicBloom::AddHashConcurrently(uint32_t hash) {
+  AddHash(hash, [](std::atomic<uint64_t>* ptr, uint64_t mask) {
+    // Happens-before between AddHash and MaybeContains is handled by
+    // access to versions_->LastSequence(), so all we have to do here is
+    // avoid races (so we don't give the compiler a license to mess up
+    // our code) and not lose bits.  std::memory_order_relaxed is enough
+    // for that.
+    if ((mask & ptr->load(std::memory_order_relaxed)) != mask) {
+      ptr->fetch_or(mask, std::memory_order_relaxed);
+    }
+  });
+}
+
+inline bool DynamicBloom::MayContain(const Slice& key) const {
+  return (MayContainHash(BloomHash(key)));
+}
+
+inline void DynamicBloom::MayContain(int num_keys, Slice* keys,
+                                     bool* may_match) const {
+  std::array<uint32_t, MultiGetContext::MAX_BATCH_SIZE> hashes;
+  std::array<size_t, MultiGetContext::MAX_BATCH_SIZE> byte_offsets;
+  for (int i = 0; i < num_keys; ++i) {
+    hashes[i] = BloomHash(keys[i]);
+    size_t a = FastRange32(kLen, hashes[i]);
+    PREFETCH(data_ + a, 0, 3);
+    byte_offsets[i] = a;
+  }
+
+  for (int i = 0; i < num_keys; i++) {
+    may_match[i] = DoubleProbe(hashes[i], byte_offsets[i]);
+  }
+}
+
+#if defined(_MSC_VER)
+#pragma warning(push)
+// local variable is initialized but not referenced
+#pragma warning(disable : 4189)
+#endif
+inline void DynamicBloom::Prefetch(uint32_t h32) {
+  size_t a = FastRange32(kLen, h32);
+  PREFETCH(data_ + a, 0, 3);
+}
+#if defined(_MSC_VER)
+#pragma warning(pop)
+#endif
+
+// Speed hacks in this implementation:
+// * Uses fastrange instead of %
+// * Minimum logic to determine first (and all) probed memory addresses.
+//   (Uses constant bit-xor offsets from the starting probe address.)
+// * (Major) Two probes per 64-bit memory fetch/write.
+//   Code simplification / optimization: only allow even number of probes.
+// * Very fast and effective (murmur-like) hash expansion/re-mixing. (At
+// least on recent CPUs, integer multiplication is very cheap. Each 64-bit
+// remix provides five pairs of bit addresses within a uint64_t.)
+//   Code simplification / optimization: only allow up to 10 probes, from a
+//   single 64-bit remix.
+//
+// The FP rate penalty for this implementation, vs. standard Bloom filter, is
+// roughly 1.12x on top of the 1.15x penalty for a 512-bit cache-local Bloom.
+// This implementation does not explicitly use the cache line size, but is
+// effectively cache-local (up to 16 probes) because of the bit-xor offsetting.
+//
+// NB: could easily be upgraded to support a 64-bit hash and
+// total_bits > 2^32 (512MB). (The latter is a bad idea without the former,
+// because of false positives.)
+
+inline bool DynamicBloom::MayContainHash(uint32_t h32) const {
+  size_t a = FastRange32(kLen, h32);
+  PREFETCH(data_ + a, 0, 3);
+  return DoubleProbe(h32, a);
+}
+
+inline bool DynamicBloom::DoubleProbe(uint32_t h32, size_t byte_offset) const {
+  // Expand/remix with 64-bit golden ratio
+  uint64_t h = 0x9e3779b97f4a7c13ULL * h32;
+  for (unsigned i = 0;; ++i) {
+    // Two bit probes per uint64_t probe
+    uint64_t mask =
+        ((uint64_t)1 << (h & 63)) | ((uint64_t)1 << ((h >> 6) & 63));
+    uint64_t val = data_[byte_offset ^ i].load(std::memory_order_relaxed);
+    if (i + 1 >= kNumDoubleProbes) {
+      return (val & mask) == mask;
+    } else if ((val & mask) != mask) {
+      return false;
+    }
+    h = (h >> 12) | (h << 52);
+  }
+}
+
+template <typename OrFunc>
+inline void DynamicBloom::AddHash(uint32_t h32, const OrFunc& or_func) {
+  size_t a = FastRange32(kLen, h32);
+  PREFETCH(data_ + a, 0, 3);
+  // Expand/remix with 64-bit golden ratio
+  uint64_t h = 0x9e3779b97f4a7c13ULL * h32;
+  for (unsigned i = 0;; ++i) {
+    // Two bit probes per uint64_t probe
+    uint64_t mask =
+        ((uint64_t)1 << (h & 63)) | ((uint64_t)1 << ((h >> 6) & 63));
+    or_func(&data_[a ^ i], mask);
+    if (i + 1 >= kNumDoubleProbes) {
+      return;
+    }
+    h = (h >> 12) | (h << 52);
+  }
+}
+
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/util/dynamic_bloom_test.cc b/src/rocksdb/util/dynamic_bloom_test.cc
new file mode 100644
index 000000000..925c5479a
--- /dev/null
+++ b/src/rocksdb/util/dynamic_bloom_test.cc
@@ -0,0 +1,325 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#ifndef GFLAGS
+#include <cstdio>
+int main() {
+  fprintf(stderr, "Please install gflags to run this test... Skipping...\n");
+  return 0;
+}
+#else
+
+#include <algorithm>
+#include <atomic>
+#include <cinttypes>
+#include <functional>
+#include <memory>
+#include <thread>
+#include <vector>
+
+#include "dynamic_bloom.h"
+#include "memory/arena.h"
+#include "port/port.h"
+#include "rocksdb/system_clock.h"
+#include "test_util/testharness.h"
+#include "test_util/testutil.h"
+#include "util/gflags_compat.h"
+#include "util/stop_watch.h"
+
+using GFLAGS_NAMESPACE::ParseCommandLineFlags;
+
+DEFINE_int32(bits_per_key, 10, "");
+DEFINE_int32(num_probes, 6, "");
+DEFINE_bool(enable_perf, false, "");
+
+namespace ROCKSDB_NAMESPACE {
+
+struct KeyMaker {
+  uint64_t a;
+  uint64_t b;
+
+  // Sequential, within a hash function block
+  inline Slice Seq(uint64_t i) {
+    a = i;
+    return Slice(reinterpret_cast<char *>(&a), sizeof(a));
+  }
+  // Not quite sequential, varies across hash function blocks
+  inline Slice Nonseq(uint64_t i) {
+    a = i;
+    b = i * 123;
+    return Slice(reinterpret_cast<char *>(this), sizeof(*this));
+  }
+  inline Slice Key(uint64_t i, bool nonseq) {
+    return nonseq ? Nonseq(i) : Seq(i);
+  }
+};
+
+class DynamicBloomTest : public testing::Test {};
+
+TEST_F(DynamicBloomTest, EmptyFilter) {
+  Arena arena;
+  DynamicBloom bloom1(&arena, 100, 2);
+  ASSERT_TRUE(!bloom1.MayContain("hello"));
+  ASSERT_TRUE(!bloom1.MayContain("world"));
+
+  DynamicBloom bloom2(&arena, CACHE_LINE_SIZE * 8 * 2 - 1, 2);
+  ASSERT_TRUE(!bloom2.MayContain("hello"));
+  ASSERT_TRUE(!bloom2.MayContain("world"));
+}
+
+TEST_F(DynamicBloomTest, Small) {
+  Arena arena;
+  DynamicBloom bloom1(&arena, 100, 2);
+  bloom1.Add("hello");
+  bloom1.Add("world");
+  ASSERT_TRUE(bloom1.MayContain("hello"));
+  ASSERT_TRUE(bloom1.MayContain("world"));
+  ASSERT_TRUE(!bloom1.MayContain("x"));
+  ASSERT_TRUE(!bloom1.MayContain("foo"));
+
+  DynamicBloom bloom2(&arena, CACHE_LINE_SIZE * 8 * 2 - 1, 2);
+  bloom2.Add("hello");
+  bloom2.Add("world");
+  ASSERT_TRUE(bloom2.MayContain("hello"));
+  ASSERT_TRUE(bloom2.MayContain("world"));
+  ASSERT_TRUE(!bloom2.MayContain("x"));
+  ASSERT_TRUE(!bloom2.MayContain("foo"));
+}
+
+TEST_F(DynamicBloomTest, SmallConcurrentAdd) {
+  Arena arena;
+  DynamicBloom bloom1(&arena, 100, 2);
+  bloom1.AddConcurrently("hello");
+  bloom1.AddConcurrently("world");
+  ASSERT_TRUE(bloom1.MayContain("hello"));
+  ASSERT_TRUE(bloom1.MayContain("world"));
+  ASSERT_TRUE(!bloom1.MayContain("x"));
+  ASSERT_TRUE(!bloom1.MayContain("foo"));
+
+  DynamicBloom bloom2(&arena, CACHE_LINE_SIZE * 8 * 2 - 1, 2);
+  bloom2.AddConcurrently("hello");
+  bloom2.AddConcurrently("world");
+  ASSERT_TRUE(bloom2.MayContain("hello"));
+  ASSERT_TRUE(bloom2.MayContain("world"));
+  ASSERT_TRUE(!bloom2.MayContain("x"));
+  ASSERT_TRUE(!bloom2.MayContain("foo"));
+}
+
+static uint32_t NextNum(uint32_t num) {
+  if (num < 10) {
+    num += 1;
+  } else if (num < 100) {
+    num += 10;
+  } else if (num < 1000) {
+    num += 100;
+  } else {
+    num = num * 26 / 10;
+  }
+  return num;
+}
+
+TEST_F(DynamicBloomTest, VaryingLengths) {
+  KeyMaker km;
+
+  // Count number of filters that significantly exceed the false positive rate
+  int mediocre_filters = 0;
+  int good_filters = 0;
+  uint32_t num_probes = static_cast<uint32_t>(FLAGS_num_probes);
+
+  fprintf(stderr, "bits_per_key: %d  num_probes: %d\n", FLAGS_bits_per_key,
+          num_probes);
+
+  // NB: FP rate impact of 32-bit hash is noticeable starting around 10M keys.
+  // But that effect is hidden if using sequential keys (unique hashes).
+  for (bool nonseq : {false, true}) {
+    const uint32_t max_num = FLAGS_enable_perf ? 40000000 : 400000;
+    for (uint32_t num = 1; num <= max_num; num = NextNum(num)) {
+      uint32_t bloom_bits = 0;
+      Arena arena;
+      bloom_bits = num * FLAGS_bits_per_key;
+      DynamicBloom bloom(&arena, bloom_bits, num_probes);
+      for (uint64_t i = 0; i < num; i++) {
+        bloom.Add(km.Key(i, nonseq));
+        ASSERT_TRUE(bloom.MayContain(km.Key(i, nonseq)));
+      }
+
+      // All added keys must match
+      for (uint64_t i = 0; i < num; i++) {
+        ASSERT_TRUE(bloom.MayContain(km.Key(i, nonseq)));
+      }
+
+      // Check false positive rate
+      int result = 0;
+      for (uint64_t i = 0; i < 30000; i++) {
+        if (bloom.MayContain(km.Key(i + 1000000000, nonseq))) {
+          result++;
+        }
+      }
+      double rate = result / 30000.0;
+
+      fprintf(stderr,
+              "False positives (%s keys): "
+              "%5.2f%% @ num = %6u, bloom_bits = %6u\n",
+              nonseq ? "nonseq" : "seq", rate * 100.0, num, bloom_bits);
+
+      if (rate > 0.0125)
+        mediocre_filters++;  // Allowed, but not too often
+      else
+        good_filters++;
+    }
+  }
+
+  fprintf(stderr, "Filters: %d good, %d mediocre\n", good_filters,
+          mediocre_filters);
+  ASSERT_LE(mediocre_filters, good_filters / 25);
+}
+
+TEST_F(DynamicBloomTest, perf) {
+  KeyMaker km;
+  StopWatchNano timer(SystemClock::Default().get());
+  uint32_t num_probes = static_cast<uint32_t>(FLAGS_num_probes);
+
+  if (!FLAGS_enable_perf) {
+    return;
+  }
+
+  for (uint32_t m = 1; m <= 8; ++m) {
+    Arena arena;
+    const uint32_t num_keys = m * 8 * 1024 * 1024;
+    fprintf(stderr, "testing %" PRIu32 "M keys\n", m * 8);
+
+    DynamicBloom std_bloom(&arena, num_keys * 10, num_probes);
+
+    timer.Start();
+    for (uint64_t i = 1; i <= num_keys; ++i) {
+      std_bloom.Add(km.Seq(i));
+    }
+
+    uint64_t elapsed = timer.ElapsedNanos();
+    fprintf(stderr, "dynamic bloom, avg add latency %3g\n",
+            static_cast<double>(elapsed) / num_keys);
+
+    uint32_t count = 0;
+    timer.Start();
+    for (uint64_t i = 1; i <= num_keys; ++i) {
+      if (std_bloom.MayContain(km.Seq(i))) {
+        ++count;
+      }
+    }
+    ASSERT_EQ(count, num_keys);
+    elapsed = timer.ElapsedNanos();
+    assert(count > 0);
+    fprintf(stderr, "dynamic bloom, avg query latency %3g\n",
+            static_cast<double>(elapsed) / count);
+  }
+}
+
+TEST_F(DynamicBloomTest, concurrent_with_perf) {
+  uint32_t num_probes = static_cast<uint32_t>(FLAGS_num_probes);
+
+  uint32_t m_limit = FLAGS_enable_perf ? 8 : 1;
+
+  uint32_t num_threads = 4;
+  std::vector<port::Thread> threads;
+
+  // NB: Uses sequential keys for speed, but that hides the FP rate
+  // impact of 32-bit hash, which is noticeable starting around 10M keys
+  // when they vary across hashing blocks.
+  for (uint32_t m = 1; m <= m_limit; ++m) {
+    Arena arena;
+    const uint32_t num_keys = m * 8 * 1024 * 1024;
+    fprintf(stderr, "testing %" PRIu32 "M keys\n", m * 8);
+
+    DynamicBloom std_bloom(&arena, num_keys * 10, num_probes);
+
+    std::atomic<uint64_t> elapsed(0);
+
+    std::function<void(size_t)> adder([&](size_t t) {
+      KeyMaker km;
+      StopWatchNano timer(SystemClock::Default().get());
+      timer.Start();
+      for (uint64_t i = 1 + t; i <= num_keys; i += num_threads) {
+        std_bloom.AddConcurrently(km.Seq(i));
+      }
+      elapsed += timer.ElapsedNanos();
+    });
+    for (size_t t = 0; t < num_threads; ++t) {
+      threads.emplace_back(adder, t);
+    }
+    while (threads.size() > 0) {
+      threads.back().join();
+      threads.pop_back();
+    }
+
+    fprintf(stderr,
+            "dynamic bloom, avg parallel add latency %3g"
+            " nanos/key\n",
+            static_cast<double>(elapsed) / num_threads / num_keys);
+
+    elapsed = 0;
+    std::function<void(size_t)> hitter([&](size_t t) {
+      KeyMaker km;
+      StopWatchNano timer(SystemClock::Default().get());
+      timer.Start();
+      for (uint64_t i = 1 + t; i <= num_keys; i += num_threads) {
+        bool f = std_bloom.MayContain(km.Seq(i));
+        ASSERT_TRUE(f);
+      }
+      elapsed += timer.ElapsedNanos();
+    });
+    for (size_t t = 0; t < num_threads; ++t) {
+      threads.emplace_back(hitter, t);
+    }
+    while (threads.size() > 0) {
+      threads.back().join();
+      threads.pop_back();
+    }
+
+    fprintf(stderr,
+            "dynamic bloom, avg parallel hit latency %3g"
+            " nanos/key\n",
+            static_cast<double>(elapsed) / num_threads / num_keys);
+
+    elapsed = 0;
+    std::atomic<uint32_t> false_positives(0);
+    std::function<void(size_t)> misser([&](size_t t) {
+      KeyMaker km;
+      StopWatchNano timer(SystemClock::Default().get());
+      timer.Start();
+      for (uint64_t i = num_keys + 1 + t; i <= 2 * num_keys; i += num_threads) {
+        bool f = std_bloom.MayContain(km.Seq(i));
+        if (f) {
+          ++false_positives;
+        }
+      }
+      elapsed += timer.ElapsedNanos();
+    });
+    for (size_t t = 0; t < num_threads; ++t) {
+      threads.emplace_back(misser, t);
+    }
+    while (threads.size() > 0) {
+      threads.back().join();
+      threads.pop_back();
+    }
+
+    fprintf(stderr,
+            "dynamic bloom, avg parallel miss latency %3g"
+            " nanos/key, %f%% false positive rate\n",
+            static_cast<double>(elapsed) / num_threads / num_keys,
+            false_positives.load() * 100.0 / num_keys);
+  }
+}
+
+}  // namespace ROCKSDB_NAMESPACE
+
+int main(int argc, char **argv) {
+  ROCKSDB_NAMESPACE::port::InstallStackTraceHandler();
+  ::testing::InitGoogleTest(&argc, argv);
+  ParseCommandLineFlags(&argc, &argv, true);
+
+  return RUN_ALL_TESTS();
+}
+
+#endif  // GFLAGS
diff --git a/src/rocksdb/util/fastrange.h b/src/rocksdb/util/fastrange.h
new file mode 100644
index 000000000..a70a980f6
--- /dev/null
+++ b/src/rocksdb/util/fastrange.h
@@ -0,0 +1,114 @@
+//  Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+// fastrange/FastRange: A faster alternative to % for mapping a hash value
+// to an arbitrary range. See https://github.com/lemire/fastrange
+//
+// Generally recommended are FastRange32 for mapping results of 32-bit
+// hash functions and FastRange64 for mapping results of 64-bit hash
+// functions. FastRange is less forgiving than % if the input hashes are
+// not well distributed over the full range of the type (32 or 64 bits).
+//
+// Also included is a templated implementation FastRangeGeneric for use
+// in generic algorithms, but not otherwise recommended because of
+// potential ambiguity. Unlike with %, it is critical to use the right
+// FastRange variant for the output size of your hash function.
+
+#pragma once
+
+#include <cstddef>
+#include <cstdint>
+#include <type_traits>
+
+#include "rocksdb/rocksdb_namespace.h"
+
+#ifdef TEST_UINT128_COMPAT
+#undef HAVE_UINT128_EXTENSION
+#endif
+
+namespace ROCKSDB_NAMESPACE {
+
+namespace detail {
+
+// Using a class template to support partial specialization
+template <typename Hash, typename Range>
+struct FastRangeGenericImpl {
+  // only reach this on no supported specialization
+};
+
+template <typename Range>
+struct FastRangeGenericImpl<uint32_t, Range> {
+  static inline Range Fn(uint32_t hash, Range range) {
+    static_assert(std::is_unsigned<Range>::value, "must be unsigned");
+    static_assert(sizeof(Range) <= sizeof(uint32_t),
+                  "cannot be larger than hash (32 bits)");
+
+    uint64_t product = uint64_t{range} * hash;
+    return static_cast<Range>(product >> 32);
+  }
+};
+
+template <typename Range>
+struct FastRangeGenericImpl<uint64_t, Range> {
+  static inline Range Fn(uint64_t hash, Range range) {
+    static_assert(std::is_unsigned<Range>::value, "must be unsigned");
+    static_assert(sizeof(Range) <= sizeof(uint64_t),
+                  "cannot be larger than hash (64 bits)");
+
+#ifdef HAVE_UINT128_EXTENSION
+    // Can use compiler's 128-bit type. Trust it to do the right thing.
+    __uint128_t wide = __uint128_t{range} * hash;
+    return static_cast<Range>(wide >> 64);
+#else
+    // Fall back: full decomposition.
+    // NOTE: GCC seems to fully understand this code as 64-bit x 64-bit
+    // -> 128-bit multiplication and optimize it appropriately
+    uint64_t range64 = range;  // ok to shift by 32, even if Range is 32-bit
+    uint64_t tmp = uint64_t{range64 & 0xffffFFFF} * uint64_t{hash & 0xffffFFFF};
+    tmp >>= 32;
+    tmp += uint64_t{range64 & 0xffffFFFF} * uint64_t{hash >> 32};
+    // Avoid overflow: first add lower 32 of tmp2, and later upper 32
+    uint64_t tmp2 = uint64_t{range64 >> 32} * uint64_t{hash & 0xffffFFFF};
+    tmp += static_cast<uint32_t>(tmp2);
+    tmp >>= 32;
+    tmp += (tmp2 >> 32);
+    tmp += uint64_t{range64 >> 32} * uint64_t{hash >> 32};
+    return static_cast<Range>(tmp);
+#endif
+  }
+};
+
+}  // namespace detail
+
+// Now an omnibus templated function (yay parameter inference).
+//
+// NOTICE:
+// This templated version is not recommended for typical use because
+// of the potential to mix a 64-bit FastRange with a 32-bit bit hash,
+// most likely because you put your 32-bit hash in an "unsigned long"
+// which is 64 bits on some platforms. That doesn't really matter for
+// an operation like %, but 64-bit FastRange gives extremely bad results,
+// mostly zero, on 32-bit hash values. And because good hashing is not
+// generally required for correctness, this kind of mistake could go
+// unnoticed with just unit tests. Plus it could vary by platform.
+template <typename Hash, typename Range>
+inline Range FastRangeGeneric(Hash hash, Range range) {
+  return detail::FastRangeGenericImpl<Hash, Range>::Fn(hash, range);
+}
+
+// The most popular / convenient / recommended variants:
+
+// Map a quality 64-bit hash value down to an arbitrary size_t range.
+// (size_t is standard for mapping to things in memory.)
+inline size_t FastRange64(uint64_t hash, size_t range) {
+  return FastRangeGeneric(hash, range);
+}
+
+// Map a quality 32-bit hash value down to an arbitrary uint32_t range.
+inline uint32_t FastRange32(uint32_t hash, uint32_t range) {
+  return FastRangeGeneric(hash, range);
+}
+
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/util/file_checksum_helper.cc b/src/rocksdb/util/file_checksum_helper.cc
new file mode 100644
index 000000000..a73920352
--- /dev/null
+++ b/src/rocksdb/util/file_checksum_helper.cc
@@ -0,0 +1,172 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+//
+//  Copyright (c) 2011 The LevelDB Authors. All rights reserved.
+//  Use of this source code is governed by a BSD-style license that can be
+//  found in the LICENSE file. See the AUTHORS file for names of contributors.
+
+#include "util/file_checksum_helper.h"
+
+#include <unordered_set>
+
+#include "db/log_reader.h"
+#include "db/version_edit.h"
+#include "db/version_edit_handler.h"
+#include "file/sequence_file_reader.h"
+#include "rocksdb/utilities/customizable_util.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+void FileChecksumListImpl::reset() { checksum_map_.clear(); }
+
+size_t FileChecksumListImpl::size() const { return checksum_map_.size(); }
+
+Status FileChecksumListImpl::GetAllFileChecksums(
+    std::vector<uint64_t>* file_numbers, std::vector<std::string>* checksums,
+    std::vector<std::string>* checksum_func_names) {
+  if (file_numbers == nullptr || checksums == nullptr ||
+      checksum_func_names == nullptr) {
+    return Status::InvalidArgument("Pointer has not been initiated");
+  }
+
+  for (auto i : checksum_map_) {
+    file_numbers->push_back(i.first);
+    checksums->push_back(i.second.first);
+    checksum_func_names->push_back(i.second.second);
+  }
+  return Status::OK();
+}
+
+Status FileChecksumListImpl::SearchOneFileChecksum(
+    uint64_t file_number, std::string* checksum,
+    std::string* checksum_func_name) {
+  if (checksum == nullptr || checksum_func_name == nullptr) {
+    return Status::InvalidArgument("Pointer has not been initiated");
+  }
+
+  auto it = checksum_map_.find(file_number);
+  if (it == checksum_map_.end()) {
+    return Status::NotFound();
+  } else {
+    *checksum = it->second.first;
+    *checksum_func_name = it->second.second;
+  }
+  return Status::OK();
+}
+
+Status FileChecksumListImpl::InsertOneFileChecksum(
+    uint64_t file_number, const std::string& checksum,
+    const std::string& checksum_func_name) {
+  auto it = checksum_map_.find(file_number);
+  if (it == checksum_map_.end()) {
+    checksum_map_.insert(std::make_pair(
+        file_number, std::make_pair(checksum, checksum_func_name)));
+  } else {
+    it->second.first = checksum;
+    it->second.second = checksum_func_name;
+  }
+  return Status::OK();
+}
+
+Status FileChecksumListImpl::RemoveOneFileChecksum(uint64_t file_number) {
+  auto it = checksum_map_.find(file_number);
+  if (it == checksum_map_.end()) {
+    return Status::NotFound();
+  } else {
+    checksum_map_.erase(it);
+  }
+  return Status::OK();
+}
+
+FileChecksumList* NewFileChecksumList() {
+  FileChecksumListImpl* checksum_list = new FileChecksumListImpl();
+  return checksum_list;
+}
+
+std::shared_ptr<FileChecksumGenFactory> GetFileChecksumGenCrc32cFactory() {
+  static std::shared_ptr<FileChecksumGenFactory> default_crc32c_gen_factory(
+      new FileChecksumGenCrc32cFactory());
+  return default_crc32c_gen_factory;
+}
+
+Status GetFileChecksumsFromManifest(Env* src_env, const std::string& abs_path,
+                                    uint64_t manifest_file_size,
+                                    FileChecksumList* checksum_list) {
+  if (checksum_list == nullptr) {
+    return Status::InvalidArgument("checksum_list is nullptr");
+  }
+  assert(checksum_list);
+  checksum_list->reset();
+  Status s;
+
+  std::unique_ptr<SequentialFileReader> file_reader;
+  {
+    std::unique_ptr<FSSequentialFile> file;
+    const std::shared_ptr<FileSystem>& fs = src_env->GetFileSystem();
+    s = fs->NewSequentialFile(abs_path,
+                              fs->OptimizeForManifestRead(FileOptions()), &file,
+                              nullptr /* dbg */);
+    if (!s.ok()) {
+      return s;
+    }
+    file_reader.reset(new SequentialFileReader(std::move(file), abs_path));
+  }
+
+  struct LogReporter : public log::Reader::Reporter {
+    Status* status_ptr;
+    virtual void Corruption(size_t /*bytes*/, const Status& st) override {
+      if (status_ptr->ok()) {
+        *status_ptr = st;
+      }
+    }
+  } reporter;
+  reporter.status_ptr = &s;
+  log::Reader reader(nullptr, std::move(file_reader), &reporter,
+                     true /* checksum */, 0 /* log_number */);
+  FileChecksumRetriever retriever(manifest_file_size, *checksum_list);
+  retriever.Iterate(reader, &s);
+  assert(!retriever.status().ok() ||
+         manifest_file_size == std::numeric_limits<uint64_t>::max() ||
+         reader.LastRecordEnd() == manifest_file_size);
+
+  return retriever.status();
+}
+
+#ifndef ROCKSDB_LITE
+namespace {
+static int RegisterFileChecksumGenFactories(ObjectLibrary& library,
+                                            const std::string& /*arg*/) {
+  library.AddFactory<FileChecksumGenFactory>(
+      FileChecksumGenCrc32cFactory::kClassName(),
+      [](const std::string& /*uri*/,
+         std::unique_ptr<FileChecksumGenFactory>* guard,
+         std::string* /* errmsg */) {
+        guard->reset(new FileChecksumGenCrc32cFactory());
+        return guard->get();
+      });
+  return 1;
+}
+}  // namespace
+#endif  // !ROCKSDB_LITE
+
+Status FileChecksumGenFactory::CreateFromString(
+    const ConfigOptions& options, const std::string& value,
+    std::shared_ptr<FileChecksumGenFactory>* result) {
+#ifndef ROCKSDB_LITE
+  static std::once_flag once;
+  std::call_once(once, [&]() {
+    RegisterFileChecksumGenFactories(*(ObjectLibrary::Default().get()), "");
+  });
+#endif  // ROCKSDB_LITE
+  if (value == FileChecksumGenCrc32cFactory::kClassName()) {
+    *result = GetFileChecksumGenCrc32cFactory();
+    return Status::OK();
+  } else {
+    Status s = LoadSharedObject<FileChecksumGenFactory>(options, value, nullptr,
+                                                        result);
+    return s;
+  }
+}
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/util/file_checksum_helper.h b/src/rocksdb/util/file_checksum_helper.h
new file mode 100644
index 000000000..d622e9bba
--- /dev/null
+++ b/src/rocksdb/util/file_checksum_helper.h
@@ -0,0 +1,100 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#pragma once
+#include <cassert>
+#include <unordered_map>
+
+#include "port/port.h"
+#include "rocksdb/file_checksum.h"
+#include "rocksdb/status.h"
+#include "util/coding.h"
+#include "util/crc32c.h"
+#include "util/math.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+// This is the class to generate the file checksum based on Crc32. It
+// will be used as the default checksum method for SST file checksum
+class FileChecksumGenCrc32c : public FileChecksumGenerator {
+ public:
+  FileChecksumGenCrc32c(const FileChecksumGenContext& /*context*/) {
+    checksum_ = 0;
+  }
+
+  void Update(const char* data, size_t n) override {
+    checksum_ = crc32c::Extend(checksum_, data, n);
+  }
+
+  void Finalize() override {
+    assert(checksum_str_.empty());
+    // Store as big endian raw bytes
+    PutFixed32(&checksum_str_, EndianSwapValue(checksum_));
+  }
+
+  std::string GetChecksum() const override {
+    assert(!checksum_str_.empty());
+    return checksum_str_;
+  }
+
+  const char* Name() const override { return "FileChecksumCrc32c"; }
+
+ private:
+  uint32_t checksum_;
+  std::string checksum_str_;
+};
+
+class FileChecksumGenCrc32cFactory : public FileChecksumGenFactory {
+ public:
+  std::unique_ptr<FileChecksumGenerator> CreateFileChecksumGenerator(
+      const FileChecksumGenContext& context) override {
+    if (context.requested_checksum_func_name.empty() ||
+        context.requested_checksum_func_name == "FileChecksumCrc32c") {
+      return std::unique_ptr<FileChecksumGenerator>(
+          new FileChecksumGenCrc32c(context));
+    } else {
+      return nullptr;
+    }
+  }
+
+  static const char* kClassName() { return "FileChecksumGenCrc32cFactory"; }
+  const char* Name() const override { return kClassName(); }
+};
+
+// The default implementaion of FileChecksumList
+class FileChecksumListImpl : public FileChecksumList {
+ public:
+  FileChecksumListImpl() {}
+  void reset() override;
+
+  size_t size() const override;
+
+  Status GetAllFileChecksums(
+      std::vector<uint64_t>* file_numbers, std::vector<std::string>* checksums,
+      std::vector<std::string>* checksum_func_names) override;
+
+  Status SearchOneFileChecksum(uint64_t file_number, std::string* checksum,
+                               std::string* checksum_func_name) override;
+
+  Status InsertOneFileChecksum(uint64_t file_number,
+                               const std::string& checksum,
+                               const std::string& checksum_func_name) override;
+
+  Status RemoveOneFileChecksum(uint64_t file_number) override;
+
+ private:
+  // Key is the file number, the first portion of the value is checksum, the
+  // second portion of the value is checksum function name.
+  std::unordered_map<uint64_t, std::pair<std::string, std::string>>
+      checksum_map_;
+};
+
+// If manifest_file_size < std::numeric_limits<uint64_t>::max(), only use
+// that length prefix of the manifest file.
+Status GetFileChecksumsFromManifest(Env* src_env, const std::string& abs_path,
+                                    uint64_t manifest_file_size,
+                                    FileChecksumList* checksum_list);
+
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/util/file_reader_writer_test.cc b/src/rocksdb/util/file_reader_writer_test.cc
new file mode 100644
index 000000000..e778efc3c
--- /dev/null
+++ b/src/rocksdb/util/file_reader_writer_test.cc
@@ -0,0 +1,1066 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+//
+#include <algorithm>
+#include <vector>
+
+#include "db/db_test_util.h"
+#include "env/mock_env.h"
+#include "file/line_file_reader.h"
+#include "file/random_access_file_reader.h"
+#include "file/read_write_util.h"
+#include "file/readahead_raf.h"
+#include "file/sequence_file_reader.h"
+#include "file/writable_file_writer.h"
+#include "rocksdb/file_system.h"
+#include "test_util/testharness.h"
+#include "test_util/testutil.h"
+#include "util/crc32c.h"
+#include "util/random.h"
+#include "utilities/fault_injection_fs.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+class WritableFileWriterTest : public testing::Test {};
+
+constexpr uint32_t kMb = static_cast<uint32_t>(1) << 20;
+
+TEST_F(WritableFileWriterTest, RangeSync) {
+  class FakeWF : public FSWritableFile {
+   public:
+    explicit FakeWF() : size_(0), last_synced_(0) {}
+    ~FakeWF() override {}
+
+    using FSWritableFile::Append;
+    IOStatus Append(const Slice& data, const IOOptions& /*options*/,
+                    IODebugContext* /*dbg*/) override {
+      size_ += data.size();
+      return IOStatus::OK();
+    }
+    IOStatus Truncate(uint64_t /*size*/, const IOOptions& /*options*/,
+                      IODebugContext* /*dbg*/) override {
+      return IOStatus::OK();
+    }
+    IOStatus Close(const IOOptions& /*options*/,
+                   IODebugContext* /*dbg*/) override {
+      EXPECT_GE(size_, last_synced_ + kMb);
+      EXPECT_LT(size_, last_synced_ + 2 * kMb);
+      // Make sure random writes generated enough writes.
+      EXPECT_GT(size_, 10 * kMb);
+      return IOStatus::OK();
+    }
+    IOStatus Flush(const IOOptions& /*options*/,
+                   IODebugContext* /*dbg*/) override {
+      return IOStatus::OK();
+    }
+    IOStatus Sync(const IOOptions& /*options*/,
+                  IODebugContext* /*dbg*/) override {
+      return IOStatus::OK();
+    }
+    IOStatus Fsync(const IOOptions& /*options*/,
+                   IODebugContext* /*dbg*/) override {
+      return IOStatus::OK();
+    }
+    void SetIOPriority(Env::IOPriority /*pri*/) override {}
+    uint64_t GetFileSize(const IOOptions& /*options*/,
+                         IODebugContext* /*dbg*/) override {
+      return size_;
+    }
+    void GetPreallocationStatus(size_t* /*block_size*/,
+                                size_t* /*last_allocated_block*/) override {}
+    size_t GetUniqueId(char* /*id*/, size_t /*max_size*/) const override {
+      return 0;
+    }
+    IOStatus InvalidateCache(size_t /*offset*/, size_t /*length*/) override {
+      return IOStatus::OK();
+    }
+
+   protected:
+    IOStatus Allocate(uint64_t /*offset*/, uint64_t /*len*/,
+                      const IOOptions& /*options*/,
+                      IODebugContext* /*dbg*/) override {
+      return IOStatus::OK();
+    }
+    IOStatus RangeSync(uint64_t offset, uint64_t nbytes,
+                       const IOOptions& /*options*/,
+                       IODebugContext* /*dbg*/) override {
+      EXPECT_EQ(offset % 4096, 0u);
+      EXPECT_EQ(nbytes % 4096, 0u);
+
+      EXPECT_EQ(offset, last_synced_);
+      last_synced_ = offset + nbytes;
+      EXPECT_GE(size_, last_synced_ + kMb);
+      if (size_ > 2 * kMb) {
+        EXPECT_LT(size_, last_synced_ + 2 * kMb);
+      }
+      return IOStatus::OK();
+    }
+
+    uint64_t size_;
+    uint64_t last_synced_;
+  };
+
+  EnvOptions env_options;
+  env_options.bytes_per_sync = kMb;
+  std::unique_ptr<FakeWF> wf(new FakeWF);
+  std::unique_ptr<WritableFileWriter> writer(
+      new WritableFileWriter(std::move(wf), "" /* don't care */, env_options));
+  Random r(301);
+  Status s;
+  std::unique_ptr<char[]> large_buf(new char[10 * kMb]);
+  for (int i = 0; i < 1000; i++) {
+    int skew_limit = (i < 700) ? 10 : 15;
+    uint32_t num = r.Skewed(skew_limit) * 100 + r.Uniform(100);
+    s = writer->Append(Slice(large_buf.get(), num));
+    ASSERT_OK(s);
+
+    // Flush in a chance of 1/10.
+    if (r.Uniform(10) == 0) {
+      s = writer->Flush();
+      ASSERT_OK(s);
+    }
+  }
+  s = writer->Close();
+  ASSERT_OK(s);
+}
+
+TEST_F(WritableFileWriterTest, IncrementalBuffer) {
+  class FakeWF : public FSWritableFile {
+   public:
+    explicit FakeWF(std::string* _file_data, bool _use_direct_io,
+                    bool _no_flush)
+        : file_data_(_file_data),
+          use_direct_io_(_use_direct_io),
+          no_flush_(_no_flush) {}
+    ~FakeWF() override {}
+
+    using FSWritableFile::Append;
+    IOStatus Append(const Slice& data, const IOOptions& /*options*/,
+                    IODebugContext* /*dbg*/) override {
+      file_data_->append(data.data(), data.size());
+      size_ += data.size();
+      return IOStatus::OK();
+    }
+    using FSWritableFile::PositionedAppend;
+    IOStatus PositionedAppend(const Slice& data, uint64_t pos,
+                              const IOOptions& /*options*/,
+                              IODebugContext* /*dbg*/) override {
+      EXPECT_TRUE(pos % 512 == 0);
+      EXPECT_TRUE(data.size() % 512 == 0);
+      file_data_->resize(pos);
+      file_data_->append(data.data(), data.size());
+      size_ += data.size();
+      return IOStatus::OK();
+    }
+
+    IOStatus Truncate(uint64_t size, const IOOptions& /*options*/,
+                      IODebugContext* /*dbg*/) override {
+      file_data_->resize(size);
+      return IOStatus::OK();
+    }
+    IOStatus Close(const IOOptions& /*options*/,
+                   IODebugContext* /*dbg*/) override {
+      return IOStatus::OK();
+    }
+    IOStatus Flush(const IOOptions& /*options*/,
+                   IODebugContext* /*dbg*/) override {
+      return IOStatus::OK();
+    }
+    IOStatus Sync(const IOOptions& /*options*/,
+                  IODebugContext* /*dbg*/) override {
+      return IOStatus::OK();
+    }
+    IOStatus Fsync(const IOOptions& /*options*/,
+                   IODebugContext* /*dbg*/) override {
+      return IOStatus::OK();
+    }
+    void SetIOPriority(Env::IOPriority /*pri*/) override {}
+    uint64_t GetFileSize(const IOOptions& /*options*/,
+                         IODebugContext* /*dbg*/) override {
+      return size_;
+    }
+    void GetPreallocationStatus(size_t* /*block_size*/,
+                                size_t* /*last_allocated_block*/) override {}
+    size_t GetUniqueId(char* /*id*/, size_t /*max_size*/) const override {
+      return 0;
+    }
+    IOStatus InvalidateCache(size_t /*offset*/, size_t /*length*/) override {
+      return IOStatus::OK();
+    }
+    bool use_direct_io() const override { return use_direct_io_; }
+
+    std::string* file_data_;
+    bool use_direct_io_;
+    bool no_flush_;
+    size_t size_ = 0;
+  };
+
+  Random r(301);
+  const int kNumAttempts = 50;
+  for (int attempt = 0; attempt < kNumAttempts; attempt++) {
+    bool no_flush = (attempt % 3 == 0);
+    EnvOptions env_options;
+    env_options.writable_file_max_buffer_size =
+        (attempt < kNumAttempts / 2) ? 512 * 1024 : 700 * 1024;
+    std::string actual;
+    std::unique_ptr<FakeWF> wf(new FakeWF(&actual,
+#ifndef ROCKSDB_LITE
+                                          attempt % 2 == 1,
+#else
+                                          false,
+#endif
+                                          no_flush));
+    std::unique_ptr<WritableFileWriter> writer(new WritableFileWriter(
+        std::move(wf), "" /* don't care */, env_options));
+
+    std::string target;
+    for (int i = 0; i < 20; i++) {
+      uint32_t num = r.Skewed(16) * 100 + r.Uniform(100);
+      std::string random_string = r.RandomString(num);
+      ASSERT_OK(writer->Append(Slice(random_string.c_str(), num)));
+      target.append(random_string.c_str(), num);
+
+      // In some attempts, flush in a chance of 1/10.
+      if (!no_flush && r.Uniform(10) == 0) {
+        ASSERT_OK(writer->Flush());
+      }
+    }
+    ASSERT_OK(writer->Flush());
+    ASSERT_OK(writer->Close());
+    ASSERT_EQ(target.size(), actual.size());
+    ASSERT_EQ(target, actual);
+  }
+}
+
+TEST_F(WritableFileWriterTest, BufferWithZeroCapacityDirectIO) {
+  EnvOptions env_opts;
+  env_opts.use_direct_writes = true;
+  env_opts.writable_file_max_buffer_size = 0;
+  {
+    std::unique_ptr<WritableFileWriter> writer;
+    const Status s =
+        WritableFileWriter::Create(FileSystem::Default(), /*fname=*/"dont_care",
+                                   FileOptions(env_opts), &writer,
+                                   /*dbg=*/nullptr);
+    ASSERT_TRUE(s.IsInvalidArgument());
+  }
+}
+
+class DBWritableFileWriterTest : public DBTestBase {
+ public:
+  DBWritableFileWriterTest()
+      : DBTestBase("db_secondary_cache_test", /*env_do_fsync=*/true) {
+    fault_fs_.reset(new FaultInjectionTestFS(env_->GetFileSystem()));
+    fault_env_.reset(new CompositeEnvWrapper(env_, fault_fs_));
+  }
+
+  std::shared_ptr<FaultInjectionTestFS> fault_fs_;
+  std::unique_ptr<Env> fault_env_;
+};
+
+TEST_F(DBWritableFileWriterTest, AppendWithChecksum) {
+  FileOptions file_options = FileOptions();
+  Options options = GetDefaultOptions();
+  options.create_if_missing = true;
+  DestroyAndReopen(options);
+  std::string fname = dbname_ + "/test_file";
+  std::unique_ptr<FSWritableFile> writable_file_ptr;
+  ASSERT_OK(fault_fs_->NewWritableFile(fname, file_options, &writable_file_ptr,
+                                       /*dbg*/ nullptr));
+  std::unique_ptr<TestFSWritableFile> file;
+  file.reset(new TestFSWritableFile(
+      fname, file_options, std::move(writable_file_ptr), fault_fs_.get()));
+  std::unique_ptr<WritableFileWriter> file_writer;
+  ImmutableOptions ioptions(options);
+  file_writer.reset(new WritableFileWriter(
+      std::move(file), fname, file_options, SystemClock::Default().get(),
+      nullptr, ioptions.stats, ioptions.listeners,
+      ioptions.file_checksum_gen_factory.get(), true, true));
+
+  Random rnd(301);
+  std::string data = rnd.RandomString(1000);
+  uint32_t data_crc32c = crc32c::Value(data.c_str(), data.size());
+  fault_fs_->SetChecksumHandoffFuncType(ChecksumType::kCRC32c);
+
+  ASSERT_OK(file_writer->Append(Slice(data.c_str()), data_crc32c));
+  ASSERT_OK(file_writer->Flush());
+  Random size_r(47);
+  for (int i = 0; i < 2000; i++) {
+    data = rnd.RandomString((static_cast<int>(size_r.Next()) % 10000));
+    data_crc32c = crc32c::Value(data.c_str(), data.size());
+    ASSERT_OK(file_writer->Append(Slice(data.c_str()), data_crc32c));
+
+    data = rnd.RandomString((static_cast<int>(size_r.Next()) % 97));
+    ASSERT_OK(file_writer->Append(Slice(data.c_str())));
+    ASSERT_OK(file_writer->Flush());
+  }
+  ASSERT_OK(file_writer->Close());
+  Destroy(options);
+}
+
+TEST_F(DBWritableFileWriterTest, AppendVerifyNoChecksum) {
+  FileOptions file_options = FileOptions();
+  Options options = GetDefaultOptions();
+  options.create_if_missing = true;
+  DestroyAndReopen(options);
+  std::string fname = dbname_ + "/test_file";
+  std::unique_ptr<FSWritableFile> writable_file_ptr;
+  ASSERT_OK(fault_fs_->NewWritableFile(fname, file_options, &writable_file_ptr,
+                                       /*dbg*/ nullptr));
+  std::unique_ptr<TestFSWritableFile> file;
+  file.reset(new TestFSWritableFile(
+      fname, file_options, std::move(writable_file_ptr), fault_fs_.get()));
+  std::unique_ptr<WritableFileWriter> file_writer;
+  ImmutableOptions ioptions(options);
+  // Enable checksum handoff for this file, but do not enable buffer checksum.
+  // So Append with checksum logic will not be triggered
+  file_writer.reset(new WritableFileWriter(
+      std::move(file), fname, file_options, SystemClock::Default().get(),
+      nullptr, ioptions.stats, ioptions.listeners,
+      ioptions.file_checksum_gen_factory.get(), true, false));
+
+  Random rnd(301);
+  std::string data = rnd.RandomString(1000);
+  uint32_t data_crc32c = crc32c::Value(data.c_str(), data.size());
+  fault_fs_->SetChecksumHandoffFuncType(ChecksumType::kCRC32c);
+
+  ASSERT_OK(file_writer->Append(Slice(data.c_str()), data_crc32c));
+  ASSERT_OK(file_writer->Flush());
+  Random size_r(47);
+  for (int i = 0; i < 1000; i++) {
+    data = rnd.RandomString((static_cast<int>(size_r.Next()) % 10000));
+    data_crc32c = crc32c::Value(data.c_str(), data.size());
+    ASSERT_OK(file_writer->Append(Slice(data.c_str()), data_crc32c));
+
+    data = rnd.RandomString((static_cast<int>(size_r.Next()) % 97));
+    ASSERT_OK(file_writer->Append(Slice(data.c_str())));
+    ASSERT_OK(file_writer->Flush());
+  }
+  ASSERT_OK(file_writer->Close());
+  Destroy(options);
+}
+
+TEST_F(DBWritableFileWriterTest, AppendWithChecksumRateLimiter) {
+  FileOptions file_options = FileOptions();
+  file_options.rate_limiter = nullptr;
+  Options options = GetDefaultOptions();
+  options.create_if_missing = true;
+  DestroyAndReopen(options);
+  std::string fname = dbname_ + "/test_file";
+  std::unique_ptr<FSWritableFile> writable_file_ptr;
+  ASSERT_OK(fault_fs_->NewWritableFile(fname, file_options, &writable_file_ptr,
+                                       /*dbg*/ nullptr));
+  std::unique_ptr<TestFSWritableFile> file;
+  file.reset(new TestFSWritableFile(
+      fname, file_options, std::move(writable_file_ptr), fault_fs_.get()));
+  std::unique_ptr<WritableFileWriter> file_writer;
+  ImmutableOptions ioptions(options);
+  // Enable checksum handoff for this file, but do not enable buffer checksum.
+  // So Append with checksum logic will not be triggered
+  file_writer.reset(new WritableFileWriter(
+      std::move(file), fname, file_options, SystemClock::Default().get(),
+      nullptr, ioptions.stats, ioptions.listeners,
+      ioptions.file_checksum_gen_factory.get(), true, true));
+  fault_fs_->SetChecksumHandoffFuncType(ChecksumType::kCRC32c);
+
+  Random rnd(301);
+  std::string data;
+  uint32_t data_crc32c;
+  uint64_t start = fault_env_->NowMicros();
+  Random size_r(47);
+  uint64_t bytes_written = 0;
+  for (int i = 0; i < 100; i++) {
+    data = rnd.RandomString((static_cast<int>(size_r.Next()) % 10000));
+    data_crc32c = crc32c::Value(data.c_str(), data.size());
+    ASSERT_OK(file_writer->Append(Slice(data.c_str()), data_crc32c));
+    bytes_written += static_cast<uint64_t>(data.size());
+
+    data = rnd.RandomString((static_cast<int>(size_r.Next()) % 97));
+    ASSERT_OK(file_writer->Append(Slice(data.c_str())));
+    ASSERT_OK(file_writer->Flush());
+    bytes_written += static_cast<uint64_t>(data.size());
+  }
+  uint64_t elapsed = fault_env_->NowMicros() - start;
+  double raw_rate = bytes_written * 1000000.0 / elapsed;
+  ASSERT_OK(file_writer->Close());
+
+  // Set the rate-limiter
+  FileOptions file_options1 = FileOptions();
+  file_options1.rate_limiter =
+      NewGenericRateLimiter(static_cast<int64_t>(0.5 * raw_rate));
+  fname = dbname_ + "/test_file_1";
+  std::unique_ptr<FSWritableFile> writable_file_ptr1;
+  ASSERT_OK(fault_fs_->NewWritableFile(fname, file_options1,
+                                       &writable_file_ptr1,
+                                       /*dbg*/ nullptr));
+  file.reset(new TestFSWritableFile(
+      fname, file_options1, std::move(writable_file_ptr1), fault_fs_.get()));
+  // Enable checksum handoff for this file, but do not enable buffer checksum.
+  // So Append with checksum logic will not be triggered
+  file_writer.reset(new WritableFileWriter(
+      std::move(file), fname, file_options1, SystemClock::Default().get(),
+      nullptr, ioptions.stats, ioptions.listeners,
+      ioptions.file_checksum_gen_factory.get(), true, true));
+
+  for (int i = 0; i < 1000; i++) {
+    data = rnd.RandomString((static_cast<int>(size_r.Next()) % 10000));
+    data_crc32c = crc32c::Value(data.c_str(), data.size());
+    ASSERT_OK(file_writer->Append(Slice(data.c_str()), data_crc32c));
+
+    data = rnd.RandomString((static_cast<int>(size_r.Next()) % 97));
+    ASSERT_OK(file_writer->Append(Slice(data.c_str())));
+    ASSERT_OK(file_writer->Flush());
+  }
+  ASSERT_OK(file_writer->Close());
+  if (file_options1.rate_limiter != nullptr) {
+    delete file_options1.rate_limiter;
+  }
+
+  Destroy(options);
+}
+
+#ifndef ROCKSDB_LITE
+TEST_F(WritableFileWriterTest, AppendStatusReturn) {
+  class FakeWF : public FSWritableFile {
+   public:
+    explicit FakeWF() : use_direct_io_(false), io_error_(false) {}
+
+    bool use_direct_io() const override { return use_direct_io_; }
+
+    using FSWritableFile::Append;
+    IOStatus Append(const Slice& /*data*/, const IOOptions& /*options*/,
+                    IODebugContext* /*dbg*/) override {
+      if (io_error_) {
+        return IOStatus::IOError("Fake IO error");
+      }
+      return IOStatus::OK();
+    }
+    using FSWritableFile::PositionedAppend;
+    IOStatus PositionedAppend(const Slice& /*data*/, uint64_t,
+                              const IOOptions& /*options*/,
+                              IODebugContext* /*dbg*/) override {
+      if (io_error_) {
+        return IOStatus::IOError("Fake IO error");
+      }
+      return IOStatus::OK();
+    }
+    IOStatus Close(const IOOptions& /*options*/,
+                   IODebugContext* /*dbg*/) override {
+      return IOStatus::OK();
+    }
+    IOStatus Flush(const IOOptions& /*options*/,
+                   IODebugContext* /*dbg*/) override {
+      return IOStatus::OK();
+    }
+    IOStatus Sync(const IOOptions& /*options*/,
+                  IODebugContext* /*dbg*/) override {
+      return IOStatus::OK();
+    }
+    void Setuse_direct_io(bool val) { use_direct_io_ = val; }
+    void SetIOError(bool val) { io_error_ = val; }
+
+   protected:
+    bool use_direct_io_;
+    bool io_error_;
+  };
+  std::unique_ptr<FakeWF> wf(new FakeWF());
+  wf->Setuse_direct_io(true);
+  std::unique_ptr<WritableFileWriter> writer(
+      new WritableFileWriter(std::move(wf), "" /* don't care */, EnvOptions()));
+
+  ASSERT_OK(writer->Append(std::string(2 * kMb, 'a')));
+
+  // Next call to WritableFile::Append() should fail
+  FakeWF* fwf = static_cast<FakeWF*>(writer->writable_file());
+  fwf->SetIOError(true);
+  ASSERT_NOK(writer->Append(std::string(2 * kMb, 'b')));
+}
+#endif
+
+class ReadaheadRandomAccessFileTest
+    : public testing::Test,
+      public testing::WithParamInterface<size_t> {
+ public:
+  static std::vector<size_t> GetReadaheadSizeList() {
+    return {1lu << 12, 1lu << 16};
+  }
+  void SetUp() override {
+    readahead_size_ = GetParam();
+    scratch_.reset(new char[2 * readahead_size_]);
+    ResetSourceStr();
+  }
+  ReadaheadRandomAccessFileTest() : control_contents_() {}
+  std::string Read(uint64_t offset, size_t n) {
+    Slice result;
+    Status s = test_read_holder_->Read(offset, n, IOOptions(), &result,
+                                       scratch_.get(), nullptr);
+    EXPECT_TRUE(s.ok() || s.IsInvalidArgument());
+    return std::string(result.data(), result.size());
+  }
+  void ResetSourceStr(const std::string& str = "") {
+    std::unique_ptr<FSWritableFile> sink(
+        new test::StringSink(&control_contents_));
+    std::unique_ptr<WritableFileWriter> write_holder(new WritableFileWriter(
+        std::move(sink), "" /* don't care */, FileOptions()));
+    Status s = write_holder->Append(Slice(str));
+    EXPECT_OK(s);
+    s = write_holder->Flush();
+    EXPECT_OK(s);
+    std::unique_ptr<FSRandomAccessFile> read_holder(
+        new test::StringSource(control_contents_));
+    test_read_holder_ =
+        NewReadaheadRandomAccessFile(std::move(read_holder), readahead_size_);
+  }
+  size_t GetReadaheadSize() const { return readahead_size_; }
+
+ private:
+  size_t readahead_size_;
+  Slice control_contents_;
+  std::unique_ptr<FSRandomAccessFile> test_read_holder_;
+  std::unique_ptr<char[]> scratch_;
+};
+
+TEST_P(ReadaheadRandomAccessFileTest, EmptySourceStr) {
+  ASSERT_EQ("", Read(0, 1));
+  ASSERT_EQ("", Read(0, 0));
+  ASSERT_EQ("", Read(13, 13));
+}
+
+TEST_P(ReadaheadRandomAccessFileTest, SourceStrLenLessThanReadaheadSize) {
+  std::string str = "abcdefghijklmnopqrs";
+  ResetSourceStr(str);
+  ASSERT_EQ(str.substr(3, 4), Read(3, 4));
+  ASSERT_EQ(str.substr(0, 3), Read(0, 3));
+  ASSERT_EQ(str, Read(0, str.size()));
+  ASSERT_EQ(str.substr(7, std::min(static_cast<int>(str.size()) - 7, 30)),
+            Read(7, 30));
+  ASSERT_EQ("", Read(100, 100));
+}
+
+TEST_P(ReadaheadRandomAccessFileTest, SourceStrLenGreaterThanReadaheadSize) {
+  Random rng(42);
+  for (int k = 0; k < 100; ++k) {
+    size_t strLen = k * GetReadaheadSize() +
+                    rng.Uniform(static_cast<int>(GetReadaheadSize()));
+    std::string str = rng.HumanReadableString(static_cast<int>(strLen));
+    ResetSourceStr(str);
+    for (int test = 1; test <= 100; ++test) {
+      size_t offset = rng.Uniform(static_cast<int>(strLen));
+      size_t n = rng.Uniform(static_cast<int>(GetReadaheadSize()));
+      ASSERT_EQ(str.substr(offset, std::min(n, strLen - offset)),
+                Read(offset, n));
+    }
+  }
+}
+
+TEST_P(ReadaheadRandomAccessFileTest, ReadExceedsReadaheadSize) {
+  Random rng(7);
+  size_t strLen = 4 * GetReadaheadSize() +
+                  rng.Uniform(static_cast<int>(GetReadaheadSize()));
+  std::string str = rng.HumanReadableString(static_cast<int>(strLen));
+  ResetSourceStr(str);
+  for (int test = 1; test <= 100; ++test) {
+    size_t offset = rng.Uniform(static_cast<int>(strLen));
+    size_t n =
+        GetReadaheadSize() + rng.Uniform(static_cast<int>(GetReadaheadSize()));
+    ASSERT_EQ(str.substr(offset, std::min(n, strLen - offset)),
+              Read(offset, n));
+  }
+}
+
+INSTANTIATE_TEST_CASE_P(
+    EmptySourceStr, ReadaheadRandomAccessFileTest,
+    ::testing::ValuesIn(ReadaheadRandomAccessFileTest::GetReadaheadSizeList()));
+INSTANTIATE_TEST_CASE_P(
+    SourceStrLenLessThanReadaheadSize, ReadaheadRandomAccessFileTest,
+    ::testing::ValuesIn(ReadaheadRandomAccessFileTest::GetReadaheadSizeList()));
+INSTANTIATE_TEST_CASE_P(
+    SourceStrLenGreaterThanReadaheadSize, ReadaheadRandomAccessFileTest,
+    ::testing::ValuesIn(ReadaheadRandomAccessFileTest::GetReadaheadSizeList()));
+INSTANTIATE_TEST_CASE_P(
+    ReadExceedsReadaheadSize, ReadaheadRandomAccessFileTest,
+    ::testing::ValuesIn(ReadaheadRandomAccessFileTest::GetReadaheadSizeList()));
+
+class ReadaheadSequentialFileTest : public testing::Test,
+                                    public testing::WithParamInterface<size_t> {
+ public:
+  static std::vector<size_t> GetReadaheadSizeList() {
+    return {1lu << 8, 1lu << 12, 1lu << 16, 1lu << 18};
+  }
+  void SetUp() override {
+    readahead_size_ = GetParam();
+    scratch_.reset(new char[2 * readahead_size_]);
+    ResetSourceStr();
+  }
+  ReadaheadSequentialFileTest() {}
+  std::string Read(size_t n) {
+    Slice result;
+    Status s = test_read_holder_->Read(
+        n, &result, scratch_.get(), Env::IO_TOTAL /* rate_limiter_priority*/);
+    EXPECT_TRUE(s.ok() || s.IsInvalidArgument());
+    return std::string(result.data(), result.size());
+  }
+  void Skip(size_t n) { test_read_holder_->Skip(n); }
+  void ResetSourceStr(const std::string& str = "") {
+    auto read_holder = std::unique_ptr<FSSequentialFile>(
+        new test::SeqStringSource(str, &seq_read_count_));
+    test_read_holder_.reset(new SequentialFileReader(std::move(read_holder),
+                                                     "test", readahead_size_));
+  }
+  size_t GetReadaheadSize() const { return readahead_size_; }
+
+ private:
+  size_t readahead_size_;
+  std::unique_ptr<SequentialFileReader> test_read_holder_;
+  std::unique_ptr<char[]> scratch_;
+  std::atomic<int> seq_read_count_;
+};
+
+TEST_P(ReadaheadSequentialFileTest, EmptySourceStr) {
+  ASSERT_EQ("", Read(0));
+  ASSERT_EQ("", Read(1));
+  ASSERT_EQ("", Read(13));
+}
+
+TEST_P(ReadaheadSequentialFileTest, SourceStrLenLessThanReadaheadSize) {
+  std::string str = "abcdefghijklmnopqrs";
+  ResetSourceStr(str);
+  ASSERT_EQ(str.substr(0, 3), Read(3));
+  ASSERT_EQ(str.substr(3, 1), Read(1));
+  ASSERT_EQ(str.substr(4), Read(str.size()));
+  ASSERT_EQ("", Read(100));
+}
+
+TEST_P(ReadaheadSequentialFileTest, SourceStrLenGreaterThanReadaheadSize) {
+  Random rng(42);
+  for (int s = 0; s < 1; ++s) {
+    for (int k = 0; k < 100; ++k) {
+      size_t strLen = k * GetReadaheadSize() +
+                      rng.Uniform(static_cast<int>(GetReadaheadSize()));
+      std::string str = rng.HumanReadableString(static_cast<int>(strLen));
+      ResetSourceStr(str);
+      size_t offset = 0;
+      for (int test = 1; test <= 100; ++test) {
+        size_t n = rng.Uniform(static_cast<int>(GetReadaheadSize()));
+        if (s && test % 2) {
+          Skip(n);
+        } else {
+          ASSERT_EQ(str.substr(offset, std::min(n, strLen - offset)), Read(n));
+        }
+        offset = std::min(offset + n, strLen);
+      }
+    }
+  }
+}
+
+TEST_P(ReadaheadSequentialFileTest, ReadExceedsReadaheadSize) {
+  Random rng(42);
+  for (int s = 0; s < 1; ++s) {
+    for (int k = 0; k < 100; ++k) {
+      size_t strLen = k * GetReadaheadSize() +
+                      rng.Uniform(static_cast<int>(GetReadaheadSize()));
+      std::string str = rng.HumanReadableString(static_cast<int>(strLen));
+      ResetSourceStr(str);
+      size_t offset = 0;
+      for (int test = 1; test <= 100; ++test) {
+        size_t n = GetReadaheadSize() +
+                   rng.Uniform(static_cast<int>(GetReadaheadSize()));
+        if (s && test % 2) {
+          Skip(n);
+        } else {
+          ASSERT_EQ(str.substr(offset, std::min(n, strLen - offset)), Read(n));
+        }
+        offset = std::min(offset + n, strLen);
+      }
+    }
+  }
+}
+
+INSTANTIATE_TEST_CASE_P(
+    EmptySourceStr, ReadaheadSequentialFileTest,
+    ::testing::ValuesIn(ReadaheadSequentialFileTest::GetReadaheadSizeList()));
+INSTANTIATE_TEST_CASE_P(
+    SourceStrLenLessThanReadaheadSize, ReadaheadSequentialFileTest,
+    ::testing::ValuesIn(ReadaheadSequentialFileTest::GetReadaheadSizeList()));
+INSTANTIATE_TEST_CASE_P(
+    SourceStrLenGreaterThanReadaheadSize, ReadaheadSequentialFileTest,
+    ::testing::ValuesIn(ReadaheadSequentialFileTest::GetReadaheadSizeList()));
+INSTANTIATE_TEST_CASE_P(
+    ReadExceedsReadaheadSize, ReadaheadSequentialFileTest,
+    ::testing::ValuesIn(ReadaheadSequentialFileTest::GetReadaheadSizeList()));
+
+namespace {
+std::string GenerateLine(int n) {
+  std::string rv;
+  // Multiples of 17 characters per line, for likely bad buffer alignment
+  for (int i = 0; i < n; ++i) {
+    rv.push_back(static_cast<char>('0' + (i % 10)));
+    rv.append("xxxxxxxxxxxxxxxx");
+  }
+  return rv;
+}
+}  // namespace
+
+TEST(LineFileReaderTest, LineFileReaderTest) {
+  const int nlines = 1000;
+
+  std::unique_ptr<Env> mem_env(MockEnv::Create(Env::Default()));
+  std::shared_ptr<FileSystem> fs = mem_env->GetFileSystem();
+  // Create an input file
+  {
+    std::unique_ptr<FSWritableFile> file;
+    ASSERT_OK(
+        fs->NewWritableFile("testfile", FileOptions(), &file, /*dbg*/ nullptr));
+
+    for (int i = 0; i < nlines; ++i) {
+      std::string line = GenerateLine(i);
+      line.push_back('\n');
+      ASSERT_OK(file->Append(line, IOOptions(), /*dbg*/ nullptr));
+    }
+  }
+
+  // Verify with no I/O errors
+  {
+    std::unique_ptr<LineFileReader> reader;
+    ASSERT_OK(LineFileReader::Create(fs, "testfile", FileOptions(), &reader,
+                                     nullptr /* dbg */,
+                                     nullptr /* rate_limiter */));
+    std::string line;
+    int count = 0;
+    while (reader->ReadLine(&line, Env::IO_TOTAL /* rate_limiter_priority */)) {
+      ASSERT_EQ(line, GenerateLine(count));
+      ++count;
+      ASSERT_EQ(static_cast<int>(reader->GetLineNumber()), count);
+    }
+    ASSERT_OK(reader->GetStatus());
+    ASSERT_EQ(count, nlines);
+    ASSERT_EQ(static_cast<int>(reader->GetLineNumber()), count);
+    // And still
+    ASSERT_FALSE(
+        reader->ReadLine(&line, Env::IO_TOTAL /* rate_limiter_priority */));
+    ASSERT_OK(reader->GetStatus());
+    ASSERT_EQ(static_cast<int>(reader->GetLineNumber()), count);
+  }
+
+  // Verify with injected I/O error
+  {
+    std::unique_ptr<LineFileReader> reader;
+    ASSERT_OK(LineFileReader::Create(fs, "testfile", FileOptions(), &reader,
+                                     nullptr /* dbg */,
+                                     nullptr /* rate_limiter */));
+    std::string line;
+    int count = 0;
+    // Read part way through the file
+    while (count < nlines / 4) {
+      ASSERT_TRUE(
+          reader->ReadLine(&line, Env::IO_TOTAL /* rate_limiter_priority */));
+      ASSERT_EQ(line, GenerateLine(count));
+      ++count;
+      ASSERT_EQ(static_cast<int>(reader->GetLineNumber()), count);
+    }
+    ASSERT_OK(reader->GetStatus());
+
+    // Inject error
+    int callback_count = 0;
+    SyncPoint::GetInstance()->SetCallBack(
+        "MemFile::Read:IOStatus", [&](void* arg) {
+          IOStatus* status = static_cast<IOStatus*>(arg);
+          *status = IOStatus::Corruption("test");
+          ++callback_count;
+        });
+    SyncPoint::GetInstance()->EnableProcessing();
+
+    while (reader->ReadLine(&line, Env::IO_TOTAL /* rate_limiter_priority */)) {
+      ASSERT_EQ(line, GenerateLine(count));
+      ++count;
+      ASSERT_EQ(static_cast<int>(reader->GetLineNumber()), count);
+    }
+    ASSERT_TRUE(reader->GetStatus().IsCorruption());
+    ASSERT_LT(count, nlines / 2);
+    ASSERT_EQ(callback_count, 1);
+
+    // Still get error & no retry
+    ASSERT_FALSE(
+        reader->ReadLine(&line, Env::IO_TOTAL /* rate_limiter_priority */));
+    ASSERT_TRUE(reader->GetStatus().IsCorruption());
+    ASSERT_EQ(callback_count, 1);
+
+    SyncPoint::GetInstance()->DisableProcessing();
+    SyncPoint::GetInstance()->ClearAllCallBacks();
+  }
+}
+
+#ifndef ROCKSDB_LITE
+class IOErrorEventListener : public EventListener {
+ public:
+  IOErrorEventListener() { notify_error_.store(0); }
+
+  void OnIOError(const IOErrorInfo& io_error_info) override {
+    notify_error_++;
+    EXPECT_FALSE(io_error_info.file_path.empty());
+    EXPECT_FALSE(io_error_info.io_status.ok());
+  }
+
+  size_t NotifyErrorCount() { return notify_error_; }
+
+  bool ShouldBeNotifiedOnFileIO() override { return true; }
+
+ private:
+  std::atomic<size_t> notify_error_;
+};
+
+TEST_F(DBWritableFileWriterTest, IOErrorNotification) {
+  class FakeWF : public FSWritableFile {
+   public:
+    explicit FakeWF() : io_error_(false) {
+      file_append_errors_.store(0);
+      file_flush_errors_.store(0);
+    }
+
+    using FSWritableFile::Append;
+    IOStatus Append(const Slice& /*data*/, const IOOptions& /*options*/,
+                    IODebugContext* /*dbg*/) override {
+      if (io_error_) {
+        file_append_errors_++;
+        return IOStatus::IOError("Fake IO error");
+      }
+      return IOStatus::OK();
+    }
+
+    using FSWritableFile::PositionedAppend;
+    IOStatus PositionedAppend(const Slice& /*data*/, uint64_t,
+                              const IOOptions& /*options*/,
+                              IODebugContext* /*dbg*/) override {
+      if (io_error_) {
+        return IOStatus::IOError("Fake IO error");
+      }
+      return IOStatus::OK();
+    }
+    IOStatus Close(const IOOptions& /*options*/,
+                   IODebugContext* /*dbg*/) override {
+      return IOStatus::OK();
+    }
+    IOStatus Flush(const IOOptions& /*options*/,
+                   IODebugContext* /*dbg*/) override {
+      if (io_error_) {
+        file_flush_errors_++;
+        return IOStatus::IOError("Fake IO error");
+      }
+      return IOStatus::OK();
+    }
+    IOStatus Sync(const IOOptions& /*options*/,
+                  IODebugContext* /*dbg*/) override {
+      return IOStatus::OK();
+    }
+
+    void SetIOError(bool val) { io_error_ = val; }
+
+    void CheckCounters(int file_append_errors, int file_flush_errors) {
+      ASSERT_EQ(file_append_errors, file_append_errors_);
+      ASSERT_EQ(file_flush_errors_, file_flush_errors);
+    }
+
+   protected:
+    bool io_error_;
+    std::atomic<size_t> file_append_errors_;
+    std::atomic<size_t> file_flush_errors_;
+  };
+
+  FileOptions file_options = FileOptions();
+  Options options = GetDefaultOptions();
+  options.create_if_missing = true;
+  IOErrorEventListener* listener = new IOErrorEventListener();
+  options.listeners.emplace_back(listener);
+
+  DestroyAndReopen(options);
+  ImmutableOptions ioptions(options);
+
+  std::string fname = dbname_ + "/test_file";
+  std::unique_ptr<FakeWF> writable_file_ptr(new FakeWF);
+
+  std::unique_ptr<WritableFileWriter> file_writer;
+  writable_file_ptr->SetIOError(true);
+
+  file_writer.reset(new WritableFileWriter(
+      std::move(writable_file_ptr), fname, file_options,
+      SystemClock::Default().get(), nullptr, ioptions.stats, ioptions.listeners,
+      ioptions.file_checksum_gen_factory.get(), true, true));
+
+  FakeWF* fwf = static_cast<FakeWF*>(file_writer->writable_file());
+
+  fwf->SetIOError(true);
+  ASSERT_NOK(file_writer->Append(std::string(2 * kMb, 'a')));
+  fwf->CheckCounters(1, 0);
+  ASSERT_EQ(listener->NotifyErrorCount(), 1);
+
+  file_writer->reset_seen_error();
+  fwf->SetIOError(true);
+  ASSERT_NOK(file_writer->Flush());
+  fwf->CheckCounters(1, 1);
+  ASSERT_EQ(listener->NotifyErrorCount(), 2);
+
+  /* No error generation */
+  file_writer->reset_seen_error();
+  fwf->SetIOError(false);
+  ASSERT_OK(file_writer->Append(std::string(2 * kMb, 'b')));
+  ASSERT_EQ(listener->NotifyErrorCount(), 2);
+  fwf->CheckCounters(1, 1);
+}
+#endif  // ROCKSDB_LITE
+
+class WritableFileWriterIOPriorityTest : public testing::Test {
+ protected:
+  // This test is to check whether the rate limiter priority can be passed
+  // correctly from WritableFileWriter functions to FSWritableFile functions.
+
+  void SetUp() override {
+    // When op_rate_limiter_priority parameter in WritableFileWriter functions
+    // is the default (Env::IO_TOTAL).
+    std::unique_ptr<FakeWF> wf{new FakeWF(Env::IO_HIGH)};
+    FileOptions file_options;
+    writer_.reset(new WritableFileWriter(std::move(wf), "" /* don't care */,
+                                         file_options));
+  }
+
+  class FakeWF : public FSWritableFile {
+   public:
+    explicit FakeWF(Env::IOPriority io_priority) { SetIOPriority(io_priority); }
+    ~FakeWF() override {}
+
+    IOStatus Append(const Slice& /*data*/, const IOOptions& options,
+                    IODebugContext* /*dbg*/) override {
+      EXPECT_EQ(options.rate_limiter_priority, io_priority_);
+      return IOStatus::OK();
+    }
+    IOStatus Append(const Slice& data, const IOOptions& options,
+                    const DataVerificationInfo& /* verification_info */,
+                    IODebugContext* dbg) override {
+      return Append(data, options, dbg);
+    }
+    IOStatus PositionedAppend(const Slice& /*data*/, uint64_t /*offset*/,
+                              const IOOptions& options,
+                              IODebugContext* /*dbg*/) override {
+      EXPECT_EQ(options.rate_limiter_priority, io_priority_);
+      return IOStatus::OK();
+    }
+    IOStatus PositionedAppend(
+        const Slice& /* data */, uint64_t /* offset */,
+        const IOOptions& options,
+        const DataVerificationInfo& /* verification_info */,
+        IODebugContext* /*dbg*/) override {
+      EXPECT_EQ(options.rate_limiter_priority, io_priority_);
+      return IOStatus::OK();
+    }
+    IOStatus Truncate(uint64_t /*size*/, const IOOptions& options,
+                      IODebugContext* /*dbg*/) override {
+      EXPECT_EQ(options.rate_limiter_priority, io_priority_);
+      return IOStatus::OK();
+    }
+    IOStatus Close(const IOOptions& options, IODebugContext* /*dbg*/) override {
+      EXPECT_EQ(options.rate_limiter_priority, io_priority_);
+      return IOStatus::OK();
+    }
+    IOStatus Flush(const IOOptions& options, IODebugContext* /*dbg*/) override {
+      EXPECT_EQ(options.rate_limiter_priority, io_priority_);
+      return IOStatus::OK();
+    }
+    IOStatus Sync(const IOOptions& options, IODebugContext* /*dbg*/) override {
+      EXPECT_EQ(options.rate_limiter_priority, io_priority_);
+      return IOStatus::OK();
+    }
+    IOStatus Fsync(const IOOptions& options, IODebugContext* /*dbg*/) override {
+      EXPECT_EQ(options.rate_limiter_priority, io_priority_);
+      return IOStatus::OK();
+    }
+    uint64_t GetFileSize(const IOOptions& options,
+                         IODebugContext* /*dbg*/) override {
+      EXPECT_EQ(options.rate_limiter_priority, io_priority_);
+      return 0;
+    }
+    void GetPreallocationStatus(size_t* /*block_size*/,
+                                size_t* /*last_allocated_block*/) override {}
+    size_t GetUniqueId(char* /*id*/, size_t /*max_size*/) const override {
+      return 0;
+    }
+    IOStatus InvalidateCache(size_t /*offset*/, size_t /*length*/) override {
+      return IOStatus::OK();
+    }
+
+    IOStatus Allocate(uint64_t /*offset*/, uint64_t /*len*/,
+                      const IOOptions& options,
+                      IODebugContext* /*dbg*/) override {
+      EXPECT_EQ(options.rate_limiter_priority, io_priority_);
+      return IOStatus::OK();
+    }
+    IOStatus RangeSync(uint64_t /*offset*/, uint64_t /*nbytes*/,
+                       const IOOptions& options,
+                       IODebugContext* /*dbg*/) override {
+      EXPECT_EQ(options.rate_limiter_priority, io_priority_);
+      return IOStatus::OK();
+    }
+
+    void PrepareWrite(size_t /*offset*/, size_t /*len*/,
+                      const IOOptions& options,
+                      IODebugContext* /*dbg*/) override {
+      EXPECT_EQ(options.rate_limiter_priority, io_priority_);
+    }
+
+    bool IsSyncThreadSafe() const override { return true; }
+  };
+
+  std::unique_ptr<WritableFileWriter> writer_;
+};
+
+TEST_F(WritableFileWriterIOPriorityTest, Append) {
+  ASSERT_OK(writer_->Append(Slice("abc")));
+}
+
+TEST_F(WritableFileWriterIOPriorityTest, Pad) { ASSERT_OK(writer_->Pad(500)); }
+
+TEST_F(WritableFileWriterIOPriorityTest, Flush) { ASSERT_OK(writer_->Flush()); }
+
+TEST_F(WritableFileWriterIOPriorityTest, Close) { ASSERT_OK(writer_->Close()); }
+
+TEST_F(WritableFileWriterIOPriorityTest, Sync) {
+  ASSERT_OK(writer_->Sync(false));
+  ASSERT_OK(writer_->Sync(true));
+}
+
+TEST_F(WritableFileWriterIOPriorityTest, SyncWithoutFlush) {
+  ASSERT_OK(writer_->SyncWithoutFlush(false));
+  ASSERT_OK(writer_->SyncWithoutFlush(true));
+}
+
+TEST_F(WritableFileWriterIOPriorityTest, BasicOp) {
+  EnvOptions env_options;
+  env_options.bytes_per_sync = kMb;
+  std::unique_ptr<FakeWF> wf(new FakeWF(Env::IO_HIGH));
+  std::unique_ptr<WritableFileWriter> writer(
+      new WritableFileWriter(std::move(wf), "" /* don't care */, env_options));
+  Random r(301);
+  Status s;
+  std::unique_ptr<char[]> large_buf(new char[10 * kMb]);
+  for (int i = 0; i < 1000; i++) {
+    int skew_limit = (i < 700) ? 10 : 15;
+    uint32_t num = r.Skewed(skew_limit) * 100 + r.Uniform(100);
+    s = writer->Append(Slice(large_buf.get(), num));
+    ASSERT_OK(s);
+
+    // Flush in a chance of 1/10.
+    if (r.Uniform(10) == 0) {
+      s = writer->Flush();
+      ASSERT_OK(s);
+    }
+  }
+  s = writer->Close();
+  ASSERT_OK(s);
+}
+}  // namespace ROCKSDB_NAMESPACE
+
+int main(int argc, char** argv) {
+  ROCKSDB_NAMESPACE::port::InstallStackTraceHandler();
+  ::testing::InitGoogleTest(&argc, argv);
+  return RUN_ALL_TESTS();
+}
diff --git a/src/rocksdb/util/filelock_test.cc b/src/rocksdb/util/filelock_test.cc
new file mode 100644
index 000000000..69947a732
--- /dev/null
+++ b/src/rocksdb/util/filelock_test.cc
@@ -0,0 +1,148 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+//
+#include <fcntl.h>
+
+#include "rocksdb/env.h"
+#include "rocksdb/status.h"
+#ifdef __FreeBSD__
+#include <sys/types.h>
+#include <sys/wait.h>
+#endif
+#include <vector>
+
+#include "test_util/testharness.h"
+#include "util/coding.h"
+#include "util/string_util.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+class LockTest : public testing::Test {
+ public:
+  static LockTest* current_;
+  std::string file_;
+  ROCKSDB_NAMESPACE::Env* env_;
+
+  LockTest()
+      : file_(test::PerThreadDBPath("db_testlock_file")),
+        env_(ROCKSDB_NAMESPACE::Env::Default()) {
+    current_ = this;
+  }
+
+  ~LockTest() override {}
+
+  Status LockFile(FileLock** db_lock) { return env_->LockFile(file_, db_lock); }
+
+  Status UnlockFile(FileLock* db_lock) { return env_->UnlockFile(db_lock); }
+
+  bool AssertFileIsLocked() {
+    return CheckFileLock(/* lock_expected = */ true);
+  }
+
+  bool AssertFileIsNotLocked() {
+    return CheckFileLock(/* lock_expected = */ false);
+  }
+
+  bool CheckFileLock(bool lock_expected) {
+    // We need to fork to check the fcntl lock as we need
+    // to open and close the file from a different process
+    // to avoid either releasing the lock on close, or not
+    // contending for it when requesting a lock.
+
+#ifdef OS_WIN
+
+    // WaitForSingleObject and GetExitCodeProcess can do what waitpid does.
+    // TODO - implement on Windows
+    return true;
+
+#else
+
+    pid_t pid = fork();
+    if (0 == pid) {
+      // child process
+      int exit_val = EXIT_FAILURE;
+      int fd = open(file_.c_str(), O_RDWR | O_CREAT, 0644);
+      if (fd < 0) {
+        // could not open file, could not check if it was locked
+        fprintf(stderr, "Open on on file %s failed.\n", file_.c_str());
+        exit(exit_val);
+      }
+
+      struct flock f;
+      memset(&f, 0, sizeof(f));
+      f.l_type = (F_WRLCK);
+      f.l_whence = SEEK_SET;
+      f.l_start = 0;
+      f.l_len = 0;  // Lock/unlock entire file
+      int value = fcntl(fd, F_SETLK, &f);
+      if (value == -1) {
+        if (lock_expected) {
+          exit_val = EXIT_SUCCESS;
+        }
+      } else {
+        if (!lock_expected) {
+          exit_val = EXIT_SUCCESS;
+        }
+      }
+      close(fd);  // lock is released for child process
+      exit(exit_val);
+    } else if (pid > 0) {
+      // parent process
+      int status;
+      while (-1 == waitpid(pid, &status, 0))
+        ;
+      if (!WIFEXITED(status) || WEXITSTATUS(status) != 0) {
+        // child process exited with non success status
+        return false;
+      } else {
+        return true;
+      }
+    } else {
+      fprintf(stderr, "Fork failed\n");
+      return false;
+    }
+    return false;
+
+#endif
+  }
+};
+LockTest* LockTest::current_;
+
+TEST_F(LockTest, LockBySameThread) {
+  FileLock* lock1;
+  FileLock* lock2;
+
+  // acquire a lock on a file
+  ASSERT_OK(LockFile(&lock1));
+
+  // check the file is locked
+  ASSERT_TRUE(AssertFileIsLocked());
+
+  // re-acquire the lock on the same file. This should fail.
+  Status s = LockFile(&lock2);
+  ASSERT_TRUE(s.IsIOError());
+#ifndef OS_WIN
+  // Validate that error message contains current thread ID.
+  ASSERT_TRUE(s.ToString().find(std::to_string(
+                  Env::Default()->GetThreadID())) != std::string::npos);
+#endif
+
+  // check the file is locked
+  ASSERT_TRUE(AssertFileIsLocked());
+
+  // release the lock
+  ASSERT_OK(UnlockFile(lock1));
+
+  // check the file is not locked
+  ASSERT_TRUE(AssertFileIsNotLocked());
+}
+
+}  // namespace ROCKSDB_NAMESPACE
+
+int main(int argc, char** argv) {
+  ROCKSDB_NAMESPACE::port::InstallStackTraceHandler();
+  ::testing::InitGoogleTest(&argc, argv);
+  return RUN_ALL_TESTS();
+}
diff --git a/src/rocksdb/util/filter_bench.cc b/src/rocksdb/util/filter_bench.cc
new file mode 100644
index 000000000..93186cd08
--- /dev/null
+++ b/src/rocksdb/util/filter_bench.cc
@@ -0,0 +1,840 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#if !defined(GFLAGS) || defined(ROCKSDB_LITE)
+#include <cstdio>
+int main() {
+  fprintf(stderr, "filter_bench requires gflags and !ROCKSDB_LITE\n");
+  return 1;
+}
+#else
+
+#include <cinttypes>
+#include <iostream>
+#include <sstream>
+#include <vector>
+
+#include "memory/arena.h"
+#include "port/port.h"
+#include "port/stack_trace.h"
+#include "rocksdb/cache.h"
+#include "rocksdb/env.h"
+#include "rocksdb/system_clock.h"
+#include "rocksdb/table.h"
+#include "table/block_based/filter_policy_internal.h"
+#include "table/block_based/full_filter_block.h"
+#include "table/block_based/mock_block_based_table.h"
+#include "table/plain/plain_table_bloom.h"
+#include "util/cast_util.h"
+#include "util/gflags_compat.h"
+#include "util/hash.h"
+#include "util/random.h"
+#include "util/stderr_logger.h"
+#include "util/stop_watch.h"
+#include "util/string_util.h"
+
+using GFLAGS_NAMESPACE::ParseCommandLineFlags;
+using GFLAGS_NAMESPACE::RegisterFlagValidator;
+using GFLAGS_NAMESPACE::SetUsageMessage;
+
+DEFINE_uint32(seed, 0, "Seed for random number generators");
+
+DEFINE_double(working_mem_size_mb, 200,
+              "MB of memory to get up to among all filters, unless "
+              "m_keys_total_max is specified.");
+
+DEFINE_uint32(average_keys_per_filter, 10000,
+              "Average number of keys per filter");
+
+DEFINE_double(vary_key_count_ratio, 0.4,
+              "Vary number of keys by up to +/- vary_key_count_ratio * "
+              "average_keys_per_filter.");
+
+DEFINE_uint32(key_size, 24, "Average number of bytes for each key");
+
+DEFINE_bool(vary_key_alignment, true,
+            "Whether to vary key alignment (default: at least 32-bit "
+            "alignment)");
+
+DEFINE_uint32(vary_key_size_log2_interval, 5,
+              "Use same key size 2^n times, then change. Key size varies from "
+              "-2 to +2 bytes vs. average, unless n>=30 to fix key size.");
+
+DEFINE_uint32(batch_size, 8, "Number of keys to group in each batch");
+
+DEFINE_double(bits_per_key, 10.0, "Bits per key setting for filters");
+
+DEFINE_double(m_queries, 200, "Millions of queries for each test mode");
+
+DEFINE_double(m_keys_total_max, 0,
+              "Maximum total keys added to filters, in millions. "
+              "0 (default) disables. Non-zero overrides working_mem_size_mb "
+              "option.");
+
+DEFINE_bool(use_full_block_reader, false,
+            "Use FullFilterBlockReader interface rather than FilterBitsReader");
+
+DEFINE_bool(use_plain_table_bloom, false,
+            "Use PlainTableBloom structure and interface rather than "
+            "FilterBitsReader/FullFilterBlockReader");
+
+DEFINE_bool(new_builder, false,
+            "Whether to create a new builder for each new filter");
+
+DEFINE_uint32(impl, 0,
+              "Select filter implementation. Without -use_plain_table_bloom:"
+              "0 = legacy full Bloom filter, "
+              "1 = format_version 5 Bloom filter, 2 = Ribbon128 filter. With "
+              "-use_plain_table_bloom: 0 = no locality, 1 = locality.");
+
+DEFINE_bool(net_includes_hashing, false,
+            "Whether query net ns/op times should include hashing. "
+            "(if not, dry run will include hashing) "
+            "(build times always include hashing)");
+
+DEFINE_bool(optimize_filters_for_memory, false,
+            "Setting for BlockBasedTableOptions::optimize_filters_for_memory");
+
+DEFINE_bool(detect_filter_construct_corruption, false,
+            "Setting for "
+            "BlockBasedTableOptions::detect_filter_construct_corruption");
+
+DEFINE_uint32(block_cache_capacity_MB, 8,
+              "Setting for "
+              "LRUCacheOptions::capacity");
+
+DEFINE_bool(charge_filter_construction, false,
+            "Setting for "
+            "CacheEntryRoleOptions::charged of"
+            "CacheEntryRole::kFilterConstruction");
+
+DEFINE_bool(strict_capacity_limit, false,
+            "Setting for "
+            "LRUCacheOptions::strict_capacity_limit");
+
+DEFINE_bool(quick, false, "Run more limited set of tests, fewer queries");
+
+DEFINE_bool(best_case, false, "Run limited tests only for best-case");
+
+DEFINE_bool(allow_bad_fp_rate, false, "Continue even if FP rate is bad");
+
+DEFINE_bool(legend, false,
+            "Print more information about interpreting results instead of "
+            "running tests");
+
+DEFINE_uint32(runs, 1, "Number of times to rebuild and run benchmark tests");
+
+void _always_assert_fail(int line, const char *file, const char *expr) {
+  fprintf(stderr, "%s: %d: Assertion %s failed\n", file, line, expr);
+  abort();
+}
+
+#define ALWAYS_ASSERT(cond) \
+  ((cond) ? (void)0 : ::_always_assert_fail(__LINE__, __FILE__, #cond))
+
+#ifndef NDEBUG
+// This could affect build times enough that we should not include it for
+// accurate speed tests
+#define PREDICT_FP_RATE
+#endif
+
+using ROCKSDB_NAMESPACE::Arena;
+using ROCKSDB_NAMESPACE::BlockContents;
+using ROCKSDB_NAMESPACE::BloomFilterPolicy;
+using ROCKSDB_NAMESPACE::BloomHash;
+using ROCKSDB_NAMESPACE::BloomLikeFilterPolicy;
+using ROCKSDB_NAMESPACE::BuiltinFilterBitsBuilder;
+using ROCKSDB_NAMESPACE::CachableEntry;
+using ROCKSDB_NAMESPACE::Cache;
+using ROCKSDB_NAMESPACE::CacheEntryRole;
+using ROCKSDB_NAMESPACE::CacheEntryRoleOptions;
+using ROCKSDB_NAMESPACE::EncodeFixed32;
+using ROCKSDB_NAMESPACE::Env;
+using ROCKSDB_NAMESPACE::FastRange32;
+using ROCKSDB_NAMESPACE::FilterBitsReader;
+using ROCKSDB_NAMESPACE::FilterBuildingContext;
+using ROCKSDB_NAMESPACE::FilterPolicy;
+using ROCKSDB_NAMESPACE::FullFilterBlockReader;
+using ROCKSDB_NAMESPACE::GetSliceHash;
+using ROCKSDB_NAMESPACE::GetSliceHash64;
+using ROCKSDB_NAMESPACE::Lower32of64;
+using ROCKSDB_NAMESPACE::LRUCacheOptions;
+using ROCKSDB_NAMESPACE::ParsedFullFilterBlock;
+using ROCKSDB_NAMESPACE::PlainTableBloomV1;
+using ROCKSDB_NAMESPACE::Random32;
+using ROCKSDB_NAMESPACE::Slice;
+using ROCKSDB_NAMESPACE::static_cast_with_check;
+using ROCKSDB_NAMESPACE::Status;
+using ROCKSDB_NAMESPACE::StderrLogger;
+using ROCKSDB_NAMESPACE::mock::MockBlockBasedTableTester;
+
+struct KeyMaker {
+  KeyMaker(size_t avg_size)
+      : smallest_size_(avg_size -
+                       (FLAGS_vary_key_size_log2_interval >= 30 ? 2 : 0)),
+        buf_size_(avg_size + 11),  // pad to vary key size and alignment
+        buf_(new char[buf_size_]) {
+    memset(buf_.get(), 0, buf_size_);
+    assert(smallest_size_ > 8);
+  }
+  size_t smallest_size_;
+  size_t buf_size_;
+  std::unique_ptr<char[]> buf_;
+
+  // Returns a unique(-ish) key based on the given parameter values. Each
+  // call returns a Slice from the same buffer so previously returned
+  // Slices should be considered invalidated.
+  Slice Get(uint32_t filter_num, uint32_t val_num) {
+    size_t start = FLAGS_vary_key_alignment ? val_num % 4 : 0;
+    size_t len = smallest_size_;
+    if (FLAGS_vary_key_size_log2_interval < 30) {
+      // To get range [avg_size - 2, avg_size + 2]
+      // use range [smallest_size, smallest_size + 4]
+      len += FastRange32(
+          (val_num >> FLAGS_vary_key_size_log2_interval) * 1234567891, 5);
+    }
+    char *data = buf_.get() + start;
+    // Populate key data such that all data makes it into a key of at
+    // least 8 bytes. We also don't want all the within-filter key
+    // variance confined to a contiguous 32 bits, because then a 32 bit
+    // hash function can "cheat" the false positive rate by
+    // approximating a perfect hash.
+    EncodeFixed32(data, val_num);
+    EncodeFixed32(data + 4, filter_num + val_num);
+    // ensure clearing leftovers from different alignment
+    EncodeFixed32(data + 8, 0);
+    return Slice(data, len);
+  }
+};
+
+void PrintWarnings() {
+#if defined(__GNUC__) && !defined(__OPTIMIZE__)
+  fprintf(stdout,
+          "WARNING: Optimization is disabled: benchmarks unnecessarily slow\n");
+#endif
+#ifndef NDEBUG
+  fprintf(stdout,
+          "WARNING: Assertions are enabled; benchmarks unnecessarily slow\n");
+#endif
+}
+
+void PrintError(const char *error) { fprintf(stderr, "ERROR: %s\n", error); }
+
+struct FilterInfo {
+  uint32_t filter_id_ = 0;
+  std::unique_ptr<const char[]> owner_;
+  Slice filter_;
+  Status filter_construction_status = Status::OK();
+  uint32_t keys_added_ = 0;
+  std::unique_ptr<FilterBitsReader> reader_;
+  std::unique_ptr<FullFilterBlockReader> full_block_reader_;
+  std::unique_ptr<PlainTableBloomV1> plain_table_bloom_;
+  uint64_t outside_queries_ = 0;
+  uint64_t false_positives_ = 0;
+};
+
+enum TestMode {
+  kSingleFilter,
+  kBatchPrepared,
+  kBatchUnprepared,
+  kFiftyOneFilter,
+  kEightyTwentyFilter,
+  kRandomFilter,
+};
+
+static const std::vector<TestMode> allTestModes = {
+    kSingleFilter,   kBatchPrepared,      kBatchUnprepared,
+    kFiftyOneFilter, kEightyTwentyFilter, kRandomFilter,
+};
+
+static const std::vector<TestMode> quickTestModes = {
+    kSingleFilter,
+    kRandomFilter,
+};
+
+static const std::vector<TestMode> bestCaseTestModes = {
+    kSingleFilter,
+};
+
+const char *TestModeToString(TestMode tm) {
+  switch (tm) {
+    case kSingleFilter:
+      return "Single filter";
+    case kBatchPrepared:
+      return "Batched, prepared";
+    case kBatchUnprepared:
+      return "Batched, unprepared";
+    case kFiftyOneFilter:
+      return "Skewed 50% in 1%";
+    case kEightyTwentyFilter:
+      return "Skewed 80% in 20%";
+    case kRandomFilter:
+      return "Random filter";
+  }
+  return "Bad TestMode";
+}
+
+// Do just enough to keep some data dependence for the
+// compiler / CPU
+static uint32_t DryRunNoHash(Slice &s) {
+  uint32_t sz = static_cast<uint32_t>(s.size());
+  if (sz >= 4) {
+    return sz + s.data()[3];
+  } else {
+    return sz;
+  }
+}
+
+static uint32_t DryRunHash32(Slice &s) {
+  // Same perf characteristics as GetSliceHash()
+  return BloomHash(s);
+}
+
+static uint32_t DryRunHash64(Slice &s) {
+  return Lower32of64(GetSliceHash64(s));
+}
+
+const std::shared_ptr<const FilterPolicy> &GetPolicy() {
+  static std::shared_ptr<const FilterPolicy> policy;
+  if (!policy) {
+    policy = BloomLikeFilterPolicy::Create(
+        BloomLikeFilterPolicy::GetAllFixedImpls().at(FLAGS_impl),
+        FLAGS_bits_per_key);
+  }
+  return policy;
+}
+
+struct FilterBench : public MockBlockBasedTableTester {
+  std::vector<KeyMaker> kms_;
+  std::vector<FilterInfo> infos_;
+  Random32 random_;
+  std::ostringstream fp_rate_report_;
+  Arena arena_;
+  double m_queries_;
+  StderrLogger stderr_logger_;
+
+  FilterBench()
+      : MockBlockBasedTableTester(GetPolicy()),
+        random_(FLAGS_seed),
+        m_queries_(0) {
+    for (uint32_t i = 0; i < FLAGS_batch_size; ++i) {
+      kms_.emplace_back(FLAGS_key_size < 8 ? 8 : FLAGS_key_size);
+    }
+    ioptions_.logger = &stderr_logger_;
+    table_options_.optimize_filters_for_memory =
+        FLAGS_optimize_filters_for_memory;
+    table_options_.detect_filter_construct_corruption =
+        FLAGS_detect_filter_construct_corruption;
+    table_options_.cache_usage_options.options_overrides.insert(
+        {CacheEntryRole::kFilterConstruction,
+         {/*.charged = */ FLAGS_charge_filter_construction
+              ? CacheEntryRoleOptions::Decision::kEnabled
+              : CacheEntryRoleOptions::Decision::kDisabled}});
+    if (FLAGS_charge_filter_construction) {
+      table_options_.no_block_cache = false;
+      LRUCacheOptions lo;
+      lo.capacity = FLAGS_block_cache_capacity_MB * 1024 * 1024;
+      lo.num_shard_bits = 0;  // 2^0 shard
+      lo.strict_capacity_limit = FLAGS_strict_capacity_limit;
+      std::shared_ptr<Cache> cache(NewLRUCache(lo));
+      table_options_.block_cache = cache;
+    }
+  }
+
+  void Go();
+
+  double RandomQueryTest(uint32_t inside_threshold, bool dry_run,
+                         TestMode mode);
+};
+
+void FilterBench::Go() {
+  if (FLAGS_use_plain_table_bloom && FLAGS_use_full_block_reader) {
+    throw std::runtime_error(
+        "Can't combine -use_plain_table_bloom and -use_full_block_reader");
+  }
+  if (FLAGS_use_plain_table_bloom) {
+    if (FLAGS_impl > 1) {
+      throw std::runtime_error(
+          "-impl must currently be >= 0 and <= 1 for Plain table");
+    }
+  } else {
+    if (FLAGS_impl > 2) {
+      throw std::runtime_error(
+          "-impl must currently be >= 0 and <= 2 for Block-based table");
+    }
+  }
+
+  if (FLAGS_vary_key_count_ratio < 0.0 || FLAGS_vary_key_count_ratio > 1.0) {
+    throw std::runtime_error("-vary_key_count_ratio must be >= 0.0 and <= 1.0");
+  }
+
+  // For example, average_keys_per_filter = 100, vary_key_count_ratio = 0.1.
+  // Varys up to +/- 10 keys. variance_range = 21 (generating value 0..20).
+  // variance_offset = 10, so value - offset average value is always 0.
+  const uint32_t variance_range =
+      1 + 2 * static_cast<uint32_t>(FLAGS_vary_key_count_ratio *
+                                    FLAGS_average_keys_per_filter);
+  const uint32_t variance_offset = variance_range / 2;
+
+  const std::vector<TestMode> &testModes = FLAGS_best_case ? bestCaseTestModes
+                                           : FLAGS_quick   ? quickTestModes
+                                                           : allTestModes;
+
+  m_queries_ = FLAGS_m_queries;
+  double working_mem_size_mb = FLAGS_working_mem_size_mb;
+  if (FLAGS_quick) {
+    m_queries_ /= 7.0;
+  } else if (FLAGS_best_case) {
+    m_queries_ /= 3.0;
+    working_mem_size_mb /= 10.0;
+  }
+
+  std::cout << "Building..." << std::endl;
+
+  std::unique_ptr<BuiltinFilterBitsBuilder> builder;
+
+  size_t total_memory_used = 0;
+  size_t total_size = 0;
+  size_t total_keys_added = 0;
+#ifdef PREDICT_FP_RATE
+  double weighted_predicted_fp_rate = 0.0;
+#endif
+  size_t max_total_keys;
+  size_t max_mem;
+  if (FLAGS_m_keys_total_max > 0) {
+    max_total_keys = static_cast<size_t>(1000000 * FLAGS_m_keys_total_max);
+    max_mem = SIZE_MAX;
+  } else {
+    max_total_keys = SIZE_MAX;
+    max_mem = static_cast<size_t>(1024 * 1024 * working_mem_size_mb);
+  }
+
+  ROCKSDB_NAMESPACE::StopWatchNano timer(
+      ROCKSDB_NAMESPACE::SystemClock::Default().get(), true);
+
+  infos_.clear();
+  while ((working_mem_size_mb == 0 || total_size < max_mem) &&
+         total_keys_added < max_total_keys) {
+    uint32_t filter_id = random_.Next();
+    uint32_t keys_to_add = FLAGS_average_keys_per_filter +
+                           FastRange32(random_.Next(), variance_range) -
+                           variance_offset;
+    if (max_total_keys - total_keys_added < keys_to_add) {
+      keys_to_add = static_cast<uint32_t>(max_total_keys - total_keys_added);
+    }
+    infos_.emplace_back();
+    FilterInfo &info = infos_.back();
+    info.filter_id_ = filter_id;
+    info.keys_added_ = keys_to_add;
+    if (FLAGS_use_plain_table_bloom) {
+      info.plain_table_bloom_.reset(new PlainTableBloomV1());
+      info.plain_table_bloom_->SetTotalBits(
+          &arena_, static_cast<uint32_t>(keys_to_add * FLAGS_bits_per_key),
+          FLAGS_impl, 0 /*huge_page*/, nullptr /*logger*/);
+      for (uint32_t i = 0; i < keys_to_add; ++i) {
+        uint32_t hash = GetSliceHash(kms_[0].Get(filter_id, i));
+        info.plain_table_bloom_->AddHash(hash);
+      }
+      info.filter_ = info.plain_table_bloom_->GetRawData();
+    } else {
+      if (!builder) {
+        builder.reset(
+            static_cast_with_check<BuiltinFilterBitsBuilder>(GetBuilder()));
+      }
+      for (uint32_t i = 0; i < keys_to_add; ++i) {
+        builder->AddKey(kms_[0].Get(filter_id, i));
+      }
+      info.filter_ =
+          builder->Finish(&info.owner_, &info.filter_construction_status);
+      if (info.filter_construction_status.ok()) {
+        info.filter_construction_status =
+            builder->MaybePostVerify(info.filter_);
+      }
+      if (!info.filter_construction_status.ok()) {
+        PrintError(info.filter_construction_status.ToString().c_str());
+      }
+#ifdef PREDICT_FP_RATE
+      weighted_predicted_fp_rate +=
+          keys_to_add *
+          builder->EstimatedFpRate(keys_to_add, info.filter_.size());
+#endif
+      if (FLAGS_new_builder) {
+        builder.reset();
+      }
+      info.reader_.reset(
+          table_options_.filter_policy->GetFilterBitsReader(info.filter_));
+      CachableEntry<ParsedFullFilterBlock> block(
+          new ParsedFullFilterBlock(table_options_.filter_policy.get(),
+                                    BlockContents(info.filter_)),
+          nullptr /* cache */, nullptr /* cache_handle */,
+          true /* own_value */);
+      info.full_block_reader_.reset(
+          new FullFilterBlockReader(table_.get(), std::move(block)));
+    }
+    total_size += info.filter_.size();
+#ifdef ROCKSDB_MALLOC_USABLE_SIZE
+    total_memory_used +=
+        malloc_usable_size(const_cast<char *>(info.filter_.data()));
+#endif  // ROCKSDB_MALLOC_USABLE_SIZE
+    total_keys_added += keys_to_add;
+  }
+
+  uint64_t elapsed_nanos = timer.ElapsedNanos();
+  double ns = double(elapsed_nanos) / total_keys_added;
+  std::cout << "Build avg ns/key: " << ns << std::endl;
+  std::cout << "Number of filters: " << infos_.size() << std::endl;
+  std::cout << "Total size (MB): " << total_size / 1024.0 / 1024.0 << std::endl;
+  if (total_memory_used > 0) {
+    std::cout << "Reported total allocated memory (MB): "
+              << total_memory_used / 1024.0 / 1024.0 << std::endl;
+    std::cout << "Reported internal fragmentation: "
+              << (total_memory_used - total_size) * 100.0 / total_size << "%"
+              << std::endl;
+  }
+
+  double bpk = total_size * 8.0 / total_keys_added;
+  std::cout << "Bits/key stored: " << bpk << std::endl;
+#ifdef PREDICT_FP_RATE
+  std::cout << "Predicted FP rate %: "
+            << 100.0 * (weighted_predicted_fp_rate / total_keys_added)
+            << std::endl;
+#endif
+  if (!FLAGS_quick && !FLAGS_best_case) {
+    double tolerable_rate = std::pow(2.0, -(bpk - 1.0) / (1.4 + bpk / 50.0));
+    std::cout << "Best possible FP rate %: " << 100.0 * std::pow(2.0, -bpk)
+              << std::endl;
+    std::cout << "Tolerable FP rate %: " << 100.0 * tolerable_rate << std::endl;
+
+    std::cout << "----------------------------" << std::endl;
+    std::cout << "Verifying..." << std::endl;
+
+    uint32_t outside_q_per_f =
+        static_cast<uint32_t>(m_queries_ * 1000000 / infos_.size());
+    uint64_t fps = 0;
+    for (uint32_t i = 0; i < infos_.size(); ++i) {
+      FilterInfo &info = infos_[i];
+      for (uint32_t j = 0; j < info.keys_added_; ++j) {
+        if (FLAGS_use_plain_table_bloom) {
+          uint32_t hash = GetSliceHash(kms_[0].Get(info.filter_id_, j));
+          ALWAYS_ASSERT(info.plain_table_bloom_->MayContainHash(hash));
+        } else {
+          ALWAYS_ASSERT(
+              info.reader_->MayMatch(kms_[0].Get(info.filter_id_, j)));
+        }
+      }
+      for (uint32_t j = 0; j < outside_q_per_f; ++j) {
+        if (FLAGS_use_plain_table_bloom) {
+          uint32_t hash =
+              GetSliceHash(kms_[0].Get(info.filter_id_, j | 0x80000000));
+          fps += info.plain_table_bloom_->MayContainHash(hash);
+        } else {
+          fps += info.reader_->MayMatch(
+              kms_[0].Get(info.filter_id_, j | 0x80000000));
+        }
+      }
+    }
+    std::cout << " No FNs :)" << std::endl;
+    double prelim_rate = double(fps) / outside_q_per_f / infos_.size();
+    std::cout << " Prelim FP rate %: " << (100.0 * prelim_rate) << std::endl;
+
+    if (!FLAGS_allow_bad_fp_rate) {
+      ALWAYS_ASSERT(prelim_rate < tolerable_rate);
+    }
+  }
+
+  std::cout << "----------------------------" << std::endl;
+  std::cout << "Mixed inside/outside queries..." << std::endl;
+  // 50% each inside and outside
+  uint32_t inside_threshold = UINT32_MAX / 2;
+  for (TestMode tm : testModes) {
+    random_.Seed(FLAGS_seed + 1);
+    double f = RandomQueryTest(inside_threshold, /*dry_run*/ false, tm);
+    random_.Seed(FLAGS_seed + 1);
+    double d = RandomQueryTest(inside_threshold, /*dry_run*/ true, tm);
+    std::cout << "  " << TestModeToString(tm) << " net ns/op: " << (f - d)
+              << std::endl;
+  }
+
+  if (!FLAGS_quick) {
+    std::cout << "----------------------------" << std::endl;
+    std::cout << "Inside queries (mostly)..." << std::endl;
+    // Do about 95% inside queries rather than 100% so that branch predictor
+    // can't give itself an artifically crazy advantage.
+    inside_threshold = UINT32_MAX / 20 * 19;
+    for (TestMode tm : testModes) {
+      random_.Seed(FLAGS_seed + 1);
+      double f = RandomQueryTest(inside_threshold, /*dry_run*/ false, tm);
+      random_.Seed(FLAGS_seed + 1);
+      double d = RandomQueryTest(inside_threshold, /*dry_run*/ true, tm);
+      std::cout << "  " << TestModeToString(tm) << " net ns/op: " << (f - d)
+                << std::endl;
+    }
+
+    std::cout << "----------------------------" << std::endl;
+    std::cout << "Outside queries (mostly)..." << std::endl;
+    // Do about 95% outside queries rather than 100% so that branch predictor
+    // can't give itself an artifically crazy advantage.
+    inside_threshold = UINT32_MAX / 20;
+    for (TestMode tm : testModes) {
+      random_.Seed(FLAGS_seed + 2);
+      double f = RandomQueryTest(inside_threshold, /*dry_run*/ false, tm);
+      random_.Seed(FLAGS_seed + 2);
+      double d = RandomQueryTest(inside_threshold, /*dry_run*/ true, tm);
+      std::cout << "  " << TestModeToString(tm) << " net ns/op: " << (f - d)
+                << std::endl;
+    }
+  }
+  std::cout << fp_rate_report_.str();
+
+  std::cout << "----------------------------" << std::endl;
+  std::cout << "Done. (For more info, run with -legend or -help.)" << std::endl;
+}
+
+double FilterBench::RandomQueryTest(uint32_t inside_threshold, bool dry_run,
+                                    TestMode mode) {
+  for (auto &info : infos_) {
+    info.outside_queries_ = 0;
+    info.false_positives_ = 0;
+  }
+
+  auto dry_run_hash_fn = DryRunNoHash;
+  if (!FLAGS_net_includes_hashing) {
+    if (FLAGS_impl == 0 || FLAGS_use_plain_table_bloom) {
+      dry_run_hash_fn = DryRunHash32;
+    } else {
+      dry_run_hash_fn = DryRunHash64;
+    }
+  }
+
+  uint32_t num_infos = static_cast<uint32_t>(infos_.size());
+  uint32_t dry_run_hash = 0;
+  uint64_t max_queries = static_cast<uint64_t>(m_queries_ * 1000000 + 0.50);
+  // Some filters may be considered secondary in order to implement skewed
+  // queries. num_primary_filters is the number that are to be treated as
+  // equal, and any remainder will be treated as secondary.
+  uint32_t num_primary_filters = num_infos;
+  // The proportion (when divided by 2^32 - 1) of filter queries going to
+  // the primary filters (default = all). The remainder of queries are
+  // against secondary filters.
+  uint32_t primary_filter_threshold = 0xffffffff;
+  if (mode == kSingleFilter) {
+    // 100% of queries to 1 filter
+    num_primary_filters = 1;
+  } else if (mode == kFiftyOneFilter) {
+    if (num_infos < 50) {
+      return 0.0;  // skip
+    }
+    // 50% of queries
+    primary_filter_threshold /= 2;
+    // to 1% of filters
+    num_primary_filters = (num_primary_filters + 99) / 100;
+  } else if (mode == kEightyTwentyFilter) {
+    if (num_infos < 5) {
+      return 0.0;  // skip
+    }
+    // 80% of queries
+    primary_filter_threshold = primary_filter_threshold / 5 * 4;
+    // to 20% of filters
+    num_primary_filters = (num_primary_filters + 4) / 5;
+  } else if (mode == kRandomFilter) {
+    if (num_infos == 1) {
+      return 0.0;  // skip
+    }
+  }
+  uint32_t batch_size = 1;
+  std::unique_ptr<Slice[]> batch_slices;
+  std::unique_ptr<Slice *[]> batch_slice_ptrs;
+  std::unique_ptr<bool[]> batch_results;
+  if (mode == kBatchPrepared || mode == kBatchUnprepared) {
+    batch_size = static_cast<uint32_t>(kms_.size());
+  }
+
+  batch_slices.reset(new Slice[batch_size]);
+  batch_slice_ptrs.reset(new Slice *[batch_size]);
+  batch_results.reset(new bool[batch_size]);
+  for (uint32_t i = 0; i < batch_size; ++i) {
+    batch_results[i] = false;
+    batch_slice_ptrs[i] = &batch_slices[i];
+  }
+
+  ROCKSDB_NAMESPACE::StopWatchNano timer(
+      ROCKSDB_NAMESPACE::SystemClock::Default().get(), true);
+
+  for (uint64_t q = 0; q < max_queries; q += batch_size) {
+    bool inside_this_time = random_.Next() <= inside_threshold;
+
+    uint32_t filter_index;
+    if (random_.Next() <= primary_filter_threshold) {
+      filter_index = random_.Uniformish(num_primary_filters);
+    } else {
+      // secondary
+      filter_index = num_primary_filters +
+                     random_.Uniformish(num_infos - num_primary_filters);
+    }
+    FilterInfo &info = infos_[filter_index];
+    for (uint32_t i = 0; i < batch_size; ++i) {
+      if (inside_this_time) {
+        batch_slices[i] =
+            kms_[i].Get(info.filter_id_, random_.Uniformish(info.keys_added_));
+      } else {
+        batch_slices[i] =
+            kms_[i].Get(info.filter_id_, random_.Uniformish(info.keys_added_) |
+                                             uint32_t{0x80000000});
+        info.outside_queries_++;
+      }
+    }
+    // TODO: implement batched interface to full block reader
+    // TODO: implement batched interface to plain table bloom
+    if (mode == kBatchPrepared && !FLAGS_use_full_block_reader &&
+        !FLAGS_use_plain_table_bloom) {
+      for (uint32_t i = 0; i < batch_size; ++i) {
+        batch_results[i] = false;
+      }
+      if (dry_run) {
+        for (uint32_t i = 0; i < batch_size; ++i) {
+          batch_results[i] = true;
+          dry_run_hash += dry_run_hash_fn(batch_slices[i]);
+        }
+      } else {
+        info.reader_->MayMatch(batch_size, batch_slice_ptrs.get(),
+                               batch_results.get());
+      }
+      for (uint32_t i = 0; i < batch_size; ++i) {
+        if (inside_this_time) {
+          ALWAYS_ASSERT(batch_results[i]);
+        } else {
+          info.false_positives_ += batch_results[i];
+        }
+      }
+    } else {
+      for (uint32_t i = 0; i < batch_size; ++i) {
+        bool may_match;
+        if (FLAGS_use_plain_table_bloom) {
+          if (dry_run) {
+            dry_run_hash += dry_run_hash_fn(batch_slices[i]);
+            may_match = true;
+          } else {
+            uint32_t hash = GetSliceHash(batch_slices[i]);
+            may_match = info.plain_table_bloom_->MayContainHash(hash);
+          }
+        } else if (FLAGS_use_full_block_reader) {
+          if (dry_run) {
+            dry_run_hash += dry_run_hash_fn(batch_slices[i]);
+            may_match = true;
+          } else {
+            may_match = info.full_block_reader_->KeyMayMatch(
+                batch_slices[i],
+                /*no_io=*/false, /*const_ikey_ptr=*/nullptr,
+                /*get_context=*/nullptr,
+                /*lookup_context=*/nullptr, Env::IO_TOTAL);
+          }
+        } else {
+          if (dry_run) {
+            dry_run_hash += dry_run_hash_fn(batch_slices[i]);
+            may_match = true;
+          } else {
+            may_match = info.reader_->MayMatch(batch_slices[i]);
+          }
+        }
+        if (inside_this_time) {
+          ALWAYS_ASSERT(may_match);
+        } else {
+          info.false_positives_ += may_match;
+        }
+      }
+    }
+  }
+
+  uint64_t elapsed_nanos = timer.ElapsedNanos();
+  double ns = double(elapsed_nanos) / max_queries;
+
+  if (!FLAGS_quick) {
+    if (dry_run) {
+      // Printing part of hash prevents dry run components from being optimized
+      // away by compiler
+      std::cout << "    Dry run (" << std::hex << (dry_run_hash & 0xfffff)
+                << std::dec << ") ";
+    } else {
+      std::cout << "    Gross filter    ";
+    }
+    std::cout << "ns/op: " << ns << std::endl;
+  }
+
+  if (!dry_run) {
+    fp_rate_report_.str("");
+    uint64_t q = 0;
+    uint64_t fp = 0;
+    double worst_fp_rate = 0.0;
+    double best_fp_rate = 1.0;
+    for (auto &info : infos_) {
+      q += info.outside_queries_;
+      fp += info.false_positives_;
+      if (info.outside_queries_ > 0) {
+        double fp_rate = double(info.false_positives_) / info.outside_queries_;
+        worst_fp_rate = std::max(worst_fp_rate, fp_rate);
+        best_fp_rate = std::min(best_fp_rate, fp_rate);
+      }
+    }
+    fp_rate_report_ << "    Average FP rate %: " << 100.0 * fp / q << std::endl;
+    if (!FLAGS_quick && !FLAGS_best_case) {
+      fp_rate_report_ << "    Worst   FP rate %: " << 100.0 * worst_fp_rate
+                      << std::endl;
+      fp_rate_report_ << "    Best    FP rate %: " << 100.0 * best_fp_rate
+                      << std::endl;
+      fp_rate_report_ << "    Best possible bits/key: "
+                      << -std::log(double(fp) / q) / std::log(2.0) << std::endl;
+    }
+  }
+  return ns;
+}
+
+int main(int argc, char **argv) {
+  ROCKSDB_NAMESPACE::port::InstallStackTraceHandler();
+  SetUsageMessage(std::string("\nUSAGE:\n") + std::string(argv[0]) +
+                  " [-quick] [OTHER OPTIONS]...");
+  ParseCommandLineFlags(&argc, &argv, true);
+
+  PrintWarnings();
+
+  if (FLAGS_legend) {
+    std::cout
+        << "Legend:" << std::endl
+        << "  \"Inside\" - key that was added to filter" << std::endl
+        << "  \"Outside\" - key that was not added to filter" << std::endl
+        << "  \"FN\" - false negative query (must not happen)" << std::endl
+        << "  \"FP\" - false positive query (OK at low rate)" << std::endl
+        << "  \"Dry run\" - cost of testing and hashing overhead." << std::endl
+        << "  \"Gross filter\" - cost of filter queries including testing "
+        << "\n     and hashing overhead." << std::endl
+        << "  \"net\" - best estimate of time in filter operation, without "
+        << "\n     testing and hashing overhead (gross filter - dry run)"
+        << std::endl
+        << "  \"ns/op\" - nanoseconds per operation (key query or add)"
+        << std::endl
+        << "  \"Single filter\" - essentially minimum cost, assuming filter"
+        << "\n     fits easily in L1 CPU cache." << std::endl
+        << "  \"Batched, prepared\" - several queries at once against a"
+        << "\n     randomly chosen filter, using multi-query interface."
+        << std::endl
+        << "  \"Batched, unprepared\" - similar, but using serial calls"
+        << "\n     to single query interface." << std::endl
+        << "  \"Random filter\" - a filter is chosen at random as target"
+        << "\n     of each query." << std::endl
+        << "  \"Skewed X% in Y%\" - like \"Random filter\" except Y% of"
+        << "\n      the filters are designated as \"hot\" and receive X%"
+        << "\n      of queries." << std::endl;
+  } else {
+    FilterBench b;
+    for (uint32_t i = 0; i < FLAGS_runs; ++i) {
+      b.Go();
+      FLAGS_seed += 100;
+      b.random_.Seed(FLAGS_seed);
+    }
+  }
+
+  return 0;
+}
+
+#endif  // !defined(GFLAGS) || defined(ROCKSDB_LITE)
diff --git a/src/rocksdb/util/gflags_compat.h b/src/rocksdb/util/gflags_compat.h
new file mode 100644
index 000000000..b6f88a5bc
--- /dev/null
+++ b/src/rocksdb/util/gflags_compat.h
@@ -0,0 +1,30 @@
+//  Copyright (c) 2017-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#pragma once
+#include <gflags/gflags.h>
+
+#include <functional>
+
+#ifndef GFLAGS_NAMESPACE
+// in case it's not defined in old versions, that's probably because it was
+// still google by default.
+#define GFLAGS_NAMESPACE google
+#endif
+
+#ifndef DEFINE_uint32
+// DEFINE_uint32 does not appear in older versions of gflags. This should be
+// a sane definition for those versions.
+#include <cstdint>
+#define DEFINE_uint32(name, val, txt)             \
+  namespace gflags_compat {                       \
+  DEFINE_int32(name, val, txt);                   \
+  }                                               \
+  std::reference_wrapper<uint32_t> FLAGS_##name = \
+      std::ref(*reinterpret_cast<uint32_t *>(&gflags_compat::FLAGS_##name));
+
+#define DECLARE_uint32(name) \
+  extern std::reference_wrapper<uint32_t> FLAGS_##name;
+#endif  // !DEFINE_uint32
diff --git a/src/rocksdb/util/hash.cc b/src/rocksdb/util/hash.cc
new file mode 100644
index 000000000..0f7f2edc1
--- /dev/null
+++ b/src/rocksdb/util/hash.cc
@@ -0,0 +1,201 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+//
+// Copyright (c) 2011 The LevelDB Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file. See the AUTHORS file for names of contributors.
+
+#include "util/hash.h"
+
+#include <string>
+
+#include "port/lang.h"
+#include "util/coding.h"
+#include "util/hash128.h"
+#include "util/math128.h"
+#include "util/xxhash.h"
+#include "util/xxph3.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+uint64_t (*kGetSliceNPHash64UnseededFnPtr)(const Slice&) = &GetSliceHash64;
+
+uint32_t Hash(const char* data, size_t n, uint32_t seed) {
+  // MurmurHash1 - fast but mediocre quality
+  // https://github.com/aappleby/smhasher/wiki/MurmurHash1
+  //
+  const uint32_t m = 0xc6a4a793;
+  const uint32_t r = 24;
+  const char* limit = data + n;
+  uint32_t h = static_cast<uint32_t>(seed ^ (n * m));
+
+  // Pick up four bytes at a time
+  while (data + 4 <= limit) {
+    uint32_t w = DecodeFixed32(data);
+    data += 4;
+    h += w;
+    h *= m;
+    h ^= (h >> 16);
+  }
+
+  // Pick up remaining bytes
+  switch (limit - data) {
+    // Note: The original hash implementation used data[i] << shift, which
+    // promotes the char to int and then performs the shift. If the char is
+    // negative, the shift is undefined behavior in C++. The hash algorithm is
+    // part of the format definition, so we cannot change it; to obtain the same
+    // behavior in a legal way we just cast to uint32_t, which will do
+    // sign-extension. To guarantee compatibility with architectures where chars
+    // are unsigned we first cast the char to int8_t.
+    case 3:
+      h += static_cast<uint32_t>(static_cast<int8_t>(data[2])) << 16;
+      FALLTHROUGH_INTENDED;
+    case 2:
+      h += static_cast<uint32_t>(static_cast<int8_t>(data[1])) << 8;
+      FALLTHROUGH_INTENDED;
+    case 1:
+      h += static_cast<uint32_t>(static_cast<int8_t>(data[0]));
+      h *= m;
+      h ^= (h >> r);
+      break;
+  }
+  return h;
+}
+
+// We are standardizing on a preview release of XXH3, because that's
+// the best available at time of standardizing.
+//
+// In testing (mostly Intel Skylake), this hash function is much more
+// thorough than Hash32 and is almost universally faster. Hash() only
+// seems faster when passing runtime-sized keys of the same small size
+// (less than about 24 bytes) thousands of times in a row; this seems
+// to allow the branch predictor to work some magic. XXH3's speed is
+// much less dependent on branch prediction.
+//
+// Hashing with a prefix extractor is potentially a common case of
+// hashing objects of small, predictable size. We could consider
+// bundling hash functions specialized for particular lengths with
+// the prefix extractors.
+uint64_t Hash64(const char* data, size_t n, uint64_t seed) {
+  return XXPH3_64bits_withSeed(data, n, seed);
+}
+
+uint64_t Hash64(const char* data, size_t n) {
+  // Same as seed = 0
+  return XXPH3_64bits(data, n);
+}
+
+uint64_t GetSlicePartsNPHash64(const SliceParts& data, uint64_t seed) {
+  // TODO(ajkr): use XXH3 streaming APIs to avoid the copy/allocation.
+  size_t concat_len = 0;
+  for (int i = 0; i < data.num_parts; ++i) {
+    concat_len += data.parts[i].size();
+  }
+  std::string concat_data;
+  concat_data.reserve(concat_len);
+  for (int i = 0; i < data.num_parts; ++i) {
+    concat_data.append(data.parts[i].data(), data.parts[i].size());
+  }
+  assert(concat_data.size() == concat_len);
+  return NPHash64(concat_data.data(), concat_len, seed);
+}
+
+Unsigned128 Hash128(const char* data, size_t n, uint64_t seed) {
+  auto h = XXH3_128bits_withSeed(data, n, seed);
+  return (Unsigned128{h.high64} << 64) | (h.low64);
+}
+
+Unsigned128 Hash128(const char* data, size_t n) {
+  // Same as seed = 0
+  auto h = XXH3_128bits(data, n);
+  return (Unsigned128{h.high64} << 64) | (h.low64);
+}
+
+void Hash2x64(const char* data, size_t n, uint64_t* high64, uint64_t* low64) {
+  // Same as seed = 0
+  auto h = XXH3_128bits(data, n);
+  *high64 = h.high64;
+  *low64 = h.low64;
+}
+
+void Hash2x64(const char* data, size_t n, uint64_t seed, uint64_t* high64,
+              uint64_t* low64) {
+  auto h = XXH3_128bits_withSeed(data, n, seed);
+  *high64 = h.high64;
+  *low64 = h.low64;
+}
+
+namespace {
+
+inline uint64_t XXH3_avalanche(uint64_t h64) {
+  h64 ^= h64 >> 37;
+  h64 *= 0x165667919E3779F9U;
+  h64 ^= h64 >> 32;
+  return h64;
+}
+
+inline uint64_t XXH3_unavalanche(uint64_t h64) {
+  h64 ^= h64 >> 32;
+  h64 *= 0x8da8ee41d6df849U;  // inverse of 0x165667919E3779F9U
+  h64 ^= h64 >> 37;
+  return h64;
+}
+
+}  // namespace
+
+void BijectiveHash2x64(uint64_t in_high64, uint64_t in_low64, uint64_t seed,
+                       uint64_t* out_high64, uint64_t* out_low64) {
+  // Adapted from XXH3_len_9to16_128b
+  const uint64_t bitflipl = /*secret part*/ 0x59973f0033362349U - seed;
+  const uint64_t bitfliph = /*secret part*/ 0xc202797692d63d58U + seed;
+  Unsigned128 tmp128 =
+      Multiply64to128(in_low64 ^ in_high64 ^ bitflipl, 0x9E3779B185EBCA87U);
+  uint64_t lo = Lower64of128(tmp128);
+  uint64_t hi = Upper64of128(tmp128);
+  lo += 0x3c0000000000000U;  // (len - 1) << 54
+  in_high64 ^= bitfliph;
+  hi += in_high64 + (Lower32of64(in_high64) * uint64_t{0x85EBCA76});
+  lo ^= EndianSwapValue(hi);
+  tmp128 = Multiply64to128(lo, 0xC2B2AE3D27D4EB4FU);
+  lo = Lower64of128(tmp128);
+  hi = Upper64of128(tmp128) + (hi * 0xC2B2AE3D27D4EB4FU);
+  *out_low64 = XXH3_avalanche(lo);
+  *out_high64 = XXH3_avalanche(hi);
+}
+
+void BijectiveUnhash2x64(uint64_t in_high64, uint64_t in_low64, uint64_t seed,
+                         uint64_t* out_high64, uint64_t* out_low64) {
+  // Inverted above (also consulting XXH3_len_9to16_128b)
+  const uint64_t bitflipl = /*secret part*/ 0x59973f0033362349U - seed;
+  const uint64_t bitfliph = /*secret part*/ 0xc202797692d63d58U + seed;
+  uint64_t lo = XXH3_unavalanche(in_low64);
+  uint64_t hi = XXH3_unavalanche(in_high64);
+  lo *= 0xba79078168d4baf;  // inverse of 0xC2B2AE3D27D4EB4FU
+  hi -= Upper64of128(Multiply64to128(lo, 0xC2B2AE3D27D4EB4FU));
+  hi *= 0xba79078168d4baf;  // inverse of 0xC2B2AE3D27D4EB4FU
+  lo ^= EndianSwapValue(hi);
+  lo -= 0x3c0000000000000U;
+  lo *= 0x887493432badb37U;  // inverse of 0x9E3779B185EBCA87U
+  hi -= Upper64of128(Multiply64to128(lo, 0x9E3779B185EBCA87U));
+  uint32_t tmp32 = Lower32of64(hi) * 0xb6c92f47;  // inverse of 0x85EBCA77
+  hi -= tmp32;
+  hi = (hi & 0xFFFFFFFF00000000U) -
+       ((tmp32 * uint64_t{0x85EBCA76}) & 0xFFFFFFFF00000000U) + tmp32;
+  hi ^= bitfliph;
+  lo ^= hi ^ bitflipl;
+  *out_high64 = hi;
+  *out_low64 = lo;
+}
+
+void BijectiveHash2x64(uint64_t in_high64, uint64_t in_low64,
+                       uint64_t* out_high64, uint64_t* out_low64) {
+  BijectiveHash2x64(in_high64, in_low64, /*seed*/ 0, out_high64, out_low64);
+}
+
+void BijectiveUnhash2x64(uint64_t in_high64, uint64_t in_low64,
+                         uint64_t* out_high64, uint64_t* out_low64) {
+  BijectiveUnhash2x64(in_high64, in_low64, /*seed*/ 0, out_high64, out_low64);
+}
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/util/hash.h b/src/rocksdb/util/hash.h
new file mode 100644
index 000000000..eafa47f34
--- /dev/null
+++ b/src/rocksdb/util/hash.h
@@ -0,0 +1,137 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+//
+// Copyright (c) 2011 The LevelDB Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file. See the AUTHORS file for names of contributors.
+//
+// Common hash functions with convenient interfaces. If hashing a
+// statically-sized input in a performance-critical context, consider
+// calling a specific hash implementation directly, such as
+// XXH3_64bits from xxhash.h.
+//
+// Since this is a very common header, implementation details are kept
+// out-of-line. Out-of-lining also aids in tracking the time spent in
+// hashing functions. Inlining is of limited benefit for runtime-sized
+// hash inputs.
+
+#pragma once
+
+#include <cstddef>
+#include <cstdint>
+
+#include "rocksdb/slice.h"
+#include "util/fastrange.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+// Stable/persistent 64-bit hash. Higher quality and generally faster than
+// Hash(), especially for inputs > 24 bytes.
+// KNOWN FLAW: incrementing seed by 1 might not give sufficiently independent
+// results from previous seed. Recommend incrementing by a large odd number.
+extern uint64_t Hash64(const char* data, size_t n, uint64_t seed);
+
+// Specific optimization without seed (same as seed = 0)
+extern uint64_t Hash64(const char* data, size_t n);
+
+// Non-persistent hash. Must only used for in-memory data structures.
+// The hash results are thus subject to change between releases,
+// architectures, build configuration, etc. (Thus, it rarely makes sense
+// to specify a seed for this function, except for a "rolling" hash.)
+// KNOWN FLAW: incrementing seed by 1 might not give sufficiently independent
+// results from previous seed. Recommend incrementing by a large odd number.
+inline uint64_t NPHash64(const char* data, size_t n, uint64_t seed) {
+#ifdef ROCKSDB_MODIFY_NPHASH
+  // For testing "subject to change"
+  return Hash64(data, n, seed + 123456789);
+#else
+  // Currently same as Hash64
+  return Hash64(data, n, seed);
+#endif
+}
+
+// Specific optimization without seed (same as seed = 0)
+inline uint64_t NPHash64(const char* data, size_t n) {
+#ifdef ROCKSDB_MODIFY_NPHASH
+  // For testing "subject to change"
+  return Hash64(data, n, 123456789);
+#else
+  // Currently same as Hash64
+  return Hash64(data, n);
+#endif
+}
+
+// Convenient and equivalent version of Hash128 without depending on 128-bit
+// scalars
+void Hash2x64(const char* data, size_t n, uint64_t* high64, uint64_t* low64);
+void Hash2x64(const char* data, size_t n, uint64_t seed, uint64_t* high64,
+              uint64_t* low64);
+
+// Hash 128 bits to 128 bits, guaranteed not to lose data (equivalent to
+// Hash2x64 on 16 bytes little endian)
+void BijectiveHash2x64(uint64_t in_high64, uint64_t in_low64,
+                       uint64_t* out_high64, uint64_t* out_low64);
+void BijectiveHash2x64(uint64_t in_high64, uint64_t in_low64, uint64_t seed,
+                       uint64_t* out_high64, uint64_t* out_low64);
+
+// Inverse of above (mostly for testing)
+void BijectiveUnhash2x64(uint64_t in_high64, uint64_t in_low64,
+                         uint64_t* out_high64, uint64_t* out_low64);
+void BijectiveUnhash2x64(uint64_t in_high64, uint64_t in_low64, uint64_t seed,
+                         uint64_t* out_high64, uint64_t* out_low64);
+
+// Stable/persistent 32-bit hash. Moderate quality and high speed on
+// small inputs.
+// TODO: consider rename to Hash32
+// KNOWN FLAW: incrementing seed by 1 might not give sufficiently independent
+// results from previous seed. Recommend pseudorandom or hashed seeds.
+extern uint32_t Hash(const char* data, size_t n, uint32_t seed);
+
+// TODO: consider rename to LegacyBloomHash32
+inline uint32_t BloomHash(const Slice& key) {
+  return Hash(key.data(), key.size(), 0xbc9f1d34);
+}
+
+inline uint64_t GetSliceHash64(const Slice& key) {
+  return Hash64(key.data(), key.size());
+}
+// Provided for convenience for use with template argument deduction, where a
+// specific overload needs to be used.
+extern uint64_t (*kGetSliceNPHash64UnseededFnPtr)(const Slice&);
+
+inline uint64_t GetSliceNPHash64(const Slice& s) {
+  return NPHash64(s.data(), s.size());
+}
+
+inline uint64_t GetSliceNPHash64(const Slice& s, uint64_t seed) {
+  return NPHash64(s.data(), s.size(), seed);
+}
+
+// Similar to `GetSliceNPHash64()` with `seed`, but input comes from
+// concatenation of `Slice`s in `data`.
+extern uint64_t GetSlicePartsNPHash64(const SliceParts& data, uint64_t seed);
+
+inline size_t GetSliceRangedNPHash(const Slice& s, size_t range) {
+  return FastRange64(NPHash64(s.data(), s.size()), range);
+}
+
+// TODO: consider rename to GetSliceHash32
+inline uint32_t GetSliceHash(const Slice& s) {
+  return Hash(s.data(), s.size(), 397);
+}
+
+// Useful for splitting up a 64-bit hash
+inline uint32_t Upper32of64(uint64_t v) {
+  return static_cast<uint32_t>(v >> 32);
+}
+inline uint32_t Lower32of64(uint64_t v) { return static_cast<uint32_t>(v); }
+
+// std::hash compatible interface.
+// TODO: consider rename to SliceHasher32
+struct SliceHasher {
+  uint32_t operator()(const Slice& s) const { return GetSliceHash(s); }
+};
+
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/util/hash128.h b/src/rocksdb/util/hash128.h
new file mode 100644
index 000000000..305caa14a
--- /dev/null
+++ b/src/rocksdb/util/hash128.h
@@ -0,0 +1,26 @@
+//  Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#pragma once
+
+// 128-bit hash gets it own header so that more popular hash.h doesn't
+// depend on math128.h
+
+#include "rocksdb/slice.h"
+#include "util/math128.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+// Stable/persistent 128-bit hash for non-cryptographic applications.
+Unsigned128 Hash128(const char* data, size_t n, uint64_t seed);
+
+// Specific optimization without seed (same as seed = 0)
+Unsigned128 Hash128(const char* data, size_t n);
+
+inline Unsigned128 GetSliceHash128(const Slice& key) {
+  return Hash128(key.data(), key.size());
+}
+
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/util/hash_containers.h b/src/rocksdb/util/hash_containers.h
new file mode 100644
index 000000000..52be3718c
--- /dev/null
+++ b/src/rocksdb/util/hash_containers.h
@@ -0,0 +1,51 @@
+//  Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+// This header establishes compile-time pluggable implementations of hashed
+// container structures, so that deployments have the option of minimal
+// dependencies with ok performance (e.g. std::unordered_map) or more
+// dependencies with optimized performance (e.g. folly::F14FastMap).
+
+#pragma once
+
+#include "rocksdb/rocksdb_namespace.h"
+
+#ifdef USE_FOLLY
+
+#include <folly/container/F14Map.h>
+#include <folly/container/F14Set.h>
+
+namespace ROCKSDB_NAMESPACE {
+
+template <typename K, typename V>
+using UnorderedMap = folly::F14FastMap<K, V>;
+
+template <typename K, typename V, typename H>
+using UnorderedMapH = folly::F14FastMap<K, V, H>;
+
+template <typename K>
+using UnorderedSet = folly::F14FastSet<K>;
+
+}  // namespace ROCKSDB_NAMESPACE
+
+#else
+
+#include <unordered_map>
+#include <unordered_set>
+
+namespace ROCKSDB_NAMESPACE {
+
+template <typename K, typename V>
+using UnorderedMap = std::unordered_map<K, V>;
+
+template <typename K, typename V, typename H>
+using UnorderedMapH = std::unordered_map<K, V, H>;
+
+template <typename K>
+using UnorderedSet = std::unordered_set<K>;
+
+}  // namespace ROCKSDB_NAMESPACE
+
+#endif
diff --git a/src/rocksdb/util/hash_map.h b/src/rocksdb/util/hash_map.h
new file mode 100644
index 000000000..e3ad2584f
--- /dev/null
+++ b/src/rocksdb/util/hash_map.h
@@ -0,0 +1,67 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+//
+
+#pragma once
+
+#include <algorithm>
+#include <array>
+#include <utility>
+
+#include "util/autovector.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+// This is similar to std::unordered_map, except that it tries to avoid
+// allocating or deallocating memory as much as possible. With
+// std::unordered_map, an allocation/deallocation is made for every insertion
+// or deletion because of the requirement that iterators remain valid even
+// with insertions or deletions. This means that the hash chains will be
+// implemented as linked lists.
+//
+// This implementation uses autovector as hash chains insteads.
+//
+template <typename K, typename V, size_t size = 128>
+class HashMap {
+  std::array<autovector<std::pair<K, V>, 1>, size> table_;
+
+ public:
+  bool Contains(K key) {
+    auto& bucket = table_[key % size];
+    auto it = std::find_if(
+        bucket.begin(), bucket.end(),
+        [key](const std::pair<K, V>& p) { return p.first == key; });
+    return it != bucket.end();
+  }
+
+  void Insert(K key, const V& value) {
+    auto& bucket = table_[key % size];
+    bucket.push_back({key, value});
+  }
+
+  void Delete(K key) {
+    auto& bucket = table_[key % size];
+    auto it = std::find_if(
+        bucket.begin(), bucket.end(),
+        [key](const std::pair<K, V>& p) { return p.first == key; });
+    if (it != bucket.end()) {
+      auto last = bucket.end() - 1;
+      if (it != last) {
+        *it = *last;
+      }
+      bucket.pop_back();
+    }
+  }
+
+  V& Get(K key) {
+    auto& bucket = table_[key % size];
+    auto it = std::find_if(
+        bucket.begin(), bucket.end(),
+        [key](const std::pair<K, V>& p) { return p.first == key; });
+    return it->second;
+  }
+};
+
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/util/hash_test.cc b/src/rocksdb/util/hash_test.cc
new file mode 100644
index 000000000..72112b044
--- /dev/null
+++ b/src/rocksdb/util/hash_test.cc
@@ -0,0 +1,853 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+//
+// Copyright (c) 2012 The LevelDB Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file. See the AUTHORS file for names of contributors.
+
+#include "util/hash.h"
+
+#include <cstring>
+#include <type_traits>
+#include <vector>
+
+#include "test_util/testharness.h"
+#include "util/coding.h"
+#include "util/coding_lean.h"
+#include "util/hash128.h"
+#include "util/math.h"
+#include "util/math128.h"
+
+using ROCKSDB_NAMESPACE::BijectiveHash2x64;
+using ROCKSDB_NAMESPACE::BijectiveUnhash2x64;
+using ROCKSDB_NAMESPACE::DecodeFixed64;
+using ROCKSDB_NAMESPACE::EncodeFixed32;
+using ROCKSDB_NAMESPACE::EndianSwapValue;
+using ROCKSDB_NAMESPACE::GetSliceHash64;
+using ROCKSDB_NAMESPACE::Hash;
+using ROCKSDB_NAMESPACE::Hash128;
+using ROCKSDB_NAMESPACE::Hash2x64;
+using ROCKSDB_NAMESPACE::Hash64;
+using ROCKSDB_NAMESPACE::Lower32of64;
+using ROCKSDB_NAMESPACE::Lower64of128;
+using ROCKSDB_NAMESPACE::ReverseBits;
+using ROCKSDB_NAMESPACE::Slice;
+using ROCKSDB_NAMESPACE::Unsigned128;
+using ROCKSDB_NAMESPACE::Upper32of64;
+using ROCKSDB_NAMESPACE::Upper64of128;
+
+// The hash algorithm is part of the file format, for example for the Bloom
+// filters. Test that the hash values are stable for a set of random strings of
+// varying lengths.
+TEST(HashTest, Values) {
+  constexpr uint32_t kSeed = 0xbc9f1d34;  // Same as BloomHash.
+
+  EXPECT_EQ(Hash("", 0, kSeed), 3164544308u);
+  EXPECT_EQ(Hash("\x08", 1, kSeed), 422599524u);
+  EXPECT_EQ(Hash("\x17", 1, kSeed), 3168152998u);
+  EXPECT_EQ(Hash("\x9a", 1, kSeed), 3195034349u);
+  EXPECT_EQ(Hash("\x1c", 1, kSeed), 2651681383u);
+  EXPECT_EQ(Hash("\x4d\x76", 2, kSeed), 2447836956u);
+  EXPECT_EQ(Hash("\x52\xd5", 2, kSeed), 3854228105u);
+  EXPECT_EQ(Hash("\x91\xf7", 2, kSeed), 31066776u);
+  EXPECT_EQ(Hash("\xd6\x27", 2, kSeed), 1806091603u);
+  EXPECT_EQ(Hash("\x30\x46\x0b", 3, kSeed), 3808221797u);
+  EXPECT_EQ(Hash("\x56\xdc\xd6", 3, kSeed), 2157698265u);
+  EXPECT_EQ(Hash("\xd4\x52\x33", 3, kSeed), 1721992661u);
+  EXPECT_EQ(Hash("\x6a\xb5\xf4", 3, kSeed), 2469105222u);
+  EXPECT_EQ(Hash("\x67\x53\x81\x1c", 4, kSeed), 118283265u);
+  EXPECT_EQ(Hash("\x69\xb8\xc0\x88", 4, kSeed), 3416318611u);
+  EXPECT_EQ(Hash("\x1e\x84\xaf\x2d", 4, kSeed), 3315003572u);
+  EXPECT_EQ(Hash("\x46\xdc\x54\xbe", 4, kSeed), 447346355u);
+  EXPECT_EQ(Hash("\xd0\x7a\x6e\xea\x56", 5, kSeed), 4255445370u);
+  EXPECT_EQ(Hash("\x86\x83\xd5\xa4\xd8", 5, kSeed), 2390603402u);
+  EXPECT_EQ(Hash("\xb7\x46\xbb\x77\xce", 5, kSeed), 2048907743u);
+  EXPECT_EQ(Hash("\x6c\xa8\xbc\xe5\x99", 5, kSeed), 2177978500u);
+  EXPECT_EQ(Hash("\x5c\x5e\xe1\xa0\x73\x81", 6, kSeed), 1036846008u);
+  EXPECT_EQ(Hash("\x08\x5d\x73\x1c\xe5\x2e", 6, kSeed), 229980482u);
+  EXPECT_EQ(Hash("\x42\xfb\xf2\x52\xb4\x10", 6, kSeed), 3655585422u);
+  EXPECT_EQ(Hash("\x73\xe1\xff\x56\x9c\xce", 6, kSeed), 3502708029u);
+  EXPECT_EQ(Hash("\x5c\xbe\x97\x75\x54\x9a\x52", 7, kSeed), 815120748u);
+  EXPECT_EQ(Hash("\x16\x82\x39\x49\x88\x2b\x36", 7, kSeed), 3056033698u);
+  EXPECT_EQ(Hash("\x59\x77\xf0\xa7\x24\xf4\x78", 7, kSeed), 587205227u);
+  EXPECT_EQ(Hash("\xd3\xa5\x7c\x0e\xc0\x02\x07", 7, kSeed), 2030937252u);
+  EXPECT_EQ(Hash("\x31\x1b\x98\x75\x96\x22\xd3\x9a", 8, kSeed), 469635402u);
+  EXPECT_EQ(Hash("\x38\xd6\xf7\x28\x20\xb4\x8a\xe9", 8, kSeed), 3530274698u);
+  EXPECT_EQ(Hash("\xbb\x18\x5d\xf4\x12\x03\xf7\x99", 8, kSeed), 1974545809u);
+  EXPECT_EQ(Hash("\x80\xd4\x3b\x3b\xae\x22\xa2\x78", 8, kSeed), 3563570120u);
+  EXPECT_EQ(Hash("\x1a\xb5\xd0\xfe\xab\xc3\x61\xb2\x99", 9, kSeed),
+            2706087434u);
+  EXPECT_EQ(Hash("\x8e\x4a\xc3\x18\x20\x2f\x06\xe6\x3c", 9, kSeed),
+            1534654151u);
+  EXPECT_EQ(Hash("\xb6\xc0\xdd\x05\x3f\xc4\x86\x4c\xef", 9, kSeed),
+            2355554696u);
+  EXPECT_EQ(Hash("\x9a\x5f\x78\x0d\xaf\x50\xe1\x1f\x55", 9, kSeed),
+            1400800912u);
+  EXPECT_EQ(Hash("\x22\x6f\x39\x1f\xf8\xdd\x4f\x52\x17\x94", 10, kSeed),
+            3420325137u);
+  EXPECT_EQ(Hash("\x32\x89\x2a\x75\x48\x3a\x4a\x02\x69\xdd", 10, kSeed),
+            3427803584u);
+  EXPECT_EQ(Hash("\x06\x92\x5c\xf4\x88\x0e\x7e\x68\x38\x3e", 10, kSeed),
+            1152407945u);
+  EXPECT_EQ(Hash("\xbd\x2c\x63\x38\xbf\xe9\x78\xb7\xbf\x15", 10, kSeed),
+            3382479516u);
+}
+
+// The hash algorithm is part of the file format, for example for the Bloom
+// filters.
+TEST(HashTest, Hash64Misc) {
+  constexpr uint32_t kSeed = 0;  // Same as GetSliceHash64
+
+  for (char fill : {'\0', 'a', '1', '\xff'}) {
+    const size_t max_size = 1000;
+    const std::string str(max_size, fill);
+
+    for (size_t size = 0; size <= max_size; ++size) {
+      uint64_t here = Hash64(str.data(), size, kSeed);
+
+      // Must be same as unseeded Hash64 and GetSliceHash64
+      EXPECT_EQ(here, Hash64(str.data(), size));
+      EXPECT_EQ(here, GetSliceHash64(Slice(str.data(), size)));
+
+      // Upper and Lower must reconstruct hash
+      EXPECT_EQ(here, (uint64_t{Upper32of64(here)} << 32) | Lower32of64(here));
+      EXPECT_EQ(here, (uint64_t{Upper32of64(here)} << 32) + Lower32of64(here));
+      EXPECT_EQ(here, (uint64_t{Upper32of64(here)} << 32) ^ Lower32of64(here));
+
+      // Seed changes hash value (with high probability)
+      for (uint64_t var_seed = 1; var_seed != 0; var_seed <<= 1) {
+        EXPECT_NE(here, Hash64(str.data(), size, var_seed));
+      }
+
+      // Size changes hash value (with high probability)
+      size_t max_smaller_by = std::min(size_t{30}, size);
+      for (size_t smaller_by = 1; smaller_by <= max_smaller_by; ++smaller_by) {
+        EXPECT_NE(here, Hash64(str.data(), size - smaller_by, kSeed));
+      }
+    }
+  }
+}
+
+// Test that hash values are "non-trivial" for "trivial" inputs
+TEST(HashTest, Hash64Trivial) {
+  // Thorough test too slow for regression testing
+  constexpr bool thorough = false;
+
+  // For various seeds, make sure hash of empty string is not zero.
+  constexpr uint64_t max_seed = thorough ? 0x1000000 : 0x10000;
+  for (uint64_t seed = 0; seed < max_seed; ++seed) {
+    uint64_t here = Hash64("", 0, seed);
+    EXPECT_NE(Lower32of64(here), 0u);
+    EXPECT_NE(Upper32of64(here), 0u);
+  }
+
+  // For standard seed, make sure hash of small strings are not zero
+  constexpr uint32_t kSeed = 0;  // Same as GetSliceHash64
+  char input[4];
+  constexpr int max_len = thorough ? 3 : 2;
+  for (int len = 1; len <= max_len; ++len) {
+    for (uint32_t i = 0; (i >> (len * 8)) == 0; ++i) {
+      EncodeFixed32(input, i);
+      uint64_t here = Hash64(input, len, kSeed);
+      EXPECT_NE(Lower32of64(here), 0u);
+      EXPECT_NE(Upper32of64(here), 0u);
+    }
+  }
+}
+
+// Test that the hash values are stable for a set of random strings of
+// varying small lengths.
+TEST(HashTest, Hash64SmallValueSchema) {
+  constexpr uint32_t kSeed = 0;  // Same as GetSliceHash64
+
+  EXPECT_EQ(Hash64("", 0, kSeed), uint64_t{5999572062939766020u});
+  EXPECT_EQ(Hash64("\x08", 1, kSeed), uint64_t{583283813901344696u});
+  EXPECT_EQ(Hash64("\x17", 1, kSeed), uint64_t{16175549975585474943u});
+  EXPECT_EQ(Hash64("\x9a", 1, kSeed), uint64_t{16322991629225003903u});
+  EXPECT_EQ(Hash64("\x1c", 1, kSeed), uint64_t{13269285487706833447u});
+  EXPECT_EQ(Hash64("\x4d\x76", 2, kSeed), uint64_t{6859542833406258115u});
+  EXPECT_EQ(Hash64("\x52\xd5", 2, kSeed), uint64_t{4919611532550636959u});
+  EXPECT_EQ(Hash64("\x91\xf7", 2, kSeed), uint64_t{14199427467559720719u});
+  EXPECT_EQ(Hash64("\xd6\x27", 2, kSeed), uint64_t{12292689282614532691u});
+  EXPECT_EQ(Hash64("\x30\x46\x0b", 3, kSeed), uint64_t{11404699285340020889u});
+  EXPECT_EQ(Hash64("\x56\xdc\xd6", 3, kSeed), uint64_t{12404347133785524237u});
+  EXPECT_EQ(Hash64("\xd4\x52\x33", 3, kSeed), uint64_t{15853805298481534034u});
+  EXPECT_EQ(Hash64("\x6a\xb5\xf4", 3, kSeed), uint64_t{16863488758399383382u});
+  EXPECT_EQ(Hash64("\x67\x53\x81\x1c", 4, kSeed),
+            uint64_t{9010661983527562386u});
+  EXPECT_EQ(Hash64("\x69\xb8\xc0\x88", 4, kSeed),
+            uint64_t{6611781377647041447u});
+  EXPECT_EQ(Hash64("\x1e\x84\xaf\x2d", 4, kSeed),
+            uint64_t{15290969111616346501u});
+  EXPECT_EQ(Hash64("\x46\xdc\x54\xbe", 4, kSeed),
+            uint64_t{7063754590279313623u});
+  EXPECT_EQ(Hash64("\xd0\x7a\x6e\xea\x56", 5, kSeed),
+            uint64_t{6384167718754869899u});
+  EXPECT_EQ(Hash64("\x86\x83\xd5\xa4\xd8", 5, kSeed),
+            uint64_t{16874407254108011067u});
+  EXPECT_EQ(Hash64("\xb7\x46\xbb\x77\xce", 5, kSeed),
+            uint64_t{16809880630149135206u});
+  EXPECT_EQ(Hash64("\x6c\xa8\xbc\xe5\x99", 5, kSeed),
+            uint64_t{1249038833153141148u});
+  EXPECT_EQ(Hash64("\x5c\x5e\xe1\xa0\x73\x81", 6, kSeed),
+            uint64_t{17358142495308219330u});
+  EXPECT_EQ(Hash64("\x08\x5d\x73\x1c\xe5\x2e", 6, kSeed),
+            uint64_t{4237646583134806322u});
+  EXPECT_EQ(Hash64("\x42\xfb\xf2\x52\xb4\x10", 6, kSeed),
+            uint64_t{4373664924115234051u});
+  EXPECT_EQ(Hash64("\x73\xe1\xff\x56\x9c\xce", 6, kSeed),
+            uint64_t{12012981210634596029u});
+  EXPECT_EQ(Hash64("\x5c\xbe\x97\x75\x54\x9a\x52", 7, kSeed),
+            uint64_t{5716522398211028826u});
+  EXPECT_EQ(Hash64("\x16\x82\x39\x49\x88\x2b\x36", 7, kSeed),
+            uint64_t{15604531309862565013u});
+  EXPECT_EQ(Hash64("\x59\x77\xf0\xa7\x24\xf4\x78", 7, kSeed),
+            uint64_t{8601330687345614172u});
+  EXPECT_EQ(Hash64("\xd3\xa5\x7c\x0e\xc0\x02\x07", 7, kSeed),
+            uint64_t{8088079329364056942u});
+  EXPECT_EQ(Hash64("\x31\x1b\x98\x75\x96\x22\xd3\x9a", 8, kSeed),
+            uint64_t{9844314944338447628u});
+  EXPECT_EQ(Hash64("\x38\xd6\xf7\x28\x20\xb4\x8a\xe9", 8, kSeed),
+            uint64_t{10973293517982163143u});
+  EXPECT_EQ(Hash64("\xbb\x18\x5d\xf4\x12\x03\xf7\x99", 8, kSeed),
+            uint64_t{9986007080564743219u});
+  EXPECT_EQ(Hash64("\x80\xd4\x3b\x3b\xae\x22\xa2\x78", 8, kSeed),
+            uint64_t{1729303145008254458u});
+  EXPECT_EQ(Hash64("\x1a\xb5\xd0\xfe\xab\xc3\x61\xb2\x99", 9, kSeed),
+            uint64_t{13253403748084181481u});
+  EXPECT_EQ(Hash64("\x8e\x4a\xc3\x18\x20\x2f\x06\xe6\x3c", 9, kSeed),
+            uint64_t{7768754303876232188u});
+  EXPECT_EQ(Hash64("\xb6\xc0\xdd\x05\x3f\xc4\x86\x4c\xef", 9, kSeed),
+            uint64_t{12439346786701492u});
+  EXPECT_EQ(Hash64("\x9a\x5f\x78\x0d\xaf\x50\xe1\x1f\x55", 9, kSeed),
+            uint64_t{10841838338450144690u});
+  EXPECT_EQ(Hash64("\x22\x6f\x39\x1f\xf8\xdd\x4f\x52\x17\x94", 10, kSeed),
+            uint64_t{12883919702069153152u});
+  EXPECT_EQ(Hash64("\x32\x89\x2a\x75\x48\x3a\x4a\x02\x69\xdd", 10, kSeed),
+            uint64_t{12692903507676842188u});
+  EXPECT_EQ(Hash64("\x06\x92\x5c\xf4\x88\x0e\x7e\x68\x38\x3e", 10, kSeed),
+            uint64_t{6540985900674032620u});
+  EXPECT_EQ(Hash64("\xbd\x2c\x63\x38\xbf\xe9\x78\xb7\xbf\x15", 10, kSeed),
+            uint64_t{10551812464348219044u});
+}
+
+std::string Hash64TestDescriptor(const char *repeat, size_t limit) {
+  const char *mod61_encode =
+      "abcdefghijklmnopqrstuvwxyz123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ";
+
+  std::string input;
+  while (input.size() < limit) {
+    input.append(repeat);
+  }
+  std::string rv;
+  for (size_t i = 0; i < limit; ++i) {
+    uint64_t h = GetSliceHash64(Slice(input.data(), i));
+    rv.append(1, mod61_encode[static_cast<size_t>(h % 61)]);
+  }
+  return rv;
+}
+
+// XXPH3 changes its algorithm for various sizes up through 250 bytes, so
+// we need to check the stability of larger sizes also.
+TEST(HashTest, Hash64LargeValueSchema) {
+  // Each of these derives a "descriptor" from the hash values for all
+  // lengths up to 430.
+  // Note that "c" is common for the zero-length string.
+  EXPECT_EQ(
+      Hash64TestDescriptor("foo", 430),
+      "cRhyWsY67B6klRA1udmOuiYuX7IthyGBKqbeosz2hzVglWCmQx8nEdnpkvPfYX56Up2OWOTV"
+      "lTzfAoYwvtqKzjD8E9xttR2unelbXbIV67NUe6bOO23BxaSFRcA3njGu5cUWfgwOqNoTsszp"
+      "uPvKRP6qaUR5VdoBkJUCFIefd7edlNK5mv6JYWaGdwxehg65hTkTmjZoPKxTZo4PLyzbL9U4"
+      "xt12ITSfeP2MfBHuLI2z2pDlBb44UQKVMx27LEoAHsdLp3WfWfgH3sdRBRCHm33UxCM4QmE2"
+      "xJ7gqSvNwTeH7v9GlC8zWbGroyD3UVNeShMLx29O7tH1biemLULwAHyIw8zdtLMDpEJ8m2ic"
+      "l6Lb4fDuuFNAs1GCVUthjK8CV8SWI8Rsz5THSwn5CGhpqUwSZcFknjwWIl5rNCvDxXJqYr");
+  // Note that "1EeRk" is common for "Rocks"
+  EXPECT_EQ(
+      Hash64TestDescriptor("Rocks", 430),
+      "c1EeRkrzgOYWLA8PuhJrwTePJewoB44WdXYDfhbk3ZxTqqg25WlPExDl7IKIQLJvnA6gJxxn"
+      "9TCSLkFGfJeXehaSS1GBqWSzfhEH4VXiXIUCuxJXxtKXcSC6FrNIQGTZbYDiUOLD6Y5inzrF"
+      "9etwQhXUBanw55xAUdNMFQAm2GjJ6UDWp2mISLiMMkLjANWMKLaZMqaFLX37qB4MRO1ooVRv"
+      "zSvaNRSCLxlggQCasQq8icWjzf3HjBlZtU6pd4rkaUxSzHqmo9oM5MghbU5Rtxg8wEfO7lVN"
+      "5wdMONYecslQTwjZUpO1K3LDf3K3XK6sUXM6ShQQ3RHmMn2acB4YtTZ3QQcHYJSOHn2DuWpa"
+      "Q8RqzX5lab92YmOLaCdOHq1BPsM7SIBzMdLgePNsJ1vvMALxAaoDUHPxoFLO2wx18IXnyX");
+  EXPECT_EQ(
+      Hash64TestDescriptor("RocksDB", 430),
+      "c1EeRkukbkb28wLTahwD2sfUhZzaBEnF8SVrxnPVB6A7b8CaAl3UKsDZISF92GSq2wDCukOq"
+      "Jgrsp7A3KZhDiLW8dFXp8UPqPxMCRlMdZeVeJ2dJxrmA6cyt99zkQFj7ELbut6jAeVqARFnw"
+      "fnWVXOsaLrq7bDCbMcns2DKvTaaqTCLMYxI7nhtLpFN1jR755FRQFcOzrrDbh7QhypjdvlYw"
+      "cdAMSZgp9JMHxbM23wPSuH6BOFgxejz35PScZfhDPvTOxIy1jc3MZsWrMC3P324zNolO7JdW"
+      "CX2I5UDKjjaEJfxbgVgJIXxtQGlmj2xkO5sPpjULQV4X2HlY7FQleJ4QRaJIB4buhCA4vUTF"
+      "eMFlxCIYUpTCsal2qsmnGOWa8WCcefrohMjDj1fjzSvSaQwlpyR1GZHF2uPOoQagiCpHpm");
+}
+
+TEST(HashTest, Hash128Misc) {
+  constexpr uint32_t kSeed = 0;  // Same as GetSliceHash128
+
+  for (char fill : {'\0', 'a', '1', '\xff', 'e'}) {
+    const size_t max_size = 1000;
+    std::string str(max_size, fill);
+
+    if (fill == 'e') {
+      // Use different characters to check endianness handling
+      for (size_t i = 0; i < str.size(); ++i) {
+        str[i] += static_cast<char>(i);
+      }
+    }
+
+    for (size_t size = 0; size <= max_size; ++size) {
+      Unsigned128 here = Hash128(str.data(), size, kSeed);
+
+      // Must be same as unseeded Hash128 and GetSliceHash128
+      EXPECT_EQ(here, Hash128(str.data(), size));
+      EXPECT_EQ(here, GetSliceHash128(Slice(str.data(), size)));
+      {
+        uint64_t hi, lo;
+        Hash2x64(str.data(), size, &hi, &lo);
+        EXPECT_EQ(Lower64of128(here), lo);
+        EXPECT_EQ(Upper64of128(here), hi);
+      }
+      if (size == 16) {
+        const uint64_t in_hi = DecodeFixed64(str.data() + 8);
+        const uint64_t in_lo = DecodeFixed64(str.data());
+        uint64_t hi, lo;
+        BijectiveHash2x64(in_hi, in_lo, &hi, &lo);
+        EXPECT_EQ(Lower64of128(here), lo);
+        EXPECT_EQ(Upper64of128(here), hi);
+        uint64_t un_hi, un_lo;
+        BijectiveUnhash2x64(hi, lo, &un_hi, &un_lo);
+        EXPECT_EQ(in_lo, un_lo);
+        EXPECT_EQ(in_hi, un_hi);
+      }
+
+      // Upper and Lower must reconstruct hash
+      EXPECT_EQ(here,
+                (Unsigned128{Upper64of128(here)} << 64) | Lower64of128(here));
+      EXPECT_EQ(here,
+                (Unsigned128{Upper64of128(here)} << 64) ^ Lower64of128(here));
+
+      // Seed changes hash value (with high probability)
+      for (uint64_t var_seed = 1; var_seed != 0; var_seed <<= 1) {
+        Unsigned128 seeded = Hash128(str.data(), size, var_seed);
+        EXPECT_NE(here, seeded);
+        // Must match seeded Hash2x64
+        {
+          uint64_t hi, lo;
+          Hash2x64(str.data(), size, var_seed, &hi, &lo);
+          EXPECT_EQ(Lower64of128(seeded), lo);
+          EXPECT_EQ(Upper64of128(seeded), hi);
+        }
+        if (size == 16) {
+          const uint64_t in_hi = DecodeFixed64(str.data() + 8);
+          const uint64_t in_lo = DecodeFixed64(str.data());
+          uint64_t hi, lo;
+          BijectiveHash2x64(in_hi, in_lo, var_seed, &hi, &lo);
+          EXPECT_EQ(Lower64of128(seeded), lo);
+          EXPECT_EQ(Upper64of128(seeded), hi);
+          uint64_t un_hi, un_lo;
+          BijectiveUnhash2x64(hi, lo, var_seed, &un_hi, &un_lo);
+          EXPECT_EQ(in_lo, un_lo);
+          EXPECT_EQ(in_hi, un_hi);
+        }
+      }
+
+      // Size changes hash value (with high probability)
+      size_t max_smaller_by = std::min(size_t{30}, size);
+      for (size_t smaller_by = 1; smaller_by <= max_smaller_by; ++smaller_by) {
+        EXPECT_NE(here, Hash128(str.data(), size - smaller_by, kSeed));
+      }
+    }
+  }
+}
+
+// Test that hash values are "non-trivial" for "trivial" inputs
+TEST(HashTest, Hash128Trivial) {
+  // Thorough test too slow for regression testing
+  constexpr bool thorough = false;
+
+  // For various seeds, make sure hash of empty string is not zero.
+  constexpr uint64_t max_seed = thorough ? 0x1000000 : 0x10000;
+  for (uint64_t seed = 0; seed < max_seed; ++seed) {
+    Unsigned128 here = Hash128("", 0, seed);
+    EXPECT_NE(Lower64of128(here), 0u);
+    EXPECT_NE(Upper64of128(here), 0u);
+  }
+
+  // For standard seed, make sure hash of small strings are not zero
+  constexpr uint32_t kSeed = 0;  // Same as GetSliceHash128
+  char input[4];
+  constexpr int max_len = thorough ? 3 : 2;
+  for (int len = 1; len <= max_len; ++len) {
+    for (uint32_t i = 0; (i >> (len * 8)) == 0; ++i) {
+      EncodeFixed32(input, i);
+      Unsigned128 here = Hash128(input, len, kSeed);
+      EXPECT_NE(Lower64of128(here), 0u);
+      EXPECT_NE(Upper64of128(here), 0u);
+    }
+  }
+}
+
+std::string Hash128TestDescriptor(const char *repeat, size_t limit) {
+  const char *mod61_encode =
+      "abcdefghijklmnopqrstuvwxyz123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ";
+
+  std::string input;
+  while (input.size() < limit) {
+    input.append(repeat);
+  }
+  std::string rv;
+  for (size_t i = 0; i < limit; ++i) {
+    auto h = GetSliceHash128(Slice(input.data(), i));
+    uint64_t h2 = Upper64of128(h) + Lower64of128(h);
+    rv.append(1, mod61_encode[static_cast<size_t>(h2 % 61)]);
+  }
+  return rv;
+}
+
+// XXH3 changes its algorithm for various sizes up through 250 bytes, so
+// we need to check the stability of larger sizes also.
+TEST(HashTest, Hash128ValueSchema) {
+  // Each of these derives a "descriptor" from the hash values for all
+  // lengths up to 430.
+  // Note that "b" is common for the zero-length string.
+  EXPECT_EQ(
+      Hash128TestDescriptor("foo", 430),
+      "bUMA3As8n9I4vNGhThXlEevxZlyMcbb6TYAlIKJ2f5ponsv99q962rYclQ7u3gfnRdCDQ5JI"
+      "2LrGUaCycbXrvLFe4SjgRb9RQwCfrnmNQ7VSEwSKMnkGCK3bDbXSrnIh5qLXdtvIZklbJpGH"
+      "Dqr93BlqF9ubTnOSYkSdx89XvQqflMIW8bjfQp9BPjQejWOeEQspnN1D3sfgVdFhpaQdHYA5"
+      "pI2XcPlCMFPxvrFuRr7joaDvjNe9IUZaunLPMewuXmC3EL95h52Ju3D7y9RNKhgYxMTrA84B"
+      "yJrMvyjdm3vlBxet4EN7v2GEyjbGuaZW9UL6lrX6PghJDg7ACfLGdxNbH3qXM4zaiG2RKnL5"
+      "S3WXKR78RBB5fRFQ8KDIEQjHFvSNsc3GrAEi6W8P2lv8JMTzjBODO2uN4wadVQFT9wpGfV");
+  // Note that "35D2v" is common for "Rocks"
+  EXPECT_EQ(
+      Hash128TestDescriptor("Rocks", 430),
+      "b35D2vzvklFVDqJmyLRXyApwGGO3EAT3swhe8XJAN3mY2UVPglzdmydxcba6JI2tSvwO6zSu"
+      "ANpjSM7tc9G5iMhsa7R8GfyCXRO1TnLg7HvdWNdgGGBirxZR68BgT7TQsYJt6zyEyISeXI1n"
+      "MXA48Xo7dWfJeYN6Z4KWlqZY7TgFXGbks9AX4ehZNSGtIhdO5i58qlgVX1bEejeOVaCcjC79"
+      "67DrMfOKds7rUQzjBa77sMPcoPW1vu6ljGJPZH3XkRyDMZ1twxXKkNxN3tE8nR7JHwyqBAxE"
+      "fTcjbOWrLZ1irWxRSombD8sGDEmclgF11IxqEhe3Rt7gyofO3nExGckKkS9KfRqsCHbiUyva"
+      "JGkJwUHRXaZnh58b4i1Ei9aQKZjXlvIVDixoZrjcNaH5XJIJlRZce9Z9t82wYapTpckYSg");
+  EXPECT_EQ(
+      Hash128TestDescriptor("RocksDB", 430),
+      "b35D2vFUst3XDZCRlSrhmYYakmqImV97LbBsV6EZlOEQpUPH1d1sD3xMKAPlA5UErHehg5O7"
+      "n966fZqhAf3hRc24kGCLfNAWjyUa7vSNOx3IcPoTyVRFZeFlcCtfl7t1QJumHOCpS33EBmBF"
+      "hvK13QjBbDWYWeHQhJhgV9Mqbx17TIcvUkEnYZxb8IzWNmjVsJG44Z7v52DjGj1ZzS62S2Vv"
+      "qWcDO7apvH5VHg68E9Wl6nXP21vlmUqEH9GeWRehfWVvY7mUpsAg5drHHQyDSdiMceiUuUxJ"
+      "XJqHFcDdzbbPk7xDvbLgWCKvH8k3MpQNWOmbSSRDdAP6nGlDjoTToYkcqVREHJzztSWAAq5h"
+      "GHSUNJ6OxsMHhf8EhXfHtKyUzRmPtjYyeckQcGmrQfFFLidc6cjMDKCdBG6c6HVBrS7H2R");
+}
+
+TEST(FastRange32Test, Values) {
+  using ROCKSDB_NAMESPACE::FastRange32;
+  // Zero range
+  EXPECT_EQ(FastRange32(0, 0), 0U);
+  EXPECT_EQ(FastRange32(123, 0), 0U);
+  EXPECT_EQ(FastRange32(0xffffffff, 0), 0U);
+
+  // One range
+  EXPECT_EQ(FastRange32(0, 1), 0U);
+  EXPECT_EQ(FastRange32(123, 1), 0U);
+  EXPECT_EQ(FastRange32(0xffffffff, 1), 0U);
+
+  // Two range
+  EXPECT_EQ(FastRange32(0, 2), 0U);
+  EXPECT_EQ(FastRange32(123, 2), 0U);
+  EXPECT_EQ(FastRange32(0x7fffffff, 2), 0U);
+  EXPECT_EQ(FastRange32(0x80000000, 2), 1U);
+  EXPECT_EQ(FastRange32(0xffffffff, 2), 1U);
+
+  // Seven range
+  EXPECT_EQ(FastRange32(0, 7), 0U);
+  EXPECT_EQ(FastRange32(123, 7), 0U);
+  EXPECT_EQ(FastRange32(613566756, 7), 0U);
+  EXPECT_EQ(FastRange32(613566757, 7), 1U);
+  EXPECT_EQ(FastRange32(1227133513, 7), 1U);
+  EXPECT_EQ(FastRange32(1227133514, 7), 2U);
+  // etc.
+  EXPECT_EQ(FastRange32(0xffffffff, 7), 6U);
+
+  // Big
+  EXPECT_EQ(FastRange32(1, 0x80000000), 0U);
+  EXPECT_EQ(FastRange32(2, 0x80000000), 1U);
+  EXPECT_EQ(FastRange32(4, 0x7fffffff), 1U);
+  EXPECT_EQ(FastRange32(4, 0x80000000), 2U);
+  EXPECT_EQ(FastRange32(0xffffffff, 0x7fffffff), 0x7ffffffeU);
+  EXPECT_EQ(FastRange32(0xffffffff, 0x80000000), 0x7fffffffU);
+}
+
+TEST(FastRange64Test, Values) {
+  using ROCKSDB_NAMESPACE::FastRange64;
+  // Zero range
+  EXPECT_EQ(FastRange64(0, 0), 0U);
+  EXPECT_EQ(FastRange64(123, 0), 0U);
+  EXPECT_EQ(FastRange64(0xffffFFFF, 0), 0U);
+  EXPECT_EQ(FastRange64(0xffffFFFFffffFFFF, 0), 0U);
+
+  // One range
+  EXPECT_EQ(FastRange64(0, 1), 0U);
+  EXPECT_EQ(FastRange64(123, 1), 0U);
+  EXPECT_EQ(FastRange64(0xffffFFFF, 1), 0U);
+  EXPECT_EQ(FastRange64(0xffffFFFFffffFFFF, 1), 0U);
+
+  // Two range
+  EXPECT_EQ(FastRange64(0, 2), 0U);
+  EXPECT_EQ(FastRange64(123, 2), 0U);
+  EXPECT_EQ(FastRange64(0xffffFFFF, 2), 0U);
+  EXPECT_EQ(FastRange64(0x7fffFFFFffffFFFF, 2), 0U);
+  EXPECT_EQ(FastRange64(0x8000000000000000, 2), 1U);
+  EXPECT_EQ(FastRange64(0xffffFFFFffffFFFF, 2), 1U);
+
+  // Seven range
+  EXPECT_EQ(FastRange64(0, 7), 0U);
+  EXPECT_EQ(FastRange64(123, 7), 0U);
+  EXPECT_EQ(FastRange64(0xffffFFFF, 7), 0U);
+  EXPECT_EQ(FastRange64(2635249153387078802, 7), 0U);
+  EXPECT_EQ(FastRange64(2635249153387078803, 7), 1U);
+  EXPECT_EQ(FastRange64(5270498306774157604, 7), 1U);
+  EXPECT_EQ(FastRange64(5270498306774157605, 7), 2U);
+  EXPECT_EQ(FastRange64(0x7fffFFFFffffFFFF, 7), 3U);
+  EXPECT_EQ(FastRange64(0x8000000000000000, 7), 3U);
+  EXPECT_EQ(FastRange64(0xffffFFFFffffFFFF, 7), 6U);
+
+  // Big but 32-bit range
+  EXPECT_EQ(FastRange64(0x100000000, 0x80000000), 0U);
+  EXPECT_EQ(FastRange64(0x200000000, 0x80000000), 1U);
+  EXPECT_EQ(FastRange64(0x400000000, 0x7fffFFFF), 1U);
+  EXPECT_EQ(FastRange64(0x400000000, 0x80000000), 2U);
+  EXPECT_EQ(FastRange64(0xffffFFFFffffFFFF, 0x7fffFFFF), 0x7fffFFFEU);
+  EXPECT_EQ(FastRange64(0xffffFFFFffffFFFF, 0x80000000), 0x7fffFFFFU);
+
+  // Big, > 32-bit range
+#if SIZE_MAX == UINT64_MAX
+  EXPECT_EQ(FastRange64(0x7fffFFFFffffFFFF, 0x4200000002), 0x2100000000U);
+  EXPECT_EQ(FastRange64(0x8000000000000000, 0x4200000002), 0x2100000001U);
+
+  EXPECT_EQ(FastRange64(0x0000000000000000, 420000000002), 0U);
+  EXPECT_EQ(FastRange64(0x7fffFFFFffffFFFF, 420000000002), 210000000000U);
+  EXPECT_EQ(FastRange64(0x8000000000000000, 420000000002), 210000000001U);
+  EXPECT_EQ(FastRange64(0xffffFFFFffffFFFF, 420000000002), 420000000001U);
+
+  EXPECT_EQ(FastRange64(0xffffFFFFffffFFFF, 0xffffFFFFffffFFFF),
+            0xffffFFFFffffFFFEU);
+#endif
+}
+
+TEST(FastRangeGenericTest, Values) {
+  using ROCKSDB_NAMESPACE::FastRangeGeneric;
+  // Generic (including big and small)
+  // Note that FastRangeGeneric is also tested indirectly above via
+  // FastRange32 and FastRange64.
+  EXPECT_EQ(
+      FastRangeGeneric(uint64_t{0x8000000000000000}, uint64_t{420000000002}),
+      uint64_t{210000000001});
+  EXPECT_EQ(FastRangeGeneric(uint64_t{0x8000000000000000}, uint16_t{12468}),
+            uint16_t{6234});
+  EXPECT_EQ(FastRangeGeneric(uint32_t{0x80000000}, uint16_t{12468}),
+            uint16_t{6234});
+  // Not recommended for typical use because for example this could fail on
+  // some platforms and pass on others:
+  // EXPECT_EQ(FastRangeGeneric(static_cast<unsigned long>(0x80000000),
+  //                           uint16_t{12468}),
+  //          uint16_t{6234});
+}
+
+// for inspection of disassembly
+uint32_t FastRange32(uint32_t hash, uint32_t range) {
+  return ROCKSDB_NAMESPACE::FastRange32(hash, range);
+}
+
+// for inspection of disassembly
+size_t FastRange64(uint64_t hash, size_t range) {
+  return ROCKSDB_NAMESPACE::FastRange64(hash, range);
+}
+
+// Tests for math.h / math128.h (not worth a separate test binary)
+using ROCKSDB_NAMESPACE::BitParity;
+using ROCKSDB_NAMESPACE::BitsSetToOne;
+using ROCKSDB_NAMESPACE::ConstexprFloorLog2;
+using ROCKSDB_NAMESPACE::CountTrailingZeroBits;
+using ROCKSDB_NAMESPACE::DecodeFixed128;
+using ROCKSDB_NAMESPACE::DecodeFixedGeneric;
+using ROCKSDB_NAMESPACE::DownwardInvolution;
+using ROCKSDB_NAMESPACE::EncodeFixed128;
+using ROCKSDB_NAMESPACE::EncodeFixedGeneric;
+using ROCKSDB_NAMESPACE::FloorLog2;
+using ROCKSDB_NAMESPACE::Lower64of128;
+using ROCKSDB_NAMESPACE::Multiply64to128;
+using ROCKSDB_NAMESPACE::Unsigned128;
+using ROCKSDB_NAMESPACE::Upper64of128;
+
+int blah(int x) { return DownwardInvolution(x); }
+
+template <typename T>
+static void test_BitOps() {
+  // This complex code is to generalize to 128-bit values. Otherwise
+  // we could just use = static_cast<T>(0x5555555555555555ULL);
+  T everyOtherBit = 0;
+  for (unsigned i = 0; i < sizeof(T); ++i) {
+    everyOtherBit = (everyOtherBit << 8) | T{0x55};
+  }
+
+  // This one built using bit operations, as our 128-bit layer
+  // might not implement arithmetic such as subtraction.
+  T vm1 = 0;  // "v minus one"
+
+  for (int i = 0; i < int{8 * sizeof(T)}; ++i) {
+    T v = T{1} << i;
+    // If we could directly use arithmetic:
+    // T vm1 = static_cast<T>(v - 1);
+
+    // FloorLog2
+    if (v > 0) {
+      EXPECT_EQ(FloorLog2(v), i);
+      EXPECT_EQ(ConstexprFloorLog2(v), i);
+    }
+    if (vm1 > 0) {
+      EXPECT_EQ(FloorLog2(vm1), i - 1);
+      EXPECT_EQ(ConstexprFloorLog2(vm1), i - 1);
+      EXPECT_EQ(FloorLog2(everyOtherBit & vm1), (i - 1) & ~1);
+      EXPECT_EQ(ConstexprFloorLog2(everyOtherBit & vm1), (i - 1) & ~1);
+    }
+
+    // CountTrailingZeroBits
+    if (v != 0) {
+      EXPECT_EQ(CountTrailingZeroBits(v), i);
+    }
+    if (vm1 != 0) {
+      EXPECT_EQ(CountTrailingZeroBits(vm1), 0);
+    }
+    if (i < int{8 * sizeof(T)} - 1) {
+      EXPECT_EQ(CountTrailingZeroBits(~vm1 & everyOtherBit), (i + 1) & ~1);
+    }
+
+    // BitsSetToOne
+    EXPECT_EQ(BitsSetToOne(v), 1);
+    EXPECT_EQ(BitsSetToOne(vm1), i);
+    EXPECT_EQ(BitsSetToOne(vm1 & everyOtherBit), (i + 1) / 2);
+
+    // BitParity
+    EXPECT_EQ(BitParity(v), 1);
+    EXPECT_EQ(BitParity(vm1), i & 1);
+    EXPECT_EQ(BitParity(vm1 & everyOtherBit), ((i + 1) / 2) & 1);
+
+    // EndianSwapValue
+    T ev = T{1} << (((sizeof(T) - 1 - (i / 8)) * 8) + i % 8);
+    EXPECT_EQ(EndianSwapValue(v), ev);
+
+    // ReverseBits
+    EXPECT_EQ(ReverseBits(v), static_cast<T>(T{1} << (8 * sizeof(T) - 1 - i)));
+#ifdef HAVE_UINT128_EXTENSION          // Uses multiplication
+    if (std::is_unsigned<T>::value) {  // Technical UB on signed type
+      T rv = T{1} << (8 * sizeof(T) - 1 - i);
+      EXPECT_EQ(ReverseBits(vm1), static_cast<T>(rv * ~T{1}));
+    }
+#endif
+
+    // DownwardInvolution
+    {
+      T misc = static_cast<T>(/*random*/ 0xc682cd153d0e3279U +
+                              i * /*random*/ 0x9b3972f3bea0baa3U);
+      if constexpr (sizeof(T) > 8) {
+        misc = (misc << 64) | (/*random*/ 0x52af031a38ced62dU +
+                               i * /*random*/ 0x936f803d9752ddc3U);
+      }
+      T misc_masked = misc & vm1;
+      EXPECT_LE(misc_masked, vm1);
+      T di_misc_masked = DownwardInvolution(misc_masked);
+      EXPECT_LE(di_misc_masked, vm1);
+      if (misc_masked > 0) {
+        // Highest-order 1 in same position
+        EXPECT_EQ(FloorLog2(misc_masked), FloorLog2(di_misc_masked));
+      }
+      // Validate involution property on short value
+      EXPECT_EQ(DownwardInvolution(di_misc_masked), misc_masked);
+
+      // Validate involution property on large value
+      T di_misc = DownwardInvolution(misc);
+      EXPECT_EQ(DownwardInvolution(di_misc), misc);
+      // Highest-order 1 in same position
+      if (misc > 0) {
+        EXPECT_EQ(FloorLog2(misc), FloorLog2(di_misc));
+      }
+
+      // Validate distributes over xor.
+      // static_casts to avoid numerical promotion effects.
+      EXPECT_EQ(DownwardInvolution(static_cast<T>(misc_masked ^ vm1)),
+                static_cast<T>(di_misc_masked ^ DownwardInvolution(vm1)));
+      T misc2 = static_cast<T>(misc >> 1);
+      EXPECT_EQ(DownwardInvolution(static_cast<T>(misc ^ misc2)),
+                static_cast<T>(di_misc ^ DownwardInvolution(misc2)));
+
+      // Choose some small number of bits to pull off to test combined
+      // uniqueness guarantee
+      int in_bits = i % 7;
+      unsigned in_mask = (unsigned{1} << in_bits) - 1U;
+      // IMPLICIT: int out_bits = 8 - in_bits;
+      std::vector<bool> seen(256, false);
+      for (int j = 0; j < 255; ++j) {
+        T t_in = misc ^ static_cast<T>(j);
+        unsigned in = static_cast<unsigned>(t_in);
+        unsigned out = static_cast<unsigned>(DownwardInvolution(t_in));
+        unsigned val = ((out << in_bits) | (in & in_mask)) & 255U;
+        EXPECT_FALSE(seen[val]);
+        seen[val] = true;
+      }
+
+      if (i + 8 < int{8 * sizeof(T)}) {
+        // Also test manipulating bits in the middle of input is
+        // bijective in bottom of output
+        seen = std::vector<bool>(256, false);
+        for (int j = 0; j < 255; ++j) {
+          T in = misc ^ (static_cast<T>(j) << i);
+          unsigned val = static_cast<unsigned>(DownwardInvolution(in)) & 255U;
+          EXPECT_FALSE(seen[val]);
+          seen[val] = true;
+        }
+      }
+    }
+
+    vm1 = (vm1 << 1) | 1;
+  }
+
+  EXPECT_EQ(ConstexprFloorLog2(T{1}), 0);
+  EXPECT_EQ(ConstexprFloorLog2(T{2}), 1);
+  EXPECT_EQ(ConstexprFloorLog2(T{3}), 1);
+  EXPECT_EQ(ConstexprFloorLog2(T{42}), 5);
+}
+
+TEST(MathTest, BitOps) {
+  test_BitOps<uint32_t>();
+  test_BitOps<uint64_t>();
+  test_BitOps<uint16_t>();
+  test_BitOps<uint8_t>();
+  test_BitOps<unsigned char>();
+  test_BitOps<unsigned short>();
+  test_BitOps<unsigned int>();
+  test_BitOps<unsigned long>();
+  test_BitOps<unsigned long long>();
+  test_BitOps<char>();
+  test_BitOps<size_t>();
+  test_BitOps<int32_t>();
+  test_BitOps<int64_t>();
+  test_BitOps<int16_t>();
+  test_BitOps<int8_t>();
+  test_BitOps<signed char>();
+  test_BitOps<short>();
+  test_BitOps<int>();
+  test_BitOps<long>();
+  test_BitOps<long long>();
+  test_BitOps<ptrdiff_t>();
+}
+
+TEST(MathTest, BitOps128) { test_BitOps<Unsigned128>(); }
+
+TEST(MathTest, Math128) {
+  const Unsigned128 sixteenHexOnes = 0x1111111111111111U;
+  const Unsigned128 thirtyHexOnes = (sixteenHexOnes << 56) | sixteenHexOnes;
+  const Unsigned128 sixteenHexTwos = 0x2222222222222222U;
+  const Unsigned128 thirtyHexTwos = (sixteenHexTwos << 56) | sixteenHexTwos;
+
+  // v will slide from all hex ones to all hex twos
+  Unsigned128 v = thirtyHexOnes;
+  for (int i = 0; i <= 30; ++i) {
+    // Test bitwise operations
+    EXPECT_EQ(BitsSetToOne(v), 30);
+    EXPECT_EQ(BitsSetToOne(~v), 128 - 30);
+    EXPECT_EQ(BitsSetToOne(v & thirtyHexOnes), 30 - i);
+    EXPECT_EQ(BitsSetToOne(v | thirtyHexOnes), 30 + i);
+    EXPECT_EQ(BitsSetToOne(v ^ thirtyHexOnes), 2 * i);
+    EXPECT_EQ(BitsSetToOne(v & thirtyHexTwos), i);
+    EXPECT_EQ(BitsSetToOne(v | thirtyHexTwos), 60 - i);
+    EXPECT_EQ(BitsSetToOne(v ^ thirtyHexTwos), 60 - 2 * i);
+
+    // Test comparisons
+    EXPECT_EQ(v == thirtyHexOnes, i == 0);
+    EXPECT_EQ(v == thirtyHexTwos, i == 30);
+    EXPECT_EQ(v > thirtyHexOnes, i > 0);
+    EXPECT_EQ(v > thirtyHexTwos, false);
+    EXPECT_EQ(v >= thirtyHexOnes, true);
+    EXPECT_EQ(v >= thirtyHexTwos, i == 30);
+    EXPECT_EQ(v < thirtyHexOnes, false);
+    EXPECT_EQ(v < thirtyHexTwos, i < 30);
+    EXPECT_EQ(v <= thirtyHexOnes, i == 0);
+    EXPECT_EQ(v <= thirtyHexTwos, true);
+
+    // Update v, clearing upper-most byte
+    v = ((v << 12) >> 8) | 0x2;
+  }
+
+  for (int i = 0; i < 128; ++i) {
+    // Test shifts
+    Unsigned128 sl = thirtyHexOnes << i;
+    Unsigned128 sr = thirtyHexOnes >> i;
+    EXPECT_EQ(BitsSetToOne(sl), std::min(30, 32 - i / 4));
+    EXPECT_EQ(BitsSetToOne(sr), std::max(0, 30 - (i + 3) / 4));
+    EXPECT_EQ(BitsSetToOne(sl & sr), i % 2 ? 0 : std::max(0, 30 - i / 2));
+  }
+
+  // Test 64x64->128 multiply
+  Unsigned128 product =
+      Multiply64to128(0x1111111111111111U, 0x2222222222222222U);
+  EXPECT_EQ(Lower64of128(product), 2295594818061633090U);
+  EXPECT_EQ(Upper64of128(product), 163971058432973792U);
+}
+
+TEST(MathTest, Coding128) {
+  const char *in = "_1234567890123456";
+  // Note: in + 1 is likely unaligned
+  Unsigned128 decoded = DecodeFixed128(in + 1);
+  EXPECT_EQ(Lower64of128(decoded), 0x3837363534333231U);
+  EXPECT_EQ(Upper64of128(decoded), 0x3635343332313039U);
+  char out[18];
+  out[0] = '_';
+  EncodeFixed128(out + 1, decoded);
+  out[17] = '\0';
+  EXPECT_EQ(std::string(in), std::string(out));
+}
+
+TEST(MathTest, CodingGeneric) {
+  const char *in = "_1234567890123456";
+  // Decode
+  // Note: in + 1 is likely unaligned
+  Unsigned128 decoded128 = DecodeFixedGeneric<Unsigned128>(in + 1);
+  EXPECT_EQ(Lower64of128(decoded128), 0x3837363534333231U);
+  EXPECT_EQ(Upper64of128(decoded128), 0x3635343332313039U);
+
+  uint64_t decoded64 = DecodeFixedGeneric<uint64_t>(in + 1);
+  EXPECT_EQ(decoded64, 0x3837363534333231U);
+
+  uint32_t decoded32 = DecodeFixedGeneric<uint32_t>(in + 1);
+  EXPECT_EQ(decoded32, 0x34333231U);
+
+  uint16_t decoded16 = DecodeFixedGeneric<uint16_t>(in + 1);
+  EXPECT_EQ(decoded16, 0x3231U);
+
+  // Encode
+  char out[18];
+  out[0] = '_';
+  memset(out + 1, '\0', 17);
+  EncodeFixedGeneric(out + 1, decoded128);
+  EXPECT_EQ(std::string(in), std::string(out));
+
+  memset(out + 1, '\0', 9);
+  EncodeFixedGeneric(out + 1, decoded64);
+  EXPECT_EQ(std::string("_12345678"), std::string(out));
+
+  memset(out + 1, '\0', 5);
+  EncodeFixedGeneric(out + 1, decoded32);
+  EXPECT_EQ(std::string("_1234"), std::string(out));
+
+  memset(out + 1, '\0', 3);
+  EncodeFixedGeneric(out + 1, decoded16);
+  EXPECT_EQ(std::string("_12"), std::string(out));
+}
+
+int main(int argc, char **argv) {
+  fprintf(stderr, "NPHash64 id: %x\n",
+          static_cast<int>(ROCKSDB_NAMESPACE::GetSliceNPHash64("RocksDB")));
+  ROCKSDB_NAMESPACE::port::InstallStackTraceHandler();
+  ::testing::InitGoogleTest(&argc, argv);
+
+  return RUN_ALL_TESTS();
+}
diff --git a/src/rocksdb/util/heap.h b/src/rocksdb/util/heap.h
new file mode 100644
index 000000000..f221fc732
--- /dev/null
+++ b/src/rocksdb/util/heap.h
@@ -0,0 +1,174 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#pragma once
+
+#include <algorithm>
+#include <cstdint>
+#include <functional>
+
+#include "port/port.h"
+#include "util/autovector.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+// Binary heap implementation optimized for use in multi-way merge sort.
+// Comparison to std::priority_queue:
+// - In libstdc++, std::priority_queue::pop() usually performs just over logN
+//   comparisons but never fewer.
+// - std::priority_queue does not have a replace-top operation, requiring a
+//   pop+push.  If the replacement element is the new top, this requires
+//   around 2logN comparisons.
+// - This heap's pop() uses a "schoolbook" downheap which requires up to ~2logN
+//   comparisons.
+// - This heap provides a replace_top() operation which requires [1, 2logN]
+//   comparisons.  When the replacement element is also the new top, this
+//   takes just 1 or 2 comparisons.
+//
+// The last property can yield an order-of-magnitude performance improvement
+// when merge-sorting real-world non-random data.  If the merge operation is
+// likely to take chunks of elements from the same input stream, only 1
+// comparison per element is needed.  In RocksDB-land, this happens when
+// compacting a database where keys are not randomly distributed across L0
+// files but nearby keys are likely to be in the same L0 file.
+//
+// The container uses the same counterintuitive ordering as
+// std::priority_queue: the comparison operator is expected to provide the
+// less-than relation, but top() will return the maximum.
+
+template <typename T, typename Compare = std::less<T>>
+class BinaryHeap {
+ public:
+  BinaryHeap() {}
+  explicit BinaryHeap(Compare cmp) : cmp_(std::move(cmp)) {}
+
+  void push(const T& value) {
+    data_.push_back(value);
+    upheap(data_.size() - 1);
+  }
+
+  void push(T&& value) {
+    data_.push_back(std::move(value));
+    upheap(data_.size() - 1);
+  }
+
+  const T& top() const {
+    assert(!empty());
+    return data_.front();
+  }
+
+  void replace_top(const T& value) {
+    assert(!empty());
+    data_.front() = value;
+    downheap(get_root());
+  }
+
+  void replace_top(T&& value) {
+    assert(!empty());
+    data_.front() = std::move(value);
+    downheap(get_root());
+  }
+
+  void pop() {
+    assert(!empty());
+    if (data_.size() > 1) {
+      // Avoid self-move-assign, because it could cause problems with
+      // classes which are not prepared for this and it trips up the
+      // STL debugger when activated.
+      data_.front() = std::move(data_.back());
+    }
+    data_.pop_back();
+    if (!empty()) {
+      downheap(get_root());
+    } else {
+      reset_root_cmp_cache();
+    }
+  }
+
+  void swap(BinaryHeap& other) {
+    std::swap(cmp_, other.cmp_);
+    data_.swap(other.data_);
+    std::swap(root_cmp_cache_, other.root_cmp_cache_);
+  }
+
+  void clear() {
+    data_.clear();
+    reset_root_cmp_cache();
+  }
+
+  bool empty() const { return data_.empty(); }
+
+  size_t size() const { return data_.size(); }
+
+  void reset_root_cmp_cache() {
+    root_cmp_cache_ = std::numeric_limits<size_t>::max();
+  }
+
+ private:
+  static inline size_t get_root() { return 0; }
+  static inline size_t get_parent(size_t index) { return (index - 1) / 2; }
+  static inline size_t get_left(size_t index) { return 2 * index + 1; }
+  static inline size_t get_right(size_t index) { return 2 * index + 2; }
+
+  void upheap(size_t index) {
+    T v = std::move(data_[index]);
+    while (index > get_root()) {
+      const size_t parent = get_parent(index);
+      if (!cmp_(data_[parent], v)) {
+        break;
+      }
+      data_[index] = std::move(data_[parent]);
+      index = parent;
+    }
+    data_[index] = std::move(v);
+    reset_root_cmp_cache();
+  }
+
+  void downheap(size_t index) {
+    T v = std::move(data_[index]);
+
+    size_t picked_child = std::numeric_limits<size_t>::max();
+    while (1) {
+      const size_t left_child = get_left(index);
+      if (get_left(index) >= data_.size()) {
+        break;
+      }
+      const size_t right_child = left_child + 1;
+      assert(right_child == get_right(index));
+      picked_child = left_child;
+      if (index == 0 && root_cmp_cache_ < data_.size()) {
+        picked_child = root_cmp_cache_;
+      } else if (right_child < data_.size() &&
+                 cmp_(data_[left_child], data_[right_child])) {
+        picked_child = right_child;
+      }
+      if (!cmp_(v, data_[picked_child])) {
+        break;
+      }
+      data_[index] = std::move(data_[picked_child]);
+      index = picked_child;
+    }
+
+    if (index == 0) {
+      // We did not change anything in the tree except for the value
+      // of the root node, left and right child did not change, we can
+      // cache that `picked_child` is the smallest child
+      // so next time we compare againist it directly
+      root_cmp_cache_ = picked_child;
+    } else {
+      // the tree changed, reset cache
+      reset_root_cmp_cache();
+    }
+
+    data_[index] = std::move(v);
+  }
+
+  Compare cmp_;
+  autovector<T> data_;
+  // Used to reduce number of cmp_ calls in downheap()
+  size_t root_cmp_cache_ = std::numeric_limits<size_t>::max();
+};
+
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/util/heap_test.cc b/src/rocksdb/util/heap_test.cc
new file mode 100644
index 000000000..bbb93324f
--- /dev/null
+++ b/src/rocksdb/util/heap_test.cc
@@ -0,0 +1,131 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#include "util/heap.h"
+
+#include <gtest/gtest.h>
+
+#include <climits>
+#include <queue>
+#include <random>
+#include <utility>
+
+#include "port/stack_trace.h"
+
+#ifndef GFLAGS
+const int64_t FLAGS_iters = 100000;
+#else
+#include "util/gflags_compat.h"
+DEFINE_int64(iters, 100000, "number of pseudo-random operations in each test");
+#endif  // GFLAGS
+
+/*
+ * Compares the custom heap implementation in util/heap.h against
+ * std::priority_queue on a pseudo-random sequence of operations.
+ */
+
+namespace ROCKSDB_NAMESPACE {
+
+using HeapTestValue = uint64_t;
+using Params = std::tuple<size_t, HeapTestValue, int64_t>;
+
+class HeapTest : public ::testing::TestWithParam<Params> {};
+
+TEST_P(HeapTest, Test) {
+  // This test performs the same pseudorandom sequence of operations on a
+  // BinaryHeap and an std::priority_queue, comparing output.  The three
+  // possible operations are insert, replace top and pop.
+  //
+  // Insert is chosen slightly more often than the others so that the size of
+  // the heap slowly grows.  Once the size heats the MAX_HEAP_SIZE limit, we
+  // disallow inserting until the heap becomes empty, testing the "draining"
+  // scenario.
+
+  const auto MAX_HEAP_SIZE = std::get<0>(GetParam());
+  const auto MAX_VALUE = std::get<1>(GetParam());
+  const auto RNG_SEED = std::get<2>(GetParam());
+
+  BinaryHeap<HeapTestValue> heap;
+  std::priority_queue<HeapTestValue> ref;
+
+  std::mt19937 rng(static_cast<unsigned int>(RNG_SEED));
+  std::uniform_int_distribution<HeapTestValue> value_dist(0, MAX_VALUE);
+  int ndrains = 0;
+  bool draining = false;  // hit max size, draining until we empty the heap
+  size_t size = 0;
+  for (int64_t i = 0; i < FLAGS_iters; ++i) {
+    if (size == 0) {
+      draining = false;
+    }
+
+    if (!draining && (size == 0 || std::bernoulli_distribution(0.4)(rng))) {
+      // insert
+      HeapTestValue val = value_dist(rng);
+      heap.push(val);
+      ref.push(val);
+      ++size;
+      if (size == MAX_HEAP_SIZE) {
+        draining = true;
+        ++ndrains;
+      }
+    } else if (std::bernoulli_distribution(0.5)(rng)) {
+      // replace top
+      HeapTestValue val = value_dist(rng);
+      heap.replace_top(val);
+      ref.pop();
+      ref.push(val);
+    } else {
+      // pop
+      assert(size > 0);
+      heap.pop();
+      ref.pop();
+      --size;
+    }
+
+    // After every operation, check that the public methods give the same
+    // results
+    assert((size == 0) == ref.empty());
+    ASSERT_EQ(size == 0, heap.empty());
+    if (size > 0) {
+      ASSERT_EQ(ref.top(), heap.top());
+    }
+  }
+
+  // Probabilities should be set up to occasionally hit the max heap size and
+  // drain it
+  assert(ndrains > 0);
+
+  heap.clear();
+  ASSERT_TRUE(heap.empty());
+}
+
+// Basic test, MAX_VALUE = 3*MAX_HEAP_SIZE (occasional duplicates)
+INSTANTIATE_TEST_CASE_P(Basic, HeapTest,
+                        ::testing::Values(Params(1000, 3000,
+                                                 0x1b575cf05b708945)));
+// Mid-size heap with small values (many duplicates)
+INSTANTIATE_TEST_CASE_P(SmallValues, HeapTest,
+                        ::testing::Values(Params(100, 10, 0x5ae213f7bd5dccd0)));
+// Small heap, large value range (no duplicates)
+INSTANTIATE_TEST_CASE_P(SmallHeap, HeapTest,
+                        ::testing::Values(Params(10, ULLONG_MAX,
+                                                 0x3e1fa8f4d01707cf)));
+// Two-element heap
+INSTANTIATE_TEST_CASE_P(TwoElementHeap, HeapTest,
+                        ::testing::Values(Params(2, 5, 0x4b5e13ea988c6abc)));
+// One-element heap
+INSTANTIATE_TEST_CASE_P(OneElementHeap, HeapTest,
+                        ::testing::Values(Params(1, 3, 0x176a1019ab0b612e)));
+
+}  // namespace ROCKSDB_NAMESPACE
+
+int main(int argc, char** argv) {
+  ROCKSDB_NAMESPACE::port::InstallStackTraceHandler();
+  ::testing::InitGoogleTest(&argc, argv);
+#ifdef GFLAGS
+  GFLAGS_NAMESPACE::ParseCommandLineFlags(&argc, &argv, true);
+#endif  // GFLAGS
+  return RUN_ALL_TESTS();
+}
diff --git a/src/rocksdb/util/kv_map.h b/src/rocksdb/util/kv_map.h
new file mode 100644
index 000000000..62be6d18e
--- /dev/null
+++ b/src/rocksdb/util/kv_map.h
@@ -0,0 +1,33 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+#pragma once
+
+#include <map>
+#include <string>
+
+#include "rocksdb/comparator.h"
+#include "rocksdb/slice.h"
+#include "util/coding.h"
+
+namespace ROCKSDB_NAMESPACE {
+namespace stl_wrappers {
+
+struct LessOfComparator {
+  explicit LessOfComparator(const Comparator* c = BytewiseComparator())
+      : cmp(c) {}
+
+  bool operator()(const std::string& a, const std::string& b) const {
+    return cmp->Compare(Slice(a), Slice(b)) < 0;
+  }
+  bool operator()(const Slice& a, const Slice& b) const {
+    return cmp->Compare(a, b) < 0;
+  }
+
+  const Comparator* cmp;
+};
+
+using KVMap = std::map<std::string, std::string, LessOfComparator>;
+}  // namespace stl_wrappers
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/util/log_write_bench.cc b/src/rocksdb/util/log_write_bench.cc
new file mode 100644
index 000000000..c1637db15
--- /dev/null
+++ b/src/rocksdb/util/log_write_bench.cc
@@ -0,0 +1,88 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#ifndef GFLAGS
+#include <cstdio>
+int main() {
+  fprintf(stderr, "Please install gflags to run rocksdb tools\n");
+  return 1;
+}
+#else
+
+#include "file/writable_file_writer.h"
+#include "monitoring/histogram.h"
+#include "rocksdb/env.h"
+#include "rocksdb/system_clock.h"
+#include "test_util/testharness.h"
+#include "test_util/testutil.h"
+#include "util/gflags_compat.h"
+
+using GFLAGS_NAMESPACE::ParseCommandLineFlags;
+using GFLAGS_NAMESPACE::SetUsageMessage;
+
+// A simple benchmark to simulate transactional logs
+
+DEFINE_int32(num_records, 6000, "Number of records.");
+DEFINE_int32(record_size, 249, "Size of each record.");
+DEFINE_int32(record_interval, 10000, "Interval between records (microSec)");
+DEFINE_int32(bytes_per_sync, 0, "bytes_per_sync parameter in EnvOptions");
+DEFINE_bool(enable_sync, false, "sync after each write.");
+
+namespace ROCKSDB_NAMESPACE {
+void RunBenchmark() {
+  std::string file_name = test::PerThreadDBPath("log_write_benchmark.log");
+  DBOptions options;
+  Env* env = Env::Default();
+  const auto& clock = env->GetSystemClock();
+  EnvOptions env_options = env->OptimizeForLogWrite(EnvOptions(), options);
+  env_options.bytes_per_sync = FLAGS_bytes_per_sync;
+  std::unique_ptr<WritableFile> file;
+  env->NewWritableFile(file_name, &file, env_options);
+  std::unique_ptr<WritableFileWriter> writer;
+  writer.reset(new WritableFileWriter(std::move(file), file_name, env_options,
+                                      clock, nullptr /* stats */,
+                                      options.listeners));
+
+  std::string record;
+  record.assign(FLAGS_record_size, 'X');
+
+  HistogramImpl hist;
+
+  uint64_t start_time = clock->NowMicros();
+  for (int i = 0; i < FLAGS_num_records; i++) {
+    uint64_t start_nanos = clock->NowNanos();
+    writer->Append(record);
+    writer->Flush();
+    if (FLAGS_enable_sync) {
+      writer->Sync(false);
+    }
+    hist.Add(clock->NowNanos() - start_nanos);
+
+    if (i % 1000 == 1) {
+      fprintf(stderr, "Wrote %d records...\n", i);
+    }
+
+    int time_to_sleep =
+        (i + 1) * FLAGS_record_interval - (clock->NowMicros() - start_time);
+    if (time_to_sleep > 0) {
+      clock->SleepForMicroseconds(time_to_sleep);
+    }
+  }
+
+  fprintf(stderr, "Distribution of latency of append+flush: \n%s",
+          hist.ToString().c_str());
+}
+}  // namespace ROCKSDB_NAMESPACE
+
+int main(int argc, char** argv) {
+  SetUsageMessage(std::string("\nUSAGE:\n") + std::string(argv[0]) +
+                  " [OPTIONS]...");
+  ParseCommandLineFlags(&argc, &argv, true);
+
+  ROCKSDB_NAMESPACE::RunBenchmark();
+  return 0;
+}
+
+#endif  // GFLAGS
diff --git a/src/rocksdb/util/math.h b/src/rocksdb/util/math.h
new file mode 100644
index 000000000..da31b43ec
--- /dev/null
+++ b/src/rocksdb/util/math.h
@@ -0,0 +1,294 @@
+//  Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#pragma once
+
+#include <assert.h>
+#ifdef _MSC_VER
+#include <intrin.h>
+#endif
+
+#include <cstdint>
+#include <type_traits>
+
+#include "rocksdb/rocksdb_namespace.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+// Fast implementation of floor(log2(v)). Undefined for 0 or negative
+// numbers (in case of signed type).
+template <typename T>
+inline int FloorLog2(T v) {
+  static_assert(std::is_integral<T>::value, "non-integral type");
+  assert(v > 0);
+#ifdef _MSC_VER
+  static_assert(sizeof(T) <= sizeof(uint64_t), "type too big");
+  unsigned long idx = 0;
+  if (sizeof(T) <= sizeof(uint32_t)) {
+    _BitScanReverse(&idx, static_cast<uint32_t>(v));
+  } else {
+#if defined(_M_X64) || defined(_M_ARM64)
+    _BitScanReverse64(&idx, static_cast<uint64_t>(v));
+#else
+    const auto vh = static_cast<uint32_t>(static_cast<uint64_t>(v) >> 32);
+    if (vh != 0) {
+      _BitScanReverse(&idx, static_cast<uint32_t>(vh));
+      idx += 32;
+    } else {
+      _BitScanReverse(&idx, static_cast<uint32_t>(v));
+    }
+#endif
+  }
+  return idx;
+#else
+  static_assert(sizeof(T) <= sizeof(unsigned long long), "type too big");
+  if (sizeof(T) <= sizeof(unsigned int)) {
+    int lz = __builtin_clz(static_cast<unsigned int>(v));
+    return int{sizeof(unsigned int)} * 8 - 1 - lz;
+  } else if (sizeof(T) <= sizeof(unsigned long)) {
+    int lz = __builtin_clzl(static_cast<unsigned long>(v));
+    return int{sizeof(unsigned long)} * 8 - 1 - lz;
+  } else {
+    int lz = __builtin_clzll(static_cast<unsigned long long>(v));
+    return int{sizeof(unsigned long long)} * 8 - 1 - lz;
+  }
+#endif
+}
+
+// Constexpr version of FloorLog2
+template <typename T>
+constexpr int ConstexprFloorLog2(T v) {
+  int rv = 0;
+  while (v > T{1}) {
+    ++rv;
+    v >>= 1;
+  }
+  return rv;
+}
+
+// Number of low-order zero bits before the first 1 bit. Undefined for 0.
+template <typename T>
+inline int CountTrailingZeroBits(T v) {
+  static_assert(std::is_integral<T>::value, "non-integral type");
+  assert(v != 0);
+#ifdef _MSC_VER
+  static_assert(sizeof(T) <= sizeof(uint64_t), "type too big");
+  unsigned long tz = 0;
+  if (sizeof(T) <= sizeof(uint32_t)) {
+    _BitScanForward(&tz, static_cast<uint32_t>(v));
+  } else {
+#if defined(_M_X64) || defined(_M_ARM64)
+    _BitScanForward64(&tz, static_cast<uint64_t>(v));
+#else
+    _BitScanForward(&tz, static_cast<uint32_t>(v));
+    if (tz == 0) {
+      _BitScanForward(&tz,
+                      static_cast<uint32_t>(static_cast<uint64_t>(v) >> 32));
+      tz += 32;
+    }
+#endif
+  }
+  return static_cast<int>(tz);
+#else
+  static_assert(sizeof(T) <= sizeof(unsigned long long), "type too big");
+  if (sizeof(T) <= sizeof(unsigned int)) {
+    return __builtin_ctz(static_cast<unsigned int>(v));
+  } else if (sizeof(T) <= sizeof(unsigned long)) {
+    return __builtin_ctzl(static_cast<unsigned long>(v));
+  } else {
+    return __builtin_ctzll(static_cast<unsigned long long>(v));
+  }
+#endif
+}
+
+// Not all MSVC compile settings will use `BitsSetToOneFallback()`. We include
+// the following code at coarse granularity for simpler macros. It's important
+// to exclude at least so our non-MSVC unit test coverage tool doesn't see it.
+#ifdef _MSC_VER
+
+namespace detail {
+
+template <typename T>
+int BitsSetToOneFallback(T v) {
+  const int kBits = static_cast<int>(sizeof(T)) * 8;
+  static_assert((kBits & (kBits - 1)) == 0, "must be power of two bits");
+  // we static_cast these bit patterns in order to truncate them to the correct
+  // size. Warning C4309 dislikes this technique, so disable it here.
+#pragma warning(disable : 4309)
+  v = static_cast<T>(v - ((v >> 1) & static_cast<T>(0x5555555555555555ull)));
+  v = static_cast<T>((v & static_cast<T>(0x3333333333333333ull)) +
+                     ((v >> 2) & static_cast<T>(0x3333333333333333ull)));
+  v = static_cast<T>((v + (v >> 4)) & static_cast<T>(0x0F0F0F0F0F0F0F0Full));
+#pragma warning(default : 4309)
+  for (int shift_bits = 8; shift_bits < kBits; shift_bits <<= 1) {
+    v += static_cast<T>(v >> shift_bits);
+  }
+  // we want the bottom "slot" that's big enough to represent a value up to
+  // (and including) kBits.
+  return static_cast<int>(v & static_cast<T>(kBits | (kBits - 1)));
+}
+
+}  // namespace detail
+
+#endif  // _MSC_VER
+
+// Number of bits set to 1. Also known as "population count".
+template <typename T>
+inline int BitsSetToOne(T v) {
+  static_assert(std::is_integral<T>::value, "non-integral type");
+#ifdef _MSC_VER
+  static_assert(sizeof(T) <= sizeof(uint64_t), "type too big");
+  if (sizeof(T) < sizeof(uint32_t)) {
+    // This bit mask is to avoid a compiler warning on unused path
+    constexpr auto mm = 8 * sizeof(uint32_t) - 1;
+    // The bit mask is to neutralize sign extension on small signed types
+    constexpr uint32_t m = (uint32_t{1} << ((8 * sizeof(T)) & mm)) - 1;
+#if defined(HAVE_SSE42) && (defined(_M_X64) || defined(_M_IX86))
+    return static_cast<int>(__popcnt(static_cast<uint32_t>(v) & m));
+#else
+    return static_cast<int>(detail::BitsSetToOneFallback(v) & m);
+#endif
+  } else if (sizeof(T) == sizeof(uint32_t)) {
+#if defined(HAVE_SSE42) && (defined(_M_X64) || defined(_M_IX86))
+    return static_cast<int>(__popcnt(static_cast<uint32_t>(v)));
+#else
+    return detail::BitsSetToOneFallback(static_cast<uint32_t>(v));
+#endif
+  } else {
+#if defined(HAVE_SSE42) && defined(_M_X64)
+    return static_cast<int>(__popcnt64(static_cast<uint64_t>(v)));
+#elif defined(HAVE_SSE42) && defined(_M_IX86)
+    return static_cast<int>(
+        __popcnt(static_cast<uint32_t>(static_cast<uint64_t>(v) >> 32) +
+                 __popcnt(static_cast<uint32_t>(v))));
+#else
+    return detail::BitsSetToOneFallback(static_cast<uint64_t>(v));
+#endif
+  }
+#else
+  static_assert(sizeof(T) <= sizeof(unsigned long long), "type too big");
+  if (sizeof(T) < sizeof(unsigned int)) {
+    // This bit mask is to avoid a compiler warning on unused path
+    constexpr auto mm = 8 * sizeof(unsigned int) - 1;
+    // This bit mask is to neutralize sign extension on small signed types
+    constexpr unsigned int m = (1U << ((8 * sizeof(T)) & mm)) - 1;
+    return __builtin_popcount(static_cast<unsigned int>(v) & m);
+  } else if (sizeof(T) == sizeof(unsigned int)) {
+    return __builtin_popcount(static_cast<unsigned int>(v));
+  } else if (sizeof(T) <= sizeof(unsigned long)) {
+    return __builtin_popcountl(static_cast<unsigned long>(v));
+  } else {
+    return __builtin_popcountll(static_cast<unsigned long long>(v));
+  }
+#endif
+}
+
+template <typename T>
+inline int BitParity(T v) {
+  static_assert(std::is_integral<T>::value, "non-integral type");
+#ifdef _MSC_VER
+  // bit parity == oddness of popcount
+  return BitsSetToOne(v) & 1;
+#else
+  static_assert(sizeof(T) <= sizeof(unsigned long long), "type too big");
+  if (sizeof(T) <= sizeof(unsigned int)) {
+    // On any sane systen, potential sign extension here won't change parity
+    return __builtin_parity(static_cast<unsigned int>(v));
+  } else if (sizeof(T) <= sizeof(unsigned long)) {
+    return __builtin_parityl(static_cast<unsigned long>(v));
+  } else {
+    return __builtin_parityll(static_cast<unsigned long long>(v));
+  }
+#endif
+}
+
+// Swaps between big and little endian. Can be used in combination with the
+// little-endian encoding/decoding functions in coding_lean.h and coding.h to
+// encode/decode big endian.
+template <typename T>
+inline T EndianSwapValue(T v) {
+  static_assert(std::is_integral<T>::value, "non-integral type");
+
+#ifdef _MSC_VER
+  if (sizeof(T) == 2) {
+    return static_cast<T>(_byteswap_ushort(static_cast<uint16_t>(v)));
+  } else if (sizeof(T) == 4) {
+    return static_cast<T>(_byteswap_ulong(static_cast<uint32_t>(v)));
+  } else if (sizeof(T) == 8) {
+    return static_cast<T>(_byteswap_uint64(static_cast<uint64_t>(v)));
+  }
+#else
+  if (sizeof(T) == 2) {
+    return static_cast<T>(__builtin_bswap16(static_cast<uint16_t>(v)));
+  } else if (sizeof(T) == 4) {
+    return static_cast<T>(__builtin_bswap32(static_cast<uint32_t>(v)));
+  } else if (sizeof(T) == 8) {
+    return static_cast<T>(__builtin_bswap64(static_cast<uint64_t>(v)));
+  }
+#endif
+  // Recognized by clang as bswap, but not by gcc :(
+  T ret_val = 0;
+  for (std::size_t i = 0; i < sizeof(T); ++i) {
+    ret_val |= ((v >> (8 * i)) & 0xff) << (8 * (sizeof(T) - 1 - i));
+  }
+  return ret_val;
+}
+
+// Reverses the order of bits in an integral value
+template <typename T>
+inline T ReverseBits(T v) {
+  T r = EndianSwapValue(v);
+  const T kHighestByte = T{1} << ((sizeof(T) - 1) * 8);
+  const T kEveryByte = kHighestByte | (kHighestByte / 255);
+
+  r = ((r & (kEveryByte * 0x0f)) << 4) | ((r >> 4) & (kEveryByte * 0x0f));
+  r = ((r & (kEveryByte * 0x33)) << 2) | ((r >> 2) & (kEveryByte * 0x33));
+  r = ((r & (kEveryByte * 0x55)) << 1) | ((r >> 1) & (kEveryByte * 0x55));
+
+  return r;
+}
+
+// Every output bit depends on many input bits in the same and higher
+// positions, but not lower positions. Specifically, this function
+// * Output highest bit set to 1 is same as input (same FloorLog2, or
+//   equivalently, same number of leading zeros)
+// * Is its own inverse (an involution)
+// * Guarantees that b bottom bits of v and c bottom bits of
+//   DownwardInvolution(v) uniquely identify b + c bottom bits of v
+//   (which is all of v if v < 2**(b + c)).
+// ** A notable special case is that modifying c adjacent bits at
+//    some chosen position in the input is bijective with the bottom c
+//    output bits.
+// * Distributes over xor, as in DI(a ^ b) == DI(a) ^ DI(b)
+//
+// This transformation is equivalent to a matrix*vector multiplication in
+// GF(2) where the matrix is recursively defined by the pattern matrix
+// P = | 1 1 |
+//     | 0 1 |
+// and replacing 1's with P and 0's with 2x2 zero matices to some depth,
+// e.g. depth of 6 for 64-bit T. An essential feature of this matrix
+// is that all square sub-matrices that include the top row are invertible.
+template <typename T>
+inline T DownwardInvolution(T v) {
+  static_assert(std::is_integral<T>::value, "non-integral type");
+  static_assert(sizeof(T) <= 8, "only supported up to 64 bits");
+
+  uint64_t r = static_cast<uint64_t>(v);
+  if constexpr (sizeof(T) > 4) {
+    r ^= r >> 32;
+  }
+  if constexpr (sizeof(T) > 2) {
+    r ^= (r & 0xffff0000ffff0000U) >> 16;
+  }
+  if constexpr (sizeof(T) > 1) {
+    r ^= (r & 0xff00ff00ff00ff00U) >> 8;
+  }
+  r ^= (r & 0xf0f0f0f0f0f0f0f0U) >> 4;
+  r ^= (r & 0xccccccccccccccccU) >> 2;
+  r ^= (r & 0xaaaaaaaaaaaaaaaaU) >> 1;
+  return static_cast<T>(r);
+}
+
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/util/math128.h b/src/rocksdb/util/math128.h
new file mode 100644
index 000000000..ae490051a
--- /dev/null
+++ b/src/rocksdb/util/math128.h
@@ -0,0 +1,316 @@
+//  Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#pragma once
+
+#include "util/coding_lean.h"
+#include "util/math.h"
+
+#ifdef TEST_UINT128_COMPAT
+#undef HAVE_UINT128_EXTENSION
+#endif
+
+namespace ROCKSDB_NAMESPACE {
+
+// Unsigned128 is a 128 bit value supporting (at least) bitwise operators,
+// shifts, and comparisons. __uint128_t is not always available.
+
+#ifdef HAVE_UINT128_EXTENSION
+using Unsigned128 = __uint128_t;
+#else
+struct Unsigned128 {
+  uint64_t lo;
+  uint64_t hi;
+
+  inline Unsigned128() {
+    static_assert(sizeof(Unsigned128) == 2 * sizeof(uint64_t),
+                  "unexpected overhead in representation");
+    lo = 0;
+    hi = 0;
+  }
+
+  inline Unsigned128(uint64_t lower) {
+    lo = lower;
+    hi = 0;
+  }
+
+  inline Unsigned128(uint64_t lower, uint64_t upper) {
+    lo = lower;
+    hi = upper;
+  }
+
+  explicit operator uint64_t() { return lo; }
+
+  explicit operator uint32_t() { return static_cast<uint32_t>(lo); }
+
+  explicit operator uint16_t() { return static_cast<uint16_t>(lo); }
+
+  explicit operator uint8_t() { return static_cast<uint8_t>(lo); }
+};
+
+inline Unsigned128 operator<<(const Unsigned128& lhs, unsigned shift) {
+  shift &= 127;
+  Unsigned128 rv;
+  if (shift >= 64) {
+    rv.lo = 0;
+    rv.hi = lhs.lo << (shift & 63);
+  } else {
+    uint64_t tmp = lhs.lo;
+    rv.lo = tmp << shift;
+    // Ensure shift==0 shifts away everything. (This avoids another
+    // conditional branch on shift == 0.)
+    tmp = tmp >> 1 >> (63 - shift);
+    rv.hi = tmp | (lhs.hi << shift);
+  }
+  return rv;
+}
+
+inline Unsigned128& operator<<=(Unsigned128& lhs, unsigned shift) {
+  lhs = lhs << shift;
+  return lhs;
+}
+
+inline Unsigned128 operator>>(const Unsigned128& lhs, unsigned shift) {
+  shift &= 127;
+  Unsigned128 rv;
+  if (shift >= 64) {
+    rv.hi = 0;
+    rv.lo = lhs.hi >> (shift & 63);
+  } else {
+    uint64_t tmp = lhs.hi;
+    rv.hi = tmp >> shift;
+    // Ensure shift==0 shifts away everything
+    tmp = tmp << 1 << (63 - shift);
+    rv.lo = tmp | (lhs.lo >> shift);
+  }
+  return rv;
+}
+
+inline Unsigned128& operator>>=(Unsigned128& lhs, unsigned shift) {
+  lhs = lhs >> shift;
+  return lhs;
+}
+
+inline Unsigned128 operator&(const Unsigned128& lhs, const Unsigned128& rhs) {
+  return Unsigned128(lhs.lo & rhs.lo, lhs.hi & rhs.hi);
+}
+
+inline Unsigned128& operator&=(Unsigned128& lhs, const Unsigned128& rhs) {
+  lhs = lhs & rhs;
+  return lhs;
+}
+
+inline Unsigned128 operator|(const Unsigned128& lhs, const Unsigned128& rhs) {
+  return Unsigned128(lhs.lo | rhs.lo, lhs.hi | rhs.hi);
+}
+
+inline Unsigned128& operator|=(Unsigned128& lhs, const Unsigned128& rhs) {
+  lhs = lhs | rhs;
+  return lhs;
+}
+
+inline Unsigned128 operator^(const Unsigned128& lhs, const Unsigned128& rhs) {
+  return Unsigned128(lhs.lo ^ rhs.lo, lhs.hi ^ rhs.hi);
+}
+
+inline Unsigned128& operator^=(Unsigned128& lhs, const Unsigned128& rhs) {
+  lhs = lhs ^ rhs;
+  return lhs;
+}
+
+inline Unsigned128 operator~(const Unsigned128& v) {
+  return Unsigned128(~v.lo, ~v.hi);
+}
+
+inline bool operator==(const Unsigned128& lhs, const Unsigned128& rhs) {
+  return lhs.lo == rhs.lo && lhs.hi == rhs.hi;
+}
+
+inline bool operator!=(const Unsigned128& lhs, const Unsigned128& rhs) {
+  return lhs.lo != rhs.lo || lhs.hi != rhs.hi;
+}
+
+inline bool operator>(const Unsigned128& lhs, const Unsigned128& rhs) {
+  return lhs.hi > rhs.hi || (lhs.hi == rhs.hi && lhs.lo > rhs.lo);
+}
+
+inline bool operator<(const Unsigned128& lhs, const Unsigned128& rhs) {
+  return lhs.hi < rhs.hi || (lhs.hi == rhs.hi && lhs.lo < rhs.lo);
+}
+
+inline bool operator>=(const Unsigned128& lhs, const Unsigned128& rhs) {
+  return lhs.hi > rhs.hi || (lhs.hi == rhs.hi && lhs.lo >= rhs.lo);
+}
+
+inline bool operator<=(const Unsigned128& lhs, const Unsigned128& rhs) {
+  return lhs.hi < rhs.hi || (lhs.hi == rhs.hi && lhs.lo <= rhs.lo);
+}
+#endif
+
+inline uint64_t Lower64of128(Unsigned128 v) {
+#ifdef HAVE_UINT128_EXTENSION
+  return static_cast<uint64_t>(v);
+#else
+  return v.lo;
+#endif
+}
+
+inline uint64_t Upper64of128(Unsigned128 v) {
+#ifdef HAVE_UINT128_EXTENSION
+  return static_cast<uint64_t>(v >> 64);
+#else
+  return v.hi;
+#endif
+}
+
+// This generally compiles down to a single fast instruction on 64-bit.
+// This doesn't really make sense as operator* because it's not a
+// general 128x128 multiply and provides more output than 64x64 multiply.
+inline Unsigned128 Multiply64to128(uint64_t a, uint64_t b) {
+#ifdef HAVE_UINT128_EXTENSION
+  return Unsigned128{a} * Unsigned128{b};
+#else
+  // Full decomposition
+  // NOTE: GCC seems to fully understand this code as 64-bit x 64-bit
+  // -> 128-bit multiplication and optimize it appropriately.
+  uint64_t tmp = uint64_t{b & 0xffffFFFF} * uint64_t{a & 0xffffFFFF};
+  uint64_t lower = tmp & 0xffffFFFF;
+  tmp >>= 32;
+  tmp += uint64_t{b & 0xffffFFFF} * uint64_t{a >> 32};
+  // Avoid overflow: first add lower 32 of tmp2, and later upper 32
+  uint64_t tmp2 = uint64_t{b >> 32} * uint64_t{a & 0xffffFFFF};
+  tmp += tmp2 & 0xffffFFFF;
+  lower |= tmp << 32;
+  tmp >>= 32;
+  tmp += tmp2 >> 32;
+  tmp += uint64_t{b >> 32} * uint64_t{a >> 32};
+  return Unsigned128(lower, tmp);
+#endif
+}
+
+template <>
+inline int FloorLog2(Unsigned128 v) {
+  if (Upper64of128(v) == 0) {
+    return FloorLog2(Lower64of128(v));
+  } else {
+    return FloorLog2(Upper64of128(v)) + 64;
+  }
+}
+
+template <>
+inline int CountTrailingZeroBits(Unsigned128 v) {
+  if (Lower64of128(v) != 0) {
+    return CountTrailingZeroBits(Lower64of128(v));
+  } else {
+    return CountTrailingZeroBits(Upper64of128(v)) + 64;
+  }
+}
+
+template <>
+inline int BitsSetToOne(Unsigned128 v) {
+  return BitsSetToOne(Lower64of128(v)) + BitsSetToOne(Upper64of128(v));
+}
+
+template <>
+inline int BitParity(Unsigned128 v) {
+  return BitParity(Lower64of128(v) ^ Upper64of128(v));
+}
+
+template <>
+inline Unsigned128 EndianSwapValue(Unsigned128 v) {
+  return (Unsigned128{EndianSwapValue(Lower64of128(v))} << 64) |
+         EndianSwapValue(Upper64of128(v));
+}
+
+template <>
+inline Unsigned128 ReverseBits(Unsigned128 v) {
+  return (Unsigned128{ReverseBits(Lower64of128(v))} << 64) |
+         ReverseBits(Upper64of128(v));
+}
+
+template <>
+inline Unsigned128 DownwardInvolution(Unsigned128 v) {
+  return (Unsigned128{DownwardInvolution(Upper64of128(v))} << 64) |
+         DownwardInvolution(Upper64of128(v) ^ Lower64of128(v));
+}
+
+template <typename T>
+struct IsUnsignedUpTo128
+    : std::integral_constant<bool, std::is_unsigned<T>::value ||
+                                       std::is_same<T, Unsigned128>::value> {};
+
+inline void EncodeFixed128(char* dst, Unsigned128 value) {
+  EncodeFixed64(dst, Lower64of128(value));
+  EncodeFixed64(dst + 8, Upper64of128(value));
+}
+
+inline Unsigned128 DecodeFixed128(const char* ptr) {
+  Unsigned128 rv = DecodeFixed64(ptr + 8);
+  return (rv << 64) | DecodeFixed64(ptr);
+}
+
+// A version of EncodeFixed* for generic algorithms. Likely to be used
+// with Unsigned128, so lives here for now.
+template <typename T>
+inline void EncodeFixedGeneric(char* /*dst*/, T /*value*/) {
+  // Unfortunately, GCC does not appear to optimize this simple code down
+  // to a trivial load on Intel:
+  //
+  // T ret_val = 0;
+  // for (size_t i = 0; i < sizeof(T); ++i) {
+  //   ret_val |= (static_cast<T>(static_cast<unsigned char>(ptr[i])) << (8 *
+  //   i));
+  // }
+  // return ret_val;
+  //
+  // But does unroll the loop, and does optimize manually unrolled version
+  // for specific sizes down to a trivial load. I have no idea why it doesn't
+  // do both on this code.
+
+  // So instead, we rely on specializations
+  static_assert(sizeof(T) == 0, "No specialization provided for this type");
+}
+
+template <>
+inline void EncodeFixedGeneric(char* dst, uint16_t value) {
+  return EncodeFixed16(dst, value);
+}
+template <>
+inline void EncodeFixedGeneric(char* dst, uint32_t value) {
+  return EncodeFixed32(dst, value);
+}
+template <>
+inline void EncodeFixedGeneric(char* dst, uint64_t value) {
+  return EncodeFixed64(dst, value);
+}
+template <>
+inline void EncodeFixedGeneric(char* dst, Unsigned128 value) {
+  return EncodeFixed128(dst, value);
+}
+
+// A version of EncodeFixed* for generic algorithms.
+template <typename T>
+inline T DecodeFixedGeneric(const char* /*dst*/) {
+  static_assert(sizeof(T) == 0, "No specialization provided for this type");
+}
+
+template <>
+inline uint16_t DecodeFixedGeneric(const char* dst) {
+  return DecodeFixed16(dst);
+}
+template <>
+inline uint32_t DecodeFixedGeneric(const char* dst) {
+  return DecodeFixed32(dst);
+}
+template <>
+inline uint64_t DecodeFixedGeneric(const char* dst) {
+  return DecodeFixed64(dst);
+}
+template <>
+inline Unsigned128 DecodeFixedGeneric(const char* dst) {
+  return DecodeFixed128(dst);
+}
+
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/util/murmurhash.cc b/src/rocksdb/util/murmurhash.cc
new file mode 100644
index 000000000..a69f3918a
--- /dev/null
+++ b/src/rocksdb/util/murmurhash.cc
@@ -0,0 +1,196 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+//
+/*
+  Murmurhash from http://sites.google.com/site/murmurhash/
+
+  All code is released to the public domain. For business purposes, Murmurhash
+  is under the MIT license.
+*/
+#include "murmurhash.h"
+
+#include "port/lang.h"
+
+#if defined(__x86_64__)
+
+// -------------------------------------------------------------------
+//
+// The same caveats as 32-bit MurmurHash2 apply here - beware of alignment
+// and endian-ness issues if used across multiple platforms.
+//
+// 64-bit hash for 64-bit platforms
+
+#ifdef ROCKSDB_UBSAN_RUN
+#if defined(__clang__)
+__attribute__((__no_sanitize__("alignment")))
+#elif defined(__GNUC__)
+__attribute__((__no_sanitize_undefined__))
+#endif
+#endif
+// clang-format off
+uint64_t MurmurHash64A ( const void * key, int len, unsigned int seed )
+{
+    const uint64_t m = 0xc6a4a7935bd1e995;
+    const int r = 47;
+
+    uint64_t h = seed ^ (len * m);
+
+    const uint64_t * data = (const uint64_t *)key;
+    const uint64_t * end = data + (len/8);
+
+    while(data != end)
+    {
+        uint64_t k = *data++;
+
+        k *= m;
+        k ^= k >> r;
+        k *= m;
+
+        h ^= k;
+        h *= m;
+    }
+
+    const unsigned char * data2 = (const unsigned char*)data;
+
+    switch(len & 7)
+    {
+    case 7: h ^= ((uint64_t)data2[6]) << 48; FALLTHROUGH_INTENDED;
+    case 6: h ^= ((uint64_t)data2[5]) << 40; FALLTHROUGH_INTENDED;
+    case 5: h ^= ((uint64_t)data2[4]) << 32; FALLTHROUGH_INTENDED;
+    case 4: h ^= ((uint64_t)data2[3]) << 24; FALLTHROUGH_INTENDED;
+    case 3: h ^= ((uint64_t)data2[2]) << 16; FALLTHROUGH_INTENDED;
+    case 2: h ^= ((uint64_t)data2[1]) << 8;  FALLTHROUGH_INTENDED;
+    case 1: h ^= ((uint64_t)data2[0]);
+        h *= m;
+    };
+
+    h ^= h >> r;
+    h *= m;
+    h ^= h >> r;
+
+    return h;
+}
+// clang-format on
+
+#elif defined(__i386__)
+
+// -------------------------------------------------------------------
+//
+// Note - This code makes a few assumptions about how your machine behaves -
+//
+// 1. We can read a 4-byte value from any address without crashing
+// 2. sizeof(int) == 4
+//
+// And it has a few limitations -
+//
+// 1. It will not work incrementally.
+// 2. It will not produce the same results on little-endian and big-endian
+//    machines.
+// clang-format off
+unsigned int MurmurHash2 ( const void * key, int len, unsigned int seed )
+{
+    // 'm' and 'r' are mixing constants generated offline.
+    // They're not really 'magic', they just happen to work well.
+
+    const unsigned int m = 0x5bd1e995;
+    const int r = 24;
+
+    // Initialize the hash to a 'random' value
+
+    unsigned int h = seed ^ len;
+
+    // Mix 4 bytes at a time into the hash
+
+    const unsigned char * data = (const unsigned char *)key;
+
+    while(len >= 4)
+    {
+        unsigned int k = *(unsigned int *)data;
+
+        k *= m;
+        k ^= k >> r;
+        k *= m;
+
+        h *= m;
+        h ^= k;
+
+        data += 4;
+        len -= 4;
+    }
+
+    // Handle the last few bytes of the input array
+
+    switch(len)
+    {
+    case 3: h ^= data[2] << 16; FALLTHROUGH_INTENDED;
+    case 2: h ^= data[1] << 8;  FALLTHROUGH_INTENDED;
+    case 1: h ^= data[0];
+        h *= m;
+    };
+
+    // Do a few final mixes of the hash to ensure the last few
+    // bytes are well-incorporated.
+
+    h ^= h >> 13;
+    h *= m;
+    h ^= h >> 15;
+
+    return h;
+}
+// clang-format on
+
+#else
+
+// -------------------------------------------------------------------
+//
+// Same as MurmurHash2, but endian- and alignment-neutral.
+// Half the speed though, alas.
+// clang-format off
+unsigned int MurmurHashNeutral2 ( const void * key, int len, unsigned int seed )
+{
+    const unsigned int m = 0x5bd1e995;
+    const int r = 24;
+
+    unsigned int h = seed ^ len;
+
+    const unsigned char * data = (const unsigned char *)key;
+
+    while(len >= 4)
+    {
+        unsigned int k;
+
+        k  = data[0];
+        k |= data[1] << 8;
+        k |= data[2] << 16;
+        k |= data[3] << 24;
+
+        k *= m;
+        k ^= k >> r;
+        k *= m;
+
+        h *= m;
+        h ^= k;
+
+        data += 4;
+        len -= 4;
+    }
+
+    switch(len)
+    {
+    case 3: h ^= data[2] << 16; FALLTHROUGH_INTENDED;
+    case 2: h ^= data[1] << 8;  FALLTHROUGH_INTENDED;
+    case 1: h ^= data[0];
+        h *= m;
+    };
+
+    h ^= h >> 13;
+    h *= m;
+    h ^= h >> 15;
+
+    return h;
+}
+// clang-format on
+
+#endif
diff --git a/src/rocksdb/util/murmurhash.h b/src/rocksdb/util/murmurhash.h
new file mode 100644
index 000000000..7ef4cbbec
--- /dev/null
+++ b/src/rocksdb/util/murmurhash.h
@@ -0,0 +1,43 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+//
+/*
+  Murmurhash from http://sites.google.com/site/murmurhash/
+
+  All code is released to the public domain. For business purposes, Murmurhash
+  is under the MIT license.
+*/
+#pragma once
+#include <stdint.h>
+
+#include "rocksdb/slice.h"
+
+#if defined(__x86_64__)
+#define MURMUR_HASH MurmurHash64A
+uint64_t MurmurHash64A(const void* key, int len, unsigned int seed);
+#define MurmurHash MurmurHash64A
+using murmur_t = uint64_t;
+
+#elif defined(__i386__)
+#define MURMUR_HASH MurmurHash2
+unsigned int MurmurHash2(const void* key, int len, unsigned int seed);
+#define MurmurHash MurmurHash2
+using murmur_t = unsigned int;
+
+#else
+#define MURMUR_HASH MurmurHashNeutral2
+unsigned int MurmurHashNeutral2(const void* key, int len, unsigned int seed);
+#define MurmurHash MurmurHashNeutral2
+using murmur_t = unsigned int;
+#endif
+
+// Allow slice to be hashable by murmur hash.
+namespace ROCKSDB_NAMESPACE {
+struct murmur_hash {
+  size_t operator()(const Slice& slice) const {
+    return MurmurHash(slice.data(), static_cast<int>(slice.size()), 0);
+  }
+};
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/util/mutexlock.h b/src/rocksdb/util/mutexlock.h
new file mode 100644
index 000000000..94066b29e
--- /dev/null
+++ b/src/rocksdb/util/mutexlock.h
@@ -0,0 +1,180 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+//
+// Copyright (c) 2011 The LevelDB Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file. See the AUTHORS file for names of contributors.
+
+#pragma once
+#include <assert.h>
+
+#include <atomic>
+#include <mutex>
+#include <thread>
+
+#include "port/port.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+// Helper class that locks a mutex on construction and unlocks the mutex when
+// the destructor of the MutexLock object is invoked.
+//
+// Typical usage:
+//
+//   void MyClass::MyMethod() {
+//     MutexLock l(&mu_);       // mu_ is an instance variable
+//     ... some complex code, possibly with multiple return paths ...
+//   }
+
+class MutexLock {
+ public:
+  explicit MutexLock(port::Mutex *mu) : mu_(mu) { this->mu_->Lock(); }
+  // No copying allowed
+  MutexLock(const MutexLock &) = delete;
+  void operator=(const MutexLock &) = delete;
+
+  ~MutexLock() { this->mu_->Unlock(); }
+
+ private:
+  port::Mutex *const mu_;
+};
+
+//
+// Acquire a ReadLock on the specified RWMutex.
+// The Lock will be automatically released when the
+// object goes out of scope.
+//
+class ReadLock {
+ public:
+  explicit ReadLock(port::RWMutex *mu) : mu_(mu) { this->mu_->ReadLock(); }
+  // No copying allowed
+  ReadLock(const ReadLock &) = delete;
+  void operator=(const ReadLock &) = delete;
+
+  ~ReadLock() { this->mu_->ReadUnlock(); }
+
+ private:
+  port::RWMutex *const mu_;
+};
+
+//
+// Automatically unlock a locked mutex when the object is destroyed
+//
+class ReadUnlock {
+ public:
+  explicit ReadUnlock(port::RWMutex *mu) : mu_(mu) { mu->AssertHeld(); }
+  // No copying allowed
+  ReadUnlock(const ReadUnlock &) = delete;
+  ReadUnlock &operator=(const ReadUnlock &) = delete;
+
+  ~ReadUnlock() { mu_->ReadUnlock(); }
+
+ private:
+  port::RWMutex *const mu_;
+};
+
+//
+// Acquire a WriteLock on the specified RWMutex.
+// The Lock will be automatically released then the
+// object goes out of scope.
+//
+class WriteLock {
+ public:
+  explicit WriteLock(port::RWMutex *mu) : mu_(mu) { this->mu_->WriteLock(); }
+  // No copying allowed
+  WriteLock(const WriteLock &) = delete;
+  void operator=(const WriteLock &) = delete;
+
+  ~WriteLock() { this->mu_->WriteUnlock(); }
+
+ private:
+  port::RWMutex *const mu_;
+};
+
+//
+// SpinMutex has very low overhead for low-contention cases.  Method names
+// are chosen so you can use std::unique_lock or std::lock_guard with it.
+//
+class SpinMutex {
+ public:
+  SpinMutex() : locked_(false) {}
+
+  bool try_lock() {
+    auto currently_locked = locked_.load(std::memory_order_relaxed);
+    return !currently_locked &&
+           locked_.compare_exchange_weak(currently_locked, true,
+                                         std::memory_order_acquire,
+                                         std::memory_order_relaxed);
+  }
+
+  void lock() {
+    for (size_t tries = 0;; ++tries) {
+      if (try_lock()) {
+        // success
+        break;
+      }
+      port::AsmVolatilePause();
+      if (tries > 100) {
+        std::this_thread::yield();
+      }
+    }
+  }
+
+  void unlock() { locked_.store(false, std::memory_order_release); }
+
+ private:
+  std::atomic<bool> locked_;
+};
+
+// We want to prevent false sharing
+template <class T>
+struct ALIGN_AS(CACHE_LINE_SIZE) LockData {
+  T lock_;
+};
+
+//
+// Inspired by Guava: https://github.com/google/guava/wiki/StripedExplained
+// A striped Lock. This offers the underlying lock striping similar
+// to that of ConcurrentHashMap in a reusable form, and extends it for
+// semaphores and read-write locks. Conceptually, lock striping is the technique
+// of dividing a lock into many <i>stripes</i>, increasing the granularity of a
+// single lock and allowing independent operations to lock different stripes and
+// proceed concurrently, instead of creating contention for a single lock.
+//
+template <class T, class P>
+class Striped {
+ public:
+  Striped(size_t stripes, std::function<uint64_t(const P &)> hash)
+      : stripes_(stripes), hash_(hash) {
+    locks_ = reinterpret_cast<LockData<T> *>(
+        port::cacheline_aligned_alloc(sizeof(LockData<T>) * stripes));
+    for (size_t i = 0; i < stripes; i++) {
+      new (&locks_[i]) LockData<T>();
+    }
+  }
+
+  virtual ~Striped() {
+    if (locks_ != nullptr) {
+      assert(stripes_ > 0);
+      for (size_t i = 0; i < stripes_; i++) {
+        locks_[i].~LockData<T>();
+      }
+      port::cacheline_aligned_free(locks_);
+    }
+  }
+
+  T *get(const P &key) {
+    uint64_t h = hash_(key);
+    size_t index = h % stripes_;
+    return &reinterpret_cast<LockData<T> *>(&locks_[index])->lock_;
+  }
+
+ private:
+  size_t stripes_;
+  LockData<T> *locks_;
+  std::function<uint64_t(const P &)> hash_;
+};
+
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/util/ppc-opcode.h b/src/rocksdb/util/ppc-opcode.h
new file mode 100644
index 000000000..5cc5af0e3
--- /dev/null
+++ b/src/rocksdb/util/ppc-opcode.h
@@ -0,0 +1,27 @@
+//  Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.
+//  Copyright (c) 2017 International Business Machines Corp.
+//  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#pragma once
+
+#define __PPC_RA(a) (((a)&0x1f) << 16)
+#define __PPC_RB(b) (((b)&0x1f) << 11)
+#define __PPC_XA(a) ((((a)&0x1f) << 16) | (((a)&0x20) >> 3))
+#define __PPC_XB(b) ((((b)&0x1f) << 11) | (((b)&0x20) >> 4))
+#define __PPC_XS(s) ((((s)&0x1f) << 21) | (((s)&0x20) >> 5))
+#define __PPC_XT(s) __PPC_XS(s)
+#define VSX_XX3(t, a, b) (__PPC_XT(t) | __PPC_XA(a) | __PPC_XB(b))
+#define VSX_XX1(s, a, b) (__PPC_XS(s) | __PPC_RA(a) | __PPC_RB(b))
+
+#define PPC_INST_VPMSUMW 0x10000488
+#define PPC_INST_VPMSUMD 0x100004c8
+#define PPC_INST_MFVSRD 0x7c000066
+#define PPC_INST_MTVSRD 0x7c000166
+
+#define VPMSUMW(t, a, b) .long PPC_INST_VPMSUMW | VSX_XX3((t), a, b)
+#define VPMSUMD(t, a, b) .long PPC_INST_VPMSUMD | VSX_XX3((t), a, b)
+#define MFVRD(a, t) .long PPC_INST_MFVSRD | VSX_XX1((t) + 32, a, 0)
+#define MTVRD(t, a) .long PPC_INST_MTVSRD | VSX_XX1((t) + 32, a, 0)
diff --git a/src/rocksdb/util/random.cc b/src/rocksdb/util/random.cc
new file mode 100644
index 000000000..c94c28dfb
--- /dev/null
+++ b/src/rocksdb/util/random.cc
@@ -0,0 +1,62 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+//
+
+#include "util/random.h"
+
+#include <stdint.h>
+#include <string.h>
+
+#include <thread>
+#include <utility>
+
+#include "port/likely.h"
+#include "util/thread_local.h"
+
+#define STORAGE_DECL static thread_local
+
+namespace ROCKSDB_NAMESPACE {
+
+Random* Random::GetTLSInstance() {
+  STORAGE_DECL Random* tls_instance;
+  STORAGE_DECL std::aligned_storage<sizeof(Random)>::type tls_instance_bytes;
+
+  auto rv = tls_instance;
+  if (UNLIKELY(rv == nullptr)) {
+    size_t seed = std::hash<std::thread::id>()(std::this_thread::get_id());
+    rv = new (&tls_instance_bytes) Random((uint32_t)seed);
+    tls_instance = rv;
+  }
+  return rv;
+}
+
+std::string Random::HumanReadableString(int len) {
+  std::string ret;
+  ret.resize(len);
+  for (int i = 0; i < len; ++i) {
+    ret[i] = static_cast<char>('a' + Uniform(26));
+  }
+  return ret;
+}
+
+std::string Random::RandomString(int len) {
+  std::string ret;
+  ret.resize(len);
+  for (int i = 0; i < len; i++) {
+    ret[i] = static_cast<char>(' ' + Uniform(95));  // ' ' .. '~'
+  }
+  return ret;
+}
+
+std::string Random::RandomBinaryString(int len) {
+  std::string ret;
+  ret.resize(len);
+  for (int i = 0; i < len; i++) {
+    ret[i] = static_cast<char>(Uniform(CHAR_MAX));
+  }
+  return ret;
+}
+
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/util/random.h b/src/rocksdb/util/random.h
new file mode 100644
index 000000000..8923bdc4f
--- /dev/null
+++ b/src/rocksdb/util/random.h
@@ -0,0 +1,190 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+//
+// Copyright (c) 2011 The LevelDB Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file. See the AUTHORS file for names of contributors.
+
+#pragma once
+#include <stdint.h>
+
+#include <algorithm>
+#include <random>
+
+#include "rocksdb/rocksdb_namespace.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+// A very simple random number generator.  Not especially good at
+// generating truly random bits, but good enough for our needs in this
+// package.
+class Random {
+ private:
+  enum : uint32_t {
+    M = 2147483647L  // 2^31-1
+  };
+  enum : uint64_t {
+    A = 16807  // bits 14, 8, 7, 5, 2, 1, 0
+  };
+
+  uint32_t seed_;
+
+  static uint32_t GoodSeed(uint32_t s) { return (s & M) != 0 ? (s & M) : 1; }
+
+ public:
+  // This is the largest value that can be returned from Next()
+  enum : uint32_t { kMaxNext = M };
+
+  explicit Random(uint32_t s) : seed_(GoodSeed(s)) {}
+
+  void Reset(uint32_t s) { seed_ = GoodSeed(s); }
+
+  uint32_t Next() {
+    // We are computing
+    //       seed_ = (seed_ * A) % M,    where M = 2^31-1
+    //
+    // seed_ must not be zero or M, or else all subsequent computed values
+    // will be zero or M respectively.  For all other values, seed_ will end
+    // up cycling through every number in [1,M-1]
+    uint64_t product = seed_ * A;
+
+    // Compute (product % M) using the fact that ((x << 31) % M) == x.
+    seed_ = static_cast<uint32_t>((product >> 31) + (product & M));
+    // The first reduction may overflow by 1 bit, so we may need to
+    // repeat.  mod == M is not possible; using > allows the faster
+    // sign-bit-based test.
+    if (seed_ > M) {
+      seed_ -= M;
+    }
+    return seed_;
+  }
+
+  uint64_t Next64() { return (uint64_t{Next()} << 32) | Next(); }
+
+  // Returns a uniformly distributed value in the range [0..n-1]
+  // REQUIRES: n > 0
+  uint32_t Uniform(int n) { return Next() % n; }
+
+  // Randomly returns true ~"1/n" of the time, and false otherwise.
+  // REQUIRES: n > 0
+  bool OneIn(int n) { return Uniform(n) == 0; }
+
+  // "Optional" one-in-n, where 0 or negative always returns false
+  // (may or may not consume a random value)
+  bool OneInOpt(int n) { return n > 0 && OneIn(n); }
+
+  // Returns random bool that is true for the given percentage of
+  // calls on average. Zero or less is always false and 100 or more
+  // is always true (may or may not consume a random value)
+  bool PercentTrue(int percentage) {
+    return static_cast<int>(Uniform(100)) < percentage;
+  }
+
+  // Skewed: pick "base" uniformly from range [0,max_log] and then
+  // return "base" random bits.  The effect is to pick a number in the
+  // range [0,2^max_log-1] with exponential bias towards smaller numbers.
+  uint32_t Skewed(int max_log) { return Uniform(1 << Uniform(max_log + 1)); }
+
+  // Returns a random string of length "len"
+  std::string RandomString(int len);
+
+  // Generates a random string of len bytes using human-readable characters
+  std::string HumanReadableString(int len);
+
+  // Generates a random binary data
+  std::string RandomBinaryString(int len);
+
+  // Returns a Random instance for use by the current thread without
+  // additional locking
+  static Random* GetTLSInstance();
+};
+
+// A good 32-bit random number generator based on std::mt19937.
+// This exists in part to avoid compiler variance in warning about coercing
+// uint_fast32_t from mt19937 to uint32_t.
+class Random32 {
+ private:
+  std::mt19937 generator_;
+
+ public:
+  explicit Random32(uint32_t s) : generator_(s) {}
+
+  // Generates the next random number
+  uint32_t Next() { return static_cast<uint32_t>(generator_()); }
+
+  // Returns a uniformly distributed value in the range [0..n-1]
+  // REQUIRES: n > 0
+  uint32_t Uniform(uint32_t n) {
+    return static_cast<uint32_t>(
+        std::uniform_int_distribution<std::mt19937::result_type>(
+            0, n - 1)(generator_));
+  }
+
+  // Returns an *almost* uniformly distributed value in the range [0..n-1].
+  // Much faster than Uniform().
+  // REQUIRES: n > 0
+  uint32_t Uniformish(uint32_t n) {
+    // fastrange (without the header)
+    return static_cast<uint32_t>((uint64_t(generator_()) * uint64_t(n)) >> 32);
+  }
+
+  // Randomly returns true ~"1/n" of the time, and false otherwise.
+  // REQUIRES: n > 0
+  bool OneIn(uint32_t n) { return Uniform(n) == 0; }
+
+  // Skewed: pick "base" uniformly from range [0,max_log] and then
+  // return "base" random bits.  The effect is to pick a number in the
+  // range [0,2^max_log-1] with exponential bias towards smaller numbers.
+  uint32_t Skewed(int max_log) {
+    return Uniform(uint32_t{1} << Uniform(max_log + 1));
+  }
+
+  // Reset the seed of the generator to the given value
+  void Seed(uint32_t new_seed) { generator_.seed(new_seed); }
+};
+
+// A good 64-bit random number generator based on std::mt19937_64
+class Random64 {
+ private:
+  std::mt19937_64 generator_;
+
+ public:
+  explicit Random64(uint64_t s) : generator_(s) {}
+
+  // Generates the next random number
+  uint64_t Next() { return generator_(); }
+
+  // Returns a uniformly distributed value in the range [0..n-1]
+  // REQUIRES: n > 0
+  uint64_t Uniform(uint64_t n) {
+    return std::uniform_int_distribution<uint64_t>(0, n - 1)(generator_);
+  }
+
+  // Randomly returns true ~"1/n" of the time, and false otherwise.
+  // REQUIRES: n > 0
+  bool OneIn(uint64_t n) { return Uniform(n) == 0; }
+
+  // Skewed: pick "base" uniformly from range [0,max_log] and then
+  // return "base" random bits.  The effect is to pick a number in the
+  // range [0,2^max_log-1] with exponential bias towards smaller numbers.
+  uint64_t Skewed(int max_log) {
+    return Uniform(uint64_t(1) << Uniform(max_log + 1));
+  }
+};
+
+// A seeded replacement for removed std::random_shuffle
+template <class RandomIt>
+void RandomShuffle(RandomIt first, RandomIt last, uint32_t seed) {
+  std::mt19937 rng(seed);
+  std::shuffle(first, last, rng);
+}
+
+// A replacement for removed std::random_shuffle
+template <class RandomIt>
+void RandomShuffle(RandomIt first, RandomIt last) {
+  RandomShuffle(first, last, std::random_device{}());
+}
+
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/util/random_test.cc b/src/rocksdb/util/random_test.cc
new file mode 100644
index 000000000..1aa62c5da
--- /dev/null
+++ b/src/rocksdb/util/random_test.cc
@@ -0,0 +1,107 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+//
+// Copyright (c) 2012 The LevelDB Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file. See the AUTHORS file for names of contributors.
+
+#include "util/random.h"
+
+#include <cstring>
+#include <vector>
+
+#include "test_util/testharness.h"
+
+using ROCKSDB_NAMESPACE::Random;
+
+TEST(RandomTest, Uniform) {
+  const int average = 20;
+  for (uint32_t seed : {0, 1, 2, 37, 4096}) {
+    Random r(seed);
+    for (int range : {1, 2, 8, 12, 100}) {
+      std::vector<int> counts(range, 0);
+
+      for (int i = 0; i < range * average; ++i) {
+        ++counts.at(r.Uniform(range));
+      }
+      int max_variance = static_cast<int>(std::sqrt(range) * 2 + 4);
+      for (int i = 0; i < range; ++i) {
+        EXPECT_GE(counts[i], std::max(1, average - max_variance));
+        EXPECT_LE(counts[i], average + max_variance + 1);
+      }
+    }
+  }
+}
+
+TEST(RandomTest, OneIn) {
+  Random r(42);
+  for (int range : {1, 2, 8, 12, 100, 1234}) {
+    const int average = 100;
+    int count = 0;
+    for (int i = 0; i < average * range; ++i) {
+      if (r.OneIn(range)) {
+        ++count;
+      }
+    }
+    if (range == 1) {
+      EXPECT_EQ(count, average);
+    } else {
+      int max_variance = static_cast<int>(std::sqrt(average) * 1.5);
+      EXPECT_GE(count, average - max_variance);
+      EXPECT_LE(count, average + max_variance);
+    }
+  }
+}
+
+TEST(RandomTest, OneInOpt) {
+  Random r(42);
+  for (int range : {-12, 0, 1, 2, 8, 12, 100, 1234}) {
+    const int average = 100;
+    int count = 0;
+    for (int i = 0; i < average * range; ++i) {
+      if (r.OneInOpt(range)) {
+        ++count;
+      }
+    }
+    if (range < 1) {
+      EXPECT_EQ(count, 0);
+    } else if (range == 1) {
+      EXPECT_EQ(count, average);
+    } else {
+      int max_variance = static_cast<int>(std::sqrt(average) * 1.5);
+      EXPECT_GE(count, average - max_variance);
+      EXPECT_LE(count, average + max_variance);
+    }
+  }
+}
+
+TEST(RandomTest, PercentTrue) {
+  Random r(42);
+  for (int pct : {-12, 0, 1, 2, 10, 50, 90, 98, 99, 100, 1234}) {
+    const int samples = 10000;
+
+    int count = 0;
+    for (int i = 0; i < samples; ++i) {
+      if (r.PercentTrue(pct)) {
+        ++count;
+      }
+    }
+    if (pct <= 0) {
+      EXPECT_EQ(count, 0);
+    } else if (pct >= 100) {
+      EXPECT_EQ(count, samples);
+    } else {
+      int est = (count * 100 + (samples / 2)) / samples;
+      EXPECT_EQ(est, pct);
+    }
+  }
+}
+
+int main(int argc, char** argv) {
+  ROCKSDB_NAMESPACE::port::InstallStackTraceHandler();
+  ::testing::InitGoogleTest(&argc, argv);
+
+  return RUN_ALL_TESTS();
+}
diff --git a/src/rocksdb/util/rate_limiter.cc b/src/rocksdb/util/rate_limiter.cc
new file mode 100644
index 000000000..6bbcabfae
--- /dev/null
+++ b/src/rocksdb/util/rate_limiter.cc
@@ -0,0 +1,378 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+//
+// Copyright (c) 2011 The LevelDB Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file. See the AUTHORS file for names of contributors.
+
+#include "util/rate_limiter.h"
+
+#include <algorithm>
+
+#include "monitoring/statistics.h"
+#include "port/port.h"
+#include "rocksdb/system_clock.h"
+#include "test_util/sync_point.h"
+#include "util/aligned_buffer.h"
+
+namespace ROCKSDB_NAMESPACE {
+size_t RateLimiter::RequestToken(size_t bytes, size_t alignment,
+                                 Env::IOPriority io_priority, Statistics* stats,
+                                 RateLimiter::OpType op_type) {
+  if (io_priority < Env::IO_TOTAL && IsRateLimited(op_type)) {
+    bytes = std::min(bytes, static_cast<size_t>(GetSingleBurstBytes()));
+
+    if (alignment > 0) {
+      // Here we may actually require more than burst and block
+      // as we can not write/read less than one page at a time on direct I/O
+      // thus we do not want to be strictly constrained by burst
+      bytes = std::max(alignment, TruncateToPageBoundary(alignment, bytes));
+    }
+    Request(bytes, io_priority, stats, op_type);
+  }
+  return bytes;
+}
+
+// Pending request
+struct GenericRateLimiter::Req {
+  explicit Req(int64_t _bytes, port::Mutex* _mu)
+      : request_bytes(_bytes), bytes(_bytes), cv(_mu), granted(false) {}
+  int64_t request_bytes;
+  int64_t bytes;
+  port::CondVar cv;
+  bool granted;
+};
+
+GenericRateLimiter::GenericRateLimiter(
+    int64_t rate_bytes_per_sec, int64_t refill_period_us, int32_t fairness,
+    RateLimiter::Mode mode, const std::shared_ptr<SystemClock>& clock,
+    bool auto_tuned)
+    : RateLimiter(mode),
+      refill_period_us_(refill_period_us),
+      rate_bytes_per_sec_(auto_tuned ? rate_bytes_per_sec / 2
+                                     : rate_bytes_per_sec),
+      refill_bytes_per_period_(
+          CalculateRefillBytesPerPeriodLocked(rate_bytes_per_sec_)),
+      clock_(clock),
+      stop_(false),
+      exit_cv_(&request_mutex_),
+      requests_to_wait_(0),
+      available_bytes_(0),
+      next_refill_us_(NowMicrosMonotonicLocked()),
+      fairness_(fairness > 100 ? 100 : fairness),
+      rnd_((uint32_t)time(nullptr)),
+      wait_until_refill_pending_(false),
+      auto_tuned_(auto_tuned),
+      num_drains_(0),
+      max_bytes_per_sec_(rate_bytes_per_sec),
+      tuned_time_(NowMicrosMonotonicLocked()) {
+  for (int i = Env::IO_LOW; i < Env::IO_TOTAL; ++i) {
+    total_requests_[i] = 0;
+    total_bytes_through_[i] = 0;
+  }
+}
+
+GenericRateLimiter::~GenericRateLimiter() {
+  MutexLock g(&request_mutex_);
+  stop_ = true;
+  std::deque<Req*>::size_type queues_size_sum = 0;
+  for (int i = Env::IO_LOW; i < Env::IO_TOTAL; ++i) {
+    queues_size_sum += queue_[i].size();
+  }
+  requests_to_wait_ = static_cast<int32_t>(queues_size_sum);
+
+  for (int i = Env::IO_TOTAL - 1; i >= Env::IO_LOW; --i) {
+    std::deque<Req*> queue = queue_[i];
+    for (auto& r : queue) {
+      r->cv.Signal();
+    }
+  }
+
+  while (requests_to_wait_ > 0) {
+    exit_cv_.Wait();
+  }
+}
+
+// This API allows user to dynamically change rate limiter's bytes per second.
+void GenericRateLimiter::SetBytesPerSecond(int64_t bytes_per_second) {
+  MutexLock g(&request_mutex_);
+  SetBytesPerSecondLocked(bytes_per_second);
+}
+
+void GenericRateLimiter::SetBytesPerSecondLocked(int64_t bytes_per_second) {
+  assert(bytes_per_second > 0);
+  rate_bytes_per_sec_.store(bytes_per_second, std::memory_order_relaxed);
+  refill_bytes_per_period_.store(
+      CalculateRefillBytesPerPeriodLocked(bytes_per_second),
+      std::memory_order_relaxed);
+}
+
+void GenericRateLimiter::Request(int64_t bytes, const Env::IOPriority pri,
+                                 Statistics* stats) {
+  assert(bytes <= refill_bytes_per_period_.load(std::memory_order_relaxed));
+  bytes = std::max(static_cast<int64_t>(0), bytes);
+  TEST_SYNC_POINT("GenericRateLimiter::Request");
+  TEST_SYNC_POINT_CALLBACK("GenericRateLimiter::Request:1",
+                           &rate_bytes_per_sec_);
+  MutexLock g(&request_mutex_);
+
+  if (auto_tuned_) {
+    static const int kRefillsPerTune = 100;
+    std::chrono::microseconds now(NowMicrosMonotonicLocked());
+    if (now - tuned_time_ >=
+        kRefillsPerTune * std::chrono::microseconds(refill_period_us_)) {
+      Status s = TuneLocked();
+      s.PermitUncheckedError();  //**TODO: What to do on error?
+    }
+  }
+
+  if (stop_) {
+    // It is now in the clean-up of ~GenericRateLimiter().
+    // Therefore any new incoming request will exit from here
+    // and not get satiesfied.
+    return;
+  }
+
+  ++total_requests_[pri];
+
+  if (available_bytes_ >= bytes) {
+    // Refill thread assigns quota and notifies requests waiting on
+    // the queue under mutex. So if we get here, that means nobody
+    // is waiting?
+    available_bytes_ -= bytes;
+    total_bytes_through_[pri] += bytes;
+    return;
+  }
+
+  // Request cannot be satisfied at this moment, enqueue
+  Req r(bytes, &request_mutex_);
+  queue_[pri].push_back(&r);
+  TEST_SYNC_POINT_CALLBACK("GenericRateLimiter::Request:PostEnqueueRequest",
+                           &request_mutex_);
+  // A thread representing a queued request coordinates with other such threads.
+  // There are two main duties.
+  //
+  // (1) Waiting for the next refill time.
+  // (2) Refilling the bytes and granting requests.
+  do {
+    int64_t time_until_refill_us = next_refill_us_ - NowMicrosMonotonicLocked();
+    if (time_until_refill_us > 0) {
+      if (wait_until_refill_pending_) {
+        // Somebody is performing (1). Trust we'll be woken up when our request
+        // is granted or we are needed for future duties.
+        r.cv.Wait();
+      } else {
+        // Whichever thread reaches here first performs duty (1) as described
+        // above.
+        int64_t wait_until = clock_->NowMicros() + time_until_refill_us;
+        RecordTick(stats, NUMBER_RATE_LIMITER_DRAINS);
+        ++num_drains_;
+        wait_until_refill_pending_ = true;
+        r.cv.TimedWait(wait_until);
+        TEST_SYNC_POINT_CALLBACK("GenericRateLimiter::Request:PostTimedWait",
+                                 &time_until_refill_us);
+        wait_until_refill_pending_ = false;
+      }
+    } else {
+      // Whichever thread reaches here first performs duty (2) as described
+      // above.
+      RefillBytesAndGrantRequestsLocked();
+      if (r.granted) {
+        // If there is any remaining requests, make sure there exists at least
+        // one candidate is awake for future duties by signaling a front request
+        // of a queue.
+        for (int i = Env::IO_TOTAL - 1; i >= Env::IO_LOW; --i) {
+          std::deque<Req*> queue = queue_[i];
+          if (!queue.empty()) {
+            queue.front()->cv.Signal();
+            break;
+          }
+        }
+      }
+    }
+    // Invariant: non-granted request is always in one queue, and granted
+    // request is always in zero queues.
+#ifndef NDEBUG
+    int num_found = 0;
+    for (int i = Env::IO_LOW; i < Env::IO_TOTAL; ++i) {
+      if (std::find(queue_[i].begin(), queue_[i].end(), &r) !=
+          queue_[i].end()) {
+        ++num_found;
+      }
+    }
+    if (r.granted) {
+      assert(num_found == 0);
+    } else {
+      assert(num_found == 1);
+    }
+#endif  // NDEBUG
+  } while (!stop_ && !r.granted);
+
+  if (stop_) {
+    // It is now in the clean-up of ~GenericRateLimiter().
+    // Therefore any woken-up request will have come out of the loop and then
+    // exit here. It might or might not have been satisfied.
+    --requests_to_wait_;
+    exit_cv_.Signal();
+  }
+}
+
+std::vector<Env::IOPriority>
+GenericRateLimiter::GeneratePriorityIterationOrderLocked() {
+  std::vector<Env::IOPriority> pri_iteration_order(Env::IO_TOTAL /* 4 */);
+  // We make Env::IO_USER a superior priority by always iterating its queue
+  // first
+  pri_iteration_order[0] = Env::IO_USER;
+
+  bool high_pri_iterated_after_mid_low_pri = rnd_.OneIn(fairness_);
+  TEST_SYNC_POINT_CALLBACK(
+      "GenericRateLimiter::GeneratePriorityIterationOrderLocked::"
+      "PostRandomOneInFairnessForHighPri",
+      &high_pri_iterated_after_mid_low_pri);
+  bool mid_pri_itereated_after_low_pri = rnd_.OneIn(fairness_);
+  TEST_SYNC_POINT_CALLBACK(
+      "GenericRateLimiter::GeneratePriorityIterationOrderLocked::"
+      "PostRandomOneInFairnessForMidPri",
+      &mid_pri_itereated_after_low_pri);
+
+  if (high_pri_iterated_after_mid_low_pri) {
+    pri_iteration_order[3] = Env::IO_HIGH;
+    pri_iteration_order[2] =
+        mid_pri_itereated_after_low_pri ? Env::IO_MID : Env::IO_LOW;
+    pri_iteration_order[1] =
+        (pri_iteration_order[2] == Env::IO_MID) ? Env::IO_LOW : Env::IO_MID;
+  } else {
+    pri_iteration_order[1] = Env::IO_HIGH;
+    pri_iteration_order[3] =
+        mid_pri_itereated_after_low_pri ? Env::IO_MID : Env::IO_LOW;
+    pri_iteration_order[2] =
+        (pri_iteration_order[3] == Env::IO_MID) ? Env::IO_LOW : Env::IO_MID;
+  }
+
+  TEST_SYNC_POINT_CALLBACK(
+      "GenericRateLimiter::GeneratePriorityIterationOrderLocked::"
+      "PreReturnPriIterationOrder",
+      &pri_iteration_order);
+  return pri_iteration_order;
+}
+
+void GenericRateLimiter::RefillBytesAndGrantRequestsLocked() {
+  TEST_SYNC_POINT_CALLBACK(
+      "GenericRateLimiter::RefillBytesAndGrantRequestsLocked", &request_mutex_);
+  next_refill_us_ = NowMicrosMonotonicLocked() + refill_period_us_;
+  // Carry over the left over quota from the last period
+  auto refill_bytes_per_period =
+      refill_bytes_per_period_.load(std::memory_order_relaxed);
+  if (available_bytes_ < refill_bytes_per_period) {
+    available_bytes_ += refill_bytes_per_period;
+  }
+
+  std::vector<Env::IOPriority> pri_iteration_order =
+      GeneratePriorityIterationOrderLocked();
+
+  for (int i = Env::IO_LOW; i < Env::IO_TOTAL; ++i) {
+    assert(!pri_iteration_order.empty());
+    Env::IOPriority current_pri = pri_iteration_order[i];
+    auto* queue = &queue_[current_pri];
+    while (!queue->empty()) {
+      auto* next_req = queue->front();
+      if (available_bytes_ < next_req->request_bytes) {
+        // Grant partial request_bytes to avoid starvation of requests
+        // that become asking for more bytes than available_bytes_
+        // due to dynamically reduced rate limiter's bytes_per_second that
+        // leads to reduced refill_bytes_per_period hence available_bytes_
+        next_req->request_bytes -= available_bytes_;
+        available_bytes_ = 0;
+        break;
+      }
+      available_bytes_ -= next_req->request_bytes;
+      next_req->request_bytes = 0;
+      total_bytes_through_[current_pri] += next_req->bytes;
+      queue->pop_front();
+
+      next_req->granted = true;
+      // Quota granted, signal the thread to exit
+      next_req->cv.Signal();
+    }
+  }
+}
+
+int64_t GenericRateLimiter::CalculateRefillBytesPerPeriodLocked(
+    int64_t rate_bytes_per_sec) {
+  if (std::numeric_limits<int64_t>::max() / rate_bytes_per_sec <
+      refill_period_us_) {
+    // Avoid unexpected result in the overflow case. The result now is still
+    // inaccurate but is a number that is large enough.
+    return std::numeric_limits<int64_t>::max() / 1000000;
+  } else {
+    return rate_bytes_per_sec * refill_period_us_ / 1000000;
+  }
+}
+
+Status GenericRateLimiter::TuneLocked() {
+  const int kLowWatermarkPct = 50;
+  const int kHighWatermarkPct = 90;
+  const int kAdjustFactorPct = 5;
+  // computed rate limit will be in
+  // `[max_bytes_per_sec_ / kAllowedRangeFactor, max_bytes_per_sec_]`.
+  const int kAllowedRangeFactor = 20;
+
+  std::chrono::microseconds prev_tuned_time = tuned_time_;
+  tuned_time_ = std::chrono::microseconds(NowMicrosMonotonicLocked());
+
+  int64_t elapsed_intervals = (tuned_time_ - prev_tuned_time +
+                               std::chrono::microseconds(refill_period_us_) -
+                               std::chrono::microseconds(1)) /
+                              std::chrono::microseconds(refill_period_us_);
+  // We tune every kRefillsPerTune intervals, so the overflow and division-by-
+  // zero conditions should never happen.
+  assert(num_drains_ <= std::numeric_limits<int64_t>::max() / 100);
+  assert(elapsed_intervals > 0);
+  int64_t drained_pct = num_drains_ * 100 / elapsed_intervals;
+
+  int64_t prev_bytes_per_sec = GetBytesPerSecond();
+  int64_t new_bytes_per_sec;
+  if (drained_pct == 0) {
+    new_bytes_per_sec = max_bytes_per_sec_ / kAllowedRangeFactor;
+  } else if (drained_pct < kLowWatermarkPct) {
+    // sanitize to prevent overflow
+    int64_t sanitized_prev_bytes_per_sec =
+        std::min(prev_bytes_per_sec, std::numeric_limits<int64_t>::max() / 100);
+    new_bytes_per_sec =
+        std::max(max_bytes_per_sec_ / kAllowedRangeFactor,
+                 sanitized_prev_bytes_per_sec * 100 / (100 + kAdjustFactorPct));
+  } else if (drained_pct > kHighWatermarkPct) {
+    // sanitize to prevent overflow
+    int64_t sanitized_prev_bytes_per_sec =
+        std::min(prev_bytes_per_sec, std::numeric_limits<int64_t>::max() /
+                                         (100 + kAdjustFactorPct));
+    new_bytes_per_sec =
+        std::min(max_bytes_per_sec_,
+                 sanitized_prev_bytes_per_sec * (100 + kAdjustFactorPct) / 100);
+  } else {
+    new_bytes_per_sec = prev_bytes_per_sec;
+  }
+  if (new_bytes_per_sec != prev_bytes_per_sec) {
+    SetBytesPerSecondLocked(new_bytes_per_sec);
+  }
+  num_drains_ = 0;
+  return Status::OK();
+}
+
+RateLimiter* NewGenericRateLimiter(
+    int64_t rate_bytes_per_sec, int64_t refill_period_us /* = 100 * 1000 */,
+    int32_t fairness /* = 10 */,
+    RateLimiter::Mode mode /* = RateLimiter::Mode::kWritesOnly */,
+    bool auto_tuned /* = false */) {
+  assert(rate_bytes_per_sec > 0);
+  assert(refill_period_us > 0);
+  assert(fairness > 0);
+  std::unique_ptr<RateLimiter> limiter(
+      new GenericRateLimiter(rate_bytes_per_sec, refill_period_us, fairness,
+                             mode, SystemClock::Default(), auto_tuned));
+  return limiter.release();
+}
+
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/util/rate_limiter.h b/src/rocksdb/util/rate_limiter.h
new file mode 100644
index 000000000..4c078f5a0
--- /dev/null
+++ b/src/rocksdb/util/rate_limiter.h
@@ -0,0 +1,146 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+//
+// Copyright (c) 2011 The LevelDB Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file. See the AUTHORS file for names of contributors.
+
+#pragma once
+
+#include <algorithm>
+#include <atomic>
+#include <chrono>
+#include <deque>
+
+#include "port/port.h"
+#include "rocksdb/env.h"
+#include "rocksdb/rate_limiter.h"
+#include "rocksdb/status.h"
+#include "rocksdb/system_clock.h"
+#include "util/mutexlock.h"
+#include "util/random.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+class GenericRateLimiter : public RateLimiter {
+ public:
+  GenericRateLimiter(int64_t refill_bytes, int64_t refill_period_us,
+                     int32_t fairness, RateLimiter::Mode mode,
+                     const std::shared_ptr<SystemClock>& clock,
+                     bool auto_tuned);
+
+  virtual ~GenericRateLimiter();
+
+  // This API allows user to dynamically change rate limiter's bytes per second.
+  virtual void SetBytesPerSecond(int64_t bytes_per_second) override;
+
+  // Request for token to write bytes. If this request can not be satisfied,
+  // the call is blocked. Caller is responsible to make sure
+  // bytes <= GetSingleBurstBytes() and bytes >= 0. Negative bytes
+  // passed in will be rounded up to 0.
+  using RateLimiter::Request;
+  virtual void Request(const int64_t bytes, const Env::IOPriority pri,
+                       Statistics* stats) override;
+
+  virtual int64_t GetSingleBurstBytes() const override {
+    return refill_bytes_per_period_.load(std::memory_order_relaxed);
+  }
+
+  virtual int64_t GetTotalBytesThrough(
+      const Env::IOPriority pri = Env::IO_TOTAL) const override {
+    MutexLock g(&request_mutex_);
+    if (pri == Env::IO_TOTAL) {
+      int64_t total_bytes_through_sum = 0;
+      for (int i = Env::IO_LOW; i < Env::IO_TOTAL; ++i) {
+        total_bytes_through_sum += total_bytes_through_[i];
+      }
+      return total_bytes_through_sum;
+    }
+    return total_bytes_through_[pri];
+  }
+
+  virtual int64_t GetTotalRequests(
+      const Env::IOPriority pri = Env::IO_TOTAL) const override {
+    MutexLock g(&request_mutex_);
+    if (pri == Env::IO_TOTAL) {
+      int64_t total_requests_sum = 0;
+      for (int i = Env::IO_LOW; i < Env::IO_TOTAL; ++i) {
+        total_requests_sum += total_requests_[i];
+      }
+      return total_requests_sum;
+    }
+    return total_requests_[pri];
+  }
+
+  virtual Status GetTotalPendingRequests(
+      int64_t* total_pending_requests,
+      const Env::IOPriority pri = Env::IO_TOTAL) const override {
+    assert(total_pending_requests != nullptr);
+    MutexLock g(&request_mutex_);
+    if (pri == Env::IO_TOTAL) {
+      int64_t total_pending_requests_sum = 0;
+      for (int i = Env::IO_LOW; i < Env::IO_TOTAL; ++i) {
+        total_pending_requests_sum += static_cast<int64_t>(queue_[i].size());
+      }
+      *total_pending_requests = total_pending_requests_sum;
+    } else {
+      *total_pending_requests = static_cast<int64_t>(queue_[pri].size());
+    }
+    return Status::OK();
+  }
+
+  virtual int64_t GetBytesPerSecond() const override {
+    return rate_bytes_per_sec_.load(std::memory_order_relaxed);
+  }
+
+  virtual void TEST_SetClock(std::shared_ptr<SystemClock> clock) {
+    MutexLock g(&request_mutex_);
+    clock_ = std::move(clock);
+    next_refill_us_ = NowMicrosMonotonicLocked();
+  }
+
+ private:
+  void RefillBytesAndGrantRequestsLocked();
+  std::vector<Env::IOPriority> GeneratePriorityIterationOrderLocked();
+  int64_t CalculateRefillBytesPerPeriodLocked(int64_t rate_bytes_per_sec);
+  Status TuneLocked();
+  void SetBytesPerSecondLocked(int64_t bytes_per_second);
+
+  uint64_t NowMicrosMonotonicLocked() {
+    return clock_->NowNanos() / std::milli::den;
+  }
+
+  // This mutex guard all internal states
+  mutable port::Mutex request_mutex_;
+
+  const int64_t refill_period_us_;
+
+  std::atomic<int64_t> rate_bytes_per_sec_;
+  std::atomic<int64_t> refill_bytes_per_period_;
+  std::shared_ptr<SystemClock> clock_;
+
+  bool stop_;
+  port::CondVar exit_cv_;
+  int32_t requests_to_wait_;
+
+  int64_t total_requests_[Env::IO_TOTAL];
+  int64_t total_bytes_through_[Env::IO_TOTAL];
+  int64_t available_bytes_;
+  int64_t next_refill_us_;
+
+  int32_t fairness_;
+  Random rnd_;
+
+  struct Req;
+  std::deque<Req*> queue_[Env::IO_TOTAL];
+  bool wait_until_refill_pending_;
+
+  bool auto_tuned_;
+  int64_t num_drains_;
+  const int64_t max_bytes_per_sec_;
+  std::chrono::microseconds tuned_time_;
+};
+
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/util/rate_limiter_test.cc b/src/rocksdb/util/rate_limiter_test.cc
new file mode 100644
index 000000000..cda134867
--- /dev/null
+++ b/src/rocksdb/util/rate_limiter_test.cc
@@ -0,0 +1,476 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+//
+// Copyright (c) 2011 The LevelDB Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file. See the AUTHORS file for names of contributors.
+
+#include "util/rate_limiter.h"
+
+#include <chrono>
+#include <cinttypes>
+#include <cstdint>
+#include <limits>
+
+#include "db/db_test_util.h"
+#include "port/port.h"
+#include "rocksdb/system_clock.h"
+#include "test_util/sync_point.h"
+#include "test_util/testharness.h"
+#include "util/random.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+// TODO(yhchiang): the rate will not be accurate when we run test in parallel.
+class RateLimiterTest : public testing::Test {
+ protected:
+  ~RateLimiterTest() override {
+    SyncPoint::GetInstance()->DisableProcessing();
+    SyncPoint::GetInstance()->ClearAllCallBacks();
+  }
+};
+
+TEST_F(RateLimiterTest, OverflowRate) {
+  GenericRateLimiter limiter(std::numeric_limits<int64_t>::max(), 1000, 10,
+                             RateLimiter::Mode::kWritesOnly,
+                             SystemClock::Default(), false /* auto_tuned */);
+  ASSERT_GT(limiter.GetSingleBurstBytes(), 1000000000ll);
+}
+
+TEST_F(RateLimiterTest, StartStop) {
+  std::unique_ptr<RateLimiter> limiter(NewGenericRateLimiter(100, 100, 10));
+}
+
+TEST_F(RateLimiterTest, GetTotalBytesThrough) {
+  std::unique_ptr<RateLimiter> limiter(NewGenericRateLimiter(
+      200 /* rate_bytes_per_sec */, 1000 * 1000 /* refill_period_us */,
+      10 /* fairness */));
+  for (int i = Env::IO_LOW; i <= Env::IO_TOTAL; ++i) {
+    ASSERT_EQ(limiter->GetTotalBytesThrough(static_cast<Env::IOPriority>(i)),
+              0);
+  }
+
+  std::int64_t request_byte = 200;
+  std::int64_t request_byte_sum = 0;
+  for (int i = Env::IO_LOW; i < Env::IO_TOTAL; ++i) {
+    limiter->Request(request_byte, static_cast<Env::IOPriority>(i),
+                     nullptr /* stats */, RateLimiter::OpType::kWrite);
+    request_byte_sum += request_byte;
+  }
+
+  for (int i = Env::IO_LOW; i < Env::IO_TOTAL; ++i) {
+    EXPECT_EQ(limiter->GetTotalBytesThrough(static_cast<Env::IOPriority>(i)),
+              request_byte)
+        << "Failed to track total_bytes_through_ correctly when IOPriority = "
+        << static_cast<Env::IOPriority>(i);
+  }
+  EXPECT_EQ(limiter->GetTotalBytesThrough(Env::IO_TOTAL), request_byte_sum)
+      << "Failed to track total_bytes_through_ correctly when IOPriority = "
+         "Env::IO_TOTAL";
+}
+
+TEST_F(RateLimiterTest, GetTotalRequests) {
+  std::unique_ptr<RateLimiter> limiter(NewGenericRateLimiter(
+      200 /* rate_bytes_per_sec */, 1000 * 1000 /* refill_period_us */,
+      10 /* fairness */));
+  for (int i = Env::IO_LOW; i <= Env::IO_TOTAL; ++i) {
+    ASSERT_EQ(limiter->GetTotalRequests(static_cast<Env::IOPriority>(i)), 0);
+  }
+
+  std::int64_t total_requests_sum = 0;
+  for (int i = Env::IO_LOW; i < Env::IO_TOTAL; ++i) {
+    limiter->Request(200, static_cast<Env::IOPriority>(i), nullptr /* stats */,
+                     RateLimiter::OpType::kWrite);
+    total_requests_sum += 1;
+  }
+
+  for (int i = Env::IO_LOW; i < Env::IO_TOTAL; ++i) {
+    EXPECT_EQ(limiter->GetTotalRequests(static_cast<Env::IOPriority>(i)), 1)
+        << "Failed to track total_requests_ correctly when IOPriority = "
+        << static_cast<Env::IOPriority>(i);
+  }
+  EXPECT_EQ(limiter->GetTotalRequests(Env::IO_TOTAL), total_requests_sum)
+      << "Failed to track total_requests_ correctly when IOPriority = "
+         "Env::IO_TOTAL";
+}
+
+TEST_F(RateLimiterTest, GetTotalPendingRequests) {
+  std::unique_ptr<RateLimiter> limiter(NewGenericRateLimiter(
+      200 /* rate_bytes_per_sec */, 1000 * 1000 /* refill_period_us */,
+      10 /* fairness */));
+  int64_t total_pending_requests = 0;
+  for (int i = Env::IO_LOW; i <= Env::IO_TOTAL; ++i) {
+    ASSERT_OK(limiter->GetTotalPendingRequests(
+        &total_pending_requests, static_cast<Env::IOPriority>(i)));
+    ASSERT_EQ(total_pending_requests, 0);
+  }
+  // This is a variable for making sure the following callback is called
+  // and the assertions in it are indeed excuted
+  bool nonzero_pending_requests_verified = false;
+  SyncPoint::GetInstance()->SetCallBack(
+      "GenericRateLimiter::Request:PostEnqueueRequest", [&](void* arg) {
+        port::Mutex* request_mutex = (port::Mutex*)arg;
+        // We temporarily unlock the mutex so that the following
+        // GetTotalPendingRequests() can acquire it
+        request_mutex->Unlock();
+        for (int i = Env::IO_LOW; i <= Env::IO_TOTAL; ++i) {
+          EXPECT_OK(limiter->GetTotalPendingRequests(
+              &total_pending_requests, static_cast<Env::IOPriority>(i)))
+              << "Failed to return total pending requests for priority level = "
+              << static_cast<Env::IOPriority>(i);
+          if (i == Env::IO_USER || i == Env::IO_TOTAL) {
+            EXPECT_EQ(total_pending_requests, 1)
+                << "Failed to correctly return total pending requests for "
+                   "priority level = "
+                << static_cast<Env::IOPriority>(i);
+          } else {
+            EXPECT_EQ(total_pending_requests, 0)
+                << "Failed to correctly return total pending requests for "
+                   "priority level = "
+                << static_cast<Env::IOPriority>(i);
+          }
+        }
+        // We lock the mutex again so that the request thread can resume running
+        // with the mutex locked
+        request_mutex->Lock();
+        nonzero_pending_requests_verified = true;
+      });
+
+  SyncPoint::GetInstance()->EnableProcessing();
+  limiter->Request(200, Env::IO_USER, nullptr /* stats */,
+                   RateLimiter::OpType::kWrite);
+  ASSERT_EQ(nonzero_pending_requests_verified, true);
+  for (int i = Env::IO_LOW; i <= Env::IO_TOTAL; ++i) {
+    EXPECT_OK(limiter->GetTotalPendingRequests(&total_pending_requests,
+                                               static_cast<Env::IOPriority>(i)))
+        << "Failed to return total pending requests for priority level = "
+        << static_cast<Env::IOPriority>(i);
+    EXPECT_EQ(total_pending_requests, 0)
+        << "Failed to correctly return total pending requests for priority "
+           "level = "
+        << static_cast<Env::IOPriority>(i);
+  }
+  SyncPoint::GetInstance()->DisableProcessing();
+  SyncPoint::GetInstance()->ClearCallBack(
+      "GenericRateLimiter::Request:PostEnqueueRequest");
+}
+
+TEST_F(RateLimiterTest, Modes) {
+  for (auto mode : {RateLimiter::Mode::kWritesOnly,
+                    RateLimiter::Mode::kReadsOnly, RateLimiter::Mode::kAllIo}) {
+    GenericRateLimiter limiter(2000 /* rate_bytes_per_sec */,
+                               1000 * 1000 /* refill_period_us */,
+                               10 /* fairness */, mode, SystemClock::Default(),
+                               false /* auto_tuned */);
+    limiter.Request(1000 /* bytes */, Env::IO_HIGH, nullptr /* stats */,
+                    RateLimiter::OpType::kRead);
+    if (mode == RateLimiter::Mode::kWritesOnly) {
+      ASSERT_EQ(0, limiter.GetTotalBytesThrough(Env::IO_HIGH));
+    } else {
+      ASSERT_EQ(1000, limiter.GetTotalBytesThrough(Env::IO_HIGH));
+    }
+
+    limiter.Request(1000 /* bytes */, Env::IO_HIGH, nullptr /* stats */,
+                    RateLimiter::OpType::kWrite);
+    if (mode == RateLimiter::Mode::kAllIo) {
+      ASSERT_EQ(2000, limiter.GetTotalBytesThrough(Env::IO_HIGH));
+    } else {
+      ASSERT_EQ(1000, limiter.GetTotalBytesThrough(Env::IO_HIGH));
+    }
+  }
+}
+
+TEST_F(RateLimiterTest, GeneratePriorityIterationOrder) {
+  std::unique_ptr<RateLimiter> limiter(NewGenericRateLimiter(
+      200 /* rate_bytes_per_sec */, 1000 * 1000 /* refill_period_us */,
+      10 /* fairness */));
+
+  bool possible_random_one_in_fairness_results_for_high_mid_pri[4][2] = {
+      {false, false}, {false, true}, {true, false}, {true, true}};
+  std::vector<Env::IOPriority> possible_priority_iteration_orders[4] = {
+      {Env::IO_USER, Env::IO_HIGH, Env::IO_MID, Env::IO_LOW},
+      {Env::IO_USER, Env::IO_HIGH, Env::IO_LOW, Env::IO_MID},
+      {Env::IO_USER, Env::IO_MID, Env::IO_LOW, Env::IO_HIGH},
+      {Env::IO_USER, Env::IO_LOW, Env::IO_MID, Env::IO_HIGH}};
+
+  for (int i = 0; i < 4; ++i) {
+    // These are variables for making sure the following callbacks are called
+    // and the assertion in the last callback is indeed excuted
+    bool high_pri_iterated_after_mid_low_pri_set = false;
+    bool mid_pri_itereated_after_low_pri_set = false;
+    bool pri_iteration_order_verified = false;
+    SyncPoint::GetInstance()->SetCallBack(
+        "GenericRateLimiter::GeneratePriorityIterationOrderLocked::"
+        "PostRandomOneInFairnessForHighPri",
+        [&](void* arg) {
+          bool* high_pri_iterated_after_mid_low_pri = (bool*)arg;
+          *high_pri_iterated_after_mid_low_pri =
+              possible_random_one_in_fairness_results_for_high_mid_pri[i][0];
+          high_pri_iterated_after_mid_low_pri_set = true;
+        });
+
+    SyncPoint::GetInstance()->SetCallBack(
+        "GenericRateLimiter::GeneratePriorityIterationOrderLocked::"
+        "PostRandomOneInFairnessForMidPri",
+        [&](void* arg) {
+          bool* mid_pri_itereated_after_low_pri = (bool*)arg;
+          *mid_pri_itereated_after_low_pri =
+              possible_random_one_in_fairness_results_for_high_mid_pri[i][1];
+          mid_pri_itereated_after_low_pri_set = true;
+        });
+
+    SyncPoint::GetInstance()->SetCallBack(
+        "GenericRateLimiter::GeneratePriorityIterationOrderLocked::"
+        "PreReturnPriIterationOrder",
+        [&](void* arg) {
+          std::vector<Env::IOPriority>* pri_iteration_order =
+              (std::vector<Env::IOPriority>*)arg;
+          EXPECT_EQ(*pri_iteration_order, possible_priority_iteration_orders[i])
+              << "Failed to generate priority iteration order correctly when "
+                 "high_pri_iterated_after_mid_low_pri = "
+              << possible_random_one_in_fairness_results_for_high_mid_pri[i][0]
+              << ", mid_pri_itereated_after_low_pri = "
+              << possible_random_one_in_fairness_results_for_high_mid_pri[i][1]
+              << std::endl;
+          pri_iteration_order_verified = true;
+        });
+
+    SyncPoint::GetInstance()->EnableProcessing();
+    limiter->Request(200 /* request max bytes to drain so that refill and order
+                           generation will be triggered every time
+                           GenericRateLimiter::Request() is called */
+                     ,
+                     Env::IO_USER, nullptr /* stats */,
+                     RateLimiter::OpType::kWrite);
+    ASSERT_EQ(high_pri_iterated_after_mid_low_pri_set, true);
+    ASSERT_EQ(mid_pri_itereated_after_low_pri_set, true);
+    ASSERT_EQ(pri_iteration_order_verified, true);
+    SyncPoint::GetInstance()->DisableProcessing();
+    SyncPoint::GetInstance()->ClearCallBack(
+        "GenericRateLimiter::GeneratePriorityIterationOrderLocked::"
+        "PreReturnPriIterationOrder");
+    SyncPoint::GetInstance()->ClearCallBack(
+        "GenericRateLimiter::GeneratePriorityIterationOrderLocked::"
+        "PostRandomOneInFairnessForMidPri");
+    SyncPoint::GetInstance()->ClearCallBack(
+        "GenericRateLimiter::GeneratePriorityIterationOrderLocked::"
+        "PostRandomOneInFairnessForHighPri");
+  }
+}
+
+TEST_F(RateLimiterTest, Rate) {
+  auto* env = Env::Default();
+  struct Arg {
+    Arg(int32_t _target_rate, int _burst)
+        : limiter(NewGenericRateLimiter(_target_rate /* rate_bytes_per_sec */,
+                                        100 * 1000 /* refill_period_us */,
+                                        10 /* fairness */)),
+          request_size(_target_rate /
+                       10 /* refill period here is 1/10 second */),
+          burst(_burst) {}
+    std::unique_ptr<RateLimiter> limiter;
+    int32_t request_size;
+    int burst;
+  };
+
+  auto writer = [](void* p) {
+    const auto& thread_clock = SystemClock::Default();
+    auto* arg = static_cast<Arg*>(p);
+    // Test for 2 seconds
+    auto until = thread_clock->NowMicros() + 2 * 1000000;
+    Random r((uint32_t)(thread_clock->NowNanos() %
+                        std::numeric_limits<uint32_t>::max()));
+    while (thread_clock->NowMicros() < until) {
+      for (int i = 0; i < static_cast<int>(r.Skewed(arg->burst * 2) + 1); ++i) {
+        arg->limiter->Request(r.Uniform(arg->request_size - 1) + 1,
+                              Env::IO_USER, nullptr /* stats */,
+                              RateLimiter::OpType::kWrite);
+      }
+
+      for (int i = 0; i < static_cast<int>(r.Skewed(arg->burst) + 1); ++i) {
+        arg->limiter->Request(r.Uniform(arg->request_size - 1) + 1,
+                              Env::IO_HIGH, nullptr /* stats */,
+                              RateLimiter::OpType::kWrite);
+      }
+
+      for (int i = 0; i < static_cast<int>(r.Skewed(arg->burst / 2 + 1) + 1);
+           ++i) {
+        arg->limiter->Request(r.Uniform(arg->request_size - 1) + 1, Env::IO_MID,
+                              nullptr /* stats */, RateLimiter::OpType::kWrite);
+      }
+
+      arg->limiter->Request(r.Uniform(arg->request_size - 1) + 1, Env::IO_LOW,
+                            nullptr /* stats */, RateLimiter::OpType::kWrite);
+    }
+  };
+
+  int samples = 0;
+  int samples_at_minimum = 0;
+
+  for (int i = 1; i <= 16; i *= 2) {
+    int32_t target = i * 1024 * 10;
+    Arg arg(target, i / 4 + 1);
+    int64_t old_total_bytes_through = 0;
+    for (int iter = 1; iter <= 2; ++iter) {
+      // second iteration changes the target dynamically
+      if (iter == 2) {
+        target *= 2;
+        arg.limiter->SetBytesPerSecond(target);
+      }
+      auto start = env->NowMicros();
+      for (int t = 0; t < i; ++t) {
+        env->StartThread(writer, &arg);
+      }
+      env->WaitForJoin();
+
+      auto elapsed = env->NowMicros() - start;
+      double rate =
+          (arg.limiter->GetTotalBytesThrough() - old_total_bytes_through) *
+          1000000.0 / elapsed;
+      old_total_bytes_through = arg.limiter->GetTotalBytesThrough();
+      fprintf(stderr,
+              "request size [1 - %" PRIi32 "], limit %" PRIi32
+              " KB/sec, actual rate: %lf KB/sec, elapsed %.2lf seconds\n",
+              arg.request_size - 1, target / 1024, rate / 1024,
+              elapsed / 1000000.0);
+
+      ++samples;
+      if (rate / target >= 0.80) {
+        ++samples_at_minimum;
+      }
+      ASSERT_LE(rate / target, 1.25);
+    }
+  }
+
+  // This can fail due to slow execution speed, like when using valgrind or in
+  // heavily loaded CI environments
+  bool skip_minimum_rate_check =
+#if (defined(CIRCLECI) && defined(OS_MACOSX)) || defined(ROCKSDB_VALGRIND_RUN)
+      true;
+#else
+      getenv("SANDCASTLE");
+#endif
+  if (skip_minimum_rate_check) {
+    fprintf(stderr, "Skipped minimum rate check (%d / %d passed)\n",
+            samples_at_minimum, samples);
+  } else {
+    ASSERT_EQ(samples_at_minimum, samples);
+  }
+}
+
+TEST_F(RateLimiterTest, LimitChangeTest) {
+  // starvation test when limit changes to a smaller value
+  int64_t refill_period = 1000 * 1000;
+  auto* env = Env::Default();
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->EnableProcessing();
+  struct Arg {
+    Arg(int32_t _request_size, Env::IOPriority _pri,
+        std::shared_ptr<RateLimiter> _limiter)
+        : request_size(_request_size), pri(_pri), limiter(_limiter) {}
+    int32_t request_size;
+    Env::IOPriority pri;
+    std::shared_ptr<RateLimiter> limiter;
+  };
+
+  auto writer = [](void* p) {
+    auto* arg = static_cast<Arg*>(p);
+    arg->limiter->Request(arg->request_size, arg->pri, nullptr /* stats */,
+                          RateLimiter::OpType::kWrite);
+  };
+
+  for (uint32_t i = 1; i <= 16; i <<= 1) {
+    int32_t target = i * 1024 * 10;
+    // refill per second
+    for (int iter = 0; iter < 2; iter++) {
+      std::shared_ptr<RateLimiter> limiter =
+          std::make_shared<GenericRateLimiter>(
+              target, refill_period, 10, RateLimiter::Mode::kWritesOnly,
+              SystemClock::Default(), false /* auto_tuned */);
+      // After "GenericRateLimiter::Request:1" the mutex is held until the bytes
+      // are refilled. This test could be improved to change the limit when lock
+      // is released in `TimedWait()`.
+      ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->LoadDependency(
+          {{"GenericRateLimiter::Request",
+            "RateLimiterTest::LimitChangeTest:changeLimitStart"},
+           {"RateLimiterTest::LimitChangeTest:changeLimitEnd",
+            "GenericRateLimiter::Request:1"}});
+      Arg arg(target, Env::IO_HIGH, limiter);
+      // The idea behind is to start a request first, then before it refills,
+      // update limit to a different value (2X/0.5X). No starvation should
+      // be guaranteed under any situation
+      // TODO(lightmark): more test cases are welcome.
+      env->StartThread(writer, &arg);
+      int32_t new_limit = (target << 1) >> (iter << 1);
+      TEST_SYNC_POINT("RateLimiterTest::LimitChangeTest:changeLimitStart");
+      arg.limiter->SetBytesPerSecond(new_limit);
+      TEST_SYNC_POINT("RateLimiterTest::LimitChangeTest:changeLimitEnd");
+      env->WaitForJoin();
+      fprintf(stderr,
+              "[COMPLETE] request size %" PRIi32 " KB, new limit %" PRIi32
+              "KB/sec, refill period %" PRIi64 " ms\n",
+              target / 1024, new_limit / 1024, refill_period / 1000);
+    }
+  }
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->DisableProcessing();
+}
+
+TEST_F(RateLimiterTest, AutoTuneIncreaseWhenFull) {
+  const std::chrono::seconds kTimePerRefill(1);
+  const int kRefillsPerTune = 100;  // needs to match util/rate_limiter.cc
+
+  SpecialEnv special_env(Env::Default(), /*time_elapse_only_sleep*/ true);
+
+  auto stats = CreateDBStatistics();
+  std::unique_ptr<RateLimiter> rate_limiter(new GenericRateLimiter(
+      1000 /* rate_bytes_per_sec */,
+      std::chrono::microseconds(kTimePerRefill).count(), 10 /* fairness */,
+      RateLimiter::Mode::kWritesOnly, special_env.GetSystemClock(),
+      true /* auto_tuned */));
+
+  // Rate limiter uses `CondVar::TimedWait()`, which does not have access to the
+  // `Env` to advance its time according to the fake wait duration. The
+  // workaround is to install a callback that advance the `Env`'s mock time.
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
+      "GenericRateLimiter::Request:PostTimedWait", [&](void* arg) {
+        int64_t time_waited_us = *static_cast<int64_t*>(arg);
+        special_env.SleepForMicroseconds(static_cast<int>(time_waited_us));
+      });
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->EnableProcessing();
+
+  // verify rate limit increases after a sequence of periods where rate limiter
+  // is always drained
+  int64_t orig_bytes_per_sec = rate_limiter->GetSingleBurstBytes();
+  rate_limiter->Request(orig_bytes_per_sec, Env::IO_HIGH, stats.get(),
+                        RateLimiter::OpType::kWrite);
+  while (std::chrono::microseconds(special_env.NowMicros()) <=
+         kRefillsPerTune * kTimePerRefill) {
+    rate_limiter->Request(orig_bytes_per_sec, Env::IO_HIGH, stats.get(),
+                          RateLimiter::OpType::kWrite);
+  }
+  int64_t new_bytes_per_sec = rate_limiter->GetSingleBurstBytes();
+  ASSERT_GT(new_bytes_per_sec, orig_bytes_per_sec);
+
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->DisableProcessing();
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->ClearCallBack(
+      "GenericRateLimiter::Request:PostTimedWait");
+
+  // decreases after a sequence of periods where rate limiter is not drained
+  orig_bytes_per_sec = new_bytes_per_sec;
+  special_env.SleepForMicroseconds(static_cast<int>(
+      kRefillsPerTune * std::chrono::microseconds(kTimePerRefill).count()));
+  // make a request so tuner can be triggered
+  rate_limiter->Request(1 /* bytes */, Env::IO_HIGH, stats.get(),
+                        RateLimiter::OpType::kWrite);
+  new_bytes_per_sec = rate_limiter->GetSingleBurstBytes();
+  ASSERT_LT(new_bytes_per_sec, orig_bytes_per_sec);
+}
+
+}  // namespace ROCKSDB_NAMESPACE
+
+int main(int argc, char** argv) {
+  ROCKSDB_NAMESPACE::port::InstallStackTraceHandler();
+  ::testing::InitGoogleTest(&argc, argv);
+  return RUN_ALL_TESTS();
+}
diff --git a/src/rocksdb/util/repeatable_thread.h b/src/rocksdb/util/repeatable_thread.h
new file mode 100644
index 000000000..c75ad7c49
--- /dev/null
+++ b/src/rocksdb/util/repeatable_thread.h
@@ -0,0 +1,149 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#pragma once
+
+#include <functional>
+#include <string>
+
+#include "monitoring/instrumented_mutex.h"
+#include "port/port.h"
+#include "rocksdb/system_clock.h"
+#include "util/mutexlock.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+// Simple wrapper around port::Thread that supports calling a callback every
+// X seconds. If you pass in 0, then it will call your callback repeatedly
+// without delay.
+class RepeatableThread {
+ public:
+  RepeatableThread(std::function<void()> function,
+                   const std::string& thread_name, SystemClock* clock,
+                   uint64_t delay_us, uint64_t initial_delay_us = 0)
+      : function_(function),
+        thread_name_("rocksdb:" + thread_name),
+        clock_(clock),
+        delay_us_(delay_us),
+        initial_delay_us_(initial_delay_us),
+        mutex_(clock),
+        cond_var_(&mutex_),
+        running_(true),
+#ifndef NDEBUG
+        waiting_(false),
+        run_count_(0),
+#endif
+        thread_([this] { thread(); }) {
+  }
+
+  void cancel() {
+    {
+      InstrumentedMutexLock l(&mutex_);
+      if (!running_) {
+        return;
+      }
+      running_ = false;
+      cond_var_.SignalAll();
+    }
+    thread_.join();
+  }
+
+  bool IsRunning() { return running_; }
+
+  ~RepeatableThread() { cancel(); }
+
+#ifndef NDEBUG
+  // Wait until RepeatableThread starting waiting, call the optional callback,
+  // then wait for one run of RepeatableThread. Tests can use provide a
+  // custom clock object to mock time, and use the callback here to bump current
+  // time and trigger RepeatableThread. See repeatable_thread_test for example.
+  //
+  // Note: only support one caller of this method.
+  void TEST_WaitForRun(std::function<void()> callback = nullptr) {
+    InstrumentedMutexLock l(&mutex_);
+    while (!waiting_) {
+      cond_var_.Wait();
+    }
+    uint64_t prev_count = run_count_;
+    if (callback != nullptr) {
+      callback();
+    }
+    cond_var_.SignalAll();
+    while (!(run_count_ > prev_count)) {
+      cond_var_.Wait();
+    }
+  }
+#endif
+
+ private:
+  bool wait(uint64_t delay) {
+    InstrumentedMutexLock l(&mutex_);
+    if (running_ && delay > 0) {
+      uint64_t wait_until = clock_->NowMicros() + delay;
+#ifndef NDEBUG
+      waiting_ = true;
+      cond_var_.SignalAll();
+#endif
+      while (running_) {
+        cond_var_.TimedWait(wait_until);
+        if (clock_->NowMicros() >= wait_until) {
+          break;
+        }
+      }
+#ifndef NDEBUG
+      waiting_ = false;
+#endif
+    }
+    return running_;
+  }
+
+  void thread() {
+#if defined(_GNU_SOURCE) && defined(__GLIBC_PREREQ)
+#if __GLIBC_PREREQ(2, 12)
+    // Set thread name.
+    auto thread_handle = thread_.native_handle();
+    int ret __attribute__((__unused__)) =
+        pthread_setname_np(thread_handle, thread_name_.c_str());
+    assert(ret == 0);
+#endif
+#endif
+
+    assert(delay_us_ > 0);
+    if (!wait(initial_delay_us_)) {
+      return;
+    }
+    do {
+      function_();
+#ifndef NDEBUG
+      {
+        InstrumentedMutexLock l(&mutex_);
+        run_count_++;
+        cond_var_.SignalAll();
+      }
+#endif
+    } while (wait(delay_us_));
+  }
+
+  const std::function<void()> function_;
+  const std::string thread_name_;
+  SystemClock* clock_;
+  const uint64_t delay_us_;
+  const uint64_t initial_delay_us_;
+
+  // Mutex lock should be held when accessing running_, waiting_
+  // and run_count_.
+  InstrumentedMutex mutex_;
+  InstrumentedCondVar cond_var_;
+  bool running_;
+#ifndef NDEBUG
+  // RepeatableThread waiting for timeout.
+  bool waiting_;
+  // Times function_ had run.
+  uint64_t run_count_;
+#endif
+  port::Thread thread_;
+};
+
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/util/repeatable_thread_test.cc b/src/rocksdb/util/repeatable_thread_test.cc
new file mode 100644
index 000000000..0b3e95464
--- /dev/null
+++ b/src/rocksdb/util/repeatable_thread_test.cc
@@ -0,0 +1,111 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#include "util/repeatable_thread.h"
+
+#include <atomic>
+#include <memory>
+
+#include "db/db_test_util.h"
+#include "test_util/mock_time_env.h"
+#include "test_util/sync_point.h"
+#include "test_util/testharness.h"
+
+class RepeatableThreadTest : public testing::Test {
+ public:
+  RepeatableThreadTest()
+      : mock_clock_(std::make_shared<ROCKSDB_NAMESPACE::MockSystemClock>(
+            ROCKSDB_NAMESPACE::SystemClock::Default())) {}
+
+ protected:
+  std::shared_ptr<ROCKSDB_NAMESPACE::MockSystemClock> mock_clock_;
+};
+
+TEST_F(RepeatableThreadTest, TimedTest) {
+  constexpr uint64_t kSecond = 1000000;  // 1s = 1000000us
+  constexpr int kIteration = 3;
+  const auto& clock = ROCKSDB_NAMESPACE::SystemClock::Default();
+  ROCKSDB_NAMESPACE::port::Mutex mutex;
+  ROCKSDB_NAMESPACE::port::CondVar test_cv(&mutex);
+  int count = 0;
+  uint64_t prev_time = clock->NowMicros();
+  ROCKSDB_NAMESPACE::RepeatableThread thread(
+      [&] {
+        ROCKSDB_NAMESPACE::MutexLock l(&mutex);
+        count++;
+        uint64_t now = clock->NowMicros();
+        assert(count == 1 || prev_time + 1 * kSecond <= now);
+        prev_time = now;
+        if (count >= kIteration) {
+          test_cv.SignalAll();
+        }
+      },
+      "rt_test", clock.get(), 1 * kSecond);
+  // Wait for execution finish.
+  {
+    ROCKSDB_NAMESPACE::MutexLock l(&mutex);
+    while (count < kIteration) {
+      test_cv.Wait();
+    }
+  }
+
+  // Test cancel
+  thread.cancel();
+}
+
+TEST_F(RepeatableThreadTest, MockEnvTest) {
+  constexpr uint64_t kSecond = 1000000;  // 1s = 1000000us
+  constexpr int kIteration = 3;
+  mock_clock_->SetCurrentTime(0);  // in seconds
+  std::atomic<int> count{0};
+
+#if defined(OS_MACOSX) && !defined(NDEBUG)
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->DisableProcessing();
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->ClearAllCallBacks();
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
+      "InstrumentedCondVar::TimedWaitInternal", [&](void* arg) {
+        // Obtain the current (real) time in seconds and add 1000 extra seconds
+        // to ensure that RepeatableThread::wait invokes TimedWait with a time
+        // greater than (real) current time. This is to prevent the TimedWait
+        // function from returning immediately without sleeping and releasing
+        // the mutex on certain platforms, e.g. OS X. If TimedWait returns
+        // immediately, the mutex will not be released, and
+        // RepeatableThread::TEST_WaitForRun never has a chance to execute the
+        // callback which, in this case, updates the result returned by
+        // mock_clock->NowMicros. Consequently, RepeatableThread::wait cannot
+        // break out of the loop, causing test to hang. The extra 1000 seconds
+        // is a best-effort approach because there seems no reliable and
+        // deterministic way to provide the aforementioned guarantee. By the
+        // time RepeatableThread::wait is called, it is no guarantee that the
+        // delay + mock_clock->NowMicros will be greater than the current real
+        // time. However, 1000 seconds should be sufficient in most cases.
+        uint64_t time_us = *reinterpret_cast<uint64_t*>(arg);
+        if (time_us < mock_clock_->RealNowMicros()) {
+          *reinterpret_cast<uint64_t*>(arg) =
+              mock_clock_->RealNowMicros() + 1000;
+        }
+      });
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->EnableProcessing();
+#endif  // OS_MACOSX && !NDEBUG
+
+  ROCKSDB_NAMESPACE::RepeatableThread thread(
+      [&] { count++; }, "rt_test", mock_clock_.get(), 1 * kSecond, 1 * kSecond);
+  for (int i = 1; i <= kIteration; i++) {
+    // Bump current time
+    thread.TEST_WaitForRun([&] { mock_clock_->SetCurrentTime(i); });
+  }
+  // Test function should be exectued exactly kIteraion times.
+  ASSERT_EQ(kIteration, count.load());
+
+  // Test cancel
+  thread.cancel();
+}
+
+int main(int argc, char** argv) {
+  ROCKSDB_NAMESPACE::port::InstallStackTraceHandler();
+  ::testing::InitGoogleTest(&argc, argv);
+
+  return RUN_ALL_TESTS();
+}
diff --git a/src/rocksdb/util/ribbon_alg.h b/src/rocksdb/util/ribbon_alg.h
new file mode 100644
index 000000000..f9afefc23
--- /dev/null
+++ b/src/rocksdb/util/ribbon_alg.h
@@ -0,0 +1,1225 @@
+//  Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#pragma once
+
+#include <array>
+#include <memory>
+
+#include "rocksdb/rocksdb_namespace.h"
+#include "util/math128.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+namespace ribbon {
+
+// RIBBON PHSF & RIBBON Filter (Rapid Incremental Boolean Banding ON-the-fly)
+//
+// ribbon_alg.h: generic versions of core algorithms.
+//
+// Ribbon is a Perfect Hash Static Function construction useful as a compact
+// static Bloom filter alternative. It combines (a) a boolean (GF(2)) linear
+// system construction that approximates a Band Matrix with hashing,
+// (b) an incremental, on-the-fly Gaussian Elimination algorithm that is
+// remarkably efficient and adaptable at constructing an upper-triangular
+// band matrix from a set of band-approximating inputs from (a), and
+// (c) a storage layout that is fast and adaptable as a filter.
+//
+// Footnotes: (a) "Efficient Gauss Elimination for Near-Quadratic Matrices
+// with One Short Random Block per Row, with Applications" by Stefan
+// Walzer and Martin Dietzfelbinger ("DW paper")
+// (b) developed by Peter C. Dillinger, though not the first on-the-fly
+// GE algorithm. See "On the fly Gaussian Elimination for LT codes" by
+// Bioglio, Grangetto, Gaeta, and Sereno.
+// (c) see "interleaved" solution storage below.
+//
+// See ribbon_impl.h for high-level behavioral summary. This file focuses
+// on the core design details.
+//
+// ######################################################################
+// ################# PHSF -> static filter reduction ####################
+//
+// A Perfect Hash Static Function is a data structure representing a
+// map from anything hashable (a "key") to values of some fixed size.
+// Crucially, it is allowed to return garbage values for anything not in
+// the original set of map keys, and it is a "static" structure: entries
+// cannot be added or deleted after construction. PHSFs representing n
+// mappings to b-bit values (assume uniformly distributed) require at least
+// n * b bits to represent, or at least b bits per entry. We typically
+// describe the compactness of a PHSF by typical bits per entry as some
+// function of b. For example, the MWHC construction (k=3 "peeling")
+// requires about 1.0222*b and a variant called Xor+ requires about
+// 1.08*b + 0.5 bits per entry.
+//
+// With more hashing, a PHSF can over-approximate a set as a Bloom filter
+// does, with no FN queries and predictable false positive (FP) query
+// rate. Instead of the user providing a value to map each input key to,
+// a hash function provides the value. Keys in the original set will
+// return a positive membership query because the underlying PHSF returns
+// the same value as hashing the key. When a key is not in the original set,
+// the PHSF returns a "garbage" value, which is only equal to the key's
+// hash with (false positive) probability 1 in 2^b.
+//
+// For a matching false positive rate, standard Bloom filters require
+// 1.44*b bits per entry. Cache-local Bloom filters (like bloom_impl.h)
+// require a bit more, around 1.5*b bits per entry. Thus, a Bloom
+// alternative could save up to or nearly 1/3rd of memory and storage
+// that RocksDB uses for SST (static) Bloom filters. (Memtable Bloom filter
+// is dynamic.)
+//
+// Recommended reading:
+// "Xor Filters: Faster and Smaller Than Bloom and Cuckoo Filters"
+// by Graf and Lemire
+// First three sections of "Fast Scalable Construction of (Minimal
+// Perfect Hash) Functions" by Genuzio, Ottaviano, and Vigna
+//
+// ######################################################################
+// ################## PHSF vs. hash table vs. Bloom #####################
+//
+// You can think of traditional hash tables and related filter variants
+// such as Cuckoo filters as utilizing an "OR" construction: a hash
+// function associates a key with some slots and the data is returned if
+// the data is found in any one of those slots. The collision resolution
+// is visible in the final data structure and requires extra information.
+// For example, Cuckoo filter uses roughly 1.05b + 2 bits per entry, and
+// Golomb-Rice code (aka "GCS") as little as b + 1.5. When the data
+// structure associates each input key with data in one slot, the
+// structure implicitly constructs a (near-)minimal (near-)perfect hash
+// (MPH) of the keys, which requires at least 1.44 bits per key to
+// represent. This is why approaches with visible collision resolution
+// have a fixed + 1.5 or more in storage overhead per entry, often in
+// addition to an overhead multiplier on b.
+//
+// By contrast Bloom filters utilize an "AND" construction: a query only
+// returns true if all bit positions associated with a key are set to 1.
+// There is no collision resolution, so Bloom filters do not suffer a
+// fixed bits per entry overhead like the above structures.
+//
+// PHSFs typically use a bitwise XOR construction: the data you want is
+// not in a single slot, but in a linear combination of several slots.
+// For static data, this gives the best of "AND" and "OR" constructions:
+// avoids the +1.44 or more fixed overhead by not approximating a MPH and
+// can do much better than Bloom's 1.44 factor on b with collision
+// resolution, which here is done ahead of time and invisible at query
+// time.
+//
+// ######################################################################
+// ######################## PHSF construction ###########################
+//
+// For a typical PHSF, construction is solving a linear system of
+// equations, typically in GF(2), which is to say that values are boolean
+// and XOR serves both as addition and subtraction. We can use matrices to
+// represent the problem:
+//
+//    C    *    S    =    R
+// (n x m)   (m x b)   (n x b)
+// where C = coefficients, S = solution, R = results
+// and solving for S given C and R.
+//
+// Note that C and R each have n rows, one for each input entry for the
+// PHSF. A row in C is given by a hash function on the PHSF input key,
+// and the corresponding row in R is the b-bit value to associate with
+// that input key. (In a filter, rows of R are given by another hash
+// function on the input key.)
+//
+// On solving, the matrix S (solution) is the final PHSF data, as it
+// maps any row from the original C to its corresponding desired result
+// in R. We just have to hash our query inputs and compute a linear
+// combination of rows in S.
+//
+// In theory, we could chose m = n and let a hash function associate
+// each input key with random rows in C. A solution exists with high
+// probability, and uses essentially minimum space, b bits per entry
+// (because we set m = n) but this has terrible scaling, something
+// like O(n^2) space and O(n^3) time during construction (Gaussian
+// elimination) and O(n) query time. But computational efficiency is
+// key, and the core of this is avoiding scanning all of S to answer
+// each query.
+//
+// The traditional approach (MWHC, aka Xor filter) starts with setting
+// only some small fixed number of columns (typically k=3) to 1 for each
+// row of C, with remaining entries implicitly 0. This is implemented as
+// three hash functions over [0,m), and S can be implemented as a vector
+// of b-bit values. Now, a query only involves looking up k rows
+// (values) in S and computing their bitwise XOR. Additionally, this
+// construction can use a linear time algorithm called "peeling" for
+// finding a solution in many cases of one existing, but peeling
+// generally requires a larger space overhead factor in the solution
+// (m/n) than is required with Gaussian elimination.
+//
+// Recommended reading:
+// "Peeling Close to the Orientability Threshold - Spatial Coupling in
+// Hashing-Based Data Structures" by Stefan Walzer
+//
+// ######################################################################
+// ##################### Ribbon PHSF construction #######################
+//
+// Ribbon constructs coefficient rows essentially the same as in the
+// Walzer/Dietzfelbinger paper cited above: for some chosen fixed width
+// r (kCoeffBits in code), each key is hashed to a starting column in
+// [0, m - r] (GetStart() in code) and an r-bit sequence of boolean
+// coefficients (GetCoeffRow() in code). If you sort the rows by start,
+// the C matrix would look something like this:
+//
+// [####00000000000000000000]
+// [####00000000000000000000]
+// [000####00000000000000000]
+// [0000####0000000000000000]
+// [0000000####0000000000000]
+// [000000000####00000000000]
+// [000000000####00000000000]
+// [0000000000000####0000000]
+// [0000000000000000####0000]
+// [00000000000000000####000]
+// [00000000000000000000####]
+//
+// where each # could be a 0 or 1, chosen uniformly by a hash function.
+// (Except we typically set the start column value to 1.) This scheme
+// uses hashing to approximate a band matrix, and it has a solution iff
+// it reduces to an upper-triangular boolean r-band matrix, like this:
+//
+// [1###00000000000000000000]
+// [01##00000000000000000000]
+// [000000000000000000000000]
+// [0001###00000000000000000]
+// [000000000000000000000000]
+// [000001##0000000000000000]
+// [000000000000000000000000]
+// [00000001###0000000000000]
+// [000000001###000000000000]
+// [0000000001##000000000000]
+// ...
+// [00000000000000000000001#]
+// [000000000000000000000001]
+//
+// where we have expanded to an m x m matrix by filling with rows of
+// all zeros as needed. As in Gaussian elimination, this form is ready for
+// generating a solution through back-substitution.
+//
+// The awesome thing about the Ribbon construction (from the DW paper) is
+// how row reductions keep each row representable as a start column and
+// r coefficients, because row reductions are only needed when two rows
+// have the same number of leading zero columns. Thus, the combination
+// of those rows, the bitwise XOR of the r-bit coefficient rows, cancels
+// out the leading 1s, so starts (at least) one column later and only
+// needs (at most) r - 1 coefficients.
+//
+// ######################################################################
+// ###################### Ribbon PHSF scalability #######################
+//
+// Although more practical detail is in ribbon_impl.h, it's worth
+// understanding some of the overall benefits and limitations of the
+// Ribbon PHSFs.
+//
+// High-end scalability is a primary issue for Ribbon PHSFs, because in
+// a single Ribbon linear system with fixed r and fixed m/n ratio, the
+// solution probability approaches zero as n approaches infinity.
+// For a given n, solution probability improves with larger r and larger
+// m/n.
+//
+// By contrast, peeling-based PHSFs have somewhat worse storage ratio
+// or solution probability for small n (less than ~1000). This is
+// especially true with spatial-coupling, where benefits are only
+// notable for n on the order of 100k or 1m or more.
+//
+// To make best use of current hardware, r=128 seems to be closest to
+// a "generally good" choice for Ribbon, at least in RocksDB where SST
+// Bloom filters typically hold around 10-100k keys, and almost always
+// less than 10m keys. r=128 ribbon has a high chance of encoding success
+// (with first hash seed) when storage overhead is around 5% (m/n ~ 1.05)
+// for roughly 10k - 10m keys in a single linear system. r=64 only scales
+// up to about 10k keys with the same storage overhead. Construction and
+// access times for r=128 are similar to r=64. r=128 tracks nearly
+// twice as much data during construction, but in most cases we expect
+// the scalability benefits of r=128 vs. r=64 to make it preferred.
+//
+// A natural approach to scaling Ribbon beyond ~10m keys is splitting
+// (or "sharding") the inputs into multiple linear systems with their
+// own hash seeds. This can also help to control peak memory consumption.
+// TODO: much more to come
+//
+// ######################################################################
+// #################### Ribbon on-the-fly banding #######################
+//
+// "Banding" is what we call the process of reducing the inputs to an
+// upper-triangular r-band matrix ready for finishing a solution with
+// back-substitution. Although the DW paper presents an algorithm for
+// this ("SGauss"), the awesome properties of their construction enable
+// an even simpler, faster, and more backtrackable algorithm. In simplest
+// terms, the SGauss algorithm requires sorting the inputs by start
+// columns, but it's possible to make Gaussian elimination resemble hash
+// table insertion!
+//
+// The enhanced algorithm is based on these observations:
+// - When processing a coefficient row with first 1 in column j,
+//   - If it's the first at column j to be processed, it can be part of
+//     the banding at row j. (And that decision never overwritten, with
+//     no loss of generality!)
+//   - Else, it can be combined with existing row j and re-processed,
+//     which will look for a later "empty" row or reach "no solution".
+//
+// We call our banding algorithm "incremental" and "on-the-fly" because
+// (like hash table insertion) we are "finished" after each input
+// processed, with respect to all inputs processed so far. Although the
+// band matrix is an intermediate step to the solution structure, we have
+// eliminated intermediate steps and unnecessary data tracking for
+// banding.
+//
+// Building on "incremental" and "on-the-fly", the banding algorithm is
+// easily backtrackable because no (non-empty) rows are overwritten in
+// the banding. Thus, if we want to "try" adding an additional set of
+// inputs to the banding, we only have to record which rows were written
+// in order to efficiently backtrack to our state before considering
+// the additional set. (TODO: how this can mitigate scalability and
+// reach sub-1% overheads)
+//
+// Like in a linear-probed hash table, as the occupancy approaches and
+// surpasses 90-95%, collision resolution dominates the construction
+// time. (Ribbon doesn't usually pay at query time; see solution
+// storage below.) This means that we can speed up construction time
+// by using a higher m/n ratio, up to negative returns around 1.2.
+// At m/n ~= 1.2, which still saves memory substantially vs. Bloom
+// filter's 1.5, construction speed (including back-substitution) is not
+// far from sorting speed, but still a few times slower than cache-local
+// Bloom construction speed.
+//
+// Back-substitution from an upper-triangular boolean band matrix is
+// especially fast and easy. All the memory accesses are sequential or at
+// least local, no random. If the number of result bits (b) is a
+// compile-time constant, the back-substitution state can even be tracked
+// in CPU registers. Regardless of the solution representation, we prefer
+// column-major representation for tracking back-substitution state, as
+// r (the band width) will typically be much larger than b (result bits
+// or columns), so better to handle r-bit values b times (per solution
+// row) than b-bit values r times.
+//
+// ######################################################################
+// ##################### Ribbon solution storage ########################
+//
+// Row-major layout is typical for boolean (bit) matrices, including for
+// MWHC (Xor) filters where a query combines k b-bit values, and k is
+// typically smaller than b. Even for k=4 and b=2, at least k=4 random
+// look-ups are required regardless of layout.
+//
+// Ribbon PHSFs are quite different, however, because
+// (a) all of the solution rows relevant to a query are within a single
+// range of r rows, and
+// (b) the number of solution rows involved (r/2 on average, or r if
+// avoiding conditional accesses) is typically much greater than
+// b, the number of solution columns.
+//
+// Row-major for Ribbon PHSFs therefore tends to incur undue CPU overhead
+// by processing (up to) r entries of b bits each, where b is typically
+// less than 10 for filter applications.
+//
+// Column-major layout has poor locality because of accessing up to b
+// memory locations in different pages (and obviously cache lines). Note
+// that negative filter queries do not typically need to access all
+// solution columns, as they can return when a mismatch is found in any
+// result/solution column. This optimization doesn't always pay off on
+// recent hardware, where the penalty for unpredictable conditional
+// branching can exceed the penalty for unnecessary work, but the
+// optimization is essentially unavailable with row-major layout.
+//
+// The best compromise seems to be interleaving column-major on the small
+// scale with row-major on the large scale. For example, let a solution
+// "block" be r rows column-major encoded as b r-bit values in sequence.
+// Each query accesses (up to) 2 adjacent blocks, which will typically
+// span 1-3 cache lines in adjacent memory. We get very close to the same
+// locality as row-major, but with much faster reconstruction of each
+// result column, at least for filter applications where b is relatively
+// small and negative queries can return early.
+//
+// ######################################################################
+// ###################### Fractional result bits ########################
+//
+// Bloom filters have great flexibility that alternatives mostly do not
+// have. One of those flexibilities is in utilizing any ratio of data
+// structure bits per key. With a typical memory allocator like jemalloc,
+// this flexibility can save roughly 10% of the filters' footprint in
+// DRAM by rounding up and down filter sizes to minimize memory internal
+// fragmentation (see optimize_filters_for_memory RocksDB option).
+//
+// At first glance, PHSFs only offer a whole number of bits per "slot"
+// (m rather than number of keys n), but coefficient locality in the
+// Ribbon construction makes fractional bits/key quite possible and
+// attractive for filter applications. This works by a prefix of the
+// structure using b-1 solution columns and the rest using b solution
+// columns. See InterleavedSolutionStorage below for more detail.
+//
+// Because false positive rates are non-linear in bits/key, this approach
+// is not quite optimal in terms of information theory. In common cases,
+// we see additional space overhead up to about 1.5% vs. theoretical
+// optimal to achieve the same FP rate. We consider this a quite acceptable
+// overhead for very efficiently utilizing space that might otherwise be
+// wasted.
+//
+// This property of Ribbon even makes it "elastic." A Ribbon filter and
+// its small metadata for answering queries can be adapted into another
+// Ribbon filter filling any smaller multiple of r bits (plus small
+// metadata), with a correspondingly higher FP rate. None of the data
+// thrown away during construction needs to be recalled for this reduction.
+// Similarly a single Ribbon construction can be separated (by solution
+// column) into two or more structures (or "layers" or "levels") with
+// independent filtering ability (no FP correlation, just as solution or
+// result columns in a single structure) despite being constructed as part
+// of a single linear system. (TODO: implement)
+// See also "ElasticBF: Fine-grained and Elastic Bloom Filter Towards
+// Efficient Read for LSM-tree-based KV Stores."
+//
+
+// ######################################################################
+// ################### CODE: Ribbon core algorithms #####################
+// ######################################################################
+//
+// These algorithms are templatized for genericity but near-maximum
+// performance in a given application. The template parameters
+// adhere to informal class/struct type concepts outlined below. (This
+// code is written for C++11 so does not use formal C++ concepts.)
+
+// Rough architecture for these algorithms:
+//
+// +-----------+     +---+     +-----------------+
+// | AddInputs | --> | H | --> | BandingStorage  |
+// +-----------+     | a |     +-----------------+
+//                   | s |             |
+//                   | h |      Back substitution
+//                   | e |             V
+// +-----------+     | r |     +-----------------+
+// | Query Key | --> |   | >+< | SolutionStorage |
+// +-----------+     +---+  |  +-----------------+
+//                          V
+//                     Query result
+
+// Common to other concepts
+// concept RibbonTypes {
+//   // An unsigned integer type for an r-bit subsequence of coefficients.
+//   // r (or kCoeffBits) is taken to be sizeof(CoeffRow) * 8, as it would
+//   // generally only hurt scalability to leave bits of CoeffRow unused.
+//   typename CoeffRow;
+//   // An unsigned integer type big enough to hold a result row (b bits,
+//   // or number of solution/result columns).
+//   // In many applications, especially filters, the number of result
+//   // columns is decided at run time, so ResultRow simply needs to be
+//   // big enough for the largest number of columns allowed.
+//   typename ResultRow;
+//   // An unsigned integer type sufficient for representing the number of
+//   // rows in the solution structure, and at least the arithmetic
+//   // promotion size (usually 32 bits). uint32_t recommended because a
+//   // single Ribbon construction doesn't really scale to billions of
+//   // entries.
+//   typename Index;
+// };
+
+// ######################################################################
+// ######################## Hashers and Banding #########################
+
+// Hasher concepts abstract out hashing details.
+
+// concept PhsfQueryHasher extends RibbonTypes {
+//   // Type for a lookup key, which is hashable.
+//   typename Key;
+//
+//   // Type for hashed summary of a Key. uint64_t is recommended.
+//   typename Hash;
+//
+//   // Compute a hash value summarizing a Key
+//   Hash GetHash(const Key &) const;
+//
+//   // Given a hash value and a number of columns that can start an
+//   // r-sequence of coefficients (== m - r + 1), return the start
+//   // column to associate with that hash value. (Starts can be chosen
+//   // uniformly or "smash" extra entries into the beginning and end for
+//   // better utilization at those extremes of the structure. Details in
+//   // ribbon.impl.h)
+//   Index GetStart(Hash, Index num_starts) const;
+//
+//   // Given a hash value, return the r-bit sequence of coefficients to
+//   // associate with it. It's generally OK if
+//   //   sizeof(CoeffRow) > sizeof(Hash)
+//   // as long as the hash itself is not too prone to collisions for the
+//   // applications and the CoeffRow is generated uniformly from
+//   // available hash data, but relatively independent of the start.
+//   //
+//   // Must be non-zero, because that's required for a solution to exist
+//   // when mapping to non-zero result row. (Note: BandingAdd could be
+//   // modified to allow 0 coeff row if that only occurs with 0 result
+//   // row, which really only makes sense for filter implementation,
+//   // where both values are hash-derived. Or BandingAdd could reject 0
+//   // coeff row, forcing next seed, but that has potential problems with
+//   // generality/scalability.)
+//   CoeffRow GetCoeffRow(Hash) const;
+// };
+
+// concept FilterQueryHasher extends PhsfQueryHasher {
+//   // For building or querying a filter, this returns the expected
+//   // result row associated with a hashed input. For general PHSF,
+//   // this must return 0.
+//   //
+//   // Although not strictly required, there's a slightly better chance of
+//   // solver success if result row is masked down here to only the bits
+//   // actually needed.
+//   ResultRow GetResultRowFromHash(Hash) const;
+// }
+
+// concept BandingHasher extends FilterQueryHasher {
+//   // For a filter, this will generally be the same as Key.
+//   // For a general PHSF, it must either
+//   // (a) include a key and a result it maps to (e.g. in a std::pair), or
+//   // (b) GetResultRowFromInput looks up the result somewhere rather than
+//   // extracting it.
+//   typename AddInput;
+//
+//   // Instead of requiring a way to extract a Key from an
+//   // AddInput, we require getting the hash of the Key part
+//   // of an AddInput, which is trivial if AddInput == Key.
+//   Hash GetHash(const AddInput &) const;
+//
+//   // For building a non-filter PHSF, this extracts or looks up the result
+//   // row to associate with an input. For filter PHSF, this must return 0.
+//   ResultRow GetResultRowFromInput(const AddInput &) const;
+//
+//   // Whether the solver can assume the lowest bit of GetCoeffRow is
+//   // always 1. When true, it should improve solver efficiency slightly.
+//   static bool kFirstCoeffAlwaysOne;
+// }
+
+// Abstract storage for the the result of "banding" the inputs (Gaussian
+// elimination to an upper-triangular boolean band matrix). Because the
+// banding is an incremental / on-the-fly algorithm, this also represents
+// all the intermediate state between input entries.
+//
+// concept BandingStorage extends RibbonTypes {
+//   // Tells the banding algorithm to prefetch memory associated with
+//   // the next input before processing the current input. Generally
+//   // recommended iff the BandingStorage doesn't easily fit in CPU
+//   // cache.
+//   bool UsePrefetch() const;
+//
+//   // Prefetches (e.g. __builtin_prefetch) memory associated with a
+//   // slot index i.
+//   void Prefetch(Index i) const;
+//
+//   // Load or store CoeffRow and ResultRow for slot index i.
+//   // (Gaussian row operations involve both sides of the equation.)
+//   // Bool `for_back_subst` indicates that customizing values for
+//   // unconstrained solution rows (cr == 0) is allowed.
+//   void LoadRow(Index i, CoeffRow *cr, ResultRow *rr, bool for_back_subst)
+//        const;
+//   void StoreRow(Index i, CoeffRow cr, ResultRow rr);
+//
+//   // Returns the number of columns that can start an r-sequence of
+//   // coefficients, which is the number of slots minus r (kCoeffBits)
+//   // plus one. (m - r + 1)
+//   Index GetNumStarts() const;
+// };
+
+// Optional storage for backtracking data in banding a set of input
+// entries. It exposes an array structure which will generally be
+// used as a stack. It must be able to accommodate as many entries
+// as are passed in as inputs to `BandingAddRange`.
+//
+// concept BacktrackStorage extends RibbonTypes {
+//   // If false, backtracking support will be disabled in the algorithm.
+//   // This should preferably be an inline compile-time constant function.
+//   bool UseBacktrack() const;
+//
+//   // Records `to_save` as the `i`th backtrack entry
+//   void BacktrackPut(Index i, Index to_save);
+//
+//   // Recalls the `i`th backtrack entry
+//   Index BacktrackGet(Index i) const;
+// }
+
+// Adds a single entry to BandingStorage (and optionally, BacktrackStorage),
+// returning true if successful or false if solution is impossible with
+// current hasher (and presumably its seed) and number of "slots" (solution
+// or banding rows). (A solution is impossible when there is a linear
+// dependence among the inputs that doesn't "cancel out".)
+//
+// Pre- and post-condition: the BandingStorage represents a band matrix
+// ready for back substitution (row echelon form except for zero rows),
+// augmented with result values such that back substitution would give a
+// solution satisfying all the cr@start -> rr entries added.
+template <bool kFirstCoeffAlwaysOne, typename BandingStorage,
+          typename BacktrackStorage>
+bool BandingAdd(BandingStorage *bs, typename BandingStorage::Index start,
+                typename BandingStorage::ResultRow rr,
+                typename BandingStorage::CoeffRow cr, BacktrackStorage *bts,
+                typename BandingStorage::Index *backtrack_pos) {
+  using CoeffRow = typename BandingStorage::CoeffRow;
+  using ResultRow = typename BandingStorage::ResultRow;
+  using Index = typename BandingStorage::Index;
+
+  Index i = start;
+
+  if (!kFirstCoeffAlwaysOne) {
+    // Requires/asserts that cr != 0
+    int tz = CountTrailingZeroBits(cr);
+    i += static_cast<Index>(tz);
+    cr >>= tz;
+  }
+
+  for (;;) {
+    assert((cr & 1) == 1);
+    CoeffRow cr_at_i;
+    ResultRow rr_at_i;
+    bs->LoadRow(i, &cr_at_i, &rr_at_i, /* for_back_subst */ false);
+    if (cr_at_i == 0) {
+      bs->StoreRow(i, cr, rr);
+      bts->BacktrackPut(*backtrack_pos, i);
+      ++*backtrack_pos;
+      return true;
+    }
+    assert((cr_at_i & 1) == 1);
+    // Gaussian row reduction
+    cr ^= cr_at_i;
+    rr ^= rr_at_i;
+    if (cr == 0) {
+      // Inconsistency or (less likely) redundancy
+      break;
+    }
+    // Find relative offset of next non-zero coefficient.
+    int tz = CountTrailingZeroBits(cr);
+    i += static_cast<Index>(tz);
+    cr >>= tz;
+  }
+
+  // Failed, unless result row == 0 because e.g. a duplicate input or a
+  // stock hash collision, with same result row. (For filter, stock hash
+  // collision implies same result row.) Or we could have a full equation
+  // equal to sum of other equations, which is very possible with
+  // small range of values for result row.
+  return rr == 0;
+}
+
+// Adds a range of entries to BandingStorage returning true if successful
+// or false if solution is impossible with current hasher (and presumably
+// its seed) and number of "slots" (solution or banding rows). (A solution
+// is impossible when there is a linear dependence among the inputs that
+// doesn't "cancel out".) Here "InputIterator" is an iterator over AddInputs.
+//
+// If UseBacktrack in the BacktrackStorage, this function call rolls back
+// to prior state on failure. If !UseBacktrack, some subset of the entries
+// will have been added to the BandingStorage, so best considered to be in
+// an indeterminate state.
+//
+template <typename BandingStorage, typename BacktrackStorage,
+          typename BandingHasher, typename InputIterator>
+bool BandingAddRange(BandingStorage *bs, BacktrackStorage *bts,
+                     const BandingHasher &bh, InputIterator begin,
+                     InputIterator end) {
+  using CoeffRow = typename BandingStorage::CoeffRow;
+  using Index = typename BandingStorage::Index;
+  using ResultRow = typename BandingStorage::ResultRow;
+  using Hash = typename BandingHasher::Hash;
+
+  static_assert(IsUnsignedUpTo128<CoeffRow>::value, "must be unsigned");
+  static_assert(IsUnsignedUpTo128<Index>::value, "must be unsigned");
+  static_assert(IsUnsignedUpTo128<ResultRow>::value, "must be unsigned");
+
+  constexpr bool kFCA1 = BandingHasher::kFirstCoeffAlwaysOne;
+
+  if (begin == end) {
+    // trivial
+    return true;
+  }
+
+  const Index num_starts = bs->GetNumStarts();
+
+  InputIterator cur = begin;
+  Index backtrack_pos = 0;
+  if (!bs->UsePrefetch()) {
+    // Simple version, no prefetch
+    for (;;) {
+      Hash h = bh.GetHash(*cur);
+      Index start = bh.GetStart(h, num_starts);
+      ResultRow rr =
+          bh.GetResultRowFromInput(*cur) | bh.GetResultRowFromHash(h);
+      CoeffRow cr = bh.GetCoeffRow(h);
+
+      if (!BandingAdd<kFCA1>(bs, start, rr, cr, bts, &backtrack_pos)) {
+        break;
+      }
+      if ((++cur) == end) {
+        return true;
+      }
+    }
+  } else {
+    // Pipelined w/prefetch
+    // Prime the pipeline
+    Hash h = bh.GetHash(*cur);
+    Index start = bh.GetStart(h, num_starts);
+    ResultRow rr = bh.GetResultRowFromInput(*cur);
+    bs->Prefetch(start);
+
+    // Pipeline
+    for (;;) {
+      rr |= bh.GetResultRowFromHash(h);
+      CoeffRow cr = bh.GetCoeffRow(h);
+      if ((++cur) == end) {
+        if (!BandingAdd<kFCA1>(bs, start, rr, cr, bts, &backtrack_pos)) {
+          break;
+        }
+        return true;
+      }
+      Hash next_h = bh.GetHash(*cur);
+      Index next_start = bh.GetStart(next_h, num_starts);
+      ResultRow next_rr = bh.GetResultRowFromInput(*cur);
+      bs->Prefetch(next_start);
+      if (!BandingAdd<kFCA1>(bs, start, rr, cr, bts, &backtrack_pos)) {
+        break;
+      }
+      h = next_h;
+      start = next_start;
+      rr = next_rr;
+    }
+  }
+  // failed; backtrack (if implemented)
+  if (bts->UseBacktrack()) {
+    while (backtrack_pos > 0) {
+      --backtrack_pos;
+      Index i = bts->BacktrackGet(backtrack_pos);
+      // Clearing the ResultRow is not strictly required, but is required
+      // for good FP rate on inputs that might have been backtracked out.
+      // (We don't want anything we've backtracked on to leak into final
+      // result, as that might not be "harmless".)
+      bs->StoreRow(i, 0, 0);
+    }
+  }
+  return false;
+}
+
+// Adds a range of entries to BandingStorage returning true if successful
+// or false if solution is impossible with current hasher (and presumably
+// its seed) and number of "slots" (solution or banding rows). (A solution
+// is impossible when there is a linear dependence among the inputs that
+// doesn't "cancel out".) Here "InputIterator" is an iterator over AddInputs.
+//
+// On failure, some subset of the entries will have been added to the
+// BandingStorage, so best considered to be in an indeterminate state.
+//
+template <typename BandingStorage, typename BandingHasher,
+          typename InputIterator>
+bool BandingAddRange(BandingStorage *bs, const BandingHasher &bh,
+                     InputIterator begin, InputIterator end) {
+  using Index = typename BandingStorage::Index;
+  struct NoopBacktrackStorage {
+    bool UseBacktrack() { return false; }
+    void BacktrackPut(Index, Index) {}
+    Index BacktrackGet(Index) {
+      assert(false);
+      return 0;
+    }
+  } nbts;
+  return BandingAddRange(bs, &nbts, bh, begin, end);
+}
+
+// ######################################################################
+// ######################### Solution Storage ###########################
+
+// Back-substitution and query algorithms unfortunately depend on some
+// details of data layout in the final data structure ("solution"). Thus,
+// there is no common SolutionStorage covering all the reasonable
+// possibilities.
+
+// ###################### SimpleSolutionStorage #########################
+
+// SimpleSolutionStorage is for a row-major storage, typically with no
+// unused bits in each ResultRow. This is mostly for demonstration
+// purposes as the simplest solution storage scheme. It is relatively slow
+// for filter queries.
+
+// concept SimpleSolutionStorage extends RibbonTypes {
+//   // This is called at the beginning of back-substitution for the
+//   // solution storage to do any remaining configuration before data
+//   // is stored to it. If configuration is previously finalized, this
+//   // could be a simple assertion or even no-op. Ribbon algorithms
+//   // only call this from back-substitution, and only once per call,
+//   // before other functions here.
+//   void PrepareForNumStarts(Index num_starts) const;
+//   // Must return num_starts passed to PrepareForNumStarts, or the most
+//   // recent call to PrepareForNumStarts if this storage object can be
+//   // reused. Note that num_starts == num_slots - kCoeffBits + 1 because
+//   // there must be a run of kCoeffBits slots starting from each start.
+//   Index GetNumStarts() const;
+//   // Load the solution row (type ResultRow) for a slot
+//   ResultRow Load(Index slot_num) const;
+//   // Store the solution row (type ResultRow) for a slot
+//   void Store(Index slot_num, ResultRow data);
+// };
+
+// Back-substitution for generating a solution from BandingStorage to
+// SimpleSolutionStorage.
+template <typename SimpleSolutionStorage, typename BandingStorage>
+void SimpleBackSubst(SimpleSolutionStorage *sss, const BandingStorage &bs) {
+  using CoeffRow = typename BandingStorage::CoeffRow;
+  using Index = typename BandingStorage::Index;
+  using ResultRow = typename BandingStorage::ResultRow;
+
+  static_assert(sizeof(Index) == sizeof(typename SimpleSolutionStorage::Index),
+                "must be same");
+  static_assert(
+      sizeof(CoeffRow) == sizeof(typename SimpleSolutionStorage::CoeffRow),
+      "must be same");
+  static_assert(
+      sizeof(ResultRow) == sizeof(typename SimpleSolutionStorage::ResultRow),
+      "must be same");
+
+  constexpr auto kCoeffBits = static_cast<Index>(sizeof(CoeffRow) * 8U);
+  constexpr auto kResultBits = static_cast<Index>(sizeof(ResultRow) * 8U);
+
+  // A column-major buffer of the solution matrix, containing enough
+  // recently-computed solution data to compute the next solution row
+  // (based also on banding data).
+  std::array<CoeffRow, kResultBits> state;
+  state.fill(0);
+
+  const Index num_starts = bs.GetNumStarts();
+  sss->PrepareForNumStarts(num_starts);
+  const Index num_slots = num_starts + kCoeffBits - 1;
+
+  for (Index i = num_slots; i > 0;) {
+    --i;
+    CoeffRow cr;
+    ResultRow rr;
+    bs.LoadRow(i, &cr, &rr, /* for_back_subst */ true);
+    // solution row
+    ResultRow sr = 0;
+    for (Index j = 0; j < kResultBits; ++j) {
+      // Compute next solution bit at row i, column j (see derivation below)
+      CoeffRow tmp = state[j] << 1;
+      bool bit = (BitParity(tmp & cr) ^ ((rr >> j) & 1)) != 0;
+      tmp |= bit ? CoeffRow{1} : CoeffRow{0};
+
+      // Now tmp is solution at column j from row i for next kCoeffBits
+      // more rows. Thus, for valid solution, the dot product of the
+      // solution column with the coefficient row has to equal the result
+      // at that column,
+      //   BitParity(tmp & cr) == ((rr >> j) & 1)
+
+      // Update state.
+      state[j] = tmp;
+      // add to solution row
+      sr |= (bit ? ResultRow{1} : ResultRow{0}) << j;
+    }
+    sss->Store(i, sr);
+  }
+}
+
+// Common functionality for querying a key (already hashed) in
+// SimpleSolutionStorage.
+template <typename SimpleSolutionStorage>
+typename SimpleSolutionStorage::ResultRow SimpleQueryHelper(
+    typename SimpleSolutionStorage::Index start_slot,
+    typename SimpleSolutionStorage::CoeffRow cr,
+    const SimpleSolutionStorage &sss) {
+  using CoeffRow = typename SimpleSolutionStorage::CoeffRow;
+  using ResultRow = typename SimpleSolutionStorage::ResultRow;
+
+  constexpr unsigned kCoeffBits = static_cast<unsigned>(sizeof(CoeffRow) * 8U);
+
+  ResultRow result = 0;
+  for (unsigned i = 0; i < kCoeffBits; ++i) {
+    // Bit masking whole value is generally faster here than 'if'
+    result ^= sss.Load(start_slot + i) &
+              (ResultRow{0} - (static_cast<ResultRow>(cr >> i) & ResultRow{1}));
+  }
+  return result;
+}
+
+// General PHSF query a key from SimpleSolutionStorage.
+template <typename SimpleSolutionStorage, typename PhsfQueryHasher>
+typename SimpleSolutionStorage::ResultRow SimplePhsfQuery(
+    const typename PhsfQueryHasher::Key &key, const PhsfQueryHasher &hasher,
+    const SimpleSolutionStorage &sss) {
+  const typename PhsfQueryHasher::Hash hash = hasher.GetHash(key);
+
+  static_assert(sizeof(typename SimpleSolutionStorage::Index) ==
+                    sizeof(typename PhsfQueryHasher::Index),
+                "must be same");
+  static_assert(sizeof(typename SimpleSolutionStorage::CoeffRow) ==
+                    sizeof(typename PhsfQueryHasher::CoeffRow),
+                "must be same");
+
+  return SimpleQueryHelper(hasher.GetStart(hash, sss.GetNumStarts()),
+                           hasher.GetCoeffRow(hash), sss);
+}
+
+// Filter query a key from SimpleSolutionStorage.
+template <typename SimpleSolutionStorage, typename FilterQueryHasher>
+bool SimpleFilterQuery(const typename FilterQueryHasher::Key &key,
+                       const FilterQueryHasher &hasher,
+                       const SimpleSolutionStorage &sss) {
+  const typename FilterQueryHasher::Hash hash = hasher.GetHash(key);
+  const typename SimpleSolutionStorage::ResultRow expected =
+      hasher.GetResultRowFromHash(hash);
+
+  static_assert(sizeof(typename SimpleSolutionStorage::Index) ==
+                    sizeof(typename FilterQueryHasher::Index),
+                "must be same");
+  static_assert(sizeof(typename SimpleSolutionStorage::CoeffRow) ==
+                    sizeof(typename FilterQueryHasher::CoeffRow),
+                "must be same");
+  static_assert(sizeof(typename SimpleSolutionStorage::ResultRow) ==
+                    sizeof(typename FilterQueryHasher::ResultRow),
+                "must be same");
+
+  return expected ==
+         SimpleQueryHelper(hasher.GetStart(hash, sss.GetNumStarts()),
+                           hasher.GetCoeffRow(hash), sss);
+}
+
+// #################### InterleavedSolutionStorage ######################
+
+// InterleavedSolutionStorage is row-major at a high level, for good
+// locality, and column-major at a low level, for CPU efficiency
+// especially in filter queries or relatively small number of result bits
+// (== solution columns). The storage is a sequence of "blocks" where a
+// block has one CoeffRow-sized segment for each solution column. Each
+// query spans at most two blocks; the starting solution row is typically
+// in the row-logical middle of a block and spans to the middle of the
+// next block. (See diagram below.)
+//
+// InterleavedSolutionStorage supports choosing b (number of result or
+// solution columns) at run time, and even supports mixing b and b-1 solution
+// columns in a single linear system solution, for filters that can
+// effectively utilize any size space (multiple of CoeffRow) for minimizing
+// FP rate for any number of added keys. To simplify query implementation
+// (with lower-index columns first), the b-bit portion comes after the b-1
+// portion of the structure.
+//
+// Diagram (=== marks logical block boundary; b=4; ### is data used by a
+// query crossing the b-1 to b boundary, each Segment has type CoeffRow):
+//  ...
+// +======================+
+// | S e g m e n t  col=0 |
+// +----------------------+
+// | S e g m e n t  col=1 |
+// +----------------------+
+// | S e g m e n t  col=2 |
+// +======================+
+// | S e g m e n #########|
+// +----------------------+
+// | S e g m e n #########|
+// +----------------------+
+// | S e g m e n #########|
+// +======================+ Result/solution columns: above = 3, below = 4
+// |#############t  col=0 |
+// +----------------------+
+// |#############t  col=1 |
+// +----------------------+
+// |#############t  col=2 |
+// +----------------------+
+// | S e g m e n t  col=3 |
+// +======================+
+// | S e g m e n t  col=0 |
+// +----------------------+
+// | S e g m e n t  col=1 |
+// +----------------------+
+// | S e g m e n t  col=2 |
+// +----------------------+
+// | S e g m e n t  col=3 |
+// +======================+
+//  ...
+//
+// InterleavedSolutionStorage will be adapted by the algorithms from
+// simple array-like segment storage. That array-like storage is templatized
+// in part so that an implementation may choose to handle byte ordering
+// at access time.
+//
+// concept InterleavedSolutionStorage extends RibbonTypes {
+//   // This is called at the beginning of back-substitution for the
+//   // solution storage to do any remaining configuration before data
+//   // is stored to it. If configuration is previously finalized, this
+//   // could be a simple assertion or even no-op. Ribbon algorithms
+//   // only call this from back-substitution, and only once per call,
+//   // before other functions here.
+//   void PrepareForNumStarts(Index num_starts) const;
+//   // Must return num_starts passed to PrepareForNumStarts, or the most
+//   // recent call to PrepareForNumStarts if this storage object can be
+//   // reused. Note that num_starts == num_slots - kCoeffBits + 1 because
+//   // there must be a run of kCoeffBits slots starting from each start.
+//   Index GetNumStarts() const;
+//   // The larger number of solution columns used (called "b" above).
+//   Index GetUpperNumColumns() const;
+//   // If returns > 0, then block numbers below that use
+//   // GetUpperNumColumns() - 1 columns per solution row, and the rest
+//   // use GetUpperNumColumns(). A block represents kCoeffBits "slots",
+//   // where all but the last kCoeffBits - 1 slots are also starts. And
+//   // a block contains a segment for each solution column.
+//   // An implementation may only support uniform columns per solution
+//   // row and return constant 0 here.
+//   Index GetUpperStartBlock() const;
+//
+//   // ### "Array of segments" portion of API ###
+//   // The number of values of type CoeffRow used in this solution
+//   // representation. (This value can be inferred from the previous
+//   // three functions, but is expected at least for sanity / assertion
+//   // checking.)
+//   Index GetNumSegments() const;
+//   // Load an entry from the logical array of segments
+//   CoeffRow LoadSegment(Index segment_num) const;
+//   // Store an entry to the logical array of segments
+//   void StoreSegment(Index segment_num, CoeffRow data);
+// };
+
+// A helper for InterleavedBackSubst.
+template <typename BandingStorage>
+inline void BackSubstBlock(typename BandingStorage::CoeffRow *state,
+                           typename BandingStorage::Index num_columns,
+                           const BandingStorage &bs,
+                           typename BandingStorage::Index start_slot) {
+  using CoeffRow = typename BandingStorage::CoeffRow;
+  using Index = typename BandingStorage::Index;
+  using ResultRow = typename BandingStorage::ResultRow;
+
+  constexpr auto kCoeffBits = static_cast<Index>(sizeof(CoeffRow) * 8U);
+
+  for (Index i = start_slot + kCoeffBits; i > start_slot;) {
+    --i;
+    CoeffRow cr;
+    ResultRow rr;
+    bs.LoadRow(i, &cr, &rr, /* for_back_subst */ true);
+    for (Index j = 0; j < num_columns; ++j) {
+      // Compute next solution bit at row i, column j (see derivation below)
+      CoeffRow tmp = state[j] << 1;
+      int bit = BitParity(tmp & cr) ^ ((rr >> j) & 1);
+      tmp |= static_cast<CoeffRow>(bit);
+
+      // Now tmp is solution at column j from row i for next kCoeffBits
+      // more rows. Thus, for valid solution, the dot product of the
+      // solution column with the coefficient row has to equal the result
+      // at that column,
+      //   BitParity(tmp & cr) == ((rr >> j) & 1)
+
+      // Update state.
+      state[j] = tmp;
+    }
+  }
+}
+
+// Back-substitution for generating a solution from BandingStorage to
+// InterleavedSolutionStorage.
+template <typename InterleavedSolutionStorage, typename BandingStorage>
+void InterleavedBackSubst(InterleavedSolutionStorage *iss,
+                          const BandingStorage &bs) {
+  using CoeffRow = typename BandingStorage::CoeffRow;
+  using Index = typename BandingStorage::Index;
+
+  static_assert(
+      sizeof(Index) == sizeof(typename InterleavedSolutionStorage::Index),
+      "must be same");
+  static_assert(
+      sizeof(CoeffRow) == sizeof(typename InterleavedSolutionStorage::CoeffRow),
+      "must be same");
+
+  constexpr auto kCoeffBits = static_cast<Index>(sizeof(CoeffRow) * 8U);
+
+  const Index num_starts = bs.GetNumStarts();
+  // Although it might be nice to have a filter that returns "always false"
+  // when no key is added, we aren't specifically supporting that here
+  // because it would require another condition branch in the query.
+  assert(num_starts > 0);
+  iss->PrepareForNumStarts(num_starts);
+
+  const Index num_slots = num_starts + kCoeffBits - 1;
+  assert(num_slots % kCoeffBits == 0);
+  const Index num_blocks = num_slots / kCoeffBits;
+  const Index num_segments = iss->GetNumSegments();
+
+  // For now upper, then lower
+  Index num_columns = iss->GetUpperNumColumns();
+  const Index upper_start_block = iss->GetUpperStartBlock();
+
+  if (num_columns == 0) {
+    // Nothing to do, presumably because there's not enough space for even
+    // a single segment.
+    assert(num_segments == 0);
+    // When num_columns == 0, a Ribbon filter query will always return true,
+    // or a PHSF query always 0.
+    return;
+  }
+
+  // We should be utilizing all available segments
+  assert(num_segments == (upper_start_block * (num_columns - 1)) +
+                             ((num_blocks - upper_start_block) * num_columns));
+
+  // TODO: consider fixed-column specializations with stack-allocated state
+
+  // A column-major buffer of the solution matrix, containing enough
+  // recently-computed solution data to compute the next solution row
+  // (based also on banding data).
+  std::unique_ptr<CoeffRow[]> state{new CoeffRow[num_columns]()};
+
+  Index block = num_blocks;
+  Index segment_num = num_segments;
+  while (block > upper_start_block) {
+    --block;
+    BackSubstBlock(state.get(), num_columns, bs, block * kCoeffBits);
+    segment_num -= num_columns;
+    for (Index i = 0; i < num_columns; ++i) {
+      iss->StoreSegment(segment_num + i, state[i]);
+    }
+  }
+  // Now (if applicable), region using lower number of columns
+  // (This should be optimized away if GetUpperStartBlock() returns
+  // constant 0.)
+  --num_columns;
+  while (block > 0) {
+    --block;
+    BackSubstBlock(state.get(), num_columns, bs, block * kCoeffBits);
+    segment_num -= num_columns;
+    for (Index i = 0; i < num_columns; ++i) {
+      iss->StoreSegment(segment_num + i, state[i]);
+    }
+  }
+  // Verify everything processed
+  assert(block == 0);
+  assert(segment_num == 0);
+}
+
+// Prefetch memory for a key in InterleavedSolutionStorage.
+template <typename InterleavedSolutionStorage, typename PhsfQueryHasher>
+inline void InterleavedPrepareQuery(
+    const typename PhsfQueryHasher::Key &key, const PhsfQueryHasher &hasher,
+    const InterleavedSolutionStorage &iss,
+    typename PhsfQueryHasher::Hash *saved_hash,
+    typename InterleavedSolutionStorage::Index *saved_segment_num,
+    typename InterleavedSolutionStorage::Index *saved_num_columns,
+    typename InterleavedSolutionStorage::Index *saved_start_bit) {
+  using Hash = typename PhsfQueryHasher::Hash;
+  using CoeffRow = typename InterleavedSolutionStorage::CoeffRow;
+  using Index = typename InterleavedSolutionStorage::Index;
+
+  static_assert(sizeof(Index) == sizeof(typename PhsfQueryHasher::Index),
+                "must be same");
+
+  const Hash hash = hasher.GetHash(key);
+  const Index start_slot = hasher.GetStart(hash, iss.GetNumStarts());
+
+  constexpr auto kCoeffBits = static_cast<Index>(sizeof(CoeffRow) * 8U);
+
+  const Index upper_start_block = iss.GetUpperStartBlock();
+  Index num_columns = iss.GetUpperNumColumns();
+  Index start_block_num = start_slot / kCoeffBits;
+  Index segment_num = start_block_num * num_columns -
+                      std::min(start_block_num, upper_start_block);
+  // Change to lower num columns if applicable.
+  // (This should not compile to a conditional branch.)
+  num_columns -= (start_block_num < upper_start_block) ? 1 : 0;
+
+  Index start_bit = start_slot % kCoeffBits;
+
+  Index segment_count = num_columns + (start_bit == 0 ? 0 : num_columns);
+
+  iss.PrefetchSegmentRange(segment_num, segment_num + segment_count);
+
+  *saved_hash = hash;
+  *saved_segment_num = segment_num;
+  *saved_num_columns = num_columns;
+  *saved_start_bit = start_bit;
+}
+
+// General PHSF query from InterleavedSolutionStorage, using data for
+// the query key from InterleavedPrepareQuery
+template <typename InterleavedSolutionStorage, typename PhsfQueryHasher>
+inline typename InterleavedSolutionStorage::ResultRow InterleavedPhsfQuery(
+    typename PhsfQueryHasher::Hash hash,
+    typename InterleavedSolutionStorage::Index segment_num,
+    typename InterleavedSolutionStorage::Index num_columns,
+    typename InterleavedSolutionStorage::Index start_bit,
+    const PhsfQueryHasher &hasher, const InterleavedSolutionStorage &iss) {
+  using CoeffRow = typename InterleavedSolutionStorage::CoeffRow;
+  using Index = typename InterleavedSolutionStorage::Index;
+  using ResultRow = typename InterleavedSolutionStorage::ResultRow;
+
+  static_assert(sizeof(Index) == sizeof(typename PhsfQueryHasher::Index),
+                "must be same");
+  static_assert(sizeof(CoeffRow) == sizeof(typename PhsfQueryHasher::CoeffRow),
+                "must be same");
+
+  constexpr auto kCoeffBits = static_cast<Index>(sizeof(CoeffRow) * 8U);
+
+  const CoeffRow cr = hasher.GetCoeffRow(hash);
+
+  ResultRow sr = 0;
+  const CoeffRow cr_left = cr << static_cast<unsigned>(start_bit);
+  for (Index i = 0; i < num_columns; ++i) {
+    sr ^= BitParity(iss.LoadSegment(segment_num + i) & cr_left) << i;
+  }
+
+  if (start_bit > 0) {
+    segment_num += num_columns;
+    const CoeffRow cr_right =
+        cr >> static_cast<unsigned>(kCoeffBits - start_bit);
+    for (Index i = 0; i < num_columns; ++i) {
+      sr ^= BitParity(iss.LoadSegment(segment_num + i) & cr_right) << i;
+    }
+  }
+
+  return sr;
+}
+
+// Filter query a key from InterleavedFilterQuery.
+template <typename InterleavedSolutionStorage, typename FilterQueryHasher>
+inline bool InterleavedFilterQuery(
+    typename FilterQueryHasher::Hash hash,
+    typename InterleavedSolutionStorage::Index segment_num,
+    typename InterleavedSolutionStorage::Index num_columns,
+    typename InterleavedSolutionStorage::Index start_bit,
+    const FilterQueryHasher &hasher, const InterleavedSolutionStorage &iss) {
+  using CoeffRow = typename InterleavedSolutionStorage::CoeffRow;
+  using Index = typename InterleavedSolutionStorage::Index;
+  using ResultRow = typename InterleavedSolutionStorage::ResultRow;
+
+  static_assert(sizeof(Index) == sizeof(typename FilterQueryHasher::Index),
+                "must be same");
+  static_assert(
+      sizeof(CoeffRow) == sizeof(typename FilterQueryHasher::CoeffRow),
+      "must be same");
+  static_assert(
+      sizeof(ResultRow) == sizeof(typename FilterQueryHasher::ResultRow),
+      "must be same");
+
+  constexpr auto kCoeffBits = static_cast<Index>(sizeof(CoeffRow) * 8U);
+
+  const CoeffRow cr = hasher.GetCoeffRow(hash);
+  const ResultRow expected = hasher.GetResultRowFromHash(hash);
+
+  // TODO: consider optimizations such as
+  // * get rid of start_bit == 0 condition with careful fetching & shifting
+  if (start_bit == 0) {
+    for (Index i = 0; i < num_columns; ++i) {
+      if (BitParity(iss.LoadSegment(segment_num + i) & cr) !=
+          (static_cast<int>(expected >> i) & 1)) {
+        return false;
+      }
+    }
+  } else {
+    const CoeffRow cr_left = cr << static_cast<unsigned>(start_bit);
+    const CoeffRow cr_right =
+        cr >> static_cast<unsigned>(kCoeffBits - start_bit);
+
+    for (Index i = 0; i < num_columns; ++i) {
+      CoeffRow soln_data =
+          (iss.LoadSegment(segment_num + i) & cr_left) ^
+          (iss.LoadSegment(segment_num + num_columns + i) & cr_right);
+      if (BitParity(soln_data) != (static_cast<int>(expected >> i) & 1)) {
+        return false;
+      }
+    }
+  }
+  // otherwise, all match
+  return true;
+}
+
+// TODO: refactor Interleaved*Query so that queries can be "prepared" by
+// prefetching memory, to hide memory latency for multiple queries in a
+// single thread.
+
+}  // namespace ribbon
+
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/util/ribbon_config.cc b/src/rocksdb/util/ribbon_config.cc
new file mode 100644
index 000000000..c1046f4aa
--- /dev/null
+++ b/src/rocksdb/util/ribbon_config.cc
@@ -0,0 +1,506 @@
+//  Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#include "util/ribbon_config.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+namespace ribbon {
+
+namespace detail {
+
+// Each instantiation of this struct is sufficiently unique for configuration
+// purposes, and is only instantiated for settings where we support the
+// configuration API. An application might only reference one instantiation,
+// meaning the rest could be pruned at link time.
+template <ConstructionFailureChance kCfc, uint64_t kCoeffBits, bool kUseSmash>
+struct BandingConfigHelperData {
+  static constexpr size_t kKnownSize = 18U;
+
+  // Because of complexity in the data, for smaller numbers of slots
+  // (powers of two up to 2^17), we record known numbers that can be added
+  // with kCfc chance of construction failure and settings in template
+  // parameters. Zero means "unsupported (too small) number of slots".
+  // (GetNumToAdd below will use interpolation for numbers of slots
+  // between powers of two; double rather than integer values here make
+  // that more accurate.)
+  static const std::array<double, kKnownSize> kKnownToAddByPow2;
+
+  // For sufficiently large number of slots, doubling the number of
+  // slots will increase the expected overhead (slots over number added)
+  // by approximately this constant.
+  // (This is roughly constant regardless of ConstructionFailureChance and
+  // smash setting.)
+  // (Would be a constant if we had partial template specialization for
+  // static const members.)
+  static inline double GetFactorPerPow2() {
+    if (kCoeffBits == 128U) {
+      return 0.0038;
+    } else {
+      assert(kCoeffBits == 64U);
+      return 0.0083;
+    }
+  }
+
+  // Overhead factor for 2^(kKnownSize-1) slots
+  // (Would be a constant if we had partial template specialization for
+  // static const members.)
+  static inline double GetFinalKnownFactor() {
+    return 1.0 * (uint32_t{1} << (kKnownSize - 1)) /
+           kKnownToAddByPow2[kKnownSize - 1];
+  }
+
+  // GetFinalKnownFactor() - (kKnownSize-1) * GetFactorPerPow2()
+  // (Would be a constant if we had partial template specialization for
+  // static const members.)
+  static inline double GetBaseFactor() {
+    return GetFinalKnownFactor() - (kKnownSize - 1) * GetFactorPerPow2();
+  }
+
+  // Get overhead factor (slots over number to add) for sufficiently large
+  // number of slots (by log base 2)
+  static inline double GetFactorForLarge(double log2_num_slots) {
+    return GetBaseFactor() + log2_num_slots * GetFactorPerPow2();
+  }
+
+  // For a given power of two number of slots (specified by whole number
+  // log base 2), implements GetNumToAdd for such limited case, returning
+  // double for better interpolation in GetNumToAdd and GetNumSlots.
+  static inline double GetNumToAddForPow2(uint32_t log2_num_slots) {
+    assert(log2_num_slots <= 32);  // help clang-analyze
+    if (log2_num_slots < kKnownSize) {
+      return kKnownToAddByPow2[log2_num_slots];
+    } else {
+      return 1.0 * (uint64_t{1} << log2_num_slots) /
+             GetFactorForLarge(1.0 * log2_num_slots);
+    }
+  }
+};
+
+// Based on data from FindOccupancy in ribbon_test
+template <>
+const std::array<double, 18>
+    BandingConfigHelperData<kOneIn2, 128U, false>::kKnownToAddByPow2{{
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,  // unsupported
+        252.984,
+        506.109,
+        1013.71,
+        2029.47,
+        4060.43,
+        8115.63,
+        16202.2,
+        32305.1,
+        64383.5,
+        128274,
+    }};
+
+template <>
+const std::array<double, 18>
+    BandingConfigHelperData<kOneIn2, 128U, /*smash*/ true>::kKnownToAddByPow2{{
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,  // unsupported
+        126.274,
+        254.279,
+        510.27,
+        1022.24,
+        2046.02,
+        4091.99,
+        8154.98,
+        16244.3,
+        32349.7,
+        64426.6,
+        128307,
+    }};
+
+template <>
+const std::array<double, 18>
+    BandingConfigHelperData<kOneIn2, 64U, false>::kKnownToAddByPow2{{
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,  // unsupported
+        124.94,
+        249.968,
+        501.234,
+        1004.06,
+        2006.15,
+        3997.89,
+        7946.99,
+        15778.4,
+        31306.9,
+        62115.3,
+        123284,
+    }};
+
+template <>
+const std::array<double, 18>
+    BandingConfigHelperData<kOneIn2, 64U, /*smash*/ true>::kKnownToAddByPow2{{
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,  // unsupported
+        62.2683,
+        126.259,
+        254.268,
+        509.975,
+        1019.98,
+        2026.16,
+        4019.75,
+        7969.8,
+        15798.2,
+        31330.3,
+        62134.2,
+        123255,
+    }};
+
+template <>
+const std::array<double, 18>
+    BandingConfigHelperData<kOneIn20, 128U, false>::kKnownToAddByPow2{{
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,  // unsupported
+        248.851,
+        499.532,
+        1001.26,
+        2003.97,
+        4005.59,
+        8000.39,
+        15966.6,
+        31828.1,
+        63447.3,
+        126506,
+    }};
+
+template <>
+const std::array<double, 18>
+    BandingConfigHelperData<kOneIn20, 128U, /*smash*/ true>::kKnownToAddByPow2{{
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,  // unsupported
+        122.637,
+        250.651,
+        506.625,
+        1018.54,
+        2036.43,
+        4041.6,
+        8039.25,
+        16005,
+        31869.6,
+        63492.8,
+        126537,
+    }};
+
+template <>
+const std::array<double, 18>
+    BandingConfigHelperData<kOneIn20, 64U, false>::kKnownToAddByPow2{{
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,  // unsupported
+        120.659,
+        243.346,
+        488.168,
+        976.373,
+        1948.86,
+        3875.85,
+        7704.97,
+        15312.4,
+        30395.1,
+        60321.8,
+        119813,
+    }};
+
+template <>
+const std::array<double, 18>
+    BandingConfigHelperData<kOneIn20, 64U, /*smash*/ true>::kKnownToAddByPow2{{
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,  // unsupported
+        58.6016,
+        122.619,
+        250.641,
+        503.595,
+        994.165,
+        1967.36,
+        3898.17,
+        7727.21,
+        15331.5,
+        30405.8,
+        60376.2,
+        119836,
+    }};
+
+template <>
+const std::array<double, 18>
+    BandingConfigHelperData<kOneIn1000, 128U, false>::kKnownToAddByPow2{{
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,  // unsupported
+        242.61,
+        491.887,
+        983.603,
+        1968.21,
+        3926.98,
+        7833.99,
+        15629,
+        31199.9,
+        62307.8,
+        123870,
+    }};
+
+template <>
+const std::array<double, 18> BandingConfigHelperData<
+    kOneIn1000, 128U, /*smash*/ true>::kKnownToAddByPow2{{
+    0,
+    0,
+    0,
+    0,
+    0,
+    0,
+    0,  // unsupported
+    117.19,
+    245.105,
+    500.748,
+    1010.67,
+    1993.4,
+    3950.01,
+    7863.31,
+    15652,
+    31262.1,
+    62462.8,
+    124095,
+}};
+
+template <>
+const std::array<double, 18>
+    BandingConfigHelperData<kOneIn1000, 64U, false>::kKnownToAddByPow2{{
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,  // unsupported
+        114,
+        234.8,
+        471.498,
+        940.165,
+        1874,
+        3721.5,
+        7387.5,
+        14592,
+        29160,
+        57745,
+        115082,
+    }};
+
+template <>
+const std::array<double, 18>
+    BandingConfigHelperData<kOneIn1000, 64U, /*smash*/ true>::kKnownToAddByPow2{
+        {
+            0,
+            0,
+            0,
+            0,
+            0,
+            0,  // unsupported
+            53.0434,
+            117,
+            245.312,
+            483.571,
+            950.251,
+            1878,
+            3736.34,
+            7387.97,
+            14618,
+            29142.9,
+            57838.8,
+            114932,
+        }};
+
+// We hide these implementation details from the .h file with explicit
+// instantiations below these partial specializations.
+
+template <ConstructionFailureChance kCfc, uint64_t kCoeffBits, bool kUseSmash,
+          bool kHomogeneous>
+uint32_t BandingConfigHelper1MaybeSupported<
+    kCfc, kCoeffBits, kUseSmash, kHomogeneous,
+    true /* kIsSupported */>::GetNumToAdd(uint32_t num_slots) {
+  using Data = detail::BandingConfigHelperData<kCfc, kCoeffBits, kUseSmash>;
+  if (num_slots == 0) {
+    return 0;
+  }
+  uint32_t num_to_add;
+  double log2_num_slots = std::log(num_slots) * 1.4426950409;
+  uint32_t floor_log2 = static_cast<uint32_t>(log2_num_slots);
+  if (floor_log2 + 1 < Data::kKnownSize) {
+    double ceil_portion = 1.0 * num_slots / (uint32_t{1} << floor_log2) - 1.0;
+    // Must be a supported number of slots
+    assert(Data::kKnownToAddByPow2[floor_log2] > 0.0);
+    // Weighted average of two nearest known data points
+    num_to_add = static_cast<uint32_t>(
+        ceil_portion * Data::kKnownToAddByPow2[floor_log2 + 1] +
+        (1.0 - ceil_portion) * Data::kKnownToAddByPow2[floor_log2]);
+  } else {
+    // Use formula for large values
+    double factor = Data::GetFactorForLarge(log2_num_slots);
+    assert(factor >= 1.0);
+    num_to_add = static_cast<uint32_t>(num_slots / factor);
+  }
+  if (kHomogeneous) {
+    // Even when standard filter construction would succeed, we might
+    // have loaded things up too much for Homogeneous filter. (Complete
+    // explanation not known but observed empirically.) This seems to
+    // correct for that, mostly affecting small filter configurations.
+    if (num_to_add >= 8) {
+      num_to_add -= 8;
+    } else {
+      assert(false);
+    }
+  }
+  return num_to_add;
+}
+
+template <ConstructionFailureChance kCfc, uint64_t kCoeffBits, bool kUseSmash,
+          bool kHomogeneous>
+uint32_t BandingConfigHelper1MaybeSupported<
+    kCfc, kCoeffBits, kUseSmash, kHomogeneous,
+    true /* kIsSupported */>::GetNumSlots(uint32_t num_to_add) {
+  using Data = detail::BandingConfigHelperData<kCfc, kCoeffBits, kUseSmash>;
+
+  if (num_to_add == 0) {
+    return 0;
+  }
+  if (kHomogeneous) {
+    // Reverse of above in GetNumToAdd
+    num_to_add += 8;
+  }
+  double log2_num_to_add = std::log(num_to_add) * 1.4426950409;
+  uint32_t approx_log2_slots = static_cast<uint32_t>(log2_num_to_add + 0.5);
+  assert(approx_log2_slots <= 32);  // help clang-analyze
+
+  double lower_num_to_add = Data::GetNumToAddForPow2(approx_log2_slots);
+  double upper_num_to_add;
+  if (approx_log2_slots == 0 || lower_num_to_add == /* unsupported */ 0) {
+    // Return minimum non-zero slots in standard implementation
+    return kUseSmash ? kCoeffBits : 2 * kCoeffBits;
+  } else if (num_to_add < lower_num_to_add) {
+    upper_num_to_add = lower_num_to_add;
+    --approx_log2_slots;
+    lower_num_to_add = Data::GetNumToAddForPow2(approx_log2_slots);
+  } else {
+    upper_num_to_add = Data::GetNumToAddForPow2(approx_log2_slots + 1);
+  }
+
+  assert(num_to_add >= lower_num_to_add);
+  assert(num_to_add < upper_num_to_add);
+
+  double upper_portion =
+      (num_to_add - lower_num_to_add) / (upper_num_to_add - lower_num_to_add);
+
+  double lower_num_slots = 1.0 * (uint64_t{1} << approx_log2_slots);
+
+  // Interpolation, round up
+  return static_cast<uint32_t>(upper_portion * lower_num_slots +
+                               lower_num_slots + 0.999999999);
+}
+
+// These explicit instantiations enable us to hide most of the
+// implementation details from the .h file. (The .h file currently
+// needs to determine whether settings are "supported" or not.)
+
+template struct BandingConfigHelper1MaybeSupported<kOneIn2, 128U, /*sm*/ false,
+                                                   /*hm*/ false, /*sup*/ true>;
+template struct BandingConfigHelper1MaybeSupported<kOneIn2, 128U, /*sm*/ true,
+                                                   /*hm*/ false, /*sup*/ true>;
+template struct BandingConfigHelper1MaybeSupported<kOneIn2, 128U, /*sm*/ false,
+                                                   /*hm*/ true, /*sup*/ true>;
+template struct BandingConfigHelper1MaybeSupported<kOneIn2, 128U, /*sm*/ true,
+                                                   /*hm*/ true, /*sup*/ true>;
+template struct BandingConfigHelper1MaybeSupported<kOneIn2, 64U, /*sm*/ false,
+                                                   /*hm*/ false, /*sup*/ true>;
+template struct BandingConfigHelper1MaybeSupported<kOneIn2, 64U, /*sm*/ true,
+                                                   /*hm*/ false, /*sup*/ true>;
+template struct BandingConfigHelper1MaybeSupported<kOneIn2, 64U, /*sm*/ false,
+                                                   /*hm*/ true, /*sup*/ true>;
+template struct BandingConfigHelper1MaybeSupported<kOneIn2, 64U, /*sm*/ true,
+                                                   /*hm*/ true, /*sup*/ true>;
+
+template struct BandingConfigHelper1MaybeSupported<kOneIn20, 128U, /*sm*/ false,
+                                                   /*hm*/ false, /*sup*/ true>;
+template struct BandingConfigHelper1MaybeSupported<kOneIn20, 128U, /*sm*/ true,
+                                                   /*hm*/ false, /*sup*/ true>;
+template struct BandingConfigHelper1MaybeSupported<kOneIn20, 128U, /*sm*/ false,
+                                                   /*hm*/ true, /*sup*/ true>;
+template struct BandingConfigHelper1MaybeSupported<kOneIn20, 128U, /*sm*/ true,
+                                                   /*hm*/ true, /*sup*/ true>;
+template struct BandingConfigHelper1MaybeSupported<kOneIn20, 64U, /*sm*/ false,
+                                                   /*hm*/ false, /*sup*/ true>;
+template struct BandingConfigHelper1MaybeSupported<kOneIn20, 64U, /*sm*/ true,
+                                                   /*hm*/ false, /*sup*/ true>;
+template struct BandingConfigHelper1MaybeSupported<kOneIn20, 64U, /*sm*/ false,
+                                                   /*hm*/ true, /*sup*/ true>;
+template struct BandingConfigHelper1MaybeSupported<kOneIn20, 64U, /*sm*/ true,
+                                                   /*hm*/ true, /*sup*/ true>;
+
+template struct BandingConfigHelper1MaybeSupported<
+    kOneIn1000, 128U, /*sm*/ false, /*hm*/ false, /*sup*/ true>;
+template struct BandingConfigHelper1MaybeSupported<
+    kOneIn1000, 128U, /*sm*/ true, /*hm*/ false, /*sup*/ true>;
+template struct BandingConfigHelper1MaybeSupported<
+    kOneIn1000, 128U, /*sm*/ false, /*hm*/ true, /*sup*/ true>;
+template struct BandingConfigHelper1MaybeSupported<
+    kOneIn1000, 128U, /*sm*/ true, /*hm*/ true, /*sup*/ true>;
+template struct BandingConfigHelper1MaybeSupported<
+    kOneIn1000, 64U, /*sm*/ false, /*hm*/ false, /*sup*/ true>;
+template struct BandingConfigHelper1MaybeSupported<kOneIn1000, 64U, /*sm*/ true,
+                                                   /*hm*/ false, /*sup*/ true>;
+template struct BandingConfigHelper1MaybeSupported<
+    kOneIn1000, 64U, /*sm*/ false, /*hm*/ true, /*sup*/ true>;
+template struct BandingConfigHelper1MaybeSupported<kOneIn1000, 64U, /*sm*/ true,
+                                                   /*hm*/ true, /*sup*/ true>;
+
+}  // namespace detail
+
+}  // namespace ribbon
+
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/util/ribbon_config.h b/src/rocksdb/util/ribbon_config.h
new file mode 100644
index 000000000..0e3edf073
--- /dev/null
+++ b/src/rocksdb/util/ribbon_config.h
@@ -0,0 +1,182 @@
+//  Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#pragma once
+
+#include <array>
+#include <cassert>
+#include <cmath>
+#include <cstdint>
+
+#include "port/lang.h"  // for FALLTHROUGH_INTENDED
+#include "rocksdb/rocksdb_namespace.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+namespace ribbon {
+
+// RIBBON PHSF & RIBBON Filter (Rapid Incremental Boolean Banding ON-the-fly)
+//
+// ribbon_config.h: APIs for relating numbers of slots with numbers of
+// additions for tolerable construction failure probabilities. This is
+// separate from ribbon_impl.h because it might not be needed for
+// some applications.
+//
+// This API assumes uint32_t for number of slots, as a single Ribbon
+// linear system should not normally overflow that without big penalties.
+//
+// Template parameter kCoeffBits uses uint64_t for convenience in case it
+// comes from size_t.
+//
+// Most of the complexity here is trying to optimize speed and
+// compiled code size, using templates to minimize table look-ups and
+// the compiled size of all linked look-up tables. Look-up tables are
+// required because we don't have good formulas, and the data comes
+// from running FindOccupancy in ribbon_test.
+
+// Represents a chosen chance of successful Ribbon construction for a single
+// seed. Allowing higher chance of failed construction can reduce space
+// overhead but takes extra time in construction.
+enum ConstructionFailureChance {
+  kOneIn2,
+  kOneIn20,
+  // When using kHomogeneous==true, construction failure chance should
+  // not generally exceed target FP rate, so it unlikely useful to
+  // allow a higher "failure" chance. In some cases, even more overhead
+  // is appropriate. (TODO)
+  kOneIn1000,
+};
+
+namespace detail {
+
+// It is useful to compile ribbon_test linking to BandingConfigHelper with
+// settings for which we do not have configuration data, as long as we don't
+// run the code. This template hack supports that.
+template <ConstructionFailureChance kCfc, uint64_t kCoeffBits, bool kUseSmash,
+          bool kHomogeneous, bool kIsSupported>
+struct BandingConfigHelper1MaybeSupported {
+ public:
+  static uint32_t GetNumToAdd(uint32_t num_slots) {
+    // Unsupported
+    assert(num_slots == 0);
+    (void)num_slots;
+    return 0;
+  }
+
+  static uint32_t GetNumSlots(uint32_t num_to_add) {
+    // Unsupported
+    assert(num_to_add == 0);
+    (void)num_to_add;
+    return 0;
+  }
+};
+
+// Base class for BandingConfigHelper1 and helper for BandingConfigHelper
+// with core implementations built on above data
+template <ConstructionFailureChance kCfc, uint64_t kCoeffBits, bool kUseSmash,
+          bool kHomogeneous>
+struct BandingConfigHelper1MaybeSupported<
+    kCfc, kCoeffBits, kUseSmash, kHomogeneous, true /* kIsSupported */> {
+ public:
+  // See BandingConfigHelper1. Implementation in ribbon_config.cc
+  static uint32_t GetNumToAdd(uint32_t num_slots);
+
+  // See BandingConfigHelper1. Implementation in ribbon_config.cc
+  static uint32_t GetNumSlots(uint32_t num_to_add);
+};
+
+}  // namespace detail
+
+template <ConstructionFailureChance kCfc, uint64_t kCoeffBits, bool kUseSmash,
+          bool kHomogeneous>
+struct BandingConfigHelper1
+    : public detail::BandingConfigHelper1MaybeSupported<
+          kCfc, kCoeffBits, kUseSmash, kHomogeneous,
+          /* kIsSupported */ kCoeffBits == 64 || kCoeffBits == 128> {
+ public:
+  // Returns a number of entries that can be added to a given number of
+  // slots, with roughly kCfc chance of construction failure per seed,
+  // or better. Does NOT do rounding for InterleavedSoln; call
+  // RoundUpNumSlots for that.
+  //
+  // inherited:
+  // static uint32_t GetNumToAdd(uint32_t num_slots);
+
+  // Returns a number of slots for a given number of entries to add
+  // that should have roughly kCfc chance of construction failure per
+  // seed, or better. Does NOT do rounding for InterleavedSoln; call
+  // RoundUpNumSlots for that.
+  //
+  // num_to_add should not exceed roughly 2/3rds of the maximum value
+  // of the uint32_t type to avoid overflow.
+  //
+  // inherited:
+  // static uint32_t GetNumSlots(uint32_t num_to_add);
+};
+
+// Configured using TypesAndSettings as in ribbon_impl.h
+template <ConstructionFailureChance kCfc, class TypesAndSettings>
+struct BandingConfigHelper1TS
+    : public BandingConfigHelper1<
+          kCfc,
+          /* kCoeffBits */ sizeof(typename TypesAndSettings::CoeffRow) * 8U,
+          TypesAndSettings::kUseSmash, TypesAndSettings::kHomogeneous> {};
+
+// Like BandingConfigHelper1TS except failure chance can be a runtime rather
+// than compile time value.
+template <class TypesAndSettings>
+struct BandingConfigHelper {
+ public:
+  static constexpr ConstructionFailureChance kDefaultFailureChance =
+      TypesAndSettings::kHomogeneous ? kOneIn1000 : kOneIn20;
+
+  static uint32_t GetNumToAdd(
+      uint32_t num_slots,
+      ConstructionFailureChance max_failure = kDefaultFailureChance) {
+    switch (max_failure) {
+      default:
+        assert(false);
+        FALLTHROUGH_INTENDED;
+      case kOneIn20: {
+        using H1 = BandingConfigHelper1TS<kOneIn20, TypesAndSettings>;
+        return H1::GetNumToAdd(num_slots);
+      }
+      case kOneIn2: {
+        using H1 = BandingConfigHelper1TS<kOneIn2, TypesAndSettings>;
+        return H1::GetNumToAdd(num_slots);
+      }
+      case kOneIn1000: {
+        using H1 = BandingConfigHelper1TS<kOneIn1000, TypesAndSettings>;
+        return H1::GetNumToAdd(num_slots);
+      }
+    }
+  }
+
+  static uint32_t GetNumSlots(
+      uint32_t num_to_add,
+      ConstructionFailureChance max_failure = kDefaultFailureChance) {
+    switch (max_failure) {
+      default:
+        assert(false);
+        FALLTHROUGH_INTENDED;
+      case kOneIn20: {
+        using H1 = BandingConfigHelper1TS<kOneIn20, TypesAndSettings>;
+        return H1::GetNumSlots(num_to_add);
+      }
+      case kOneIn2: {
+        using H1 = BandingConfigHelper1TS<kOneIn2, TypesAndSettings>;
+        return H1::GetNumSlots(num_to_add);
+      }
+      case kOneIn1000: {
+        using H1 = BandingConfigHelper1TS<kOneIn1000, TypesAndSettings>;
+        return H1::GetNumSlots(num_to_add);
+      }
+    }
+  }
+};
+
+}  // namespace ribbon
+
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/util/ribbon_impl.h b/src/rocksdb/util/ribbon_impl.h
new file mode 100644
index 000000000..0afecc67d
--- /dev/null
+++ b/src/rocksdb/util/ribbon_impl.h
@@ -0,0 +1,1137 @@
+//  Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#pragma once
+
+#include <cmath>
+
+#include "port/port.h"  // for PREFETCH
+#include "util/fastrange.h"
+#include "util/ribbon_alg.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+namespace ribbon {
+
+// RIBBON PHSF & RIBBON Filter (Rapid Incremental Boolean Banding ON-the-fly)
+//
+// ribbon_impl.h: templated (parameterized) standard implementations
+//
+// Ribbon is a Perfect Hash Static Function construction useful as a compact
+// static Bloom filter alternative. See ribbon_alg.h for core algorithms
+// and core design details.
+//
+// TODO: more details on trade-offs and practical issues.
+//
+// APIs for configuring Ribbon are in ribbon_config.h
+
+// Ribbon implementations in this file take these parameters, which must be
+// provided in a class/struct type with members expressed in this concept:
+
+// concept TypesAndSettings {
+//   // See RibbonTypes and *Hasher in ribbon_alg.h, except here we have
+//   // the added constraint that Hash be equivalent to either uint32_t or
+//   // uint64_t.
+//   typename Hash;
+//   typename CoeffRow;
+//   typename ResultRow;
+//   typename Index;
+//   typename Key;
+//   static constexpr bool kFirstCoeffAlwaysOne;
+//
+//   // An unsigned integer type for identifying a hash seed, typically
+//   // uint32_t or uint64_t. Importantly, this is the amount of data
+//   // stored in memory for identifying a raw seed. See StandardHasher.
+//   typename Seed;
+//
+//   // When true, the PHSF implements a static filter, expecting just
+//   // keys as inputs for construction. When false, implements a general
+//   // PHSF and expects std::pair<Key, ResultRow> as inputs for
+//   // construction.
+//   static constexpr bool kIsFilter;
+//
+//   // When true, enables a special "homogeneous" filter implementation that
+//   // is slightly faster to construct, and never fails to construct though
+//   // FP rate can quickly explode in cases where corresponding
+//   // non-homogeneous filter would fail (or nearly fail?) to construct.
+//   // For smaller filters, you can configure with ConstructionFailureChance
+//   // smaller than desired FP rate to largely counteract this effect.
+//   // TODO: configuring Homogeneous Ribbon for arbitrarily large filters
+//   // based on data from OptimizeHomogAtScale
+//   static constexpr bool kHomogeneous;
+//
+//   // When true, adds a tiny bit more hashing logic on queries and
+//   // construction to improve utilization at the beginning and end of
+//   // the structure.  Recommended when CoeffRow is only 64 bits (or
+//   // less), so typical num_starts < 10k. Although this is compatible
+//   // with kHomogeneous, the competing space vs. time priorities might
+//   // not be useful.
+//   static constexpr bool kUseSmash;
+//
+//   // When true, allows number of "starts" to be zero, for best support
+//   // of the "no keys to add" case by always returning false for filter
+//   // queries. (This is distinct from the "keys added but no space for
+//   // any data" case, in which a filter always returns true.) The cost
+//   // supporting this is a conditional branch (probably predictable) in
+//   // queries.
+//   static constexpr bool kAllowZeroStarts;
+//
+//   // A seedable stock hash function on Keys. All bits of Hash must
+//   // be reasonably high quality. XXH functions recommended, but
+//   // Murmur, City, Farm, etc. also work.
+//   static Hash HashFn(const Key &, Seed raw_seed);
+// };
+
+// A bit of a hack to automatically construct the type for
+// AddInput based on a constexpr bool.
+template <typename Key, typename ResultRow, bool IsFilter>
+struct AddInputSelector {
+  // For general PHSF, not filter
+  using T = std::pair<Key, ResultRow>;
+};
+
+template <typename Key, typename ResultRow>
+struct AddInputSelector<Key, ResultRow, true /*IsFilter*/> {
+  // For Filter
+  using T = Key;
+};
+
+// To avoid writing 'typename' everywhere that we use types like 'Index'
+#define IMPORT_RIBBON_TYPES_AND_SETTINGS(TypesAndSettings)                   \
+  using CoeffRow = typename TypesAndSettings::CoeffRow;                      \
+  using ResultRow = typename TypesAndSettings::ResultRow;                    \
+  using Index = typename TypesAndSettings::Index;                            \
+  using Hash = typename TypesAndSettings::Hash;                              \
+  using Key = typename TypesAndSettings::Key;                                \
+  using Seed = typename TypesAndSettings::Seed;                              \
+                                                                             \
+  /* Some more additions */                                                  \
+  using QueryInput = Key;                                                    \
+  using AddInput = typename ROCKSDB_NAMESPACE::ribbon::AddInputSelector<     \
+      Key, ResultRow, TypesAndSettings::kIsFilter>::T;                       \
+  static constexpr auto kCoeffBits =                                         \
+      static_cast<Index>(sizeof(CoeffRow) * 8U);                             \
+                                                                             \
+  /* Export to algorithm */                                                  \
+  static constexpr bool kFirstCoeffAlwaysOne =                               \
+      TypesAndSettings::kFirstCoeffAlwaysOne;                                \
+                                                                             \
+  static_assert(sizeof(CoeffRow) + sizeof(ResultRow) + sizeof(Index) +       \
+                        sizeof(Hash) + sizeof(Key) + sizeof(Seed) +          \
+                        sizeof(QueryInput) + sizeof(AddInput) + kCoeffBits + \
+                        kFirstCoeffAlwaysOne >                               \
+                    0,                                                       \
+                "avoid unused warnings, semicolon expected after macro call")
+
+#ifdef _MSC_VER
+#pragma warning(push)
+#pragma warning(disable : 4309)  // cast truncating constant
+#pragma warning(disable : 4307)  // arithmetic constant overflow
+#endif
+
+// StandardHasher: A standard implementation of concepts RibbonTypes,
+// PhsfQueryHasher, FilterQueryHasher, and BandingHasher from ribbon_alg.h.
+//
+// This implementation should be suitable for most all practical purposes
+// as it "behaves" across a wide range of settings, with little room left
+// for improvement. The key functionality in this hasher is generating
+// CoeffRows, starts, and (for filters) ResultRows, which could be ~150
+// bits of data or more, from a modest hash of 64 or even just 32 bits, with
+// enough uniformity and bitwise independence to be close to "the best you
+// can do" with available hash information in terms of FP rate and
+// compactness. (64 bits recommended and sufficient for PHSF practical
+// purposes.)
+//
+// Another feature of this hasher is a minimal "premixing" of seeds before
+// they are provided to TypesAndSettings::HashFn in case that function does
+// not provide sufficiently independent hashes when iterating merely
+// sequentially on seeds. (This for example works around a problem with the
+// preview version 0.7.2 of XXH3 used in RocksDB, a.k.a. XXPH3 or Hash64, and
+// MurmurHash1 used in RocksDB, a.k.a. Hash.) We say this pre-mixing step
+// translates "ordinal seeds," which we iterate sequentially to find a
+// solution, into "raw seeds," with many more bits changing for each
+// iteration. The translation is an easily reversible lightweight mixing,
+// not suitable for hashing on its own. An advantage of this approach is that
+// StandardHasher can store just the raw seed (e.g. 64 bits) for fast query
+// times, while from the application perspective, we can limit to a small
+// number of ordinal keys (e.g. 64 in 6 bits) for saving in metadata.
+//
+// The default constructor initializes the seed to ordinal seed zero, which
+// is equal to raw seed zero.
+//
+template <class TypesAndSettings>
+class StandardHasher {
+ public:
+  IMPORT_RIBBON_TYPES_AND_SETTINGS(TypesAndSettings);
+
+  inline Hash GetHash(const Key& key) const {
+    return TypesAndSettings::HashFn(key, raw_seed_);
+  };
+  // For when AddInput == pair<Key, ResultRow> (kIsFilter == false)
+  inline Hash GetHash(const std::pair<Key, ResultRow>& bi) const {
+    return GetHash(bi.first);
+  };
+  inline Index GetStart(Hash h, Index num_starts) const {
+    // This is "critical path" code because it's required before memory
+    // lookup.
+    //
+    // FastRange gives us a fast and effective mapping from h to the
+    // appropriate range. This depends most, sometimes exclusively, on
+    // upper bits of h.
+    //
+    if (TypesAndSettings::kUseSmash) {
+      // Extra logic to "smash" entries at beginning and end, for
+      // better utilization. For example, without smash and with
+      // kFirstCoeffAlwaysOne, there's about a 30% chance that the
+      // first slot in the banding will be unused, and worse without
+      // kFirstCoeffAlwaysOne. The ending slots are even less utilized
+      // without smash.
+      //
+      // But since this only affects roughly kCoeffBits of the slots,
+      // it's usually small enough to be ignorable (less computation in
+      // this function) when number of slots is roughly 10k or larger.
+      //
+      // The best values for these smash weights might depend on how
+      // densely you're packing entries, and also kCoeffBits, but this
+      // seems to work well for roughly 95% success probability.
+      //
+      constexpr Index kFrontSmash = kCoeffBits / 4;
+      constexpr Index kBackSmash = kCoeffBits / 4;
+      Index start = FastRangeGeneric(h, num_starts + kFrontSmash + kBackSmash);
+      start = std::max(start, kFrontSmash);
+      start -= kFrontSmash;
+      start = std::min(start, num_starts - 1);
+      return start;
+    } else {
+      // For query speed, we allow small number of initial and final
+      // entries to be under-utilized.
+      // NOTE: This call statically enforces that Hash is equivalent to
+      // either uint32_t or uint64_t.
+      return FastRangeGeneric(h, num_starts);
+    }
+  }
+  inline CoeffRow GetCoeffRow(Hash h) const {
+    // This is not so much "critical path" code because it can be done in
+    // parallel (instruction level) with memory lookup.
+    //
+    // When we might have many entries squeezed into a single start,
+    // we need reasonably good remixing for CoeffRow.
+    if (TypesAndSettings::kUseSmash) {
+      // Reasonably good, reasonably fast, reasonably general.
+      // Probably not 1:1 but probably close enough.
+      Unsigned128 a = Multiply64to128(h, kAltCoeffFactor1);
+      Unsigned128 b = Multiply64to128(h, kAltCoeffFactor2);
+      auto cr = static_cast<CoeffRow>(b ^ (a << 64) ^ (a >> 64));
+
+      // Now ensure the value is non-zero
+      if (kFirstCoeffAlwaysOne) {
+        cr |= 1;
+      } else {
+        // Still have to ensure some bit is non-zero
+        cr |= (cr == 0) ? 1 : 0;
+      }
+      return cr;
+    }
+    // If not kUseSmash, we ensure we're not squeezing many entries into a
+    // single start, in part by ensuring num_starts > num_slots / 2. Thus,
+    // here we do not need good remixing for CoeffRow, but just enough that
+    // (a) every bit is reasonably independent from Start.
+    // (b) every Hash-length bit subsequence of the CoeffRow has full or
+    // nearly full entropy from h.
+    // (c) if nontrivial bit subsequences within are correlated, it needs to
+    // be more complicated than exact copy or bitwise not (at least without
+    // kFirstCoeffAlwaysOne), or else there seems to be a kind of
+    // correlated clustering effect.
+    // (d) the CoeffRow is not zero, so that no one input on its own can
+    // doom construction success. (Preferably a mix of 1's and 0's if
+    // satisfying above.)
+
+    // First, establish sufficient bitwise independence from Start, with
+    // multiplication by a large random prime.
+    // Note that we cast to Hash because if we use product bits beyond
+    // original input size, that's going to correlate with Start (FastRange)
+    // even with a (likely) different multiplier here.
+    Hash a = h * kCoeffAndResultFactor;
+
+    static_assert(
+        sizeof(Hash) == sizeof(uint64_t) || sizeof(Hash) == sizeof(uint32_t),
+        "Supported sizes");
+    // If that's big enough, we're done. If not, we have to expand it,
+    // maybe up to 4x size.
+    uint64_t b;
+    if (sizeof(Hash) < sizeof(uint64_t)) {
+      // Almost-trivial hash expansion (OK - see above), favoring roughly
+      // equal number of 1's and 0's in result
+      b = (uint64_t{a} << 32) ^ (a ^ kCoeffXor32);
+    } else {
+      b = a;
+    }
+    static_assert(sizeof(CoeffRow) <= sizeof(Unsigned128), "Supported sizes");
+    Unsigned128 c;
+    if (sizeof(uint64_t) < sizeof(CoeffRow)) {
+      // Almost-trivial hash expansion (OK - see above), favoring roughly
+      // equal number of 1's and 0's in result
+      c = (Unsigned128{b} << 64) ^ (b ^ kCoeffXor64);
+    } else {
+      c = b;
+    }
+    auto cr = static_cast<CoeffRow>(c);
+
+    // Now ensure the value is non-zero
+    if (kFirstCoeffAlwaysOne) {
+      cr |= 1;
+    } else if (sizeof(CoeffRow) == sizeof(Hash)) {
+      // Still have to ensure some bit is non-zero
+      cr |= (cr == 0) ? 1 : 0;
+    } else {
+      // (We did trivial expansion with constant xor, which ensures some
+      // bits are non-zero.)
+    }
+    return cr;
+  }
+  inline ResultRow GetResultRowMask() const {
+    // TODO: will be used with InterleavedSolutionStorage?
+    // For now, all bits set (note: might be a small type so might need to
+    // narrow after promotion)
+    return static_cast<ResultRow>(~ResultRow{0});
+  }
+  inline ResultRow GetResultRowFromHash(Hash h) const {
+    if (TypesAndSettings::kIsFilter && !TypesAndSettings::kHomogeneous) {
+      // This is not so much "critical path" code because it can be done in
+      // parallel (instruction level) with memory lookup.
+      //
+      // ResultRow bits only needs to be independent from CoeffRow bits if
+      // many entries might have the same start location, where "many" is
+      // comparable to number of hash bits or kCoeffBits. If !kUseSmash
+      // and num_starts > kCoeffBits, it is safe and efficient to draw from
+      // the same bits computed for CoeffRow, which are reasonably
+      // independent from Start. (Inlining and common subexpression
+      // elimination with GetCoeffRow should make this
+      // a single shared multiplication in generated code when !kUseSmash.)
+      Hash a = h * kCoeffAndResultFactor;
+
+      // The bits here that are *most* independent of Start are the highest
+      // order bits (as in Knuth multiplicative hash). To make those the
+      // most preferred for use in the result row, we do a bswap here.
+      auto rr = static_cast<ResultRow>(EndianSwapValue(a));
+      return rr & GetResultRowMask();
+    } else {
+      // Must be zero
+      return 0;
+    }
+  }
+  // For when AddInput == Key (kIsFilter == true)
+  inline ResultRow GetResultRowFromInput(const Key&) const {
+    // Must be zero
+    return 0;
+  }
+  // For when AddInput == pair<Key, ResultRow> (kIsFilter == false)
+  inline ResultRow GetResultRowFromInput(
+      const std::pair<Key, ResultRow>& bi) const {
+    // Simple extraction
+    return bi.second;
+  }
+
+  // Seed tracking APIs - see class comment
+  void SetRawSeed(Seed seed) { raw_seed_ = seed; }
+  Seed GetRawSeed() { return raw_seed_; }
+  void SetOrdinalSeed(Seed count) {
+    // A simple, reversible mixing of any size (whole bytes) up to 64 bits.
+    // This allows casting the raw seed to any smaller size we use for
+    // ordinal seeds without risk of duplicate raw seeds for unique ordinal
+    // seeds.
+
+    // Seed type might be smaller than numerical promotion size, but Hash
+    // should be at least that size, so we use Hash as intermediate type.
+    static_assert(sizeof(Seed) <= sizeof(Hash),
+                  "Hash must be at least size of Seed");
+
+    // Multiply by a large random prime (one-to-one for any prefix of bits)
+    Hash tmp = count * kToRawSeedFactor;
+    // Within-byte one-to-one mixing
+    static_assert((kSeedMixMask & (kSeedMixMask >> kSeedMixShift)) == 0,
+                  "Illegal mask+shift");
+    tmp ^= (tmp & kSeedMixMask) >> kSeedMixShift;
+    raw_seed_ = static_cast<Seed>(tmp);
+    // dynamic verification
+    assert(GetOrdinalSeed() == count);
+  }
+  Seed GetOrdinalSeed() {
+    Hash tmp = raw_seed_;
+    // Within-byte one-to-one mixing (its own inverse)
+    tmp ^= (tmp & kSeedMixMask) >> kSeedMixShift;
+    // Multiply by 64-bit multiplicative inverse
+    static_assert(kToRawSeedFactor * kFromRawSeedFactor == Hash{1},
+                  "Must be inverses");
+    return static_cast<Seed>(tmp * kFromRawSeedFactor);
+  }
+
+ protected:
+  // For expanding hash:
+  // large random prime
+  static constexpr Hash kCoeffAndResultFactor =
+      static_cast<Hash>(0xc28f82822b650bedULL);
+  static constexpr uint64_t kAltCoeffFactor1 = 0x876f170be4f1fcb9U;
+  static constexpr uint64_t kAltCoeffFactor2 = 0xf0433a4aecda4c5fU;
+  // random-ish data
+  static constexpr uint32_t kCoeffXor32 = 0xa6293635U;
+  static constexpr uint64_t kCoeffXor64 = 0xc367844a6e52731dU;
+
+  // For pre-mixing seeds
+  static constexpr Hash kSeedMixMask = static_cast<Hash>(0xf0f0f0f0f0f0f0f0ULL);
+  static constexpr unsigned kSeedMixShift = 4U;
+  static constexpr Hash kToRawSeedFactor =
+      static_cast<Hash>(0xc78219a23eeadd03ULL);
+  static constexpr Hash kFromRawSeedFactor =
+      static_cast<Hash>(0xfe1a137d14b475abULL);
+
+  // See class description
+  Seed raw_seed_ = 0;
+};
+
+// StandardRehasher (and StandardRehasherAdapter): A variant of
+// StandardHasher that uses the same type for keys as for hashes.
+// This is primarily intended for building a Ribbon filter
+// from existing hashes without going back to original inputs in
+// order to apply a different seed. This hasher seeds a 1-to-1 mixing
+// transformation to apply a seed to an existing hash. (Untested for
+// hash-sized keys that are not already uniformly distributed.) This
+// transformation builds on the seed pre-mixing done in StandardHasher.
+//
+// Testing suggests essentially no degradation of solution success rate
+// vs. going back to original inputs when changing hash seeds. For example:
+// Average re-seeds for solution with r=128, 1.02x overhead, and ~100k keys
+// is about 1.10 for both StandardHasher and StandardRehasher.
+//
+// StandardRehasher is not really recommended for general PHSFs (not
+// filters) because a collision in the original hash could prevent
+// construction despite re-seeding the Rehasher. (Such collisions
+// do not interfere with filter construction.)
+//
+// concept RehasherTypesAndSettings: like TypesAndSettings but
+// does not require Key or HashFn.
+template <class RehasherTypesAndSettings>
+class StandardRehasherAdapter : public RehasherTypesAndSettings {
+ public:
+  using Hash = typename RehasherTypesAndSettings::Hash;
+  using Key = Hash;
+  using Seed = typename RehasherTypesAndSettings::Seed;
+
+  static Hash HashFn(const Hash& input, Seed raw_seed) {
+    // Note: raw_seed is already lightly pre-mixed, and this multiplication
+    // by a large prime is sufficient mixing (low-to-high bits) on top of
+    // that for good FastRange results, which depends primarily on highest
+    // bits. (The hashed CoeffRow and ResultRow are less sensitive to
+    // mixing than Start.)
+    // Also note: did consider adding ^ (input >> some) before the
+    // multiplication, but doesn't appear to be necessary.
+    return (input ^ raw_seed) * kRehashFactor;
+  }
+
+ private:
+  static constexpr Hash kRehashFactor =
+      static_cast<Hash>(0x6193d459236a3a0dULL);
+};
+
+// See comment on StandardRehasherAdapter
+template <class RehasherTypesAndSettings>
+using StandardRehasher =
+    StandardHasher<StandardRehasherAdapter<RehasherTypesAndSettings>>;
+
+#ifdef _MSC_VER
+#pragma warning(pop)
+#endif
+
+// Especially with smaller hashes (e.g. 32 bit), there can be noticeable
+// false positives due to collisions in the Hash returned by GetHash.
+// This function returns the expected FP rate due to those collisions,
+// which can be added to the expected FP rate from the underlying data
+// structure. (Note: technically, a + b is only a good approximation of
+// 1-(1-a)(1-b) == a + b - a*b, if a and b are much closer to 0 than to 1.)
+// The number of entries added can be a double here in case it's an
+// average.
+template <class Hasher, typename Numerical>
+double ExpectedCollisionFpRate(const Hasher& hasher, Numerical added) {
+  // Standardize on the 'double' specialization
+  return ExpectedCollisionFpRate(hasher, 1.0 * added);
+}
+template <class Hasher>
+double ExpectedCollisionFpRate(const Hasher& /*hasher*/, double added) {
+  // Technically, there could be overlap among the added, but ignoring that
+  // is typically close enough.
+  return added / std::pow(256.0, sizeof(typename Hasher::Hash));
+}
+
+// StandardBanding: a canonical implementation of BandingStorage and
+// BacktrackStorage, with convenience API for banding (solving with on-the-fly
+// Gaussian elimination) with and without backtracking.
+template <class TypesAndSettings>
+class StandardBanding : public StandardHasher<TypesAndSettings> {
+ public:
+  IMPORT_RIBBON_TYPES_AND_SETTINGS(TypesAndSettings);
+
+  StandardBanding(Index num_slots = 0, Index backtrack_size = 0) {
+    Reset(num_slots, backtrack_size);
+  }
+
+  void Reset(Index num_slots, Index backtrack_size = 0) {
+    if (num_slots == 0) {
+      // Unusual (TypesAndSettings::kAllowZeroStarts) or "uninitialized"
+      num_starts_ = 0;
+    } else {
+      // Normal
+      assert(num_slots >= kCoeffBits);
+      if (num_slots > num_slots_allocated_) {
+        coeff_rows_.reset(new CoeffRow[num_slots]());
+        if (!TypesAndSettings::kHomogeneous) {
+          // Note: don't strictly have to zero-init result_rows,
+          // except possible information leakage, etc ;)
+          result_rows_.reset(new ResultRow[num_slots]());
+        }
+        num_slots_allocated_ = num_slots;
+      } else {
+        for (Index i = 0; i < num_slots; ++i) {
+          coeff_rows_[i] = 0;
+          if (!TypesAndSettings::kHomogeneous) {
+            // Note: don't strictly have to zero-init result_rows,
+            // except possible information leakage, etc ;)
+            result_rows_[i] = 0;
+          }
+        }
+      }
+      num_starts_ = num_slots - kCoeffBits + 1;
+    }
+    EnsureBacktrackSize(backtrack_size);
+  }
+
+  void EnsureBacktrackSize(Index backtrack_size) {
+    if (backtrack_size > backtrack_size_) {
+      backtrack_.reset(new Index[backtrack_size]);
+      backtrack_size_ = backtrack_size;
+    }
+  }
+
+  // ********************************************************************
+  // From concept BandingStorage
+
+  inline bool UsePrefetch() const {
+    // A rough guesstimate of when prefetching during construction pays off.
+    // TODO: verify/validate
+    return num_starts_ > 1500;
+  }
+  inline void Prefetch(Index i) const {
+    PREFETCH(&coeff_rows_[i], 1 /* rw */, 1 /* locality */);
+    if (!TypesAndSettings::kHomogeneous) {
+      PREFETCH(&result_rows_[i], 1 /* rw */, 1 /* locality */);
+    }
+  }
+  inline void LoadRow(Index i, CoeffRow* cr, ResultRow* rr,
+                      bool for_back_subst) const {
+    *cr = coeff_rows_[i];
+    if (TypesAndSettings::kHomogeneous) {
+      if (for_back_subst && *cr == 0) {
+        // Cheap pseudorandom data to fill unconstrained solution rows
+        *rr = static_cast<ResultRow>(i * 0x9E3779B185EBCA87ULL);
+      } else {
+        *rr = 0;
+      }
+    } else {
+      *rr = result_rows_[i];
+    }
+  }
+  inline void StoreRow(Index i, CoeffRow cr, ResultRow rr) {
+    coeff_rows_[i] = cr;
+    if (TypesAndSettings::kHomogeneous) {
+      assert(rr == 0);
+    } else {
+      result_rows_[i] = rr;
+    }
+  }
+  inline Index GetNumStarts() const { return num_starts_; }
+
+  // from concept BacktrackStorage, for when backtracking is used
+  inline bool UseBacktrack() const { return true; }
+  inline void BacktrackPut(Index i, Index to_save) { backtrack_[i] = to_save; }
+  inline Index BacktrackGet(Index i) const { return backtrack_[i]; }
+
+  // ********************************************************************
+  // Some useful API, still somewhat low level. Here an input is
+  // a Key for filters, or std::pair<Key, ResultRow> for general PHSF.
+
+  // Adds a range of inputs to the banding, returning true if successful.
+  // False means none or some may have been successfully added, so it's
+  // best to Reset this banding before any further use.
+  //
+  // Adding can fail even before all the "slots" are completely "full".
+  //
+  template <typename InputIterator>
+  bool AddRange(InputIterator begin, InputIterator end) {
+    assert(num_starts_ > 0 || TypesAndSettings::kAllowZeroStarts);
+    if (TypesAndSettings::kAllowZeroStarts && num_starts_ == 0) {
+      // Unusual. Can't add any in this case.
+      return begin == end;
+    }
+    // Normal
+    return BandingAddRange(this, *this, begin, end);
+  }
+
+  // Adds a range of inputs to the banding, returning true if successful,
+  // or if unsuccessful, rolls back to state before this call and returns
+  // false. Caller guarantees that the number of inputs in this batch
+  // does not exceed `backtrack_size` provided to Reset.
+  //
+  // Adding can fail even before all the "slots" are completely "full".
+  //
+  template <typename InputIterator>
+  bool AddRangeOrRollBack(InputIterator begin, InputIterator end) {
+    assert(num_starts_ > 0 || TypesAndSettings::kAllowZeroStarts);
+    if (TypesAndSettings::kAllowZeroStarts && num_starts_ == 0) {
+      // Unusual. Can't add any in this case.
+      return begin == end;
+    }
+    // else Normal
+    return BandingAddRange(this, this, *this, begin, end);
+  }
+
+  // Adds a single input to the banding, returning true if successful.
+  // If unsuccessful, returns false and banding state is unchanged.
+  //
+  // Adding can fail even before all the "slots" are completely "full".
+  //
+  bool Add(const AddInput& input) {
+    // Pointer can act as iterator
+    return AddRange(&input, &input + 1);
+  }
+
+  // Return the number of "occupied" rows (with non-zero coefficients stored).
+  Index GetOccupiedCount() const {
+    Index count = 0;
+    if (num_starts_ > 0) {
+      const Index num_slots = num_starts_ + kCoeffBits - 1;
+      for (Index i = 0; i < num_slots; ++i) {
+        if (coeff_rows_[i] != 0) {
+          ++count;
+        }
+      }
+    }
+    return count;
+  }
+
+  // Returns whether a row is "occupied" in the banding (non-zero
+  // coefficients stored). (Only recommended for debug/test)
+  bool IsOccupied(Index i) { return coeff_rows_[i] != 0; }
+
+  // ********************************************************************
+  // High-level API
+
+  // Iteratively (a) resets the structure for `num_slots`, (b) attempts
+  // to add the range of inputs, and (c) if unsuccessful, chooses next
+  // hash seed, until either successful or unsuccessful with all the
+  // allowed seeds. Returns true if successful. In that case, use
+  // GetOrdinalSeed() or GetRawSeed() to get the successful seed.
+  //
+  // The allowed sequence of hash seeds is determined by
+  // `starting_ordinal_seed,` the first ordinal seed to be attempted
+  // (see StandardHasher), and `ordinal_seed_mask,` a bit mask (power of
+  // two minus one) for the range of ordinal seeds to consider. The
+  // max number of seeds considered will be ordinal_seed_mask + 1.
+  // For filters we suggest `starting_ordinal_seed` be chosen randomly
+  // or round-robin, to minimize false positive correlations between keys.
+  //
+  // If unsuccessful, how best to continue is going to be application
+  // specific. It should be possible to choose parameters such that
+  // failure is extremely unlikely, using max_seed around 32 to 64.
+  // (TODO: APIs to help choose parameters) One option for fallback in
+  // constructing a filter is to construct a Bloom filter instead.
+  // Increasing num_slots is an option, but should not be used often
+  // unless construction maximum latency is a concern (rather than
+  // average running time of construction). Instead, choose parameters
+  // appropriately and trust that seeds are independent. (Also,
+  // increasing num_slots without changing hash seed would have a
+  // significant correlation in success, rather than independence.)
+  template <typename InputIterator>
+  bool ResetAndFindSeedToSolve(Index num_slots, InputIterator begin,
+                               InputIterator end,
+                               Seed starting_ordinal_seed = 0U,
+                               Seed ordinal_seed_mask = 63U) {
+    // power of 2 minus 1
+    assert((ordinal_seed_mask & (ordinal_seed_mask + 1)) == 0);
+    // starting seed is within mask
+    assert((starting_ordinal_seed & ordinal_seed_mask) ==
+           starting_ordinal_seed);
+    starting_ordinal_seed &= ordinal_seed_mask;  // if not debug
+
+    Seed cur_ordinal_seed = starting_ordinal_seed;
+    do {
+      StandardHasher<TypesAndSettings>::SetOrdinalSeed(cur_ordinal_seed);
+      Reset(num_slots);
+      bool success = AddRange(begin, end);
+      if (success) {
+        return true;
+      }
+      cur_ordinal_seed = (cur_ordinal_seed + 1) & ordinal_seed_mask;
+    } while (cur_ordinal_seed != starting_ordinal_seed);
+    // Reached limit by circling around
+    return false;
+  }
+
+  static std::size_t EstimateMemoryUsage(uint32_t num_slots) {
+    std::size_t bytes_coeff_rows = num_slots * sizeof(CoeffRow);
+    std::size_t bytes_result_rows = num_slots * sizeof(ResultRow);
+    std::size_t bytes_backtrack = 0;
+    std::size_t bytes_banding =
+        bytes_coeff_rows + bytes_result_rows + bytes_backtrack;
+
+    return bytes_banding;
+  }
+
+ protected:
+  // TODO: explore combining in a struct
+  std::unique_ptr<CoeffRow[]> coeff_rows_;
+  std::unique_ptr<ResultRow[]> result_rows_;
+  // We generally store "starts" instead of slots for speed of GetStart(),
+  // as in StandardHasher.
+  Index num_starts_ = 0;
+  Index num_slots_allocated_ = 0;
+  std::unique_ptr<Index[]> backtrack_;
+  Index backtrack_size_ = 0;
+};
+
+// Implements concept SimpleSolutionStorage, mostly for demonstration
+// purposes. This is "in memory" only because it does not handle byte
+// ordering issues for serialization.
+template <class TypesAndSettings>
+class InMemSimpleSolution {
+ public:
+  IMPORT_RIBBON_TYPES_AND_SETTINGS(TypesAndSettings);
+
+  void PrepareForNumStarts(Index num_starts) {
+    if (TypesAndSettings::kAllowZeroStarts && num_starts == 0) {
+      // Unusual
+      num_starts_ = 0;
+    } else {
+      // Normal
+      const Index num_slots = num_starts + kCoeffBits - 1;
+      assert(num_slots >= kCoeffBits);
+      if (num_slots > num_slots_allocated_) {
+        // Do not need to init the memory
+        solution_rows_.reset(new ResultRow[num_slots]);
+        num_slots_allocated_ = num_slots;
+      }
+      num_starts_ = num_starts;
+    }
+  }
+
+  Index GetNumStarts() const { return num_starts_; }
+
+  ResultRow Load(Index slot_num) const { return solution_rows_[slot_num]; }
+
+  void Store(Index slot_num, ResultRow solution_row) {
+    solution_rows_[slot_num] = solution_row;
+  }
+
+  // ********************************************************************
+  // High-level API
+
+  template <typename BandingStorage>
+  void BackSubstFrom(const BandingStorage& bs) {
+    if (TypesAndSettings::kAllowZeroStarts && bs.GetNumStarts() == 0) {
+      // Unusual
+      PrepareForNumStarts(0);
+    } else {
+      // Normal
+      SimpleBackSubst(this, bs);
+    }
+  }
+
+  template <typename PhsfQueryHasher>
+  ResultRow PhsfQuery(const Key& input, const PhsfQueryHasher& hasher) const {
+    // assert(!TypesAndSettings::kIsFilter);  Can be useful in testing
+    if (TypesAndSettings::kAllowZeroStarts && num_starts_ == 0) {
+      // Unusual
+      return 0;
+    } else {
+      // Normal
+      return SimplePhsfQuery(input, hasher, *this);
+    }
+  }
+
+  template <typename FilterQueryHasher>
+  bool FilterQuery(const Key& input, const FilterQueryHasher& hasher) const {
+    assert(TypesAndSettings::kIsFilter);
+    if (TypesAndSettings::kAllowZeroStarts && num_starts_ == 0) {
+      // Unusual. Zero starts presumes no keys added -> always false
+      return false;
+    } else {
+      // Normal, or upper_num_columns_ == 0 means "no space for data" and
+      // thus will always return true.
+      return SimpleFilterQuery(input, hasher, *this);
+    }
+  }
+
+  double ExpectedFpRate() const {
+    assert(TypesAndSettings::kIsFilter);
+    if (TypesAndSettings::kAllowZeroStarts && num_starts_ == 0) {
+      // Unusual, but we don't have FPs if we always return false.
+      return 0.0;
+    }
+    // else Normal
+
+    // Each result (solution) bit (column) cuts FP rate in half
+    return std::pow(0.5, 8U * sizeof(ResultRow));
+  }
+
+  // ********************************************************************
+  // Static high-level API
+
+  // Round up to a number of slots supported by this structure. Note that
+  // this needs to be must be taken into account for the banding if this
+  // solution layout/storage is to be used.
+  static Index RoundUpNumSlots(Index num_slots) {
+    // Must be at least kCoeffBits for at least one start
+    // Or if not smash, even more because hashing not equipped
+    // for stacking up so many entries on a single start location
+    auto min_slots = kCoeffBits * (TypesAndSettings::kUseSmash ? 1 : 2);
+    return std::max(num_slots, static_cast<Index>(min_slots));
+  }
+
+ protected:
+  // We generally store "starts" instead of slots for speed of GetStart(),
+  // as in StandardHasher.
+  Index num_starts_ = 0;
+  Index num_slots_allocated_ = 0;
+  std::unique_ptr<ResultRow[]> solution_rows_;
+};
+
+// Implements concept InterleavedSolutionStorage always using little-endian
+// byte order, so easy for serialization/deserialization. This implementation
+// fully supports fractional bits per key, where any number of segments
+// (number of bytes multiple of sizeof(CoeffRow)) can be used with any number
+// of slots that is a multiple of kCoeffBits.
+//
+// The structure is passed an externally allocated/de-allocated byte buffer
+// that is optionally pre-populated (from storage) for answering queries,
+// or can be populated by BackSubstFrom.
+//
+template <class TypesAndSettings>
+class SerializableInterleavedSolution {
+ public:
+  IMPORT_RIBBON_TYPES_AND_SETTINGS(TypesAndSettings);
+
+  // Does not take ownership of `data` but uses it (up to `data_len` bytes)
+  // throughout lifetime
+  SerializableInterleavedSolution(char* data, size_t data_len)
+      : data_(data), data_len_(data_len) {}
+
+  void PrepareForNumStarts(Index num_starts) {
+    assert(num_starts == 0 || (num_starts % kCoeffBits == 1));
+    num_starts_ = num_starts;
+
+    InternalConfigure();
+  }
+
+  Index GetNumStarts() const { return num_starts_; }
+
+  Index GetNumBlocks() const {
+    const Index num_slots = num_starts_ + kCoeffBits - 1;
+    return num_slots / kCoeffBits;
+  }
+
+  Index GetUpperNumColumns() const { return upper_num_columns_; }
+
+  Index GetUpperStartBlock() const { return upper_start_block_; }
+
+  Index GetNumSegments() const {
+    return static_cast<Index>(data_len_ / sizeof(CoeffRow));
+  }
+
+  CoeffRow LoadSegment(Index segment_num) const {
+    assert(data_ != nullptr);  // suppress clang analyzer report
+    return DecodeFixedGeneric<CoeffRow>(data_ + segment_num * sizeof(CoeffRow));
+  }
+  void StoreSegment(Index segment_num, CoeffRow val) {
+    assert(data_ != nullptr);  // suppress clang analyzer report
+    EncodeFixedGeneric(data_ + segment_num * sizeof(CoeffRow), val);
+  }
+  void PrefetchSegmentRange(Index begin_segment_num,
+                            Index end_segment_num) const {
+    if (end_segment_num == begin_segment_num) {
+      // Nothing to do
+      return;
+    }
+    char* cur = data_ + begin_segment_num * sizeof(CoeffRow);
+    char* last = data_ + (end_segment_num - 1) * sizeof(CoeffRow);
+    while (cur < last) {
+      PREFETCH(cur, 0 /* rw */, 1 /* locality */);
+      cur += CACHE_LINE_SIZE;
+    }
+    PREFETCH(last, 0 /* rw */, 1 /* locality */);
+  }
+
+  // ********************************************************************
+  // High-level API
+
+  void ConfigureForNumBlocks(Index num_blocks) {
+    if (num_blocks == 0) {
+      PrepareForNumStarts(0);
+    } else {
+      PrepareForNumStarts(num_blocks * kCoeffBits - kCoeffBits + 1);
+    }
+  }
+
+  void ConfigureForNumSlots(Index num_slots) {
+    assert(num_slots % kCoeffBits == 0);
+    ConfigureForNumBlocks(num_slots / kCoeffBits);
+  }
+
+  template <typename BandingStorage>
+  void BackSubstFrom(const BandingStorage& bs) {
+    if (TypesAndSettings::kAllowZeroStarts && bs.GetNumStarts() == 0) {
+      // Unusual
+      PrepareForNumStarts(0);
+    } else {
+      // Normal
+      InterleavedBackSubst(this, bs);
+    }
+  }
+
+  template <typename PhsfQueryHasher>
+  ResultRow PhsfQuery(const Key& input, const PhsfQueryHasher& hasher) const {
+    // assert(!TypesAndSettings::kIsFilter);  Can be useful in testing
+    if (TypesAndSettings::kAllowZeroStarts && num_starts_ == 0) {
+      // Unusual
+      return 0;
+    } else {
+      // Normal
+      // NOTE: not using a struct to encourage compiler optimization
+      Hash hash;
+      Index segment_num;
+      Index num_columns;
+      Index start_bit;
+      InterleavedPrepareQuery(input, hasher, *this, &hash, &segment_num,
+                              &num_columns, &start_bit);
+      return InterleavedPhsfQuery(hash, segment_num, num_columns, start_bit,
+                                  hasher, *this);
+    }
+  }
+
+  template <typename FilterQueryHasher>
+  bool FilterQuery(const Key& input, const FilterQueryHasher& hasher) const {
+    assert(TypesAndSettings::kIsFilter);
+    if (TypesAndSettings::kAllowZeroStarts && num_starts_ == 0) {
+      // Unusual. Zero starts presumes no keys added -> always false
+      return false;
+    } else {
+      // Normal, or upper_num_columns_ == 0 means "no space for data" and
+      // thus will always return true.
+      // NOTE: not using a struct to encourage compiler optimization
+      Hash hash;
+      Index segment_num;
+      Index num_columns;
+      Index start_bit;
+      InterleavedPrepareQuery(input, hasher, *this, &hash, &segment_num,
+                              &num_columns, &start_bit);
+      return InterleavedFilterQuery(hash, segment_num, num_columns, start_bit,
+                                    hasher, *this);
+    }
+  }
+
+  double ExpectedFpRate() const {
+    assert(TypesAndSettings::kIsFilter);
+    if (TypesAndSettings::kAllowZeroStarts && num_starts_ == 0) {
+      // Unusual. Zero starts presumes no keys added -> always false
+      return 0.0;
+    }
+    // else Normal
+
+    // Note: Ignoring smash setting; still close enough in that case
+    double lower_portion =
+        (upper_start_block_ * 1.0 * kCoeffBits) / num_starts_;
+
+    // Each result (solution) bit (column) cuts FP rate in half. Weight that
+    // for upper and lower number of bits (columns).
+    return lower_portion * std::pow(0.5, upper_num_columns_ - 1) +
+           (1.0 - lower_portion) * std::pow(0.5, upper_num_columns_);
+  }
+
+  // ********************************************************************
+  // Static high-level API
+
+  // Round up to a number of slots supported by this structure. Note that
+  // this needs to be must be taken into account for the banding if this
+  // solution layout/storage is to be used.
+  static Index RoundUpNumSlots(Index num_slots) {
+    // Must be multiple of kCoeffBits
+    Index corrected = (num_slots + kCoeffBits - 1) / kCoeffBits * kCoeffBits;
+
+    // Do not use num_starts==1 unless kUseSmash, because the hashing
+    // might not be equipped for stacking up so many entries on a
+    // single start location.
+    if (!TypesAndSettings::kUseSmash && corrected == kCoeffBits) {
+      corrected += kCoeffBits;
+    }
+    return corrected;
+  }
+
+  // Round down to a number of slots supported by this structure. Note that
+  // this needs to be must be taken into account for the banding if this
+  // solution layout/storage is to be used.
+  static Index RoundDownNumSlots(Index num_slots) {
+    // Must be multiple of kCoeffBits
+    Index corrected = num_slots / kCoeffBits * kCoeffBits;
+
+    // Do not use num_starts==1 unless kUseSmash, because the hashing
+    // might not be equipped for stacking up so many entries on a
+    // single start location.
+    if (!TypesAndSettings::kUseSmash && corrected == kCoeffBits) {
+      corrected = 0;
+    }
+    return corrected;
+  }
+
+  // Compute the number of bytes for a given number of slots and desired
+  // FP rate. Since desired FP rate might not be exactly achievable,
+  // rounding_bias32==0 means to always round toward lower FP rate
+  // than desired (more bytes); rounding_bias32==max uint32_t means always
+  // round toward higher FP rate than desired (fewer bytes); other values
+  // act as a proportional threshold or bias between the two.
+  static size_t GetBytesForFpRate(Index num_slots, double desired_fp_rate,
+                                  uint32_t rounding_bias32) {
+    return InternalGetBytesForFpRate(num_slots, desired_fp_rate,
+                                     1.0 / desired_fp_rate, rounding_bias32);
+  }
+
+  // The same, but specifying desired accuracy as 1.0 / FP rate, or
+  // one_in_fp_rate. E.g. desired_one_in_fp_rate=100 means 1% FP rate.
+  static size_t GetBytesForOneInFpRate(Index num_slots,
+                                       double desired_one_in_fp_rate,
+                                       uint32_t rounding_bias32) {
+    return InternalGetBytesForFpRate(num_slots, 1.0 / desired_one_in_fp_rate,
+                                     desired_one_in_fp_rate, rounding_bias32);
+  }
+
+ protected:
+  static size_t InternalGetBytesForFpRate(Index num_slots,
+                                          double desired_fp_rate,
+                                          double desired_one_in_fp_rate,
+                                          uint32_t rounding_bias32) {
+    assert(TypesAndSettings::kIsFilter);
+    if (TypesAndSettings::kAllowZeroStarts) {
+      if (num_slots == 0) {
+        // Unusual. Zero starts presumes no keys added -> always false (no FPs)
+        return 0U;
+      }
+    } else {
+      assert(num_slots > 0);
+    }
+    // Must be rounded up already.
+    assert(RoundUpNumSlots(num_slots) == num_slots);
+
+    if (desired_one_in_fp_rate > 1.0 && desired_fp_rate < 1.0) {
+      // Typical: less than 100% FP rate
+      if (desired_one_in_fp_rate <= static_cast<ResultRow>(-1)) {
+        // Typical: Less than maximum result row entropy
+        ResultRow rounded = static_cast<ResultRow>(desired_one_in_fp_rate);
+        int lower_columns = FloorLog2(rounded);
+        double lower_columns_fp_rate = std::pow(2.0, -lower_columns);
+        double upper_columns_fp_rate = std::pow(2.0, -(lower_columns + 1));
+        // Floating point don't let me down!
+        assert(lower_columns_fp_rate >= desired_fp_rate);
+        assert(upper_columns_fp_rate <= desired_fp_rate);
+
+        double lower_portion = (desired_fp_rate - upper_columns_fp_rate) /
+                               (lower_columns_fp_rate - upper_columns_fp_rate);
+        // Floating point don't let me down!
+        assert(lower_portion >= 0.0);
+        assert(lower_portion <= 1.0);
+
+        double rounding_bias = (rounding_bias32 + 0.5) / double{0x100000000};
+        assert(rounding_bias > 0.0);
+        assert(rounding_bias < 1.0);
+
+        // Note: Ignoring smash setting; still close enough in that case
+        Index num_starts = num_slots - kCoeffBits + 1;
+        // Lower upper_start_block means lower FP rate (higher accuracy)
+        Index upper_start_block = static_cast<Index>(
+            (lower_portion * num_starts + rounding_bias) / kCoeffBits);
+        Index num_blocks = num_slots / kCoeffBits;
+        assert(upper_start_block < num_blocks);
+
+        // Start by assuming all blocks use lower number of columns
+        Index num_segments = num_blocks * static_cast<Index>(lower_columns);
+        // Correct by 1 each for blocks using upper number of columns
+        num_segments += (num_blocks - upper_start_block);
+        // Total bytes
+        return num_segments * sizeof(CoeffRow);
+      } else {
+        // one_in_fp_rate too big, thus requested FP rate is smaller than
+        // supported. Use max number of columns for minimum supported FP rate.
+        return num_slots * sizeof(ResultRow);
+      }
+    } else {
+      // Effectively asking for 100% FP rate, or NaN etc.
+      if (TypesAndSettings::kAllowZeroStarts) {
+        // Zero segments
+        return 0U;
+      } else {
+        // One segment (minimum size, maximizing FP rate)
+        return sizeof(CoeffRow);
+      }
+    }
+  }
+
+  void InternalConfigure() {
+    const Index num_blocks = GetNumBlocks();
+    Index num_segments = GetNumSegments();
+
+    if (num_blocks == 0) {
+      // Exceptional
+      upper_num_columns_ = 0;
+      upper_start_block_ = 0;
+    } else {
+      // Normal
+      upper_num_columns_ =
+          (num_segments + /*round up*/ num_blocks - 1) / num_blocks;
+      upper_start_block_ = upper_num_columns_ * num_blocks - num_segments;
+      // Unless that's more columns than supported by ResultRow data type
+      if (upper_num_columns_ > 8U * sizeof(ResultRow)) {
+        // Use maximum columns (there will be space unused)
+        upper_num_columns_ = static_cast<Index>(8U * sizeof(ResultRow));
+        upper_start_block_ = 0;
+        num_segments = num_blocks * upper_num_columns_;
+      }
+    }
+    // Update data_len_ for correct rounding and/or unused space
+    // NOTE: unused space stays gone if we PrepareForNumStarts again.
+    // We are prioritizing minimizing the number of fields over making
+    // the "unusued space" feature work well.
+    data_len_ = num_segments * sizeof(CoeffRow);
+  }
+
+  char* const data_;
+  size_t data_len_;
+  Index num_starts_ = 0;
+  Index upper_num_columns_ = 0;
+  Index upper_start_block_ = 0;
+};
+
+}  // namespace ribbon
+
+}  // namespace ROCKSDB_NAMESPACE
+
+// For convenience working with templates
+#define IMPORT_RIBBON_IMPL_TYPES(TypesAndSettings)                            \
+  using Hasher = ROCKSDB_NAMESPACE::ribbon::StandardHasher<TypesAndSettings>; \
+  using Banding =                                                             \
+      ROCKSDB_NAMESPACE::ribbon::StandardBanding<TypesAndSettings>;           \
+  using SimpleSoln =                                                          \
+      ROCKSDB_NAMESPACE::ribbon::InMemSimpleSolution<TypesAndSettings>;       \
+  using InterleavedSoln =                                                     \
+      ROCKSDB_NAMESPACE::ribbon::SerializableInterleavedSolution<             \
+          TypesAndSettings>;                                                  \
+  static_assert(sizeof(Hasher) + sizeof(Banding) + sizeof(SimpleSoln) +       \
+                        sizeof(InterleavedSoln) >                             \
+                    0,                                                        \
+                "avoid unused warnings, semicolon expected after macro call")
diff --git a/src/rocksdb/util/ribbon_test.cc b/src/rocksdb/util/ribbon_test.cc
new file mode 100644
index 000000000..6519df3d5
--- /dev/null
+++ b/src/rocksdb/util/ribbon_test.cc
@@ -0,0 +1,1308 @@
+//  Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#include "rocksdb/system_clock.h"
+#include "test_util/testharness.h"
+#include "util/bloom_impl.h"
+#include "util/coding.h"
+#include "util/hash.h"
+#include "util/ribbon_config.h"
+#include "util/ribbon_impl.h"
+#include "util/stop_watch.h"
+#include "util/string_util.h"
+
+#ifndef GFLAGS
+uint32_t FLAGS_thoroughness = 5;
+uint32_t FLAGS_max_add = 0;
+uint32_t FLAGS_min_check = 4000;
+uint32_t FLAGS_max_check = 100000;
+bool FLAGS_verbose = false;
+
+bool FLAGS_find_occ = false;
+bool FLAGS_find_slot_occ = false;
+double FLAGS_find_next_factor = 1.618;
+uint32_t FLAGS_find_iters = 10000;
+uint32_t FLAGS_find_min_slots = 128;
+uint32_t FLAGS_find_max_slots = 1000000;
+
+bool FLAGS_optimize_homog = false;
+uint32_t FLAGS_optimize_homog_slots = 30000000;
+uint32_t FLAGS_optimize_homog_check = 200000;
+double FLAGS_optimize_homog_granularity = 0.002;
+#else
+#include "util/gflags_compat.h"
+using GFLAGS_NAMESPACE::ParseCommandLineFlags;
+// Using 500 is a good test when you have time to be thorough.
+// Default is for general RocksDB regression test runs.
+DEFINE_uint32(thoroughness, 5, "iterations per configuration");
+DEFINE_uint32(max_add, 0,
+              "Add up to this number of entries to a single filter in "
+              "CompactnessAndBacktrackAndFpRate; 0 == reasonable default");
+DEFINE_uint32(min_check, 4000,
+              "Minimum number of novel entries for testing FP rate");
+DEFINE_uint32(max_check, 10000,
+              "Maximum number of novel entries for testing FP rate");
+DEFINE_bool(verbose, false, "Print extra details");
+
+// Options for FindOccupancy, which is more of a tool than a test.
+DEFINE_bool(find_occ, false, "whether to run the FindOccupancy tool");
+DEFINE_bool(find_slot_occ, false,
+            "whether to show individual slot occupancies with "
+            "FindOccupancy tool");
+DEFINE_double(find_next_factor, 1.618,
+              "factor to next num_slots for FindOccupancy");
+DEFINE_uint32(find_iters, 10000, "number of samples for FindOccupancy");
+DEFINE_uint32(find_min_slots, 128, "number of slots for FindOccupancy");
+DEFINE_uint32(find_max_slots, 1000000, "number of slots for FindOccupancy");
+
+// Options for OptimizeHomogAtScale, which is more of a tool than a test.
+DEFINE_bool(optimize_homog, false,
+            "whether to run the OptimizeHomogAtScale tool");
+DEFINE_uint32(optimize_homog_slots, 30000000,
+              "number of slots for OptimizeHomogAtScale");
+DEFINE_uint32(optimize_homog_check, 200000,
+              "number of queries for checking FP rate in OptimizeHomogAtScale");
+DEFINE_double(
+    optimize_homog_granularity, 0.002,
+    "overhead change between FP rate checking in OptimizeHomogAtScale");
+
+#endif  // GFLAGS
+
+template <typename TypesAndSettings>
+class RibbonTypeParamTest : public ::testing::Test {};
+
+class RibbonTest : public ::testing::Test {};
+
+namespace {
+
+// Different ways of generating keys for testing
+
+// Generate semi-sequential keys
+struct StandardKeyGen {
+  StandardKeyGen(const std::string& prefix, uint64_t id)
+      : id_(id), str_(prefix) {
+    ROCKSDB_NAMESPACE::PutFixed64(&str_, /*placeholder*/ 0);
+  }
+
+  // Prefix (only one required)
+  StandardKeyGen& operator++() {
+    ++id_;
+    return *this;
+  }
+
+  StandardKeyGen& operator+=(uint64_t i) {
+    id_ += i;
+    return *this;
+  }
+
+  const std::string& operator*() {
+    // Use multiplication to mix things up a little in the key
+    ROCKSDB_NAMESPACE::EncodeFixed64(&str_[str_.size() - 8],
+                                     id_ * uint64_t{0x1500000001});
+    return str_;
+  }
+
+  bool operator==(const StandardKeyGen& other) {
+    // Same prefix is assumed
+    return id_ == other.id_;
+  }
+  bool operator!=(const StandardKeyGen& other) {
+    // Same prefix is assumed
+    return id_ != other.id_;
+  }
+
+  uint64_t id_;
+  std::string str_;
+};
+
+// Generate small sequential keys, that can misbehave with sequential seeds
+// as in https://github.com/Cyan4973/xxHash/issues/469.
+// These keys are only heuristically unique, but that's OK with 64 bits,
+// for testing purposes.
+struct SmallKeyGen {
+  SmallKeyGen(const std::string& prefix, uint64_t id) : id_(id) {
+    // Hash the prefix for a heuristically unique offset
+    id_ += ROCKSDB_NAMESPACE::GetSliceHash64(prefix);
+    ROCKSDB_NAMESPACE::PutFixed64(&str_, id_);
+  }
+
+  // Prefix (only one required)
+  SmallKeyGen& operator++() {
+    ++id_;
+    return *this;
+  }
+
+  SmallKeyGen& operator+=(uint64_t i) {
+    id_ += i;
+    return *this;
+  }
+
+  const std::string& operator*() {
+    ROCKSDB_NAMESPACE::EncodeFixed64(&str_[str_.size() - 8], id_);
+    return str_;
+  }
+
+  bool operator==(const SmallKeyGen& other) { return id_ == other.id_; }
+  bool operator!=(const SmallKeyGen& other) { return id_ != other.id_; }
+
+  uint64_t id_;
+  std::string str_;
+};
+
+template <typename KeyGen>
+struct Hash32KeyGenWrapper : public KeyGen {
+  Hash32KeyGenWrapper(const std::string& prefix, uint64_t id)
+      : KeyGen(prefix, id) {}
+  uint32_t operator*() {
+    auto& key = *static_cast<KeyGen&>(*this);
+    // unseeded
+    return ROCKSDB_NAMESPACE::GetSliceHash(key);
+  }
+};
+
+template <typename KeyGen>
+struct Hash64KeyGenWrapper : public KeyGen {
+  Hash64KeyGenWrapper(const std::string& prefix, uint64_t id)
+      : KeyGen(prefix, id) {}
+  uint64_t operator*() {
+    auto& key = *static_cast<KeyGen&>(*this);
+    // unseeded
+    return ROCKSDB_NAMESPACE::GetSliceHash64(key);
+  }
+};
+
+using ROCKSDB_NAMESPACE::ribbon::ConstructionFailureChance;
+
+const std::vector<ConstructionFailureChance> kFailureOnly50Pct = {
+    ROCKSDB_NAMESPACE::ribbon::kOneIn2};
+
+const std::vector<ConstructionFailureChance> kFailureOnlyRare = {
+    ROCKSDB_NAMESPACE::ribbon::kOneIn1000};
+
+const std::vector<ConstructionFailureChance> kFailureAll = {
+    ROCKSDB_NAMESPACE::ribbon::kOneIn2, ROCKSDB_NAMESPACE::ribbon::kOneIn20,
+    ROCKSDB_NAMESPACE::ribbon::kOneIn1000};
+
+}  // namespace
+
+using ROCKSDB_NAMESPACE::ribbon::ExpectedCollisionFpRate;
+using ROCKSDB_NAMESPACE::ribbon::StandardHasher;
+using ROCKSDB_NAMESPACE::ribbon::StandardRehasherAdapter;
+
+struct DefaultTypesAndSettings {
+  using CoeffRow = ROCKSDB_NAMESPACE::Unsigned128;
+  using ResultRow = uint8_t;
+  using Index = uint32_t;
+  using Hash = uint64_t;
+  using Seed = uint32_t;
+  using Key = ROCKSDB_NAMESPACE::Slice;
+  static constexpr bool kIsFilter = true;
+  static constexpr bool kHomogeneous = false;
+  static constexpr bool kFirstCoeffAlwaysOne = true;
+  static constexpr bool kUseSmash = false;
+  static constexpr bool kAllowZeroStarts = false;
+  static Hash HashFn(const Key& key, uint64_t raw_seed) {
+    // This version 0.7.2 preview of XXH3 (a.k.a. XXPH3) function does
+    // not pass SmallKeyGen tests below without some seed premixing from
+    // StandardHasher. See https://github.com/Cyan4973/xxHash/issues/469
+    return ROCKSDB_NAMESPACE::Hash64(key.data(), key.size(), raw_seed);
+  }
+  // For testing
+  using KeyGen = StandardKeyGen;
+  static const std::vector<ConstructionFailureChance>& FailureChanceToTest() {
+    return kFailureAll;
+  }
+};
+
+using TypesAndSettings_Coeff128 = DefaultTypesAndSettings;
+struct TypesAndSettings_Coeff128Smash : public DefaultTypesAndSettings {
+  static constexpr bool kUseSmash = true;
+};
+struct TypesAndSettings_Coeff64 : public DefaultTypesAndSettings {
+  using CoeffRow = uint64_t;
+};
+struct TypesAndSettings_Coeff64Smash : public TypesAndSettings_Coeff64 {
+  static constexpr bool kUseSmash = true;
+};
+struct TypesAndSettings_Coeff64Smash0 : public TypesAndSettings_Coeff64Smash {
+  static constexpr bool kFirstCoeffAlwaysOne = false;
+};
+
+// Homogeneous Ribbon configurations
+struct TypesAndSettings_Coeff128_Homog : public DefaultTypesAndSettings {
+  static constexpr bool kHomogeneous = true;
+  // Since our best construction success setting still has 1/1000 failure
+  // rate, the best FP rate we test is 1/256
+  using ResultRow = uint8_t;
+  // Homogeneous only makes sense with sufficient slots for equivalent of
+  // almost sure construction success
+  static const std::vector<ConstructionFailureChance>& FailureChanceToTest() {
+    return kFailureOnlyRare;
+  }
+};
+struct TypesAndSettings_Coeff128Smash_Homog
+    : public TypesAndSettings_Coeff128_Homog {
+  // Smash (extra time to save space) + Homog (extra space to save time)
+  // doesn't make much sense in practice, but we minimally test it
+  static constexpr bool kUseSmash = true;
+};
+struct TypesAndSettings_Coeff64_Homog : public TypesAndSettings_Coeff128_Homog {
+  using CoeffRow = uint64_t;
+};
+struct TypesAndSettings_Coeff64Smash_Homog
+    : public TypesAndSettings_Coeff64_Homog {
+  // Smash (extra time to save space) + Homog (extra space to save time)
+  // doesn't make much sense in practice, but we minimally test it
+  static constexpr bool kUseSmash = true;
+};
+
+// Less exhaustive mix of coverage, but still covering the most stressful case
+// (only 50% construction success)
+struct AbridgedTypesAndSettings : public DefaultTypesAndSettings {
+  static const std::vector<ConstructionFailureChance>& FailureChanceToTest() {
+    return kFailureOnly50Pct;
+  }
+};
+struct TypesAndSettings_Result16 : public AbridgedTypesAndSettings {
+  using ResultRow = uint16_t;
+};
+struct TypesAndSettings_Result32 : public AbridgedTypesAndSettings {
+  using ResultRow = uint32_t;
+};
+struct TypesAndSettings_IndexSizeT : public AbridgedTypesAndSettings {
+  using Index = size_t;
+};
+struct TypesAndSettings_Hash32 : public AbridgedTypesAndSettings {
+  using Hash = uint32_t;
+  static Hash HashFn(const Key& key, Hash raw_seed) {
+    // This MurmurHash1 function does not pass tests below without the
+    // seed premixing from StandardHasher. In fact, it needs more than
+    // just a multiplication mixer on the ordinal seed.
+    return ROCKSDB_NAMESPACE::Hash(key.data(), key.size(), raw_seed);
+  }
+};
+struct TypesAndSettings_Hash32_Result16 : public AbridgedTypesAndSettings {
+  using ResultRow = uint16_t;
+};
+struct TypesAndSettings_KeyString : public AbridgedTypesAndSettings {
+  using Key = std::string;
+};
+struct TypesAndSettings_Seed8 : public AbridgedTypesAndSettings {
+  // This is not a generally recommended configuration. With the configured
+  // hash function, it would fail with SmallKeyGen due to insufficient
+  // independence among the seeds.
+  using Seed = uint8_t;
+};
+struct TypesAndSettings_NoAlwaysOne : public AbridgedTypesAndSettings {
+  static constexpr bool kFirstCoeffAlwaysOne = false;
+};
+struct TypesAndSettings_AllowZeroStarts : public AbridgedTypesAndSettings {
+  static constexpr bool kAllowZeroStarts = true;
+};
+struct TypesAndSettings_Seed64 : public AbridgedTypesAndSettings {
+  using Seed = uint64_t;
+};
+struct TypesAndSettings_Rehasher
+    : public StandardRehasherAdapter<AbridgedTypesAndSettings> {
+  using KeyGen = Hash64KeyGenWrapper<StandardKeyGen>;
+};
+struct TypesAndSettings_Rehasher_Result16 : public TypesAndSettings_Rehasher {
+  using ResultRow = uint16_t;
+};
+struct TypesAndSettings_Rehasher_Result32 : public TypesAndSettings_Rehasher {
+  using ResultRow = uint32_t;
+};
+struct TypesAndSettings_Rehasher_Seed64
+    : public StandardRehasherAdapter<TypesAndSettings_Seed64> {
+  using KeyGen = Hash64KeyGenWrapper<StandardKeyGen>;
+  // Note: 64-bit seed with Rehasher gives slightly better average reseeds
+};
+struct TypesAndSettings_Rehasher32
+    : public StandardRehasherAdapter<TypesAndSettings_Hash32> {
+  using KeyGen = Hash32KeyGenWrapper<StandardKeyGen>;
+};
+struct TypesAndSettings_Rehasher32_Coeff64
+    : public TypesAndSettings_Rehasher32 {
+  using CoeffRow = uint64_t;
+};
+struct TypesAndSettings_SmallKeyGen : public AbridgedTypesAndSettings {
+  // SmallKeyGen stresses the independence of different hash seeds
+  using KeyGen = SmallKeyGen;
+};
+struct TypesAndSettings_Hash32_SmallKeyGen : public TypesAndSettings_Hash32 {
+  // SmallKeyGen stresses the independence of different hash seeds
+  using KeyGen = SmallKeyGen;
+};
+struct TypesAndSettings_Coeff32 : public DefaultTypesAndSettings {
+  using CoeffRow = uint32_t;
+};
+struct TypesAndSettings_Coeff32Smash : public TypesAndSettings_Coeff32 {
+  static constexpr bool kUseSmash = true;
+};
+struct TypesAndSettings_Coeff16 : public DefaultTypesAndSettings {
+  using CoeffRow = uint16_t;
+};
+struct TypesAndSettings_Coeff16Smash : public TypesAndSettings_Coeff16 {
+  static constexpr bool kUseSmash = true;
+};
+
+using TestTypesAndSettings = ::testing::Types<
+    TypesAndSettings_Coeff128, TypesAndSettings_Coeff128Smash,
+    TypesAndSettings_Coeff64, TypesAndSettings_Coeff64Smash,
+    TypesAndSettings_Coeff64Smash0, TypesAndSettings_Coeff128_Homog,
+    TypesAndSettings_Coeff128Smash_Homog, TypesAndSettings_Coeff64_Homog,
+    TypesAndSettings_Coeff64Smash_Homog, TypesAndSettings_Result16,
+    TypesAndSettings_Result32, TypesAndSettings_IndexSizeT,
+    TypesAndSettings_Hash32, TypesAndSettings_Hash32_Result16,
+    TypesAndSettings_KeyString, TypesAndSettings_Seed8,
+    TypesAndSettings_NoAlwaysOne, TypesAndSettings_AllowZeroStarts,
+    TypesAndSettings_Seed64, TypesAndSettings_Rehasher,
+    TypesAndSettings_Rehasher_Result16, TypesAndSettings_Rehasher_Result32,
+    TypesAndSettings_Rehasher_Seed64, TypesAndSettings_Rehasher32,
+    TypesAndSettings_Rehasher32_Coeff64, TypesAndSettings_SmallKeyGen,
+    TypesAndSettings_Hash32_SmallKeyGen, TypesAndSettings_Coeff32,
+    TypesAndSettings_Coeff32Smash, TypesAndSettings_Coeff16,
+    TypesAndSettings_Coeff16Smash>;
+TYPED_TEST_CASE(RibbonTypeParamTest, TestTypesAndSettings);
+
+namespace {
+
+// For testing Poisson-distributed (or similar) statistics, get value for
+// `stddevs_allowed` standard deviations above expected mean
+// `expected_count`.
+// (Poisson approximates Binomial only if probability of a trial being
+// in the count is low.)
+uint64_t PoissonUpperBound(double expected_count, double stddevs_allowed) {
+  return static_cast<uint64_t>(
+      expected_count + stddevs_allowed * std::sqrt(expected_count) + 1.0);
+}
+
+uint64_t PoissonLowerBound(double expected_count, double stddevs_allowed) {
+  return static_cast<uint64_t>(std::max(
+      0.0, expected_count - stddevs_allowed * std::sqrt(expected_count)));
+}
+
+uint64_t FrequentPoissonUpperBound(double expected_count) {
+  // Allow up to 5.0 standard deviations for frequently checked statistics
+  return PoissonUpperBound(expected_count, 5.0);
+}
+
+uint64_t FrequentPoissonLowerBound(double expected_count) {
+  return PoissonLowerBound(expected_count, 5.0);
+}
+
+uint64_t InfrequentPoissonUpperBound(double expected_count) {
+  // Allow up to 3 standard deviations for infrequently checked statistics
+  return PoissonUpperBound(expected_count, 3.0);
+}
+
+uint64_t InfrequentPoissonLowerBound(double expected_count) {
+  return PoissonLowerBound(expected_count, 3.0);
+}
+
+}  // namespace
+
+TYPED_TEST(RibbonTypeParamTest, CompactnessAndBacktrackAndFpRate) {
+  IMPORT_RIBBON_TYPES_AND_SETTINGS(TypeParam);
+  IMPORT_RIBBON_IMPL_TYPES(TypeParam);
+  using KeyGen = typename TypeParam::KeyGen;
+  using ConfigHelper =
+      ROCKSDB_NAMESPACE::ribbon::BandingConfigHelper<TypeParam>;
+
+  if (sizeof(CoeffRow) < 8) {
+    ROCKSDB_GTEST_BYPASS("Not fully supported");
+    return;
+  }
+
+  const auto log2_thoroughness =
+      static_cast<uint32_t>(ROCKSDB_NAMESPACE::FloorLog2(FLAGS_thoroughness));
+
+  // We are going to choose num_to_add using an exponential distribution,
+  // so that we have good representation of small-to-medium filters.
+  // Here we just pick some reasonable, practical upper bound based on
+  // kCoeffBits or option.
+  const double log_max_add = std::log(
+      FLAGS_max_add > 0 ? FLAGS_max_add
+                        : static_cast<uint32_t>(kCoeffBits * kCoeffBits) *
+                              std::max(FLAGS_thoroughness, uint32_t{32}));
+
+  // This needs to be enough below the minimum number of slots to get a
+  // reasonable number of samples with the minimum number of slots.
+  const double log_min_add = std::log(0.66 * SimpleSoln::RoundUpNumSlots(1));
+
+  ASSERT_GT(log_max_add, log_min_add);
+
+  const double diff_log_add = log_max_add - log_min_add;
+
+  for (ConstructionFailureChance cs : TypeParam::FailureChanceToTest()) {
+    double expected_reseeds;
+    switch (cs) {
+      default:
+        assert(false);
+        FALLTHROUGH_INTENDED;
+      case ROCKSDB_NAMESPACE::ribbon::kOneIn2:
+        fprintf(stderr, "== Failure: 50 percent\n");
+        expected_reseeds = 1.0;
+        break;
+      case ROCKSDB_NAMESPACE::ribbon::kOneIn20:
+        fprintf(stderr, "== Failure: 95 percent\n");
+        expected_reseeds = 0.053;
+        break;
+      case ROCKSDB_NAMESPACE::ribbon::kOneIn1000:
+        fprintf(stderr, "== Failure: 1/1000\n");
+        expected_reseeds = 0.001;
+        break;
+    }
+
+    uint64_t total_reseeds = 0;
+    uint64_t total_singles = 0;
+    uint64_t total_single_failures = 0;
+    uint64_t total_batch = 0;
+    uint64_t total_batch_successes = 0;
+    uint64_t total_fp_count = 0;
+    uint64_t total_added = 0;
+    uint64_t total_expand_trials = 0;
+    uint64_t total_expand_failures = 0;
+    double total_expand_overhead = 0.0;
+
+    uint64_t soln_query_nanos = 0;
+    uint64_t soln_query_count = 0;
+    uint64_t bloom_query_nanos = 0;
+    uint64_t isoln_query_nanos = 0;
+    uint64_t isoln_query_count = 0;
+
+    // Take different samples if you change thoroughness
+    ROCKSDB_NAMESPACE::Random32 rnd(FLAGS_thoroughness);
+
+    for (uint32_t i = 0; i < FLAGS_thoroughness; ++i) {
+      // We are going to choose num_to_add using an exponential distribution
+      // as noted above, but instead of randomly choosing them, we generate
+      // samples linearly using the golden ratio, which ensures a nice spread
+      // even for a small number of samples, and starting with the minimum
+      // number of slots to ensure it is tested.
+      double log_add =
+          std::fmod(0.6180339887498948482 * diff_log_add * i, diff_log_add) +
+          log_min_add;
+      uint32_t num_to_add = static_cast<uint32_t>(std::exp(log_add));
+
+      // Most of the time, test the Interleaved solution storage, but when
+      // we do we have to make num_slots a multiple of kCoeffBits. So
+      // sometimes we want to test without that limitation.
+      bool test_interleaved = (i % 7) != 6;
+
+      // Compute num_slots, and re-adjust num_to_add to get as close as possible
+      // to next num_slots, to stress that num_slots in terms of construction
+      // success. Ensure at least one iteration:
+      Index num_slots = Index{0} - 1;
+      --num_to_add;
+      for (;;) {
+        Index next_num_slots = SimpleSoln::RoundUpNumSlots(
+            ConfigHelper::GetNumSlots(num_to_add + 1, cs));
+        if (test_interleaved) {
+          next_num_slots = InterleavedSoln::RoundUpNumSlots(next_num_slots);
+          // assert idempotent
+          EXPECT_EQ(next_num_slots,
+                    InterleavedSoln::RoundUpNumSlots(next_num_slots));
+        }
+        // assert idempotent with InterleavedSoln::RoundUpNumSlots
+        EXPECT_EQ(next_num_slots, SimpleSoln::RoundUpNumSlots(next_num_slots));
+
+        if (next_num_slots > num_slots) {
+          break;
+        }
+        num_slots = next_num_slots;
+        ++num_to_add;
+      }
+      assert(num_slots < Index{0} - 1);
+
+      total_added += num_to_add;
+
+      std::string prefix;
+      ROCKSDB_NAMESPACE::PutFixed32(&prefix, rnd.Next());
+
+      // Batch that must be added
+      std::string added_str = prefix + "added";
+      KeyGen keys_begin(added_str, 0);
+      KeyGen keys_end(added_str, num_to_add);
+
+      // A couple more that will probably be added
+      KeyGen one_more(prefix + "more", 1);
+      KeyGen two_more(prefix + "more", 2);
+
+      // Batch that may or may not be added
+      uint32_t batch_size =
+          static_cast<uint32_t>(2.0 * std::sqrt(num_slots - num_to_add));
+      if (batch_size < 10U) {
+        batch_size = 0;
+      }
+      std::string batch_str = prefix + "batch";
+      KeyGen batch_begin(batch_str, 0);
+      KeyGen batch_end(batch_str, batch_size);
+
+      // Batch never (successfully) added, but used for querying FP rate
+      std::string not_str = prefix + "not";
+      KeyGen other_keys_begin(not_str, 0);
+      KeyGen other_keys_end(not_str, FLAGS_max_check);
+
+      double overhead_ratio = 1.0 * num_slots / num_to_add;
+      if (FLAGS_verbose) {
+        fprintf(stderr, "Adding(%s) %u / %u   Overhead: %g   Batch size: %u\n",
+                test_interleaved ? "i" : "s", (unsigned)num_to_add,
+                (unsigned)num_slots, overhead_ratio, (unsigned)batch_size);
+      }
+
+      // Vary bytes for InterleavedSoln to use number of solution columns
+      // from 0 to max allowed by ResultRow type (and used by SimpleSoln).
+      // Specifically include 0 and max, and otherwise skew toward max.
+      uint32_t max_ibytes =
+          static_cast<uint32_t>(sizeof(ResultRow) * num_slots);
+      size_t ibytes;
+      if (i == 0) {
+        ibytes = 0;
+      } else if (i == 1) {
+        ibytes = max_ibytes;
+      } else {
+        // Skewed
+        ibytes =
+            std::max(rnd.Uniformish(max_ibytes), rnd.Uniformish(max_ibytes));
+      }
+      std::unique_ptr<char[]> idata(new char[ibytes]);
+      InterleavedSoln isoln(idata.get(), ibytes);
+
+      SimpleSoln soln;
+      Hasher hasher;
+      bool first_single;
+      bool second_single;
+      bool batch_success;
+      {
+        Banding banding;
+        // Traditional solve for a fixed set.
+        ASSERT_TRUE(
+            banding.ResetAndFindSeedToSolve(num_slots, keys_begin, keys_end));
+
+        Index occupied_count = banding.GetOccupiedCount();
+        Index more_added = 0;
+
+        if (TypeParam::kHomogeneous || overhead_ratio < 1.01 ||
+            batch_size == 0) {
+          // Homogeneous not compatible with backtracking because add
+          // doesn't fail. Small overhead ratio too packed to expect more
+          first_single = false;
+          second_single = false;
+          batch_success = false;
+        } else {
+          // Now to test backtracking, starting with guaranteed fail. By using
+          // the keys that will be used to test FP rate, we are then doing an
+          // extra check that after backtracking there are no remnants (e.g. in
+          // result side of banding) of these entries.
+          KeyGen other_keys_too_big_end = other_keys_begin;
+          other_keys_too_big_end += num_to_add;
+          banding.EnsureBacktrackSize(std::max(num_to_add, batch_size));
+          EXPECT_FALSE(banding.AddRangeOrRollBack(other_keys_begin,
+                                                  other_keys_too_big_end));
+          EXPECT_EQ(occupied_count, banding.GetOccupiedCount());
+
+          // Check that we still have a good chance of adding a couple more
+          // individually
+          first_single = banding.Add(*one_more);
+          second_single = banding.Add(*two_more);
+          more_added += (first_single ? 1 : 0) + (second_single ? 1 : 0);
+          total_singles += 2U;
+          total_single_failures += 2U - more_added;
+
+          // Or as a batch
+          batch_success = banding.AddRangeOrRollBack(batch_begin, batch_end);
+          ++total_batch;
+          if (batch_success) {
+            more_added += batch_size;
+            ++total_batch_successes;
+          }
+          EXPECT_LE(banding.GetOccupiedCount(), occupied_count + more_added);
+        }
+
+        // Also verify that redundant adds are OK (no effect)
+        ASSERT_TRUE(
+            banding.AddRange(keys_begin, KeyGen(added_str, num_to_add / 8)));
+        EXPECT_LE(banding.GetOccupiedCount(), occupied_count + more_added);
+
+        // Now back-substitution
+        soln.BackSubstFrom(banding);
+        if (test_interleaved) {
+          isoln.BackSubstFrom(banding);
+        }
+
+        Seed reseeds = banding.GetOrdinalSeed();
+        total_reseeds += reseeds;
+
+        EXPECT_LE(reseeds, 8 + log2_thoroughness);
+        if (reseeds > log2_thoroughness + 1) {
+          fprintf(
+              stderr, "%s high reseeds at %u, %u/%u: %u\n",
+              reseeds > log2_thoroughness + 8 ? "ERROR Extremely" : "Somewhat",
+              static_cast<unsigned>(i), static_cast<unsigned>(num_to_add),
+              static_cast<unsigned>(num_slots), static_cast<unsigned>(reseeds));
+        }
+
+        if (reseeds > 0) {
+          // "Expand" test: given a failed construction, how likely is it to
+          // pass with same seed and more slots. At each step, we increase
+          // enough to ensure there is at least one shift within each coeff
+          // block.
+          ++total_expand_trials;
+          Index expand_count = 0;
+          Index ex_slots = num_slots;
+          banding.SetOrdinalSeed(0);
+          for (;; ++expand_count) {
+            ASSERT_LE(expand_count, log2_thoroughness);
+            ex_slots += ex_slots / kCoeffBits;
+            if (test_interleaved) {
+              ex_slots = InterleavedSoln::RoundUpNumSlots(ex_slots);
+            }
+            banding.Reset(ex_slots);
+            bool success = banding.AddRange(keys_begin, keys_end);
+            if (success) {
+              break;
+            }
+          }
+          total_expand_failures += expand_count;
+          total_expand_overhead += 1.0 * (ex_slots - num_slots) / num_slots;
+        }
+
+        hasher.SetOrdinalSeed(reseeds);
+      }
+      // soln and hasher now independent of Banding object
+
+      // Verify keys added
+      KeyGen cur = keys_begin;
+      while (cur != keys_end) {
+        ASSERT_TRUE(soln.FilterQuery(*cur, hasher));
+        ASSERT_TRUE(!test_interleaved || isoln.FilterQuery(*cur, hasher));
+        ++cur;
+      }
+      // We (maybe) snuck these in!
+      if (first_single) {
+        ASSERT_TRUE(soln.FilterQuery(*one_more, hasher));
+        ASSERT_TRUE(!test_interleaved || isoln.FilterQuery(*one_more, hasher));
+      }
+      if (second_single) {
+        ASSERT_TRUE(soln.FilterQuery(*two_more, hasher));
+        ASSERT_TRUE(!test_interleaved || isoln.FilterQuery(*two_more, hasher));
+      }
+      if (batch_success) {
+        cur = batch_begin;
+        while (cur != batch_end) {
+          ASSERT_TRUE(soln.FilterQuery(*cur, hasher));
+          ASSERT_TRUE(!test_interleaved || isoln.FilterQuery(*cur, hasher));
+          ++cur;
+        }
+      }
+
+      // Check FP rate (depends only on number of result bits == solution
+      // columns)
+      Index fp_count = 0;
+      cur = other_keys_begin;
+      {
+        ROCKSDB_NAMESPACE::StopWatchNano timer(
+            ROCKSDB_NAMESPACE::SystemClock::Default().get(), true);
+        while (cur != other_keys_end) {
+          bool fp = soln.FilterQuery(*cur, hasher);
+          fp_count += fp ? 1 : 0;
+          ++cur;
+        }
+        soln_query_nanos += timer.ElapsedNanos();
+        soln_query_count += FLAGS_max_check;
+      }
+      {
+        double expected_fp_count = soln.ExpectedFpRate() * FLAGS_max_check;
+        // For expected FP rate, also include false positives due to collisions
+        // in Hash value. (Negligible for 64-bit, can matter for 32-bit.)
+        double correction =
+            FLAGS_max_check * ExpectedCollisionFpRate(hasher, num_to_add);
+
+        // NOTE: rare violations expected with kHomogeneous
+        EXPECT_LE(fp_count,
+                  FrequentPoissonUpperBound(expected_fp_count + correction));
+        EXPECT_GE(fp_count,
+                  FrequentPoissonLowerBound(expected_fp_count + correction));
+      }
+      total_fp_count += fp_count;
+
+      // And also check FP rate for isoln
+      if (test_interleaved) {
+        Index ifp_count = 0;
+        cur = other_keys_begin;
+        ROCKSDB_NAMESPACE::StopWatchNano timer(
+            ROCKSDB_NAMESPACE::SystemClock::Default().get(), true);
+        while (cur != other_keys_end) {
+          ifp_count += isoln.FilterQuery(*cur, hasher) ? 1 : 0;
+          ++cur;
+        }
+        isoln_query_nanos += timer.ElapsedNanos();
+        isoln_query_count += FLAGS_max_check;
+        {
+          double expected_fp_count = isoln.ExpectedFpRate() * FLAGS_max_check;
+          // For expected FP rate, also include false positives due to
+          // collisions in Hash value. (Negligible for 64-bit, can matter for
+          // 32-bit.)
+          double correction =
+              FLAGS_max_check * ExpectedCollisionFpRate(hasher, num_to_add);
+
+          // NOTE: rare violations expected with kHomogeneous
+          EXPECT_LE(ifp_count,
+                    FrequentPoissonUpperBound(expected_fp_count + correction));
+
+          // FIXME: why sometimes can we slightly "beat the odds"?
+          // (0.95 factor should not be needed)
+          EXPECT_GE(ifp_count, FrequentPoissonLowerBound(
+                                   0.95 * expected_fp_count + correction));
+        }
+        // Since the bits used in isoln are a subset of the bits used in soln,
+        // it cannot have fewer FPs
+        EXPECT_GE(ifp_count, fp_count);
+      }
+
+      // And compare to Bloom time, for fun
+      if (ibytes >= /* minimum Bloom impl bytes*/ 64) {
+        Index bfp_count = 0;
+        cur = other_keys_begin;
+        ROCKSDB_NAMESPACE::StopWatchNano timer(
+            ROCKSDB_NAMESPACE::SystemClock::Default().get(), true);
+        while (cur != other_keys_end) {
+          uint64_t h = hasher.GetHash(*cur);
+          uint32_t h1 = ROCKSDB_NAMESPACE::Lower32of64(h);
+          uint32_t h2 = sizeof(Hash) >= 8 ? ROCKSDB_NAMESPACE::Upper32of64(h)
+                                          : h1 * 0x9e3779b9;
+          bfp_count +=
+              ROCKSDB_NAMESPACE::FastLocalBloomImpl::HashMayMatch(
+                  h1, h2, static_cast<uint32_t>(ibytes), 6, idata.get())
+                  ? 1
+                  : 0;
+          ++cur;
+        }
+        bloom_query_nanos += timer.ElapsedNanos();
+        // ensure bfp_count is used
+        ASSERT_LT(bfp_count, FLAGS_max_check);
+      }
+    }
+
+    // "outside" == key not in original set so either negative or false positive
+    fprintf(stderr,
+            "Simple      outside query, hot, incl hashing, ns/key: %g\n",
+            1.0 * soln_query_nanos / soln_query_count);
+    fprintf(stderr,
+            "Interleaved outside query, hot, incl hashing, ns/key: %g\n",
+            1.0 * isoln_query_nanos / isoln_query_count);
+    fprintf(stderr,
+            "Bloom       outside query, hot, incl hashing, ns/key: %g\n",
+            1.0 * bloom_query_nanos / soln_query_count);
+
+    if (TypeParam::kHomogeneous) {
+      EXPECT_EQ(total_reseeds, 0U);
+    } else {
+      double average_reseeds = 1.0 * total_reseeds / FLAGS_thoroughness;
+      fprintf(stderr, "Average re-seeds: %g\n", average_reseeds);
+      // Values above were chosen to target around 50% chance of encoding
+      // success rate (average of 1.0 re-seeds) or slightly better. But 1.15 is
+      // also close enough.
+      EXPECT_LE(total_reseeds,
+                InfrequentPoissonUpperBound(1.15 * expected_reseeds *
+                                            FLAGS_thoroughness));
+      // Would use 0.85 here instead of 0.75, but
+      // TypesAndSettings_Hash32_SmallKeyGen can "beat the odds" because of
+      // sequential keys with a small, cheap hash function. We accept that
+      // there are surely inputs that are somewhat bad for this setup, but
+      // these somewhat good inputs are probably more likely.
+      EXPECT_GE(total_reseeds,
+                InfrequentPoissonLowerBound(0.75 * expected_reseeds *
+                                            FLAGS_thoroughness));
+    }
+
+    if (total_expand_trials > 0) {
+      double average_expand_failures =
+          1.0 * total_expand_failures / total_expand_trials;
+      fprintf(stderr, "Average expand failures, and overhead: %g, %g\n",
+              average_expand_failures,
+              total_expand_overhead / total_expand_trials);
+      // Seems to be a generous allowance
+      EXPECT_LE(total_expand_failures,
+                InfrequentPoissonUpperBound(1.0 * total_expand_trials));
+    } else {
+      fprintf(stderr, "Average expand failures: N/A\n");
+    }
+
+    if (total_singles > 0) {
+      double single_failure_rate = 1.0 * total_single_failures / total_singles;
+      fprintf(stderr, "Add'l single, failure rate: %g\n", single_failure_rate);
+      // A rough bound (one sided) based on nothing in particular
+      double expected_single_failures = 1.0 * total_singles /
+                                        (sizeof(CoeffRow) == 16 ? 128
+                                         : TypeParam::kUseSmash ? 64
+                                                                : 32);
+      EXPECT_LE(total_single_failures,
+                InfrequentPoissonUpperBound(expected_single_failures));
+    }
+
+    if (total_batch > 0) {
+      // Counting successes here for Poisson to approximate the Binomial
+      // distribution.
+      // A rough bound (one sided) based on nothing in particular.
+      double expected_batch_successes = 1.0 * total_batch / 2;
+      uint64_t lower_bound =
+          InfrequentPoissonLowerBound(expected_batch_successes);
+      fprintf(stderr, "Add'l batch, success rate: %g (>= %g)\n",
+              1.0 * total_batch_successes / total_batch,
+              1.0 * lower_bound / total_batch);
+      EXPECT_GE(total_batch_successes, lower_bound);
+    }
+
+    {
+      uint64_t total_checked = uint64_t{FLAGS_max_check} * FLAGS_thoroughness;
+      double expected_total_fp_count =
+          total_checked * std::pow(0.5, 8U * sizeof(ResultRow));
+      // For expected FP rate, also include false positives due to collisions
+      // in Hash value. (Negligible for 64-bit, can matter for 32-bit.)
+      double average_added = 1.0 * total_added / FLAGS_thoroughness;
+      expected_total_fp_count +=
+          total_checked * ExpectedCollisionFpRate(Hasher(), average_added);
+
+      uint64_t upper_bound =
+          InfrequentPoissonUpperBound(expected_total_fp_count);
+      uint64_t lower_bound =
+          InfrequentPoissonLowerBound(expected_total_fp_count);
+      fprintf(stderr, "Average FP rate: %g (~= %g, <= %g, >= %g)\n",
+              1.0 * total_fp_count / total_checked,
+              expected_total_fp_count / total_checked,
+              1.0 * upper_bound / total_checked,
+              1.0 * lower_bound / total_checked);
+      EXPECT_LE(total_fp_count, upper_bound);
+      EXPECT_GE(total_fp_count, lower_bound);
+    }
+  }
+}
+
+TYPED_TEST(RibbonTypeParamTest, Extremes) {
+  IMPORT_RIBBON_TYPES_AND_SETTINGS(TypeParam);
+  IMPORT_RIBBON_IMPL_TYPES(TypeParam);
+  using KeyGen = typename TypeParam::KeyGen;
+
+  size_t bytes = 128 * 1024;
+  std::unique_ptr<char[]> buf(new char[bytes]);
+  InterleavedSoln isoln(buf.get(), bytes);
+  SimpleSoln soln;
+  Hasher hasher;
+  Banding banding;
+
+  // ########################################
+  // Add zero keys to minimal number of slots
+  KeyGen begin_and_end("foo", 123);
+  ASSERT_TRUE(banding.ResetAndFindSeedToSolve(
+      /*slots*/ kCoeffBits, begin_and_end, begin_and_end, /*first seed*/ 0,
+      /* seed mask*/ 0));
+
+  soln.BackSubstFrom(banding);
+  isoln.BackSubstFrom(banding);
+
+  // Because there's plenty of memory, we expect the interleaved solution to
+  // use maximum supported columns (same as simple solution)
+  ASSERT_EQ(isoln.GetUpperNumColumns(), 8U * sizeof(ResultRow));
+  ASSERT_EQ(isoln.GetUpperStartBlock(), 0U);
+
+  // Somewhat oddly, we expect same FP rate as if we had essentially filled
+  // up the slots.
+  KeyGen other_keys_begin("not", 0);
+  KeyGen other_keys_end("not", FLAGS_max_check);
+
+  Index fp_count = 0;
+  KeyGen cur = other_keys_begin;
+  while (cur != other_keys_end) {
+    bool isoln_query_result = isoln.FilterQuery(*cur, hasher);
+    bool soln_query_result = soln.FilterQuery(*cur, hasher);
+    // Solutions are equivalent
+    ASSERT_EQ(isoln_query_result, soln_query_result);
+    if (!TypeParam::kHomogeneous) {
+      // And in fact we only expect an FP when ResultRow is 0
+      // (except Homogeneous)
+      ASSERT_EQ(soln_query_result, hasher.GetResultRowFromHash(
+                                       hasher.GetHash(*cur)) == ResultRow{0});
+    }
+    fp_count += soln_query_result ? 1 : 0;
+    ++cur;
+  }
+  {
+    ASSERT_EQ(isoln.ExpectedFpRate(), soln.ExpectedFpRate());
+    double expected_fp_count = isoln.ExpectedFpRate() * FLAGS_max_check;
+    EXPECT_LE(fp_count, InfrequentPoissonUpperBound(expected_fp_count));
+    if (TypeParam::kHomogeneous) {
+      // Pseudorandom garbage in Homogeneous filter can "beat the odds" if
+      // nothing added
+    } else {
+      EXPECT_GE(fp_count, InfrequentPoissonLowerBound(expected_fp_count));
+    }
+  }
+
+  // ######################################################
+  // Use zero bytes for interleaved solution (key(s) added)
+
+  // Add one key
+  KeyGen key_begin("added", 0);
+  KeyGen key_end("added", 1);
+  ASSERT_TRUE(banding.ResetAndFindSeedToSolve(
+      /*slots*/ kCoeffBits, key_begin, key_end, /*first seed*/ 0,
+      /* seed mask*/ 0));
+
+  InterleavedSoln isoln2(nullptr, /*bytes*/ 0);
+
+  isoln2.BackSubstFrom(banding);
+
+  ASSERT_EQ(isoln2.GetUpperNumColumns(), 0U);
+  ASSERT_EQ(isoln2.GetUpperStartBlock(), 0U);
+
+  // All queries return true
+  ASSERT_TRUE(isoln2.FilterQuery(*other_keys_begin, hasher));
+  ASSERT_EQ(isoln2.ExpectedFpRate(), 1.0);
+}
+
+TEST(RibbonTest, AllowZeroStarts) {
+  IMPORT_RIBBON_TYPES_AND_SETTINGS(TypesAndSettings_AllowZeroStarts);
+  IMPORT_RIBBON_IMPL_TYPES(TypesAndSettings_AllowZeroStarts);
+  using KeyGen = StandardKeyGen;
+
+  InterleavedSoln isoln(nullptr, /*bytes*/ 0);
+  SimpleSoln soln;
+  Hasher hasher;
+  Banding banding;
+
+  KeyGen begin("foo", 0);
+  KeyGen end("foo", 1);
+  // Can't add 1 entry
+  ASSERT_FALSE(banding.ResetAndFindSeedToSolve(/*slots*/ 0, begin, end));
+
+  KeyGen begin_and_end("foo", 123);
+  // Can add 0 entries
+  ASSERT_TRUE(banding.ResetAndFindSeedToSolve(/*slots*/ 0, begin_and_end,
+                                              begin_and_end));
+
+  Seed reseeds = banding.GetOrdinalSeed();
+  ASSERT_EQ(reseeds, 0U);
+  hasher.SetOrdinalSeed(reseeds);
+
+  // Can construct 0-slot solutions
+  isoln.BackSubstFrom(banding);
+  soln.BackSubstFrom(banding);
+
+  // Should always return false
+  ASSERT_FALSE(isoln.FilterQuery(*begin, hasher));
+  ASSERT_FALSE(soln.FilterQuery(*begin, hasher));
+
+  // And report that in FP rate
+  ASSERT_EQ(isoln.ExpectedFpRate(), 0.0);
+  ASSERT_EQ(soln.ExpectedFpRate(), 0.0);
+}
+
+TEST(RibbonTest, RawAndOrdinalSeeds) {
+  StandardHasher<TypesAndSettings_Seed64> hasher64;
+  StandardHasher<DefaultTypesAndSettings> hasher64_32;
+  StandardHasher<TypesAndSettings_Hash32> hasher32;
+  StandardHasher<TypesAndSettings_Seed8> hasher8;
+
+  for (uint32_t limit : {0xffU, 0xffffU}) {
+    std::vector<bool> seen(limit + 1);
+    for (uint32_t i = 0; i < limit; ++i) {
+      hasher64.SetOrdinalSeed(i);
+      auto raw64 = hasher64.GetRawSeed();
+      hasher32.SetOrdinalSeed(i);
+      auto raw32 = hasher32.GetRawSeed();
+      hasher8.SetOrdinalSeed(static_cast<uint8_t>(i));
+      auto raw8 = hasher8.GetRawSeed();
+      {
+        hasher64_32.SetOrdinalSeed(i);
+        auto raw64_32 = hasher64_32.GetRawSeed();
+        ASSERT_EQ(raw64_32, raw32);  // Same size seed
+      }
+      if (i == 0) {
+        // Documented that ordinal seed 0 == raw seed 0
+        ASSERT_EQ(raw64, 0U);
+        ASSERT_EQ(raw32, 0U);
+        ASSERT_EQ(raw8, 0U);
+      } else {
+        // Extremely likely that upper bits are set
+        ASSERT_GT(raw64, raw32);
+        ASSERT_GT(raw32, raw8);
+      }
+      // Hashers agree on lower bits
+      ASSERT_EQ(static_cast<uint32_t>(raw64), raw32);
+      ASSERT_EQ(static_cast<uint8_t>(raw32), raw8);
+
+      // The translation is one-to-one for this size prefix
+      uint32_t v = static_cast<uint32_t>(raw32 & limit);
+      ASSERT_EQ(raw64 & limit, v);
+      ASSERT_FALSE(seen[v]);
+      seen[v] = true;
+    }
+  }
+}
+
+namespace {
+
+struct PhsfInputGen {
+  PhsfInputGen(const std::string& prefix, uint64_t id) : id_(id) {
+    val_.first = prefix;
+    ROCKSDB_NAMESPACE::PutFixed64(&val_.first, /*placeholder*/ 0);
+  }
+
+  // Prefix (only one required)
+  PhsfInputGen& operator++() {
+    ++id_;
+    return *this;
+  }
+
+  const std::pair<std::string, uint8_t>& operator*() {
+    // Use multiplication to mix things up a little in the key
+    ROCKSDB_NAMESPACE::EncodeFixed64(&val_.first[val_.first.size() - 8],
+                                     id_ * uint64_t{0x1500000001});
+    // Occasionally repeat values etc.
+    val_.second = static_cast<uint8_t>(id_ * 7 / 8);
+    return val_;
+  }
+
+  const std::pair<std::string, uint8_t>* operator->() { return &**this; }
+
+  bool operator==(const PhsfInputGen& other) {
+    // Same prefix is assumed
+    return id_ == other.id_;
+  }
+  bool operator!=(const PhsfInputGen& other) {
+    // Same prefix is assumed
+    return id_ != other.id_;
+  }
+
+  uint64_t id_;
+  std::pair<std::string, uint8_t> val_;
+};
+
+struct PhsfTypesAndSettings : public DefaultTypesAndSettings {
+  static constexpr bool kIsFilter = false;
+};
+}  // namespace
+
+TEST(RibbonTest, PhsfBasic) {
+  IMPORT_RIBBON_TYPES_AND_SETTINGS(PhsfTypesAndSettings);
+  IMPORT_RIBBON_IMPL_TYPES(PhsfTypesAndSettings);
+
+  Index num_slots = 12800;
+  Index num_to_add = static_cast<Index>(num_slots / 1.02);
+
+  PhsfInputGen begin("in", 0);
+  PhsfInputGen end("in", num_to_add);
+
+  std::unique_ptr<char[]> idata(new char[/*bytes*/ num_slots]);
+  InterleavedSoln isoln(idata.get(), /*bytes*/ num_slots);
+  SimpleSoln soln;
+  Hasher hasher;
+
+  {
+    Banding banding;
+    ASSERT_TRUE(banding.ResetAndFindSeedToSolve(num_slots, begin, end));
+
+    soln.BackSubstFrom(banding);
+    isoln.BackSubstFrom(banding);
+
+    hasher.SetOrdinalSeed(banding.GetOrdinalSeed());
+  }
+
+  for (PhsfInputGen cur = begin; cur != end; ++cur) {
+    ASSERT_EQ(cur->second, soln.PhsfQuery(cur->first, hasher));
+    ASSERT_EQ(cur->second, isoln.PhsfQuery(cur->first, hasher));
+  }
+}
+
+// Not a real test, but a tool used to build APIs in ribbon_config.h
+TYPED_TEST(RibbonTypeParamTest, FindOccupancy) {
+  IMPORT_RIBBON_TYPES_AND_SETTINGS(TypeParam);
+  IMPORT_RIBBON_IMPL_TYPES(TypeParam);
+  using KeyGen = typename TypeParam::KeyGen;
+
+  if (!FLAGS_find_occ) {
+    ROCKSDB_GTEST_BYPASS("Tool disabled during unit test runs");
+    return;
+  }
+
+  KeyGen cur(std::to_string(testing::UnitTest::GetInstance()->random_seed()),
+             0);
+
+  Banding banding;
+  Index num_slots = InterleavedSoln::RoundUpNumSlots(FLAGS_find_min_slots);
+  Index max_slots = InterleavedSoln::RoundUpNumSlots(FLAGS_find_max_slots);
+  while (num_slots <= max_slots) {
+    std::map<int32_t, uint32_t> rem_histogram;
+    std::map<Index, uint32_t> slot_histogram;
+    if (FLAGS_find_slot_occ) {
+      for (Index i = 0; i < kCoeffBits; ++i) {
+        slot_histogram[i] = 0;
+        slot_histogram[num_slots - 1 - i] = 0;
+        slot_histogram[num_slots / 2 - kCoeffBits / 2 + i] = 0;
+      }
+    }
+    uint64_t total_added = 0;
+    for (uint32_t i = 0; i < FLAGS_find_iters; ++i) {
+      banding.Reset(num_slots);
+      uint32_t j = 0;
+      KeyGen end = cur;
+      end += num_slots + num_slots / 10;
+      for (; cur != end; ++cur) {
+        if (banding.Add(*cur)) {
+          ++j;
+        } else {
+          break;
+        }
+      }
+      total_added += j;
+      for (auto& slot : slot_histogram) {
+        slot.second += banding.IsOccupied(slot.first);
+      }
+
+      int32_t bucket =
+          static_cast<int32_t>(num_slots) - static_cast<int32_t>(j);
+      rem_histogram[bucket]++;
+      if (FLAGS_verbose) {
+        fprintf(stderr, "num_slots: %u i: %u / %u avg_overhead: %g\r",
+                static_cast<unsigned>(num_slots), static_cast<unsigned>(i),
+                static_cast<unsigned>(FLAGS_find_iters),
+                1.0 * (i + 1) * num_slots / total_added);
+      }
+    }
+    if (FLAGS_verbose) {
+      fprintf(stderr, "\n");
+    }
+
+    uint32_t cumulative = 0;
+
+    double p50_rem = 0;
+    double p95_rem = 0;
+    double p99_9_rem = 0;
+
+    for (auto& h : rem_histogram) {
+      double before = 1.0 * cumulative / FLAGS_find_iters;
+      double not_after = 1.0 * (cumulative + h.second) / FLAGS_find_iters;
+      if (FLAGS_verbose) {
+        fprintf(stderr, "overhead: %g before: %g not_after: %g\n",
+                1.0 * num_slots / (num_slots - h.first), before, not_after);
+      }
+      cumulative += h.second;
+      if (before < 0.5 && 0.5 <= not_after) {
+        // fake it with linear interpolation
+        double portion = (0.5 - before) / (not_after - before);
+        p50_rem = h.first + portion;
+      } else if (before < 0.95 && 0.95 <= not_after) {
+        // fake it with linear interpolation
+        double portion = (0.95 - before) / (not_after - before);
+        p95_rem = h.first + portion;
+      } else if (before < 0.999 && 0.999 <= not_after) {
+        // fake it with linear interpolation
+        double portion = (0.999 - before) / (not_after - before);
+        p99_9_rem = h.first + portion;
+      }
+    }
+    for (auto& slot : slot_histogram) {
+      fprintf(stderr, "slot[%u] occupied: %g\n", (unsigned)slot.first,
+              1.0 * slot.second / FLAGS_find_iters);
+    }
+
+    double mean_rem =
+        (1.0 * FLAGS_find_iters * num_slots - total_added) / FLAGS_find_iters;
+    fprintf(
+        stderr,
+        "num_slots: %u iters: %u mean_ovr: %g p50_ovr: %g p95_ovr: %g "
+        "p99.9_ovr: %g mean_rem: %g p50_rem: %g p95_rem: %g p99.9_rem: %g\n",
+        static_cast<unsigned>(num_slots),
+        static_cast<unsigned>(FLAGS_find_iters),
+        1.0 * num_slots / (num_slots - mean_rem),
+        1.0 * num_slots / (num_slots - p50_rem),
+        1.0 * num_slots / (num_slots - p95_rem),
+        1.0 * num_slots / (num_slots - p99_9_rem), mean_rem, p50_rem, p95_rem,
+        p99_9_rem);
+
+    num_slots = std::max(
+        num_slots + 1, static_cast<Index>(num_slots * FLAGS_find_next_factor));
+    num_slots = InterleavedSoln::RoundUpNumSlots(num_slots);
+  }
+}
+
+// Not a real test, but a tool to understand Homogeneous Ribbon
+// behavior (TODO: configuration APIs & tests)
+TYPED_TEST(RibbonTypeParamTest, OptimizeHomogAtScale) {
+  IMPORT_RIBBON_TYPES_AND_SETTINGS(TypeParam);
+  IMPORT_RIBBON_IMPL_TYPES(TypeParam);
+  using KeyGen = typename TypeParam::KeyGen;
+
+  if (!FLAGS_optimize_homog) {
+    ROCKSDB_GTEST_BYPASS("Tool disabled during unit test runs");
+    return;
+  }
+
+  if (!TypeParam::kHomogeneous) {
+    ROCKSDB_GTEST_BYPASS("Only for Homogeneous Ribbon");
+    return;
+  }
+
+  KeyGen cur(std::to_string(testing::UnitTest::GetInstance()->random_seed()),
+             0);
+
+  Banding banding;
+  Index num_slots = SimpleSoln::RoundUpNumSlots(FLAGS_optimize_homog_slots);
+  banding.Reset(num_slots);
+
+  // This and "band_ovr" is the "allocated overhead", or slots over added.
+  // It does not take into account FP rates.
+  double target_overhead = 1.20;
+  uint32_t num_added = 0;
+
+  do {
+    do {
+      (void)banding.Add(*cur);
+      ++cur;
+      ++num_added;
+    } while (1.0 * num_slots / num_added > target_overhead);
+
+    SimpleSoln soln;
+    soln.BackSubstFrom(banding);
+
+    std::array<uint32_t, 8U * sizeof(ResultRow)> fp_counts_by_cols;
+    fp_counts_by_cols.fill(0U);
+    for (uint32_t i = 0; i < FLAGS_optimize_homog_check; ++i) {
+      ResultRow r = soln.PhsfQuery(*cur, banding);
+      ++cur;
+      for (size_t j = 0; j < fp_counts_by_cols.size(); ++j) {
+        if ((r & 1) == 1) {
+          break;
+        }
+        fp_counts_by_cols[j]++;
+        r /= 2;
+      }
+    }
+    fprintf(stderr, "band_ovr: %g ", 1.0 * num_slots / num_added);
+    for (unsigned j = 0; j < fp_counts_by_cols.size(); ++j) {
+      double inv_fp_rate =
+          1.0 * FLAGS_optimize_homog_check / fp_counts_by_cols[j];
+      double equiv_cols = std::log(inv_fp_rate) * 1.4426950409;
+      // Overhead vs. information-theoretic minimum based on observed
+      // FP rate (subject to sampling error, especially for low FP rates)
+      double actual_overhead =
+          1.0 * (j + 1) * num_slots / (equiv_cols * num_added);
+      fprintf(stderr, "ovr_%u: %g ", j + 1, actual_overhead);
+    }
+    fprintf(stderr, "\n");
+    target_overhead -= FLAGS_optimize_homog_granularity;
+  } while (target_overhead > 1.0);
+}
+
+int main(int argc, char** argv) {
+  ROCKSDB_NAMESPACE::port::InstallStackTraceHandler();
+  ::testing::InitGoogleTest(&argc, argv);
+#ifdef GFLAGS
+  ParseCommandLineFlags(&argc, &argv, true);
+#endif  // GFLAGS
+  return RUN_ALL_TESTS();
+}
diff --git a/src/rocksdb/util/set_comparator.h b/src/rocksdb/util/set_comparator.h
new file mode 100644
index 000000000..e0e64436a
--- /dev/null
+++ b/src/rocksdb/util/set_comparator.h
@@ -0,0 +1,24 @@
+// Copyright (c) 2011-present, Facebook, Inc. All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#pragma once
+
+#include "rocksdb/comparator.h"
+
+namespace ROCKSDB_NAMESPACE {
+// A comparator to be used in std::set
+struct SetComparator {
+  explicit SetComparator() : user_comparator_(BytewiseComparator()) {}
+  explicit SetComparator(const Comparator* user_comparator)
+      : user_comparator_(user_comparator ? user_comparator
+                                         : BytewiseComparator()) {}
+  bool operator()(const Slice& lhs, const Slice& rhs) const {
+    return user_comparator_->Compare(lhs, rhs) < 0;
+  }
+
+ private:
+  const Comparator* user_comparator_;
+};
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/util/single_thread_executor.h b/src/rocksdb/util/single_thread_executor.h
new file mode 100644
index 000000000..c69f2a292
--- /dev/null
+++ b/src/rocksdb/util/single_thread_executor.h
@@ -0,0 +1,56 @@
+//  Copyright (c) Meta Platforms, Inc. and affiliates.
+//
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+//
+#pragma once
+
+#if USE_COROUTINES
+#include <atomic>
+
+#include "folly/CPortability.h"
+#include "folly/CppAttributes.h"
+#include "folly/Executor.h"
+#include "util/async_file_reader.h"
+
+namespace ROCKSDB_NAMESPACE {
+// Implements a simple executor that runs callback functions in the same
+// thread, unlike CPUThreadExecutor which may schedule the callback on
+// another thread. Runs in a tight loop calling the queued callbacks,
+// and polls for async IO completions when idle. The completions will
+// resume suspended coroutines and they get added to the queue, which
+// will get picked up by this loop.
+// Any possibility of deadlock is precluded because the file system
+// guarantees that async IO completion callbacks will not be scheduled
+// to run in this thread or this executor.
+class SingleThreadExecutor : public folly::Executor {
+ public:
+  explicit SingleThreadExecutor(AsyncFileReader& reader)
+      : reader_(reader), busy_(false) {}
+
+  void add(folly::Func callback) override {
+    auto& q = q_;
+    q.push(std::move(callback));
+    if (q.size() == 1 && !busy_) {
+      while (!q.empty()) {
+        q.front()();
+        q.pop();
+
+        if (q.empty()) {
+          // Prevent recursion, as the Wait may queue resumed coroutines
+          busy_ = true;
+          reader_.Wait();
+          busy_ = false;
+        }
+      }
+    }
+  }
+
+ private:
+  std::queue<folly::Func> q_;
+  AsyncFileReader& reader_;
+  bool busy_;
+};
+}  // namespace ROCKSDB_NAMESPACE
+#endif  // USE_COROUTINES
diff --git a/src/rocksdb/util/slice.cc b/src/rocksdb/util/slice.cc
new file mode 100644
index 000000000..1fa21afcb
--- /dev/null
+++ b/src/rocksdb/util/slice.cc
@@ -0,0 +1,405 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+//
+// Copyright (c) 2012 The LevelDB Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file. See the AUTHORS file for names of contributors.
+
+#include "rocksdb/slice.h"
+
+#include <stdio.h>
+
+#include <algorithm>
+
+#include "rocksdb/convenience.h"
+#include "rocksdb/slice_transform.h"
+#include "rocksdb/utilities/object_registry.h"
+#include "rocksdb/utilities/options_type.h"
+#include "util/string_util.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+namespace {
+
+class FixedPrefixTransform : public SliceTransform {
+ private:
+  size_t prefix_len_;
+  std::string id_;
+
+ public:
+  explicit FixedPrefixTransform(size_t prefix_len) : prefix_len_(prefix_len) {
+    id_ = std::string(kClassName()) + "." + std::to_string(prefix_len_);
+  }
+
+  static const char* kClassName() { return "rocksdb.FixedPrefix"; }
+  static const char* kNickName() { return "fixed"; }
+  const char* Name() const override { return kClassName(); }
+  const char* NickName() const override { return kNickName(); }
+
+  bool IsInstanceOf(const std::string& name) const override {
+    if (name == id_) {
+      return true;
+    } else if (StartsWith(name, kNickName())) {
+      std::string alt_id =
+          std::string(kNickName()) + ":" + std::to_string(prefix_len_);
+      if (name == alt_id) {
+        return true;
+      }
+    }
+    return SliceTransform::IsInstanceOf(name);
+  }
+
+  std::string GetId() const override { return id_; }
+
+  Slice Transform(const Slice& src) const override {
+    assert(InDomain(src));
+    return Slice(src.data(), prefix_len_);
+  }
+
+  bool InDomain(const Slice& src) const override {
+    return (src.size() >= prefix_len_);
+  }
+
+  bool InRange(const Slice& dst) const override {
+    return (dst.size() == prefix_len_);
+  }
+
+  bool FullLengthEnabled(size_t* len) const override {
+    *len = prefix_len_;
+    return true;
+  }
+
+  bool SameResultWhenAppended(const Slice& prefix) const override {
+    return InDomain(prefix);
+  }
+};
+
+class CappedPrefixTransform : public SliceTransform {
+ private:
+  size_t cap_len_;
+  std::string id_;
+
+ public:
+  explicit CappedPrefixTransform(size_t cap_len) : cap_len_(cap_len) {
+    id_ = std::string(kClassName()) + "." + std::to_string(cap_len_);
+  }
+
+  static const char* kClassName() { return "rocksdb.CappedPrefix"; }
+  static const char* kNickName() { return "capped"; }
+  const char* Name() const override { return kClassName(); }
+  const char* NickName() const override { return kNickName(); }
+  std::string GetId() const override { return id_; }
+
+  bool IsInstanceOf(const std::string& name) const override {
+    if (name == id_) {
+      return true;
+    } else if (StartsWith(name, kNickName())) {
+      std::string alt_id =
+          std::string(kNickName()) + ":" + std::to_string(cap_len_);
+      if (name == alt_id) {
+        return true;
+      }
+    }
+    return SliceTransform::IsInstanceOf(name);
+  }
+
+  Slice Transform(const Slice& src) const override {
+    assert(InDomain(src));
+    return Slice(src.data(), std::min(cap_len_, src.size()));
+  }
+
+  bool InDomain(const Slice& /*src*/) const override { return true; }
+
+  bool InRange(const Slice& dst) const override {
+    return (dst.size() <= cap_len_);
+  }
+
+  bool FullLengthEnabled(size_t* len) const override {
+    *len = cap_len_;
+    return true;
+  }
+
+  bool SameResultWhenAppended(const Slice& prefix) const override {
+    return prefix.size() >= cap_len_;
+  }
+};
+
+class NoopTransform : public SliceTransform {
+ public:
+  explicit NoopTransform() {}
+
+  static const char* kClassName() { return "rocksdb.Noop"; }
+  const char* Name() const override { return kClassName(); }
+
+  Slice Transform(const Slice& src) const override { return src; }
+
+  bool InDomain(const Slice& /*src*/) const override { return true; }
+
+  bool InRange(const Slice& /*dst*/) const override { return true; }
+
+  bool SameResultWhenAppended(const Slice& /*prefix*/) const override {
+    return false;
+  }
+};
+
+}  // end namespace
+
+const SliceTransform* NewFixedPrefixTransform(size_t prefix_len) {
+  return new FixedPrefixTransform(prefix_len);
+}
+
+const SliceTransform* NewCappedPrefixTransform(size_t cap_len) {
+  return new CappedPrefixTransform(cap_len);
+}
+
+const SliceTransform* NewNoopTransform() { return new NoopTransform; }
+
+#ifndef ROCKSDB_LITE
+static int RegisterBuiltinSliceTransform(ObjectLibrary& library,
+                                         const std::string& /*arg*/) {
+  // For the builtin transforms, the format is typically
+  // [Name].[0-9]+ or [NickName]:[0-9]+
+  library.AddFactory<const SliceTransform>(
+      NoopTransform::kClassName(),
+      [](const std::string& /*uri*/,
+         std::unique_ptr<const SliceTransform>* guard,
+         std::string* /*errmsg*/) {
+        guard->reset(NewNoopTransform());
+        return guard->get();
+      });
+  library.AddFactory<const SliceTransform>(
+      ObjectLibrary::PatternEntry(FixedPrefixTransform::kNickName(), false)
+          .AddNumber(":"),
+      [](const std::string& uri, std::unique_ptr<const SliceTransform>* guard,
+         std::string* /*errmsg*/) {
+        auto colon = uri.find(":");
+        auto len = ParseSizeT(uri.substr(colon + 1));
+        guard->reset(NewFixedPrefixTransform(len));
+        return guard->get();
+      });
+  library.AddFactory<const SliceTransform>(
+      ObjectLibrary::PatternEntry(FixedPrefixTransform::kClassName(), false)
+          .AddNumber("."),
+      [](const std::string& uri, std::unique_ptr<const SliceTransform>* guard,
+         std::string* /*errmsg*/) {
+        auto len = ParseSizeT(
+            uri.substr(strlen(FixedPrefixTransform::kClassName()) + 1));
+        guard->reset(NewFixedPrefixTransform(len));
+        return guard->get();
+      });
+  library.AddFactory<const SliceTransform>(
+      ObjectLibrary::PatternEntry(CappedPrefixTransform::kNickName(), false)
+          .AddNumber(":"),
+      [](const std::string& uri, std::unique_ptr<const SliceTransform>* guard,
+         std::string* /*errmsg*/) {
+        auto colon = uri.find(":");
+        auto len = ParseSizeT(uri.substr(colon + 1));
+        guard->reset(NewCappedPrefixTransform(len));
+        return guard->get();
+      });
+  library.AddFactory<const SliceTransform>(
+      ObjectLibrary::PatternEntry(CappedPrefixTransform::kClassName(), false)
+          .AddNumber("."),
+      [](const std::string& uri, std::unique_ptr<const SliceTransform>* guard,
+         std::string* /*errmsg*/) {
+        auto len = ParseSizeT(
+            uri.substr(strlen(CappedPrefixTransform::kClassName()) + 1));
+        guard->reset(NewCappedPrefixTransform(len));
+        return guard->get();
+      });
+  size_t num_types;
+  return static_cast<int>(library.GetFactoryCount(&num_types));
+}
+#endif  // ROCKSDB_LITE
+
+Status SliceTransform::CreateFromString(
+    const ConfigOptions& config_options, const std::string& value,
+    std::shared_ptr<const SliceTransform>* result) {
+#ifndef ROCKSDB_LITE
+  static std::once_flag once;
+  std::call_once(once, [&]() {
+    RegisterBuiltinSliceTransform(*(ObjectLibrary::Default().get()), "");
+  });
+#endif  // ROCKSDB_LITE
+  std::string id;
+  std::unordered_map<std::string, std::string> opt_map;
+  Status status = Customizable::GetOptionsMap(config_options, result->get(),
+                                              value, &id, &opt_map);
+  if (!status.ok()) {  // GetOptionsMap failed
+    return status;
+  } else if (id.empty() && opt_map.empty()) {
+    result->reset();
+  } else {
+#ifndef ROCKSDB_LITE
+    status = config_options.registry->NewSharedObject(id, result);
+#else
+    auto Matches = [](const std::string& input, size_t size,
+                      const char* pattern, char sep) {
+      auto plen = strlen(pattern);
+      return (size > plen + 2 && input[plen] == sep &&
+              StartsWith(input, pattern));
+    };
+
+    auto size = id.size();
+    if (id == NoopTransform::kClassName()) {
+      result->reset(NewNoopTransform());
+    } else if (Matches(id, size, FixedPrefixTransform::kNickName(), ':')) {
+      auto fixed = strlen(FixedPrefixTransform::kNickName());
+      auto len = ParseSizeT(id.substr(fixed + 1));
+      result->reset(NewFixedPrefixTransform(len));
+    } else if (Matches(id, size, CappedPrefixTransform::kNickName(), ':')) {
+      auto capped = strlen(CappedPrefixTransform::kNickName());
+      auto len = ParseSizeT(id.substr(capped + 1));
+      result->reset(NewCappedPrefixTransform(len));
+    } else if (Matches(id, size, CappedPrefixTransform::kClassName(), '.')) {
+      auto capped = strlen(CappedPrefixTransform::kClassName());
+      auto len = ParseSizeT(id.substr(capped + 1));
+      result->reset(NewCappedPrefixTransform(len));
+    } else if (Matches(id, size, FixedPrefixTransform::kClassName(), '.')) {
+      auto fixed = strlen(FixedPrefixTransform::kClassName());
+      auto len = ParseSizeT(id.substr(fixed + 1));
+      result->reset(NewFixedPrefixTransform(len));
+    } else {
+      status = Status::NotSupported("Cannot load object in LITE mode ", id);
+    }
+#endif  // ROCKSDB_LITE
+    if (config_options.ignore_unsupported_options && status.IsNotSupported()) {
+      return Status::OK();
+    } else if (status.ok()) {
+      SliceTransform* transform = const_cast<SliceTransform*>(result->get());
+      status =
+          Customizable::ConfigureNewObject(config_options, transform, opt_map);
+    }
+  }
+  return status;
+}
+
+std::string SliceTransform::AsString() const {
+#ifndef ROCKSDB_LITE
+  if (HasRegisteredOptions()) {
+    ConfigOptions opts;
+    opts.delimiter = ";";
+    return ToString(opts);
+  }
+#endif  // ROCKSDB_LITE
+  return GetId();
+}
+
+// 2 small internal utility functions, for efficient hex conversions
+// and no need for snprintf, toupper etc...
+// Originally from wdt/util/EncryptionUtils.cpp - for
+// std::to_string(true)/DecodeHex:
+char toHex(unsigned char v) {
+  if (v <= 9) {
+    return '0' + v;
+  }
+  return 'A' + v - 10;
+}
+// most of the code is for validation/error check
+int fromHex(char c) {
+  // toupper:
+  if (c >= 'a' && c <= 'f') {
+    c -= ('a' - 'A');  // aka 0x20
+  }
+  // validation
+  if (c < '0' || (c > '9' && (c < 'A' || c > 'F'))) {
+    return -1;  // invalid not 0-9A-F hex char
+  }
+  if (c <= '9') {
+    return c - '0';
+  }
+  return c - 'A' + 10;
+}
+
+Slice::Slice(const SliceParts& parts, std::string* buf) {
+  size_t length = 0;
+  for (int i = 0; i < parts.num_parts; ++i) {
+    length += parts.parts[i].size();
+  }
+  buf->reserve(length);
+
+  for (int i = 0; i < parts.num_parts; ++i) {
+    buf->append(parts.parts[i].data(), parts.parts[i].size());
+  }
+  data_ = buf->data();
+  size_ = buf->size();
+}
+
+// Return a string that contains the copy of the referenced data.
+std::string Slice::ToString(bool hex) const {
+  std::string result;  // RVO/NRVO/move
+  if (hex) {
+    result.reserve(2 * size_);
+    for (size_t i = 0; i < size_; ++i) {
+      unsigned char c = data_[i];
+      result.push_back(toHex(c >> 4));
+      result.push_back(toHex(c & 0xf));
+    }
+    return result;
+  } else {
+    result.assign(data_, size_);
+    return result;
+  }
+}
+
+// Originally from rocksdb/utilities/ldb_cmd.h
+bool Slice::DecodeHex(std::string* result) const {
+  std::string::size_type len = size_;
+  if (len % 2) {
+    // Hex string must be even number of hex digits to get complete bytes back
+    return false;
+  }
+  if (!result) {
+    return false;
+  }
+  result->clear();
+  result->reserve(len / 2);
+
+  for (size_t i = 0; i < len;) {
+    int h1 = fromHex(data_[i++]);
+    if (h1 < 0) {
+      return false;
+    }
+    int h2 = fromHex(data_[i++]);
+    if (h2 < 0) {
+      return false;
+    }
+    result->push_back(static_cast<char>((h1 << 4) | h2));
+  }
+  return true;
+}
+
+PinnableSlice::PinnableSlice(PinnableSlice&& other) {
+  *this = std::move(other);
+}
+
+PinnableSlice& PinnableSlice::operator=(PinnableSlice&& other) {
+  if (this != &other) {
+    Cleanable::Reset();
+    Cleanable::operator=(std::move(other));
+    size_ = other.size_;
+    pinned_ = other.pinned_;
+    if (pinned_) {
+      data_ = other.data_;
+      // When it's pinned, buf should no longer be of use.
+    } else {
+      if (other.buf_ == &other.self_space_) {
+        self_space_ = std::move(other.self_space_);
+        buf_ = &self_space_;
+        data_ = buf_->data();
+      } else {
+        buf_ = other.buf_;
+        data_ = other.data_;
+      }
+    }
+    other.self_space_.clear();
+    other.buf_ = &other.self_space_;
+    other.pinned_ = false;
+    other.PinSelf();
+  }
+  return *this;
+}
+
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/util/slice_test.cc b/src/rocksdb/util/slice_test.cc
new file mode 100644
index 000000000..e1c35d567
--- /dev/null
+++ b/src/rocksdb/util/slice_test.cc
@@ -0,0 +1,191 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#include "rocksdb/slice.h"
+
+#include <gtest/gtest.h>
+
+#include "port/port.h"
+#include "port/stack_trace.h"
+#include "rocksdb/data_structure.h"
+#include "rocksdb/types.h"
+#include "test_util/testharness.h"
+#include "test_util/testutil.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+TEST(SliceTest, StringView) {
+  std::string s = "foo";
+  std::string_view sv = s;
+  ASSERT_EQ(Slice(s), Slice(sv));
+  ASSERT_EQ(Slice(s), Slice(std::move(sv)));
+}
+
+// Use this to keep track of the cleanups that were actually performed
+void Multiplier(void* arg1, void* arg2) {
+  int* res = reinterpret_cast<int*>(arg1);
+  int* num = reinterpret_cast<int*>(arg2);
+  *res *= *num;
+}
+
+class PinnableSliceTest : public testing::Test {
+ public:
+  void AssertSameData(const std::string& expected, const PinnableSlice& slice) {
+    std::string got;
+    got.assign(slice.data(), slice.size());
+    ASSERT_EQ(expected, got);
+  }
+};
+
+// Test that the external buffer is moved instead of being copied.
+TEST_F(PinnableSliceTest, MoveExternalBuffer) {
+  Slice s("123");
+  std::string buf;
+  PinnableSlice v1(&buf);
+  v1.PinSelf(s);
+
+  PinnableSlice v2(std::move(v1));
+  ASSERT_EQ(buf.data(), v2.data());
+  ASSERT_EQ(&buf, v2.GetSelf());
+
+  PinnableSlice v3;
+  v3 = std::move(v2);
+  ASSERT_EQ(buf.data(), v3.data());
+  ASSERT_EQ(&buf, v3.GetSelf());
+}
+
+TEST_F(PinnableSliceTest, Move) {
+  int n2 = 2;
+  int res = 1;
+  const std::string const_str1 = "123";
+  const std::string const_str2 = "ABC";
+  Slice slice1(const_str1);
+  Slice slice2(const_str2);
+
+  {
+    // Test move constructor on a pinned slice.
+    res = 1;
+    PinnableSlice v1;
+    v1.PinSlice(slice1, Multiplier, &res, &n2);
+    PinnableSlice v2(std::move(v1));
+
+    // Since v1's Cleanable has been moved to v2,
+    // no cleanup should happen in Reset.
+    v1.Reset();
+    ASSERT_EQ(1, res);
+
+    AssertSameData(const_str1, v2);
+  }
+  // v2 is cleaned up.
+  ASSERT_EQ(2, res);
+
+  {
+    // Test move constructor on an unpinned slice.
+    PinnableSlice v1;
+    v1.PinSelf(slice1);
+    PinnableSlice v2(std::move(v1));
+
+    AssertSameData(const_str1, v2);
+  }
+
+  {
+    // Test move assignment from a pinned slice to
+    // another pinned slice.
+    res = 1;
+    PinnableSlice v1;
+    v1.PinSlice(slice1, Multiplier, &res, &n2);
+    PinnableSlice v2;
+    v2.PinSlice(slice2, Multiplier, &res, &n2);
+    v2 = std::move(v1);
+
+    // v2's Cleanable will be Reset before moving
+    // anything from v1.
+    ASSERT_EQ(2, res);
+    // Since v1's Cleanable has been moved to v2,
+    // no cleanup should happen in Reset.
+    v1.Reset();
+    ASSERT_EQ(2, res);
+
+    AssertSameData(const_str1, v2);
+  }
+  // The Cleanable moved from v1 to v2 will be Reset.
+  ASSERT_EQ(4, res);
+
+  {
+    // Test move assignment from a pinned slice to
+    // an unpinned slice.
+    res = 1;
+    PinnableSlice v1;
+    v1.PinSlice(slice1, Multiplier, &res, &n2);
+    PinnableSlice v2;
+    v2.PinSelf(slice2);
+    v2 = std::move(v1);
+
+    // Since v1's Cleanable has been moved to v2,
+    // no cleanup should happen in Reset.
+    v1.Reset();
+    ASSERT_EQ(1, res);
+
+    AssertSameData(const_str1, v2);
+  }
+  // The Cleanable moved from v1 to v2 will be Reset.
+  ASSERT_EQ(2, res);
+
+  {
+    // Test move assignment from an upinned slice to
+    // another unpinned slice.
+    PinnableSlice v1;
+    v1.PinSelf(slice1);
+    PinnableSlice v2;
+    v2.PinSelf(slice2);
+    v2 = std::move(v1);
+
+    AssertSameData(const_str1, v2);
+  }
+
+  {
+    // Test move assignment from an upinned slice to
+    // a pinned slice.
+    res = 1;
+    PinnableSlice v1;
+    v1.PinSelf(slice1);
+    PinnableSlice v2;
+    v2.PinSlice(slice2, Multiplier, &res, &n2);
+    v2 = std::move(v1);
+
+    // v2's Cleanable will be Reset before moving
+    // anything from v1.
+    ASSERT_EQ(2, res);
+
+    AssertSameData(const_str1, v2);
+  }
+  // No Cleanable is moved from v1 to v2, so no more cleanup.
+  ASSERT_EQ(2, res);
+}
+
+// ***************************************************************** //
+// Unit test for SmallEnumSet
+class SmallEnumSetTest : public testing::Test {
+ public:
+  SmallEnumSetTest() {}
+  ~SmallEnumSetTest() {}
+};
+
+TEST_F(SmallEnumSetTest, SmallSetTest) {
+  FileTypeSet fs;
+  ASSERT_TRUE(fs.Add(FileType::kIdentityFile));
+  ASSERT_FALSE(fs.Add(FileType::kIdentityFile));
+  ASSERT_TRUE(fs.Add(FileType::kInfoLogFile));
+  ASSERT_TRUE(fs.Contains(FileType::kIdentityFile));
+  ASSERT_FALSE(fs.Contains(FileType::kDBLockFile));
+}
+
+}  // namespace ROCKSDB_NAMESPACE
+
+int main(int argc, char** argv) {
+  ROCKSDB_NAMESPACE::port::InstallStackTraceHandler();
+  ::testing::InitGoogleTest(&argc, argv);
+  return RUN_ALL_TESTS();
+}
diff --git a/src/rocksdb/util/slice_transform_test.cc b/src/rocksdb/util/slice_transform_test.cc
new file mode 100644
index 000000000..64ac8bb1f
--- /dev/null
+++ b/src/rocksdb/util/slice_transform_test.cc
@@ -0,0 +1,154 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+//
+// Copyright (c) 2011 The LevelDB Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file. See the AUTHORS file for names of contributors.
+
+#include "rocksdb/slice_transform.h"
+
+#include "rocksdb/db.h"
+#include "rocksdb/env.h"
+#include "rocksdb/filter_policy.h"
+#include "rocksdb/statistics.h"
+#include "rocksdb/table.h"
+#include "test_util/testharness.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+class SliceTransformTest : public testing::Test {};
+
+TEST_F(SliceTransformTest, CapPrefixTransform) {
+  std::string s;
+  s = "abcdefge";
+
+  std::unique_ptr<const SliceTransform> transform;
+
+  transform.reset(NewCappedPrefixTransform(6));
+  ASSERT_EQ(transform->Transform(s).ToString(), "abcdef");
+  ASSERT_TRUE(transform->SameResultWhenAppended("123456"));
+  ASSERT_TRUE(transform->SameResultWhenAppended("1234567"));
+  ASSERT_TRUE(!transform->SameResultWhenAppended("12345"));
+
+  transform.reset(NewCappedPrefixTransform(8));
+  ASSERT_EQ(transform->Transform(s).ToString(), "abcdefge");
+
+  transform.reset(NewCappedPrefixTransform(10));
+  ASSERT_EQ(transform->Transform(s).ToString(), "abcdefge");
+
+  transform.reset(NewCappedPrefixTransform(0));
+  ASSERT_EQ(transform->Transform(s).ToString(), "");
+
+  transform.reset(NewCappedPrefixTransform(0));
+  ASSERT_EQ(transform->Transform("").ToString(), "");
+}
+
+class SliceTransformDBTest : public testing::Test {
+ private:
+  std::string dbname_;
+  Env* env_;
+  DB* db_;
+
+ public:
+  SliceTransformDBTest() : env_(Env::Default()), db_(nullptr) {
+    dbname_ = test::PerThreadDBPath("slice_transform_db_test");
+    EXPECT_OK(DestroyDB(dbname_, last_options_));
+  }
+
+  ~SliceTransformDBTest() override {
+    delete db_;
+    EXPECT_OK(DestroyDB(dbname_, last_options_));
+  }
+
+  DB* db() { return db_; }
+
+  // Return the current option configuration.
+  Options* GetOptions() { return &last_options_; }
+
+  void DestroyAndReopen() {
+    // Destroy using last options
+    Destroy();
+    ASSERT_OK(TryReopen());
+  }
+
+  void Destroy() {
+    delete db_;
+    db_ = nullptr;
+    ASSERT_OK(DestroyDB(dbname_, last_options_));
+  }
+
+  Status TryReopen() {
+    delete db_;
+    db_ = nullptr;
+    last_options_.create_if_missing = true;
+
+    return DB::Open(last_options_, dbname_, &db_);
+  }
+
+  Options last_options_;
+};
+
+namespace {
+uint64_t TestGetTickerCount(const Options& options, Tickers ticker_type) {
+  return options.statistics->getTickerCount(ticker_type);
+}
+}  // namespace
+
+TEST_F(SliceTransformDBTest, CapPrefix) {
+  last_options_.prefix_extractor.reset(NewCappedPrefixTransform(8));
+  last_options_.statistics = ROCKSDB_NAMESPACE::CreateDBStatistics();
+  BlockBasedTableOptions bbto;
+  bbto.filter_policy.reset(NewBloomFilterPolicy(10, false));
+  bbto.whole_key_filtering = false;
+  last_options_.table_factory.reset(NewBlockBasedTableFactory(bbto));
+  ASSERT_OK(TryReopen());
+
+  ReadOptions ro;
+  FlushOptions fo;
+  WriteOptions wo;
+
+  ASSERT_OK(db()->Put(wo, "barbarbar", "foo"));
+  ASSERT_OK(db()->Put(wo, "barbarbar2", "foo2"));
+  ASSERT_OK(db()->Put(wo, "foo", "bar"));
+  ASSERT_OK(db()->Put(wo, "foo3", "bar3"));
+  ASSERT_OK(db()->Flush(fo));
+
+  std::unique_ptr<Iterator> iter(db()->NewIterator(ro));
+
+  iter->Seek("foo");
+  ASSERT_OK(iter->status());
+  ASSERT_TRUE(iter->Valid());
+  ASSERT_EQ(iter->value().ToString(), "bar");
+  ASSERT_EQ(TestGetTickerCount(last_options_, BLOOM_FILTER_PREFIX_USEFUL), 0U);
+
+  iter->Seek("foo2");
+  ASSERT_OK(iter->status());
+  ASSERT_TRUE(!iter->Valid());
+  ASSERT_EQ(TestGetTickerCount(last_options_, BLOOM_FILTER_PREFIX_USEFUL), 1U);
+
+  iter->Seek("barbarbar");
+  ASSERT_OK(iter->status());
+  ASSERT_TRUE(iter->Valid());
+  ASSERT_EQ(iter->value().ToString(), "foo");
+  ASSERT_EQ(TestGetTickerCount(last_options_, BLOOM_FILTER_PREFIX_USEFUL), 1U);
+
+  iter->Seek("barfoofoo");
+  ASSERT_OK(iter->status());
+  ASSERT_TRUE(!iter->Valid());
+  ASSERT_EQ(TestGetTickerCount(last_options_, BLOOM_FILTER_PREFIX_USEFUL), 2U);
+
+  iter->Seek("foobarbar");
+  ASSERT_OK(iter->status());
+  ASSERT_TRUE(!iter->Valid());
+  ASSERT_EQ(TestGetTickerCount(last_options_, BLOOM_FILTER_PREFIX_USEFUL), 3U);
+}
+
+}  // namespace ROCKSDB_NAMESPACE
+
+int main(int argc, char** argv) {
+  ROCKSDB_NAMESPACE::port::InstallStackTraceHandler();
+  ::testing::InitGoogleTest(&argc, argv);
+  return RUN_ALL_TESTS();
+}
diff --git a/src/rocksdb/util/status.cc b/src/rocksdb/util/status.cc
new file mode 100644
index 000000000..72fdfdbcc
--- /dev/null
+++ b/src/rocksdb/util/status.cc
@@ -0,0 +1,154 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+//
+// Copyright (c) 2011 The LevelDB Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file. See the AUTHORS file for names of contributors.
+
+#include "rocksdb/status.h"
+
+#include <stdio.h>
+#ifdef OS_WIN
+#include <string.h>
+#endif
+#include <cstring>
+
+#include "port/port.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+std::unique_ptr<const char[]> Status::CopyState(const char* s) {
+  const size_t cch = std::strlen(s) + 1;  // +1 for the null terminator
+  char* rv = new char[cch];
+  std::strncpy(rv, s, cch);
+  return std::unique_ptr<const char[]>(rv);
+}
+
+static const char* msgs[static_cast<int>(Status::kMaxSubCode)] = {
+    "",                                                   // kNone
+    "Timeout Acquiring Mutex",                            // kMutexTimeout
+    "Timeout waiting to lock key",                        // kLockTimeout
+    "Failed to acquire lock due to max_num_locks limit",  // kLockLimit
+    "No space left on device",                            // kNoSpace
+    "Deadlock",                                           // kDeadlock
+    "Stale file handle",                                  // kStaleFile
+    "Memory limit reached",                               // kMemoryLimit
+    "Space limit reached",                                // kSpaceLimit
+    "No such file or directory",                          // kPathNotFound
+    // KMergeOperandsInsufficientCapacity
+    "Insufficient capacity for merge operands",
+    // kManualCompactionPaused
+    "Manual compaction paused",
+    " (overwritten)",    // kOverwritten, subcode of OK
+    "Txn not prepared",  // kTxnNotPrepared
+    "IO fenced off",     // kIOFenced
+};
+
+Status::Status(Code _code, SubCode _subcode, const Slice& msg,
+               const Slice& msg2, Severity sev)
+    : code_(_code),
+      subcode_(_subcode),
+      sev_(sev),
+      retryable_(false),
+      data_loss_(false),
+      scope_(0) {
+  assert(subcode_ != kMaxSubCode);
+  const size_t len1 = msg.size();
+  const size_t len2 = msg2.size();
+  const size_t size = len1 + (len2 ? (2 + len2) : 0);
+  char* const result = new char[size + 1];  // +1 for null terminator
+  memcpy(result, msg.data(), len1);
+  if (len2) {
+    result[len1] = ':';
+    result[len1 + 1] = ' ';
+    memcpy(result + len1 + 2, msg2.data(), len2);
+  }
+  result[size] = '\0';  // null terminator for C style string
+  state_.reset(result);
+}
+
+std::string Status::ToString() const {
+#ifdef ROCKSDB_ASSERT_STATUS_CHECKED
+  checked_ = true;
+#endif  // ROCKSDB_ASSERT_STATUS_CHECKED
+  const char* type = nullptr;
+  switch (code_) {
+    case kOk:
+      return "OK";
+    case kNotFound:
+      type = "NotFound: ";
+      break;
+    case kCorruption:
+      type = "Corruption: ";
+      break;
+    case kNotSupported:
+      type = "Not implemented: ";
+      break;
+    case kInvalidArgument:
+      type = "Invalid argument: ";
+      break;
+    case kIOError:
+      type = "IO error: ";
+      break;
+    case kMergeInProgress:
+      type = "Merge in progress: ";
+      break;
+    case kIncomplete:
+      type = "Result incomplete: ";
+      break;
+    case kShutdownInProgress:
+      type = "Shutdown in progress: ";
+      break;
+    case kTimedOut:
+      type = "Operation timed out: ";
+      break;
+    case kAborted:
+      type = "Operation aborted: ";
+      break;
+    case kBusy:
+      type = "Resource busy: ";
+      break;
+    case kExpired:
+      type = "Operation expired: ";
+      break;
+    case kTryAgain:
+      type = "Operation failed. Try again.: ";
+      break;
+    case kCompactionTooLarge:
+      type = "Compaction too large: ";
+      break;
+    case kColumnFamilyDropped:
+      type = "Column family dropped: ";
+      break;
+    case kMaxCode:
+      assert(false);
+      break;
+  }
+  char tmp[30];
+  if (type == nullptr) {
+    // This should not happen since `code_` should be a valid non-`kMaxCode`
+    // member of the `Code` enum. The above switch-statement should have had a
+    // case assigning `type` to a corresponding string.
+    assert(false);
+    snprintf(tmp, sizeof(tmp), "Unknown code(%d): ", static_cast<int>(code()));
+    type = tmp;
+  }
+  std::string result(type);
+  if (subcode_ != kNone) {
+    uint32_t index = static_cast<int32_t>(subcode_);
+    assert(sizeof(msgs) / sizeof(msgs[0]) > index);
+    result.append(msgs[index]);
+  }
+
+  if (state_ != nullptr) {
+    if (subcode_ != kNone) {
+      result.append(": ");
+    }
+    result.append(state_.get());
+  }
+  return result;
+}
+
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/util/stderr_logger.cc b/src/rocksdb/util/stderr_logger.cc
new file mode 100644
index 000000000..6044b8b93
--- /dev/null
+++ b/src/rocksdb/util/stderr_logger.cc
@@ -0,0 +1,30 @@
+//  Copyright (c) Meta Platforms, Inc. and affiliates.
+//
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#include "util/stderr_logger.h"
+
+#include "port/sys_time.h"
+
+namespace ROCKSDB_NAMESPACE {
+StderrLogger::~StderrLogger() {}
+
+void StderrLogger::Logv(const char* format, va_list ap) {
+  const uint64_t thread_id = Env::Default()->GetThreadID();
+
+  port::TimeVal now_tv;
+  port::GetTimeOfDay(&now_tv, nullptr);
+  const time_t seconds = now_tv.tv_sec;
+  struct tm t;
+  port::LocalTimeR(&seconds, &t);
+  fprintf(stderr, "%04d/%02d/%02d-%02d:%02d:%02d.%06d %llx ", t.tm_year + 1900,
+          t.tm_mon + 1, t.tm_mday, t.tm_hour, t.tm_min, t.tm_sec,
+          static_cast<int>(now_tv.tv_usec),
+          static_cast<long long unsigned int>(thread_id));
+
+  vfprintf(stderr, format, ap);
+  fprintf(stderr, "\n");
+}
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/util/stderr_logger.h b/src/rocksdb/util/stderr_logger.h
new file mode 100644
index 000000000..c3b01210c
--- /dev/null
+++ b/src/rocksdb/util/stderr_logger.h
@@ -0,0 +1,31 @@
+//  Copyright (c) Meta Platforms, Inc. and affiliates.
+//
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#pragma once
+
+#include <stdarg.h>
+#include <stdio.h>
+
+#include "rocksdb/env.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+// Prints logs to stderr for faster debugging
+class StderrLogger : public Logger {
+ public:
+  explicit StderrLogger(const InfoLogLevel log_level = InfoLogLevel::INFO_LEVEL)
+      : Logger(log_level) {}
+
+  ~StderrLogger() override;
+
+  // Brings overloaded Logv()s into scope so they're not hidden when we override
+  // a subset of them.
+  using Logger::Logv;
+
+  virtual void Logv(const char* format, va_list ap) override;
+};
+
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/util/stop_watch.h b/src/rocksdb/util/stop_watch.h
new file mode 100644
index 000000000..e26380d97
--- /dev/null
+++ b/src/rocksdb/util/stop_watch.h
@@ -0,0 +1,118 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+//
+#pragma once
+#include "monitoring/statistics.h"
+#include "rocksdb/system_clock.h"
+
+namespace ROCKSDB_NAMESPACE {
+// Auto-scoped.
+// Records the measure time into the corresponding histogram if statistics
+// is not nullptr. It is also saved into *elapsed if the pointer is not nullptr
+// and overwrite is true, it will be added to *elapsed if overwrite is false.
+class StopWatch {
+ public:
+  StopWatch(SystemClock* clock, Statistics* statistics,
+            const uint32_t hist_type, uint64_t* elapsed = nullptr,
+            bool overwrite = true, bool delay_enabled = false)
+      : clock_(clock),
+        statistics_(statistics),
+        hist_type_(hist_type),
+        elapsed_(elapsed),
+        overwrite_(overwrite),
+        stats_enabled_(statistics &&
+                       statistics->get_stats_level() >=
+                           StatsLevel::kExceptTimers &&
+                       statistics->HistEnabledForType(hist_type)),
+        delay_enabled_(delay_enabled),
+        total_delay_(0),
+        delay_start_time_(0),
+        start_time_((stats_enabled_ || elapsed != nullptr) ? clock->NowMicros()
+                                                           : 0) {}
+
+  ~StopWatch() {
+    if (elapsed_) {
+      if (overwrite_) {
+        *elapsed_ = clock_->NowMicros() - start_time_;
+      } else {
+        *elapsed_ += clock_->NowMicros() - start_time_;
+      }
+    }
+    if (elapsed_ && delay_enabled_) {
+      *elapsed_ -= total_delay_;
+    }
+    if (stats_enabled_) {
+      statistics_->reportTimeToHistogram(
+          hist_type_, (elapsed_ != nullptr)
+                          ? *elapsed_
+                          : (clock_->NowMicros() - start_time_));
+    }
+  }
+
+  void DelayStart() {
+    // if delay_start_time_ is not 0, it means we are already tracking delay,
+    // so delay_start_time_ should not be overwritten
+    if (elapsed_ && delay_enabled_ && delay_start_time_ == 0) {
+      delay_start_time_ = clock_->NowMicros();
+    }
+  }
+
+  void DelayStop() {
+    if (elapsed_ && delay_enabled_ && delay_start_time_ != 0) {
+      total_delay_ += clock_->NowMicros() - delay_start_time_;
+    }
+    // reset to 0 means currently no delay is being tracked, so two consecutive
+    // calls to DelayStop will not increase total_delay_
+    delay_start_time_ = 0;
+  }
+
+  uint64_t GetDelay() const { return delay_enabled_ ? total_delay_ : 0; }
+
+  uint64_t start_time() const { return start_time_; }
+
+ private:
+  SystemClock* clock_;
+  Statistics* statistics_;
+  const uint32_t hist_type_;
+  uint64_t* elapsed_;
+  bool overwrite_;
+  bool stats_enabled_;
+  bool delay_enabled_;
+  uint64_t total_delay_;
+  uint64_t delay_start_time_;
+  const uint64_t start_time_;
+};
+
+// a nano second precision stopwatch
+class StopWatchNano {
+ public:
+  explicit StopWatchNano(SystemClock* clock, bool auto_start = false)
+      : clock_(clock), start_(0) {
+    if (auto_start) {
+      Start();
+    }
+  }
+
+  void Start() { start_ = clock_->NowNanos(); }
+
+  uint64_t ElapsedNanos(bool reset = false) {
+    auto now = clock_->NowNanos();
+    auto elapsed = now - start_;
+    if (reset) {
+      start_ = now;
+    }
+    return elapsed;
+  }
+
+  uint64_t ElapsedNanosSafe(bool reset = false) {
+    return (clock_ != nullptr) ? ElapsedNanos(reset) : 0U;
+  }
+
+ private:
+  SystemClock* clock_;
+  uint64_t start_;
+};
+
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/util/string_util.cc b/src/rocksdb/util/string_util.cc
new file mode 100644
index 000000000..324482a4c
--- /dev/null
+++ b/src/rocksdb/util/string_util.cc
@@ -0,0 +1,504 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+//
+#include "util/string_util.h"
+
+#include <errno.h>
+#include <stdio.h>
+#include <stdlib.h>
+
+#include <algorithm>
+#include <cinttypes>
+#include <cmath>
+#include <sstream>
+#include <string>
+#include <utility>
+#include <vector>
+
+#include "port/port.h"
+#include "port/sys_time.h"
+#include "rocksdb/slice.h"
+
+#ifndef __has_cpp_attribute
+#define ROCKSDB_HAS_CPP_ATTRIBUTE(x) 0
+#else
+#define ROCKSDB_HAS_CPP_ATTRIBUTE(x) __has_cpp_attribute(x)
+#endif
+
+#if ROCKSDB_HAS_CPP_ATTRIBUTE(maybe_unused) && __cplusplus >= 201703L
+#define ROCKSDB_MAYBE_UNUSED [[maybe_unused]]
+#elif ROCKSDB_HAS_CPP_ATTRIBUTE(gnu::unused) || __GNUC__
+#define ROCKSDB_MAYBE_UNUSED [[gnu::unused]]
+#else
+#define ROCKSDB_MAYBE_UNUSED
+#endif
+
+namespace ROCKSDB_NAMESPACE {
+
+const std::string kNullptrString = "nullptr";
+
+std::vector<std::string> StringSplit(const std::string& arg, char delim) {
+  std::vector<std::string> splits;
+  std::stringstream ss(arg);
+  std::string item;
+  while (std::getline(ss, item, delim)) {
+    splits.push_back(item);
+  }
+  return splits;
+}
+
+// for micros < 10ms, print "XX us".
+// for micros < 10sec, print "XX ms".
+// for micros >= 10 sec, print "XX sec".
+// for micros <= 1 hour, print Y:X M:S".
+// for micros > 1 hour, print Z:Y:X H:M:S".
+int AppendHumanMicros(uint64_t micros, char* output, int len,
+                      bool fixed_format) {
+  if (micros < 10000 && !fixed_format) {
+    return snprintf(output, len, "%" PRIu64 " us", micros);
+  } else if (micros < 10000000 && !fixed_format) {
+    return snprintf(output, len, "%.3lf ms",
+                    static_cast<double>(micros) / 1000);
+  } else if (micros < 1000000l * 60 && !fixed_format) {
+    return snprintf(output, len, "%.3lf sec",
+                    static_cast<double>(micros) / 1000000);
+  } else if (micros < 1000000ll * 60 * 60 && !fixed_format) {
+    return snprintf(output, len, "%02" PRIu64 ":%05.3f M:S",
+                    micros / 1000000 / 60,
+                    static_cast<double>(micros % 60000000) / 1000000);
+  } else {
+    return snprintf(output, len, "%02" PRIu64 ":%02" PRIu64 ":%05.3f H:M:S",
+                    micros / 1000000 / 3600, (micros / 1000000 / 60) % 60,
+                    static_cast<double>(micros % 60000000) / 1000000);
+  }
+}
+
+// for sizes >=10TB, print "XXTB"
+// for sizes >=10GB, print "XXGB"
+// etc.
+// append file size summary to output and return the len
+int AppendHumanBytes(uint64_t bytes, char* output, int len) {
+  const uint64_t ull10 = 10;
+  if (bytes >= ull10 << 40) {
+    return snprintf(output, len, "%" PRIu64 "TB", bytes >> 40);
+  } else if (bytes >= ull10 << 30) {
+    return snprintf(output, len, "%" PRIu64 "GB", bytes >> 30);
+  } else if (bytes >= ull10 << 20) {
+    return snprintf(output, len, "%" PRIu64 "MB", bytes >> 20);
+  } else if (bytes >= ull10 << 10) {
+    return snprintf(output, len, "%" PRIu64 "KB", bytes >> 10);
+  } else {
+    return snprintf(output, len, "%" PRIu64 "B", bytes);
+  }
+}
+
+void AppendNumberTo(std::string* str, uint64_t num) {
+  char buf[30];
+  snprintf(buf, sizeof(buf), "%" PRIu64, num);
+  str->append(buf);
+}
+
+void AppendEscapedStringTo(std::string* str, const Slice& value) {
+  for (size_t i = 0; i < value.size(); i++) {
+    char c = value[i];
+    if (c >= ' ' && c <= '~') {
+      str->push_back(c);
+    } else {
+      char buf[10];
+      snprintf(buf, sizeof(buf), "\\x%02x",
+               static_cast<unsigned int>(c) & 0xff);
+      str->append(buf);
+    }
+  }
+}
+
+std::string NumberToHumanString(int64_t num) {
+  char buf[19];
+  int64_t absnum = num < 0 ? -num : num;
+  if (absnum < 10000) {
+    snprintf(buf, sizeof(buf), "%" PRIi64, num);
+  } else if (absnum < 10000000) {
+    snprintf(buf, sizeof(buf), "%" PRIi64 "K", num / 1000);
+  } else if (absnum < 10000000000LL) {
+    snprintf(buf, sizeof(buf), "%" PRIi64 "M", num / 1000000);
+  } else {
+    snprintf(buf, sizeof(buf), "%" PRIi64 "G", num / 1000000000);
+  }
+  return std::string(buf);
+}
+
+std::string BytesToHumanString(uint64_t bytes) {
+  const char* size_name[] = {"KB", "MB", "GB", "TB"};
+  double final_size = static_cast<double>(bytes);
+  size_t size_idx;
+
+  // always start with KB
+  final_size /= 1024;
+  size_idx = 0;
+
+  while (size_idx < 3 && final_size >= 1024) {
+    final_size /= 1024;
+    size_idx++;
+  }
+
+  char buf[20];
+  snprintf(buf, sizeof(buf), "%.2f %s", final_size, size_name[size_idx]);
+  return std::string(buf);
+}
+
+std::string TimeToHumanString(int unixtime) {
+  char time_buffer[80];
+  time_t rawtime = unixtime;
+  struct tm tInfo;
+  struct tm* timeinfo = port::LocalTimeR(&rawtime, &tInfo);
+  assert(timeinfo == &tInfo);
+  strftime(time_buffer, 80, "%c", timeinfo);
+  return std::string(time_buffer);
+}
+
+std::string EscapeString(const Slice& value) {
+  std::string r;
+  AppendEscapedStringTo(&r, value);
+  return r;
+}
+
+bool ConsumeDecimalNumber(Slice* in, uint64_t* val) {
+  uint64_t v = 0;
+  int digits = 0;
+  while (!in->empty()) {
+    char c = (*in)[0];
+    if (c >= '0' && c <= '9') {
+      ++digits;
+      const unsigned int delta = (c - '0');
+      static const uint64_t kMaxUint64 = ~static_cast<uint64_t>(0);
+      if (v > kMaxUint64 / 10 ||
+          (v == kMaxUint64 / 10 && delta > kMaxUint64 % 10)) {
+        // Overflow
+        return false;
+      }
+      v = (v * 10) + delta;
+      in->remove_prefix(1);
+    } else {
+      break;
+    }
+  }
+  *val = v;
+  return (digits > 0);
+}
+
+bool isSpecialChar(const char c) {
+  if (c == '\\' || c == '#' || c == ':' || c == '\r' || c == '\n') {
+    return true;
+  }
+  return false;
+}
+
+namespace {
+using CharMap = std::pair<char, char>;
+}
+
+char UnescapeChar(const char c) {
+  static const CharMap convert_map[] = {{'r', '\r'}, {'n', '\n'}};
+
+  auto iter = std::find_if(std::begin(convert_map), std::end(convert_map),
+                           [c](const CharMap& p) { return p.first == c; });
+
+  if (iter == std::end(convert_map)) {
+    return c;
+  }
+  return iter->second;
+}
+
+char EscapeChar(const char c) {
+  static const CharMap convert_map[] = {{'\n', 'n'}, {'\r', 'r'}};
+
+  auto iter = std::find_if(std::begin(convert_map), std::end(convert_map),
+                           [c](const CharMap& p) { return p.first == c; });
+
+  if (iter == std::end(convert_map)) {
+    return c;
+  }
+  return iter->second;
+}
+
+std::string EscapeOptionString(const std::string& raw_string) {
+  std::string output;
+  for (auto c : raw_string) {
+    if (isSpecialChar(c)) {
+      output += '\\';
+      output += EscapeChar(c);
+    } else {
+      output += c;
+    }
+  }
+
+  return output;
+}
+
+std::string UnescapeOptionString(const std::string& escaped_string) {
+  bool escaped = false;
+  std::string output;
+
+  for (auto c : escaped_string) {
+    if (escaped) {
+      output += UnescapeChar(c);
+      escaped = false;
+    } else {
+      if (c == '\\') {
+        escaped = true;
+        continue;
+      }
+      output += c;
+    }
+  }
+  return output;
+}
+
+std::string trim(const std::string& str) {
+  if (str.empty()) return std::string();
+  size_t start = 0;
+  size_t end = str.size() - 1;
+  while (isspace(str[start]) != 0 && start < end) {
+    ++start;
+  }
+  while (isspace(str[end]) != 0 && start < end) {
+    --end;
+  }
+  if (start <= end) {
+    return str.substr(start, end - start + 1);
+  }
+  return std::string();
+}
+
+bool EndsWith(const std::string& string, const std::string& pattern) {
+  size_t plen = pattern.size();
+  size_t slen = string.size();
+  if (plen <= slen) {
+    return string.compare(slen - plen, plen, pattern) == 0;
+  } else {
+    return false;
+  }
+}
+
+bool StartsWith(const std::string& string, const std::string& pattern) {
+  return string.compare(0, pattern.size(), pattern) == 0;
+}
+
+#ifndef ROCKSDB_LITE
+
+bool ParseBoolean(const std::string& type, const std::string& value) {
+  if (value == "true" || value == "1") {
+    return true;
+  } else if (value == "false" || value == "0") {
+    return false;
+  }
+  throw std::invalid_argument(type);
+}
+
+uint8_t ParseUint8(const std::string& value) {
+  uint64_t num = ParseUint64(value);
+  if ((num >> 8LL) == 0) {
+    return static_cast<uint8_t>(num);
+  } else {
+    throw std::out_of_range(value);
+  }
+}
+
+uint32_t ParseUint32(const std::string& value) {
+  uint64_t num = ParseUint64(value);
+  if ((num >> 32LL) == 0) {
+    return static_cast<uint32_t>(num);
+  } else {
+    throw std::out_of_range(value);
+  }
+}
+
+int32_t ParseInt32(const std::string& value) {
+  int64_t num = ParseInt64(value);
+  if (num <= std::numeric_limits<int32_t>::max() &&
+      num >= std::numeric_limits<int32_t>::min()) {
+    return static_cast<int32_t>(num);
+  } else {
+    throw std::out_of_range(value);
+  }
+}
+
+#endif
+
+uint64_t ParseUint64(const std::string& value) {
+  size_t endchar;
+#ifndef CYGWIN
+  uint64_t num = std::stoull(value.c_str(), &endchar);
+#else
+  char* endptr;
+  uint64_t num = std::strtoul(value.c_str(), &endptr, 0);
+  endchar = endptr - value.c_str();
+#endif
+
+  if (endchar < value.length()) {
+    char c = value[endchar];
+    if (c == 'k' || c == 'K')
+      num <<= 10LL;
+    else if (c == 'm' || c == 'M')
+      num <<= 20LL;
+    else if (c == 'g' || c == 'G')
+      num <<= 30LL;
+    else if (c == 't' || c == 'T')
+      num <<= 40LL;
+  }
+
+  return num;
+}
+
+int64_t ParseInt64(const std::string& value) {
+  size_t endchar;
+#ifndef CYGWIN
+  int64_t num = std::stoll(value.c_str(), &endchar);
+#else
+  char* endptr;
+  int64_t num = std::strtoll(value.c_str(), &endptr, 0);
+  endchar = endptr - value.c_str();
+#endif
+
+  if (endchar < value.length()) {
+    char c = value[endchar];
+    if (c == 'k' || c == 'K')
+      num <<= 10LL;
+    else if (c == 'm' || c == 'M')
+      num <<= 20LL;
+    else if (c == 'g' || c == 'G')
+      num <<= 30LL;
+    else if (c == 't' || c == 'T')
+      num <<= 40LL;
+  }
+
+  return num;
+}
+
+int ParseInt(const std::string& value) {
+  size_t endchar;
+#ifndef CYGWIN
+  int num = std::stoi(value.c_str(), &endchar);
+#else
+  char* endptr;
+  int num = std::strtoul(value.c_str(), &endptr, 0);
+  endchar = endptr - value.c_str();
+#endif
+
+  if (endchar < value.length()) {
+    char c = value[endchar];
+    if (c == 'k' || c == 'K')
+      num <<= 10;
+    else if (c == 'm' || c == 'M')
+      num <<= 20;
+    else if (c == 'g' || c == 'G')
+      num <<= 30;
+  }
+
+  return num;
+}
+
+double ParseDouble(const std::string& value) {
+#ifndef CYGWIN
+  return std::stod(value);
+#else
+  return std::strtod(value.c_str(), 0);
+#endif
+}
+
+size_t ParseSizeT(const std::string& value) {
+  return static_cast<size_t>(ParseUint64(value));
+}
+
+std::vector<int> ParseVectorInt(const std::string& value) {
+  std::vector<int> result;
+  size_t start = 0;
+  while (start < value.size()) {
+    size_t end = value.find(':', start);
+    if (end == std::string::npos) {
+      result.push_back(ParseInt(value.substr(start)));
+      break;
+    } else {
+      result.push_back(ParseInt(value.substr(start, end - start)));
+      start = end + 1;
+    }
+  }
+  return result;
+}
+
+bool SerializeIntVector(const std::vector<int>& vec, std::string* value) {
+  *value = "";
+  for (size_t i = 0; i < vec.size(); ++i) {
+    if (i > 0) {
+      *value += ":";
+    }
+    *value += std::to_string(vec[i]);
+  }
+  return true;
+}
+
+// Copied from folly/string.cpp:
+// https://github.com/facebook/folly/blob/0deef031cb8aab76dc7e736f8b7c22d701d5f36b/folly/String.cpp#L457
+// There are two variants of `strerror_r` function, one returns
+// `int`, and another returns `char*`. Selecting proper version using
+// preprocessor macros portably is extremely hard.
+//
+// For example, on Android function signature depends on `__USE_GNU` and
+// `__ANDROID_API__` macros (https://git.io/fjBBE).
+//
+// So we are using C++ overloading trick: we pass a pointer of
+// `strerror_r` to `invoke_strerror_r` function, and C++ compiler
+// selects proper function.
+
+#if !(defined(_WIN32) && (defined(__MINGW32__) || defined(_MSC_VER)))
+ROCKSDB_MAYBE_UNUSED
+static std::string invoke_strerror_r(int (*strerror_r)(int, char*, size_t),
+                                     int err, char* buf, size_t buflen) {
+  // Using XSI-compatible strerror_r
+  int r = strerror_r(err, buf, buflen);
+
+  // OSX/FreeBSD use EINVAL and Linux uses -1 so just check for non-zero
+  if (r != 0) {
+    snprintf(buf, buflen, "Unknown error %d (strerror_r failed with error %d)",
+             err, errno);
+  }
+  return buf;
+}
+
+ROCKSDB_MAYBE_UNUSED
+static std::string invoke_strerror_r(char* (*strerror_r)(int, char*, size_t),
+                                     int err, char* buf, size_t buflen) {
+  // Using GNU strerror_r
+  return strerror_r(err, buf, buflen);
+}
+#endif  // !(defined(_WIN32) && (defined(__MINGW32__) || defined(_MSC_VER)))
+
+std::string errnoStr(int err) {
+  char buf[1024];
+  buf[0] = '\0';
+
+  std::string result;
+
+  // https://developer.apple.com/library/mac/documentation/Darwin/Reference/ManPages/man3/strerror_r.3.html
+  // http://www.kernel.org/doc/man-pages/online/pages/man3/strerror.3.html
+#if defined(_WIN32) && (defined(__MINGW32__) || defined(_MSC_VER))
+  // mingw64 has no strerror_r, but Windows has strerror_s, which C11 added
+  // as well. So maybe we should use this across all platforms (together
+  // with strerrorlen_s). Note strerror_r and _s have swapped args.
+  int r = strerror_s(buf, sizeof(buf), err);
+  if (r != 0) {
+    snprintf(buf, sizeof(buf),
+             "Unknown error %d (strerror_r failed with error %d)", err, errno);
+  }
+  result.assign(buf);
+#else
+  // Using any strerror_r
+  result.assign(invoke_strerror_r(strerror_r, err, buf, sizeof(buf)));
+#endif
+
+  return result;
+}
+
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/util/string_util.h b/src/rocksdb/util/string_util.h
new file mode 100644
index 000000000..11178fd1d
--- /dev/null
+++ b/src/rocksdb/util/string_util.h
@@ -0,0 +1,177 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+//
+
+#pragma once
+
+#include <cstdint>
+#include <sstream>
+#include <string>
+#include <unordered_map>
+#include <vector>
+
+#include "rocksdb/rocksdb_namespace.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+class Slice;
+
+extern std::vector<std::string> StringSplit(const std::string& arg, char delim);
+
+// Append a human-readable printout of "num" to *str
+extern void AppendNumberTo(std::string* str, uint64_t num);
+
+// Append a human-readable printout of "value" to *str.
+// Escapes any non-printable characters found in "value".
+extern void AppendEscapedStringTo(std::string* str, const Slice& value);
+
+// Put n digits from v in base kBase to (*buf)[0] to (*buf)[n-1] and
+// advance *buf to the position after what was written.
+template <size_t kBase>
+inline void PutBaseChars(char** buf, size_t n, uint64_t v, bool uppercase) {
+  const char* digitChars = uppercase ? "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
+                                     : "0123456789abcdefghijklmnopqrstuvwxyz";
+  for (size_t i = n; i > 0; --i) {
+    (*buf)[i - 1] = digitChars[static_cast<size_t>(v % kBase)];
+    v /= kBase;
+  }
+  *buf += n;
+}
+
+// Parse n digits from *buf in base kBase to *v and advance *buf to the
+// position after what was read. On success, true is returned. On failure,
+// false is returned, *buf is placed at the first bad character, and *v
+// contains the partial parsed data. Overflow is not checked but the
+// result is accurate mod 2^64. Requires the starting value of *v to be
+// zero or previously accumulated parsed digits, i.e.
+//   ParseBaseChars(&b, n, &v);
+// is equivalent to n calls to
+//   ParseBaseChars(&b, 1, &v);
+template <int kBase>
+inline bool ParseBaseChars(const char** buf, size_t n, uint64_t* v) {
+  while (n) {
+    char c = **buf;
+    *v *= static_cast<uint64_t>(kBase);
+    if (c >= '0' && (kBase >= 10 ? c <= '9' : c < '0' + kBase)) {
+      *v += static_cast<uint64_t>(c - '0');
+    } else if (kBase > 10 && c >= 'A' && c < 'A' + kBase - 10) {
+      *v += static_cast<uint64_t>(c - 'A' + 10);
+    } else if (kBase > 10 && c >= 'a' && c < 'a' + kBase - 10) {
+      *v += static_cast<uint64_t>(c - 'a' + 10);
+    } else {
+      return false;
+    }
+    --n;
+    ++*buf;
+  }
+  return true;
+}
+
+// Return a human-readable version of num.
+// for num >= 10.000, prints "xxK"
+// for num >= 10.000.000, prints "xxM"
+// for num >= 10.000.000.000, prints "xxG"
+extern std::string NumberToHumanString(int64_t num);
+
+// Return a human-readable version of bytes
+// ex: 1048576 -> 1.00 GB
+extern std::string BytesToHumanString(uint64_t bytes);
+
+// Return a human-readable version of unix time
+// ex: 1562116015 -> "Tue Jul  2 18:06:55 2019"
+extern std::string TimeToHumanString(int unixtime);
+
+// Append a human-readable time in micros.
+int AppendHumanMicros(uint64_t micros, char* output, int len,
+                      bool fixed_format);
+
+// Append a human-readable size in bytes
+int AppendHumanBytes(uint64_t bytes, char* output, int len);
+
+// Return a human-readable version of "value".
+// Escapes any non-printable characters found in "value".
+extern std::string EscapeString(const Slice& value);
+
+// Parse a human-readable number from "*in" into *value.  On success,
+// advances "*in" past the consumed number and sets "*val" to the
+// numeric value.  Otherwise, returns false and leaves *in in an
+// unspecified state.
+extern bool ConsumeDecimalNumber(Slice* in, uint64_t* val);
+
+// Returns true if the input char "c" is considered as a special character
+// that will be escaped when EscapeOptionString() is called.
+//
+// @param c the input char
+// @return true if the input char "c" is considered as a special character.
+// @see EscapeOptionString
+bool isSpecialChar(const char c);
+
+// If the input char is an escaped char, it will return the its
+// associated raw-char.  Otherwise, the function will simply return
+// the original input char.
+char UnescapeChar(const char c);
+
+// If the input char is a control char, it will return the its
+// associated escaped char.  Otherwise, the function will simply return
+// the original input char.
+char EscapeChar(const char c);
+
+// Converts a raw string to an escaped string.  Escaped-characters are
+// defined via the isSpecialChar() function.  When a char in the input
+// string "raw_string" is classified as a special characters, then it
+// will be prefixed by '\' in the output.
+//
+// It's inverse function is UnescapeOptionString().
+// @param raw_string the input string
+// @return the '\' escaped string of the input "raw_string"
+// @see isSpecialChar, UnescapeOptionString
+std::string EscapeOptionString(const std::string& raw_string);
+
+// The inverse function of EscapeOptionString.  It converts
+// an '\' escaped string back to a raw string.
+//
+// @param escaped_string the input '\' escaped string
+// @return the raw string of the input "escaped_string"
+std::string UnescapeOptionString(const std::string& escaped_string);
+
+std::string trim(const std::string& str);
+
+// Returns true if "string" ends with "pattern"
+bool EndsWith(const std::string& string, const std::string& pattern);
+
+// Returns true if "string" starts with "pattern"
+bool StartsWith(const std::string& string, const std::string& pattern);
+
+#ifndef ROCKSDB_LITE
+bool ParseBoolean(const std::string& type, const std::string& value);
+
+uint8_t ParseUint8(const std::string& value);
+
+uint32_t ParseUint32(const std::string& value);
+
+int32_t ParseInt32(const std::string& value);
+#endif
+
+uint64_t ParseUint64(const std::string& value);
+
+int ParseInt(const std::string& value);
+
+int64_t ParseInt64(const std::string& value);
+
+double ParseDouble(const std::string& value);
+
+size_t ParseSizeT(const std::string& value);
+
+std::vector<int> ParseVectorInt(const std::string& value);
+
+bool SerializeIntVector(const std::vector<int>& vec, std::string* value);
+
+extern const std::string kNullptrString;
+
+// errnoStr() function returns a string that describes the error code passed in
+// the argument err
+extern std::string errnoStr(int err);
+
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/util/thread_guard.h b/src/rocksdb/util/thread_guard.h
new file mode 100644
index 000000000..b2bb06a1b
--- /dev/null
+++ b/src/rocksdb/util/thread_guard.h
@@ -0,0 +1,41 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#pragma once
+
+#include "port/port.h"
+#include "rocksdb/rocksdb_namespace.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+// Resource management object for threads that joins the thread upon
+// destruction. Has unique ownership of the thread object, so copying it is not
+// allowed, while moving it transfers ownership.
+class ThreadGuard {
+ public:
+  ThreadGuard() = default;
+
+  explicit ThreadGuard(port::Thread&& thread) : thread_(std::move(thread)) {}
+
+  ThreadGuard(const ThreadGuard&) = delete;
+  ThreadGuard& operator=(const ThreadGuard&) = delete;
+
+  ThreadGuard(ThreadGuard&&) noexcept = default;
+  ThreadGuard& operator=(ThreadGuard&&) noexcept = default;
+
+  ~ThreadGuard() {
+    if (thread_.joinable()) {
+      thread_.join();
+    }
+  }
+
+  const port::Thread& GetThread() const { return thread_; }
+  port::Thread& GetThread() { return thread_; }
+
+ private:
+  port::Thread thread_;
+};
+
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/util/thread_list_test.cc b/src/rocksdb/util/thread_list_test.cc
new file mode 100644
index 000000000..af4e62355
--- /dev/null
+++ b/src/rocksdb/util/thread_list_test.cc
@@ -0,0 +1,360 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#include <condition_variable>
+#include <mutex>
+
+#include "monitoring/thread_status_updater.h"
+#include "rocksdb/db.h"
+#include "test_util/testharness.h"
+
+#ifdef ROCKSDB_USING_THREAD_STATUS
+
+namespace ROCKSDB_NAMESPACE {
+
+class SimulatedBackgroundTask {
+ public:
+  SimulatedBackgroundTask(
+      const void* db_key, const std::string& db_name, const void* cf_key,
+      const std::string& cf_name,
+      const ThreadStatus::OperationType operation_type =
+          ThreadStatus::OP_UNKNOWN,
+      const ThreadStatus::StateType state_type = ThreadStatus::STATE_UNKNOWN)
+      : db_key_(db_key),
+        db_name_(db_name),
+        cf_key_(cf_key),
+        cf_name_(cf_name),
+        operation_type_(operation_type),
+        state_type_(state_type),
+        should_run_(true),
+        running_count_(0) {
+    Env::Default()->GetThreadStatusUpdater()->NewColumnFamilyInfo(
+        db_key_, db_name_, cf_key_, cf_name_);
+  }
+
+  ~SimulatedBackgroundTask() {
+    Env::Default()->GetThreadStatusUpdater()->EraseDatabaseInfo(db_key_);
+  }
+
+  void Run() {
+    std::unique_lock<std::mutex> l(mutex_);
+    running_count_++;
+    bg_cv_.notify_all();
+    Env::Default()->GetThreadStatusUpdater()->SetColumnFamilyInfoKey(cf_key_);
+    Env::Default()->GetThreadStatusUpdater()->SetThreadOperation(
+        operation_type_);
+    Env::Default()->GetThreadStatusUpdater()->SetThreadState(state_type_);
+    while (should_run_) {
+      bg_cv_.wait(l);
+    }
+    Env::Default()->GetThreadStatusUpdater()->ClearThreadState();
+    Env::Default()->GetThreadStatusUpdater()->ClearThreadOperation();
+    Env::Default()->GetThreadStatusUpdater()->SetColumnFamilyInfoKey(nullptr);
+    running_count_--;
+    bg_cv_.notify_all();
+  }
+
+  void FinishAllTasks() {
+    std::unique_lock<std::mutex> l(mutex_);
+    should_run_ = false;
+    bg_cv_.notify_all();
+  }
+
+  void WaitUntilScheduled(int job_count) {
+    std::unique_lock<std::mutex> l(mutex_);
+    while (running_count_ < job_count) {
+      bg_cv_.wait(l);
+    }
+  }
+
+  void WaitUntilDone() {
+    std::unique_lock<std::mutex> l(mutex_);
+    while (running_count_ > 0) {
+      bg_cv_.wait(l);
+    }
+  }
+
+  static void DoSimulatedTask(void* arg) {
+    reinterpret_cast<SimulatedBackgroundTask*>(arg)->Run();
+  }
+
+ private:
+  const void* db_key_;
+  const std::string db_name_;
+  const void* cf_key_;
+  const std::string cf_name_;
+  const ThreadStatus::OperationType operation_type_;
+  const ThreadStatus::StateType state_type_;
+  std::mutex mutex_;
+  std::condition_variable bg_cv_;
+  bool should_run_;
+  std::atomic<int> running_count_;
+};
+
+class ThreadListTest : public testing::Test {
+ public:
+  ThreadListTest() {}
+};
+
+TEST_F(ThreadListTest, GlobalTables) {
+  // verify the global tables for operations and states are properly indexed.
+  for (int type = 0; type != ThreadStatus::NUM_OP_TYPES; ++type) {
+    ASSERT_EQ(global_operation_table[type].type, type);
+    ASSERT_EQ(
+        global_operation_table[type].name,
+        ThreadStatus::GetOperationName(ThreadStatus::OperationType(type)));
+  }
+
+  for (int type = 0; type != ThreadStatus::NUM_STATE_TYPES; ++type) {
+    ASSERT_EQ(global_state_table[type].type, type);
+    ASSERT_EQ(global_state_table[type].name,
+              ThreadStatus::GetStateName(ThreadStatus::StateType(type)));
+  }
+
+  for (int stage = 0; stage != ThreadStatus::NUM_OP_STAGES; ++stage) {
+    ASSERT_EQ(global_op_stage_table[stage].stage, stage);
+    ASSERT_EQ(global_op_stage_table[stage].name,
+              ThreadStatus::GetOperationStageName(
+                  ThreadStatus::OperationStage(stage)));
+  }
+}
+
+TEST_F(ThreadListTest, SimpleColumnFamilyInfoTest) {
+  Env* env = Env::Default();
+  const int kHighPriorityThreads = 3;
+  const int kLowPriorityThreads = 5;
+  const int kSimulatedHighPriThreads = kHighPriorityThreads - 1;
+  const int kSimulatedLowPriThreads = kLowPriorityThreads / 3;
+  const int kDelayMicros = 1000000;
+  env->SetBackgroundThreads(kHighPriorityThreads, Env::HIGH);
+  env->SetBackgroundThreads(kLowPriorityThreads, Env::LOW);
+  // Wait 1 second so that threads start
+  Env::Default()->SleepForMicroseconds(kDelayMicros);
+  SimulatedBackgroundTask running_task(reinterpret_cast<void*>(1234), "running",
+                                       reinterpret_cast<void*>(5678),
+                                       "pikachu");
+
+  for (int test = 0; test < kSimulatedHighPriThreads; ++test) {
+    env->Schedule(&SimulatedBackgroundTask::DoSimulatedTask, &running_task,
+                  Env::Priority::HIGH);
+  }
+
+  for (int test = 0; test < kSimulatedLowPriThreads; ++test) {
+    env->Schedule(&SimulatedBackgroundTask::DoSimulatedTask, &running_task,
+                  Env::Priority::LOW);
+  }
+  running_task.WaitUntilScheduled(kSimulatedHighPriThreads +
+                                  kSimulatedLowPriThreads);
+  // We can only reserve limited number of waiting threads
+  ASSERT_EQ(kHighPriorityThreads - kSimulatedHighPriThreads,
+            env->ReserveThreads(kHighPriorityThreads, Env::Priority::HIGH));
+  ASSERT_EQ(kLowPriorityThreads - kSimulatedLowPriThreads,
+            env->ReserveThreads(kLowPriorityThreads, Env::Priority::LOW));
+
+  // Reservation shall not affect the existing thread list
+  std::vector<ThreadStatus> thread_list;
+
+  // Verify the number of running threads in each pool.
+  ASSERT_OK(env->GetThreadList(&thread_list));
+  int running_count[ThreadStatus::NUM_THREAD_TYPES] = {0};
+  for (auto thread_status : thread_list) {
+    if (thread_status.cf_name == "pikachu" &&
+        thread_status.db_name == "running") {
+      running_count[thread_status.thread_type]++;
+    }
+  }
+  // Cannot reserve more threads
+  ASSERT_EQ(0, env->ReserveThreads(kHighPriorityThreads, Env::Priority::HIGH));
+  ASSERT_EQ(0, env->ReserveThreads(kLowPriorityThreads, Env::Priority::LOW));
+
+  ASSERT_EQ(running_count[ThreadStatus::HIGH_PRIORITY],
+            kSimulatedHighPriThreads);
+  ASSERT_EQ(running_count[ThreadStatus::LOW_PRIORITY], kSimulatedLowPriThreads);
+  ASSERT_EQ(running_count[ThreadStatus::USER], 0);
+
+  running_task.FinishAllTasks();
+  running_task.WaitUntilDone();
+
+  ASSERT_EQ(kHighPriorityThreads - kSimulatedHighPriThreads,
+            env->ReleaseThreads(kHighPriorityThreads, Env::Priority::HIGH));
+  ASSERT_EQ(kLowPriorityThreads - kSimulatedLowPriThreads,
+            env->ReleaseThreads(kLowPriorityThreads, Env::Priority::LOW));
+  // Verify none of the threads are running
+  ASSERT_OK(env->GetThreadList(&thread_list));
+
+  for (int i = 0; i < ThreadStatus::NUM_THREAD_TYPES; ++i) {
+    running_count[i] = 0;
+  }
+  for (auto thread_status : thread_list) {
+    if (thread_status.cf_name == "pikachu" &&
+        thread_status.db_name == "running") {
+      running_count[thread_status.thread_type]++;
+    }
+  }
+
+  ASSERT_EQ(running_count[ThreadStatus::HIGH_PRIORITY], 0);
+  ASSERT_EQ(running_count[ThreadStatus::LOW_PRIORITY], 0);
+  ASSERT_EQ(running_count[ThreadStatus::USER], 0);
+}
+
+namespace {
+void UpdateStatusCounts(const std::vector<ThreadStatus>& thread_list,
+                        int operation_counts[], int state_counts[]) {
+  for (auto thread_status : thread_list) {
+    operation_counts[thread_status.operation_type]++;
+    state_counts[thread_status.state_type]++;
+  }
+}
+
+void VerifyAndResetCounts(const int correct_counts[], int collected_counts[],
+                          int size) {
+  for (int i = 0; i < size; ++i) {
+    ASSERT_EQ(collected_counts[i], correct_counts[i]);
+    collected_counts[i] = 0;
+  }
+}
+
+void UpdateCount(int operation_counts[], int from_event, int to_event,
+                 int amount) {
+  operation_counts[from_event] -= amount;
+  operation_counts[to_event] += amount;
+}
+}  // namespace
+
+TEST_F(ThreadListTest, SimpleEventTest) {
+  Env* env = Env::Default();
+
+  // simulated tasks
+  const int kFlushWriteTasks = 3;
+  SimulatedBackgroundTask flush_write_task(
+      reinterpret_cast<void*>(1234), "running", reinterpret_cast<void*>(5678),
+      "pikachu", ThreadStatus::OP_FLUSH);
+
+  const int kCompactionWriteTasks = 4;
+  SimulatedBackgroundTask compaction_write_task(
+      reinterpret_cast<void*>(1234), "running", reinterpret_cast<void*>(5678),
+      "pikachu", ThreadStatus::OP_COMPACTION);
+
+  const int kCompactionReadTasks = 5;
+  SimulatedBackgroundTask compaction_read_task(
+      reinterpret_cast<void*>(1234), "running", reinterpret_cast<void*>(5678),
+      "pikachu", ThreadStatus::OP_COMPACTION);
+
+  const int kCompactionWaitTasks = 6;
+  SimulatedBackgroundTask compaction_wait_task(
+      reinterpret_cast<void*>(1234), "running", reinterpret_cast<void*>(5678),
+      "pikachu", ThreadStatus::OP_COMPACTION);
+
+  // setup right answers
+  int correct_operation_counts[ThreadStatus::NUM_OP_TYPES] = {0};
+  correct_operation_counts[ThreadStatus::OP_FLUSH] = kFlushWriteTasks;
+  correct_operation_counts[ThreadStatus::OP_COMPACTION] =
+      kCompactionWriteTasks + kCompactionReadTasks + kCompactionWaitTasks;
+
+  env->SetBackgroundThreads(correct_operation_counts[ThreadStatus::OP_FLUSH],
+                            Env::HIGH);
+  env->SetBackgroundThreads(
+      correct_operation_counts[ThreadStatus::OP_COMPACTION], Env::LOW);
+
+  // schedule the simulated tasks
+  for (int t = 0; t < kFlushWriteTasks; ++t) {
+    env->Schedule(&SimulatedBackgroundTask::DoSimulatedTask, &flush_write_task,
+                  Env::Priority::HIGH);
+  }
+  flush_write_task.WaitUntilScheduled(kFlushWriteTasks);
+
+  for (int t = 0; t < kCompactionWriteTasks; ++t) {
+    env->Schedule(&SimulatedBackgroundTask::DoSimulatedTask,
+                  &compaction_write_task, Env::Priority::LOW);
+  }
+  compaction_write_task.WaitUntilScheduled(kCompactionWriteTasks);
+
+  for (int t = 0; t < kCompactionReadTasks; ++t) {
+    env->Schedule(&SimulatedBackgroundTask::DoSimulatedTask,
+                  &compaction_read_task, Env::Priority::LOW);
+  }
+  compaction_read_task.WaitUntilScheduled(kCompactionReadTasks);
+
+  for (int t = 0; t < kCompactionWaitTasks; ++t) {
+    env->Schedule(&SimulatedBackgroundTask::DoSimulatedTask,
+                  &compaction_wait_task, Env::Priority::LOW);
+  }
+  compaction_wait_task.WaitUntilScheduled(kCompactionWaitTasks);
+
+  // verify the thread-status
+  int operation_counts[ThreadStatus::NUM_OP_TYPES] = {0};
+  int state_counts[ThreadStatus::NUM_STATE_TYPES] = {0};
+
+  std::vector<ThreadStatus> thread_list;
+  ASSERT_OK(env->GetThreadList(&thread_list));
+  UpdateStatusCounts(thread_list, operation_counts, state_counts);
+  VerifyAndResetCounts(correct_operation_counts, operation_counts,
+                       ThreadStatus::NUM_OP_TYPES);
+
+  // terminate compaction-wait tasks and see if the thread-status
+  // reflects this update
+  compaction_wait_task.FinishAllTasks();
+  compaction_wait_task.WaitUntilDone();
+  UpdateCount(correct_operation_counts, ThreadStatus::OP_COMPACTION,
+              ThreadStatus::OP_UNKNOWN, kCompactionWaitTasks);
+
+  ASSERT_OK(env->GetThreadList(&thread_list));
+  UpdateStatusCounts(thread_list, operation_counts, state_counts);
+  VerifyAndResetCounts(correct_operation_counts, operation_counts,
+                       ThreadStatus::NUM_OP_TYPES);
+
+  // terminate flush-write tasks and see if the thread-status
+  // reflects this update
+  flush_write_task.FinishAllTasks();
+  flush_write_task.WaitUntilDone();
+  UpdateCount(correct_operation_counts, ThreadStatus::OP_FLUSH,
+              ThreadStatus::OP_UNKNOWN, kFlushWriteTasks);
+
+  ASSERT_OK(env->GetThreadList(&thread_list));
+  UpdateStatusCounts(thread_list, operation_counts, state_counts);
+  VerifyAndResetCounts(correct_operation_counts, operation_counts,
+                       ThreadStatus::NUM_OP_TYPES);
+
+  // terminate compaction-write tasks and see if the thread-status
+  // reflects this update
+  compaction_write_task.FinishAllTasks();
+  compaction_write_task.WaitUntilDone();
+  UpdateCount(correct_operation_counts, ThreadStatus::OP_COMPACTION,
+              ThreadStatus::OP_UNKNOWN, kCompactionWriteTasks);
+
+  ASSERT_OK(env->GetThreadList(&thread_list));
+  UpdateStatusCounts(thread_list, operation_counts, state_counts);
+  VerifyAndResetCounts(correct_operation_counts, operation_counts,
+                       ThreadStatus::NUM_OP_TYPES);
+
+  // terminate compaction-write tasks and see if the thread-status
+  // reflects this update
+  compaction_read_task.FinishAllTasks();
+  compaction_read_task.WaitUntilDone();
+  UpdateCount(correct_operation_counts, ThreadStatus::OP_COMPACTION,
+              ThreadStatus::OP_UNKNOWN, kCompactionReadTasks);
+
+  ASSERT_OK(env->GetThreadList(&thread_list));
+  UpdateStatusCounts(thread_list, operation_counts, state_counts);
+  VerifyAndResetCounts(correct_operation_counts, operation_counts,
+                       ThreadStatus::NUM_OP_TYPES);
+}
+
+}  // namespace ROCKSDB_NAMESPACE
+
+int main(int argc, char** argv) {
+  ROCKSDB_NAMESPACE::port::InstallStackTraceHandler();
+  ::testing::InitGoogleTest(&argc, argv);
+  return RUN_ALL_TESTS();
+}
+
+#else
+
+int main(int argc, char** argv) {
+  ::testing::InitGoogleTest(&argc, argv);
+  return 0;
+}
+
+#endif  // ROCKSDB_USING_THREAD_STATUS
diff --git a/src/rocksdb/util/thread_local.cc b/src/rocksdb/util/thread_local.cc
new file mode 100644
index 000000000..969639d9b
--- /dev/null
+++ b/src/rocksdb/util/thread_local.cc
@@ -0,0 +1,521 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+//
+// Copyright (c) 2011 The LevelDB Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file. See the AUTHORS file for names of contributors.
+
+#include "util/thread_local.h"
+
+#include <stdlib.h>
+
+#include "port/likely.h"
+#include "util/mutexlock.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+struct Entry {
+  Entry() : ptr(nullptr) {}
+  Entry(const Entry& e) : ptr(e.ptr.load(std::memory_order_relaxed)) {}
+  std::atomic<void*> ptr;
+};
+
+class StaticMeta;
+
+// This is the structure that is declared as "thread_local" storage.
+// The vector keep list of atomic pointer for all instances for "current"
+// thread. The vector is indexed by an Id that is unique in process and
+// associated with one ThreadLocalPtr instance. The Id is assigned by a
+// global StaticMeta singleton. So if we instantiated 3 ThreadLocalPtr
+// instances, each thread will have a ThreadData with a vector of size 3:
+//     ---------------------------------------------------
+//     |          | instance 1 | instance 2 | instance 3 |
+//     ---------------------------------------------------
+//     | thread 1 |    void*   |    void*   |    void*   | <- ThreadData
+//     ---------------------------------------------------
+//     | thread 2 |    void*   |    void*   |    void*   | <- ThreadData
+//     ---------------------------------------------------
+//     | thread 3 |    void*   |    void*   |    void*   | <- ThreadData
+//     ---------------------------------------------------
+struct ThreadData {
+  explicit ThreadData(ThreadLocalPtr::StaticMeta* _inst)
+      : entries(), next(nullptr), prev(nullptr), inst(_inst) {}
+  std::vector<Entry> entries;
+  ThreadData* next;
+  ThreadData* prev;
+  ThreadLocalPtr::StaticMeta* inst;
+};
+
+class ThreadLocalPtr::StaticMeta {
+ public:
+  StaticMeta();
+
+  // Return the next available Id
+  uint32_t GetId();
+  // Return the next available Id without claiming it
+  uint32_t PeekId() const;
+  // Return the given Id back to the free pool. This also triggers
+  // UnrefHandler for associated pointer value (if not NULL) for all threads.
+  void ReclaimId(uint32_t id);
+
+  // Return the pointer value for the given id for the current thread.
+  void* Get(uint32_t id) const;
+  // Reset the pointer value for the given id for the current thread.
+  void Reset(uint32_t id, void* ptr);
+  // Atomically swap the supplied ptr and return the previous value
+  void* Swap(uint32_t id, void* ptr);
+  // Atomically compare and swap the provided value only if it equals
+  // to expected value.
+  bool CompareAndSwap(uint32_t id, void* ptr, void*& expected);
+  // Reset all thread local data to replacement, and return non-nullptr
+  // data for all existing threads
+  void Scrape(uint32_t id, autovector<void*>* ptrs, void* const replacement);
+  // Update res by applying func on each thread-local value. Holds a lock that
+  // prevents unref handler from running during this call, but clients must
+  // still provide external synchronization since the owning thread can
+  // access the values without internal locking, e.g., via Get() and Reset().
+  void Fold(uint32_t id, FoldFunc func, void* res);
+
+  // Register the UnrefHandler for id
+  void SetHandler(uint32_t id, UnrefHandler handler);
+
+  // protect inst, next_instance_id_, free_instance_ids_, head_,
+  // ThreadData.entries
+  //
+  // Note that here we prefer function static variable instead of the usual
+  // global static variable.  The reason is that c++ destruction order of
+  // static variables in the reverse order of their construction order.
+  // However, C++ does not guarantee any construction order when global
+  // static variables are defined in different files, while the function
+  // static variables are initialized when their function are first called.
+  // As a result, the construction order of the function static variables
+  // can be controlled by properly invoke their first function calls in
+  // the right order.
+  //
+  // For instance, the following function contains a function static
+  // variable.  We place a dummy function call of this inside
+  // Env::Default() to ensure the construction order of the construction
+  // order.
+  static port::Mutex* Mutex();
+
+  // Returns the member mutex of the current StaticMeta.  In general,
+  // Mutex() should be used instead of this one.  However, in case where
+  // the static variable inside Instance() goes out of scope, MemberMutex()
+  // should be used.  One example is OnThreadExit() function.
+  port::Mutex* MemberMutex() { return &mutex_; }
+
+ private:
+  // Get UnrefHandler for id with acquiring mutex
+  // REQUIRES: mutex locked
+  UnrefHandler GetHandler(uint32_t id);
+
+  // Triggered before a thread terminates
+  static void OnThreadExit(void* ptr);
+
+  // Add current thread's ThreadData to the global chain
+  // REQUIRES: mutex locked
+  void AddThreadData(ThreadData* d);
+
+  // Remove current thread's ThreadData from the global chain
+  // REQUIRES: mutex locked
+  void RemoveThreadData(ThreadData* d);
+
+  static ThreadData* GetThreadLocal();
+
+  uint32_t next_instance_id_;
+  // Used to recycle Ids in case ThreadLocalPtr is instantiated and destroyed
+  // frequently. This also prevents it from blowing up the vector space.
+  autovector<uint32_t> free_instance_ids_;
+  // Chain all thread local structure together. This is necessary since
+  // when one ThreadLocalPtr gets destroyed, we need to loop over each
+  // thread's version of pointer corresponding to that instance and
+  // call UnrefHandler for it.
+  ThreadData head_;
+
+  std::unordered_map<uint32_t, UnrefHandler> handler_map_;
+
+  // The private mutex.  Developers should always use Mutex() instead of
+  // using this variable directly.
+  port::Mutex mutex_;
+  // Thread local storage
+  static thread_local ThreadData* tls_;
+
+  // Used to make thread exit trigger possible if !defined(OS_MACOSX).
+  // Otherwise, used to retrieve thread data.
+  pthread_key_t pthread_key_;
+};
+
+thread_local ThreadData* ThreadLocalPtr::StaticMeta::tls_ = nullptr;
+
+// Windows doesn't support a per-thread destructor with its
+// TLS primitives.  So, we build it manually by inserting a
+// function to be called on each thread's exit.
+// See http://www.codeproject.com/Articles/8113/Thread-Local-Storage-The-C-Way
+// and http://www.nynaeve.net/?p=183
+//
+// really we do this to have clear conscience since using TLS with thread-pools
+// is iffy
+// although OK within a request. But otherwise, threads have no identity in its
+// modern use.
+
+// This runs on windows only called from the System Loader
+#ifdef OS_WIN
+
+// Windows cleanup routine is invoked from a System Loader with a different
+// signature so we can not directly hookup the original OnThreadExit which is
+// private member
+// so we make StaticMeta class share with the us the address of the function so
+// we can invoke it.
+namespace wintlscleanup {
+
+// This is set to OnThreadExit in StaticMeta singleton constructor
+UnrefHandler thread_local_inclass_routine = nullptr;
+pthread_key_t thread_local_key = pthread_key_t(-1);
+
+// Static callback function to call with each thread termination.
+void NTAPI WinOnThreadExit(PVOID module, DWORD reason, PVOID reserved) {
+  // We decided to punt on PROCESS_EXIT
+  if (DLL_THREAD_DETACH == reason) {
+    if (thread_local_key != pthread_key_t(-1) &&
+        thread_local_inclass_routine != nullptr) {
+      void* tls = TlsGetValue(thread_local_key);
+      if (tls != nullptr) {
+        thread_local_inclass_routine(tls);
+      }
+    }
+  }
+}
+
+}  // namespace wintlscleanup
+
+// extern "C" suppresses C++ name mangling so we know the symbol name for the
+// linker /INCLUDE:symbol pragma above.
+extern "C" {
+
+#ifdef _MSC_VER
+// The linker must not discard thread_callback_on_exit.  (We force a reference
+// to this variable with a linker /include:symbol pragma to ensure that.) If
+// this variable is discarded, the OnThreadExit function will never be called.
+#ifndef _X86_
+
+// .CRT section is merged with .rdata on x64 so it must be constant data.
+#pragma const_seg(".CRT$XLB")
+// When defining a const variable, it must have external linkage to be sure the
+// linker doesn't discard it.
+extern const PIMAGE_TLS_CALLBACK p_thread_callback_on_exit;
+const PIMAGE_TLS_CALLBACK p_thread_callback_on_exit =
+    wintlscleanup::WinOnThreadExit;
+// Reset the default section.
+#pragma const_seg()
+
+#pragma comment(linker, "/include:_tls_used")
+#pragma comment(linker, "/include:p_thread_callback_on_exit")
+
+#else  // _X86_
+
+#pragma data_seg(".CRT$XLB")
+PIMAGE_TLS_CALLBACK p_thread_callback_on_exit = wintlscleanup::WinOnThreadExit;
+// Reset the default section.
+#pragma data_seg()
+
+#pragma comment(linker, "/INCLUDE:__tls_used")
+#pragma comment(linker, "/INCLUDE:_p_thread_callback_on_exit")
+
+#endif  // _X86_
+
+#else
+// https://github.com/couchbase/gperftools/blob/master/src/windows/port.cc
+BOOL WINAPI DllMain(HINSTANCE h, DWORD dwReason, PVOID pv) {
+  if (dwReason == DLL_THREAD_DETACH)
+    wintlscleanup::WinOnThreadExit(h, dwReason, pv);
+  return TRUE;
+}
+#endif
+}  // extern "C"
+
+#endif  // OS_WIN
+
+void ThreadLocalPtr::InitSingletons() { ThreadLocalPtr::Instance(); }
+
+ThreadLocalPtr::StaticMeta* ThreadLocalPtr::Instance() {
+  // Here we prefer function static variable instead of global
+  // static variable as function static variable is initialized
+  // when the function is first call.  As a result, we can properly
+  // control their construction order by properly preparing their
+  // first function call.
+  //
+  // Note that here we decide to make "inst" a static pointer w/o deleting
+  // it at the end instead of a static variable.  This is to avoid the following
+  // destruction order disaster happens when a child thread using ThreadLocalPtr
+  // dies AFTER the main thread dies:  When a child thread happens to use
+  // ThreadLocalPtr, it will try to delete its thread-local data on its
+  // OnThreadExit when the child thread dies.  However, OnThreadExit depends
+  // on the following variable.  As a result, if the main thread dies before any
+  // child thread happen to use ThreadLocalPtr dies, then the destruction of
+  // the following variable will go first, then OnThreadExit, therefore causing
+  // invalid access.
+  //
+  // The above problem can be solved by using thread_local to store tls_.
+  // thread_local supports dynamic construction and destruction of
+  // non-primitive typed variables.  As a result, we can guarantee the
+  // destruction order even when the main thread dies before any child threads.
+  static ThreadLocalPtr::StaticMeta* inst = new ThreadLocalPtr::StaticMeta();
+  return inst;
+}
+
+port::Mutex* ThreadLocalPtr::StaticMeta::Mutex() { return &Instance()->mutex_; }
+
+void ThreadLocalPtr::StaticMeta::OnThreadExit(void* ptr) {
+  auto* tls = static_cast<ThreadData*>(ptr);
+  assert(tls != nullptr);
+
+  // Use the cached StaticMeta::Instance() instead of directly calling
+  // the variable inside StaticMeta::Instance() might already go out of
+  // scope here in case this OnThreadExit is called after the main thread
+  // dies.
+  auto* inst = tls->inst;
+  pthread_setspecific(inst->pthread_key_, nullptr);
+
+  MutexLock l(inst->MemberMutex());
+  inst->RemoveThreadData(tls);
+  // Unref stored pointers of current thread from all instances
+  uint32_t id = 0;
+  for (auto& e : tls->entries) {
+    void* raw = e.ptr.load();
+    if (raw != nullptr) {
+      auto unref = inst->GetHandler(id);
+      if (unref != nullptr) {
+        unref(raw);
+      }
+    }
+    ++id;
+  }
+  // Delete thread local structure no matter if it is Mac platform
+  delete tls;
+}
+
+ThreadLocalPtr::StaticMeta::StaticMeta()
+    : next_instance_id_(0), head_(this), pthread_key_(0) {
+  if (pthread_key_create(&pthread_key_, &OnThreadExit) != 0) {
+    abort();
+  }
+
+  // OnThreadExit is not getting called on the main thread.
+  // Call through the static destructor mechanism to avoid memory leak.
+  //
+  // Caveats: ~A() will be invoked _after_ ~StaticMeta for the global
+  // singleton (destructors are invoked in reverse order of constructor
+  // _completion_); the latter must not mutate internal members. This
+  // cleanup mechanism inherently relies on use-after-release of the
+  // StaticMeta, and is brittle with respect to compiler-specific handling
+  // of memory backing destructed statically-scoped objects. Perhaps
+  // registering with atexit(3) would be more robust.
+  //
+// This is not required on Windows.
+#if !defined(OS_WIN)
+  static struct A {
+    ~A() {
+      if (tls_) {
+        OnThreadExit(tls_);
+      }
+    }
+  } a;
+#endif  // !defined(OS_WIN)
+
+  head_.next = &head_;
+  head_.prev = &head_;
+
+#ifdef OS_WIN
+  // Share with Windows its cleanup routine and the key
+  wintlscleanup::thread_local_inclass_routine = OnThreadExit;
+  wintlscleanup::thread_local_key = pthread_key_;
+#endif
+}
+
+void ThreadLocalPtr::StaticMeta::AddThreadData(ThreadData* d) {
+  Mutex()->AssertHeld();
+  d->next = &head_;
+  d->prev = head_.prev;
+  head_.prev->next = d;
+  head_.prev = d;
+}
+
+void ThreadLocalPtr::StaticMeta::RemoveThreadData(ThreadData* d) {
+  Mutex()->AssertHeld();
+  d->next->prev = d->prev;
+  d->prev->next = d->next;
+  d->next = d->prev = d;
+}
+
+ThreadData* ThreadLocalPtr::StaticMeta::GetThreadLocal() {
+  if (UNLIKELY(tls_ == nullptr)) {
+    auto* inst = Instance();
+    tls_ = new ThreadData(inst);
+    {
+      // Register it in the global chain, needs to be done before thread exit
+      // handler registration
+      MutexLock l(Mutex());
+      inst->AddThreadData(tls_);
+    }
+    // Even it is not OS_MACOSX, need to register value for pthread_key_ so that
+    // its exit handler will be triggered.
+    if (pthread_setspecific(inst->pthread_key_, tls_) != 0) {
+      {
+        MutexLock l(Mutex());
+        inst->RemoveThreadData(tls_);
+      }
+      delete tls_;
+      abort();
+    }
+  }
+  return tls_;
+}
+
+void* ThreadLocalPtr::StaticMeta::Get(uint32_t id) const {
+  auto* tls = GetThreadLocal();
+  if (UNLIKELY(id >= tls->entries.size())) {
+    return nullptr;
+  }
+  return tls->entries[id].ptr.load(std::memory_order_acquire);
+}
+
+void ThreadLocalPtr::StaticMeta::Reset(uint32_t id, void* ptr) {
+  auto* tls = GetThreadLocal();
+  if (UNLIKELY(id >= tls->entries.size())) {
+    // Need mutex to protect entries access within ReclaimId
+    MutexLock l(Mutex());
+    tls->entries.resize(id + 1);
+  }
+  tls->entries[id].ptr.store(ptr, std::memory_order_release);
+}
+
+void* ThreadLocalPtr::StaticMeta::Swap(uint32_t id, void* ptr) {
+  auto* tls = GetThreadLocal();
+  if (UNLIKELY(id >= tls->entries.size())) {
+    // Need mutex to protect entries access within ReclaimId
+    MutexLock l(Mutex());
+    tls->entries.resize(id + 1);
+  }
+  return tls->entries[id].ptr.exchange(ptr, std::memory_order_acquire);
+}
+
+bool ThreadLocalPtr::StaticMeta::CompareAndSwap(uint32_t id, void* ptr,
+                                                void*& expected) {
+  auto* tls = GetThreadLocal();
+  if (UNLIKELY(id >= tls->entries.size())) {
+    // Need mutex to protect entries access within ReclaimId
+    MutexLock l(Mutex());
+    tls->entries.resize(id + 1);
+  }
+  return tls->entries[id].ptr.compare_exchange_strong(
+      expected, ptr, std::memory_order_release, std::memory_order_relaxed);
+}
+
+void ThreadLocalPtr::StaticMeta::Scrape(uint32_t id, autovector<void*>* ptrs,
+                                        void* const replacement) {
+  MutexLock l(Mutex());
+  for (ThreadData* t = head_.next; t != &head_; t = t->next) {
+    if (id < t->entries.size()) {
+      void* ptr =
+          t->entries[id].ptr.exchange(replacement, std::memory_order_acquire);
+      if (ptr != nullptr) {
+        ptrs->push_back(ptr);
+      }
+    }
+  }
+}
+
+void ThreadLocalPtr::StaticMeta::Fold(uint32_t id, FoldFunc func, void* res) {
+  MutexLock l(Mutex());
+  for (ThreadData* t = head_.next; t != &head_; t = t->next) {
+    if (id < t->entries.size()) {
+      void* ptr = t->entries[id].ptr.load();
+      if (ptr != nullptr) {
+        func(ptr, res);
+      }
+    }
+  }
+}
+
+uint32_t ThreadLocalPtr::TEST_PeekId() { return Instance()->PeekId(); }
+
+void ThreadLocalPtr::StaticMeta::SetHandler(uint32_t id, UnrefHandler handler) {
+  MutexLock l(Mutex());
+  handler_map_[id] = handler;
+}
+
+UnrefHandler ThreadLocalPtr::StaticMeta::GetHandler(uint32_t id) {
+  Mutex()->AssertHeld();
+  auto iter = handler_map_.find(id);
+  if (iter == handler_map_.end()) {
+    return nullptr;
+  }
+  return iter->second;
+}
+
+uint32_t ThreadLocalPtr::StaticMeta::GetId() {
+  MutexLock l(Mutex());
+  if (free_instance_ids_.empty()) {
+    return next_instance_id_++;
+  }
+
+  uint32_t id = free_instance_ids_.back();
+  free_instance_ids_.pop_back();
+  return id;
+}
+
+uint32_t ThreadLocalPtr::StaticMeta::PeekId() const {
+  MutexLock l(Mutex());
+  if (!free_instance_ids_.empty()) {
+    return free_instance_ids_.back();
+  }
+  return next_instance_id_;
+}
+
+void ThreadLocalPtr::StaticMeta::ReclaimId(uint32_t id) {
+  // This id is not used, go through all thread local data and release
+  // corresponding value
+  MutexLock l(Mutex());
+  auto unref = GetHandler(id);
+  for (ThreadData* t = head_.next; t != &head_; t = t->next) {
+    if (id < t->entries.size()) {
+      void* ptr = t->entries[id].ptr.exchange(nullptr);
+      if (ptr != nullptr && unref != nullptr) {
+        unref(ptr);
+      }
+    }
+  }
+  handler_map_[id] = nullptr;
+  free_instance_ids_.push_back(id);
+}
+
+ThreadLocalPtr::ThreadLocalPtr(UnrefHandler handler)
+    : id_(Instance()->GetId()) {
+  if (handler != nullptr) {
+    Instance()->SetHandler(id_, handler);
+  }
+}
+
+ThreadLocalPtr::~ThreadLocalPtr() { Instance()->ReclaimId(id_); }
+
+void* ThreadLocalPtr::Get() const { return Instance()->Get(id_); }
+
+void ThreadLocalPtr::Reset(void* ptr) { Instance()->Reset(id_, ptr); }
+
+void* ThreadLocalPtr::Swap(void* ptr) { return Instance()->Swap(id_, ptr); }
+
+bool ThreadLocalPtr::CompareAndSwap(void* ptr, void*& expected) {
+  return Instance()->CompareAndSwap(id_, ptr, expected);
+}
+
+void ThreadLocalPtr::Scrape(autovector<void*>* ptrs, void* const replacement) {
+  Instance()->Scrape(id_, ptrs, replacement);
+}
+
+void ThreadLocalPtr::Fold(FoldFunc func, void* res) {
+  Instance()->Fold(id_, func, res);
+}
+
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/util/thread_local.h b/src/rocksdb/util/thread_local.h
new file mode 100644
index 000000000..fde68f86f
--- /dev/null
+++ b/src/rocksdb/util/thread_local.h
@@ -0,0 +1,100 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+//
+// Copyright (c) 2011 The LevelDB Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file. See the AUTHORS file for names of contributors.
+
+#pragma once
+
+#include <atomic>
+#include <functional>
+#include <memory>
+#include <unordered_map>
+#include <vector>
+
+#include "port/port.h"
+#include "util/autovector.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+// Cleanup function that will be called for a stored thread local
+// pointer (if not NULL) when one of the following happens:
+// (1) a thread terminates
+// (2) a ThreadLocalPtr is destroyed
+//
+// Warning: this function is called while holding a global mutex. The same mutex
+// is used (at least in some cases) by most methods of ThreadLocalPtr, and it's
+// shared across all instances of ThreadLocalPtr. Thereforere extra care
+// is needed to avoid deadlocks. In particular, the handler shouldn't lock any
+// mutexes and shouldn't call any methods of any ThreadLocalPtr instances,
+// unless you know what you're doing.
+using UnrefHandler = void (*)(void* ptr);
+
+// ThreadLocalPtr stores only values of pointer type.  Different from
+// the usual thread-local-storage, ThreadLocalPtr has the ability to
+// distinguish data coming from different threads and different
+// ThreadLocalPtr instances.  For example, if a regular thread_local
+// variable A is declared in DBImpl, two DBImpl objects would share
+// the same A.  However, a ThreadLocalPtr that is defined under the
+// scope of DBImpl can avoid such confliction.  As a result, its memory
+// usage would be O(# of threads * # of ThreadLocalPtr instances).
+class ThreadLocalPtr {
+ public:
+  explicit ThreadLocalPtr(UnrefHandler handler = nullptr);
+
+  ThreadLocalPtr(const ThreadLocalPtr&) = delete;
+  ThreadLocalPtr& operator=(const ThreadLocalPtr&) = delete;
+
+  ~ThreadLocalPtr();
+
+  // Return the current pointer stored in thread local
+  void* Get() const;
+
+  // Set a new pointer value to the thread local storage.
+  void Reset(void* ptr);
+
+  // Atomically swap the supplied ptr and return the previous value
+  void* Swap(void* ptr);
+
+  // Atomically compare the stored value with expected. Set the new
+  // pointer value to thread local only if the comparison is true.
+  // Otherwise, expected returns the stored value.
+  // Return true on success, false on failure
+  bool CompareAndSwap(void* ptr, void*& expected);
+
+  // Reset all thread local data to replacement, and return non-nullptr
+  // data for all existing threads
+  void Scrape(autovector<void*>* ptrs, void* const replacement);
+
+  using FoldFunc = std::function<void(void*, void*)>;
+  // Update res by applying func on each thread-local value. Holds a lock that
+  // prevents unref handler from running during this call, but clients must
+  // still provide external synchronization since the owning thread can
+  // access the values without internal locking, e.g., via Get() and Reset().
+  void Fold(FoldFunc func, void* res);
+
+  // Add here for testing
+  // Return the next available Id without claiming it
+  static uint32_t TEST_PeekId();
+
+  // Initialize the static singletons of the ThreadLocalPtr.
+  //
+  // If this function is not called, then the singletons will be
+  // automatically initialized when they are used.
+  //
+  // Calling this function twice or after the singletons have been
+  // initialized will be no-op.
+  static void InitSingletons();
+
+  class StaticMeta;
+
+ private:
+  static StaticMeta* Instance();
+
+  const uint32_t id_;
+};
+
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/util/thread_local_test.cc b/src/rocksdb/util/thread_local_test.cc
new file mode 100644
index 000000000..25ef5c0ee
--- /dev/null
+++ b/src/rocksdb/util/thread_local_test.cc
@@ -0,0 +1,582 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#include "util/thread_local.h"
+
+#include <atomic>
+#include <string>
+#include <thread>
+
+#include "port/port.h"
+#include "rocksdb/env.h"
+#include "test_util/sync_point.h"
+#include "test_util/testharness.h"
+#include "test_util/testutil.h"
+#include "util/autovector.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+class ThreadLocalTest : public testing::Test {
+ public:
+  ThreadLocalTest() : env_(Env::Default()) {}
+
+  Env* env_;
+};
+
+namespace {
+
+struct Params {
+  Params(port::Mutex* m, port::CondVar* c, int* u, int n,
+         UnrefHandler handler = nullptr)
+      : mu(m),
+        cv(c),
+        unref(u),
+        total(n),
+        started(0),
+        completed(0),
+        doWrite(false),
+        tls1(handler),
+        tls2(nullptr) {}
+
+  port::Mutex* mu;
+  port::CondVar* cv;
+  int* unref;
+  int total;
+  int started;
+  int completed;
+  bool doWrite;
+  ThreadLocalPtr tls1;
+  ThreadLocalPtr* tls2;
+};
+
+class IDChecker : public ThreadLocalPtr {
+ public:
+  static uint32_t PeekId() { return TEST_PeekId(); }
+};
+
+}  // anonymous namespace
+
+// Suppress false positive clang analyzer warnings.
+#ifndef __clang_analyzer__
+TEST_F(ThreadLocalTest, UniqueIdTest) {
+  port::Mutex mu;
+  port::CondVar cv(&mu);
+
+  uint32_t base_id = IDChecker::PeekId();
+  // New ThreadLocal instance bumps id by 1
+  {
+    // Id used 0
+    Params p1(&mu, &cv, nullptr, 1u);
+    ASSERT_EQ(IDChecker::PeekId(), base_id + 1u);
+    // Id used 1
+    Params p2(&mu, &cv, nullptr, 1u);
+    ASSERT_EQ(IDChecker::PeekId(), base_id + 2u);
+    // Id used 2
+    Params p3(&mu, &cv, nullptr, 1u);
+    ASSERT_EQ(IDChecker::PeekId(), base_id + 3u);
+    // Id used 3
+    Params p4(&mu, &cv, nullptr, 1u);
+    ASSERT_EQ(IDChecker::PeekId(), base_id + 4u);
+  }
+  // id 3, 2, 1, 0 are in the free queue in order
+  ASSERT_EQ(IDChecker::PeekId(), base_id + 0u);
+
+  // pick up 0
+  Params p1(&mu, &cv, nullptr, 1u);
+  ASSERT_EQ(IDChecker::PeekId(), base_id + 1u);
+  // pick up 1
+  Params* p2 = new Params(&mu, &cv, nullptr, 1u);
+  ASSERT_EQ(IDChecker::PeekId(), base_id + 2u);
+  // pick up 2
+  Params p3(&mu, &cv, nullptr, 1u);
+  ASSERT_EQ(IDChecker::PeekId(), base_id + 3u);
+  // return up 1
+  delete p2;
+  ASSERT_EQ(IDChecker::PeekId(), base_id + 1u);
+  // Now we have 3, 1 in queue
+  // pick up 1
+  Params p4(&mu, &cv, nullptr, 1u);
+  ASSERT_EQ(IDChecker::PeekId(), base_id + 3u);
+  // pick up 3
+  Params p5(&mu, &cv, nullptr, 1u);
+  // next new id
+  ASSERT_EQ(IDChecker::PeekId(), base_id + 4u);
+  // After exit, id sequence in queue:
+  // 3, 1, 2, 0
+}
+#endif  // __clang_analyzer__
+
+TEST_F(ThreadLocalTest, SequentialReadWriteTest) {
+  // global id list carries over 3, 1, 2, 0
+  uint32_t base_id = IDChecker::PeekId();
+
+  port::Mutex mu;
+  port::CondVar cv(&mu);
+  Params p(&mu, &cv, nullptr, 1);
+  ThreadLocalPtr tls2;
+  p.tls2 = &tls2;
+
+  ASSERT_GT(IDChecker::PeekId(), base_id);
+  base_id = IDChecker::PeekId();
+
+  auto func = [](Params* ptr) {
+    Params& params = *ptr;
+    ASSERT_TRUE(params.tls1.Get() == nullptr);
+    params.tls1.Reset(reinterpret_cast<int*>(1));
+    ASSERT_TRUE(params.tls1.Get() == reinterpret_cast<int*>(1));
+    params.tls1.Reset(reinterpret_cast<int*>(2));
+    ASSERT_TRUE(params.tls1.Get() == reinterpret_cast<int*>(2));
+
+    ASSERT_TRUE(params.tls2->Get() == nullptr);
+    params.tls2->Reset(reinterpret_cast<int*>(1));
+    ASSERT_TRUE(params.tls2->Get() == reinterpret_cast<int*>(1));
+    params.tls2->Reset(reinterpret_cast<int*>(2));
+    ASSERT_TRUE(params.tls2->Get() == reinterpret_cast<int*>(2));
+
+    params.mu->Lock();
+    ++(params.completed);
+    params.cv->SignalAll();
+    params.mu->Unlock();
+  };
+
+  for (int iter = 0; iter < 1024; ++iter) {
+    ASSERT_EQ(IDChecker::PeekId(), base_id);
+    // Another new thread, read/write should not see value from previous thread
+    env_->StartThreadTyped(func, &p);
+
+    mu.Lock();
+    while (p.completed != iter + 1) {
+      cv.Wait();
+    }
+    mu.Unlock();
+    ASSERT_EQ(IDChecker::PeekId(), base_id);
+  }
+}
+
+TEST_F(ThreadLocalTest, ConcurrentReadWriteTest) {
+  // global id list carries over 3, 1, 2, 0
+  uint32_t base_id = IDChecker::PeekId();
+
+  ThreadLocalPtr tls2;
+  port::Mutex mu1;
+  port::CondVar cv1(&mu1);
+  Params p1(&mu1, &cv1, nullptr, 16);
+  p1.tls2 = &tls2;
+
+  port::Mutex mu2;
+  port::CondVar cv2(&mu2);
+  Params p2(&mu2, &cv2, nullptr, 16);
+  p2.doWrite = true;
+  p2.tls2 = &tls2;
+
+  auto func = [](void* ptr) {
+    auto& p = *static_cast<Params*>(ptr);
+
+    p.mu->Lock();
+    // Size_T switches size along with the ptr size
+    // we want to cast to.
+    size_t own = ++(p.started);
+    p.cv->SignalAll();
+    while (p.started != p.total) {
+      p.cv->Wait();
+    }
+    p.mu->Unlock();
+
+    // Let write threads write a different value from the read threads
+    if (p.doWrite) {
+      own += 8192;
+    }
+
+    ASSERT_TRUE(p.tls1.Get() == nullptr);
+    ASSERT_TRUE(p.tls2->Get() == nullptr);
+
+    auto* env = Env::Default();
+    auto start = env->NowMicros();
+
+    p.tls1.Reset(reinterpret_cast<size_t*>(own));
+    p.tls2->Reset(reinterpret_cast<size_t*>(own + 1));
+    // Loop for 1 second
+    while (env->NowMicros() - start < 1000 * 1000) {
+      for (int iter = 0; iter < 100000; ++iter) {
+        ASSERT_TRUE(p.tls1.Get() == reinterpret_cast<size_t*>(own));
+        ASSERT_TRUE(p.tls2->Get() == reinterpret_cast<size_t*>(own + 1));
+        if (p.doWrite) {
+          p.tls1.Reset(reinterpret_cast<size_t*>(own));
+          p.tls2->Reset(reinterpret_cast<size_t*>(own + 1));
+        }
+      }
+    }
+
+    p.mu->Lock();
+    ++(p.completed);
+    p.cv->SignalAll();
+    p.mu->Unlock();
+  };
+
+  // Initiate 2 instnaces: one keeps writing and one keeps reading.
+  // The read instance should not see data from the write instance.
+  // Each thread local copy of the value are also different from each
+  // other.
+  for (int th = 0; th < p1.total; ++th) {
+    env_->StartThreadTyped(func, &p1);
+  }
+  for (int th = 0; th < p2.total; ++th) {
+    env_->StartThreadTyped(func, &p2);
+  }
+
+  mu1.Lock();
+  while (p1.completed != p1.total) {
+    cv1.Wait();
+  }
+  mu1.Unlock();
+
+  mu2.Lock();
+  while (p2.completed != p2.total) {
+    cv2.Wait();
+  }
+  mu2.Unlock();
+
+  ASSERT_EQ(IDChecker::PeekId(), base_id + 3u);
+}
+
+TEST_F(ThreadLocalTest, Unref) {
+  auto unref = [](void* ptr) {
+    auto& p = *static_cast<Params*>(ptr);
+    p.mu->Lock();
+    ++(*p.unref);
+    p.mu->Unlock();
+  };
+
+  // Case 0: no unref triggered if ThreadLocalPtr is never accessed
+  auto func0 = [](Params* ptr) {
+    auto& p = *ptr;
+    p.mu->Lock();
+    ++(p.started);
+    p.cv->SignalAll();
+    while (p.started != p.total) {
+      p.cv->Wait();
+    }
+    p.mu->Unlock();
+  };
+
+  for (int th = 1; th <= 128; th += th) {
+    port::Mutex mu;
+    port::CondVar cv(&mu);
+    int unref_count = 0;
+    Params p(&mu, &cv, &unref_count, th, unref);
+
+    for (int i = 0; i < p.total; ++i) {
+      env_->StartThreadTyped(func0, &p);
+    }
+    env_->WaitForJoin();
+    ASSERT_EQ(unref_count, 0);
+  }
+
+  // Case 1: unref triggered by thread exit
+  auto func1 = [](Params* ptr) {
+    auto& p = *ptr;
+
+    p.mu->Lock();
+    ++(p.started);
+    p.cv->SignalAll();
+    while (p.started != p.total) {
+      p.cv->Wait();
+    }
+    p.mu->Unlock();
+
+    ASSERT_TRUE(p.tls1.Get() == nullptr);
+    ASSERT_TRUE(p.tls2->Get() == nullptr);
+
+    p.tls1.Reset(ptr);
+    p.tls2->Reset(ptr);
+
+    p.tls1.Reset(ptr);
+    p.tls2->Reset(ptr);
+  };
+
+  for (int th = 1; th <= 128; th += th) {
+    port::Mutex mu;
+    port::CondVar cv(&mu);
+    int unref_count = 0;
+    ThreadLocalPtr tls2(unref);
+    Params p(&mu, &cv, &unref_count, th, unref);
+    p.tls2 = &tls2;
+
+    for (int i = 0; i < p.total; ++i) {
+      env_->StartThreadTyped(func1, &p);
+    }
+
+    env_->WaitForJoin();
+
+    // N threads x 2 ThreadLocal instance cleanup on thread exit
+    ASSERT_EQ(unref_count, 2 * p.total);
+  }
+
+  // Case 2: unref triggered by ThreadLocal instance destruction
+  auto func2 = [](Params* ptr) {
+    auto& p = *ptr;
+
+    p.mu->Lock();
+    ++(p.started);
+    p.cv->SignalAll();
+    while (p.started != p.total) {
+      p.cv->Wait();
+    }
+    p.mu->Unlock();
+
+    ASSERT_TRUE(p.tls1.Get() == nullptr);
+    ASSERT_TRUE(p.tls2->Get() == nullptr);
+
+    p.tls1.Reset(ptr);
+    p.tls2->Reset(ptr);
+
+    p.tls1.Reset(ptr);
+    p.tls2->Reset(ptr);
+
+    p.mu->Lock();
+    ++(p.completed);
+    p.cv->SignalAll();
+
+    // Waiting for instruction to exit thread
+    while (p.completed != 0) {
+      p.cv->Wait();
+    }
+    p.mu->Unlock();
+  };
+
+  for (int th = 1; th <= 128; th += th) {
+    port::Mutex mu;
+    port::CondVar cv(&mu);
+    int unref_count = 0;
+    Params p(&mu, &cv, &unref_count, th, unref);
+    p.tls2 = new ThreadLocalPtr(unref);
+
+    for (int i = 0; i < p.total; ++i) {
+      env_->StartThreadTyped(func2, &p);
+    }
+
+    // Wait for all threads to finish using Params
+    mu.Lock();
+    while (p.completed != p.total) {
+      cv.Wait();
+    }
+    mu.Unlock();
+
+    // Now destroy one ThreadLocal instance
+    delete p.tls2;
+    p.tls2 = nullptr;
+    // instance destroy for N threads
+    ASSERT_EQ(unref_count, p.total);
+
+    // Signal to exit
+    mu.Lock();
+    p.completed = 0;
+    cv.SignalAll();
+    mu.Unlock();
+    env_->WaitForJoin();
+    // additional N threads exit unref for the left instance
+    ASSERT_EQ(unref_count, 2 * p.total);
+  }
+}
+
+TEST_F(ThreadLocalTest, Swap) {
+  ThreadLocalPtr tls;
+  tls.Reset(reinterpret_cast<void*>(1));
+  ASSERT_EQ(reinterpret_cast<int64_t>(tls.Swap(nullptr)), 1);
+  ASSERT_TRUE(tls.Swap(reinterpret_cast<void*>(2)) == nullptr);
+  ASSERT_EQ(reinterpret_cast<int64_t>(tls.Get()), 2);
+  ASSERT_EQ(reinterpret_cast<int64_t>(tls.Swap(reinterpret_cast<void*>(3))), 2);
+}
+
+TEST_F(ThreadLocalTest, Scrape) {
+  auto unref = [](void* ptr) {
+    auto& p = *static_cast<Params*>(ptr);
+    p.mu->Lock();
+    ++(*p.unref);
+    p.mu->Unlock();
+  };
+
+  auto func = [](void* ptr) {
+    auto& p = *static_cast<Params*>(ptr);
+
+    ASSERT_TRUE(p.tls1.Get() == nullptr);
+    ASSERT_TRUE(p.tls2->Get() == nullptr);
+
+    p.tls1.Reset(ptr);
+    p.tls2->Reset(ptr);
+
+    p.tls1.Reset(ptr);
+    p.tls2->Reset(ptr);
+
+    p.mu->Lock();
+    ++(p.completed);
+    p.cv->SignalAll();
+
+    // Waiting for instruction to exit thread
+    while (p.completed != 0) {
+      p.cv->Wait();
+    }
+    p.mu->Unlock();
+  };
+
+  for (int th = 1; th <= 128; th += th) {
+    port::Mutex mu;
+    port::CondVar cv(&mu);
+    int unref_count = 0;
+    Params p(&mu, &cv, &unref_count, th, unref);
+    p.tls2 = new ThreadLocalPtr(unref);
+
+    for (int i = 0; i < p.total; ++i) {
+      env_->StartThreadTyped(func, &p);
+    }
+
+    // Wait for all threads to finish using Params
+    mu.Lock();
+    while (p.completed != p.total) {
+      cv.Wait();
+    }
+    mu.Unlock();
+
+    ASSERT_EQ(unref_count, 0);
+
+    // Scrape all thread local data. No unref at thread
+    // exit or ThreadLocalPtr destruction
+    autovector<void*> ptrs;
+    p.tls1.Scrape(&ptrs, nullptr);
+    p.tls2->Scrape(&ptrs, nullptr);
+    delete p.tls2;
+    // Signal to exit
+    mu.Lock();
+    p.completed = 0;
+    cv.SignalAll();
+    mu.Unlock();
+    env_->WaitForJoin();
+
+    ASSERT_EQ(unref_count, 0);
+  }
+}
+
+TEST_F(ThreadLocalTest, Fold) {
+  auto unref = [](void* ptr) {
+    delete static_cast<std::atomic<int64_t>*>(ptr);
+  };
+  static const int kNumThreads = 16;
+  static const int kItersPerThread = 10;
+  port::Mutex mu;
+  port::CondVar cv(&mu);
+  Params params(&mu, &cv, nullptr, kNumThreads, unref);
+  auto func = [](void* ptr) {
+    auto& p = *static_cast<Params*>(ptr);
+    ASSERT_TRUE(p.tls1.Get() == nullptr);
+    p.tls1.Reset(new std::atomic<int64_t>(0));
+
+    for (int i = 0; i < kItersPerThread; ++i) {
+      static_cast<std::atomic<int64_t>*>(p.tls1.Get())->fetch_add(1);
+    }
+
+    p.mu->Lock();
+    ++(p.completed);
+    p.cv->SignalAll();
+
+    // Waiting for instruction to exit thread
+    while (p.completed != 0) {
+      p.cv->Wait();
+    }
+    p.mu->Unlock();
+  };
+
+  for (int th = 0; th < params.total; ++th) {
+    env_->StartThread(func, &params);
+  }
+
+  // Wait for all threads to finish using Params
+  mu.Lock();
+  while (params.completed != params.total) {
+    cv.Wait();
+  }
+  mu.Unlock();
+
+  // Verify Fold() behavior
+  int64_t sum = 0;
+  params.tls1.Fold(
+      [](void* ptr, void* res) {
+        auto sum_ptr = static_cast<int64_t*>(res);
+        *sum_ptr += static_cast<std::atomic<int64_t>*>(ptr)->load();
+      },
+      &sum);
+  ASSERT_EQ(sum, kNumThreads * kItersPerThread);
+
+  // Signal to exit
+  mu.Lock();
+  params.completed = 0;
+  cv.SignalAll();
+  mu.Unlock();
+  env_->WaitForJoin();
+}
+
+TEST_F(ThreadLocalTest, CompareAndSwap) {
+  ThreadLocalPtr tls;
+  ASSERT_TRUE(tls.Swap(reinterpret_cast<void*>(1)) == nullptr);
+  void* expected = reinterpret_cast<void*>(1);
+  // Swap in 2
+  ASSERT_TRUE(tls.CompareAndSwap(reinterpret_cast<void*>(2), expected));
+  expected = reinterpret_cast<void*>(100);
+  // Fail Swap, still 2
+  ASSERT_TRUE(!tls.CompareAndSwap(reinterpret_cast<void*>(2), expected));
+  ASSERT_EQ(expected, reinterpret_cast<void*>(2));
+  // Swap in 3
+  expected = reinterpret_cast<void*>(2);
+  ASSERT_TRUE(tls.CompareAndSwap(reinterpret_cast<void*>(3), expected));
+  ASSERT_EQ(tls.Get(), reinterpret_cast<void*>(3));
+}
+
+namespace {
+
+void* AccessThreadLocal(void* /*arg*/) {
+  TEST_SYNC_POINT("AccessThreadLocal:Start");
+  ThreadLocalPtr tlp;
+  tlp.Reset(new std::string("hello RocksDB"));
+  TEST_SYNC_POINT("AccessThreadLocal:End");
+  return nullptr;
+}
+
+}  // namespace
+
+// The following test is disabled as it requires manual steps to run it
+// correctly.
+//
+// Currently we have no way to acess SyncPoint w/o ASAN error when the
+// child thread dies after the main thread dies.  So if you manually enable
+// this test and only see an ASAN error on SyncPoint, it means you pass the
+// test.
+TEST_F(ThreadLocalTest, DISABLED_MainThreadDiesFirst) {
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->LoadDependency(
+      {{"AccessThreadLocal:Start", "MainThreadDiesFirst:End"},
+       {"PosixEnv::~PosixEnv():End", "AccessThreadLocal:End"}});
+
+  // Triggers the initialization of singletons.
+  Env::Default();
+
+#ifndef ROCKSDB_LITE
+  try {
+#endif  // ROCKSDB_LITE
+    ROCKSDB_NAMESPACE::port::Thread th(&AccessThreadLocal, nullptr);
+    th.detach();
+    TEST_SYNC_POINT("MainThreadDiesFirst:End");
+#ifndef ROCKSDB_LITE
+  } catch (const std::system_error& ex) {
+    std::cerr << "Start thread: " << ex.code() << std::endl;
+    FAIL();
+  }
+#endif  // ROCKSDB_LITE
+}
+
+}  // namespace ROCKSDB_NAMESPACE
+
+int main(int argc, char** argv) {
+  ROCKSDB_NAMESPACE::port::InstallStackTraceHandler();
+  ::testing::InitGoogleTest(&argc, argv);
+  return RUN_ALL_TESTS();
+}
diff --git a/src/rocksdb/util/thread_operation.h b/src/rocksdb/util/thread_operation.h
new file mode 100644
index 000000000..c24fccd5c
--- /dev/null
+++ b/src/rocksdb/util/thread_operation.h
@@ -0,0 +1,112 @@
+// Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+//
+// This file defines the structures for thread operation and state.
+// Thread operations are used to describe high level action of a
+// thread such as doing compaction or flush, while thread state
+// are used to describe lower-level action such as reading /
+// writing a file or waiting for a mutex.  Operations and states
+// are designed to be independent.  Typically, a thread usually involves
+// in one operation and one state at any specific point in time.
+
+#pragma once
+
+#include <string>
+
+#include "rocksdb/thread_status.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+#ifdef ROCKSDB_USING_THREAD_STATUS
+
+// The structure that describes a major thread operation.
+struct OperationInfo {
+  const ThreadStatus::OperationType type;
+  const std::string name;
+};
+
+// The global operation table.
+//
+// When updating a status of a thread, the pointer of the OperationInfo
+// of the current ThreadStatusData will be pointing to one of the
+// rows in this global table.
+//
+// Note that it's not designed to be constant as in the future we
+// might consider adding global count to the OperationInfo.
+static OperationInfo global_operation_table[] = {
+    {ThreadStatus::OP_UNKNOWN, ""},
+    {ThreadStatus::OP_COMPACTION, "Compaction"},
+    {ThreadStatus::OP_FLUSH, "Flush"}};
+
+struct OperationStageInfo {
+  const ThreadStatus::OperationStage stage;
+  const std::string name;
+};
+
+// A table maintains the mapping from stage type to stage string.
+// Note that the string must be changed accordingly when the
+// associated function name changed.
+static OperationStageInfo global_op_stage_table[] = {
+    {ThreadStatus::STAGE_UNKNOWN, ""},
+    {ThreadStatus::STAGE_FLUSH_RUN, "FlushJob::Run"},
+    {ThreadStatus::STAGE_FLUSH_WRITE_L0, "FlushJob::WriteLevel0Table"},
+    {ThreadStatus::STAGE_COMPACTION_PREPARE, "CompactionJob::Prepare"},
+    {ThreadStatus::STAGE_COMPACTION_RUN, "CompactionJob::Run"},
+    {ThreadStatus::STAGE_COMPACTION_PROCESS_KV,
+     "CompactionJob::ProcessKeyValueCompaction"},
+    {ThreadStatus::STAGE_COMPACTION_INSTALL, "CompactionJob::Install"},
+    {ThreadStatus::STAGE_COMPACTION_SYNC_FILE,
+     "CompactionJob::FinishCompactionOutputFile"},
+    {ThreadStatus::STAGE_PICK_MEMTABLES_TO_FLUSH,
+     "MemTableList::PickMemtablesToFlush"},
+    {ThreadStatus::STAGE_MEMTABLE_ROLLBACK,
+     "MemTableList::RollbackMemtableFlush"},
+    {ThreadStatus::STAGE_MEMTABLE_INSTALL_FLUSH_RESULTS,
+     "MemTableList::TryInstallMemtableFlushResults"},
+};
+
+// The structure that describes a state.
+struct StateInfo {
+  const ThreadStatus::StateType type;
+  const std::string name;
+};
+
+// The global state table.
+//
+// When updating a status of a thread, the pointer of the StateInfo
+// of the current ThreadStatusData will be pointing to one of the
+// rows in this global table.
+static StateInfo global_state_table[] = {
+    {ThreadStatus::STATE_UNKNOWN, ""},
+    {ThreadStatus::STATE_MUTEX_WAIT, "Mutex Wait"},
+};
+
+struct OperationProperty {
+  int code;
+  std::string name;
+};
+
+static OperationProperty compaction_operation_properties[] = {
+    {ThreadStatus::COMPACTION_JOB_ID, "JobID"},
+    {ThreadStatus::COMPACTION_INPUT_OUTPUT_LEVEL, "InputOutputLevel"},
+    {ThreadStatus::COMPACTION_PROP_FLAGS, "Manual/Deletion/Trivial"},
+    {ThreadStatus::COMPACTION_TOTAL_INPUT_BYTES, "TotalInputBytes"},
+    {ThreadStatus::COMPACTION_BYTES_READ, "BytesRead"},
+    {ThreadStatus::COMPACTION_BYTES_WRITTEN, "BytesWritten"},
+};
+
+static OperationProperty flush_operation_properties[] = {
+    {ThreadStatus::FLUSH_JOB_ID, "JobID"},
+    {ThreadStatus::FLUSH_BYTES_MEMTABLES, "BytesMemtables"},
+    {ThreadStatus::FLUSH_BYTES_WRITTEN, "BytesWritten"}};
+
+#else
+
+struct OperationInfo {};
+
+struct StateInfo {};
+
+#endif  // ROCKSDB_USING_THREAD_STATUS
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/util/threadpool_imp.cc b/src/rocksdb/util/threadpool_imp.cc
new file mode 100644
index 000000000..09706cac5
--- /dev/null
+++ b/src/rocksdb/util/threadpool_imp.cc
@@ -0,0 +1,551 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+//
+// Copyright (c) 2011 The LevelDB Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file. See the AUTHORS file for names of contributors.
+
+#include "util/threadpool_imp.h"
+
+#ifndef OS_WIN
+#include <unistd.h>
+#endif
+
+#ifdef OS_LINUX
+#include <sys/resource.h>
+#include <sys/syscall.h>
+#endif
+
+#include <stdlib.h>
+
+#include <algorithm>
+#include <atomic>
+#include <condition_variable>
+#include <deque>
+#include <mutex>
+#include <sstream>
+#include <thread>
+#include <vector>
+
+#include "monitoring/thread_status_util.h"
+#include "port/port.h"
+#include "test_util/sync_point.h"
+#include "util/string_util.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+void ThreadPoolImpl::PthreadCall(const char* label, int result) {
+  if (result != 0) {
+    fprintf(stderr, "pthread %s: %s\n", label, errnoStr(result).c_str());
+    abort();
+  }
+}
+
+struct ThreadPoolImpl::Impl {
+  Impl();
+  ~Impl();
+
+  void JoinThreads(bool wait_for_jobs_to_complete);
+
+  void SetBackgroundThreadsInternal(int num, bool allow_reduce);
+  int GetBackgroundThreads();
+
+  unsigned int GetQueueLen() const {
+    return queue_len_.load(std::memory_order_relaxed);
+  }
+
+  void LowerIOPriority();
+
+  void LowerCPUPriority(CpuPriority pri);
+
+  void WakeUpAllThreads() { bgsignal_.notify_all(); }
+
+  void BGThread(size_t thread_id);
+
+  void StartBGThreads();
+
+  void Submit(std::function<void()>&& schedule,
+              std::function<void()>&& unschedule, void* tag);
+
+  int UnSchedule(void* arg);
+
+  void SetHostEnv(Env* env) { env_ = env; }
+
+  Env* GetHostEnv() const { return env_; }
+
+  bool HasExcessiveThread() const {
+    return static_cast<int>(bgthreads_.size()) > total_threads_limit_;
+  }
+
+  // Return true iff the current thread is the excessive thread to terminate.
+  // Always terminate the running thread that is added last, even if there are
+  // more than one thread to terminate.
+  bool IsLastExcessiveThread(size_t thread_id) const {
+    return HasExcessiveThread() && thread_id == bgthreads_.size() - 1;
+  }
+
+  bool IsExcessiveThread(size_t thread_id) const {
+    return static_cast<int>(thread_id) >= total_threads_limit_;
+  }
+
+  // Return the thread priority.
+  // This would allow its member-thread to know its priority.
+  Env::Priority GetThreadPriority() const { return priority_; }
+
+  // Set the thread priority.
+  void SetThreadPriority(Env::Priority priority) { priority_ = priority; }
+
+  int ReserveThreads(int threads_to_be_reserved) {
+    std::unique_lock<std::mutex> lock(mu_);
+    // We can reserve at most num_waiting_threads_ in total so the number of
+    // threads that can be reserved might be fewer than the desired one. In
+    // rare cases, num_waiting_threads_ could be less than reserved_threads
+    // due to SetBackgroundThreadInternal or last excessive threads. If that
+    // happens, we cannot reserve any other threads.
+    int reserved_threads_in_success =
+        std::min(std::max(num_waiting_threads_ - reserved_threads_, 0),
+                 threads_to_be_reserved);
+    reserved_threads_ += reserved_threads_in_success;
+    return reserved_threads_in_success;
+  }
+
+  int ReleaseThreads(int threads_to_be_released) {
+    std::unique_lock<std::mutex> lock(mu_);
+    // We cannot release more than reserved_threads_
+    int released_threads_in_success =
+        std::min(reserved_threads_, threads_to_be_released);
+    reserved_threads_ -= released_threads_in_success;
+    WakeUpAllThreads();
+    return released_threads_in_success;
+  }
+
+ private:
+  static void BGThreadWrapper(void* arg);
+
+  bool low_io_priority_;
+  CpuPriority cpu_priority_;
+  Env::Priority priority_;
+  Env* env_;
+
+  int total_threads_limit_;
+  std::atomic_uint queue_len_;  // Queue length. Used for stats reporting
+  // Number of reserved threads, managed by ReserveThreads(..) and
+  // ReleaseThreads(..), if num_waiting_threads_ is no larger than
+  // reserved_threads_, its thread will be blocked to ensure the reservation
+  // mechanism
+  int reserved_threads_;
+  // Number of waiting threads (Maximum number of threads that can be
+  // reserved), in rare cases, num_waiting_threads_ could be less than
+  // reserved_threads due to SetBackgroundThreadInternal or last
+  // excessive threads.
+  int num_waiting_threads_;
+  bool exit_all_threads_;
+  bool wait_for_jobs_to_complete_;
+
+  // Entry per Schedule()/Submit() call
+  struct BGItem {
+    void* tag = nullptr;
+    std::function<void()> function;
+    std::function<void()> unschedFunction;
+  };
+
+  using BGQueue = std::deque<BGItem>;
+  BGQueue queue_;
+
+  std::mutex mu_;
+  std::condition_variable bgsignal_;
+  std::vector<port::Thread> bgthreads_;
+};
+
+inline ThreadPoolImpl::Impl::Impl()
+    : low_io_priority_(false),
+      cpu_priority_(CpuPriority::kNormal),
+      priority_(Env::LOW),
+      env_(nullptr),
+      total_threads_limit_(0),
+      queue_len_(),
+      reserved_threads_(0),
+      num_waiting_threads_(0),
+      exit_all_threads_(false),
+      wait_for_jobs_to_complete_(false),
+      queue_(),
+      mu_(),
+      bgsignal_(),
+      bgthreads_() {}
+
+inline ThreadPoolImpl::Impl::~Impl() { assert(bgthreads_.size() == 0U); }
+
+void ThreadPoolImpl::Impl::JoinThreads(bool wait_for_jobs_to_complete) {
+  std::unique_lock<std::mutex> lock(mu_);
+  assert(!exit_all_threads_);
+
+  wait_for_jobs_to_complete_ = wait_for_jobs_to_complete;
+  exit_all_threads_ = true;
+  // prevent threads from being recreated right after they're joined, in case
+  // the user is concurrently submitting jobs.
+  total_threads_limit_ = 0;
+  reserved_threads_ = 0;
+  num_waiting_threads_ = 0;
+
+  lock.unlock();
+
+  bgsignal_.notify_all();
+
+  for (auto& th : bgthreads_) {
+    th.join();
+  }
+
+  bgthreads_.clear();
+
+  exit_all_threads_ = false;
+  wait_for_jobs_to_complete_ = false;
+}
+
+inline void ThreadPoolImpl::Impl::LowerIOPriority() {
+  std::lock_guard<std::mutex> lock(mu_);
+  low_io_priority_ = true;
+}
+
+inline void ThreadPoolImpl::Impl::LowerCPUPriority(CpuPriority pri) {
+  std::lock_guard<std::mutex> lock(mu_);
+  cpu_priority_ = pri;
+}
+
+void ThreadPoolImpl::Impl::BGThread(size_t thread_id) {
+  bool low_io_priority = false;
+  CpuPriority current_cpu_priority = CpuPriority::kNormal;
+
+  while (true) {
+    // Wait until there is an item that is ready to run
+    std::unique_lock<std::mutex> lock(mu_);
+    // Stop waiting if the thread needs to do work or needs to terminate.
+    // Increase num_waiting_threads_ once this task has started waiting
+    num_waiting_threads_++;
+
+    TEST_SYNC_POINT("ThreadPoolImpl::BGThread::WaitingThreadsInc");
+    TEST_IDX_SYNC_POINT("ThreadPoolImpl::BGThread::Start:th", thread_id);
+    // When not exist_all_threads and the current thread id is not the last
+    // excessive thread, it may be blocked due to 3 reasons: 1) queue is empty
+    // 2) it is the excessive thread (not the last one)
+    // 3) the number of waiting threads is not greater than reserved threads
+    // (i.e, no available threads due to full reservation")
+    while (!exit_all_threads_ && !IsLastExcessiveThread(thread_id) &&
+           (queue_.empty() || IsExcessiveThread(thread_id) ||
+            num_waiting_threads_ <= reserved_threads_)) {
+      bgsignal_.wait(lock);
+    }
+    // Decrease num_waiting_threads_ once the thread is not waiting
+    num_waiting_threads_--;
+
+    if (exit_all_threads_) {  // mechanism to let BG threads exit safely
+
+      if (!wait_for_jobs_to_complete_ || queue_.empty()) {
+        break;
+      }
+    } else if (IsLastExcessiveThread(thread_id)) {
+      // Current thread is the last generated one and is excessive.
+      // We always terminate excessive thread in the reverse order of
+      // generation time. But not when `exit_all_threads_ == true`,
+      // otherwise `JoinThreads()` could try to `join()` a `detach()`ed
+      // thread.
+      auto& terminating_thread = bgthreads_.back();
+      terminating_thread.detach();
+      bgthreads_.pop_back();
+      if (HasExcessiveThread()) {
+        // There is still at least more excessive thread to terminate.
+        WakeUpAllThreads();
+      }
+      TEST_IDX_SYNC_POINT("ThreadPoolImpl::BGThread::Termination:th",
+                          thread_id);
+      TEST_SYNC_POINT("ThreadPoolImpl::BGThread::Termination");
+      break;
+    }
+
+    auto func = std::move(queue_.front().function);
+    queue_.pop_front();
+
+    queue_len_.store(static_cast<unsigned int>(queue_.size()),
+                     std::memory_order_relaxed);
+
+    bool decrease_io_priority = (low_io_priority != low_io_priority_);
+    CpuPriority cpu_priority = cpu_priority_;
+    lock.unlock();
+
+    if (cpu_priority < current_cpu_priority) {
+      TEST_SYNC_POINT_CALLBACK("ThreadPoolImpl::BGThread::BeforeSetCpuPriority",
+                               &current_cpu_priority);
+      // 0 means current thread.
+      port::SetCpuPriority(0, cpu_priority);
+      current_cpu_priority = cpu_priority;
+      TEST_SYNC_POINT_CALLBACK("ThreadPoolImpl::BGThread::AfterSetCpuPriority",
+                               &current_cpu_priority);
+    }
+
+#ifdef OS_LINUX
+    if (decrease_io_priority) {
+#define IOPRIO_CLASS_SHIFT (13)
+#define IOPRIO_PRIO_VALUE(class, data) (((class) << IOPRIO_CLASS_SHIFT) | data)
+      // Put schedule into IOPRIO_CLASS_IDLE class (lowest)
+      // These system calls only have an effect when used in conjunction
+      // with an I/O scheduler that supports I/O priorities. As at
+      // kernel 2.6.17 the only such scheduler is the Completely
+      // Fair Queuing (CFQ) I/O scheduler.
+      // To change scheduler:
+      //  echo cfq > /sys/block/<device_name>/queue/schedule
+      // Tunables to consider:
+      //  /sys/block/<device_name>/queue/slice_idle
+      //  /sys/block/<device_name>/queue/slice_sync
+      syscall(SYS_ioprio_set, 1,  // IOPRIO_WHO_PROCESS
+              0,                  // current thread
+              IOPRIO_PRIO_VALUE(3, 0));
+      low_io_priority = true;
+    }
+#else
+    (void)decrease_io_priority;  // avoid 'unused variable' error
+#endif
+
+    TEST_SYNC_POINT_CALLBACK("ThreadPoolImpl::Impl::BGThread:BeforeRun",
+                             &priority_);
+
+    func();
+  }
+}
+
+// Helper struct for passing arguments when creating threads.
+struct BGThreadMetadata {
+  ThreadPoolImpl::Impl* thread_pool_;
+  size_t thread_id_;  // Thread count in the thread.
+  BGThreadMetadata(ThreadPoolImpl::Impl* thread_pool, size_t thread_id)
+      : thread_pool_(thread_pool), thread_id_(thread_id) {}
+};
+
+void ThreadPoolImpl::Impl::BGThreadWrapper(void* arg) {
+  BGThreadMetadata* meta = reinterpret_cast<BGThreadMetadata*>(arg);
+  size_t thread_id = meta->thread_id_;
+  ThreadPoolImpl::Impl* tp = meta->thread_pool_;
+#ifdef ROCKSDB_USING_THREAD_STATUS
+  // initialize it because compiler isn't good enough to see we don't use it
+  // uninitialized
+  ThreadStatus::ThreadType thread_type = ThreadStatus::NUM_THREAD_TYPES;
+  switch (tp->GetThreadPriority()) {
+    case Env::Priority::HIGH:
+      thread_type = ThreadStatus::HIGH_PRIORITY;
+      break;
+    case Env::Priority::LOW:
+      thread_type = ThreadStatus::LOW_PRIORITY;
+      break;
+    case Env::Priority::BOTTOM:
+      thread_type = ThreadStatus::BOTTOM_PRIORITY;
+      break;
+    case Env::Priority::USER:
+      thread_type = ThreadStatus::USER;
+      break;
+    case Env::Priority::TOTAL:
+      assert(false);
+      return;
+  }
+  assert(thread_type != ThreadStatus::NUM_THREAD_TYPES);
+  ThreadStatusUtil::RegisterThread(tp->GetHostEnv(), thread_type);
+#endif
+  delete meta;
+  tp->BGThread(thread_id);
+#ifdef ROCKSDB_USING_THREAD_STATUS
+  ThreadStatusUtil::UnregisterThread();
+#endif
+  return;
+}
+
+void ThreadPoolImpl::Impl::SetBackgroundThreadsInternal(int num,
+                                                        bool allow_reduce) {
+  std::lock_guard<std::mutex> lock(mu_);
+  if (exit_all_threads_) {
+    return;
+  }
+  if (num > total_threads_limit_ ||
+      (num < total_threads_limit_ && allow_reduce)) {
+    total_threads_limit_ = std::max(0, num);
+    WakeUpAllThreads();
+    StartBGThreads();
+  }
+}
+
+int ThreadPoolImpl::Impl::GetBackgroundThreads() {
+  std::unique_lock<std::mutex> lock(mu_);
+  return total_threads_limit_;
+}
+
+void ThreadPoolImpl::Impl::StartBGThreads() {
+  // Start background thread if necessary
+  while ((int)bgthreads_.size() < total_threads_limit_) {
+    port::Thread p_t(&BGThreadWrapper,
+                     new BGThreadMetadata(this, bgthreads_.size()));
+
+// Set the thread name to aid debugging
+#if defined(_GNU_SOURCE) && defined(__GLIBC_PREREQ)
+#if __GLIBC_PREREQ(2, 12)
+    auto th_handle = p_t.native_handle();
+    std::string thread_priority = Env::PriorityToString(GetThreadPriority());
+    std::ostringstream thread_name_stream;
+    thread_name_stream << "rocksdb:";
+    for (char c : thread_priority) {
+      thread_name_stream << static_cast<char>(tolower(c));
+    }
+    pthread_setname_np(th_handle, thread_name_stream.str().c_str());
+#endif
+#endif
+    bgthreads_.push_back(std::move(p_t));
+  }
+}
+
+void ThreadPoolImpl::Impl::Submit(std::function<void()>&& schedule,
+                                  std::function<void()>&& unschedule,
+                                  void* tag) {
+  std::lock_guard<std::mutex> lock(mu_);
+
+  if (exit_all_threads_) {
+    return;
+  }
+
+  StartBGThreads();
+
+  // Add to priority queue
+  queue_.push_back(BGItem());
+  TEST_SYNC_POINT("ThreadPoolImpl::Submit::Enqueue");
+  auto& item = queue_.back();
+  item.tag = tag;
+  item.function = std::move(schedule);
+  item.unschedFunction = std::move(unschedule);
+
+  queue_len_.store(static_cast<unsigned int>(queue_.size()),
+                   std::memory_order_relaxed);
+
+  if (!HasExcessiveThread()) {
+    // Wake up at least one waiting thread.
+    bgsignal_.notify_one();
+  } else {
+    // Need to wake up all threads to make sure the one woken
+    // up is not the one to terminate.
+    WakeUpAllThreads();
+  }
+}
+
+int ThreadPoolImpl::Impl::UnSchedule(void* arg) {
+  int count = 0;
+
+  std::vector<std::function<void()>> candidates;
+  {
+    std::lock_guard<std::mutex> lock(mu_);
+
+    // Remove from priority queue
+    BGQueue::iterator it = queue_.begin();
+    while (it != queue_.end()) {
+      if (arg == (*it).tag) {
+        if (it->unschedFunction) {
+          candidates.push_back(std::move(it->unschedFunction));
+        }
+        it = queue_.erase(it);
+        count++;
+      } else {
+        ++it;
+      }
+    }
+    queue_len_.store(static_cast<unsigned int>(queue_.size()),
+                     std::memory_order_relaxed);
+  }
+
+  // Run unschedule functions outside the mutex
+  for (auto& f : candidates) {
+    f();
+  }
+
+  return count;
+}
+
+ThreadPoolImpl::ThreadPoolImpl() : impl_(new Impl()) {}
+
+ThreadPoolImpl::~ThreadPoolImpl() {}
+
+void ThreadPoolImpl::JoinAllThreads() { impl_->JoinThreads(false); }
+
+void ThreadPoolImpl::SetBackgroundThreads(int num) {
+  impl_->SetBackgroundThreadsInternal(num, true);
+}
+
+int ThreadPoolImpl::GetBackgroundThreads() {
+  return impl_->GetBackgroundThreads();
+}
+
+unsigned int ThreadPoolImpl::GetQueueLen() const {
+  return impl_->GetQueueLen();
+}
+
+void ThreadPoolImpl::WaitForJobsAndJoinAllThreads() {
+  impl_->JoinThreads(true);
+}
+
+void ThreadPoolImpl::LowerIOPriority() { impl_->LowerIOPriority(); }
+
+void ThreadPoolImpl::LowerCPUPriority(CpuPriority pri) {
+  impl_->LowerCPUPriority(pri);
+}
+
+void ThreadPoolImpl::IncBackgroundThreadsIfNeeded(int num) {
+  impl_->SetBackgroundThreadsInternal(num, false);
+}
+
+void ThreadPoolImpl::SubmitJob(const std::function<void()>& job) {
+  auto copy(job);
+  impl_->Submit(std::move(copy), std::function<void()>(), nullptr);
+}
+
+void ThreadPoolImpl::SubmitJob(std::function<void()>&& job) {
+  impl_->Submit(std::move(job), std::function<void()>(), nullptr);
+}
+
+void ThreadPoolImpl::Schedule(void (*function)(void* arg1), void* arg,
+                              void* tag, void (*unschedFunction)(void* arg)) {
+  if (unschedFunction == nullptr) {
+    impl_->Submit(std::bind(function, arg), std::function<void()>(), tag);
+  } else {
+    impl_->Submit(std::bind(function, arg), std::bind(unschedFunction, arg),
+                  tag);
+  }
+}
+
+int ThreadPoolImpl::UnSchedule(void* arg) { return impl_->UnSchedule(arg); }
+
+void ThreadPoolImpl::SetHostEnv(Env* env) { impl_->SetHostEnv(env); }
+
+Env* ThreadPoolImpl::GetHostEnv() const { return impl_->GetHostEnv(); }
+
+// Return the thread priority.
+// This would allow its member-thread to know its priority.
+Env::Priority ThreadPoolImpl::GetThreadPriority() const {
+  return impl_->GetThreadPriority();
+}
+
+// Set the thread priority.
+void ThreadPoolImpl::SetThreadPriority(Env::Priority priority) {
+  impl_->SetThreadPriority(priority);
+}
+
+// Reserve a specific number of threads, prevent them from running other
+// functions The number of reserved threads could be fewer than the desired one
+int ThreadPoolImpl::ReserveThreads(int threads_to_be_reserved) {
+  return impl_->ReserveThreads(threads_to_be_reserved);
+}
+
+// Release a specific number of threads
+int ThreadPoolImpl::ReleaseThreads(int threads_to_be_released) {
+  return impl_->ReleaseThreads(threads_to_be_released);
+}
+
+ThreadPool* NewThreadPool(int num_threads) {
+  ThreadPoolImpl* thread_pool = new ThreadPoolImpl();
+  thread_pool->SetBackgroundThreads(num_threads);
+  return thread_pool;
+}
+
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/util/threadpool_imp.h b/src/rocksdb/util/threadpool_imp.h
new file mode 100644
index 000000000..a5109e38f
--- /dev/null
+++ b/src/rocksdb/util/threadpool_imp.h
@@ -0,0 +1,120 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+//
+// Copyright (c) 2011 The LevelDB Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file. See the AUTHORS file for names of contributors.
+#pragma once
+
+#include <functional>
+#include <memory>
+
+#include "rocksdb/env.h"
+#include "rocksdb/threadpool.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+class ThreadPoolImpl : public ThreadPool {
+ public:
+  ThreadPoolImpl();
+  ~ThreadPoolImpl();
+
+  ThreadPoolImpl(ThreadPoolImpl&&) = delete;
+  ThreadPoolImpl& operator=(ThreadPoolImpl&&) = delete;
+
+  // Implement ThreadPool interfaces
+
+  // Wait for all threads to finish.
+  // Discards all the jobs that did not
+  // start executing and waits for those running
+  // to complete
+  void JoinAllThreads() override;
+
+  // Set the number of background threads that will be executing the
+  // scheduled jobs.
+  void SetBackgroundThreads(int num) override;
+  int GetBackgroundThreads() override;
+
+  // Get the number of jobs scheduled in the ThreadPool queue.
+  unsigned int GetQueueLen() const override;
+
+  // Waits for all jobs to complete those
+  // that already started running and those that did not
+  // start yet
+  void WaitForJobsAndJoinAllThreads() override;
+
+  // Make threads to run at a lower kernel IO priority
+  // Currently only has effect on Linux
+  void LowerIOPriority();
+
+  // Make threads to run at a lower kernel CPU priority
+  // Currently only has effect on Linux
+  void LowerCPUPriority(CpuPriority pri);
+
+  // Ensure there is at aleast num threads in the pool
+  // but do not kill threads if there are more
+  void IncBackgroundThreadsIfNeeded(int num);
+
+  // Submit a fire and forget job
+  // These jobs can not be unscheduled
+
+  // This allows to submit the same job multiple times
+  void SubmitJob(const std::function<void()>&) override;
+  // This moves the function in for efficiency
+  void SubmitJob(std::function<void()>&&) override;
+
+  // Schedule a job with an unschedule tag and unschedule function
+  // Can be used to filter and unschedule jobs by a tag
+  // that are still in the queue and did not start running
+  void Schedule(void (*function)(void* arg1), void* arg, void* tag,
+                void (*unschedFunction)(void* arg));
+
+  // Filter jobs that are still in a queue and match
+  // the given tag. Remove them from a queue if any
+  // and for each such job execute an unschedule function
+  // if such was given at scheduling time.
+  int UnSchedule(void* tag);
+
+  void SetHostEnv(Env* env);
+
+  Env* GetHostEnv() const;
+
+  // Return the thread priority.
+  // This would allow its member-thread to know its priority.
+  Env::Priority GetThreadPriority() const;
+
+  // Set the thread priority.
+  void SetThreadPriority(Env::Priority priority);
+
+  // Reserve a specific number of threads, prevent them from running other
+  // functions The number of reserved threads could be fewer than the desired
+  // one
+  int ReserveThreads(int threads_to_be_reserved) override;
+
+  // Release a specific number of threads
+  int ReleaseThreads(int threads_to_be_released) override;
+
+  static void PthreadCall(const char* label, int result);
+
+  struct Impl;
+
+ private:
+  // Current public virtual interface does not provide usable
+  // functionality and thus can not be used internally to
+  // facade different implementations.
+  //
+  // We propose a pimpl idiom in order to easily replace the thread pool impl
+  // w/o touching the header file but providing a different .cc potentially
+  // CMake option driven.
+  //
+  // Another option is to introduce a Env::MakeThreadPool() virtual interface
+  // and override the environment. This would require refactoring ThreadPool
+  // usage.
+  //
+  // We can also combine these two approaches
+  std::unique_ptr<Impl> impl_;
+};
+
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/util/timer.h b/src/rocksdb/util/timer.h
new file mode 100644
index 000000000..db71cefaf
--- /dev/null
+++ b/src/rocksdb/util/timer.h
@@ -0,0 +1,340 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+//
+
+#pragma once
+
+#include <functional>
+#include <memory>
+#include <queue>
+#include <unordered_map>
+#include <utility>
+#include <vector>
+
+#include "monitoring/instrumented_mutex.h"
+#include "rocksdb/system_clock.h"
+#include "test_util/sync_point.h"
+#include "util/mutexlock.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+// A Timer class to handle repeated work.
+//
+// `Start()` and `Shutdown()` are currently not thread-safe. The client must
+// serialize calls to these two member functions.
+//
+// A single timer instance can handle multiple functions via a single thread.
+// It is better to leave long running work to a dedicated thread pool.
+//
+// Timer can be started by calling `Start()`, and ended by calling `Shutdown()`.
+// Work (in terms of a `void function`) can be scheduled by calling `Add` with
+// a unique function name and de-scheduled by calling `Cancel`.
+// Many functions can be added.
+//
+// Impl Details:
+// A heap is used to keep track of when the next timer goes off.
+// A map from a function name to the function keeps track of all the functions.
+class Timer {
+ public:
+  explicit Timer(SystemClock* clock)
+      : clock_(clock),
+        mutex_(clock),
+        cond_var_(&mutex_),
+        running_(false),
+        executing_task_(false) {}
+
+  ~Timer() { Shutdown(); }
+
+  // Add a new function to run.
+  // fn_name has to be identical, otherwise it will fail to add and return false
+  // start_after_us is the initial delay.
+  // repeat_every_us is the interval between ending time of the last call and
+  // starting time of the next call. For example, repeat_every_us = 2000 and
+  // the function takes 1000us to run. If it starts at time [now]us, then it
+  // finishes at [now]+1000us, 2nd run starting time will be at [now]+3000us.
+  // repeat_every_us == 0 means do not repeat.
+  bool Add(std::function<void()> fn, const std::string& fn_name,
+           uint64_t start_after_us, uint64_t repeat_every_us) {
+    auto fn_info = std::make_unique<FunctionInfo>(std::move(fn), fn_name, 0,
+                                                  repeat_every_us);
+    InstrumentedMutexLock l(&mutex_);
+    // Assign time within mutex to make sure the next_run_time is larger than
+    // the current running one
+    fn_info->next_run_time_us = clock_->NowMicros() + start_after_us;
+    // the new task start time should never before the current task executing
+    // time, as the executing task can only be running if it's next_run_time_us
+    // is due (<= clock_->NowMicros()).
+    if (executing_task_ &&
+        fn_info->next_run_time_us < heap_.top()->next_run_time_us) {
+      return false;
+    }
+    auto it = map_.find(fn_name);
+    if (it == map_.end()) {
+      heap_.push(fn_info.get());
+      map_.try_emplace(fn_name, std::move(fn_info));
+    } else {
+      // timer doesn't support duplicated function name
+      return false;
+    }
+    cond_var_.SignalAll();
+    return true;
+  }
+
+  void Cancel(const std::string& fn_name) {
+    InstrumentedMutexLock l(&mutex_);
+
+    // Mark the function with fn_name as invalid so that it will not be
+    // requeued.
+    auto it = map_.find(fn_name);
+    if (it != map_.end() && it->second) {
+      it->second->Cancel();
+    }
+
+    // If the currently running function is fn_name, then we need to wait
+    // until it finishes before returning to caller.
+    while (!heap_.empty() && executing_task_) {
+      FunctionInfo* func_info = heap_.top();
+      assert(func_info);
+      if (func_info->name == fn_name) {
+        WaitForTaskCompleteIfNecessary();
+      } else {
+        break;
+      }
+    }
+  }
+
+  void CancelAll() {
+    InstrumentedMutexLock l(&mutex_);
+    CancelAllWithLock();
+  }
+
+  // Start the Timer
+  bool Start() {
+    InstrumentedMutexLock l(&mutex_);
+    if (running_) {
+      return false;
+    }
+
+    running_ = true;
+    thread_ = std::make_unique<port::Thread>(&Timer::Run, this);
+    return true;
+  }
+
+  // Shutdown the Timer
+  bool Shutdown() {
+    {
+      InstrumentedMutexLock l(&mutex_);
+      if (!running_) {
+        return false;
+      }
+      running_ = false;
+      CancelAllWithLock();
+      cond_var_.SignalAll();
+    }
+
+    if (thread_) {
+      thread_->join();
+    }
+    return true;
+  }
+
+  bool HasPendingTask() const {
+    InstrumentedMutexLock l(&mutex_);
+    for (const auto& fn_info : map_) {
+      if (fn_info.second->IsValid()) {
+        return true;
+      }
+    }
+    return false;
+  }
+
+#ifndef NDEBUG
+  // Wait until Timer starting waiting, call the optional callback, then wait
+  // for Timer waiting again.
+  // Tests can provide a custom Clock object to mock time, and use the callback
+  // here to bump current time and trigger Timer. See timer_test for example.
+  //
+  // Note: only support one caller of this method.
+  void TEST_WaitForRun(const std::function<void()>& callback = nullptr) {
+    InstrumentedMutexLock l(&mutex_);
+    // It act as a spin lock
+    while (executing_task_ ||
+           (!heap_.empty() &&
+            heap_.top()->next_run_time_us <= clock_->NowMicros())) {
+      cond_var_.TimedWait(clock_->NowMicros() + 1000);
+    }
+    if (callback != nullptr) {
+      callback();
+    }
+    cond_var_.SignalAll();
+    do {
+      cond_var_.TimedWait(clock_->NowMicros() + 1000);
+    } while (executing_task_ ||
+             (!heap_.empty() &&
+              heap_.top()->next_run_time_us <= clock_->NowMicros()));
+  }
+
+  size_t TEST_GetPendingTaskNum() const {
+    InstrumentedMutexLock l(&mutex_);
+    size_t ret = 0;
+    for (const auto& fn_info : map_) {
+      if (fn_info.second->IsValid()) {
+        ret++;
+      }
+    }
+    return ret;
+  }
+
+  void TEST_OverrideTimer(SystemClock* clock) {
+    InstrumentedMutexLock l(&mutex_);
+    clock_ = clock;
+  }
+#endif  // NDEBUG
+
+ private:
+  void Run() {
+    InstrumentedMutexLock l(&mutex_);
+
+    while (running_) {
+      if (heap_.empty()) {
+        // wait
+        TEST_SYNC_POINT("Timer::Run::Waiting");
+        cond_var_.Wait();
+        continue;
+      }
+
+      FunctionInfo* current_fn = heap_.top();
+      assert(current_fn);
+
+      if (!current_fn->IsValid()) {
+        heap_.pop();
+        map_.erase(current_fn->name);
+        continue;
+      }
+
+      if (current_fn->next_run_time_us <= clock_->NowMicros()) {
+        // make a copy of the function so it won't be changed after
+        // mutex_.unlock.
+        std::function<void()> fn = current_fn->fn;
+        executing_task_ = true;
+        mutex_.Unlock();
+        // Execute the work
+        fn();
+        mutex_.Lock();
+        executing_task_ = false;
+        cond_var_.SignalAll();
+
+        // Remove the work from the heap once it is done executing, make sure
+        // it's the same function after executing the work while mutex is
+        // released.
+        // Note that we are just removing the pointer from the heap. Its
+        // memory is still managed in the map (as it holds a unique ptr).
+        // So current_fn is still a valid ptr.
+        assert(heap_.top() == current_fn);
+        heap_.pop();
+
+        // current_fn may be cancelled already.
+        if (current_fn->IsValid() && current_fn->repeat_every_us > 0) {
+          assert(running_);
+          current_fn->next_run_time_us =
+              clock_->NowMicros() + current_fn->repeat_every_us;
+
+          // Schedule new work into the heap with new time.
+          heap_.push(current_fn);
+        } else {
+          // if current_fn is cancelled or no need to repeat, remove it from the
+          // map to avoid leak.
+          map_.erase(current_fn->name);
+        }
+      } else {
+        cond_var_.TimedWait(current_fn->next_run_time_us);
+      }
+    }
+  }
+
+  void CancelAllWithLock() {
+    mutex_.AssertHeld();
+    if (map_.empty() && heap_.empty()) {
+      return;
+    }
+
+    // With mutex_ held, set all tasks to invalid so that they will not be
+    // re-queued.
+    for (auto& elem : map_) {
+      auto& func_info = elem.second;
+      assert(func_info);
+      func_info->Cancel();
+    }
+
+    // WaitForTaskCompleteIfNecessary() may release mutex_
+    WaitForTaskCompleteIfNecessary();
+
+    while (!heap_.empty()) {
+      heap_.pop();
+    }
+    map_.clear();
+  }
+
+  // A wrapper around std::function to keep track when it should run next
+  // and at what frequency.
+  struct FunctionInfo {
+    // the actual work
+    std::function<void()> fn;
+    // name of the function
+    std::string name;
+    // when the function should run next
+    uint64_t next_run_time_us;
+    // repeat interval
+    uint64_t repeat_every_us;
+    // controls whether this function is valid.
+    // A function is valid upon construction and until someone explicitly
+    // calls `Cancel()`.
+    bool valid;
+
+    FunctionInfo(std::function<void()>&& _fn, std::string _name,
+                 const uint64_t _next_run_time_us, uint64_t _repeat_every_us)
+        : fn(std::move(_fn)),
+          name(std::move(_name)),
+          next_run_time_us(_next_run_time_us),
+          repeat_every_us(_repeat_every_us),
+          valid(true) {}
+
+    void Cancel() { valid = false; }
+
+    bool IsValid() const { return valid; }
+  };
+
+  void WaitForTaskCompleteIfNecessary() {
+    mutex_.AssertHeld();
+    while (executing_task_) {
+      TEST_SYNC_POINT("Timer::WaitForTaskCompleteIfNecessary:TaskExecuting");
+      cond_var_.Wait();
+    }
+  }
+
+  struct RunTimeOrder {
+    bool operator()(const FunctionInfo* f1, const FunctionInfo* f2) {
+      return f1->next_run_time_us > f2->next_run_time_us;
+    }
+  };
+
+  SystemClock* clock_;
+  // This mutex controls both the heap_ and the map_. It needs to be held for
+  // making any changes in them.
+  mutable InstrumentedMutex mutex_;
+  InstrumentedCondVar cond_var_;
+  std::unique_ptr<port::Thread> thread_;
+  bool running_;
+  bool executing_task_;
+
+  std::priority_queue<FunctionInfo*, std::vector<FunctionInfo*>, RunTimeOrder>
+      heap_;
+
+  // In addition to providing a mapping from a function name to a function,
+  // it is also responsible for memory management.
+  std::unordered_map<std::string, std::unique_ptr<FunctionInfo>> map_;
+};
+
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/util/timer_queue.h b/src/rocksdb/util/timer_queue.h
new file mode 100644
index 000000000..36a1744ac
--- /dev/null
+++ b/src/rocksdb/util/timer_queue.h
@@ -0,0 +1,231 @@
+//  Portions Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+//
+// Borrowed from
+// http://www.crazygaze.com/blog/2016/03/24/portable-c-timer-queue/
+// Timer Queue
+//
+// License
+//
+// The source code in this article is licensed under the CC0 license, so feel
+// free to copy, modify, share, do whatever you want with it.
+// No attribution is required, but Ill be happy if you do.
+// CC0 license
+
+// The person who associated a work with this deed has dedicated the work to the
+// public domain by waiving all of his or her rights to the work worldwide
+// under copyright law, including all related and neighboring rights, to the
+// extent allowed by law.  You can copy, modify, distribute and perform the
+// work, even for commercial purposes, all without asking permission.
+
+#pragma once
+
+#include <assert.h>
+
+#include <chrono>
+#include <condition_variable>
+#include <functional>
+#include <queue>
+#include <thread>
+#include <utility>
+#include <vector>
+
+#include "port/port.h"
+#include "test_util/sync_point.h"
+
+// Allows execution of handlers at a specified time in the future
+// Guarantees:
+//  - All handlers are executed ONCE, even if cancelled (aborted parameter will
+// be set to true)
+//      - If TimerQueue is destroyed, it will cancel all handlers.
+//  - Handlers are ALWAYS executed in the Timer Queue worker thread.
+//  - Handlers execution order is NOT guaranteed
+//
+////////////////////////////////////////////////////////////////////////////////
+// borrowed from
+// http://www.crazygaze.com/blog/2016/03/24/portable-c-timer-queue/
+class TimerQueue {
+ public:
+  TimerQueue() : m_th(&TimerQueue::run, this) {}
+
+  ~TimerQueue() { shutdown(); }
+
+  // This function is not thread-safe.
+  void shutdown() {
+    if (closed_) {
+      return;
+    }
+    cancelAll();
+    // Abusing the timer queue to trigger the shutdown.
+    add(0, [this](bool) {
+      m_finish = true;
+      return std::make_pair(false, 0);
+    });
+    m_th.join();
+    closed_ = true;
+  }
+
+  // Adds a new timer
+  // \return
+  //  Returns the ID of the new timer. You can use this ID to cancel the
+  // timer
+  uint64_t add(int64_t milliseconds,
+               std::function<std::pair<bool, int64_t>(bool)> handler) {
+    WorkItem item;
+    Clock::time_point tp = Clock::now();
+    item.end = tp + std::chrono::milliseconds(milliseconds);
+    TEST_SYNC_POINT_CALLBACK("TimeQueue::Add:item.end", &item.end);
+    item.period = milliseconds;
+    item.handler = std::move(handler);
+
+    std::unique_lock<std::mutex> lk(m_mtx);
+    uint64_t id = ++m_idcounter;
+    item.id = id;
+    m_items.push(std::move(item));
+
+    // Something changed, so wake up timer thread
+    m_checkWork.notify_one();
+    return id;
+  }
+
+  // Cancels the specified timer
+  // \return
+  //  1 if the timer was cancelled.
+  //  0 if you were too late to cancel (or the timer ID was never valid to
+  // start with)
+  size_t cancel(uint64_t id) {
+    // Instead of removing the item from the container (thus breaking the
+    // heap integrity), we set the item as having no handler, and put
+    // that handler on a new item at the top for immediate execution
+    // The timer thread will then ignore the original item, since it has no
+    // handler.
+    std::unique_lock<std::mutex> lk(m_mtx);
+    for (auto&& item : m_items.getContainer()) {
+      if (item.id == id && item.handler) {
+        WorkItem newItem;
+        // Zero time, so it stays at the top for immediate execution
+        newItem.end = Clock::time_point();
+        newItem.id = 0;  // Means it is a canceled item
+        // Move the handler from item to newitem (thus clearing item)
+        newItem.handler = std::move(item.handler);
+        m_items.push(std::move(newItem));
+
+        // Something changed, so wake up timer thread
+        m_checkWork.notify_one();
+        return 1;
+      }
+    }
+    return 0;
+  }
+
+  // Cancels all timers
+  // \return
+  //  The number of timers cancelled
+  size_t cancelAll() {
+    // Setting all "end" to 0 (for immediate execution) is ok,
+    // since it maintains the heap integrity
+    std::unique_lock<std::mutex> lk(m_mtx);
+    m_cancel = true;
+    for (auto&& item : m_items.getContainer()) {
+      if (item.id && item.handler) {
+        item.end = Clock::time_point();
+        item.id = 0;
+      }
+    }
+    auto ret = m_items.size();
+
+    m_checkWork.notify_one();
+    return ret;
+  }
+
+ private:
+  using Clock = std::chrono::steady_clock;
+  TimerQueue(const TimerQueue&) = delete;
+  TimerQueue& operator=(const TimerQueue&) = delete;
+
+  void run() {
+    std::unique_lock<std::mutex> lk(m_mtx);
+    while (!m_finish) {
+      auto end = calcWaitTime_lock();
+      if (end.first) {
+        // Timers found, so wait until it expires (or something else
+        // changes)
+        m_checkWork.wait_until(lk, end.second);
+      } else {
+        // No timers exist, so wait forever until something changes
+        m_checkWork.wait(lk);
+      }
+
+      // Check and execute as much work as possible, such as, all expired
+      // timers
+      checkWork(&lk);
+    }
+
+    // If we are shutting down, we should not have any items left,
+    // since the shutdown cancels all items
+    assert(m_items.size() == 0);
+  }
+
+  std::pair<bool, Clock::time_point> calcWaitTime_lock() {
+    while (m_items.size()) {
+      if (m_items.top().handler) {
+        // Item present, so return the new wait time
+        return std::make_pair(true, m_items.top().end);
+      } else {
+        // Discard empty handlers (they were cancelled)
+        m_items.pop();
+      }
+    }
+
+    // No items found, so return no wait time (causes the thread to wait
+    // indefinitely)
+    return std::make_pair(false, Clock::time_point());
+  }
+
+  void checkWork(std::unique_lock<std::mutex>* lk) {
+    while (m_items.size() && m_items.top().end <= Clock::now()) {
+      WorkItem item(m_items.top());
+      m_items.pop();
+
+      if (item.handler) {
+        (*lk).unlock();
+        auto reschedule_pair = item.handler(item.id == 0);
+        (*lk).lock();
+        if (!m_cancel && reschedule_pair.first) {
+          int64_t new_period = (reschedule_pair.second == -1)
+                                   ? item.period
+                                   : reschedule_pair.second;
+
+          item.period = new_period;
+          item.end = Clock::now() + std::chrono::milliseconds(new_period);
+          m_items.push(std::move(item));
+        }
+      }
+    }
+  }
+
+  bool m_finish = false;
+  bool m_cancel = false;
+  uint64_t m_idcounter = 0;
+  std::condition_variable m_checkWork;
+
+  struct WorkItem {
+    Clock::time_point end;
+    int64_t period;
+    uint64_t id;  // id==0 means it was cancelled
+    std::function<std::pair<bool, int64_t>(bool)> handler;
+    bool operator>(const WorkItem& other) const { return end > other.end; }
+  };
+
+  std::mutex m_mtx;
+  // Inheriting from priority_queue, so we can access the internal container
+  class Queue : public std::priority_queue<WorkItem, std::vector<WorkItem>,
+                                           std::greater<WorkItem>> {
+   public:
+    std::vector<WorkItem>& getContainer() { return this->c; }
+  } m_items;
+  ROCKSDB_NAMESPACE::port::Thread m_th;
+  bool closed_ = false;
+};
diff --git a/src/rocksdb/util/timer_queue_test.cc b/src/rocksdb/util/timer_queue_test.cc
new file mode 100644
index 000000000..b3c3768ec
--- /dev/null
+++ b/src/rocksdb/util/timer_queue_test.cc
@@ -0,0 +1,73 @@
+//  Portions Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+// borrowed from
+// http://www.crazygaze.com/blog/2016/03/24/portable-c-timer-queue/
+// Timer Queue
+//
+// License
+//
+// The source code in this article is licensed under the CC0 license, so feel
+// free
+// to copy, modify, share, do whatever you want with it.
+// No attribution is required, but Ill be happy if you do.
+// CC0 license
+
+// The person who associated a work with this deed has dedicated the work to the
+// public domain by waiving all of his or her rights to the work worldwide
+// under copyright law, including all related and neighboring rights, to the
+// extent allowed by law.  You can copy, modify, distribute and perform the
+// work, even for
+// commercial purposes, all without asking permission. See Other Information
+// below.
+//
+
+#include "util/timer_queue.h"
+
+#include <future>
+
+namespace Timing {
+
+using Clock = std::chrono::high_resolution_clock;
+double now() {
+  static auto start = Clock::now();
+  return std::chrono::duration<double, std::milli>(Clock::now() - start)
+      .count();
+}
+
+}  // namespace Timing
+
+int main() {
+  TimerQueue q;
+
+  double tnow = Timing::now();
+
+  q.add(10000, [tnow](bool aborted) mutable {
+    printf("T 1: %d, Elapsed %4.2fms\n", aborted, Timing::now() - tnow);
+    return std::make_pair(false, 0);
+  });
+  q.add(10001, [tnow](bool aborted) mutable {
+    printf("T 2: %d, Elapsed %4.2fms\n", aborted, Timing::now() - tnow);
+    return std::make_pair(false, 0);
+  });
+
+  q.add(1000, [tnow](bool aborted) mutable {
+    printf("T 3: %d, Elapsed %4.2fms\n", aborted, Timing::now() - tnow);
+    return std::make_pair(!aborted, 1000);
+  });
+
+  auto id = q.add(2000, [tnow](bool aborted) mutable {
+    printf("T 4: %d, Elapsed %4.2fms\n", aborted, Timing::now() - tnow);
+    return std::make_pair(!aborted, 2000);
+  });
+
+  (void)id;
+  // auto ret = q.cancel(id);
+  // assert(ret == 1);
+  // q.cancelAll();
+
+  return 0;
+}
+//////////////////////////////////////////
diff --git a/src/rocksdb/util/timer_test.cc b/src/rocksdb/util/timer_test.cc
new file mode 100644
index 000000000..0ebfa9f3d
--- /dev/null
+++ b/src/rocksdb/util/timer_test.cc
@@ -0,0 +1,402 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#include "util/timer.h"
+
+#include "db/db_test_util.h"
+#include "rocksdb/file_system.h"
+#include "test_util/mock_time_env.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+class TimerTest : public testing::Test {
+ public:
+  TimerTest()
+      : mock_clock_(std::make_shared<MockSystemClock>(SystemClock::Default())) {
+  }
+
+ protected:
+  std::shared_ptr<MockSystemClock> mock_clock_;
+
+  void SetUp() override { mock_clock_->InstallTimedWaitFixCallback(); }
+
+  const int kUsPerSec = 1000000;
+};
+
+TEST_F(TimerTest, SingleScheduleOnce) {
+  const int kInitDelayUs = 1 * kUsPerSec;
+  Timer timer(mock_clock_.get());
+
+  int count = 0;
+  timer.Add([&] { count++; }, "fn_sch_test", kInitDelayUs, 0);
+
+  ASSERT_TRUE(timer.Start());
+
+  ASSERT_EQ(0, count);
+  // Wait for execution to finish
+  timer.TEST_WaitForRun(
+      [&] { mock_clock_->SleepForMicroseconds(kInitDelayUs); });
+  ASSERT_EQ(1, count);
+
+  ASSERT_TRUE(timer.Shutdown());
+}
+
+TEST_F(TimerTest, MultipleScheduleOnce) {
+  const int kInitDelay1Us = 1 * kUsPerSec;
+  const int kInitDelay2Us = 3 * kUsPerSec;
+  Timer timer(mock_clock_.get());
+
+  int count1 = 0;
+  timer.Add([&] { count1++; }, "fn_sch_test1", kInitDelay1Us, 0);
+
+  int count2 = 0;
+  timer.Add([&] { count2++; }, "fn_sch_test2", kInitDelay2Us, 0);
+
+  ASSERT_TRUE(timer.Start());
+  ASSERT_EQ(0, count1);
+  ASSERT_EQ(0, count2);
+
+  timer.TEST_WaitForRun(
+      [&] { mock_clock_->SleepForMicroseconds(kInitDelay1Us); });
+
+  ASSERT_EQ(1, count1);
+  ASSERT_EQ(0, count2);
+
+  timer.TEST_WaitForRun([&] {
+    mock_clock_->SleepForMicroseconds(kInitDelay2Us - kInitDelay1Us);
+  });
+
+  ASSERT_EQ(1, count1);
+  ASSERT_EQ(1, count2);
+
+  ASSERT_TRUE(timer.Shutdown());
+}
+
+TEST_F(TimerTest, SingleScheduleRepeatedly) {
+  const int kIterations = 5;
+  const int kInitDelayUs = 1 * kUsPerSec;
+  const int kRepeatUs = 1 * kUsPerSec;
+
+  Timer timer(mock_clock_.get());
+  int count = 0;
+  timer.Add([&] { count++; }, "fn_sch_test", kInitDelayUs, kRepeatUs);
+
+  ASSERT_TRUE(timer.Start());
+  ASSERT_EQ(0, count);
+
+  timer.TEST_WaitForRun(
+      [&] { mock_clock_->SleepForMicroseconds(kInitDelayUs); });
+
+  ASSERT_EQ(1, count);
+
+  // Wait for execution to finish
+  for (int i = 1; i < kIterations; i++) {
+    timer.TEST_WaitForRun(
+        [&] { mock_clock_->SleepForMicroseconds(kRepeatUs); });
+  }
+  ASSERT_EQ(kIterations, count);
+
+  ASSERT_TRUE(timer.Shutdown());
+}
+
+TEST_F(TimerTest, MultipleScheduleRepeatedly) {
+  const int kIterations = 5;
+  const int kInitDelay1Us = 0 * kUsPerSec;
+  const int kInitDelay2Us = 1 * kUsPerSec;
+  const int kInitDelay3Us = 0 * kUsPerSec;
+  const int kRepeatUs = 2 * kUsPerSec;
+  const int kLargeRepeatUs = 100 * kUsPerSec;
+
+  Timer timer(mock_clock_.get());
+
+  int count1 = 0;
+  timer.Add([&] { count1++; }, "fn_sch_test1", kInitDelay1Us, kRepeatUs);
+
+  int count2 = 0;
+  timer.Add([&] { count2++; }, "fn_sch_test2", kInitDelay2Us, kRepeatUs);
+
+  // Add a function with relatively large repeat interval
+  int count3 = 0;
+  timer.Add([&] { count3++; }, "fn_sch_test3", kInitDelay3Us, kLargeRepeatUs);
+
+  ASSERT_TRUE(timer.Start());
+
+  ASSERT_EQ(0, count2);
+  // Wait for execution to finish
+  for (int i = 1; i < kIterations * (kRepeatUs / kUsPerSec); i++) {
+    timer.TEST_WaitForRun(
+        [&] { mock_clock_->SleepForMicroseconds(1 * kUsPerSec); });
+    ASSERT_EQ((i + 2) / (kRepeatUs / kUsPerSec), count1);
+    ASSERT_EQ((i + 1) / (kRepeatUs / kUsPerSec), count2);
+
+    // large interval function should only run once (the first one).
+    ASSERT_EQ(1, count3);
+  }
+
+  timer.Cancel("fn_sch_test1");
+
+  // Wait for execution to finish
+  timer.TEST_WaitForRun(
+      [&] { mock_clock_->SleepForMicroseconds(1 * kUsPerSec); });
+  ASSERT_EQ(kIterations, count1);
+  ASSERT_EQ(kIterations, count2);
+  ASSERT_EQ(1, count3);
+
+  timer.Cancel("fn_sch_test2");
+
+  ASSERT_EQ(kIterations, count1);
+  ASSERT_EQ(kIterations, count2);
+
+  // execute the long interval one
+  timer.TEST_WaitForRun([&] {
+    mock_clock_->SleepForMicroseconds(
+        kLargeRepeatUs - static_cast<int>(mock_clock_->NowMicros()));
+  });
+  ASSERT_EQ(2, count3);
+
+  ASSERT_TRUE(timer.Shutdown());
+}
+
+TEST_F(TimerTest, AddAfterStartTest) {
+  const int kIterations = 5;
+  const int kInitDelayUs = 1 * kUsPerSec;
+  const int kRepeatUs = 1 * kUsPerSec;
+
+  // wait timer to run and then add a new job
+  SyncPoint::GetInstance()->LoadDependency(
+      {{"Timer::Run::Waiting", "TimerTest:AddAfterStartTest:1"}});
+  SyncPoint::GetInstance()->EnableProcessing();
+
+  Timer timer(mock_clock_.get());
+
+  ASSERT_TRUE(timer.Start());
+
+  TEST_SYNC_POINT("TimerTest:AddAfterStartTest:1");
+  int count = 0;
+  timer.Add([&] { count++; }, "fn_sch_test", kInitDelayUs, kRepeatUs);
+  ASSERT_EQ(0, count);
+  // Wait for execution to finish
+  timer.TEST_WaitForRun(
+      [&] { mock_clock_->SleepForMicroseconds(kInitDelayUs); });
+  ASSERT_EQ(1, count);
+
+  for (int i = 1; i < kIterations; i++) {
+    timer.TEST_WaitForRun(
+        [&] { mock_clock_->SleepForMicroseconds(kRepeatUs); });
+  }
+  ASSERT_EQ(kIterations, count);
+
+  ASSERT_TRUE(timer.Shutdown());
+}
+
+TEST_F(TimerTest, CancelRunningTask) {
+  static constexpr char kTestFuncName[] = "test_func";
+  const int kRepeatUs = 1 * kUsPerSec;
+  Timer timer(mock_clock_.get());
+  ASSERT_TRUE(timer.Start());
+  int* value = new int;
+  *value = 0;
+  SyncPoint::GetInstance()->DisableProcessing();
+  SyncPoint::GetInstance()->LoadDependency({
+      {"TimerTest::CancelRunningTask:test_func:0",
+       "TimerTest::CancelRunningTask:BeforeCancel"},
+      {"Timer::WaitForTaskCompleteIfNecessary:TaskExecuting",
+       "TimerTest::CancelRunningTask:test_func:1"},
+  });
+  SyncPoint::GetInstance()->EnableProcessing();
+  timer.Add(
+      [&]() {
+        *value = 1;
+        TEST_SYNC_POINT("TimerTest::CancelRunningTask:test_func:0");
+        TEST_SYNC_POINT("TimerTest::CancelRunningTask:test_func:1");
+      },
+      kTestFuncName, 0, kRepeatUs);
+  port::Thread control_thr([&]() {
+    TEST_SYNC_POINT("TimerTest::CancelRunningTask:BeforeCancel");
+    timer.Cancel(kTestFuncName);
+    // Verify that *value has been set to 1.
+    ASSERT_EQ(1, *value);
+    delete value;
+    value = nullptr;
+  });
+  mock_clock_->SleepForMicroseconds(kRepeatUs);
+  control_thr.join();
+  ASSERT_TRUE(timer.Shutdown());
+}
+
+TEST_F(TimerTest, ShutdownRunningTask) {
+  const int kRepeatUs = 1 * kUsPerSec;
+  constexpr char kTestFunc1Name[] = "test_func1";
+  constexpr char kTestFunc2Name[] = "test_func2";
+  Timer timer(mock_clock_.get());
+
+  SyncPoint::GetInstance()->DisableProcessing();
+  SyncPoint::GetInstance()->LoadDependency({
+      {"TimerTest::ShutdownRunningTest:test_func:0",
+       "TimerTest::ShutdownRunningTest:BeforeShutdown"},
+      {"Timer::WaitForTaskCompleteIfNecessary:TaskExecuting",
+       "TimerTest::ShutdownRunningTest:test_func:1"},
+  });
+  SyncPoint::GetInstance()->EnableProcessing();
+
+  ASSERT_TRUE(timer.Start());
+
+  int* value = new int;
+  *value = 0;
+  timer.Add(
+      [&]() {
+        TEST_SYNC_POINT("TimerTest::ShutdownRunningTest:test_func:0");
+        *value = 1;
+        TEST_SYNC_POINT("TimerTest::ShutdownRunningTest:test_func:1");
+      },
+      kTestFunc1Name, 0, kRepeatUs);
+
+  timer.Add([&]() { ++(*value); }, kTestFunc2Name, 0, kRepeatUs);
+
+  port::Thread control_thr([&]() {
+    TEST_SYNC_POINT("TimerTest::ShutdownRunningTest:BeforeShutdown");
+    timer.Shutdown();
+  });
+  mock_clock_->SleepForMicroseconds(kRepeatUs);
+  control_thr.join();
+  delete value;
+}
+
+TEST_F(TimerTest, AddSameFuncName) {
+  const int kInitDelayUs = 1 * kUsPerSec;
+  const int kRepeat1Us = 5 * kUsPerSec;
+  const int kRepeat2Us = 4 * kUsPerSec;
+
+  Timer timer(mock_clock_.get());
+  ASSERT_TRUE(timer.Start());
+
+  int func_counter1 = 0;
+  ASSERT_TRUE(timer.Add([&] { func_counter1++; }, "duplicated_func",
+                        kInitDelayUs, kRepeat1Us));
+
+  int func2_counter = 0;
+  ASSERT_TRUE(
+      timer.Add([&] { func2_counter++; }, "func2", kInitDelayUs, kRepeat2Us));
+
+  // New function with the same name should fail to add
+  int func_counter2 = 0;
+  ASSERT_FALSE(timer.Add([&] { func_counter2++; }, "duplicated_func",
+                         kInitDelayUs, kRepeat1Us));
+
+  ASSERT_EQ(0, func_counter1);
+  ASSERT_EQ(0, func2_counter);
+
+  timer.TEST_WaitForRun(
+      [&] { mock_clock_->SleepForMicroseconds(kInitDelayUs); });
+
+  ASSERT_EQ(1, func_counter1);
+  ASSERT_EQ(1, func2_counter);
+
+  timer.TEST_WaitForRun([&] { mock_clock_->SleepForMicroseconds(kRepeat1Us); });
+
+  ASSERT_EQ(2, func_counter1);
+  ASSERT_EQ(2, func2_counter);
+  ASSERT_EQ(0, func_counter2);
+
+  ASSERT_TRUE(timer.Shutdown());
+}
+
+TEST_F(TimerTest, RepeatIntervalWithFuncRunningTime) {
+  const int kInitDelayUs = 1 * kUsPerSec;
+  const int kRepeatUs = 5 * kUsPerSec;
+  const int kFuncRunningTimeUs = 1 * kUsPerSec;
+
+  Timer timer(mock_clock_.get());
+  ASSERT_TRUE(timer.Start());
+
+  int func_counter = 0;
+  timer.Add(
+      [&] {
+        mock_clock_->SleepForMicroseconds(kFuncRunningTimeUs);
+        func_counter++;
+      },
+      "func", kInitDelayUs, kRepeatUs);
+
+  ASSERT_EQ(0, func_counter);
+  timer.TEST_WaitForRun(
+      [&] { mock_clock_->SleepForMicroseconds(kInitDelayUs); });
+  ASSERT_EQ(1, func_counter);
+  ASSERT_EQ(kInitDelayUs + kFuncRunningTimeUs, mock_clock_->NowMicros());
+
+  // After repeat interval time, the function is not executed, as running
+  // the function takes some time (`kFuncRunningTimeSec`). The repeat interval
+  // is the time between ending time of the last call and starting time of the
+  // next call.
+  uint64_t next_abs_interval_time_us = kInitDelayUs + kRepeatUs;
+  timer.TEST_WaitForRun([&] {
+    mock_clock_->SetCurrentTime(next_abs_interval_time_us / kUsPerSec);
+  });
+  ASSERT_EQ(1, func_counter);
+
+  // After the function running time, it's executed again
+  timer.TEST_WaitForRun(
+      [&] { mock_clock_->SleepForMicroseconds(kFuncRunningTimeUs); });
+  ASSERT_EQ(2, func_counter);
+
+  ASSERT_TRUE(timer.Shutdown());
+}
+
+TEST_F(TimerTest, DestroyRunningTimer) {
+  const int kInitDelayUs = 1 * kUsPerSec;
+  const int kRepeatUs = 1 * kUsPerSec;
+
+  auto timer_ptr = new Timer(mock_clock_.get());
+
+  int count = 0;
+  timer_ptr->Add([&] { count++; }, "fn_sch_test", kInitDelayUs, kRepeatUs);
+  ASSERT_TRUE(timer_ptr->Start());
+
+  timer_ptr->TEST_WaitForRun(
+      [&] { mock_clock_->SleepForMicroseconds(kInitDelayUs); });
+
+  // delete a running timer should not cause any exception
+  delete timer_ptr;
+}
+
+TEST_F(TimerTest, DestroyTimerWithRunningFunc) {
+  const int kRepeatUs = 1 * kUsPerSec;
+  auto timer_ptr = new Timer(mock_clock_.get());
+
+  SyncPoint::GetInstance()->DisableProcessing();
+  SyncPoint::GetInstance()->LoadDependency({
+      {"TimerTest::DestroyTimerWithRunningFunc:test_func:0",
+       "TimerTest::DestroyTimerWithRunningFunc:BeforeDelete"},
+      {"Timer::WaitForTaskCompleteIfNecessary:TaskExecuting",
+       "TimerTest::DestroyTimerWithRunningFunc:test_func:1"},
+  });
+  SyncPoint::GetInstance()->EnableProcessing();
+
+  ASSERT_TRUE(timer_ptr->Start());
+
+  int count = 0;
+  timer_ptr->Add(
+      [&]() {
+        TEST_SYNC_POINT("TimerTest::DestroyTimerWithRunningFunc:test_func:0");
+        count++;
+        TEST_SYNC_POINT("TimerTest::DestroyTimerWithRunningFunc:test_func:1");
+      },
+      "fn_running_test", 0, kRepeatUs);
+
+  port::Thread control_thr([&] {
+    TEST_SYNC_POINT("TimerTest::DestroyTimerWithRunningFunc:BeforeDelete");
+    delete timer_ptr;
+  });
+  mock_clock_->SleepForMicroseconds(kRepeatUs);
+  control_thr.join();
+}
+
+}  // namespace ROCKSDB_NAMESPACE
+
+int main(int argc, char** argv) {
+  ROCKSDB_NAMESPACE::port::InstallStackTraceHandler();
+  ::testing::InitGoogleTest(&argc, argv);
+
+  return RUN_ALL_TESTS();
+}
diff --git a/src/rocksdb/util/user_comparator_wrapper.h b/src/rocksdb/util/user_comparator_wrapper.h
new file mode 100644
index 000000000..59ebada12
--- /dev/null
+++ b/src/rocksdb/util/user_comparator_wrapper.h
@@ -0,0 +1,64 @@
+// Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+// Copyright (c) 2011 The LevelDB Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file. See the AUTHORS file for names of contributors.
+
+#pragma once
+
+#include "monitoring/perf_context_imp.h"
+#include "rocksdb/comparator.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+// Wrapper of user comparator, with auto increment to
+// perf_context.user_key_comparison_count.
+class UserComparatorWrapper {
+ public:
+  // `UserComparatorWrapper`s constructed with the default constructor are not
+  // usable and will segfault on any attempt to use them for comparisons.
+  UserComparatorWrapper() : user_comparator_(nullptr) {}
+
+  explicit UserComparatorWrapper(const Comparator* const user_cmp)
+      : user_comparator_(user_cmp) {}
+
+  ~UserComparatorWrapper() = default;
+
+  const Comparator* user_comparator() const { return user_comparator_; }
+
+  int Compare(const Slice& a, const Slice& b) const {
+    PERF_COUNTER_ADD(user_key_comparison_count, 1);
+    return user_comparator_->Compare(a, b);
+  }
+
+  bool Equal(const Slice& a, const Slice& b) const {
+    PERF_COUNTER_ADD(user_key_comparison_count, 1);
+    return user_comparator_->Equal(a, b);
+  }
+
+  int CompareTimestamp(const Slice& ts1, const Slice& ts2) const {
+    return user_comparator_->CompareTimestamp(ts1, ts2);
+  }
+
+  int CompareWithoutTimestamp(const Slice& a, const Slice& b) const {
+    PERF_COUNTER_ADD(user_key_comparison_count, 1);
+    return user_comparator_->CompareWithoutTimestamp(a, b);
+  }
+
+  int CompareWithoutTimestamp(const Slice& a, bool a_has_ts, const Slice& b,
+                              bool b_has_ts) const {
+    PERF_COUNTER_ADD(user_key_comparison_count, 1);
+    return user_comparator_->CompareWithoutTimestamp(a, a_has_ts, b, b_has_ts);
+  }
+
+  bool EqualWithoutTimestamp(const Slice& a, const Slice& b) const {
+    return user_comparator_->EqualWithoutTimestamp(a, b);
+  }
+
+ private:
+  const Comparator* user_comparator_;
+};
+
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/util/vector_iterator.h b/src/rocksdb/util/vector_iterator.h
new file mode 100644
index 000000000..c4cc01d56
--- /dev/null
+++ b/src/rocksdb/util/vector_iterator.h
@@ -0,0 +1,118 @@
+// Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.
+#pragma once
+
+#include <algorithm>
+#include <string>
+#include <vector>
+
+#include "db/dbformat.h"
+#include "rocksdb/comparator.h"
+#include "rocksdb/iterator.h"
+#include "rocksdb/slice.h"
+#include "table/internal_iterator.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+// Iterator over a vector of keys/values
+class VectorIterator : public InternalIterator {
+ public:
+  VectorIterator(std::vector<std::string> keys, std::vector<std::string> values,
+                 const CompareInterface* icmp = nullptr)
+      : keys_(std::move(keys)),
+        values_(std::move(values)),
+        current_(keys_.size()),
+        indexed_cmp_(icmp, &keys_) {
+    assert(keys_.size() == values_.size());
+
+    indices_.reserve(keys_.size());
+    for (size_t i = 0; i < keys_.size(); i++) {
+      indices_.push_back(i);
+    }
+    if (icmp != nullptr) {
+      std::sort(indices_.begin(), indices_.end(), indexed_cmp_);
+    }
+  }
+
+  virtual bool Valid() const override {
+    return !indices_.empty() && current_ < indices_.size();
+  }
+
+  virtual void SeekToFirst() override { current_ = 0; }
+  virtual void SeekToLast() override { current_ = indices_.size() - 1; }
+
+  virtual void Seek(const Slice& target) override {
+    if (indexed_cmp_.cmp != nullptr) {
+      current_ = std::lower_bound(indices_.begin(), indices_.end(), target,
+                                  indexed_cmp_) -
+                 indices_.begin();
+    } else {
+      current_ =
+          std::lower_bound(keys_.begin(), keys_.end(), target.ToString()) -
+          keys_.begin();
+    }
+  }
+
+  virtual void SeekForPrev(const Slice& target) override {
+    if (indexed_cmp_.cmp != nullptr) {
+      current_ = std::upper_bound(indices_.begin(), indices_.end(), target,
+                                  indexed_cmp_) -
+                 indices_.begin();
+    } else {
+      current_ =
+          std::upper_bound(keys_.begin(), keys_.end(), target.ToString()) -
+          keys_.begin();
+    }
+    if (!Valid()) {
+      SeekToLast();
+    } else {
+      Prev();
+    }
+  }
+
+  virtual void Next() override { current_++; }
+  virtual void Prev() override { current_--; }
+
+  virtual Slice key() const override {
+    return Slice(keys_[indices_[current_]]);
+  }
+  virtual Slice value() const override {
+    return Slice(values_[indices_[current_]]);
+  }
+
+  virtual Status status() const override { return Status::OK(); }
+
+  virtual bool IsKeyPinned() const override { return true; }
+  virtual bool IsValuePinned() const override { return true; }
+
+ protected:
+  std::vector<std::string> keys_;
+  std::vector<std::string> values_;
+  size_t current_;
+
+ private:
+  struct IndexedKeyComparator {
+    IndexedKeyComparator(const CompareInterface* c,
+                         const std::vector<std::string>* ks)
+        : cmp(c), keys(ks) {}
+
+    bool operator()(size_t a, size_t b) const {
+      return cmp->Compare((*keys)[a], (*keys)[b]) < 0;
+    }
+
+    bool operator()(size_t a, const Slice& b) const {
+      return cmp->Compare((*keys)[a], b) < 0;
+    }
+
+    bool operator()(const Slice& a, size_t b) const {
+      return cmp->Compare(a, (*keys)[b]) < 0;
+    }
+
+    const CompareInterface* cmp;
+    const std::vector<std::string>* keys;
+  };
+
+  IndexedKeyComparator indexed_cmp_;
+  std::vector<size_t> indices_;
+};
+
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/util/work_queue.h b/src/rocksdb/util/work_queue.h
new file mode 100644
index 000000000..94ece85d9
--- /dev/null
+++ b/src/rocksdb/util/work_queue.h
@@ -0,0 +1,150 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+/*
+ * Copyright (c) 2016-present, Facebook, Inc.
+ * All rights reserved.
+ *
+ * This source code is licensed under both the BSD-style license (found in the
+ * LICENSE file in the root directory of this source tree) and the GPLv2 (found
+ * in the COPYING file in the root directory of this source tree).
+ */
+#pragma once
+
+#include <atomic>
+#include <cassert>
+#include <condition_variable>
+#include <cstddef>
+#include <functional>
+#include <mutex>
+#include <queue>
+
+#include "rocksdb/rocksdb_namespace.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+/// Unbounded thread-safe work queue.
+//
+// This file is an excerpt from Facebook's zstd repo at
+// https://github.com/facebook/zstd/. The relevant file is
+// contrib/pzstd/utils/WorkQueue.h.
+
+template <typename T>
+class WorkQueue {
+  // Protects all member variable access
+  std::mutex mutex_;
+  std::condition_variable readerCv_;
+  std::condition_variable writerCv_;
+  std::condition_variable finishCv_;
+
+  std::queue<T> queue_;
+  bool done_;
+  std::size_t maxSize_;
+
+  // Must have lock to call this function
+  bool full() const {
+    if (maxSize_ == 0) {
+      return false;
+    }
+    return queue_.size() >= maxSize_;
+  }
+
+ public:
+  /**
+   * Constructs an empty work queue with an optional max size.
+   * If `maxSize == 0` the queue size is unbounded.
+   *
+   * @param maxSize The maximum allowed size of the work queue.
+   */
+  WorkQueue(std::size_t maxSize = 0) : done_(false), maxSize_(maxSize) {}
+
+  /**
+   * Push an item onto the work queue.  Notify a single thread that work is
+   * available.  If `finish()` has been called, do nothing and return false.
+   * If `push()` returns false, then `item` has not been copied from.
+   *
+   * @param item  Item to push onto the queue.
+   * @returns     True upon success, false if `finish()` has been called.  An
+   *               item was pushed iff `push()` returns true.
+   */
+  template <typename U>
+  bool push(U&& item) {
+    {
+      std::unique_lock<std::mutex> lock(mutex_);
+      while (full() && !done_) {
+        writerCv_.wait(lock);
+      }
+      if (done_) {
+        return false;
+      }
+      queue_.push(std::forward<U>(item));
+    }
+    readerCv_.notify_one();
+    return true;
+  }
+
+  /**
+   * Attempts to pop an item off the work queue.  It will block until data is
+   * available or `finish()` has been called.
+   *
+   * @param[out] item  If `pop` returns `true`, it contains the popped item.
+   *                    If `pop` returns `false`, it is unmodified.
+   * @returns          True upon success.  False if the queue is empty and
+   *                    `finish()` has been called.
+   */
+  bool pop(T& item) {
+    {
+      std::unique_lock<std::mutex> lock(mutex_);
+      while (queue_.empty() && !done_) {
+        readerCv_.wait(lock);
+      }
+      if (queue_.empty()) {
+        assert(done_);
+        return false;
+      }
+      item = queue_.front();
+      queue_.pop();
+    }
+    writerCv_.notify_one();
+    return true;
+  }
+
+  /**
+   * Sets the maximum queue size.  If `maxSize == 0` then it is unbounded.
+   *
+   * @param maxSize The new maximum queue size.
+   */
+  void setMaxSize(std::size_t maxSize) {
+    {
+      std::lock_guard<std::mutex> lock(mutex_);
+      maxSize_ = maxSize;
+    }
+    writerCv_.notify_all();
+  }
+
+  /**
+   * Promise that `push()` won't be called again, so once the queue is empty
+   * there will never any more work.
+   */
+  void finish() {
+    {
+      std::lock_guard<std::mutex> lock(mutex_);
+      assert(!done_);
+      done_ = true;
+    }
+    readerCv_.notify_all();
+    writerCv_.notify_all();
+    finishCv_.notify_all();
+  }
+
+  /// Blocks until `finish()` has been called (but the queue may not be empty).
+  void waitUntilFinished() {
+    std::unique_lock<std::mutex> lock(mutex_);
+    while (!done_) {
+      finishCv_.wait(lock);
+    }
+  }
+};
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/util/work_queue_test.cc b/src/rocksdb/util/work_queue_test.cc
new file mode 100644
index 000000000..c23a51279
--- /dev/null
+++ b/src/rocksdb/util/work_queue_test.cc
@@ -0,0 +1,272 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+/*
+ * Copyright (c) 2016-present, Facebook, Inc.
+ * All rights reserved.
+ *
+ * This source code is licensed under both the BSD-style license (found in the
+ * LICENSE file in the root directory of this source tree) and the GPLv2 (found
+ * in the COPYING file in the root directory of this source tree).
+ */
+#include "util/work_queue.h"
+
+#include <gtest/gtest.h>
+
+#include <iostream>
+#include <memory>
+#include <mutex>
+#include <thread>
+#include <vector>
+
+#include "port/stack_trace.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+// Unit test for work_queue.h.
+//
+// This file is an excerpt from Facebook's zstd repo at
+// https://github.com/facebook/zstd/. The relevant file is
+// contrib/pzstd/utils/test/WorkQueueTest.cpp.
+
+struct Popper {
+  WorkQueue<int>* queue;
+  int* results;
+  std::mutex* mutex;
+
+  void operator()() {
+    int result;
+    while (queue->pop(result)) {
+      std::lock_guard<std::mutex> lock(*mutex);
+      results[result] = result;
+    }
+  }
+};
+
+TEST(WorkQueue, SingleThreaded) {
+  WorkQueue<int> queue;
+  int result;
+
+  queue.push(5);
+  EXPECT_TRUE(queue.pop(result));
+  EXPECT_EQ(5, result);
+
+  queue.push(1);
+  queue.push(2);
+  EXPECT_TRUE(queue.pop(result));
+  EXPECT_EQ(1, result);
+  EXPECT_TRUE(queue.pop(result));
+  EXPECT_EQ(2, result);
+
+  queue.push(1);
+  queue.push(2);
+  queue.finish();
+  EXPECT_TRUE(queue.pop(result));
+  EXPECT_EQ(1, result);
+  EXPECT_TRUE(queue.pop(result));
+  EXPECT_EQ(2, result);
+  EXPECT_FALSE(queue.pop(result));
+
+  queue.waitUntilFinished();
+}
+
+TEST(WorkQueue, SPSC) {
+  WorkQueue<int> queue;
+  const int max = 100;
+
+  for (int i = 0; i < 10; ++i) {
+    queue.push(i);
+  }
+
+  std::thread thread([&queue, max] {
+    int result;
+    for (int i = 0;; ++i) {
+      if (!queue.pop(result)) {
+        EXPECT_EQ(i, max);
+        break;
+      }
+      EXPECT_EQ(i, result);
+    }
+  });
+
+  std::this_thread::yield();
+  for (int i = 10; i < max; ++i) {
+    queue.push(i);
+  }
+  queue.finish();
+
+  thread.join();
+}
+
+TEST(WorkQueue, SPMC) {
+  WorkQueue<int> queue;
+  std::vector<int> results(50, -1);
+  std::mutex mutex;
+  std::vector<std::thread> threads;
+  for (int i = 0; i < 5; ++i) {
+    threads.emplace_back(Popper{&queue, results.data(), &mutex});
+  }
+
+  for (int i = 0; i < 50; ++i) {
+    queue.push(i);
+  }
+  queue.finish();
+
+  for (auto& thread : threads) {
+    thread.join();
+  }
+
+  for (int i = 0; i < 50; ++i) {
+    EXPECT_EQ(i, results[i]);
+  }
+}
+
+TEST(WorkQueue, MPMC) {
+  WorkQueue<int> queue;
+  std::vector<int> results(100, -1);
+  std::mutex mutex;
+  std::vector<std::thread> popperThreads;
+  for (int i = 0; i < 4; ++i) {
+    popperThreads.emplace_back(Popper{&queue, results.data(), &mutex});
+  }
+
+  std::vector<std::thread> pusherThreads;
+  for (int i = 0; i < 2; ++i) {
+    auto min = i * 50;
+    auto max = (i + 1) * 50;
+    pusherThreads.emplace_back([&queue, min, max] {
+      for (int j = min; j < max; ++j) {
+        queue.push(j);
+      }
+    });
+  }
+
+  for (auto& thread : pusherThreads) {
+    thread.join();
+  }
+  queue.finish();
+
+  for (auto& thread : popperThreads) {
+    thread.join();
+  }
+
+  for (int i = 0; i < 100; ++i) {
+    EXPECT_EQ(i, results[i]);
+  }
+}
+
+TEST(WorkQueue, BoundedSizeWorks) {
+  WorkQueue<int> queue(1);
+  int result;
+  queue.push(5);
+  queue.pop(result);
+  queue.push(5);
+  queue.pop(result);
+  queue.push(5);
+  queue.finish();
+  queue.pop(result);
+  EXPECT_EQ(5, result);
+}
+
+TEST(WorkQueue, BoundedSizePushAfterFinish) {
+  WorkQueue<int> queue(1);
+  int result;
+  queue.push(5);
+  std::thread pusher([&queue] { queue.push(6); });
+  // Dirtily try and make sure that pusher has run.
+  std::this_thread::sleep_for(std::chrono::seconds(1));
+  queue.finish();
+  EXPECT_TRUE(queue.pop(result));
+  EXPECT_EQ(5, result);
+  EXPECT_FALSE(queue.pop(result));
+
+  pusher.join();
+}
+
+TEST(WorkQueue, SetMaxSize) {
+  WorkQueue<int> queue(2);
+  int result;
+  queue.push(5);
+  queue.push(6);
+  queue.setMaxSize(1);
+  std::thread pusher([&queue] { queue.push(7); });
+  // Dirtily try and make sure that pusher has run.
+  std::this_thread::sleep_for(std::chrono::seconds(1));
+  queue.finish();
+  EXPECT_TRUE(queue.pop(result));
+  EXPECT_EQ(5, result);
+  EXPECT_TRUE(queue.pop(result));
+  EXPECT_EQ(6, result);
+  EXPECT_FALSE(queue.pop(result));
+
+  pusher.join();
+}
+
+TEST(WorkQueue, BoundedSizeMPMC) {
+  WorkQueue<int> queue(10);
+  std::vector<int> results(200, -1);
+  std::mutex mutex;
+  std::cerr << "Creating popperThreads" << std::endl;
+  std::vector<std::thread> popperThreads;
+  for (int i = 0; i < 4; ++i) {
+    popperThreads.emplace_back(Popper{&queue, results.data(), &mutex});
+  }
+
+  std::cerr << "Creating pusherThreads" << std::endl;
+  std::vector<std::thread> pusherThreads;
+  for (int i = 0; i < 2; ++i) {
+    auto min = i * 100;
+    auto max = (i + 1) * 100;
+    pusherThreads.emplace_back([&queue, min, max] {
+      for (int j = min; j < max; ++j) {
+        queue.push(j);
+      }
+    });
+  }
+
+  std::cerr << "Joining pusherThreads" << std::endl;
+  for (auto& thread : pusherThreads) {
+    thread.join();
+  }
+  std::cerr << "Finishing queue" << std::endl;
+  queue.finish();
+
+  std::cerr << "Joining popperThreads" << std::endl;
+  for (auto& thread : popperThreads) {
+    thread.join();
+  }
+
+  std::cerr << "Inspecting results" << std::endl;
+  for (int i = 0; i < 200; ++i) {
+    EXPECT_EQ(i, results[i]);
+  }
+}
+
+TEST(WorkQueue, FailedPush) {
+  WorkQueue<int> queue;
+  EXPECT_TRUE(queue.push(1));
+  queue.finish();
+  EXPECT_FALSE(queue.push(1));
+}
+
+TEST(WorkQueue, FailedPop) {
+  WorkQueue<int> queue;
+  int x = 5;
+  EXPECT_TRUE(queue.push(x));
+  queue.finish();
+  x = 0;
+  EXPECT_TRUE(queue.pop(x));
+  EXPECT_EQ(5, x);
+  EXPECT_FALSE(queue.pop(x));
+  EXPECT_EQ(5, x);
+}
+
+}  // namespace ROCKSDB_NAMESPACE
+
+int main(int argc, char** argv) {
+  ROCKSDB_NAMESPACE::port::InstallStackTraceHandler();
+  ::testing::InitGoogleTest(&argc, argv);
+  return RUN_ALL_TESTS();
+}
diff --git a/src/rocksdb/util/xxhash.cc b/src/rocksdb/util/xxhash.cc
new file mode 100644
index 000000000..88852c330
--- /dev/null
+++ b/src/rocksdb/util/xxhash.cc
@@ -0,0 +1,48 @@
+//  Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+/*
+ * xxHash - Extremely Fast Hash algorithm
+ * Copyright (C) 2012-2020 Yann Collet
+ *
+ * BSD 2-Clause License (https://www.opensource.org/licenses/bsd-license.php)
+ *
+ * Redistribution and use in source and binary forms, with or without
+ * modification, are permitted provided that the following conditions are
+ * met:
+ *
+ *    * Redistributions of source code must retain the above copyright
+ *      notice, this list of conditions and the following disclaimer.
+ *    * Redistributions in binary form must reproduce the above
+ *      copyright notice, this list of conditions and the following disclaimer
+ *      in the documentation and/or other materials provided with the
+ *      distribution.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+ * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+ * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+ * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+ * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+ * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+ * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+ * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+ * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ * You can contact the author at:
+ *   - xxHash homepage: https://www.xxhash.com
+ *   - xxHash source repository: https://github.com/Cyan4973/xxHash
+ */
+
+/*
+ * xxhash.c instantiates functions defined in xxhash.h
+ */
+// clang-format off
+#ifndef XXH_STATIC_LINKING_ONLY
+#define XXH_STATIC_LINKING_ONLY   /* access advanced declarations */
+#endif // !defined(XXH_STATIC_LINKING_ONLY)
+#define XXH_IMPLEMENTATION   /* access definitions */
+
+#include "xxhash.h"
diff --git a/src/rocksdb/util/xxhash.h b/src/rocksdb/util/xxhash.h
new file mode 100644
index 000000000..195f06b39
--- /dev/null
+++ b/src/rocksdb/util/xxhash.h
@@ -0,0 +1,5346 @@
+//  Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+/* BEGIN RocksDB customizations */
+#ifndef XXH_STATIC_LINKING_ONLY
+// Using compiled xxhash.cc
+#define XXH_STATIC_LINKING_ONLY 1
+#endif  // !defined(XXH_STATIC_LINKING_ONLY)
+#ifndef XXH_NAMESPACE
+#define XXH_NAMESPACE ROCKSDB_
+#endif  // !defined(XXH_NAMESPACE)
+
+// for FALLTHROUGH_INTENDED, inserted as appropriate
+#include "port/lang.h"
+/* END RocksDB customizations */
+
+// clang-format off
+/*
+ * xxHash - Extremely Fast Hash algorithm
+ * Header File
+ * Copyright (C) 2012-2020 Yann Collet
+ *
+ * BSD 2-Clause License (https://www.opensource.org/licenses/bsd-license.php)
+ *
+ * Redistribution and use in source and binary forms, with or without
+ * modification, are permitted provided that the following conditions are
+ * met:
+ *
+ *    * Redistributions of source code must retain the above copyright
+ *      notice, this list of conditions and the following disclaimer.
+ *    * Redistributions in binary form must reproduce the above
+ *      copyright notice, this list of conditions and the following disclaimer
+ *      in the documentation and/or other materials provided with the
+ *      distribution.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+ * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+ * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+ * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+ * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+ * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+ * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+ * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+ * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ * You can contact the author at:
+ *   - xxHash homepage: https://www.xxhash.com
+ *   - xxHash source repository: https://github.com/Cyan4973/xxHash
+ */
+/*!
+ * @mainpage xxHash
+ *
+ * @file xxhash.h
+ * xxHash prototypes and implementation
+ */
+/* TODO: update */
+/* Notice extracted from xxHash homepage:
+
+xxHash is an extremely fast hash algorithm, running at RAM speed limits.
+It also successfully passes all tests from the SMHasher suite.
+
+Comparison (single thread, Windows Seven 32 bits, using SMHasher on a Core 2 Duo @3GHz)
+
+Name            Speed       Q.Score   Author
+xxHash          5.4 GB/s     10
+CrapWow         3.2 GB/s      2       Andrew
+MurmurHash 3a   2.7 GB/s     10       Austin Appleby
+SpookyHash      2.0 GB/s     10       Bob Jenkins
+SBox            1.4 GB/s      9       Bret Mulvey
+Lookup3         1.2 GB/s      9       Bob Jenkins
+SuperFastHash   1.2 GB/s      1       Paul Hsieh
+CityHash64      1.05 GB/s    10       Pike & Alakuijala
+FNV             0.55 GB/s     5       Fowler, Noll, Vo
+CRC32           0.43 GB/s     9
+MD5-32          0.33 GB/s    10       Ronald L. Rivest
+SHA1-32         0.28 GB/s    10
+
+Q.Score is a measure of quality of the hash function.
+It depends on successfully passing SMHasher test set.
+10 is a perfect score.
+
+Note: SMHasher's CRC32 implementation is not the fastest one.
+Other speed-oriented implementations can be faster,
+especially in combination with PCLMUL instruction:
+https://fastcompression.blogspot.com/2019/03/presenting-xxh3.html?showComment=1552696407071#c3490092340461170735
+
+A 64-bit version, named XXH64, is available since r35.
+It offers much better speed, but for 64-bit applications only.
+Name     Speed on 64 bits    Speed on 32 bits
+XXH64       13.8 GB/s            1.9 GB/s
+XXH32        6.8 GB/s            6.0 GB/s
+*/
+
+#if defined (__cplusplus)
+extern "C" {
+#endif
+
+/* ****************************
+ *  INLINE mode
+ ******************************/
+/*!
+ * XXH_INLINE_ALL (and XXH_PRIVATE_API)
+ * Use these build macros to inline xxhash into the target unit.
+ * Inlining improves performance on small inputs, especially when the length is
+ * expressed as a compile-time constant:
+ *
+ *      https://fastcompression.blogspot.com/2018/03/xxhash-for-small-keys-impressive-power.html
+ *
+ * It also keeps xxHash symbols private to the unit, so they are not exported.
+ *
+ * Usage:
+ *     #define XXH_INLINE_ALL
+ *     #include "xxhash.h"
+ *
+ * Do not compile and link xxhash.o as a separate object, as it is not useful.
+ */
+#if (defined(XXH_INLINE_ALL) || defined(XXH_PRIVATE_API)) \
+    && !defined(XXH_INLINE_ALL_31684351384)
+   /* this section should be traversed only once */
+#  define XXH_INLINE_ALL_31684351384
+   /* give access to the advanced API, required to compile implementations */
+#  undef XXH_STATIC_LINKING_ONLY   /* avoid macro redef */
+#  define XXH_STATIC_LINKING_ONLY
+   /* make all functions private */
+#  undef XXH_PUBLIC_API
+#  if defined(__GNUC__)
+#    define XXH_PUBLIC_API static __inline __attribute__((unused))
+#  elif defined (__cplusplus) || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */)
+#    define XXH_PUBLIC_API static inline
+#  elif defined(_MSC_VER)
+#    define XXH_PUBLIC_API static __inline
+#  else
+     /* note: this version may generate warnings for unused static functions */
+#    define XXH_PUBLIC_API static
+#  endif
+
+   /*
+    * This part deals with the special case where a unit wants to inline xxHash,
+    * but "xxhash.h" has previously been included without XXH_INLINE_ALL, such
+    * as part of some previously included *.h header file.
+    * Without further action, the new include would just be ignored,
+    * and functions would effectively _not_ be inlined (silent failure).
+    * The following macros solve this situation by prefixing all inlined names,
+    * avoiding naming collision with previous inclusions.
+    */
+#  ifdef XXH_NAMESPACE
+#    error "XXH_INLINE_ALL with XXH_NAMESPACE is not supported"
+     /*
+      * Note: Alternative: #undef all symbols (it's a pretty large list).
+      * Without #error: it compiles, but functions are actually not inlined.
+      */
+#  endif
+#  define XXH_NAMESPACE XXH_INLINE_
+   /*
+    * Some identifiers (enums, type names) are not symbols, but they must
+    * still be renamed to avoid redeclaration.
+    * Alternative solution: do not redeclare them.
+    * However, this requires some #ifdefs, and is a more dispersed action.
+    * Meanwhile, renaming can be achieved in a single block
+    */
+#  define XXH_IPREF(Id)   XXH_INLINE_ ## Id
+#  define XXH_OK XXH_IPREF(XXH_OK)
+#  define XXH_ERROR XXH_IPREF(XXH_ERROR)
+#  define XXH_errorcode XXH_IPREF(XXH_errorcode)
+#  define XXH32_canonical_t  XXH_IPREF(XXH32_canonical_t)
+#  define XXH64_canonical_t  XXH_IPREF(XXH64_canonical_t)
+#  define XXH128_canonical_t XXH_IPREF(XXH128_canonical_t)
+#  define XXH32_state_s XXH_IPREF(XXH32_state_s)
+#  define XXH32_state_t XXH_IPREF(XXH32_state_t)
+#  define XXH64_state_s XXH_IPREF(XXH64_state_s)
+#  define XXH64_state_t XXH_IPREF(XXH64_state_t)
+#  define XXH3_state_s  XXH_IPREF(XXH3_state_s)
+#  define XXH3_state_t  XXH_IPREF(XXH3_state_t)
+#  define XXH128_hash_t XXH_IPREF(XXH128_hash_t)
+   /* Ensure the header is parsed again, even if it was previously included */
+#  undef XXHASH_H_5627135585666179
+#  undef XXHASH_H_STATIC_13879238742
+#endif /* XXH_INLINE_ALL || XXH_PRIVATE_API */
+
+
+
+/* ****************************************************************
+ *  Stable API
+ *****************************************************************/
+#ifndef XXHASH_H_5627135585666179
+#define XXHASH_H_5627135585666179 1
+
+
+/*!
+ * @defgroup public Public API
+ * Contains details on the public xxHash functions.
+ * @{
+ */
+/* specific declaration modes for Windows */
+#if !defined(XXH_INLINE_ALL) && !defined(XXH_PRIVATE_API)
+#  if defined(WIN32) && defined(_MSC_VER) && (defined(XXH_IMPORT) || defined(XXH_EXPORT))
+#    ifdef XXH_EXPORT
+#      define XXH_PUBLIC_API __declspec(dllexport)
+#    elif XXH_IMPORT
+#      define XXH_PUBLIC_API __declspec(dllimport)
+#    endif
+#  else
+#    define XXH_PUBLIC_API   /* do nothing */
+#  endif
+#endif
+
+#ifdef XXH_DOXYGEN
+/*!
+ * @brief Emulate a namespace by transparently prefixing all symbols.
+ *
+ * If you want to include _and expose_ xxHash functions from within your own
+ * library, but also want to avoid symbol collisions with other libraries which
+ * may also include xxHash, you can use XXH_NAMESPACE to automatically prefix
+ * any public symbol from xxhash library with the value of XXH_NAMESPACE
+ * (therefore, avoid empty or numeric values).
+ *
+ * Note that no change is required within the calling program as long as it
+ * includes `xxhash.h`: Regular symbol names will be automatically translated
+ * by this header.
+ */
+#  define XXH_NAMESPACE /* YOUR NAME HERE */
+#  undef XXH_NAMESPACE
+#endif
+
+#ifdef XXH_NAMESPACE
+#  define XXH_CAT(A,B) A##B
+#  define XXH_NAME2(A,B) XXH_CAT(A,B)
+#  define XXH_versionNumber XXH_NAME2(XXH_NAMESPACE, XXH_versionNumber)
+/* XXH32 */
+#  define XXH32 XXH_NAME2(XXH_NAMESPACE, XXH32)
+#  define XXH32_createState XXH_NAME2(XXH_NAMESPACE, XXH32_createState)
+#  define XXH32_freeState XXH_NAME2(XXH_NAMESPACE, XXH32_freeState)
+#  define XXH32_reset XXH_NAME2(XXH_NAMESPACE, XXH32_reset)
+#  define XXH32_update XXH_NAME2(XXH_NAMESPACE, XXH32_update)
+#  define XXH32_digest XXH_NAME2(XXH_NAMESPACE, XXH32_digest)
+#  define XXH32_copyState XXH_NAME2(XXH_NAMESPACE, XXH32_copyState)
+#  define XXH32_canonicalFromHash XXH_NAME2(XXH_NAMESPACE, XXH32_canonicalFromHash)
+#  define XXH32_hashFromCanonical XXH_NAME2(XXH_NAMESPACE, XXH32_hashFromCanonical)
+/* XXH64 */
+#  define XXH64 XXH_NAME2(XXH_NAMESPACE, XXH64)
+#  define XXH64_createState XXH_NAME2(XXH_NAMESPACE, XXH64_createState)
+#  define XXH64_freeState XXH_NAME2(XXH_NAMESPACE, XXH64_freeState)
+#  define XXH64_reset XXH_NAME2(XXH_NAMESPACE, XXH64_reset)
+#  define XXH64_update XXH_NAME2(XXH_NAMESPACE, XXH64_update)
+#  define XXH64_digest XXH_NAME2(XXH_NAMESPACE, XXH64_digest)
+#  define XXH64_copyState XXH_NAME2(XXH_NAMESPACE, XXH64_copyState)
+#  define XXH64_canonicalFromHash XXH_NAME2(XXH_NAMESPACE, XXH64_canonicalFromHash)
+#  define XXH64_hashFromCanonical XXH_NAME2(XXH_NAMESPACE, XXH64_hashFromCanonical)
+/* XXH3_64bits */
+#  define XXH3_64bits XXH_NAME2(XXH_NAMESPACE, XXH3_64bits)
+#  define XXH3_64bits_withSecret XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_withSecret)
+#  define XXH3_64bits_withSeed XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_withSeed)
+#  define XXH3_createState XXH_NAME2(XXH_NAMESPACE, XXH3_createState)
+#  define XXH3_freeState XXH_NAME2(XXH_NAMESPACE, XXH3_freeState)
+#  define XXH3_copyState XXH_NAME2(XXH_NAMESPACE, XXH3_copyState)
+#  define XXH3_64bits_reset XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_reset)
+#  define XXH3_64bits_reset_withSeed XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_reset_withSeed)
+#  define XXH3_64bits_reset_withSecret XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_reset_withSecret)
+#  define XXH3_64bits_update XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_update)
+#  define XXH3_64bits_digest XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_digest)
+#  define XXH3_generateSecret XXH_NAME2(XXH_NAMESPACE, XXH3_generateSecret)
+/* XXH3_128bits */
+#  define XXH128 XXH_NAME2(XXH_NAMESPACE, XXH128)
+#  define XXH3_128bits XXH_NAME2(XXH_NAMESPACE, XXH3_128bits)
+#  define XXH3_128bits_withSeed XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_withSeed)
+#  define XXH3_128bits_withSecret XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_withSecret)
+#  define XXH3_128bits_reset XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_reset)
+#  define XXH3_128bits_reset_withSeed XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_reset_withSeed)
+#  define XXH3_128bits_reset_withSecret XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_reset_withSecret)
+#  define XXH3_128bits_update XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_update)
+#  define XXH3_128bits_digest XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_digest)
+#  define XXH128_isEqual XXH_NAME2(XXH_NAMESPACE, XXH128_isEqual)
+#  define XXH128_cmp     XXH_NAME2(XXH_NAMESPACE, XXH128_cmp)
+#  define XXH128_canonicalFromHash XXH_NAME2(XXH_NAMESPACE, XXH128_canonicalFromHash)
+#  define XXH128_hashFromCanonical XXH_NAME2(XXH_NAMESPACE, XXH128_hashFromCanonical)
+#endif
+
+
+/* *************************************
+*  Version
+***************************************/
+#define XXH_VERSION_MAJOR    0
+#define XXH_VERSION_MINOR    8
+#define XXH_VERSION_RELEASE  1
+#define XXH_VERSION_NUMBER  (XXH_VERSION_MAJOR *100*100 + XXH_VERSION_MINOR *100 + XXH_VERSION_RELEASE)
+
+/*!
+ * @brief Obtains the xxHash version.
+ *
+ * This is only useful when xxHash is compiled as a shared library, as it is
+ * independent of the version defined in the header.
+ *
+ * @return `XXH_VERSION_NUMBER` as of when the libray was compiled.
+ */
+XXH_PUBLIC_API unsigned XXH_versionNumber (void);
+
+
+/* ****************************
+*  Definitions
+******************************/
+#include <stddef.h>   /* size_t */
+typedef enum { XXH_OK=0, XXH_ERROR } XXH_errorcode;
+
+
+/*-**********************************************************************
+*  32-bit hash
+************************************************************************/
+#if defined(XXH_DOXYGEN) /* Don't show <stdint.h> include */
+/*!
+ * @brief An unsigned 32-bit integer.
+ *
+ * Not necessarily defined to `uint32_t` but functionally equivalent.
+ */
+typedef uint32_t XXH32_hash_t;
+#elif !defined (__VMS) \
+  && (defined (__cplusplus) \
+  || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */) )
+#   include <stdint.h>
+    typedef uint32_t XXH32_hash_t;
+#else
+#   include <limits.h>
+#   if UINT_MAX == 0xFFFFFFFFUL
+      typedef unsigned int XXH32_hash_t;
+#   else
+#     if ULONG_MAX == 0xFFFFFFFFUL
+        typedef unsigned long XXH32_hash_t;
+#     else
+#       error "unsupported platform: need a 32-bit type"
+#     endif
+#   endif
+#endif
+
+/*!
+ * @}
+ *
+ * @defgroup xxh32_family XXH32 family
+ * @ingroup public
+ * Contains functions used in the classic 32-bit xxHash algorithm.
+ *
+ * @note
+ *   XXH32 is considered rather weak by today's standards.
+ *   The @ref xxh3_family provides competitive speed for both 32-bit and 64-bit
+ *   systems, and offers true 64/128 bit hash results. It provides a superior
+ *   level of dispersion, and greatly reduces the risks of collisions.
+ *
+ * @see @ref xxh64_family, @ref xxh3_family : Other xxHash families
+ * @see @ref xxh32_impl for implementation details
+ * @{
+ */
+
+/*!
+ * @brief Calculates the 32-bit hash of @p input using xxHash32.
+ *
+ * Speed on Core 2 Duo @ 3 GHz (single thread, SMHasher benchmark): 5.4 GB/s
+ *
+ * @param input The block of data to be hashed, at least @p length bytes in size.
+ * @param length The length of @p input, in bytes.
+ * @param seed The 32-bit seed to alter the hash's output predictably.
+ *
+ * @pre
+ *   The memory between @p input and @p input + @p length must be valid,
+ *   readable, contiguous memory. However, if @p length is `0`, @p input may be
+ *   `NULL`. In C++, this also must be *TriviallyCopyable*.
+ *
+ * @return The calculated 32-bit hash value.
+ *
+ * @see
+ *    XXH64(), XXH3_64bits_withSeed(), XXH3_128bits_withSeed(), XXH128():
+ *    Direct equivalents for the other variants of xxHash.
+ * @see
+ *    XXH32_createState(), XXH32_update(), XXH32_digest(): Streaming version.
+ */
+XXH_PUBLIC_API XXH32_hash_t XXH32 (const void* input, size_t length, XXH32_hash_t seed);
+
+/*!
+ * Streaming functions generate the xxHash value from an incremental input.
+ * This method is slower than single-call functions, due to state management.
+ * For small inputs, prefer `XXH32()` and `XXH64()`, which are better optimized.
+ *
+ * An XXH state must first be allocated using `XXH*_createState()`.
+ *
+ * Start a new hash by initializing the state with a seed using `XXH*_reset()`.
+ *
+ * Then, feed the hash state by calling `XXH*_update()` as many times as necessary.
+ *
+ * The function returns an error code, with 0 meaning OK, and any other value
+ * meaning there is an error.
+ *
+ * Finally, a hash value can be produced anytime, by using `XXH*_digest()`.
+ * This function returns the nn-bits hash as an int or long long.
+ *
+ * It's still possible to continue inserting input into the hash state after a
+ * digest, and generate new hash values later on by invoking `XXH*_digest()`.
+ *
+ * When done, release the state using `XXH*_freeState()`.
+ *
+ * Example code for incrementally hashing a file:
+ * @code{.c}
+ *    #include <stdio.h>
+ *    #include <xxhash.h>
+ *    #define BUFFER_SIZE 256
+ *
+ *    // Note: XXH64 and XXH3 use the same interface.
+ *    XXH32_hash_t
+ *    hashFile(FILE* stream)
+ *    {
+ *        XXH32_state_t* state;
+ *        unsigned char buf[BUFFER_SIZE];
+ *        size_t amt;
+ *        XXH32_hash_t hash;
+ *
+ *        state = XXH32_createState();       // Create a state
+ *        assert(state != NULL);             // Error check here
+ *        XXH32_reset(state, 0xbaad5eed);    // Reset state with our seed
+ *        while ((amt = fread(buf, 1, sizeof(buf), stream)) != 0) {
+ *            XXH32_update(state, buf, amt); // Hash the file in chunks
+ *        }
+ *        hash = XXH32_digest(state);        // Finalize the hash
+ *        XXH32_freeState(state);            // Clean up
+ *        return hash;
+ *    }
+ * @endcode
+ */
+
+/*!
+ * @typedef struct XXH32_state_s XXH32_state_t
+ * @brief The opaque state struct for the XXH32 streaming API.
+ *
+ * @see XXH32_state_s for details.
+ */
+typedef struct XXH32_state_s XXH32_state_t;
+
+/*!
+ * @brief Allocates an @ref XXH32_state_t.
+ *
+ * Must be freed with XXH32_freeState().
+ * @return An allocated XXH32_state_t on success, `NULL` on failure.
+ */
+XXH_PUBLIC_API XXH32_state_t* XXH32_createState(void);
+/*!
+ * @brief Frees an @ref XXH32_state_t.
+ *
+ * Must be allocated with XXH32_createState().
+ * @param statePtr A pointer to an @ref XXH32_state_t allocated with @ref XXH32_createState().
+ * @return XXH_OK.
+ */
+XXH_PUBLIC_API XXH_errorcode  XXH32_freeState(XXH32_state_t* statePtr);
+/*!
+ * @brief Copies one @ref XXH32_state_t to another.
+ *
+ * @param dst_state The state to copy to.
+ * @param src_state The state to copy from.
+ * @pre
+ *   @p dst_state and @p src_state must not be `NULL` and must not overlap.
+ */
+XXH_PUBLIC_API void XXH32_copyState(XXH32_state_t* dst_state, const XXH32_state_t* src_state);
+
+/*!
+ * @brief Resets an @ref XXH32_state_t to begin a new hash.
+ *
+ * This function resets and seeds a state. Call it before @ref XXH32_update().
+ *
+ * @param statePtr The state struct to reset.
+ * @param seed The 32-bit seed to alter the hash result predictably.
+ *
+ * @pre
+ *   @p statePtr must not be `NULL`.
+ *
+ * @return @ref XXH_OK on success, @ref XXH_ERROR on failure.
+ */
+XXH_PUBLIC_API XXH_errorcode XXH32_reset  (XXH32_state_t* statePtr, XXH32_hash_t seed);
+
+/*!
+ * @brief Consumes a block of @p input to an @ref XXH32_state_t.
+ *
+ * Call this to incrementally consume blocks of data.
+ *
+ * @param statePtr The state struct to update.
+ * @param input The block of data to be hashed, at least @p length bytes in size.
+ * @param length The length of @p input, in bytes.
+ *
+ * @pre
+ *   @p statePtr must not be `NULL`.
+ * @pre
+ *   The memory between @p input and @p input + @p length must be valid,
+ *   readable, contiguous memory. However, if @p length is `0`, @p input may be
+ *   `NULL`. In C++, this also must be *TriviallyCopyable*.
+ *
+ * @return @ref XXH_OK on success, @ref XXH_ERROR on failure.
+ */
+XXH_PUBLIC_API XXH_errorcode XXH32_update (XXH32_state_t* statePtr, const void* input, size_t length);
+
+/*!
+ * @brief Returns the calculated hash value from an @ref XXH32_state_t.
+ *
+ * @note
+ *   Calling XXH32_digest() will not affect @p statePtr, so you can update,
+ *   digest, and update again.
+ *
+ * @param statePtr The state struct to calculate the hash from.
+ *
+ * @pre
+ *  @p statePtr must not be `NULL`.
+ *
+ * @return The calculated xxHash32 value from that state.
+ */
+XXH_PUBLIC_API XXH32_hash_t  XXH32_digest (const XXH32_state_t* statePtr);
+
+/*******   Canonical representation   *******/
+
+/*
+ * The default return values from XXH functions are unsigned 32 and 64 bit
+ * integers.
+ * This the simplest and fastest format for further post-processing.
+ *
+ * However, this leaves open the question of what is the order on the byte level,
+ * since little and big endian conventions will store the same number differently.
+ *
+ * The canonical representation settles this issue by mandating big-endian
+ * convention, the same convention as human-readable numbers (large digits first).
+ *
+ * When writing hash values to storage, sending them over a network, or printing
+ * them, it's highly recommended to use the canonical representation to ensure
+ * portability across a wider range of systems, present and future.
+ *
+ * The following functions allow transformation of hash values to and from
+ * canonical format.
+ */
+
+/*!
+ * @brief Canonical (big endian) representation of @ref XXH32_hash_t.
+ */
+typedef struct {
+    unsigned char digest[4]; /*!< Hash bytes, big endian */
+} XXH32_canonical_t;
+
+/*!
+ * @brief Converts an @ref XXH32_hash_t to a big endian @ref XXH32_canonical_t.
+ *
+ * @param dst The @ref XXH32_canonical_t pointer to be stored to.
+ * @param hash The @ref XXH32_hash_t to be converted.
+ *
+ * @pre
+ *   @p dst must not be `NULL`.
+ */
+XXH_PUBLIC_API void XXH32_canonicalFromHash(XXH32_canonical_t* dst, XXH32_hash_t hash);
+
+/*!
+ * @brief Converts an @ref XXH32_canonical_t to a native @ref XXH32_hash_t.
+ *
+ * @param src The @ref XXH32_canonical_t to convert.
+ *
+ * @pre
+ *   @p src must not be `NULL`.
+ *
+ * @return The converted hash.
+ */
+XXH_PUBLIC_API XXH32_hash_t XXH32_hashFromCanonical(const XXH32_canonical_t* src);
+
+
+/*!
+ * @}
+ * @ingroup public
+ * @{
+ */
+
+#ifndef XXH_NO_LONG_LONG
+/*-**********************************************************************
+*  64-bit hash
+************************************************************************/
+#if defined(XXH_DOXYGEN) /* don't include <stdint.h> */
+/*!
+ * @brief An unsigned 64-bit integer.
+ *
+ * Not necessarily defined to `uint64_t` but functionally equivalent.
+ */
+typedef uint64_t XXH64_hash_t;
+#elif !defined (__VMS) \
+  && (defined (__cplusplus) \
+  || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */) )
+#  include <stdint.h>
+   typedef uint64_t XXH64_hash_t;
+#else
+#  include <limits.h>
+#  if defined(__LP64__) && ULONG_MAX == 0xFFFFFFFFFFFFFFFFULL
+     /* LP64 ABI says uint64_t is unsigned long */
+     typedef unsigned long XXH64_hash_t;
+#  else
+     /* the following type must have a width of 64-bit */
+     typedef unsigned long long XXH64_hash_t;
+#  endif
+#endif
+
+/*!
+ * @}
+ *
+ * @defgroup xxh64_family XXH64 family
+ * @ingroup public
+ * @{
+ * Contains functions used in the classic 64-bit xxHash algorithm.
+ *
+ * @note
+ *   XXH3 provides competitive speed for both 32-bit and 64-bit systems,
+ *   and offers true 64/128 bit hash results. It provides a superior level of
+ *   dispersion, and greatly reduces the risks of collisions.
+ */
+
+
+/*!
+ * @brief Calculates the 64-bit hash of @p input using xxHash64.
+ *
+ * This function usually runs faster on 64-bit systems, but slower on 32-bit
+ * systems (see benchmark).
+ *
+ * @param input The block of data to be hashed, at least @p length bytes in size.
+ * @param length The length of @p input, in bytes.
+ * @param seed The 64-bit seed to alter the hash's output predictably.
+ *
+ * @pre
+ *   The memory between @p input and @p input + @p length must be valid,
+ *   readable, contiguous memory. However, if @p length is `0`, @p input may be
+ *   `NULL`. In C++, this also must be *TriviallyCopyable*.
+ *
+ * @return The calculated 64-bit hash.
+ *
+ * @see
+ *    XXH32(), XXH3_64bits_withSeed(), XXH3_128bits_withSeed(), XXH128():
+ *    Direct equivalents for the other variants of xxHash.
+ * @see
+ *    XXH64_createState(), XXH64_update(), XXH64_digest(): Streaming version.
+ */
+XXH_PUBLIC_API XXH64_hash_t XXH64(const void* input, size_t length, XXH64_hash_t seed);
+
+/*******   Streaming   *******/
+/*!
+ * @brief The opaque state struct for the XXH64 streaming API.
+ *
+ * @see XXH64_state_s for details.
+ */
+typedef struct XXH64_state_s XXH64_state_t;   /* incomplete type */
+XXH_PUBLIC_API XXH64_state_t* XXH64_createState(void);
+XXH_PUBLIC_API XXH_errorcode  XXH64_freeState(XXH64_state_t* statePtr);
+XXH_PUBLIC_API void XXH64_copyState(XXH64_state_t* dst_state, const XXH64_state_t* src_state);
+
+XXH_PUBLIC_API XXH_errorcode XXH64_reset  (XXH64_state_t* statePtr, XXH64_hash_t seed);
+XXH_PUBLIC_API XXH_errorcode XXH64_update (XXH64_state_t* statePtr, const void* input, size_t length);
+XXH_PUBLIC_API XXH64_hash_t  XXH64_digest (const XXH64_state_t* statePtr);
+
+/*******   Canonical representation   *******/
+typedef struct { unsigned char digest[sizeof(XXH64_hash_t)]; } XXH64_canonical_t;
+XXH_PUBLIC_API void XXH64_canonicalFromHash(XXH64_canonical_t* dst, XXH64_hash_t hash);
+XXH_PUBLIC_API XXH64_hash_t XXH64_hashFromCanonical(const XXH64_canonical_t* src);
+
+/*!
+ * @}
+ * ************************************************************************
+ * @defgroup xxh3_family XXH3 family
+ * @ingroup public
+ * @{
+ *
+ * XXH3 is a more recent hash algorithm featuring:
+ *  - Improved speed for both small and large inputs
+ *  - True 64-bit and 128-bit outputs
+ *  - SIMD acceleration
+ *  - Improved 32-bit viability
+ *
+ * Speed analysis methodology is explained here:
+ *
+ *    https://fastcompression.blogspot.com/2019/03/presenting-xxh3.html
+ *
+ * Compared to XXH64, expect XXH3 to run approximately
+ * ~2x faster on large inputs and >3x faster on small ones,
+ * exact differences vary depending on platform.
+ *
+ * XXH3's speed benefits greatly from SIMD and 64-bit arithmetic,
+ * but does not require it.
+ * Any 32-bit and 64-bit targets that can run XXH32 smoothly
+ * can run XXH3 at competitive speeds, even without vector support.
+ * Further details are explained in the implementation.
+ *
+ * Optimized implementations are provided for AVX512, AVX2, SSE2, NEON, POWER8,
+ * ZVector and scalar targets. This can be controlled via the XXH_VECTOR macro.
+ *
+ * XXH3 implementation is portable:
+ * it has a generic C90 formulation that can be compiled on any platform,
+ * all implementations generage exactly the same hash value on all platforms.
+ * Starting from v0.8.0, it's also labelled "stable", meaning that
+ * any future version will also generate the same hash value.
+ *
+ * XXH3 offers 2 variants, _64bits and _128bits.
+ *
+ * When only 64 bits are needed, prefer invoking the _64bits variant, as it
+ * reduces the amount of mixing, resulting in faster speed on small inputs.
+ * It's also generally simpler to manipulate a scalar return type than a struct.
+ *
+ * The API supports one-shot hashing, streaming mode, and custom secrets.
+ */
+
+/*-**********************************************************************
+*  XXH3 64-bit variant
+************************************************************************/
+
+/* XXH3_64bits():
+ * default 64-bit variant, using default secret and default seed of 0.
+ * It's the fastest variant. */
+XXH_PUBLIC_API XXH64_hash_t XXH3_64bits(const void* data, size_t len);
+
+/*
+ * XXH3_64bits_withSeed():
+ * This variant generates a custom secret on the fly
+ * based on default secret altered using the `seed` value.
+ * While this operation is decently fast, note that it's not completely free.
+ * Note: seed==0 produces the same results as XXH3_64bits().
+ */
+XXH_PUBLIC_API XXH64_hash_t XXH3_64bits_withSeed(const void* data, size_t len, XXH64_hash_t seed);
+
+/*!
+ * The bare minimum size for a custom secret.
+ *
+ * @see
+ *  XXH3_64bits_withSecret(), XXH3_64bits_reset_withSecret(),
+ *  XXH3_128bits_withSecret(), XXH3_128bits_reset_withSecret().
+ */
+#define XXH3_SECRET_SIZE_MIN 136
+
+/*
+ * XXH3_64bits_withSecret():
+ * It's possible to provide any blob of bytes as a "secret" to generate the hash.
+ * This makes it more difficult for an external actor to prepare an intentional collision.
+ * The main condition is that secretSize *must* be large enough (>= XXH3_SECRET_SIZE_MIN).
+ * However, the quality of produced hash values depends on secret's entropy.
+ * Technically, the secret must look like a bunch of random bytes.
+ * Avoid "trivial" or structured data such as repeated sequences or a text document.
+ * Whenever unsure about the "randomness" of the blob of bytes,
+ * consider relabelling it as a "custom seed" instead,
+ * and employ "XXH3_generateSecret()" (see below)
+ * to generate a high entropy secret derived from the custom seed.
+ */
+XXH_PUBLIC_API XXH64_hash_t XXH3_64bits_withSecret(const void* data, size_t len, const void* secret, size_t secretSize);
+
+
+/*******   Streaming   *******/
+/*
+ * Streaming requires state maintenance.
+ * This operation costs memory and CPU.
+ * As a consequence, streaming is slower than one-shot hashing.
+ * For better performance, prefer one-shot functions whenever applicable.
+ */
+
+/*!
+ * @brief The state struct for the XXH3 streaming API.
+ *
+ * @see XXH3_state_s for details.
+ */
+typedef struct XXH3_state_s XXH3_state_t;
+XXH_PUBLIC_API XXH3_state_t* XXH3_createState(void);
+XXH_PUBLIC_API XXH_errorcode XXH3_freeState(XXH3_state_t* statePtr);
+XXH_PUBLIC_API void XXH3_copyState(XXH3_state_t* dst_state, const XXH3_state_t* src_state);
+
+/*
+ * XXH3_64bits_reset():
+ * Initialize with default parameters.
+ * digest will be equivalent to `XXH3_64bits()`.
+ */
+XXH_PUBLIC_API XXH_errorcode XXH3_64bits_reset(XXH3_state_t* statePtr);
+/*
+ * XXH3_64bits_reset_withSeed():
+ * Generate a custom secret from `seed`, and store it into `statePtr`.
+ * digest will be equivalent to `XXH3_64bits_withSeed()`.
+ */
+XXH_PUBLIC_API XXH_errorcode XXH3_64bits_reset_withSeed(XXH3_state_t* statePtr, XXH64_hash_t seed);
+/*
+ * XXH3_64bits_reset_withSecret():
+ * `secret` is referenced, it _must outlive_ the hash streaming session.
+ * Similar to one-shot API, `secretSize` must be >= `XXH3_SECRET_SIZE_MIN`,
+ * and the quality of produced hash values depends on secret's entropy
+ * (secret's content should look like a bunch of random bytes).
+ * When in doubt about the randomness of a candidate `secret`,
+ * consider employing `XXH3_generateSecret()` instead (see below).
+ */
+XXH_PUBLIC_API XXH_errorcode XXH3_64bits_reset_withSecret(XXH3_state_t* statePtr, const void* secret, size_t secretSize);
+
+XXH_PUBLIC_API XXH_errorcode XXH3_64bits_update (XXH3_state_t* statePtr, const void* input, size_t length);
+XXH_PUBLIC_API XXH64_hash_t  XXH3_64bits_digest (const XXH3_state_t* statePtr);
+
+/* note : canonical representation of XXH3 is the same as XXH64
+ * since they both produce XXH64_hash_t values */
+
+
+/*-**********************************************************************
+*  XXH3 128-bit variant
+************************************************************************/
+
+/*!
+ * @brief The return value from 128-bit hashes.
+ *
+ * Stored in little endian order, although the fields themselves are in native
+ * endianness.
+ */
+typedef struct {
+    XXH64_hash_t low64;   /*!< `value & 0xFFFFFFFFFFFFFFFF` */
+    XXH64_hash_t high64;  /*!< `value >> 64` */
+} XXH128_hash_t;
+
+XXH_PUBLIC_API XXH128_hash_t XXH3_128bits(const void* data, size_t len);
+XXH_PUBLIC_API XXH128_hash_t XXH3_128bits_withSeed(const void* data, size_t len, XXH64_hash_t seed);
+XXH_PUBLIC_API XXH128_hash_t XXH3_128bits_withSecret(const void* data, size_t len, const void* secret, size_t secretSize);
+
+/*******   Streaming   *******/
+/*
+ * Streaming requires state maintenance.
+ * This operation costs memory and CPU.
+ * As a consequence, streaming is slower than one-shot hashing.
+ * For better performance, prefer one-shot functions whenever applicable.
+ *
+ * XXH3_128bits uses the same XXH3_state_t as XXH3_64bits().
+ * Use already declared XXH3_createState() and XXH3_freeState().
+ *
+ * All reset and streaming functions have same meaning as their 64-bit counterpart.
+ */
+
+XXH_PUBLIC_API XXH_errorcode XXH3_128bits_reset(XXH3_state_t* statePtr);
+XXH_PUBLIC_API XXH_errorcode XXH3_128bits_reset_withSeed(XXH3_state_t* statePtr, XXH64_hash_t seed);
+XXH_PUBLIC_API XXH_errorcode XXH3_128bits_reset_withSecret(XXH3_state_t* statePtr, const void* secret, size_t secretSize);
+
+XXH_PUBLIC_API XXH_errorcode XXH3_128bits_update (XXH3_state_t* statePtr, const void* input, size_t length);
+XXH_PUBLIC_API XXH128_hash_t XXH3_128bits_digest (const XXH3_state_t* statePtr);
+
+/* Following helper functions make it possible to compare XXH128_hast_t values.
+ * Since XXH128_hash_t is a structure, this capability is not offered by the language.
+ * Note: For better performance, these functions can be inlined using XXH_INLINE_ALL */
+
+/*!
+ * XXH128_isEqual():
+ * Return: 1 if `h1` and `h2` are equal, 0 if they are not.
+ */
+XXH_PUBLIC_API int XXH128_isEqual(XXH128_hash_t h1, XXH128_hash_t h2);
+
+/*!
+ * XXH128_cmp():
+ *
+ * This comparator is compatible with stdlib's `qsort()`/`bsearch()`.
+ *
+ * return: >0 if *h128_1  > *h128_2
+ *         =0 if *h128_1 == *h128_2
+ *         <0 if *h128_1  < *h128_2
+ */
+XXH_PUBLIC_API int XXH128_cmp(const void* h128_1, const void* h128_2);
+
+
+/*******   Canonical representation   *******/
+typedef struct { unsigned char digest[sizeof(XXH128_hash_t)]; } XXH128_canonical_t;
+XXH_PUBLIC_API void XXH128_canonicalFromHash(XXH128_canonical_t* dst, XXH128_hash_t hash);
+XXH_PUBLIC_API XXH128_hash_t XXH128_hashFromCanonical(const XXH128_canonical_t* src);
+
+
+#endif  /* XXH_NO_LONG_LONG */
+
+/*!
+ * @}
+ */
+#endif /* XXHASH_H_5627135585666179 */
+
+
+
+#if defined(XXH_STATIC_LINKING_ONLY) && !defined(XXHASH_H_STATIC_13879238742)
+#define XXHASH_H_STATIC_13879238742
+/* ****************************************************************************
+ * This section contains declarations which are not guaranteed to remain stable.
+ * They may change in future versions, becoming incompatible with a different
+ * version of the library.
+ * These declarations should only be used with static linking.
+ * Never use them in association with dynamic linking!
+ ***************************************************************************** */
+
+/*
+ * These definitions are only present to allow static allocation
+ * of XXH states, on stack or in a struct, for example.
+ * Never **ever** access their members directly.
+ */
+
+/*!
+ * @internal
+ * @brief Structure for XXH32 streaming API.
+ *
+ * @note This is only defined when @ref XXH_STATIC_LINKING_ONLY,
+ * @ref XXH_INLINE_ALL, or @ref XXH_IMPLEMENTATION is defined. Otherwise it is
+ * an opaque type. This allows fields to safely be changed.
+ *
+ * Typedef'd to @ref XXH32_state_t.
+ * Do not access the members of this struct directly.
+ * @see XXH64_state_s, XXH3_state_s
+ */
+struct XXH32_state_s {
+   XXH32_hash_t total_len_32; /*!< Total length hashed, modulo 2^32 */
+   XXH32_hash_t large_len;    /*!< Whether the hash is >= 16 (handles @ref total_len_32 overflow) */
+   XXH32_hash_t v1;           /*!< First accumulator lane */
+   XXH32_hash_t v2;           /*!< Second accumulator lane */
+   XXH32_hash_t v3;           /*!< Third accumulator lane */
+   XXH32_hash_t v4;           /*!< Fourth accumulator lane */
+   XXH32_hash_t mem32[4];     /*!< Internal buffer for partial reads. Treated as unsigned char[16]. */
+   XXH32_hash_t memsize;      /*!< Amount of data in @ref mem32 */
+   XXH32_hash_t reserved;     /*!< Reserved field. Do not read or write to it, it may be removed. */
+};   /* typedef'd to XXH32_state_t */
+
+
+#ifndef XXH_NO_LONG_LONG  /* defined when there is no 64-bit support */
+
+/*!
+ * @internal
+ * @brief Structure for XXH64 streaming API.
+ *
+ * @note This is only defined when @ref XXH_STATIC_LINKING_ONLY,
+ * @ref XXH_INLINE_ALL, or @ref XXH_IMPLEMENTATION is defined. Otherwise it is
+ * an opaque type. This allows fields to safely be changed.
+ *
+ * Typedef'd to @ref XXH64_state_t.
+ * Do not access the members of this struct directly.
+ * @see XXH32_state_s, XXH3_state_s
+ */
+struct XXH64_state_s {
+   XXH64_hash_t total_len;    /*!< Total length hashed. This is always 64-bit. */
+   XXH64_hash_t v1;           /*!< First accumulator lane */
+   XXH64_hash_t v2;           /*!< Second accumulator lane */
+   XXH64_hash_t v3;           /*!< Third accumulator lane */
+   XXH64_hash_t v4;           /*!< Fourth accumulator lane */
+   XXH64_hash_t mem64[4];     /*!< Internal buffer for partial reads. Treated as unsigned char[32]. */
+   XXH32_hash_t memsize;      /*!< Amount of data in @ref mem64 */
+   XXH32_hash_t reserved32;   /*!< Reserved field, needed for padding anyways*/
+   XXH64_hash_t reserved64;   /*!< Reserved field. Do not read or write to it, it may be removed. */
+};   /* typedef'd to XXH64_state_t */
+
+#if defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 201112L)   /* C11+ */
+#  include <stdalign.h>
+#  define XXH_ALIGN(n)      alignas(n)
+#elif defined(__GNUC__)
+#  define XXH_ALIGN(n)      __attribute__ ((aligned(n)))
+#elif defined(_MSC_VER)
+#  define XXH_ALIGN(n)      __declspec(align(n))
+#else
+#  define XXH_ALIGN(n)   /* disabled */
+#endif
+
+/* Old GCC versions only accept the attribute after the type in structures. */
+#if !(defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201112L))   /* C11+ */ \
+    && defined(__GNUC__)
+#   define XXH_ALIGN_MEMBER(align, type) type XXH_ALIGN(align)
+#else
+#   define XXH_ALIGN_MEMBER(align, type) XXH_ALIGN(align) type
+#endif
+
+/*!
+ * @brief The size of the internal XXH3 buffer.
+ *
+ * This is the optimal update size for incremental hashing.
+ *
+ * @see XXH3_64b_update(), XXH3_128b_update().
+ */
+#define XXH3_INTERNALBUFFER_SIZE 256
+
+/*!
+ * @brief Default size of the secret buffer (and @ref XXH3_kSecret).
+ *
+ * This is the size used in @ref XXH3_kSecret and the seeded functions.
+ *
+ * Not to be confused with @ref XXH3_SECRET_SIZE_MIN.
+ */
+#define XXH3_SECRET_DEFAULT_SIZE 192
+
+/*!
+ * @internal
+ * @brief Structure for XXH3 streaming API.
+ *
+ * @note This is only defined when @ref XXH_STATIC_LINKING_ONLY,
+ * @ref XXH_INLINE_ALL, or @ref XXH_IMPLEMENTATION is defined. Otherwise it is
+ * an opaque type. This allows fields to safely be changed.
+ *
+ * @note **This structure has a strict alignment requirement of 64 bytes.** Do
+ * not allocate this with `malloc()` or `new`, it will not be sufficiently
+ * aligned. Use @ref XXH3_createState() and @ref XXH3_freeState(), or stack
+ * allocation.
+ *
+ * Typedef'd to @ref XXH3_state_t.
+ * Do not access the members of this struct directly.
+ *
+ * @see XXH3_INITSTATE() for stack initialization.
+ * @see XXH3_createState(), XXH3_freeState().
+ * @see XXH32_state_s, XXH64_state_s
+ */
+struct XXH3_state_s {
+   XXH_ALIGN_MEMBER(64, XXH64_hash_t acc[8]);
+       /*!< The 8 accumulators. Similar to `vN` in @ref XXH32_state_s::v1 and @ref XXH64_state_s */
+   XXH_ALIGN_MEMBER(64, unsigned char customSecret[XXH3_SECRET_DEFAULT_SIZE]);
+       /*!< Used to store a custom secret generated from a seed. */
+   XXH_ALIGN_MEMBER(64, unsigned char buffer[XXH3_INTERNALBUFFER_SIZE]);
+       /*!< The internal buffer. @see XXH32_state_s::mem32 */
+   XXH32_hash_t bufferedSize;
+       /*!< The amount of memory in @ref buffer, @see XXH32_state_s::memsize */
+   XXH32_hash_t reserved32;
+       /*!< Reserved field. Needed for padding on 64-bit. */
+   size_t nbStripesSoFar;
+       /*!< Number or stripes processed. */
+   XXH64_hash_t totalLen;
+       /*!< Total length hashed. 64-bit even on 32-bit targets. */
+   size_t nbStripesPerBlock;
+       /*!< Number of stripes per block. */
+   size_t secretLimit;
+       /*!< Size of @ref customSecret or @ref extSecret */
+   XXH64_hash_t seed;
+       /*!< Seed for _withSeed variants. Must be zero otherwise, @see XXH3_INITSTATE() */
+   XXH64_hash_t reserved64;
+       /*!< Reserved field. */
+   const unsigned char* extSecret;
+       /*!< Reference to an external secret for the _withSecret variants, NULL
+        *   for other variants. */
+   /* note: there may be some padding at the end due to alignment on 64 bytes */
+}; /* typedef'd to XXH3_state_t */
+
+#undef XXH_ALIGN_MEMBER
+
+/*!
+ * @brief Initializes a stack-allocated `XXH3_state_s`.
+ *
+ * When the @ref XXH3_state_t structure is merely emplaced on stack,
+ * it should be initialized with XXH3_INITSTATE() or a memset()
+ * in case its first reset uses XXH3_NNbits_reset_withSeed().
+ * This init can be omitted if the first reset uses default or _withSecret mode.
+ * This operation isn't necessary when the state is created with XXH3_createState().
+ * Note that this doesn't prepare the state for a streaming operation,
+ * it's still necessary to use XXH3_NNbits_reset*() afterwards.
+ */
+#define XXH3_INITSTATE(XXH3_state_ptr)   { (XXH3_state_ptr)->seed = 0; }
+
+
+/* ===   Experimental API   === */
+/* Symbols defined below must be considered tied to a specific library version. */
+
+/*
+ * XXH3_generateSecret():
+ *
+ * Derive a high-entropy secret from any user-defined content, named customSeed.
+ * The generated secret can be used in combination with `*_withSecret()` functions.
+ * The `_withSecret()` variants are useful to provide a higher level of protection than 64-bit seed,
+ * as it becomes much more difficult for an external actor to guess how to impact the calculation logic.
+ *
+ * The function accepts as input a custom seed of any length and any content,
+ * and derives from it a high-entropy secret of length XXH3_SECRET_DEFAULT_SIZE
+ * into an already allocated buffer secretBuffer.
+ * The generated secret is _always_ XXH_SECRET_DEFAULT_SIZE bytes long.
+ *
+ * The generated secret can then be used with any `*_withSecret()` variant.
+ * Functions `XXH3_128bits_withSecret()`, `XXH3_64bits_withSecret()`,
+ * `XXH3_128bits_reset_withSecret()` and `XXH3_64bits_reset_withSecret()`
+ * are part of this list. They all accept a `secret` parameter
+ * which must be very long for implementation reasons (>= XXH3_SECRET_SIZE_MIN)
+ * _and_ feature very high entropy (consist of random-looking bytes).
+ * These conditions can be a high bar to meet, so
+ * this function can be used to generate a secret of proper quality.
+ *
+ * customSeed can be anything. It can have any size, even small ones,
+ * and its content can be anything, even stupidly "low entropy" source such as a bunch of zeroes.
+ * The resulting `secret` will nonetheless provide all expected qualities.
+ *
+ * Supplying NULL as the customSeed copies the default secret into `secretBuffer`.
+ * When customSeedSize > 0, supplying NULL as customSeed is undefined behavior.
+ */
+XXH_PUBLIC_API void XXH3_generateSecret(void* secretBuffer, const void* customSeed, size_t customSeedSize);
+
+
+/* simple short-cut to pre-selected XXH3_128bits variant */
+XXH_PUBLIC_API XXH128_hash_t XXH128(const void* data, size_t len, XXH64_hash_t seed);
+
+
+#endif  /* XXH_NO_LONG_LONG */
+#if defined(XXH_INLINE_ALL) || defined(XXH_PRIVATE_API)
+#  define XXH_IMPLEMENTATION
+#endif
+
+#endif  /* defined(XXH_STATIC_LINKING_ONLY) && !defined(XXHASH_H_STATIC_13879238742) */
+
+
+/* ======================================================================== */
+/* ======================================================================== */
+/* ======================================================================== */
+
+
+/*-**********************************************************************
+ * xxHash implementation
+ *-**********************************************************************
+ * xxHash's implementation used to be hosted inside xxhash.c.
+ *
+ * However, inlining requires implementation to be visible to the compiler,
+ * hence be included alongside the header.
+ * Previously, implementation was hosted inside xxhash.c,
+ * which was then #included when inlining was activated.
+ * This construction created issues with a few build and install systems,
+ * as it required xxhash.c to be stored in /include directory.
+ *
+ * xxHash implementation is now directly integrated within xxhash.h.
+ * As a consequence, xxhash.c is no longer needed in /include.
+ *
+ * xxhash.c is still available and is still useful.
+ * In a "normal" setup, when xxhash is not inlined,
+ * xxhash.h only exposes the prototypes and public symbols,
+ * while xxhash.c can be built into an object file xxhash.o
+ * which can then be linked into the final binary.
+ ************************************************************************/
+
+#if ( defined(XXH_INLINE_ALL) || defined(XXH_PRIVATE_API) \
+   || defined(XXH_IMPLEMENTATION) ) && !defined(XXH_IMPLEM_13a8737387)
+#  define XXH_IMPLEM_13a8737387
+
+/* *************************************
+*  Tuning parameters
+***************************************/
+
+/*!
+ * @defgroup tuning Tuning parameters
+ * @{
+ *
+ * Various macros to control xxHash's behavior.
+ */
+#ifdef XXH_DOXYGEN
+/*!
+ * @brief Define this to disable 64-bit code.
+ *
+ * Useful if only using the @ref xxh32_family and you have a strict C90 compiler.
+ */
+#  define XXH_NO_LONG_LONG
+#  undef XXH_NO_LONG_LONG /* don't actually */
+/*!
+ * @brief Controls how unaligned memory is accessed.
+ *
+ * By default, access to unaligned memory is controlled by `memcpy()`, which is
+ * safe and portable.
+ *
+ * Unfortunately, on some target/compiler combinations, the generated assembly
+ * is sub-optimal.
+ *
+ * The below switch allow selection of a different access method
+ * in the search for improved performance.
+ *
+ * @par Possible options:
+ *
+ *  - `XXH_FORCE_MEMORY_ACCESS=0` (default): `memcpy`
+ *   @par
+ *     Use `memcpy()`. Safe and portable. Note that most modern compilers will
+ *     eliminate the function call and treat it as an unaligned access.
+ *
+ *  - `XXH_FORCE_MEMORY_ACCESS=1`: `__attribute__((packed))`
+ *   @par
+ *     Depends on compiler extensions and is therefore not portable.
+ *     This method is safe _if_ your compiler supports it,
+ *     and *generally* as fast or faster than `memcpy`.
+ *
+ *  - `XXH_FORCE_MEMORY_ACCESS=2`: Direct cast
+ *  @par
+ *     Casts directly and dereferences. This method doesn't depend on the
+ *     compiler, but it violates the C standard as it directly dereferences an
+ *     unaligned pointer. It can generate buggy code on targets which do not
+ *     support unaligned memory accesses, but in some circumstances, it's the
+ *     only known way to get the most performance.
+ *
+ *  - `XXH_FORCE_MEMORY_ACCESS=3`: Byteshift
+ *  @par
+ *     Also portable. This can generate the best code on old compilers which don't
+ *     inline small `memcpy()` calls, and it might also be faster on big-endian
+ *     systems which lack a native byteswap instruction. However, some compilers
+ *     will emit literal byteshifts even if the target supports unaligned access.
+ *  .
+ *
+ * @warning
+ *   Methods 1 and 2 rely on implementation-defined behavior. Use these with
+ *   care, as what works on one compiler/platform/optimization level may cause
+ *   another to read garbage data or even crash.
+ *
+ * See https://stackoverflow.com/a/32095106/646947 for details.
+ *
+ * Prefer these methods in priority order (0 > 3 > 1 > 2)
+ */
+#  define XXH_FORCE_MEMORY_ACCESS 0
+/*!
+ * @def XXH_ACCEPT_NULL_INPUT_POINTER
+ * @brief Whether to add explicit `NULL` checks.
+ *
+ * If the input pointer is `NULL` and the length is non-zero, xxHash's default
+ * behavior is to dereference it, triggering a segfault.
+ *
+ * When this macro is enabled, xxHash actively checks the input for a null pointer.
+ * If it is, the result for null input pointers is the same as a zero-length input.
+ */
+#  define XXH_ACCEPT_NULL_INPUT_POINTER 0
+/*!
+ * @def XXH_FORCE_ALIGN_CHECK
+ * @brief If defined to non-zero, adds a special path for aligned inputs (XXH32()
+ * and XXH64() only).
+ *
+ * This is an important performance trick for architectures without decent
+ * unaligned memory access performance.
+ *
+ * It checks for input alignment, and when conditions are met, uses a "fast
+ * path" employing direct 32-bit/64-bit reads, resulting in _dramatically
+ * faster_ read speed.
+ *
+ * The check costs one initial branch per hash, which is generally negligible,
+ * but not zero.
+ *
+ * Moreover, it's not useful to generate an additional code path if memory
+ * access uses the same instruction for both aligned and unaligned
+ * addresses (e.g. x86 and aarch64).
+ *
+ * In these cases, the alignment check can be removed by setting this macro to 0.
+ * Then the code will always use unaligned memory access.
+ * Align check is automatically disabled on x86, x64 & arm64,
+ * which are platforms known to offer good unaligned memory accesses performance.
+ *
+ * This option does not affect XXH3 (only XXH32 and XXH64).
+ */
+#  define XXH_FORCE_ALIGN_CHECK 0
+
+/*!
+ * @def XXH_NO_INLINE_HINTS
+ * @brief When non-zero, sets all functions to `static`.
+ *
+ * By default, xxHash tries to force the compiler to inline almost all internal
+ * functions.
+ *
+ * This can usually improve performance due to reduced jumping and improved
+ * constant folding, but significantly increases the size of the binary which
+ * might not be favorable.
+ *
+ * Additionally, sometimes the forced inlining can be detrimental to performance,
+ * depending on the architecture.
+ *
+ * XXH_NO_INLINE_HINTS marks all internal functions as static, giving the
+ * compiler full control on whether to inline or not.
+ *
+ * When not optimizing (-O0), optimizing for size (-Os, -Oz), or using
+ * -fno-inline with GCC or Clang, this will automatically be defined.
+ */
+#  define XXH_NO_INLINE_HINTS 0
+
+/*!
+ * @def XXH_REROLL
+ * @brief Whether to reroll `XXH32_finalize` and `XXH64_finalize`.
+ *
+ * For performance, `XXH32_finalize` and `XXH64_finalize` use an unrolled loop
+ * in the form of a switch statement.
+ *
+ * This is not always desirable, as it generates larger code, and depending on
+ * the architecture, may even be slower
+ *
+ * This is automatically defined with `-Os`/`-Oz` on GCC and Clang.
+ */
+#  define XXH_REROLL 0
+
+/*!
+ * @internal
+ * @brief Redefines old internal names.
+ *
+ * For compatibility with code that uses xxHash's internals before the names
+ * were changed to improve namespacing. There is no other reason to use this.
+ */
+#  define XXH_OLD_NAMES
+#  undef XXH_OLD_NAMES /* don't actually use, it is ugly. */
+#endif /* XXH_DOXYGEN */
+/*!
+ * @}
+ */
+
+#ifndef XXH_FORCE_MEMORY_ACCESS   /* can be defined externally, on command line for example */
+   /* prefer __packed__ structures (method 1) for gcc on armv7 and armv8 */
+#  if !defined(__clang__) && ( \
+    (defined(__INTEL_COMPILER) && !defined(_WIN32)) || \
+    (defined(__GNUC__) && (defined(__ARM_ARCH) && __ARM_ARCH >= 7)) )
+#    define XXH_FORCE_MEMORY_ACCESS 1
+#  endif
+#endif
+
+#ifndef XXH_ACCEPT_NULL_INPUT_POINTER   /* can be defined externally */
+#  define XXH_ACCEPT_NULL_INPUT_POINTER 0
+#endif
+
+#ifndef XXH_FORCE_ALIGN_CHECK  /* can be defined externally */
+#  if defined(__i386)  || defined(__x86_64__) || defined(__aarch64__) \
+   || defined(_M_IX86) || defined(_M_X64)     || defined(_M_ARM64) /* visual */
+#    define XXH_FORCE_ALIGN_CHECK 0
+#  else
+#    define XXH_FORCE_ALIGN_CHECK 1
+#  endif
+#endif
+
+#ifndef XXH_NO_INLINE_HINTS
+#  if defined(__OPTIMIZE_SIZE__) /* -Os, -Oz */ \
+   || defined(__NO_INLINE__)     /* -O0, -fno-inline */
+#    define XXH_NO_INLINE_HINTS 1
+#  else
+#    define XXH_NO_INLINE_HINTS 0
+#  endif
+#endif
+
+#ifndef XXH_REROLL
+#  if defined(__OPTIMIZE_SIZE__)
+#    define XXH_REROLL 1
+#  else
+#    define XXH_REROLL 0
+#  endif
+#endif
+
+/*!
+ * @defgroup impl Implementation
+ * @{
+ */
+
+
+/* *************************************
+*  Includes & Memory related functions
+***************************************/
+/*
+ * Modify the local functions below should you wish to use
+ * different memory routines for malloc() and free()
+ */
+#include <stdlib.h>
+
+/*!
+ * @internal
+ * @brief Modify this function to use a different routine than malloc().
+ */
+static void* XXH_malloc(size_t s) { return malloc(s); }
+
+/*!
+ * @internal
+ * @brief Modify this function to use a different routine than free().
+ */
+static void XXH_free(void* p) { free(p); }
+
+#include <string.h>
+
+/*!
+ * @internal
+ * @brief Modify this function to use a different routine than memcpy().
+ */
+static void* XXH_memcpy(void* dest, const void* src, size_t size)
+{
+    return memcpy(dest,src,size);
+}
+
+#include <limits.h>   /* ULLONG_MAX */
+
+
+/* *************************************
+*  Compiler Specific Options
+***************************************/
+#ifdef _MSC_VER /* Visual Studio warning fix */
+#  pragma warning(disable : 4127) /* disable: C4127: conditional expression is constant */
+#endif
+
+#if XXH_NO_INLINE_HINTS  /* disable inlining hints */
+#  if defined(__GNUC__)
+#    define XXH_FORCE_INLINE static __attribute__((unused))
+#  else
+#    define XXH_FORCE_INLINE static
+#  endif
+#  define XXH_NO_INLINE static
+/* enable inlining hints */
+#elif defined(_MSC_VER)  /* Visual Studio */
+#  define XXH_FORCE_INLINE static __forceinline
+#  define XXH_NO_INLINE static __declspec(noinline)
+#elif defined(__GNUC__)
+#  define XXH_FORCE_INLINE static __inline__ __attribute__((always_inline, unused))
+#  define XXH_NO_INLINE static __attribute__((noinline))
+#elif defined (__cplusplus) \
+  || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L))   /* C99 */
+#  define XXH_FORCE_INLINE static inline
+#  define XXH_NO_INLINE static
+#else
+#  define XXH_FORCE_INLINE static
+#  define XXH_NO_INLINE static
+#endif
+
+
+
+/* *************************************
+*  Debug
+***************************************/
+/*!
+ * @ingroup tuning
+ * @def XXH_DEBUGLEVEL
+ * @brief Sets the debugging level.
+ *
+ * XXH_DEBUGLEVEL is expected to be defined externally, typically via the
+ * compiler's command line options. The value must be a number.
+ */
+#ifndef XXH_DEBUGLEVEL
+#  ifdef DEBUGLEVEL /* backwards compat */
+#    define XXH_DEBUGLEVEL DEBUGLEVEL
+#  else
+#    define XXH_DEBUGLEVEL 0
+#  endif
+#endif
+
+#if (XXH_DEBUGLEVEL>=1)
+#  include <assert.h>   /* note: can still be disabled with NDEBUG */
+#  define XXH_ASSERT(c)   assert(c)
+#else
+#  define XXH_ASSERT(c)   ((void)0)
+#endif
+
+/* note: use after variable declarations */
+#define XXH_STATIC_ASSERT(c)  do { enum { XXH_sa = 1/(int)(!!(c)) }; } while (0)
+
+/*!
+ * @internal
+ * @def XXH_COMPILER_GUARD(var)
+ * @brief Used to prevent unwanted optimizations for @p var.
+ *
+ * It uses an empty GCC inline assembly statement with a register constraint
+ * which forces @p var into a general purpose register (eg eax, ebx, ecx
+ * on x86) and marks it as modified.
+ *
+ * This is used in a few places to avoid unwanted autovectorization (e.g.
+ * XXH32_round()). All vectorization we want is explicit via intrinsics,
+ * and _usually_ isn't wanted elsewhere.
+ *
+ * We also use it to prevent unwanted constant folding for AArch64 in
+ * XXH3_initCustomSecret_scalar().
+ */
+#ifdef __GNUC__
+#  define XXH_COMPILER_GUARD(var) __asm__ __volatile__("" : "+r" (var))
+#else
+#  define XXH_COMPILER_GUARD(var) ((void)0)
+#endif
+
+/* *************************************
+*  Basic Types
+***************************************/
+#if !defined (__VMS) \
+ && (defined (__cplusplus) \
+ || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */) )
+# include <stdint.h>
+  typedef uint8_t xxh_u8;
+#else
+  typedef unsigned char xxh_u8;
+#endif
+typedef XXH32_hash_t xxh_u32;
+
+#ifdef XXH_OLD_NAMES
+#  define BYTE xxh_u8
+#  define U8   xxh_u8
+#  define U32  xxh_u32
+#endif
+
+/* ***   Memory access   *** */
+
+/*!
+ * @internal
+ * @fn xxh_u32 XXH_read32(const void* ptr)
+ * @brief Reads an unaligned 32-bit integer from @p ptr in native endianness.
+ *
+ * Affected by @ref XXH_FORCE_MEMORY_ACCESS.
+ *
+ * @param ptr The pointer to read from.
+ * @return The 32-bit native endian integer from the bytes at @p ptr.
+ */
+
+/*!
+ * @internal
+ * @fn xxh_u32 XXH_readLE32(const void* ptr)
+ * @brief Reads an unaligned 32-bit little endian integer from @p ptr.
+ *
+ * Affected by @ref XXH_FORCE_MEMORY_ACCESS.
+ *
+ * @param ptr The pointer to read from.
+ * @return The 32-bit little endian integer from the bytes at @p ptr.
+ */
+
+/*!
+ * @internal
+ * @fn xxh_u32 XXH_readBE32(const void* ptr)
+ * @brief Reads an unaligned 32-bit big endian integer from @p ptr.
+ *
+ * Affected by @ref XXH_FORCE_MEMORY_ACCESS.
+ *
+ * @param ptr The pointer to read from.
+ * @return The 32-bit big endian integer from the bytes at @p ptr.
+ */
+
+/*!
+ * @internal
+ * @fn xxh_u32 XXH_readLE32_align(const void* ptr, XXH_alignment align)
+ * @brief Like @ref XXH_readLE32(), but has an option for aligned reads.
+ *
+ * Affected by @ref XXH_FORCE_MEMORY_ACCESS.
+ * Note that when @ref XXH_FORCE_ALIGN_CHECK == 0, the @p align parameter is
+ * always @ref XXH_alignment::XXH_unaligned.
+ *
+ * @param ptr The pointer to read from.
+ * @param align Whether @p ptr is aligned.
+ * @pre
+ *   If @p align == @ref XXH_alignment::XXH_aligned, @p ptr must be 4 byte
+ *   aligned.
+ * @return The 32-bit little endian integer from the bytes at @p ptr.
+ */
+
+#if (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==3))
+/*
+ * Manual byteshift. Best for old compilers which don't inline memcpy.
+ * We actually directly use XXH_readLE32 and XXH_readBE32.
+ */
+#elif (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==2))
+
+/*
+ * Force direct memory access. Only works on CPU which support unaligned memory
+ * access in hardware.
+ */
+static xxh_u32 XXH_read32(const void* memPtr) { return *(const xxh_u32*) memPtr; }
+
+#elif (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==1))
+
+/*
+ * __pack instructions are safer but compiler specific, hence potentially
+ * problematic for some compilers.
+ *
+ * Currently only defined for GCC and ICC.
+ */
+#ifdef XXH_OLD_NAMES
+typedef union { xxh_u32 u32; } __attribute__((packed)) unalign;
+#endif
+static xxh_u32 XXH_read32(const void* ptr)
+{
+    typedef union { xxh_u32 u32; } __attribute__((packed)) xxh_unalign;
+    return ((const xxh_unalign*)ptr)->u32;
+}
+
+#else
+
+/*
+ * Portable and safe solution. Generally efficient.
+ * see: https://stackoverflow.com/a/32095106/646947
+ */
+static xxh_u32 XXH_read32(const void* memPtr)
+{
+    xxh_u32 val;
+    memcpy(&val, memPtr, sizeof(val));
+    return val;
+}
+
+#endif   /* XXH_FORCE_DIRECT_MEMORY_ACCESS */
+
+
+/* ***   Endianness   *** */
+/*!
+ * @ingroup tuning
+ * @def XXH_CPU_LITTLE_ENDIAN
+ * @brief Whether the target is little endian.
+ *
+ * Defined to 1 if the target is little endian, or 0 if it is big endian.
+ * It can be defined externally, for example on the compiler command line.
+ *
+ * If it is not defined, a runtime check (which is usually constant folded)
+ * is used instead.
+ *
+ * @note
+ *   This is not necessarily defined to an integer constant.
+ *
+ * @see XXH_isLittleEndian() for the runtime check.
+ */
+#ifndef XXH_CPU_LITTLE_ENDIAN
+/*
+ * Try to detect endianness automatically, to avoid the nonstandard behavior
+ * in `XXH_isLittleEndian()`
+ */
+#  if defined(_WIN32) /* Windows is always little endian */ \
+     || defined(__LITTLE_ENDIAN__) \
+     || (defined(__BYTE_ORDER__) && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__)
+#    define XXH_CPU_LITTLE_ENDIAN 1
+#  elif defined(__BIG_ENDIAN__) \
+     || (defined(__BYTE_ORDER__) && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__)
+#    define XXH_CPU_LITTLE_ENDIAN 0
+#  else
+/*!
+ * @internal
+ * @brief Runtime check for @ref XXH_CPU_LITTLE_ENDIAN.
+ *
+ * Most compilers will constant fold this.
+ */
+static int XXH_isLittleEndian(void)
+{
+    /*
+     * Portable and well-defined behavior.
+     * Don't use static: it is detrimental to performance.
+     */
+    const union { xxh_u32 u; xxh_u8 c[4]; } one = { 1 };
+    return one.c[0];
+}
+#   define XXH_CPU_LITTLE_ENDIAN   XXH_isLittleEndian()
+#  endif
+#endif
+
+
+
+
+/* ****************************************
+*  Compiler-specific Functions and Macros
+******************************************/
+#define XXH_GCC_VERSION (__GNUC__ * 100 + __GNUC_MINOR__)
+
+#ifdef __has_builtin
+#  define XXH_HAS_BUILTIN(x) __has_builtin(x)
+#else
+#  define XXH_HAS_BUILTIN(x) 0
+#endif
+
+/*!
+ * @internal
+ * @def XXH_rotl32(x,r)
+ * @brief 32-bit rotate left.
+ *
+ * @param x The 32-bit integer to be rotated.
+ * @param r The number of bits to rotate.
+ * @pre
+ *   @p r > 0 && @p r < 32
+ * @note
+ *   @p x and @p r may be evaluated multiple times.
+ * @return The rotated result.
+ */
+#if !defined(NO_CLANG_BUILTIN) && XXH_HAS_BUILTIN(__builtin_rotateleft32) \
+                               && XXH_HAS_BUILTIN(__builtin_rotateleft64)
+#  define XXH_rotl32 __builtin_rotateleft32
+#  define XXH_rotl64 __builtin_rotateleft64
+/* Note: although _rotl exists for minGW (GCC under windows), performance seems poor */
+#elif defined(_MSC_VER)
+#  define XXH_rotl32(x,r) _rotl(x,r)
+#  define XXH_rotl64(x,r) _rotl64(x,r)
+#else
+#  define XXH_rotl32(x,r) (((x) << (r)) | ((x) >> (32 - (r))))
+#  define XXH_rotl64(x,r) (((x) << (r)) | ((x) >> (64 - (r))))
+#endif
+
+/*!
+ * @internal
+ * @fn xxh_u32 XXH_swap32(xxh_u32 x)
+ * @brief A 32-bit byteswap.
+ *
+ * @param x The 32-bit integer to byteswap.
+ * @return @p x, byteswapped.
+ */
+#if defined(_MSC_VER)     /* Visual Studio */
+#  define XXH_swap32 _byteswap_ulong
+#elif XXH_GCC_VERSION >= 403
+#  define XXH_swap32 __builtin_bswap32
+#else
+static xxh_u32 XXH_swap32 (xxh_u32 x)
+{
+    return  ((x << 24) & 0xff000000 ) |
+            ((x <<  8) & 0x00ff0000 ) |
+            ((x >>  8) & 0x0000ff00 ) |
+            ((x >> 24) & 0x000000ff );
+}
+#endif
+
+
+/* ***************************
+*  Memory reads
+*****************************/
+
+/*!
+ * @internal
+ * @brief Enum to indicate whether a pointer is aligned.
+ */
+typedef enum {
+    XXH_aligned,  /*!< Aligned */
+    XXH_unaligned /*!< Possibly unaligned */
+} XXH_alignment;
+
+/*
+ * XXH_FORCE_MEMORY_ACCESS==3 is an endian-independent byteshift load.
+ *
+ * This is ideal for older compilers which don't inline memcpy.
+ */
+#if (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==3))
+
+XXH_FORCE_INLINE xxh_u32 XXH_readLE32(const void* memPtr)
+{
+    const xxh_u8* bytePtr = (const xxh_u8 *)memPtr;
+    return bytePtr[0]
+         | ((xxh_u32)bytePtr[1] << 8)
+         | ((xxh_u32)bytePtr[2] << 16)
+         | ((xxh_u32)bytePtr[3] << 24);
+}
+
+XXH_FORCE_INLINE xxh_u32 XXH_readBE32(const void* memPtr)
+{
+    const xxh_u8* bytePtr = (const xxh_u8 *)memPtr;
+    return bytePtr[3]
+         | ((xxh_u32)bytePtr[2] << 8)
+         | ((xxh_u32)bytePtr[1] << 16)
+         | ((xxh_u32)bytePtr[0] << 24);
+}
+
+#else
+XXH_FORCE_INLINE xxh_u32 XXH_readLE32(const void* ptr)
+{
+    return XXH_CPU_LITTLE_ENDIAN ? XXH_read32(ptr) : XXH_swap32(XXH_read32(ptr));
+}
+
+static xxh_u32 XXH_readBE32(const void* ptr)
+{
+    return XXH_CPU_LITTLE_ENDIAN ? XXH_swap32(XXH_read32(ptr)) : XXH_read32(ptr);
+}
+#endif
+
+XXH_FORCE_INLINE xxh_u32
+XXH_readLE32_align(const void* ptr, XXH_alignment align)
+{
+    if (align==XXH_unaligned) {
+        return XXH_readLE32(ptr);
+    } else {
+        return XXH_CPU_LITTLE_ENDIAN ? *(const xxh_u32*)ptr : XXH_swap32(*(const xxh_u32*)ptr);
+    }
+}
+
+
+/* *************************************
+*  Misc
+***************************************/
+/*! @ingroup public */
+XXH_PUBLIC_API unsigned XXH_versionNumber (void) { return XXH_VERSION_NUMBER; }
+
+
+/* *******************************************************************
+*  32-bit hash functions
+*********************************************************************/
+/*!
+ * @}
+ * @defgroup xxh32_impl XXH32 implementation
+ * @ingroup impl
+ * @{
+ */
+ /* #define instead of static const, to be used as initializers */
+#define XXH_PRIME32_1  0x9E3779B1U  /*!< 0b10011110001101110111100110110001 */
+#define XXH_PRIME32_2  0x85EBCA77U  /*!< 0b10000101111010111100101001110111 */
+#define XXH_PRIME32_3  0xC2B2AE3DU  /*!< 0b11000010101100101010111000111101 */
+#define XXH_PRIME32_4  0x27D4EB2FU  /*!< 0b00100111110101001110101100101111 */
+#define XXH_PRIME32_5  0x165667B1U  /*!< 0b00010110010101100110011110110001 */
+
+#ifdef XXH_OLD_NAMES
+#  define PRIME32_1 XXH_PRIME32_1
+#  define PRIME32_2 XXH_PRIME32_2
+#  define PRIME32_3 XXH_PRIME32_3
+#  define PRIME32_4 XXH_PRIME32_4
+#  define PRIME32_5 XXH_PRIME32_5
+#endif
+
+/*!
+ * @internal
+ * @brief Normal stripe processing routine.
+ *
+ * This shuffles the bits so that any bit from @p input impacts several bits in
+ * @p acc.
+ *
+ * @param acc The accumulator lane.
+ * @param input The stripe of input to mix.
+ * @return The mixed accumulator lane.
+ */
+static xxh_u32 XXH32_round(xxh_u32 acc, xxh_u32 input)
+{
+    acc += input * XXH_PRIME32_2;
+    acc  = XXH_rotl32(acc, 13);
+    acc *= XXH_PRIME32_1;
+#if (defined(__SSE4_1__) || defined(__aarch64__)) && !defined(XXH_ENABLE_AUTOVECTORIZE)
+    /*
+     * UGLY HACK:
+     * A compiler fence is the only thing that prevents GCC and Clang from
+     * autovectorizing the XXH32 loop (pragmas and attributes don't work for some
+     * reason) without globally disabling SSE4.1.
+     *
+     * The reason we want to avoid vectorization is because despite working on
+     * 4 integers at a time, there are multiple factors slowing XXH32 down on
+     * SSE4:
+     * - There's a ridiculous amount of lag from pmulld (10 cycles of latency on
+     *   newer chips!) making it slightly slower to multiply four integers at
+     *   once compared to four integers independently. Even when pmulld was
+     *   fastest, Sandy/Ivy Bridge, it is still not worth it to go into SSE
+     *   just to multiply unless doing a long operation.
+     *
+     * - Four instructions are required to rotate,
+     *      movqda tmp,  v // not required with VEX encoding
+     *      pslld  tmp, 13 // tmp <<= 13
+     *      psrld  v,   19 // x >>= 19
+     *      por    v,  tmp // x |= tmp
+     *   compared to one for scalar:
+     *      roll   v, 13    // reliably fast across the board
+     *      shldl  v, v, 13 // Sandy Bridge and later prefer this for some reason
+     *
+     * - Instruction level parallelism is actually more beneficial here because
+     *   the SIMD actually serializes this operation: While v1 is rotating, v2
+     *   can load data, while v3 can multiply. SSE forces them to operate
+     *   together.
+     *
+     * This is also enabled on AArch64, as Clang autovectorizes it incorrectly
+     * and it is pointless writing a NEON implementation that is basically the
+     * same speed as scalar for XXH32.
+     */
+    XXH_COMPILER_GUARD(acc);
+#endif
+    return acc;
+}
+
+/*!
+ * @internal
+ * @brief Mixes all bits to finalize the hash.
+ *
+ * The final mix ensures that all input bits have a chance to impact any bit in
+ * the output digest, resulting in an unbiased distribution.
+ *
+ * @param h32 The hash to avalanche.
+ * @return The avalanched hash.
+ */
+static xxh_u32 XXH32_avalanche(xxh_u32 h32)
+{
+    h32 ^= h32 >> 15;
+    h32 *= XXH_PRIME32_2;
+    h32 ^= h32 >> 13;
+    h32 *= XXH_PRIME32_3;
+    h32 ^= h32 >> 16;
+    return(h32);
+}
+
+#define XXH_get32bits(p) XXH_readLE32_align(p, align)
+
+/*!
+ * @internal
+ * @brief Processes the last 0-15 bytes of @p ptr.
+ *
+ * There may be up to 15 bytes remaining to consume from the input.
+ * This final stage will digest them to ensure that all input bytes are present
+ * in the final mix.
+ *
+ * @param h32 The hash to finalize.
+ * @param ptr The pointer to the remaining input.
+ * @param len The remaining length, modulo 16.
+ * @param align Whether @p ptr is aligned.
+ * @return The finalized hash.
+ */
+static xxh_u32
+XXH32_finalize(xxh_u32 h32, const xxh_u8* ptr, size_t len, XXH_alignment align)
+{
+#define XXH_PROCESS1 do {                           \
+    h32 += (*ptr++) * XXH_PRIME32_5;                \
+    h32 = XXH_rotl32(h32, 11) * XXH_PRIME32_1;      \
+} while (0)
+
+#define XXH_PROCESS4 do {                           \
+    h32 += XXH_get32bits(ptr) * XXH_PRIME32_3;      \
+    ptr += 4;                                   \
+    h32  = XXH_rotl32(h32, 17) * XXH_PRIME32_4;     \
+} while (0)
+
+    /* Compact rerolled version */
+    if (XXH_REROLL) {
+        len &= 15;
+        while (len >= 4) {
+            XXH_PROCESS4;
+            len -= 4;
+        }
+        while (len > 0) {
+            XXH_PROCESS1;
+            --len;
+        }
+        return XXH32_avalanche(h32);
+    } else {
+         switch(len&15) /* or switch(bEnd - p) */ {
+           case 12:      XXH_PROCESS4;
+                         FALLTHROUGH_INTENDED;
+           case 8:       XXH_PROCESS4;
+                         FALLTHROUGH_INTENDED;
+           case 4:       XXH_PROCESS4;
+                         return XXH32_avalanche(h32);
+
+           case 13:      XXH_PROCESS4;
+                         FALLTHROUGH_INTENDED;
+           case 9:       XXH_PROCESS4;
+                         FALLTHROUGH_INTENDED;
+           case 5:       XXH_PROCESS4;
+                         XXH_PROCESS1;
+                         return XXH32_avalanche(h32);
+
+           case 14:      XXH_PROCESS4;
+                         FALLTHROUGH_INTENDED;
+           case 10:      XXH_PROCESS4;
+                         FALLTHROUGH_INTENDED;
+           case 6:       XXH_PROCESS4;
+                         XXH_PROCESS1;
+                         XXH_PROCESS1;
+                         return XXH32_avalanche(h32);
+
+           case 15:      XXH_PROCESS4;
+                         FALLTHROUGH_INTENDED;
+           case 11:      XXH_PROCESS4;
+                         FALLTHROUGH_INTENDED;
+           case 7:       XXH_PROCESS4;
+                         FALLTHROUGH_INTENDED;
+           case 3:       XXH_PROCESS1;
+                         FALLTHROUGH_INTENDED;
+           case 2:       XXH_PROCESS1;
+                         FALLTHROUGH_INTENDED;
+           case 1:       XXH_PROCESS1;
+                         FALLTHROUGH_INTENDED;
+           case 0:       return XXH32_avalanche(h32);
+        }
+        XXH_ASSERT(0);
+        return h32;   /* reaching this point is deemed impossible */
+    }
+}
+
+#ifdef XXH_OLD_NAMES
+#  define PROCESS1 XXH_PROCESS1
+#  define PROCESS4 XXH_PROCESS4
+#else
+#  undef XXH_PROCESS1
+#  undef XXH_PROCESS4
+#endif
+
+/*!
+ * @internal
+ * @brief The implementation for @ref XXH32().
+ *
+ * @param input, len, seed Directly passed from @ref XXH32().
+ * @param align Whether @p input is aligned.
+ * @return The calculated hash.
+ */
+XXH_FORCE_INLINE xxh_u32
+XXH32_endian_align(const xxh_u8* input, size_t len, xxh_u32 seed, XXH_alignment align)
+{
+    const xxh_u8* bEnd = input ? input + len : NULL;
+    xxh_u32 h32;
+
+#if defined(XXH_ACCEPT_NULL_INPUT_POINTER) && (XXH_ACCEPT_NULL_INPUT_POINTER>=1)
+    if (input==NULL) {
+        len=0;
+        bEnd=input=(const xxh_u8*)(size_t)16;
+    }
+#endif
+
+    if (len>=16) {
+        const xxh_u8* const limit = bEnd - 15;
+        xxh_u32 v1 = seed + XXH_PRIME32_1 + XXH_PRIME32_2;
+        xxh_u32 v2 = seed + XXH_PRIME32_2;
+        xxh_u32 v3 = seed + 0;
+        xxh_u32 v4 = seed - XXH_PRIME32_1;
+
+        do {
+            v1 = XXH32_round(v1, XXH_get32bits(input)); input += 4;
+            v2 = XXH32_round(v2, XXH_get32bits(input)); input += 4;
+            v3 = XXH32_round(v3, XXH_get32bits(input)); input += 4;
+            v4 = XXH32_round(v4, XXH_get32bits(input)); input += 4;
+        } while (input < limit);
+
+        h32 = XXH_rotl32(v1, 1)  + XXH_rotl32(v2, 7)
+            + XXH_rotl32(v3, 12) + XXH_rotl32(v4, 18);
+    } else {
+        h32  = seed + XXH_PRIME32_5;
+    }
+
+    h32 += (xxh_u32)len;
+
+    return XXH32_finalize(h32, input, len&15, align);
+}
+
+/*! @ingroup xxh32_family */
+XXH_PUBLIC_API XXH32_hash_t XXH32 (const void* input, size_t len, XXH32_hash_t seed)
+{
+#if 0
+    /* Simple version, good for code maintenance, but unfortunately slow for small inputs */
+    XXH32_state_t state;
+    XXH32_reset(&state, seed);
+    XXH32_update(&state, (const xxh_u8*)input, len);
+    return XXH32_digest(&state);
+#else
+    if (XXH_FORCE_ALIGN_CHECK) {
+        if ((((size_t)input) & 3) == 0) {   /* Input is 4-bytes aligned, leverage the speed benefit */
+            return XXH32_endian_align((const xxh_u8*)input, len, seed, XXH_aligned);
+    }   }
+
+    return XXH32_endian_align((const xxh_u8*)input, len, seed, XXH_unaligned);
+#endif
+}
+
+
+
+/*******   Hash streaming   *******/
+/*!
+ * @ingroup xxh32_family
+ */
+XXH_PUBLIC_API XXH32_state_t* XXH32_createState(void)
+{
+    return (XXH32_state_t*)XXH_malloc(sizeof(XXH32_state_t));
+}
+/*! @ingroup xxh32_family */
+XXH_PUBLIC_API XXH_errorcode XXH32_freeState(XXH32_state_t* statePtr)
+{
+    XXH_free(statePtr);
+    return XXH_OK;
+}
+
+/*! @ingroup xxh32_family */
+XXH_PUBLIC_API void XXH32_copyState(XXH32_state_t* dstState, const XXH32_state_t* srcState)
+{
+    memcpy(dstState, srcState, sizeof(*dstState));
+}
+
+/*! @ingroup xxh32_family */
+XXH_PUBLIC_API XXH_errorcode XXH32_reset(XXH32_state_t* statePtr, XXH32_hash_t seed)
+{
+    XXH32_state_t state;   /* using a local state to memcpy() in order to avoid strict-aliasing warnings */
+    memset(&state, 0, sizeof(state));
+    state.v1 = seed + XXH_PRIME32_1 + XXH_PRIME32_2;
+    state.v2 = seed + XXH_PRIME32_2;
+    state.v3 = seed + 0;
+    state.v4 = seed - XXH_PRIME32_1;
+    /* do not write into reserved, planned to be removed in a future version */
+    memcpy(statePtr, &state, sizeof(state) - sizeof(state.reserved));
+    return XXH_OK;
+}
+
+
+/*! @ingroup xxh32_family */
+XXH_PUBLIC_API XXH_errorcode
+XXH32_update(XXH32_state_t* state, const void* input, size_t len)
+{
+    if (input==NULL)
+#if defined(XXH_ACCEPT_NULL_INPUT_POINTER) && (XXH_ACCEPT_NULL_INPUT_POINTER>=1)
+        return XXH_OK;
+#else
+        return XXH_ERROR;
+#endif
+
+    {   const xxh_u8* p = (const xxh_u8*)input;
+        const xxh_u8* const bEnd = p + len;
+
+        state->total_len_32 += (XXH32_hash_t)len;
+        state->large_len |= (XXH32_hash_t)((len>=16) | (state->total_len_32>=16));
+
+        if (state->memsize + len < 16)  {   /* fill in tmp buffer */
+            XXH_memcpy((xxh_u8*)(state->mem32) + state->memsize, input, len);
+            state->memsize += (XXH32_hash_t)len;
+            return XXH_OK;
+        }
+
+        if (state->memsize) {   /* some data left from previous update */
+            XXH_memcpy((xxh_u8*)(state->mem32) + state->memsize, input, 16-state->memsize);
+            {   const xxh_u32* p32 = state->mem32;
+                state->v1 = XXH32_round(state->v1, XXH_readLE32(p32)); p32++;
+                state->v2 = XXH32_round(state->v2, XXH_readLE32(p32)); p32++;
+                state->v3 = XXH32_round(state->v3, XXH_readLE32(p32)); p32++;
+                state->v4 = XXH32_round(state->v4, XXH_readLE32(p32));
+            }
+            p += 16-state->memsize;
+            state->memsize = 0;
+        }
+
+        /* uintptr_t casts avoid UB or compiler warning on out-of-bounds
+         * pointer arithmetic */
+        if ((uintptr_t)p <= (uintptr_t)bEnd - 16) {
+            const uintptr_t limit = (uintptr_t)bEnd - 16;
+            xxh_u32 v1 = state->v1;
+            xxh_u32 v2 = state->v2;
+            xxh_u32 v3 = state->v3;
+            xxh_u32 v4 = state->v4;
+
+            do {
+                v1 = XXH32_round(v1, XXH_readLE32(p)); p+=4;
+                v2 = XXH32_round(v2, XXH_readLE32(p)); p+=4;
+                v3 = XXH32_round(v3, XXH_readLE32(p)); p+=4;
+                v4 = XXH32_round(v4, XXH_readLE32(p)); p+=4;
+            } while ((uintptr_t)p<=limit);
+
+            state->v1 = v1;
+            state->v2 = v2;
+            state->v3 = v3;
+            state->v4 = v4;
+        }
+
+        if (p < bEnd) {
+            XXH_memcpy(state->mem32, p, (size_t)(bEnd-p));
+            state->memsize = (unsigned)(bEnd-p);
+        }
+    }
+
+    return XXH_OK;
+}
+
+
+/*! @ingroup xxh32_family */
+XXH_PUBLIC_API XXH32_hash_t XXH32_digest(const XXH32_state_t* state)
+{
+    xxh_u32 h32;
+
+    if (state->large_len) {
+        h32 = XXH_rotl32(state->v1, 1)
+            + XXH_rotl32(state->v2, 7)
+            + XXH_rotl32(state->v3, 12)
+            + XXH_rotl32(state->v4, 18);
+    } else {
+        h32 = state->v3 /* == seed */ + XXH_PRIME32_5;
+    }
+
+    h32 += state->total_len_32;
+
+    return XXH32_finalize(h32, (const xxh_u8*)state->mem32, state->memsize, XXH_aligned);
+}
+
+
+/*******   Canonical representation   *******/
+
+/*!
+ * @ingroup xxh32_family
+ * The default return values from XXH functions are unsigned 32 and 64 bit
+ * integers.
+ *
+ * The canonical representation uses big endian convention, the same convention
+ * as human-readable numbers (large digits first).
+ *
+ * This way, hash values can be written into a file or buffer, remaining
+ * comparable across different systems.
+ *
+ * The following functions allow transformation of hash values to and from their
+ * canonical format.
+ */
+XXH_PUBLIC_API void XXH32_canonicalFromHash(XXH32_canonical_t* dst, XXH32_hash_t hash)
+{
+    XXH_STATIC_ASSERT(sizeof(XXH32_canonical_t) == sizeof(XXH32_hash_t));
+    if (XXH_CPU_LITTLE_ENDIAN) hash = XXH_swap32(hash);
+    memcpy(dst, &hash, sizeof(*dst));
+}
+/*! @ingroup xxh32_family */
+XXH_PUBLIC_API XXH32_hash_t XXH32_hashFromCanonical(const XXH32_canonical_t* src)
+{
+    return XXH_readBE32(src);
+}
+
+
+#ifndef XXH_NO_LONG_LONG
+
+/* *******************************************************************
+*  64-bit hash functions
+*********************************************************************/
+/*!
+ * @}
+ * @ingroup impl
+ * @{
+ */
+/*******   Memory access   *******/
+
+typedef XXH64_hash_t xxh_u64;
+
+#ifdef XXH_OLD_NAMES
+#  define U64 xxh_u64
+#endif
+
+#if (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==3))
+/*
+ * Manual byteshift. Best for old compilers which don't inline memcpy.
+ * We actually directly use XXH_readLE64 and XXH_readBE64.
+ */
+#elif (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==2))
+
+/* Force direct memory access. Only works on CPU which support unaligned memory access in hardware */
+static xxh_u64 XXH_read64(const void* memPtr)
+{
+    return *(const xxh_u64*) memPtr;
+}
+
+#elif (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==1))
+
+/*
+ * __pack instructions are safer, but compiler specific, hence potentially
+ * problematic for some compilers.
+ *
+ * Currently only defined for GCC and ICC.
+ */
+#ifdef XXH_OLD_NAMES
+typedef union { xxh_u32 u32; xxh_u64 u64; } __attribute__((packed)) unalign64;
+#endif
+static xxh_u64 XXH_read64(const void* ptr)
+{
+    typedef union { xxh_u32 u32; xxh_u64 u64; } __attribute__((packed)) xxh_unalign64;
+    return ((const xxh_unalign64*)ptr)->u64;
+}
+
+#else
+
+/*
+ * Portable and safe solution. Generally efficient.
+ * see: https://stackoverflow.com/a/32095106/646947
+ */
+static xxh_u64 XXH_read64(const void* memPtr)
+{
+    xxh_u64 val;
+    memcpy(&val, memPtr, sizeof(val));
+    return val;
+}
+
+#endif   /* XXH_FORCE_DIRECT_MEMORY_ACCESS */
+
+#if defined(_MSC_VER)     /* Visual Studio */
+#  define XXH_swap64 _byteswap_uint64
+#elif XXH_GCC_VERSION >= 403
+#  define XXH_swap64 __builtin_bswap64
+#else
+static xxh_u64 XXH_swap64(xxh_u64 x)
+{
+    return  ((x << 56) & 0xff00000000000000ULL) |
+            ((x << 40) & 0x00ff000000000000ULL) |
+            ((x << 24) & 0x0000ff0000000000ULL) |
+            ((x << 8)  & 0x000000ff00000000ULL) |
+            ((x >> 8)  & 0x00000000ff000000ULL) |
+            ((x >> 24) & 0x0000000000ff0000ULL) |
+            ((x >> 40) & 0x000000000000ff00ULL) |
+            ((x >> 56) & 0x00000000000000ffULL);
+}
+#endif
+
+
+/* XXH_FORCE_MEMORY_ACCESS==3 is an endian-independent byteshift load. */
+#if (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==3))
+
+XXH_FORCE_INLINE xxh_u64 XXH_readLE64(const void* memPtr)
+{
+    const xxh_u8* bytePtr = (const xxh_u8 *)memPtr;
+    return bytePtr[0]
+         | ((xxh_u64)bytePtr[1] << 8)
+         | ((xxh_u64)bytePtr[2] << 16)
+         | ((xxh_u64)bytePtr[3] << 24)
+         | ((xxh_u64)bytePtr[4] << 32)
+         | ((xxh_u64)bytePtr[5] << 40)
+         | ((xxh_u64)bytePtr[6] << 48)
+         | ((xxh_u64)bytePtr[7] << 56);
+}
+
+XXH_FORCE_INLINE xxh_u64 XXH_readBE64(const void* memPtr)
+{
+    const xxh_u8* bytePtr = (const xxh_u8 *)memPtr;
+    return bytePtr[7]
+         | ((xxh_u64)bytePtr[6] << 8)
+         | ((xxh_u64)bytePtr[5] << 16)
+         | ((xxh_u64)bytePtr[4] << 24)
+         | ((xxh_u64)bytePtr[3] << 32)
+         | ((xxh_u64)bytePtr[2] << 40)
+         | ((xxh_u64)bytePtr[1] << 48)
+         | ((xxh_u64)bytePtr[0] << 56);
+}
+
+#else
+XXH_FORCE_INLINE xxh_u64 XXH_readLE64(const void* ptr)
+{
+    return XXH_CPU_LITTLE_ENDIAN ? XXH_read64(ptr) : XXH_swap64(XXH_read64(ptr));
+}
+
+static xxh_u64 XXH_readBE64(const void* ptr)
+{
+    return XXH_CPU_LITTLE_ENDIAN ? XXH_swap64(XXH_read64(ptr)) : XXH_read64(ptr);
+}
+#endif
+
+XXH_FORCE_INLINE xxh_u64
+XXH_readLE64_align(const void* ptr, XXH_alignment align)
+{
+    if (align==XXH_unaligned)
+        return XXH_readLE64(ptr);
+    else
+        return XXH_CPU_LITTLE_ENDIAN ? *(const xxh_u64*)ptr : XXH_swap64(*(const xxh_u64*)ptr);
+}
+
+
+/*******   xxh64   *******/
+/*!
+ * @}
+ * @defgroup xxh64_impl XXH64 implementation
+ * @ingroup impl
+ * @{
+ */
+/* #define rather that static const, to be used as initializers */
+#define XXH_PRIME64_1  0x9E3779B185EBCA87ULL  /*!< 0b1001111000110111011110011011000110000101111010111100101010000111 */
+#define XXH_PRIME64_2  0xC2B2AE3D27D4EB4FULL  /*!< 0b1100001010110010101011100011110100100111110101001110101101001111 */
+#define XXH_PRIME64_3  0x165667B19E3779F9ULL  /*!< 0b0001011001010110011001111011000110011110001101110111100111111001 */
+#define XXH_PRIME64_4  0x85EBCA77C2B2AE63ULL  /*!< 0b1000010111101011110010100111011111000010101100101010111001100011 */
+#define XXH_PRIME64_5  0x27D4EB2F165667C5ULL  /*!< 0b0010011111010100111010110010111100010110010101100110011111000101 */
+
+#ifdef XXH_OLD_NAMES
+#  define PRIME64_1 XXH_PRIME64_1
+#  define PRIME64_2 XXH_PRIME64_2
+#  define PRIME64_3 XXH_PRIME64_3
+#  define PRIME64_4 XXH_PRIME64_4
+#  define PRIME64_5 XXH_PRIME64_5
+#endif
+
+static xxh_u64 XXH64_round(xxh_u64 acc, xxh_u64 input)
+{
+    acc += input * XXH_PRIME64_2;
+    acc  = XXH_rotl64(acc, 31);
+    acc *= XXH_PRIME64_1;
+    return acc;
+}
+
+static xxh_u64 XXH64_mergeRound(xxh_u64 acc, xxh_u64 val)
+{
+    val  = XXH64_round(0, val);
+    acc ^= val;
+    acc  = acc * XXH_PRIME64_1 + XXH_PRIME64_4;
+    return acc;
+}
+
+static xxh_u64 XXH64_avalanche(xxh_u64 h64)
+{
+    h64 ^= h64 >> 33;
+    h64 *= XXH_PRIME64_2;
+    h64 ^= h64 >> 29;
+    h64 *= XXH_PRIME64_3;
+    h64 ^= h64 >> 32;
+    return h64;
+}
+
+
+#define XXH_get64bits(p) XXH_readLE64_align(p, align)
+
+static xxh_u64
+XXH64_finalize(xxh_u64 h64, const xxh_u8* ptr, size_t len, XXH_alignment align)
+{
+    len &= 31;
+    while (len >= 8) {
+        xxh_u64 const k1 = XXH64_round(0, XXH_get64bits(ptr));
+        ptr += 8;
+        h64 ^= k1;
+        h64  = XXH_rotl64(h64,27) * XXH_PRIME64_1 + XXH_PRIME64_4;
+        len -= 8;
+    }
+    if (len >= 4) {
+        h64 ^= (xxh_u64)(XXH_get32bits(ptr)) * XXH_PRIME64_1;
+        ptr += 4;
+        h64 = XXH_rotl64(h64, 23) * XXH_PRIME64_2 + XXH_PRIME64_3;
+        len -= 4;
+    }
+    while (len > 0) {
+        h64 ^= (*ptr++) * XXH_PRIME64_5;
+        h64 = XXH_rotl64(h64, 11) * XXH_PRIME64_1;
+        --len;
+    }
+    return  XXH64_avalanche(h64);
+}
+
+#ifdef XXH_OLD_NAMES
+#  define PROCESS1_64 XXH_PROCESS1_64
+#  define PROCESS4_64 XXH_PROCESS4_64
+#  define PROCESS8_64 XXH_PROCESS8_64
+#else
+#  undef XXH_PROCESS1_64
+#  undef XXH_PROCESS4_64
+#  undef XXH_PROCESS8_64
+#endif
+
+XXH_FORCE_INLINE xxh_u64
+XXH64_endian_align(const xxh_u8* input, size_t len, xxh_u64 seed, XXH_alignment align)
+{
+    const xxh_u8* bEnd = input ? input + len : NULL;
+    xxh_u64 h64;
+
+#if defined(XXH_ACCEPT_NULL_INPUT_POINTER) && (XXH_ACCEPT_NULL_INPUT_POINTER>=1)
+    if (input==NULL) {
+        len=0;
+        bEnd=input=(const xxh_u8*)(size_t)32;
+    }
+#endif
+
+    if (len>=32) {
+        const xxh_u8* const limit = bEnd - 32;
+        xxh_u64 v1 = seed + XXH_PRIME64_1 + XXH_PRIME64_2;
+        xxh_u64 v2 = seed + XXH_PRIME64_2;
+        xxh_u64 v3 = seed + 0;
+        xxh_u64 v4 = seed - XXH_PRIME64_1;
+
+        do {
+            v1 = XXH64_round(v1, XXH_get64bits(input)); input+=8;
+            v2 = XXH64_round(v2, XXH_get64bits(input)); input+=8;
+            v3 = XXH64_round(v3, XXH_get64bits(input)); input+=8;
+            v4 = XXH64_round(v4, XXH_get64bits(input)); input+=8;
+        } while (input<=limit);
+
+        h64 = XXH_rotl64(v1, 1) + XXH_rotl64(v2, 7) + XXH_rotl64(v3, 12) + XXH_rotl64(v4, 18);
+        h64 = XXH64_mergeRound(h64, v1);
+        h64 = XXH64_mergeRound(h64, v2);
+        h64 = XXH64_mergeRound(h64, v3);
+        h64 = XXH64_mergeRound(h64, v4);
+
+    } else {
+        h64  = seed + XXH_PRIME64_5;
+    }
+
+    h64 += (xxh_u64) len;
+
+    return XXH64_finalize(h64, input, len, align);
+}
+
+
+/*! @ingroup xxh64_family */
+XXH_PUBLIC_API XXH64_hash_t XXH64 (const void* input, size_t len, XXH64_hash_t seed)
+{
+#if 0
+    /* Simple version, good for code maintenance, but unfortunately slow for small inputs */
+    XXH64_state_t state;
+    XXH64_reset(&state, seed);
+    XXH64_update(&state, (const xxh_u8*)input, len);
+    return XXH64_digest(&state);
+#else
+    if (XXH_FORCE_ALIGN_CHECK) {
+        if ((((size_t)input) & 7)==0) {  /* Input is aligned, let's leverage the speed advantage */
+            return XXH64_endian_align((const xxh_u8*)input, len, seed, XXH_aligned);
+    }   }
+
+    return XXH64_endian_align((const xxh_u8*)input, len, seed, XXH_unaligned);
+
+#endif
+}
+
+/*******   Hash Streaming   *******/
+
+/*! @ingroup xxh64_family*/
+XXH_PUBLIC_API XXH64_state_t* XXH64_createState(void)
+{
+    return (XXH64_state_t*)XXH_malloc(sizeof(XXH64_state_t));
+}
+/*! @ingroup xxh64_family */
+XXH_PUBLIC_API XXH_errorcode XXH64_freeState(XXH64_state_t* statePtr)
+{
+    XXH_free(statePtr);
+    return XXH_OK;
+}
+
+/*! @ingroup xxh64_family */
+XXH_PUBLIC_API void XXH64_copyState(XXH64_state_t* dstState, const XXH64_state_t* srcState)
+{
+    memcpy(dstState, srcState, sizeof(*dstState));
+}
+
+/*! @ingroup xxh64_family */
+XXH_PUBLIC_API XXH_errorcode XXH64_reset(XXH64_state_t* statePtr, XXH64_hash_t seed)
+{
+    XXH64_state_t state;   /* use a local state to memcpy() in order to avoid strict-aliasing warnings */
+    memset(&state, 0, sizeof(state));
+    state.v1 = seed + XXH_PRIME64_1 + XXH_PRIME64_2;
+    state.v2 = seed + XXH_PRIME64_2;
+    state.v3 = seed + 0;
+    state.v4 = seed - XXH_PRIME64_1;
+     /* do not write into reserved64, might be removed in a future version */
+    memcpy(statePtr, &state, sizeof(state) - sizeof(state.reserved64));
+    return XXH_OK;
+}
+
+/*! @ingroup xxh64_family */
+XXH_PUBLIC_API XXH_errorcode
+XXH64_update (XXH64_state_t* state, const void* input, size_t len)
+{
+    if (input==NULL)
+#if defined(XXH_ACCEPT_NULL_INPUT_POINTER) && (XXH_ACCEPT_NULL_INPUT_POINTER>=1)
+        return XXH_OK;
+#else
+        return XXH_ERROR;
+#endif
+
+    {   const xxh_u8* p = (const xxh_u8*)input;
+        const xxh_u8* const bEnd = p + len;
+
+        state->total_len += len;
+
+        if (state->memsize + len < 32) {  /* fill in tmp buffer */
+            XXH_memcpy(((xxh_u8*)state->mem64) + state->memsize, input, len);
+            state->memsize += (xxh_u32)len;
+            return XXH_OK;
+        }
+
+        if (state->memsize) {   /* tmp buffer is full */
+            XXH_memcpy(((xxh_u8*)state->mem64) + state->memsize, input, 32-state->memsize);
+            state->v1 = XXH64_round(state->v1, XXH_readLE64(state->mem64+0));
+            state->v2 = XXH64_round(state->v2, XXH_readLE64(state->mem64+1));
+            state->v3 = XXH64_round(state->v3, XXH_readLE64(state->mem64+2));
+            state->v4 = XXH64_round(state->v4, XXH_readLE64(state->mem64+3));
+            p += 32 - state->memsize;
+            state->memsize = 0;
+        }
+
+        /* uintptr_t casts avoid UB or compiler warning on out-of-bounds
+         * pointer arithmetic */
+        if ((uintptr_t)p + 32 <= (uintptr_t)bEnd) {
+            const uintptr_t limit = (uintptr_t)bEnd - 32;
+            xxh_u64 v1 = state->v1;
+            xxh_u64 v2 = state->v2;
+            xxh_u64 v3 = state->v3;
+            xxh_u64 v4 = state->v4;
+
+            do {
+                v1 = XXH64_round(v1, XXH_readLE64(p)); p+=8;
+                v2 = XXH64_round(v2, XXH_readLE64(p)); p+=8;
+                v3 = XXH64_round(v3, XXH_readLE64(p)); p+=8;
+                v4 = XXH64_round(v4, XXH_readLE64(p)); p+=8;
+            } while ((uintptr_t)p<=limit);
+
+            state->v1 = v1;
+            state->v2 = v2;
+            state->v3 = v3;
+            state->v4 = v4;
+        }
+
+        if (p < bEnd) {
+            XXH_memcpy(state->mem64, p, (size_t)(bEnd-p));
+            state->memsize = (unsigned)(bEnd-p);
+        }
+    }
+
+    return XXH_OK;
+}
+
+
+/*! @ingroup xxh64_family */
+XXH_PUBLIC_API XXH64_hash_t XXH64_digest(const XXH64_state_t* state)
+{
+    xxh_u64 h64;
+
+    if (state->total_len >= 32) {
+        xxh_u64 const v1 = state->v1;
+        xxh_u64 const v2 = state->v2;
+        xxh_u64 const v3 = state->v3;
+        xxh_u64 const v4 = state->v4;
+
+        h64 = XXH_rotl64(v1, 1) + XXH_rotl64(v2, 7) + XXH_rotl64(v3, 12) + XXH_rotl64(v4, 18);
+        h64 = XXH64_mergeRound(h64, v1);
+        h64 = XXH64_mergeRound(h64, v2);
+        h64 = XXH64_mergeRound(h64, v3);
+        h64 = XXH64_mergeRound(h64, v4);
+    } else {
+        h64  = state->v3 /*seed*/ + XXH_PRIME64_5;
+    }
+
+    h64 += (xxh_u64) state->total_len;
+
+    return XXH64_finalize(h64, (const xxh_u8*)state->mem64, (size_t)state->total_len, XXH_aligned);
+}
+
+
+/******* Canonical representation   *******/
+
+/*! @ingroup xxh64_family */
+XXH_PUBLIC_API void XXH64_canonicalFromHash(XXH64_canonical_t* dst, XXH64_hash_t hash)
+{
+    XXH_STATIC_ASSERT(sizeof(XXH64_canonical_t) == sizeof(XXH64_hash_t));
+    if (XXH_CPU_LITTLE_ENDIAN) hash = XXH_swap64(hash);
+    memcpy(dst, &hash, sizeof(*dst));
+}
+
+/*! @ingroup xxh64_family */
+XXH_PUBLIC_API XXH64_hash_t XXH64_hashFromCanonical(const XXH64_canonical_t* src)
+{
+    return XXH_readBE64(src);
+}
+
+#ifndef XXH_NO_XXH3
+
+/* *********************************************************************
+*  XXH3
+*  New generation hash designed for speed on small keys and vectorization
+************************************************************************ */
+/*!
+ * @}
+ * @defgroup xxh3_impl XXH3 implementation
+ * @ingroup impl
+ * @{
+ */
+
+/* ===   Compiler specifics   === */
+
+#if ((defined(sun) || defined(__sun)) && __cplusplus) /* Solaris includes __STDC_VERSION__ with C++. Tested with GCC 5.5 */
+#  define XXH_RESTRICT /* disable */
+#elif defined (__STDC_VERSION__) && __STDC_VERSION__ >= 199901L   /* >= C99 */
+#  define XXH_RESTRICT   restrict
+#else
+/* Note: it might be useful to define __restrict or __restrict__ for some C++ compilers */
+#  define XXH_RESTRICT   /* disable */
+#endif
+
+#if (defined(__GNUC__) && (__GNUC__ >= 3))  \
+  || (defined(__INTEL_COMPILER) && (__INTEL_COMPILER >= 800)) \
+  || defined(__clang__)
+#    define XXH_likely(x) __builtin_expect(x, 1)
+#    define XXH_unlikely(x) __builtin_expect(x, 0)
+#else
+#    define XXH_likely(x) (x)
+#    define XXH_unlikely(x) (x)
+#endif
+
+#if defined(__GNUC__)
+#  if defined(__AVX2__)
+#    include <immintrin.h>
+#  elif defined(__SSE2__)
+#    include <emmintrin.h>
+#  elif defined(__ARM_NEON__) || defined(__ARM_NEON)
+#    define inline __inline__  /* circumvent a clang bug */
+#    include <arm_neon.h>
+#    undef inline
+#  endif
+#elif defined(_MSC_VER)
+#  include <intrin.h>
+#endif
+
+/*
+ * One goal of XXH3 is to make it fast on both 32-bit and 64-bit, while
+ * remaining a true 64-bit/128-bit hash function.
+ *
+ * This is done by prioritizing a subset of 64-bit operations that can be
+ * emulated without too many steps on the average 32-bit machine.
+ *
+ * For example, these two lines seem similar, and run equally fast on 64-bit:
+ *
+ *   xxh_u64 x;
+ *   x ^= (x >> 47); // good
+ *   x ^= (x >> 13); // bad
+ *
+ * However, to a 32-bit machine, there is a major difference.
+ *
+ * x ^= (x >> 47) looks like this:
+ *
+ *   x.lo ^= (x.hi >> (47 - 32));
+ *
+ * while x ^= (x >> 13) looks like this:
+ *
+ *   // note: funnel shifts are not usually cheap.
+ *   x.lo ^= (x.lo >> 13) | (x.hi << (32 - 13));
+ *   x.hi ^= (x.hi >> 13);
+ *
+ * The first one is significantly faster than the second, simply because the
+ * shift is larger than 32. This means:
+ *  - All the bits we need are in the upper 32 bits, so we can ignore the lower
+ *    32 bits in the shift.
+ *  - The shift result will always fit in the lower 32 bits, and therefore,
+ *    we can ignore the upper 32 bits in the xor.
+ *
+ * Thanks to this optimization, XXH3 only requires these features to be efficient:
+ *
+ *  - Usable unaligned access
+ *  - A 32-bit or 64-bit ALU
+ *      - If 32-bit, a decent ADC instruction
+ *  - A 32 or 64-bit multiply with a 64-bit result
+ *  - For the 128-bit variant, a decent byteswap helps short inputs.
+ *
+ * The first two are already required by XXH32, and almost all 32-bit and 64-bit
+ * platforms which can run XXH32 can run XXH3 efficiently.
+ *
+ * Thumb-1, the classic 16-bit only subset of ARM's instruction set, is one
+ * notable exception.
+ *
+ * First of all, Thumb-1 lacks support for the UMULL instruction which
+ * performs the important long multiply. This means numerous __aeabi_lmul
+ * calls.
+ *
+ * Second of all, the 8 functional registers are just not enough.
+ * Setup for __aeabi_lmul, byteshift loads, pointers, and all arithmetic need
+ * Lo registers, and this shuffling results in thousands more MOVs than A32.
+ *
+ * A32 and T32 don't have this limitation. They can access all 14 registers,
+ * do a 32->64 multiply with UMULL, and the flexible operand allowing free
+ * shifts is helpful, too.
+ *
+ * Therefore, we do a quick sanity check.
+ *
+ * If compiling Thumb-1 for a target which supports ARM instructions, we will
+ * emit a warning, as it is not a "sane" platform to compile for.
+ *
+ * Usually, if this happens, it is because of an accident and you probably need
+ * to specify -march, as you likely meant to compile for a newer architecture.
+ *
+ * Credit: large sections of the vectorial and asm source code paths
+ *         have been contributed by @easyaspi314
+ */
+#if defined(__thumb__) && !defined(__thumb2__) && defined(__ARM_ARCH_ISA_ARM)
+#   warning "XXH3 is highly inefficient without ARM or Thumb-2."
+#endif
+
+/* ==========================================
+ * Vectorization detection
+ * ========================================== */
+
+#ifdef XXH_DOXYGEN
+/*!
+ * @ingroup tuning
+ * @brief Overrides the vectorization implementation chosen for XXH3.
+ *
+ * Can be defined to 0 to disable SIMD or any of the values mentioned in
+ * @ref XXH_VECTOR_TYPE.
+ *
+ * If this is not defined, it uses predefined macros to determine the best
+ * implementation.
+ */
+#  define XXH_VECTOR XXH_SCALAR
+/*!
+ * @ingroup tuning
+ * @brief Possible values for @ref XXH_VECTOR.
+ *
+ * Note that these are actually implemented as macros.
+ *
+ * If this is not defined, it is detected automatically.
+ * @ref XXH_X86DISPATCH overrides this.
+ */
+enum XXH_VECTOR_TYPE /* fake enum */ {
+    XXH_SCALAR = 0,  /*!< Portable scalar version */
+    XXH_SSE2   = 1,  /*!<
+                      * SSE2 for Pentium 4, Opteron, all x86_64.
+                      *
+                      * @note SSE2 is also guaranteed on Windows 10, macOS, and
+                      * Android x86.
+                      */
+    XXH_AVX2   = 2,  /*!< AVX2 for Haswell and Bulldozer */
+    XXH_AVX512 = 3,  /*!< AVX512 for Skylake and Icelake */
+    XXH_NEON   = 4,  /*!< NEON for most ARMv7-A and all AArch64 */
+    XXH_VSX    = 5,  /*!< VSX and ZVector for POWER8/z13 (64-bit) */
+};
+/*!
+ * @ingroup tuning
+ * @brief Selects the minimum alignment for XXH3's accumulators.
+ *
+ * When using SIMD, this should match the alignment reqired for said vector
+ * type, so, for example, 32 for AVX2.
+ *
+ * Default: Auto detected.
+ */
+#  define XXH_ACC_ALIGN 8
+#endif
+
+/* Actual definition */
+#ifndef XXH_DOXYGEN
+#  define XXH_SCALAR 0
+#  define XXH_SSE2   1
+#  define XXH_AVX2   2
+#  define XXH_AVX512 3
+#  define XXH_NEON   4
+#  define XXH_VSX    5
+#endif
+
+#ifndef XXH_VECTOR    /* can be defined on command line */
+#  if defined(__AVX512F__)
+#    define XXH_VECTOR XXH_AVX512
+#  elif defined(__AVX2__)
+#    define XXH_VECTOR XXH_AVX2
+#  elif defined(__SSE2__) || defined(_M_AMD64) || defined(_M_X64) || (defined(_M_IX86_FP) && (_M_IX86_FP == 2))
+#    define XXH_VECTOR XXH_SSE2
+#  elif defined(__GNUC__) /* msvc support maybe later */ \
+  && (defined(__ARM_NEON__) || defined(__ARM_NEON)) \
+  && (defined(__LITTLE_ENDIAN__) /* We only support little endian NEON */ \
+    || (defined(__BYTE_ORDER__) && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__))
+#    define XXH_VECTOR XXH_NEON
+#  elif (defined(__PPC64__) && defined(__POWER8_VECTOR__)) \
+     || (defined(__s390x__) && defined(__VEC__)) \
+     && defined(__GNUC__) /* TODO: IBM XL */
+#    define XXH_VECTOR XXH_VSX
+#  else
+#    define XXH_VECTOR XXH_SCALAR
+#  endif
+#endif
+
+/*
+ * Controls the alignment of the accumulator,
+ * for compatibility with aligned vector loads, which are usually faster.
+ */
+#ifndef XXH_ACC_ALIGN
+#  if defined(XXH_X86DISPATCH)
+#     define XXH_ACC_ALIGN 64  /* for compatibility with avx512 */
+#  elif XXH_VECTOR == XXH_SCALAR  /* scalar */
+#     define XXH_ACC_ALIGN 8
+#  elif XXH_VECTOR == XXH_SSE2  /* sse2 */
+#     define XXH_ACC_ALIGN 16
+#  elif XXH_VECTOR == XXH_AVX2  /* avx2 */
+#     define XXH_ACC_ALIGN 32
+#  elif XXH_VECTOR == XXH_NEON  /* neon */
+#     define XXH_ACC_ALIGN 16
+#  elif XXH_VECTOR == XXH_VSX   /* vsx */
+#     define XXH_ACC_ALIGN 16
+#  elif XXH_VECTOR == XXH_AVX512  /* avx512 */
+#     define XXH_ACC_ALIGN 64
+#  endif
+#endif
+
+#if defined(XXH_X86DISPATCH) || XXH_VECTOR == XXH_SSE2 \
+    || XXH_VECTOR == XXH_AVX2 || XXH_VECTOR == XXH_AVX512
+#  define XXH_SEC_ALIGN XXH_ACC_ALIGN
+#else
+#  define XXH_SEC_ALIGN 8
+#endif
+
+/*
+ * UGLY HACK:
+ * GCC usually generates the best code with -O3 for xxHash.
+ *
+ * However, when targeting AVX2, it is overzealous in its unrolling resulting
+ * in code roughly 3/4 the speed of Clang.
+ *
+ * There are other issues, such as GCC splitting _mm256_loadu_si256 into
+ * _mm_loadu_si128 + _mm256_inserti128_si256. This is an optimization which
+ * only applies to Sandy and Ivy Bridge... which don't even support AVX2.
+ *
+ * That is why when compiling the AVX2 version, it is recommended to use either
+ *   -O2 -mavx2 -march=haswell
+ * or
+ *   -O2 -mavx2 -mno-avx256-split-unaligned-load
+ * for decent performance, or to use Clang instead.
+ *
+ * Fortunately, we can control the first one with a pragma that forces GCC into
+ * -O2, but the other one we can't control without "failed to inline always
+ * inline function due to target mismatch" warnings.
+ */
+#if XXH_VECTOR == XXH_AVX2 /* AVX2 */ \
+  && defined(__GNUC__) && !defined(__clang__) /* GCC, not Clang */ \
+  && defined(__OPTIMIZE__) && !defined(__OPTIMIZE_SIZE__) /* respect -O0 and -Os */
+#  pragma GCC push_options
+#  pragma GCC optimize("-O2")
+#endif
+
+
+#if XXH_VECTOR == XXH_NEON
+/*
+ * NEON's setup for vmlal_u32 is a little more complicated than it is on
+ * SSE2, AVX2, and VSX.
+ *
+ * While PMULUDQ and VMULEUW both perform a mask, VMLAL.U32 performs an upcast.
+ *
+ * To do the same operation, the 128-bit 'Q' register needs to be split into
+ * two 64-bit 'D' registers, performing this operation::
+ *
+ *   [                a                 |                 b                ]
+ *            |              '---------. .--------'                |
+ *            |                         x                          |
+ *            |              .---------' '--------.                |
+ *   [ a & 0xFFFFFFFF | b & 0xFFFFFFFF ],[    a >> 32     |     b >> 32    ]
+ *
+ * Due to significant changes in aarch64, the fastest method for aarch64 is
+ * completely different than the fastest method for ARMv7-A.
+ *
+ * ARMv7-A treats D registers as unions overlaying Q registers, so modifying
+ * D11 will modify the high half of Q5. This is similar to how modifying AH
+ * will only affect bits 8-15 of AX on x86.
+ *
+ * VZIP takes two registers, and puts even lanes in one register and odd lanes
+ * in the other.
+ *
+ * On ARMv7-A, this strangely modifies both parameters in place instead of
+ * taking the usual 3-operand form.
+ *
+ * Therefore, if we want to do this, we can simply use a D-form VZIP.32 on the
+ * lower and upper halves of the Q register to end up with the high and low
+ * halves where we want - all in one instruction.
+ *
+ *   vzip.32   d10, d11       @ d10 = { d10[0], d11[0] }; d11 = { d10[1], d11[1] }
+ *
+ * Unfortunately we need inline assembly for this: Instructions modifying two
+ * registers at once is not possible in GCC or Clang's IR, and they have to
+ * create a copy.
+ *
+ * aarch64 requires a different approach.
+ *
+ * In order to make it easier to write a decent compiler for aarch64, many
+ * quirks were removed, such as conditional execution.
+ *
+ * NEON was also affected by this.
+ *
+ * aarch64 cannot access the high bits of a Q-form register, and writes to a
+ * D-form register zero the high bits, similar to how writes to W-form scalar
+ * registers (or DWORD registers on x86_64) work.
+ *
+ * The formerly free vget_high intrinsics now require a vext (with a few
+ * exceptions)
+ *
+ * Additionally, VZIP was replaced by ZIP1 and ZIP2, which are the equivalent
+ * of PUNPCKL* and PUNPCKH* in SSE, respectively, in order to only modify one
+ * operand.
+ *
+ * The equivalent of the VZIP.32 on the lower and upper halves would be this
+ * mess:
+ *
+ *   ext     v2.4s, v0.4s, v0.4s, #2 // v2 = { v0[2], v0[3], v0[0], v0[1] }
+ *   zip1    v1.2s, v0.2s, v2.2s     // v1 = { v0[0], v2[0] }
+ *   zip2    v0.2s, v0.2s, v1.2s     // v0 = { v0[1], v2[1] }
+ *
+ * Instead, we use a literal downcast, vmovn_u64 (XTN), and vshrn_n_u64 (SHRN):
+ *
+ *   shrn    v1.2s, v0.2d, #32  // v1 = (uint32x2_t)(v0 >> 32);
+ *   xtn     v0.2s, v0.2d       // v0 = (uint32x2_t)(v0 & 0xFFFFFFFF);
+ *
+ * This is available on ARMv7-A, but is less efficient than a single VZIP.32.
+ */
+
+/*!
+ * Function-like macro:
+ * void XXH_SPLIT_IN_PLACE(uint64x2_t &in, uint32x2_t &outLo, uint32x2_t &outHi)
+ * {
+ *     outLo = (uint32x2_t)(in & 0xFFFFFFFF);
+ *     outHi = (uint32x2_t)(in >> 32);
+ *     in = UNDEFINED;
+ * }
+ */
+# if !defined(XXH_NO_VZIP_HACK) /* define to disable */ \
+   && defined(__GNUC__) \
+   && !defined(__aarch64__) && !defined(__arm64__)
+#  define XXH_SPLIT_IN_PLACE(in, outLo, outHi)                                              \
+    do {                                                                                    \
+      /* Undocumented GCC/Clang operand modifier: %e0 = lower D half, %f0 = upper D half */ \
+      /* https://github.com/gcc-mirror/gcc/blob/38cf91e5/gcc/config/arm/arm.c#L22486 */     \
+      /* https://github.com/llvm-mirror/llvm/blob/2c4ca683/lib/Target/ARM/ARMAsmPrinter.cpp#L399 */ \
+      __asm__("vzip.32  %e0, %f0" : "+w" (in));                                             \
+      (outLo) = vget_low_u32 (vreinterpretq_u32_u64(in));                                   \
+      (outHi) = vget_high_u32(vreinterpretq_u32_u64(in));                                   \
+   } while (0)
+# else
+#  define XXH_SPLIT_IN_PLACE(in, outLo, outHi)                                            \
+    do {                                                                                  \
+      (outLo) = vmovn_u64    (in);                                                        \
+      (outHi) = vshrn_n_u64  ((in), 32);                                                  \
+    } while (0)
+# endif
+#endif  /* XXH_VECTOR == XXH_NEON */
+
+/*
+ * VSX and Z Vector helpers.
+ *
+ * This is very messy, and any pull requests to clean this up are welcome.
+ *
+ * There are a lot of problems with supporting VSX and s390x, due to
+ * inconsistent intrinsics, spotty coverage, and multiple endiannesses.
+ */
+#if XXH_VECTOR == XXH_VSX
+#  if defined(__s390x__)
+#    include <s390intrin.h>
+#  else
+/* gcc's altivec.h can have the unwanted consequence to unconditionally
+ * #define bool, vector, and pixel keywords,
+ * with bad consequences for programs already using these keywords for other purposes.
+ * The paragraph defining these macros is skipped when __APPLE_ALTIVEC__ is defined.
+ * __APPLE_ALTIVEC__ is _generally_ defined automatically by the compiler,
+ * but it seems that, in some cases, it isn't.
+ * Force the build macro to be defined, so that keywords are not altered.
+ */
+#    if defined(__GNUC__) && !defined(__APPLE_ALTIVEC__)
+#      define __APPLE_ALTIVEC__
+#    endif
+#    include <altivec.h>
+#  endif
+
+typedef __vector unsigned long long xxh_u64x2;
+typedef __vector unsigned char xxh_u8x16;
+typedef __vector unsigned xxh_u32x4;
+
+# ifndef XXH_VSX_BE
+#  if defined(__BIG_ENDIAN__) \
+  || (defined(__BYTE_ORDER__) && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__)
+#    define XXH_VSX_BE 1
+#  elif defined(__VEC_ELEMENT_REG_ORDER__) && __VEC_ELEMENT_REG_ORDER__ == __ORDER_BIG_ENDIAN__
+#    warning "-maltivec=be is not recommended. Please use native endianness."
+#    define XXH_VSX_BE 1
+#  else
+#    define XXH_VSX_BE 0
+#  endif
+# endif /* !defined(XXH_VSX_BE) */
+
+# if XXH_VSX_BE
+#  if defined(__POWER9_VECTOR__) || (defined(__clang__) && defined(__s390x__))
+#    define XXH_vec_revb vec_revb
+#  else
+/*!
+ * A polyfill for POWER9's vec_revb().
+ */
+XXH_FORCE_INLINE xxh_u64x2 XXH_vec_revb(xxh_u64x2 val)
+{
+    xxh_u8x16 const vByteSwap = { 0x07, 0x06, 0x05, 0x04, 0x03, 0x02, 0x01, 0x00,
+                                  0x0F, 0x0E, 0x0D, 0x0C, 0x0B, 0x0A, 0x09, 0x08 };
+    return vec_perm(val, val, vByteSwap);
+}
+#  endif
+# endif /* XXH_VSX_BE */
+
+/*!
+ * Performs an unaligned vector load and byte swaps it on big endian.
+ */
+XXH_FORCE_INLINE xxh_u64x2 XXH_vec_loadu(const void *ptr)
+{
+    xxh_u64x2 ret;
+    memcpy(&ret, ptr, sizeof(xxh_u64x2));
+# if XXH_VSX_BE
+    ret = XXH_vec_revb(ret);
+# endif
+    return ret;
+}
+
+/*
+ * vec_mulo and vec_mule are very problematic intrinsics on PowerPC
+ *
+ * These intrinsics weren't added until GCC 8, despite existing for a while,
+ * and they are endian dependent. Also, their meaning swap depending on version.
+ * */
+# if defined(__s390x__)
+ /* s390x is always big endian, no issue on this platform */
+#  define XXH_vec_mulo vec_mulo
+#  define XXH_vec_mule vec_mule
+# elif defined(__clang__) && XXH_HAS_BUILTIN(__builtin_altivec_vmuleuw)
+/* Clang has a better way to control this, we can just use the builtin which doesn't swap. */
+#  define XXH_vec_mulo __builtin_altivec_vmulouw
+#  define XXH_vec_mule __builtin_altivec_vmuleuw
+# else
+/* gcc needs inline assembly */
+/* Adapted from https://github.com/google/highwayhash/blob/master/highwayhash/hh_vsx.h. */
+XXH_FORCE_INLINE xxh_u64x2 XXH_vec_mulo(xxh_u32x4 a, xxh_u32x4 b)
+{
+    xxh_u64x2 result;
+    __asm__("vmulouw %0, %1, %2" : "=v" (result) : "v" (a), "v" (b));
+    return result;
+}
+XXH_FORCE_INLINE xxh_u64x2 XXH_vec_mule(xxh_u32x4 a, xxh_u32x4 b)
+{
+    xxh_u64x2 result;
+    __asm__("vmuleuw %0, %1, %2" : "=v" (result) : "v" (a), "v" (b));
+    return result;
+}
+# endif /* XXH_vec_mulo, XXH_vec_mule */
+#endif /* XXH_VECTOR == XXH_VSX */
+
+
+/* prefetch
+ * can be disabled, by declaring XXH_NO_PREFETCH build macro */
+#if defined(XXH_NO_PREFETCH)
+#  define XXH_PREFETCH(ptr)  (void)(ptr)  /* disabled */
+#else
+#  if defined(_MSC_VER) && (defined(_M_X64) || defined(_M_IX86))  /* _mm_prefetch() not defined outside of x86/x64 */
+#    include <mmintrin.h>   /* https://msdn.microsoft.com/fr-fr/library/84szxsww(v=vs.90).aspx */
+#    define XXH_PREFETCH(ptr)  _mm_prefetch((const char*)(ptr), _MM_HINT_T0)
+#  elif defined(__GNUC__) && ( (__GNUC__ >= 4) || ( (__GNUC__ == 3) && (__GNUC_MINOR__ >= 1) ) )
+#    define XXH_PREFETCH(ptr)  __builtin_prefetch((ptr), 0 /* rw==read */, 3 /* locality */)
+#  else
+#    define XXH_PREFETCH(ptr) (void)(ptr)  /* disabled */
+#  endif
+#endif  /* XXH_NO_PREFETCH */
+
+
+/* ==========================================
+ * XXH3 default settings
+ * ========================================== */
+
+#define XXH_SECRET_DEFAULT_SIZE 192   /* minimum XXH3_SECRET_SIZE_MIN */
+
+#if (XXH_SECRET_DEFAULT_SIZE < XXH3_SECRET_SIZE_MIN)
+#  error "default keyset is not large enough"
+#endif
+
+/*! Pseudorandom secret taken directly from FARSH. */
+XXH_ALIGN(64) static const xxh_u8 XXH3_kSecret[XXH_SECRET_DEFAULT_SIZE] = {
+    0xb8, 0xfe, 0x6c, 0x39, 0x23, 0xa4, 0x4b, 0xbe, 0x7c, 0x01, 0x81, 0x2c, 0xf7, 0x21, 0xad, 0x1c,
+    0xde, 0xd4, 0x6d, 0xe9, 0x83, 0x90, 0x97, 0xdb, 0x72, 0x40, 0xa4, 0xa4, 0xb7, 0xb3, 0x67, 0x1f,
+    0xcb, 0x79, 0xe6, 0x4e, 0xcc, 0xc0, 0xe5, 0x78, 0x82, 0x5a, 0xd0, 0x7d, 0xcc, 0xff, 0x72, 0x21,
+    0xb8, 0x08, 0x46, 0x74, 0xf7, 0x43, 0x24, 0x8e, 0xe0, 0x35, 0x90, 0xe6, 0x81, 0x3a, 0x26, 0x4c,
+    0x3c, 0x28, 0x52, 0xbb, 0x91, 0xc3, 0x00, 0xcb, 0x88, 0xd0, 0x65, 0x8b, 0x1b, 0x53, 0x2e, 0xa3,
+    0x71, 0x64, 0x48, 0x97, 0xa2, 0x0d, 0xf9, 0x4e, 0x38, 0x19, 0xef, 0x46, 0xa9, 0xde, 0xac, 0xd8,
+    0xa8, 0xfa, 0x76, 0x3f, 0xe3, 0x9c, 0x34, 0x3f, 0xf9, 0xdc, 0xbb, 0xc7, 0xc7, 0x0b, 0x4f, 0x1d,
+    0x8a, 0x51, 0xe0, 0x4b, 0xcd, 0xb4, 0x59, 0x31, 0xc8, 0x9f, 0x7e, 0xc9, 0xd9, 0x78, 0x73, 0x64,
+    0xea, 0xc5, 0xac, 0x83, 0x34, 0xd3, 0xeb, 0xc3, 0xc5, 0x81, 0xa0, 0xff, 0xfa, 0x13, 0x63, 0xeb,
+    0x17, 0x0d, 0xdd, 0x51, 0xb7, 0xf0, 0xda, 0x49, 0xd3, 0x16, 0x55, 0x26, 0x29, 0xd4, 0x68, 0x9e,
+    0x2b, 0x16, 0xbe, 0x58, 0x7d, 0x47, 0xa1, 0xfc, 0x8f, 0xf8, 0xb8, 0xd1, 0x7a, 0xd0, 0x31, 0xce,
+    0x45, 0xcb, 0x3a, 0x8f, 0x95, 0x16, 0x04, 0x28, 0xaf, 0xd7, 0xfb, 0xca, 0xbb, 0x4b, 0x40, 0x7e,
+};
+
+
+#ifdef XXH_OLD_NAMES
+#  define kSecret XXH3_kSecret
+#endif
+
+#ifdef XXH_DOXYGEN
+/*!
+ * @brief Calculates a 32-bit to 64-bit long multiply.
+ *
+ * Implemented as a macro.
+ *
+ * Wraps `__emulu` on MSVC x86 because it tends to call `__allmul` when it doesn't
+ * need to (but it shouldn't need to anyways, it is about 7 instructions to do
+ * a 64x64 multiply...). Since we know that this will _always_ emit `MULL`, we
+ * use that instead of the normal method.
+ *
+ * If you are compiling for platforms like Thumb-1 and don't have a better option,
+ * you may also want to write your own long multiply routine here.
+ *
+ * @param x, y Numbers to be multiplied
+ * @return 64-bit product of the low 32 bits of @p x and @p y.
+ */
+XXH_FORCE_INLINE xxh_u64
+XXH_mult32to64(xxh_u64 x, xxh_u64 y)
+{
+   return (x & 0xFFFFFFFF) * (y & 0xFFFFFFFF);
+}
+#elif defined(_MSC_VER) && defined(_M_IX86)
+#    include <intrin.h>
+#    define XXH_mult32to64(x, y) __emulu((unsigned)(x), (unsigned)(y))
+#else
+/*
+ * Downcast + upcast is usually better than masking on older compilers like
+ * GCC 4.2 (especially 32-bit ones), all without affecting newer compilers.
+ *
+ * The other method, (x & 0xFFFFFFFF) * (y & 0xFFFFFFFF), will AND both operands
+ * and perform a full 64x64 multiply -- entirely redundant on 32-bit.
+ */
+#    define XXH_mult32to64(x, y) ((xxh_u64)(xxh_u32)(x) * (xxh_u64)(xxh_u32)(y))
+#endif
+
+/*!
+ * @brief Calculates a 64->128-bit long multiply.
+ *
+ * Uses `__uint128_t` and `_umul128` if available, otherwise uses a scalar
+ * version.
+ *
+ * @param lhs, rhs The 64-bit integers to be multiplied
+ * @return The 128-bit result represented in an @ref XXH128_hash_t.
+ */
+static XXH128_hash_t
+XXH_mult64to128(xxh_u64 lhs, xxh_u64 rhs)
+{
+    /*
+     * GCC/Clang __uint128_t method.
+     *
+     * On most 64-bit targets, GCC and Clang define a __uint128_t type.
+     * This is usually the best way as it usually uses a native long 64-bit
+     * multiply, such as MULQ on x86_64 or MUL + UMULH on aarch64.
+     *
+     * Usually.
+     *
+     * Despite being a 32-bit platform, Clang (and emscripten) define this type
+     * despite not having the arithmetic for it. This results in a laggy
+     * compiler builtin call which calculates a full 128-bit multiply.
+     * In that case it is best to use the portable one.
+     * https://github.com/Cyan4973/xxHash/issues/211#issuecomment-515575677
+     */
+#if defined(__GNUC__) && !defined(__wasm__) \
+    && defined(__SIZEOF_INT128__) \
+    || (defined(_INTEGRAL_MAX_BITS) && _INTEGRAL_MAX_BITS >= 128)
+
+    __uint128_t const product = (__uint128_t)lhs * (__uint128_t)rhs;
+    XXH128_hash_t r128;
+    r128.low64  = (xxh_u64)(product);
+    r128.high64 = (xxh_u64)(product >> 64);
+    return r128;
+
+    /*
+     * MSVC for x64's _umul128 method.
+     *
+     * xxh_u64 _umul128(xxh_u64 Multiplier, xxh_u64 Multiplicand, xxh_u64 *HighProduct);
+     *
+     * This compiles to single operand MUL on x64.
+     */
+#elif defined(_M_X64) || defined(_M_IA64)
+
+#ifndef _MSC_VER
+#   pragma intrinsic(_umul128)
+#endif
+    xxh_u64 product_high;
+    xxh_u64 const product_low = _umul128(lhs, rhs, &product_high);
+    XXH128_hash_t r128;
+    r128.low64  = product_low;
+    r128.high64 = product_high;
+    return r128;
+
+#else
+    /*
+     * Portable scalar method. Optimized for 32-bit and 64-bit ALUs.
+     *
+     * This is a fast and simple grade school multiply, which is shown below
+     * with base 10 arithmetic instead of base 0x100000000.
+     *
+     *           9 3 // D2 lhs = 93
+     *         x 7 5 // D2 rhs = 75
+     *     ----------
+     *           1 5 // D2 lo_lo = (93 % 10) * (75 % 10) = 15
+     *         4 5 | // D2 hi_lo = (93 / 10) * (75 % 10) = 45
+     *         2 1 | // D2 lo_hi = (93 % 10) * (75 / 10) = 21
+     *     + 6 3 | | // D2 hi_hi = (93 / 10) * (75 / 10) = 63
+     *     ---------
+     *         2 7 | // D2 cross = (15 / 10) + (45 % 10) + 21 = 27
+     *     + 6 7 | | // D2 upper = (27 / 10) + (45 / 10) + 63 = 67
+     *     ---------
+     *       6 9 7 5 // D4 res = (27 * 10) + (15 % 10) + (67 * 100) = 6975
+     *
+     * The reasons for adding the products like this are:
+     *  1. It avoids manual carry tracking. Just like how
+     *     (9 * 9) + 9 + 9 = 99, the same applies with this for UINT64_MAX.
+     *     This avoids a lot of complexity.
+     *
+     *  2. It hints for, and on Clang, compiles to, the powerful UMAAL
+     *     instruction available in ARM's Digital Signal Processing extension
+     *     in 32-bit ARMv6 and later, which is shown below:
+     *
+     *         void UMAAL(xxh_u32 *RdLo, xxh_u32 *RdHi, xxh_u32 Rn, xxh_u32 Rm)
+     *         {
+     *             xxh_u64 product = (xxh_u64)*RdLo * (xxh_u64)*RdHi + Rn + Rm;
+     *             *RdLo = (xxh_u32)(product & 0xFFFFFFFF);
+     *             *RdHi = (xxh_u32)(product >> 32);
+     *         }
+     *
+     *     This instruction was designed for efficient long multiplication, and
+     *     allows this to be calculated in only 4 instructions at speeds
+     *     comparable to some 64-bit ALUs.
+     *
+     *  3. It isn't terrible on other platforms. Usually this will be a couple
+     *     of 32-bit ADD/ADCs.
+     */
+
+    /* First calculate all of the cross products. */
+    xxh_u64 const lo_lo = XXH_mult32to64(lhs & 0xFFFFFFFF, rhs & 0xFFFFFFFF);
+    xxh_u64 const hi_lo = XXH_mult32to64(lhs >> 32,        rhs & 0xFFFFFFFF);
+    xxh_u64 const lo_hi = XXH_mult32to64(lhs & 0xFFFFFFFF, rhs >> 32);
+    xxh_u64 const hi_hi = XXH_mult32to64(lhs >> 32,        rhs >> 32);
+
+    /* Now add the products together. These will never overflow. */
+    xxh_u64 const cross = (lo_lo >> 32) + (hi_lo & 0xFFFFFFFF) + lo_hi;
+    xxh_u64 const upper = (hi_lo >> 32) + (cross >> 32)        + hi_hi;
+    xxh_u64 const lower = (cross << 32) | (lo_lo & 0xFFFFFFFF);
+
+    XXH128_hash_t r128;
+    r128.low64  = lower;
+    r128.high64 = upper;
+    return r128;
+#endif
+}
+
+/*!
+ * @brief Calculates a 64-bit to 128-bit multiply, then XOR folds it.
+ *
+ * The reason for the separate function is to prevent passing too many structs
+ * around by value. This will hopefully inline the multiply, but we don't force it.
+ *
+ * @param lhs, rhs The 64-bit integers to multiply
+ * @return The low 64 bits of the product XOR'd by the high 64 bits.
+ * @see XXH_mult64to128()
+ */
+static xxh_u64
+XXH3_mul128_fold64(xxh_u64 lhs, xxh_u64 rhs)
+{
+    XXH128_hash_t product = XXH_mult64to128(lhs, rhs);
+    return product.low64 ^ product.high64;
+}
+
+/*! Seems to produce slightly better code on GCC for some reason. */
+XXH_FORCE_INLINE xxh_u64 XXH_xorshift64(xxh_u64 v64, int shift)
+{
+    XXH_ASSERT(0 <= shift && shift < 64);
+    return v64 ^ (v64 >> shift);
+}
+
+/*
+ * This is a fast avalanche stage,
+ * suitable when input bits are already partially mixed
+ */
+static XXH64_hash_t XXH3_avalanche(xxh_u64 h64)
+{
+    h64 = XXH_xorshift64(h64, 37);
+    h64 *= 0x165667919E3779F9ULL;
+    h64 = XXH_xorshift64(h64, 32);
+    return h64;
+}
+
+/*
+ * This is a stronger avalanche,
+ * inspired by Pelle Evensen's rrmxmx
+ * preferable when input has not been previously mixed
+ */
+static XXH64_hash_t XXH3_rrmxmx(xxh_u64 h64, xxh_u64 len)
+{
+    /* this mix is inspired by Pelle Evensen's rrmxmx */
+    h64 ^= XXH_rotl64(h64, 49) ^ XXH_rotl64(h64, 24);
+    h64 *= 0x9FB21C651E98DF25ULL;
+    h64 ^= (h64 >> 35) + len ;
+    h64 *= 0x9FB21C651E98DF25ULL;
+    return XXH_xorshift64(h64, 28);
+}
+
+
+/* ==========================================
+ * Short keys
+ * ==========================================
+ * One of the shortcomings of XXH32 and XXH64 was that their performance was
+ * sub-optimal on short lengths. It used an iterative algorithm which strongly
+ * favored lengths that were a multiple of 4 or 8.
+ *
+ * Instead of iterating over individual inputs, we use a set of single shot
+ * functions which piece together a range of lengths and operate in constant time.
+ *
+ * Additionally, the number of multiplies has been significantly reduced. This
+ * reduces latency, especially when emulating 64-bit multiplies on 32-bit.
+ *
+ * Depending on the platform, this may or may not be faster than XXH32, but it
+ * is almost guaranteed to be faster than XXH64.
+ */
+
+/*
+ * At very short lengths, there isn't enough input to fully hide secrets, or use
+ * the entire secret.
+ *
+ * There is also only a limited amount of mixing we can do before significantly
+ * impacting performance.
+ *
+ * Therefore, we use different sections of the secret and always mix two secret
+ * samples with an XOR. This should have no effect on performance on the
+ * seedless or withSeed variants because everything _should_ be constant folded
+ * by modern compilers.
+ *
+ * The XOR mixing hides individual parts of the secret and increases entropy.
+ *
+ * This adds an extra layer of strength for custom secrets.
+ */
+XXH_FORCE_INLINE XXH64_hash_t
+XXH3_len_1to3_64b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed)
+{
+    XXH_ASSERT(input != NULL);
+    XXH_ASSERT(1 <= len && len <= 3);
+    XXH_ASSERT(secret != NULL);
+    /*
+     * len = 1: combined = { input[0], 0x01, input[0], input[0] }
+     * len = 2: combined = { input[1], 0x02, input[0], input[1] }
+     * len = 3: combined = { input[2], 0x03, input[0], input[1] }
+     */
+    {   xxh_u8  const c1 = input[0];
+        xxh_u8  const c2 = input[len >> 1];
+        xxh_u8  const c3 = input[len - 1];
+        xxh_u32 const combined = ((xxh_u32)c1 << 16) | ((xxh_u32)c2  << 24)
+                               | ((xxh_u32)c3 <<  0) | ((xxh_u32)len << 8);
+        xxh_u64 const bitflip = (XXH_readLE32(secret) ^ XXH_readLE32(secret+4)) + seed;
+        xxh_u64 const keyed = (xxh_u64)combined ^ bitflip;
+        return XXH64_avalanche(keyed);
+    }
+}
+
+XXH_FORCE_INLINE XXH64_hash_t
+XXH3_len_4to8_64b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed)
+{
+    XXH_ASSERT(input != NULL);
+    XXH_ASSERT(secret != NULL);
+    XXH_ASSERT(4 <= len && len <= 8);
+    seed ^= (xxh_u64)XXH_swap32((xxh_u32)seed) << 32;
+    {   xxh_u32 const input1 = XXH_readLE32(input);
+        xxh_u32 const input2 = XXH_readLE32(input + len - 4);
+        xxh_u64 const bitflip = (XXH_readLE64(secret+8) ^ XXH_readLE64(secret+16)) - seed;
+        xxh_u64 const input64 = input2 + (((xxh_u64)input1) << 32);
+        xxh_u64 const keyed = input64 ^ bitflip;
+        return XXH3_rrmxmx(keyed, len);
+    }
+}
+
+XXH_FORCE_INLINE XXH64_hash_t
+XXH3_len_9to16_64b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed)
+{
+    XXH_ASSERT(input != NULL);
+    XXH_ASSERT(secret != NULL);
+    XXH_ASSERT(9 <= len && len <= 16);
+    {   xxh_u64 const bitflip1 = (XXH_readLE64(secret+24) ^ XXH_readLE64(secret+32)) + seed;
+        xxh_u64 const bitflip2 = (XXH_readLE64(secret+40) ^ XXH_readLE64(secret+48)) - seed;
+        xxh_u64 const input_lo = XXH_readLE64(input)           ^ bitflip1;
+        xxh_u64 const input_hi = XXH_readLE64(input + len - 8) ^ bitflip2;
+        xxh_u64 const acc = len
+                          + XXH_swap64(input_lo) + input_hi
+                          + XXH3_mul128_fold64(input_lo, input_hi);
+        return XXH3_avalanche(acc);
+    }
+}
+
+XXH_FORCE_INLINE XXH64_hash_t
+XXH3_len_0to16_64b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed)
+{
+    XXH_ASSERT(len <= 16);
+    {   if (XXH_likely(len >  8)) return XXH3_len_9to16_64b(input, len, secret, seed);
+        if (XXH_likely(len >= 4)) return XXH3_len_4to8_64b(input, len, secret, seed);
+        if (len) return XXH3_len_1to3_64b(input, len, secret, seed);
+        return XXH64_avalanche(seed ^ (XXH_readLE64(secret+56) ^ XXH_readLE64(secret+64)));
+    }
+}
+
+/*
+ * DISCLAIMER: There are known *seed-dependent* multicollisions here due to
+ * multiplication by zero, affecting hashes of lengths 17 to 240.
+ *
+ * However, they are very unlikely.
+ *
+ * Keep this in mind when using the unseeded XXH3_64bits() variant: As with all
+ * unseeded non-cryptographic hashes, it does not attempt to defend itself
+ * against specially crafted inputs, only random inputs.
+ *
+ * Compared to classic UMAC where a 1 in 2^31 chance of 4 consecutive bytes
+ * cancelling out the secret is taken an arbitrary number of times (addressed
+ * in XXH3_accumulate_512), this collision is very unlikely with random inputs
+ * and/or proper seeding:
+ *
+ * This only has a 1 in 2^63 chance of 8 consecutive bytes cancelling out, in a
+ * function that is only called up to 16 times per hash with up to 240 bytes of
+ * input.
+ *
+ * This is not too bad for a non-cryptographic hash function, especially with
+ * only 64 bit outputs.
+ *
+ * The 128-bit variant (which trades some speed for strength) is NOT affected
+ * by this, although it is always a good idea to use a proper seed if you care
+ * about strength.
+ */
+XXH_FORCE_INLINE xxh_u64 XXH3_mix16B(const xxh_u8* XXH_RESTRICT input,
+                                     const xxh_u8* XXH_RESTRICT secret, xxh_u64 seed64)
+{
+#if defined(__GNUC__) && !defined(__clang__) /* GCC, not Clang */ \
+  && defined(__i386__) && defined(__SSE2__)  /* x86 + SSE2 */ \
+  && !defined(XXH_ENABLE_AUTOVECTORIZE)      /* Define to disable like XXH32 hack */
+    /*
+     * UGLY HACK:
+     * GCC for x86 tends to autovectorize the 128-bit multiply, resulting in
+     * slower code.
+     *
+     * By forcing seed64 into a register, we disrupt the cost model and
+     * cause it to scalarize. See `XXH32_round()`
+     *
+     * FIXME: Clang's output is still _much_ faster -- On an AMD Ryzen 3600,
+     * XXH3_64bits @ len=240 runs at 4.6 GB/s with Clang 9, but 3.3 GB/s on
+     * GCC 9.2, despite both emitting scalar code.
+     *
+     * GCC generates much better scalar code than Clang for the rest of XXH3,
+     * which is why finding a more optimal codepath is an interest.
+     */
+    XXH_COMPILER_GUARD(seed64);
+#endif
+    {   xxh_u64 const input_lo = XXH_readLE64(input);
+        xxh_u64 const input_hi = XXH_readLE64(input+8);
+        return XXH3_mul128_fold64(
+            input_lo ^ (XXH_readLE64(secret)   + seed64),
+            input_hi ^ (XXH_readLE64(secret+8) - seed64)
+        );
+    }
+}
+
+/* For mid range keys, XXH3 uses a Mum-hash variant. */
+XXH_FORCE_INLINE XXH64_hash_t
+XXH3_len_17to128_64b(const xxh_u8* XXH_RESTRICT input, size_t len,
+                     const xxh_u8* XXH_RESTRICT secret, size_t secretSize,
+                     XXH64_hash_t seed)
+{
+    XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN); (void)secretSize;
+    XXH_ASSERT(16 < len && len <= 128);
+
+    {   xxh_u64 acc = len * XXH_PRIME64_1;
+        if (len > 32) {
+            if (len > 64) {
+                if (len > 96) {
+                    acc += XXH3_mix16B(input+48, secret+96, seed);
+                    acc += XXH3_mix16B(input+len-64, secret+112, seed);
+                }
+                acc += XXH3_mix16B(input+32, secret+64, seed);
+                acc += XXH3_mix16B(input+len-48, secret+80, seed);
+            }
+            acc += XXH3_mix16B(input+16, secret+32, seed);
+            acc += XXH3_mix16B(input+len-32, secret+48, seed);
+        }
+        acc += XXH3_mix16B(input+0, secret+0, seed);
+        acc += XXH3_mix16B(input+len-16, secret+16, seed);
+
+        return XXH3_avalanche(acc);
+    }
+}
+
+#define XXH3_MIDSIZE_MAX 240
+
+XXH_NO_INLINE XXH64_hash_t
+XXH3_len_129to240_64b(const xxh_u8* XXH_RESTRICT input, size_t len,
+                      const xxh_u8* XXH_RESTRICT secret, size_t secretSize,
+                      XXH64_hash_t seed)
+{
+    XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN); (void)secretSize;
+    XXH_ASSERT(128 < len && len <= XXH3_MIDSIZE_MAX);
+
+    #define XXH3_MIDSIZE_STARTOFFSET 3
+    #define XXH3_MIDSIZE_LASTOFFSET  17
+
+    {   xxh_u64 acc = len * XXH_PRIME64_1;
+        int const nbRounds = (int)len / 16;
+        int i;
+        for (i=0; i<8; i++) {
+            acc += XXH3_mix16B(input+(16*i), secret+(16*i), seed);
+        }
+        acc = XXH3_avalanche(acc);
+        XXH_ASSERT(nbRounds >= 8);
+#if defined(__clang__)                                /* Clang */ \
+    && (defined(__ARM_NEON) || defined(__ARM_NEON__)) /* NEON */ \
+    && !defined(XXH_ENABLE_AUTOVECTORIZE)             /* Define to disable */
+        /*
+         * UGLY HACK:
+         * Clang for ARMv7-A tries to vectorize this loop, similar to GCC x86.
+         * In everywhere else, it uses scalar code.
+         *
+         * For 64->128-bit multiplies, even if the NEON was 100% optimal, it
+         * would still be slower than UMAAL (see XXH_mult64to128).
+         *
+         * Unfortunately, Clang doesn't handle the long multiplies properly and
+         * converts them to the nonexistent "vmulq_u64" intrinsic, which is then
+         * scalarized into an ugly mess of VMOV.32 instructions.
+         *
+         * This mess is difficult to avoid without turning autovectorization
+         * off completely, but they are usually relatively minor and/or not
+         * worth it to fix.
+         *
+         * This loop is the easiest to fix, as unlike XXH32, this pragma
+         * _actually works_ because it is a loop vectorization instead of an
+         * SLP vectorization.
+         */
+        #pragma clang loop vectorize(disable)
+#endif
+        for (i=8 ; i < nbRounds; i++) {
+            acc += XXH3_mix16B(input+(16*i), secret+(16*(i-8)) + XXH3_MIDSIZE_STARTOFFSET, seed);
+        }
+        /* last bytes */
+        acc += XXH3_mix16B(input + len - 16, secret + XXH3_SECRET_SIZE_MIN - XXH3_MIDSIZE_LASTOFFSET, seed);
+        return XXH3_avalanche(acc);
+    }
+}
+
+
+/* =======     Long Keys     ======= */
+
+#define XXH_STRIPE_LEN 64
+#define XXH_SECRET_CONSUME_RATE 8   /* nb of secret bytes consumed at each accumulation */
+#define XXH_ACC_NB (XXH_STRIPE_LEN / sizeof(xxh_u64))
+
+#ifdef XXH_OLD_NAMES
+#  define STRIPE_LEN XXH_STRIPE_LEN
+#  define ACC_NB XXH_ACC_NB
+#endif
+
+XXH_FORCE_INLINE void XXH_writeLE64(void* dst, xxh_u64 v64)
+{
+    if (!XXH_CPU_LITTLE_ENDIAN) v64 = XXH_swap64(v64);
+    memcpy(dst, &v64, sizeof(v64));
+}
+
+/* Several intrinsic functions below are supposed to accept __int64 as argument,
+ * as documented in https://software.intel.com/sites/landingpage/IntrinsicsGuide/ .
+ * However, several environments do not define __int64 type,
+ * requiring a workaround.
+ */
+#if !defined (__VMS) \
+  && (defined (__cplusplus) \
+  || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */) )
+    typedef int64_t xxh_i64;
+#else
+    /* the following type must have a width of 64-bit */
+    typedef long long xxh_i64;
+#endif
+
+/*
+ * XXH3_accumulate_512 is the tightest loop for long inputs, and it is the most optimized.
+ *
+ * It is a hardened version of UMAC, based off of FARSH's implementation.
+ *
+ * This was chosen because it adapts quite well to 32-bit, 64-bit, and SIMD
+ * implementations, and it is ridiculously fast.
+ *
+ * We harden it by mixing the original input to the accumulators as well as the product.
+ *
+ * This means that in the (relatively likely) case of a multiply by zero, the
+ * original input is preserved.
+ *
+ * On 128-bit inputs, we swap 64-bit pairs when we add the input to improve
+ * cross-pollination, as otherwise the upper and lower halves would be
+ * essentially independent.
+ *
+ * This doesn't matter on 64-bit hashes since they all get merged together in
+ * the end, so we skip the extra step.
+ *
+ * Both XXH3_64bits and XXH3_128bits use this subroutine.
+ */
+
+#if (XXH_VECTOR == XXH_AVX512) \
+     || (defined(XXH_DISPATCH_AVX512) && XXH_DISPATCH_AVX512 != 0)
+
+#ifndef XXH_TARGET_AVX512
+# define XXH_TARGET_AVX512  /* disable attribute target */
+#endif
+
+XXH_FORCE_INLINE XXH_TARGET_AVX512 void
+XXH3_accumulate_512_avx512(void* XXH_RESTRICT acc,
+                     const void* XXH_RESTRICT input,
+                     const void* XXH_RESTRICT secret)
+{
+    XXH_ALIGN(64) __m512i* const xacc = (__m512i *) acc;
+    XXH_ASSERT((((size_t)acc) & 63) == 0);
+    XXH_STATIC_ASSERT(XXH_STRIPE_LEN == sizeof(__m512i));
+
+    {
+        /* data_vec    = input[0]; */
+        __m512i const data_vec    = _mm512_loadu_si512   (input);
+        /* key_vec     = secret[0]; */
+        __m512i const key_vec     = _mm512_loadu_si512   (secret);
+        /* data_key    = data_vec ^ key_vec; */
+        __m512i const data_key    = _mm512_xor_si512     (data_vec, key_vec);
+        /* data_key_lo = data_key >> 32; */
+        __m512i const data_key_lo = _mm512_shuffle_epi32 (data_key, (_MM_PERM_ENUM)_MM_SHUFFLE(0, 3, 0, 1));
+        /* product     = (data_key & 0xffffffff) * (data_key_lo & 0xffffffff); */
+        __m512i const product     = _mm512_mul_epu32     (data_key, data_key_lo);
+        /* xacc[0] += swap(data_vec); */
+        __m512i const data_swap = _mm512_shuffle_epi32(data_vec, (_MM_PERM_ENUM)_MM_SHUFFLE(1, 0, 3, 2));
+        __m512i const sum       = _mm512_add_epi64(*xacc, data_swap);
+        /* xacc[0] += product; */
+        *xacc = _mm512_add_epi64(product, sum);
+    }
+}
+
+/*
+ * XXH3_scrambleAcc: Scrambles the accumulators to improve mixing.
+ *
+ * Multiplication isn't perfect, as explained by Google in HighwayHash:
+ *
+ *  // Multiplication mixes/scrambles bytes 0-7 of the 64-bit result to
+ *  // varying degrees. In descending order of goodness, bytes
+ *  // 3 4 2 5 1 6 0 7 have quality 228 224 164 160 100 96 36 32.
+ *  // As expected, the upper and lower bytes are much worse.
+ *
+ * Source: https://github.com/google/highwayhash/blob/0aaf66b/highwayhash/hh_avx2.h#L291
+ *
+ * Since our algorithm uses a pseudorandom secret to add some variance into the
+ * mix, we don't need to (or want to) mix as often or as much as HighwayHash does.
+ *
+ * This isn't as tight as XXH3_accumulate, but still written in SIMD to avoid
+ * extraction.
+ *
+ * Both XXH3_64bits and XXH3_128bits use this subroutine.
+ */
+
+XXH_FORCE_INLINE XXH_TARGET_AVX512 void
+XXH3_scrambleAcc_avx512(void* XXH_RESTRICT acc, const void* XXH_RESTRICT secret)
+{
+    XXH_ASSERT((((size_t)acc) & 63) == 0);
+    XXH_STATIC_ASSERT(XXH_STRIPE_LEN == sizeof(__m512i));
+    {   XXH_ALIGN(64) __m512i* const xacc = (__m512i*) acc;
+        const __m512i prime32 = _mm512_set1_epi32((int)XXH_PRIME32_1);
+
+        /* xacc[0] ^= (xacc[0] >> 47) */
+        __m512i const acc_vec     = *xacc;
+        __m512i const shifted     = _mm512_srli_epi64    (acc_vec, 47);
+        __m512i const data_vec    = _mm512_xor_si512     (acc_vec, shifted);
+        /* xacc[0] ^= secret; */
+        __m512i const key_vec     = _mm512_loadu_si512   (secret);
+        __m512i const data_key    = _mm512_xor_si512     (data_vec, key_vec);
+
+        /* xacc[0] *= XXH_PRIME32_1; */
+        __m512i const data_key_hi = _mm512_shuffle_epi32 (data_key, (_MM_PERM_ENUM)_MM_SHUFFLE(0, 3, 0, 1));
+        __m512i const prod_lo     = _mm512_mul_epu32     (data_key, prime32);
+        __m512i const prod_hi     = _mm512_mul_epu32     (data_key_hi, prime32);
+        *xacc = _mm512_add_epi64(prod_lo, _mm512_slli_epi64(prod_hi, 32));
+    }
+}
+
+XXH_FORCE_INLINE XXH_TARGET_AVX512 void
+XXH3_initCustomSecret_avx512(void* XXH_RESTRICT customSecret, xxh_u64 seed64)
+{
+    XXH_STATIC_ASSERT((XXH_SECRET_DEFAULT_SIZE & 63) == 0);
+    XXH_STATIC_ASSERT(XXH_SEC_ALIGN == 64);
+    XXH_ASSERT(((size_t)customSecret & 63) == 0);
+    (void)(&XXH_writeLE64);
+    {   int const nbRounds = XXH_SECRET_DEFAULT_SIZE / sizeof(__m512i);
+        __m512i const seed = _mm512_mask_set1_epi64(_mm512_set1_epi64((xxh_i64)seed64), 0xAA, (xxh_i64)(0U - seed64));
+
+        XXH_ALIGN(64) const __m512i* const src  = (const __m512i*) XXH3_kSecret;
+        XXH_ALIGN(64)       __m512i* const dest = (      __m512i*) customSecret;
+        int i;
+        for (i=0; i < nbRounds; ++i) {
+            /* GCC has a bug, _mm512_stream_load_si512 accepts 'void*', not 'void const*',
+             * this will warn "discards 'const' qualifier". */
+            union {
+                XXH_ALIGN(64) const __m512i* cp;
+                XXH_ALIGN(64) void* p;
+            } remote_const_void;
+            remote_const_void.cp = src + i;
+            dest[i] = _mm512_add_epi64(_mm512_stream_load_si512(remote_const_void.p), seed);
+    }   }
+}
+
+#endif
+
+#if (XXH_VECTOR == XXH_AVX2) \
+    || (defined(XXH_DISPATCH_AVX2) && XXH_DISPATCH_AVX2 != 0)
+
+#ifndef XXH_TARGET_AVX2
+# define XXH_TARGET_AVX2  /* disable attribute target */
+#endif
+
+XXH_FORCE_INLINE XXH_TARGET_AVX2 void
+XXH3_accumulate_512_avx2( void* XXH_RESTRICT acc,
+                    const void* XXH_RESTRICT input,
+                    const void* XXH_RESTRICT secret)
+{
+    XXH_ASSERT((((size_t)acc) & 31) == 0);
+    {   XXH_ALIGN(32) __m256i* const xacc    =       (__m256i *) acc;
+        /* Unaligned. This is mainly for pointer arithmetic, and because
+         * _mm256_loadu_si256 requires  a const __m256i * pointer for some reason. */
+        const         __m256i* const xinput  = (const __m256i *) input;
+        /* Unaligned. This is mainly for pointer arithmetic, and because
+         * _mm256_loadu_si256 requires a const __m256i * pointer for some reason. */
+        const         __m256i* const xsecret = (const __m256i *) secret;
+
+        size_t i;
+        for (i=0; i < XXH_STRIPE_LEN/sizeof(__m256i); i++) {
+            /* data_vec    = xinput[i]; */
+            __m256i const data_vec    = _mm256_loadu_si256    (xinput+i);
+            /* key_vec     = xsecret[i]; */
+            __m256i const key_vec     = _mm256_loadu_si256   (xsecret+i);
+            /* data_key    = data_vec ^ key_vec; */
+            __m256i const data_key    = _mm256_xor_si256     (data_vec, key_vec);
+            /* data_key_lo = data_key >> 32; */
+            __m256i const data_key_lo = _mm256_shuffle_epi32 (data_key, _MM_SHUFFLE(0, 3, 0, 1));
+            /* product     = (data_key & 0xffffffff) * (data_key_lo & 0xffffffff); */
+            __m256i const product     = _mm256_mul_epu32     (data_key, data_key_lo);
+            /* xacc[i] += swap(data_vec); */
+            __m256i const data_swap = _mm256_shuffle_epi32(data_vec, _MM_SHUFFLE(1, 0, 3, 2));
+            __m256i const sum       = _mm256_add_epi64(xacc[i], data_swap);
+            /* xacc[i] += product; */
+            xacc[i] = _mm256_add_epi64(product, sum);
+    }   }
+}
+
+XXH_FORCE_INLINE XXH_TARGET_AVX2 void
+XXH3_scrambleAcc_avx2(void* XXH_RESTRICT acc, const void* XXH_RESTRICT secret)
+{
+    XXH_ASSERT((((size_t)acc) & 31) == 0);
+    {   XXH_ALIGN(32) __m256i* const xacc = (__m256i*) acc;
+        /* Unaligned. This is mainly for pointer arithmetic, and because
+         * _mm256_loadu_si256 requires a const __m256i * pointer for some reason. */
+        const         __m256i* const xsecret = (const __m256i *) secret;
+        const __m256i prime32 = _mm256_set1_epi32((int)XXH_PRIME32_1);
+
+        size_t i;
+        for (i=0; i < XXH_STRIPE_LEN/sizeof(__m256i); i++) {
+            /* xacc[i] ^= (xacc[i] >> 47) */
+            __m256i const acc_vec     = xacc[i];
+            __m256i const shifted     = _mm256_srli_epi64    (acc_vec, 47);
+            __m256i const data_vec    = _mm256_xor_si256     (acc_vec, shifted);
+            /* xacc[i] ^= xsecret; */
+            __m256i const key_vec     = _mm256_loadu_si256   (xsecret+i);
+            __m256i const data_key    = _mm256_xor_si256     (data_vec, key_vec);
+
+            /* xacc[i] *= XXH_PRIME32_1; */
+            __m256i const data_key_hi = _mm256_shuffle_epi32 (data_key, _MM_SHUFFLE(0, 3, 0, 1));
+            __m256i const prod_lo     = _mm256_mul_epu32     (data_key, prime32);
+            __m256i const prod_hi     = _mm256_mul_epu32     (data_key_hi, prime32);
+            xacc[i] = _mm256_add_epi64(prod_lo, _mm256_slli_epi64(prod_hi, 32));
+        }
+    }
+}
+
+XXH_FORCE_INLINE XXH_TARGET_AVX2 void XXH3_initCustomSecret_avx2(void* XXH_RESTRICT customSecret, xxh_u64 seed64)
+{
+    XXH_STATIC_ASSERT((XXH_SECRET_DEFAULT_SIZE & 31) == 0);
+    XXH_STATIC_ASSERT((XXH_SECRET_DEFAULT_SIZE / sizeof(__m256i)) == 6);
+    XXH_STATIC_ASSERT(XXH_SEC_ALIGN <= 64);
+    (void)(&XXH_writeLE64);
+    XXH_PREFETCH(customSecret);
+    {   __m256i const seed = _mm256_set_epi64x((xxh_i64)(0U - seed64), (xxh_i64)seed64, (xxh_i64)(0U - seed64), (xxh_i64)seed64);
+
+        XXH_ALIGN(64) const __m256i* const src  = (const __m256i*) XXH3_kSecret;
+        XXH_ALIGN(64)       __m256i*       dest = (      __m256i*) customSecret;
+
+#       if defined(__GNUC__) || defined(__clang__)
+        /*
+         * On GCC & Clang, marking 'dest' as modified will cause the compiler:
+         *   - do not extract the secret from sse registers in the internal loop
+         *   - use less common registers, and avoid pushing these reg into stack
+         */
+        XXH_COMPILER_GUARD(dest);
+#       endif
+
+        /* GCC -O2 need unroll loop manually */
+        dest[0] = _mm256_add_epi64(_mm256_stream_load_si256(src+0), seed);
+        dest[1] = _mm256_add_epi64(_mm256_stream_load_si256(src+1), seed);
+        dest[2] = _mm256_add_epi64(_mm256_stream_load_si256(src+2), seed);
+        dest[3] = _mm256_add_epi64(_mm256_stream_load_si256(src+3), seed);
+        dest[4] = _mm256_add_epi64(_mm256_stream_load_si256(src+4), seed);
+        dest[5] = _mm256_add_epi64(_mm256_stream_load_si256(src+5), seed);
+    }
+}
+
+#endif
+
+/* x86dispatch always generates SSE2 */
+#if (XXH_VECTOR == XXH_SSE2) || defined(XXH_X86DISPATCH)
+
+#ifndef XXH_TARGET_SSE2
+# define XXH_TARGET_SSE2  /* disable attribute target */
+#endif
+
+XXH_FORCE_INLINE XXH_TARGET_SSE2 void
+XXH3_accumulate_512_sse2( void* XXH_RESTRICT acc,
+                    const void* XXH_RESTRICT input,
+                    const void* XXH_RESTRICT secret)
+{
+    /* SSE2 is just a half-scale version of the AVX2 version. */
+    XXH_ASSERT((((size_t)acc) & 15) == 0);
+    {   XXH_ALIGN(16) __m128i* const xacc    =       (__m128i *) acc;
+        /* Unaligned. This is mainly for pointer arithmetic, and because
+         * _mm_loadu_si128 requires a const __m128i * pointer for some reason. */
+        const         __m128i* const xinput  = (const __m128i *) input;
+        /* Unaligned. This is mainly for pointer arithmetic, and because
+         * _mm_loadu_si128 requires a const __m128i * pointer for some reason. */
+        const         __m128i* const xsecret = (const __m128i *) secret;
+
+        size_t i;
+        for (i=0; i < XXH_STRIPE_LEN/sizeof(__m128i); i++) {
+            /* data_vec    = xinput[i]; */
+            __m128i const data_vec    = _mm_loadu_si128   (xinput+i);
+            /* key_vec     = xsecret[i]; */
+            __m128i const key_vec     = _mm_loadu_si128   (xsecret+i);
+            /* data_key    = data_vec ^ key_vec; */
+            __m128i const data_key    = _mm_xor_si128     (data_vec, key_vec);
+            /* data_key_lo = data_key >> 32; */
+            __m128i const data_key_lo = _mm_shuffle_epi32 (data_key, _MM_SHUFFLE(0, 3, 0, 1));
+            /* product     = (data_key & 0xffffffff) * (data_key_lo & 0xffffffff); */
+            __m128i const product     = _mm_mul_epu32     (data_key, data_key_lo);
+            /* xacc[i] += swap(data_vec); */
+            __m128i const data_swap = _mm_shuffle_epi32(data_vec, _MM_SHUFFLE(1,0,3,2));
+            __m128i const sum       = _mm_add_epi64(xacc[i], data_swap);
+            /* xacc[i] += product; */
+            xacc[i] = _mm_add_epi64(product, sum);
+    }   }
+}
+
+XXH_FORCE_INLINE XXH_TARGET_SSE2 void
+XXH3_scrambleAcc_sse2(void* XXH_RESTRICT acc, const void* XXH_RESTRICT secret)
+{
+    XXH_ASSERT((((size_t)acc) & 15) == 0);
+    {   XXH_ALIGN(16) __m128i* const xacc = (__m128i*) acc;
+        /* Unaligned. This is mainly for pointer arithmetic, and because
+         * _mm_loadu_si128 requires a const __m128i * pointer for some reason. */
+        const         __m128i* const xsecret = (const __m128i *) secret;
+        const __m128i prime32 = _mm_set1_epi32((int)XXH_PRIME32_1);
+
+        size_t i;
+        for (i=0; i < XXH_STRIPE_LEN/sizeof(__m128i); i++) {
+            /* xacc[i] ^= (xacc[i] >> 47) */
+            __m128i const acc_vec     = xacc[i];
+            __m128i const shifted     = _mm_srli_epi64    (acc_vec, 47);
+            __m128i const data_vec    = _mm_xor_si128     (acc_vec, shifted);
+            /* xacc[i] ^= xsecret[i]; */
+            __m128i const key_vec     = _mm_loadu_si128   (xsecret+i);
+            __m128i const data_key    = _mm_xor_si128     (data_vec, key_vec);
+
+            /* xacc[i] *= XXH_PRIME32_1; */
+            __m128i const data_key_hi = _mm_shuffle_epi32 (data_key, _MM_SHUFFLE(0, 3, 0, 1));
+            __m128i const prod_lo     = _mm_mul_epu32     (data_key, prime32);
+            __m128i const prod_hi     = _mm_mul_epu32     (data_key_hi, prime32);
+            xacc[i] = _mm_add_epi64(prod_lo, _mm_slli_epi64(prod_hi, 32));
+        }
+    }
+}
+
+XXH_FORCE_INLINE XXH_TARGET_SSE2 void XXH3_initCustomSecret_sse2(void* XXH_RESTRICT customSecret, xxh_u64 seed64)
+{
+    XXH_STATIC_ASSERT((XXH_SECRET_DEFAULT_SIZE & 15) == 0);
+    (void)(&XXH_writeLE64);
+    {   int const nbRounds = XXH_SECRET_DEFAULT_SIZE / sizeof(__m128i);
+
+#       if defined(_MSC_VER) && defined(_M_IX86) && _MSC_VER < 1900
+        // MSVC 32bit mode does not support _mm_set_epi64x before 2015
+        XXH_ALIGN(16) const xxh_i64 seed64x2[2] = { (xxh_i64)seed64, (xxh_i64)(0U - seed64) };
+        __m128i const seed = _mm_load_si128((__m128i const*)seed64x2);
+#       else
+        __m128i const seed = _mm_set_epi64x((xxh_i64)(0U - seed64), (xxh_i64)seed64);
+#       endif
+        int i;
+
+        XXH_ALIGN(64)        const float* const src  = (float const*) XXH3_kSecret;
+        XXH_ALIGN(XXH_SEC_ALIGN) __m128i*       dest = (__m128i*) customSecret;
+#       if defined(__GNUC__) || defined(__clang__)
+        /*
+         * On GCC & Clang, marking 'dest' as modified will cause the compiler:
+         *   - do not extract the secret from sse registers in the internal loop
+         *   - use less common registers, and avoid pushing these reg into stack
+         */
+        XXH_COMPILER_GUARD(dest);
+#       endif
+
+        for (i=0; i < nbRounds; ++i) {
+            dest[i] = _mm_add_epi64(_mm_castps_si128(_mm_load_ps(src+i*4)), seed);
+    }   }
+}
+
+#endif
+
+#if (XXH_VECTOR == XXH_NEON)
+
+XXH_FORCE_INLINE void
+XXH3_accumulate_512_neon( void* XXH_RESTRICT acc,
+                    const void* XXH_RESTRICT input,
+                    const void* XXH_RESTRICT secret)
+{
+    XXH_ASSERT((((size_t)acc) & 15) == 0);
+    {
+        XXH_ALIGN(16) uint64x2_t* const xacc = (uint64x2_t *) acc;
+        /* We don't use a uint32x4_t pointer because it causes bus errors on ARMv7. */
+        uint8_t const* const xinput = (const uint8_t *) input;
+        uint8_t const* const xsecret  = (const uint8_t *) secret;
+
+        size_t i;
+        for (i=0; i < XXH_STRIPE_LEN / sizeof(uint64x2_t); i++) {
+            /* data_vec = xinput[i]; */
+            uint8x16_t data_vec    = vld1q_u8(xinput  + (i * 16));
+            /* key_vec  = xsecret[i];  */
+            uint8x16_t key_vec     = vld1q_u8(xsecret + (i * 16));
+            uint64x2_t data_key;
+            uint32x2_t data_key_lo, data_key_hi;
+            /* xacc[i] += swap(data_vec); */
+            uint64x2_t const data64  = vreinterpretq_u64_u8(data_vec);
+            uint64x2_t const swapped = vextq_u64(data64, data64, 1);
+            xacc[i] = vaddq_u64 (xacc[i], swapped);
+            /* data_key = data_vec ^ key_vec; */
+            data_key = vreinterpretq_u64_u8(veorq_u8(data_vec, key_vec));
+            /* data_key_lo = (uint32x2_t) (data_key & 0xFFFFFFFF);
+             * data_key_hi = (uint32x2_t) (data_key >> 32);
+             * data_key = UNDEFINED; */
+            XXH_SPLIT_IN_PLACE(data_key, data_key_lo, data_key_hi);
+            /* xacc[i] += (uint64x2_t) data_key_lo * (uint64x2_t) data_key_hi; */
+            xacc[i] = vmlal_u32 (xacc[i], data_key_lo, data_key_hi);
+
+        }
+    }
+}
+
+XXH_FORCE_INLINE void
+XXH3_scrambleAcc_neon(void* XXH_RESTRICT acc, const void* XXH_RESTRICT secret)
+{
+    XXH_ASSERT((((size_t)acc) & 15) == 0);
+
+    {   uint64x2_t* xacc       = (uint64x2_t*) acc;
+        uint8_t const* xsecret = (uint8_t const*) secret;
+        uint32x2_t prime       = vdup_n_u32 (XXH_PRIME32_1);
+
+        size_t i;
+        for (i=0; i < XXH_STRIPE_LEN/sizeof(uint64x2_t); i++) {
+            /* xacc[i] ^= (xacc[i] >> 47); */
+            uint64x2_t acc_vec  = xacc[i];
+            uint64x2_t shifted  = vshrq_n_u64 (acc_vec, 47);
+            uint64x2_t data_vec = veorq_u64   (acc_vec, shifted);
+
+            /* xacc[i] ^= xsecret[i]; */
+            uint8x16_t key_vec  = vld1q_u8(xsecret + (i * 16));
+            uint64x2_t data_key = veorq_u64(data_vec, vreinterpretq_u64_u8(key_vec));
+
+            /* xacc[i] *= XXH_PRIME32_1 */
+            uint32x2_t data_key_lo, data_key_hi;
+            /* data_key_lo = (uint32x2_t) (xacc[i] & 0xFFFFFFFF);
+             * data_key_hi = (uint32x2_t) (xacc[i] >> 32);
+             * xacc[i] = UNDEFINED; */
+            XXH_SPLIT_IN_PLACE(data_key, data_key_lo, data_key_hi);
+            {   /*
+                 * prod_hi = (data_key >> 32) * XXH_PRIME32_1;
+                 *
+                 * Avoid vmul_u32 + vshll_n_u32 since Clang 6 and 7 will
+                 * incorrectly "optimize" this:
+                 *   tmp     = vmul_u32(vmovn_u64(a), vmovn_u64(b));
+                 *   shifted = vshll_n_u32(tmp, 32);
+                 * to this:
+                 *   tmp     = "vmulq_u64"(a, b); // no such thing!
+                 *   shifted = vshlq_n_u64(tmp, 32);
+                 *
+                 * However, unlike SSE, Clang lacks a 64-bit multiply routine
+                 * for NEON, and it scalarizes two 64-bit multiplies instead.
+                 *
+                 * vmull_u32 has the same timing as vmul_u32, and it avoids
+                 * this bug completely.
+                 * See https://bugs.llvm.org/show_bug.cgi?id=39967
+                 */
+                uint64x2_t prod_hi = vmull_u32 (data_key_hi, prime);
+                /* xacc[i] = prod_hi << 32; */
+                xacc[i] = vshlq_n_u64(prod_hi, 32);
+                /* xacc[i] += (prod_hi & 0xFFFFFFFF) * XXH_PRIME32_1; */
+                xacc[i] = vmlal_u32(xacc[i], data_key_lo, prime);
+            }
+    }   }
+}
+
+#endif
+
+#if (XXH_VECTOR == XXH_VSX)
+
+XXH_FORCE_INLINE void
+XXH3_accumulate_512_vsx(  void* XXH_RESTRICT acc,
+                    const void* XXH_RESTRICT input,
+                    const void* XXH_RESTRICT secret)
+{
+          xxh_u64x2* const xacc     =       (xxh_u64x2*) acc;    /* presumed aligned */
+    xxh_u64x2 const* const xinput   = (xxh_u64x2 const*) input;   /* no alignment restriction */
+    xxh_u64x2 const* const xsecret  = (xxh_u64x2 const*) secret;    /* no alignment restriction */
+    xxh_u64x2 const v32 = { 32, 32 };
+    size_t i;
+    for (i = 0; i < XXH_STRIPE_LEN / sizeof(xxh_u64x2); i++) {
+        /* data_vec = xinput[i]; */
+        xxh_u64x2 const data_vec = XXH_vec_loadu(xinput + i);
+        /* key_vec = xsecret[i]; */
+        xxh_u64x2 const key_vec  = XXH_vec_loadu(xsecret + i);
+        xxh_u64x2 const data_key = data_vec ^ key_vec;
+        /* shuffled = (data_key << 32) | (data_key >> 32); */
+        xxh_u32x4 const shuffled = (xxh_u32x4)vec_rl(data_key, v32);
+        /* product = ((xxh_u64x2)data_key & 0xFFFFFFFF) * ((xxh_u64x2)shuffled & 0xFFFFFFFF); */
+        xxh_u64x2 const product  = XXH_vec_mulo((xxh_u32x4)data_key, shuffled);
+        xacc[i] += product;
+
+        /* swap high and low halves */
+#ifdef __s390x__
+        xacc[i] += vec_permi(data_vec, data_vec, 2);
+#else
+        xacc[i] += vec_xxpermdi(data_vec, data_vec, 2);
+#endif
+    }
+}
+
+XXH_FORCE_INLINE void
+XXH3_scrambleAcc_vsx(void* XXH_RESTRICT acc, const void* XXH_RESTRICT secret)
+{
+    XXH_ASSERT((((size_t)acc) & 15) == 0);
+
+    {         xxh_u64x2* const xacc    =       (xxh_u64x2*) acc;
+        const xxh_u64x2* const xsecret = (const xxh_u64x2*) secret;
+        /* constants */
+        xxh_u64x2 const v32  = { 32, 32 };
+        xxh_u64x2 const v47 = { 47, 47 };
+        xxh_u32x4 const prime = { XXH_PRIME32_1, XXH_PRIME32_1, XXH_PRIME32_1, XXH_PRIME32_1 };
+        size_t i;
+        for (i = 0; i < XXH_STRIPE_LEN / sizeof(xxh_u64x2); i++) {
+            /* xacc[i] ^= (xacc[i] >> 47); */
+            xxh_u64x2 const acc_vec  = xacc[i];
+            xxh_u64x2 const data_vec = acc_vec ^ (acc_vec >> v47);
+
+            /* xacc[i] ^= xsecret[i]; */
+            xxh_u64x2 const key_vec  = XXH_vec_loadu(xsecret + i);
+            xxh_u64x2 const data_key = data_vec ^ key_vec;
+
+            /* xacc[i] *= XXH_PRIME32_1 */
+            /* prod_lo = ((xxh_u64x2)data_key & 0xFFFFFFFF) * ((xxh_u64x2)prime & 0xFFFFFFFF);  */
+            xxh_u64x2 const prod_even  = XXH_vec_mule((xxh_u32x4)data_key, prime);
+            /* prod_hi = ((xxh_u64x2)data_key >> 32) * ((xxh_u64x2)prime >> 32);  */
+            xxh_u64x2 const prod_odd  = XXH_vec_mulo((xxh_u32x4)data_key, prime);
+            xacc[i] = prod_odd + (prod_even << v32);
+    }   }
+}
+
+#endif
+
+/* scalar variants - universal */
+
+XXH_FORCE_INLINE void
+XXH3_accumulate_512_scalar(void* XXH_RESTRICT acc,
+                     const void* XXH_RESTRICT input,
+                     const void* XXH_RESTRICT secret)
+{
+    XXH_ALIGN(XXH_ACC_ALIGN) xxh_u64* const xacc = (xxh_u64*) acc; /* presumed aligned */
+    const xxh_u8* const xinput  = (const xxh_u8*) input;  /* no alignment restriction */
+    const xxh_u8* const xsecret = (const xxh_u8*) secret;   /* no alignment restriction */
+    size_t i;
+    XXH_ASSERT(((size_t)acc & (XXH_ACC_ALIGN-1)) == 0);
+    for (i=0; i < XXH_ACC_NB; i++) {
+        xxh_u64 const data_val = XXH_readLE64(xinput + 8*i);
+        xxh_u64 const data_key = data_val ^ XXH_readLE64(xsecret + i*8);
+        xacc[i ^ 1] += data_val; /* swap adjacent lanes */
+        xacc[i] += XXH_mult32to64(data_key & 0xFFFFFFFF, data_key >> 32);
+    }
+}
+
+XXH_FORCE_INLINE void
+XXH3_scrambleAcc_scalar(void* XXH_RESTRICT acc, const void* XXH_RESTRICT secret)
+{
+    XXH_ALIGN(XXH_ACC_ALIGN) xxh_u64* const xacc = (xxh_u64*) acc;   /* presumed aligned */
+    const xxh_u8* const xsecret = (const xxh_u8*) secret;   /* no alignment restriction */
+    size_t i;
+    XXH_ASSERT((((size_t)acc) & (XXH_ACC_ALIGN-1)) == 0);
+    for (i=0; i < XXH_ACC_NB; i++) {
+        xxh_u64 const key64 = XXH_readLE64(xsecret + 8*i);
+        xxh_u64 acc64 = xacc[i];
+        acc64 = XXH_xorshift64(acc64, 47);
+        acc64 ^= key64;
+        acc64 *= XXH_PRIME32_1;
+        xacc[i] = acc64;
+    }
+}
+
+XXH_FORCE_INLINE void
+XXH3_initCustomSecret_scalar(void* XXH_RESTRICT customSecret, xxh_u64 seed64)
+{
+    /*
+     * We need a separate pointer for the hack below,
+     * which requires a non-const pointer.
+     * Any decent compiler will optimize this out otherwise.
+     */
+    const xxh_u8* kSecretPtr = XXH3_kSecret;
+    XXH_STATIC_ASSERT((XXH_SECRET_DEFAULT_SIZE & 15) == 0);
+
+#if defined(__clang__) && defined(__aarch64__)
+    /*
+     * UGLY HACK:
+     * Clang generates a bunch of MOV/MOVK pairs for aarch64, and they are
+     * placed sequentially, in order, at the top of the unrolled loop.
+     *
+     * While MOVK is great for generating constants (2 cycles for a 64-bit
+     * constant compared to 4 cycles for LDR), long MOVK chains stall the
+     * integer pipelines:
+     *   I   L   S
+     * MOVK
+     * MOVK
+     * MOVK
+     * MOVK
+     * ADD
+     * SUB      STR
+     *          STR
+     * By forcing loads from memory (as the asm line causes Clang to assume
+     * that XXH3_kSecretPtr has been changed), the pipelines are used more
+     * efficiently:
+     *   I   L   S
+     *      LDR
+     *  ADD LDR
+     *  SUB     STR
+     *          STR
+     * XXH3_64bits_withSeed, len == 256, Snapdragon 835
+     *   without hack: 2654.4 MB/s
+     *   with hack:    3202.9 MB/s
+     */
+    XXH_COMPILER_GUARD(kSecretPtr);
+#endif
+    /*
+     * Note: in debug mode, this overrides the asm optimization
+     * and Clang will emit MOVK chains again.
+     */
+    XXH_ASSERT(kSecretPtr == XXH3_kSecret);
+
+    {   int const nbRounds = XXH_SECRET_DEFAULT_SIZE / 16;
+        int i;
+        for (i=0; i < nbRounds; i++) {
+            /*
+             * The asm hack causes Clang to assume that kSecretPtr aliases with
+             * customSecret, and on aarch64, this prevented LDP from merging two
+             * loads together for free. Putting the loads together before the stores
+             * properly generates LDP.
+             */
+            xxh_u64 lo = XXH_readLE64(kSecretPtr + 16*i)     + seed64;
+            xxh_u64 hi = XXH_readLE64(kSecretPtr + 16*i + 8) - seed64;
+            XXH_writeLE64((xxh_u8*)customSecret + 16*i,     lo);
+            XXH_writeLE64((xxh_u8*)customSecret + 16*i + 8, hi);
+    }   }
+}
+
+
+typedef void (*XXH3_f_accumulate_512)(void* XXH_RESTRICT, const void*, const void*);
+typedef void (*XXH3_f_scrambleAcc)(void* XXH_RESTRICT, const void*);
+typedef void (*XXH3_f_initCustomSecret)(void* XXH_RESTRICT, xxh_u64);
+
+
+#if (XXH_VECTOR == XXH_AVX512)
+
+#define XXH3_accumulate_512 XXH3_accumulate_512_avx512
+#define XXH3_scrambleAcc    XXH3_scrambleAcc_avx512
+#define XXH3_initCustomSecret XXH3_initCustomSecret_avx512
+
+#elif (XXH_VECTOR == XXH_AVX2)
+
+#define XXH3_accumulate_512 XXH3_accumulate_512_avx2
+#define XXH3_scrambleAcc    XXH3_scrambleAcc_avx2
+#define XXH3_initCustomSecret XXH3_initCustomSecret_avx2
+
+#elif (XXH_VECTOR == XXH_SSE2)
+
+#define XXH3_accumulate_512 XXH3_accumulate_512_sse2
+#define XXH3_scrambleAcc    XXH3_scrambleAcc_sse2
+#define XXH3_initCustomSecret XXH3_initCustomSecret_sse2
+
+#elif (XXH_VECTOR == XXH_NEON)
+
+#define XXH3_accumulate_512 XXH3_accumulate_512_neon
+#define XXH3_scrambleAcc    XXH3_scrambleAcc_neon
+#define XXH3_initCustomSecret XXH3_initCustomSecret_scalar
+
+#elif (XXH_VECTOR == XXH_VSX)
+
+#define XXH3_accumulate_512 XXH3_accumulate_512_vsx
+#define XXH3_scrambleAcc    XXH3_scrambleAcc_vsx
+#define XXH3_initCustomSecret XXH3_initCustomSecret_scalar
+
+#else /* scalar */
+
+#define XXH3_accumulate_512 XXH3_accumulate_512_scalar
+#define XXH3_scrambleAcc    XXH3_scrambleAcc_scalar
+#define XXH3_initCustomSecret XXH3_initCustomSecret_scalar
+
+#endif
+
+
+
+#ifndef XXH_PREFETCH_DIST
+#  ifdef __clang__
+#    define XXH_PREFETCH_DIST 320
+#  else
+#    if (XXH_VECTOR == XXH_AVX512)
+#      define XXH_PREFETCH_DIST 512
+#    else
+#      define XXH_PREFETCH_DIST 384
+#    endif
+#  endif  /* __clang__ */
+#endif  /* XXH_PREFETCH_DIST */
+
+/*
+ * XXH3_accumulate()
+ * Loops over XXH3_accumulate_512().
+ * Assumption: nbStripes will not overflow the secret size
+ */
+XXH_FORCE_INLINE void
+XXH3_accumulate(     xxh_u64* XXH_RESTRICT acc,
+                const xxh_u8* XXH_RESTRICT input,
+                const xxh_u8* XXH_RESTRICT secret,
+                      size_t nbStripes,
+                      XXH3_f_accumulate_512 f_acc512)
+{
+    size_t n;
+    for (n = 0; n < nbStripes; n++ ) {
+        const xxh_u8* const in = input + n*XXH_STRIPE_LEN;
+        XXH_PREFETCH(in + XXH_PREFETCH_DIST);
+        f_acc512(acc,
+                 in,
+                 secret + n*XXH_SECRET_CONSUME_RATE);
+    }
+}
+
+XXH_FORCE_INLINE void
+XXH3_hashLong_internal_loop(xxh_u64* XXH_RESTRICT acc,
+                      const xxh_u8* XXH_RESTRICT input, size_t len,
+                      const xxh_u8* XXH_RESTRICT secret, size_t secretSize,
+                            XXH3_f_accumulate_512 f_acc512,
+                            XXH3_f_scrambleAcc f_scramble)
+{
+    size_t const nbStripesPerBlock = (secretSize - XXH_STRIPE_LEN) / XXH_SECRET_CONSUME_RATE;
+    size_t const block_len = XXH_STRIPE_LEN * nbStripesPerBlock;
+    size_t const nb_blocks = (len - 1) / block_len;
+
+    size_t n;
+
+    XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN);
+
+    for (n = 0; n < nb_blocks; n++) {
+        XXH3_accumulate(acc, input + n*block_len, secret, nbStripesPerBlock, f_acc512);
+        f_scramble(acc, secret + secretSize - XXH_STRIPE_LEN);
+    }
+
+    /* last partial block */
+    XXH_ASSERT(len > XXH_STRIPE_LEN);
+    {   size_t const nbStripes = ((len - 1) - (block_len * nb_blocks)) / XXH_STRIPE_LEN;
+        XXH_ASSERT(nbStripes <= (secretSize / XXH_SECRET_CONSUME_RATE));
+        XXH3_accumulate(acc, input + nb_blocks*block_len, secret, nbStripes, f_acc512);
+
+        /* last stripe */
+        {   const xxh_u8* const p = input + len - XXH_STRIPE_LEN;
+#define XXH_SECRET_LASTACC_START 7  /* not aligned on 8, last secret is different from acc & scrambler */
+            f_acc512(acc, p, secret + secretSize - XXH_STRIPE_LEN - XXH_SECRET_LASTACC_START);
+    }   }
+}
+
+XXH_FORCE_INLINE xxh_u64
+XXH3_mix2Accs(const xxh_u64* XXH_RESTRICT acc, const xxh_u8* XXH_RESTRICT secret)
+{
+    return XXH3_mul128_fold64(
+               acc[0] ^ XXH_readLE64(secret),
+               acc[1] ^ XXH_readLE64(secret+8) );
+}
+
+static XXH64_hash_t
+XXH3_mergeAccs(const xxh_u64* XXH_RESTRICT acc, const xxh_u8* XXH_RESTRICT secret, xxh_u64 start)
+{
+    xxh_u64 result64 = start;
+    size_t i = 0;
+
+    for (i = 0; i < 4; i++) {
+        result64 += XXH3_mix2Accs(acc+2*i, secret + 16*i);
+#if defined(__clang__)                                /* Clang */ \
+    && (defined(__arm__) || defined(__thumb__))       /* ARMv7 */ \
+    && (defined(__ARM_NEON) || defined(__ARM_NEON__)) /* NEON */  \
+    && !defined(XXH_ENABLE_AUTOVECTORIZE)             /* Define to disable */
+        /*
+         * UGLY HACK:
+         * Prevent autovectorization on Clang ARMv7-a. Exact same problem as
+         * the one in XXH3_len_129to240_64b. Speeds up shorter keys > 240b.
+         * XXH3_64bits, len == 256, Snapdragon 835:
+         *   without hack: 2063.7 MB/s
+         *   with hack:    2560.7 MB/s
+         */
+        XXH_COMPILER_GUARD(result64);
+#endif
+    }
+
+    return XXH3_avalanche(result64);
+}
+
+#define XXH3_INIT_ACC { XXH_PRIME32_3, XXH_PRIME64_1, XXH_PRIME64_2, XXH_PRIME64_3, \
+                        XXH_PRIME64_4, XXH_PRIME32_2, XXH_PRIME64_5, XXH_PRIME32_1 }
+
+XXH_FORCE_INLINE XXH64_hash_t
+XXH3_hashLong_64b_internal(const void* XXH_RESTRICT input, size_t len,
+                           const void* XXH_RESTRICT secret, size_t secretSize,
+                           XXH3_f_accumulate_512 f_acc512,
+                           XXH3_f_scrambleAcc f_scramble)
+{
+    XXH_ALIGN(XXH_ACC_ALIGN) xxh_u64 acc[XXH_ACC_NB] = XXH3_INIT_ACC;
+
+    XXH3_hashLong_internal_loop(acc, (const xxh_u8*)input, len, (const xxh_u8*)secret, secretSize, f_acc512, f_scramble);
+
+    /* converge into final hash */
+    XXH_STATIC_ASSERT(sizeof(acc) == 64);
+    /* do not align on 8, so that the secret is different from the accumulator */
+#define XXH_SECRET_MERGEACCS_START 11
+    XXH_ASSERT(secretSize >= sizeof(acc) + XXH_SECRET_MERGEACCS_START);
+    return XXH3_mergeAccs(acc, (const xxh_u8*)secret + XXH_SECRET_MERGEACCS_START, (xxh_u64)len * XXH_PRIME64_1);
+}
+
+/*
+ * It's important for performance that XXH3_hashLong is not inlined.
+ */
+XXH_NO_INLINE XXH64_hash_t
+XXH3_hashLong_64b_withSecret(const void* XXH_RESTRICT input, size_t len,
+                             XXH64_hash_t seed64, const xxh_u8* XXH_RESTRICT secret, size_t secretLen)
+{
+    (void)seed64;
+    return XXH3_hashLong_64b_internal(input, len, secret, secretLen, XXH3_accumulate_512, XXH3_scrambleAcc);
+}
+
+/*
+ * It's important for performance that XXH3_hashLong is not inlined.
+ * Since the function is not inlined, the compiler may not be able to understand that,
+ * in some scenarios, its `secret` argument is actually a compile time constant.
+ * This variant enforces that the compiler can detect that,
+ * and uses this opportunity to streamline the generated code for better performance.
+ */
+XXH_NO_INLINE XXH64_hash_t
+XXH3_hashLong_64b_default(const void* XXH_RESTRICT input, size_t len,
+                          XXH64_hash_t seed64, const xxh_u8* XXH_RESTRICT secret, size_t secretLen)
+{
+    (void)seed64; (void)secret; (void)secretLen;
+    return XXH3_hashLong_64b_internal(input, len, XXH3_kSecret, sizeof(XXH3_kSecret), XXH3_accumulate_512, XXH3_scrambleAcc);
+}
+
+/*
+ * XXH3_hashLong_64b_withSeed():
+ * Generate a custom key based on alteration of default XXH3_kSecret with the seed,
+ * and then use this key for long mode hashing.
+ *
+ * This operation is decently fast but nonetheless costs a little bit of time.
+ * Try to avoid it whenever possible (typically when seed==0).
+ *
+ * It's important for performance that XXH3_hashLong is not inlined. Not sure
+ * why (uop cache maybe?), but the difference is large and easily measurable.
+ */
+XXH_FORCE_INLINE XXH64_hash_t
+XXH3_hashLong_64b_withSeed_internal(const void* input, size_t len,
+                                    XXH64_hash_t seed,
+                                    XXH3_f_accumulate_512 f_acc512,
+                                    XXH3_f_scrambleAcc f_scramble,
+                                    XXH3_f_initCustomSecret f_initSec)
+{
+    if (seed == 0)
+        return XXH3_hashLong_64b_internal(input, len,
+                                          XXH3_kSecret, sizeof(XXH3_kSecret),
+                                          f_acc512, f_scramble);
+    {   XXH_ALIGN(XXH_SEC_ALIGN) xxh_u8 secret[XXH_SECRET_DEFAULT_SIZE];
+        f_initSec(secret, seed);
+        return XXH3_hashLong_64b_internal(input, len, secret, sizeof(secret),
+                                          f_acc512, f_scramble);
+    }
+}
+
+/*
+ * It's important for performance that XXH3_hashLong is not inlined.
+ */
+XXH_NO_INLINE XXH64_hash_t
+XXH3_hashLong_64b_withSeed(const void* input, size_t len,
+                           XXH64_hash_t seed, const xxh_u8* secret, size_t secretLen)
+{
+    (void)secret; (void)secretLen;
+    return XXH3_hashLong_64b_withSeed_internal(input, len, seed,
+                XXH3_accumulate_512, XXH3_scrambleAcc, XXH3_initCustomSecret);
+}
+
+
+typedef XXH64_hash_t (*XXH3_hashLong64_f)(const void* XXH_RESTRICT, size_t,
+                                          XXH64_hash_t, const xxh_u8* XXH_RESTRICT, size_t);
+
+XXH_FORCE_INLINE XXH64_hash_t
+XXH3_64bits_internal(const void* XXH_RESTRICT input, size_t len,
+                     XXH64_hash_t seed64, const void* XXH_RESTRICT secret, size_t secretLen,
+                     XXH3_hashLong64_f f_hashLong)
+{
+    XXH_ASSERT(secretLen >= XXH3_SECRET_SIZE_MIN);
+    /*
+     * If an action is to be taken if `secretLen` condition is not respected,
+     * it should be done here.
+     * For now, it's a contract pre-condition.
+     * Adding a check and a branch here would cost performance at every hash.
+     * Also, note that function signature doesn't offer room to return an error.
+     */
+    if (len <= 16)
+        return XXH3_len_0to16_64b((const xxh_u8*)input, len, (const xxh_u8*)secret, seed64);
+    if (len <= 128)
+        return XXH3_len_17to128_64b((const xxh_u8*)input, len, (const xxh_u8*)secret, secretLen, seed64);
+    if (len <= XXH3_MIDSIZE_MAX)
+        return XXH3_len_129to240_64b((const xxh_u8*)input, len, (const xxh_u8*)secret, secretLen, seed64);
+    return f_hashLong(input, len, seed64, (const xxh_u8*)secret, secretLen);
+}
+
+
+/* ===   Public entry point   === */
+
+/*! @ingroup xxh3_family */
+XXH_PUBLIC_API XXH64_hash_t XXH3_64bits(const void* input, size_t len)
+{
+    return XXH3_64bits_internal(input, len, 0, XXH3_kSecret, sizeof(XXH3_kSecret), XXH3_hashLong_64b_default);
+}
+
+/*! @ingroup xxh3_family */
+XXH_PUBLIC_API XXH64_hash_t
+XXH3_64bits_withSecret(const void* input, size_t len, const void* secret, size_t secretSize)
+{
+    return XXH3_64bits_internal(input, len, 0, secret, secretSize, XXH3_hashLong_64b_withSecret);
+}
+
+/*! @ingroup xxh3_family */
+XXH_PUBLIC_API XXH64_hash_t
+XXH3_64bits_withSeed(const void* input, size_t len, XXH64_hash_t seed)
+{
+    return XXH3_64bits_internal(input, len, seed, XXH3_kSecret, sizeof(XXH3_kSecret), XXH3_hashLong_64b_withSeed);
+}
+
+
+/* ===   XXH3 streaming   === */
+
+/*
+ * Malloc's a pointer that is always aligned to align.
+ *
+ * This must be freed with `XXH_alignedFree()`.
+ *
+ * malloc typically guarantees 16 byte alignment on 64-bit systems and 8 byte
+ * alignment on 32-bit. This isn't enough for the 32 byte aligned loads in AVX2
+ * or on 32-bit, the 16 byte aligned loads in SSE2 and NEON.
+ *
+ * This underalignment previously caused a rather obvious crash which went
+ * completely unnoticed due to XXH3_createState() not actually being tested.
+ * Credit to RedSpah for noticing this bug.
+ *
+ * The alignment is done manually: Functions like posix_memalign or _mm_malloc
+ * are avoided: To maintain portability, we would have to write a fallback
+ * like this anyways, and besides, testing for the existence of library
+ * functions without relying on external build tools is impossible.
+ *
+ * The method is simple: Overallocate, manually align, and store the offset
+ * to the original behind the returned pointer.
+ *
+ * Align must be a power of 2 and 8 <= align <= 128.
+ */
+static void* XXH_alignedMalloc(size_t s, size_t align)
+{
+    XXH_ASSERT(align <= 128 && align >= 8); /* range check */
+    XXH_ASSERT((align & (align-1)) == 0);   /* power of 2 */
+    XXH_ASSERT(s != 0 && s < (s + align));  /* empty/overflow */
+    {   /* Overallocate to make room for manual realignment and an offset byte */
+        xxh_u8* base = (xxh_u8*)XXH_malloc(s + align);
+        if (base != NULL) {
+            /*
+             * Get the offset needed to align this pointer.
+             *
+             * Even if the returned pointer is aligned, there will always be
+             * at least one byte to store the offset to the original pointer.
+             */
+            size_t offset = align - ((size_t)base & (align - 1)); /* base % align */
+            /* Add the offset for the now-aligned pointer */
+            xxh_u8* ptr = base + offset;
+
+            XXH_ASSERT((size_t)ptr % align == 0);
+
+            /* Store the offset immediately before the returned pointer. */
+            ptr[-1] = (xxh_u8)offset;
+            return ptr;
+        }
+        return NULL;
+    }
+}
+/*
+ * Frees an aligned pointer allocated by XXH_alignedMalloc(). Don't pass
+ * normal malloc'd pointers, XXH_alignedMalloc has a specific data layout.
+ */
+static void XXH_alignedFree(void* p)
+{
+    if (p != NULL) {
+        xxh_u8* ptr = (xxh_u8*)p;
+        /* Get the offset byte we added in XXH_malloc. */
+        xxh_u8 offset = ptr[-1];
+        /* Free the original malloc'd pointer */
+        xxh_u8* base = ptr - offset;
+        XXH_free(base);
+    }
+}
+/*! @ingroup xxh3_family */
+XXH_PUBLIC_API XXH3_state_t* XXH3_createState(void)
+{
+    XXH3_state_t* const state = (XXH3_state_t*)XXH_alignedMalloc(sizeof(XXH3_state_t), 64);
+    if (state==NULL) return NULL;
+    XXH3_INITSTATE(state);
+    return state;
+}
+
+/*! @ingroup xxh3_family */
+XXH_PUBLIC_API XXH_errorcode XXH3_freeState(XXH3_state_t* statePtr)
+{
+    XXH_alignedFree(statePtr);
+    return XXH_OK;
+}
+
+/*! @ingroup xxh3_family */
+XXH_PUBLIC_API void
+XXH3_copyState(XXH3_state_t* dst_state, const XXH3_state_t* src_state)
+{
+    memcpy(dst_state, src_state, sizeof(*dst_state));
+}
+
+static void
+XXH3_reset_internal(XXH3_state_t* statePtr,
+                           XXH64_hash_t seed,
+                           const void* secret, size_t secretSize)
+{
+    size_t const initStart = offsetof(XXH3_state_t, bufferedSize);
+    size_t const initLength = offsetof(XXH3_state_t, nbStripesPerBlock) - initStart;
+    XXH_ASSERT(offsetof(XXH3_state_t, nbStripesPerBlock) > initStart);
+    XXH_ASSERT(statePtr != NULL);
+    /* set members from bufferedSize to nbStripesPerBlock (excluded) to 0 */
+    memset((char*)statePtr + initStart, 0, initLength);
+    statePtr->acc[0] = XXH_PRIME32_3;
+    statePtr->acc[1] = XXH_PRIME64_1;
+    statePtr->acc[2] = XXH_PRIME64_2;
+    statePtr->acc[3] = XXH_PRIME64_3;
+    statePtr->acc[4] = XXH_PRIME64_4;
+    statePtr->acc[5] = XXH_PRIME32_2;
+    statePtr->acc[6] = XXH_PRIME64_5;
+    statePtr->acc[7] = XXH_PRIME32_1;
+    statePtr->seed = seed;
+    statePtr->extSecret = (const unsigned char*)secret;
+    XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN);
+    statePtr->secretLimit = secretSize - XXH_STRIPE_LEN;
+    statePtr->nbStripesPerBlock = statePtr->secretLimit / XXH_SECRET_CONSUME_RATE;
+}
+
+/*! @ingroup xxh3_family */
+XXH_PUBLIC_API XXH_errorcode
+XXH3_64bits_reset(XXH3_state_t* statePtr)
+{
+    if (statePtr == NULL) return XXH_ERROR;
+    XXH3_reset_internal(statePtr, 0, XXH3_kSecret, XXH_SECRET_DEFAULT_SIZE);
+    return XXH_OK;
+}
+
+/*! @ingroup xxh3_family */
+XXH_PUBLIC_API XXH_errorcode
+XXH3_64bits_reset_withSecret(XXH3_state_t* statePtr, const void* secret, size_t secretSize)
+{
+    if (statePtr == NULL) return XXH_ERROR;
+    XXH3_reset_internal(statePtr, 0, secret, secretSize);
+    if (secret == NULL) return XXH_ERROR;
+    if (secretSize < XXH3_SECRET_SIZE_MIN) return XXH_ERROR;
+    return XXH_OK;
+}
+
+/*! @ingroup xxh3_family */
+XXH_PUBLIC_API XXH_errorcode
+XXH3_64bits_reset_withSeed(XXH3_state_t* statePtr, XXH64_hash_t seed)
+{
+    if (statePtr == NULL) return XXH_ERROR;
+    if (seed==0) return XXH3_64bits_reset(statePtr);
+    if (seed != statePtr->seed) XXH3_initCustomSecret(statePtr->customSecret, seed);
+    XXH3_reset_internal(statePtr, seed, NULL, XXH_SECRET_DEFAULT_SIZE);
+    return XXH_OK;
+}
+
+/* Note : when XXH3_consumeStripes() is invoked,
+ * there must be a guarantee that at least one more byte must be consumed from input
+ * so that the function can blindly consume all stripes using the "normal" secret segment */
+XXH_FORCE_INLINE void
+XXH3_consumeStripes(xxh_u64* XXH_RESTRICT acc,
+                    size_t* XXH_RESTRICT nbStripesSoFarPtr, size_t nbStripesPerBlock,
+                    const xxh_u8* XXH_RESTRICT input, size_t nbStripes,
+                    const xxh_u8* XXH_RESTRICT secret, size_t secretLimit,
+                    XXH3_f_accumulate_512 f_acc512,
+                    XXH3_f_scrambleAcc f_scramble)
+{
+    XXH_ASSERT(nbStripes <= nbStripesPerBlock);  /* can handle max 1 scramble per invocation */
+    XXH_ASSERT(*nbStripesSoFarPtr < nbStripesPerBlock);
+    if (nbStripesPerBlock - *nbStripesSoFarPtr <= nbStripes) {
+        /* need a scrambling operation */
+        size_t const nbStripesToEndofBlock = nbStripesPerBlock - *nbStripesSoFarPtr;
+        size_t const nbStripesAfterBlock = nbStripes - nbStripesToEndofBlock;
+        XXH3_accumulate(acc, input, secret + nbStripesSoFarPtr[0] * XXH_SECRET_CONSUME_RATE, nbStripesToEndofBlock, f_acc512);
+        f_scramble(acc, secret + secretLimit);
+        XXH3_accumulate(acc, input + nbStripesToEndofBlock * XXH_STRIPE_LEN, secret, nbStripesAfterBlock, f_acc512);
+        *nbStripesSoFarPtr = nbStripesAfterBlock;
+    } else {
+        XXH3_accumulate(acc, input, secret + nbStripesSoFarPtr[0] * XXH_SECRET_CONSUME_RATE, nbStripes, f_acc512);
+        *nbStripesSoFarPtr += nbStripes;
+    }
+}
+
+/*
+ * Both XXH3_64bits_update and XXH3_128bits_update use this routine.
+ */
+XXH_FORCE_INLINE XXH_errorcode
+XXH3_update(XXH3_state_t* state,
+            const xxh_u8* input, size_t len,
+            XXH3_f_accumulate_512 f_acc512,
+            XXH3_f_scrambleAcc f_scramble)
+{
+    if (input==NULL)
+#if defined(XXH_ACCEPT_NULL_INPUT_POINTER) && (XXH_ACCEPT_NULL_INPUT_POINTER>=1)
+        return XXH_OK;
+#else
+        return XXH_ERROR;
+#endif
+
+    {   const xxh_u8* const bEnd = input + len;
+        const unsigned char* const secret = (state->extSecret == NULL) ? state->customSecret : state->extSecret;
+
+        state->totalLen += len;
+        XXH_ASSERT(state->bufferedSize <= XXH3_INTERNALBUFFER_SIZE);
+
+        if (state->bufferedSize + len <= XXH3_INTERNALBUFFER_SIZE) {  /* fill in tmp buffer */
+            XXH_memcpy(state->buffer + state->bufferedSize, input, len);
+            state->bufferedSize += (XXH32_hash_t)len;
+            return XXH_OK;
+        }
+        /* total input is now > XXH3_INTERNALBUFFER_SIZE */
+
+        #define XXH3_INTERNALBUFFER_STRIPES (XXH3_INTERNALBUFFER_SIZE / XXH_STRIPE_LEN)
+        XXH_STATIC_ASSERT(XXH3_INTERNALBUFFER_SIZE % XXH_STRIPE_LEN == 0);   /* clean multiple */
+
+        /*
+         * Internal buffer is partially filled (always, except at beginning)
+         * Complete it, then consume it.
+         */
+        if (state->bufferedSize) {
+            size_t const loadSize = XXH3_INTERNALBUFFER_SIZE - state->bufferedSize;
+            XXH_memcpy(state->buffer + state->bufferedSize, input, loadSize);
+            input += loadSize;
+            XXH3_consumeStripes(state->acc,
+                               &state->nbStripesSoFar, state->nbStripesPerBlock,
+                                state->buffer, XXH3_INTERNALBUFFER_STRIPES,
+                                secret, state->secretLimit,
+                                f_acc512, f_scramble);
+            state->bufferedSize = 0;
+        }
+        XXH_ASSERT(input < bEnd);
+
+        /* Consume input by a multiple of internal buffer size */
+        if (input+XXH3_INTERNALBUFFER_SIZE < bEnd) {
+            const xxh_u8* const limit = bEnd - XXH3_INTERNALBUFFER_SIZE;
+            do {
+                XXH3_consumeStripes(state->acc,
+                                   &state->nbStripesSoFar, state->nbStripesPerBlock,
+                                    input, XXH3_INTERNALBUFFER_STRIPES,
+                                    secret, state->secretLimit,
+                                    f_acc512, f_scramble);
+                input += XXH3_INTERNALBUFFER_SIZE;
+            } while (input<limit);
+            /* for last partial stripe */
+            memcpy(state->buffer + sizeof(state->buffer) - XXH_STRIPE_LEN, input - XXH_STRIPE_LEN, XXH_STRIPE_LEN);
+        }
+        XXH_ASSERT(input < bEnd);
+
+        /* Some remaining input (always) : buffer it */
+        XXH_memcpy(state->buffer, input, (size_t)(bEnd-input));
+        state->bufferedSize = (XXH32_hash_t)(bEnd-input);
+    }
+
+    return XXH_OK;
+}
+
+/*! @ingroup xxh3_family */
+XXH_PUBLIC_API XXH_errorcode
+XXH3_64bits_update(XXH3_state_t* state, const void* input, size_t len)
+{
+    return XXH3_update(state, (const xxh_u8*)input, len,
+                       XXH3_accumulate_512, XXH3_scrambleAcc);
+}
+
+
+XXH_FORCE_INLINE void
+XXH3_digest_long (XXH64_hash_t* acc,
+                  const XXH3_state_t* state,
+                  const unsigned char* secret)
+{
+    /*
+     * Digest on a local copy. This way, the state remains unaltered, and it can
+     * continue ingesting more input afterwards.
+     */
+    memcpy(acc, state->acc, sizeof(state->acc));
+    if (state->bufferedSize >= XXH_STRIPE_LEN) {
+        size_t const nbStripes = (state->bufferedSize - 1) / XXH_STRIPE_LEN;
+        size_t nbStripesSoFar = state->nbStripesSoFar;
+        XXH3_consumeStripes(acc,
+                           &nbStripesSoFar, state->nbStripesPerBlock,
+                            state->buffer, nbStripes,
+                            secret, state->secretLimit,
+                            XXH3_accumulate_512, XXH3_scrambleAcc);
+        /* last stripe */
+        XXH3_accumulate_512(acc,
+                            state->buffer + state->bufferedSize - XXH_STRIPE_LEN,
+                            secret + state->secretLimit - XXH_SECRET_LASTACC_START);
+    } else {  /* bufferedSize < XXH_STRIPE_LEN */
+        xxh_u8 lastStripe[XXH_STRIPE_LEN];
+        size_t const catchupSize = XXH_STRIPE_LEN - state->bufferedSize;
+        XXH_ASSERT(state->bufferedSize > 0);  /* there is always some input buffered */
+        memcpy(lastStripe, state->buffer + sizeof(state->buffer) - catchupSize, catchupSize);
+        memcpy(lastStripe + catchupSize, state->buffer, state->bufferedSize);
+        XXH3_accumulate_512(acc,
+                            lastStripe,
+                            secret + state->secretLimit - XXH_SECRET_LASTACC_START);
+    }
+}
+
+/*! @ingroup xxh3_family */
+XXH_PUBLIC_API XXH64_hash_t XXH3_64bits_digest (const XXH3_state_t* state)
+{
+    const unsigned char* const secret = (state->extSecret == NULL) ? state->customSecret : state->extSecret;
+    if (state->totalLen > XXH3_MIDSIZE_MAX) {
+        XXH_ALIGN(XXH_ACC_ALIGN) XXH64_hash_t acc[XXH_ACC_NB];
+        XXH3_digest_long(acc, state, secret);
+        return XXH3_mergeAccs(acc,
+                              secret + XXH_SECRET_MERGEACCS_START,
+                              (xxh_u64)state->totalLen * XXH_PRIME64_1);
+    }
+    /* totalLen <= XXH3_MIDSIZE_MAX: digesting a short input */
+    if (state->seed)
+        return XXH3_64bits_withSeed(state->buffer, (size_t)state->totalLen, state->seed);
+    return XXH3_64bits_withSecret(state->buffer, (size_t)(state->totalLen),
+                                  secret, state->secretLimit + XXH_STRIPE_LEN);
+}
+
+
+#define XXH_MIN(x, y) (((x) > (y)) ? (y) : (x))
+
+/*! @ingroup xxh3_family */
+XXH_PUBLIC_API void
+XXH3_generateSecret(void* secretBuffer, const void* customSeed, size_t customSeedSize)
+{
+    XXH_ASSERT(secretBuffer != NULL);
+    if (customSeedSize == 0) {
+        memcpy(secretBuffer, XXH3_kSecret, XXH_SECRET_DEFAULT_SIZE);
+        return;
+    }
+    XXH_ASSERT(customSeed != NULL);
+
+    {   size_t const segmentSize = sizeof(XXH128_hash_t);
+        size_t const nbSegments = XXH_SECRET_DEFAULT_SIZE / segmentSize;
+        XXH128_canonical_t scrambler;
+        XXH64_hash_t seeds[12];
+        size_t segnb;
+        XXH_ASSERT(nbSegments == 12);
+        XXH_ASSERT(segmentSize * nbSegments == XXH_SECRET_DEFAULT_SIZE); /* exact multiple */
+        XXH128_canonicalFromHash(&scrambler, XXH128(customSeed, customSeedSize, 0));
+
+        /*
+        * Copy customSeed to seeds[], truncating or repeating as necessary.
+        */
+        {   size_t toFill = XXH_MIN(customSeedSize, sizeof(seeds));
+            size_t filled = toFill;
+            memcpy(seeds, customSeed, toFill);
+            while (filled < sizeof(seeds)) {
+                toFill = XXH_MIN(filled, sizeof(seeds) - filled);
+                memcpy((char*)seeds + filled, seeds, toFill);
+                filled += toFill;
+        }   }
+
+        /* generate secret */
+        memcpy(secretBuffer, &scrambler, sizeof(scrambler));
+        for (segnb=1; segnb < nbSegments; segnb++) {
+            size_t const segmentStart = segnb * segmentSize;
+            XXH128_canonical_t segment;
+            XXH128_canonicalFromHash(&segment,
+                XXH128(&scrambler, sizeof(scrambler), XXH_readLE64(seeds + segnb) + segnb) );
+            memcpy((char*)secretBuffer + segmentStart, &segment, sizeof(segment));
+    }   }
+}
+
+
+/* ==========================================
+ * XXH3 128 bits (a.k.a XXH128)
+ * ==========================================
+ * XXH3's 128-bit variant has better mixing and strength than the 64-bit variant,
+ * even without counting the significantly larger output size.
+ *
+ * For example, extra steps are taken to avoid the seed-dependent collisions
+ * in 17-240 byte inputs (See XXH3_mix16B and XXH128_mix32B).
+ *
+ * This strength naturally comes at the cost of some speed, especially on short
+ * lengths. Note that longer hashes are about as fast as the 64-bit version
+ * due to it using only a slight modification of the 64-bit loop.
+ *
+ * XXH128 is also more oriented towards 64-bit machines. It is still extremely
+ * fast for a _128-bit_ hash on 32-bit (it usually clears XXH64).
+ */
+
+XXH_FORCE_INLINE XXH128_hash_t
+XXH3_len_1to3_128b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed)
+{
+    /* A doubled version of 1to3_64b with different constants. */
+    XXH_ASSERT(input != NULL);
+    XXH_ASSERT(1 <= len && len <= 3);
+    XXH_ASSERT(secret != NULL);
+    /*
+     * len = 1: combinedl = { input[0], 0x01, input[0], input[0] }
+     * len = 2: combinedl = { input[1], 0x02, input[0], input[1] }
+     * len = 3: combinedl = { input[2], 0x03, input[0], input[1] }
+     */
+    {   xxh_u8 const c1 = input[0];
+        xxh_u8 const c2 = input[len >> 1];
+        xxh_u8 const c3 = input[len - 1];
+        xxh_u32 const combinedl = ((xxh_u32)c1 <<16) | ((xxh_u32)c2 << 24)
+                                | ((xxh_u32)c3 << 0) | ((xxh_u32)len << 8);
+        xxh_u32 const combinedh = XXH_rotl32(XXH_swap32(combinedl), 13);
+        xxh_u64 const bitflipl = (XXH_readLE32(secret) ^ XXH_readLE32(secret+4)) + seed;
+        xxh_u64 const bitfliph = (XXH_readLE32(secret+8) ^ XXH_readLE32(secret+12)) - seed;
+        xxh_u64 const keyed_lo = (xxh_u64)combinedl ^ bitflipl;
+        xxh_u64 const keyed_hi = (xxh_u64)combinedh ^ bitfliph;
+        XXH128_hash_t h128;
+        h128.low64  = XXH64_avalanche(keyed_lo);
+        h128.high64 = XXH64_avalanche(keyed_hi);
+        return h128;
+    }
+}
+
+XXH_FORCE_INLINE XXH128_hash_t
+XXH3_len_4to8_128b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed)
+{
+    XXH_ASSERT(input != NULL);
+    XXH_ASSERT(secret != NULL);
+    XXH_ASSERT(4 <= len && len <= 8);
+    seed ^= (xxh_u64)XXH_swap32((xxh_u32)seed) << 32;
+    {   xxh_u32 const input_lo = XXH_readLE32(input);
+        xxh_u32 const input_hi = XXH_readLE32(input + len - 4);
+        xxh_u64 const input_64 = input_lo + ((xxh_u64)input_hi << 32);
+        xxh_u64 const bitflip = (XXH_readLE64(secret+16) ^ XXH_readLE64(secret+24)) + seed;
+        xxh_u64 const keyed = input_64 ^ bitflip;
+
+        /* Shift len to the left to ensure it is even, this avoids even multiplies. */
+        XXH128_hash_t m128 = XXH_mult64to128(keyed, XXH_PRIME64_1 + (len << 2));
+
+        m128.high64 += (m128.low64 << 1);
+        m128.low64  ^= (m128.high64 >> 3);
+
+        m128.low64   = XXH_xorshift64(m128.low64, 35);
+        m128.low64  *= 0x9FB21C651E98DF25ULL;
+        m128.low64   = XXH_xorshift64(m128.low64, 28);
+        m128.high64  = XXH3_avalanche(m128.high64);
+        return m128;
+    }
+}
+
+XXH_FORCE_INLINE XXH128_hash_t
+XXH3_len_9to16_128b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed)
+{
+    XXH_ASSERT(input != NULL);
+    XXH_ASSERT(secret != NULL);
+    XXH_ASSERT(9 <= len && len <= 16);
+    {   xxh_u64 const bitflipl = (XXH_readLE64(secret+32) ^ XXH_readLE64(secret+40)) - seed;
+        xxh_u64 const bitfliph = (XXH_readLE64(secret+48) ^ XXH_readLE64(secret+56)) + seed;
+        xxh_u64 const input_lo = XXH_readLE64(input);
+        xxh_u64       input_hi = XXH_readLE64(input + len - 8);
+        XXH128_hash_t m128 = XXH_mult64to128(input_lo ^ input_hi ^ bitflipl, XXH_PRIME64_1);
+        /*
+         * Put len in the middle of m128 to ensure that the length gets mixed to
+         * both the low and high bits in the 128x64 multiply below.
+         */
+        m128.low64 += (xxh_u64)(len - 1) << 54;
+        input_hi   ^= bitfliph;
+        /*
+         * Add the high 32 bits of input_hi to the high 32 bits of m128, then
+         * add the long product of the low 32 bits of input_hi and XXH_PRIME32_2 to
+         * the high 64 bits of m128.
+         *
+         * The best approach to this operation is different on 32-bit and 64-bit.
+         */
+        if (sizeof(void *) < sizeof(xxh_u64)) { /* 32-bit */
+            /*
+             * 32-bit optimized version, which is more readable.
+             *
+             * On 32-bit, it removes an ADC and delays a dependency between the two
+             * halves of m128.high64, but it generates an extra mask on 64-bit.
+             */
+            m128.high64 += (input_hi & 0xFFFFFFFF00000000ULL) + XXH_mult32to64((xxh_u32)input_hi, XXH_PRIME32_2);
+        } else {
+            /*
+             * 64-bit optimized (albeit more confusing) version.
+             *
+             * Uses some properties of addition and multiplication to remove the mask:
+             *
+             * Let:
+             *    a = input_hi.lo = (input_hi & 0x00000000FFFFFFFF)
+             *    b = input_hi.hi = (input_hi & 0xFFFFFFFF00000000)
+             *    c = XXH_PRIME32_2
+             *
+             *    a + (b * c)
+             * Inverse Property: x + y - x == y
+             *    a + (b * (1 + c - 1))
+             * Distributive Property: x * (y + z) == (x * y) + (x * z)
+             *    a + (b * 1) + (b * (c - 1))
+             * Identity Property: x * 1 == x
+             *    a + b + (b * (c - 1))
+             *
+             * Substitute a, b, and c:
+             *    input_hi.hi + input_hi.lo + ((xxh_u64)input_hi.lo * (XXH_PRIME32_2 - 1))
+             *
+             * Since input_hi.hi + input_hi.lo == input_hi, we get this:
+             *    input_hi + ((xxh_u64)input_hi.lo * (XXH_PRIME32_2 - 1))
+             */
+            m128.high64 += input_hi + XXH_mult32to64((xxh_u32)input_hi, XXH_PRIME32_2 - 1);
+        }
+        /* m128 ^= XXH_swap64(m128 >> 64); */
+        m128.low64  ^= XXH_swap64(m128.high64);
+
+        {   /* 128x64 multiply: h128 = m128 * XXH_PRIME64_2; */
+            XXH128_hash_t h128 = XXH_mult64to128(m128.low64, XXH_PRIME64_2);
+            h128.high64 += m128.high64 * XXH_PRIME64_2;
+
+            h128.low64   = XXH3_avalanche(h128.low64);
+            h128.high64  = XXH3_avalanche(h128.high64);
+            return h128;
+    }   }
+}
+
+/*
+ * Assumption: `secret` size is >= XXH3_SECRET_SIZE_MIN
+ */
+XXH_FORCE_INLINE XXH128_hash_t
+XXH3_len_0to16_128b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed)
+{
+    XXH_ASSERT(len <= 16);
+    {   if (len > 8) return XXH3_len_9to16_128b(input, len, secret, seed);
+        if (len >= 4) return XXH3_len_4to8_128b(input, len, secret, seed);
+        if (len) return XXH3_len_1to3_128b(input, len, secret, seed);
+        {   XXH128_hash_t h128;
+            xxh_u64 const bitflipl = XXH_readLE64(secret+64) ^ XXH_readLE64(secret+72);
+            xxh_u64 const bitfliph = XXH_readLE64(secret+80) ^ XXH_readLE64(secret+88);
+            h128.low64 = XXH64_avalanche(seed ^ bitflipl);
+            h128.high64 = XXH64_avalanche( seed ^ bitfliph);
+            return h128;
+    }   }
+}
+
+/*
+ * A bit slower than XXH3_mix16B, but handles multiply by zero better.
+ */
+XXH_FORCE_INLINE XXH128_hash_t
+XXH128_mix32B(XXH128_hash_t acc, const xxh_u8* input_1, const xxh_u8* input_2,
+              const xxh_u8* secret, XXH64_hash_t seed)
+{
+    acc.low64  += XXH3_mix16B (input_1, secret+0, seed);
+    acc.low64  ^= XXH_readLE64(input_2) + XXH_readLE64(input_2 + 8);
+    acc.high64 += XXH3_mix16B (input_2, secret+16, seed);
+    acc.high64 ^= XXH_readLE64(input_1) + XXH_readLE64(input_1 + 8);
+    return acc;
+}
+
+
+XXH_FORCE_INLINE XXH128_hash_t
+XXH3_len_17to128_128b(const xxh_u8* XXH_RESTRICT input, size_t len,
+                      const xxh_u8* XXH_RESTRICT secret, size_t secretSize,
+                      XXH64_hash_t seed)
+{
+    XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN); (void)secretSize;
+    XXH_ASSERT(16 < len && len <= 128);
+
+    {   XXH128_hash_t acc;
+        acc.low64 = len * XXH_PRIME64_1;
+        acc.high64 = 0;
+        if (len > 32) {
+            if (len > 64) {
+                if (len > 96) {
+                    acc = XXH128_mix32B(acc, input+48, input+len-64, secret+96, seed);
+                }
+                acc = XXH128_mix32B(acc, input+32, input+len-48, secret+64, seed);
+            }
+            acc = XXH128_mix32B(acc, input+16, input+len-32, secret+32, seed);
+        }
+        acc = XXH128_mix32B(acc, input, input+len-16, secret, seed);
+        {   XXH128_hash_t h128;
+            h128.low64  = acc.low64 + acc.high64;
+            h128.high64 = (acc.low64    * XXH_PRIME64_1)
+                        + (acc.high64   * XXH_PRIME64_4)
+                        + ((len - seed) * XXH_PRIME64_2);
+            h128.low64  = XXH3_avalanche(h128.low64);
+            h128.high64 = (XXH64_hash_t)0 - XXH3_avalanche(h128.high64);
+            return h128;
+        }
+    }
+}
+
+XXH_NO_INLINE XXH128_hash_t
+XXH3_len_129to240_128b(const xxh_u8* XXH_RESTRICT input, size_t len,
+                       const xxh_u8* XXH_RESTRICT secret, size_t secretSize,
+                       XXH64_hash_t seed)
+{
+    XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN); (void)secretSize;
+    XXH_ASSERT(128 < len && len <= XXH3_MIDSIZE_MAX);
+
+    {   XXH128_hash_t acc;
+        int const nbRounds = (int)len / 32;
+        int i;
+        acc.low64 = len * XXH_PRIME64_1;
+        acc.high64 = 0;
+        for (i=0; i<4; i++) {
+            acc = XXH128_mix32B(acc,
+                                input  + (32 * i),
+                                input  + (32 * i) + 16,
+                                secret + (32 * i),
+                                seed);
+        }
+        acc.low64 = XXH3_avalanche(acc.low64);
+        acc.high64 = XXH3_avalanche(acc.high64);
+        XXH_ASSERT(nbRounds >= 4);
+        for (i=4 ; i < nbRounds; i++) {
+            acc = XXH128_mix32B(acc,
+                                input + (32 * i),
+                                input + (32 * i) + 16,
+                                secret + XXH3_MIDSIZE_STARTOFFSET + (32 * (i - 4)),
+                                seed);
+        }
+        /* last bytes */
+        acc = XXH128_mix32B(acc,
+                            input + len - 16,
+                            input + len - 32,
+                            secret + XXH3_SECRET_SIZE_MIN - XXH3_MIDSIZE_LASTOFFSET - 16,
+                            0ULL - seed);
+
+        {   XXH128_hash_t h128;
+            h128.low64  = acc.low64 + acc.high64;
+            h128.high64 = (acc.low64    * XXH_PRIME64_1)
+                        + (acc.high64   * XXH_PRIME64_4)
+                        + ((len - seed) * XXH_PRIME64_2);
+            h128.low64  = XXH3_avalanche(h128.low64);
+            h128.high64 = (XXH64_hash_t)0 - XXH3_avalanche(h128.high64);
+            return h128;
+        }
+    }
+}
+
+XXH_FORCE_INLINE XXH128_hash_t
+XXH3_hashLong_128b_internal(const void* XXH_RESTRICT input, size_t len,
+                            const xxh_u8* XXH_RESTRICT secret, size_t secretSize,
+                            XXH3_f_accumulate_512 f_acc512,
+                            XXH3_f_scrambleAcc f_scramble)
+{
+    XXH_ALIGN(XXH_ACC_ALIGN) xxh_u64 acc[XXH_ACC_NB] = XXH3_INIT_ACC;
+
+    XXH3_hashLong_internal_loop(acc, (const xxh_u8*)input, len, secret, secretSize, f_acc512, f_scramble);
+
+    /* converge into final hash */
+    XXH_STATIC_ASSERT(sizeof(acc) == 64);
+    XXH_ASSERT(secretSize >= sizeof(acc) + XXH_SECRET_MERGEACCS_START);
+    {   XXH128_hash_t h128;
+        h128.low64  = XXH3_mergeAccs(acc,
+                                     secret + XXH_SECRET_MERGEACCS_START,
+                                     (xxh_u64)len * XXH_PRIME64_1);
+        h128.high64 = XXH3_mergeAccs(acc,
+                                     secret + secretSize
+                                            - sizeof(acc) - XXH_SECRET_MERGEACCS_START,
+                                     ~((xxh_u64)len * XXH_PRIME64_2));
+        return h128;
+    }
+}
+
+/*
+ * It's important for performance that XXH3_hashLong is not inlined.
+ */
+XXH_NO_INLINE XXH128_hash_t
+XXH3_hashLong_128b_default(const void* XXH_RESTRICT input, size_t len,
+                           XXH64_hash_t seed64,
+                           const void* XXH_RESTRICT secret, size_t secretLen)
+{
+    (void)seed64; (void)secret; (void)secretLen;
+    return XXH3_hashLong_128b_internal(input, len, XXH3_kSecret, sizeof(XXH3_kSecret),
+                                       XXH3_accumulate_512, XXH3_scrambleAcc);
+}
+
+/*
+ * It's important for performance that XXH3_hashLong is not inlined.
+ */
+XXH_NO_INLINE XXH128_hash_t
+XXH3_hashLong_128b_withSecret(const void* XXH_RESTRICT input, size_t len,
+                              XXH64_hash_t seed64,
+                              const void* XXH_RESTRICT secret, size_t secretLen)
+{
+    (void)seed64;
+    return XXH3_hashLong_128b_internal(input, len, (const xxh_u8*)secret, secretLen,
+                                       XXH3_accumulate_512, XXH3_scrambleAcc);
+}
+
+XXH_FORCE_INLINE XXH128_hash_t
+XXH3_hashLong_128b_withSeed_internal(const void* XXH_RESTRICT input, size_t len,
+                                XXH64_hash_t seed64,
+                                XXH3_f_accumulate_512 f_acc512,
+                                XXH3_f_scrambleAcc f_scramble,
+                                XXH3_f_initCustomSecret f_initSec)
+{
+    if (seed64 == 0)
+        return XXH3_hashLong_128b_internal(input, len,
+                                           XXH3_kSecret, sizeof(XXH3_kSecret),
+                                           f_acc512, f_scramble);
+    {   XXH_ALIGN(XXH_SEC_ALIGN) xxh_u8 secret[XXH_SECRET_DEFAULT_SIZE];
+        f_initSec(secret, seed64);
+        return XXH3_hashLong_128b_internal(input, len, (const xxh_u8*)secret, sizeof(secret),
+                                           f_acc512, f_scramble);
+    }
+}
+
+/*
+ * It's important for performance that XXH3_hashLong is not inlined.
+ */
+XXH_NO_INLINE XXH128_hash_t
+XXH3_hashLong_128b_withSeed(const void* input, size_t len,
+                            XXH64_hash_t seed64, const void* XXH_RESTRICT secret, size_t secretLen)
+{
+    (void)secret; (void)secretLen;
+    return XXH3_hashLong_128b_withSeed_internal(input, len, seed64,
+                XXH3_accumulate_512, XXH3_scrambleAcc, XXH3_initCustomSecret);
+}
+
+typedef XXH128_hash_t (*XXH3_hashLong128_f)(const void* XXH_RESTRICT, size_t,
+                                            XXH64_hash_t, const void* XXH_RESTRICT, size_t);
+
+XXH_FORCE_INLINE XXH128_hash_t
+XXH3_128bits_internal(const void* input, size_t len,
+                      XXH64_hash_t seed64, const void* XXH_RESTRICT secret, size_t secretLen,
+                      XXH3_hashLong128_f f_hl128)
+{
+    XXH_ASSERT(secretLen >= XXH3_SECRET_SIZE_MIN);
+    /*
+     * If an action is to be taken if `secret` conditions are not respected,
+     * it should be done here.
+     * For now, it's a contract pre-condition.
+     * Adding a check and a branch here would cost performance at every hash.
+     */
+    if (len <= 16)
+        return XXH3_len_0to16_128b((const xxh_u8*)input, len, (const xxh_u8*)secret, seed64);
+    if (len <= 128)
+        return XXH3_len_17to128_128b((const xxh_u8*)input, len, (const xxh_u8*)secret, secretLen, seed64);
+    if (len <= XXH3_MIDSIZE_MAX)
+        return XXH3_len_129to240_128b((const xxh_u8*)input, len, (const xxh_u8*)secret, secretLen, seed64);
+    return f_hl128(input, len, seed64, secret, secretLen);
+}
+
+
+/* ===   Public XXH128 API   === */
+
+/*! @ingroup xxh3_family */
+XXH_PUBLIC_API XXH128_hash_t XXH3_128bits(const void* input, size_t len)
+{
+    return XXH3_128bits_internal(input, len, 0,
+                                 XXH3_kSecret, sizeof(XXH3_kSecret),
+                                 XXH3_hashLong_128b_default);
+}
+
+/*! @ingroup xxh3_family */
+XXH_PUBLIC_API XXH128_hash_t
+XXH3_128bits_withSecret(const void* input, size_t len, const void* secret, size_t secretSize)
+{
+    return XXH3_128bits_internal(input, len, 0,
+                                 (const xxh_u8*)secret, secretSize,
+                                 XXH3_hashLong_128b_withSecret);
+}
+
+/*! @ingroup xxh3_family */
+XXH_PUBLIC_API XXH128_hash_t
+XXH3_128bits_withSeed(const void* input, size_t len, XXH64_hash_t seed)
+{
+    return XXH3_128bits_internal(input, len, seed,
+                                 XXH3_kSecret, sizeof(XXH3_kSecret),
+                                 XXH3_hashLong_128b_withSeed);
+}
+
+/*! @ingroup xxh3_family */
+XXH_PUBLIC_API XXH128_hash_t
+XXH128(const void* input, size_t len, XXH64_hash_t seed)
+{
+    return XXH3_128bits_withSeed(input, len, seed);
+}
+
+
+/* ===   XXH3 128-bit streaming   === */
+
+/*
+ * All the functions are actually the same as for 64-bit streaming variant.
+ * The only difference is the finalization routine.
+ */
+
+/*! @ingroup xxh3_family */
+XXH_PUBLIC_API XXH_errorcode
+XXH3_128bits_reset(XXH3_state_t* statePtr)
+{
+    if (statePtr == NULL) return XXH_ERROR;
+    XXH3_reset_internal(statePtr, 0, XXH3_kSecret, XXH_SECRET_DEFAULT_SIZE);
+    return XXH_OK;
+}
+
+/*! @ingroup xxh3_family */
+XXH_PUBLIC_API XXH_errorcode
+XXH3_128bits_reset_withSecret(XXH3_state_t* statePtr, const void* secret, size_t secretSize)
+{
+    if (statePtr == NULL) return XXH_ERROR;
+    XXH3_reset_internal(statePtr, 0, secret, secretSize);
+    if (secret == NULL) return XXH_ERROR;
+    if (secretSize < XXH3_SECRET_SIZE_MIN) return XXH_ERROR;
+    return XXH_OK;
+}
+
+/*! @ingroup xxh3_family */
+XXH_PUBLIC_API XXH_errorcode
+XXH3_128bits_reset_withSeed(XXH3_state_t* statePtr, XXH64_hash_t seed)
+{
+    if (statePtr == NULL) return XXH_ERROR;
+    if (seed==0) return XXH3_128bits_reset(statePtr);
+    if (seed != statePtr->seed) XXH3_initCustomSecret(statePtr->customSecret, seed);
+    XXH3_reset_internal(statePtr, seed, NULL, XXH_SECRET_DEFAULT_SIZE);
+    return XXH_OK;
+}
+
+/*! @ingroup xxh3_family */
+XXH_PUBLIC_API XXH_errorcode
+XXH3_128bits_update(XXH3_state_t* state, const void* input, size_t len)
+{
+    return XXH3_update(state, (const xxh_u8*)input, len,
+                       XXH3_accumulate_512, XXH3_scrambleAcc);
+}
+
+/*! @ingroup xxh3_family */
+XXH_PUBLIC_API XXH128_hash_t XXH3_128bits_digest (const XXH3_state_t* state)
+{
+    const unsigned char* const secret = (state->extSecret == NULL) ? state->customSecret : state->extSecret;
+    if (state->totalLen > XXH3_MIDSIZE_MAX) {
+        XXH_ALIGN(XXH_ACC_ALIGN) XXH64_hash_t acc[XXH_ACC_NB];
+        XXH3_digest_long(acc, state, secret);
+        XXH_ASSERT(state->secretLimit + XXH_STRIPE_LEN >= sizeof(acc) + XXH_SECRET_MERGEACCS_START);
+        {   XXH128_hash_t h128;
+            h128.low64  = XXH3_mergeAccs(acc,
+                                         secret + XXH_SECRET_MERGEACCS_START,
+                                         (xxh_u64)state->totalLen * XXH_PRIME64_1);
+            h128.high64 = XXH3_mergeAccs(acc,
+                                         secret + state->secretLimit + XXH_STRIPE_LEN
+                                                - sizeof(acc) - XXH_SECRET_MERGEACCS_START,
+                                         ~((xxh_u64)state->totalLen * XXH_PRIME64_2));
+            return h128;
+        }
+    }
+    /* len <= XXH3_MIDSIZE_MAX : short code */
+    if (state->seed)
+        return XXH3_128bits_withSeed(state->buffer, (size_t)state->totalLen, state->seed);
+    return XXH3_128bits_withSecret(state->buffer, (size_t)(state->totalLen),
+                                   secret, state->secretLimit + XXH_STRIPE_LEN);
+}
+
+/* 128-bit utility functions */
+
+#include <string.h>   /* memcmp, memcpy */
+
+/* return : 1 is equal, 0 if different */
+/*! @ingroup xxh3_family */
+XXH_PUBLIC_API int XXH128_isEqual(XXH128_hash_t h1, XXH128_hash_t h2)
+{
+    /* note : XXH128_hash_t is compact, it has no padding byte */
+    return !(memcmp(&h1, &h2, sizeof(h1)));
+}
+
+/* This prototype is compatible with stdlib's qsort().
+ * return : >0 if *h128_1  > *h128_2
+ *          <0 if *h128_1  < *h128_2
+ *          =0 if *h128_1 == *h128_2  */
+/*! @ingroup xxh3_family */
+XXH_PUBLIC_API int XXH128_cmp(const void* h128_1, const void* h128_2)
+{
+    XXH128_hash_t const h1 = *(const XXH128_hash_t*)h128_1;
+    XXH128_hash_t const h2 = *(const XXH128_hash_t*)h128_2;
+    int const hcmp = (h1.high64 > h2.high64) - (h2.high64 > h1.high64);
+    /* note : bets that, in most cases, hash values are different */
+    if (hcmp) return hcmp;
+    return (h1.low64 > h2.low64) - (h2.low64 > h1.low64);
+}
+
+
+/*======   Canonical representation   ======*/
+/*! @ingroup xxh3_family */
+XXH_PUBLIC_API void
+XXH128_canonicalFromHash(XXH128_canonical_t* dst, XXH128_hash_t hash)
+{
+    XXH_STATIC_ASSERT(sizeof(XXH128_canonical_t) == sizeof(XXH128_hash_t));
+    if (XXH_CPU_LITTLE_ENDIAN) {
+        hash.high64 = XXH_swap64(hash.high64);
+        hash.low64  = XXH_swap64(hash.low64);
+    }
+    memcpy(dst, &hash.high64, sizeof(hash.high64));
+    memcpy((char*)dst + sizeof(hash.high64), &hash.low64, sizeof(hash.low64));
+}
+
+/*! @ingroup xxh3_family */
+XXH_PUBLIC_API XXH128_hash_t
+XXH128_hashFromCanonical(const XXH128_canonical_t* src)
+{
+    XXH128_hash_t h;
+    h.high64 = XXH_readBE64(src);
+    h.low64  = XXH_readBE64(src->digest + 8);
+    return h;
+}
+
+/* Pop our optimization override from above */
+#if XXH_VECTOR == XXH_AVX2 /* AVX2 */ \
+  && defined(__GNUC__) && !defined(__clang__) /* GCC, not Clang */ \
+  && defined(__OPTIMIZE__) && !defined(__OPTIMIZE_SIZE__) /* respect -O0 and -Os */
+#  pragma GCC pop_options
+#endif
+
+#endif  /* XXH_NO_LONG_LONG */
+
+#endif  /* XXH_NO_XXH3 */
+
+/*!
+ * @}
+ */
+#endif  /* XXH_IMPLEMENTATION */
+
+
+#if defined (__cplusplus)
+}
+#endif
diff --git a/src/rocksdb/util/xxph3.h b/src/rocksdb/util/xxph3.h
new file mode 100644
index 000000000..968000c3a
--- /dev/null
+++ b/src/rocksdb/util/xxph3.h
@@ -0,0 +1,1764 @@
+//  Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+/*
+   xxHash - Extremely Fast Hash algorithm
+   Header File
+   Copyright (C) 2012-2016, Yann Collet.
+
+   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions are
+   met:
+
+       * Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+       * Redistributions in binary form must reproduce the above
+   copyright notice, this list of conditions and the following disclaimer
+   in the documentation and/or other materials provided with the
+   distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+   You can contact the author at :
+   - xxHash source repository : https://github.com/Cyan4973/xxHash
+*/
+
+// This is a fork of a preview version of xxHash, as RocksDB depends on
+// this preview version of XXH3. To allow this to coexist with the
+// standard xxHash, including in the "unity" build where all source files
+// and headers go into a single translation unit, here "XXH" has been
+// replaced with "XXPH" for XX Preview Hash.
+
+#ifndef XXPHASH_H_5627135585666179
+#define XXPHASH_H_5627135585666179 1
+
+/* BEGIN RocksDB customizations */
+#ifndef XXPH_STATIC_LINKING_ONLY
+// Access experimental APIs
+#define XXPH_STATIC_LINKING_ONLY 1
+#endif
+#define XXPH_NAMESPACE ROCKSDB_
+#define XXPH_INLINE_ALL
+#include <cstring>
+/* END RocksDB customizations */
+
+// clang-format off
+#if defined (__cplusplus)
+extern "C" {
+#endif
+
+
+/* ****************************
+*  Definitions
+******************************/
+#include <stddef.h>   /* size_t */
+typedef enum { XXPH_OK=0, XXPH_ERROR } XXPH_errorcode;
+
+
+/* ****************************
+ *  API modifier
+ ******************************/
+/** XXPH_INLINE_ALL (and XXPH_PRIVATE_API)
+ *  This build macro includes xxhash functions in `static` mode
+ *  in order to inline them, and remove their symbol from the public list.
+ *  Inlining offers great performance improvement on small keys,
+ *  and dramatic ones when length is expressed as a compile-time constant.
+ *  See https://fastcompression.blogspot.com/2018/03/xxhash-for-small-keys-impressive-power.html .
+ *  Methodology :
+ *     #define XXPH_INLINE_ALL
+ *     #include "xxhash.h"
+ * `xxhash.c` is automatically included.
+ *  It's not useful to compile and link it as a separate object.
+ */
+#if defined(XXPH_INLINE_ALL) || defined(XXPH_PRIVATE_API)
+#  ifndef XXPH_STATIC_LINKING_ONLY
+#    define XXPH_STATIC_LINKING_ONLY
+#  endif
+#  if defined(__GNUC__)
+#    define XXPH_PUBLIC_API static __inline __attribute__((unused))
+#  elif defined (__cplusplus) || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */)
+#    define XXPH_PUBLIC_API static inline
+#  elif defined(_MSC_VER)
+#    define XXPH_PUBLIC_API static __inline
+#  else
+     /* this version may generate warnings for unused static functions */
+#    define XXPH_PUBLIC_API static
+#  endif
+#else
+#  if defined(WIN32) && defined(_MSC_VER) && (defined(XXPH_IMPORT) || defined(XXPH_EXPORT))
+#    ifdef XXPH_EXPORT
+#      define XXPH_PUBLIC_API __declspec(dllexport)
+#    elif XXPH_IMPORT
+#      define XXPH_PUBLIC_API __declspec(dllimport)
+#    endif
+#  else
+#    define XXPH_PUBLIC_API   /* do nothing */
+#  endif
+#endif /* XXPH_INLINE_ALL || XXPH_PRIVATE_API */
+
+/*! XXPH_NAMESPACE, aka Namespace Emulation :
+ *
+ * If you want to include _and expose_ xxHash functions from within your own library,
+ * but also want to avoid symbol collisions with other libraries which may also include xxHash,
+ *
+ * you can use XXPH_NAMESPACE, to automatically prefix any public symbol from xxhash library
+ * with the value of XXPH_NAMESPACE (therefore, avoid NULL and numeric values).
+ *
+ * Note that no change is required within the calling program as long as it includes `xxhash.h` :
+ * regular symbol name will be automatically translated by this header.
+ */
+#ifdef XXPH_NAMESPACE
+#  define XXPH_CAT(A,B) A##B
+#  define XXPH_NAME2(A,B) XXPH_CAT(A,B)
+#  define XXPH_versionNumber XXPH_NAME2(XXPH_NAMESPACE, XXPH_versionNumber)
+#endif
+
+
+/* *************************************
+*  Version
+***************************************/
+#define XXPH_VERSION_MAJOR    0
+#define XXPH_VERSION_MINOR    7
+#define XXPH_VERSION_RELEASE  2
+#define XXPH_VERSION_NUMBER  (XXPH_VERSION_MAJOR *100*100 + XXPH_VERSION_MINOR *100 + XXPH_VERSION_RELEASE)
+XXPH_PUBLIC_API unsigned XXPH_versionNumber (void);
+
+
+/*-**********************************************************************
+*  32-bit hash
+************************************************************************/
+#if !defined (__VMS) \
+  && (defined (__cplusplus) \
+  || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */) )
+#   include <stdint.h>
+    typedef uint32_t XXPH32_hash_t;
+#else
+#   include <limits.h>
+#   if UINT_MAX == 0xFFFFFFFFUL
+      typedef unsigned int XXPH32_hash_t;
+#   else
+#     if ULONG_MAX == 0xFFFFFFFFUL
+        typedef unsigned long XXPH32_hash_t;
+#     else
+#       error "unsupported platform : need a 32-bit type"
+#     endif
+#   endif
+#endif
+
+#ifndef XXPH_NO_LONG_LONG
+/*-**********************************************************************
+*  64-bit hash
+************************************************************************/
+#if !defined (__VMS) \
+  && (defined (__cplusplus) \
+  || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */) )
+#   include <stdint.h>
+    typedef uint64_t XXPH64_hash_t;
+#else
+    /* the following type must have a width of 64-bit */
+    typedef unsigned long long XXPH64_hash_t;
+#endif
+
+#endif  /* XXPH_NO_LONG_LONG */
+
+
+
+#ifdef XXPH_STATIC_LINKING_ONLY
+
+/* ================================================================================================
+   This section contains declarations which are not guaranteed to remain stable.
+   They may change in future versions, becoming incompatible with a different version of the library.
+   These declarations should only be used with static linking.
+   Never use them in association with dynamic linking !
+=================================================================================================== */
+
+
+/*-**********************************************************************
+*  XXPH3
+*  New experimental hash
+************************************************************************/
+#ifndef XXPH_NO_LONG_LONG
+
+
+/* ============================================
+ * XXPH3 is a new hash algorithm,
+ * featuring improved speed performance for both small and large inputs.
+ * See full speed analysis at : http://fastcompression.blogspot.com/2019/03/presenting-xxh3.html
+ * In general, expect XXPH3 to run about ~2x faster on large inputs,
+ * and >3x faster on small ones, though exact differences depend on platform.
+ *
+ * The algorithm is portable, will generate the same hash on all platforms.
+ * It benefits greatly from vectorization units, but does not require it.
+ *
+ * XXPH3 offers 2 variants, _64bits and _128bits.
+ * When only 64 bits are needed, prefer calling the _64bits variant :
+ * it reduces the amount of mixing, resulting in faster speed on small inputs.
+ * It's also generally simpler to manipulate a scalar return type than a struct.
+ *
+ * The XXPH3 algorithm is still considered experimental.
+ * Produced results can still change between versions.
+ * Results produced by v0.7.x are not comparable with results from v0.7.y .
+ * It's nonetheless possible to use XXPH3 for ephemeral data (local sessions),
+ * but avoid storing values in long-term storage for later reads.
+ *
+ * The API supports one-shot hashing, streaming mode, and custom secrets.
+ *
+ * There are still a number of opened questions that community can influence during the experimental period.
+ * I'm trying to list a few of them below, though don't consider this list as complete.
+ *
+ * - 128-bits output type : currently defined as a structure of two 64-bits fields.
+ *                          That's because 128-bit values do not exist in C standard.
+ *                          Note that it means that, at byte level, result is not identical depending on endianess.
+ *                          However, at field level, they are identical on all platforms.
+ *                          The canonical representation solves the issue of identical byte-level representation across platforms,
+ *                          which is necessary for serialization.
+ *                          Q1 : Would there be a better representation for a 128-bit hash result ?
+ *                          Q2 : Are the names of the inner 64-bit fields important ? Should they be changed ?
+ *
+ * - Prototype XXPH128() :   XXPH128() uses the same arguments as XXPH64(), for consistency.
+ *                          It means it maps to XXPH3_128bits_withSeed().
+ *                          This variant is slightly slower than XXPH3_128bits(),
+ *                          because the seed is now part of the algorithm, and can't be simplified.
+ *                          Is that a good idea ?
+ *
+ * - Seed type for XXPH128() : currently, it's a single 64-bit value, like the 64-bit variant.
+ *                          It could be argued that it's more logical to offer a 128-bit seed input parameter for a 128-bit hash.
+ *                          But 128-bit seed is more difficult to use, since it requires to pass a structure instead of a scalar value.
+ *                          Such a variant could either replace current one, or become an additional one.
+ *                          Farmhash, for example, offers both variants (the 128-bits seed variant is called `doubleSeed`).
+ *                          Follow up question : if both 64-bit and 128-bit seeds are allowed, which variant should be called XXPH128 ?
+ *
+ * - Result for len==0 :    Currently, the result of hashing a zero-length input is always `0`.
+ *                          It seems okay as a return value when using "default" secret and seed.
+ *                          But is it still fine to return `0` when secret or seed are non-default ?
+ *                          Are there use cases which could depend on generating a different hash result for zero-length input when the secret is different ?
+ *
+ * - Consistency (1) :      Streaming XXPH128 uses an XXPH3 state, which is the same state as XXPH3_64bits().
+ *                          It means a 128bit streaming loop must invoke the following symbols :
+ *                          XXPH3_createState(), XXPH3_128bits_reset(), XXPH3_128bits_update() (loop), XXPH3_128bits_digest(), XXPH3_freeState().
+ *                          Is that consistent enough ?
+ *
+ * - Consistency (2) :      The canonical representation of `XXPH3_64bits` is provided by existing functions
+ *                          XXPH64_canonicalFromHash(), and reverse operation XXPH64_hashFromCanonical().
+ *                          As a mirror, canonical functions for XXPH128_hash_t results generated by `XXPH3_128bits`
+ *                          are XXPH128_canonicalFromHash() and XXPH128_hashFromCanonical().
+ *                          Which means, `XXPH3` doesn't appear in the names, because canonical functions operate on a type,
+ *                          independently of which algorithm was used to generate that type.
+ *                          Is that consistent enough ?
+ */
+
+#ifdef XXPH_NAMESPACE
+#  define XXPH3_64bits XXPH_NAME2(XXPH_NAMESPACE, XXPH3_64bits)
+#  define XXPH3_64bits_withSecret XXPH_NAME2(XXPH_NAMESPACE, XXPH3_64bits_withSecret)
+#  define XXPH3_64bits_withSeed XXPH_NAME2(XXPH_NAMESPACE, XXPH3_64bits_withSeed)
+#endif
+
+/* XXPH3_64bits() :
+ * default 64-bit variant, using default secret and default seed of 0.
+ * It's the fastest variant. */
+XXPH_PUBLIC_API XXPH64_hash_t XXPH3_64bits(const void* data, size_t len);
+
+/* XXPH3_64bits_withSecret() :
+ * It's possible to provide any blob of bytes as a "secret" to generate the hash.
+ * This makes it more difficult for an external actor to prepare an intentional collision.
+ * The secret *must* be large enough (>= XXPH3_SECRET_SIZE_MIN).
+ * It should consist of random bytes.
+ * Avoid repeating same character, or sequences of bytes,
+ * and especially avoid swathes of \0.
+ * Failure to respect these conditions will result in a poor quality hash.
+ */
+#define XXPH3_SECRET_SIZE_MIN 136
+XXPH_PUBLIC_API XXPH64_hash_t XXPH3_64bits_withSecret(const void* data, size_t len, const void* secret, size_t secretSize);
+
+/* XXPH3_64bits_withSeed() :
+ * This variant generates on the fly a custom secret,
+ * based on the default secret, altered using the `seed` value.
+ * While this operation is decently fast, note that it's not completely free.
+ * note : seed==0 produces same results as XXPH3_64bits() */
+XXPH_PUBLIC_API XXPH64_hash_t XXPH3_64bits_withSeed(const void* data, size_t len, XXPH64_hash_t seed);
+
+#if defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 201112L)   /* C11+ */
+#  include <stdalign.h>
+#  define XXPH_ALIGN(n)      alignas(n)
+#elif defined(__GNUC__)
+#  define XXPH_ALIGN(n)      __attribute__ ((aligned(n)))
+#elif defined(_MSC_VER)
+#  define XXPH_ALIGN(n)      __declspec(align(n))
+#else
+#  define XXPH_ALIGN(n)   /* disabled */
+#endif
+
+#define XXPH3_SECRET_DEFAULT_SIZE 192   /* minimum XXPH3_SECRET_SIZE_MIN */
+
+#endif  /* XXPH_NO_LONG_LONG */
+
+
+/*-**********************************************************************
+*  XXPH_INLINE_ALL
+************************************************************************/
+#if defined(XXPH_INLINE_ALL) || defined(XXPH_PRIVATE_API)
+
+/* === RocksDB modification: was #include here but permanently inlining === */
+
+typedef struct {
+    XXPH64_hash_t low64;
+    XXPH64_hash_t high64;
+} XXPH128_hash_t;
+
+/* *************************************
+*  Tuning parameters
+***************************************/
+/*!XXPH_FORCE_MEMORY_ACCESS :
+ * By default, access to unaligned memory is controlled by `memcpy()`, which is safe and portable.
+ * Unfortunately, on some target/compiler combinations, the generated assembly is sub-optimal.
+ * The below switch allow to select different access method for improved performance.
+ * Method 0 (default) : use `memcpy()`. Safe and portable.
+ * Method 1 : `__packed` statement. It depends on compiler extension (ie, not portable).
+ *            This method is safe if your compiler supports it, and *generally* as fast or faster than `memcpy`.
+ * Method 2 : direct access. This method doesn't depend on compiler but violate C standard.
+ *            It can generate buggy code on targets which do not support unaligned memory accesses.
+ *            But in some circumstances, it's the only known way to get the most performance (ie GCC + ARMv6)
+ * See http://stackoverflow.com/a/32095106/646947 for details.
+ * Prefer these methods in priority order (0 > 1 > 2)
+ */
+#ifndef XXPH_FORCE_MEMORY_ACCESS   /* can be defined externally, on command line for example */
+#  if !defined(__clang__) && defined(__GNUC__) && defined(__ARM_FEATURE_UNALIGNED) && defined(__ARM_ARCH) && (__ARM_ARCH == 6)
+#    define XXPH_FORCE_MEMORY_ACCESS 2
+#  elif !defined(__clang__) && ((defined(__INTEL_COMPILER) && !defined(_WIN32)) || \
+  (defined(__GNUC__) && (defined(__ARM_ARCH) && __ARM_ARCH >= 7)))
+#    define XXPH_FORCE_MEMORY_ACCESS 1
+#  endif
+#endif
+
+/*!XXPH_ACCEPT_NULL_INPUT_POINTER :
+ * If input pointer is NULL, xxHash default behavior is to dereference it, triggering a segfault.
+ * When this macro is enabled, xxHash actively checks input for null pointer.
+ * It it is, result for null input pointers is the same as a null-length input.
+ */
+#ifndef XXPH_ACCEPT_NULL_INPUT_POINTER   /* can be defined externally */
+#  define XXPH_ACCEPT_NULL_INPUT_POINTER 0
+#endif
+
+/*!XXPH_FORCE_ALIGN_CHECK :
+ * This is a minor performance trick, only useful with lots of very small keys.
+ * It means : check for aligned/unaligned input.
+ * The check costs one initial branch per hash;
+ * set it to 0 when the input is guaranteed to be aligned,
+ * or when alignment doesn't matter for performance.
+ */
+#ifndef XXPH_FORCE_ALIGN_CHECK /* can be defined externally */
+#  if defined(__i386) || defined(_M_IX86) || defined(__x86_64__) || defined(_M_X64)
+#    define XXPH_FORCE_ALIGN_CHECK 0
+#  else
+#    define XXPH_FORCE_ALIGN_CHECK 1
+#  endif
+#endif
+
+/*!XXPH_REROLL:
+ * Whether to reroll XXPH32_finalize, and XXPH64_finalize,
+ * instead of using an unrolled jump table/if statement loop.
+ *
+ * This is automatically defined on -Os/-Oz on GCC and Clang. */
+#ifndef XXPH_REROLL
+#  if defined(__OPTIMIZE_SIZE__)
+#    define XXPH_REROLL 1
+#  else
+#    define XXPH_REROLL 0
+#  endif
+#endif
+
+#include <limits.h>   /* ULLONG_MAX */
+
+#ifndef XXPH_STATIC_LINKING_ONLY
+#define XXPH_STATIC_LINKING_ONLY
+#endif
+
+/* BEGIN RocksDB customizations */
+#include "port/lang.h" /* for FALLTHROUGH_INTENDED, inserted as appropriate */
+/* END RocksDB customizations */
+
+/* *************************************
+*  Compiler Specific Options
+***************************************/
+#ifdef _MSC_VER    /* Visual Studio */
+#  pragma warning(disable : 4127)      /* disable: C4127: conditional expression is constant */
+#  define XXPH_FORCE_INLINE static __forceinline
+#  define XXPH_NO_INLINE static __declspec(noinline)
+#else
+#  if defined (__cplusplus) || defined (__STDC_VERSION__) && __STDC_VERSION__ >= 199901L   /* C99 */
+#    ifdef __GNUC__
+#      define XXPH_FORCE_INLINE static inline __attribute__((always_inline))
+#      define XXPH_NO_INLINE static __attribute__((noinline))
+#    else
+#      define XXPH_FORCE_INLINE static inline
+#      define XXPH_NO_INLINE static
+#    endif
+#  else
+#    define XXPH_FORCE_INLINE static
+#    define XXPH_NO_INLINE static
+#  endif /* __STDC_VERSION__ */
+#endif
+
+
+
+/* *************************************
+*  Debug
+***************************************/
+/* DEBUGLEVEL is expected to be defined externally,
+ * typically through compiler command line.
+ * Value must be a number. */
+#ifndef DEBUGLEVEL
+#  define DEBUGLEVEL 0
+#endif
+
+#if (DEBUGLEVEL>=1)
+#  include <assert.h>   /* note : can still be disabled with NDEBUG */
+#  define XXPH_ASSERT(c)   assert(c)
+#else
+#  define XXPH_ASSERT(c)   ((void)0)
+#endif
+
+/* note : use after variable declarations */
+#define XXPH_STATIC_ASSERT(c)  { enum { XXPH_sa = 1/(int)(!!(c)) }; }
+
+
+/* *************************************
+*  Basic Types
+***************************************/
+#if !defined (__VMS) \
+ && (defined (__cplusplus) \
+ || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */) )
+# include <stdint.h>
+  typedef uint8_t  xxh_u8;
+#else
+  typedef unsigned char      xxh_u8;
+#endif
+typedef XXPH32_hash_t xxh_u32;
+
+
+/* ===   Memory access   === */
+
+#if (defined(XXPH_FORCE_MEMORY_ACCESS) && (XXPH_FORCE_MEMORY_ACCESS==2))
+
+/* Force direct memory access. Only works on CPU which support unaligned memory access in hardware */
+static xxh_u32 XXPH_read32(const void* memPtr) { return *(const xxh_u32*) memPtr; }
+
+#elif (defined(XXPH_FORCE_MEMORY_ACCESS) && (XXPH_FORCE_MEMORY_ACCESS==1))
+
+/* __pack instructions are safer, but compiler specific, hence potentially problematic for some compilers */
+/* currently only defined for gcc and icc */
+typedef union { xxh_u32 u32; } __attribute__((packed)) unalign;
+static xxh_u32 XXPH_read32(const void* ptr) { return ((const unalign*)ptr)->u32; }
+
+#else
+
+/* portable and safe solution. Generally efficient.
+ * see : http://stackoverflow.com/a/32095106/646947
+ */
+static xxh_u32 XXPH_read32(const void* memPtr)
+{
+    xxh_u32 val;
+    memcpy(&val, memPtr, sizeof(val));
+    return val;
+}
+
+#endif   /* XXPH_FORCE_DIRECT_MEMORY_ACCESS */
+
+
+/* ===   Endianess   === */
+
+/* XXPH_CPU_LITTLE_ENDIAN can be defined externally, for example on the compiler command line */
+#ifndef XXPH_CPU_LITTLE_ENDIAN
+#  if defined(_WIN32) /* Windows is always little endian */ \
+     || defined(__LITTLE_ENDIAN__) \
+     || (defined(__BYTE_ORDER__) && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__)
+#    define XXPH_CPU_LITTLE_ENDIAN 1
+#  elif defined(__BIG_ENDIAN__) \
+     || (defined(__BYTE_ORDER__) && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__)
+#    define XXPH_CPU_LITTLE_ENDIAN 0
+#  else
+static int XXPH_isLittleEndian(void)
+{
+    const union { xxh_u32 u; xxh_u8 c[4]; } one = { 1 };   /* don't use static : performance detrimental  */
+    return one.c[0];
+}
+#   define XXPH_CPU_LITTLE_ENDIAN   XXPH_isLittleEndian()
+#  endif
+#endif
+
+
+
+
+/* ****************************************
+*  Compiler-specific Functions and Macros
+******************************************/
+#define XXPH_GCC_VERSION (__GNUC__ * 100 + __GNUC_MINOR__)
+
+#ifndef __has_builtin
+#  define __has_builtin(x) 0
+#endif
+
+#if !defined(NO_CLANG_BUILTIN) && __has_builtin(__builtin_rotateleft32) && __has_builtin(__builtin_rotateleft64)
+#  define XXPH_rotl32 __builtin_rotateleft32
+#  define XXPH_rotl64 __builtin_rotateleft64
+/* Note : although _rotl exists for minGW (GCC under windows), performance seems poor */
+#elif defined(_MSC_VER)
+#  define XXPH_rotl32(x,r) _rotl(x,r)
+#  define XXPH_rotl64(x,r) _rotl64(x,r)
+#else
+#  define XXPH_rotl32(x,r) (((x) << (r)) | ((x) >> (32 - (r))))
+#  define XXPH_rotl64(x,r) (((x) << (r)) | ((x) >> (64 - (r))))
+#endif
+
+#if defined(_MSC_VER)     /* Visual Studio */
+#  define XXPH_swap32 _byteswap_ulong
+#elif XXPH_GCC_VERSION >= 403
+#  define XXPH_swap32 __builtin_bswap32
+#else
+static xxh_u32 XXPH_swap32 (xxh_u32 x)
+{
+    return  ((x << 24) & 0xff000000 ) |
+            ((x <<  8) & 0x00ff0000 ) |
+            ((x >>  8) & 0x0000ff00 ) |
+            ((x >> 24) & 0x000000ff );
+}
+#endif
+
+
+/* ***************************
+*  Memory reads
+*****************************/
+typedef enum { XXPH_aligned, XXPH_unaligned } XXPH_alignment;
+
+XXPH_FORCE_INLINE xxh_u32 XXPH_readLE32(const void* ptr)
+{
+    return XXPH_CPU_LITTLE_ENDIAN ? XXPH_read32(ptr) : XXPH_swap32(XXPH_read32(ptr));
+}
+
+XXPH_FORCE_INLINE xxh_u32
+XXPH_readLE32_align(const void* ptr, XXPH_alignment align)
+{
+    if (align==XXPH_unaligned) {
+        return XXPH_readLE32(ptr);
+    } else {
+        return XXPH_CPU_LITTLE_ENDIAN ? *(const xxh_u32*)ptr : XXPH_swap32(*(const xxh_u32*)ptr);
+    }
+}
+
+
+/* *************************************
+*  Misc
+***************************************/
+XXPH_PUBLIC_API unsigned XXPH_versionNumber (void) { return XXPH_VERSION_NUMBER; }
+
+
+static const xxh_u32 PRIME32_1 = 0x9E3779B1U;   /* 0b10011110001101110111100110110001 */
+static const xxh_u32 PRIME32_2 = 0x85EBCA77U;   /* 0b10000101111010111100101001110111 */
+static const xxh_u32 PRIME32_3 = 0xC2B2AE3DU;   /* 0b11000010101100101010111000111101 */
+static const xxh_u32 PRIME32_4 = 0x27D4EB2FU;   /* 0b00100111110101001110101100101111 */
+static const xxh_u32 PRIME32_5 = 0x165667B1U;   /* 0b00010110010101100110011110110001 */
+
+#ifndef XXPH_NO_LONG_LONG
+
+/* *******************************************************************
+*  64-bit hash functions
+*********************************************************************/
+
+/*======   Memory access   ======*/
+
+typedef XXPH64_hash_t xxh_u64;
+
+#if (defined(XXPH_FORCE_MEMORY_ACCESS) && (XXPH_FORCE_MEMORY_ACCESS==2))
+
+/* Force direct memory access. Only works on CPU which support unaligned memory access in hardware */
+static xxh_u64 XXPH_read64(const void* memPtr) { return *(const xxh_u64*) memPtr; }
+
+#elif (defined(XXPH_FORCE_MEMORY_ACCESS) && (XXPH_FORCE_MEMORY_ACCESS==1))
+
+/* __pack instructions are safer, but compiler specific, hence potentially problematic for some compilers */
+/* currently only defined for gcc and icc */
+typedef union { xxh_u32 u32; xxh_u64 u64; } __attribute__((packed)) unalign64;
+static xxh_u64 XXPH_read64(const void* ptr) { return ((const unalign64*)ptr)->u64; }
+
+#else
+
+/* portable and safe solution. Generally efficient.
+ * see : http://stackoverflow.com/a/32095106/646947
+ */
+
+static xxh_u64 XXPH_read64(const void* memPtr)
+{
+    xxh_u64 val;
+    memcpy(&val, memPtr, sizeof(val));
+    return val;
+}
+
+#endif   /* XXPH_FORCE_DIRECT_MEMORY_ACCESS */
+
+#if defined(_MSC_VER)     /* Visual Studio */
+#  define XXPH_swap64 _byteswap_uint64
+#elif XXPH_GCC_VERSION >= 403
+#  define XXPH_swap64 __builtin_bswap64
+#else
+static xxh_u64 XXPH_swap64 (xxh_u64 x)
+{
+    return  ((x << 56) & 0xff00000000000000ULL) |
+            ((x << 40) & 0x00ff000000000000ULL) |
+            ((x << 24) & 0x0000ff0000000000ULL) |
+            ((x << 8)  & 0x000000ff00000000ULL) |
+            ((x >> 8)  & 0x00000000ff000000ULL) |
+            ((x >> 24) & 0x0000000000ff0000ULL) |
+            ((x >> 40) & 0x000000000000ff00ULL) |
+            ((x >> 56) & 0x00000000000000ffULL);
+}
+#endif
+
+XXPH_FORCE_INLINE xxh_u64 XXPH_readLE64(const void* ptr)
+{
+    return XXPH_CPU_LITTLE_ENDIAN ? XXPH_read64(ptr) : XXPH_swap64(XXPH_read64(ptr));
+}
+
+XXPH_FORCE_INLINE xxh_u64
+XXPH_readLE64_align(const void* ptr, XXPH_alignment align)
+{
+    if (align==XXPH_unaligned)
+        return XXPH_readLE64(ptr);
+    else
+        return XXPH_CPU_LITTLE_ENDIAN ? *(const xxh_u64*)ptr : XXPH_swap64(*(const xxh_u64*)ptr);
+}
+
+
+/*======   xxh64   ======*/
+
+static const xxh_u64 PRIME64_1 = 0x9E3779B185EBCA87ULL;   /* 0b1001111000110111011110011011000110000101111010111100101010000111 */
+static const xxh_u64 PRIME64_2 = 0xC2B2AE3D27D4EB4FULL;   /* 0b1100001010110010101011100011110100100111110101001110101101001111 */
+static const xxh_u64 PRIME64_3 = 0x165667B19E3779F9ULL;   /* 0b0001011001010110011001111011000110011110001101110111100111111001 */
+static const xxh_u64 PRIME64_4 = 0x85EBCA77C2B2AE63ULL;   /* 0b1000010111101011110010100111011111000010101100101010111001100011 */
+static const xxh_u64 PRIME64_5 = 0x27D4EB2F165667C5ULL;   /* 0b0010011111010100111010110010111100010110010101100110011111000101 */
+
+
+/* *********************************************************************
+*  XXPH3
+*  New generation hash designed for speed on small keys and vectorization
+************************************************************************ */
+
+/*======== Was #include "xxh3.h", now inlined below ==========*/
+
+/*
+   xxHash - Extremely Fast Hash algorithm
+   Development source file for `xxh3`
+   Copyright (C) 2019-present, Yann Collet.
+
+   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions are
+   met:
+
+       * Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+       * Redistributions in binary form must reproduce the above
+   copyright notice, this list of conditions and the following disclaimer
+   in the documentation and/or other materials provided with the
+   distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+   You can contact the author at :
+   - xxHash source repository : https://github.com/Cyan4973/xxHash
+*/
+
+/* RocksDB Note: This file contains a preview release (xxhash repository
+   version 0.7.2) of XXPH3 that is unlikely to be compatible with the final
+   version of XXPH3. We have therefore renamed this XXPH3 ("preview"), for
+   clarity so that we can continue to use this version even after
+   integrating a newer incompatible version.
+*/
+
+/* ===   Dependencies   === */
+
+#undef XXPH_INLINE_ALL   /* in case it's already defined */
+#define XXPH_INLINE_ALL
+
+
+/* ===   Compiler specifics   === */
+
+#if defined (__STDC_VERSION__) && __STDC_VERSION__ >= 199901L   /* >= C99 */
+#  define XXPH_RESTRICT   restrict
+#else
+/* note : it might be useful to define __restrict or __restrict__ for some C++ compilers */
+#  define XXPH_RESTRICT   /* disable */
+#endif
+
+#if defined(__GNUC__)
+#  if defined(__AVX2__)
+#    include <immintrin.h>
+#  elif defined(__SSE2__)
+#    include <emmintrin.h>
+#  elif defined(__ARM_NEON__) || defined(__ARM_NEON)
+#    define inline __inline__  /* clang bug */
+#    include <arm_neon.h>
+#    undef inline
+#  endif
+#elif defined(_MSC_VER)
+#  include <intrin.h>
+#endif
+
+/*
+ * Sanity check.
+ *
+ * XXPH3 only requires these features to be efficient:
+ *
+ *  - Usable unaligned access
+ *  - A 32-bit or 64-bit ALU
+ *      - If 32-bit, a decent ADC instruction
+ *  - A 32 or 64-bit multiply with a 64-bit result
+ *
+ * Almost all 32-bit and 64-bit targets meet this, except for Thumb-1, the
+ * classic 16-bit only subset of ARM's instruction set.
+ *
+ * First of all, Thumb-1 lacks support for the UMULL instruction which
+ * performs the important long multiply. This means numerous __aeabi_lmul
+ * calls.
+ *
+ * Second of all, the 8 functional registers are just not enough.
+ * Setup for __aeabi_lmul, byteshift loads, pointers, and all arithmetic need
+ * Lo registers, and this shuffling results in thousands more MOVs than A32.
+ *
+ * A32 and T32 don't have this limitation. They can access all 14 registers,
+ * do a 32->64 multiply with UMULL, and the flexible operand is helpful too.
+ *
+ * If compiling Thumb-1 for a target which supports ARM instructions, we
+ * will give a warning.
+ *
+ * Usually, if this happens, it is because of an accident and you probably
+ * need to specify -march, as you probably meant to compileh for a newer
+ * architecture.
+ */
+#if defined(__thumb__) && !defined(__thumb2__) && defined(__ARM_ARCH_ISA_ARM)
+#   warning "XXPH3 is highly inefficient without ARM or Thumb-2."
+#endif
+
+/* ==========================================
+ * Vectorization detection
+ * ========================================== */
+#define XXPH_SCALAR 0
+#define XXPH_SSE2   1
+#define XXPH_AVX2   2
+#define XXPH_NEON   3
+#define XXPH_VSX    4
+
+#ifndef XXPH_VECTOR    /* can be defined on command line */
+#  if defined(__AVX2__)
+#    define XXPH_VECTOR XXPH_AVX2
+#  elif defined(__SSE2__) || defined(_M_AMD64) || defined(_M_X64) || (defined(_M_IX86_FP) && (_M_IX86_FP == 2))
+#    define XXPH_VECTOR XXPH_SSE2
+#  elif defined(__GNUC__) /* msvc support maybe later */ \
+  && (defined(__ARM_NEON__) || defined(__ARM_NEON)) \
+  && (defined(__LITTLE_ENDIAN__) /* We only support little endian NEON */ \
+    || (defined(__BYTE_ORDER__) && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__))
+#    define XXPH_VECTOR XXPH_NEON
+#  elif defined(__PPC64__) && defined(__POWER8_VECTOR__) && defined(__GNUC__)
+#    define XXPH_VECTOR XXPH_VSX
+#  else
+#    define XXPH_VECTOR XXPH_SCALAR
+#  endif
+#endif
+
+/* control alignment of accumulator,
+ * for compatibility with fast vector loads */
+#ifndef XXPH_ACC_ALIGN
+#  if XXPH_VECTOR == 0   /* scalar */
+#     define XXPH_ACC_ALIGN 8
+#  elif XXPH_VECTOR == 1  /* sse2 */
+#     define XXPH_ACC_ALIGN 16
+#  elif XXPH_VECTOR == 2  /* avx2 */
+#     define XXPH_ACC_ALIGN 32
+#  elif XXPH_VECTOR == 3  /* neon */
+#     define XXPH_ACC_ALIGN 16
+#  elif XXPH_VECTOR == 4  /* vsx */
+#     define XXPH_ACC_ALIGN 16
+#  endif
+#endif
+
+/* xxh_u64 XXPH_mult32to64(xxh_u32 a, xxh_u64 b) { return (xxh_u64)a * (xxh_u64)b; } */
+#if defined(_MSC_VER) && defined(_M_IX86)
+#    include <intrin.h>
+#    define XXPH_mult32to64(x, y) __emulu(x, y)
+#else
+#    define XXPH_mult32to64(x, y) ((xxh_u64)((x) & 0xFFFFFFFF) * (xxh_u64)((y) & 0xFFFFFFFF))
+#endif
+
+/* VSX stuff. It's a lot because VSX support is mediocre across compilers and
+ * there is a lot of mischief with endianness. */
+#if XXPH_VECTOR == XXPH_VSX
+#  include <altivec.h>
+#  undef vector
+typedef __vector unsigned long long U64x2;
+typedef __vector unsigned char U8x16;
+typedef __vector unsigned U32x4;
+
+#ifndef XXPH_VSX_BE
+#  if defined(__BIG_ENDIAN__) \
+  || (defined(__BYTE_ORDER__) && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__)
+#    define XXPH_VSX_BE 1
+#  elif defined(__VEC_ELEMENT_REG_ORDER__) && __VEC_ELEMENT_REG_ORDER__ == __ORDER_BIG_ENDIAN__
+#    warning "-maltivec=be is not recommended. Please use native endianness."
+#    define XXPH_VSX_BE 1
+#  else
+#    define XXPH_VSX_BE 0
+#  endif
+#endif
+
+/* We need some helpers for big endian mode. */
+#if XXPH_VSX_BE
+/* A wrapper for POWER9's vec_revb. */
+#  ifdef __POWER9_VECTOR__
+#    define XXPH_vec_revb vec_revb
+#  else
+XXPH_FORCE_INLINE U64x2 XXPH_vec_revb(U64x2 val)
+{
+    U8x16 const vByteSwap = { 0x07, 0x06, 0x05, 0x04, 0x03, 0x02, 0x01, 0x00,
+                              0x0F, 0x0E, 0x0D, 0x0C, 0x0B, 0x0A, 0x09, 0x08 };
+    return vec_perm(val, val, vByteSwap);
+}
+#  endif
+
+/* Power8 Crypto gives us vpermxor which is very handy for
+ * PPC64EB.
+ *
+ * U8x16 vpermxor(U8x16 a, U8x16 b, U8x16 mask)
+ * {
+ *     U8x16 ret;
+ *     for (int i = 0; i < 16; i++) {
+ *         ret[i] = a[mask[i] & 0xF] ^ b[mask[i] >> 4];
+ *     }
+ *     return ret;
+ * }
+ *
+ * Because both of the main loops load the key, swap, and xor it with input,
+ * we can combine the key swap into this instruction.
+ */
+#  ifdef vec_permxor
+#    define XXPH_vec_permxor vec_permxor
+#  else
+#    define XXPH_vec_permxor __builtin_crypto_vpermxor
+#  endif
+#endif  /* XXPH_VSX_BE */
+/*
+ * Because we reinterpret the multiply, there are endian memes: vec_mulo actually becomes
+ * vec_mule.
+ *
+ * Additionally, the intrinsic wasn't added until GCC 8, despite existing for a while.
+ * Clang has an easy way to control this, we can just use the builtin which doesn't swap.
+ * GCC needs inline assembly. */
+#if __has_builtin(__builtin_altivec_vmuleuw)
+#  define XXPH_vec_mulo __builtin_altivec_vmulouw
+#  define XXPH_vec_mule __builtin_altivec_vmuleuw
+#else
+/* Adapted from https://github.com/google/highwayhash/blob/master/highwayhash/hh_vsx.h. */
+XXPH_FORCE_INLINE U64x2 XXPH_vec_mulo(U32x4 a, U32x4 b) {
+    U64x2 result;
+    __asm__("vmulouw %0, %1, %2" : "=v" (result) : "v" (a), "v" (b));
+    return result;
+}
+XXPH_FORCE_INLINE U64x2 XXPH_vec_mule(U32x4 a, U32x4 b) {
+    U64x2 result;
+    __asm__("vmuleuw %0, %1, %2" : "=v" (result) : "v" (a), "v" (b));
+    return result;
+}
+#endif /* __has_builtin(__builtin_altivec_vmuleuw) */
+#endif /* XXPH_VECTOR == XXPH_VSX */
+
+/* prefetch
+ * can be disabled, by declaring XXPH_NO_PREFETCH build macro */
+#if defined(XXPH_NO_PREFETCH)
+#  define XXPH_PREFETCH(ptr)  (void)(ptr)  /* disabled */
+#else
+#if defined(_MSC_VER) && \
+    (defined(_M_X64) ||  \
+     defined(_M_IX86)) /* _mm_prefetch() is not defined outside of x86/x64 */
+#    include <mmintrin.h>   /* https://msdn.microsoft.com/fr-fr/library/84szxsww(v=vs.90).aspx */
+#    define XXPH_PREFETCH(ptr)  _mm_prefetch((const char*)(ptr), _MM_HINT_T0)
+#  elif defined(__GNUC__) && ( (__GNUC__ >= 4) || ( (__GNUC__ == 3) && (__GNUC_MINOR__ >= 1) ) )
+#    define XXPH_PREFETCH(ptr)  __builtin_prefetch((ptr), 0 /* rw==read */, 3 /* locality */)
+#  else
+#    define XXPH_PREFETCH(ptr) (void)(ptr)  /* disabled */
+#  endif
+#endif  /* XXPH_NO_PREFETCH */
+
+
+/* ==========================================
+ * XXPH3 default settings
+ * ========================================== */
+
+#define XXPH_SECRET_DEFAULT_SIZE 192   /* minimum XXPH3_SECRET_SIZE_MIN */
+
+#if (XXPH_SECRET_DEFAULT_SIZE < XXPH3_SECRET_SIZE_MIN)
+#  error "default keyset is not large enough"
+#endif
+
+XXPH_ALIGN(64) static const xxh_u8 kSecret[XXPH_SECRET_DEFAULT_SIZE] = {
+    0xb8, 0xfe, 0x6c, 0x39, 0x23, 0xa4, 0x4b, 0xbe, 0x7c, 0x01, 0x81, 0x2c, 0xf7, 0x21, 0xad, 0x1c,
+    0xde, 0xd4, 0x6d, 0xe9, 0x83, 0x90, 0x97, 0xdb, 0x72, 0x40, 0xa4, 0xa4, 0xb7, 0xb3, 0x67, 0x1f,
+    0xcb, 0x79, 0xe6, 0x4e, 0xcc, 0xc0, 0xe5, 0x78, 0x82, 0x5a, 0xd0, 0x7d, 0xcc, 0xff, 0x72, 0x21,
+    0xb8, 0x08, 0x46, 0x74, 0xf7, 0x43, 0x24, 0x8e, 0xe0, 0x35, 0x90, 0xe6, 0x81, 0x3a, 0x26, 0x4c,
+    0x3c, 0x28, 0x52, 0xbb, 0x91, 0xc3, 0x00, 0xcb, 0x88, 0xd0, 0x65, 0x8b, 0x1b, 0x53, 0x2e, 0xa3,
+    0x71, 0x64, 0x48, 0x97, 0xa2, 0x0d, 0xf9, 0x4e, 0x38, 0x19, 0xef, 0x46, 0xa9, 0xde, 0xac, 0xd8,
+    0xa8, 0xfa, 0x76, 0x3f, 0xe3, 0x9c, 0x34, 0x3f, 0xf9, 0xdc, 0xbb, 0xc7, 0xc7, 0x0b, 0x4f, 0x1d,
+    0x8a, 0x51, 0xe0, 0x4b, 0xcd, 0xb4, 0x59, 0x31, 0xc8, 0x9f, 0x7e, 0xc9, 0xd9, 0x78, 0x73, 0x64,
+
+    0xea, 0xc5, 0xac, 0x83, 0x34, 0xd3, 0xeb, 0xc3, 0xc5, 0x81, 0xa0, 0xff, 0xfa, 0x13, 0x63, 0xeb,
+    0x17, 0x0d, 0xdd, 0x51, 0xb7, 0xf0, 0xda, 0x49, 0xd3, 0x16, 0x55, 0x26, 0x29, 0xd4, 0x68, 0x9e,
+    0x2b, 0x16, 0xbe, 0x58, 0x7d, 0x47, 0xa1, 0xfc, 0x8f, 0xf8, 0xb8, 0xd1, 0x7a, 0xd0, 0x31, 0xce,
+    0x45, 0xcb, 0x3a, 0x8f, 0x95, 0x16, 0x04, 0x28, 0xaf, 0xd7, 0xfb, 0xca, 0xbb, 0x4b, 0x40, 0x7e,
+};
+
+/*
+ * GCC for x86 has a tendency to use SSE in this loop. While it
+ * successfully avoids swapping (as MUL overwrites EAX and EDX), it
+ * slows it down because instead of free register swap shifts, it
+ * must use pshufd and punpckl/hd.
+ *
+ * To prevent this, we use this attribute to shut off SSE.
+ */
+#if defined(__GNUC__) && !defined(__clang__) && defined(__i386__)
+__attribute__((__target__("no-sse")))
+#endif
+static XXPH128_hash_t
+XXPH_mult64to128(xxh_u64 lhs, xxh_u64 rhs)
+{
+    /*
+     * GCC/Clang __uint128_t method.
+     *
+     * On most 64-bit targets, GCC and Clang define a __uint128_t type.
+     * This is usually the best way as it usually uses a native long 64-bit
+     * multiply, such as MULQ on x86_64 or MUL + UMULH on aarch64.
+     *
+     * Usually.
+     *
+     * Despite being a 32-bit platform, Clang (and emscripten) define this
+     * type despite not having the arithmetic for it. This results in a
+     * laggy compiler builtin call which calculates a full 128-bit multiply.
+     * In that case it is best to use the portable one.
+     * https://github.com/Cyan4973/xxHash/issues/211#issuecomment-515575677
+     */
+#if defined(__GNUC__) && !defined(__wasm__) \
+    && defined(__SIZEOF_INT128__) \
+    || (defined(_INTEGRAL_MAX_BITS) && _INTEGRAL_MAX_BITS >= 128)
+
+    __uint128_t product = (__uint128_t)lhs * (__uint128_t)rhs;
+    XXPH128_hash_t const r128 = { (xxh_u64)(product), (xxh_u64)(product >> 64) };
+    return r128;
+
+    /*
+     * MSVC for x64's _umul128 method.
+     *
+     * xxh_u64 _umul128(xxh_u64 Multiplier, xxh_u64 Multiplicand, xxh_u64 *HighProduct);
+     *
+     * This compiles to single operand MUL on x64.
+     */
+#elif defined(_M_X64) || defined(_M_IA64)
+
+#ifndef _MSC_VER
+#   pragma intrinsic(_umul128)
+#endif
+    xxh_u64 product_high;
+    xxh_u64 const product_low = _umul128(lhs, rhs, &product_high);
+    XXPH128_hash_t const r128 = { product_low, product_high };
+    return r128;
+
+#else
+    /*
+     * Portable scalar method. Optimized for 32-bit and 64-bit ALUs.
+     *
+     * This is a fast and simple grade school multiply, which is shown
+     * below with base 10 arithmetic instead of base 0x100000000.
+     *
+     *           9 3 // D2 lhs = 93
+     *         x 7 5 // D2 rhs = 75
+     *     ----------
+     *           1 5 // D2 lo_lo = (93 % 10) * (75 % 10)
+     *         4 5 | // D2 hi_lo = (93 / 10) * (75 % 10)
+     *         2 1 | // D2 lo_hi = (93 % 10) * (75 / 10)
+     *     + 6 3 | | // D2 hi_hi = (93 / 10) * (75 / 10)
+     *     ---------
+     *         2 7 | // D2 cross  = (15 / 10) + (45 % 10) + 21
+     *     + 6 7 | | // D2 upper  = (27 / 10) + (45 / 10) + 63
+     *     ---------
+     *       6 9 7 5
+     *
+     * The reasons for adding the products like this are:
+     *  1. It avoids manual carry tracking. Just like how
+     *     (9 * 9) + 9 + 9 = 99, the same applies with this for
+     *     UINT64_MAX. This avoids a lot of complexity.
+     *
+     *  2. It hints for, and on Clang, compiles to, the powerful UMAAL
+     *     instruction available in ARMv6+ A32/T32, which is shown below:
+     *
+     *         void UMAAL(xxh_u32 *RdLo, xxh_u32 *RdHi, xxh_u32 Rn, xxh_u32 Rm)
+     *         {
+     *             xxh_u64 product = (xxh_u64)*RdLo * (xxh_u64)*RdHi + Rn + Rm;
+     *             *RdLo = (xxh_u32)(product & 0xFFFFFFFF);
+     *             *RdHi = (xxh_u32)(product >> 32);
+     *         }
+     *
+     *     This instruction was designed for efficient long multiplication,
+     *     and allows this to be calculated in only 4 instructions which
+     *     is comparable to some 64-bit ALUs.
+     *
+     *  3. It isn't terrible on other platforms. Usually this will be
+     *     a couple of 32-bit ADD/ADCs.
+     */
+
+    /* First calculate all of the cross products. */
+    xxh_u64 const lo_lo = XXPH_mult32to64(lhs & 0xFFFFFFFF, rhs & 0xFFFFFFFF);
+    xxh_u64 const hi_lo = XXPH_mult32to64(lhs >> 32,        rhs & 0xFFFFFFFF);
+    xxh_u64 const lo_hi = XXPH_mult32to64(lhs & 0xFFFFFFFF, rhs >> 32);
+    xxh_u64 const hi_hi = XXPH_mult32to64(lhs >> 32,        rhs >> 32);
+
+    /* Now add the products together. These will never overflow. */
+    xxh_u64 const cross = (lo_lo >> 32) + (hi_lo & 0xFFFFFFFF) + lo_hi;
+    xxh_u64 const upper = (hi_lo >> 32) + (cross >> 32)        + hi_hi;
+    xxh_u64 const lower = (cross << 32) | (lo_lo & 0xFFFFFFFF);
+
+    XXPH128_hash_t r128 = { lower, upper };
+    return r128;
+#endif
+}
+
+/*
+ * We want to keep the attribute here because a target switch
+ * disables inlining.
+ *
+ * Does a 64-bit to 128-bit multiply, then XOR folds it.
+ * The reason for the separate function is to prevent passing
+ * too many structs around by value. This will hopefully inline
+ * the multiply, but we don't force it.
+ */
+#if defined(__GNUC__) && !defined(__clang__) && defined(__i386__)
+__attribute__((__target__("no-sse")))
+#endif
+static xxh_u64
+XXPH3_mul128_fold64(xxh_u64 lhs, xxh_u64 rhs)
+{
+    XXPH128_hash_t product = XXPH_mult64to128(lhs, rhs);
+    return product.low64 ^ product.high64;
+}
+
+
+static XXPH64_hash_t XXPH3_avalanche(xxh_u64 h64)
+{
+    h64 ^= h64 >> 37;
+    h64 *= PRIME64_3;
+    h64 ^= h64 >> 32;
+    return h64;
+}
+
+
+/* ==========================================
+ * Short keys
+ * ========================================== */
+
+XXPH_FORCE_INLINE XXPH64_hash_t
+XXPH3_len_1to3_64b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXPH64_hash_t seed)
+{
+    XXPH_ASSERT(input != NULL);
+    XXPH_ASSERT(1 <= len && len <= 3);
+    XXPH_ASSERT(secret != NULL);
+    {   xxh_u8 const c1 = input[0];
+        xxh_u8 const c2 = input[len >> 1];
+        xxh_u8 const c3 = input[len - 1];
+        xxh_u32  const combined = ((xxh_u32)c1) | (((xxh_u32)c2) << 8) | (((xxh_u32)c3) << 16) | (((xxh_u32)len) << 24);
+        xxh_u64  const keyed = (xxh_u64)combined ^ (XXPH_readLE32(secret) + seed);
+        xxh_u64  const mixed = keyed * PRIME64_1;
+        return XXPH3_avalanche(mixed);
+    }
+}
+
+XXPH_FORCE_INLINE XXPH64_hash_t
+XXPH3_len_4to8_64b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXPH64_hash_t seed)
+{
+    XXPH_ASSERT(input != NULL);
+    XXPH_ASSERT(secret != NULL);
+    XXPH_ASSERT(4 <= len && len <= 8);
+    {   xxh_u32 const input_lo = XXPH_readLE32(input);
+        xxh_u32 const input_hi = XXPH_readLE32(input + len - 4);
+        xxh_u64 const input_64 = input_lo | ((xxh_u64)input_hi << 32);
+        xxh_u64 const keyed = input_64 ^ (XXPH_readLE64(secret) + seed);
+        xxh_u64 const mix64 = len + ((keyed ^ (keyed >> 51)) * PRIME32_1);
+        return XXPH3_avalanche((mix64 ^ (mix64 >> 47)) * PRIME64_2);
+    }
+}
+
+XXPH_FORCE_INLINE XXPH64_hash_t
+XXPH3_len_9to16_64b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXPH64_hash_t seed)
+{
+    XXPH_ASSERT(input != NULL);
+    XXPH_ASSERT(secret != NULL);
+    XXPH_ASSERT(9 <= len && len <= 16);
+    {   xxh_u64 const input_lo = XXPH_readLE64(input)           ^ (XXPH_readLE64(secret)     + seed);
+        xxh_u64 const input_hi = XXPH_readLE64(input + len - 8) ^ (XXPH_readLE64(secret + 8) - seed);
+        xxh_u64 const acc = len + (input_lo + input_hi) + XXPH3_mul128_fold64(input_lo, input_hi);
+        return XXPH3_avalanche(acc);
+    }
+}
+
+XXPH_FORCE_INLINE XXPH64_hash_t
+XXPH3_len_0to16_64b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXPH64_hash_t seed)
+{
+    XXPH_ASSERT(len <= 16);
+    {   if (len > 8) return XXPH3_len_9to16_64b(input, len, secret, seed);
+        if (len >= 4) return XXPH3_len_4to8_64b(input, len, secret, seed);
+        if (len) return XXPH3_len_1to3_64b(input, len, secret, seed);
+        /*
+         * RocksDB modification from XXPH3 preview: zero result for empty
+         * string can be problematic for multiplication-based algorithms.
+         * Return a hash of the seed instead.
+         */
+        return XXPH3_mul128_fold64(seed + XXPH_readLE64(secret), PRIME64_2);
+    }
+}
+
+
+/* ===    Long Keys    === */
+
+#define STRIPE_LEN 64
+#define XXPH_SECRET_CONSUME_RATE 8   /* nb of secret bytes consumed at each accumulation */
+#define ACC_NB (STRIPE_LEN / sizeof(xxh_u64))
+
+typedef enum { XXPH3_acc_64bits, XXPH3_acc_128bits } XXPH3_accWidth_e;
+
+XXPH_FORCE_INLINE void
+XXPH3_accumulate_512(      void* XXPH_RESTRICT acc,
+                    const void* XXPH_RESTRICT input,
+                    const void* XXPH_RESTRICT secret,
+                    XXPH3_accWidth_e accWidth)
+{
+#if (XXPH_VECTOR == XXPH_AVX2)
+
+    XXPH_ASSERT((((size_t)acc) & 31) == 0);
+    {   XXPH_ALIGN(32) __m256i* const xacc  =       (__m256i *) acc;
+        const         __m256i* const xinput = (const __m256i *) input;  /* not really aligned, just for ptr arithmetic, and because _mm256_loadu_si256() requires this type */
+        const         __m256i* const xsecret = (const __m256i *) secret;   /* not really aligned, just for ptr arithmetic, and because _mm256_loadu_si256() requires this type */
+
+        size_t i;
+        for (i=0; i < STRIPE_LEN/sizeof(__m256i); i++) {
+            __m256i const data_vec = _mm256_loadu_si256 (xinput+i);
+            __m256i const key_vec = _mm256_loadu_si256 (xsecret+i);
+            __m256i const data_key = _mm256_xor_si256 (data_vec, key_vec);                                  /* uint32 dk[8]  = {d0+k0, d1+k1, d2+k2, d3+k3, ...} */
+            __m256i const product = _mm256_mul_epu32 (data_key, _mm256_shuffle_epi32 (data_key, 0x31));  /* uint64 mul[4] = {dk0*dk1, dk2*dk3, ...} */
+            if (accWidth == XXPH3_acc_128bits) {
+                __m256i const data_swap = _mm256_shuffle_epi32(data_vec, _MM_SHUFFLE(1,0,3,2));
+                __m256i const sum = _mm256_add_epi64(xacc[i], data_swap);
+                xacc[i]  = _mm256_add_epi64(product, sum);
+            } else {  /* XXPH3_acc_64bits */
+                __m256i const sum = _mm256_add_epi64(xacc[i], data_vec);
+                xacc[i]  = _mm256_add_epi64(product, sum);
+            }
+    }   }
+
+#elif (XXPH_VECTOR == XXPH_SSE2)
+
+    XXPH_ASSERT((((size_t)acc) & 15) == 0);
+    {   XXPH_ALIGN(16) __m128i* const xacc  =       (__m128i *) acc;
+        const         __m128i* const xinput = (const __m128i *) input;  /* not really aligned, just for ptr arithmetic, and because _mm_loadu_si128() requires this type */
+        const         __m128i* const xsecret = (const __m128i *) secret;   /* not really aligned, just for ptr arithmetic, and because _mm_loadu_si128() requires this type */
+
+        size_t i;
+        for (i=0; i < STRIPE_LEN/sizeof(__m128i); i++) {
+            __m128i const data_vec = _mm_loadu_si128 (xinput+i);
+            __m128i const key_vec = _mm_loadu_si128 (xsecret+i);
+            __m128i const data_key = _mm_xor_si128 (data_vec, key_vec);                                  /* uint32 dk[8]  = {d0+k0, d1+k1, d2+k2, d3+k3, ...} */
+            __m128i const product = _mm_mul_epu32 (data_key, _mm_shuffle_epi32 (data_key, 0x31));  /* uint64 mul[4] = {dk0*dk1, dk2*dk3, ...} */
+            if (accWidth == XXPH3_acc_128bits) {
+                __m128i const data_swap = _mm_shuffle_epi32(data_vec, _MM_SHUFFLE(1,0,3,2));
+                __m128i const sum = _mm_add_epi64(xacc[i], data_swap);
+                xacc[i]  = _mm_add_epi64(product, sum);
+            } else {  /* XXPH3_acc_64bits */
+                __m128i const sum = _mm_add_epi64(xacc[i], data_vec);
+                xacc[i]  = _mm_add_epi64(product, sum);
+            }
+    }   }
+
+#elif (XXPH_VECTOR == XXPH_NEON)
+
+    XXPH_ASSERT((((size_t)acc) & 15) == 0);
+    {
+        XXPH_ALIGN(16) uint64x2_t* const xacc = (uint64x2_t *) acc;
+        /* We don't use a uint32x4_t pointer because it causes bus errors on ARMv7. */
+        uint8_t const* const xinput = (const uint8_t *) input;
+        uint8_t const* const xsecret  = (const uint8_t *) secret;
+
+        size_t i;
+        for (i=0; i < STRIPE_LEN / sizeof(uint64x2_t); i++) {
+#if !defined(__aarch64__) && !defined(__arm64__) && defined(__GNUC__) /* ARM32-specific hack */
+            /* vzip on ARMv7 Clang generates a lot of vmovs (technically vorrs) without this.
+             * vzip on 32-bit ARM NEON will overwrite the original register, and I think that Clang
+             * assumes I don't want to destroy it and tries to make a copy. This slows down the code
+             * a lot.
+             * aarch64 not only uses an entirely different syntax, but it requires three
+             * instructions...
+             *    ext    v1.16B, v0.16B, #8    // select high bits because aarch64 can't address them directly
+             *    zip1   v3.2s, v0.2s, v1.2s   // first zip
+             *    zip2   v2.2s, v0.2s, v1.2s   // second zip
+             * ...to do what ARM does in one:
+             *    vzip.32 d0, d1               // Interleave high and low bits and overwrite. */
+
+            /* data_vec = xsecret[i]; */
+            uint8x16_t const data_vec    = vld1q_u8(xinput + (i * 16));
+            /* key_vec  = xsecret[i];  */
+            uint8x16_t const key_vec     = vld1q_u8(xsecret  + (i * 16));
+            /* data_key = data_vec ^ key_vec; */
+            uint32x4_t       data_key;
+
+            if (accWidth == XXPH3_acc_64bits) {
+                /* Add first to prevent register swaps */
+                /* xacc[i] += data_vec; */
+                xacc[i] = vaddq_u64 (xacc[i], vreinterpretq_u64_u8(data_vec));
+            } else {  /* XXPH3_acc_128bits */
+                /* xacc[i] += swap(data_vec); */
+                /* can probably be optimized better */
+                uint64x2_t const data64 = vreinterpretq_u64_u8(data_vec);
+                uint64x2_t const swapped= vextq_u64(data64, data64, 1);
+                xacc[i] = vaddq_u64 (xacc[i], swapped);
+            }
+
+            data_key = vreinterpretq_u32_u8(veorq_u8(data_vec, key_vec));
+
+            /* Here's the magic. We use the quirkiness of vzip to shuffle data_key in place.
+             * shuffle: data_key[0, 1, 2, 3] = data_key[0, 2, 1, 3] */
+            __asm__("vzip.32 %e0, %f0" : "+w" (data_key));
+            /* xacc[i] += (uint64x2_t) data_key[0, 1] * (uint64x2_t) data_key[2, 3]; */
+            xacc[i] = vmlal_u32(xacc[i], vget_low_u32(data_key), vget_high_u32(data_key));
+
+#else
+            /* On aarch64, vshrn/vmovn seems to be equivalent to, if not faster than, the vzip method. */
+
+            /* data_vec = xsecret[i]; */
+            uint8x16_t const data_vec    = vld1q_u8(xinput + (i * 16));
+            /* key_vec  = xsecret[i];  */
+            uint8x16_t const key_vec     = vld1q_u8(xsecret  + (i * 16));
+            /* data_key = data_vec ^ key_vec; */
+            uint64x2_t const data_key    = vreinterpretq_u64_u8(veorq_u8(data_vec, key_vec));
+            /* data_key_lo = (uint32x2_t) (data_key & 0xFFFFFFFF); */
+            uint32x2_t const data_key_lo = vmovn_u64  (data_key);
+            /* data_key_hi = (uint32x2_t) (data_key >> 32); */
+            uint32x2_t const data_key_hi = vshrn_n_u64 (data_key, 32);
+            if (accWidth == XXPH3_acc_64bits) {
+                /* xacc[i] += data_vec; */
+                xacc[i] = vaddq_u64 (xacc[i], vreinterpretq_u64_u8(data_vec));
+            } else {  /* XXPH3_acc_128bits */
+                /* xacc[i] += swap(data_vec); */
+                uint64x2_t const data64 = vreinterpretq_u64_u8(data_vec);
+                uint64x2_t const swapped= vextq_u64(data64, data64, 1);
+                xacc[i] = vaddq_u64 (xacc[i], swapped);
+            }
+            /* xacc[i] += (uint64x2_t) data_key_lo * (uint64x2_t) data_key_hi; */
+            xacc[i] = vmlal_u32 (xacc[i], data_key_lo, data_key_hi);
+
+#endif
+        }
+    }
+
+#elif (XXPH_VECTOR == XXPH_VSX) && /* work around a compiler bug */ (__GNUC__ > 5)
+          U64x2* const xacc =        (U64x2*) acc;    /* presumed aligned */
+    U64x2 const* const xinput = (U64x2 const*) input;   /* no alignment restriction */
+    U64x2 const* const xsecret  = (U64x2 const*) secret;    /* no alignment restriction */
+    U64x2 const v32 = { 32,  32 };
+#if XXPH_VSX_BE
+    U8x16 const vXorSwap  = { 0x07, 0x16, 0x25, 0x34, 0x43, 0x52, 0x61, 0x70,
+                              0x8F, 0x9E, 0xAD, 0xBC, 0xCB, 0xDA, 0xE9, 0xF8 };
+#endif
+    size_t i;
+    for (i = 0; i < STRIPE_LEN / sizeof(U64x2); i++) {
+        /* data_vec = xinput[i]; */
+        /* key_vec = xsecret[i]; */
+#if XXPH_VSX_BE
+        /* byteswap */
+        U64x2 const data_vec = XXPH_vec_revb(vec_vsx_ld(0, xinput + i));
+        U64x2 const key_raw = vec_vsx_ld(0, xsecret + i);
+        /* See comment above. data_key = data_vec ^ swap(xsecret[i]); */
+        U64x2 const data_key = (U64x2)XXPH_vec_permxor((U8x16)data_vec, (U8x16)key_raw, vXorSwap);
+#else
+        U64x2 const data_vec = vec_vsx_ld(0, xinput + i);
+        U64x2 const key_vec = vec_vsx_ld(0, xsecret + i);
+        U64x2 const data_key = data_vec ^ key_vec;
+#endif
+        /* shuffled = (data_key << 32) | (data_key >> 32); */
+        U32x4 const shuffled = (U32x4)vec_rl(data_key, v32);
+        /* product = ((U64x2)data_key & 0xFFFFFFFF) * ((U64x2)shuffled & 0xFFFFFFFF); */
+        U64x2 const product = XXPH_vec_mulo((U32x4)data_key, shuffled);
+        xacc[i] += product;
+
+        if (accWidth == XXPH3_acc_64bits) {
+            xacc[i] += data_vec;
+        } else {  /* XXPH3_acc_128bits */
+            /* swap high and low halves */
+            U64x2 const data_swapped = vec_xxpermdi(data_vec, data_vec, 2);
+            xacc[i] += data_swapped;
+        }
+    }
+
+#else   /* scalar variant of Accumulator - universal */
+
+    XXPH_ALIGN(XXPH_ACC_ALIGN) xxh_u64* const xacc = (xxh_u64*) acc;    /* presumed aligned on 32-bytes boundaries, little hint for the auto-vectorizer */
+    const xxh_u8* const xinput = (const xxh_u8*) input;  /* no alignment restriction */
+    const xxh_u8* const xsecret  = (const xxh_u8*) secret;   /* no alignment restriction */
+    size_t i;
+    XXPH_ASSERT(((size_t)acc & (XXPH_ACC_ALIGN-1)) == 0);
+    for (i=0; i < ACC_NB; i++) {
+        xxh_u64 const data_val = XXPH_readLE64(xinput + 8*i);
+        xxh_u64 const data_key = data_val ^ XXPH_readLE64(xsecret + i*8);
+
+        if (accWidth == XXPH3_acc_64bits) {
+            xacc[i] += data_val;
+        } else {
+            xacc[i ^ 1] += data_val; /* swap adjacent lanes */
+        }
+        xacc[i] += XXPH_mult32to64(data_key & 0xFFFFFFFF, data_key >> 32);
+    }
+#endif
+}
+
+XXPH_FORCE_INLINE void
+XXPH3_scrambleAcc(void* XXPH_RESTRICT acc, const void* XXPH_RESTRICT secret)
+{
+#if (XXPH_VECTOR == XXPH_AVX2)
+
+    XXPH_ASSERT((((size_t)acc) & 31) == 0);
+    {   XXPH_ALIGN(32) __m256i* const xacc = (__m256i*) acc;
+        const         __m256i* const xsecret = (const __m256i *) secret;   /* not really aligned, just for ptr arithmetic, and because _mm256_loadu_si256() requires this argument type */
+        const __m256i prime32 = _mm256_set1_epi32((int)PRIME32_1);
+
+        size_t i;
+        for (i=0; i < STRIPE_LEN/sizeof(__m256i); i++) {
+            /* xacc[i] ^= (xacc[i] >> 47) */
+            __m256i const acc_vec     = xacc[i];
+            __m256i const shifted     = _mm256_srli_epi64    (acc_vec, 47);
+            __m256i const data_vec    = _mm256_xor_si256     (acc_vec, shifted);
+            /* xacc[i] ^= xsecret; */
+            __m256i const key_vec     = _mm256_loadu_si256   (xsecret+i);
+            __m256i const data_key    = _mm256_xor_si256     (data_vec, key_vec);
+
+            /* xacc[i] *= PRIME32_1; */
+            __m256i const data_key_hi = _mm256_shuffle_epi32 (data_key, 0x31);
+            __m256i const prod_lo     = _mm256_mul_epu32     (data_key, prime32);
+            __m256i const prod_hi     = _mm256_mul_epu32     (data_key_hi, prime32);
+            xacc[i] = _mm256_add_epi64(prod_lo, _mm256_slli_epi64(prod_hi, 32));
+        }
+    }
+
+#elif (XXPH_VECTOR == XXPH_SSE2)
+
+    XXPH_ASSERT((((size_t)acc) & 15) == 0);
+    {   XXPH_ALIGN(16) __m128i* const xacc = (__m128i*) acc;
+        const         __m128i* const xsecret = (const __m128i *) secret;   /* not really aligned, just for ptr arithmetic, and because _mm_loadu_si128() requires this argument type */
+        const __m128i prime32 = _mm_set1_epi32((int)PRIME32_1);
+
+        size_t i;
+        for (i=0; i < STRIPE_LEN/sizeof(__m128i); i++) {
+            /* xacc[i] ^= (xacc[i] >> 47) */
+            __m128i const acc_vec     = xacc[i];
+            __m128i const shifted     = _mm_srli_epi64    (acc_vec, 47);
+            __m128i const data_vec    = _mm_xor_si128     (acc_vec, shifted);
+            /* xacc[i] ^= xsecret; */
+            __m128i const key_vec     = _mm_loadu_si128   (xsecret+i);
+            __m128i const data_key    = _mm_xor_si128     (data_vec, key_vec);
+
+            /* xacc[i] *= PRIME32_1; */
+            __m128i const data_key_hi = _mm_shuffle_epi32 (data_key, 0x31);
+            __m128i const prod_lo     = _mm_mul_epu32     (data_key, prime32);
+            __m128i const prod_hi     = _mm_mul_epu32     (data_key_hi, prime32);
+            xacc[i] = _mm_add_epi64(prod_lo, _mm_slli_epi64(prod_hi, 32));
+        }
+    }
+
+#elif (XXPH_VECTOR == XXPH_NEON)
+
+    XXPH_ASSERT((((size_t)acc) & 15) == 0);
+
+    {   uint64x2_t* const xacc =     (uint64x2_t*) acc;
+        uint8_t const* const xsecret = (uint8_t const*) secret;
+        uint32x2_t const prime     = vdup_n_u32 (PRIME32_1);
+
+        size_t i;
+        for (i=0; i < STRIPE_LEN/sizeof(uint64x2_t); i++) {
+            /* data_vec = xacc[i] ^ (xacc[i] >> 47); */
+            uint64x2_t const   acc_vec  = xacc[i];
+            uint64x2_t const   shifted  = vshrq_n_u64 (acc_vec, 47);
+            uint64x2_t const   data_vec = veorq_u64   (acc_vec, shifted);
+
+            /* key_vec  = xsecret[i]; */
+            uint32x4_t const   key_vec  = vreinterpretq_u32_u8(vld1q_u8(xsecret + (i * 16)));
+            /* data_key = data_vec ^ key_vec; */
+            uint32x4_t const   data_key = veorq_u32   (vreinterpretq_u32_u64(data_vec), key_vec);
+            /* shuffled = { data_key[0, 2], data_key[1, 3] }; */
+            uint32x2x2_t const shuffled = vzip_u32    (vget_low_u32(data_key), vget_high_u32(data_key));
+
+            /* data_key *= PRIME32_1 */
+
+            /* prod_hi = (data_key >> 32) * PRIME32_1; */
+            uint64x2_t const   prod_hi = vmull_u32    (shuffled.val[1], prime);
+            /* xacc[i] = prod_hi << 32; */
+            xacc[i] = vshlq_n_u64(prod_hi, 32);
+            /* xacc[i] += (prod_hi & 0xFFFFFFFF) * PRIME32_1; */
+            xacc[i] = vmlal_u32(xacc[i], shuffled.val[0], prime);
+    }   }
+
+#elif (XXPH_VECTOR == XXPH_VSX) && /* work around a compiler bug */ (__GNUC__ > 5)
+
+          U64x2* const xacc =       (U64x2*) acc;
+    const U64x2* const xsecret = (const U64x2*) secret;
+    /* constants */
+    U64x2 const v32  = { 32, 32 };
+    U64x2 const v47 = { 47, 47 };
+    U32x4 const prime = { PRIME32_1, PRIME32_1, PRIME32_1, PRIME32_1 };
+    size_t i;
+#if XXPH_VSX_BE
+    /* endian swap */
+    U8x16 const vXorSwap  = { 0x07, 0x16, 0x25, 0x34, 0x43, 0x52, 0x61, 0x70,
+                              0x8F, 0x9E, 0xAD, 0xBC, 0xCB, 0xDA, 0xE9, 0xF8 };
+#endif
+    for (i = 0; i < STRIPE_LEN / sizeof(U64x2); i++) {
+        U64x2 const acc_vec  = xacc[i];
+        U64x2 const data_vec = acc_vec ^ (acc_vec >> v47);
+        /* key_vec = xsecret[i]; */
+#if XXPH_VSX_BE
+        /* swap bytes words */
+        U64x2 const key_raw  = vec_vsx_ld(0, xsecret + i);
+        U64x2 const data_key = (U64x2)XXPH_vec_permxor((U8x16)data_vec, (U8x16)key_raw, vXorSwap);
+#else
+        U64x2 const key_vec  = vec_vsx_ld(0, xsecret + i);
+        U64x2 const data_key = data_vec ^ key_vec;
+#endif
+
+        /* data_key *= PRIME32_1 */
+
+        /* prod_lo = ((U64x2)data_key & 0xFFFFFFFF) * ((U64x2)prime & 0xFFFFFFFF);  */
+        U64x2 const prod_even  = XXPH_vec_mule((U32x4)data_key, prime);
+        /* prod_hi = ((U64x2)data_key >> 32) * ((U64x2)prime >> 32);  */
+        U64x2 const prod_odd  = XXPH_vec_mulo((U32x4)data_key, prime);
+        xacc[i] = prod_odd + (prod_even << v32);
+    }
+
+#else   /* scalar variant of Scrambler - universal */
+
+    XXPH_ALIGN(XXPH_ACC_ALIGN) xxh_u64* const xacc = (xxh_u64*) acc;   /* presumed aligned on 32-bytes boundaries, little hint for the auto-vectorizer */
+    const xxh_u8* const xsecret = (const xxh_u8*) secret;   /* no alignment restriction */
+    size_t i;
+    XXPH_ASSERT((((size_t)acc) & (XXPH_ACC_ALIGN-1)) == 0);
+    for (i=0; i < ACC_NB; i++) {
+        xxh_u64 const key64 = XXPH_readLE64(xsecret + 8*i);
+        xxh_u64 acc64 = xacc[i];
+        acc64 ^= acc64 >> 47;
+        acc64 ^= key64;
+        acc64 *= PRIME32_1;
+        xacc[i] = acc64;
+    }
+
+#endif
+}
+
+#define XXPH_PREFETCH_DIST 384
+
+/* assumption : nbStripes will not overflow secret size */
+XXPH_FORCE_INLINE void
+XXPH3_accumulate(       xxh_u64* XXPH_RESTRICT acc,
+                const xxh_u8* XXPH_RESTRICT input,
+                const xxh_u8* XXPH_RESTRICT secret,
+                      size_t nbStripes,
+                      XXPH3_accWidth_e accWidth)
+{
+    size_t n;
+    for (n = 0; n < nbStripes; n++ ) {
+        const xxh_u8* const in = input + n*STRIPE_LEN;
+        XXPH_PREFETCH(in + XXPH_PREFETCH_DIST);
+        XXPH3_accumulate_512(acc,
+                            in,
+                            secret + n*XXPH_SECRET_CONSUME_RATE,
+                            accWidth);
+    }
+}
+
+/* note : clang auto-vectorizes well in SS2 mode _if_ this function is `static`,
+ *        and doesn't auto-vectorize it at all if it is `FORCE_INLINE`.
+ *        However, it auto-vectorizes better AVX2 if it is `FORCE_INLINE`
+ *        Pretty much every other modes and compilers prefer `FORCE_INLINE`.
+ */
+
+#if defined(__clang__) && (XXPH_VECTOR==0) && !defined(__AVX2__) && !defined(__arm__) && !defined(__thumb__)
+static void
+#else
+XXPH_FORCE_INLINE void
+#endif
+XXPH3_hashLong_internal_loop( xxh_u64* XXPH_RESTRICT acc,
+                      const xxh_u8* XXPH_RESTRICT input, size_t len,
+                      const xxh_u8* XXPH_RESTRICT secret, size_t secretSize,
+                            XXPH3_accWidth_e accWidth)
+{
+    size_t const nb_rounds = (secretSize - STRIPE_LEN) / XXPH_SECRET_CONSUME_RATE;
+    size_t const block_len = STRIPE_LEN * nb_rounds;
+    size_t const nb_blocks = len / block_len;
+
+    size_t n;
+
+    XXPH_ASSERT(secretSize >= XXPH3_SECRET_SIZE_MIN);
+
+    for (n = 0; n < nb_blocks; n++) {
+        XXPH3_accumulate(acc, input + n*block_len, secret, nb_rounds, accWidth);
+        XXPH3_scrambleAcc(acc, secret + secretSize - STRIPE_LEN);
+    }
+
+    /* last partial block */
+    XXPH_ASSERT(len > STRIPE_LEN);
+    {   size_t const nbStripes = (len - (block_len * nb_blocks)) / STRIPE_LEN;
+        XXPH_ASSERT(nbStripes <= (secretSize / XXPH_SECRET_CONSUME_RATE));
+        XXPH3_accumulate(acc, input + nb_blocks*block_len, secret, nbStripes, accWidth);
+
+        /* last stripe */
+        if (len & (STRIPE_LEN - 1)) {
+            const xxh_u8* const p = input + len - STRIPE_LEN;
+#define XXPH_SECRET_LASTACC_START 7  /* do not align on 8, so that secret is different from scrambler */
+            XXPH3_accumulate_512(acc, p, secret + secretSize - STRIPE_LEN - XXPH_SECRET_LASTACC_START, accWidth);
+    }   }
+}
+
+XXPH_FORCE_INLINE xxh_u64
+XXPH3_mix2Accs(const xxh_u64* XXPH_RESTRICT acc, const xxh_u8* XXPH_RESTRICT secret)
+{
+    return XXPH3_mul128_fold64(
+               acc[0] ^ XXPH_readLE64(secret),
+               acc[1] ^ XXPH_readLE64(secret+8) );
+}
+
+static XXPH64_hash_t
+XXPH3_mergeAccs(const xxh_u64* XXPH_RESTRICT acc, const xxh_u8* XXPH_RESTRICT secret, xxh_u64 start)
+{
+    xxh_u64 result64 = start;
+
+    result64 += XXPH3_mix2Accs(acc+0, secret +  0);
+    result64 += XXPH3_mix2Accs(acc+2, secret + 16);
+    result64 += XXPH3_mix2Accs(acc+4, secret + 32);
+    result64 += XXPH3_mix2Accs(acc+6, secret + 48);
+
+    return XXPH3_avalanche(result64);
+}
+
+#define XXPH3_INIT_ACC { PRIME32_3, PRIME64_1, PRIME64_2, PRIME64_3, \
+                        PRIME64_4, PRIME32_2, PRIME64_5, PRIME32_1 };
+
+XXPH_FORCE_INLINE XXPH64_hash_t
+XXPH3_hashLong_internal(const xxh_u8* XXPH_RESTRICT input, size_t len,
+                       const xxh_u8* XXPH_RESTRICT secret, size_t secretSize)
+{
+    XXPH_ALIGN(XXPH_ACC_ALIGN) xxh_u64 acc[ACC_NB] = XXPH3_INIT_ACC;
+
+    XXPH3_hashLong_internal_loop(acc, input, len, secret, secretSize, XXPH3_acc_64bits);
+
+    /* converge into final hash */
+    XXPH_STATIC_ASSERT(sizeof(acc) == 64);
+#define XXPH_SECRET_MERGEACCS_START 11  /* do not align on 8, so that secret is different from accumulator */
+    XXPH_ASSERT(secretSize >= sizeof(acc) + XXPH_SECRET_MERGEACCS_START);
+    return XXPH3_mergeAccs(acc, secret + XXPH_SECRET_MERGEACCS_START, (xxh_u64)len * PRIME64_1);
+}
+
+
+XXPH_NO_INLINE XXPH64_hash_t    /* It's important for performance that XXPH3_hashLong is not inlined. Not sure why (uop cache maybe ?), but difference is large and easily measurable */
+XXPH3_hashLong_64b_defaultSecret(const xxh_u8* XXPH_RESTRICT input, size_t len)
+{
+    return XXPH3_hashLong_internal(input, len, kSecret, sizeof(kSecret));
+}
+
+XXPH_NO_INLINE XXPH64_hash_t    /* It's important for performance that XXPH3_hashLong is not inlined. Not sure why (uop cache maybe ?), but difference is large and easily measurable */
+XXPH3_hashLong_64b_withSecret(const xxh_u8* XXPH_RESTRICT input, size_t len,
+                             const xxh_u8* XXPH_RESTRICT secret, size_t secretSize)
+{
+    return XXPH3_hashLong_internal(input, len, secret, secretSize);
+}
+
+
+XXPH_FORCE_INLINE void XXPH_writeLE64(void* dst, xxh_u64 v64)
+{
+    if (!XXPH_CPU_LITTLE_ENDIAN) v64 = XXPH_swap64(v64);
+    memcpy(dst, &v64, sizeof(v64));
+}
+
+/* XXPH3_initCustomSecret() :
+ * destination `customSecret` is presumed allocated and same size as `kSecret`.
+ */
+XXPH_FORCE_INLINE void XXPH3_initCustomSecret(xxh_u8* customSecret, xxh_u64 seed64)
+{
+    int const nbRounds = XXPH_SECRET_DEFAULT_SIZE / 16;
+    int i;
+
+    XXPH_STATIC_ASSERT((XXPH_SECRET_DEFAULT_SIZE & 15) == 0);
+
+    for (i=0; i < nbRounds; i++) {
+        XXPH_writeLE64(customSecret + 16*i,     XXPH_readLE64(kSecret + 16*i)     + seed64);
+        XXPH_writeLE64(customSecret + 16*i + 8, XXPH_readLE64(kSecret + 16*i + 8) - seed64);
+    }
+}
+
+
+/* XXPH3_hashLong_64b_withSeed() :
+ * Generate a custom key,
+ * based on alteration of default kSecret with the seed,
+ * and then use this key for long mode hashing.
+ * This operation is decently fast but nonetheless costs a little bit of time.
+ * Try to avoid it whenever possible (typically when seed==0).
+ */
+XXPH_NO_INLINE XXPH64_hash_t    /* It's important for performance that XXPH3_hashLong is not inlined. Not sure why (uop cache maybe ?), but difference is large and easily measurable */
+XXPH3_hashLong_64b_withSeed(const xxh_u8* input, size_t len, XXPH64_hash_t seed)
+{
+    XXPH_ALIGN(8) xxh_u8 secret[XXPH_SECRET_DEFAULT_SIZE];
+    if (seed==0) return XXPH3_hashLong_64b_defaultSecret(input, len);
+    XXPH3_initCustomSecret(secret, seed);
+    return XXPH3_hashLong_internal(input, len, secret, sizeof(secret));
+}
+
+
+XXPH_FORCE_INLINE xxh_u64 XXPH3_mix16B(const xxh_u8* XXPH_RESTRICT input,
+                                 const xxh_u8* XXPH_RESTRICT secret, xxh_u64 seed64)
+{
+    xxh_u64 const input_lo = XXPH_readLE64(input);
+    xxh_u64 const input_hi = XXPH_readLE64(input+8);
+    return XXPH3_mul128_fold64(
+               input_lo ^ (XXPH_readLE64(secret)   + seed64),
+               input_hi ^ (XXPH_readLE64(secret+8) - seed64) );
+}
+
+
+XXPH_FORCE_INLINE XXPH64_hash_t
+XXPH3_len_17to128_64b(const xxh_u8* XXPH_RESTRICT input, size_t len,
+                     const xxh_u8* XXPH_RESTRICT secret, size_t secretSize,
+                     XXPH64_hash_t seed)
+{
+    XXPH_ASSERT(secretSize >= XXPH3_SECRET_SIZE_MIN); (void)secretSize;
+    XXPH_ASSERT(16 < len && len <= 128);
+
+    {   xxh_u64 acc = len * PRIME64_1;
+        if (len > 32) {
+            if (len > 64) {
+                if (len > 96) {
+                    acc += XXPH3_mix16B(input+48, secret+96, seed);
+                    acc += XXPH3_mix16B(input+len-64, secret+112, seed);
+                }
+                acc += XXPH3_mix16B(input+32, secret+64, seed);
+                acc += XXPH3_mix16B(input+len-48, secret+80, seed);
+            }
+            acc += XXPH3_mix16B(input+16, secret+32, seed);
+            acc += XXPH3_mix16B(input+len-32, secret+48, seed);
+        }
+        acc += XXPH3_mix16B(input+0, secret+0, seed);
+        acc += XXPH3_mix16B(input+len-16, secret+16, seed);
+
+        return XXPH3_avalanche(acc);
+    }
+}
+
+#define XXPH3_MIDSIZE_MAX 240
+
+XXPH_NO_INLINE XXPH64_hash_t
+XXPH3_len_129to240_64b(const xxh_u8* XXPH_RESTRICT input, size_t len,
+                      const xxh_u8* XXPH_RESTRICT secret, size_t secretSize,
+                      XXPH64_hash_t seed)
+{
+    XXPH_ASSERT(secretSize >= XXPH3_SECRET_SIZE_MIN); (void)secretSize;
+    XXPH_ASSERT(128 < len && len <= XXPH3_MIDSIZE_MAX);
+
+    #define XXPH3_MIDSIZE_STARTOFFSET 3
+    #define XXPH3_MIDSIZE_LASTOFFSET  17
+
+    {   xxh_u64 acc = len * PRIME64_1;
+        int const nbRounds = (int)len / 16;
+        int i;
+        for (i=0; i<8; i++) {
+            acc += XXPH3_mix16B(input+(16*i), secret+(16*i), seed);
+        }
+        acc = XXPH3_avalanche(acc);
+        XXPH_ASSERT(nbRounds >= 8);
+        for (i=8 ; i < nbRounds; i++) {
+            acc += XXPH3_mix16B(input+(16*i), secret+(16*(i-8)) + XXPH3_MIDSIZE_STARTOFFSET, seed);
+        }
+        /* last bytes */
+        acc += XXPH3_mix16B(input + len - 16, secret + XXPH3_SECRET_SIZE_MIN - XXPH3_MIDSIZE_LASTOFFSET, seed);
+        return XXPH3_avalanche(acc);
+    }
+}
+
+/* ===   Public entry point   === */
+
+XXPH_PUBLIC_API XXPH64_hash_t XXPH3_64bits(const void* input, size_t len)
+{
+    if (len <= 16) return XXPH3_len_0to16_64b((const xxh_u8*)input, len, kSecret, 0);
+    if (len <= 128) return XXPH3_len_17to128_64b((const xxh_u8*)input, len, kSecret, sizeof(kSecret), 0);
+    if (len <= XXPH3_MIDSIZE_MAX) return XXPH3_len_129to240_64b((const xxh_u8*)input, len, kSecret, sizeof(kSecret), 0);
+    return XXPH3_hashLong_64b_defaultSecret((const xxh_u8*)input, len);
+}
+
+XXPH_PUBLIC_API XXPH64_hash_t
+XXPH3_64bits_withSecret(const void* input, size_t len, const void* secret, size_t secretSize)
+{
+    XXPH_ASSERT(secretSize >= XXPH3_SECRET_SIZE_MIN);
+    /* if an action must be taken should `secret` conditions not be respected,
+     * it should be done here.
+     * For now, it's a contract pre-condition.
+     * Adding a check and a branch here would cost performance at every hash */
+     if (len <= 16) return XXPH3_len_0to16_64b((const xxh_u8*)input, len, (const xxh_u8*)secret, 0);
+     if (len <= 128) return XXPH3_len_17to128_64b((const xxh_u8*)input, len, (const xxh_u8*)secret, secretSize, 0);
+     if (len <= XXPH3_MIDSIZE_MAX) return XXPH3_len_129to240_64b((const xxh_u8*)input, len, (const xxh_u8*)secret, secretSize, 0);
+     return XXPH3_hashLong_64b_withSecret((const xxh_u8*)input, len, (const xxh_u8*)secret, secretSize);
+}
+
+XXPH_PUBLIC_API XXPH64_hash_t
+XXPH3_64bits_withSeed(const void* input, size_t len, XXPH64_hash_t seed)
+{
+    if (len <= 16) return XXPH3_len_0to16_64b((const xxh_u8*)input, len, kSecret, seed);
+    if (len <= 128) return XXPH3_len_17to128_64b((const xxh_u8*)input, len, kSecret, sizeof(kSecret), seed);
+    if (len <= XXPH3_MIDSIZE_MAX) return XXPH3_len_129to240_64b((const xxh_u8*)input, len, kSecret, sizeof(kSecret), seed);
+    return XXPH3_hashLong_64b_withSeed((const xxh_u8*)input, len, seed);
+}
+
+/* ===   XXPH3 streaming   === */
+
+/* RocksDB Note: unused & removed due to bug in preview version */
+
+/*======== END #include "xxh3.h", now inlined above ==========*/
+
+#endif  /* XXPH_NO_LONG_LONG */
+
+/* === END RocksDB modification of permanently inlining === */
+
+#endif /*  defined(XXPH_INLINE_ALL) || defined(XXPH_PRIVATE_API) */
+
+#endif /* XXPH_STATIC_LINKING_ONLY */
+
+#if defined (__cplusplus)
+}
+#endif
+
+#endif /* XXPHASH_H_5627135585666179 */
diff --git a/src/rocksdb/utilities/agg_merge/agg_merge.cc b/src/rocksdb/utilities/agg_merge/agg_merge.cc
new file mode 100644
index 000000000..a7eab1f12
--- /dev/null
+++ b/src/rocksdb/utilities/agg_merge/agg_merge.cc
@@ -0,0 +1,238 @@
+//  Copyright (c) 2017-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#include "utilities/agg_merge/agg_merge.h"
+
+#include <assert.h>
+
+#include <deque>
+#include <memory>
+#include <type_traits>
+#include <utility>
+#include <vector>
+
+#include "port/lang.h"
+#include "port/likely.h"
+#include "rocksdb/merge_operator.h"
+#include "rocksdb/slice.h"
+#include "rocksdb/utilities/agg_merge.h"
+#include "rocksdb/utilities/options_type.h"
+#include "util/coding.h"
+#include "utilities/merge_operators.h"
+
+namespace ROCKSDB_NAMESPACE {
+static std::unordered_map<std::string, std::unique_ptr<Aggregator>> func_map;
+const std::string kUnnamedFuncName = "";
+const std::string kErrorFuncName = "kErrorFuncName";
+
+Status AddAggregator(const std::string& function_name,
+                     std::unique_ptr<Aggregator>&& agg) {
+  if (function_name == kErrorFuncName) {
+    return Status::InvalidArgument(
+        "Cannot register function name kErrorFuncName");
+  }
+  func_map.emplace(function_name, std::move(agg));
+  return Status::OK();
+}
+
+AggMergeOperator::AggMergeOperator() {}
+
+std::string EncodeAggFuncAndPayloadNoCheck(const Slice& function_name,
+                                           const Slice& value) {
+  std::string result;
+  PutLengthPrefixedSlice(&result, function_name);
+  result += value.ToString();
+  return result;
+}
+
+Status EncodeAggFuncAndPayload(const Slice& function_name, const Slice& payload,
+                               std::string& output) {
+  if (function_name == kErrorFuncName) {
+    return Status::InvalidArgument("Cannot use error function name");
+  }
+  if (function_name != kUnnamedFuncName &&
+      func_map.find(function_name.ToString()) == func_map.end()) {
+    return Status::InvalidArgument("Function name not registered");
+  }
+  output = EncodeAggFuncAndPayloadNoCheck(function_name, payload);
+  return Status::OK();
+}
+
+bool ExtractAggFuncAndValue(const Slice& op, Slice& func, Slice& value) {
+  value = op;
+  return GetLengthPrefixedSlice(&value, &func);
+}
+
+bool ExtractList(const Slice& encoded_list, std::vector<Slice>& decoded_list) {
+  decoded_list.clear();
+  Slice list_slice = encoded_list;
+  Slice item;
+  while (GetLengthPrefixedSlice(&list_slice, &item)) {
+    decoded_list.push_back(item);
+  }
+  return list_slice.empty();
+}
+
+class AggMergeOperator::Accumulator {
+ public:
+  bool Add(const Slice& op, bool is_partial_aggregation) {
+    if (ignore_operands_) {
+      return true;
+    }
+    Slice my_func;
+    Slice my_value;
+    bool ret = ExtractAggFuncAndValue(op, my_func, my_value);
+    if (!ret) {
+      ignore_operands_ = true;
+      return true;
+    }
+
+    // Determine whether we need to do partial merge.
+    if (is_partial_aggregation && !my_func.empty()) {
+      auto f = func_map.find(my_func.ToString());
+      if (f == func_map.end() || !f->second->DoPartialAggregate()) {
+        return false;
+      }
+    }
+
+    if (!func_valid_) {
+      if (my_func != kUnnamedFuncName) {
+        func_ = my_func;
+        func_valid_ = true;
+      }
+    } else if (func_ != my_func) {
+      // User switched aggregation function. Need to aggregate the older
+      // one first.
+
+      // Previous aggreagion can't be done in partial merge
+      if (is_partial_aggregation) {
+        func_valid_ = false;
+        ignore_operands_ = true;
+        return false;
+      }
+
+      // We could consider stashing an iterator into the hash of aggregators
+      // to avoid repeated lookups when the aggregator doesn't change.
+      auto f = func_map.find(func_.ToString());
+      if (f == func_map.end() || !f->second->Aggregate(values_, scratch_)) {
+        func_valid_ = false;
+        ignore_operands_ = true;
+        return true;
+      }
+      std::swap(scratch_, aggregated_);
+      values_.clear();
+      values_.push_back(aggregated_);
+      func_ = my_func;
+    }
+    values_.push_back(my_value);
+    return true;
+  }
+
+  // Return false if aggregation fails.
+  // One possible reason
+  bool GetResult(std::string& result) {
+    if (!func_valid_) {
+      return false;
+    }
+    auto f = func_map.find(func_.ToString());
+    if (f == func_map.end()) {
+      return false;
+    }
+    if (!f->second->Aggregate(values_, scratch_)) {
+      return false;
+    }
+    result = EncodeAggFuncAndPayloadNoCheck(func_, scratch_);
+    return true;
+  }
+
+  void Clear() {
+    func_.clear();
+    values_.clear();
+    aggregated_.clear();
+    scratch_.clear();
+    ignore_operands_ = false;
+    func_valid_ = false;
+  }
+
+ private:
+  Slice func_;
+  std::vector<Slice> values_;
+  std::string aggregated_;
+  std::string scratch_;
+  bool ignore_operands_ = false;
+  bool func_valid_ = false;
+};
+
+// Creating and using a new Accumulator might invoke multiple malloc and is
+// expensive if it needs to be done when processing each merge operation.
+// AggMergeOperator's merge operators can be invoked concurrently by multiple
+// threads so we cannot simply create one Aggregator and reuse.
+// We use thread local instances instead.
+AggMergeOperator::Accumulator& AggMergeOperator::GetTLSAccumulator() {
+  static thread_local Accumulator tls_acc;
+  tls_acc.Clear();
+  return tls_acc;
+}
+
+void AggMergeOperator::PackAllMergeOperands(const MergeOperationInput& merge_in,
+                                            MergeOperationOutput& merge_out) {
+  merge_out.new_value = "";
+  PutLengthPrefixedSlice(&merge_out.new_value, kErrorFuncName);
+  if (merge_in.existing_value != nullptr) {
+    PutLengthPrefixedSlice(&merge_out.new_value, *merge_in.existing_value);
+  }
+  for (const Slice& op : merge_in.operand_list) {
+    PutLengthPrefixedSlice(&merge_out.new_value, op);
+  }
+}
+
+bool AggMergeOperator::FullMergeV2(const MergeOperationInput& merge_in,
+                                   MergeOperationOutput* merge_out) const {
+  Accumulator& agg = GetTLSAccumulator();
+  if (merge_in.existing_value != nullptr) {
+    agg.Add(*merge_in.existing_value, /*is_partial_aggregation=*/false);
+  }
+  for (const Slice& e : merge_in.operand_list) {
+    agg.Add(e, /*is_partial_aggregation=*/false);
+  }
+
+  bool succ = agg.GetResult(merge_out->new_value);
+  if (!succ) {
+    // If aggregation can't happen, pack all merge operands. In contrast to
+    // merge operator, we don't want to fail the DB. If users insert wrong
+    // format or call unregistered an aggregation function, we still hope
+    // the DB can continue functioning with other keys.
+    PackAllMergeOperands(merge_in, *merge_out);
+  }
+  agg.Clear();
+  return true;
+}
+
+bool AggMergeOperator::PartialMergeMulti(const Slice& /*key*/,
+                                         const std::deque<Slice>& operand_list,
+                                         std::string* new_value,
+                                         Logger* /*logger*/) const {
+  Accumulator& agg = GetTLSAccumulator();
+  bool do_aggregation = true;
+  for (const Slice& item : operand_list) {
+    do_aggregation = agg.Add(item, /*is_partial_aggregation=*/true);
+    if (!do_aggregation) {
+      break;
+    }
+  }
+  if (do_aggregation) {
+    do_aggregation = agg.GetResult(*new_value);
+  }
+  agg.Clear();
+  return do_aggregation;
+}
+
+std::shared_ptr<MergeOperator> GetAggMergeOperator() {
+  STATIC_AVOID_DESTRUCTION(std::shared_ptr<MergeOperator>, instance)
+  (std::make_shared<AggMergeOperator>());
+  assert(instance);
+  return instance;
+}
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/utilities/agg_merge/agg_merge.h b/src/rocksdb/utilities/agg_merge/agg_merge.h
new file mode 100644
index 000000000..00e58de08
--- /dev/null
+++ b/src/rocksdb/utilities/agg_merge/agg_merge.h
@@ -0,0 +1,49 @@
+//  Copyright (c) 2017-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#pragma once
+#include <algorithm>
+#include <cstddef>
+#include <memory>
+#include <unordered_map>
+
+#include "rocksdb/merge_operator.h"
+#include "rocksdb/slice.h"
+#include "rocksdb/utilities/agg_merge.h"
+#include "utilities/cassandra/cassandra_options.h"
+
+namespace ROCKSDB_NAMESPACE {
+class AggMergeOperator : public MergeOperator {
+ public:
+  explicit AggMergeOperator();
+
+  bool FullMergeV2(const MergeOperationInput& merge_in,
+                   MergeOperationOutput* merge_out) const override;
+
+  bool PartialMergeMulti(const Slice& key,
+                         const std::deque<Slice>& operand_list,
+                         std::string* new_value, Logger* logger) const override;
+
+  const char* Name() const override { return kClassName(); }
+  static const char* kClassName() { return "AggMergeOperator.v1"; }
+
+  bool AllowSingleOperand() const override { return true; }
+
+  bool ShouldMerge(const std::vector<Slice>&) const override { return false; }
+
+ private:
+  class Accumulator;
+
+  // Pack all merge operands into one value. This is called when aggregation
+  // fails. The existing values are preserved and returned so that users can
+  // debug the problem.
+  static void PackAllMergeOperands(const MergeOperationInput& merge_in,
+                                   MergeOperationOutput& merge_out);
+  static Accumulator& GetTLSAccumulator();
+};
+
+extern std::string EncodeAggFuncAndPayloadNoCheck(const Slice& function_name,
+                                                  const Slice& value);
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/utilities/agg_merge/agg_merge_test.cc b/src/rocksdb/utilities/agg_merge/agg_merge_test.cc
new file mode 100644
index 000000000..a65441cd0
--- /dev/null
+++ b/src/rocksdb/utilities/agg_merge/agg_merge_test.cc
@@ -0,0 +1,135 @@
+// Copyright (c) 2017-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#include "rocksdb/utilities/agg_merge.h"
+
+#include <gtest/gtest.h>
+
+#include <memory>
+
+#include "db/db_test_util.h"
+#include "rocksdb/options.h"
+#include "test_util/testharness.h"
+#include "utilities/agg_merge/agg_merge.h"
+#include "utilities/agg_merge/test_agg_merge.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+class AggMergeTest : public DBTestBase {
+ public:
+  AggMergeTest() : DBTestBase("agg_merge_db_test", /*env_do_fsync=*/true) {}
+};
+
+TEST_F(AggMergeTest, TestUsingMergeOperator) {
+  ASSERT_OK(AddAggregator("sum", std::make_unique<SumAggregator>()));
+  ASSERT_OK(AddAggregator("last3", std::make_unique<Last3Aggregator>()));
+  ASSERT_OK(AddAggregator("mul", std::make_unique<MultipleAggregator>()));
+
+  Options options = CurrentOptions();
+  options.merge_operator = GetAggMergeOperator();
+  Reopen(options);
+  std::string v = EncodeHelper::EncodeFuncAndInt("sum", 10);
+  ASSERT_OK(Merge("foo", v));
+  v = EncodeHelper::EncodeFuncAndInt("sum", 20);
+  ASSERT_OK(Merge("foo", v));
+  v = EncodeHelper::EncodeFuncAndInt("sum", 15);
+  ASSERT_OK(Merge("foo", v));
+
+  v = EncodeHelper::EncodeFuncAndList("last3", {"a", "b"});
+  ASSERT_OK(Merge("bar", v));
+  v = EncodeHelper::EncodeFuncAndList("last3", {"c", "d", "e"});
+  ASSERT_OK(Merge("bar", v));
+  ASSERT_OK(Flush());
+  v = EncodeHelper::EncodeFuncAndList("last3", {"f"});
+  ASSERT_OK(Merge("bar", v));
+
+  // Test Put() without aggregation type.
+  v = EncodeHelper::EncodeFuncAndInt(kUnnamedFuncName, 30);
+  ASSERT_OK(Put("foo2", v));
+  v = EncodeHelper::EncodeFuncAndInt("sum", 10);
+  ASSERT_OK(Merge("foo2", v));
+  v = EncodeHelper::EncodeFuncAndInt("sum", 20);
+  ASSERT_OK(Merge("foo2", v));
+
+  EXPECT_EQ(EncodeHelper::EncodeFuncAndInt("sum", 45), Get("foo"));
+  EXPECT_EQ(EncodeHelper::EncodeFuncAndList("last3", {"f", "c", "d"}),
+            Get("bar"));
+  EXPECT_EQ(EncodeHelper::EncodeFuncAndInt("sum", 60), Get("foo2"));
+
+  // Test changing aggregation type
+  v = EncodeHelper::EncodeFuncAndInt("mul", 10);
+  ASSERT_OK(Put("bar2", v));
+  v = EncodeHelper::EncodeFuncAndInt("mul", 20);
+  ASSERT_OK(Merge("bar2", v));
+  v = EncodeHelper::EncodeFuncAndInt("sum", 30);
+  ASSERT_OK(Merge("bar2", v));
+  v = EncodeHelper::EncodeFuncAndInt("sum", 40);
+  ASSERT_OK(Merge("bar2", v));
+  EXPECT_EQ(EncodeHelper::EncodeFuncAndInt("sum", 10 * 20 + 30 + 40),
+            Get("bar2"));
+
+  // Changing aggregation type with partial merge
+  v = EncodeHelper::EncodeFuncAndInt("mul", 10);
+  ASSERT_OK(Merge("foo3", v));
+  ASSERT_OK(Flush());
+  v = EncodeHelper::EncodeFuncAndInt("mul", 10);
+  ASSERT_OK(Merge("foo3", v));
+  v = EncodeHelper::EncodeFuncAndInt("mul", 10);
+  ASSERT_OK(Merge("foo3", v));
+  v = EncodeHelper::EncodeFuncAndInt("sum", 10);
+  ASSERT_OK(Merge("foo3", v));
+  ASSERT_OK(Flush());
+  EXPECT_EQ(EncodeHelper::EncodeFuncAndInt("sum", 10 * 10 * 10 + 10),
+            Get("foo3"));
+
+  // Merge after full merge
+  v = EncodeHelper::EncodeFuncAndInt("sum", 1);
+  ASSERT_OK(Merge("foo4", v));
+  v = EncodeHelper::EncodeFuncAndInt("sum", 2);
+  ASSERT_OK(Merge("foo4", v));
+  ASSERT_OK(Flush());
+  v = EncodeHelper::EncodeFuncAndInt("sum", 3);
+  ASSERT_OK(Merge("foo4", v));
+  v = EncodeHelper::EncodeFuncAndInt("sum", 4);
+  ASSERT_OK(Merge("foo4", v));
+  ASSERT_OK(Flush());
+  ASSERT_OK(db_->CompactRange(CompactRangeOptions(), nullptr, nullptr));
+  v = EncodeHelper::EncodeFuncAndInt("sum", 5);
+  ASSERT_OK(Merge("foo4", v));
+  EXPECT_EQ(EncodeHelper::EncodeFuncAndInt("sum", 15), Get("foo4"));
+
+  // Test unregistered function name
+  v = EncodeAggFuncAndPayloadNoCheck("non_existing", "1");
+  ASSERT_OK(Merge("bar3", v));
+  std::string v1;
+  v1 = EncodeAggFuncAndPayloadNoCheck("non_existing", "invalid");
+  ;
+  ASSERT_OK(Merge("bar3", v1));
+  EXPECT_EQ(EncodeAggFuncAndPayloadNoCheck(kErrorFuncName,
+                                           EncodeHelper::EncodeList({v, v1})),
+            Get("bar3"));
+
+  // invalidate input
+  ASSERT_OK(EncodeAggFuncAndPayload("sum", "invalid", v));
+  ASSERT_OK(Merge("bar4", v));
+  v1 = EncodeHelper::EncodeFuncAndInt("sum", 20);
+  ASSERT_OK(Merge("bar4", v1));
+  std::string aggregated_value = Get("bar4");
+  Slice func, payload;
+  ASSERT_TRUE(ExtractAggFuncAndValue(aggregated_value, func, payload));
+  EXPECT_EQ(kErrorFuncName, func);
+  std::vector<Slice> decoded_list;
+  ASSERT_TRUE(ExtractList(payload, decoded_list));
+  ASSERT_EQ(2, decoded_list.size());
+  ASSERT_EQ(v, decoded_list[0]);
+  ASSERT_EQ(v1, decoded_list[1]);
+}
+}  // namespace ROCKSDB_NAMESPACE
+
+int main(int argc, char** argv) {
+  ROCKSDB_NAMESPACE::port::InstallStackTraceHandler();
+  ::testing::InitGoogleTest(&argc, argv);
+  return RUN_ALL_TESTS();
+}
diff --git a/src/rocksdb/utilities/agg_merge/test_agg_merge.cc b/src/rocksdb/utilities/agg_merge/test_agg_merge.cc
new file mode 100644
index 000000000..06e5b5697
--- /dev/null
+++ b/src/rocksdb/utilities/agg_merge/test_agg_merge.cc
@@ -0,0 +1,104 @@
+//  Copyright (c) 2017-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#include "test_agg_merge.h"
+
+#include <assert.h>
+
+#include <deque>
+#include <vector>
+
+#include "util/coding.h"
+#include "utilities/agg_merge/agg_merge.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+std::string EncodeHelper::EncodeFuncAndInt(const Slice& function_name,
+                                           int64_t value) {
+  std::string encoded_value;
+  PutVarsignedint64(&encoded_value, value);
+  std::string ret;
+  Status s = EncodeAggFuncAndPayload(function_name, encoded_value, ret);
+  assert(s.ok());
+  return ret;
+}
+
+std::string EncodeHelper::EncodeInt(int64_t value) {
+  std::string encoded_value;
+  PutVarsignedint64(&encoded_value, value);
+  return encoded_value;
+}
+
+std::string EncodeHelper::EncodeFuncAndList(const Slice& function_name,
+                                            const std::vector<Slice>& list) {
+  std::string ret;
+  Status s = EncodeAggFuncAndPayload(function_name, EncodeList(list), ret);
+  assert(s.ok());
+  return ret;
+}
+
+std::string EncodeHelper::EncodeList(const std::vector<Slice>& list) {
+  std::string result;
+  for (const Slice& entity : list) {
+    PutLengthPrefixedSlice(&result, entity);
+  }
+  return result;
+}
+
+bool SumAggregator::Aggregate(const std::vector<Slice>& item_list,
+                              std::string& result) const {
+  int64_t sum = 0;
+  for (const Slice& item : item_list) {
+    int64_t ivalue;
+    Slice v = item;
+    if (!GetVarsignedint64(&v, &ivalue) || !v.empty()) {
+      return false;
+    }
+    sum += ivalue;
+  }
+  result = EncodeHelper::EncodeInt(sum);
+  return true;
+}
+
+bool MultipleAggregator::Aggregate(const std::vector<Slice>& item_list,
+                                   std::string& result) const {
+  int64_t mresult = 1;
+  for (const Slice& item : item_list) {
+    int64_t ivalue;
+    Slice v = item;
+    if (!GetVarsignedint64(&v, &ivalue) || !v.empty()) {
+      return false;
+    }
+    mresult *= ivalue;
+  }
+  result = EncodeHelper::EncodeInt(mresult);
+  return true;
+}
+
+bool Last3Aggregator::Aggregate(const std::vector<Slice>& item_list,
+                                std::string& result) const {
+  std::vector<Slice> last3;
+  last3.reserve(3);
+  for (auto it = item_list.rbegin(); it != item_list.rend(); ++it) {
+    Slice input = *it;
+    Slice entity;
+    bool ret;
+    while ((ret = GetLengthPrefixedSlice(&input, &entity)) == true) {
+      last3.push_back(entity);
+      if (last3.size() >= 3) {
+        break;
+      }
+    }
+    if (last3.size() >= 3) {
+      break;
+    }
+    if (!ret) {
+      continue;
+    }
+  }
+  result = EncodeHelper::EncodeList(last3);
+  return true;
+}
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/utilities/agg_merge/test_agg_merge.h b/src/rocksdb/utilities/agg_merge/test_agg_merge.h
new file mode 100644
index 000000000..5bdf8b9cc
--- /dev/null
+++ b/src/rocksdb/utilities/agg_merge/test_agg_merge.h
@@ -0,0 +1,47 @@
+//  Copyright (c) 2017-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#pragma once
+#include <algorithm>
+#include <cstddef>
+#include <memory>
+#include <unordered_map>
+
+#include "rocksdb/merge_operator.h"
+#include "rocksdb/slice.h"
+#include "rocksdb/utilities/agg_merge.h"
+#include "utilities/cassandra/cassandra_options.h"
+
+namespace ROCKSDB_NAMESPACE {
+class SumAggregator : public Aggregator {
+ public:
+  ~SumAggregator() override {}
+  bool Aggregate(const std::vector<Slice>&, std::string& result) const override;
+  bool DoPartialAggregate() const override { return true; }
+};
+
+class MultipleAggregator : public Aggregator {
+ public:
+  ~MultipleAggregator() override {}
+  bool Aggregate(const std::vector<Slice>&, std::string& result) const override;
+  bool DoPartialAggregate() const override { return true; }
+};
+
+class Last3Aggregator : public Aggregator {
+ public:
+  ~Last3Aggregator() override {}
+  bool Aggregate(const std::vector<Slice>&, std::string& result) const override;
+};
+
+class EncodeHelper {
+ public:
+  static std::string EncodeFuncAndInt(const Slice& function_name,
+                                      int64_t value);
+  static std::string EncodeInt(int64_t value);
+  static std::string EncodeList(const std::vector<Slice>& list);
+  static std::string EncodeFuncAndList(const Slice& function_name,
+                                       const std::vector<Slice>& list);
+};
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/utilities/backup/backup_engine.cc b/src/rocksdb/utilities/backup/backup_engine.cc
new file mode 100644
index 000000000..81b4a6629
--- /dev/null
+++ b/src/rocksdb/utilities/backup/backup_engine.cc
@@ -0,0 +1,3181 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+//
+// Copyright (c) 2011 The LevelDB Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file. See the AUTHORS file for names of contributors.
+
+#ifndef ROCKSDB_LITE
+
+#include <algorithm>
+#include <atomic>
+#include <cinttypes>
+#include <cstdlib>
+#include <functional>
+#include <future>
+#include <limits>
+#include <map>
+#include <mutex>
+#include <sstream>
+#include <string>
+#include <thread>
+#include <unordered_map>
+#include <unordered_set>
+#include <vector>
+
+#include "env/composite_env_wrapper.h"
+#include "env/fs_readonly.h"
+#include "env/fs_remap.h"
+#include "file/filename.h"
+#include "file/line_file_reader.h"
+#include "file/sequence_file_reader.h"
+#include "file/writable_file_writer.h"
+#include "logging/logging.h"
+#include "monitoring/iostats_context_imp.h"
+#include "options/options_helper.h"
+#include "port/port.h"
+#include "rocksdb/advanced_options.h"
+#include "rocksdb/env.h"
+#include "rocksdb/rate_limiter.h"
+#include "rocksdb/statistics.h"
+#include "rocksdb/transaction_log.h"
+#include "table/sst_file_dumper.h"
+#include "test_util/sync_point.h"
+#include "util/cast_util.h"
+#include "util/channel.h"
+#include "util/coding.h"
+#include "util/crc32c.h"
+#include "util/math.h"
+#include "util/rate_limiter.h"
+#include "util/string_util.h"
+#include "utilities/backup/backup_engine_impl.h"
+#include "utilities/checkpoint/checkpoint_impl.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+namespace {
+using ShareFilesNaming = BackupEngineOptions::ShareFilesNaming;
+
+constexpr BackupID kLatestBackupIDMarker = static_cast<BackupID>(-2);
+
+inline uint32_t ChecksumHexToInt32(const std::string& checksum_hex) {
+  std::string checksum_str;
+  Slice(checksum_hex).DecodeHex(&checksum_str);
+  return EndianSwapValue(DecodeFixed32(checksum_str.c_str()));
+}
+inline std::string ChecksumStrToHex(const std::string& checksum_str) {
+  return Slice(checksum_str).ToString(true);
+}
+inline std::string ChecksumInt32ToHex(const uint32_t& checksum_value) {
+  std::string checksum_str;
+  PutFixed32(&checksum_str, EndianSwapValue(checksum_value));
+  return ChecksumStrToHex(checksum_str);
+}
+
+const std::string kPrivateDirName = "private";
+const std::string kMetaDirName = "meta";
+const std::string kSharedDirName = "shared";
+const std::string kSharedChecksumDirName = "shared_checksum";
+const std::string kPrivateDirSlash = kPrivateDirName + "/";
+const std::string kMetaDirSlash = kMetaDirName + "/";
+const std::string kSharedDirSlash = kSharedDirName + "/";
+const std::string kSharedChecksumDirSlash = kSharedChecksumDirName + "/";
+
+}  // namespace
+
+void BackupStatistics::IncrementNumberSuccessBackup() {
+  number_success_backup++;
+}
+void BackupStatistics::IncrementNumberFailBackup() { number_fail_backup++; }
+
+uint32_t BackupStatistics::GetNumberSuccessBackup() const {
+  return number_success_backup;
+}
+uint32_t BackupStatistics::GetNumberFailBackup() const {
+  return number_fail_backup;
+}
+
+std::string BackupStatistics::ToString() const {
+  char result[50];
+  snprintf(result, sizeof(result), "# success backup: %u, # fail backup: %u",
+           GetNumberSuccessBackup(), GetNumberFailBackup());
+  return result;
+}
+
+void BackupEngineOptions::Dump(Logger* logger) const {
+  ROCKS_LOG_INFO(logger, "               Options.backup_dir: %s",
+                 backup_dir.c_str());
+  ROCKS_LOG_INFO(logger, "               Options.backup_env: %p", backup_env);
+  ROCKS_LOG_INFO(logger, "        Options.share_table_files: %d",
+                 static_cast<int>(share_table_files));
+  ROCKS_LOG_INFO(logger, "                 Options.info_log: %p", info_log);
+  ROCKS_LOG_INFO(logger, "                     Options.sync: %d",
+                 static_cast<int>(sync));
+  ROCKS_LOG_INFO(logger, "         Options.destroy_old_data: %d",
+                 static_cast<int>(destroy_old_data));
+  ROCKS_LOG_INFO(logger, "         Options.backup_log_files: %d",
+                 static_cast<int>(backup_log_files));
+  ROCKS_LOG_INFO(logger, "        Options.backup_rate_limit: %" PRIu64,
+                 backup_rate_limit);
+  ROCKS_LOG_INFO(logger, "       Options.restore_rate_limit: %" PRIu64,
+                 restore_rate_limit);
+  ROCKS_LOG_INFO(logger, "Options.max_background_operations: %d",
+                 max_background_operations);
+}
+
+namespace {
+// -------- BackupEngineImpl class ---------
+class BackupEngineImpl {
+ public:
+  BackupEngineImpl(const BackupEngineOptions& options, Env* db_env,
+                   bool read_only = false);
+  ~BackupEngineImpl();
+
+  IOStatus CreateNewBackupWithMetadata(const CreateBackupOptions& options,
+                                       DB* db, const std::string& app_metadata,
+                                       BackupID* new_backup_id_ptr);
+
+  IOStatus PurgeOldBackups(uint32_t num_backups_to_keep);
+
+  IOStatus DeleteBackup(BackupID backup_id);
+
+  void StopBackup() { stop_backup_.store(true, std::memory_order_release); }
+
+  IOStatus GarbageCollect();
+
+  // The returned BackupInfos are in chronological order, which means the
+  // latest backup comes last.
+  void GetBackupInfo(std::vector<BackupInfo>* backup_info,
+                     bool include_file_details) const;
+
+  Status GetBackupInfo(BackupID backup_id, BackupInfo* backup_info,
+                       bool include_file_details = false) const;
+
+  void GetCorruptedBackups(std::vector<BackupID>* corrupt_backup_ids) const;
+
+  IOStatus RestoreDBFromBackup(const RestoreOptions& options,
+                               BackupID backup_id, const std::string& db_dir,
+                               const std::string& wal_dir) const;
+
+  IOStatus RestoreDBFromLatestBackup(const RestoreOptions& options,
+                                     const std::string& db_dir,
+                                     const std::string& wal_dir) const {
+    // Note: don't read latest_valid_backup_id_ outside of lock
+    return RestoreDBFromBackup(options, kLatestBackupIDMarker, db_dir, wal_dir);
+  }
+
+  IOStatus VerifyBackup(BackupID backup_id,
+                        bool verify_with_checksum = false) const;
+
+  IOStatus Initialize();
+
+  ShareFilesNaming GetNamingNoFlags() const {
+    return options_.share_files_with_checksum_naming &
+           BackupEngineOptions::kMaskNoNamingFlags;
+  }
+  ShareFilesNaming GetNamingFlags() const {
+    return options_.share_files_with_checksum_naming &
+           BackupEngineOptions::kMaskNamingFlags;
+  }
+
+  void TEST_SetDefaultRateLimitersClock(
+      const std::shared_ptr<SystemClock>& backup_rate_limiter_clock,
+      const std::shared_ptr<SystemClock>& restore_rate_limiter_clock) {
+    if (backup_rate_limiter_clock) {
+      static_cast<GenericRateLimiter*>(options_.backup_rate_limiter.get())
+          ->TEST_SetClock(backup_rate_limiter_clock);
+    }
+
+    if (restore_rate_limiter_clock) {
+      static_cast<GenericRateLimiter*>(options_.restore_rate_limiter.get())
+          ->TEST_SetClock(restore_rate_limiter_clock);
+    }
+  }
+
+ private:
+  void DeleteChildren(const std::string& dir,
+                      uint32_t file_type_filter = 0) const;
+  IOStatus DeleteBackupNoGC(BackupID backup_id);
+
+  // Extends the "result" map with pathname->size mappings for the contents of
+  // "dir" in "env". Pathnames are prefixed with "dir".
+  IOStatus ReadChildFileCurrentSizes(
+      const std::string& dir, const std::shared_ptr<FileSystem>&,
+      std::unordered_map<std::string, uint64_t>* result) const;
+
+  struct FileInfo {
+    FileInfo(const std::string& fname, uint64_t sz, const std::string& checksum,
+             const std::string& id, const std::string& sid, Temperature _temp)
+        : refs(0),
+          filename(fname),
+          size(sz),
+          checksum_hex(checksum),
+          db_id(id),
+          db_session_id(sid),
+          temp(_temp) {}
+
+    FileInfo(const FileInfo&) = delete;
+    FileInfo& operator=(const FileInfo&) = delete;
+
+    int refs;
+    const std::string filename;
+    const uint64_t size;
+    // crc32c checksum as hex. empty == unknown / unavailable
+    std::string checksum_hex;
+    // DB identities
+    // db_id is obtained for potential usage in the future but not used
+    // currently
+    const std::string db_id;
+    // db_session_id appears in the backup SST filename if the table naming
+    // option is kUseDbSessionId
+    const std::string db_session_id;
+    Temperature temp;
+
+    std::string GetDbFileName() {
+      std::string rv;
+      // extract the filename part
+      size_t slash = filename.find_last_of('/');
+      // file will either be shared/<file>, shared_checksum/<file_crc32c_size>,
+      // shared_checksum/<file_session>, shared_checksum/<file_crc32c_session>,
+      // or private/<number>/<file>
+      assert(slash != std::string::npos);
+      rv = filename.substr(slash + 1);
+
+      // if the file was in shared_checksum, extract the real file name
+      // in this case the file is <number>_<checksum>_<size>.<type>,
+      // <number>_<session>.<type>, or <number>_<checksum>_<session>.<type>
+      if (filename.substr(0, slash) == kSharedChecksumDirName) {
+        rv = GetFileFromChecksumFile(rv);
+      }
+      return rv;
+    }
+  };
+
+  // TODO: deprecate this function once we migrate all BackupEngine's rate
+  // limiting to lower-level ones (i.e, ones in file access wrapper level like
+  // `WritableFileWriter`)
+  static void LoopRateLimitRequestHelper(const size_t total_bytes_to_request,
+                                         RateLimiter* rate_limiter,
+                                         const Env::IOPriority pri,
+                                         Statistics* stats,
+                                         const RateLimiter::OpType op_type);
+
+  static inline std::string WithoutTrailingSlash(const std::string& path) {
+    if (path.empty() || path.back() != '/') {
+      return path;
+    } else {
+      return path.substr(path.size() - 1);
+    }
+  }
+
+  static inline std::string WithTrailingSlash(const std::string& path) {
+    if (path.empty() || path.back() != '/') {
+      return path + '/';
+    } else {
+      return path;
+    }
+  }
+
+  // A filesystem wrapper that makes shared backup files appear to be in the
+  // private backup directory (dst_dir), so that the private backup dir can
+  // be opened as a read-only DB.
+  class RemapSharedFileSystem : public RemapFileSystem {
+   public:
+    RemapSharedFileSystem(const std::shared_ptr<FileSystem>& base,
+                          const std::string& dst_dir,
+                          const std::string& src_base_dir,
+                          const std::vector<std::shared_ptr<FileInfo>>& files)
+        : RemapFileSystem(base),
+          dst_dir_(WithoutTrailingSlash(dst_dir)),
+          dst_dir_slash_(WithTrailingSlash(dst_dir)),
+          src_base_dir_(WithTrailingSlash(src_base_dir)) {
+      for (auto& info : files) {
+        if (!StartsWith(info->filename, kPrivateDirSlash)) {
+          assert(StartsWith(info->filename, kSharedDirSlash) ||
+                 StartsWith(info->filename, kSharedChecksumDirSlash));
+          remaps_[info->GetDbFileName()] = info;
+        }
+      }
+    }
+
+    const char* Name() const override {
+      return "BackupEngineImpl::RemapSharedFileSystem";
+    }
+
+    // Sometimes a directory listing is required in opening a DB
+    IOStatus GetChildren(const std::string& dir, const IOOptions& options,
+                         std::vector<std::string>* result,
+                         IODebugContext* dbg) override {
+      IOStatus s = RemapFileSystem::GetChildren(dir, options, result, dbg);
+      if (s.ok() && (dir == dst_dir_ || dir == dst_dir_slash_)) {
+        // Assume remapped files exist
+        for (auto& r : remaps_) {
+          result->push_back(r.first);
+        }
+      }
+      return s;
+    }
+
+    // Sometimes a directory listing is required in opening a DB
+    IOStatus GetChildrenFileAttributes(const std::string& dir,
+                                       const IOOptions& options,
+                                       std::vector<FileAttributes>* result,
+                                       IODebugContext* dbg) override {
+      IOStatus s =
+          RemapFileSystem::GetChildrenFileAttributes(dir, options, result, dbg);
+      if (s.ok() && (dir == dst_dir_ || dir == dst_dir_slash_)) {
+        // Assume remapped files exist with recorded size
+        for (auto& r : remaps_) {
+          result->emplace_back();  // clean up with C++20
+          FileAttributes& attr = result->back();
+          attr.name = r.first;
+          attr.size_bytes = r.second->size;
+        }
+      }
+      return s;
+    }
+
+   protected:
+    // When a file in dst_dir is requested, see if we need to remap to shared
+    // file path.
+    std::pair<IOStatus, std::string> EncodePath(
+        const std::string& path) override {
+      if (path.empty() || path[0] != '/') {
+        return {IOStatus::InvalidArgument(path, "Not an absolute path"), ""};
+      }
+      std::pair<IOStatus, std::string> rv{IOStatus(), path};
+      if (StartsWith(path, dst_dir_slash_)) {
+        std::string relative = path.substr(dst_dir_slash_.size());
+        auto it = remaps_.find(relative);
+        if (it != remaps_.end()) {
+          rv.second = src_base_dir_ + it->second->filename;
+        }
+      }
+      return rv;
+    }
+
+   private:
+    // Absolute path to a directory that some extra files will be mapped into.
+    const std::string dst_dir_;
+    // Includes a trailing slash.
+    const std::string dst_dir_slash_;
+    // Absolute path to a directory containing some files to be mapped into
+    // dst_dir_. Includes a trailing slash.
+    const std::string src_base_dir_;
+    // If remaps_[x] exists, attempt to read dst_dir_ / x should instead read
+    // src_base_dir_ / remaps_[x]->filename. FileInfo is used to maximize
+    // sharing with other backup data in memory.
+    std::unordered_map<std::string, std::shared_ptr<FileInfo>> remaps_;
+  };
+
+  class BackupMeta {
+   public:
+    BackupMeta(
+        const std::string& meta_filename, const std::string& meta_tmp_filename,
+        std::unordered_map<std::string, std::shared_ptr<FileInfo>>* file_infos,
+        Env* env, const std::shared_ptr<FileSystem>& fs)
+        : timestamp_(0),
+          sequence_number_(0),
+          size_(0),
+          meta_filename_(meta_filename),
+          meta_tmp_filename_(meta_tmp_filename),
+          file_infos_(file_infos),
+          env_(env),
+          fs_(fs) {}
+
+    BackupMeta(const BackupMeta&) = delete;
+    BackupMeta& operator=(const BackupMeta&) = delete;
+
+    ~BackupMeta() {}
+
+    void RecordTimestamp() {
+      // Best effort
+      Status s = env_->GetCurrentTime(&timestamp_);
+      if (!s.ok()) {
+        timestamp_ = /* something clearly fabricated */ 1;
+      }
+    }
+    int64_t GetTimestamp() const { return timestamp_; }
+    uint64_t GetSize() const { return size_; }
+    uint32_t GetNumberFiles() const {
+      return static_cast<uint32_t>(files_.size());
+    }
+    void SetSequenceNumber(uint64_t sequence_number) {
+      sequence_number_ = sequence_number;
+    }
+    uint64_t GetSequenceNumber() const { return sequence_number_; }
+
+    const std::string& GetAppMetadata() const { return app_metadata_; }
+
+    void SetAppMetadata(const std::string& app_metadata) {
+      app_metadata_ = app_metadata;
+    }
+
+    IOStatus AddFile(std::shared_ptr<FileInfo> file_info);
+
+    IOStatus Delete(bool delete_meta = true);
+
+    bool Empty() const { return files_.empty(); }
+
+    std::shared_ptr<FileInfo> GetFile(const std::string& filename) const {
+      auto it = file_infos_->find(filename);
+      if (it == file_infos_->end()) {
+        return nullptr;
+      }
+      return it->second;
+    }
+
+    const std::vector<std::shared_ptr<FileInfo>>& GetFiles() const {
+      return files_;
+    }
+
+    // @param abs_path_to_size Pre-fetched file sizes (bytes).
+    IOStatus LoadFromFile(
+        const std::string& backup_dir,
+        const std::unordered_map<std::string, uint64_t>& abs_path_to_size,
+        RateLimiter* rate_limiter, Logger* info_log,
+        std::unordered_set<std::string>* reported_ignored_fields);
+    IOStatus StoreToFile(
+        bool sync, int schema_version,
+        const TEST_BackupMetaSchemaOptions* schema_test_options);
+
+    std::string GetInfoString() {
+      std::ostringstream ss;
+      ss << "Timestamp: " << timestamp_ << std::endl;
+      char human_size[16];
+      AppendHumanBytes(size_, human_size, sizeof(human_size));
+      ss << "Size: " << human_size << std::endl;
+      ss << "Files:" << std::endl;
+      for (const auto& file : files_) {
+        AppendHumanBytes(file->size, human_size, sizeof(human_size));
+        ss << file->filename << ", size " << human_size << ", refs "
+           << file->refs << std::endl;
+      }
+      return ss.str();
+    }
+
+    const std::shared_ptr<Env>& GetEnvForOpen() const {
+      if (!env_for_open_) {
+        // Lazy initialize
+        // Find directories
+        std::string dst_dir = meta_filename_;
+        auto i = dst_dir.rfind(kMetaDirSlash);
+        assert(i != std::string::npos);
+        std::string src_base_dir = dst_dir.substr(0, i);
+        dst_dir.replace(i, kMetaDirSlash.size(), kPrivateDirSlash);
+        // Make the RemapSharedFileSystem
+        std::shared_ptr<FileSystem> remap_fs =
+            std::make_shared<RemapSharedFileSystem>(fs_, dst_dir, src_base_dir,
+                                                    files_);
+        // Make it read-only for safety
+        remap_fs = std::make_shared<ReadOnlyFileSystem>(remap_fs);
+        // Make an Env wrapper
+        env_for_open_ = std::make_shared<CompositeEnvWrapper>(env_, remap_fs);
+      }
+      return env_for_open_;
+    }
+
+   private:
+    int64_t timestamp_;
+    // sequence number is only approximate, should not be used
+    // by clients
+    uint64_t sequence_number_;
+    uint64_t size_;
+    std::string app_metadata_;
+    std::string const meta_filename_;
+    std::string const meta_tmp_filename_;
+    // files with relative paths (without "/" prefix!!)
+    std::vector<std::shared_ptr<FileInfo>> files_;
+    std::unordered_map<std::string, std::shared_ptr<FileInfo>>* file_infos_;
+    Env* env_;
+    mutable std::shared_ptr<Env> env_for_open_;
+    std::shared_ptr<FileSystem> fs_;
+    IOOptions iooptions_ = IOOptions();
+  };  // BackupMeta
+
+  void SetBackupInfoFromBackupMeta(BackupID id, const BackupMeta& meta,
+                                   BackupInfo* backup_info,
+                                   bool include_file_details) const;
+
+  inline std::string GetAbsolutePath(
+      const std::string& relative_path = "") const {
+    assert(relative_path.size() == 0 || relative_path[0] != '/');
+    return options_.backup_dir + "/" + relative_path;
+  }
+  inline std::string GetPrivateFileRel(BackupID backup_id, bool tmp = false,
+                                       const std::string& file = "") const {
+    assert(file.size() == 0 || file[0] != '/');
+    return kPrivateDirSlash + std::to_string(backup_id) + (tmp ? ".tmp" : "") +
+           "/" + file;
+  }
+  inline std::string GetSharedFileRel(const std::string& file = "",
+                                      bool tmp = false) const {
+    assert(file.size() == 0 || file[0] != '/');
+    return kSharedDirSlash + std::string(tmp ? "." : "") + file +
+           (tmp ? ".tmp" : "");
+  }
+  inline std::string GetSharedFileWithChecksumRel(const std::string& file = "",
+                                                  bool tmp = false) const {
+    assert(file.size() == 0 || file[0] != '/');
+    return kSharedChecksumDirSlash + std::string(tmp ? "." : "") + file +
+           (tmp ? ".tmp" : "");
+  }
+  inline bool UseLegacyNaming(const std::string& sid) const {
+    return GetNamingNoFlags() ==
+               BackupEngineOptions::kLegacyCrc32cAndFileSize ||
+           sid.empty();
+  }
+  inline std::string GetSharedFileWithChecksum(
+      const std::string& file, const std::string& checksum_hex,
+      const uint64_t file_size, const std::string& db_session_id) const {
+    assert(file.size() == 0 || file[0] != '/');
+    std::string file_copy = file;
+    if (UseLegacyNaming(db_session_id)) {
+      assert(!checksum_hex.empty());
+      file_copy.insert(file_copy.find_last_of('.'),
+                       "_" + std::to_string(ChecksumHexToInt32(checksum_hex)) +
+                           "_" + std::to_string(file_size));
+    } else {
+      file_copy.insert(file_copy.find_last_of('.'), "_s" + db_session_id);
+      if (GetNamingFlags() & BackupEngineOptions::kFlagIncludeFileSize) {
+        file_copy.insert(file_copy.find_last_of('.'),
+                         "_" + std::to_string(file_size));
+      }
+    }
+    return file_copy;
+  }
+  static inline std::string GetFileFromChecksumFile(const std::string& file) {
+    assert(file.size() == 0 || file[0] != '/');
+    std::string file_copy = file;
+    size_t first_underscore = file_copy.find_first_of('_');
+    return file_copy.erase(first_underscore,
+                           file_copy.find_last_of('.') - first_underscore);
+  }
+  inline std::string GetBackupMetaFile(BackupID backup_id, bool tmp) const {
+    return GetAbsolutePath(kMetaDirName) + "/" + (tmp ? "." : "") +
+           std::to_string(backup_id) + (tmp ? ".tmp" : "");
+  }
+
+  // If size_limit == 0, there is no size limit, copy everything.
+  //
+  // Exactly one of src and contents must be non-empty.
+  //
+  // @param src If non-empty, the file is copied from this pathname.
+  // @param contents If non-empty, the file will be created with these contents.
+  // @param src_temperature Pass in expected temperature of src, return back
+  // temperature reported by FileSystem
+  IOStatus CopyOrCreateFile(const std::string& src, const std::string& dst,
+                            const std::string& contents, uint64_t size_limit,
+                            Env* src_env, Env* dst_env,
+                            const EnvOptions& src_env_options, bool sync,
+                            RateLimiter* rate_limiter,
+                            std::function<void()> progress_callback,
+                            Temperature* src_temperature,
+                            Temperature dst_temperature,
+                            uint64_t* bytes_toward_next_callback,
+                            uint64_t* size, std::string* checksum_hex);
+
+  IOStatus ReadFileAndComputeChecksum(const std::string& src,
+                                      const std::shared_ptr<FileSystem>& src_fs,
+                                      const EnvOptions& src_env_options,
+                                      uint64_t size_limit,
+                                      std::string* checksum_hex,
+                                      const Temperature src_temperature) const;
+
+  // Obtain db_id and db_session_id from the table properties of file_path
+  Status GetFileDbIdentities(Env* src_env, const EnvOptions& src_env_options,
+                             const std::string& file_path,
+                             Temperature file_temp, RateLimiter* rate_limiter,
+                             std::string* db_id, std::string* db_session_id);
+
+  struct CopyOrCreateResult {
+    ~CopyOrCreateResult() {
+      // The Status needs to be ignored here for two reasons.
+      // First, if the BackupEngineImpl shuts down with jobs outstanding, then
+      // it is possible that the Status in the future/promise is never read,
+      // resulting in an unchecked Status. Second, if there are items in the
+      // channel when the BackupEngineImpl is shutdown, these will also have
+      // Status that have not been checked.  This
+      // TODO: Fix those issues so that the Status
+      io_status.PermitUncheckedError();
+    }
+    uint64_t size;
+    std::string checksum_hex;
+    std::string db_id;
+    std::string db_session_id;
+    IOStatus io_status;
+    Temperature expected_src_temperature = Temperature::kUnknown;
+    Temperature current_src_temperature = Temperature::kUnknown;
+  };
+
+  // Exactly one of src_path and contents must be non-empty. If src_path is
+  // non-empty, the file is copied from this pathname. Otherwise, if contents is
+  // non-empty, the file will be created at dst_path with these contents.
+  struct CopyOrCreateWorkItem {
+    std::string src_path;
+    std::string dst_path;
+    Temperature src_temperature;
+    Temperature dst_temperature;
+    std::string contents;
+    Env* src_env;
+    Env* dst_env;
+    EnvOptions src_env_options;
+    bool sync;
+    RateLimiter* rate_limiter;
+    uint64_t size_limit;
+    Statistics* stats;
+    std::promise<CopyOrCreateResult> result;
+    std::function<void()> progress_callback;
+    std::string src_checksum_func_name;
+    std::string src_checksum_hex;
+    std::string db_id;
+    std::string db_session_id;
+
+    CopyOrCreateWorkItem()
+        : src_path(""),
+          dst_path(""),
+          src_temperature(Temperature::kUnknown),
+          dst_temperature(Temperature::kUnknown),
+          contents(""),
+          src_env(nullptr),
+          dst_env(nullptr),
+          src_env_options(),
+          sync(false),
+          rate_limiter(nullptr),
+          size_limit(0),
+          stats(nullptr),
+          src_checksum_func_name(kUnknownFileChecksumFuncName),
+          src_checksum_hex(""),
+          db_id(""),
+          db_session_id("") {}
+
+    CopyOrCreateWorkItem(const CopyOrCreateWorkItem&) = delete;
+    CopyOrCreateWorkItem& operator=(const CopyOrCreateWorkItem&) = delete;
+
+    CopyOrCreateWorkItem(CopyOrCreateWorkItem&& o) noexcept {
+      *this = std::move(o);
+    }
+
+    CopyOrCreateWorkItem& operator=(CopyOrCreateWorkItem&& o) noexcept {
+      src_path = std::move(o.src_path);
+      dst_path = std::move(o.dst_path);
+      src_temperature = std::move(o.src_temperature);
+      dst_temperature = std::move(o.dst_temperature);
+      contents = std::move(o.contents);
+      src_env = o.src_env;
+      dst_env = o.dst_env;
+      src_env_options = std::move(o.src_env_options);
+      sync = o.sync;
+      rate_limiter = o.rate_limiter;
+      size_limit = o.size_limit;
+      stats = o.stats;
+      result = std::move(o.result);
+      progress_callback = std::move(o.progress_callback);
+      src_checksum_func_name = std::move(o.src_checksum_func_name);
+      src_checksum_hex = std::move(o.src_checksum_hex);
+      db_id = std::move(o.db_id);
+      db_session_id = std::move(o.db_session_id);
+      src_temperature = o.src_temperature;
+      return *this;
+    }
+
+    CopyOrCreateWorkItem(
+        std::string _src_path, std::string _dst_path,
+        const Temperature _src_temperature, const Temperature _dst_temperature,
+        std::string _contents, Env* _src_env, Env* _dst_env,
+        EnvOptions _src_env_options, bool _sync, RateLimiter* _rate_limiter,
+        uint64_t _size_limit, Statistics* _stats,
+        std::function<void()> _progress_callback = []() {},
+        const std::string& _src_checksum_func_name =
+            kUnknownFileChecksumFuncName,
+        const std::string& _src_checksum_hex = "",
+        const std::string& _db_id = "", const std::string& _db_session_id = "")
+        : src_path(std::move(_src_path)),
+          dst_path(std::move(_dst_path)),
+          src_temperature(_src_temperature),
+          dst_temperature(_dst_temperature),
+          contents(std::move(_contents)),
+          src_env(_src_env),
+          dst_env(_dst_env),
+          src_env_options(std::move(_src_env_options)),
+          sync(_sync),
+          rate_limiter(_rate_limiter),
+          size_limit(_size_limit),
+          stats(_stats),
+          progress_callback(_progress_callback),
+          src_checksum_func_name(_src_checksum_func_name),
+          src_checksum_hex(_src_checksum_hex),
+          db_id(_db_id),
+          db_session_id(_db_session_id) {}
+  };
+
+  struct BackupAfterCopyOrCreateWorkItem {
+    std::future<CopyOrCreateResult> result;
+    bool shared;
+    bool needed_to_copy;
+    Env* backup_env;
+    std::string dst_path_tmp;
+    std::string dst_path;
+    std::string dst_relative;
+    BackupAfterCopyOrCreateWorkItem()
+        : shared(false),
+          needed_to_copy(false),
+          backup_env(nullptr),
+          dst_path_tmp(""),
+          dst_path(""),
+          dst_relative("") {}
+
+    BackupAfterCopyOrCreateWorkItem(
+        BackupAfterCopyOrCreateWorkItem&& o) noexcept {
+      *this = std::move(o);
+    }
+
+    BackupAfterCopyOrCreateWorkItem& operator=(
+        BackupAfterCopyOrCreateWorkItem&& o) noexcept {
+      result = std::move(o.result);
+      shared = o.shared;
+      needed_to_copy = o.needed_to_copy;
+      backup_env = o.backup_env;
+      dst_path_tmp = std::move(o.dst_path_tmp);
+      dst_path = std::move(o.dst_path);
+      dst_relative = std::move(o.dst_relative);
+      return *this;
+    }
+
+    BackupAfterCopyOrCreateWorkItem(std::future<CopyOrCreateResult>&& _result,
+                                    bool _shared, bool _needed_to_copy,
+                                    Env* _backup_env, std::string _dst_path_tmp,
+                                    std::string _dst_path,
+                                    std::string _dst_relative)
+        : result(std::move(_result)),
+          shared(_shared),
+          needed_to_copy(_needed_to_copy),
+          backup_env(_backup_env),
+          dst_path_tmp(std::move(_dst_path_tmp)),
+          dst_path(std::move(_dst_path)),
+          dst_relative(std::move(_dst_relative)) {}
+  };
+
+  struct RestoreAfterCopyOrCreateWorkItem {
+    std::future<CopyOrCreateResult> result;
+    std::string from_file;
+    std::string to_file;
+    std::string checksum_hex;
+    RestoreAfterCopyOrCreateWorkItem() : checksum_hex("") {}
+    RestoreAfterCopyOrCreateWorkItem(std::future<CopyOrCreateResult>&& _result,
+                                     const std::string& _from_file,
+                                     const std::string& _to_file,
+                                     const std::string& _checksum_hex)
+        : result(std::move(_result)),
+          from_file(_from_file),
+          to_file(_to_file),
+          checksum_hex(_checksum_hex) {}
+    RestoreAfterCopyOrCreateWorkItem(
+        RestoreAfterCopyOrCreateWorkItem&& o) noexcept {
+      *this = std::move(o);
+    }
+
+    RestoreAfterCopyOrCreateWorkItem& operator=(
+        RestoreAfterCopyOrCreateWorkItem&& o) noexcept {
+      result = std::move(o.result);
+      checksum_hex = std::move(o.checksum_hex);
+      return *this;
+    }
+  };
+
+  bool initialized_;
+  std::mutex byte_report_mutex_;
+  mutable channel<CopyOrCreateWorkItem> files_to_copy_or_create_;
+  std::vector<port::Thread> threads_;
+  std::atomic<CpuPriority> threads_cpu_priority_;
+
+  // Certain operations like PurgeOldBackups and DeleteBackup will trigger
+  // automatic GarbageCollect (true) unless we've already done one in this
+  // session and have not failed to delete backup files since then (false).
+  bool might_need_garbage_collect_ = true;
+
+  // Adds a file to the backup work queue to be copied or created if it doesn't
+  // already exist.
+  //
+  // Exactly one of src_dir and contents must be non-empty.
+  //
+  // @param src_dir If non-empty, the file in this directory named fname will be
+  //    copied.
+  // @param fname Name of destination file and, in case of copy, source file.
+  // @param contents If non-empty, the file will be created with these contents.
+  IOStatus AddBackupFileWorkItem(
+      std::unordered_set<std::string>& live_dst_paths,
+      std::vector<BackupAfterCopyOrCreateWorkItem>& backup_items_to_finish,
+      BackupID backup_id, bool shared, const std::string& src_dir,
+      const std::string& fname,  // starts with "/"
+      const EnvOptions& src_env_options, RateLimiter* rate_limiter,
+      FileType file_type, uint64_t size_bytes, Statistics* stats,
+      uint64_t size_limit = 0, bool shared_checksum = false,
+      std::function<void()> progress_callback = []() {},
+      const std::string& contents = std::string(),
+      const std::string& src_checksum_func_name = kUnknownFileChecksumFuncName,
+      const std::string& src_checksum_str = kUnknownFileChecksum,
+      const Temperature src_temperature = Temperature::kUnknown);
+
+  // backup state data
+  BackupID latest_backup_id_;
+  BackupID latest_valid_backup_id_;
+  std::map<BackupID, std::unique_ptr<BackupMeta>> backups_;
+  std::map<BackupID, std::pair<IOStatus, std::unique_ptr<BackupMeta>>>
+      corrupt_backups_;
+  std::unordered_map<std::string, std::shared_ptr<FileInfo>>
+      backuped_file_infos_;
+  std::atomic<bool> stop_backup_;
+
+  // options data
+  BackupEngineOptions options_;
+  Env* db_env_;
+  Env* backup_env_;
+
+  // directories
+  std::unique_ptr<FSDirectory> backup_directory_;
+  std::unique_ptr<FSDirectory> shared_directory_;
+  std::unique_ptr<FSDirectory> meta_directory_;
+  std::unique_ptr<FSDirectory> private_directory_;
+
+  static const size_t kDefaultCopyFileBufferSize = 5 * 1024 * 1024LL;  // 5MB
+  bool read_only_;
+  BackupStatistics backup_statistics_;
+  std::unordered_set<std::string> reported_ignored_fields_;
+  static const size_t kMaxAppMetaSize = 1024 * 1024;  // 1MB
+  std::shared_ptr<FileSystem> db_fs_;
+  std::shared_ptr<FileSystem> backup_fs_;
+  IOOptions io_options_ = IOOptions();
+
+ public:
+  std::unique_ptr<TEST_BackupMetaSchemaOptions> schema_test_options_;
+};
+
+// -------- BackupEngineImplThreadSafe class ---------
+// This locking layer for thread safety in the public API is layered on
+// top to prevent accidental recursive locking with RWMutex, which is UB.
+// Note: BackupEngineReadOnlyBase inherited twice, but has no fields
+class BackupEngineImplThreadSafe : public BackupEngine,
+                                   public BackupEngineReadOnly {
+ public:
+  BackupEngineImplThreadSafe(const BackupEngineOptions& options, Env* db_env,
+                             bool read_only = false)
+      : impl_(options, db_env, read_only) {}
+  ~BackupEngineImplThreadSafe() override {}
+
+  using BackupEngine::CreateNewBackupWithMetadata;
+  IOStatus CreateNewBackupWithMetadata(const CreateBackupOptions& options,
+                                       DB* db, const std::string& app_metadata,
+                                       BackupID* new_backup_id) override {
+    WriteLock lock(&mutex_);
+    return impl_.CreateNewBackupWithMetadata(options, db, app_metadata,
+                                             new_backup_id);
+  }
+
+  IOStatus PurgeOldBackups(uint32_t num_backups_to_keep) override {
+    WriteLock lock(&mutex_);
+    return impl_.PurgeOldBackups(num_backups_to_keep);
+  }
+
+  IOStatus DeleteBackup(BackupID backup_id) override {
+    WriteLock lock(&mutex_);
+    return impl_.DeleteBackup(backup_id);
+  }
+
+  void StopBackup() override {
+    // No locking needed
+    impl_.StopBackup();
+  }
+
+  IOStatus GarbageCollect() override {
+    WriteLock lock(&mutex_);
+    return impl_.GarbageCollect();
+  }
+
+  Status GetLatestBackupInfo(BackupInfo* backup_info,
+                             bool include_file_details = false) const override {
+    ReadLock lock(&mutex_);
+    return impl_.GetBackupInfo(kLatestBackupIDMarker, backup_info,
+                               include_file_details);
+  }
+
+  Status GetBackupInfo(BackupID backup_id, BackupInfo* backup_info,
+                       bool include_file_details = false) const override {
+    ReadLock lock(&mutex_);
+    return impl_.GetBackupInfo(backup_id, backup_info, include_file_details);
+  }
+
+  void GetBackupInfo(std::vector<BackupInfo>* backup_info,
+                     bool include_file_details) const override {
+    ReadLock lock(&mutex_);
+    impl_.GetBackupInfo(backup_info, include_file_details);
+  }
+
+  void GetCorruptedBackups(
+      std::vector<BackupID>* corrupt_backup_ids) const override {
+    ReadLock lock(&mutex_);
+    impl_.GetCorruptedBackups(corrupt_backup_ids);
+  }
+
+  using BackupEngine::RestoreDBFromBackup;
+  IOStatus RestoreDBFromBackup(const RestoreOptions& options,
+                               BackupID backup_id, const std::string& db_dir,
+                               const std::string& wal_dir) const override {
+    ReadLock lock(&mutex_);
+    return impl_.RestoreDBFromBackup(options, backup_id, db_dir, wal_dir);
+  }
+
+  using BackupEngine::RestoreDBFromLatestBackup;
+  IOStatus RestoreDBFromLatestBackup(
+      const RestoreOptions& options, const std::string& db_dir,
+      const std::string& wal_dir) const override {
+    // Defer to above function, which locks
+    return RestoreDBFromBackup(options, kLatestBackupIDMarker, db_dir, wal_dir);
+  }
+
+  IOStatus VerifyBackup(BackupID backup_id,
+                        bool verify_with_checksum = false) const override {
+    ReadLock lock(&mutex_);
+    return impl_.VerifyBackup(backup_id, verify_with_checksum);
+  }
+
+  // Not public API but needed
+  IOStatus Initialize() {
+    // No locking needed
+    return impl_.Initialize();
+  }
+
+  // Not public API but used in testing
+  void TEST_SetBackupMetaSchemaOptions(
+      const TEST_BackupMetaSchemaOptions& options) {
+    impl_.schema_test_options_.reset(new TEST_BackupMetaSchemaOptions(options));
+  }
+
+  // Not public API but used in testing
+  void TEST_SetDefaultRateLimitersClock(
+      const std::shared_ptr<SystemClock>& backup_rate_limiter_clock = nullptr,
+      const std::shared_ptr<SystemClock>& restore_rate_limiter_clock =
+          nullptr) {
+    impl_.TEST_SetDefaultRateLimitersClock(backup_rate_limiter_clock,
+                                           restore_rate_limiter_clock);
+  }
+
+ private:
+  mutable port::RWMutex mutex_;
+  BackupEngineImpl impl_;
+};
+}  // namespace
+
+IOStatus BackupEngine::Open(const BackupEngineOptions& options, Env* env,
+                            BackupEngine** backup_engine_ptr) {
+  std::unique_ptr<BackupEngineImplThreadSafe> backup_engine(
+      new BackupEngineImplThreadSafe(options, env));
+  auto s = backup_engine->Initialize();
+  if (!s.ok()) {
+    *backup_engine_ptr = nullptr;
+    return s;
+  }
+  *backup_engine_ptr = backup_engine.release();
+  return IOStatus::OK();
+}
+
+namespace {
+BackupEngineImpl::BackupEngineImpl(const BackupEngineOptions& options,
+                                   Env* db_env, bool read_only)
+    : initialized_(false),
+      threads_cpu_priority_(),
+      latest_backup_id_(0),
+      latest_valid_backup_id_(0),
+      stop_backup_(false),
+      options_(options),
+      db_env_(db_env),
+      backup_env_(options.backup_env != nullptr ? options.backup_env : db_env_),
+      read_only_(read_only) {
+  if (options_.backup_rate_limiter == nullptr &&
+      options_.backup_rate_limit > 0) {
+    options_.backup_rate_limiter.reset(
+        NewGenericRateLimiter(options_.backup_rate_limit));
+  }
+  if (options_.restore_rate_limiter == nullptr &&
+      options_.restore_rate_limit > 0) {
+    options_.restore_rate_limiter.reset(
+        NewGenericRateLimiter(options_.restore_rate_limit));
+  }
+  db_fs_ = db_env_->GetFileSystem();
+  backup_fs_ = backup_env_->GetFileSystem();
+}
+
+BackupEngineImpl::~BackupEngineImpl() {
+  files_to_copy_or_create_.sendEof();
+  for (auto& t : threads_) {
+    t.join();
+  }
+  LogFlush(options_.info_log);
+  for (const auto& it : corrupt_backups_) {
+    it.second.first.PermitUncheckedError();
+  }
+}
+
+IOStatus BackupEngineImpl::Initialize() {
+  assert(!initialized_);
+  initialized_ = true;
+  if (read_only_) {
+    ROCKS_LOG_INFO(options_.info_log, "Starting read_only backup engine");
+  }
+  options_.Dump(options_.info_log);
+
+  auto meta_path = GetAbsolutePath(kMetaDirName);
+
+  if (!read_only_) {
+    // we might need to clean up from previous crash or I/O errors
+    might_need_garbage_collect_ = true;
+
+    if (options_.max_valid_backups_to_open !=
+        std::numeric_limits<int32_t>::max()) {
+      options_.max_valid_backups_to_open = std::numeric_limits<int32_t>::max();
+      ROCKS_LOG_WARN(
+          options_.info_log,
+          "`max_valid_backups_to_open` is not set to the default value. "
+          "Ignoring its value since BackupEngine is not read-only.");
+    }
+
+    // gather the list of directories that we need to create
+    std::vector<std::pair<std::string, std::unique_ptr<FSDirectory>*>>
+        directories;
+    directories.emplace_back(GetAbsolutePath(), &backup_directory_);
+    if (options_.share_table_files) {
+      if (options_.share_files_with_checksum) {
+        directories.emplace_back(
+            GetAbsolutePath(GetSharedFileWithChecksumRel()),
+            &shared_directory_);
+      } else {
+        directories.emplace_back(GetAbsolutePath(GetSharedFileRel()),
+                                 &shared_directory_);
+      }
+    }
+    directories.emplace_back(GetAbsolutePath(kPrivateDirName),
+                             &private_directory_);
+    directories.emplace_back(meta_path, &meta_directory_);
+    // create all the dirs we need
+    for (const auto& d : directories) {
+      IOStatus io_s =
+          backup_fs_->CreateDirIfMissing(d.first, io_options_, nullptr);
+      if (io_s.ok()) {
+        io_s =
+            backup_fs_->NewDirectory(d.first, io_options_, d.second, nullptr);
+      }
+      if (!io_s.ok()) {
+        return io_s;
+      }
+    }
+  }
+
+  std::vector<std::string> backup_meta_files;
+  {
+    IOStatus io_s = backup_fs_->GetChildren(meta_path, io_options_,
+                                            &backup_meta_files, nullptr);
+    if (io_s.IsNotFound()) {
+      return IOStatus::NotFound(meta_path + " is missing");
+    } else if (!io_s.ok()) {
+      return io_s;
+    }
+  }
+  // create backups_ structure
+  for (auto& file : backup_meta_files) {
+    ROCKS_LOG_INFO(options_.info_log, "Detected backup %s", file.c_str());
+    BackupID backup_id = 0;
+    sscanf(file.c_str(), "%u", &backup_id);
+    if (backup_id == 0 || file != std::to_string(backup_id)) {
+      // Invalid file name, will be deleted with auto-GC when user
+      // initiates an append or write operation. (Behave as read-only until
+      // then.)
+      ROCKS_LOG_INFO(options_.info_log, "Skipping unrecognized meta file %s",
+                     file.c_str());
+      continue;
+    }
+    assert(backups_.find(backup_id) == backups_.end());
+    // Insert all the (backup_id, BackupMeta) that will be loaded later
+    // The loading performed later will check whether there are corrupt backups
+    // and move the corrupt backups to corrupt_backups_
+    backups_.insert(std::make_pair(
+        backup_id, std::unique_ptr<BackupMeta>(new BackupMeta(
+                       GetBackupMetaFile(backup_id, false /* tmp */),
+                       GetBackupMetaFile(backup_id, true /* tmp */),
+                       &backuped_file_infos_, backup_env_, backup_fs_))));
+  }
+
+  latest_backup_id_ = 0;
+  latest_valid_backup_id_ = 0;
+  if (options_.destroy_old_data) {  // Destroy old data
+    assert(!read_only_);
+    ROCKS_LOG_INFO(
+        options_.info_log,
+        "Backup Engine started with destroy_old_data == true, deleting all "
+        "backups");
+    IOStatus io_s = PurgeOldBackups(0);
+    if (io_s.ok()) {
+      io_s = GarbageCollect();
+    }
+    if (!io_s.ok()) {
+      return io_s;
+    }
+  } else {  // Load data from storage
+    // abs_path_to_size: maps absolute paths of files in backup directory to
+    // their corresponding sizes
+    std::unordered_map<std::string, uint64_t> abs_path_to_size;
+    // Insert files and their sizes in backup sub-directories (shared and
+    // shared_checksum) to abs_path_to_size
+    for (const auto& rel_dir :
+         {GetSharedFileRel(), GetSharedFileWithChecksumRel()}) {
+      const auto abs_dir = GetAbsolutePath(rel_dir);
+      IOStatus io_s =
+          ReadChildFileCurrentSizes(abs_dir, backup_fs_, &abs_path_to_size);
+      if (!io_s.ok()) {
+        // I/O error likely impacting all backups
+        return io_s;
+      }
+    }
+    // load the backups if any, until valid_backups_to_open of the latest
+    // non-corrupted backups have been successfully opened.
+    int valid_backups_to_open = options_.max_valid_backups_to_open;
+    for (auto backup_iter = backups_.rbegin(); backup_iter != backups_.rend();
+         ++backup_iter) {
+      assert(latest_backup_id_ == 0 || latest_backup_id_ > backup_iter->first);
+      if (latest_backup_id_ == 0) {
+        latest_backup_id_ = backup_iter->first;
+      }
+      if (valid_backups_to_open == 0) {
+        break;
+      }
+
+      // Insert files and their sizes in backup sub-directories
+      // (private/backup_id) to abs_path_to_size
+      IOStatus io_s = ReadChildFileCurrentSizes(
+          GetAbsolutePath(GetPrivateFileRel(backup_iter->first)), backup_fs_,
+          &abs_path_to_size);
+      if (io_s.ok()) {
+        io_s = backup_iter->second->LoadFromFile(
+            options_.backup_dir, abs_path_to_size,
+            options_.backup_rate_limiter.get(), options_.info_log,
+            &reported_ignored_fields_);
+      }
+      if (io_s.IsCorruption() || io_s.IsNotSupported()) {
+        ROCKS_LOG_INFO(options_.info_log, "Backup %u corrupted -- %s",
+                       backup_iter->first, io_s.ToString().c_str());
+        corrupt_backups_.insert(std::make_pair(
+            backup_iter->first,
+            std::make_pair(io_s, std::move(backup_iter->second))));
+      } else if (!io_s.ok()) {
+        // Distinguish corruption errors from errors in the backup Env.
+        // Errors in the backup Env (i.e., this code path) will cause Open() to
+        // fail, whereas corruption errors would not cause Open() failures.
+        return io_s;
+      } else {
+        ROCKS_LOG_INFO(options_.info_log, "Loading backup %" PRIu32 " OK:\n%s",
+                       backup_iter->first,
+                       backup_iter->second->GetInfoString().c_str());
+        assert(latest_valid_backup_id_ == 0 ||
+               latest_valid_backup_id_ > backup_iter->first);
+        if (latest_valid_backup_id_ == 0) {
+          latest_valid_backup_id_ = backup_iter->first;
+        }
+        --valid_backups_to_open;
+      }
+    }
+
+    for (const auto& corrupt : corrupt_backups_) {
+      backups_.erase(backups_.find(corrupt.first));
+    }
+    // erase the backups before max_valid_backups_to_open
+    int num_unopened_backups;
+    if (options_.max_valid_backups_to_open == 0) {
+      num_unopened_backups = 0;
+    } else {
+      num_unopened_backups =
+          std::max(0, static_cast<int>(backups_.size()) -
+                          options_.max_valid_backups_to_open);
+    }
+    for (int i = 0; i < num_unopened_backups; ++i) {
+      assert(backups_.begin()->second->Empty());
+      backups_.erase(backups_.begin());
+    }
+  }
+
+  ROCKS_LOG_INFO(options_.info_log, "Latest backup is %u", latest_backup_id_);
+  ROCKS_LOG_INFO(options_.info_log, "Latest valid backup is %u",
+                 latest_valid_backup_id_);
+
+  // set up threads perform copies from files_to_copy_or_create_ in the
+  // background
+  threads_cpu_priority_ = CpuPriority::kNormal;
+  threads_.reserve(options_.max_background_operations);
+  for (int t = 0; t < options_.max_background_operations; t++) {
+    threads_.emplace_back([this]() {
+#if defined(_GNU_SOURCE) && defined(__GLIBC_PREREQ)
+#if __GLIBC_PREREQ(2, 12)
+      pthread_setname_np(pthread_self(), "backup_engine");
+#endif
+#endif
+      CpuPriority current_priority = CpuPriority::kNormal;
+      CopyOrCreateWorkItem work_item;
+      uint64_t bytes_toward_next_callback = 0;
+      while (files_to_copy_or_create_.read(work_item)) {
+        CpuPriority priority = threads_cpu_priority_;
+        if (current_priority != priority) {
+          TEST_SYNC_POINT_CALLBACK(
+              "BackupEngineImpl::Initialize:SetCpuPriority", &priority);
+          port::SetCpuPriority(0, priority);
+          current_priority = priority;
+        }
+        // `bytes_read` and `bytes_written` stats are enabled based on
+        // compile-time support and cannot be dynamically toggled. So we do not
+        // need to worry about `PerfLevel` here, unlike many other
+        // `IOStatsContext` / `PerfContext` stats.
+        uint64_t prev_bytes_read = IOSTATS(bytes_read);
+        uint64_t prev_bytes_written = IOSTATS(bytes_written);
+
+        CopyOrCreateResult result;
+        Temperature temp = work_item.src_temperature;
+        result.io_status = CopyOrCreateFile(
+            work_item.src_path, work_item.dst_path, work_item.contents,
+            work_item.size_limit, work_item.src_env, work_item.dst_env,
+            work_item.src_env_options, work_item.sync, work_item.rate_limiter,
+            work_item.progress_callback, &temp, work_item.dst_temperature,
+            &bytes_toward_next_callback, &result.size, &result.checksum_hex);
+
+        RecordTick(work_item.stats, BACKUP_READ_BYTES,
+                   IOSTATS(bytes_read) - prev_bytes_read);
+        RecordTick(work_item.stats, BACKUP_WRITE_BYTES,
+                   IOSTATS(bytes_written) - prev_bytes_written);
+
+        result.db_id = work_item.db_id;
+        result.db_session_id = work_item.db_session_id;
+        result.expected_src_temperature = work_item.src_temperature;
+        result.current_src_temperature = temp;
+        if (result.io_status.ok() && !work_item.src_checksum_hex.empty()) {
+          // unknown checksum function name implies no db table file checksum in
+          // db manifest; work_item.src_checksum_hex not empty means
+          // backup engine has calculated its crc32c checksum for the table
+          // file; therefore, we are able to compare the checksums.
+          if (work_item.src_checksum_func_name ==
+                  kUnknownFileChecksumFuncName ||
+              work_item.src_checksum_func_name == kDbFileChecksumFuncName) {
+            if (work_item.src_checksum_hex != result.checksum_hex) {
+              std::string checksum_info(
+                  "Expected checksum is " + work_item.src_checksum_hex +
+                  " while computed checksum is " + result.checksum_hex);
+              result.io_status = IOStatus::Corruption(
+                  "Checksum mismatch after copying to " + work_item.dst_path +
+                  ": " + checksum_info);
+            }
+          } else {
+            // FIXME(peterd): dead code?
+            std::string checksum_function_info(
+                "Existing checksum function is " +
+                work_item.src_checksum_func_name +
+                " while provided checksum function is " +
+                kBackupFileChecksumFuncName);
+            ROCKS_LOG_INFO(
+                options_.info_log,
+                "Unable to verify checksum after copying to %s: %s\n",
+                work_item.dst_path.c_str(), checksum_function_info.c_str());
+          }
+        }
+        work_item.result.set_value(std::move(result));
+      }
+    });
+  }
+  ROCKS_LOG_INFO(options_.info_log, "Initialized BackupEngine");
+  return IOStatus::OK();
+}
+
+IOStatus BackupEngineImpl::CreateNewBackupWithMetadata(
+    const CreateBackupOptions& options, DB* db, const std::string& app_metadata,
+    BackupID* new_backup_id_ptr) {
+  assert(initialized_);
+  assert(!read_only_);
+  if (app_metadata.size() > kMaxAppMetaSize) {
+    return IOStatus::InvalidArgument("App metadata too large");
+  }
+
+  if (options.decrease_background_thread_cpu_priority) {
+    if (options.background_thread_cpu_priority < threads_cpu_priority_) {
+      threads_cpu_priority_.store(options.background_thread_cpu_priority);
+    }
+  }
+
+  BackupID new_backup_id = latest_backup_id_ + 1;
+
+  // `bytes_read` and `bytes_written` stats are enabled based on compile-time
+  // support and cannot be dynamically toggled. So we do not need to worry about
+  // `PerfLevel` here, unlike many other `IOStatsContext` / `PerfContext` stats.
+  uint64_t prev_bytes_read = IOSTATS(bytes_read);
+  uint64_t prev_bytes_written = IOSTATS(bytes_written);
+
+  assert(backups_.find(new_backup_id) == backups_.end());
+
+  auto private_dir = GetAbsolutePath(GetPrivateFileRel(new_backup_id));
+  IOStatus io_s = backup_fs_->FileExists(private_dir, io_options_, nullptr);
+  if (io_s.ok()) {
+    // maybe last backup failed and left partial state behind, clean it up.
+    // need to do this before updating backups_ such that a private dir
+    // named after new_backup_id will be cleaned up.
+    // (If an incomplete new backup is followed by an incomplete delete
+    // of the latest full backup, then there could be more than one next
+    // id with a private dir, the last thing to be deleted in delete
+    // backup, but all will be cleaned up with a GarbageCollect.)
+    io_s = GarbageCollect();
+  } else if (io_s.IsNotFound()) {
+    // normal case, the new backup's private dir doesn't exist yet
+    io_s = IOStatus::OK();
+  }
+
+  auto ret = backups_.insert(std::make_pair(
+      new_backup_id, std::unique_ptr<BackupMeta>(new BackupMeta(
+                         GetBackupMetaFile(new_backup_id, false /* tmp */),
+                         GetBackupMetaFile(new_backup_id, true /* tmp */),
+                         &backuped_file_infos_, backup_env_, backup_fs_))));
+  assert(ret.second == true);
+  auto& new_backup = ret.first->second;
+  new_backup->RecordTimestamp();
+  new_backup->SetAppMetadata(app_metadata);
+
+  auto start_backup = backup_env_->NowMicros();
+
+  ROCKS_LOG_INFO(options_.info_log,
+                 "Started the backup process -- creating backup %u",
+                 new_backup_id);
+
+  if (options_.share_table_files && !options_.share_files_with_checksum) {
+    ROCKS_LOG_WARN(options_.info_log,
+                   "BackupEngineOptions::share_files_with_checksum=false is "
+                   "DEPRECATED and could lead to data loss.");
+  }
+
+  if (io_s.ok()) {
+    io_s = backup_fs_->CreateDir(private_dir, io_options_, nullptr);
+  }
+
+  // A set into which we will insert the dst_paths that are calculated for live
+  // files and live WAL files.
+  // This is used to check whether a live files shares a dst_path with another
+  // live file.
+  std::unordered_set<std::string> live_dst_paths;
+
+  std::vector<BackupAfterCopyOrCreateWorkItem> backup_items_to_finish;
+  // Add a CopyOrCreateWorkItem to the channel for each live file
+  Status disabled = db->DisableFileDeletions();
+  DBOptions db_options = db->GetDBOptions();
+  Statistics* stats = db_options.statistics.get();
+  if (io_s.ok()) {
+    CheckpointImpl checkpoint(db);
+    uint64_t sequence_number = 0;
+    FileChecksumGenFactory* db_checksum_factory =
+        db_options.file_checksum_gen_factory.get();
+    const std::string kFileChecksumGenFactoryName =
+        "FileChecksumGenCrc32cFactory";
+    bool compare_checksum =
+        db_checksum_factory != nullptr &&
+                db_checksum_factory->Name() == kFileChecksumGenFactoryName
+            ? true
+            : false;
+    EnvOptions src_raw_env_options(db_options);
+    RateLimiter* rate_limiter = options_.backup_rate_limiter.get();
+    io_s = status_to_io_status(checkpoint.CreateCustomCheckpoint(
+        [&](const std::string& /*src_dirname*/, const std::string& /*fname*/,
+            FileType) {
+          // custom checkpoint will switch to calling copy_file_cb after it sees
+          // NotSupported returned from link_file_cb.
+          return IOStatus::NotSupported();
+        } /* link_file_cb */,
+        [&](const std::string& src_dirname, const std::string& fname,
+            uint64_t size_limit_bytes, FileType type,
+            const std::string& checksum_func_name,
+            const std::string& checksum_val,
+            const Temperature src_temperature) {
+          if (type == kWalFile && !options_.backup_log_files) {
+            return IOStatus::OK();
+          }
+          Log(options_.info_log, "add file for backup %s", fname.c_str());
+          uint64_t size_bytes = 0;
+          IOStatus io_st;
+          if (type == kTableFile || type == kBlobFile) {
+            io_st = db_fs_->GetFileSize(src_dirname + "/" + fname, io_options_,
+                                        &size_bytes, nullptr);
+            if (!io_st.ok()) {
+              Log(options_.info_log, "GetFileSize is failed: %s",
+                  io_st.ToString().c_str());
+              return io_st;
+            }
+          }
+          EnvOptions src_env_options;
+          switch (type) {
+            case kWalFile:
+              src_env_options =
+                  db_env_->OptimizeForLogRead(src_raw_env_options);
+              break;
+            case kTableFile:
+              src_env_options = db_env_->OptimizeForCompactionTableRead(
+                  src_raw_env_options, ImmutableDBOptions(db_options));
+              break;
+            case kDescriptorFile:
+              src_env_options =
+                  db_env_->OptimizeForManifestRead(src_raw_env_options);
+              break;
+            case kBlobFile:
+              src_env_options = db_env_->OptimizeForBlobFileRead(
+                  src_raw_env_options, ImmutableDBOptions(db_options));
+              break;
+            default:
+              // Other backed up files (like options file) are not read by live
+              // DB, so don't need to worry about avoiding mixing buffered and
+              // direct I/O. Just use plain defaults.
+              src_env_options = src_raw_env_options;
+              break;
+          }
+          io_st = AddBackupFileWorkItem(
+              live_dst_paths, backup_items_to_finish, new_backup_id,
+              options_.share_table_files &&
+                  (type == kTableFile || type == kBlobFile),
+              src_dirname, fname, src_env_options, rate_limiter, type,
+              size_bytes, db_options.statistics.get(), size_limit_bytes,
+              options_.share_files_with_checksum &&
+                  (type == kTableFile || type == kBlobFile),
+              options.progress_callback, "" /* contents */, checksum_func_name,
+              checksum_val, src_temperature);
+          return io_st;
+        } /* copy_file_cb */,
+        [&](const std::string& fname, const std::string& contents,
+            FileType type) {
+          Log(options_.info_log, "add file for backup %s", fname.c_str());
+          return AddBackupFileWorkItem(
+              live_dst_paths, backup_items_to_finish, new_backup_id,
+              false /* shared */, "" /* src_dir */, fname,
+              EnvOptions() /* src_env_options */, rate_limiter, type,
+              contents.size(), db_options.statistics.get(), 0 /* size_limit */,
+              false /* shared_checksum */, options.progress_callback, contents);
+        } /* create_file_cb */,
+        &sequence_number,
+        options.flush_before_backup ? 0 : std::numeric_limits<uint64_t>::max(),
+        compare_checksum));
+    if (io_s.ok()) {
+      new_backup->SetSequenceNumber(sequence_number);
+    }
+  }
+  ROCKS_LOG_INFO(options_.info_log, "add files for backup done, wait finish.");
+  IOStatus item_io_status;
+  for (auto& item : backup_items_to_finish) {
+    item.result.wait();
+    auto result = item.result.get();
+    item_io_status = result.io_status;
+    Temperature temp = result.expected_src_temperature;
+    if (result.current_src_temperature != Temperature::kUnknown &&
+        (temp == Temperature::kUnknown ||
+         options_.current_temperatures_override_manifest)) {
+      temp = result.current_src_temperature;
+    }
+    if (item_io_status.ok() && item.shared && item.needed_to_copy) {
+      item_io_status = item.backup_env->GetFileSystem()->RenameFile(
+          item.dst_path_tmp, item.dst_path, io_options_, nullptr);
+    }
+    if (item_io_status.ok()) {
+      item_io_status = new_backup.get()->AddFile(std::make_shared<FileInfo>(
+          item.dst_relative, result.size, result.checksum_hex, result.db_id,
+          result.db_session_id, temp));
+    }
+    if (!item_io_status.ok()) {
+      io_s = item_io_status;
+    }
+  }
+
+  // we copied all the files, enable file deletions
+  if (disabled.ok()) {  // If we successfully disabled file deletions
+    db->EnableFileDeletions(false).PermitUncheckedError();
+  }
+  auto backup_time = backup_env_->NowMicros() - start_backup;
+
+  if (io_s.ok()) {
+    // persist the backup metadata on the disk
+    io_s = new_backup->StoreToFile(options_.sync, options_.schema_version,
+                                   schema_test_options_.get());
+  }
+  if (io_s.ok() && options_.sync) {
+    std::unique_ptr<FSDirectory> backup_private_directory;
+    backup_fs_
+        ->NewDirectory(GetAbsolutePath(GetPrivateFileRel(new_backup_id, false)),
+                       io_options_, &backup_private_directory, nullptr)
+        .PermitUncheckedError();
+    if (backup_private_directory != nullptr) {
+      io_s = backup_private_directory->FsyncWithDirOptions(io_options_, nullptr,
+                                                           DirFsyncOptions());
+    }
+    if (io_s.ok() && private_directory_ != nullptr) {
+      io_s = private_directory_->FsyncWithDirOptions(io_options_, nullptr,
+                                                     DirFsyncOptions());
+    }
+    if (io_s.ok() && meta_directory_ != nullptr) {
+      io_s = meta_directory_->FsyncWithDirOptions(io_options_, nullptr,
+                                                  DirFsyncOptions());
+    }
+    if (io_s.ok() && shared_directory_ != nullptr) {
+      io_s = shared_directory_->FsyncWithDirOptions(io_options_, nullptr,
+                                                    DirFsyncOptions());
+    }
+    if (io_s.ok() && backup_directory_ != nullptr) {
+      io_s = backup_directory_->FsyncWithDirOptions(io_options_, nullptr,
+                                                    DirFsyncOptions());
+    }
+  }
+
+  if (io_s.ok()) {
+    backup_statistics_.IncrementNumberSuccessBackup();
+    // here we know that we succeeded and installed the new backup
+    latest_backup_id_ = new_backup_id;
+    latest_valid_backup_id_ = new_backup_id;
+    if (new_backup_id_ptr) {
+      *new_backup_id_ptr = new_backup_id;
+    }
+    ROCKS_LOG_INFO(options_.info_log, "Backup DONE. All is good");
+
+    // backup_speed is in byte/second
+    double backup_speed = new_backup->GetSize() / (1.048576 * backup_time);
+    ROCKS_LOG_INFO(options_.info_log, "Backup number of files: %u",
+                   new_backup->GetNumberFiles());
+    char human_size[16];
+    AppendHumanBytes(new_backup->GetSize(), human_size, sizeof(human_size));
+    ROCKS_LOG_INFO(options_.info_log, "Backup size: %s", human_size);
+    ROCKS_LOG_INFO(options_.info_log, "Backup time: %" PRIu64 " microseconds",
+                   backup_time);
+    ROCKS_LOG_INFO(options_.info_log, "Backup speed: %.3f MB/s", backup_speed);
+    ROCKS_LOG_INFO(options_.info_log, "Backup Statistics %s",
+                   backup_statistics_.ToString().c_str());
+  } else {
+    backup_statistics_.IncrementNumberFailBackup();
+    // clean all the files we might have created
+    ROCKS_LOG_INFO(options_.info_log, "Backup failed -- %s",
+                   io_s.ToString().c_str());
+    ROCKS_LOG_INFO(options_.info_log, "Backup Statistics %s\n",
+                   backup_statistics_.ToString().c_str());
+    // delete files that we might have already written
+    might_need_garbage_collect_ = true;
+    DeleteBackup(new_backup_id).PermitUncheckedError();
+  }
+
+  RecordTick(stats, BACKUP_READ_BYTES, IOSTATS(bytes_read) - prev_bytes_read);
+  RecordTick(stats, BACKUP_WRITE_BYTES,
+             IOSTATS(bytes_written) - prev_bytes_written);
+  return io_s;
+}
+
+IOStatus BackupEngineImpl::PurgeOldBackups(uint32_t num_backups_to_keep) {
+  assert(initialized_);
+  assert(!read_only_);
+
+  // Best effort deletion even with errors
+  IOStatus overall_status = IOStatus::OK();
+
+  ROCKS_LOG_INFO(options_.info_log, "Purging old backups, keeping %u",
+                 num_backups_to_keep);
+  std::vector<BackupID> to_delete;
+  auto itr = backups_.begin();
+  while ((backups_.size() - to_delete.size()) > num_backups_to_keep) {
+    to_delete.push_back(itr->first);
+    itr++;
+  }
+  for (auto backup_id : to_delete) {
+    // Do not GC until end
+    IOStatus io_s = DeleteBackupNoGC(backup_id);
+    if (!io_s.ok()) {
+      overall_status = io_s;
+    }
+  }
+  // Clean up after any incomplete backup deletion, potentially from
+  // earlier session.
+  if (might_need_garbage_collect_) {
+    IOStatus io_s = GarbageCollect();
+    if (!io_s.ok() && overall_status.ok()) {
+      overall_status = io_s;
+    }
+  }
+  return overall_status;
+}
+
+IOStatus BackupEngineImpl::DeleteBackup(BackupID backup_id) {
+  IOStatus s1 = DeleteBackupNoGC(backup_id);
+  IOStatus s2 = IOStatus::OK();
+
+  // Clean up after any incomplete backup deletion, potentially from
+  // earlier session.
+  if (might_need_garbage_collect_) {
+    s2 = GarbageCollect();
+  }
+
+  if (!s1.ok()) {
+    // Any failure in the primary objective trumps any failure in the
+    // secondary objective.
+    s2.PermitUncheckedError();
+    return s1;
+  } else {
+    return s2;
+  }
+}
+
+// Does not auto-GarbageCollect nor lock
+IOStatus BackupEngineImpl::DeleteBackupNoGC(BackupID backup_id) {
+  assert(initialized_);
+  assert(!read_only_);
+
+  ROCKS_LOG_INFO(options_.info_log, "Deleting backup %u", backup_id);
+  auto backup = backups_.find(backup_id);
+  if (backup != backups_.end()) {
+    IOStatus io_s = backup->second->Delete();
+    if (!io_s.ok()) {
+      return io_s;
+    }
+    backups_.erase(backup);
+  } else {
+    auto corrupt = corrupt_backups_.find(backup_id);
+    if (corrupt == corrupt_backups_.end()) {
+      return IOStatus::NotFound("Backup not found");
+    }
+    IOStatus io_s = corrupt->second.second->Delete();
+    if (!io_s.ok()) {
+      return io_s;
+    }
+    corrupt->second.first.PermitUncheckedError();
+    corrupt_backups_.erase(corrupt);
+  }
+
+  // After removing meta file, best effort deletion even with errors.
+  // (Don't delete other files if we can't delete the meta file right
+  // now.)
+  std::vector<std::string> to_delete;
+  for (auto& itr : backuped_file_infos_) {
+    if (itr.second->refs == 0) {
+      IOStatus io_s = backup_fs_->DeleteFile(GetAbsolutePath(itr.first),
+                                             io_options_, nullptr);
+      ROCKS_LOG_INFO(options_.info_log, "Deleting %s -- %s", itr.first.c_str(),
+                     io_s.ToString().c_str());
+      to_delete.push_back(itr.first);
+      if (!io_s.ok()) {
+        // Trying again later might work
+        might_need_garbage_collect_ = true;
+      }
+    }
+  }
+  for (auto& td : to_delete) {
+    backuped_file_infos_.erase(td);
+  }
+
+  // take care of private dirs -- GarbageCollect() will take care of them
+  // if they are not empty
+  std::string private_dir = GetPrivateFileRel(backup_id);
+  IOStatus io_s =
+      backup_fs_->DeleteDir(GetAbsolutePath(private_dir), io_options_, nullptr);
+  ROCKS_LOG_INFO(options_.info_log, "Deleting private dir %s -- %s",
+                 private_dir.c_str(), io_s.ToString().c_str());
+  if (!io_s.ok()) {
+    // Full gc or trying again later might work
+    might_need_garbage_collect_ = true;
+  }
+  return IOStatus::OK();
+}
+
+void BackupEngineImpl::SetBackupInfoFromBackupMeta(
+    BackupID id, const BackupMeta& meta, BackupInfo* backup_info,
+    bool include_file_details) const {
+  *backup_info = BackupInfo(id, meta.GetTimestamp(), meta.GetSize(),
+                            meta.GetNumberFiles(), meta.GetAppMetadata());
+  std::string dir =
+      options_.backup_dir + "/" + kPrivateDirSlash + std::to_string(id);
+  if (include_file_details) {
+    auto& file_details = backup_info->file_details;
+    file_details.reserve(meta.GetFiles().size());
+    for (auto& file_ptr : meta.GetFiles()) {
+      BackupFileInfo& finfo = *file_details.emplace(file_details.end());
+      finfo.relative_filename = file_ptr->filename;
+      finfo.size = file_ptr->size;
+      finfo.directory = dir;
+      uint64_t number;
+      FileType type;
+      bool ok = ParseFileName(file_ptr->filename, &number, &type);
+      if (ok) {
+        finfo.file_number = number;
+        finfo.file_type = type;
+      }
+      // TODO: temperature, file_checksum, file_checksum_func_name
+    }
+    backup_info->name_for_open = GetAbsolutePath(GetPrivateFileRel(id));
+    backup_info->name_for_open.pop_back();  // remove trailing '/'
+    backup_info->env_for_open = meta.GetEnvForOpen();
+  }
+}
+
+Status BackupEngineImpl::GetBackupInfo(BackupID backup_id,
+                                       BackupInfo* backup_info,
+                                       bool include_file_details) const {
+  assert(initialized_);
+  if (backup_id == kLatestBackupIDMarker) {
+    // Note: Read latest_valid_backup_id_ inside of lock
+    backup_id = latest_valid_backup_id_;
+  }
+  auto corrupt_itr = corrupt_backups_.find(backup_id);
+  if (corrupt_itr != corrupt_backups_.end()) {
+    return Status::Corruption(corrupt_itr->second.first.ToString());
+  }
+  auto backup_itr = backups_.find(backup_id);
+  if (backup_itr == backups_.end()) {
+    return Status::NotFound("Backup not found");
+  }
+  auto& backup = backup_itr->second;
+  if (backup->Empty()) {
+    return Status::NotFound("Backup not found");
+  }
+
+  SetBackupInfoFromBackupMeta(backup_id, *backup, backup_info,
+                              include_file_details);
+  return Status::OK();
+}
+
+void BackupEngineImpl::GetBackupInfo(std::vector<BackupInfo>* backup_info,
+                                     bool include_file_details) const {
+  assert(initialized_);
+  backup_info->resize(backups_.size());
+  size_t i = 0;
+  for (auto& backup : backups_) {
+    const BackupMeta& meta = *backup.second;
+    if (!meta.Empty()) {
+      SetBackupInfoFromBackupMeta(backup.first, meta, &backup_info->at(i++),
+                                  include_file_details);
+    }
+  }
+}
+
+void BackupEngineImpl::GetCorruptedBackups(
+    std::vector<BackupID>* corrupt_backup_ids) const {
+  assert(initialized_);
+  corrupt_backup_ids->reserve(corrupt_backups_.size());
+  for (auto& backup : corrupt_backups_) {
+    corrupt_backup_ids->push_back(backup.first);
+  }
+}
+
+IOStatus BackupEngineImpl::RestoreDBFromBackup(
+    const RestoreOptions& options, BackupID backup_id,
+    const std::string& db_dir, const std::string& wal_dir) const {
+  assert(initialized_);
+  if (backup_id == kLatestBackupIDMarker) {
+    // Note: Read latest_valid_backup_id_ inside of lock
+    backup_id = latest_valid_backup_id_;
+  }
+  auto corrupt_itr = corrupt_backups_.find(backup_id);
+  if (corrupt_itr != corrupt_backups_.end()) {
+    return corrupt_itr->second.first;
+  }
+  auto backup_itr = backups_.find(backup_id);
+  if (backup_itr == backups_.end()) {
+    return IOStatus::NotFound("Backup not found");
+  }
+  auto& backup = backup_itr->second;
+  if (backup->Empty()) {
+    return IOStatus::NotFound("Backup not found");
+  }
+
+  ROCKS_LOG_INFO(options_.info_log, "Restoring backup id %u\n", backup_id);
+  ROCKS_LOG_INFO(options_.info_log, "keep_log_files: %d\n",
+                 static_cast<int>(options.keep_log_files));
+
+  // just in case. Ignore errors
+  db_fs_->CreateDirIfMissing(db_dir, io_options_, nullptr)
+      .PermitUncheckedError();
+  db_fs_->CreateDirIfMissing(wal_dir, io_options_, nullptr)
+      .PermitUncheckedError();
+
+  if (options.keep_log_files) {
+    // delete files in db_dir, but keep all the log files
+    DeleteChildren(db_dir, 1 << kWalFile);
+    // move all the files from archive dir to wal_dir
+    std::string archive_dir = ArchivalDirectory(wal_dir);
+    std::vector<std::string> archive_files;
+    db_fs_->GetChildren(archive_dir, io_options_, &archive_files, nullptr)
+        .PermitUncheckedError();  // ignore errors
+    for (const auto& f : archive_files) {
+      uint64_t number;
+      FileType type;
+      bool ok = ParseFileName(f, &number, &type);
+      if (ok && type == kWalFile) {
+        ROCKS_LOG_INFO(options_.info_log,
+                       "Moving log file from archive/ to wal_dir: %s",
+                       f.c_str());
+        IOStatus io_s = db_fs_->RenameFile(
+            archive_dir + "/" + f, wal_dir + "/" + f, io_options_, nullptr);
+        if (!io_s.ok()) {
+          // if we can't move log file from archive_dir to wal_dir,
+          // we should fail, since it might mean data loss
+          return io_s;
+        }
+      }
+    }
+  } else {
+    DeleteChildren(wal_dir);
+    DeleteChildren(ArchivalDirectory(wal_dir));
+    DeleteChildren(db_dir);
+  }
+
+  IOStatus io_s;
+  std::vector<RestoreAfterCopyOrCreateWorkItem> restore_items_to_finish;
+  std::string temporary_current_file;
+  std::string final_current_file;
+  std::unique_ptr<FSDirectory> db_dir_for_fsync;
+  std::unique_ptr<FSDirectory> wal_dir_for_fsync;
+
+  for (const auto& file_info : backup->GetFiles()) {
+    const std::string& file = file_info->filename;
+    // 1. get DB filename
+    std::string dst = file_info->GetDbFileName();
+
+    // 2. find the filetype
+    uint64_t number;
+    FileType type;
+    bool ok = ParseFileName(dst, &number, &type);
+    if (!ok) {
+      return IOStatus::Corruption("Backup corrupted: Fail to parse filename " +
+                                  dst);
+    }
+    // 3. Construct the final path
+    // kWalFile lives in wal_dir and all the rest live in db_dir
+    if (type == kWalFile) {
+      dst = wal_dir + "/" + dst;
+      if (options_.sync && !wal_dir_for_fsync) {
+        io_s = db_fs_->NewDirectory(wal_dir, io_options_, &wal_dir_for_fsync,
+                                    nullptr);
+        if (!io_s.ok()) {
+          return io_s;
+        }
+      }
+    } else {
+      dst = db_dir + "/" + dst;
+      if (options_.sync && !db_dir_for_fsync) {
+        io_s = db_fs_->NewDirectory(db_dir, io_options_, &db_dir_for_fsync,
+                                    nullptr);
+        if (!io_s.ok()) {
+          return io_s;
+        }
+      }
+    }
+    // For atomicity, initially restore CURRENT file to a temporary name.
+    // This is useful even without options_.sync e.g. in case the restore
+    // process is interrupted.
+    if (type == kCurrentFile) {
+      final_current_file = dst;
+      dst = temporary_current_file = dst + ".tmp";
+    }
+
+    ROCKS_LOG_INFO(options_.info_log, "Restoring %s to %s\n", file.c_str(),
+                   dst.c_str());
+    CopyOrCreateWorkItem copy_or_create_work_item(
+        GetAbsolutePath(file), dst, Temperature::kUnknown /* src_temp */,
+        file_info->temp, "" /* contents */, backup_env_, db_env_,
+        EnvOptions() /* src_env_options */, options_.sync,
+        options_.restore_rate_limiter.get(), file_info->size,
+        nullptr /* stats */);
+    RestoreAfterCopyOrCreateWorkItem after_copy_or_create_work_item(
+        copy_or_create_work_item.result.get_future(), file, dst,
+        file_info->checksum_hex);
+    files_to_copy_or_create_.write(std::move(copy_or_create_work_item));
+    restore_items_to_finish.push_back(
+        std::move(after_copy_or_create_work_item));
+  }
+  IOStatus item_io_status;
+  for (auto& item : restore_items_to_finish) {
+    item.result.wait();
+    auto result = item.result.get();
+    item_io_status = result.io_status;
+    // Note: It is possible that both of the following bad-status cases occur
+    // during copying. But, we only return one status.
+    if (!item_io_status.ok()) {
+      io_s = item_io_status;
+      break;
+    } else if (!item.checksum_hex.empty() &&
+               item.checksum_hex != result.checksum_hex) {
+      io_s = IOStatus::Corruption(
+          "While restoring " + item.from_file + " -> " + item.to_file +
+          ": expected checksum is " + item.checksum_hex +
+          " while computed checksum is " + result.checksum_hex);
+      break;
+    }
+  }
+
+  // When enabled, the first FsyncWithDirOptions is to ensure all files are
+  // fully persisted before renaming CURRENT.tmp
+  if (io_s.ok() && db_dir_for_fsync) {
+    ROCKS_LOG_INFO(options_.info_log, "Restore: fsync\n");
+    io_s = db_dir_for_fsync->FsyncWithDirOptions(io_options_, nullptr,
+                                                 DirFsyncOptions());
+  }
+
+  if (io_s.ok() && wal_dir_for_fsync) {
+    io_s = wal_dir_for_fsync->FsyncWithDirOptions(io_options_, nullptr,
+                                                  DirFsyncOptions());
+  }
+
+  if (io_s.ok() && !temporary_current_file.empty()) {
+    ROCKS_LOG_INFO(options_.info_log, "Restore: atomic rename CURRENT.tmp\n");
+    assert(!final_current_file.empty());
+    io_s = db_fs_->RenameFile(temporary_current_file, final_current_file,
+                              io_options_, nullptr);
+  }
+
+  if (io_s.ok() && db_dir_for_fsync && !temporary_current_file.empty()) {
+    // Second FsyncWithDirOptions is to ensure the final atomic rename of DB
+    // restore is fully persisted even if power goes out right after restore
+    // operation returns success
+    assert(db_dir_for_fsync);
+    io_s = db_dir_for_fsync->FsyncWithDirOptions(
+        io_options_, nullptr, DirFsyncOptions(final_current_file));
+  }
+
+  ROCKS_LOG_INFO(options_.info_log, "Restoring done -- %s\n",
+                 io_s.ToString().c_str());
+  return io_s;
+}
+
+IOStatus BackupEngineImpl::VerifyBackup(BackupID backup_id,
+                                        bool verify_with_checksum) const {
+  assert(initialized_);
+  // Check if backup_id is corrupted, or valid and registered
+  auto corrupt_itr = corrupt_backups_.find(backup_id);
+  if (corrupt_itr != corrupt_backups_.end()) {
+    return corrupt_itr->second.first;
+  }
+
+  auto backup_itr = backups_.find(backup_id);
+  if (backup_itr == backups_.end()) {
+    return IOStatus::NotFound();
+  }
+
+  auto& backup = backup_itr->second;
+  if (backup->Empty()) {
+    return IOStatus::NotFound();
+  }
+
+  ROCKS_LOG_INFO(options_.info_log, "Verifying backup id %u\n", backup_id);
+
+  // Find all existing backup files belong to backup_id
+  std::unordered_map<std::string, uint64_t> curr_abs_path_to_size;
+  for (const auto& rel_dir : {GetPrivateFileRel(backup_id), GetSharedFileRel(),
+                              GetSharedFileWithChecksumRel()}) {
+    const auto abs_dir = GetAbsolutePath(rel_dir);
+    // Shared directories allowed to be missing in some cases. Expected but
+    // missing files will be reported a few lines down.
+    ReadChildFileCurrentSizes(abs_dir, backup_fs_, &curr_abs_path_to_size)
+        .PermitUncheckedError();
+  }
+
+  // For all files registered in backup
+  for (const auto& file_info : backup->GetFiles()) {
+    const auto abs_path = GetAbsolutePath(file_info->filename);
+    // check existence of the file
+    if (curr_abs_path_to_size.find(abs_path) == curr_abs_path_to_size.end()) {
+      return IOStatus::NotFound("File missing: " + abs_path);
+    }
+    // verify file size
+    if (file_info->size != curr_abs_path_to_size[abs_path]) {
+      std::string size_info("Expected file size is " +
+                            std::to_string(file_info->size) +
+                            " while found file size is " +
+                            std::to_string(curr_abs_path_to_size[abs_path]));
+      return IOStatus::Corruption("File corrupted: File size mismatch for " +
+                                  abs_path + ": " + size_info);
+    }
+    if (verify_with_checksum && !file_info->checksum_hex.empty()) {
+      // verify file checksum
+      std::string checksum_hex;
+      ROCKS_LOG_INFO(options_.info_log, "Verifying %s checksum...\n",
+                     abs_path.c_str());
+      IOStatus io_s = ReadFileAndComputeChecksum(
+          abs_path, backup_fs_, EnvOptions(), 0 /* size_limit */, &checksum_hex,
+          Temperature::kUnknown);
+      if (!io_s.ok()) {
+        return io_s;
+      } else if (file_info->checksum_hex != checksum_hex) {
+        std::string checksum_info(
+            "Expected checksum is " + file_info->checksum_hex +
+            " while computed checksum is " + checksum_hex);
+        return IOStatus::Corruption("File corrupted: Checksum mismatch for " +
+                                    abs_path + ": " + checksum_info);
+      }
+    }
+  }
+  return IOStatus::OK();
+}
+
+IOStatus BackupEngineImpl::CopyOrCreateFile(
+    const std::string& src, const std::string& dst, const std::string& contents,
+    uint64_t size_limit, Env* src_env, Env* dst_env,
+    const EnvOptions& src_env_options, bool sync, RateLimiter* rate_limiter,
+    std::function<void()> progress_callback, Temperature* src_temperature,
+    Temperature dst_temperature, uint64_t* bytes_toward_next_callback,
+    uint64_t* size, std::string* checksum_hex) {
+  assert(src.empty() != contents.empty());
+  IOStatus io_s;
+  std::unique_ptr<FSWritableFile> dst_file;
+  std::unique_ptr<FSSequentialFile> src_file;
+  FileOptions dst_file_options;
+  dst_file_options.use_mmap_writes = false;
+  dst_file_options.temperature = dst_temperature;
+  // TODO:(gzh) maybe use direct reads/writes here if possible
+  if (size != nullptr) {
+    *size = 0;
+  }
+  uint32_t checksum_value = 0;
+
+  // Check if size limit is set. if not, set it to very big number
+  if (size_limit == 0) {
+    size_limit = std::numeric_limits<uint64_t>::max();
+  }
+
+  io_s = dst_env->GetFileSystem()->NewWritableFile(dst, dst_file_options,
+                                                   &dst_file, nullptr);
+  if (io_s.ok() && !src.empty()) {
+    auto src_file_options = FileOptions(src_env_options);
+    src_file_options.temperature = *src_temperature;
+    io_s = src_env->GetFileSystem()->NewSequentialFile(src, src_file_options,
+                                                       &src_file, nullptr);
+  }
+  if (io_s.IsPathNotFound() && *src_temperature != Temperature::kUnknown) {
+    // Retry without temperature hint in case the FileSystem is strict with
+    // non-kUnknown temperature option
+    io_s = src_env->GetFileSystem()->NewSequentialFile(
+        src, FileOptions(src_env_options), &src_file, nullptr);
+  }
+  if (!io_s.ok()) {
+    return io_s;
+  }
+
+  size_t buf_size =
+      rate_limiter ? static_cast<size_t>(rate_limiter->GetSingleBurstBytes())
+                   : kDefaultCopyFileBufferSize;
+
+  std::unique_ptr<WritableFileWriter> dest_writer(
+      new WritableFileWriter(std::move(dst_file), dst, dst_file_options));
+  std::unique_ptr<SequentialFileReader> src_reader;
+  std::unique_ptr<char[]> buf;
+  if (!src.empty()) {
+    // Return back current temperature in FileSystem
+    *src_temperature = src_file->GetTemperature();
+
+    src_reader.reset(new SequentialFileReader(
+        std::move(src_file), src, nullptr /* io_tracer */, {}, rate_limiter));
+    buf.reset(new char[buf_size]);
+  }
+
+  Slice data;
+  do {
+    if (stop_backup_.load(std::memory_order_acquire)) {
+      return status_to_io_status(Status::Incomplete("Backup stopped"));
+    }
+    if (!src.empty()) {
+      size_t buffer_to_read =
+          (buf_size < size_limit) ? buf_size : static_cast<size_t>(size_limit);
+      io_s = src_reader->Read(buffer_to_read, &data, buf.get(),
+                              Env::IO_LOW /* rate_limiter_priority */);
+      *bytes_toward_next_callback += data.size();
+    } else {
+      data = contents;
+    }
+    size_limit -= data.size();
+    TEST_SYNC_POINT_CALLBACK(
+        "BackupEngineImpl::CopyOrCreateFile:CorruptionDuringBackup",
+        (src.length() > 4 && src.rfind(".sst") == src.length() - 4) ? &data
+                                                                    : nullptr);
+
+    if (!io_s.ok()) {
+      return io_s;
+    }
+
+    if (size != nullptr) {
+      *size += data.size();
+    }
+    if (checksum_hex != nullptr) {
+      checksum_value = crc32c::Extend(checksum_value, data.data(), data.size());
+    }
+    io_s = dest_writer->Append(data);
+
+    if (rate_limiter != nullptr) {
+      if (!src.empty()) {
+        rate_limiter->Request(data.size(), Env::IO_LOW, nullptr /* stats */,
+                              RateLimiter::OpType::kWrite);
+      } else {
+        LoopRateLimitRequestHelper(data.size(), rate_limiter, Env::IO_LOW,
+                                   nullptr /* stats */,
+                                   RateLimiter::OpType::kWrite);
+      }
+    }
+    while (*bytes_toward_next_callback >=
+           options_.callback_trigger_interval_size) {
+      *bytes_toward_next_callback -= options_.callback_trigger_interval_size;
+      std::lock_guard<std::mutex> lock(byte_report_mutex_);
+      progress_callback();
+    }
+  } while (io_s.ok() && contents.empty() && data.size() > 0 && size_limit > 0);
+
+  // Convert uint32_t checksum to hex checksum
+  if (checksum_hex != nullptr) {
+    checksum_hex->assign(ChecksumInt32ToHex(checksum_value));
+  }
+
+  if (io_s.ok() && sync) {
+    io_s = dest_writer->Sync(false);
+  }
+  if (io_s.ok()) {
+    io_s = dest_writer->Close();
+  }
+  return io_s;
+}
+
+// fname will always start with "/"
+IOStatus BackupEngineImpl::AddBackupFileWorkItem(
+    std::unordered_set<std::string>& live_dst_paths,
+    std::vector<BackupAfterCopyOrCreateWorkItem>& backup_items_to_finish,
+    BackupID backup_id, bool shared, const std::string& src_dir,
+    const std::string& fname, const EnvOptions& src_env_options,
+    RateLimiter* rate_limiter, FileType file_type, uint64_t size_bytes,
+    Statistics* stats, uint64_t size_limit, bool shared_checksum,
+    std::function<void()> progress_callback, const std::string& contents,
+    const std::string& src_checksum_func_name,
+    const std::string& src_checksum_str, const Temperature src_temperature) {
+  assert(contents.empty() != src_dir.empty());
+
+  std::string src_path = src_dir + "/" + fname;
+  std::string dst_relative;
+  std::string dst_relative_tmp;
+  std::string db_id;
+  std::string db_session_id;
+  // crc32c checksum in hex. empty == unavailable / unknown
+  std::string checksum_hex;
+
+  // Whenever a default checksum function name is passed in, we will compares
+  // the corresponding checksum values after copying. Note that only table and
+  // blob files may have a known checksum function name passed in.
+  //
+  // If no default checksum function name is passed in and db session id is not
+  // available, we will calculate the checksum *before* copying in two cases
+  // (we always calcuate checksums when copying or creating for any file types):
+  // a) share_files_with_checksum is true and file type is table;
+  // b) share_table_files is true and the file exists already.
+  //
+  // Step 0: Check if default checksum function name is passed in
+  if (kDbFileChecksumFuncName == src_checksum_func_name) {
+    if (src_checksum_str == kUnknownFileChecksum) {
+      return status_to_io_status(
+          Status::Aborted("Unknown checksum value for " + fname));
+    }
+    checksum_hex = ChecksumStrToHex(src_checksum_str);
+  }
+
+  // Step 1: Prepare the relative path to destination
+  if (shared && shared_checksum) {
+    if (GetNamingNoFlags() != BackupEngineOptions::kLegacyCrc32cAndFileSize &&
+        file_type != kBlobFile) {
+      // Prepare db_session_id to add to the file name
+      // Ignore the returned status
+      // In the failed cases, db_id and db_session_id will be empty
+      GetFileDbIdentities(db_env_, src_env_options, src_path, src_temperature,
+                          rate_limiter, &db_id, &db_session_id)
+          .PermitUncheckedError();
+    }
+    // Calculate checksum if checksum and db session id are not available.
+    // If db session id is available, we will not calculate the checksum
+    // since the session id should suffice to avoid file name collision in
+    // the shared_checksum directory.
+    if (checksum_hex.empty() && db_session_id.empty()) {
+      IOStatus io_s = ReadFileAndComputeChecksum(
+          src_path, db_fs_, src_env_options, size_limit, &checksum_hex,
+          src_temperature);
+      if (!io_s.ok()) {
+        return io_s;
+      }
+    }
+    if (size_bytes == std::numeric_limits<uint64_t>::max()) {
+      return IOStatus::NotFound("File missing: " + src_path);
+    }
+    // dst_relative depends on the following conditions:
+    // 1) the naming scheme is kUseDbSessionId,
+    // 2) db_session_id is not empty,
+    // 3) checksum is available in the DB manifest.
+    // If 1,2,3) are satisfied, then dst_relative will be of the form:
+    // shared_checksum/<file_number>_<checksum>_<db_session_id>.sst
+    // If 1,2) are satisfied, then dst_relative will be of the form:
+    // shared_checksum/<file_number>_<db_session_id>.sst
+    // Otherwise, dst_relative is of the form
+    // shared_checksum/<file_number>_<checksum>_<size>.sst
+    //
+    // For blob files, db_session_id is not supported with the blob file format.
+    // It uses original/legacy naming scheme.
+    // dst_relative will be of the form:
+    // shared_checksum/<file_number>_<checksum>_<size>.blob
+    dst_relative = GetSharedFileWithChecksum(fname, checksum_hex, size_bytes,
+                                             db_session_id);
+    dst_relative_tmp = GetSharedFileWithChecksumRel(dst_relative, true);
+    dst_relative = GetSharedFileWithChecksumRel(dst_relative, false);
+  } else if (shared) {
+    dst_relative_tmp = GetSharedFileRel(fname, true);
+    dst_relative = GetSharedFileRel(fname, false);
+  } else {
+    dst_relative = GetPrivateFileRel(backup_id, false, fname);
+  }
+
+  // We copy into `temp_dest_path` and, once finished, rename it to
+  // `final_dest_path`. This allows files to atomically appear at
+  // `final_dest_path`. We can copy directly to the final path when atomicity
+  // is unnecessary, like for files in private backup directories.
+  const std::string* copy_dest_path;
+  std::string temp_dest_path;
+  std::string final_dest_path = GetAbsolutePath(dst_relative);
+  if (!dst_relative_tmp.empty()) {
+    temp_dest_path = GetAbsolutePath(dst_relative_tmp);
+    copy_dest_path = &temp_dest_path;
+  } else {
+    copy_dest_path = &final_dest_path;
+  }
+
+  // Step 2: Determine whether to copy or not
+  // if it's shared, we also need to check if it exists -- if it does, no need
+  // to copy it again.
+  bool need_to_copy = true;
+  // true if final_dest_path is the same path as another live file
+  const bool same_path =
+      live_dst_paths.find(final_dest_path) != live_dst_paths.end();
+
+  bool file_exists = false;
+  if (shared && !same_path) {
+    // Should be in shared directory but not a live path, check existence in
+    // shared directory
+    IOStatus exist =
+        backup_fs_->FileExists(final_dest_path, io_options_, nullptr);
+    if (exist.ok()) {
+      file_exists = true;
+    } else if (exist.IsNotFound()) {
+      file_exists = false;
+    } else {
+      return exist;
+    }
+  }
+
+  if (!contents.empty()) {
+    need_to_copy = false;
+  } else if (shared && (same_path || file_exists)) {
+    need_to_copy = false;
+    auto find_result = backuped_file_infos_.find(dst_relative);
+    if (find_result == backuped_file_infos_.end() && !same_path) {
+      // file exists but not referenced
+      ROCKS_LOG_INFO(
+          options_.info_log,
+          "%s already present, but not referenced by any backup. We will "
+          "overwrite the file.",
+          fname.c_str());
+      need_to_copy = true;
+      // Defer any failure reporting to when we try to write the file
+      backup_fs_->DeleteFile(final_dest_path, io_options_, nullptr)
+          .PermitUncheckedError();
+    } else {
+      // file exists and referenced
+      if (checksum_hex.empty()) {
+        // same_path should not happen for a standard DB, so OK to
+        // read file contents to check for checksum mismatch between
+        // two files from same DB getting same name.
+        // For compatibility with future meta file that might not have
+        // crc32c checksum available, consider it might be empty, but
+        // we don't currently generate meta file without crc32c checksum.
+        // Therefore we have to read & compute it if we don't have it.
+        if (!same_path && !find_result->second->checksum_hex.empty()) {
+          assert(find_result != backuped_file_infos_.end());
+          // Note: to save I/O on incremental backups, we copy prior known
+          // checksum of the file instead of reading entire file contents
+          // to recompute it.
+          checksum_hex = find_result->second->checksum_hex;
+          // Regarding corruption detection, consider:
+          // (a) the DB file is corrupt (since previous backup) and the backup
+          // file is OK: we failed to detect, but the backup is safe. DB can
+          // be repaired/restored once its corruption is detected.
+          // (b) the backup file is corrupt (since previous backup) and the
+          // db file is OK: we failed to detect, but the backup is corrupt.
+          // CreateNewBackup should support fast incremental backups and
+          // there's no way to support that without reading all the files.
+          // We might add an option for extra checks on incremental backup,
+          // but until then, use VerifyBackups to check existing backup data.
+          // (c) file name collision with legitimately different content.
+          // This is almost inconceivable with a well-generated DB session
+          // ID, but even in that case, we double check the file sizes in
+          // BackupMeta::AddFile.
+        } else {
+          IOStatus io_s = ReadFileAndComputeChecksum(
+              src_path, db_fs_, src_env_options, size_limit, &checksum_hex,
+              src_temperature);
+          if (!io_s.ok()) {
+            return io_s;
+          }
+        }
+      }
+      if (!db_session_id.empty()) {
+        ROCKS_LOG_INFO(options_.info_log,
+                       "%s already present, with checksum %s, size %" PRIu64
+                       " and DB session identity %s",
+                       fname.c_str(), checksum_hex.c_str(), size_bytes,
+                       db_session_id.c_str());
+      } else {
+        ROCKS_LOG_INFO(options_.info_log,
+                       "%s already present, with checksum %s and size %" PRIu64,
+                       fname.c_str(), checksum_hex.c_str(), size_bytes);
+      }
+    }
+  }
+  live_dst_paths.insert(final_dest_path);
+
+  // Step 3: Add work item
+  if (!contents.empty() || need_to_copy) {
+    ROCKS_LOG_INFO(options_.info_log, "Copying %s to %s", fname.c_str(),
+                   copy_dest_path->c_str());
+    CopyOrCreateWorkItem copy_or_create_work_item(
+        src_dir.empty() ? "" : src_path, *copy_dest_path, src_temperature,
+        Temperature::kUnknown /*dst_temp*/, contents, db_env_, backup_env_,
+        src_env_options, options_.sync, rate_limiter, size_limit, stats,
+        progress_callback, src_checksum_func_name, checksum_hex, db_id,
+        db_session_id);
+    BackupAfterCopyOrCreateWorkItem after_copy_or_create_work_item(
+        copy_or_create_work_item.result.get_future(), shared, need_to_copy,
+        backup_env_, temp_dest_path, final_dest_path, dst_relative);
+    files_to_copy_or_create_.write(std::move(copy_or_create_work_item));
+    backup_items_to_finish.push_back(std::move(after_copy_or_create_work_item));
+  } else {
+    std::promise<CopyOrCreateResult> promise_result;
+    BackupAfterCopyOrCreateWorkItem after_copy_or_create_work_item(
+        promise_result.get_future(), shared, need_to_copy, backup_env_,
+        temp_dest_path, final_dest_path, dst_relative);
+    backup_items_to_finish.push_back(std::move(after_copy_or_create_work_item));
+    CopyOrCreateResult result;
+    result.io_status = IOStatus::OK();
+    result.size = size_bytes;
+    result.checksum_hex = std::move(checksum_hex);
+    result.db_id = std::move(db_id);
+    result.db_session_id = std::move(db_session_id);
+    promise_result.set_value(std::move(result));
+  }
+  return IOStatus::OK();
+}
+
+IOStatus BackupEngineImpl::ReadFileAndComputeChecksum(
+    const std::string& src, const std::shared_ptr<FileSystem>& src_fs,
+    const EnvOptions& src_env_options, uint64_t size_limit,
+    std::string* checksum_hex, const Temperature src_temperature) const {
+  if (checksum_hex == nullptr) {
+    return status_to_io_status(Status::Aborted("Checksum pointer is null"));
+  }
+  uint32_t checksum_value = 0;
+  if (size_limit == 0) {
+    size_limit = std::numeric_limits<uint64_t>::max();
+  }
+
+  std::unique_ptr<SequentialFileReader> src_reader;
+  auto file_options = FileOptions(src_env_options);
+  file_options.temperature = src_temperature;
+  RateLimiter* rate_limiter = options_.backup_rate_limiter.get();
+  IOStatus io_s = SequentialFileReader::Create(
+      src_fs, src, file_options, &src_reader, nullptr /* dbg */, rate_limiter);
+  if (io_s.IsPathNotFound() && src_temperature != Temperature::kUnknown) {
+    // Retry without temperature hint in case the FileSystem is strict with
+    // non-kUnknown temperature option
+    file_options.temperature = Temperature::kUnknown;
+    io_s = SequentialFileReader::Create(src_fs, src, file_options, &src_reader,
+                                        nullptr /* dbg */, rate_limiter);
+  }
+  if (!io_s.ok()) {
+    return io_s;
+  }
+
+  size_t buf_size = kDefaultCopyFileBufferSize;
+  std::unique_ptr<char[]> buf(new char[buf_size]);
+  Slice data;
+
+  do {
+    if (stop_backup_.load(std::memory_order_acquire)) {
+      return status_to_io_status(Status::Incomplete("Backup stopped"));
+    }
+    size_t buffer_to_read =
+        (buf_size < size_limit) ? buf_size : static_cast<size_t>(size_limit);
+    io_s = src_reader->Read(buffer_to_read, &data, buf.get(),
+                            Env::IO_LOW /* rate_limiter_priority */);
+    if (!io_s.ok()) {
+      return io_s;
+    }
+
+    size_limit -= data.size();
+    checksum_value = crc32c::Extend(checksum_value, data.data(), data.size());
+  } while (data.size() > 0 && size_limit > 0);
+
+  checksum_hex->assign(ChecksumInt32ToHex(checksum_value));
+
+  return io_s;
+}
+
+Status BackupEngineImpl::GetFileDbIdentities(
+    Env* src_env, const EnvOptions& src_env_options,
+    const std::string& file_path, Temperature file_temp,
+    RateLimiter* rate_limiter, std::string* db_id, std::string* db_session_id) {
+  assert(db_id != nullptr || db_session_id != nullptr);
+
+  Options options;
+  options.env = src_env;
+  SstFileDumper sst_reader(options, file_path, file_temp,
+                           2 * 1024 * 1024
+                           /* readahead_size */,
+                           false /* verify_checksum */, false /* output_hex */,
+                           false /* decode_blob_index */, src_env_options,
+                           true /* silent */);
+
+  const TableProperties* table_properties = nullptr;
+  std::shared_ptr<const TableProperties> tp;
+  Status s = sst_reader.getStatus();
+
+  if (s.ok()) {
+    // Try to get table properties from the table reader of sst_reader
+    if (!sst_reader.ReadTableProperties(&tp).ok()) {
+      // Try to use table properites from the initialization of sst_reader
+      table_properties = sst_reader.GetInitTableProperties();
+    } else {
+      table_properties = tp.get();
+      if (table_properties != nullptr && rate_limiter != nullptr) {
+        // sizeof(*table_properties) is a sufficent but far-from-exact
+        // approximation of read bytes due to metaindex block, std::string
+        // properties and varint compression
+        LoopRateLimitRequestHelper(sizeof(*table_properties), rate_limiter,
+                                   Env::IO_LOW, nullptr /* stats */,
+                                   RateLimiter::OpType::kRead);
+      }
+    }
+  } else {
+    ROCKS_LOG_INFO(options_.info_log, "Failed to read %s: %s",
+                   file_path.c_str(), s.ToString().c_str());
+    return s;
+  }
+
+  if (table_properties != nullptr) {
+    if (db_id != nullptr) {
+      db_id->assign(table_properties->db_id);
+    }
+    if (db_session_id != nullptr) {
+      db_session_id->assign(table_properties->db_session_id);
+      if (db_session_id->empty()) {
+        s = Status::NotFound("DB session identity not found in " + file_path);
+        ROCKS_LOG_INFO(options_.info_log, "%s", s.ToString().c_str());
+        return s;
+      }
+    }
+    return Status::OK();
+  } else {
+    s = Status::Corruption("Table properties missing in " + file_path);
+    ROCKS_LOG_INFO(options_.info_log, "%s", s.ToString().c_str());
+    return s;
+  }
+}
+
+void BackupEngineImpl::LoopRateLimitRequestHelper(
+    const size_t total_bytes_to_request, RateLimiter* rate_limiter,
+    const Env::IOPriority pri, Statistics* stats,
+    const RateLimiter::OpType op_type) {
+  assert(rate_limiter != nullptr);
+  size_t remaining_bytes = total_bytes_to_request;
+  size_t request_bytes = 0;
+  while (remaining_bytes > 0) {
+    request_bytes =
+        std::min(static_cast<size_t>(rate_limiter->GetSingleBurstBytes()),
+                 remaining_bytes);
+    rate_limiter->Request(request_bytes, pri, stats, op_type);
+    remaining_bytes -= request_bytes;
+  }
+}
+
+void BackupEngineImpl::DeleteChildren(const std::string& dir,
+                                      uint32_t file_type_filter) const {
+  std::vector<std::string> children;
+  db_fs_->GetChildren(dir, io_options_, &children, nullptr)
+      .PermitUncheckedError();  // ignore errors
+
+  for (const auto& f : children) {
+    uint64_t number;
+    FileType type;
+    bool ok = ParseFileName(f, &number, &type);
+    if (ok && (file_type_filter & (1 << type))) {
+      // don't delete this file
+      continue;
+    }
+    db_fs_->DeleteFile(dir + "/" + f, io_options_, nullptr)
+        .PermitUncheckedError();  // ignore errors
+  }
+}
+
+IOStatus BackupEngineImpl::ReadChildFileCurrentSizes(
+    const std::string& dir, const std::shared_ptr<FileSystem>& fs,
+    std::unordered_map<std::string, uint64_t>* result) const {
+  assert(result != nullptr);
+  std::vector<Env::FileAttributes> files_attrs;
+  IOStatus io_status = fs->FileExists(dir, io_options_, nullptr);
+  if (io_status.ok()) {
+    io_status =
+        fs->GetChildrenFileAttributes(dir, io_options_, &files_attrs, nullptr);
+  } else if (io_status.IsNotFound()) {
+    // Insert no entries can be considered success
+    io_status = IOStatus::OK();
+  }
+  const bool slash_needed = dir.empty() || dir.back() != '/';
+  for (const auto& file_attrs : files_attrs) {
+    result->emplace(dir + (slash_needed ? "/" : "") + file_attrs.name,
+                    file_attrs.size_bytes);
+  }
+  return io_status;
+}
+
+IOStatus BackupEngineImpl::GarbageCollect() {
+  assert(!read_only_);
+
+  // We will make a best effort to remove all garbage even in the presence
+  // of inconsistencies or I/O failures that inhibit finding garbage.
+  IOStatus overall_status = IOStatus::OK();
+  // If all goes well, we don't need another auto-GC this session
+  might_need_garbage_collect_ = false;
+
+  ROCKS_LOG_INFO(options_.info_log, "Starting garbage collection");
+
+  // delete obsolete shared files
+  for (bool with_checksum : {false, true}) {
+    std::vector<std::string> shared_children;
+    {
+      std::string shared_path;
+      if (with_checksum) {
+        shared_path = GetAbsolutePath(GetSharedFileWithChecksumRel());
+      } else {
+        shared_path = GetAbsolutePath(GetSharedFileRel());
+      }
+      IOStatus io_s = backup_fs_->FileExists(shared_path, io_options_, nullptr);
+      if (io_s.ok()) {
+        io_s = backup_fs_->GetChildren(shared_path, io_options_,
+                                       &shared_children, nullptr);
+      } else if (io_s.IsNotFound()) {
+        io_s = IOStatus::OK();
+      }
+      if (!io_s.ok()) {
+        overall_status = io_s;
+        // Trying again later might work
+        might_need_garbage_collect_ = true;
+      }
+    }
+    for (auto& child : shared_children) {
+      std::string rel_fname;
+      if (with_checksum) {
+        rel_fname = GetSharedFileWithChecksumRel(child);
+      } else {
+        rel_fname = GetSharedFileRel(child);
+      }
+      auto child_itr = backuped_file_infos_.find(rel_fname);
+      // if it's not refcounted, delete it
+      if (child_itr == backuped_file_infos_.end() ||
+          child_itr->second->refs == 0) {
+        // this might be a directory, but DeleteFile will just fail in that
+        // case, so we're good
+        IOStatus io_s = backup_fs_->DeleteFile(GetAbsolutePath(rel_fname),
+                                               io_options_, nullptr);
+        ROCKS_LOG_INFO(options_.info_log, "Deleting %s -- %s",
+                       rel_fname.c_str(), io_s.ToString().c_str());
+        backuped_file_infos_.erase(rel_fname);
+        if (!io_s.ok()) {
+          // Trying again later might work
+          might_need_garbage_collect_ = true;
+        }
+      }
+    }
+  }
+
+  // delete obsolete private files
+  std::vector<std::string> private_children;
+  {
+    IOStatus io_s =
+        backup_fs_->GetChildren(GetAbsolutePath(kPrivateDirName), io_options_,
+                                &private_children, nullptr);
+    if (!io_s.ok()) {
+      overall_status = io_s;
+      // Trying again later might work
+      might_need_garbage_collect_ = true;
+    }
+  }
+  for (auto& child : private_children) {
+    BackupID backup_id = 0;
+    bool tmp_dir = child.find(".tmp") != std::string::npos;
+    sscanf(child.c_str(), "%u", &backup_id);
+    if (!tmp_dir &&  // if it's tmp_dir, delete it
+        (backup_id == 0 || backups_.find(backup_id) != backups_.end())) {
+      // it's either not a number or it's still alive. continue
+      continue;
+    }
+    // here we have to delete the dir and all its children
+    std::string full_private_path =
+        GetAbsolutePath(GetPrivateFileRel(backup_id));
+    std::vector<std::string> subchildren;
+    if (backup_fs_
+            ->GetChildren(full_private_path, io_options_, &subchildren, nullptr)
+            .ok()) {
+      for (auto& subchild : subchildren) {
+        IOStatus io_s = backup_fs_->DeleteFile(full_private_path + subchild,
+                                               io_options_, nullptr);
+        ROCKS_LOG_INFO(options_.info_log, "Deleting %s -- %s",
+                       (full_private_path + subchild).c_str(),
+                       io_s.ToString().c_str());
+        if (!io_s.ok()) {
+          // Trying again later might work
+          might_need_garbage_collect_ = true;
+        }
+      }
+    }
+    // finally delete the private dir
+    IOStatus io_s =
+        backup_fs_->DeleteDir(full_private_path, io_options_, nullptr);
+    ROCKS_LOG_INFO(options_.info_log, "Deleting dir %s -- %s",
+                   full_private_path.c_str(), io_s.ToString().c_str());
+    if (!io_s.ok()) {
+      // Trying again later might work
+      might_need_garbage_collect_ = true;
+    }
+  }
+
+  assert(overall_status.ok() || might_need_garbage_collect_);
+  return overall_status;
+}
+
+// ------- BackupMeta class --------
+
+IOStatus BackupEngineImpl::BackupMeta::AddFile(
+    std::shared_ptr<FileInfo> file_info) {
+  auto itr = file_infos_->find(file_info->filename);
+  if (itr == file_infos_->end()) {
+    auto ret = file_infos_->insert({file_info->filename, file_info});
+    if (ret.second) {
+      itr = ret.first;
+      itr->second->refs = 1;
+    } else {
+      // if this happens, something is seriously wrong
+      return IOStatus::Corruption("In memory metadata insertion error");
+    }
+  } else {
+    // Compare sizes, because we scanned that off the filesystem on both
+    // ends. This is like a check in VerifyBackup.
+    if (itr->second->size != file_info->size) {
+      std::string msg = "Size mismatch for existing backup file: ";
+      msg.append(file_info->filename);
+      msg.append(" Size in backup is " + std::to_string(itr->second->size) +
+                 " while size in DB is " + std::to_string(file_info->size));
+      msg.append(
+          " If this DB file checks as not corrupt, try deleting old"
+          " backups or backing up to a different backup directory.");
+      return IOStatus::Corruption(msg);
+    }
+    if (file_info->checksum_hex.empty()) {
+      // No checksum available to check
+    } else if (itr->second->checksum_hex.empty()) {
+      // Remember checksum if newly acquired
+      itr->second->checksum_hex = file_info->checksum_hex;
+    } else if (itr->second->checksum_hex != file_info->checksum_hex) {
+      // Note: to save I/O, these will be equal trivially on already backed
+      // up files that don't have the checksum in their name. And it should
+      // never fail for files that do have checksum in their name.
+
+      // Should never reach here, but produce an appropriate corruption
+      // message in case we do in a release build.
+      assert(false);
+      std::string msg = "Checksum mismatch for existing backup file: ";
+      msg.append(file_info->filename);
+      msg.append(" Expected checksum is " + itr->second->checksum_hex +
+                 " while computed checksum is " + file_info->checksum_hex);
+      msg.append(
+          " If this DB file checks as not corrupt, try deleting old"
+          " backups or backing up to a different backup directory.");
+      return IOStatus::Corruption(msg);
+    }
+    ++itr->second->refs;  // increase refcount if already present
+  }
+
+  size_ += file_info->size;
+  files_.push_back(itr->second);
+
+  return IOStatus::OK();
+}
+
+IOStatus BackupEngineImpl::BackupMeta::Delete(bool delete_meta) {
+  IOStatus io_s;
+  for (const auto& file : files_) {
+    --file->refs;  // decrease refcount
+  }
+  files_.clear();
+  // delete meta file
+  if (delete_meta) {
+    io_s = fs_->FileExists(meta_filename_, iooptions_, nullptr);
+    if (io_s.ok()) {
+      io_s = fs_->DeleteFile(meta_filename_, iooptions_, nullptr);
+    } else if (io_s.IsNotFound()) {
+      io_s = IOStatus::OK();  // nothing to delete
+    }
+  }
+  timestamp_ = 0;
+  return io_s;
+}
+
+// Constants for backup meta file schema (see LoadFromFile)
+const std::string kSchemaVersionPrefix{"schema_version "};
+const std::string kFooterMarker{"// FOOTER"};
+
+const std::string kAppMetaDataFieldName{"metadata"};
+
+// WART: The checksums are crc32c but named "crc32"
+const std::string kFileCrc32cFieldName{"crc32"};
+const std::string kFileSizeFieldName{"size"};
+const std::string kTemperatureFieldName{"temp"};
+
+// Marks a (future) field that should cause failure if not recognized.
+// Other fields are assumed to be ignorable. For example, in the future
+// we might add
+//  ni::file_name_escape uri_percent
+// to indicate all file names have had spaces and special characters
+// escaped using a URI percent encoding.
+const std::string kNonIgnorableFieldPrefix{"ni::"};
+
+// Each backup meta file is of the format (schema version 1):
+//----------------------------------------------------------
+// <timestamp>
+// <seq number>
+// metadata <metadata> (optional)
+// <number of files>
+// <file1> crc32 <crc32c_as_unsigned_decimal>
+// <file2> crc32 <crc32c_as_unsigned_decimal>
+// ...
+//----------------------------------------------------------
+//
+// For schema version 2.x (not in public APIs, but
+// forward-compatibility started):
+//----------------------------------------------------------
+// schema_version <ver>
+// <timestamp>
+// <seq number>
+// [<field name> <field data>]
+// ...
+// <number of files>
+// <file1>( <field name> <field data no spaces>)*
+// <file2>( <field name> <field data no spaces>)*
+// ...
+// [// FOOTER]
+// [<field name> <field data>]
+// ...
+//----------------------------------------------------------
+// where
+// <ver> ::= [0-9]+([.][0-9]+)
+// <field name> ::= [A-Za-z_][A-Za-z_0-9.]+
+// <field data> is anything but newline
+// <field data no spaces> is anything but space and newline
+// Although "// FOOTER" wouldn't strictly be required as a delimiter
+// given the number of files is included, it is there for parsing
+// sanity in case of corruption. It is only required if followed
+// by footer fields, such as a checksum of the meta file (so far).
+// Unrecognized fields are ignored, to support schema evolution on
+// non-critical features with forward compatibility. Update schema
+// major version for breaking changes. Schema minor versions are indicated
+// only for diagnostic/debugging purposes.
+//
+// Fields in schema version 2.0:
+// * Top-level meta fields:
+//   * Only "metadata" as in schema version 1
+// * File meta fields:
+//   * "crc32" - a crc32c checksum as in schema version 1
+//   * "size" - the size of the file (new)
+// * Footer meta fields:
+//   * None yet (future use for meta file checksum anticipated)
+//
+IOStatus BackupEngineImpl::BackupMeta::LoadFromFile(
+    const std::string& backup_dir,
+    const std::unordered_map<std::string, uint64_t>& abs_path_to_size,
+    RateLimiter* rate_limiter, Logger* info_log,
+    std::unordered_set<std::string>* reported_ignored_fields) {
+  assert(reported_ignored_fields);
+  assert(Empty());
+
+  std::unique_ptr<LineFileReader> backup_meta_reader;
+  {
+    IOStatus io_s = LineFileReader::Create(fs_, meta_filename_, FileOptions(),
+                                           &backup_meta_reader,
+                                           nullptr /* dbg */, rate_limiter);
+    if (!io_s.ok()) {
+      return io_s;
+    }
+  }
+
+  // If we don't read an explicit schema_version, that implies version 1,
+  // which is what we call the original backup meta schema.
+  int schema_major_version = 1;
+
+  // Failures handled at the end
+  std::string line;
+  if (backup_meta_reader->ReadLine(&line,
+                                   Env::IO_LOW /* rate_limiter_priority */)) {
+    if (StartsWith(line, kSchemaVersionPrefix)) {
+      std::string ver = line.substr(kSchemaVersionPrefix.size());
+      if (ver == "2" || StartsWith(ver, "2.")) {
+        schema_major_version = 2;
+      } else {
+        return IOStatus::NotSupported(
+            "Unsupported/unrecognized schema version: " + ver);
+      }
+      line.clear();
+    } else if (line.empty()) {
+      return IOStatus::Corruption("Unexpected empty line");
+    }
+  }
+  if (!line.empty()) {
+    timestamp_ = std::strtoull(line.c_str(), nullptr, /*base*/ 10);
+  } else if (backup_meta_reader->ReadLine(
+                 &line, Env::IO_LOW /* rate_limiter_priority */)) {
+    timestamp_ = std::strtoull(line.c_str(), nullptr, /*base*/ 10);
+  }
+  if (backup_meta_reader->ReadLine(&line,
+                                   Env::IO_LOW /* rate_limiter_priority */)) {
+    sequence_number_ = std::strtoull(line.c_str(), nullptr, /*base*/ 10);
+  }
+  uint32_t num_files = UINT32_MAX;
+  while (backup_meta_reader->ReadLine(
+      &line, Env::IO_LOW /* rate_limiter_priority */)) {
+    if (line.empty()) {
+      return IOStatus::Corruption("Unexpected empty line");
+    }
+    // Number -> number of files -> exit loop reading optional meta fields
+    if (line[0] >= '0' && line[0] <= '9') {
+      num_files = static_cast<uint32_t>(strtoul(line.c_str(), nullptr, 10));
+      break;
+    }
+    // else, must be a meta field assignment
+    auto space_pos = line.find_first_of(' ');
+    if (space_pos == std::string::npos) {
+      return IOStatus::Corruption("Expected number of files or meta field");
+    }
+    std::string field_name = line.substr(0, space_pos);
+    std::string field_data = line.substr(space_pos + 1);
+    if (field_name == kAppMetaDataFieldName) {
+      // app metadata present
+      bool decode_success = Slice(field_data).DecodeHex(&app_metadata_);
+      if (!decode_success) {
+        return IOStatus::Corruption(
+            "Failed to decode stored hex encoded app metadata");
+      }
+    } else if (schema_major_version < 2) {
+      return IOStatus::Corruption("Expected number of files or \"" +
+                                  kAppMetaDataFieldName + "\" field");
+    } else if (StartsWith(field_name, kNonIgnorableFieldPrefix)) {
+      return IOStatus::NotSupported("Unrecognized non-ignorable meta field " +
+                                    field_name + " (from future version?)");
+    } else {
+      // Warn the first time we see any particular unrecognized meta field
+      if (reported_ignored_fields->insert("meta:" + field_name).second) {
+        ROCKS_LOG_WARN(info_log, "Ignoring unrecognized backup meta field %s",
+                       field_name.c_str());
+      }
+    }
+  }
+  std::vector<std::shared_ptr<FileInfo>> files;
+  bool footer_present = false;
+  while (backup_meta_reader->ReadLine(
+      &line, Env::IO_LOW /* rate_limiter_priority */)) {
+    std::vector<std::string> components = StringSplit(line, ' ');
+
+    if (components.size() < 1) {
+      return IOStatus::Corruption("Empty line instead of file entry.");
+    }
+    if (schema_major_version >= 2 && components.size() == 2 &&
+        line == kFooterMarker) {
+      footer_present = true;
+      break;
+    }
+
+    const std::string& filename = components[0];
+
+    uint64_t actual_size;
+    const std::shared_ptr<FileInfo> file_info = GetFile(filename);
+    if (file_info) {
+      actual_size = file_info->size;
+    } else {
+      std::string abs_path = backup_dir + "/" + filename;
+      auto e = abs_path_to_size.find(abs_path);
+      if (e == abs_path_to_size.end()) {
+        return IOStatus::Corruption(
+            "Pathname in meta file not found on disk: " + abs_path);
+      }
+      actual_size = e->second;
+    }
+
+    if (schema_major_version >= 2) {
+      if (components.size() % 2 != 1) {
+        return IOStatus::Corruption(
+            "Bad number of line components for file entry.");
+      }
+    } else {
+      // Check restricted original schema
+      if (components.size() < 3) {
+        return IOStatus::Corruption("File checksum is missing for " + filename +
+                                    " in " + meta_filename_);
+      }
+      if (components[1] != kFileCrc32cFieldName) {
+        return IOStatus::Corruption("Unknown checksum type for " + filename +
+                                    " in " + meta_filename_);
+      }
+      if (components.size() > 3) {
+        return IOStatus::Corruption("Extra data for entry " + filename +
+                                    " in " + meta_filename_);
+      }
+    }
+
+    std::string checksum_hex;
+    Temperature temp = Temperature::kUnknown;
+    for (unsigned i = 1; i < components.size(); i += 2) {
+      const std::string& field_name = components[i];
+      const std::string& field_data = components[i + 1];
+
+      if (field_name == kFileCrc32cFieldName) {
+        uint32_t checksum_value =
+            static_cast<uint32_t>(strtoul(field_data.c_str(), nullptr, 10));
+        if (field_data != std::to_string(checksum_value)) {
+          return IOStatus::Corruption("Invalid checksum value for " + filename +
+                                      " in " + meta_filename_);
+        }
+        checksum_hex = ChecksumInt32ToHex(checksum_value);
+      } else if (field_name == kFileSizeFieldName) {
+        uint64_t ex_size =
+            std::strtoull(field_data.c_str(), nullptr, /*base*/ 10);
+        if (ex_size != actual_size) {
+          return IOStatus::Corruption(
+              "For file " + filename + " expected size " +
+              std::to_string(ex_size) + " but found size" +
+              std::to_string(actual_size));
+        }
+      } else if (field_name == kTemperatureFieldName) {
+        auto iter = temperature_string_map.find(field_data);
+        if (iter != temperature_string_map.end()) {
+          temp = iter->second;
+        } else {
+          // Could report corruption, but in case of new temperatures added
+          // in future, letting those map to kUnknown which should generally
+          // be safe.
+          temp = Temperature::kUnknown;
+        }
+      } else if (StartsWith(field_name, kNonIgnorableFieldPrefix)) {
+        return IOStatus::NotSupported("Unrecognized non-ignorable file field " +
+                                      field_name + " (from future version?)");
+      } else {
+        // Warn the first time we see any particular unrecognized file field
+        if (reported_ignored_fields->insert("file:" + field_name).second) {
+          ROCKS_LOG_WARN(info_log, "Ignoring unrecognized backup file field %s",
+                         field_name.c_str());
+        }
+      }
+    }
+
+    files.emplace_back(new FileInfo(filename, actual_size, checksum_hex,
+                                    /*id*/ "", /*sid*/ "", temp));
+  }
+
+  if (footer_present) {
+    assert(schema_major_version >= 2);
+    while (backup_meta_reader->ReadLine(
+        &line, Env::IO_LOW /* rate_limiter_priority */)) {
+      if (line.empty()) {
+        return IOStatus::Corruption("Unexpected empty line");
+      }
+      auto space_pos = line.find_first_of(' ');
+      if (space_pos == std::string::npos) {
+        return IOStatus::Corruption("Expected footer field");
+      }
+      std::string field_name = line.substr(0, space_pos);
+      std::string field_data = line.substr(space_pos + 1);
+      if (StartsWith(field_name, kNonIgnorableFieldPrefix)) {
+        return IOStatus::NotSupported("Unrecognized non-ignorable field " +
+                                      field_name + " (from future version?)");
+      } else if (reported_ignored_fields->insert("footer:" + field_name)
+                     .second) {
+        // Warn the first time we see any particular unrecognized footer field
+        ROCKS_LOG_WARN(info_log,
+                       "Ignoring unrecognized backup meta footer field %s",
+                       field_name.c_str());
+      }
+    }
+  }
+
+  {
+    IOStatus io_s = backup_meta_reader->GetStatus();
+    if (!io_s.ok()) {
+      return io_s;
+    }
+  }
+
+  if (num_files != files.size()) {
+    return IOStatus::Corruption(
+        "Inconsistent number of files or missing/incomplete header in " +
+        meta_filename_);
+  }
+
+  files_.reserve(files.size());
+  for (const auto& file_info : files) {
+    IOStatus io_s = AddFile(file_info);
+    if (!io_s.ok()) {
+      return io_s;
+    }
+  }
+
+  return IOStatus::OK();
+}
+
+const std::vector<std::string> minor_version_strings{
+    "",  // invalid major version 0
+    "",  // implicit major version 1
+    "2.0",
+};
+
+IOStatus BackupEngineImpl::BackupMeta::StoreToFile(
+    bool sync, int schema_version,
+    const TEST_BackupMetaSchemaOptions* schema_test_options) {
+  if (schema_version < 1) {
+    return IOStatus::InvalidArgument(
+        "BackupEngineOptions::schema_version must be >= 1");
+  }
+  if (schema_version > static_cast<int>(minor_version_strings.size() - 1)) {
+    return IOStatus::NotSupported(
+        "Only BackupEngineOptions::schema_version <= " +
+        std::to_string(minor_version_strings.size() - 1) + " is supported");
+  }
+  std::string ver = minor_version_strings[schema_version];
+
+  // Need schema_version >= 2 for TEST_BackupMetaSchemaOptions
+  assert(schema_version >= 2 || schema_test_options == nullptr);
+
+  IOStatus io_s;
+  std::unique_ptr<FSWritableFile> backup_meta_file;
+  FileOptions file_options;
+  file_options.use_mmap_writes = false;
+  file_options.use_direct_writes = false;
+  io_s = fs_->NewWritableFile(meta_tmp_filename_, file_options,
+                              &backup_meta_file, nullptr);
+  if (!io_s.ok()) {
+    return io_s;
+  }
+
+  std::ostringstream buf;
+  if (schema_test_options) {
+    // override for testing
+    ver = schema_test_options->version;
+  }
+  if (!ver.empty()) {
+    assert(schema_version >= 2);
+    buf << kSchemaVersionPrefix << ver << "\n";
+  }
+  buf << static_cast<unsigned long long>(timestamp_) << "\n";
+  buf << sequence_number_ << "\n";
+
+  if (!app_metadata_.empty()) {
+    std::string hex_encoded_metadata =
+        Slice(app_metadata_).ToString(/* hex */ true);
+    buf << kAppMetaDataFieldName << " " << hex_encoded_metadata << "\n";
+  }
+  if (schema_test_options) {
+    for (auto& e : schema_test_options->meta_fields) {
+      buf << e.first << " " << e.second << "\n";
+    }
+  }
+  buf << files_.size() << "\n";
+
+  for (const auto& file : files_) {
+    buf << file->filename;
+    if (schema_test_options == nullptr ||
+        schema_test_options->crc32c_checksums) {
+      // use crc32c for now, switch to something else if needed
+      buf << " " << kFileCrc32cFieldName << " "
+          << ChecksumHexToInt32(file->checksum_hex);
+    }
+    if (schema_version >= 2 && file->temp != Temperature::kUnknown) {
+      buf << " " << kTemperatureFieldName << " "
+          << temperature_to_string[file->temp];
+    }
+    if (schema_test_options && schema_test_options->file_sizes) {
+      buf << " " << kFileSizeFieldName << " " << std::to_string(file->size);
+    }
+    if (schema_test_options) {
+      for (auto& e : schema_test_options->file_fields) {
+        buf << " " << e.first << " " << e.second;
+      }
+    }
+    buf << "\n";
+  }
+
+  if (schema_test_options && !schema_test_options->footer_fields.empty()) {
+    buf << kFooterMarker << "\n";
+    for (auto& e : schema_test_options->footer_fields) {
+      buf << e.first << " " << e.second << "\n";
+    }
+  }
+
+  io_s = backup_meta_file->Append(Slice(buf.str()), iooptions_, nullptr);
+  IOSTATS_ADD(bytes_written, buf.str().size());
+  if (io_s.ok() && sync) {
+    io_s = backup_meta_file->Sync(iooptions_, nullptr);
+  }
+  if (io_s.ok()) {
+    io_s = backup_meta_file->Close(iooptions_, nullptr);
+  }
+  if (io_s.ok()) {
+    io_s = fs_->RenameFile(meta_tmp_filename_, meta_filename_, iooptions_,
+                           nullptr);
+  }
+  return io_s;
+}
+}  // namespace
+
+IOStatus BackupEngineReadOnly::Open(const BackupEngineOptions& options,
+                                    Env* env,
+                                    BackupEngineReadOnly** backup_engine_ptr) {
+  if (options.destroy_old_data) {
+    return IOStatus::InvalidArgument(
+        "Can't destroy old data with ReadOnly BackupEngine");
+  }
+  std::unique_ptr<BackupEngineImplThreadSafe> backup_engine(
+      new BackupEngineImplThreadSafe(options, env, true /*read_only*/));
+  auto s = backup_engine->Initialize();
+  if (!s.ok()) {
+    *backup_engine_ptr = nullptr;
+    return s;
+  }
+  *backup_engine_ptr = backup_engine.release();
+  return IOStatus::OK();
+}
+
+void TEST_SetBackupMetaSchemaOptions(
+    BackupEngine* engine, const TEST_BackupMetaSchemaOptions& options) {
+  BackupEngineImplThreadSafe* impl =
+      static_cast_with_check<BackupEngineImplThreadSafe>(engine);
+  impl->TEST_SetBackupMetaSchemaOptions(options);
+}
+
+void TEST_SetDefaultRateLimitersClock(
+    BackupEngine* engine,
+    const std::shared_ptr<SystemClock>& backup_rate_limiter_clock,
+    const std::shared_ptr<SystemClock>& restore_rate_limiter_clock) {
+  BackupEngineImplThreadSafe* impl =
+      static_cast_with_check<BackupEngineImplThreadSafe>(engine);
+  impl->TEST_SetDefaultRateLimitersClock(backup_rate_limiter_clock,
+                                         restore_rate_limiter_clock);
+}
+}  // namespace ROCKSDB_NAMESPACE
+
+#endif  // ROCKSDB_LITE
diff --git a/src/rocksdb/utilities/backup/backup_engine_impl.h b/src/rocksdb/utilities/backup/backup_engine_impl.h
new file mode 100644
index 000000000..398f47f27
--- /dev/null
+++ b/src/rocksdb/utilities/backup/backup_engine_impl.h
@@ -0,0 +1,36 @@
+//  Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#pragma once
+#ifndef ROCKSDB_LITE
+
+#include "rocksdb/utilities/backup_engine.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+struct TEST_BackupMetaSchemaOptions {
+  std::string version = "2";
+  bool crc32c_checksums = false;
+  bool file_sizes = true;
+  std::map<std::string, std::string> meta_fields;
+  std::map<std::string, std::string> file_fields;
+  std::map<std::string, std::string> footer_fields;
+};
+
+// Modifies the BackupEngine(Impl) to write backup meta files using the
+// unpublished schema version 2, for the life of this object (not backup_dir).
+// TEST_BackupMetaSchemaOptions offers some customization for testing.
+void TEST_SetBackupMetaSchemaOptions(
+    BackupEngine* engine, const TEST_BackupMetaSchemaOptions& options);
+
+// Modifies the BackupEngine(Impl) to use specified clocks for backup and
+// restore rate limiters created by default if not specified by users for
+// test speedup.
+void TEST_SetDefaultRateLimitersClock(
+    BackupEngine* engine,
+    const std::shared_ptr<SystemClock>& backup_rate_limiter_clock = nullptr,
+    const std::shared_ptr<SystemClock>& restore_rate_limiter_clock = nullptr);
+}  // namespace ROCKSDB_NAMESPACE
+#endif  // ROCKSDB_LITE
diff --git a/src/rocksdb/utilities/backup/backup_engine_test.cc b/src/rocksdb/utilities/backup/backup_engine_test.cc
new file mode 100644
index 000000000..d1f74f769
--- /dev/null
+++ b/src/rocksdb/utilities/backup/backup_engine_test.cc
@@ -0,0 +1,4219 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+//
+// Copyright (c) 2011 The LevelDB Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file. See the AUTHORS file for names of contributors.
+
+#if !defined(ROCKSDB_LITE) && !defined(OS_WIN)
+
+#include "rocksdb/utilities/backup_engine.h"
+
+#include <algorithm>
+#include <array>
+#include <atomic>
+#include <cstddef>
+#include <cstdint>
+#include <limits>
+#include <memory>
+#include <random>
+#include <string>
+#include <utility>
+
+#include "db/db_impl/db_impl.h"
+#include "db/db_test_util.h"
+#include "env/composite_env_wrapper.h"
+#include "env/env_chroot.h"
+#include "file/filename.h"
+#include "port/port.h"
+#include "port/stack_trace.h"
+#include "rocksdb/advanced_options.h"
+#include "rocksdb/env.h"
+#include "rocksdb/file_checksum.h"
+#include "rocksdb/rate_limiter.h"
+#include "rocksdb/statistics.h"
+#include "rocksdb/transaction_log.h"
+#include "rocksdb/types.h"
+#include "rocksdb/utilities/options_util.h"
+#include "rocksdb/utilities/stackable_db.h"
+#include "test_util/sync_point.h"
+#include "test_util/testharness.h"
+#include "test_util/testutil.h"
+#include "util/cast_util.h"
+#include "util/mutexlock.h"
+#include "util/random.h"
+#include "util/rate_limiter.h"
+#include "util/stderr_logger.h"
+#include "util/string_util.h"
+#include "utilities/backup/backup_engine_impl.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+namespace {
+using ShareFilesNaming = BackupEngineOptions::ShareFilesNaming;
+const auto kLegacyCrc32cAndFileSize =
+    BackupEngineOptions::kLegacyCrc32cAndFileSize;
+const auto kUseDbSessionId = BackupEngineOptions::kUseDbSessionId;
+const auto kFlagIncludeFileSize = BackupEngineOptions::kFlagIncludeFileSize;
+const auto kNamingDefault = kUseDbSessionId | kFlagIncludeFileSize;
+
+class DummyDB : public StackableDB {
+ public:
+  /* implicit */
+  DummyDB(const Options& options, const std::string& dbname)
+      : StackableDB(nullptr),
+        options_(options),
+        dbname_(dbname),
+        deletions_enabled_(true),
+        sequence_number_(0) {}
+
+  SequenceNumber GetLatestSequenceNumber() const override {
+    return ++sequence_number_;
+  }
+
+  const std::string& GetName() const override { return dbname_; }
+
+  Env* GetEnv() const override { return options_.env; }
+
+  using DB::GetOptions;
+  Options GetOptions(ColumnFamilyHandle* /*column_family*/) const override {
+    return options_;
+  }
+
+  DBOptions GetDBOptions() const override { return DBOptions(options_); }
+
+  Status EnableFileDeletions(bool /*force*/) override {
+    EXPECT_TRUE(!deletions_enabled_);
+    deletions_enabled_ = true;
+    return Status::OK();
+  }
+
+  Status DisableFileDeletions() override {
+    EXPECT_TRUE(deletions_enabled_);
+    deletions_enabled_ = false;
+    return Status::OK();
+  }
+
+  ColumnFamilyHandle* DefaultColumnFamily() const override { return nullptr; }
+
+  Status GetLiveFilesStorageInfo(
+      const LiveFilesStorageInfoOptions& opts,
+      std::vector<LiveFileStorageInfo>* files) override {
+    uint64_t number;
+    FileType type;
+    files->clear();
+    for (auto& f : live_files_) {
+      bool success = ParseFileName(f, &number, &type);
+      if (!success) {
+        return Status::InvalidArgument("Bad file name: " + f);
+      }
+      files->emplace_back();
+      LiveFileStorageInfo& info = files->back();
+      info.relative_filename = f;
+      info.directory = dbname_;
+      info.file_number = number;
+      info.file_type = type;
+      if (type == kDescriptorFile) {
+        info.size = 100;  // See TestFs::GetChildrenFileAttributes below
+        info.trim_to_size = true;
+      } else if (type == kCurrentFile) {
+        info.size = 0;
+        info.trim_to_size = true;
+      } else {
+        info.size = 200;  // See TestFs::GetChildrenFileAttributes below
+      }
+      if (opts.include_checksum_info) {
+        info.file_checksum = kUnknownFileChecksum;
+        info.file_checksum_func_name = kUnknownFileChecksumFuncName;
+      }
+    }
+    return Status::OK();
+  }
+
+  // To avoid FlushWAL called on stacked db which is nullptr
+  Status FlushWAL(bool /*sync*/) override { return Status::OK(); }
+
+  std::vector<std::string> live_files_;
+
+ private:
+  Options options_;
+  std::string dbname_;
+  bool deletions_enabled_;
+  mutable SequenceNumber sequence_number_;
+};  // DummyDB
+
+class TestFs : public FileSystemWrapper {
+ public:
+  explicit TestFs(const std::shared_ptr<FileSystem>& t)
+      : FileSystemWrapper(t) {}
+  const char* Name() const override { return "TestFs"; }
+
+  class DummySequentialFile : public FSSequentialFile {
+   public:
+    explicit DummySequentialFile(bool fail_reads)
+        : FSSequentialFile(), rnd_(5), fail_reads_(fail_reads) {}
+    IOStatus Read(size_t n, const IOOptions&, Slice* result, char* scratch,
+                  IODebugContext*) override {
+      if (fail_reads_) {
+        return IOStatus::IOError();
+      }
+      size_t read_size = (n > size_left) ? size_left : n;
+      for (size_t i = 0; i < read_size; ++i) {
+        scratch[i] = rnd_.Next() & 255;
+      }
+      *result = Slice(scratch, read_size);
+      size_left -= read_size;
+      return IOStatus::OK();
+    }
+
+    IOStatus Skip(uint64_t n) override {
+      size_left = (n > size_left) ? size_left - n : 0;
+      return IOStatus::OK();
+    }
+
+   private:
+    size_t size_left = 200;
+    Random rnd_;
+    bool fail_reads_;
+  };
+
+  IOStatus NewSequentialFile(const std::string& f, const FileOptions& file_opts,
+                             std::unique_ptr<FSSequentialFile>* r,
+                             IODebugContext* dbg) override {
+    MutexLock l(&mutex_);
+    if (dummy_sequential_file_) {
+      r->reset(
+          new TestFs::DummySequentialFile(dummy_sequential_file_fail_reads_));
+      return IOStatus::OK();
+    } else {
+      IOStatus s = FileSystemWrapper::NewSequentialFile(f, file_opts, r, dbg);
+      if (s.ok()) {
+        if ((*r)->use_direct_io()) {
+          ++num_direct_seq_readers_;
+        }
+        ++num_seq_readers_;
+      }
+      return s;
+    }
+  }
+
+  IOStatus NewWritableFile(const std::string& f, const FileOptions& file_opts,
+                           std::unique_ptr<FSWritableFile>* r,
+                           IODebugContext* dbg) override {
+    MutexLock l(&mutex_);
+    written_files_.push_back(f);
+    if (limit_written_files_ == 0) {
+      return IOStatus::NotSupported("Limit on written files reached");
+    }
+    limit_written_files_--;
+    IOStatus s = FileSystemWrapper::NewWritableFile(f, file_opts, r, dbg);
+    if (s.ok()) {
+      if ((*r)->use_direct_io()) {
+        ++num_direct_writers_;
+      }
+      ++num_writers_;
+    }
+    return s;
+  }
+
+  IOStatus NewRandomAccessFile(const std::string& f,
+                               const FileOptions& file_opts,
+                               std::unique_ptr<FSRandomAccessFile>* r,
+                               IODebugContext* dbg) override {
+    MutexLock l(&mutex_);
+    IOStatus s = FileSystemWrapper::NewRandomAccessFile(f, file_opts, r, dbg);
+    if (s.ok()) {
+      if ((*r)->use_direct_io()) {
+        ++num_direct_rand_readers_;
+      }
+      ++num_rand_readers_;
+    }
+    return s;
+  }
+
+  IOStatus DeleteFile(const std::string& f, const IOOptions& options,
+                      IODebugContext* dbg) override {
+    MutexLock l(&mutex_);
+    if (fail_delete_files_) {
+      return IOStatus::IOError();
+    }
+    EXPECT_GT(limit_delete_files_, 0U);
+    limit_delete_files_--;
+    return FileSystemWrapper::DeleteFile(f, options, dbg);
+  }
+
+  IOStatus DeleteDir(const std::string& d, const IOOptions& options,
+                     IODebugContext* dbg) override {
+    MutexLock l(&mutex_);
+    if (fail_delete_files_) {
+      return IOStatus::IOError();
+    }
+    return FileSystemWrapper::DeleteDir(d, options, dbg);
+  }
+
+  void AssertWrittenFiles(std::vector<std::string>& should_have_written) {
+    MutexLock l(&mutex_);
+    std::sort(should_have_written.begin(), should_have_written.end());
+    std::sort(written_files_.begin(), written_files_.end());
+
+    ASSERT_EQ(should_have_written, written_files_);
+  }
+
+  void ClearWrittenFiles() {
+    MutexLock l(&mutex_);
+    written_files_.clear();
+  }
+
+  void SetLimitWrittenFiles(uint64_t limit) {
+    MutexLock l(&mutex_);
+    limit_written_files_ = limit;
+  }
+
+  void SetLimitDeleteFiles(uint64_t limit) {
+    MutexLock l(&mutex_);
+    limit_delete_files_ = limit;
+  }
+
+  void SetDeleteFileFailure(bool fail) {
+    MutexLock l(&mutex_);
+    fail_delete_files_ = fail;
+  }
+
+  void SetDummySequentialFile(bool dummy_sequential_file) {
+    MutexLock l(&mutex_);
+    dummy_sequential_file_ = dummy_sequential_file;
+  }
+  void SetDummySequentialFileFailReads(bool dummy_sequential_file_fail_reads) {
+    MutexLock l(&mutex_);
+    dummy_sequential_file_fail_reads_ = dummy_sequential_file_fail_reads;
+  }
+
+  void SetGetChildrenFailure(bool fail) { get_children_failure_ = fail; }
+  IOStatus GetChildren(const std::string& dir, const IOOptions& io_opts,
+                       std::vector<std::string>* r,
+                       IODebugContext* dbg) override {
+    if (get_children_failure_) {
+      return IOStatus::IOError("SimulatedFailure");
+    }
+    return FileSystemWrapper::GetChildren(dir, io_opts, r, dbg);
+  }
+
+  // Some test cases do not actually create the test files (e.g., see
+  // DummyDB::live_files_) - for those cases, we mock those files' attributes
+  // so CreateNewBackup() can get their attributes.
+  void SetFilenamesForMockedAttrs(const std::vector<std::string>& filenames) {
+    filenames_for_mocked_attrs_ = filenames;
+  }
+  IOStatus GetChildrenFileAttributes(const std::string& dir,
+                                     const IOOptions& options,
+                                     std::vector<FileAttributes>* result,
+                                     IODebugContext* dbg) override {
+    if (filenames_for_mocked_attrs_.size() > 0) {
+      for (const auto& filename : filenames_for_mocked_attrs_) {
+        uint64_t size_bytes = 200;  // Match TestFs
+        if (filename.find("MANIFEST") == 0) {
+          size_bytes = 100;  // Match DummyDB::GetLiveFiles
+        }
+        result->push_back({dir + "/" + filename, size_bytes});
+      }
+      return IOStatus::OK();
+    }
+    return FileSystemWrapper::GetChildrenFileAttributes(dir, options, result,
+                                                        dbg);
+  }
+
+  IOStatus GetFileSize(const std::string& f, const IOOptions& options,
+                       uint64_t* s, IODebugContext* dbg) override {
+    if (filenames_for_mocked_attrs_.size() > 0) {
+      auto fname = f.substr(f.find_last_of('/') + 1);
+      auto filename_iter = std::find(filenames_for_mocked_attrs_.begin(),
+                                     filenames_for_mocked_attrs_.end(), fname);
+      if (filename_iter != filenames_for_mocked_attrs_.end()) {
+        *s = 200;  // Match TestFs
+        if (fname.find("MANIFEST") == 0) {
+          *s = 100;  // Match DummyDB::GetLiveFiles
+        }
+        return IOStatus::OK();
+      }
+      return IOStatus::NotFound(fname);
+    }
+    return FileSystemWrapper::GetFileSize(f, options, s, dbg);
+  }
+
+  void SetCreateDirIfMissingFailure(bool fail) {
+    create_dir_if_missing_failure_ = fail;
+  }
+  IOStatus CreateDirIfMissing(const std::string& d, const IOOptions& options,
+                              IODebugContext* dbg) override {
+    if (create_dir_if_missing_failure_) {
+      return IOStatus::IOError("SimulatedFailure");
+    }
+    return FileSystemWrapper::CreateDirIfMissing(d, options, dbg);
+  }
+
+  void SetNewDirectoryFailure(bool fail) { new_directory_failure_ = fail; }
+  IOStatus NewDirectory(const std::string& name, const IOOptions& io_opts,
+                        std::unique_ptr<FSDirectory>* result,
+                        IODebugContext* dbg) override {
+    if (new_directory_failure_) {
+      return IOStatus::IOError("SimulatedFailure");
+    }
+    return FileSystemWrapper::NewDirectory(name, io_opts, result, dbg);
+  }
+
+  void ClearFileOpenCounters() {
+    MutexLock l(&mutex_);
+    num_rand_readers_ = 0;
+    num_direct_rand_readers_ = 0;
+    num_seq_readers_ = 0;
+    num_direct_seq_readers_ = 0;
+    num_writers_ = 0;
+    num_direct_writers_ = 0;
+  }
+
+  int num_rand_readers() { return num_rand_readers_; }
+  int num_direct_rand_readers() { return num_direct_rand_readers_; }
+  int num_seq_readers() { return num_seq_readers_; }
+  int num_direct_seq_readers() { return num_direct_seq_readers_; }
+  int num_writers() { return num_writers_; }
+  // FIXME(?): unused
+  int num_direct_writers() { return num_direct_writers_; }
+
+ private:
+  port::Mutex mutex_;
+  bool dummy_sequential_file_ = false;
+  bool dummy_sequential_file_fail_reads_ = false;
+  std::vector<std::string> written_files_;
+  std::vector<std::string> filenames_for_mocked_attrs_;
+  uint64_t limit_written_files_ = 1000000;
+  uint64_t limit_delete_files_ = 1000000;
+  bool fail_delete_files_ = false;
+
+  bool get_children_failure_ = false;
+  bool create_dir_if_missing_failure_ = false;
+  bool new_directory_failure_ = false;
+
+  // Keeps track of how many files of each type were successfully opened, and
+  // out of those, how many were opened with direct I/O.
+  std::atomic<int> num_rand_readers_{};
+  std::atomic<int> num_direct_rand_readers_{};
+  std::atomic<int> num_seq_readers_{};
+  std::atomic<int> num_direct_seq_readers_{};
+  std::atomic<int> num_writers_{};
+  std::atomic<int> num_direct_writers_{};
+};  // TestFs
+
+class FileManager : public EnvWrapper {
+ public:
+  explicit FileManager(Env* t) : EnvWrapper(t), rnd_(5) {}
+  const char* Name() const override { return "FileManager"; }
+
+  Status GetRandomFileInDir(const std::string& dir, std::string* fname,
+                            uint64_t* fsize) {
+    std::vector<FileAttributes> children;
+    auto s = GetChildrenFileAttributes(dir, &children);
+    if (!s.ok()) {
+      return s;
+    } else if (children.size() <= 2) {  // . and ..
+      return Status::NotFound("Empty directory: " + dir);
+    }
+    assert(fname != nullptr);
+    while (true) {
+      int i = rnd_.Next() % children.size();
+      fname->assign(dir + "/" + children[i].name);
+      *fsize = children[i].size_bytes;
+      return Status::OK();
+    }
+    // should never get here
+    assert(false);
+    return Status::NotFound("");
+  }
+
+  Status DeleteRandomFileInDir(const std::string& dir) {
+    std::vector<std::string> children;
+    Status s = GetChildren(dir, &children);
+    if (!s.ok()) {
+      return s;
+    }
+    while (true) {
+      int i = rnd_.Next() % children.size();
+      return DeleteFile(dir + "/" + children[i]);
+    }
+    // should never get here
+    assert(false);
+    return Status::NotFound("");
+  }
+
+  Status AppendToRandomFileInDir(const std::string& dir,
+                                 const std::string& data) {
+    std::vector<std::string> children;
+    Status s = GetChildren(dir, &children);
+    if (!s.ok()) {
+      return s;
+    }
+    while (true) {
+      int i = rnd_.Next() % children.size();
+      return WriteToFile(dir + "/" + children[i], data);
+    }
+    // should never get here
+    assert(false);
+    return Status::NotFound("");
+  }
+
+  Status CorruptFile(const std::string& fname, uint64_t bytes_to_corrupt) {
+    std::string file_contents;
+    Status s = ReadFileToString(this, fname, &file_contents);
+    if (!s.ok()) {
+      return s;
+    }
+    s = DeleteFile(fname);
+    if (!s.ok()) {
+      return s;
+    }
+
+    for (uint64_t i = 0; i < bytes_to_corrupt; ++i) {
+      std::string tmp = rnd_.RandomString(1);
+      file_contents[rnd_.Next() % file_contents.size()] = tmp[0];
+    }
+    return WriteToFile(fname, file_contents);
+  }
+
+  Status CorruptFileStart(const std::string& fname) {
+    std::string to_xor = "blah";
+    std::string file_contents;
+    Status s = ReadFileToString(this, fname, &file_contents);
+    if (!s.ok()) {
+      return s;
+    }
+    s = DeleteFile(fname);
+    if (!s.ok()) {
+      return s;
+    }
+    for (size_t i = 0; i < to_xor.size(); ++i) {
+      file_contents[i] ^= to_xor[i];
+    }
+    return WriteToFile(fname, file_contents);
+  }
+
+  Status CorruptChecksum(const std::string& fname, bool appear_valid) {
+    std::string metadata;
+    Status s = ReadFileToString(this, fname, &metadata);
+    if (!s.ok()) {
+      return s;
+    }
+    s = DeleteFile(fname);
+    if (!s.ok()) {
+      return s;
+    }
+
+    auto pos = metadata.find("private");
+    if (pos == std::string::npos) {
+      return Status::Corruption("private file is expected");
+    }
+    pos = metadata.find(" crc32 ", pos + 6);
+    if (pos == std::string::npos) {
+      return Status::Corruption("checksum not found");
+    }
+
+    if (metadata.size() < pos + 7) {
+      return Status::Corruption("bad CRC32 checksum value");
+    }
+
+    if (appear_valid) {
+      if (metadata[pos + 8] == '\n') {
+        // single digit value, safe to insert one more digit
+        metadata.insert(pos + 8, 1, '0');
+      } else {
+        metadata.erase(pos + 8, 1);
+      }
+    } else {
+      metadata[pos + 7] = 'a';
+    }
+
+    return WriteToFile(fname, metadata);
+  }
+
+  Status WriteToFile(const std::string& fname, const std::string& data) {
+    std::unique_ptr<WritableFile> file;
+    EnvOptions env_options;
+    env_options.use_mmap_writes = false;
+    Status s = EnvWrapper::NewWritableFile(fname, &file, env_options);
+    if (!s.ok()) {
+      return s;
+    }
+    return file->Append(Slice(data));
+  }
+
+ private:
+  Random rnd_;
+};  // FileManager
+
+// utility functions
+namespace {
+
+enum FillDBFlushAction {
+  kFlushMost,
+  kFlushAll,
+  kAutoFlushOnly,
+};
+
+// Many tests in this file expect FillDB to write at least one sst file,
+// so the default behavior (if not kAutoFlushOnly) of FillDB is to force
+// a flush. But to ensure coverage of the WAL file case, we also (by default)
+// do one Put after the Flush (kFlushMost).
+size_t FillDB(DB* db, int from, int to,
+              FillDBFlushAction flush_action = kFlushMost) {
+  size_t bytes_written = 0;
+  for (int i = from; i < to; ++i) {
+    std::string key = "testkey" + std::to_string(i);
+    std::string value = "testvalue" + std::to_string(i);
+    bytes_written += key.size() + value.size();
+
+    EXPECT_OK(db->Put(WriteOptions(), Slice(key), Slice(value)));
+
+    if (flush_action == kFlushMost && i == to - 2) {
+      EXPECT_OK(db->Flush(FlushOptions()));
+    }
+  }
+  if (flush_action == kFlushAll) {
+    EXPECT_OK(db->Flush(FlushOptions()));
+  }
+  return bytes_written;
+}
+
+void AssertExists(DB* db, int from, int to) {
+  for (int i = from; i < to; ++i) {
+    std::string key = "testkey" + std::to_string(i);
+    std::string value;
+    Status s = db->Get(ReadOptions(), Slice(key), &value);
+    ASSERT_EQ(value, "testvalue" + std::to_string(i));
+  }
+}
+
+void AssertEmpty(DB* db, int from, int to) {
+  for (int i = from; i < to; ++i) {
+    std::string key = "testkey" + std::to_string(i);
+    std::string value = "testvalue" + std::to_string(i);
+
+    Status s = db->Get(ReadOptions(), Slice(key), &value);
+    ASSERT_TRUE(s.IsNotFound());
+  }
+}
+}  // namespace
+
+class BackupEngineTest : public testing::Test {
+ public:
+  enum ShareOption {
+    kNoShare,
+    kShareNoChecksum,
+    kShareWithChecksum,
+  };
+
+  const std::vector<ShareOption> kAllShareOptions = {kNoShare, kShareNoChecksum,
+                                                     kShareWithChecksum};
+
+  BackupEngineTest() {
+    // set up files
+    std::string db_chroot = test::PerThreadDBPath("db_for_backup");
+    std::string backup_chroot = test::PerThreadDBPath("db_backups");
+    EXPECT_OK(Env::Default()->CreateDirIfMissing(db_chroot));
+    EXPECT_OK(Env::Default()->CreateDirIfMissing(backup_chroot));
+    dbname_ = "/tempdb";
+    backupdir_ = "/tempbk";
+    latest_backup_ = backupdir_ + "/LATEST_BACKUP";
+
+    // set up FileSystem & Envs
+    db_chroot_fs_ = NewChrootFileSystem(FileSystem::Default(), db_chroot);
+    backup_chroot_fs_ =
+        NewChrootFileSystem(FileSystem::Default(), backup_chroot);
+    test_db_fs_ = std::make_shared<TestFs>(db_chroot_fs_);
+    test_backup_fs_ = std::make_shared<TestFs>(backup_chroot_fs_);
+    SetEnvsFromFileSystems();
+
+    // set up db options
+    options_.create_if_missing = true;
+    options_.paranoid_checks = true;
+    options_.write_buffer_size = 1 << 17;  // 128KB
+    options_.wal_dir = dbname_;
+    options_.enable_blob_files = true;
+
+    // The sync option is not easily testable in unit tests, but should be
+    // smoke tested across all the other backup tests. However, it is
+    // certainly not worth doubling the runtime of backup tests for it.
+    // Thus, we can enable sync for one of our alternate testing
+    // configurations.
+    constexpr bool kUseSync =
+#ifdef ROCKSDB_MODIFY_NPHASH
+        true;
+#else
+        false;
+#endif  // ROCKSDB_MODIFY_NPHASH
+
+    // set up backup db options
+    engine_options_.reset(new BackupEngineOptions(
+        backupdir_, test_backup_env_.get(), /*share_table_files*/ true,
+        logger_.get(), kUseSync));
+
+    // most tests will use multi-threaded backups
+    engine_options_->max_background_operations = 7;
+
+    // delete old files in db
+    DestroyDBWithoutCheck(dbname_, options_);
+
+    // delete old LATEST_BACKUP file, which some tests create for compatibility
+    // testing.
+    backup_chroot_env_->DeleteFile(latest_backup_).PermitUncheckedError();
+  }
+
+  void SetEnvsFromFileSystems() {
+    db_chroot_env_.reset(
+        new CompositeEnvWrapper(Env::Default(), db_chroot_fs_));
+    backup_chroot_env_.reset(
+        new CompositeEnvWrapper(Env::Default(), backup_chroot_fs_));
+    test_db_env_.reset(new CompositeEnvWrapper(Env::Default(), test_db_fs_));
+    options_.env = test_db_env_.get();
+    test_backup_env_.reset(
+        new CompositeEnvWrapper(Env::Default(), test_backup_fs_));
+    if (engine_options_) {
+      engine_options_->backup_env = test_backup_env_.get();
+    }
+    file_manager_.reset(new FileManager(backup_chroot_env_.get()));
+    db_file_manager_.reset(new FileManager(db_chroot_env_.get()));
+
+    // Create logger
+    DBOptions logger_options;
+    logger_options.env = db_chroot_env_.get();
+    ASSERT_OK(CreateLoggerFromOptions(dbname_, logger_options, &logger_));
+  }
+
+  DB* OpenDB() {
+    DB* db;
+    EXPECT_OK(DB::Open(options_, dbname_, &db));
+    return db;
+  }
+
+  void CloseAndReopenDB(bool read_only = false) {
+    // Close DB
+    db_.reset();
+
+    // Open DB
+    test_db_fs_->SetLimitWrittenFiles(1000000);
+    DB* db;
+    if (read_only) {
+      ASSERT_OK(DB::OpenForReadOnly(options_, dbname_, &db));
+    } else {
+      ASSERT_OK(DB::Open(options_, dbname_, &db));
+    }
+    db_.reset(db);
+  }
+
+  void InitializeDBAndBackupEngine(bool dummy = false) {
+    // reset all the db env defaults
+    test_db_fs_->SetLimitWrittenFiles(1000000);
+    test_db_fs_->SetDummySequentialFile(dummy);
+
+    DB* db;
+    if (dummy) {
+      dummy_db_ = new DummyDB(options_, dbname_);
+      db = dummy_db_;
+    } else {
+      ASSERT_OK(DB::Open(options_, dbname_, &db));
+    }
+    db_.reset(db);
+  }
+
+  virtual void OpenDBAndBackupEngine(
+      bool destroy_old_data = false, bool dummy = false,
+      ShareOption shared_option = kShareNoChecksum) {
+    InitializeDBAndBackupEngine(dummy);
+    // reset backup env defaults
+    test_backup_fs_->SetLimitWrittenFiles(1000000);
+    engine_options_->destroy_old_data = destroy_old_data;
+    engine_options_->share_table_files = shared_option != kNoShare;
+    engine_options_->share_files_with_checksum =
+        shared_option == kShareWithChecksum;
+    OpenBackupEngine(destroy_old_data);
+  }
+
+  void CloseDBAndBackupEngine() {
+    db_.reset();
+    backup_engine_.reset();
+  }
+
+  void OpenBackupEngine(bool destroy_old_data = false) {
+    engine_options_->destroy_old_data = destroy_old_data;
+    engine_options_->info_log = logger_.get();
+    BackupEngine* backup_engine;
+    ASSERT_OK(BackupEngine::Open(test_db_env_.get(), *engine_options_,
+                                 &backup_engine));
+    backup_engine_.reset(backup_engine);
+  }
+
+  void CloseBackupEngine() { backup_engine_.reset(nullptr); }
+
+  // cross-cutting test of GetBackupInfo
+  void AssertBackupInfoConsistency() {
+    std::vector<BackupInfo> backup_info;
+    backup_engine_->GetBackupInfo(&backup_info, /*with file details*/ true);
+    std::map<std::string, uint64_t> file_sizes;
+
+    // Find the files that are supposed to be there
+    for (auto& backup : backup_info) {
+      uint64_t sum_for_backup = 0;
+      for (auto& file : backup.file_details) {
+        auto e = file_sizes.find(file.relative_filename);
+        if (e == file_sizes.end()) {
+          // fprintf(stderr, "Adding %s -> %u\n",
+          // file.relative_filename.c_str(), (unsigned)file.size);
+          file_sizes[file.relative_filename] = file.size;
+        } else {
+          ASSERT_EQ(file_sizes[file.relative_filename], file.size);
+        }
+        sum_for_backup += file.size;
+      }
+      ASSERT_EQ(backup.size, sum_for_backup);
+    }
+
+    std::vector<BackupID> corrupt_backup_ids;
+    backup_engine_->GetCorruptedBackups(&corrupt_backup_ids);
+    bool has_corrupt = corrupt_backup_ids.size() > 0;
+
+    // Compare with what's in backup dir
+    std::vector<std::string> child_dirs;
+    ASSERT_OK(
+        test_backup_env_->GetChildren(backupdir_ + "/private", &child_dirs));
+    for (auto& dir : child_dirs) {
+      dir = "private/" + dir;
+    }
+    child_dirs.push_back("shared");           // might not exist
+    child_dirs.push_back("shared_checksum");  // might not exist
+    for (auto& dir : child_dirs) {
+      std::vector<std::string> children;
+      test_backup_env_->GetChildren(backupdir_ + "/" + dir, &children)
+          .PermitUncheckedError();
+      // fprintf(stderr, "ls %s\n", (backupdir_ + "/" + dir).c_str());
+      for (auto& file : children) {
+        uint64_t size;
+        size = UINT64_MAX;  // appease clang-analyze
+        std::string rel_file = dir + "/" + file;
+        // fprintf(stderr, "stat %s\n", (backupdir_ + "/" + rel_file).c_str());
+        ASSERT_OK(
+            test_backup_env_->GetFileSize(backupdir_ + "/" + rel_file, &size));
+        auto e = file_sizes.find(rel_file);
+        if (e == file_sizes.end()) {
+          // The only case in which we should find files not reported
+          ASSERT_TRUE(has_corrupt);
+        } else {
+          ASSERT_EQ(e->second, size);
+          file_sizes.erase(e);
+        }
+      }
+    }
+
+    // Everything should have been matched
+    ASSERT_EQ(file_sizes.size(), 0);
+  }
+
+  // restores backup backup_id and asserts the existence of
+  // [start_exist, end_exist> and not-existence of
+  // [end_exist, end>
+  //
+  // if backup_id == 0, it means restore from latest
+  // if end == 0, don't check AssertEmpty
+  void AssertBackupConsistency(BackupID backup_id, uint32_t start_exist,
+                               uint32_t end_exist, uint32_t end = 0,
+                               bool keep_log_files = false) {
+    RestoreOptions restore_options(keep_log_files);
+    bool opened_backup_engine = false;
+    if (backup_engine_.get() == nullptr) {
+      opened_backup_engine = true;
+      OpenBackupEngine();
+    }
+    AssertBackupInfoConsistency();
+
+    // Now perform restore
+    if (backup_id > 0) {
+      ASSERT_OK(backup_engine_->RestoreDBFromBackup(backup_id, dbname_, dbname_,
+                                                    restore_options));
+    } else {
+      ASSERT_OK(backup_engine_->RestoreDBFromLatestBackup(dbname_, dbname_,
+                                                          restore_options));
+    }
+    DB* db = OpenDB();
+    // Check DB contents
+    AssertExists(db, start_exist, end_exist);
+    if (end != 0) {
+      AssertEmpty(db, end_exist, end);
+    }
+    delete db;
+    if (opened_backup_engine) {
+      CloseBackupEngine();
+    }
+  }
+
+  void DeleteLogFiles() {
+    std::vector<std::string> delete_logs;
+    ASSERT_OK(db_chroot_env_->GetChildren(dbname_, &delete_logs));
+    for (auto f : delete_logs) {
+      uint64_t number;
+      FileType type;
+      bool ok = ParseFileName(f, &number, &type);
+      if (ok && type == kWalFile) {
+        ASSERT_OK(db_chroot_env_->DeleteFile(dbname_ + "/" + f));
+      }
+    }
+  }
+
+  Status GetDataFilesInDB(const FileType& file_type,
+                          std::vector<FileAttributes>* files) {
+    std::vector<std::string> live;
+    uint64_t ignore_manifest_size;
+    Status s = db_->GetLiveFiles(live, &ignore_manifest_size, /*flush*/ false);
+    if (!s.ok()) {
+      return s;
+    }
+    std::vector<FileAttributes> children;
+    s = test_db_env_->GetChildrenFileAttributes(dbname_, &children);
+    for (const auto& child : children) {
+      FileType type;
+      uint64_t number = 0;
+      if (ParseFileName(child.name, &number, &type) && type == file_type &&
+          std::find(live.begin(), live.end(), "/" + child.name) != live.end()) {
+        files->push_back(child);
+      }
+    }
+    return s;
+  }
+
+  Status GetRandomDataFileInDB(const FileType& file_type,
+                               std::string* fname_out,
+                               uint64_t* fsize_out = nullptr) {
+    Random rnd(6);  // NB: hardly "random"
+    std::vector<FileAttributes> files;
+    Status s = GetDataFilesInDB(file_type, &files);
+    if (!s.ok()) {
+      return s;
+    }
+    if (files.empty()) {
+      return Status::NotFound("");
+    }
+    size_t i = rnd.Uniform(static_cast<int>(files.size()));
+    *fname_out = dbname_ + "/" + files[i].name;
+    if (fsize_out) {
+      *fsize_out = files[i].size_bytes;
+    }
+    return Status::OK();
+  }
+
+  Status CorruptRandomDataFileInDB(const FileType& file_type) {
+    std::string fname;
+    uint64_t fsize = 0;
+    Status s = GetRandomDataFileInDB(file_type, &fname, &fsize);
+    if (!s.ok()) {
+      return s;
+    }
+
+    std::string file_contents;
+    s = ReadFileToString(test_db_env_.get(), fname, &file_contents);
+    if (!s.ok()) {
+      return s;
+    }
+    s = test_db_env_->DeleteFile(fname);
+    if (!s.ok()) {
+      return s;
+    }
+
+    file_contents[0] = (file_contents[0] + 257) % 256;
+    return WriteStringToFile(test_db_env_.get(), file_contents, fname);
+  }
+
+  void AssertDirectoryFilesMatchRegex(const std::string& dir,
+                                      const TestRegex& pattern,
+                                      const std::string& file_type,
+                                      int minimum_count) {
+    std::vector<FileAttributes> children;
+    ASSERT_OK(file_manager_->GetChildrenFileAttributes(dir, &children));
+    int found_count = 0;
+    for (const auto& child : children) {
+      if (EndsWith(child.name, file_type)) {
+        ASSERT_MATCHES_REGEX(child.name, pattern);
+        ++found_count;
+      }
+    }
+    ASSERT_GE(found_count, minimum_count);
+  }
+
+  void AssertDirectoryFilesSizeIndicators(const std::string& dir,
+                                          int minimum_count) {
+    std::vector<FileAttributes> children;
+    ASSERT_OK(file_manager_->GetChildrenFileAttributes(dir, &children));
+    int found_count = 0;
+    for (const auto& child : children) {
+      auto last_underscore = child.name.find_last_of('_');
+      auto last_dot = child.name.find_last_of('.');
+      ASSERT_NE(child.name, child.name.substr(0, last_underscore));
+      ASSERT_NE(child.name, child.name.substr(0, last_dot));
+      ASSERT_LT(last_underscore, last_dot);
+      std::string s = child.name.substr(last_underscore + 1,
+                                        last_dot - (last_underscore + 1));
+      ASSERT_EQ(s, std::to_string(child.size_bytes));
+      ++found_count;
+    }
+    ASSERT_GE(found_count, minimum_count);
+  }
+
+  // files
+  std::string dbname_;
+  std::string backupdir_;
+  std::string latest_backup_;
+
+  // logger_ must be above backup_engine_ such that the engine's destructor,
+  // which uses a raw pointer to the logger, executes first.
+  std::shared_ptr<Logger> logger_;
+
+  // FileSystems
+  std::shared_ptr<FileSystem> db_chroot_fs_;
+  std::shared_ptr<FileSystem> backup_chroot_fs_;
+  std::shared_ptr<TestFs> test_db_fs_;
+  std::shared_ptr<TestFs> test_backup_fs_;
+
+  // Env wrappers
+  std::unique_ptr<Env> db_chroot_env_;
+  std::unique_ptr<Env> backup_chroot_env_;
+  std::unique_ptr<Env> test_db_env_;
+  std::unique_ptr<Env> test_backup_env_;
+  std::unique_ptr<FileManager> file_manager_;
+  std::unique_ptr<FileManager> db_file_manager_;
+
+  // all the dbs!
+  DummyDB* dummy_db_;  // owned as db_ when present
+  std::unique_ptr<DB> db_;
+  std::unique_ptr<BackupEngine> backup_engine_;
+
+  // options
+  Options options_;
+
+ protected:
+  void DestroyDBWithoutCheck(const std::string& dbname,
+                             const Options& options) {
+    // DestroyDB may fail because the db might not be existed for some tests
+    DestroyDB(dbname, options).PermitUncheckedError();
+  }
+
+  std::unique_ptr<BackupEngineOptions> engine_options_;
+};  // BackupEngineTest
+
+void AppendPath(const std::string& path, std::vector<std::string>& v) {
+  for (auto& f : v) {
+    f = path + f;
+  }
+}
+
+class BackupEngineTestWithParam : public BackupEngineTest,
+                                  public testing::WithParamInterface<bool> {
+ public:
+  BackupEngineTestWithParam() {
+    engine_options_->share_files_with_checksum = GetParam();
+  }
+  void OpenDBAndBackupEngine(
+      bool destroy_old_data = false, bool dummy = false,
+      ShareOption shared_option = kShareNoChecksum) override {
+    BackupEngineTest::InitializeDBAndBackupEngine(dummy);
+    // reset backup env defaults
+    test_backup_fs_->SetLimitWrittenFiles(1000000);
+    engine_options_->destroy_old_data = destroy_old_data;
+    engine_options_->share_table_files = shared_option != kNoShare;
+    // NOTE: keep share_files_with_checksum setting from constructor
+    OpenBackupEngine(destroy_old_data);
+  }
+};
+
+TEST_F(BackupEngineTest, FileCollision) {
+  const int keys_iteration = 100;
+  for (const auto& sopt : kAllShareOptions) {
+    OpenDBAndBackupEngine(true /* destroy_old_data */, false /* dummy */, sopt);
+    FillDB(db_.get(), 0, keys_iteration);
+    ASSERT_OK(backup_engine_->CreateNewBackup(db_.get()));
+    FillDB(db_.get(), keys_iteration, keys_iteration * 2);
+    ASSERT_OK(backup_engine_->CreateNewBackup(db_.get()));
+    CloseDBAndBackupEngine();
+
+    // If the db directory has been cleaned up, it is sensitive to file
+    // collision.
+    DestroyDBWithoutCheck(dbname_, options_);
+
+    // open fresh DB, but old backups present
+    OpenDBAndBackupEngine(false /* destroy_old_data */, false /* dummy */,
+                          sopt);
+    FillDB(db_.get(), 0, keys_iteration);
+    ASSERT_OK(db_->Flush(FlushOptions()));  // like backup would do
+    FillDB(db_.get(), keys_iteration, keys_iteration * 2);
+    if (sopt != kShareNoChecksum) {
+      ASSERT_OK(backup_engine_->CreateNewBackup(db_.get()));
+    } else {
+      // The new table files created in FillDB() will clash with the old
+      // backup and sharing tables with no checksum will have the file
+      // collision problem.
+      ASSERT_NOK(backup_engine_->CreateNewBackup(db_.get()));
+      ASSERT_OK(backup_engine_->PurgeOldBackups(0));
+      ASSERT_OK(backup_engine_->CreateNewBackup(db_.get()));
+    }
+    CloseDBAndBackupEngine();
+
+    // delete old data
+    DestroyDBWithoutCheck(dbname_, options_);
+  }
+}
+
+// This test verifies that the verifyBackup method correctly identifies
+// invalid backups
+TEST_P(BackupEngineTestWithParam, VerifyBackup) {
+  const int keys_iteration = 5000;
+  OpenDBAndBackupEngine(true);
+  // create five backups
+  for (int i = 0; i < 5; ++i) {
+    FillDB(db_.get(), keys_iteration * i, keys_iteration * (i + 1));
+    ASSERT_OK(backup_engine_->CreateNewBackup(db_.get(), true));
+  }
+  CloseDBAndBackupEngine();
+
+  OpenDBAndBackupEngine();
+  // ---------- case 1. - valid backup -----------
+  ASSERT_TRUE(backup_engine_->VerifyBackup(1).ok());
+
+  // ---------- case 2. - delete a file -----------i
+  ASSERT_OK(file_manager_->DeleteRandomFileInDir(backupdir_ + "/private/1"));
+  ASSERT_TRUE(backup_engine_->VerifyBackup(1).IsNotFound());
+
+  // ---------- case 3. - corrupt a file -----------
+  std::string append_data = "Corrupting a random file";
+  ASSERT_OK(file_manager_->AppendToRandomFileInDir(backupdir_ + "/private/2",
+                                                   append_data));
+  ASSERT_TRUE(backup_engine_->VerifyBackup(2).IsCorruption());
+
+  // ---------- case 4. - invalid backup -----------
+  ASSERT_TRUE(backup_engine_->VerifyBackup(6).IsNotFound());
+  CloseDBAndBackupEngine();
+}
+
+#if !defined(ROCKSDB_VALGRIND_RUN) || defined(ROCKSDB_FULL_VALGRIND_RUN)
+// open DB, write, close DB, backup, restore, repeat
+TEST_P(BackupEngineTestWithParam, OfflineIntegrationTest) {
+  // has to be a big number, so that it triggers the memtable flush
+  const int keys_iteration = 5000;
+  const int max_key = keys_iteration * 4 + 10;
+  // first iter -- flush before backup
+  // second iter -- don't flush before backup
+  for (int iter = 0; iter < 2; ++iter) {
+    // delete old data
+    DestroyDBWithoutCheck(dbname_, options_);
+    bool destroy_data = true;
+
+    // every iteration --
+    // 1. insert new data in the DB
+    // 2. backup the DB
+    // 3. destroy the db
+    // 4. restore the db, check everything is still there
+    for (int i = 0; i < 5; ++i) {
+      // in last iteration, put smaller amount of data,
+      int fill_up_to = std::min(keys_iteration * (i + 1), max_key);
+      // ---- insert new data and back up ----
+      OpenDBAndBackupEngine(destroy_data);
+      destroy_data = false;
+      // kAutoFlushOnly to preserve legacy test behavior (consider updating)
+      FillDB(db_.get(), keys_iteration * i, fill_up_to, kAutoFlushOnly);
+      ASSERT_OK(backup_engine_->CreateNewBackup(db_.get(), iter == 0))
+          << "iter: " << iter << ", idx: " << i;
+      CloseDBAndBackupEngine();
+      DestroyDBWithoutCheck(dbname_, options_);
+
+      // ---- make sure it's empty ----
+      DB* db = OpenDB();
+      AssertEmpty(db, 0, fill_up_to);
+      delete db;
+
+      // ---- restore the DB ----
+      OpenBackupEngine();
+      if (i >= 3) {  // test purge old backups
+        // when i == 4, purge to only 1 backup
+        // when i == 3, purge to 2 backups
+        ASSERT_OK(backup_engine_->PurgeOldBackups(5 - i));
+      }
+      // ---- make sure the data is there ---
+      AssertBackupConsistency(0, 0, fill_up_to, max_key);
+      CloseBackupEngine();
+    }
+  }
+}
+
+// open DB, write, backup, write, backup, close, restore
+TEST_P(BackupEngineTestWithParam, OnlineIntegrationTest) {
+  // has to be a big number, so that it triggers the memtable flush
+  const int keys_iteration = 5000;
+  const int max_key = keys_iteration * 4 + 10;
+  Random rnd(7);
+  // delete old data
+  DestroyDBWithoutCheck(dbname_, options_);
+
+  // TODO: Implement & test db_paths support in backup (not supported in
+  // restore)
+  // options_.db_paths.emplace_back(dbname_, 500 * 1024);
+  // options_.db_paths.emplace_back(dbname_ + "_2", 1024 * 1024 * 1024);
+
+  OpenDBAndBackupEngine(true);
+  // write some data, backup, repeat
+  for (int i = 0; i < 5; ++i) {
+    if (i == 4) {
+      // delete backup number 2, online delete!
+      ASSERT_OK(backup_engine_->DeleteBackup(2));
+    }
+    // in last iteration, put smaller amount of data,
+    // so that backups can share sst files
+    int fill_up_to = std::min(keys_iteration * (i + 1), max_key);
+    // kAutoFlushOnly to preserve legacy test behavior (consider updating)
+    FillDB(db_.get(), keys_iteration * i, fill_up_to, kAutoFlushOnly);
+    // we should get consistent results with flush_before_backup
+    // set to both true and false
+    ASSERT_OK(backup_engine_->CreateNewBackup(db_.get(), !!(rnd.Next() % 2)));
+  }
+  // close and destroy
+  CloseDBAndBackupEngine();
+  DestroyDBWithoutCheck(dbname_, options_);
+
+  // ---- make sure it's empty ----
+  DB* db = OpenDB();
+  AssertEmpty(db, 0, max_key);
+  delete db;
+
+  // ---- restore every backup and verify all the data is there ----
+  OpenBackupEngine();
+  for (int i = 1; i <= 5; ++i) {
+    if (i == 2) {
+      // we deleted backup 2
+      Status s = backup_engine_->RestoreDBFromBackup(2, dbname_, dbname_);
+      ASSERT_TRUE(!s.ok());
+    } else {
+      int fill_up_to = std::min(keys_iteration * i, max_key);
+      AssertBackupConsistency(i, 0, fill_up_to, max_key);
+    }
+  }
+
+  // delete some backups -- this should leave only backups 3 and 5 alive
+  ASSERT_OK(backup_engine_->DeleteBackup(4));
+  ASSERT_OK(backup_engine_->PurgeOldBackups(2));
+
+  std::vector<BackupInfo> backup_info;
+  backup_engine_->GetBackupInfo(&backup_info);
+  ASSERT_EQ(2UL, backup_info.size());
+
+  // check backup 3
+  AssertBackupConsistency(3, 0, 3 * keys_iteration, max_key);
+  // check backup 5
+  AssertBackupConsistency(5, 0, max_key);
+
+  CloseBackupEngine();
+}
+#endif  // !defined(ROCKSDB_VALGRIND_RUN) || defined(ROCKSDB_FULL_VALGRIND_RUN)
+
+INSTANTIATE_TEST_CASE_P(BackupEngineTestWithParam, BackupEngineTestWithParam,
+                        ::testing::Bool());
+
+// this will make sure that backup does not copy the same file twice
+TEST_F(BackupEngineTest, NoDoubleCopy_And_AutoGC) {
+  OpenDBAndBackupEngine(true, true);
+
+  // should write 5 DB files + one meta file
+  test_backup_fs_->SetLimitWrittenFiles(7);
+  test_backup_fs_->ClearWrittenFiles();
+  test_db_fs_->SetLimitWrittenFiles(0);
+  dummy_db_->live_files_ = {"00010.sst", "00011.sst", "CURRENT", "MANIFEST-01",
+                            "00011.log"};
+  test_db_fs_->SetFilenamesForMockedAttrs(dummy_db_->live_files_);
+  ASSERT_OK(backup_engine_->CreateNewBackup(db_.get(), false));
+  std::vector<std::string> should_have_written = {
+      "/shared/.00010.sst.tmp", "/shared/.00011.sst.tmp", "/private/1/CURRENT",
+      "/private/1/MANIFEST-01", "/private/1/00011.log",   "/meta/.1.tmp"};
+  AppendPath(backupdir_, should_have_written);
+  test_backup_fs_->AssertWrittenFiles(should_have_written);
+
+  char db_number = '1';
+
+  for (std::string other_sst : {"00015.sst", "00017.sst", "00019.sst"}) {
+    // should write 4 new DB files + one meta file
+    // should not write/copy 00010.sst, since it's already there!
+    test_backup_fs_->SetLimitWrittenFiles(6);
+    test_backup_fs_->ClearWrittenFiles();
+
+    dummy_db_->live_files_ = {"00010.sst", other_sst, "CURRENT", "MANIFEST-01",
+                              "00011.log"};
+    test_db_fs_->SetFilenamesForMockedAttrs(dummy_db_->live_files_);
+    ASSERT_OK(backup_engine_->CreateNewBackup(db_.get(), false));
+    // should not open 00010.sst - it's already there
+
+    ++db_number;
+    std::string private_dir = std::string("/private/") + db_number;
+    should_have_written = {
+        "/shared/." + other_sst + ".tmp", private_dir + "/CURRENT",
+        private_dir + "/MANIFEST-01", private_dir + "/00011.log",
+        std::string("/meta/.") + db_number + ".tmp"};
+    AppendPath(backupdir_, should_have_written);
+    test_backup_fs_->AssertWrittenFiles(should_have_written);
+  }
+
+  ASSERT_OK(backup_engine_->DeleteBackup(1));
+  ASSERT_OK(test_backup_env_->FileExists(backupdir_ + "/shared/00010.sst"));
+
+  // 00011.sst was only in backup 1, should be deleted
+  ASSERT_EQ(Status::NotFound(),
+            test_backup_env_->FileExists(backupdir_ + "/shared/00011.sst"));
+  ASSERT_OK(test_backup_env_->FileExists(backupdir_ + "/shared/00015.sst"));
+
+  // MANIFEST file size should be only 100
+  uint64_t size = 0;
+  ASSERT_OK(test_backup_env_->GetFileSize(backupdir_ + "/private/2/MANIFEST-01",
+                                          &size));
+  ASSERT_EQ(100UL, size);
+  ASSERT_OK(
+      test_backup_env_->GetFileSize(backupdir_ + "/shared/00015.sst", &size));
+  ASSERT_EQ(200UL, size);
+
+  CloseBackupEngine();
+
+  //
+  // Now simulate incomplete delete by removing just meta
+  //
+  ASSERT_OK(test_backup_env_->DeleteFile(backupdir_ + "/meta/2"));
+
+  OpenBackupEngine();
+
+  // 1 appears to be removed, so
+  // 2 non-corrupt and 0 corrupt seen
+  std::vector<BackupInfo> backup_info;
+  std::vector<BackupID> corrupt_backup_ids;
+  backup_engine_->GetBackupInfo(&backup_info);
+  backup_engine_->GetCorruptedBackups(&corrupt_backup_ids);
+  ASSERT_EQ(2UL, backup_info.size());
+  ASSERT_EQ(0UL, corrupt_backup_ids.size());
+
+  // Keep the two we see, but this should suffice to purge unreferenced
+  // shared files from incomplete delete.
+  ASSERT_OK(backup_engine_->PurgeOldBackups(2));
+
+  // Make sure dangling sst file has been removed (somewhere along this
+  // process). GarbageCollect should not be needed.
+  ASSERT_EQ(Status::NotFound(),
+            test_backup_env_->FileExists(backupdir_ + "/shared/00015.sst"));
+  ASSERT_OK(test_backup_env_->FileExists(backupdir_ + "/shared/00017.sst"));
+  ASSERT_OK(test_backup_env_->FileExists(backupdir_ + "/shared/00019.sst"));
+
+  // Now actually purge a good one
+  ASSERT_OK(backup_engine_->PurgeOldBackups(1));
+
+  ASSERT_EQ(Status::NotFound(),
+            test_backup_env_->FileExists(backupdir_ + "/shared/00017.sst"));
+  ASSERT_OK(test_backup_env_->FileExists(backupdir_ + "/shared/00019.sst"));
+
+  CloseDBAndBackupEngine();
+}
+
+// test various kind of corruptions that may happen:
+// 1. Not able to write a file for backup - that backup should fail,
+//      everything else should work
+// 2. Corrupted backup meta file or missing backuped file - we should
+//      not be able to open that backup, but all other backups should be
+//      fine
+// 3. Corrupted checksum value - if the checksum is not a valid uint32_t,
+//      db open should fail, otherwise, it aborts during the restore process.
+TEST_F(BackupEngineTest, CorruptionsTest) {
+  const int keys_iteration = 5000;
+  Random rnd(6);
+  Status s;
+
+  OpenDBAndBackupEngine(true);
+  // create five backups
+  for (int i = 0; i < 5; ++i) {
+    FillDB(db_.get(), keys_iteration * i, keys_iteration * (i + 1));
+    ASSERT_OK(backup_engine_->CreateNewBackup(db_.get(), !!(rnd.Next() % 2)));
+  }
+
+  // ---------- case 1. - fail a write -----------
+  // try creating backup 6, but fail a write
+  FillDB(db_.get(), keys_iteration * 5, keys_iteration * 6);
+  test_backup_fs_->SetLimitWrittenFiles(2);
+  // should fail
+  s = backup_engine_->CreateNewBackup(db_.get(), !!(rnd.Next() % 2));
+  ASSERT_NOK(s);
+  test_backup_fs_->SetLimitWrittenFiles(1000000);
+  // latest backup should have all the keys
+  CloseDBAndBackupEngine();
+  AssertBackupConsistency(0, 0, keys_iteration * 5, keys_iteration * 6);
+
+  // --------- case 2. corrupted backup meta or missing backuped file ----
+  ASSERT_OK(file_manager_->CorruptFile(backupdir_ + "/meta/5", 3));
+  // since 5 meta is now corrupted, latest backup should be 4
+  AssertBackupConsistency(0, 0, keys_iteration * 4, keys_iteration * 5);
+  OpenBackupEngine();
+  s = backup_engine_->RestoreDBFromBackup(5, dbname_, dbname_);
+  ASSERT_NOK(s);
+  CloseBackupEngine();
+  ASSERT_OK(file_manager_->DeleteRandomFileInDir(backupdir_ + "/private/4"));
+  // 4 is corrupted, 3 is the latest backup now
+  AssertBackupConsistency(0, 0, keys_iteration * 3, keys_iteration * 5);
+  OpenBackupEngine();
+  s = backup_engine_->RestoreDBFromBackup(4, dbname_, dbname_);
+  CloseBackupEngine();
+  ASSERT_NOK(s);
+
+  // --------- case 3. corrupted checksum value ----
+  ASSERT_OK(file_manager_->CorruptChecksum(backupdir_ + "/meta/3", false));
+  // checksum of backup 3 is an invalid value, this can be detected at
+  // db open time, and it reverts to the previous backup automatically
+  AssertBackupConsistency(0, 0, keys_iteration * 2, keys_iteration * 5);
+  // checksum of the backup 2 appears to be valid, this can cause checksum
+  // mismatch and abort restore process
+  ASSERT_OK(file_manager_->CorruptChecksum(backupdir_ + "/meta/2", true));
+  ASSERT_OK(file_manager_->FileExists(backupdir_ + "/meta/2"));
+  OpenBackupEngine();
+  ASSERT_OK(file_manager_->FileExists(backupdir_ + "/meta/2"));
+  s = backup_engine_->RestoreDBFromBackup(2, dbname_, dbname_);
+  ASSERT_NOK(s);
+
+  // make sure that no corrupt backups have actually been deleted!
+  ASSERT_OK(file_manager_->FileExists(backupdir_ + "/meta/1"));
+  ASSERT_OK(file_manager_->FileExists(backupdir_ + "/meta/2"));
+  ASSERT_OK(file_manager_->FileExists(backupdir_ + "/meta/3"));
+  ASSERT_OK(file_manager_->FileExists(backupdir_ + "/meta/4"));
+  ASSERT_OK(file_manager_->FileExists(backupdir_ + "/meta/5"));
+  ASSERT_OK(file_manager_->FileExists(backupdir_ + "/private/1"));
+  ASSERT_OK(file_manager_->FileExists(backupdir_ + "/private/2"));
+  ASSERT_OK(file_manager_->FileExists(backupdir_ + "/private/3"));
+  ASSERT_OK(file_manager_->FileExists(backupdir_ + "/private/4"));
+  ASSERT_OK(file_manager_->FileExists(backupdir_ + "/private/5"));
+
+  // delete the corrupt backups and then make sure they're actually deleted
+  ASSERT_OK(backup_engine_->DeleteBackup(5));
+  ASSERT_OK(backup_engine_->DeleteBackup(4));
+  ASSERT_OK(backup_engine_->DeleteBackup(3));
+  ASSERT_OK(backup_engine_->DeleteBackup(2));
+  // Should not be needed anymore with auto-GC on DeleteBackup
+  //(void)backup_engine_->GarbageCollect();
+  ASSERT_EQ(Status::NotFound(),
+            file_manager_->FileExists(backupdir_ + "/meta/5"));
+  ASSERT_EQ(Status::NotFound(),
+            file_manager_->FileExists(backupdir_ + "/private/5"));
+  ASSERT_EQ(Status::NotFound(),
+            file_manager_->FileExists(backupdir_ + "/meta/4"));
+  ASSERT_EQ(Status::NotFound(),
+            file_manager_->FileExists(backupdir_ + "/private/4"));
+  ASSERT_EQ(Status::NotFound(),
+            file_manager_->FileExists(backupdir_ + "/meta/3"));
+  ASSERT_EQ(Status::NotFound(),
+            file_manager_->FileExists(backupdir_ + "/private/3"));
+  ASSERT_EQ(Status::NotFound(),
+            file_manager_->FileExists(backupdir_ + "/meta/2"));
+  ASSERT_EQ(Status::NotFound(),
+            file_manager_->FileExists(backupdir_ + "/private/2"));
+  CloseBackupEngine();
+  AssertBackupConsistency(0, 0, keys_iteration * 1, keys_iteration * 5);
+
+  // new backup should be 2!
+  OpenDBAndBackupEngine();
+  FillDB(db_.get(), keys_iteration * 1, keys_iteration * 2);
+  ASSERT_OK(backup_engine_->CreateNewBackup(db_.get(), !!(rnd.Next() % 2)));
+  CloseDBAndBackupEngine();
+  AssertBackupConsistency(2, 0, keys_iteration * 2, keys_iteration * 5);
+}
+
+// Corrupt a file but maintain its size
+TEST_F(BackupEngineTest, CorruptFileMaintainSize) {
+  const int keys_iteration = 5000;
+  OpenDBAndBackupEngine(true);
+  // create a backup
+  FillDB(db_.get(), 0, keys_iteration);
+  ASSERT_OK(backup_engine_->CreateNewBackup(db_.get(), true));
+  CloseDBAndBackupEngine();
+
+  OpenDBAndBackupEngine();
+  // verify with file size
+  ASSERT_OK(backup_engine_->VerifyBackup(1, false));
+  // verify with file checksum
+  ASSERT_OK(backup_engine_->VerifyBackup(1, true));
+
+  std::string file_to_corrupt;
+  uint64_t file_size = 0;
+  // under normal circumstance, there should be at least one nonempty file
+  while (file_size == 0) {
+    // get a random file in /private/1
+    assert(file_manager_
+               ->GetRandomFileInDir(backupdir_ + "/private/1", &file_to_corrupt,
+                                    &file_size)
+               .ok());
+    // corrupt the file by replacing its content by file_size random bytes
+    ASSERT_OK(file_manager_->CorruptFile(file_to_corrupt, file_size));
+  }
+  // file sizes match
+  ASSERT_OK(backup_engine_->VerifyBackup(1, false));
+  // file checksums mismatch
+  ASSERT_NOK(backup_engine_->VerifyBackup(1, true));
+  // sanity check, use default second argument
+  ASSERT_OK(backup_engine_->VerifyBackup(1));
+  CloseDBAndBackupEngine();
+
+  // an extra challenge
+  // set share_files_with_checksum to true and do two more backups
+  // corrupt all the table files in shared_checksum but maintain their sizes
+  OpenDBAndBackupEngine(true /* destroy_old_data */, false /* dummy */,
+                        kShareWithChecksum);
+  // creat two backups
+  for (int i = 1; i < 3; ++i) {
+    FillDB(db_.get(), keys_iteration * i, keys_iteration * (i + 1));
+    ASSERT_OK(backup_engine_->CreateNewBackup(db_.get(), true));
+  }
+  CloseDBAndBackupEngine();
+
+  OpenDBAndBackupEngine();
+  std::vector<FileAttributes> children;
+  const std::string dir = backupdir_ + "/shared_checksum";
+  ASSERT_OK(file_manager_->GetChildrenFileAttributes(dir, &children));
+  for (const auto& child : children) {
+    if (child.size_bytes == 0) {
+      continue;
+    }
+    // corrupt the file by replacing its content by file_size random bytes
+    ASSERT_OK(
+        file_manager_->CorruptFile(dir + "/" + child.name, child.size_bytes));
+  }
+  // file sizes match
+  ASSERT_OK(backup_engine_->VerifyBackup(1, false));
+  ASSERT_OK(backup_engine_->VerifyBackup(2, false));
+  // file checksums mismatch
+  ASSERT_NOK(backup_engine_->VerifyBackup(1, true));
+  ASSERT_NOK(backup_engine_->VerifyBackup(2, true));
+  CloseDBAndBackupEngine();
+}
+
+// Corrupt a blob file but maintain its size
+TEST_P(BackupEngineTestWithParam, CorruptBlobFileMaintainSize) {
+  const int keys_iteration = 5000;
+  OpenDBAndBackupEngine(true);
+  // create a backup
+  FillDB(db_.get(), 0, keys_iteration);
+  ASSERT_OK(backup_engine_->CreateNewBackup(db_.get(), true));
+  CloseDBAndBackupEngine();
+
+  OpenDBAndBackupEngine();
+  // verify with file size
+  ASSERT_OK(backup_engine_->VerifyBackup(1, false));
+  // verify with file checksum
+  ASSERT_OK(backup_engine_->VerifyBackup(1, true));
+
+  std::string file_to_corrupt;
+  std::vector<FileAttributes> children;
+
+  std::string dir = backupdir_;
+  if (engine_options_->share_files_with_checksum) {
+    dir += "/shared_checksum";
+  } else {
+    dir += "/shared";
+  }
+
+  ASSERT_OK(file_manager_->GetChildrenFileAttributes(dir, &children));
+
+  for (const auto& child : children) {
+    if (EndsWith(child.name, ".blob") && child.size_bytes != 0) {
+      // corrupt the blob files by replacing its content by file_size random
+      // bytes
+      ASSERT_OK(
+          file_manager_->CorruptFile(dir + "/" + child.name, child.size_bytes));
+    }
+  }
+
+  // file sizes match
+  ASSERT_OK(backup_engine_->VerifyBackup(1, false));
+  // file checksums mismatch
+  ASSERT_NOK(backup_engine_->VerifyBackup(1, true));
+  // sanity check, use default second argument
+  ASSERT_OK(backup_engine_->VerifyBackup(1));
+  CloseDBAndBackupEngine();
+}
+
+// Test if BackupEngine will fail to create new backup if some table has been
+// corrupted and the table file checksum is stored in the DB manifest
+TEST_F(BackupEngineTest, TableFileCorruptedBeforeBackup) {
+  const int keys_iteration = 50000;
+
+  OpenDBAndBackupEngine(true /* destroy_old_data */, false /* dummy */,
+                        kNoShare);
+  FillDB(db_.get(), 0, keys_iteration);
+  CloseAndReopenDB(/*read_only*/ true);
+  // corrupt a random table file in the DB directory
+  ASSERT_OK(CorruptRandomDataFileInDB(kTableFile));
+  // file_checksum_gen_factory is null, and thus table checksum is not
+  // verified for creating a new backup; no correction is detected
+  ASSERT_OK(backup_engine_->CreateNewBackup(db_.get()));
+  CloseDBAndBackupEngine();
+
+  // delete old files in db
+  DestroyDBWithoutCheck(dbname_, options_);
+
+  // Enable table file checksum in DB manifest
+  options_.file_checksum_gen_factory = GetFileChecksumGenCrc32cFactory();
+  OpenDBAndBackupEngine(true /* destroy_old_data */, false /* dummy */,
+                        kNoShare);
+  FillDB(db_.get(), 0, keys_iteration);
+  CloseAndReopenDB(/*read_only*/ true);
+  // corrupt a random table file in the DB directory
+  ASSERT_OK(CorruptRandomDataFileInDB(kTableFile));
+  // table file checksum is enabled so we should be able to detect any
+  // corruption
+  ASSERT_NOK(backup_engine_->CreateNewBackup(db_.get()));
+  CloseDBAndBackupEngine();
+}
+
+// Test if BackupEngine will fail to create new backup if some blob files has
+// been corrupted and the blob file checksum is stored in the DB manifest
+TEST_F(BackupEngineTest, BlobFileCorruptedBeforeBackup) {
+  const int keys_iteration = 50000;
+
+  OpenDBAndBackupEngine(true /* destroy_old_data */, false /* dummy */,
+                        kNoShare);
+  FillDB(db_.get(), 0, keys_iteration);
+  CloseAndReopenDB(/*read_only*/ true);
+  // corrupt a random blob file in the DB directory
+  ASSERT_OK(CorruptRandomDataFileInDB(kBlobFile));
+  // file_checksum_gen_factory is null, and thus blob checksum is not
+  // verified for creating a new backup; no correction is detected
+  ASSERT_OK(backup_engine_->CreateNewBackup(db_.get()));
+  CloseDBAndBackupEngine();
+
+  // delete old files in db
+  DestroyDBWithoutCheck(dbname_, options_);
+
+  // Enable file checksum in DB manifest
+  options_.file_checksum_gen_factory = GetFileChecksumGenCrc32cFactory();
+  OpenDBAndBackupEngine(true /* destroy_old_data */, false /* dummy */,
+                        kNoShare);
+  FillDB(db_.get(), 0, keys_iteration);
+  CloseAndReopenDB(/*read_only*/ true);
+  // corrupt a random blob file in the DB directory
+  ASSERT_OK(CorruptRandomDataFileInDB(kBlobFile));
+
+  // file checksum is enabled so we should be able to detect any
+  // corruption
+  ASSERT_NOK(backup_engine_->CreateNewBackup(db_.get()));
+  CloseDBAndBackupEngine();
+}
+
+#if !defined(ROCKSDB_VALGRIND_RUN) || defined(ROCKSDB_FULL_VALGRIND_RUN)
+// Test if BackupEngine will fail to create new backup if some table has been
+// corrupted and the table file checksum is stored in the DB manifest for the
+// case when backup table files will be stored in a shared directory
+TEST_P(BackupEngineTestWithParam, TableFileCorruptedBeforeBackup) {
+  const int keys_iteration = 50000;
+
+  OpenDBAndBackupEngine(true /* destroy_old_data */);
+  FillDB(db_.get(), 0, keys_iteration);
+  CloseAndReopenDB(/*read_only*/ true);
+  // corrupt a random table file in the DB directory
+  ASSERT_OK(CorruptRandomDataFileInDB(kTableFile));
+  // cannot detect corruption since DB manifest has no table checksums
+  ASSERT_OK(backup_engine_->CreateNewBackup(db_.get()));
+  CloseDBAndBackupEngine();
+
+  // delete old files in db
+  DestroyDBWithoutCheck(dbname_, options_);
+
+  // Enable table checksums in DB manifest
+  options_.file_checksum_gen_factory = GetFileChecksumGenCrc32cFactory();
+  OpenDBAndBackupEngine(true /* destroy_old_data */);
+  FillDB(db_.get(), 0, keys_iteration);
+  CloseAndReopenDB(/*read_only*/ true);
+  // corrupt a random table file in the DB directory
+  ASSERT_OK(CorruptRandomDataFileInDB(kTableFile));
+  // corruption is detected
+  ASSERT_NOK(backup_engine_->CreateNewBackup(db_.get()));
+  CloseDBAndBackupEngine();
+}
+
+// Test if BackupEngine will fail to create new backup if some blob files have
+// been corrupted and the blob file checksum is stored in the DB manifest for
+// the case when backup blob files will be stored in a shared directory
+TEST_P(BackupEngineTestWithParam, BlobFileCorruptedBeforeBackup) {
+  const int keys_iteration = 50000;
+  OpenDBAndBackupEngine(true /* destroy_old_data */);
+  FillDB(db_.get(), 0, keys_iteration);
+  CloseAndReopenDB(/*read_only*/ true);
+  // corrupt a random blob file in the DB directory
+  ASSERT_OK(CorruptRandomDataFileInDB(kBlobFile));
+  // cannot detect corruption since DB manifest has no blob file checksums
+  ASSERT_OK(backup_engine_->CreateNewBackup(db_.get()));
+  CloseDBAndBackupEngine();
+
+  // delete old files in db
+  DestroyDBWithoutCheck(dbname_, options_);
+
+  // Enable blob file checksums in DB manifest
+  options_.file_checksum_gen_factory = GetFileChecksumGenCrc32cFactory();
+  OpenDBAndBackupEngine(true /* destroy_old_data */);
+  FillDB(db_.get(), 0, keys_iteration);
+  CloseAndReopenDB(/*read_only*/ true);
+  // corrupt a random blob file in the DB directory
+  ASSERT_OK(CorruptRandomDataFileInDB(kBlobFile));
+  // corruption is detected
+  ASSERT_NOK(backup_engine_->CreateNewBackup(db_.get()));
+  CloseDBAndBackupEngine();
+}
+#endif  // !defined(ROCKSDB_VALGRIND_RUN) || defined(ROCKSDB_FULL_VALGRIND_RUN)
+
+TEST_F(BackupEngineTest, TableFileWithoutDbChecksumCorruptedDuringBackup) {
+  const int keys_iteration = 50000;
+  engine_options_->share_files_with_checksum_naming = kLegacyCrc32cAndFileSize;
+  // When share_files_with_checksum is on, we calculate checksums of table
+  // files before and after copying. So we can test whether a corruption has
+  // happened during the file is copied to backup directory.
+  OpenDBAndBackupEngine(true /* destroy_old_data */, false /* dummy */,
+                        kShareWithChecksum);
+
+  FillDB(db_.get(), 0, keys_iteration);
+  std::atomic<bool> corrupted{false};
+  // corrupt files when copying to the backup directory
+  SyncPoint::GetInstance()->SetCallBack(
+      "BackupEngineImpl::CopyOrCreateFile:CorruptionDuringBackup",
+      [&](void* data) {
+        if (data != nullptr) {
+          Slice* d = reinterpret_cast<Slice*>(data);
+          if (!d->empty()) {
+            d->remove_suffix(1);
+            corrupted = true;
+          }
+        }
+      });
+  SyncPoint::GetInstance()->EnableProcessing();
+  Status s = backup_engine_->CreateNewBackup(db_.get());
+  if (corrupted) {
+    ASSERT_NOK(s);
+  } else {
+    // should not in this path in normal cases
+    ASSERT_OK(s);
+  }
+
+  SyncPoint::GetInstance()->DisableProcessing();
+  SyncPoint::GetInstance()->ClearAllCallBacks();
+
+  CloseDBAndBackupEngine();
+  // delete old files in db
+  DestroyDBWithoutCheck(dbname_, options_);
+}
+
+TEST_F(BackupEngineTest, TableFileWithDbChecksumCorruptedDuringBackup) {
+  const int keys_iteration = 50000;
+  options_.file_checksum_gen_factory = GetFileChecksumGenCrc32cFactory();
+  for (auto& sopt : kAllShareOptions) {
+    // Since the default DB table file checksum is on, we obtain checksums of
+    // table files from the DB manifest before copying and verify it with the
+    // one calculated during copying.
+    // Therefore, we can test whether a corruption has happened during the file
+    // being copied to backup directory.
+    OpenDBAndBackupEngine(true /* destroy_old_data */, false /* dummy */, sopt);
+
+    FillDB(db_.get(), 0, keys_iteration);
+
+    // corrupt files when copying to the backup directory
+    SyncPoint::GetInstance()->SetCallBack(
+        "BackupEngineImpl::CopyOrCreateFile:CorruptionDuringBackup",
+        [&](void* data) {
+          if (data != nullptr) {
+            Slice* d = reinterpret_cast<Slice*>(data);
+            if (!d->empty()) {
+              d->remove_suffix(1);
+            }
+          }
+        });
+    SyncPoint::GetInstance()->EnableProcessing();
+    // The only case that we can't detect a corruption is when the file
+    // being backed up is empty. But as keys_iteration is large, such
+    // a case shouldn't have happened and we should be able to detect
+    // the corruption.
+    ASSERT_NOK(backup_engine_->CreateNewBackup(db_.get()));
+
+    SyncPoint::GetInstance()->DisableProcessing();
+    SyncPoint::GetInstance()->ClearAllCallBacks();
+
+    CloseDBAndBackupEngine();
+    // delete old files in db
+    DestroyDBWithoutCheck(dbname_, options_);
+  }
+}
+
+TEST_F(BackupEngineTest, InterruptCreationTest) {
+  // Interrupt backup creation by failing new writes and failing cleanup of the
+  // partial state. Then verify a subsequent backup can still succeed.
+  const int keys_iteration = 5000;
+  Random rnd(6);
+
+  OpenDBAndBackupEngine(true /* destroy_old_data */);
+  FillDB(db_.get(), 0, keys_iteration);
+  test_backup_fs_->SetLimitWrittenFiles(2);
+  test_backup_fs_->SetDeleteFileFailure(true);
+  // should fail creation
+  ASSERT_NOK(backup_engine_->CreateNewBackup(db_.get(), !!(rnd.Next() % 2)));
+  CloseDBAndBackupEngine();
+  // should also fail cleanup so the tmp directory stays behind
+  ASSERT_OK(backup_chroot_env_->FileExists(backupdir_ + "/private/1/"));
+
+  OpenDBAndBackupEngine(false /* destroy_old_data */);
+  test_backup_fs_->SetLimitWrittenFiles(1000000);
+  test_backup_fs_->SetDeleteFileFailure(false);
+  ASSERT_OK(backup_engine_->CreateNewBackup(db_.get(), !!(rnd.Next() % 2)));
+  // latest backup should have all the keys
+  CloseDBAndBackupEngine();
+  AssertBackupConsistency(0, 0, keys_iteration);
+}
+
+TEST_F(BackupEngineTest, FlushCompactDuringBackupCheckpoint) {
+  const int keys_iteration = 5000;
+  options_.file_checksum_gen_factory = GetFileChecksumGenCrc32cFactory();
+  for (const auto& sopt : kAllShareOptions) {
+    OpenDBAndBackupEngine(true /* destroy_old_data */, false /* dummy */, sopt);
+    FillDB(db_.get(), 0, keys_iteration);
+    // That FillDB leaves a mix of flushed and unflushed data
+    SyncPoint::GetInstance()->LoadDependency(
+        {{"CheckpointImpl::CreateCustomCheckpoint:AfterGetLive1",
+          "BackupEngineTest::FlushCompactDuringBackupCheckpoint:Before"},
+         {"BackupEngineTest::FlushCompactDuringBackupCheckpoint:After",
+          "CheckpointImpl::CreateCustomCheckpoint:AfterGetLive2"}});
+    SyncPoint::GetInstance()->EnableProcessing();
+    ROCKSDB_NAMESPACE::port::Thread flush_thread{[this]() {
+      TEST_SYNC_POINT(
+          "BackupEngineTest::FlushCompactDuringBackupCheckpoint:Before");
+      FillDB(db_.get(), keys_iteration, 2 * keys_iteration);
+      ASSERT_OK(db_->Flush(FlushOptions()));
+      DBImpl* dbi = static_cast<DBImpl*>(db_.get());
+      ASSERT_OK(dbi->TEST_WaitForFlushMemTable());
+      ASSERT_OK(db_->CompactRange(CompactRangeOptions(), nullptr, nullptr));
+      ASSERT_OK(dbi->TEST_WaitForCompact());
+      TEST_SYNC_POINT(
+          "BackupEngineTest::FlushCompactDuringBackupCheckpoint:After");
+    }};
+    ASSERT_OK(backup_engine_->CreateNewBackup(db_.get()));
+    flush_thread.join();
+    CloseDBAndBackupEngine();
+    SyncPoint::GetInstance()->DisableProcessing();
+    SyncPoint::GetInstance()->ClearAllCallBacks();
+    /* FIXME(peterd): reinstate with option for checksum in file names
+    if (sopt == kShareWithChecksum) {
+      // Ensure we actually got DB manifest checksums by inspecting
+      // shared_checksum file names for hex checksum component
+      TestRegex expected("[^_]+_[0-9A-F]{8}_[^_]+.sst");
+      std::vector<FileAttributes> children;
+      const std::string dir = backupdir_ + "/shared_checksum";
+      ASSERT_OK(file_manager_->GetChildrenFileAttributes(dir, &children));
+      for (const auto& child : children) {
+        if (child.size_bytes == 0) {
+          continue;
+        }
+        EXPECT_MATCHES_REGEX(child.name, expected);
+      }
+    }
+    */
+    AssertBackupConsistency(0, 0, keys_iteration);
+  }
+}
+
+inline std::string OptionsPath(std::string ret, int backupID) {
+  ret += "/private/";
+  ret += std::to_string(backupID);
+  ret += "/";
+  return ret;
+}
+
+// Backup the LATEST options file to
+// "<backup_dir>/private/<backup_id>/OPTIONS<number>"
+
+TEST_F(BackupEngineTest, BackupOptions) {
+  OpenDBAndBackupEngine(true);
+  for (int i = 1; i < 5; i++) {
+    std::string name;
+    std::vector<std::string> filenames;
+    // Must reset() before reset(OpenDB()) again.
+    // Calling OpenDB() while *db_ is existing will cause LOCK issue
+    db_.reset();
+    db_.reset(OpenDB());
+    ASSERT_OK(backup_engine_->CreateNewBackup(db_.get(), true));
+    ASSERT_OK(ROCKSDB_NAMESPACE::GetLatestOptionsFileName(db_->GetName(),
+                                                          options_.env, &name));
+    ASSERT_OK(file_manager_->FileExists(OptionsPath(backupdir_, i) + name));
+    ASSERT_OK(backup_chroot_env_->GetChildren(OptionsPath(backupdir_, i),
+                                              &filenames));
+    for (auto fn : filenames) {
+      if (fn.compare(0, 7, "OPTIONS") == 0) {
+        ASSERT_EQ(name, fn);
+      }
+    }
+  }
+
+  CloseDBAndBackupEngine();
+}
+
+TEST_F(BackupEngineTest, SetOptionsBackupRaceCondition) {
+  OpenDBAndBackupEngine(true);
+  SyncPoint::GetInstance()->LoadDependency(
+      {{"CheckpointImpl::CreateCheckpoint:SavedLiveFiles1",
+        "BackupEngineTest::SetOptionsBackupRaceCondition:BeforeSetOptions"},
+       {"BackupEngineTest::SetOptionsBackupRaceCondition:AfterSetOptions",
+        "CheckpointImpl::CreateCheckpoint:SavedLiveFiles2"}});
+  SyncPoint::GetInstance()->EnableProcessing();
+  ROCKSDB_NAMESPACE::port::Thread setoptions_thread{[this]() {
+    TEST_SYNC_POINT(
+        "BackupEngineTest::SetOptionsBackupRaceCondition:BeforeSetOptions");
+    DBImpl* dbi = static_cast<DBImpl*>(db_.get());
+    // Change arbitrary option to trigger OPTIONS file deletion
+    ASSERT_OK(dbi->SetOptions(dbi->DefaultColumnFamily(),
+                              {{"paranoid_file_checks", "false"}}));
+    ASSERT_OK(dbi->SetOptions(dbi->DefaultColumnFamily(),
+                              {{"paranoid_file_checks", "true"}}));
+    ASSERT_OK(dbi->SetOptions(dbi->DefaultColumnFamily(),
+                              {{"paranoid_file_checks", "false"}}));
+    TEST_SYNC_POINT(
+        "BackupEngineTest::SetOptionsBackupRaceCondition:AfterSetOptions");
+  }};
+  ASSERT_OK(backup_engine_->CreateNewBackup(db_.get()));
+  setoptions_thread.join();
+  CloseDBAndBackupEngine();
+}
+
+// This test verifies we don't delete the latest backup when read-only option is
+// set
+TEST_F(BackupEngineTest, NoDeleteWithReadOnly) {
+  const int keys_iteration = 5000;
+  Random rnd(6);
+
+  OpenDBAndBackupEngine(true);
+  // create five backups
+  for (int i = 0; i < 5; ++i) {
+    FillDB(db_.get(), keys_iteration * i, keys_iteration * (i + 1));
+    ASSERT_OK(backup_engine_->CreateNewBackup(db_.get(), !!(rnd.Next() % 2)));
+  }
+  CloseDBAndBackupEngine();
+  ASSERT_OK(file_manager_->WriteToFile(latest_backup_, "4"));
+
+  engine_options_->destroy_old_data = false;
+  BackupEngineReadOnly* read_only_backup_engine;
+  ASSERT_OK(BackupEngineReadOnly::Open(
+      backup_chroot_env_.get(), *engine_options_, &read_only_backup_engine));
+
+  // assert that data from backup 5 is still here (even though LATEST_BACKUP
+  // says 4 is latest)
+  ASSERT_OK(file_manager_->FileExists(backupdir_ + "/meta/5"));
+  ASSERT_OK(file_manager_->FileExists(backupdir_ + "/private/5"));
+
+  // Behavior change: We now ignore LATEST_BACKUP contents. This means that
+  // we should have 5 backups, even if LATEST_BACKUP says 4.
+  std::vector<BackupInfo> backup_info;
+  read_only_backup_engine->GetBackupInfo(&backup_info);
+  ASSERT_EQ(5UL, backup_info.size());
+  delete read_only_backup_engine;
+}
+
+TEST_F(BackupEngineTest, FailOverwritingBackups) {
+  options_.write_buffer_size = 1024 * 1024 * 1024;  // 1GB
+  options_.disable_auto_compactions = true;
+
+  // create backups 1, 2, 3, 4, 5
+  OpenDBAndBackupEngine(true);
+  for (int i = 0; i < 5; ++i) {
+    CloseDBAndBackupEngine();
+    DeleteLogFiles();
+    OpenDBAndBackupEngine(false);
+    FillDB(db_.get(), 100 * i, 100 * (i + 1), kFlushAll);
+    ASSERT_OK(backup_engine_->CreateNewBackup(db_.get()));
+  }
+  CloseDBAndBackupEngine();
+
+  // restore 3
+  OpenBackupEngine();
+  ASSERT_OK(backup_engine_->RestoreDBFromBackup(3, dbname_, dbname_));
+  CloseBackupEngine();
+
+  OpenDBAndBackupEngine(false);
+  // More data, bigger SST
+  FillDB(db_.get(), 1000, 1300, kFlushAll);
+  Status s = backup_engine_->CreateNewBackup(db_.get());
+  // the new backup fails because new table files
+  // clash with old table files from backups 4 and 5
+  // (since write_buffer_size is huge, we can be sure that
+  // each backup will generate only one sst file and that
+  // a file generated here would have the same name as an
+  // sst file generated by backup 4, and will be bigger)
+  ASSERT_TRUE(s.IsCorruption());
+  ASSERT_OK(backup_engine_->DeleteBackup(4));
+  ASSERT_OK(backup_engine_->DeleteBackup(5));
+  // now, the backup can succeed
+  ASSERT_OK(backup_engine_->CreateNewBackup(db_.get()));
+  CloseDBAndBackupEngine();
+}
+
+TEST_F(BackupEngineTest, NoShareTableFiles) {
+  const int keys_iteration = 5000;
+  OpenDBAndBackupEngine(true, false, kNoShare);
+  for (int i = 0; i < 5; ++i) {
+    FillDB(db_.get(), keys_iteration * i, keys_iteration * (i + 1));
+    ASSERT_OK(backup_engine_->CreateNewBackup(db_.get(), !!(i % 2)));
+  }
+  CloseDBAndBackupEngine();
+
+  for (int i = 0; i < 5; ++i) {
+    AssertBackupConsistency(i + 1, 0, keys_iteration * (i + 1),
+                            keys_iteration * 6);
+  }
+}
+
+// Verify that you can backup and restore with share_files_with_checksum on
+TEST_F(BackupEngineTest, ShareTableFilesWithChecksums) {
+  const int keys_iteration = 5000;
+  OpenDBAndBackupEngine(true, false, kShareWithChecksum);
+  for (int i = 0; i < 5; ++i) {
+    FillDB(db_.get(), keys_iteration * i, keys_iteration * (i + 1));
+    ASSERT_OK(backup_engine_->CreateNewBackup(db_.get(), !!(i % 2)));
+  }
+  CloseDBAndBackupEngine();
+
+  for (int i = 0; i < 5; ++i) {
+    AssertBackupConsistency(i + 1, 0, keys_iteration * (i + 1),
+                            keys_iteration * 6);
+  }
+}
+
+// Verify that you can backup and restore using share_files_with_checksum set to
+// false and then transition this option to true
+TEST_F(BackupEngineTest, ShareTableFilesWithChecksumsTransition) {
+  const int keys_iteration = 5000;
+  // set share_files_with_checksum to false
+  OpenDBAndBackupEngine(true, false, kShareNoChecksum);
+  for (int i = 0; i < 5; ++i) {
+    FillDB(db_.get(), keys_iteration * i, keys_iteration * (i + 1));
+    ASSERT_OK(backup_engine_->CreateNewBackup(db_.get(), true));
+  }
+  CloseDBAndBackupEngine();
+
+  for (int i = 0; i < 5; ++i) {
+    AssertBackupConsistency(i + 1, 0, keys_iteration * (i + 1),
+                            keys_iteration * 6);
+  }
+
+  // set share_files_with_checksum to true and do some more backups
+  OpenDBAndBackupEngine(false /* destroy_old_data */, false,
+                        kShareWithChecksum);
+  for (int i = 5; i < 10; ++i) {
+    FillDB(db_.get(), keys_iteration * i, keys_iteration * (i + 1));
+    ASSERT_OK(backup_engine_->CreateNewBackup(db_.get(), true));
+  }
+  CloseDBAndBackupEngine();
+
+  // Verify first (about to delete)
+  AssertBackupConsistency(1, 0, keys_iteration, keys_iteration * 11);
+
+  // For an extra challenge, make sure that GarbageCollect / DeleteBackup
+  // is OK even if we open without share_table_files
+  OpenDBAndBackupEngine(false /* destroy_old_data */, false, kNoShare);
+  ASSERT_OK(backup_engine_->DeleteBackup(1));
+  ASSERT_OK(backup_engine_->GarbageCollect());
+  CloseDBAndBackupEngine();
+
+  // Verify rest (not deleted)
+  for (int i = 1; i < 10; ++i) {
+    AssertBackupConsistency(i + 1, 0, keys_iteration * (i + 1),
+                            keys_iteration * 11);
+  }
+}
+
+// Verify backup and restore with various naming options, check names
+TEST_F(BackupEngineTest, ShareTableFilesWithChecksumsNewNaming) {
+  ASSERT_TRUE(engine_options_->share_files_with_checksum_naming ==
+              kNamingDefault);
+
+  const int keys_iteration = 5000;
+
+  OpenDBAndBackupEngine(true, false, kShareWithChecksum);
+  FillDB(db_.get(), 0, keys_iteration);
+  CloseDBAndBackupEngine();
+
+  static const std::map<ShareFilesNaming, TestRegex> option_to_expected = {
+      {kLegacyCrc32cAndFileSize, "[0-9]+_[0-9]+_[0-9]+[.]sst"},
+      // kFlagIncludeFileSize redundant here
+      {kLegacyCrc32cAndFileSize | kFlagIncludeFileSize,
+       "[0-9]+_[0-9]+_[0-9]+[.]sst"},
+      {kUseDbSessionId, "[0-9]+_s[0-9A-Z]{20}[.]sst"},
+      {kUseDbSessionId | kFlagIncludeFileSize,
+       "[0-9]+_s[0-9A-Z]{20}_[0-9]+[.]sst"},
+  };
+
+  const TestRegex blobfile_pattern = "[0-9]+_[0-9]+_[0-9]+[.]blob";
+
+  for (const auto& pair : option_to_expected) {
+    CloseAndReopenDB();
+    engine_options_->share_files_with_checksum_naming = pair.first;
+    OpenBackupEngine(true /*destroy_old_data*/);
+    ASSERT_OK(backup_engine_->CreateNewBackup(db_.get()));
+    CloseDBAndBackupEngine();
+    AssertBackupConsistency(1, 0, keys_iteration, keys_iteration * 2);
+    AssertDirectoryFilesMatchRegex(backupdir_ + "/shared_checksum", pair.second,
+                                   ".sst", 1 /* minimum_count */);
+    if (std::string::npos != pair.second.GetPattern().find("_[0-9]+[.]sst")) {
+      AssertDirectoryFilesSizeIndicators(backupdir_ + "/shared_checksum",
+                                         1 /* minimum_count */);
+    }
+
+    AssertDirectoryFilesMatchRegex(backupdir_ + "/shared_checksum",
+                                   blobfile_pattern, ".blob",
+                                   1 /* minimum_count */);
+  }
+}
+
+// Mimic SST file generated by pre-6.12 releases and verify that
+// old names are always used regardless of naming option.
+TEST_F(BackupEngineTest, ShareTableFilesWithChecksumsOldFileNaming) {
+  const int keys_iteration = 5000;
+
+  // Pre-6.12 release did not include db id and db session id properties.
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
+      "PropertyBlockBuilder::AddTableProperty:Start", [&](void* props_vs) {
+        auto props = static_cast<TableProperties*>(props_vs);
+        props->db_id = "";
+        props->db_session_id = "";
+      });
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->EnableProcessing();
+
+  // Corrupting the table properties corrupts the unique id.
+  // Ignore the unique id recorded in the manifest.
+  options_.verify_sst_unique_id_in_manifest = false;
+
+  OpenDBAndBackupEngine(true, false, kShareWithChecksum);
+  FillDB(db_.get(), 0, keys_iteration);
+  CloseDBAndBackupEngine();
+
+  // Old names should always be used on old files
+  const TestRegex sstfile_pattern("[0-9]+_[0-9]+_[0-9]+[.]sst");
+
+  const TestRegex blobfile_pattern = "[0-9]+_[0-9]+_[0-9]+[.]blob";
+
+  for (ShareFilesNaming option : {kNamingDefault, kUseDbSessionId}) {
+    CloseAndReopenDB();
+    engine_options_->share_files_with_checksum_naming = option;
+    OpenBackupEngine(true /*destroy_old_data*/);
+    ASSERT_OK(backup_engine_->CreateNewBackup(db_.get()));
+    CloseDBAndBackupEngine();
+    AssertBackupConsistency(1, 0, keys_iteration, keys_iteration * 2);
+    AssertDirectoryFilesMatchRegex(backupdir_ + "/shared_checksum",
+                                   sstfile_pattern, ".sst",
+                                   1 /* minimum_count */);
+    AssertDirectoryFilesMatchRegex(backupdir_ + "/shared_checksum",
+                                   blobfile_pattern, ".blob",
+                                   1 /* minimum_count */);
+  }
+
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->DisableProcessing();
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->ClearAllCallBacks();
+}
+
+// Test how naming options interact with detecting DB corruption
+// between incremental backups
+TEST_F(BackupEngineTest, TableFileCorruptionBeforeIncremental) {
+  const auto share_no_checksum = static_cast<ShareFilesNaming>(0);
+
+  for (bool corrupt_before_first_backup : {false, true}) {
+    for (ShareFilesNaming option :
+         {share_no_checksum, kLegacyCrc32cAndFileSize, kNamingDefault}) {
+      auto share =
+          option == share_no_checksum ? kShareNoChecksum : kShareWithChecksum;
+      if (option != share_no_checksum) {
+        engine_options_->share_files_with_checksum_naming = option;
+      }
+      OpenDBAndBackupEngine(true, false, share);
+      DBImpl* dbi = static_cast<DBImpl*>(db_.get());
+      // A small SST file
+      ASSERT_OK(dbi->Put(WriteOptions(), "x", "y"));
+      ASSERT_OK(dbi->Flush(FlushOptions()));
+      // And a bigger one
+      ASSERT_OK(dbi->Put(WriteOptions(), "y", Random(42).RandomString(500)));
+      ASSERT_OK(dbi->Flush(FlushOptions()));
+      ASSERT_OK(dbi->TEST_WaitForFlushMemTable());
+      CloseAndReopenDB(/*read_only*/ true);
+
+      std::vector<FileAttributes> table_files;
+      ASSERT_OK(GetDataFilesInDB(kTableFile, &table_files));
+      ASSERT_EQ(table_files.size(), 2);
+      std::string tf0 = dbname_ + "/" + table_files[0].name;
+      std::string tf1 = dbname_ + "/" + table_files[1].name;
+
+      CloseDBAndBackupEngine();
+
+      if (corrupt_before_first_backup) {
+        // This corrupts a data block, which does not cause DB open
+        // failure, only failure on accessing the block.
+        ASSERT_OK(db_file_manager_->CorruptFileStart(tf0));
+      }
+
+      OpenDBAndBackupEngine(false, false, share);
+      ASSERT_OK(backup_engine_->CreateNewBackup(db_.get()));
+      CloseDBAndBackupEngine();
+
+      // if corrupt_before_first_backup, this undoes the initial corruption
+      ASSERT_OK(db_file_manager_->CorruptFileStart(tf0));
+
+      OpenDBAndBackupEngine(false, false, share);
+      Status s = backup_engine_->CreateNewBackup(db_.get());
+
+      // Even though none of the naming options catch the inconsistency
+      // between the first and second time backing up fname, in the case
+      // of kUseDbSessionId (kNamingDefault), this is an intentional
+      // trade-off to avoid full scan of files from the DB that are
+      // already backed up. If we did the scan, kUseDbSessionId could catch
+      // the corruption. kLegacyCrc32cAndFileSize does the scan (to
+      // compute checksum for name) without catching the corruption,
+      // because the corruption means the names don't merge.
+      EXPECT_OK(s);
+
+      // VerifyBackup doesn't check DB integrity or table file internal
+      // checksums
+      EXPECT_OK(backup_engine_->VerifyBackup(1, true));
+      EXPECT_OK(backup_engine_->VerifyBackup(2, true));
+
+      db_.reset();
+      ASSERT_OK(backup_engine_->RestoreDBFromBackup(2, dbname_, dbname_));
+      {
+        DB* db = OpenDB();
+        s = db->VerifyChecksum();
+        delete db;
+      }
+      if (option != kLegacyCrc32cAndFileSize && !corrupt_before_first_backup) {
+        // Second backup is OK because it used (uncorrupt) file from first
+        // backup instead of (corrupt) file from DB.
+        // This is arguably a good trade-off vs. treating the file as distinct
+        // from the old version, because a file should be more likely to be
+        // corrupt as it ages. Although the backed-up file might also corrupt
+        // with age, the alternative approach (checksum in file name computed
+        // from current DB file contents) wouldn't detect that case at backup
+        // time either. Although you would have both copies of the file with
+        // the alternative approach, that would only last until the older
+        // backup is deleted.
+        ASSERT_OK(s);
+      } else if (option == kLegacyCrc32cAndFileSize &&
+                 corrupt_before_first_backup) {
+        // Second backup is OK because it saved the updated (uncorrupt)
+        // file from DB, instead of the sharing with first backup.
+        // Recall: if corrupt_before_first_backup, [second CorruptFileStart]
+        // undoes the initial corruption.
+        // This is arguably a bad trade-off vs. sharing the old version of the
+        // file because a file should be more likely to corrupt as it ages.
+        // (Not likely that the previously backed-up version was already
+        // corrupt and the new version is non-corrupt. This approach doesn't
+        // help if backed-up version is corrupted after taking the backup.)
+        ASSERT_OK(s);
+      } else {
+        // Something is legitimately corrupted, but we can't be sure what
+        // with information available (TODO? unless one passes block checksum
+        // test and other doesn't. Probably better to use end-to-end full file
+        // checksum anyway.)
+        ASSERT_TRUE(s.IsCorruption());
+      }
+
+      CloseDBAndBackupEngine();
+      DestroyDBWithoutCheck(dbname_, options_);
+    }
+  }
+}
+
+// Test how naming options interact with detecting file size corruption
+// between incremental backups
+TEST_F(BackupEngineTest, FileSizeForIncremental) {
+  const auto share_no_checksum = static_cast<ShareFilesNaming>(0);
+  // TODO: enable blob files once Integrated BlobDB supports DB session id.
+  options_.enable_blob_files = false;
+
+  for (ShareFilesNaming option : {share_no_checksum, kLegacyCrc32cAndFileSize,
+                                  kNamingDefault, kUseDbSessionId}) {
+    auto share =
+        option == share_no_checksum ? kShareNoChecksum : kShareWithChecksum;
+    if (option != share_no_checksum) {
+      engine_options_->share_files_with_checksum_naming = option;
+    }
+    OpenDBAndBackupEngine(true, false, share);
+
+    std::vector<FileAttributes> children;
+    const std::string shared_dir =
+        backupdir_ +
+        (option == share_no_checksum ? "/shared" : "/shared_checksum");
+
+    // A single small SST file
+    ASSERT_OK(db_->Put(WriteOptions(), "x", "y"));
+
+    // First, test that we always detect file size corruption on the shared
+    // backup side on incremental. (Since sizes aren't really part of backup
+    // meta file, this works by querying the filesystem for the sizes.)
+    ASSERT_OK(backup_engine_->CreateNewBackup(db_.get(), true /*flush*/));
+    CloseDBAndBackupEngine();
+
+    // Corrupt backup SST file
+    ASSERT_OK(file_manager_->GetChildrenFileAttributes(shared_dir, &children));
+    ASSERT_EQ(children.size(), 1U);  // one sst
+    for (const auto& child : children) {
+      if (child.name.size() > 4 && child.size_bytes > 0) {
+        ASSERT_OK(
+            file_manager_->WriteToFile(shared_dir + "/" + child.name, "asdf"));
+        break;
+      }
+    }
+
+    OpenDBAndBackupEngine(false, false, share);
+    Status s = backup_engine_->CreateNewBackup(db_.get());
+    EXPECT_TRUE(s.IsCorruption());
+
+    ASSERT_OK(backup_engine_->PurgeOldBackups(0));
+    CloseDBAndBackupEngine();
+
+    // Second, test that a hypothetical db session id collision would likely
+    // not suffice to corrupt a backup, because there's a good chance of
+    // file size difference (in this test, guaranteed) so either no name
+    // collision or detected collision.
+
+    // Create backup 1
+    OpenDBAndBackupEngine(false, false, share);
+    ASSERT_OK(backup_engine_->CreateNewBackup(db_.get()));
+
+    // Even though we have "the same" DB state as backup 1, we need
+    // to restore to recreate the same conditions as later restore.
+    db_.reset();
+    DestroyDBWithoutCheck(dbname_, options_);
+    ASSERT_OK(backup_engine_->RestoreDBFromBackup(1, dbname_, dbname_));
+    CloseDBAndBackupEngine();
+
+    // Forge session id
+    ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
+        "DBImpl::SetDbSessionId", [](void* sid_void_star) {
+          std::string* sid = static_cast<std::string*>(sid_void_star);
+          *sid = "01234567890123456789";
+        });
+    ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->EnableProcessing();
+
+    // Create another SST file
+    OpenDBAndBackupEngine(false, false, share);
+    ASSERT_OK(db_->Put(WriteOptions(), "y", "x"));
+
+    // Create backup 2
+    ASSERT_OK(backup_engine_->CreateNewBackup(db_.get(), true /*flush*/));
+
+    // Restore backup 1 (again)
+    db_.reset();
+    DestroyDBWithoutCheck(dbname_, options_);
+    ASSERT_OK(backup_engine_->RestoreDBFromBackup(1, dbname_, dbname_));
+    CloseDBAndBackupEngine();
+
+    // Create another SST file with same number and db session id, only bigger
+    OpenDBAndBackupEngine(false, false, share);
+    ASSERT_OK(db_->Put(WriteOptions(), "y", Random(42).RandomString(500)));
+
+    // Count backup SSTs files.
+    children.clear();
+    ASSERT_OK(file_manager_->GetChildrenFileAttributes(shared_dir, &children));
+    ASSERT_EQ(children.size(), 2U);  // two sst files
+
+    // Try create backup 3
+    s = backup_engine_->CreateNewBackup(db_.get(), true /*flush*/);
+
+    // Re-count backup SSTs
+    children.clear();
+    ASSERT_OK(file_manager_->GetChildrenFileAttributes(shared_dir, &children));
+
+    if (option == kUseDbSessionId) {
+      // Acceptable to call it corruption if size is not in name and
+      // db session id collision is practically impossible.
+      EXPECT_TRUE(s.IsCorruption());
+      EXPECT_EQ(children.size(), 2U);  // no SST file added
+    } else if (option == share_no_checksum) {
+      // Good to call it corruption if both backups cannot be
+      // accommodated.
+      EXPECT_TRUE(s.IsCorruption());
+      EXPECT_EQ(children.size(), 2U);  // no SST file added
+    } else {
+      // Since opening a DB seems sufficient for detecting size corruption
+      // on the DB side, this should be a good thing, ...
+      EXPECT_OK(s);
+      // ... as long as we did actually treat it as a distinct SST file.
+      EXPECT_EQ(children.size(), 3U);  // Another SST added
+    }
+    CloseDBAndBackupEngine();
+    DestroyDBWithoutCheck(dbname_, options_);
+    ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->DisableProcessing();
+    ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->ClearAllCallBacks();
+  }
+}
+
+// Verify backup and restore with share_files_with_checksum off and then
+// transition this option to on and share_files_with_checksum_naming to be
+// based on kUseDbSessionId
+TEST_F(BackupEngineTest, ShareTableFilesWithChecksumsNewNamingTransition) {
+  const int keys_iteration = 5000;
+  // We may set share_files_with_checksum_naming to kLegacyCrc32cAndFileSize
+  // here but even if we don't, it should have no effect when
+  // share_files_with_checksum is false
+  ASSERT_TRUE(engine_options_->share_files_with_checksum_naming ==
+              kNamingDefault);
+  // set share_files_with_checksum to false
+  OpenDBAndBackupEngine(true, false, kShareNoChecksum);
+  int j = 3;
+  for (int i = 0; i < j; ++i) {
+    FillDB(db_.get(), keys_iteration * i, keys_iteration * (i + 1));
+    ASSERT_OK(backup_engine_->CreateNewBackup(db_.get(), true));
+  }
+  CloseDBAndBackupEngine();
+
+  for (int i = 0; i < j; ++i) {
+    AssertBackupConsistency(i + 1, 0, keys_iteration * (i + 1),
+                            keys_iteration * (j + 1));
+  }
+
+  // set share_files_with_checksum to true and do some more backups
+  // and use session id in the name of SST file backup
+  ASSERT_TRUE(engine_options_->share_files_with_checksum_naming ==
+              kNamingDefault);
+  OpenDBAndBackupEngine(false /* destroy_old_data */, false,
+                        kShareWithChecksum);
+  FillDB(db_.get(), keys_iteration * j, keys_iteration * (j + 1));
+  ASSERT_OK(backup_engine_->CreateNewBackup(db_.get(), true));
+  CloseDBAndBackupEngine();
+  // Use checksum in the name as well
+  ++j;
+  options_.file_checksum_gen_factory = GetFileChecksumGenCrc32cFactory();
+  OpenDBAndBackupEngine(false /* destroy_old_data */, false,
+                        kShareWithChecksum);
+  FillDB(db_.get(), keys_iteration * j, keys_iteration * (j + 1));
+  ASSERT_OK(backup_engine_->CreateNewBackup(db_.get(), true));
+  CloseDBAndBackupEngine();
+
+  // Verify first (about to delete)
+  AssertBackupConsistency(1, 0, keys_iteration, keys_iteration * (j + 1));
+
+  // For an extra challenge, make sure that GarbageCollect / DeleteBackup
+  // is OK even if we open without share_table_files but with
+  // share_files_with_checksum_naming based on kUseDbSessionId
+  ASSERT_TRUE(engine_options_->share_files_with_checksum_naming ==
+              kNamingDefault);
+  OpenDBAndBackupEngine(false /* destroy_old_data */, false, kNoShare);
+  ASSERT_OK(backup_engine_->DeleteBackup(1));
+  ASSERT_OK(backup_engine_->GarbageCollect());
+  CloseDBAndBackupEngine();
+
+  // Verify second (about to delete)
+  AssertBackupConsistency(2, 0, keys_iteration * 2, keys_iteration * (j + 1));
+
+  // Use checksum and file size for backup table file names and open without
+  // share_table_files
+  // Again, make sure that GarbageCollect / DeleteBackup is OK
+  engine_options_->share_files_with_checksum_naming = kLegacyCrc32cAndFileSize;
+  OpenDBAndBackupEngine(false /* destroy_old_data */, false, kNoShare);
+  ASSERT_OK(backup_engine_->DeleteBackup(2));
+  ASSERT_OK(backup_engine_->GarbageCollect());
+  CloseDBAndBackupEngine();
+
+  // Verify rest (not deleted)
+  for (int i = 2; i < j; ++i) {
+    AssertBackupConsistency(i + 1, 0, keys_iteration * (i + 1),
+                            keys_iteration * (j + 1));
+  }
+}
+
+// Verify backup and restore with share_files_with_checksum on and transition
+// from kLegacyCrc32cAndFileSize to kUseDbSessionId
+TEST_F(BackupEngineTest, ShareTableFilesWithChecksumsNewNamingUpgrade) {
+  engine_options_->share_files_with_checksum_naming = kLegacyCrc32cAndFileSize;
+  const int keys_iteration = 5000;
+  // set share_files_with_checksum to true
+  OpenDBAndBackupEngine(true, false, kShareWithChecksum);
+  int j = 3;
+  for (int i = 0; i < j; ++i) {
+    FillDB(db_.get(), keys_iteration * i, keys_iteration * (i + 1));
+    ASSERT_OK(backup_engine_->CreateNewBackup(db_.get(), true));
+  }
+  CloseDBAndBackupEngine();
+
+  for (int i = 0; i < j; ++i) {
+    AssertBackupConsistency(i + 1, 0, keys_iteration * (i + 1),
+                            keys_iteration * (j + 1));
+  }
+
+  engine_options_->share_files_with_checksum_naming = kUseDbSessionId;
+  OpenDBAndBackupEngine(false /* destroy_old_data */, false,
+                        kShareWithChecksum);
+  FillDB(db_.get(), keys_iteration * j, keys_iteration * (j + 1));
+  ASSERT_OK(backup_engine_->CreateNewBackup(db_.get(), true));
+  CloseDBAndBackupEngine();
+
+  ++j;
+  options_.file_checksum_gen_factory = GetFileChecksumGenCrc32cFactory();
+  OpenDBAndBackupEngine(false /* destroy_old_data */, false,
+                        kShareWithChecksum);
+  FillDB(db_.get(), keys_iteration * j, keys_iteration * (j + 1));
+  ASSERT_OK(backup_engine_->CreateNewBackup(db_.get(), true));
+  CloseDBAndBackupEngine();
+
+  // Verify first (about to delete)
+  AssertBackupConsistency(1, 0, keys_iteration, keys_iteration * (j + 1));
+
+  // For an extra challenge, make sure that GarbageCollect / DeleteBackup
+  // is OK even if we open without share_table_files
+  OpenDBAndBackupEngine(false /* destroy_old_data */, false, kNoShare);
+  ASSERT_OK(backup_engine_->DeleteBackup(1));
+  ASSERT_OK(backup_engine_->GarbageCollect());
+  CloseDBAndBackupEngine();
+
+  // Verify second (about to delete)
+  AssertBackupConsistency(2, 0, keys_iteration * 2, keys_iteration * (j + 1));
+
+  // Use checksum and file size for backup table file names and open without
+  // share_table_files
+  // Again, make sure that GarbageCollect / DeleteBackup is OK
+  engine_options_->share_files_with_checksum_naming = kLegacyCrc32cAndFileSize;
+  OpenDBAndBackupEngine(false /* destroy_old_data */, false, kNoShare);
+  ASSERT_OK(backup_engine_->DeleteBackup(2));
+  ASSERT_OK(backup_engine_->GarbageCollect());
+  CloseDBAndBackupEngine();
+
+  // Verify rest (not deleted)
+  for (int i = 2; i < j; ++i) {
+    AssertBackupConsistency(i + 1, 0, keys_iteration * (i + 1),
+                            keys_iteration * (j + 1));
+  }
+}
+
+// This test simulates cleaning up after aborted or incomplete creation
+// of a new backup.
+TEST_F(BackupEngineTest, DeleteTmpFiles) {
+  for (int cleanup_fn : {1, 2, 3, 4}) {
+    for (ShareOption shared_option : kAllShareOptions) {
+      OpenDBAndBackupEngine(false /* destroy_old_data */, false /* dummy */,
+                            shared_option);
+      ASSERT_OK(backup_engine_->CreateNewBackup(db_.get()));
+      BackupID next_id = 1;
+      BackupID oldest_id = std::numeric_limits<BackupID>::max();
+      {
+        std::vector<BackupInfo> backup_info;
+        backup_engine_->GetBackupInfo(&backup_info);
+        for (const auto& bi : backup_info) {
+          next_id = std::max(next_id, bi.backup_id + 1);
+          oldest_id = std::min(oldest_id, bi.backup_id);
+        }
+      }
+      CloseDBAndBackupEngine();
+
+      // An aborted or incomplete new backup will always be in the next
+      // id (maybe more)
+      std::string next_private = "private/" + std::to_string(next_id);
+
+      // NOTE: both shared and shared_checksum should be cleaned up
+      // regardless of how the backup engine is opened.
+      std::vector<std::string> tmp_files_and_dirs;
+      for (const auto& dir_and_file : {
+               std::make_pair(std::string("shared"),
+                              std::string(".00006.sst.tmp")),
+               std::make_pair(std::string("shared_checksum"),
+                              std::string(".00007.sst.tmp")),
+               std::make_pair(next_private, std::string("00003.sst")),
+           }) {
+        std::string dir = backupdir_ + "/" + dir_and_file.first;
+        ASSERT_OK(file_manager_->CreateDirIfMissing(dir));
+        ASSERT_OK(file_manager_->FileExists(dir));
+
+        std::string file = dir + "/" + dir_and_file.second;
+        ASSERT_OK(file_manager_->WriteToFile(file, "tmp"));
+        ASSERT_OK(file_manager_->FileExists(file));
+
+        tmp_files_and_dirs.push_back(file);
+      }
+      if (cleanup_fn != /*CreateNewBackup*/ 4) {
+        // This exists after CreateNewBackup because it's deleted then
+        // re-created.
+        tmp_files_and_dirs.push_back(backupdir_ + "/" + next_private);
+      }
+
+      OpenDBAndBackupEngine(false /* destroy_old_data */, false /* dummy */,
+                            shared_option);
+      // Need to call one of these explicitly to delete tmp files
+      switch (cleanup_fn) {
+        case 1:
+          ASSERT_OK(backup_engine_->GarbageCollect());
+          break;
+        case 2:
+          ASSERT_OK(backup_engine_->DeleteBackup(oldest_id));
+          break;
+        case 3:
+          ASSERT_OK(backup_engine_->PurgeOldBackups(1));
+          break;
+        case 4:
+          // Does a garbage collect if it sees that next private dir exists
+          ASSERT_OK(backup_engine_->CreateNewBackup(db_.get()));
+          break;
+        default:
+          assert(false);
+      }
+      CloseDBAndBackupEngine();
+      for (std::string file_or_dir : tmp_files_and_dirs) {
+        if (file_manager_->FileExists(file_or_dir) != Status::NotFound()) {
+          FAIL() << file_or_dir << " was expected to be deleted." << cleanup_fn;
+        }
+      }
+    }
+  }
+}
+
+TEST_F(BackupEngineTest, KeepLogFiles) {
+  engine_options_->backup_log_files = false;
+  // basically infinite
+  options_.WAL_ttl_seconds = 24 * 60 * 60;
+  OpenDBAndBackupEngine(true);
+  FillDB(db_.get(), 0, 100, kFlushAll);
+  FillDB(db_.get(), 100, 200, kFlushAll);
+  ASSERT_OK(backup_engine_->CreateNewBackup(db_.get(), false));
+  FillDB(db_.get(), 200, 300, kFlushAll);
+  FillDB(db_.get(), 300, 400, kFlushAll);
+  FillDB(db_.get(), 400, 500, kFlushAll);
+  CloseDBAndBackupEngine();
+
+  // all data should be there if we call with keep_log_files = true
+  AssertBackupConsistency(0, 0, 500, 600, true);
+}
+
+#if !defined(ROCKSDB_VALGRIND_RUN) || defined(ROCKSDB_FULL_VALGRIND_RUN)
+class BackupEngineRateLimitingTestWithParam
+    : public BackupEngineTest,
+      public testing::WithParamInterface<
+          std::tuple<bool /* make throttle */,
+                     int /* 0 = single threaded, 1 = multi threaded*/,
+                     std::pair<uint64_t, uint64_t> /* limits */>> {
+ public:
+  BackupEngineRateLimitingTestWithParam() {}
+};
+
+uint64_t const MB = 1024 * 1024;
+
+INSTANTIATE_TEST_CASE_P(
+    RateLimiting, BackupEngineRateLimitingTestWithParam,
+    ::testing::Values(std::make_tuple(false, 0, std::make_pair(1 * MB, 5 * MB)),
+                      std::make_tuple(false, 0, std::make_pair(2 * MB, 3 * MB)),
+                      std::make_tuple(false, 1, std::make_pair(1 * MB, 5 * MB)),
+                      std::make_tuple(false, 1, std::make_pair(2 * MB, 3 * MB)),
+                      std::make_tuple(true, 0, std::make_pair(1 * MB, 5 * MB)),
+                      std::make_tuple(true, 0, std::make_pair(2 * MB, 3 * MB)),
+                      std::make_tuple(true, 1, std::make_pair(1 * MB, 5 * MB)),
+                      std::make_tuple(true, 1,
+                                      std::make_pair(2 * MB, 3 * MB))));
+
+TEST_P(BackupEngineRateLimitingTestWithParam, RateLimiting) {
+  size_t const kMicrosPerSec = 1000 * 1000LL;
+  const bool custom_rate_limiter = std::get<0>(GetParam());
+  // iter 0 -- single threaded
+  // iter 1 -- multi threaded
+  const int iter = std::get<1>(GetParam());
+  const std::pair<uint64_t, uint64_t> limit = std::get<2>(GetParam());
+  std::unique_ptr<Env> special_env(
+      new SpecialEnv(db_chroot_env_.get(), /*time_elapse_only_sleep*/ true));
+  // destroy old data
+  Options options;
+  options.env = special_env.get();
+  DestroyDBWithoutCheck(dbname_, options);
+
+  if (custom_rate_limiter) {
+    std::shared_ptr<RateLimiter> backup_rate_limiter =
+        std::make_shared<GenericRateLimiter>(
+            limit.first, 100 * 1000 /* refill_period_us */, 10 /* fairness */,
+            RateLimiter::Mode::kWritesOnly /* mode */,
+            special_env->GetSystemClock(), false /* auto_tuned */);
+    std::shared_ptr<RateLimiter> restore_rate_limiter =
+        std::make_shared<GenericRateLimiter>(
+            limit.second, 100 * 1000 /* refill_period_us */, 10 /* fairness */,
+            RateLimiter::Mode::kWritesOnly /* mode */,
+            special_env->GetSystemClock(), false /* auto_tuned */);
+    engine_options_->backup_rate_limiter = backup_rate_limiter;
+    engine_options_->restore_rate_limiter = restore_rate_limiter;
+  } else {
+    engine_options_->backup_rate_limit = limit.first;
+    engine_options_->restore_rate_limit = limit.second;
+  }
+
+  engine_options_->max_background_operations = (iter == 0) ? 1 : 10;
+  options_.compression = kNoCompression;
+
+  // Rate limiter uses `CondVar::TimedWait()`, which does not have access to the
+  // `Env` to advance its time according to the fake wait duration. The
+  // workaround is to install a callback that advance the `Env`'s mock time.
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
+      "GenericRateLimiter::Request:PostTimedWait", [&](void* arg) {
+        int64_t time_waited_us = *static_cast<int64_t*>(arg);
+        special_env->SleepForMicroseconds(static_cast<int>(time_waited_us));
+      });
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->EnableProcessing();
+
+  OpenDBAndBackupEngine(true);
+  TEST_SetDefaultRateLimitersClock(backup_engine_.get(),
+                                   special_env->GetSystemClock());
+
+  size_t bytes_written = FillDB(db_.get(), 0, 10000);
+
+  auto start_backup = special_env->NowMicros();
+  ASSERT_OK(backup_engine_->CreateNewBackup(db_.get(), false));
+  auto backup_time = special_env->NowMicros() - start_backup;
+  CloseDBAndBackupEngine();
+  auto rate_limited_backup_time = (bytes_written * kMicrosPerSec) / limit.first;
+  ASSERT_GT(backup_time, 0.8 * rate_limited_backup_time);
+
+  OpenBackupEngine();
+  TEST_SetDefaultRateLimitersClock(
+      backup_engine_.get(),
+      special_env->GetSystemClock() /* backup_rate_limiter_clock */,
+      special_env->GetSystemClock() /* restore_rate_limiter_clock */);
+
+  auto start_restore = special_env->NowMicros();
+  ASSERT_OK(backup_engine_->RestoreDBFromLatestBackup(dbname_, dbname_));
+  auto restore_time = special_env->NowMicros() - start_restore;
+  CloseBackupEngine();
+  auto rate_limited_restore_time =
+      (bytes_written * kMicrosPerSec) / limit.second;
+  ASSERT_GT(restore_time, 0.8 * rate_limited_restore_time);
+
+  AssertBackupConsistency(0, 0, 10000, 10100);
+
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->DisableProcessing();
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->ClearCallBack(
+      "GenericRateLimiter::Request:PostTimedWait");
+}
+
+TEST_P(BackupEngineRateLimitingTestWithParam, RateLimitingVerifyBackup) {
+  const std::size_t kMicrosPerSec = 1000 * 1000LL;
+  const bool custom_rate_limiter = std::get<0>(GetParam());
+  const std::uint64_t backup_rate_limiter_limit = std::get<2>(GetParam()).first;
+  const bool is_single_threaded = std::get<1>(GetParam()) == 0 ? true : false;
+  std::unique_ptr<Env> special_env(
+      new SpecialEnv(db_chroot_env_.get(), /*time_elapse_only_sleep*/ true));
+
+  if (custom_rate_limiter) {
+    std::shared_ptr<RateLimiter> backup_rate_limiter =
+        std::make_shared<GenericRateLimiter>(
+            backup_rate_limiter_limit, 100 * 1000 /* refill_period_us */,
+            10 /* fairness */, RateLimiter::Mode::kAllIo /* mode */,
+            special_env->GetSystemClock(), false /* auto_tuned */);
+    engine_options_->backup_rate_limiter = backup_rate_limiter;
+  } else {
+    engine_options_->backup_rate_limit = backup_rate_limiter_limit;
+  }
+
+  engine_options_->max_background_operations = is_single_threaded ? 1 : 10;
+
+  Options options;
+  options.env = special_env.get();
+  DestroyDBWithoutCheck(dbname_, options);
+  // Rate limiter uses `CondVar::TimedWait()`, which does not have access to the
+  // `Env` to advance its time according to the fake wait duration. The
+  // workaround is to install a callback that advance the `Env`'s mock time.
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
+      "GenericRateLimiter::Request:PostTimedWait", [&](void* arg) {
+        int64_t time_waited_us = *static_cast<int64_t*>(arg);
+        special_env->SleepForMicroseconds(static_cast<int>(time_waited_us));
+      });
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->EnableProcessing();
+  OpenDBAndBackupEngine(true /* destroy_old_data */);
+  TEST_SetDefaultRateLimitersClock(backup_engine_.get(),
+                                   special_env->GetSystemClock(), nullptr);
+  FillDB(db_.get(), 0, 10000);
+
+  ASSERT_OK(backup_engine_->CreateNewBackup(db_.get(),
+                                            false /* flush_before_backup */));
+
+  std::vector<BackupInfo> backup_infos;
+  BackupInfo backup_info;
+  backup_engine_->GetBackupInfo(&backup_infos);
+  ASSERT_EQ(1, backup_infos.size());
+  const int backup_id = 1;
+  ASSERT_EQ(backup_id, backup_infos[0].backup_id);
+  ASSERT_OK(backup_engine_->GetBackupInfo(backup_id, &backup_info,
+                                          true /* include_file_details */));
+
+  std::uint64_t bytes_read_during_verify_backup = 0;
+  for (BackupFileInfo backup_file_info : backup_info.file_details) {
+    bytes_read_during_verify_backup += backup_file_info.size;
+  }
+  auto start_verify_backup = special_env->NowMicros();
+  ASSERT_OK(
+      backup_engine_->VerifyBackup(backup_id, true /* verify_with_checksum */));
+  auto verify_backup_time = special_env->NowMicros() - start_verify_backup;
+  auto rate_limited_verify_backup_time =
+      (bytes_read_during_verify_backup * kMicrosPerSec) /
+      backup_rate_limiter_limit;
+  if (custom_rate_limiter) {
+    EXPECT_GE(verify_backup_time, 0.8 * rate_limited_verify_backup_time);
+  }
+
+  CloseDBAndBackupEngine();
+  AssertBackupConsistency(backup_id, 0, 10000, 10010);
+  DestroyDBWithoutCheck(dbname_, options);
+
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->DisableProcessing();
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->ClearCallBack(
+      "GenericRateLimiter::Request:PostTimedWait");
+}
+
+TEST_P(BackupEngineRateLimitingTestWithParam, RateLimitingChargeReadInBackup) {
+  bool is_single_threaded = std::get<1>(GetParam()) == 0 ? true : false;
+  engine_options_->max_background_operations = is_single_threaded ? 1 : 10;
+
+  const std::uint64_t backup_rate_limiter_limit = std::get<2>(GetParam()).first;
+  std::shared_ptr<RateLimiter> backup_rate_limiter(NewGenericRateLimiter(
+      backup_rate_limiter_limit, 100 * 1000 /* refill_period_us */,
+      10 /* fairness */, RateLimiter::Mode::kWritesOnly /* mode */));
+  engine_options_->backup_rate_limiter = backup_rate_limiter;
+
+  DestroyDBWithoutCheck(dbname_, Options());
+  OpenDBAndBackupEngine(true /* destroy_old_data */, false /* dummy */,
+                        kShareWithChecksum /* shared_option */);
+  FillDB(db_.get(), 0, 10);
+  ASSERT_OK(backup_engine_->CreateNewBackup(db_.get(),
+                                            false /* flush_before_backup */));
+  std::int64_t total_bytes_through_with_no_read_charged =
+      backup_rate_limiter->GetTotalBytesThrough();
+  CloseBackupEngine();
+
+  backup_rate_limiter.reset(NewGenericRateLimiter(
+      backup_rate_limiter_limit, 100 * 1000 /* refill_period_us */,
+      10 /* fairness */, RateLimiter::Mode::kAllIo /* mode */));
+  engine_options_->backup_rate_limiter = backup_rate_limiter;
+
+  OpenBackupEngine(true);
+  ASSERT_OK(backup_engine_->CreateNewBackup(db_.get(),
+                                            false /* flush_before_backup */));
+  std::int64_t total_bytes_through_with_read_charged =
+      backup_rate_limiter->GetTotalBytesThrough();
+  EXPECT_GT(total_bytes_through_with_read_charged,
+            total_bytes_through_with_no_read_charged);
+  CloseDBAndBackupEngine();
+  AssertBackupConsistency(1, 0, 10, 20);
+  DestroyDBWithoutCheck(dbname_, Options());
+}
+
+TEST_P(BackupEngineRateLimitingTestWithParam, RateLimitingChargeReadInRestore) {
+  bool is_single_threaded = std::get<1>(GetParam()) == 0 ? true : false;
+  engine_options_->max_background_operations = is_single_threaded ? 1 : 10;
+
+  const std::uint64_t restore_rate_limiter_limit =
+      std::get<2>(GetParam()).second;
+  std::shared_ptr<RateLimiter> restore_rate_limiter(NewGenericRateLimiter(
+      restore_rate_limiter_limit, 100 * 1000 /* refill_period_us */,
+      10 /* fairness */, RateLimiter::Mode::kWritesOnly /* mode */));
+  engine_options_->restore_rate_limiter = restore_rate_limiter;
+
+  DestroyDBWithoutCheck(dbname_, Options());
+  OpenDBAndBackupEngine(true /* destroy_old_data */);
+  FillDB(db_.get(), 0, 10);
+  ASSERT_OK(backup_engine_->CreateNewBackup(db_.get(),
+                                            false /* flush_before_backup */));
+  CloseDBAndBackupEngine();
+  DestroyDBWithoutCheck(dbname_, Options());
+
+  OpenBackupEngine(false /* destroy_old_data */);
+  ASSERT_OK(backup_engine_->RestoreDBFromLatestBackup(dbname_, dbname_));
+  std::int64_t total_bytes_through_with_no_read_charged =
+      restore_rate_limiter->GetTotalBytesThrough();
+  CloseBackupEngine();
+  DestroyDBWithoutCheck(dbname_, Options());
+
+  restore_rate_limiter.reset(NewGenericRateLimiter(
+      restore_rate_limiter_limit, 100 * 1000 /* refill_period_us */,
+      10 /* fairness */, RateLimiter::Mode::kAllIo /* mode */));
+  engine_options_->restore_rate_limiter = restore_rate_limiter;
+
+  OpenBackupEngine(false /* destroy_old_data */);
+  ASSERT_OK(backup_engine_->RestoreDBFromLatestBackup(dbname_, dbname_));
+  std::int64_t total_bytes_through_with_read_charged =
+      restore_rate_limiter->GetTotalBytesThrough();
+  EXPECT_EQ(total_bytes_through_with_read_charged,
+            total_bytes_through_with_no_read_charged * 2);
+  CloseBackupEngine();
+  AssertBackupConsistency(1, 0, 10, 20);
+  DestroyDBWithoutCheck(dbname_, Options());
+}
+
+TEST_P(BackupEngineRateLimitingTestWithParam,
+       RateLimitingChargeReadInInitialize) {
+  bool is_single_threaded = std::get<1>(GetParam()) == 0 ? true : false;
+  engine_options_->max_background_operations = is_single_threaded ? 1 : 10;
+
+  const std::uint64_t backup_rate_limiter_limit = std::get<2>(GetParam()).first;
+  std::shared_ptr<RateLimiter> backup_rate_limiter(NewGenericRateLimiter(
+      backup_rate_limiter_limit, 100 * 1000 /* refill_period_us */,
+      10 /* fairness */, RateLimiter::Mode::kAllIo /* mode */));
+  engine_options_->backup_rate_limiter = backup_rate_limiter;
+
+  DestroyDBWithoutCheck(dbname_, Options());
+  OpenDBAndBackupEngine(true /* destroy_old_data */);
+  FillDB(db_.get(), 0, 10);
+  ASSERT_OK(backup_engine_->CreateNewBackup(db_.get(),
+                                            false /* flush_before_backup */));
+  CloseDBAndBackupEngine();
+  AssertBackupConsistency(1, 0, 10, 20);
+
+  std::int64_t total_bytes_through_before_initialize =
+      engine_options_->backup_rate_limiter->GetTotalBytesThrough();
+  OpenDBAndBackupEngine(false /* destroy_old_data */);
+  // We charge read in BackupEngineImpl::BackupMeta::LoadFromFile,
+  // which is called in BackupEngineImpl::Initialize() during
+  // OpenBackupEngine(false)
+  EXPECT_GT(engine_options_->backup_rate_limiter->GetTotalBytesThrough(),
+            total_bytes_through_before_initialize);
+  CloseDBAndBackupEngine();
+  DestroyDBWithoutCheck(dbname_, Options());
+}
+
+class BackupEngineRateLimitingTestWithParam2
+    : public BackupEngineTest,
+      public testing::WithParamInterface<
+          std::tuple<std::pair<uint64_t, uint64_t> /* limits */>> {
+ public:
+  BackupEngineRateLimitingTestWithParam2() {}
+};
+
+INSTANTIATE_TEST_CASE_P(
+    LowRefillBytesPerPeriod, BackupEngineRateLimitingTestWithParam2,
+    ::testing::Values(std::make_tuple(std::make_pair(1, 1))));
+// To verify we don't request over-sized bytes relative to
+// refill_bytes_per_period_ in each RateLimiter::Request() called in
+// BackupEngine through verifying we don't trigger assertion
+// failure on over-sized request in GenericRateLimiter in debug builds
+TEST_P(BackupEngineRateLimitingTestWithParam2,
+       RateLimitingWithLowRefillBytesPerPeriod) {
+  SpecialEnv special_env(Env::Default(), /*time_elapse_only_sleep*/ true);
+
+  engine_options_->max_background_operations = 1;
+  const uint64_t backup_rate_limiter_limit = std::get<0>(GetParam()).first;
+  std::shared_ptr<RateLimiter> backup_rate_limiter(
+      std::make_shared<GenericRateLimiter>(
+          backup_rate_limiter_limit, 1000 * 1000 /* refill_period_us */,
+          10 /* fairness */, RateLimiter::Mode::kAllIo /* mode */,
+          special_env.GetSystemClock(), false /* auto_tuned */));
+
+  engine_options_->backup_rate_limiter = backup_rate_limiter;
+
+  const uint64_t restore_rate_limiter_limit = std::get<0>(GetParam()).second;
+  std::shared_ptr<RateLimiter> restore_rate_limiter(
+      std::make_shared<GenericRateLimiter>(
+          restore_rate_limiter_limit, 1000 * 1000 /* refill_period_us */,
+          10 /* fairness */, RateLimiter::Mode::kAllIo /* mode */,
+          special_env.GetSystemClock(), false /* auto_tuned */));
+
+  engine_options_->restore_rate_limiter = restore_rate_limiter;
+
+  // Rate limiter uses `CondVar::TimedWait()`, which does not have access to the
+  // `Env` to advance its time according to the fake wait duration. The
+  // workaround is to install a callback that advance the `Env`'s mock time.
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
+      "GenericRateLimiter::Request:PostTimedWait", [&](void* arg) {
+        int64_t time_waited_us = *static_cast<int64_t*>(arg);
+        special_env.SleepForMicroseconds(static_cast<int>(time_waited_us));
+      });
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->EnableProcessing();
+
+  DestroyDBWithoutCheck(dbname_, Options());
+  OpenDBAndBackupEngine(true /* destroy_old_data */, false /* dummy */,
+                        kShareWithChecksum /* shared_option */);
+
+  FillDB(db_.get(), 0, 100);
+  int64_t total_bytes_through_before_backup =
+      engine_options_->backup_rate_limiter->GetTotalBytesThrough();
+  EXPECT_OK(backup_engine_->CreateNewBackup(db_.get(),
+                                            false /* flush_before_backup */));
+  int64_t total_bytes_through_after_backup =
+      engine_options_->backup_rate_limiter->GetTotalBytesThrough();
+  ASSERT_GT(total_bytes_through_after_backup,
+            total_bytes_through_before_backup);
+
+  std::vector<BackupInfo> backup_infos;
+  BackupInfo backup_info;
+  backup_engine_->GetBackupInfo(&backup_infos);
+  ASSERT_EQ(1, backup_infos.size());
+  const int backup_id = 1;
+  ASSERT_EQ(backup_id, backup_infos[0].backup_id);
+  ASSERT_OK(backup_engine_->GetBackupInfo(backup_id, &backup_info,
+                                          true /* include_file_details */));
+  int64_t total_bytes_through_before_verify_backup =
+      engine_options_->backup_rate_limiter->GetTotalBytesThrough();
+  EXPECT_OK(
+      backup_engine_->VerifyBackup(backup_id, true /* verify_with_checksum */));
+  int64_t total_bytes_through_after_verify_backup =
+      engine_options_->backup_rate_limiter->GetTotalBytesThrough();
+  ASSERT_GT(total_bytes_through_after_verify_backup,
+            total_bytes_through_before_verify_backup);
+
+  CloseDBAndBackupEngine();
+  AssertBackupConsistency(backup_id, 0, 100, 101);
+
+  int64_t total_bytes_through_before_initialize =
+      engine_options_->backup_rate_limiter->GetTotalBytesThrough();
+  OpenDBAndBackupEngine(false /* destroy_old_data */);
+  // We charge read in BackupEngineImpl::BackupMeta::LoadFromFile,
+  // which is called in BackupEngineImpl::Initialize() during
+  // OpenBackupEngine(false)
+  int64_t total_bytes_through_after_initialize =
+      engine_options_->backup_rate_limiter->GetTotalBytesThrough();
+  ASSERT_GT(total_bytes_through_after_initialize,
+            total_bytes_through_before_initialize);
+  CloseDBAndBackupEngine();
+
+  DestroyDBWithoutCheck(dbname_, Options());
+  OpenBackupEngine(false /* destroy_old_data */);
+  int64_t total_bytes_through_before_restore =
+      engine_options_->restore_rate_limiter->GetTotalBytesThrough();
+  EXPECT_OK(backup_engine_->RestoreDBFromLatestBackup(dbname_, dbname_));
+  int64_t total_bytes_through_after_restore =
+      engine_options_->restore_rate_limiter->GetTotalBytesThrough();
+  ASSERT_GT(total_bytes_through_after_restore,
+            total_bytes_through_before_restore);
+  CloseBackupEngine();
+
+  DestroyDBWithoutCheck(dbname_, Options());
+
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->DisableProcessing();
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->ClearCallBack(
+      "GenericRateLimiter::Request:PostTimedWait");
+}
+
+#endif  // !defined(ROCKSDB_VALGRIND_RUN) || defined(ROCKSDB_FULL_VALGRIND_RUN)
+
+TEST_F(BackupEngineTest, ReadOnlyBackupEngine) {
+  DestroyDBWithoutCheck(dbname_, options_);
+  OpenDBAndBackupEngine(true);
+  FillDB(db_.get(), 0, 100);
+  // Also test read-only DB with CreateNewBackup and flush=true (no flush)
+  CloseAndReopenDB(/*read_only*/ true);
+  ASSERT_OK(backup_engine_->CreateNewBackup(db_.get(), /*flush*/ true));
+  CloseAndReopenDB(/*read_only*/ false);
+  FillDB(db_.get(), 100, 200);
+  ASSERT_OK(backup_engine_->CreateNewBackup(db_.get(), /*flush*/ true));
+  CloseDBAndBackupEngine();
+  DestroyDBWithoutCheck(dbname_, options_);
+
+  engine_options_->destroy_old_data = false;
+  test_backup_fs_->ClearWrittenFiles();
+  test_backup_fs_->SetLimitDeleteFiles(0);
+  BackupEngineReadOnly* read_only_backup_engine;
+  ASSERT_OK(BackupEngineReadOnly::Open(db_chroot_env_.get(), *engine_options_,
+                                       &read_only_backup_engine));
+  std::vector<BackupInfo> backup_info;
+  read_only_backup_engine->GetBackupInfo(&backup_info);
+  ASSERT_EQ(backup_info.size(), 2U);
+
+  RestoreOptions restore_options(false);
+  ASSERT_OK(read_only_backup_engine->RestoreDBFromLatestBackup(
+      dbname_, dbname_, restore_options));
+  delete read_only_backup_engine;
+  std::vector<std::string> should_have_written;
+  test_backup_fs_->AssertWrittenFiles(should_have_written);
+
+  DB* db = OpenDB();
+  AssertExists(db, 0, 200);
+  delete db;
+}
+
+TEST_F(BackupEngineTest, OpenBackupAsReadOnlyDB) {
+  DestroyDBWithoutCheck(dbname_, options_);
+  options_.write_dbid_to_manifest = false;
+
+  OpenDBAndBackupEngine(true);
+  FillDB(db_.get(), 0, 100);
+  ASSERT_OK(backup_engine_->CreateNewBackup(db_.get(), /*flush*/ false));
+
+  options_.write_dbid_to_manifest = true;  // exercises some read-only DB code
+  CloseAndReopenDB();
+
+  FillDB(db_.get(), 100, 200);
+  ASSERT_OK(backup_engine_->CreateNewBackup(db_.get(), /*flush*/ false));
+  db_.reset();  // CloseDB
+  DestroyDBWithoutCheck(dbname_, options_);
+  BackupInfo backup_info;
+  // First, check that we get empty fields without include_file_details
+  ASSERT_OK(backup_engine_->GetBackupInfo(/*id*/ 1U, &backup_info,
+                                          /*with file details*/ false));
+  ASSERT_EQ(backup_info.name_for_open, "");
+  ASSERT_FALSE(backup_info.env_for_open);
+
+  // Now for the real test
+  backup_info = BackupInfo();
+  ASSERT_OK(backup_engine_->GetBackupInfo(/*id*/ 1U, &backup_info,
+                                          /*with file details*/ true));
+
+  // Caution: DBOptions only holds a raw pointer to Env, so something else
+  // must keep it alive.
+  // Case 1: Keeping BackupEngine open suffices to keep Env alive
+  DB* db = nullptr;
+  Options opts = options_;
+  // Ensure some key defaults are set
+  opts.wal_dir = "";
+  opts.create_if_missing = false;
+  opts.info_log.reset();
+
+  opts.env = backup_info.env_for_open.get();
+  std::string name = backup_info.name_for_open;
+  backup_info = BackupInfo();
+  ASSERT_OK(DB::OpenForReadOnly(opts, name, &db));
+
+  AssertExists(db, 0, 100);
+  AssertEmpty(db, 100, 200);
+
+  delete db;
+  db = nullptr;
+
+  // Case 2: Keeping BackupInfo alive rather than BackupEngine also suffices
+  ASSERT_OK(backup_engine_->GetBackupInfo(/*id*/ 2U, &backup_info,
+                                          /*with file details*/ true));
+  CloseBackupEngine();
+  opts.create_if_missing = true;  // check also OK (though pointless)
+  opts.env = backup_info.env_for_open.get();
+  name = backup_info.name_for_open;
+  // Note: keeping backup_info alive
+  ASSERT_OK(DB::OpenForReadOnly(opts, name, &db));
+
+  AssertExists(db, 0, 200);
+  delete db;
+  db = nullptr;
+
+  // Now try opening read-write and make sure it fails, for safety.
+  ASSERT_TRUE(DB::Open(opts, name, &db).IsIOError());
+}
+
+TEST_F(BackupEngineTest, ProgressCallbackDuringBackup) {
+  DestroyDBWithoutCheck(dbname_, options_);
+  // Too big for this small DB
+  engine_options_->callback_trigger_interval_size = 100000;
+  OpenDBAndBackupEngine(true);
+  FillDB(db_.get(), 0, 100);
+  bool is_callback_invoked = false;
+  ASSERT_OK(backup_engine_->CreateNewBackup(
+      db_.get(), true,
+      [&is_callback_invoked]() { is_callback_invoked = true; }));
+  ASSERT_FALSE(is_callback_invoked);
+  CloseBackupEngine();
+
+  // Easily small enough for this small DB
+  engine_options_->callback_trigger_interval_size = 1000;
+  OpenBackupEngine();
+  ASSERT_OK(backup_engine_->CreateNewBackup(
+      db_.get(), true,
+      [&is_callback_invoked]() { is_callback_invoked = true; }));
+  ASSERT_TRUE(is_callback_invoked);
+  CloseDBAndBackupEngine();
+  DestroyDBWithoutCheck(dbname_, options_);
+}
+
+TEST_F(BackupEngineTest, GarbageCollectionBeforeBackup) {
+  DestroyDBWithoutCheck(dbname_, options_);
+  OpenDBAndBackupEngine(true);
+
+  ASSERT_OK(backup_chroot_env_->CreateDirIfMissing(backupdir_ + "/shared"));
+  std::string file_five = backupdir_ + "/shared/000009.sst";
+  std::string file_five_contents = "I'm not really a sst file";
+  // this depends on the fact that 00009.sst is the first file created by the DB
+  ASSERT_OK(file_manager_->WriteToFile(file_five, file_five_contents));
+
+  FillDB(db_.get(), 0, 100);
+  // backup overwrites file 000009.sst
+  ASSERT_OK(backup_engine_->CreateNewBackup(db_.get(), true));
+
+  std::string new_file_five_contents;
+  ASSERT_OK(ReadFileToString(backup_chroot_env_.get(), file_five,
+                             &new_file_five_contents));
+  // file 000009.sst was overwritten
+  ASSERT_TRUE(new_file_five_contents != file_five_contents);
+
+  CloseDBAndBackupEngine();
+
+  AssertBackupConsistency(0, 0, 100);
+}
+
+// Test that we properly propagate Env failures
+TEST_F(BackupEngineTest, EnvFailures) {
+  BackupEngine* backup_engine;
+
+  // get children failure
+  {
+    test_backup_fs_->SetGetChildrenFailure(true);
+    ASSERT_NOK(BackupEngine::Open(test_db_env_.get(), *engine_options_,
+                                  &backup_engine));
+    test_backup_fs_->SetGetChildrenFailure(false);
+  }
+
+  // created dir failure
+  {
+    test_backup_fs_->SetCreateDirIfMissingFailure(true);
+    ASSERT_NOK(BackupEngine::Open(test_db_env_.get(), *engine_options_,
+                                  &backup_engine));
+    test_backup_fs_->SetCreateDirIfMissingFailure(false);
+  }
+
+  // new directory failure
+  {
+    test_backup_fs_->SetNewDirectoryFailure(true);
+    ASSERT_NOK(BackupEngine::Open(test_db_env_.get(), *engine_options_,
+                                  &backup_engine));
+    test_backup_fs_->SetNewDirectoryFailure(false);
+  }
+
+  // Read from meta-file failure
+  {
+    DestroyDBWithoutCheck(dbname_, options_);
+    OpenDBAndBackupEngine(true);
+    FillDB(db_.get(), 0, 100);
+    ASSERT_OK(backup_engine_->CreateNewBackup(db_.get(), true));
+    CloseDBAndBackupEngine();
+    test_backup_fs_->SetDummySequentialFile(true);
+    test_backup_fs_->SetDummySequentialFileFailReads(true);
+    engine_options_->destroy_old_data = false;
+    ASSERT_NOK(BackupEngine::Open(test_db_env_.get(), *engine_options_,
+                                  &backup_engine));
+    test_backup_fs_->SetDummySequentialFile(false);
+    test_backup_fs_->SetDummySequentialFileFailReads(false);
+  }
+
+  // no failure
+  {
+    ASSERT_OK(BackupEngine::Open(test_db_env_.get(), *engine_options_,
+                                 &backup_engine));
+    delete backup_engine;
+  }
+}
+
+// Verify manifest can roll while a backup is being created with the old
+// manifest.
+TEST_F(BackupEngineTest, ChangeManifestDuringBackupCreation) {
+  DestroyDBWithoutCheck(dbname_, options_);
+  options_.max_manifest_file_size = 0;  // always rollover manifest for file add
+  OpenDBAndBackupEngine(true);
+  FillDB(db_.get(), 0, 100, kAutoFlushOnly);
+
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->LoadDependency({
+      {"CheckpointImpl::CreateCheckpoint:SavedLiveFiles1",
+       "VersionSet::LogAndApply:WriteManifest"},
+      {"VersionSet::LogAndApply:WriteManifestDone",
+       "CheckpointImpl::CreateCheckpoint:SavedLiveFiles2"},
+  });
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->EnableProcessing();
+
+  ROCKSDB_NAMESPACE::port::Thread flush_thread{
+      [this]() { ASSERT_OK(db_->Flush(FlushOptions())); }};
+
+  ASSERT_OK(backup_engine_->CreateNewBackup(db_.get(), false));
+
+  flush_thread.join();
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->DisableProcessing();
+
+  // The last manifest roll would've already been cleaned up by the full scan
+  // that happens when CreateNewBackup invokes EnableFileDeletions. We need to
+  // trigger another roll to verify non-full scan purges stale manifests.
+  DBImpl* db_impl = static_cast_with_check<DBImpl>(db_.get());
+  std::string prev_manifest_path =
+      DescriptorFileName(dbname_, db_impl->TEST_Current_Manifest_FileNo());
+  FillDB(db_.get(), 0, 100, kAutoFlushOnly);
+  ASSERT_OK(db_chroot_env_->FileExists(prev_manifest_path));
+  ASSERT_OK(db_->Flush(FlushOptions()));
+  // Even though manual flush completed above, the background thread may not
+  // have finished its cleanup work. `TEST_WaitForBackgroundWork()` will wait
+  // until all the background thread's work has completed, including cleanup.
+  ASSERT_OK(db_impl->TEST_WaitForBackgroundWork());
+  ASSERT_TRUE(db_chroot_env_->FileExists(prev_manifest_path).IsNotFound());
+
+  CloseDBAndBackupEngine();
+  DestroyDBWithoutCheck(dbname_, options_);
+  AssertBackupConsistency(0, 0, 100);
+}
+
+// see https://github.com/facebook/rocksdb/issues/921
+TEST_F(BackupEngineTest, Issue921Test) {
+  BackupEngine* backup_engine;
+  engine_options_->share_table_files = false;
+  ASSERT_OK(
+      backup_chroot_env_->CreateDirIfMissing(engine_options_->backup_dir));
+  engine_options_->backup_dir += "/new_dir";
+  ASSERT_OK(BackupEngine::Open(backup_chroot_env_.get(), *engine_options_,
+                               &backup_engine));
+
+  delete backup_engine;
+}
+
+TEST_F(BackupEngineTest, BackupWithMetadata) {
+  const int keys_iteration = 5000;
+  OpenDBAndBackupEngine(true);
+  // create five backups
+  for (int i = 0; i < 5; ++i) {
+    const std::string metadata = std::to_string(i);
+    FillDB(db_.get(), keys_iteration * i, keys_iteration * (i + 1));
+    // Here also test CreateNewBackupWithMetadata with CreateBackupOptions
+    // and outputting saved BackupID.
+    CreateBackupOptions opts;
+    opts.flush_before_backup = true;
+    BackupID new_id = 0;
+    ASSERT_OK(backup_engine_->CreateNewBackupWithMetadata(opts, db_.get(),
+                                                          metadata, &new_id));
+    ASSERT_EQ(new_id, static_cast<BackupID>(i + 1));
+  }
+  CloseDBAndBackupEngine();
+
+  OpenDBAndBackupEngine();
+  {  // Verify in bulk BackupInfo
+    std::vector<BackupInfo> backup_infos;
+    backup_engine_->GetBackupInfo(&backup_infos);
+    ASSERT_EQ(5, backup_infos.size());
+    for (int i = 0; i < 5; i++) {
+      ASSERT_EQ(std::to_string(i), backup_infos[i].app_metadata);
+    }
+  }
+  // Also verify in individual BackupInfo
+  for (int i = 0; i < 5; i++) {
+    BackupInfo backup_info;
+    ASSERT_OK(backup_engine_->GetBackupInfo(static_cast<BackupID>(i + 1),
+                                            &backup_info));
+    ASSERT_EQ(std::to_string(i), backup_info.app_metadata);
+  }
+  CloseDBAndBackupEngine();
+  DestroyDBWithoutCheck(dbname_, options_);
+}
+
+TEST_F(BackupEngineTest, BinaryMetadata) {
+  OpenDBAndBackupEngine(true);
+  std::string binaryMetadata = "abc\ndef";
+  binaryMetadata.push_back('\0');
+  binaryMetadata.append("ghi");
+  ASSERT_OK(
+      backup_engine_->CreateNewBackupWithMetadata(db_.get(), binaryMetadata));
+  CloseDBAndBackupEngine();
+
+  OpenDBAndBackupEngine();
+  std::vector<BackupInfo> backup_infos;
+  backup_engine_->GetBackupInfo(&backup_infos);
+  ASSERT_EQ(1, backup_infos.size());
+  ASSERT_EQ(binaryMetadata, backup_infos[0].app_metadata);
+  CloseDBAndBackupEngine();
+  DestroyDBWithoutCheck(dbname_, options_);
+}
+
+TEST_F(BackupEngineTest, MetadataTooLarge) {
+  OpenDBAndBackupEngine(true);
+  std::string largeMetadata(1024 * 1024 + 1, 0);
+  ASSERT_NOK(
+      backup_engine_->CreateNewBackupWithMetadata(db_.get(), largeMetadata));
+  CloseDBAndBackupEngine();
+  DestroyDBWithoutCheck(dbname_, options_);
+}
+
+TEST_F(BackupEngineTest, MetaSchemaVersion2_SizeCorruption) {
+  engine_options_->schema_version = 1;
+  OpenDBAndBackupEngine(/*destroy_old_data*/ true);
+
+  // Backup 1: no future schema, no sizes, with checksums
+  ASSERT_OK(backup_engine_->CreateNewBackup(db_.get()));
+
+  CloseDBAndBackupEngine();
+  engine_options_->schema_version = 2;
+  OpenDBAndBackupEngine(/*destroy_old_data*/ false);
+
+  // Backup 2: no checksums, no sizes
+  TEST_BackupMetaSchemaOptions test_opts;
+  test_opts.crc32c_checksums = false;
+  test_opts.file_sizes = false;
+  TEST_SetBackupMetaSchemaOptions(backup_engine_.get(), test_opts);
+  ASSERT_OK(backup_engine_->CreateNewBackup(db_.get()));
+
+  // Backup 3: no checksums, with sizes
+  test_opts.file_sizes = true;
+  TEST_SetBackupMetaSchemaOptions(backup_engine_.get(), test_opts);
+  ASSERT_OK(backup_engine_->CreateNewBackup(db_.get()));
+
+  // Backup 4: with checksums and sizes
+  test_opts.crc32c_checksums = true;
+  TEST_SetBackupMetaSchemaOptions(backup_engine_.get(), test_opts);
+  ASSERT_OK(backup_engine_->CreateNewBackup(db_.get()));
+
+  CloseDBAndBackupEngine();
+
+  // Corrupt all the CURRENT files with the wrong size
+  const std::string private_dir = backupdir_ + "/private";
+
+  for (int id = 1; id <= 3; ++id) {
+    ASSERT_OK(file_manager_->WriteToFile(
+        private_dir + "/" + std::to_string(id) + "/CURRENT", "x"));
+  }
+  // Except corrupt Backup 4 with same size CURRENT file
+  {
+    uint64_t size = 0;
+    ASSERT_OK(test_backup_env_->GetFileSize(private_dir + "/4/CURRENT", &size));
+    ASSERT_OK(file_manager_->WriteToFile(private_dir + "/4/CURRENT",
+                                         std::string(size, 'x')));
+  }
+
+  OpenBackupEngine();
+
+  // Only the one with sizes in metadata will be immediately detected
+  // as corrupt
+  std::vector<BackupID> corrupted;
+  backup_engine_->GetCorruptedBackups(&corrupted);
+  ASSERT_EQ(corrupted.size(), 1);
+  ASSERT_EQ(corrupted[0], 3);
+
+  // Size corruption detected on Restore with checksum
+  ASSERT_TRUE(backup_engine_->RestoreDBFromBackup(1 /*id*/, dbname_, dbname_)
+                  .IsCorruption());
+
+  // Size corruption not detected without checksums nor sizes
+  ASSERT_OK(backup_engine_->RestoreDBFromBackup(2 /*id*/, dbname_, dbname_));
+
+  // Non-size corruption detected on Restore with checksum
+  ASSERT_TRUE(backup_engine_->RestoreDBFromBackup(4 /*id*/, dbname_, dbname_)
+                  .IsCorruption());
+
+  CloseBackupEngine();
+}
+
+TEST_F(BackupEngineTest, MetaSchemaVersion2_NotSupported) {
+  engine_options_->schema_version = 2;
+  TEST_BackupMetaSchemaOptions test_opts;
+  std::string app_metadata = "abc\ndef";
+
+  OpenDBAndBackupEngine(true);
+  // Start with supported
+  TEST_SetBackupMetaSchemaOptions(backup_engine_.get(), test_opts);
+  ASSERT_OK(
+      backup_engine_->CreateNewBackupWithMetadata(db_.get(), app_metadata));
+
+  // Because we are injecting badness with a TEST API, the badness is only
+  // detected on attempt to restore.
+  // Not supported versions
+  test_opts.version = "3";
+  TEST_SetBackupMetaSchemaOptions(backup_engine_.get(), test_opts);
+  ASSERT_OK(
+      backup_engine_->CreateNewBackupWithMetadata(db_.get(), app_metadata));
+  test_opts.version = "23.45.67";
+  TEST_SetBackupMetaSchemaOptions(backup_engine_.get(), test_opts);
+  ASSERT_OK(
+      backup_engine_->CreateNewBackupWithMetadata(db_.get(), app_metadata));
+  test_opts.version = "2";
+
+  // Non-ignorable fields
+  test_opts.meta_fields["ni::blah"] = "123";
+  TEST_SetBackupMetaSchemaOptions(backup_engine_.get(), test_opts);
+  ASSERT_OK(
+      backup_engine_->CreateNewBackupWithMetadata(db_.get(), app_metadata));
+  test_opts.meta_fields.clear();
+
+  test_opts.file_fields["ni::123"] = "xyz";
+  TEST_SetBackupMetaSchemaOptions(backup_engine_.get(), test_opts);
+  ASSERT_OK(
+      backup_engine_->CreateNewBackupWithMetadata(db_.get(), app_metadata));
+  test_opts.file_fields.clear();
+
+  test_opts.footer_fields["ni::123"] = "xyz";
+  TEST_SetBackupMetaSchemaOptions(backup_engine_.get(), test_opts);
+  ASSERT_OK(
+      backup_engine_->CreateNewBackupWithMetadata(db_.get(), app_metadata));
+  test_opts.footer_fields.clear();
+  CloseDBAndBackupEngine();
+
+  OpenBackupEngine();
+  std::vector<BackupID> corrupted;
+  backup_engine_->GetCorruptedBackups(&corrupted);
+  ASSERT_EQ(corrupted.size(), 5);
+
+  ASSERT_OK(backup_engine_->RestoreDBFromLatestBackup(dbname_, dbname_));
+  CloseBackupEngine();
+}
+
+TEST_F(BackupEngineTest, MetaSchemaVersion2_Restore) {
+  engine_options_->schema_version = 2;
+  TEST_BackupMetaSchemaOptions test_opts;
+  const int keys_iteration = 5000;
+
+  OpenDBAndBackupEngine(true, false, kShareWithChecksum);
+  FillDB(db_.get(), 0, keys_iteration);
+  // Start with minimum metadata to ensure it works without it being filled
+  // based on shared files also in other backups with the metadata.
+  test_opts.crc32c_checksums = false;
+  test_opts.file_sizes = false;
+  TEST_SetBackupMetaSchemaOptions(backup_engine_.get(), test_opts);
+  ASSERT_OK(backup_engine_->CreateNewBackup(db_.get(), true));
+  CloseDBAndBackupEngine();
+
+  AssertBackupConsistency(1 /* id */, 0, keys_iteration, keys_iteration * 2);
+
+  OpenDBAndBackupEngine(false /* destroy_old_data */, false,
+                        kShareWithChecksum);
+  test_opts.file_sizes = true;
+  TEST_SetBackupMetaSchemaOptions(backup_engine_.get(), test_opts);
+  ASSERT_OK(backup_engine_->CreateNewBackup(db_.get(), true));
+  CloseDBAndBackupEngine();
+
+  for (int id = 1; id <= 2; ++id) {
+    AssertBackupConsistency(id, 0, keys_iteration, keys_iteration * 2);
+  }
+
+  OpenDBAndBackupEngine(false /* destroy_old_data */, false,
+                        kShareWithChecksum);
+  test_opts.crc32c_checksums = true;
+  TEST_SetBackupMetaSchemaOptions(backup_engine_.get(), test_opts);
+  ASSERT_OK(backup_engine_->CreateNewBackup(db_.get(), true));
+  CloseDBAndBackupEngine();
+
+  for (int id = 1; id <= 3; ++id) {
+    AssertBackupConsistency(id, 0, keys_iteration, keys_iteration * 2);
+  }
+
+  OpenDBAndBackupEngine(false /* destroy_old_data */, false,
+                        kShareWithChecksum);
+  // No TEST_EnableWriteFutureSchemaVersion2
+  ASSERT_OK(backup_engine_->CreateNewBackup(db_.get(), true));
+  CloseDBAndBackupEngine();
+
+  for (int id = 1; id <= 4; ++id) {
+    AssertBackupConsistency(id, 0, keys_iteration, keys_iteration * 2);
+  }
+
+  OpenDBAndBackupEngine(false /* destroy_old_data */, false,
+                        kShareWithChecksum);
+  // Minor version updates should be forward-compatible
+  test_opts.version = "2.5.70";
+  test_opts.meta_fields["asdf.3456"] = "-42";
+  test_opts.meta_fields["__QRST"] = " 1 $ %%& ";
+  test_opts.file_fields["z94._"] = "^\\";
+  test_opts.file_fields["_7yyyyyyyyy"] = "111111111111";
+  test_opts.footer_fields["Qwzn.tz89"] = "ASDF!!@# ##=\t ";
+  test_opts.footer_fields["yes"] = "no!";
+  TEST_SetBackupMetaSchemaOptions(backup_engine_.get(), test_opts);
+  ASSERT_OK(backup_engine_->CreateNewBackup(db_.get(), true));
+  CloseDBAndBackupEngine();
+
+  for (int id = 1; id <= 5; ++id) {
+    AssertBackupConsistency(id, 0, keys_iteration, keys_iteration * 2);
+  }
+}
+
+TEST_F(BackupEngineTest, Concurrency) {
+  // Check that we can simultaneously:
+  // * Run several read operations in different threads on a single
+  // BackupEngine object, and
+  // * With another BackupEngine object on the same
+  // backup_dir, run the same read operations in another thread, and
+  // * With yet another BackupEngine object on the same
+  // backup_dir, create two new backups in parallel threads.
+  //
+  // Because of the challenges of integrating this into db_stress,
+  // this is a non-deterministic mini-stress test here instead.
+
+  // To check for a race condition in handling buffer size based on byte
+  // burst limit, we need a (generous) rate limiter
+  std::shared_ptr<RateLimiter> limiter{NewGenericRateLimiter(1000000000)};
+  engine_options_->backup_rate_limiter = limiter;
+  engine_options_->restore_rate_limiter = limiter;
+
+  OpenDBAndBackupEngine(true, false, kShareWithChecksum);
+
+  static constexpr int keys_iteration = 5000;
+  FillDB(db_.get(), 0, keys_iteration);
+  ASSERT_OK(backup_engine_->CreateNewBackup(db_.get()));
+
+  FillDB(db_.get(), keys_iteration, 2 * keys_iteration);
+  ASSERT_OK(backup_engine_->CreateNewBackup(db_.get()));
+
+  static constexpr int max_factor = 3;
+  FillDB(db_.get(), 2 * keys_iteration, max_factor * keys_iteration);
+  // will create another backup soon...
+
+  Options db_opts = options_;
+  db_opts.wal_dir = "";
+  db_opts.create_if_missing = false;
+  BackupEngineOptions be_opts = *engine_options_;
+  be_opts.destroy_old_data = false;
+
+  std::mt19937 rng{std::random_device()()};
+
+  std::array<std::thread, 4> read_threads;
+  std::array<std::thread, 4> restore_verify_threads;
+  for (uint32_t i = 0; i < read_threads.size(); ++i) {
+    uint32_t sleep_micros = rng() % 100000;
+    read_threads[i] = std::thread([this, i, sleep_micros, &db_opts, &be_opts,
+                                   &restore_verify_threads, &limiter] {
+      test_db_env_->SleepForMicroseconds(sleep_micros);
+
+      // Whether to also re-open the BackupEngine, potentially seeing
+      // additional backups
+      bool reopen = i == 3;
+      // Whether we are going to restore "latest"
+      bool latest = i > 1;
+
+      BackupEngine* my_be;
+      if (reopen) {
+        ASSERT_OK(BackupEngine::Open(test_db_env_.get(), be_opts, &my_be));
+      } else {
+        my_be = backup_engine_.get();
+      }
+
+      // Verify metadata (we don't receive updates from concurrently
+      // creating a new backup)
+      std::vector<BackupInfo> infos;
+      my_be->GetBackupInfo(&infos);
+      const uint32_t count = static_cast<uint32_t>(infos.size());
+      infos.clear();
+      if (reopen) {
+        ASSERT_GE(count, 2U);
+        ASSERT_LE(count, 4U);
+        fprintf(stderr, "Reopen saw %u backups\n", count);
+      } else {
+        ASSERT_EQ(count, 2U);
+      }
+      std::vector<BackupID> ids;
+      my_be->GetCorruptedBackups(&ids);
+      ASSERT_EQ(ids.size(), 0U);
+
+      // (Eventually, see below) Restore one of the backups, or "latest"
+      std::string restore_db_dir = dbname_ + "/restore" + std::to_string(i);
+      DestroyDir(test_db_env_.get(), restore_db_dir).PermitUncheckedError();
+      BackupID to_restore;
+      if (latest) {
+        to_restore = count;
+      } else {
+        to_restore = i + 1;
+      }
+
+      // Open restored DB to verify its contents, but test atomic restore
+      // by doing it async and ensuring we either get OK or InvalidArgument
+      restore_verify_threads[i] =
+          std::thread([this, &db_opts, restore_db_dir, to_restore] {
+            DB* restored;
+            Status s;
+            for (;;) {
+              s = DB::Open(db_opts, restore_db_dir, &restored);
+              if (s.IsInvalidArgument()) {
+                // Restore hasn't finished
+                test_db_env_->SleepForMicroseconds(1000);
+                continue;
+              } else {
+                // We should only get InvalidArgument if restore is
+                // incomplete, or OK if complete
+                ASSERT_OK(s);
+                break;
+              }
+            }
+            int factor = std::min(static_cast<int>(to_restore), max_factor);
+            AssertExists(restored, 0, factor * keys_iteration);
+            AssertEmpty(restored, factor * keys_iteration,
+                        (factor + 1) * keys_iteration);
+            delete restored;
+          });
+
+      // (Ok now) Restore one of the backups, or "latest"
+      if (latest) {
+        ASSERT_OK(
+            my_be->RestoreDBFromLatestBackup(restore_db_dir, restore_db_dir));
+      } else {
+        ASSERT_OK(my_be->VerifyBackup(to_restore, true));
+        ASSERT_OK(my_be->RestoreDBFromBackup(to_restore, restore_db_dir,
+                                             restore_db_dir));
+      }
+
+      // Test for race condition in reconfiguring limiter
+      // FIXME: this could set to a different value in all threads, except
+      // GenericRateLimiter::SetBytesPerSecond has a write-write race
+      // reported by TSAN
+      if (i == 0) {
+        limiter->SetBytesPerSecond(2000000000);
+      }
+
+      // Re-verify metadata (we don't receive updates from concurrently
+      // creating a new backup)
+      my_be->GetBackupInfo(&infos);
+      ASSERT_EQ(infos.size(), count);
+      my_be->GetCorruptedBackups(&ids);
+      ASSERT_EQ(ids.size(), 0);
+      // fprintf(stderr, "Finished read thread\n");
+
+      if (reopen) {
+        delete my_be;
+      }
+    });
+  }
+
+  BackupEngine* alt_be;
+  ASSERT_OK(BackupEngine::Open(test_db_env_.get(), be_opts, &alt_be));
+
+  std::array<std::thread, 2> append_threads;
+  for (unsigned i = 0; i < append_threads.size(); ++i) {
+    uint32_t sleep_micros = rng() % 100000;
+    append_threads[i] = std::thread([this, sleep_micros, alt_be] {
+      test_db_env_->SleepForMicroseconds(sleep_micros);
+      // WART: CreateNewBackup doesn't tell you the BackupID it just created,
+      // which is ugly for multithreaded setting.
+      // TODO: add delete backup also when that is added
+      ASSERT_OK(alt_be->CreateNewBackup(db_.get()));
+      // fprintf(stderr, "Finished append thread\n");
+    });
+  }
+
+  for (auto& t : append_threads) {
+    t.join();
+  }
+  // Verify metadata
+  std::vector<BackupInfo> infos;
+  alt_be->GetBackupInfo(&infos);
+  ASSERT_EQ(infos.size(), 2 + append_threads.size());
+
+  for (auto& t : read_threads) {
+    t.join();
+  }
+
+  delete alt_be;
+
+  for (auto& t : restore_verify_threads) {
+    t.join();
+  }
+
+  CloseDBAndBackupEngine();
+}
+
+TEST_F(BackupEngineTest, LimitBackupsOpened) {
+  // Verify the specified max backups are opened, including skipping over
+  // corrupted backups.
+  //
+  // Setup:
+  // - backups 1, 2, and 4 are valid
+  // - backup 3 is corrupt
+  // - max_valid_backups_to_open == 2
+  //
+  // Expectation: the engine opens backups 4 and 2 since those are latest two
+  // non-corrupt backups.
+  const int kNumKeys = 5000;
+  OpenDBAndBackupEngine(true);
+  for (int i = 1; i <= 4; ++i) {
+    FillDB(db_.get(), kNumKeys * i, kNumKeys * (i + 1));
+    ASSERT_OK(backup_engine_->CreateNewBackup(db_.get(), true));
+    if (i == 3) {
+      ASSERT_OK(file_manager_->CorruptFile(backupdir_ + "/meta/3", 3));
+    }
+  }
+  CloseDBAndBackupEngine();
+
+  engine_options_->max_valid_backups_to_open = 2;
+  engine_options_->destroy_old_data = false;
+  BackupEngineReadOnly* read_only_backup_engine;
+  ASSERT_OK(BackupEngineReadOnly::Open(
+      backup_chroot_env_.get(), *engine_options_, &read_only_backup_engine));
+
+  std::vector<BackupInfo> backup_infos;
+  read_only_backup_engine->GetBackupInfo(&backup_infos);
+  ASSERT_EQ(2, backup_infos.size());
+  ASSERT_EQ(2, backup_infos[0].backup_id);
+  ASSERT_EQ(4, backup_infos[1].backup_id);
+  delete read_only_backup_engine;
+}
+
+TEST_F(BackupEngineTest, IgnoreLimitBackupsOpenedWhenNotReadOnly) {
+  // Verify the specified max_valid_backups_to_open is ignored if the engine
+  // is not read-only.
+  //
+  // Setup:
+  // - backups 1, 2, and 4 are valid
+  // - backup 3 is corrupt
+  // - max_valid_backups_to_open == 2
+  //
+  // Expectation: the engine opens backups 4, 2, and 1 since those are latest
+  // non-corrupt backups, by ignoring max_valid_backups_to_open == 2.
+  const int kNumKeys = 5000;
+  OpenDBAndBackupEngine(true);
+  for (int i = 1; i <= 4; ++i) {
+    FillDB(db_.get(), kNumKeys * i, kNumKeys * (i + 1));
+    ASSERT_OK(backup_engine_->CreateNewBackup(db_.get(), true));
+    if (i == 3) {
+      ASSERT_OK(file_manager_->CorruptFile(backupdir_ + "/meta/3", 3));
+    }
+  }
+  CloseDBAndBackupEngine();
+
+  engine_options_->max_valid_backups_to_open = 2;
+  OpenDBAndBackupEngine();
+  std::vector<BackupInfo> backup_infos;
+  backup_engine_->GetBackupInfo(&backup_infos);
+  ASSERT_EQ(3, backup_infos.size());
+  ASSERT_EQ(1, backup_infos[0].backup_id);
+  ASSERT_EQ(2, backup_infos[1].backup_id);
+  ASSERT_EQ(4, backup_infos[2].backup_id);
+  CloseDBAndBackupEngine();
+  DestroyDBWithoutCheck(dbname_, options_);
+}
+
+TEST_F(BackupEngineTest, CreateWhenLatestBackupCorrupted) {
+  // we should pick an ID greater than corrupted backups' IDs so creation can
+  // succeed even when latest backup is corrupted.
+  const int kNumKeys = 5000;
+  OpenDBAndBackupEngine(true /* destroy_old_data */);
+  BackupInfo backup_info;
+  ASSERT_TRUE(backup_engine_->GetLatestBackupInfo(&backup_info).IsNotFound());
+  FillDB(db_.get(), 0 /* from */, kNumKeys);
+  ASSERT_OK(backup_engine_->CreateNewBackup(db_.get(),
+                                            true /* flush_before_backup */));
+  ASSERT_OK(file_manager_->CorruptFile(backupdir_ + "/meta/1",
+                                       3 /* bytes_to_corrupt */));
+  CloseDBAndBackupEngine();
+
+  OpenDBAndBackupEngine();
+  ASSERT_TRUE(backup_engine_->GetLatestBackupInfo(&backup_info).IsNotFound());
+
+  ASSERT_OK(backup_engine_->CreateNewBackup(db_.get(),
+                                            true /* flush_before_backup */));
+
+  ASSERT_TRUE(backup_engine_->GetLatestBackupInfo(&backup_info).ok());
+  ASSERT_EQ(2, backup_info.backup_id);
+
+  std::vector<BackupInfo> backup_infos;
+  backup_engine_->GetBackupInfo(&backup_infos);
+  ASSERT_EQ(1, backup_infos.size());
+  ASSERT_EQ(2, backup_infos[0].backup_id);
+
+  // Verify individual GetBackupInfo by ID
+  ASSERT_TRUE(backup_engine_->GetBackupInfo(0U, &backup_info).IsNotFound());
+  ASSERT_TRUE(backup_engine_->GetBackupInfo(1U, &backup_info).IsCorruption());
+  ASSERT_TRUE(backup_engine_->GetBackupInfo(2U, &backup_info).ok());
+  ASSERT_TRUE(backup_engine_->GetBackupInfo(3U, &backup_info).IsNotFound());
+  ASSERT_TRUE(
+      backup_engine_->GetBackupInfo(999999U, &backup_info).IsNotFound());
+}
+
+TEST_F(BackupEngineTest, WriteOnlyEngineNoSharedFileDeletion) {
+  // Verifies a write-only BackupEngine does not delete files belonging to valid
+  // backups when GarbageCollect, PurgeOldBackups, or DeleteBackup are called.
+  const int kNumKeys = 5000;
+  for (int i = 0; i < 3; ++i) {
+    OpenDBAndBackupEngine(i == 0 /* destroy_old_data */);
+    FillDB(db_.get(), i * kNumKeys, (i + 1) * kNumKeys);
+    ASSERT_OK(backup_engine_->CreateNewBackup(db_.get(), true));
+    CloseDBAndBackupEngine();
+
+    engine_options_->max_valid_backups_to_open = 0;
+    OpenDBAndBackupEngine();
+    switch (i) {
+      case 0:
+        ASSERT_OK(backup_engine_->GarbageCollect());
+        break;
+      case 1:
+        ASSERT_OK(backup_engine_->PurgeOldBackups(1 /* num_backups_to_keep */));
+        break;
+      case 2:
+        ASSERT_OK(backup_engine_->DeleteBackup(2 /* backup_id */));
+        break;
+      default:
+        assert(false);
+    }
+    CloseDBAndBackupEngine();
+
+    engine_options_->max_valid_backups_to_open =
+        std::numeric_limits<int32_t>::max();
+    AssertBackupConsistency(i + 1, 0, (i + 1) * kNumKeys);
+  }
+}
+
+TEST_P(BackupEngineTestWithParam, BackupUsingDirectIO) {
+  // Tests direct I/O on the backup engine's reads and writes on the DB env and
+  // backup env
+  // We use ChrootEnv underneath so the below line checks for direct I/O support
+  // in the chroot directory, not the true filesystem root.
+  if (!test::IsDirectIOSupported(test_db_env_.get(), "/")) {
+    ROCKSDB_GTEST_SKIP("Test requires Direct I/O Support");
+    return;
+  }
+  const int kNumKeysPerBackup = 100;
+  const int kNumBackups = 3;
+  options_.use_direct_reads = true;
+  OpenDBAndBackupEngine(true /* destroy_old_data */);
+  for (int i = 0; i < kNumBackups; ++i) {
+    FillDB(db_.get(), i * kNumKeysPerBackup /* from */,
+           (i + 1) * kNumKeysPerBackup /* to */, kFlushAll);
+
+    // Clear the file open counters and then do a bunch of backup engine ops.
+    // For all ops, files should be opened in direct mode.
+    test_backup_fs_->ClearFileOpenCounters();
+    test_db_fs_->ClearFileOpenCounters();
+    CloseBackupEngine();
+    OpenBackupEngine();
+    ASSERT_OK(backup_engine_->CreateNewBackup(db_.get(),
+                                              false /* flush_before_backup */));
+    ASSERT_OK(backup_engine_->VerifyBackup(i + 1));
+    CloseBackupEngine();
+    OpenBackupEngine();
+    std::vector<BackupInfo> backup_infos;
+    backup_engine_->GetBackupInfo(&backup_infos);
+    ASSERT_EQ(static_cast<size_t>(i + 1), backup_infos.size());
+
+    // Verify backup engine always opened files with direct I/O
+    ASSERT_EQ(0, test_db_fs_->num_writers());
+    ASSERT_GE(test_db_fs_->num_direct_rand_readers(), 0);
+    ASSERT_GT(test_db_fs_->num_direct_seq_readers(), 0);
+    // Currently the DB doesn't support reading WALs or manifest with direct
+    // I/O, so subtract two.
+    ASSERT_EQ(test_db_fs_->num_seq_readers() - 2,
+              test_db_fs_->num_direct_seq_readers());
+    ASSERT_EQ(test_db_fs_->num_rand_readers(),
+              test_db_fs_->num_direct_rand_readers());
+  }
+  CloseDBAndBackupEngine();
+
+  for (int i = 0; i < kNumBackups; ++i) {
+    AssertBackupConsistency(i + 1 /* backup_id */,
+                            i * kNumKeysPerBackup /* start_exist */,
+                            (i + 1) * kNumKeysPerBackup /* end_exist */,
+                            (i + 2) * kNumKeysPerBackup /* end */);
+  }
+}
+
+TEST_F(BackupEngineTest, BackgroundThreadCpuPriority) {
+  std::atomic<CpuPriority> priority(CpuPriority::kNormal);
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
+      "BackupEngineImpl::Initialize:SetCpuPriority", [&](void* new_priority) {
+        priority.store(*reinterpret_cast<CpuPriority*>(new_priority));
+      });
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->EnableProcessing();
+
+  // 1 thread is easier to test, otherwise, we may not be sure which thread
+  // actually does the work during CreateNewBackup.
+  engine_options_->max_background_operations = 1;
+  OpenDBAndBackupEngine(true);
+
+  {
+    FillDB(db_.get(), 0, 100);
+
+    // by default, cpu priority is not changed.
+    CreateBackupOptions options;
+    ASSERT_OK(backup_engine_->CreateNewBackup(options, db_.get()));
+
+    ASSERT_EQ(priority, CpuPriority::kNormal);
+  }
+
+  {
+    FillDB(db_.get(), 101, 200);
+
+    // decrease cpu priority from normal to low.
+    CreateBackupOptions options;
+    options.decrease_background_thread_cpu_priority = true;
+    options.background_thread_cpu_priority = CpuPriority::kLow;
+    ASSERT_OK(backup_engine_->CreateNewBackup(options, db_.get()));
+
+    ASSERT_EQ(priority, CpuPriority::kLow);
+  }
+
+  {
+    FillDB(db_.get(), 201, 300);
+
+    // try to upgrade cpu priority back to normal,
+    // the priority should still low.
+    CreateBackupOptions options;
+    options.decrease_background_thread_cpu_priority = true;
+    options.background_thread_cpu_priority = CpuPriority::kNormal;
+    ASSERT_OK(backup_engine_->CreateNewBackup(options, db_.get()));
+
+    ASSERT_EQ(priority, CpuPriority::kLow);
+  }
+
+  {
+    FillDB(db_.get(), 301, 400);
+
+    // decrease cpu priority from low to idle.
+    CreateBackupOptions options;
+    options.decrease_background_thread_cpu_priority = true;
+    options.background_thread_cpu_priority = CpuPriority::kIdle;
+    ASSERT_OK(backup_engine_->CreateNewBackup(options, db_.get()));
+
+    ASSERT_EQ(priority, CpuPriority::kIdle);
+  }
+
+  {
+    FillDB(db_.get(), 301, 400);
+
+    // reset priority to later verify that it's not updated by SetCpuPriority.
+    priority = CpuPriority::kNormal;
+
+    // setting the same cpu priority won't call SetCpuPriority.
+    CreateBackupOptions options;
+    options.decrease_background_thread_cpu_priority = true;
+    options.background_thread_cpu_priority = CpuPriority::kIdle;
+
+    // Also check output backup_id with CreateNewBackup
+    BackupID new_id = 0;
+    ASSERT_OK(backup_engine_->CreateNewBackup(options, db_.get(), &new_id));
+    ASSERT_EQ(new_id, 5U);
+
+    ASSERT_EQ(priority, CpuPriority::kNormal);
+  }
+
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->DisableProcessing();
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->ClearAllCallBacks();
+  CloseDBAndBackupEngine();
+  DestroyDBWithoutCheck(dbname_, options_);
+}
+
+// Populates `*total_size` with the size of all files under `backup_dir`.
+// We don't go through `BackupEngine` currently because it's hard to figure out
+// the metadata file size.
+Status GetSizeOfBackupFiles(FileSystem* backup_fs,
+                            const std::string& backup_dir, size_t* total_size) {
+  *total_size = 0;
+  std::vector<std::string> dir_stack = {backup_dir};
+  Status s;
+  while (s.ok() && !dir_stack.empty()) {
+    std::string dir = std::move(dir_stack.back());
+    dir_stack.pop_back();
+    std::vector<std::string> children;
+    s = backup_fs->GetChildren(dir, IOOptions(), &children, nullptr /* dbg */);
+    for (size_t i = 0; s.ok() && i < children.size(); ++i) {
+      std::string path = dir + "/" + children[i];
+      bool is_dir;
+      s = backup_fs->IsDirectory(path, IOOptions(), &is_dir, nullptr /* dbg */);
+      uint64_t file_size = 0;
+      if (s.ok()) {
+        if (is_dir) {
+          dir_stack.emplace_back(std::move(path));
+        } else {
+          s = backup_fs->GetFileSize(path, IOOptions(), &file_size,
+                                     nullptr /* dbg */);
+        }
+      }
+      if (s.ok()) {
+        *total_size += file_size;
+      }
+    }
+  }
+  return s;
+}
+
+TEST_F(BackupEngineTest, IOStats) {
+  // Tests the `BACKUP_READ_BYTES` and `BACKUP_WRITE_BYTES` ticker stats have
+  // the expected values according to the files in the backups.
+
+  // These ticker stats are expected to be populated regardless of `PerfLevel`
+  // in user thread
+  SetPerfLevel(kDisable);
+
+  options_.statistics = CreateDBStatistics();
+  OpenDBAndBackupEngine(true /* destroy_old_data */, false /* dummy */,
+                        kShareWithChecksum);
+
+  FillDB(db_.get(), 0 /* from */, 100 /* to */, kFlushMost);
+
+  ASSERT_EQ(0, options_.statistics->getTickerCount(BACKUP_READ_BYTES));
+  ASSERT_EQ(0, options_.statistics->getTickerCount(BACKUP_WRITE_BYTES));
+  ASSERT_OK(backup_engine_->CreateNewBackup(db_.get(),
+                                            false /* flush_before_backup */));
+
+  size_t orig_backup_files_size;
+  ASSERT_OK(GetSizeOfBackupFiles(test_backup_env_->GetFileSystem().get(),
+                                 backupdir_, &orig_backup_files_size));
+  size_t expected_bytes_written = orig_backup_files_size;
+  ASSERT_EQ(expected_bytes_written,
+            options_.statistics->getTickerCount(BACKUP_WRITE_BYTES));
+  // Bytes read is more difficult to pin down since there are reads for many
+  // purposes other than creating file, like `GetSortedWalFiles()` to find first
+  // sequence number, or `CreateNewBackup()` thread to find SST file session ID.
+  // So we loosely require there are at least as many reads as needed for
+  // copying, but not as many as twice that.
+  ASSERT_GE(options_.statistics->getTickerCount(BACKUP_READ_BYTES),
+            expected_bytes_written);
+  ASSERT_LT(expected_bytes_written,
+            2 * options_.statistics->getTickerCount(BACKUP_READ_BYTES));
+
+  FillDB(db_.get(), 100 /* from */, 200 /* to */, kFlushMost);
+
+  ASSERT_OK(options_.statistics->Reset());
+  ASSERT_OK(backup_engine_->CreateNewBackup(db_.get(),
+                                            false /* flush_before_backup */));
+  size_t final_backup_files_size;
+  ASSERT_OK(GetSizeOfBackupFiles(test_backup_env_->GetFileSystem().get(),
+                                 backupdir_, &final_backup_files_size));
+  expected_bytes_written = final_backup_files_size - orig_backup_files_size;
+  ASSERT_EQ(expected_bytes_written,
+            options_.statistics->getTickerCount(BACKUP_WRITE_BYTES));
+  // See above for why these bounds were chosen.
+  ASSERT_GE(options_.statistics->getTickerCount(BACKUP_READ_BYTES),
+            expected_bytes_written);
+  ASSERT_LT(expected_bytes_written,
+            2 * options_.statistics->getTickerCount(BACKUP_READ_BYTES));
+}
+
+TEST_F(BackupEngineTest, FileTemperatures) {
+  CloseDBAndBackupEngine();
+
+  // Required for recording+restoring temperatures
+  engine_options_->schema_version = 2;
+
+  // More file IO instrumentation
+  auto my_db_fs = std::make_shared<FileTemperatureTestFS>(db_chroot_fs_);
+  test_db_fs_ = std::make_shared<TestFs>(my_db_fs);
+  SetEnvsFromFileSystems();
+
+  // Use temperatures
+  options_.bottommost_temperature = Temperature::kWarm;
+  options_.level0_file_num_compaction_trigger = 2;
+  // set dynamic_level to true so the compaction would compact the data to the
+  // last level directly which will have the last_level_temperature
+  options_.level_compaction_dynamic_level_bytes = true;
+
+  OpenDBAndBackupEngine(true /* destroy_old_data */, false /* dummy */,
+                        kShareWithChecksum);
+
+  // generate a bottommost file (combined from 2) and a non-bottommost file
+  DBImpl* dbi = static_cast_with_check<DBImpl>(db_.get());
+  ASSERT_OK(db_->Put(WriteOptions(), "a", "val"));
+  ASSERT_OK(db_->Put(WriteOptions(), "c", "val"));
+  ASSERT_OK(db_->Flush(FlushOptions()));
+  ASSERT_OK(db_->Put(WriteOptions(), "b", "val"));
+  ASSERT_OK(db_->Put(WriteOptions(), "d", "val"));
+  ASSERT_OK(db_->Flush(FlushOptions()));
+  ASSERT_OK(dbi->TEST_WaitForCompact());
+  ASSERT_OK(db_->Put(WriteOptions(), "e", "val"));
+  ASSERT_OK(db_->Flush(FlushOptions()));
+
+  // Get temperatures from manifest
+  std::map<uint64_t, Temperature> manifest_temps;
+  std::map<Temperature, int> manifest_temp_counts;
+  {
+    std::vector<LiveFileStorageInfo> infos;
+    ASSERT_OK(
+        db_->GetLiveFilesStorageInfo(LiveFilesStorageInfoOptions(), &infos));
+    for (auto info : infos) {
+      if (info.file_type == kTableFile) {
+        manifest_temps.emplace(info.file_number, info.temperature);
+        manifest_temp_counts[info.temperature]++;
+      }
+    }
+  }
+
+  // Verify expected manifest temperatures
+  ASSERT_EQ(manifest_temp_counts.size(), 2);
+  ASSERT_EQ(manifest_temp_counts[Temperature::kWarm], 1);
+  ASSERT_EQ(manifest_temp_counts[Temperature::kUnknown], 1);
+
+  // Verify manifest temperatures match FS temperatures
+  std::map<uint64_t, Temperature> current_temps;
+  my_db_fs->CopyCurrentSstFileTemperatures(&current_temps);
+  for (const auto& manifest_temp : manifest_temps) {
+    ASSERT_EQ(current_temps[manifest_temp.first], manifest_temp.second);
+  }
+
+  // Try a few different things
+  for (int i = 1; i <= 5; ++i) {
+    // Expected temperatures after restore are based on manifest temperatures
+    std::map<uint64_t, Temperature> expected_temps = manifest_temps;
+
+    if (i >= 2) {
+      // For iterations 2 & 3, override current temperature of one file
+      // and vary which temperature is authoritative (current or manifest).
+      // For iterations 4 & 5, override current temperature of both files
+      // but make sure an current temperate always takes precedence over
+      // unknown regardless of current_temperatures_override_manifest setting.
+      bool use_current = ((i % 2) == 1);
+      engine_options_->current_temperatures_override_manifest = use_current;
+      CloseBackupEngine();
+      OpenBackupEngine();
+      for (const auto& manifest_temp : manifest_temps) {
+        if (i <= 3) {
+          if (manifest_temp.second == Temperature::kWarm) {
+            my_db_fs->OverrideSstFileTemperature(manifest_temp.first,
+                                                 Temperature::kCold);
+            if (use_current) {
+              expected_temps[manifest_temp.first] = Temperature::kCold;
+            }
+          }
+        } else {
+          assert(i <= 5);
+          if (manifest_temp.second == Temperature::kWarm) {
+            my_db_fs->OverrideSstFileTemperature(manifest_temp.first,
+                                                 Temperature::kUnknown);
+          } else {
+            ASSERT_EQ(manifest_temp.second, Temperature::kUnknown);
+            my_db_fs->OverrideSstFileTemperature(manifest_temp.first,
+                                                 Temperature::kHot);
+            // regardless of use_current
+            expected_temps[manifest_temp.first] = Temperature::kHot;
+          }
+        }
+      }
+    }
+
+    // Sample requested temperatures in opening files for backup
+    my_db_fs->PopRequestedSstFileTemperatures();
+    ASSERT_OK(backup_engine_->CreateNewBackup(db_.get()));
+
+    // Verify requested temperatures against manifest temperatures (before
+    // retry with kUnknown if needed, and before backup finds out current
+    // temperatures in FileSystem)
+    std::vector<std::pair<uint64_t, Temperature>> requested_temps;
+    my_db_fs->PopRequestedSstFileTemperatures(&requested_temps);
+    std::set<uint64_t> distinct_requests;
+    for (const auto& requested_temp : requested_temps) {
+      // Matching manifest temperatures, except allow retry request with
+      // kUnknown
+      auto manifest_temp = manifest_temps.at(requested_temp.first);
+      if (manifest_temp == Temperature::kUnknown ||
+          requested_temp.second != Temperature::kUnknown) {
+        ASSERT_EQ(manifest_temp, requested_temp.second);
+      }
+      distinct_requests.insert(requested_temp.first);
+    }
+    // Two distinct requests
+    ASSERT_EQ(distinct_requests.size(), 2);
+
+    // Verify against backup info file details API
+    BackupInfo info;
+    ASSERT_OK(backup_engine_->GetLatestBackupInfo(
+        &info, /*include_file_details*/ true));
+    ASSERT_GT(info.file_details.size(), 2);
+    for (auto& e : info.file_details) {
+      ASSERT_EQ(expected_temps[e.file_number], e.temperature);
+    }
+
+    // Restore backup to another virtual (tiered) dir
+    const std::string restore_dir = "/restore" + std::to_string(i);
+    ASSERT_OK(backup_engine_->RestoreDBFromLatestBackup(
+        RestoreOptions(), restore_dir, restore_dir));
+
+    // Verify restored FS temperatures match expectation
+    // (FileTemperatureTestFS doesn't distinguish directories when reporting
+    // current temperatures, just whatever SST was written or overridden last
+    // with that file number.)
+    my_db_fs->CopyCurrentSstFileTemperatures(&current_temps);
+    for (const auto& expected_temp : expected_temps) {
+      ASSERT_EQ(current_temps[expected_temp.first], expected_temp.second);
+    }
+
+    // Delete backup to force next backup to copy files
+    ASSERT_OK(backup_engine_->PurgeOldBackups(0));
+  }
+}
+
+}  // namespace
+
+}  // namespace ROCKSDB_NAMESPACE
+
+int main(int argc, char** argv) {
+  ROCKSDB_NAMESPACE::port::InstallStackTraceHandler();
+  ::testing::InitGoogleTest(&argc, argv);
+  return RUN_ALL_TESTS();
+}
+
+#else
+#include <stdio.h>
+
+int main(int /*argc*/, char** /*argv*/) {
+  fprintf(stderr, "SKIPPED as BackupEngine is not supported in ROCKSDB_LITE\n");
+  return 0;
+}
+
+#endif  // !defined(ROCKSDB_LITE) && !defined(OS_WIN)
diff --git a/src/rocksdb/utilities/blob_db/blob_compaction_filter.cc b/src/rocksdb/utilities/blob_db/blob_compaction_filter.cc
new file mode 100644
index 000000000..86907e979
--- /dev/null
+++ b/src/rocksdb/utilities/blob_db/blob_compaction_filter.cc
@@ -0,0 +1,490 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#ifndef ROCKSDB_LITE
+
+#include "utilities/blob_db/blob_compaction_filter.h"
+
+#include <cinttypes>
+
+#include "db/dbformat.h"
+#include "logging/logging.h"
+#include "rocksdb/system_clock.h"
+#include "test_util/sync_point.h"
+
+namespace ROCKSDB_NAMESPACE {
+namespace blob_db {
+
+BlobIndexCompactionFilterBase::~BlobIndexCompactionFilterBase() {
+  if (blob_file_) {
+    CloseAndRegisterNewBlobFile();
+  }
+  RecordTick(statistics_, BLOB_DB_BLOB_INDEX_EXPIRED_COUNT, expired_count_);
+  RecordTick(statistics_, BLOB_DB_BLOB_INDEX_EXPIRED_SIZE, expired_size_);
+  RecordTick(statistics_, BLOB_DB_BLOB_INDEX_EVICTED_COUNT, evicted_count_);
+  RecordTick(statistics_, BLOB_DB_BLOB_INDEX_EVICTED_SIZE, evicted_size_);
+}
+
+CompactionFilter::Decision BlobIndexCompactionFilterBase::FilterV2(
+    int level, const Slice& key, ValueType value_type, const Slice& value,
+    std::string* new_value, std::string* skip_until) const {
+  const CompactionFilter* ucf = user_comp_filter();
+  if (value_type != kBlobIndex) {
+    if (ucf == nullptr) {
+      return Decision::kKeep;
+    }
+    // Apply user compaction filter for inlined data.
+    CompactionFilter::Decision decision =
+        ucf->FilterV2(level, key, value_type, value, new_value, skip_until);
+    if (decision == Decision::kChangeValue) {
+      return HandleValueChange(key, new_value);
+    }
+    return decision;
+  }
+  BlobIndex blob_index;
+  Status s = blob_index.DecodeFrom(value);
+  if (!s.ok()) {
+    // Unable to decode blob index. Keeping the value.
+    return Decision::kKeep;
+  }
+  if (blob_index.HasTTL() && blob_index.expiration() <= current_time_) {
+    // Expired
+    expired_count_++;
+    expired_size_ += key.size() + value.size();
+    return Decision::kRemove;
+  }
+  if (!blob_index.IsInlined() &&
+      blob_index.file_number() < context_.next_file_number &&
+      context_.current_blob_files.count(blob_index.file_number()) == 0) {
+    // Corresponding blob file gone (most likely, evicted by FIFO eviction).
+    evicted_count_++;
+    evicted_size_ += key.size() + value.size();
+    return Decision::kRemove;
+  }
+  if (context_.fifo_eviction_seq > 0 && blob_index.HasTTL() &&
+      blob_index.expiration() < context_.evict_expiration_up_to) {
+    // Hack: Internal key is passed to BlobIndexCompactionFilter for it to
+    // get sequence number.
+    ParsedInternalKey ikey;
+    if (!ParseInternalKey(
+             key, &ikey,
+             context_.blob_db_impl->db_options_.allow_data_in_errors)
+             .ok()) {
+      assert(false);
+      return Decision::kKeep;
+    }
+    // Remove keys that could have been remove by last FIFO eviction.
+    // If get error while parsing key, ignore and continue.
+    if (ikey.sequence < context_.fifo_eviction_seq) {
+      evicted_count_++;
+      evicted_size_ += key.size() + value.size();
+      return Decision::kRemove;
+    }
+  }
+  // Apply user compaction filter for all non-TTL blob data.
+  if (ucf != nullptr && !blob_index.HasTTL()) {
+    // Hack: Internal key is passed to BlobIndexCompactionFilter for it to
+    // get sequence number.
+    ParsedInternalKey ikey;
+    if (!ParseInternalKey(
+             key, &ikey,
+             context_.blob_db_impl->db_options_.allow_data_in_errors)
+             .ok()) {
+      assert(false);
+      return Decision::kKeep;
+    }
+    // Read value from blob file.
+    PinnableSlice blob;
+    CompressionType compression_type = kNoCompression;
+    constexpr bool need_decompress = true;
+    if (!ReadBlobFromOldFile(ikey.user_key, blob_index, &blob, need_decompress,
+                             &compression_type)) {
+      return Decision::kIOError;
+    }
+    CompactionFilter::Decision decision = ucf->FilterV2(
+        level, ikey.user_key, kValue, blob, new_value, skip_until);
+    if (decision == Decision::kChangeValue) {
+      return HandleValueChange(ikey.user_key, new_value);
+    }
+    return decision;
+  }
+  return Decision::kKeep;
+}
+
+CompactionFilter::Decision BlobIndexCompactionFilterBase::HandleValueChange(
+    const Slice& key, std::string* new_value) const {
+  BlobDBImpl* const blob_db_impl = context_.blob_db_impl;
+  assert(blob_db_impl);
+
+  if (new_value->size() < blob_db_impl->bdb_options_.min_blob_size) {
+    // Keep new_value inlined.
+    return Decision::kChangeValue;
+  }
+  if (!OpenNewBlobFileIfNeeded()) {
+    return Decision::kIOError;
+  }
+  Slice new_blob_value(*new_value);
+  std::string compression_output;
+  if (blob_db_impl->bdb_options_.compression != kNoCompression) {
+    new_blob_value =
+        blob_db_impl->GetCompressedSlice(new_blob_value, &compression_output);
+  }
+  uint64_t new_blob_file_number = 0;
+  uint64_t new_blob_offset = 0;
+  if (!WriteBlobToNewFile(key, new_blob_value, &new_blob_file_number,
+                          &new_blob_offset)) {
+    return Decision::kIOError;
+  }
+  if (!CloseAndRegisterNewBlobFileIfNeeded()) {
+    return Decision::kIOError;
+  }
+  BlobIndex::EncodeBlob(new_value, new_blob_file_number, new_blob_offset,
+                        new_blob_value.size(),
+                        blob_db_impl->bdb_options_.compression);
+  return Decision::kChangeBlobIndex;
+}
+
+BlobIndexCompactionFilterGC::~BlobIndexCompactionFilterGC() {
+  assert(context().blob_db_impl);
+
+  ROCKS_LOG_INFO(context().blob_db_impl->db_options_.info_log,
+                 "GC pass finished %s: encountered %" PRIu64 " blobs (%" PRIu64
+                 " bytes), relocated %" PRIu64 " blobs (%" PRIu64
+                 " bytes), created %" PRIu64 " new blob file(s)",
+                 !gc_stats_.HasError() ? "successfully" : "with failure",
+                 gc_stats_.AllBlobs(), gc_stats_.AllBytes(),
+                 gc_stats_.RelocatedBlobs(), gc_stats_.RelocatedBytes(),
+                 gc_stats_.NewFiles());
+
+  RecordTick(statistics(), BLOB_DB_GC_NUM_KEYS_RELOCATED,
+             gc_stats_.RelocatedBlobs());
+  RecordTick(statistics(), BLOB_DB_GC_BYTES_RELOCATED,
+             gc_stats_.RelocatedBytes());
+  RecordTick(statistics(), BLOB_DB_GC_NUM_NEW_FILES, gc_stats_.NewFiles());
+  RecordTick(statistics(), BLOB_DB_GC_FAILURES, gc_stats_.HasError());
+}
+
+bool BlobIndexCompactionFilterBase::IsBlobFileOpened() const {
+  if (blob_file_) {
+    assert(writer_);
+    return true;
+  }
+  return false;
+}
+
+bool BlobIndexCompactionFilterBase::OpenNewBlobFileIfNeeded() const {
+  if (IsBlobFileOpened()) {
+    return true;
+  }
+
+  BlobDBImpl* const blob_db_impl = context_.blob_db_impl;
+  assert(blob_db_impl);
+
+  const Status s = blob_db_impl->CreateBlobFileAndWriter(
+      /* has_ttl */ false, ExpirationRange(), "compaction/GC", &blob_file_,
+      &writer_);
+  if (!s.ok()) {
+    ROCKS_LOG_ERROR(
+        blob_db_impl->db_options_.info_log,
+        "Error opening new blob file during compaction/GC, status: %s",
+        s.ToString().c_str());
+    blob_file_.reset();
+    writer_.reset();
+    return false;
+  }
+
+  assert(blob_file_);
+  assert(writer_);
+
+  return true;
+}
+
+bool BlobIndexCompactionFilterBase::ReadBlobFromOldFile(
+    const Slice& key, const BlobIndex& blob_index, PinnableSlice* blob,
+    bool need_decompress, CompressionType* compression_type) const {
+  BlobDBImpl* const blob_db_impl = context_.blob_db_impl;
+  assert(blob_db_impl);
+
+  Status s = blob_db_impl->GetRawBlobFromFile(
+      key, blob_index.file_number(), blob_index.offset(), blob_index.size(),
+      blob, compression_type);
+
+  if (!s.ok()) {
+    ROCKS_LOG_ERROR(
+        blob_db_impl->db_options_.info_log,
+        "Error reading blob during compaction/GC, key: %s (%s), status: %s",
+        key.ToString(/* output_hex */ true).c_str(),
+        blob_index.DebugString(/* output_hex */ true).c_str(),
+        s.ToString().c_str());
+
+    return false;
+  }
+
+  if (need_decompress && *compression_type != kNoCompression) {
+    s = blob_db_impl->DecompressSlice(*blob, *compression_type, blob);
+    if (!s.ok()) {
+      ROCKS_LOG_ERROR(
+          blob_db_impl->db_options_.info_log,
+          "Uncompression error during blob read from file: %" PRIu64
+          " blob_offset: %" PRIu64 " blob_size: %" PRIu64
+          " key: %s status: '%s'",
+          blob_index.file_number(), blob_index.offset(), blob_index.size(),
+          key.ToString(/* output_hex */ true).c_str(), s.ToString().c_str());
+
+      return false;
+    }
+  }
+
+  return true;
+}
+
+bool BlobIndexCompactionFilterBase::WriteBlobToNewFile(
+    const Slice& key, const Slice& blob, uint64_t* new_blob_file_number,
+    uint64_t* new_blob_offset) const {
+  TEST_SYNC_POINT("BlobIndexCompactionFilterBase::WriteBlobToNewFile");
+  assert(new_blob_file_number);
+  assert(new_blob_offset);
+
+  assert(blob_file_);
+  *new_blob_file_number = blob_file_->BlobFileNumber();
+
+  assert(writer_);
+  uint64_t new_key_offset = 0;
+  const Status s = writer_->AddRecord(key, blob, kNoExpiration, &new_key_offset,
+                                      new_blob_offset);
+
+  if (!s.ok()) {
+    const BlobDBImpl* const blob_db_impl = context_.blob_db_impl;
+    assert(blob_db_impl);
+
+    ROCKS_LOG_ERROR(blob_db_impl->db_options_.info_log,
+                    "Error writing blob to new file %s during compaction/GC, "
+                    "key: %s, status: %s",
+                    blob_file_->PathName().c_str(),
+                    key.ToString(/* output_hex */ true).c_str(),
+                    s.ToString().c_str());
+    return false;
+  }
+
+  const uint64_t new_size =
+      BlobLogRecord::kHeaderSize + key.size() + blob.size();
+  blob_file_->BlobRecordAdded(new_size);
+
+  BlobDBImpl* const blob_db_impl = context_.blob_db_impl;
+  assert(blob_db_impl);
+
+  blob_db_impl->total_blob_size_ += new_size;
+
+  return true;
+}
+
+bool BlobIndexCompactionFilterBase::CloseAndRegisterNewBlobFileIfNeeded()
+    const {
+  const BlobDBImpl* const blob_db_impl = context_.blob_db_impl;
+  assert(blob_db_impl);
+
+  assert(blob_file_);
+  if (blob_file_->GetFileSize() < blob_db_impl->bdb_options_.blob_file_size) {
+    return true;
+  }
+
+  return CloseAndRegisterNewBlobFile();
+}
+
+bool BlobIndexCompactionFilterBase::CloseAndRegisterNewBlobFile() const {
+  BlobDBImpl* const blob_db_impl = context_.blob_db_impl;
+  assert(blob_db_impl);
+  assert(blob_file_);
+
+  Status s;
+
+  {
+    WriteLock wl(&blob_db_impl->mutex_);
+
+    s = blob_db_impl->CloseBlobFile(blob_file_);
+
+    // Note: we delay registering the new blob file until it's closed to
+    // prevent FIFO eviction from processing it during compaction/GC.
+    blob_db_impl->RegisterBlobFile(blob_file_);
+  }
+
+  assert(blob_file_->Immutable());
+
+  if (!s.ok()) {
+    ROCKS_LOG_ERROR(
+        blob_db_impl->db_options_.info_log,
+        "Error closing new blob file %s during compaction/GC, status: %s",
+        blob_file_->PathName().c_str(), s.ToString().c_str());
+  }
+
+  blob_file_.reset();
+  return s.ok();
+}
+
+CompactionFilter::BlobDecision BlobIndexCompactionFilterGC::PrepareBlobOutput(
+    const Slice& key, const Slice& existing_value,
+    std::string* new_value) const {
+  assert(new_value);
+
+  const BlobDBImpl* const blob_db_impl = context().blob_db_impl;
+  (void)blob_db_impl;
+
+  assert(blob_db_impl);
+  assert(blob_db_impl->bdb_options_.enable_garbage_collection);
+
+  BlobIndex blob_index;
+  const Status s = blob_index.DecodeFrom(existing_value);
+  if (!s.ok()) {
+    gc_stats_.SetError();
+    return BlobDecision::kCorruption;
+  }
+
+  if (blob_index.IsInlined()) {
+    gc_stats_.AddBlob(blob_index.value().size());
+
+    return BlobDecision::kKeep;
+  }
+
+  gc_stats_.AddBlob(blob_index.size());
+
+  if (blob_index.HasTTL()) {
+    return BlobDecision::kKeep;
+  }
+
+  if (blob_index.file_number() >= context_gc_.cutoff_file_number) {
+    return BlobDecision::kKeep;
+  }
+
+  // Note: each compaction generates its own blob files, which, depending on the
+  // workload, might result in many small blob files. The total number of files
+  // is bounded though (determined by the number of compactions and the blob
+  // file size option).
+  if (!OpenNewBlobFileIfNeeded()) {
+    gc_stats_.SetError();
+    return BlobDecision::kIOError;
+  }
+
+  PinnableSlice blob;
+  CompressionType compression_type = kNoCompression;
+  std::string compression_output;
+  if (!ReadBlobFromOldFile(key, blob_index, &blob, false, &compression_type)) {
+    gc_stats_.SetError();
+    return BlobDecision::kIOError;
+  }
+
+  // If the compression_type is changed, re-compress it with the new compression
+  // type.
+  if (compression_type != blob_db_impl->bdb_options_.compression) {
+    if (compression_type != kNoCompression) {
+      const Status status =
+          blob_db_impl->DecompressSlice(blob, compression_type, &blob);
+      if (!status.ok()) {
+        gc_stats_.SetError();
+        return BlobDecision::kCorruption;
+      }
+    }
+    if (blob_db_impl->bdb_options_.compression != kNoCompression) {
+      blob_db_impl->GetCompressedSlice(blob, &compression_output);
+      blob = PinnableSlice(&compression_output);
+      blob.PinSelf();
+    }
+  }
+
+  uint64_t new_blob_file_number = 0;
+  uint64_t new_blob_offset = 0;
+  if (!WriteBlobToNewFile(key, blob, &new_blob_file_number, &new_blob_offset)) {
+    gc_stats_.SetError();
+    return BlobDecision::kIOError;
+  }
+
+  if (!CloseAndRegisterNewBlobFileIfNeeded()) {
+    gc_stats_.SetError();
+    return BlobDecision::kIOError;
+  }
+
+  BlobIndex::EncodeBlob(new_value, new_blob_file_number, new_blob_offset,
+                        blob.size(), compression_type);
+
+  gc_stats_.AddRelocatedBlob(blob_index.size());
+
+  return BlobDecision::kChangeValue;
+}
+
+bool BlobIndexCompactionFilterGC::OpenNewBlobFileIfNeeded() const {
+  if (IsBlobFileOpened()) {
+    return true;
+  }
+  bool result = BlobIndexCompactionFilterBase::OpenNewBlobFileIfNeeded();
+  if (result) {
+    gc_stats_.AddNewFile();
+  }
+  return result;
+}
+
+std::unique_ptr<CompactionFilter>
+BlobIndexCompactionFilterFactoryBase::CreateUserCompactionFilterFromFactory(
+    const CompactionFilter::Context& context) const {
+  std::unique_ptr<CompactionFilter> user_comp_filter_from_factory;
+  if (user_comp_filter_factory_) {
+    user_comp_filter_from_factory =
+        user_comp_filter_factory_->CreateCompactionFilter(context);
+  }
+  return user_comp_filter_from_factory;
+}
+
+std::unique_ptr<CompactionFilter>
+BlobIndexCompactionFilterFactory::CreateCompactionFilter(
+    const CompactionFilter::Context& _context) {
+  assert(clock());
+
+  int64_t current_time = 0;
+  Status s = clock()->GetCurrentTime(&current_time);
+  if (!s.ok()) {
+    return nullptr;
+  }
+  assert(current_time >= 0);
+
+  assert(blob_db_impl());
+
+  BlobCompactionContext context;
+  blob_db_impl()->GetCompactionContext(&context);
+
+  std::unique_ptr<CompactionFilter> user_comp_filter_from_factory =
+      CreateUserCompactionFilterFromFactory(_context);
+
+  return std::unique_ptr<CompactionFilter>(new BlobIndexCompactionFilter(
+      std::move(context), user_comp_filter(),
+      std::move(user_comp_filter_from_factory), current_time, statistics()));
+}
+
+std::unique_ptr<CompactionFilter>
+BlobIndexCompactionFilterFactoryGC::CreateCompactionFilter(
+    const CompactionFilter::Context& _context) {
+  assert(clock());
+
+  int64_t current_time = 0;
+  Status s = clock()->GetCurrentTime(&current_time);
+  if (!s.ok()) {
+    return nullptr;
+  }
+  assert(current_time >= 0);
+
+  assert(blob_db_impl());
+
+  BlobCompactionContext context;
+  BlobCompactionContextGC context_gc;
+  blob_db_impl()->GetCompactionContext(&context, &context_gc);
+
+  std::unique_ptr<CompactionFilter> user_comp_filter_from_factory =
+      CreateUserCompactionFilterFromFactory(_context);
+
+  return std::unique_ptr<CompactionFilter>(new BlobIndexCompactionFilterGC(
+      std::move(context), std::move(context_gc), user_comp_filter(),
+      std::move(user_comp_filter_from_factory), current_time, statistics()));
+}
+
+}  // namespace blob_db
+}  // namespace ROCKSDB_NAMESPACE
+#endif  // ROCKSDB_LITE
diff --git a/src/rocksdb/utilities/blob_db/blob_compaction_filter.h b/src/rocksdb/utilities/blob_db/blob_compaction_filter.h
new file mode 100644
index 000000000..1493cfc1a
--- /dev/null
+++ b/src/rocksdb/utilities/blob_db/blob_compaction_filter.h
@@ -0,0 +1,204 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+#pragma once
+#ifndef ROCKSDB_LITE
+
+#include <unordered_set>
+
+#include "db/blob/blob_index.h"
+#include "monitoring/statistics.h"
+#include "rocksdb/compaction_filter.h"
+#include "utilities/blob_db/blob_db_gc_stats.h"
+#include "utilities/blob_db/blob_db_impl.h"
+#include "utilities/compaction_filters/layered_compaction_filter_base.h"
+
+namespace ROCKSDB_NAMESPACE {
+class SystemClock;
+namespace blob_db {
+
+struct BlobCompactionContext {
+  BlobDBImpl* blob_db_impl = nullptr;
+  uint64_t next_file_number = 0;
+  std::unordered_set<uint64_t> current_blob_files;
+  SequenceNumber fifo_eviction_seq = 0;
+  uint64_t evict_expiration_up_to = 0;
+};
+
+struct BlobCompactionContextGC {
+  uint64_t cutoff_file_number = 0;
+};
+
+// Compaction filter that deletes expired blob indexes from the base DB.
+// Comes into two varieties, one for the non-GC case and one for the GC case.
+class BlobIndexCompactionFilterBase : public LayeredCompactionFilterBase {
+ public:
+  BlobIndexCompactionFilterBase(
+      BlobCompactionContext&& _context,
+      const CompactionFilter* _user_comp_filter,
+      std::unique_ptr<const CompactionFilter> _user_comp_filter_from_factory,
+      uint64_t current_time, Statistics* stats)
+      : LayeredCompactionFilterBase(_user_comp_filter,
+                                    std::move(_user_comp_filter_from_factory)),
+        context_(std::move(_context)),
+        current_time_(current_time),
+        statistics_(stats) {}
+
+  ~BlobIndexCompactionFilterBase() override;
+
+  // Filter expired blob indexes regardless of snapshots.
+  bool IgnoreSnapshots() const override { return true; }
+
+  Decision FilterV2(int level, const Slice& key, ValueType value_type,
+                    const Slice& value, std::string* new_value,
+                    std::string* skip_until) const override;
+
+  bool IsStackedBlobDbInternalCompactionFilter() const override { return true; }
+
+ protected:
+  bool IsBlobFileOpened() const;
+  virtual bool OpenNewBlobFileIfNeeded() const;
+  bool ReadBlobFromOldFile(const Slice& key, const BlobIndex& blob_index,
+                           PinnableSlice* blob, bool need_decompress,
+                           CompressionType* compression_type) const;
+  bool WriteBlobToNewFile(const Slice& key, const Slice& blob,
+                          uint64_t* new_blob_file_number,
+                          uint64_t* new_blob_offset) const;
+  bool CloseAndRegisterNewBlobFileIfNeeded() const;
+  bool CloseAndRegisterNewBlobFile() const;
+
+  Statistics* statistics() const { return statistics_; }
+  const BlobCompactionContext& context() const { return context_; }
+
+ private:
+  Decision HandleValueChange(const Slice& key, std::string* new_value) const;
+
+ private:
+  BlobCompactionContext context_;
+  const uint64_t current_time_;
+  Statistics* statistics_;
+
+  mutable std::shared_ptr<BlobFile> blob_file_;
+  mutable std::shared_ptr<BlobLogWriter> writer_;
+
+  // It is safe to not using std::atomic since the compaction filter, created
+  // from a compaction filter factroy, will not be called from multiple threads.
+  mutable uint64_t expired_count_ = 0;
+  mutable uint64_t expired_size_ = 0;
+  mutable uint64_t evicted_count_ = 0;
+  mutable uint64_t evicted_size_ = 0;
+};
+
+class BlobIndexCompactionFilter : public BlobIndexCompactionFilterBase {
+ public:
+  BlobIndexCompactionFilter(
+      BlobCompactionContext&& _context,
+      const CompactionFilter* _user_comp_filter,
+      std::unique_ptr<const CompactionFilter> _user_comp_filter_from_factory,
+      uint64_t current_time, Statistics* stats)
+      : BlobIndexCompactionFilterBase(std::move(_context), _user_comp_filter,
+                                      std::move(_user_comp_filter_from_factory),
+                                      current_time, stats) {}
+
+  const char* Name() const override { return "BlobIndexCompactionFilter"; }
+};
+
+class BlobIndexCompactionFilterGC : public BlobIndexCompactionFilterBase {
+ public:
+  BlobIndexCompactionFilterGC(
+      BlobCompactionContext&& _context, BlobCompactionContextGC&& context_gc,
+      const CompactionFilter* _user_comp_filter,
+      std::unique_ptr<const CompactionFilter> _user_comp_filter_from_factory,
+      uint64_t current_time, Statistics* stats)
+      : BlobIndexCompactionFilterBase(std::move(_context), _user_comp_filter,
+                                      std::move(_user_comp_filter_from_factory),
+                                      current_time, stats),
+        context_gc_(std::move(context_gc)) {}
+
+  ~BlobIndexCompactionFilterGC() override;
+
+  const char* Name() const override { return "BlobIndexCompactionFilterGC"; }
+
+  BlobDecision PrepareBlobOutput(const Slice& key, const Slice& existing_value,
+                                 std::string* new_value) const override;
+
+ private:
+  bool OpenNewBlobFileIfNeeded() const override;
+
+ private:
+  BlobCompactionContextGC context_gc_;
+  mutable BlobDBGarbageCollectionStats gc_stats_;
+};
+
+// Compaction filter factory; similarly to the filters above, it comes
+// in two flavors, one that creates filters that support GC, and one
+// that creates non-GC filters.
+class BlobIndexCompactionFilterFactoryBase : public CompactionFilterFactory {
+ public:
+  BlobIndexCompactionFilterFactoryBase(BlobDBImpl* _blob_db_impl,
+                                       SystemClock* _clock,
+                                       const ColumnFamilyOptions& _cf_options,
+                                       Statistics* _statistics)
+      : blob_db_impl_(_blob_db_impl),
+        clock_(_clock),
+        statistics_(_statistics),
+        user_comp_filter_(_cf_options.compaction_filter),
+        user_comp_filter_factory_(_cf_options.compaction_filter_factory) {}
+
+ protected:
+  std::unique_ptr<CompactionFilter> CreateUserCompactionFilterFromFactory(
+      const CompactionFilter::Context& context) const;
+
+  BlobDBImpl* blob_db_impl() const { return blob_db_impl_; }
+  SystemClock* clock() const { return clock_; }
+  Statistics* statistics() const { return statistics_; }
+  const CompactionFilter* user_comp_filter() const { return user_comp_filter_; }
+
+ private:
+  BlobDBImpl* blob_db_impl_;
+  SystemClock* clock_;
+  Statistics* statistics_;
+  const CompactionFilter* user_comp_filter_;
+  std::shared_ptr<CompactionFilterFactory> user_comp_filter_factory_;
+};
+
+class BlobIndexCompactionFilterFactory
+    : public BlobIndexCompactionFilterFactoryBase {
+ public:
+  BlobIndexCompactionFilterFactory(BlobDBImpl* _blob_db_impl,
+                                   SystemClock* _clock,
+                                   const ColumnFamilyOptions& _cf_options,
+                                   Statistics* _statistics)
+      : BlobIndexCompactionFilterFactoryBase(_blob_db_impl, _clock, _cf_options,
+                                             _statistics) {}
+
+  const char* Name() const override {
+    return "BlobIndexCompactionFilterFactory";
+  }
+
+  std::unique_ptr<CompactionFilter> CreateCompactionFilter(
+      const CompactionFilter::Context& context) override;
+};
+
+class BlobIndexCompactionFilterFactoryGC
+    : public BlobIndexCompactionFilterFactoryBase {
+ public:
+  BlobIndexCompactionFilterFactoryGC(BlobDBImpl* _blob_db_impl,
+                                     SystemClock* _clock,
+                                     const ColumnFamilyOptions& _cf_options,
+                                     Statistics* _statistics)
+      : BlobIndexCompactionFilterFactoryBase(_blob_db_impl, _clock, _cf_options,
+                                             _statistics) {}
+
+  const char* Name() const override {
+    return "BlobIndexCompactionFilterFactoryGC";
+  }
+
+  std::unique_ptr<CompactionFilter> CreateCompactionFilter(
+      const CompactionFilter::Context& context) override;
+};
+
+}  // namespace blob_db
+}  // namespace ROCKSDB_NAMESPACE
+#endif  // ROCKSDB_LITE
diff --git a/src/rocksdb/utilities/blob_db/blob_db.cc b/src/rocksdb/utilities/blob_db/blob_db.cc
new file mode 100644
index 000000000..cbd02e68e
--- /dev/null
+++ b/src/rocksdb/utilities/blob_db/blob_db.cc
@@ -0,0 +1,114 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+//
+#ifndef ROCKSDB_LITE
+
+#include "utilities/blob_db/blob_db.h"
+
+#include <cinttypes>
+
+#include "logging/logging.h"
+#include "utilities/blob_db/blob_db_impl.h"
+
+namespace ROCKSDB_NAMESPACE {
+namespace blob_db {
+
+Status BlobDB::Open(const Options& options, const BlobDBOptions& bdb_options,
+                    const std::string& dbname, BlobDB** blob_db) {
+  *blob_db = nullptr;
+  DBOptions db_options(options);
+  ColumnFamilyOptions cf_options(options);
+  std::vector<ColumnFamilyDescriptor> column_families;
+  column_families.push_back(
+      ColumnFamilyDescriptor(kDefaultColumnFamilyName, cf_options));
+  std::vector<ColumnFamilyHandle*> handles;
+  Status s = BlobDB::Open(db_options, bdb_options, dbname, column_families,
+                          &handles, blob_db);
+  if (s.ok()) {
+    assert(handles.size() == 1);
+    // i can delete the handle since DBImpl is always holding a reference to
+    // default column family
+    delete handles[0];
+  }
+  return s;
+}
+
+Status BlobDB::Open(const DBOptions& db_options,
+                    const BlobDBOptions& bdb_options, const std::string& dbname,
+                    const std::vector<ColumnFamilyDescriptor>& column_families,
+                    std::vector<ColumnFamilyHandle*>* handles,
+                    BlobDB** blob_db) {
+  assert(handles);
+
+  if (column_families.size() != 1 ||
+      column_families[0].name != kDefaultColumnFamilyName) {
+    return Status::NotSupported(
+        "Blob DB doesn't support non-default column family.");
+  }
+
+  BlobDBImpl* blob_db_impl = new BlobDBImpl(dbname, bdb_options, db_options,
+                                            column_families[0].options);
+  Status s = blob_db_impl->Open(handles);
+  if (s.ok()) {
+    *blob_db = static_cast<BlobDB*>(blob_db_impl);
+  } else {
+    if (!handles->empty()) {
+      for (ColumnFamilyHandle* cfh : *handles) {
+        blob_db_impl->DestroyColumnFamilyHandle(cfh);
+      }
+
+      handles->clear();
+    }
+
+    delete blob_db_impl;
+    *blob_db = nullptr;
+  }
+  return s;
+}
+
+BlobDB::BlobDB() : StackableDB(nullptr) {}
+
+void BlobDBOptions::Dump(Logger* log) const {
+  ROCKS_LOG_HEADER(
+      log, "                                  BlobDBOptions.blob_dir: %s",
+      blob_dir.c_str());
+  ROCKS_LOG_HEADER(
+      log, "                             BlobDBOptions.path_relative: %d",
+      path_relative);
+  ROCKS_LOG_HEADER(
+      log, "                                   BlobDBOptions.is_fifo: %d",
+      is_fifo);
+  ROCKS_LOG_HEADER(
+      log, "                               BlobDBOptions.max_db_size: %" PRIu64,
+      max_db_size);
+  ROCKS_LOG_HEADER(
+      log, "                            BlobDBOptions.ttl_range_secs: %" PRIu64,
+      ttl_range_secs);
+  ROCKS_LOG_HEADER(
+      log, "                             BlobDBOptions.min_blob_size: %" PRIu64,
+      min_blob_size);
+  ROCKS_LOG_HEADER(
+      log, "                            BlobDBOptions.bytes_per_sync: %" PRIu64,
+      bytes_per_sync);
+  ROCKS_LOG_HEADER(
+      log, "                            BlobDBOptions.blob_file_size: %" PRIu64,
+      blob_file_size);
+  ROCKS_LOG_HEADER(
+      log, "                               BlobDBOptions.compression: %d",
+      static_cast<int>(compression));
+  ROCKS_LOG_HEADER(
+      log, "                 BlobDBOptions.enable_garbage_collection: %d",
+      enable_garbage_collection);
+  ROCKS_LOG_HEADER(
+      log, "                 BlobDBOptions.garbage_collection_cutoff: %f",
+      garbage_collection_cutoff);
+  ROCKS_LOG_HEADER(
+      log, "                  BlobDBOptions.disable_background_tasks: %d",
+      disable_background_tasks);
+}
+
+}  // namespace blob_db
+}  // namespace ROCKSDB_NAMESPACE
+#endif
diff --git a/src/rocksdb/utilities/blob_db/blob_db.h b/src/rocksdb/utilities/blob_db/blob_db.h
new file mode 100644
index 000000000..e9d92486f
--- /dev/null
+++ b/src/rocksdb/utilities/blob_db/blob_db.h
@@ -0,0 +1,266 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#pragma once
+
+#ifndef ROCKSDB_LITE
+
+#include <functional>
+#include <limits>
+#include <string>
+#include <vector>
+
+#include "rocksdb/db.h"
+#include "rocksdb/status.h"
+#include "rocksdb/utilities/stackable_db.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+namespace blob_db {
+
+// A wrapped database which puts values of KV pairs in a separate log
+// and store location to the log in the underlying DB.
+//
+// The factory needs to be moved to include/rocksdb/utilities to allow
+// users to use blob DB.
+
+constexpr uint64_t kNoExpiration = std::numeric_limits<uint64_t>::max();
+
+struct BlobDBOptions {
+  // Name of the directory under the base DB where blobs will be stored. Using
+  // a directory where the base DB stores its SST files is not supported.
+  // Default is "blob_dir"
+  std::string blob_dir = "blob_dir";
+
+  // whether the blob_dir path is relative or absolute.
+  bool path_relative = true;
+
+  // When max_db_size is reached, evict blob files to free up space
+  // instead of returnning NoSpace error on write. Blob files will be
+  // evicted from oldest to newest, based on file creation time.
+  bool is_fifo = false;
+
+  // Maximum size of the database (including SST files and blob files).
+  //
+  // Default: 0 (no limits)
+  uint64_t max_db_size = 0;
+
+  // a new bucket is opened, for ttl_range. So if ttl_range is 600seconds
+  // (10 minutes), and the first bucket starts at 1471542000
+  // then the blob buckets will be
+  // first bucket is 1471542000 - 1471542600
+  // second bucket is 1471542600 - 1471543200
+  // and so on
+  uint64_t ttl_range_secs = 3600;
+
+  // The smallest value to store in blob log. Values smaller than this threshold
+  // will be inlined in base DB together with the key.
+  uint64_t min_blob_size = 0;
+
+  // Allows OS to incrementally sync blob files to disk for every
+  // bytes_per_sync bytes written. Users shouldn't rely on it for
+  // persistency guarantee.
+  uint64_t bytes_per_sync = 512 * 1024;
+
+  // the target size of each blob file. File will become immutable
+  // after it exceeds that size
+  uint64_t blob_file_size = 256 * 1024 * 1024;
+
+  // what compression to use for Blob's
+  CompressionType compression = kNoCompression;
+
+  // If enabled, BlobDB cleans up stale blobs in non-TTL files during compaction
+  // by rewriting the remaining live blobs to new files.
+  bool enable_garbage_collection = false;
+
+  // The cutoff in terms of blob file age for garbage collection. Blobs in
+  // the oldest N non-TTL blob files will be rewritten when encountered during
+  // compaction, where N = garbage_collection_cutoff * number_of_non_TTL_files.
+  double garbage_collection_cutoff = 0.25;
+
+  // Disable all background job. Used for test only.
+  bool disable_background_tasks = false;
+
+  void Dump(Logger* log) const;
+};
+
+class BlobDB : public StackableDB {
+ public:
+  using ROCKSDB_NAMESPACE::StackableDB::Put;
+  virtual Status Put(const WriteOptions& options, const Slice& key,
+                     const Slice& value) override = 0;
+  virtual Status Put(const WriteOptions& options,
+                     ColumnFamilyHandle* column_family, const Slice& key,
+                     const Slice& value) override {
+    if (column_family->GetID() != DefaultColumnFamily()->GetID()) {
+      return Status::NotSupported(
+          "Blob DB doesn't support non-default column family.");
+    }
+    return Put(options, key, value);
+  }
+
+  using ROCKSDB_NAMESPACE::StackableDB::Delete;
+  virtual Status Delete(const WriteOptions& options,
+                        ColumnFamilyHandle* column_family,
+                        const Slice& key) override {
+    if (column_family->GetID() != DefaultColumnFamily()->GetID()) {
+      return Status::NotSupported(
+          "Blob DB doesn't support non-default column family.");
+    }
+    assert(db_ != nullptr);
+    return db_->Delete(options, column_family, key);
+  }
+
+  virtual Status PutWithTTL(const WriteOptions& options, const Slice& key,
+                            const Slice& value, uint64_t ttl) = 0;
+  virtual Status PutWithTTL(const WriteOptions& options,
+                            ColumnFamilyHandle* column_family, const Slice& key,
+                            const Slice& value, uint64_t ttl) {
+    if (column_family->GetID() != DefaultColumnFamily()->GetID()) {
+      return Status::NotSupported(
+          "Blob DB doesn't support non-default column family.");
+    }
+    return PutWithTTL(options, key, value, ttl);
+  }
+
+  // Put with expiration. Key with expiration time equal to
+  // std::numeric_limits<uint64_t>::max() means the key don't expire.
+  virtual Status PutUntil(const WriteOptions& options, const Slice& key,
+                          const Slice& value, uint64_t expiration) = 0;
+  virtual Status PutUntil(const WriteOptions& options,
+                          ColumnFamilyHandle* column_family, const Slice& key,
+                          const Slice& value, uint64_t expiration) {
+    if (column_family->GetID() != DefaultColumnFamily()->GetID()) {
+      return Status::NotSupported(
+          "Blob DB doesn't support non-default column family.");
+    }
+    return PutUntil(options, key, value, expiration);
+  }
+
+  using ROCKSDB_NAMESPACE::StackableDB::Get;
+  virtual Status Get(const ReadOptions& options,
+                     ColumnFamilyHandle* column_family, const Slice& key,
+                     PinnableSlice* value) override = 0;
+
+  // Get value and expiration.
+  virtual Status Get(const ReadOptions& options,
+                     ColumnFamilyHandle* column_family, const Slice& key,
+                     PinnableSlice* value, uint64_t* expiration) = 0;
+  virtual Status Get(const ReadOptions& options, const Slice& key,
+                     PinnableSlice* value, uint64_t* expiration) {
+    return Get(options, DefaultColumnFamily(), key, value, expiration);
+  }
+
+  using ROCKSDB_NAMESPACE::StackableDB::MultiGet;
+  virtual std::vector<Status> MultiGet(
+      const ReadOptions& options, const std::vector<Slice>& keys,
+      std::vector<std::string>* values) override = 0;
+  virtual std::vector<Status> MultiGet(
+      const ReadOptions& options,
+      const std::vector<ColumnFamilyHandle*>& column_families,
+      const std::vector<Slice>& keys,
+      std::vector<std::string>* values) override {
+    for (auto column_family : column_families) {
+      if (column_family->GetID() != DefaultColumnFamily()->GetID()) {
+        return std::vector<Status>(
+            column_families.size(),
+            Status::NotSupported(
+                "Blob DB doesn't support non-default column family."));
+      }
+    }
+    return MultiGet(options, keys, values);
+  }
+  virtual void MultiGet(const ReadOptions& /*options*/,
+                        ColumnFamilyHandle* /*column_family*/,
+                        const size_t num_keys, const Slice* /*keys*/,
+                        PinnableSlice* /*values*/, Status* statuses,
+                        const bool /*sorted_input*/ = false) override {
+    for (size_t i = 0; i < num_keys; ++i) {
+      statuses[i] =
+          Status::NotSupported("Blob DB doesn't support batched MultiGet");
+    }
+  }
+
+  using ROCKSDB_NAMESPACE::StackableDB::SingleDelete;
+  virtual Status SingleDelete(const WriteOptions& /*wopts*/,
+                              ColumnFamilyHandle* /*column_family*/,
+                              const Slice& /*key*/) override {
+    return Status::NotSupported("Not supported operation in blob db.");
+  }
+
+  using ROCKSDB_NAMESPACE::StackableDB::Merge;
+  virtual Status Merge(const WriteOptions& /*options*/,
+                       ColumnFamilyHandle* /*column_family*/,
+                       const Slice& /*key*/, const Slice& /*value*/) override {
+    return Status::NotSupported("Not supported operation in blob db.");
+  }
+
+  virtual Status Write(const WriteOptions& opts,
+                       WriteBatch* updates) override = 0;
+
+  using ROCKSDB_NAMESPACE::StackableDB::NewIterator;
+  virtual Iterator* NewIterator(const ReadOptions& options) override = 0;
+  virtual Iterator* NewIterator(const ReadOptions& options,
+                                ColumnFamilyHandle* column_family) override {
+    if (column_family->GetID() != DefaultColumnFamily()->GetID()) {
+      // Blob DB doesn't support non-default column family.
+      return nullptr;
+    }
+    return NewIterator(options);
+  }
+
+  Status CompactFiles(
+      const CompactionOptions& compact_options,
+      const std::vector<std::string>& input_file_names, const int output_level,
+      const int output_path_id = -1,
+      std::vector<std::string>* const output_file_names = nullptr,
+      CompactionJobInfo* compaction_job_info = nullptr) override = 0;
+  Status CompactFiles(
+      const CompactionOptions& compact_options,
+      ColumnFamilyHandle* column_family,
+      const std::vector<std::string>& input_file_names, const int output_level,
+      const int output_path_id = -1,
+      std::vector<std::string>* const output_file_names = nullptr,
+      CompactionJobInfo* compaction_job_info = nullptr) override {
+    if (column_family->GetID() != DefaultColumnFamily()->GetID()) {
+      return Status::NotSupported(
+          "Blob DB doesn't support non-default column family.");
+    }
+
+    return CompactFiles(compact_options, input_file_names, output_level,
+                        output_path_id, output_file_names, compaction_job_info);
+  }
+
+  using ROCKSDB_NAMESPACE::StackableDB::Close;
+  virtual Status Close() override = 0;
+
+  // Opening blob db.
+  static Status Open(const Options& options, const BlobDBOptions& bdb_options,
+                     const std::string& dbname, BlobDB** blob_db);
+
+  static Status Open(const DBOptions& db_options,
+                     const BlobDBOptions& bdb_options,
+                     const std::string& dbname,
+                     const std::vector<ColumnFamilyDescriptor>& column_families,
+                     std::vector<ColumnFamilyHandle*>* handles,
+                     BlobDB** blob_db);
+
+  virtual BlobDBOptions GetBlobDBOptions() const = 0;
+
+  virtual Status SyncBlobFiles() = 0;
+
+  virtual ~BlobDB() {}
+
+ protected:
+  explicit BlobDB();
+};
+
+// Destroy the content of the database.
+Status DestroyBlobDB(const std::string& dbname, const Options& options,
+                     const BlobDBOptions& bdb_options);
+
+}  // namespace blob_db
+}  // namespace ROCKSDB_NAMESPACE
+#endif  // ROCKSDB_LITE
diff --git a/src/rocksdb/utilities/blob_db/blob_db_gc_stats.h b/src/rocksdb/utilities/blob_db/blob_db_gc_stats.h
new file mode 100644
index 000000000..fea6b0032
--- /dev/null
+++ b/src/rocksdb/utilities/blob_db/blob_db_gc_stats.h
@@ -0,0 +1,56 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+//
+#pragma once
+
+#include <cstdint>
+
+#include "rocksdb/rocksdb_namespace.h"
+
+#ifndef ROCKSDB_LITE
+
+namespace ROCKSDB_NAMESPACE {
+
+namespace blob_db {
+
+/**
+ * Statistics related to a single garbage collection pass (i.e. a single
+ * (sub)compaction).
+ */
+class BlobDBGarbageCollectionStats {
+ public:
+  uint64_t AllBlobs() const { return all_blobs_; }
+  uint64_t AllBytes() const { return all_bytes_; }
+  uint64_t RelocatedBlobs() const { return relocated_blobs_; }
+  uint64_t RelocatedBytes() const { return relocated_bytes_; }
+  uint64_t NewFiles() const { return new_files_; }
+  bool HasError() const { return error_; }
+
+  void AddBlob(uint64_t size) {
+    ++all_blobs_;
+    all_bytes_ += size;
+  }
+
+  void AddRelocatedBlob(uint64_t size) {
+    ++relocated_blobs_;
+    relocated_bytes_ += size;
+  }
+
+  void AddNewFile() { ++new_files_; }
+
+  void SetError() { error_ = true; }
+
+ private:
+  uint64_t all_blobs_ = 0;
+  uint64_t all_bytes_ = 0;
+  uint64_t relocated_blobs_ = 0;
+  uint64_t relocated_bytes_ = 0;
+  uint64_t new_files_ = 0;
+  bool error_ = false;
+};
+
+}  // namespace blob_db
+}  // namespace ROCKSDB_NAMESPACE
+#endif  // ROCKSDB_LITE
diff --git a/src/rocksdb/utilities/blob_db/blob_db_impl.cc b/src/rocksdb/utilities/blob_db/blob_db_impl.cc
new file mode 100644
index 000000000..87e294c5c
--- /dev/null
+++ b/src/rocksdb/utilities/blob_db/blob_db_impl.cc
@@ -0,0 +1,2177 @@
+
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+#ifndef ROCKSDB_LITE
+
+#include "utilities/blob_db/blob_db_impl.h"
+
+#include <algorithm>
+#include <cinttypes>
+#include <iomanip>
+#include <memory>
+#include <sstream>
+
+#include "db/blob/blob_index.h"
+#include "db/db_impl/db_impl.h"
+#include "db/write_batch_internal.h"
+#include "file/file_util.h"
+#include "file/filename.h"
+#include "file/random_access_file_reader.h"
+#include "file/sst_file_manager_impl.h"
+#include "file/writable_file_writer.h"
+#include "logging/logging.h"
+#include "monitoring/instrumented_mutex.h"
+#include "monitoring/statistics.h"
+#include "rocksdb/convenience.h"
+#include "rocksdb/env.h"
+#include "rocksdb/iterator.h"
+#include "rocksdb/utilities/stackable_db.h"
+#include "rocksdb/utilities/transaction.h"
+#include "table/block_based/block.h"
+#include "table/block_based/block_based_table_builder.h"
+#include "table/block_based/block_builder.h"
+#include "table/meta_blocks.h"
+#include "test_util/sync_point.h"
+#include "util/cast_util.h"
+#include "util/crc32c.h"
+#include "util/mutexlock.h"
+#include "util/random.h"
+#include "util/stop_watch.h"
+#include "util/timer_queue.h"
+#include "utilities/blob_db/blob_compaction_filter.h"
+#include "utilities/blob_db/blob_db_iterator.h"
+#include "utilities/blob_db/blob_db_listener.h"
+
+namespace {
+int kBlockBasedTableVersionFormat = 2;
+}  // end namespace
+
+namespace ROCKSDB_NAMESPACE {
+namespace blob_db {
+
+bool BlobFileComparator::operator()(
+    const std::shared_ptr<BlobFile>& lhs,
+    const std::shared_ptr<BlobFile>& rhs) const {
+  return lhs->BlobFileNumber() > rhs->BlobFileNumber();
+}
+
+bool BlobFileComparatorTTL::operator()(
+    const std::shared_ptr<BlobFile>& lhs,
+    const std::shared_ptr<BlobFile>& rhs) const {
+  assert(lhs->HasTTL() && rhs->HasTTL());
+  if (lhs->expiration_range_.first < rhs->expiration_range_.first) {
+    return true;
+  }
+  if (lhs->expiration_range_.first > rhs->expiration_range_.first) {
+    return false;
+  }
+  return lhs->BlobFileNumber() < rhs->BlobFileNumber();
+}
+
+BlobDBImpl::BlobDBImpl(const std::string& dbname,
+                       const BlobDBOptions& blob_db_options,
+                       const DBOptions& db_options,
+                       const ColumnFamilyOptions& cf_options)
+    : BlobDB(),
+      dbname_(dbname),
+      db_impl_(nullptr),
+      env_(db_options.env),
+      bdb_options_(blob_db_options),
+      db_options_(db_options),
+      cf_options_(cf_options),
+      file_options_(db_options),
+      statistics_(db_options_.statistics.get()),
+      next_file_number_(1),
+      flush_sequence_(0),
+      closed_(true),
+      open_file_count_(0),
+      total_blob_size_(0),
+      live_sst_size_(0),
+      fifo_eviction_seq_(0),
+      evict_expiration_up_to_(0),
+      debug_level_(0) {
+  clock_ = env_->GetSystemClock().get();
+  blob_dir_ = (bdb_options_.path_relative)
+                  ? dbname + "/" + bdb_options_.blob_dir
+                  : bdb_options_.blob_dir;
+  file_options_.bytes_per_sync = blob_db_options.bytes_per_sync;
+}
+
+BlobDBImpl::~BlobDBImpl() {
+  tqueue_.shutdown();
+  // CancelAllBackgroundWork(db_, true);
+  Status s __attribute__((__unused__)) = Close();
+  assert(s.ok());
+}
+
+Status BlobDBImpl::Close() {
+  if (closed_) {
+    return Status::OK();
+  }
+  closed_ = true;
+
+  // Close base DB before BlobDBImpl destructs to stop event listener and
+  // compaction filter call.
+  Status s = db_->Close();
+  // delete db_ anyway even if close failed.
+  delete db_;
+  // Reset pointers to avoid StackableDB delete the pointer again.
+  db_ = nullptr;
+  db_impl_ = nullptr;
+  if (!s.ok()) {
+    return s;
+  }
+
+  s = SyncBlobFiles();
+  return s;
+}
+
+BlobDBOptions BlobDBImpl::GetBlobDBOptions() const { return bdb_options_; }
+
+Status BlobDBImpl::Open(std::vector<ColumnFamilyHandle*>* handles) {
+  assert(handles != nullptr);
+  assert(db_ == nullptr);
+
+  if (blob_dir_.empty()) {
+    return Status::NotSupported("No blob directory in options");
+  }
+
+  if (bdb_options_.garbage_collection_cutoff < 0.0 ||
+      bdb_options_.garbage_collection_cutoff > 1.0) {
+    return Status::InvalidArgument(
+        "Garbage collection cutoff must be in the interval [0.0, 1.0]");
+  }
+
+  // Temporarily disable compactions in the base DB during open; save the user
+  // defined value beforehand so we can restore it once BlobDB is initialized.
+  // Note: this is only needed if garbage collection is enabled.
+  const bool disable_auto_compactions = cf_options_.disable_auto_compactions;
+
+  if (bdb_options_.enable_garbage_collection) {
+    cf_options_.disable_auto_compactions = true;
+  }
+
+  Status s;
+
+  // Create info log.
+  if (db_options_.info_log == nullptr) {
+    s = CreateLoggerFromOptions(dbname_, db_options_, &db_options_.info_log);
+    if (!s.ok()) {
+      return s;
+    }
+  }
+
+  ROCKS_LOG_INFO(db_options_.info_log, "Opening BlobDB...");
+
+  if ((cf_options_.compaction_filter != nullptr ||
+       cf_options_.compaction_filter_factory != nullptr)) {
+    ROCKS_LOG_INFO(db_options_.info_log,
+                   "BlobDB only support compaction filter on non-TTL values.");
+  }
+
+  // Open blob directory.
+  s = env_->CreateDirIfMissing(blob_dir_);
+  if (!s.ok()) {
+    ROCKS_LOG_ERROR(db_options_.info_log,
+                    "Failed to create blob_dir %s, status: %s",
+                    blob_dir_.c_str(), s.ToString().c_str());
+  }
+  s = env_->GetFileSystem()->NewDirectory(blob_dir_, IOOptions(), &dir_ent_,
+                                          nullptr);
+  if (!s.ok()) {
+    ROCKS_LOG_ERROR(db_options_.info_log,
+                    "Failed to open blob_dir %s, status: %s", blob_dir_.c_str(),
+                    s.ToString().c_str());
+    return s;
+  }
+
+  // Open blob files.
+  s = OpenAllBlobFiles();
+  if (!s.ok()) {
+    return s;
+  }
+
+  // Update options
+  if (bdb_options_.enable_garbage_collection) {
+    db_options_.listeners.push_back(std::make_shared<BlobDBListenerGC>(this));
+    cf_options_.compaction_filter_factory =
+        std::make_shared<BlobIndexCompactionFilterFactoryGC>(
+            this, clock_, cf_options_, statistics_);
+  } else {
+    db_options_.listeners.push_back(std::make_shared<BlobDBListener>(this));
+    cf_options_.compaction_filter_factory =
+        std::make_shared<BlobIndexCompactionFilterFactory>(
+            this, clock_, cf_options_, statistics_);
+  }
+
+  // Reset user compaction filter after building into compaction factory.
+  cf_options_.compaction_filter = nullptr;
+
+  // Open base db.
+  ColumnFamilyDescriptor cf_descriptor(kDefaultColumnFamilyName, cf_options_);
+  s = DB::Open(db_options_, dbname_, {cf_descriptor}, handles, &db_);
+  if (!s.ok()) {
+    return s;
+  }
+  db_impl_ = static_cast_with_check<DBImpl>(db_->GetRootDB());
+
+  // Sanitize the blob_dir provided. Using a directory where the
+  // base DB stores its files for the default CF is not supported.
+  const ColumnFamilyData* const cfd =
+      static_cast<ColumnFamilyHandleImpl*>(DefaultColumnFamily())->cfd();
+  assert(cfd);
+
+  const ImmutableCFOptions* const ioptions = cfd->ioptions();
+  assert(ioptions);
+
+  assert(env_);
+
+  for (const auto& cf_path : ioptions->cf_paths) {
+    bool blob_dir_same_as_cf_dir = false;
+    s = env_->AreFilesSame(blob_dir_, cf_path.path, &blob_dir_same_as_cf_dir);
+    if (!s.ok()) {
+      ROCKS_LOG_ERROR(db_options_.info_log,
+                      "Error while sanitizing blob_dir %s, status: %s",
+                      blob_dir_.c_str(), s.ToString().c_str());
+      return s;
+    }
+
+    if (blob_dir_same_as_cf_dir) {
+      return Status::NotSupported(
+          "Using the base DB's storage directories for BlobDB files is not "
+          "supported.");
+    }
+  }
+
+  // Initialize SST file <-> oldest blob file mapping if garbage collection
+  // is enabled.
+  if (bdb_options_.enable_garbage_collection) {
+    std::vector<LiveFileMetaData> live_files;
+    db_->GetLiveFilesMetaData(&live_files);
+
+    InitializeBlobFileToSstMapping(live_files);
+
+    MarkUnreferencedBlobFilesObsoleteDuringOpen();
+
+    if (!disable_auto_compactions) {
+      s = db_->EnableAutoCompaction(*handles);
+      if (!s.ok()) {
+        ROCKS_LOG_ERROR(
+            db_options_.info_log,
+            "Failed to enable automatic compactions during open, status: %s",
+            s.ToString().c_str());
+        return s;
+      }
+    }
+  }
+
+  // Add trash files in blob dir to file delete scheduler.
+  SstFileManagerImpl* sfm = static_cast<SstFileManagerImpl*>(
+      db_impl_->immutable_db_options().sst_file_manager.get());
+  DeleteScheduler::CleanupDirectory(env_, sfm, blob_dir_);
+
+  UpdateLiveSSTSize();
+
+  // Start background jobs.
+  if (!bdb_options_.disable_background_tasks) {
+    StartBackgroundTasks();
+  }
+
+  ROCKS_LOG_INFO(db_options_.info_log, "BlobDB pointer %p", this);
+  bdb_options_.Dump(db_options_.info_log.get());
+  closed_ = false;
+  return s;
+}
+
+void BlobDBImpl::StartBackgroundTasks() {
+  // store a call to a member function and object
+  tqueue_.add(
+      kReclaimOpenFilesPeriodMillisecs,
+      std::bind(&BlobDBImpl::ReclaimOpenFiles, this, std::placeholders::_1));
+  tqueue_.add(
+      kDeleteObsoleteFilesPeriodMillisecs,
+      std::bind(&BlobDBImpl::DeleteObsoleteFiles, this, std::placeholders::_1));
+  tqueue_.add(kSanityCheckPeriodMillisecs,
+              std::bind(&BlobDBImpl::SanityCheck, this, std::placeholders::_1));
+  tqueue_.add(
+      kEvictExpiredFilesPeriodMillisecs,
+      std::bind(&BlobDBImpl::EvictExpiredFiles, this, std::placeholders::_1));
+}
+
+Status BlobDBImpl::GetAllBlobFiles(std::set<uint64_t>* file_numbers) {
+  assert(file_numbers != nullptr);
+  std::vector<std::string> all_files;
+  Status s = env_->GetChildren(blob_dir_, &all_files);
+  if (!s.ok()) {
+    ROCKS_LOG_ERROR(db_options_.info_log,
+                    "Failed to get list of blob files, status: %s",
+                    s.ToString().c_str());
+    return s;
+  }
+
+  for (const auto& file_name : all_files) {
+    uint64_t file_number;
+    FileType type;
+    bool success = ParseFileName(file_name, &file_number, &type);
+    if (success && type == kBlobFile) {
+      file_numbers->insert(file_number);
+    } else {
+      ROCKS_LOG_WARN(db_options_.info_log,
+                     "Skipping file in blob directory: %s", file_name.c_str());
+    }
+  }
+
+  return s;
+}
+
+Status BlobDBImpl::OpenAllBlobFiles() {
+  std::set<uint64_t> file_numbers;
+  Status s = GetAllBlobFiles(&file_numbers);
+  if (!s.ok()) {
+    return s;
+  }
+
+  if (!file_numbers.empty()) {
+    next_file_number_.store(*file_numbers.rbegin() + 1);
+  }
+
+  std::ostringstream blob_file_oss;
+  std::ostringstream live_imm_oss;
+  std::ostringstream obsolete_file_oss;
+
+  for (auto& file_number : file_numbers) {
+    std::shared_ptr<BlobFile> blob_file = std::make_shared<BlobFile>(
+        this, blob_dir_, file_number, db_options_.info_log.get());
+    blob_file->MarkImmutable(/* sequence */ 0);
+
+    // Read file header and footer
+    Status read_metadata_status =
+        blob_file->ReadMetadata(env_->GetFileSystem(), file_options_);
+    if (read_metadata_status.IsCorruption()) {
+      // Remove incomplete file.
+      if (!obsolete_files_.empty()) {
+        obsolete_file_oss << ", ";
+      }
+      obsolete_file_oss << file_number;
+
+      ObsoleteBlobFile(blob_file, 0 /*obsolete_seq*/, false /*update_size*/);
+      continue;
+    } else if (!read_metadata_status.ok()) {
+      ROCKS_LOG_ERROR(db_options_.info_log,
+                      "Unable to read metadata of blob file %" PRIu64
+                      ", status: '%s'",
+                      file_number, read_metadata_status.ToString().c_str());
+      return read_metadata_status;
+    }
+
+    total_blob_size_ += blob_file->GetFileSize();
+
+    if (!blob_files_.empty()) {
+      blob_file_oss << ", ";
+    }
+    blob_file_oss << file_number;
+
+    blob_files_[file_number] = blob_file;
+
+    if (!blob_file->HasTTL()) {
+      if (!live_imm_non_ttl_blob_files_.empty()) {
+        live_imm_oss << ", ";
+      }
+      live_imm_oss << file_number;
+
+      live_imm_non_ttl_blob_files_[file_number] = blob_file;
+    }
+  }
+
+  ROCKS_LOG_INFO(db_options_.info_log,
+                 "Found %" ROCKSDB_PRIszt " blob files: %s", blob_files_.size(),
+                 blob_file_oss.str().c_str());
+  ROCKS_LOG_INFO(
+      db_options_.info_log, "Found %" ROCKSDB_PRIszt " non-TTL blob files: %s",
+      live_imm_non_ttl_blob_files_.size(), live_imm_oss.str().c_str());
+  ROCKS_LOG_INFO(db_options_.info_log,
+                 "Found %" ROCKSDB_PRIszt
+                 " incomplete or corrupted blob files: %s",
+                 obsolete_files_.size(), obsolete_file_oss.str().c_str());
+  return s;
+}
+
+template <typename Linker>
+void BlobDBImpl::LinkSstToBlobFileImpl(uint64_t sst_file_number,
+                                       uint64_t blob_file_number,
+                                       Linker linker) {
+  assert(bdb_options_.enable_garbage_collection);
+  assert(blob_file_number != kInvalidBlobFileNumber);
+
+  auto it = blob_files_.find(blob_file_number);
+  if (it == blob_files_.end()) {
+    ROCKS_LOG_WARN(db_options_.info_log,
+                   "Blob file %" PRIu64
+                   " not found while trying to link "
+                   "SST file %" PRIu64,
+                   blob_file_number, sst_file_number);
+    return;
+  }
+
+  BlobFile* const blob_file = it->second.get();
+  assert(blob_file);
+
+  linker(blob_file, sst_file_number);
+
+  ROCKS_LOG_INFO(db_options_.info_log,
+                 "Blob file %" PRIu64 " linked to SST file %" PRIu64,
+                 blob_file_number, sst_file_number);
+}
+
+void BlobDBImpl::LinkSstToBlobFile(uint64_t sst_file_number,
+                                   uint64_t blob_file_number) {
+  auto linker = [](BlobFile* blob_file, uint64_t sst_file) {
+    WriteLock file_lock(&blob_file->mutex_);
+    blob_file->LinkSstFile(sst_file);
+  };
+
+  LinkSstToBlobFileImpl(sst_file_number, blob_file_number, linker);
+}
+
+void BlobDBImpl::LinkSstToBlobFileNoLock(uint64_t sst_file_number,
+                                         uint64_t blob_file_number) {
+  auto linker = [](BlobFile* blob_file, uint64_t sst_file) {
+    blob_file->LinkSstFile(sst_file);
+  };
+
+  LinkSstToBlobFileImpl(sst_file_number, blob_file_number, linker);
+}
+
+void BlobDBImpl::UnlinkSstFromBlobFile(uint64_t sst_file_number,
+                                       uint64_t blob_file_number) {
+  assert(bdb_options_.enable_garbage_collection);
+  assert(blob_file_number != kInvalidBlobFileNumber);
+
+  auto it = blob_files_.find(blob_file_number);
+  if (it == blob_files_.end()) {
+    ROCKS_LOG_WARN(db_options_.info_log,
+                   "Blob file %" PRIu64
+                   " not found while trying to unlink "
+                   "SST file %" PRIu64,
+                   blob_file_number, sst_file_number);
+    return;
+  }
+
+  BlobFile* const blob_file = it->second.get();
+  assert(blob_file);
+
+  {
+    WriteLock file_lock(&blob_file->mutex_);
+    blob_file->UnlinkSstFile(sst_file_number);
+  }
+
+  ROCKS_LOG_INFO(db_options_.info_log,
+                 "Blob file %" PRIu64 " unlinked from SST file %" PRIu64,
+                 blob_file_number, sst_file_number);
+}
+
+void BlobDBImpl::InitializeBlobFileToSstMapping(
+    const std::vector<LiveFileMetaData>& live_files) {
+  assert(bdb_options_.enable_garbage_collection);
+
+  for (const auto& live_file : live_files) {
+    const uint64_t sst_file_number = live_file.file_number;
+    const uint64_t blob_file_number = live_file.oldest_blob_file_number;
+
+    if (blob_file_number == kInvalidBlobFileNumber) {
+      continue;
+    }
+
+    LinkSstToBlobFileNoLock(sst_file_number, blob_file_number);
+  }
+}
+
+void BlobDBImpl::ProcessFlushJobInfo(const FlushJobInfo& info) {
+  assert(bdb_options_.enable_garbage_collection);
+
+  WriteLock lock(&mutex_);
+
+  if (info.oldest_blob_file_number != kInvalidBlobFileNumber) {
+    LinkSstToBlobFile(info.file_number, info.oldest_blob_file_number);
+  }
+
+  assert(flush_sequence_ < info.largest_seqno);
+  flush_sequence_ = info.largest_seqno;
+
+  MarkUnreferencedBlobFilesObsolete();
+}
+
+void BlobDBImpl::ProcessCompactionJobInfo(const CompactionJobInfo& info) {
+  assert(bdb_options_.enable_garbage_collection);
+
+  if (!info.status.ok()) {
+    return;
+  }
+
+  // Note: the same SST file may appear in both the input and the output
+  // file list in case of a trivial move. We walk through the two lists
+  // below in a fashion that's similar to merge sort to detect this.
+
+  auto cmp = [](const CompactionFileInfo& lhs, const CompactionFileInfo& rhs) {
+    return lhs.file_number < rhs.file_number;
+  };
+
+  auto inputs = info.input_file_infos;
+  auto iit = inputs.begin();
+  const auto iit_end = inputs.end();
+
+  std::sort(iit, iit_end, cmp);
+
+  auto outputs = info.output_file_infos;
+  auto oit = outputs.begin();
+  const auto oit_end = outputs.end();
+
+  std::sort(oit, oit_end, cmp);
+
+  WriteLock lock(&mutex_);
+
+  while (iit != iit_end && oit != oit_end) {
+    const auto& input = *iit;
+    const auto& output = *oit;
+
+    if (input.file_number == output.file_number) {
+      ++iit;
+      ++oit;
+    } else if (input.file_number < output.file_number) {
+      if (input.oldest_blob_file_number != kInvalidBlobFileNumber) {
+        UnlinkSstFromBlobFile(input.file_number, input.oldest_blob_file_number);
+      }
+
+      ++iit;
+    } else {
+      assert(output.file_number < input.file_number);
+
+      if (output.oldest_blob_file_number != kInvalidBlobFileNumber) {
+        LinkSstToBlobFile(output.file_number, output.oldest_blob_file_number);
+      }
+
+      ++oit;
+    }
+  }
+
+  while (iit != iit_end) {
+    const auto& input = *iit;
+
+    if (input.oldest_blob_file_number != kInvalidBlobFileNumber) {
+      UnlinkSstFromBlobFile(input.file_number, input.oldest_blob_file_number);
+    }
+
+    ++iit;
+  }
+
+  while (oit != oit_end) {
+    const auto& output = *oit;
+
+    if (output.oldest_blob_file_number != kInvalidBlobFileNumber) {
+      LinkSstToBlobFile(output.file_number, output.oldest_blob_file_number);
+    }
+
+    ++oit;
+  }
+
+  MarkUnreferencedBlobFilesObsolete();
+}
+
+bool BlobDBImpl::MarkBlobFileObsoleteIfNeeded(
+    const std::shared_ptr<BlobFile>& blob_file, SequenceNumber obsolete_seq) {
+  assert(blob_file);
+  assert(!blob_file->HasTTL());
+  assert(blob_file->Immutable());
+  assert(bdb_options_.enable_garbage_collection);
+
+  // Note: FIFO eviction could have marked this file obsolete already.
+  if (blob_file->Obsolete()) {
+    return true;
+  }
+
+  // We cannot mark this file (or any higher-numbered files for that matter)
+  // obsolete if it is referenced by any memtables or SSTs. We keep track of
+  // the SSTs explicitly. To account for memtables, we keep track of the highest
+  // sequence number received in flush notifications, and we do not mark the
+  // blob file obsolete if there are still unflushed memtables from before
+  // the time the blob file was closed.
+  if (blob_file->GetImmutableSequence() > flush_sequence_ ||
+      !blob_file->GetLinkedSstFiles().empty()) {
+    return false;
+  }
+
+  ROCKS_LOG_INFO(db_options_.info_log,
+                 "Blob file %" PRIu64 " is no longer needed, marking obsolete",
+                 blob_file->BlobFileNumber());
+
+  ObsoleteBlobFile(blob_file, obsolete_seq, /* update_size */ true);
+  return true;
+}
+
+template <class Functor>
+void BlobDBImpl::MarkUnreferencedBlobFilesObsoleteImpl(Functor mark_if_needed) {
+  assert(bdb_options_.enable_garbage_collection);
+
+  // Iterate through all live immutable non-TTL blob files, and mark them
+  // obsolete assuming no SST files or memtables rely on the blobs in them.
+  // Note: we need to stop as soon as we find a blob file that has any
+  // linked SSTs (or one potentially referenced by memtables).
+
+  uint64_t obsoleted_files = 0;
+
+  auto it = live_imm_non_ttl_blob_files_.begin();
+  while (it != live_imm_non_ttl_blob_files_.end()) {
+    const auto& blob_file = it->second;
+    assert(blob_file);
+    assert(blob_file->BlobFileNumber() == it->first);
+    assert(!blob_file->HasTTL());
+    assert(blob_file->Immutable());
+
+    // Small optimization: Obsolete() does an atomic read, so we can do
+    // this check without taking a lock on the blob file's mutex.
+    if (blob_file->Obsolete()) {
+      it = live_imm_non_ttl_blob_files_.erase(it);
+      continue;
+    }
+
+    if (!mark_if_needed(blob_file)) {
+      break;
+    }
+
+    it = live_imm_non_ttl_blob_files_.erase(it);
+
+    ++obsoleted_files;
+  }
+
+  if (obsoleted_files > 0) {
+    ROCKS_LOG_INFO(db_options_.info_log,
+                   "%" PRIu64 " blob file(s) marked obsolete by GC",
+                   obsoleted_files);
+    RecordTick(statistics_, BLOB_DB_GC_NUM_FILES, obsoleted_files);
+  }
+}
+
+void BlobDBImpl::MarkUnreferencedBlobFilesObsolete() {
+  const SequenceNumber obsolete_seq = GetLatestSequenceNumber();
+
+  MarkUnreferencedBlobFilesObsoleteImpl(
+      [this, obsolete_seq](const std::shared_ptr<BlobFile>& blob_file) {
+        WriteLock file_lock(&blob_file->mutex_);
+        return MarkBlobFileObsoleteIfNeeded(blob_file, obsolete_seq);
+      });
+}
+
+void BlobDBImpl::MarkUnreferencedBlobFilesObsoleteDuringOpen() {
+  MarkUnreferencedBlobFilesObsoleteImpl(
+      [this](const std::shared_ptr<BlobFile>& blob_file) {
+        return MarkBlobFileObsoleteIfNeeded(blob_file, /* obsolete_seq */ 0);
+      });
+}
+
+void BlobDBImpl::CloseRandomAccessLocked(
+    const std::shared_ptr<BlobFile>& bfile) {
+  bfile->CloseRandomAccessLocked();
+  open_file_count_--;
+}
+
+Status BlobDBImpl::GetBlobFileReader(
+    const std::shared_ptr<BlobFile>& blob_file,
+    std::shared_ptr<RandomAccessFileReader>* reader) {
+  assert(reader != nullptr);
+  bool fresh_open = false;
+  Status s = blob_file->GetReader(env_, file_options_, reader, &fresh_open);
+  if (s.ok() && fresh_open) {
+    assert(*reader != nullptr);
+    open_file_count_++;
+  }
+  return s;
+}
+
+std::shared_ptr<BlobFile> BlobDBImpl::NewBlobFile(
+    bool has_ttl, const ExpirationRange& expiration_range,
+    const std::string& reason) {
+  assert(has_ttl == (expiration_range.first || expiration_range.second));
+
+  uint64_t file_num = next_file_number_++;
+
+  const uint32_t column_family_id =
+      static_cast<ColumnFamilyHandleImpl*>(DefaultColumnFamily())->GetID();
+  auto blob_file = std::make_shared<BlobFile>(
+      this, blob_dir_, file_num, db_options_.info_log.get(), column_family_id,
+      bdb_options_.compression, has_ttl, expiration_range);
+
+  ROCKS_LOG_DEBUG(db_options_.info_log, "New blob file created: %s reason='%s'",
+                  blob_file->PathName().c_str(), reason.c_str());
+  LogFlush(db_options_.info_log);
+
+  return blob_file;
+}
+
+void BlobDBImpl::RegisterBlobFile(std::shared_ptr<BlobFile> blob_file) {
+  const uint64_t blob_file_number = blob_file->BlobFileNumber();
+
+  auto it = blob_files_.lower_bound(blob_file_number);
+  assert(it == blob_files_.end() || it->first != blob_file_number);
+
+  blob_files_.insert(it,
+                     std::map<uint64_t, std::shared_ptr<BlobFile>>::value_type(
+                         blob_file_number, std::move(blob_file)));
+}
+
+Status BlobDBImpl::CreateWriterLocked(const std::shared_ptr<BlobFile>& bfile) {
+  std::string fpath(bfile->PathName());
+  std::unique_ptr<FSWritableFile> wfile;
+  const auto& fs = env_->GetFileSystem();
+
+  Status s = fs->ReopenWritableFile(fpath, file_options_, &wfile, nullptr);
+  if (!s.ok()) {
+    ROCKS_LOG_ERROR(db_options_.info_log,
+                    "Failed to open blob file for write: %s status: '%s'"
+                    " exists: '%s'",
+                    fpath.c_str(), s.ToString().c_str(),
+                    fs->FileExists(fpath, file_options_.io_options, nullptr)
+                        .ToString()
+                        .c_str());
+    return s;
+  }
+
+  std::unique_ptr<WritableFileWriter> fwriter;
+  fwriter.reset(new WritableFileWriter(std::move(wfile), fpath, file_options_));
+
+  uint64_t boffset = bfile->GetFileSize();
+  if (debug_level_ >= 2 && boffset) {
+    ROCKS_LOG_DEBUG(db_options_.info_log,
+                    "Open blob file: %s with offset: %" PRIu64, fpath.c_str(),
+                    boffset);
+  }
+
+  BlobLogWriter::ElemType et = BlobLogWriter::kEtNone;
+  if (bfile->file_size_ == BlobLogHeader::kSize) {
+    et = BlobLogWriter::kEtFileHdr;
+  } else if (bfile->file_size_ > BlobLogHeader::kSize) {
+    et = BlobLogWriter::kEtRecord;
+  } else if (bfile->file_size_) {
+    ROCKS_LOG_WARN(db_options_.info_log,
+                   "Open blob file: %s with wrong size: %" PRIu64,
+                   fpath.c_str(), boffset);
+    return Status::Corruption("Invalid blob file size");
+  }
+
+  constexpr bool do_flush = true;
+
+  bfile->log_writer_ = std::make_shared<BlobLogWriter>(
+      std::move(fwriter), clock_, statistics_, bfile->file_number_,
+      db_options_.use_fsync, do_flush, boffset);
+  bfile->log_writer_->last_elem_type_ = et;
+
+  return s;
+}
+
+std::shared_ptr<BlobFile> BlobDBImpl::FindBlobFileLocked(
+    uint64_t expiration) const {
+  if (open_ttl_files_.empty()) {
+    return nullptr;
+  }
+
+  std::shared_ptr<BlobFile> tmp = std::make_shared<BlobFile>();
+  tmp->SetHasTTL(true);
+  tmp->expiration_range_ = std::make_pair(expiration, 0);
+  tmp->file_number_ = std::numeric_limits<uint64_t>::max();
+
+  auto citr = open_ttl_files_.equal_range(tmp);
+  if (citr.first == open_ttl_files_.end()) {
+    assert(citr.second == open_ttl_files_.end());
+
+    std::shared_ptr<BlobFile> check = *(open_ttl_files_.rbegin());
+    return (check->expiration_range_.second <= expiration) ? nullptr : check;
+  }
+
+  if (citr.first != citr.second) {
+    return *(citr.first);
+  }
+
+  auto finditr = citr.second;
+  if (finditr != open_ttl_files_.begin()) {
+    --finditr;
+  }
+
+  bool b2 = (*finditr)->expiration_range_.second <= expiration;
+  bool b1 = (*finditr)->expiration_range_.first > expiration;
+
+  return (b1 || b2) ? nullptr : (*finditr);
+}
+
+Status BlobDBImpl::CheckOrCreateWriterLocked(
+    const std::shared_ptr<BlobFile>& blob_file,
+    std::shared_ptr<BlobLogWriter>* writer) {
+  assert(writer != nullptr);
+  *writer = blob_file->GetWriter();
+  if (*writer != nullptr) {
+    return Status::OK();
+  }
+  Status s = CreateWriterLocked(blob_file);
+  if (s.ok()) {
+    *writer = blob_file->GetWriter();
+  }
+  return s;
+}
+
+Status BlobDBImpl::CreateBlobFileAndWriter(
+    bool has_ttl, const ExpirationRange& expiration_range,
+    const std::string& reason, std::shared_ptr<BlobFile>* blob_file,
+    std::shared_ptr<BlobLogWriter>* writer) {
+  TEST_SYNC_POINT("BlobDBImpl::CreateBlobFileAndWriter");
+  assert(has_ttl == (expiration_range.first || expiration_range.second));
+  assert(blob_file);
+  assert(writer);
+
+  *blob_file = NewBlobFile(has_ttl, expiration_range, reason);
+  assert(*blob_file);
+
+  // file not visible, hence no lock
+  Status s = CheckOrCreateWriterLocked(*blob_file, writer);
+  if (!s.ok()) {
+    ROCKS_LOG_ERROR(db_options_.info_log,
+                    "Failed to get writer for blob file: %s, error: %s",
+                    (*blob_file)->PathName().c_str(), s.ToString().c_str());
+    return s;
+  }
+
+  assert(*writer);
+
+  s = (*writer)->WriteHeader((*blob_file)->header_);
+  if (!s.ok()) {
+    ROCKS_LOG_ERROR(db_options_.info_log,
+                    "Failed to write header to new blob file: %s"
+                    " status: '%s'",
+                    (*blob_file)->PathName().c_str(), s.ToString().c_str());
+    return s;
+  }
+
+  (*blob_file)->SetFileSize(BlobLogHeader::kSize);
+  total_blob_size_ += BlobLogHeader::kSize;
+
+  return s;
+}
+
+Status BlobDBImpl::SelectBlobFile(std::shared_ptr<BlobFile>* blob_file) {
+  assert(blob_file);
+
+  {
+    ReadLock rl(&mutex_);
+
+    if (open_non_ttl_file_) {
+      assert(!open_non_ttl_file_->Immutable());
+      *blob_file = open_non_ttl_file_;
+      return Status::OK();
+    }
+  }
+
+  // Check again
+  WriteLock wl(&mutex_);
+
+  if (open_non_ttl_file_) {
+    assert(!open_non_ttl_file_->Immutable());
+    *blob_file = open_non_ttl_file_;
+    return Status::OK();
+  }
+
+  std::shared_ptr<BlobLogWriter> writer;
+  const Status s = CreateBlobFileAndWriter(
+      /* has_ttl */ false, ExpirationRange(),
+      /* reason */ "SelectBlobFile", blob_file, &writer);
+  if (!s.ok()) {
+    return s;
+  }
+
+  RegisterBlobFile(*blob_file);
+  open_non_ttl_file_ = *blob_file;
+
+  return s;
+}
+
+Status BlobDBImpl::SelectBlobFileTTL(uint64_t expiration,
+                                     std::shared_ptr<BlobFile>* blob_file) {
+  assert(blob_file);
+  assert(expiration != kNoExpiration);
+
+  {
+    ReadLock rl(&mutex_);
+
+    *blob_file = FindBlobFileLocked(expiration);
+    if (*blob_file != nullptr) {
+      assert(!(*blob_file)->Immutable());
+      return Status::OK();
+    }
+  }
+
+  // Check again
+  WriteLock wl(&mutex_);
+
+  *blob_file = FindBlobFileLocked(expiration);
+  if (*blob_file != nullptr) {
+    assert(!(*blob_file)->Immutable());
+    return Status::OK();
+  }
+
+  const uint64_t exp_low =
+      (expiration / bdb_options_.ttl_range_secs) * bdb_options_.ttl_range_secs;
+  const uint64_t exp_high = exp_low + bdb_options_.ttl_range_secs;
+  const ExpirationRange expiration_range(exp_low, exp_high);
+
+  std::ostringstream oss;
+  oss << "SelectBlobFileTTL range: [" << exp_low << ',' << exp_high << ')';
+
+  std::shared_ptr<BlobLogWriter> writer;
+  const Status s =
+      CreateBlobFileAndWriter(/* has_ttl */ true, expiration_range,
+                              /* reason */ oss.str(), blob_file, &writer);
+  if (!s.ok()) {
+    return s;
+  }
+
+  RegisterBlobFile(*blob_file);
+  open_ttl_files_.insert(*blob_file);
+
+  return s;
+}
+
+class BlobDBImpl::BlobInserter : public WriteBatch::Handler {
+ private:
+  const WriteOptions& options_;
+  BlobDBImpl* blob_db_impl_;
+  uint32_t default_cf_id_;
+  WriteBatch batch_;
+
+ public:
+  BlobInserter(const WriteOptions& options, BlobDBImpl* blob_db_impl,
+               uint32_t default_cf_id)
+      : options_(options),
+        blob_db_impl_(blob_db_impl),
+        default_cf_id_(default_cf_id) {}
+
+  WriteBatch* batch() { return &batch_; }
+
+  Status PutCF(uint32_t column_family_id, const Slice& key,
+               const Slice& value) override {
+    if (column_family_id != default_cf_id_) {
+      return Status::NotSupported(
+          "Blob DB doesn't support non-default column family.");
+    }
+    Status s = blob_db_impl_->PutBlobValue(options_, key, value, kNoExpiration,
+                                           &batch_);
+    return s;
+  }
+
+  Status DeleteCF(uint32_t column_family_id, const Slice& key) override {
+    if (column_family_id != default_cf_id_) {
+      return Status::NotSupported(
+          "Blob DB doesn't support non-default column family.");
+    }
+    Status s = WriteBatchInternal::Delete(&batch_, column_family_id, key);
+    return s;
+  }
+
+  virtual Status DeleteRange(uint32_t column_family_id, const Slice& begin_key,
+                             const Slice& end_key) {
+    if (column_family_id != default_cf_id_) {
+      return Status::NotSupported(
+          "Blob DB doesn't support non-default column family.");
+    }
+    Status s = WriteBatchInternal::DeleteRange(&batch_, column_family_id,
+                                               begin_key, end_key);
+    return s;
+  }
+
+  Status SingleDeleteCF(uint32_t /*column_family_id*/,
+                        const Slice& /*key*/) override {
+    return Status::NotSupported("Not supported operation in blob db.");
+  }
+
+  Status MergeCF(uint32_t /*column_family_id*/, const Slice& /*key*/,
+                 const Slice& /*value*/) override {
+    return Status::NotSupported("Not supported operation in blob db.");
+  }
+
+  void LogData(const Slice& blob) override { batch_.PutLogData(blob); }
+};
+
+Status BlobDBImpl::Write(const WriteOptions& options, WriteBatch* updates) {
+  StopWatch write_sw(clock_, statistics_, BLOB_DB_WRITE_MICROS);
+  RecordTick(statistics_, BLOB_DB_NUM_WRITE);
+  uint32_t default_cf_id =
+      static_cast_with_check<ColumnFamilyHandleImpl>(DefaultColumnFamily())
+          ->GetID();
+  Status s;
+  BlobInserter blob_inserter(options, this, default_cf_id);
+  {
+    // Release write_mutex_ before DB write to avoid race condition with
+    // flush begin listener, which also require write_mutex_ to sync
+    // blob files.
+    MutexLock l(&write_mutex_);
+    s = updates->Iterate(&blob_inserter);
+  }
+  if (!s.ok()) {
+    return s;
+  }
+  return db_->Write(options, blob_inserter.batch());
+}
+
+Status BlobDBImpl::Put(const WriteOptions& options, const Slice& key,
+                       const Slice& value) {
+  return PutUntil(options, key, value, kNoExpiration);
+}
+
+Status BlobDBImpl::PutWithTTL(const WriteOptions& options, const Slice& key,
+                              const Slice& value, uint64_t ttl) {
+  uint64_t now = EpochNow();
+  uint64_t expiration = kNoExpiration - now > ttl ? now + ttl : kNoExpiration;
+  return PutUntil(options, key, value, expiration);
+}
+
+Status BlobDBImpl::PutUntil(const WriteOptions& options, const Slice& key,
+                            const Slice& value, uint64_t expiration) {
+  StopWatch write_sw(clock_, statistics_, BLOB_DB_WRITE_MICROS);
+  RecordTick(statistics_, BLOB_DB_NUM_PUT);
+  Status s;
+  WriteBatch batch;
+  {
+    // Release write_mutex_ before DB write to avoid race condition with
+    // flush begin listener, which also require write_mutex_ to sync
+    // blob files.
+    MutexLock l(&write_mutex_);
+    s = PutBlobValue(options, key, value, expiration, &batch);
+  }
+  if (s.ok()) {
+    s = db_->Write(options, &batch);
+  }
+  return s;
+}
+
+Status BlobDBImpl::PutBlobValue(const WriteOptions& /*options*/,
+                                const Slice& key, const Slice& value,
+                                uint64_t expiration, WriteBatch* batch) {
+  write_mutex_.AssertHeld();
+  Status s;
+  std::string index_entry;
+  uint32_t column_family_id =
+      static_cast_with_check<ColumnFamilyHandleImpl>(DefaultColumnFamily())
+          ->GetID();
+  if (value.size() < bdb_options_.min_blob_size) {
+    if (expiration == kNoExpiration) {
+      // Put as normal value
+      s = batch->Put(key, value);
+      RecordTick(statistics_, BLOB_DB_WRITE_INLINED);
+    } else {
+      // Inlined with TTL
+      BlobIndex::EncodeInlinedTTL(&index_entry, expiration, value);
+      s = WriteBatchInternal::PutBlobIndex(batch, column_family_id, key,
+                                           index_entry);
+      RecordTick(statistics_, BLOB_DB_WRITE_INLINED_TTL);
+    }
+  } else {
+    std::string compression_output;
+    Slice value_compressed = GetCompressedSlice(value, &compression_output);
+
+    std::string headerbuf;
+    BlobLogWriter::ConstructBlobHeader(&headerbuf, key, value_compressed,
+                                       expiration);
+
+    // Check DB size limit before selecting blob file to
+    // Since CheckSizeAndEvictBlobFiles() can close blob files, it needs to be
+    // done before calling SelectBlobFile().
+    s = CheckSizeAndEvictBlobFiles(headerbuf.size() + key.size() +
+                                   value_compressed.size());
+    if (!s.ok()) {
+      return s;
+    }
+
+    std::shared_ptr<BlobFile> blob_file;
+    if (expiration != kNoExpiration) {
+      s = SelectBlobFileTTL(expiration, &blob_file);
+    } else {
+      s = SelectBlobFile(&blob_file);
+    }
+    if (s.ok()) {
+      assert(blob_file != nullptr);
+      assert(blob_file->GetCompressionType() == bdb_options_.compression);
+      s = AppendBlob(blob_file, headerbuf, key, value_compressed, expiration,
+                     &index_entry);
+    }
+    if (s.ok()) {
+      if (expiration != kNoExpiration) {
+        WriteLock file_lock(&blob_file->mutex_);
+        blob_file->ExtendExpirationRange(expiration);
+      }
+      s = CloseBlobFileIfNeeded(blob_file);
+    }
+    if (s.ok()) {
+      s = WriteBatchInternal::PutBlobIndex(batch, column_family_id, key,
+                                           index_entry);
+    }
+    if (s.ok()) {
+      if (expiration == kNoExpiration) {
+        RecordTick(statistics_, BLOB_DB_WRITE_BLOB);
+      } else {
+        RecordTick(statistics_, BLOB_DB_WRITE_BLOB_TTL);
+      }
+    } else {
+      ROCKS_LOG_ERROR(
+          db_options_.info_log,
+          "Failed to append blob to FILE: %s: KEY: %s VALSZ: %" ROCKSDB_PRIszt
+          " status: '%s' blob_file: '%s'",
+          blob_file->PathName().c_str(), key.ToString().c_str(), value.size(),
+          s.ToString().c_str(), blob_file->DumpState().c_str());
+    }
+  }
+
+  RecordTick(statistics_, BLOB_DB_NUM_KEYS_WRITTEN);
+  RecordTick(statistics_, BLOB_DB_BYTES_WRITTEN, key.size() + value.size());
+  RecordInHistogram(statistics_, BLOB_DB_KEY_SIZE, key.size());
+  RecordInHistogram(statistics_, BLOB_DB_VALUE_SIZE, value.size());
+
+  return s;
+}
+
+Slice BlobDBImpl::GetCompressedSlice(const Slice& raw,
+                                     std::string* compression_output) const {
+  if (bdb_options_.compression == kNoCompression) {
+    return raw;
+  }
+  StopWatch compression_sw(clock_, statistics_, BLOB_DB_COMPRESSION_MICROS);
+  CompressionType type = bdb_options_.compression;
+  CompressionOptions opts;
+  CompressionContext context(type);
+  CompressionInfo info(opts, context, CompressionDict::GetEmptyDict(), type,
+                       0 /* sample_for_compression */);
+  CompressBlock(raw, info, &type, kBlockBasedTableVersionFormat, false,
+                compression_output, nullptr, nullptr);
+  return *compression_output;
+}
+
+Status BlobDBImpl::DecompressSlice(const Slice& compressed_value,
+                                   CompressionType compression_type,
+                                   PinnableSlice* value_output) const {
+  assert(compression_type != kNoCompression);
+
+  BlockContents contents;
+  auto cfh = static_cast<ColumnFamilyHandleImpl*>(DefaultColumnFamily());
+
+  {
+    StopWatch decompression_sw(clock_, statistics_,
+                               BLOB_DB_DECOMPRESSION_MICROS);
+    UncompressionContext context(compression_type);
+    UncompressionInfo info(context, UncompressionDict::GetEmptyDict(),
+                           compression_type);
+    Status s = UncompressBlockData(
+        info, compressed_value.data(), compressed_value.size(), &contents,
+        kBlockBasedTableVersionFormat, *(cfh->cfd()->ioptions()));
+    if (!s.ok()) {
+      return Status::Corruption("Unable to decompress blob.");
+    }
+  }
+
+  value_output->PinSelf(contents.data);
+
+  return Status::OK();
+}
+
+Status BlobDBImpl::CompactFiles(
+    const CompactionOptions& compact_options,
+    const std::vector<std::string>& input_file_names, const int output_level,
+    const int output_path_id, std::vector<std::string>* const output_file_names,
+    CompactionJobInfo* compaction_job_info) {
+  // Note: we need CompactionJobInfo to be able to track updates to the
+  // blob file <-> SST mappings, so we provide one if the user hasn't,
+  // assuming that GC is enabled.
+  CompactionJobInfo info{};
+  if (bdb_options_.enable_garbage_collection && !compaction_job_info) {
+    compaction_job_info = &info;
+  }
+
+  const Status s =
+      db_->CompactFiles(compact_options, input_file_names, output_level,
+                        output_path_id, output_file_names, compaction_job_info);
+  if (!s.ok()) {
+    return s;
+  }
+
+  if (bdb_options_.enable_garbage_collection) {
+    assert(compaction_job_info);
+    ProcessCompactionJobInfo(*compaction_job_info);
+  }
+
+  return s;
+}
+
+void BlobDBImpl::GetCompactionContextCommon(BlobCompactionContext* context) {
+  assert(context);
+
+  context->blob_db_impl = this;
+  context->next_file_number = next_file_number_.load();
+  context->current_blob_files.clear();
+  for (auto& p : blob_files_) {
+    context->current_blob_files.insert(p.first);
+  }
+  context->fifo_eviction_seq = fifo_eviction_seq_;
+  context->evict_expiration_up_to = evict_expiration_up_to_;
+}
+
+void BlobDBImpl::GetCompactionContext(BlobCompactionContext* context) {
+  assert(context);
+
+  ReadLock l(&mutex_);
+  GetCompactionContextCommon(context);
+}
+
+void BlobDBImpl::GetCompactionContext(BlobCompactionContext* context,
+                                      BlobCompactionContextGC* context_gc) {
+  assert(context);
+  assert(context_gc);
+
+  ReadLock l(&mutex_);
+  GetCompactionContextCommon(context);
+
+  if (!live_imm_non_ttl_blob_files_.empty()) {
+    auto it = live_imm_non_ttl_blob_files_.begin();
+    std::advance(it, bdb_options_.garbage_collection_cutoff *
+                         live_imm_non_ttl_blob_files_.size());
+    context_gc->cutoff_file_number = it != live_imm_non_ttl_blob_files_.end()
+                                         ? it->first
+                                         : std::numeric_limits<uint64_t>::max();
+  }
+}
+
+void BlobDBImpl::UpdateLiveSSTSize() {
+  uint64_t live_sst_size = 0;
+  bool ok = GetIntProperty(DB::Properties::kLiveSstFilesSize, &live_sst_size);
+  if (ok) {
+    live_sst_size_.store(live_sst_size);
+    ROCKS_LOG_INFO(db_options_.info_log,
+                   "Updated total SST file size: %" PRIu64 " bytes.",
+                   live_sst_size);
+  } else {
+    ROCKS_LOG_ERROR(
+        db_options_.info_log,
+        "Failed to update total SST file size after flush or compaction.");
+  }
+  {
+    // Trigger FIFO eviction if needed.
+    MutexLock l(&write_mutex_);
+    Status s = CheckSizeAndEvictBlobFiles(0, true /*force*/);
+    if (s.IsNoSpace()) {
+      ROCKS_LOG_WARN(db_options_.info_log,
+                     "DB grow out-of-space after SST size updated. Current live"
+                     " SST size: %" PRIu64
+                     " , current blob files size: %" PRIu64 ".",
+                     live_sst_size_.load(), total_blob_size_.load());
+    }
+  }
+}
+
+Status BlobDBImpl::CheckSizeAndEvictBlobFiles(uint64_t blob_size,
+                                              bool force_evict) {
+  write_mutex_.AssertHeld();
+
+  uint64_t live_sst_size = live_sst_size_.load();
+  if (bdb_options_.max_db_size == 0 ||
+      live_sst_size + total_blob_size_.load() + blob_size <=
+          bdb_options_.max_db_size) {
+    return Status::OK();
+  }
+
+  if (bdb_options_.is_fifo == false ||
+      (!force_evict && live_sst_size + blob_size > bdb_options_.max_db_size)) {
+    // FIFO eviction is disabled, or no space to insert new blob even we evict
+    // all blob files.
+    return Status::NoSpace(
+        "Write failed, as writing it would exceed max_db_size limit.");
+  }
+
+  std::vector<std::shared_ptr<BlobFile>> candidate_files;
+  CopyBlobFiles(&candidate_files);
+  std::sort(candidate_files.begin(), candidate_files.end(),
+            BlobFileComparator());
+  fifo_eviction_seq_ = GetLatestSequenceNumber();
+
+  WriteLock l(&mutex_);
+
+  while (!candidate_files.empty() &&
+         live_sst_size + total_blob_size_.load() + blob_size >
+             bdb_options_.max_db_size) {
+    std::shared_ptr<BlobFile> blob_file = candidate_files.back();
+    candidate_files.pop_back();
+    WriteLock file_lock(&blob_file->mutex_);
+    if (blob_file->Obsolete()) {
+      // File already obsoleted by someone else.
+      assert(blob_file->Immutable());
+      continue;
+    }
+    // FIFO eviction can evict open blob files.
+    if (!blob_file->Immutable()) {
+      Status s = CloseBlobFile(blob_file);
+      if (!s.ok()) {
+        return s;
+      }
+    }
+    assert(blob_file->Immutable());
+    auto expiration_range = blob_file->GetExpirationRange();
+    ROCKS_LOG_INFO(db_options_.info_log,
+                   "Evict oldest blob file since DB out of space. Current "
+                   "live SST file size: %" PRIu64 ", total blob size: %" PRIu64
+                   ", max db size: %" PRIu64 ", evicted blob file #%" PRIu64
+                   ".",
+                   live_sst_size, total_blob_size_.load(),
+                   bdb_options_.max_db_size, blob_file->BlobFileNumber());
+    ObsoleteBlobFile(blob_file, fifo_eviction_seq_, true /*update_size*/);
+    evict_expiration_up_to_ = expiration_range.first;
+    RecordTick(statistics_, BLOB_DB_FIFO_NUM_FILES_EVICTED);
+    RecordTick(statistics_, BLOB_DB_FIFO_NUM_KEYS_EVICTED,
+               blob_file->BlobCount());
+    RecordTick(statistics_, BLOB_DB_FIFO_BYTES_EVICTED,
+               blob_file->GetFileSize());
+    TEST_SYNC_POINT("BlobDBImpl::EvictOldestBlobFile:Evicted");
+  }
+  if (live_sst_size + total_blob_size_.load() + blob_size >
+      bdb_options_.max_db_size) {
+    return Status::NoSpace(
+        "Write failed, as writing it would exceed max_db_size limit.");
+  }
+  return Status::OK();
+}
+
+Status BlobDBImpl::AppendBlob(const std::shared_ptr<BlobFile>& bfile,
+                              const std::string& headerbuf, const Slice& key,
+                              const Slice& value, uint64_t expiration,
+                              std::string* index_entry) {
+  Status s;
+  uint64_t blob_offset = 0;
+  uint64_t key_offset = 0;
+  {
+    WriteLock lockbfile_w(&bfile->mutex_);
+    std::shared_ptr<BlobLogWriter> writer;
+    s = CheckOrCreateWriterLocked(bfile, &writer);
+    if (!s.ok()) {
+      return s;
+    }
+
+    // write the blob to the blob log.
+    s = writer->EmitPhysicalRecord(headerbuf, key, value, &key_offset,
+                                   &blob_offset);
+  }
+
+  if (!s.ok()) {
+    ROCKS_LOG_ERROR(db_options_.info_log,
+                    "Invalid status in AppendBlob: %s status: '%s'",
+                    bfile->PathName().c_str(), s.ToString().c_str());
+    return s;
+  }
+
+  uint64_t size_put = headerbuf.size() + key.size() + value.size();
+  bfile->BlobRecordAdded(size_put);
+  total_blob_size_ += size_put;
+
+  if (expiration == kNoExpiration) {
+    BlobIndex::EncodeBlob(index_entry, bfile->BlobFileNumber(), blob_offset,
+                          value.size(), bdb_options_.compression);
+  } else {
+    BlobIndex::EncodeBlobTTL(index_entry, expiration, bfile->BlobFileNumber(),
+                             blob_offset, value.size(),
+                             bdb_options_.compression);
+  }
+
+  return s;
+}
+
+std::vector<Status> BlobDBImpl::MultiGet(const ReadOptions& read_options,
+                                         const std::vector<Slice>& keys,
+                                         std::vector<std::string>* values) {
+  StopWatch multiget_sw(clock_, statistics_, BLOB_DB_MULTIGET_MICROS);
+  RecordTick(statistics_, BLOB_DB_NUM_MULTIGET);
+  // Get a snapshot to avoid blob file get deleted between we
+  // fetch and index entry and reading from the file.
+  ReadOptions ro(read_options);
+  bool snapshot_created = SetSnapshotIfNeeded(&ro);
+
+  std::vector<Status> statuses;
+  statuses.reserve(keys.size());
+  values->clear();
+  values->reserve(keys.size());
+  PinnableSlice value;
+  for (size_t i = 0; i < keys.size(); i++) {
+    statuses.push_back(Get(ro, DefaultColumnFamily(), keys[i], &value));
+    values->push_back(value.ToString());
+    value.Reset();
+  }
+  if (snapshot_created) {
+    db_->ReleaseSnapshot(ro.snapshot);
+  }
+  return statuses;
+}
+
+bool BlobDBImpl::SetSnapshotIfNeeded(ReadOptions* read_options) {
+  assert(read_options != nullptr);
+  if (read_options->snapshot != nullptr) {
+    return false;
+  }
+  read_options->snapshot = db_->GetSnapshot();
+  return true;
+}
+
+Status BlobDBImpl::GetBlobValue(const Slice& key, const Slice& index_entry,
+                                PinnableSlice* value, uint64_t* expiration) {
+  assert(value);
+
+  BlobIndex blob_index;
+  Status s = blob_index.DecodeFrom(index_entry);
+  if (!s.ok()) {
+    return s;
+  }
+
+  if (blob_index.HasTTL() && blob_index.expiration() <= EpochNow()) {
+    return Status::NotFound("Key expired");
+  }
+
+  if (expiration != nullptr) {
+    if (blob_index.HasTTL()) {
+      *expiration = blob_index.expiration();
+    } else {
+      *expiration = kNoExpiration;
+    }
+  }
+
+  if (blob_index.IsInlined()) {
+    // TODO(yiwu): If index_entry is a PinnableSlice, we can also pin the same
+    // memory buffer to avoid extra copy.
+    value->PinSelf(blob_index.value());
+    return Status::OK();
+  }
+
+  CompressionType compression_type = kNoCompression;
+  s = GetRawBlobFromFile(key, blob_index.file_number(), blob_index.offset(),
+                         blob_index.size(), value, &compression_type);
+  if (!s.ok()) {
+    return s;
+  }
+
+  if (compression_type != kNoCompression) {
+    s = DecompressSlice(*value, compression_type, value);
+    if (!s.ok()) {
+      if (debug_level_ >= 2) {
+        ROCKS_LOG_ERROR(
+            db_options_.info_log,
+            "Uncompression error during blob read from file: %" PRIu64
+            " blob_offset: %" PRIu64 " blob_size: %" PRIu64
+            " key: %s status: '%s'",
+            blob_index.file_number(), blob_index.offset(), blob_index.size(),
+            key.ToString(/* output_hex */ true).c_str(), s.ToString().c_str());
+      }
+      return s;
+    }
+  }
+
+  return Status::OK();
+}
+
+Status BlobDBImpl::GetRawBlobFromFile(const Slice& key, uint64_t file_number,
+                                      uint64_t offset, uint64_t size,
+                                      PinnableSlice* value,
+                                      CompressionType* compression_type) {
+  assert(value);
+  assert(compression_type);
+  assert(*compression_type == kNoCompression);
+
+  if (!size) {
+    value->PinSelf("");
+    return Status::OK();
+  }
+
+  // offset has to have certain min, as we will read CRC
+  // later from the Blob Header, which needs to be also a
+  // valid offset.
+  if (offset <
+      (BlobLogHeader::kSize + BlobLogRecord::kHeaderSize + key.size())) {
+    if (debug_level_ >= 2) {
+      ROCKS_LOG_ERROR(db_options_.info_log,
+                      "Invalid blob index file_number: %" PRIu64
+                      " blob_offset: %" PRIu64 " blob_size: %" PRIu64
+                      " key: %s",
+                      file_number, offset, size,
+                      key.ToString(/* output_hex */ true).c_str());
+    }
+
+    return Status::NotFound("Invalid blob offset");
+  }
+
+  std::shared_ptr<BlobFile> blob_file;
+
+  {
+    ReadLock rl(&mutex_);
+    auto it = blob_files_.find(file_number);
+
+    // file was deleted
+    if (it == blob_files_.end()) {
+      return Status::NotFound("Blob Not Found as blob file missing");
+    }
+
+    blob_file = it->second;
+  }
+
+  *compression_type = blob_file->GetCompressionType();
+
+  // takes locks when called
+  std::shared_ptr<RandomAccessFileReader> reader;
+  Status s = GetBlobFileReader(blob_file, &reader);
+  if (!s.ok()) {
+    return s;
+  }
+
+  assert(offset >= key.size() + sizeof(uint32_t));
+  const uint64_t record_offset = offset - key.size() - sizeof(uint32_t);
+  const uint64_t record_size = sizeof(uint32_t) + key.size() + size;
+
+  // Allocate the buffer. This is safe in C++11
+  std::string buf;
+  AlignedBuf aligned_buf;
+
+  // A partial blob record contain checksum, key and value.
+  Slice blob_record;
+
+  {
+    StopWatch read_sw(clock_, statistics_, BLOB_DB_BLOB_FILE_READ_MICROS);
+    // TODO: rate limit old blob DB file reads.
+    if (reader->use_direct_io()) {
+      s = reader->Read(IOOptions(), record_offset,
+                       static_cast<size_t>(record_size), &blob_record, nullptr,
+                       &aligned_buf, Env::IO_TOTAL /* rate_limiter_priority */);
+    } else {
+      buf.reserve(static_cast<size_t>(record_size));
+      s = reader->Read(IOOptions(), record_offset,
+                       static_cast<size_t>(record_size), &blob_record, &buf[0],
+                       nullptr, Env::IO_TOTAL /* rate_limiter_priority */);
+    }
+    RecordTick(statistics_, BLOB_DB_BLOB_FILE_BYTES_READ, blob_record.size());
+  }
+
+  if (!s.ok()) {
+    ROCKS_LOG_DEBUG(
+        db_options_.info_log,
+        "Failed to read blob from blob file %" PRIu64 ", blob_offset: %" PRIu64
+        ", blob_size: %" PRIu64 ", key_size: %" ROCKSDB_PRIszt ", status: '%s'",
+        file_number, offset, size, key.size(), s.ToString().c_str());
+    return s;
+  }
+
+  if (blob_record.size() != record_size) {
+    ROCKS_LOG_DEBUG(
+        db_options_.info_log,
+        "Failed to read blob from blob file %" PRIu64 ", blob_offset: %" PRIu64
+        ", blob_size: %" PRIu64 ", key_size: %" ROCKSDB_PRIszt
+        ", read %" ROCKSDB_PRIszt " bytes, expected %" PRIu64 " bytes",
+        file_number, offset, size, key.size(), blob_record.size(), record_size);
+
+    return Status::Corruption("Failed to retrieve blob from blob index.");
+  }
+
+  Slice crc_slice(blob_record.data(), sizeof(uint32_t));
+  Slice blob_value(blob_record.data() + sizeof(uint32_t) + key.size(),
+                   static_cast<size_t>(size));
+
+  uint32_t crc_exp = 0;
+  if (!GetFixed32(&crc_slice, &crc_exp)) {
+    ROCKS_LOG_DEBUG(
+        db_options_.info_log,
+        "Unable to decode CRC from blob file %" PRIu64 ", blob_offset: %" PRIu64
+        ", blob_size: %" PRIu64 ", key size: %" ROCKSDB_PRIszt ", status: '%s'",
+        file_number, offset, size, key.size(), s.ToString().c_str());
+    return Status::Corruption("Unable to decode checksum.");
+  }
+
+  uint32_t crc = crc32c::Value(blob_record.data() + sizeof(uint32_t),
+                               blob_record.size() - sizeof(uint32_t));
+  crc = crc32c::Mask(crc);  // Adjust for storage
+  if (crc != crc_exp) {
+    if (debug_level_ >= 2) {
+      ROCKS_LOG_ERROR(
+          db_options_.info_log,
+          "Blob crc mismatch file: %" PRIu64 " blob_offset: %" PRIu64
+          " blob_size: %" PRIu64 " key: %s status: '%s'",
+          file_number, offset, size,
+          key.ToString(/* output_hex */ true).c_str(), s.ToString().c_str());
+    }
+
+    return Status::Corruption("Corruption. Blob CRC mismatch");
+  }
+
+  value->PinSelf(blob_value);
+
+  return Status::OK();
+}
+
+Status BlobDBImpl::Get(const ReadOptions& read_options,
+                       ColumnFamilyHandle* column_family, const Slice& key,
+                       PinnableSlice* value) {
+  return Get(read_options, column_family, key, value,
+             static_cast<uint64_t*>(nullptr) /*expiration*/);
+}
+
+Status BlobDBImpl::Get(const ReadOptions& read_options,
+                       ColumnFamilyHandle* column_family, const Slice& key,
+                       PinnableSlice* value, uint64_t* expiration) {
+  StopWatch get_sw(clock_, statistics_, BLOB_DB_GET_MICROS);
+  RecordTick(statistics_, BLOB_DB_NUM_GET);
+  return GetImpl(read_options, column_family, key, value, expiration);
+}
+
+Status BlobDBImpl::GetImpl(const ReadOptions& read_options,
+                           ColumnFamilyHandle* column_family, const Slice& key,
+                           PinnableSlice* value, uint64_t* expiration) {
+  if (column_family->GetID() != DefaultColumnFamily()->GetID()) {
+    return Status::NotSupported(
+        "Blob DB doesn't support non-default column family.");
+  }
+  // Get a snapshot to avoid blob file get deleted between we
+  // fetch and index entry and reading from the file.
+  // TODO(yiwu): For Get() retry if file not found would be a simpler strategy.
+  ReadOptions ro(read_options);
+  bool snapshot_created = SetSnapshotIfNeeded(&ro);
+
+  PinnableSlice index_entry;
+  Status s;
+  bool is_blob_index = false;
+  DBImpl::GetImplOptions get_impl_options;
+  get_impl_options.column_family = column_family;
+  get_impl_options.value = &index_entry;
+  get_impl_options.is_blob_index = &is_blob_index;
+  s = db_impl_->GetImpl(ro, key, get_impl_options);
+  if (expiration != nullptr) {
+    *expiration = kNoExpiration;
+  }
+  RecordTick(statistics_, BLOB_DB_NUM_KEYS_READ);
+  if (s.ok()) {
+    if (is_blob_index) {
+      s = GetBlobValue(key, index_entry, value, expiration);
+    } else {
+      // The index entry is the value itself in this case.
+      value->PinSelf(index_entry);
+    }
+    RecordTick(statistics_, BLOB_DB_BYTES_READ, value->size());
+  }
+  if (snapshot_created) {
+    db_->ReleaseSnapshot(ro.snapshot);
+  }
+  return s;
+}
+
+std::pair<bool, int64_t> BlobDBImpl::SanityCheck(bool aborted) {
+  if (aborted) {
+    return std::make_pair(false, -1);
+  }
+
+  ReadLock rl(&mutex_);
+
+  ROCKS_LOG_INFO(db_options_.info_log, "Starting Sanity Check");
+  ROCKS_LOG_INFO(db_options_.info_log, "Number of files %" ROCKSDB_PRIszt,
+                 blob_files_.size());
+  ROCKS_LOG_INFO(db_options_.info_log, "Number of open files %" ROCKSDB_PRIszt,
+                 open_ttl_files_.size());
+
+  for (const auto& blob_file : open_ttl_files_) {
+    (void)blob_file;
+    assert(!blob_file->Immutable());
+  }
+
+  for (const auto& pair : live_imm_non_ttl_blob_files_) {
+    const auto& blob_file = pair.second;
+    (void)blob_file;
+    assert(!blob_file->HasTTL());
+    assert(blob_file->Immutable());
+  }
+
+  uint64_t now = EpochNow();
+
+  for (auto blob_file_pair : blob_files_) {
+    auto blob_file = blob_file_pair.second;
+    std::ostringstream buf;
+
+    buf << "Blob file " << blob_file->BlobFileNumber() << ", size "
+        << blob_file->GetFileSize() << ", blob count " << blob_file->BlobCount()
+        << ", immutable " << blob_file->Immutable();
+
+    if (blob_file->HasTTL()) {
+      ExpirationRange expiration_range;
+      {
+        ReadLock file_lock(&blob_file->mutex_);
+        expiration_range = blob_file->GetExpirationRange();
+      }
+      buf << ", expiration range (" << expiration_range.first << ", "
+          << expiration_range.second << ")";
+
+      if (!blob_file->Obsolete()) {
+        buf << ", expire in " << (expiration_range.second - now) << "seconds";
+      }
+    }
+    if (blob_file->Obsolete()) {
+      buf << ", obsolete at " << blob_file->GetObsoleteSequence();
+    }
+    buf << ".";
+    ROCKS_LOG_INFO(db_options_.info_log, "%s", buf.str().c_str());
+  }
+
+  // reschedule
+  return std::make_pair(true, -1);
+}
+
+Status BlobDBImpl::CloseBlobFile(std::shared_ptr<BlobFile> bfile) {
+  TEST_SYNC_POINT("BlobDBImpl::CloseBlobFile");
+  assert(bfile);
+  assert(!bfile->Immutable());
+  assert(!bfile->Obsolete());
+
+  if (bfile->HasTTL() || bfile == open_non_ttl_file_) {
+    write_mutex_.AssertHeld();
+  }
+
+  ROCKS_LOG_INFO(db_options_.info_log,
+                 "Closing blob file %" PRIu64 ". Path: %s",
+                 bfile->BlobFileNumber(), bfile->PathName().c_str());
+
+  const SequenceNumber sequence = GetLatestSequenceNumber();
+
+  const Status s = bfile->WriteFooterAndCloseLocked(sequence);
+
+  if (s.ok()) {
+    total_blob_size_ += BlobLogFooter::kSize;
+  } else {
+    bfile->MarkImmutable(sequence);
+
+    ROCKS_LOG_ERROR(db_options_.info_log,
+                    "Failed to close blob file %" PRIu64 "with error: %s",
+                    bfile->BlobFileNumber(), s.ToString().c_str());
+  }
+
+  if (bfile->HasTTL()) {
+    size_t erased __attribute__((__unused__));
+    erased = open_ttl_files_.erase(bfile);
+  } else {
+    if (bfile == open_non_ttl_file_) {
+      open_non_ttl_file_ = nullptr;
+    }
+
+    const uint64_t blob_file_number = bfile->BlobFileNumber();
+    auto it = live_imm_non_ttl_blob_files_.lower_bound(blob_file_number);
+    assert(it == live_imm_non_ttl_blob_files_.end() ||
+           it->first != blob_file_number);
+    live_imm_non_ttl_blob_files_.insert(
+        it, std::map<uint64_t, std::shared_ptr<BlobFile>>::value_type(
+                blob_file_number, bfile));
+  }
+
+  return s;
+}
+
+Status BlobDBImpl::CloseBlobFileIfNeeded(std::shared_ptr<BlobFile>& bfile) {
+  write_mutex_.AssertHeld();
+
+  // atomic read
+  if (bfile->GetFileSize() < bdb_options_.blob_file_size) {
+    return Status::OK();
+  }
+
+  WriteLock lock(&mutex_);
+  WriteLock file_lock(&bfile->mutex_);
+
+  assert(!bfile->Obsolete() || bfile->Immutable());
+  if (bfile->Immutable()) {
+    return Status::OK();
+  }
+
+  return CloseBlobFile(bfile);
+}
+
+void BlobDBImpl::ObsoleteBlobFile(std::shared_ptr<BlobFile> blob_file,
+                                  SequenceNumber obsolete_seq,
+                                  bool update_size) {
+  assert(blob_file->Immutable());
+  assert(!blob_file->Obsolete());
+
+  // Should hold write lock of mutex_ or during DB open.
+  blob_file->MarkObsolete(obsolete_seq);
+  obsolete_files_.push_back(blob_file);
+  assert(total_blob_size_.load() >= blob_file->GetFileSize());
+  if (update_size) {
+    total_blob_size_ -= blob_file->GetFileSize();
+  }
+}
+
+bool BlobDBImpl::VisibleToActiveSnapshot(
+    const std::shared_ptr<BlobFile>& bfile) {
+  assert(bfile->Obsolete());
+
+  // We check whether the oldest snapshot is no less than the last sequence
+  // by the time the blob file become obsolete. If so, the blob file is not
+  // visible to all existing snapshots.
+  //
+  // If we keep track of the earliest sequence of the keys in the blob file,
+  // we could instead check if there's a snapshot falls in range
+  // [earliest_sequence, obsolete_sequence). But doing so will make the
+  // implementation more complicated.
+  SequenceNumber obsolete_sequence = bfile->GetObsoleteSequence();
+  SequenceNumber oldest_snapshot = kMaxSequenceNumber;
+  {
+    // Need to lock DBImpl mutex before access snapshot list.
+    InstrumentedMutexLock l(db_impl_->mutex());
+    auto& snapshots = db_impl_->snapshots();
+    if (!snapshots.empty()) {
+      oldest_snapshot = snapshots.oldest()->GetSequenceNumber();
+    }
+  }
+  bool visible = oldest_snapshot < obsolete_sequence;
+  if (visible) {
+    ROCKS_LOG_INFO(db_options_.info_log,
+                   "Obsolete blob file %" PRIu64 " (obsolete at %" PRIu64
+                   ") visible to oldest snapshot %" PRIu64 ".",
+                   bfile->BlobFileNumber(), obsolete_sequence, oldest_snapshot);
+  }
+  return visible;
+}
+
+std::pair<bool, int64_t> BlobDBImpl::EvictExpiredFiles(bool aborted) {
+  if (aborted) {
+    return std::make_pair(false, -1);
+  }
+
+  TEST_SYNC_POINT("BlobDBImpl::EvictExpiredFiles:0");
+  TEST_SYNC_POINT("BlobDBImpl::EvictExpiredFiles:1");
+
+  std::vector<std::shared_ptr<BlobFile>> process_files;
+  uint64_t now = EpochNow();
+  {
+    ReadLock rl(&mutex_);
+    for (auto p : blob_files_) {
+      auto& blob_file = p.second;
+      ReadLock file_lock(&blob_file->mutex_);
+      if (blob_file->HasTTL() && !blob_file->Obsolete() &&
+          blob_file->GetExpirationRange().second <= now) {
+        process_files.push_back(blob_file);
+      }
+    }
+  }
+
+  TEST_SYNC_POINT("BlobDBImpl::EvictExpiredFiles:2");
+  TEST_SYNC_POINT("BlobDBImpl::EvictExpiredFiles:3");
+  TEST_SYNC_POINT_CALLBACK("BlobDBImpl::EvictExpiredFiles:cb", nullptr);
+
+  SequenceNumber seq = GetLatestSequenceNumber();
+  {
+    MutexLock l(&write_mutex_);
+    WriteLock lock(&mutex_);
+    for (auto& blob_file : process_files) {
+      WriteLock file_lock(&blob_file->mutex_);
+
+      // Need to double check if the file is obsolete.
+      if (blob_file->Obsolete()) {
+        assert(blob_file->Immutable());
+        continue;
+      }
+
+      if (!blob_file->Immutable()) {
+        CloseBlobFile(blob_file);
+      }
+
+      assert(blob_file->Immutable());
+
+      ObsoleteBlobFile(blob_file, seq, true /*update_size*/);
+    }
+  }
+
+  return std::make_pair(true, -1);
+}
+
+Status BlobDBImpl::SyncBlobFiles() {
+  MutexLock l(&write_mutex_);
+
+  std::vector<std::shared_ptr<BlobFile>> process_files;
+  {
+    ReadLock rl(&mutex_);
+    for (auto fitr : open_ttl_files_) {
+      process_files.push_back(fitr);
+    }
+    if (open_non_ttl_file_ != nullptr) {
+      process_files.push_back(open_non_ttl_file_);
+    }
+  }
+
+  Status s;
+  for (auto& blob_file : process_files) {
+    s = blob_file->Fsync();
+    if (!s.ok()) {
+      ROCKS_LOG_ERROR(db_options_.info_log,
+                      "Failed to sync blob file %" PRIu64 ", status: %s",
+                      blob_file->BlobFileNumber(), s.ToString().c_str());
+      return s;
+    }
+  }
+
+  s = dir_ent_->FsyncWithDirOptions(IOOptions(), nullptr, DirFsyncOptions());
+  if (!s.ok()) {
+    ROCKS_LOG_ERROR(db_options_.info_log,
+                    "Failed to sync blob directory, status: %s",
+                    s.ToString().c_str());
+  }
+  return s;
+}
+
+std::pair<bool, int64_t> BlobDBImpl::ReclaimOpenFiles(bool aborted) {
+  if (aborted) return std::make_pair(false, -1);
+
+  if (open_file_count_.load() < kOpenFilesTrigger) {
+    return std::make_pair(true, -1);
+  }
+
+  // in the future, we should sort by last_access_
+  // instead of closing every file
+  ReadLock rl(&mutex_);
+  for (auto const& ent : blob_files_) {
+    auto bfile = ent.second;
+    if (bfile->last_access_.load() == -1) continue;
+
+    WriteLock lockbfile_w(&bfile->mutex_);
+    CloseRandomAccessLocked(bfile);
+  }
+
+  return std::make_pair(true, -1);
+}
+
+std::pair<bool, int64_t> BlobDBImpl::DeleteObsoleteFiles(bool aborted) {
+  if (aborted) {
+    return std::make_pair(false, -1);
+  }
+
+  MutexLock delete_file_lock(&delete_file_mutex_);
+  if (disable_file_deletions_ > 0) {
+    return std::make_pair(true, -1);
+  }
+
+  std::list<std::shared_ptr<BlobFile>> tobsolete;
+  {
+    WriteLock wl(&mutex_);
+    if (obsolete_files_.empty()) {
+      return std::make_pair(true, -1);
+    }
+    tobsolete.swap(obsolete_files_);
+  }
+
+  bool file_deleted = false;
+  for (auto iter = tobsolete.begin(); iter != tobsolete.end();) {
+    auto bfile = *iter;
+    {
+      ReadLock lockbfile_r(&bfile->mutex_);
+      if (VisibleToActiveSnapshot(bfile)) {
+        ROCKS_LOG_INFO(db_options_.info_log,
+                       "Could not delete file due to snapshot failure %s",
+                       bfile->PathName().c_str());
+        ++iter;
+        continue;
+      }
+    }
+    ROCKS_LOG_INFO(db_options_.info_log,
+                   "Will delete file due to snapshot success %s",
+                   bfile->PathName().c_str());
+
+    {
+      WriteLock wl(&mutex_);
+      blob_files_.erase(bfile->BlobFileNumber());
+    }
+
+    Status s = DeleteDBFile(&(db_impl_->immutable_db_options()),
+                            bfile->PathName(), blob_dir_, true,
+                            /*force_fg=*/false);
+    if (!s.ok()) {
+      ROCKS_LOG_ERROR(db_options_.info_log,
+                      "File failed to be deleted as obsolete %s",
+                      bfile->PathName().c_str());
+      ++iter;
+      continue;
+    }
+
+    file_deleted = true;
+    ROCKS_LOG_INFO(db_options_.info_log,
+                   "File deleted as obsolete from blob dir %s",
+                   bfile->PathName().c_str());
+
+    iter = tobsolete.erase(iter);
+  }
+
+  // directory change. Fsync
+  if (file_deleted) {
+    Status s = dir_ent_->FsyncWithDirOptions(
+        IOOptions(), nullptr,
+        DirFsyncOptions(DirFsyncOptions::FsyncReason::kFileDeleted));
+    if (!s.ok()) {
+      ROCKS_LOG_ERROR(db_options_.info_log, "Failed to sync dir %s: %s",
+                      blob_dir_.c_str(), s.ToString().c_str());
+    }
+  }
+
+  // put files back into obsolete if for some reason, delete failed
+  if (!tobsolete.empty()) {
+    WriteLock wl(&mutex_);
+    for (auto bfile : tobsolete) {
+      blob_files_.insert(std::make_pair(bfile->BlobFileNumber(), bfile));
+      obsolete_files_.push_front(bfile);
+    }
+  }
+
+  return std::make_pair(!aborted, -1);
+}
+
+void BlobDBImpl::CopyBlobFiles(
+    std::vector<std::shared_ptr<BlobFile>>* bfiles_copy) {
+  ReadLock rl(&mutex_);
+  for (auto const& p : blob_files_) {
+    bfiles_copy->push_back(p.second);
+  }
+}
+
+Iterator* BlobDBImpl::NewIterator(const ReadOptions& read_options) {
+  auto* cfd =
+      static_cast_with_check<ColumnFamilyHandleImpl>(DefaultColumnFamily())
+          ->cfd();
+  // Get a snapshot to avoid blob file get deleted between we
+  // fetch and index entry and reading from the file.
+  ManagedSnapshot* own_snapshot = nullptr;
+  const Snapshot* snapshot = read_options.snapshot;
+  if (snapshot == nullptr) {
+    own_snapshot = new ManagedSnapshot(db_);
+    snapshot = own_snapshot->snapshot();
+  }
+  auto* iter = db_impl_->NewIteratorImpl(
+      read_options, cfd, snapshot->GetSequenceNumber(),
+      nullptr /*read_callback*/, true /*expose_blob_index*/);
+  return new BlobDBIterator(own_snapshot, iter, this, clock_, statistics_);
+}
+
+Status DestroyBlobDB(const std::string& dbname, const Options& options,
+                     const BlobDBOptions& bdb_options) {
+  const ImmutableDBOptions soptions(SanitizeOptions(dbname, options));
+  Env* env = soptions.env;
+
+  Status status;
+  std::string blobdir;
+  blobdir = (bdb_options.path_relative) ? dbname + "/" + bdb_options.blob_dir
+                                        : bdb_options.blob_dir;
+
+  std::vector<std::string> filenames;
+  if (env->GetChildren(blobdir, &filenames).ok()) {
+    for (const auto& f : filenames) {
+      uint64_t number;
+      FileType type;
+      if (ParseFileName(f, &number, &type) && type == kBlobFile) {
+        Status del = DeleteDBFile(&soptions, blobdir + "/" + f, blobdir, true,
+                                  /*force_fg=*/false);
+        if (status.ok() && !del.ok()) {
+          status = del;
+        }
+      }
+    }
+    // TODO: What to do if we cannot delete the directory?
+    env->DeleteDir(blobdir).PermitUncheckedError();
+  }
+  Status destroy = DestroyDB(dbname, options);
+  if (status.ok() && !destroy.ok()) {
+    status = destroy;
+  }
+
+  return status;
+}
+
+#ifndef NDEBUG
+Status BlobDBImpl::TEST_GetBlobValue(const Slice& key, const Slice& index_entry,
+                                     PinnableSlice* value) {
+  return GetBlobValue(key, index_entry, value);
+}
+
+void BlobDBImpl::TEST_AddDummyBlobFile(uint64_t blob_file_number,
+                                       SequenceNumber immutable_sequence) {
+  auto blob_file = std::make_shared<BlobFile>(this, blob_dir_, blob_file_number,
+                                              db_options_.info_log.get());
+  blob_file->MarkImmutable(immutable_sequence);
+
+  blob_files_[blob_file_number] = blob_file;
+  live_imm_non_ttl_blob_files_[blob_file_number] = blob_file;
+}
+
+std::vector<std::shared_ptr<BlobFile>> BlobDBImpl::TEST_GetBlobFiles() const {
+  ReadLock l(&mutex_);
+  std::vector<std::shared_ptr<BlobFile>> blob_files;
+  for (auto& p : blob_files_) {
+    blob_files.emplace_back(p.second);
+  }
+  return blob_files;
+}
+
+std::vector<std::shared_ptr<BlobFile>> BlobDBImpl::TEST_GetLiveImmNonTTLFiles()
+    const {
+  ReadLock l(&mutex_);
+  std::vector<std::shared_ptr<BlobFile>> live_imm_non_ttl_files;
+  for (const auto& pair : live_imm_non_ttl_blob_files_) {
+    live_imm_non_ttl_files.emplace_back(pair.second);
+  }
+  return live_imm_non_ttl_files;
+}
+
+std::vector<std::shared_ptr<BlobFile>> BlobDBImpl::TEST_GetObsoleteFiles()
+    const {
+  ReadLock l(&mutex_);
+  std::vector<std::shared_ptr<BlobFile>> obsolete_files;
+  for (auto& bfile : obsolete_files_) {
+    obsolete_files.emplace_back(bfile);
+  }
+  return obsolete_files;
+}
+
+void BlobDBImpl::TEST_DeleteObsoleteFiles() {
+  DeleteObsoleteFiles(false /*abort*/);
+}
+
+Status BlobDBImpl::TEST_CloseBlobFile(std::shared_ptr<BlobFile>& bfile) {
+  MutexLock l(&write_mutex_);
+  WriteLock lock(&mutex_);
+  WriteLock file_lock(&bfile->mutex_);
+
+  return CloseBlobFile(bfile);
+}
+
+void BlobDBImpl::TEST_ObsoleteBlobFile(std::shared_ptr<BlobFile>& blob_file,
+                                       SequenceNumber obsolete_seq,
+                                       bool update_size) {
+  return ObsoleteBlobFile(blob_file, obsolete_seq, update_size);
+}
+
+void BlobDBImpl::TEST_EvictExpiredFiles() {
+  EvictExpiredFiles(false /*abort*/);
+}
+
+uint64_t BlobDBImpl::TEST_live_sst_size() { return live_sst_size_.load(); }
+
+void BlobDBImpl::TEST_InitializeBlobFileToSstMapping(
+    const std::vector<LiveFileMetaData>& live_files) {
+  InitializeBlobFileToSstMapping(live_files);
+}
+
+void BlobDBImpl::TEST_ProcessFlushJobInfo(const FlushJobInfo& info) {
+  ProcessFlushJobInfo(info);
+}
+
+void BlobDBImpl::TEST_ProcessCompactionJobInfo(const CompactionJobInfo& info) {
+  ProcessCompactionJobInfo(info);
+}
+
+#endif  //  !NDEBUG
+
+}  // namespace blob_db
+}  // namespace ROCKSDB_NAMESPACE
+#endif  // ROCKSDB_LITE
diff --git a/src/rocksdb/utilities/blob_db/blob_db_impl.h b/src/rocksdb/utilities/blob_db/blob_db_impl.h
new file mode 100644
index 000000000..0b4dbf5e5
--- /dev/null
+++ b/src/rocksdb/utilities/blob_db/blob_db_impl.h
@@ -0,0 +1,503 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#pragma once
+
+#ifndef ROCKSDB_LITE
+
+#include <atomic>
+#include <condition_variable>
+#include <limits>
+#include <list>
+#include <memory>
+#include <set>
+#include <string>
+#include <thread>
+#include <unordered_map>
+#include <utility>
+#include <vector>
+
+#include "db/blob/blob_log_format.h"
+#include "db/blob/blob_log_writer.h"
+#include "db/db_iter.h"
+#include "rocksdb/compaction_filter.h"
+#include "rocksdb/db.h"
+#include "rocksdb/file_system.h"
+#include "rocksdb/listener.h"
+#include "rocksdb/options.h"
+#include "rocksdb/statistics.h"
+#include "rocksdb/wal_filter.h"
+#include "util/mutexlock.h"
+#include "util/timer_queue.h"
+#include "utilities/blob_db/blob_db.h"
+#include "utilities/blob_db/blob_file.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+class DBImpl;
+class ColumnFamilyHandle;
+class ColumnFamilyData;
+class SystemClock;
+
+struct FlushJobInfo;
+
+namespace blob_db {
+
+struct BlobCompactionContext;
+struct BlobCompactionContextGC;
+class BlobDBImpl;
+class BlobFile;
+
+// Comparator to sort "TTL" aware Blob files based on the lower value of
+// TTL range.
+struct BlobFileComparatorTTL {
+  bool operator()(const std::shared_ptr<BlobFile>& lhs,
+                  const std::shared_ptr<BlobFile>& rhs) const;
+};
+
+struct BlobFileComparator {
+  bool operator()(const std::shared_ptr<BlobFile>& lhs,
+                  const std::shared_ptr<BlobFile>& rhs) const;
+};
+
+/**
+ * The implementation class for BlobDB. It manages the blob logs, which
+ * are sequentially written files. Blob logs can be of the TTL or non-TTL
+ * varieties; the former are cleaned up when they expire, while the latter
+ * are (optionally) garbage collected.
+ */
+class BlobDBImpl : public BlobDB {
+  friend class BlobFile;
+  friend class BlobDBIterator;
+  friend class BlobDBListener;
+  friend class BlobDBListenerGC;
+  friend class BlobIndexCompactionFilterBase;
+  friend class BlobIndexCompactionFilterGC;
+
+ public:
+  // deletions check period
+  static constexpr uint32_t kDeleteCheckPeriodMillisecs = 2 * 1000;
+
+  // sanity check task
+  static constexpr uint32_t kSanityCheckPeriodMillisecs = 20 * 60 * 1000;
+
+  // how many random access open files can we tolerate
+  static constexpr uint32_t kOpenFilesTrigger = 100;
+
+  // how often to schedule reclaim open files.
+  static constexpr uint32_t kReclaimOpenFilesPeriodMillisecs = 1 * 1000;
+
+  // how often to schedule delete obs files periods
+  static constexpr uint32_t kDeleteObsoleteFilesPeriodMillisecs = 10 * 1000;
+
+  // how often to schedule expired files eviction.
+  static constexpr uint32_t kEvictExpiredFilesPeriodMillisecs = 10 * 1000;
+
+  // when should oldest file be evicted:
+  // on reaching 90% of blob_dir_size
+  static constexpr double kEvictOldestFileAtSize = 0.9;
+
+  using BlobDB::Put;
+  Status Put(const WriteOptions& options, const Slice& key,
+             const Slice& value) override;
+
+  using BlobDB::Get;
+  Status Get(const ReadOptions& read_options, ColumnFamilyHandle* column_family,
+             const Slice& key, PinnableSlice* value) override;
+
+  Status Get(const ReadOptions& read_options, ColumnFamilyHandle* column_family,
+             const Slice& key, PinnableSlice* value,
+             uint64_t* expiration) override;
+
+  using BlobDB::NewIterator;
+  virtual Iterator* NewIterator(const ReadOptions& read_options) override;
+
+  using BlobDB::NewIterators;
+  virtual Status NewIterators(
+      const ReadOptions& /*read_options*/,
+      const std::vector<ColumnFamilyHandle*>& /*column_families*/,
+      std::vector<Iterator*>* /*iterators*/) override {
+    return Status::NotSupported("Not implemented");
+  }
+
+  using BlobDB::MultiGet;
+  virtual std::vector<Status> MultiGet(
+      const ReadOptions& read_options, const std::vector<Slice>& keys,
+      std::vector<std::string>* values) override;
+
+  using BlobDB::Write;
+  virtual Status Write(const WriteOptions& opts, WriteBatch* updates) override;
+
+  virtual Status Close() override;
+
+  using BlobDB::PutWithTTL;
+  Status PutWithTTL(const WriteOptions& options, const Slice& key,
+                    const Slice& value, uint64_t ttl) override;
+
+  using BlobDB::PutUntil;
+  Status PutUntil(const WriteOptions& options, const Slice& key,
+                  const Slice& value, uint64_t expiration) override;
+
+  using BlobDB::CompactFiles;
+  Status CompactFiles(
+      const CompactionOptions& compact_options,
+      const std::vector<std::string>& input_file_names, const int output_level,
+      const int output_path_id = -1,
+      std::vector<std::string>* const output_file_names = nullptr,
+      CompactionJobInfo* compaction_job_info = nullptr) override;
+
+  BlobDBOptions GetBlobDBOptions() const override;
+
+  BlobDBImpl(const std::string& dbname, const BlobDBOptions& bdb_options,
+             const DBOptions& db_options,
+             const ColumnFamilyOptions& cf_options);
+
+  virtual Status DisableFileDeletions() override;
+
+  virtual Status EnableFileDeletions(bool force) override;
+
+  virtual Status GetLiveFiles(std::vector<std::string>&,
+                              uint64_t* manifest_file_size,
+                              bool flush_memtable = true) override;
+  virtual void GetLiveFilesMetaData(std::vector<LiveFileMetaData>*) override;
+
+  ~BlobDBImpl();
+
+  Status Open(std::vector<ColumnFamilyHandle*>* handles);
+
+  Status SyncBlobFiles() override;
+
+  // Common part of the two GetCompactionContext methods below.
+  // REQUIRES: read lock on mutex_
+  void GetCompactionContextCommon(BlobCompactionContext* context);
+
+  void GetCompactionContext(BlobCompactionContext* context);
+  void GetCompactionContext(BlobCompactionContext* context,
+                            BlobCompactionContextGC* context_gc);
+
+#ifndef NDEBUG
+  Status TEST_GetBlobValue(const Slice& key, const Slice& index_entry,
+                           PinnableSlice* value);
+
+  void TEST_AddDummyBlobFile(uint64_t blob_file_number,
+                             SequenceNumber immutable_sequence);
+
+  std::vector<std::shared_ptr<BlobFile>> TEST_GetBlobFiles() const;
+
+  std::vector<std::shared_ptr<BlobFile>> TEST_GetLiveImmNonTTLFiles() const;
+
+  std::vector<std::shared_ptr<BlobFile>> TEST_GetObsoleteFiles() const;
+
+  Status TEST_CloseBlobFile(std::shared_ptr<BlobFile>& bfile);
+
+  void TEST_ObsoleteBlobFile(std::shared_ptr<BlobFile>& blob_file,
+                             SequenceNumber obsolete_seq = 0,
+                             bool update_size = true);
+
+  void TEST_EvictExpiredFiles();
+
+  void TEST_DeleteObsoleteFiles();
+
+  uint64_t TEST_live_sst_size();
+
+  const std::string& TEST_blob_dir() const { return blob_dir_; }
+
+  void TEST_InitializeBlobFileToSstMapping(
+      const std::vector<LiveFileMetaData>& live_files);
+
+  void TEST_ProcessFlushJobInfo(const FlushJobInfo& info);
+
+  void TEST_ProcessCompactionJobInfo(const CompactionJobInfo& info);
+
+#endif  //  !NDEBUG
+
+ private:
+  class BlobInserter;
+
+  // Create a snapshot if there isn't one in read options.
+  // Return true if a snapshot is created.
+  bool SetSnapshotIfNeeded(ReadOptions* read_options);
+
+  Status GetImpl(const ReadOptions& read_options,
+                 ColumnFamilyHandle* column_family, const Slice& key,
+                 PinnableSlice* value, uint64_t* expiration = nullptr);
+
+  Status GetBlobValue(const Slice& key, const Slice& index_entry,
+                      PinnableSlice* value, uint64_t* expiration = nullptr);
+
+  Status GetRawBlobFromFile(const Slice& key, uint64_t file_number,
+                            uint64_t offset, uint64_t size,
+                            PinnableSlice* value,
+                            CompressionType* compression_type);
+
+  Slice GetCompressedSlice(const Slice& raw,
+                           std::string* compression_output) const;
+
+  Status DecompressSlice(const Slice& compressed_value,
+                         CompressionType compression_type,
+                         PinnableSlice* value_output) const;
+
+  // Close a file by appending a footer, and removes file from open files list.
+  // REQUIRES: lock held on write_mutex_, write lock held on both the db mutex_
+  // and the blob file's mutex_. If called on a blob file which is visible only
+  // to a single thread (like in the case of new files written during
+  // compaction/GC), the locks on write_mutex_ and the blob file's mutex_ can be
+  // avoided.
+  Status CloseBlobFile(std::shared_ptr<BlobFile> bfile);
+
+  // Close a file if its size exceeds blob_file_size
+  // REQUIRES: lock held on write_mutex_.
+  Status CloseBlobFileIfNeeded(std::shared_ptr<BlobFile>& bfile);
+
+  // Mark file as obsolete and move the file to obsolete file list.
+  //
+  // REQUIRED: hold write lock of mutex_ or during DB open.
+  void ObsoleteBlobFile(std::shared_ptr<BlobFile> blob_file,
+                        SequenceNumber obsolete_seq, bool update_size);
+
+  Status PutBlobValue(const WriteOptions& options, const Slice& key,
+                      const Slice& value, uint64_t expiration,
+                      WriteBatch* batch);
+
+  Status AppendBlob(const std::shared_ptr<BlobFile>& bfile,
+                    const std::string& headerbuf, const Slice& key,
+                    const Slice& value, uint64_t expiration,
+                    std::string* index_entry);
+
+  // Create a new blob file and associated writer.
+  Status CreateBlobFileAndWriter(bool has_ttl,
+                                 const ExpirationRange& expiration_range,
+                                 const std::string& reason,
+                                 std::shared_ptr<BlobFile>* blob_file,
+                                 std::shared_ptr<BlobLogWriter>* writer);
+
+  // Get the open non-TTL blob log file, or create a new one if no such file
+  // exists.
+  Status SelectBlobFile(std::shared_ptr<BlobFile>* blob_file);
+
+  // Get the open TTL blob log file for a certain expiration, or create a new
+  // one if no such file exists.
+  Status SelectBlobFileTTL(uint64_t expiration,
+                           std::shared_ptr<BlobFile>* blob_file);
+
+  std::shared_ptr<BlobFile> FindBlobFileLocked(uint64_t expiration) const;
+
+  // periodic sanity check. Bunch of checks
+  std::pair<bool, int64_t> SanityCheck(bool aborted);
+
+  // Delete files that have been marked obsolete (either because of TTL
+  // or GC). Check whether any snapshots exist which refer to the same.
+  std::pair<bool, int64_t> DeleteObsoleteFiles(bool aborted);
+
+  // periodically check if open blob files and their TTL's has expired
+  // if expired, close the sequential writer and make the file immutable
+  std::pair<bool, int64_t> EvictExpiredFiles(bool aborted);
+
+  // if the number of open files, approaches ULIMIT's this
+  // task will close random readers, which are kept around for
+  // efficiency
+  std::pair<bool, int64_t> ReclaimOpenFiles(bool aborted);
+
+  std::pair<bool, int64_t> RemoveTimerQ(TimerQueue* tq, bool aborted);
+
+  // Adds the background tasks to the timer queue
+  void StartBackgroundTasks();
+
+  // add a new Blob File
+  std::shared_ptr<BlobFile> NewBlobFile(bool has_ttl,
+                                        const ExpirationRange& expiration_range,
+                                        const std::string& reason);
+
+  // Register a new blob file.
+  // REQUIRES: write lock on mutex_.
+  void RegisterBlobFile(std::shared_ptr<BlobFile> blob_file);
+
+  // collect all the blob log files from the blob directory
+  Status GetAllBlobFiles(std::set<uint64_t>* file_numbers);
+
+  // Open all blob files found in blob_dir.
+  Status OpenAllBlobFiles();
+
+  // Link an SST to a blob file. Comes in locking and non-locking varieties
+  // (the latter is used during Open).
+  template <typename Linker>
+  void LinkSstToBlobFileImpl(uint64_t sst_file_number,
+                             uint64_t blob_file_number, Linker linker);
+
+  void LinkSstToBlobFile(uint64_t sst_file_number, uint64_t blob_file_number);
+
+  void LinkSstToBlobFileNoLock(uint64_t sst_file_number,
+                               uint64_t blob_file_number);
+
+  // Unlink an SST from a blob file.
+  void UnlinkSstFromBlobFile(uint64_t sst_file_number,
+                             uint64_t blob_file_number);
+
+  // Initialize the mapping between blob files and SSTs during Open.
+  void InitializeBlobFileToSstMapping(
+      const std::vector<LiveFileMetaData>& live_files);
+
+  // Update the mapping between blob files and SSTs after a flush and mark
+  // any unneeded blob files obsolete.
+  void ProcessFlushJobInfo(const FlushJobInfo& info);
+
+  // Update the mapping between blob files and SSTs after a compaction and
+  // mark any unneeded blob files obsolete.
+  void ProcessCompactionJobInfo(const CompactionJobInfo& info);
+
+  // Mark an immutable non-TTL blob file obsolete assuming it has no more SSTs
+  // linked to it, and all memtables from before the blob file became immutable
+  // have been flushed. Note: should only be called if the condition holds for
+  // all lower-numbered non-TTL blob files as well.
+  bool MarkBlobFileObsoleteIfNeeded(const std::shared_ptr<BlobFile>& blob_file,
+                                    SequenceNumber obsolete_seq);
+
+  // Mark all immutable non-TTL blob files that aren't needed by any SSTs as
+  // obsolete. Comes in two varieties; the version used during Open need not
+  // worry about locking or snapshots.
+  template <class Functor>
+  void MarkUnreferencedBlobFilesObsoleteImpl(Functor mark_if_needed);
+
+  void MarkUnreferencedBlobFilesObsolete();
+  void MarkUnreferencedBlobFilesObsoleteDuringOpen();
+
+  void UpdateLiveSSTSize();
+
+  Status GetBlobFileReader(const std::shared_ptr<BlobFile>& blob_file,
+                           std::shared_ptr<RandomAccessFileReader>* reader);
+
+  // hold write mutex on file and call.
+  // Close the above Random Access reader
+  void CloseRandomAccessLocked(const std::shared_ptr<BlobFile>& bfile);
+
+  // hold write mutex on file and call
+  // creates a sequential (append) writer for this blobfile
+  Status CreateWriterLocked(const std::shared_ptr<BlobFile>& bfile);
+
+  // returns a BlobLogWriter object for the file. If writer is not
+  // already present, creates one. Needs Write Mutex to be held
+  Status CheckOrCreateWriterLocked(const std::shared_ptr<BlobFile>& blob_file,
+                                   std::shared_ptr<BlobLogWriter>* writer);
+
+  // checks if there is no snapshot which is referencing the
+  // blobs
+  bool VisibleToActiveSnapshot(const std::shared_ptr<BlobFile>& file);
+  bool FileDeleteOk_SnapshotCheckLocked(const std::shared_ptr<BlobFile>& bfile);
+
+  void CopyBlobFiles(std::vector<std::shared_ptr<BlobFile>>* bfiles_copy);
+
+  uint64_t EpochNow() { return clock_->NowMicros() / 1000000; }
+
+  // Check if inserting a new blob will make DB grow out of space.
+  // If is_fifo = true, FIFO eviction will be triggered to make room for the
+  // new blob. If force_evict = true, FIFO eviction will evict blob files
+  // even eviction will not make enough room for the new blob.
+  Status CheckSizeAndEvictBlobFiles(uint64_t blob_size,
+                                    bool force_evict = false);
+
+  // name of the database directory
+  std::string dbname_;
+
+  // the base DB
+  DBImpl* db_impl_;
+  Env* env_;
+  SystemClock* clock_;
+  // the options that govern the behavior of Blob Storage
+  BlobDBOptions bdb_options_;
+  DBOptions db_options_;
+  ColumnFamilyOptions cf_options_;
+  FileOptions file_options_;
+
+  // Raw pointer of statistic. db_options_ has a std::shared_ptr to hold
+  // ownership.
+  Statistics* statistics_;
+
+  // by default this is "blob_dir" under dbname_
+  // but can be configured
+  std::string blob_dir_;
+
+  // pointer to directory
+  std::unique_ptr<FSDirectory> dir_ent_;
+
+  // Read Write Mutex, which protects all the data structures
+  // HEAVILY TRAFFICKED
+  mutable port::RWMutex mutex_;
+
+  // Writers has to hold write_mutex_ before writing.
+  mutable port::Mutex write_mutex_;
+
+  // counter for blob file number
+  std::atomic<uint64_t> next_file_number_;
+
+  // entire metadata of all the BLOB files memory
+  std::map<uint64_t, std::shared_ptr<BlobFile>> blob_files_;
+
+  // All live immutable non-TTL blob files.
+  std::map<uint64_t, std::shared_ptr<BlobFile>> live_imm_non_ttl_blob_files_;
+
+  // The largest sequence number that has been flushed.
+  SequenceNumber flush_sequence_;
+
+  // opened non-TTL blob file.
+  std::shared_ptr<BlobFile> open_non_ttl_file_;
+
+  // all the blob files which are currently being appended to based
+  // on variety of incoming TTL's
+  std::set<std::shared_ptr<BlobFile>, BlobFileComparatorTTL> open_ttl_files_;
+
+  // Flag to check whether Close() has been called on this DB
+  bool closed_;
+
+  // timer based queue to execute tasks
+  TimerQueue tqueue_;
+
+  // number of files opened for random access/GET
+  // counter is used to monitor and close excess RA files.
+  std::atomic<uint32_t> open_file_count_;
+
+  // Total size of all live blob files (i.e. exclude obsolete files).
+  std::atomic<uint64_t> total_blob_size_;
+
+  // total size of SST files.
+  std::atomic<uint64_t> live_sst_size_;
+
+  // Latest FIFO eviction timestamp
+  //
+  // REQUIRES: access with metex_ lock held.
+  uint64_t fifo_eviction_seq_;
+
+  // The expiration up to which latest FIFO eviction evicts.
+  //
+  // REQUIRES: access with metex_ lock held.
+  uint64_t evict_expiration_up_to_;
+
+  std::list<std::shared_ptr<BlobFile>> obsolete_files_;
+
+  // DeleteObsoleteFiles, DiableFileDeletions and EnableFileDeletions block
+  // on the mutex to avoid contention.
+  //
+  // While DeleteObsoleteFiles hold both mutex_ and delete_file_mutex_, note
+  // the difference. mutex_ only needs to be held when access the
+  // data-structure, and delete_file_mutex_ needs to be held the whole time
+  // during DeleteObsoleteFiles to avoid being run simultaneously with
+  // DisableFileDeletions.
+  //
+  // If both of mutex_ and delete_file_mutex_ needs to be held, it is adviced
+  // to hold delete_file_mutex_ first to avoid deadlock.
+  mutable port::Mutex delete_file_mutex_;
+
+  // Each call of DisableFileDeletions will increase disable_file_deletion_
+  // by 1. EnableFileDeletions will either decrease the count by 1 or reset
+  // it to zeor, depending on the force flag.
+  //
+  // REQUIRES: access with delete_file_mutex_ held.
+  int disable_file_deletions_ = 0;
+
+  uint32_t debug_level_;
+};
+
+}  // namespace blob_db
+}  // namespace ROCKSDB_NAMESPACE
+#endif  // ROCKSDB_LITE
diff --git a/src/rocksdb/utilities/blob_db/blob_db_impl_filesnapshot.cc b/src/rocksdb/utilities/blob_db/blob_db_impl_filesnapshot.cc
new file mode 100644
index 000000000..87e3f33cc
--- /dev/null
+++ b/src/rocksdb/utilities/blob_db/blob_db_impl_filesnapshot.cc
@@ -0,0 +1,113 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#ifndef ROCKSDB_LITE
+
+#include "file/filename.h"
+#include "logging/logging.h"
+#include "util/cast_util.h"
+#include "util/mutexlock.h"
+#include "utilities/blob_db/blob_db_impl.h"
+
+// BlobDBImpl methods to get snapshot of files, e.g. for replication.
+
+namespace ROCKSDB_NAMESPACE {
+namespace blob_db {
+
+Status BlobDBImpl::DisableFileDeletions() {
+  // Disable base DB file deletions.
+  Status s = db_impl_->DisableFileDeletions();
+  if (!s.ok()) {
+    return s;
+  }
+
+  int count = 0;
+  {
+    // Hold delete_file_mutex_ to make sure no DeleteObsoleteFiles job
+    // is running.
+    MutexLock l(&delete_file_mutex_);
+    count = ++disable_file_deletions_;
+  }
+
+  ROCKS_LOG_INFO(db_options_.info_log,
+                 "Disabled blob file deletions. count: %d", count);
+  return Status::OK();
+}
+
+Status BlobDBImpl::EnableFileDeletions(bool force) {
+  // Enable base DB file deletions.
+  Status s = db_impl_->EnableFileDeletions(force);
+  if (!s.ok()) {
+    return s;
+  }
+
+  int count = 0;
+  {
+    MutexLock l(&delete_file_mutex_);
+    if (force) {
+      disable_file_deletions_ = 0;
+    } else if (disable_file_deletions_ > 0) {
+      count = --disable_file_deletions_;
+    }
+    assert(count >= 0);
+  }
+
+  ROCKS_LOG_INFO(db_options_.info_log, "Enabled blob file deletions. count: %d",
+                 count);
+  // Consider trigger DeleteobsoleteFiles once after re-enabled, if we are to
+  // make DeleteobsoleteFiles re-run interval configuration.
+  return Status::OK();
+}
+
+Status BlobDBImpl::GetLiveFiles(std::vector<std::string>& ret,
+                                uint64_t* manifest_file_size,
+                                bool flush_memtable) {
+  if (!bdb_options_.path_relative) {
+    return Status::NotSupported(
+        "Not able to get relative blob file path from absolute blob_dir.");
+  }
+  // Hold a lock in the beginning to avoid updates to base DB during the call
+  ReadLock rl(&mutex_);
+  Status s = db_->GetLiveFiles(ret, manifest_file_size, flush_memtable);
+  if (!s.ok()) {
+    return s;
+  }
+  ret.reserve(ret.size() + blob_files_.size());
+  for (auto bfile_pair : blob_files_) {
+    auto blob_file = bfile_pair.second;
+    // Path should be relative to db_name, but begin with slash.
+    ret.emplace_back(
+        BlobFileName("", bdb_options_.blob_dir, blob_file->BlobFileNumber()));
+  }
+  return Status::OK();
+}
+
+void BlobDBImpl::GetLiveFilesMetaData(std::vector<LiveFileMetaData>* metadata) {
+  // Path should be relative to db_name.
+  assert(bdb_options_.path_relative);
+  // Hold a lock in the beginning to avoid updates to base DB during the call
+  ReadLock rl(&mutex_);
+  db_->GetLiveFilesMetaData(metadata);
+  for (auto bfile_pair : blob_files_) {
+    auto blob_file = bfile_pair.second;
+    LiveFileMetaData filemetadata;
+    filemetadata.size = blob_file->GetFileSize();
+    const uint64_t file_number = blob_file->BlobFileNumber();
+    // Path should be relative to db_name, but begin with slash.
+    filemetadata.name = BlobFileName("", bdb_options_.blob_dir, file_number);
+    filemetadata.file_number = file_number;
+    if (blob_file->HasTTL()) {
+      filemetadata.oldest_ancester_time = blob_file->GetExpirationRange().first;
+    }
+    auto cfh =
+        static_cast_with_check<ColumnFamilyHandleImpl>(DefaultColumnFamily());
+    filemetadata.column_family_name = cfh->GetName();
+    metadata->emplace_back(filemetadata);
+  }
+}
+
+}  // namespace blob_db
+}  // namespace ROCKSDB_NAMESPACE
+#endif  // !ROCKSDB_LITE
diff --git a/src/rocksdb/utilities/blob_db/blob_db_iterator.h b/src/rocksdb/utilities/blob_db/blob_db_iterator.h
new file mode 100644
index 000000000..fd2b2f8f5
--- /dev/null
+++ b/src/rocksdb/utilities/blob_db/blob_db_iterator.h
@@ -0,0 +1,150 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#pragma once
+#ifndef ROCKSDB_LITE
+
+#include "db/arena_wrapped_db_iter.h"
+#include "rocksdb/iterator.h"
+#include "util/stop_watch.h"
+#include "utilities/blob_db/blob_db_impl.h"
+
+namespace ROCKSDB_NAMESPACE {
+class Statistics;
+class SystemClock;
+
+namespace blob_db {
+
+using ROCKSDB_NAMESPACE::ManagedSnapshot;
+
+class BlobDBIterator : public Iterator {
+ public:
+  BlobDBIterator(ManagedSnapshot* snapshot, ArenaWrappedDBIter* iter,
+                 BlobDBImpl* blob_db, SystemClock* clock,
+                 Statistics* statistics)
+      : snapshot_(snapshot),
+        iter_(iter),
+        blob_db_(blob_db),
+        clock_(clock),
+        statistics_(statistics) {}
+
+  virtual ~BlobDBIterator() = default;
+
+  bool Valid() const override {
+    if (!iter_->Valid()) {
+      return false;
+    }
+    return status_.ok();
+  }
+
+  Status status() const override {
+    if (!iter_->status().ok()) {
+      return iter_->status();
+    }
+    return status_;
+  }
+
+  void SeekToFirst() override {
+    StopWatch seek_sw(clock_, statistics_, BLOB_DB_SEEK_MICROS);
+    RecordTick(statistics_, BLOB_DB_NUM_SEEK);
+    iter_->SeekToFirst();
+    while (UpdateBlobValue()) {
+      iter_->Next();
+    }
+  }
+
+  void SeekToLast() override {
+    StopWatch seek_sw(clock_, statistics_, BLOB_DB_SEEK_MICROS);
+    RecordTick(statistics_, BLOB_DB_NUM_SEEK);
+    iter_->SeekToLast();
+    while (UpdateBlobValue()) {
+      iter_->Prev();
+    }
+  }
+
+  void Seek(const Slice& target) override {
+    StopWatch seek_sw(clock_, statistics_, BLOB_DB_SEEK_MICROS);
+    RecordTick(statistics_, BLOB_DB_NUM_SEEK);
+    iter_->Seek(target);
+    while (UpdateBlobValue()) {
+      iter_->Next();
+    }
+  }
+
+  void SeekForPrev(const Slice& target) override {
+    StopWatch seek_sw(clock_, statistics_, BLOB_DB_SEEK_MICROS);
+    RecordTick(statistics_, BLOB_DB_NUM_SEEK);
+    iter_->SeekForPrev(target);
+    while (UpdateBlobValue()) {
+      iter_->Prev();
+    }
+  }
+
+  void Next() override {
+    assert(Valid());
+    StopWatch next_sw(clock_, statistics_, BLOB_DB_NEXT_MICROS);
+    RecordTick(statistics_, BLOB_DB_NUM_NEXT);
+    iter_->Next();
+    while (UpdateBlobValue()) {
+      iter_->Next();
+    }
+  }
+
+  void Prev() override {
+    assert(Valid());
+    StopWatch prev_sw(clock_, statistics_, BLOB_DB_PREV_MICROS);
+    RecordTick(statistics_, BLOB_DB_NUM_PREV);
+    iter_->Prev();
+    while (UpdateBlobValue()) {
+      iter_->Prev();
+    }
+  }
+
+  Slice key() const override {
+    assert(Valid());
+    return iter_->key();
+  }
+
+  Slice value() const override {
+    assert(Valid());
+    if (!iter_->IsBlob()) {
+      return iter_->value();
+    }
+    return value_;
+  }
+
+  // Iterator::Refresh() not supported.
+
+ private:
+  // Return true if caller should continue to next value.
+  bool UpdateBlobValue() {
+    value_.Reset();
+    status_ = Status::OK();
+    if (iter_->Valid() && iter_->status().ok() && iter_->IsBlob()) {
+      Status s = blob_db_->GetBlobValue(iter_->key(), iter_->value(), &value_);
+      if (s.IsNotFound()) {
+        return true;
+      } else {
+        if (!s.ok()) {
+          status_ = s;
+        }
+        return false;
+      }
+    } else {
+      return false;
+    }
+  }
+
+  std::unique_ptr<ManagedSnapshot> snapshot_;
+  std::unique_ptr<ArenaWrappedDBIter> iter_;
+  BlobDBImpl* blob_db_;
+  SystemClock* clock_;
+  Statistics* statistics_;
+  Status status_;
+  PinnableSlice value_;
+};
+}  // namespace blob_db
+}  // namespace ROCKSDB_NAMESPACE
+#endif  // !ROCKSDB_LITE
diff --git a/src/rocksdb/utilities/blob_db/blob_db_listener.h b/src/rocksdb/utilities/blob_db/blob_db_listener.h
new file mode 100644
index 000000000..d17d29853
--- /dev/null
+++ b/src/rocksdb/utilities/blob_db/blob_db_listener.h
@@ -0,0 +1,71 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#pragma once
+
+#ifndef ROCKSDB_LITE
+
+#include <atomic>
+
+#include "rocksdb/listener.h"
+#include "util/mutexlock.h"
+#include "utilities/blob_db/blob_db_impl.h"
+
+namespace ROCKSDB_NAMESPACE {
+namespace blob_db {
+
+class BlobDBListener : public EventListener {
+ public:
+  explicit BlobDBListener(BlobDBImpl* blob_db_impl)
+      : blob_db_impl_(blob_db_impl) {}
+
+  void OnFlushBegin(DB* /*db*/, const FlushJobInfo& /*info*/) override {
+    assert(blob_db_impl_ != nullptr);
+    blob_db_impl_->SyncBlobFiles();
+  }
+
+  void OnFlushCompleted(DB* /*db*/, const FlushJobInfo& /*info*/) override {
+    assert(blob_db_impl_ != nullptr);
+    blob_db_impl_->UpdateLiveSSTSize();
+  }
+
+  void OnCompactionCompleted(DB* /*db*/,
+                             const CompactionJobInfo& /*info*/) override {
+    assert(blob_db_impl_ != nullptr);
+    blob_db_impl_->UpdateLiveSSTSize();
+  }
+
+  const char* Name() const override { return kClassName(); }
+  static const char* kClassName() { return "BlobDBListener"; }
+
+ protected:
+  BlobDBImpl* blob_db_impl_;
+};
+
+class BlobDBListenerGC : public BlobDBListener {
+ public:
+  explicit BlobDBListenerGC(BlobDBImpl* blob_db_impl)
+      : BlobDBListener(blob_db_impl) {}
+
+  const char* Name() const override { return kClassName(); }
+  static const char* kClassName() { return "BlobDBListenerGC"; }
+  void OnFlushCompleted(DB* db, const FlushJobInfo& info) override {
+    BlobDBListener::OnFlushCompleted(db, info);
+
+    assert(blob_db_impl_);
+    blob_db_impl_->ProcessFlushJobInfo(info);
+  }
+
+  void OnCompactionCompleted(DB* db, const CompactionJobInfo& info) override {
+    BlobDBListener::OnCompactionCompleted(db, info);
+
+    assert(blob_db_impl_);
+    blob_db_impl_->ProcessCompactionJobInfo(info);
+  }
+};
+
+}  // namespace blob_db
+}  // namespace ROCKSDB_NAMESPACE
+#endif  // !ROCKSDB_LITE
diff --git a/src/rocksdb/utilities/blob_db/blob_db_test.cc b/src/rocksdb/utilities/blob_db/blob_db_test.cc
new file mode 100644
index 000000000..e392962b2
--- /dev/null
+++ b/src/rocksdb/utilities/blob_db/blob_db_test.cc
@@ -0,0 +1,2407 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#ifndef ROCKSDB_LITE
+
+#include "utilities/blob_db/blob_db.h"
+
+#include <algorithm>
+#include <chrono>
+#include <cstdlib>
+#include <iomanip>
+#include <map>
+#include <memory>
+#include <sstream>
+#include <string>
+#include <vector>
+
+#include "db/blob/blob_index.h"
+#include "db/db_test_util.h"
+#include "env/composite_env_wrapper.h"
+#include "file/file_util.h"
+#include "file/sst_file_manager_impl.h"
+#include "port/port.h"
+#include "rocksdb/utilities/debug.h"
+#include "test_util/mock_time_env.h"
+#include "test_util/sync_point.h"
+#include "test_util/testharness.h"
+#include "util/random.h"
+#include "util/string_util.h"
+#include "utilities/blob_db/blob_db_impl.h"
+#include "utilities/fault_injection_env.h"
+
+namespace ROCKSDB_NAMESPACE {
+namespace blob_db {
+
+class BlobDBTest : public testing::Test {
+ public:
+  const int kMaxBlobSize = 1 << 14;
+
+  struct BlobIndexVersion {
+    BlobIndexVersion() = default;
+    BlobIndexVersion(std::string _user_key, uint64_t _file_number,
+                     uint64_t _expiration, SequenceNumber _sequence,
+                     ValueType _type)
+        : user_key(std::move(_user_key)),
+          file_number(_file_number),
+          expiration(_expiration),
+          sequence(_sequence),
+          type(_type) {}
+
+    std::string user_key;
+    uint64_t file_number = kInvalidBlobFileNumber;
+    uint64_t expiration = kNoExpiration;
+    SequenceNumber sequence = 0;
+    ValueType type = kTypeValue;
+  };
+
+  BlobDBTest()
+      : dbname_(test::PerThreadDBPath("blob_db_test")), blob_db_(nullptr) {
+    mock_clock_ = std::make_shared<MockSystemClock>(SystemClock::Default());
+    mock_env_.reset(new CompositeEnvWrapper(Env::Default(), mock_clock_));
+    fault_injection_env_.reset(new FaultInjectionTestEnv(Env::Default()));
+
+    Status s = DestroyBlobDB(dbname_, Options(), BlobDBOptions());
+    assert(s.ok());
+  }
+
+  ~BlobDBTest() override {
+    SyncPoint::GetInstance()->ClearAllCallBacks();
+    Destroy();
+  }
+
+  Status TryOpen(BlobDBOptions bdb_options = BlobDBOptions(),
+                 Options options = Options()) {
+    options.create_if_missing = true;
+    if (options.env == mock_env_.get()) {
+      // Need to disable stats dumping and persisting which also use
+      // RepeatableThread, which uses InstrumentedCondVar::TimedWaitInternal.
+      // With mocked time, this can hang on some platforms (MacOS)
+      // because (a) on some platforms, pthread_cond_timedwait does not appear
+      // to release the lock for other threads to operate if the deadline time
+      // is already passed, and (b) TimedWait calls are currently a bad
+      // abstraction because the deadline parameter is usually computed from
+      // Env time, but is interpreted in real clock time.
+      options.stats_dump_period_sec = 0;
+      options.stats_persist_period_sec = 0;
+    }
+    return BlobDB::Open(options, bdb_options, dbname_, &blob_db_);
+  }
+
+  void Open(BlobDBOptions bdb_options = BlobDBOptions(),
+            Options options = Options()) {
+    ASSERT_OK(TryOpen(bdb_options, options));
+  }
+
+  void Reopen(BlobDBOptions bdb_options = BlobDBOptions(),
+              Options options = Options()) {
+    assert(blob_db_ != nullptr);
+    delete blob_db_;
+    blob_db_ = nullptr;
+    Open(bdb_options, options);
+  }
+
+  void Close() {
+    assert(blob_db_ != nullptr);
+    delete blob_db_;
+    blob_db_ = nullptr;
+  }
+
+  void Destroy() {
+    if (blob_db_) {
+      Options options = blob_db_->GetOptions();
+      BlobDBOptions bdb_options = blob_db_->GetBlobDBOptions();
+      delete blob_db_;
+      blob_db_ = nullptr;
+      ASSERT_OK(DestroyBlobDB(dbname_, options, bdb_options));
+    }
+  }
+
+  BlobDBImpl *blob_db_impl() {
+    return reinterpret_cast<BlobDBImpl *>(blob_db_);
+  }
+
+  Status Put(const Slice &key, const Slice &value,
+             std::map<std::string, std::string> *data = nullptr) {
+    Status s = blob_db_->Put(WriteOptions(), key, value);
+    if (data != nullptr) {
+      (*data)[key.ToString()] = value.ToString();
+    }
+    return s;
+  }
+
+  void Delete(const std::string &key,
+              std::map<std::string, std::string> *data = nullptr) {
+    ASSERT_OK(blob_db_->Delete(WriteOptions(), key));
+    if (data != nullptr) {
+      data->erase(key);
+    }
+  }
+
+  Status PutWithTTL(const Slice &key, const Slice &value, uint64_t ttl,
+                    std::map<std::string, std::string> *data = nullptr) {
+    Status s = blob_db_->PutWithTTL(WriteOptions(), key, value, ttl);
+    if (data != nullptr) {
+      (*data)[key.ToString()] = value.ToString();
+    }
+    return s;
+  }
+
+  Status PutUntil(const Slice &key, const Slice &value, uint64_t expiration) {
+    return blob_db_->PutUntil(WriteOptions(), key, value, expiration);
+  }
+
+  void PutRandomWithTTL(const std::string &key, uint64_t ttl, Random *rnd,
+                        std::map<std::string, std::string> *data = nullptr) {
+    int len = rnd->Next() % kMaxBlobSize + 1;
+    std::string value = rnd->HumanReadableString(len);
+    ASSERT_OK(
+        blob_db_->PutWithTTL(WriteOptions(), Slice(key), Slice(value), ttl));
+    if (data != nullptr) {
+      (*data)[key] = value;
+    }
+  }
+
+  void PutRandomUntil(const std::string &key, uint64_t expiration, Random *rnd,
+                      std::map<std::string, std::string> *data = nullptr) {
+    int len = rnd->Next() % kMaxBlobSize + 1;
+    std::string value = rnd->HumanReadableString(len);
+    ASSERT_OK(blob_db_->PutUntil(WriteOptions(), Slice(key), Slice(value),
+                                 expiration));
+    if (data != nullptr) {
+      (*data)[key] = value;
+    }
+  }
+
+  void PutRandom(const std::string &key, Random *rnd,
+                 std::map<std::string, std::string> *data = nullptr) {
+    PutRandom(blob_db_, key, rnd, data);
+  }
+
+  void PutRandom(DB *db, const std::string &key, Random *rnd,
+                 std::map<std::string, std::string> *data = nullptr) {
+    int len = rnd->Next() % kMaxBlobSize + 1;
+    std::string value = rnd->HumanReadableString(len);
+    ASSERT_OK(db->Put(WriteOptions(), Slice(key), Slice(value)));
+    if (data != nullptr) {
+      (*data)[key] = value;
+    }
+  }
+
+  void PutRandomToWriteBatch(
+      const std::string &key, Random *rnd, WriteBatch *batch,
+      std::map<std::string, std::string> *data = nullptr) {
+    int len = rnd->Next() % kMaxBlobSize + 1;
+    std::string value = rnd->HumanReadableString(len);
+    ASSERT_OK(batch->Put(key, value));
+    if (data != nullptr) {
+      (*data)[key] = value;
+    }
+  }
+
+  // Verify blob db contain expected data and nothing more.
+  void VerifyDB(const std::map<std::string, std::string> &data) {
+    VerifyDB(blob_db_, data);
+  }
+
+  void VerifyDB(DB *db, const std::map<std::string, std::string> &data) {
+    // Verify normal Get
+    auto *cfh = db->DefaultColumnFamily();
+    for (auto &p : data) {
+      PinnableSlice value_slice;
+      ASSERT_OK(db->Get(ReadOptions(), cfh, p.first, &value_slice));
+      ASSERT_EQ(p.second, value_slice.ToString());
+      std::string value;
+      ASSERT_OK(db->Get(ReadOptions(), cfh, p.first, &value));
+      ASSERT_EQ(p.second, value);
+    }
+
+    // Verify iterators
+    Iterator *iter = db->NewIterator(ReadOptions());
+    iter->SeekToFirst();
+    for (auto &p : data) {
+      ASSERT_TRUE(iter->Valid());
+      ASSERT_EQ(p.first, iter->key().ToString());
+      ASSERT_EQ(p.second, iter->value().ToString());
+      iter->Next();
+    }
+    ASSERT_FALSE(iter->Valid());
+    ASSERT_OK(iter->status());
+    delete iter;
+  }
+
+  void VerifyBaseDB(
+      const std::map<std::string, KeyVersion> &expected_versions) {
+    auto *bdb_impl = static_cast<BlobDBImpl *>(blob_db_);
+    DB *db = blob_db_->GetRootDB();
+    const size_t kMaxKeys = 10000;
+    std::vector<KeyVersion> versions;
+    ASSERT_OK(GetAllKeyVersions(db, "", "", kMaxKeys, &versions));
+    ASSERT_EQ(expected_versions.size(), versions.size());
+    size_t i = 0;
+    for (auto &key_version : expected_versions) {
+      const KeyVersion &expected_version = key_version.second;
+      ASSERT_EQ(expected_version.user_key, versions[i].user_key);
+      ASSERT_EQ(expected_version.sequence, versions[i].sequence);
+      ASSERT_EQ(expected_version.type, versions[i].type);
+      if (versions[i].type == kTypeValue) {
+        ASSERT_EQ(expected_version.value, versions[i].value);
+      } else {
+        ASSERT_EQ(kTypeBlobIndex, versions[i].type);
+        PinnableSlice value;
+        ASSERT_OK(bdb_impl->TEST_GetBlobValue(versions[i].user_key,
+                                              versions[i].value, &value));
+        ASSERT_EQ(expected_version.value, value.ToString());
+      }
+      i++;
+    }
+  }
+
+  void VerifyBaseDBBlobIndex(
+      const std::map<std::string, BlobIndexVersion> &expected_versions) {
+    const size_t kMaxKeys = 10000;
+    std::vector<KeyVersion> versions;
+    ASSERT_OK(
+        GetAllKeyVersions(blob_db_->GetRootDB(), "", "", kMaxKeys, &versions));
+    ASSERT_EQ(versions.size(), expected_versions.size());
+
+    size_t i = 0;
+    for (const auto &expected_pair : expected_versions) {
+      const BlobIndexVersion &expected_version = expected_pair.second;
+
+      ASSERT_EQ(versions[i].user_key, expected_version.user_key);
+      ASSERT_EQ(versions[i].sequence, expected_version.sequence);
+      ASSERT_EQ(versions[i].type, expected_version.type);
+      if (versions[i].type != kTypeBlobIndex) {
+        ASSERT_EQ(kInvalidBlobFileNumber, expected_version.file_number);
+        ASSERT_EQ(kNoExpiration, expected_version.expiration);
+
+        ++i;
+        continue;
+      }
+
+      BlobIndex blob_index;
+      ASSERT_OK(blob_index.DecodeFrom(versions[i].value));
+
+      const uint64_t file_number = !blob_index.IsInlined()
+                                       ? blob_index.file_number()
+                                       : kInvalidBlobFileNumber;
+      ASSERT_EQ(file_number, expected_version.file_number);
+
+      const uint64_t expiration =
+          blob_index.HasTTL() ? blob_index.expiration() : kNoExpiration;
+      ASSERT_EQ(expiration, expected_version.expiration);
+
+      ++i;
+    }
+  }
+
+  void InsertBlobs() {
+    WriteOptions wo;
+    std::string value;
+
+    Random rnd(301);
+    for (size_t i = 0; i < 100000; i++) {
+      uint64_t ttl = rnd.Next() % 86400;
+      PutRandomWithTTL("key" + std::to_string(i % 500), ttl, &rnd, nullptr);
+    }
+
+    for (size_t i = 0; i < 10; i++) {
+      Delete("key" + std::to_string(i % 500));
+    }
+  }
+
+  const std::string dbname_;
+  std::shared_ptr<MockSystemClock> mock_clock_;
+  std::unique_ptr<Env> mock_env_;
+  std::unique_ptr<FaultInjectionTestEnv> fault_injection_env_;
+  BlobDB *blob_db_;
+};  // class BlobDBTest
+
+TEST_F(BlobDBTest, Put) {
+  Random rnd(301);
+  BlobDBOptions bdb_options;
+  bdb_options.min_blob_size = 0;
+  bdb_options.disable_background_tasks = true;
+  Open(bdb_options);
+  std::map<std::string, std::string> data;
+  for (size_t i = 0; i < 100; i++) {
+    PutRandom("key" + std::to_string(i), &rnd, &data);
+  }
+  VerifyDB(data);
+}
+
+TEST_F(BlobDBTest, PutWithTTL) {
+  Random rnd(301);
+  Options options;
+  options.env = mock_env_.get();
+  BlobDBOptions bdb_options;
+  bdb_options.ttl_range_secs = 1000;
+  bdb_options.min_blob_size = 0;
+  bdb_options.blob_file_size = 256 * 1000 * 1000;
+  bdb_options.disable_background_tasks = true;
+  Open(bdb_options, options);
+  std::map<std::string, std::string> data;
+  mock_clock_->SetCurrentTime(50);
+  for (size_t i = 0; i < 100; i++) {
+    uint64_t ttl = rnd.Next() % 100;
+    PutRandomWithTTL("key" + std::to_string(i), ttl, &rnd,
+                     (ttl <= 50 ? nullptr : &data));
+  }
+  mock_clock_->SetCurrentTime(100);
+  auto *bdb_impl = static_cast<BlobDBImpl *>(blob_db_);
+  auto blob_files = bdb_impl->TEST_GetBlobFiles();
+  ASSERT_EQ(1, blob_files.size());
+  ASSERT_TRUE(blob_files[0]->HasTTL());
+  ASSERT_OK(bdb_impl->TEST_CloseBlobFile(blob_files[0]));
+  VerifyDB(data);
+}
+
+TEST_F(BlobDBTest, PutUntil) {
+  Random rnd(301);
+  Options options;
+  options.env = mock_env_.get();
+  BlobDBOptions bdb_options;
+  bdb_options.ttl_range_secs = 1000;
+  bdb_options.min_blob_size = 0;
+  bdb_options.blob_file_size = 256 * 1000 * 1000;
+  bdb_options.disable_background_tasks = true;
+  Open(bdb_options, options);
+  std::map<std::string, std::string> data;
+  mock_clock_->SetCurrentTime(50);
+  for (size_t i = 0; i < 100; i++) {
+    uint64_t expiration = rnd.Next() % 100 + 50;
+    PutRandomUntil("key" + std::to_string(i), expiration, &rnd,
+                   (expiration <= 100 ? nullptr : &data));
+  }
+  mock_clock_->SetCurrentTime(100);
+  auto *bdb_impl = static_cast<BlobDBImpl *>(blob_db_);
+  auto blob_files = bdb_impl->TEST_GetBlobFiles();
+  ASSERT_EQ(1, blob_files.size());
+  ASSERT_TRUE(blob_files[0]->HasTTL());
+  ASSERT_OK(bdb_impl->TEST_CloseBlobFile(blob_files[0]));
+  VerifyDB(data);
+}
+
+TEST_F(BlobDBTest, StackableDBGet) {
+  Random rnd(301);
+  BlobDBOptions bdb_options;
+  bdb_options.min_blob_size = 0;
+  bdb_options.disable_background_tasks = true;
+  Open(bdb_options);
+  std::map<std::string, std::string> data;
+  for (size_t i = 0; i < 100; i++) {
+    PutRandom("key" + std::to_string(i), &rnd, &data);
+  }
+  for (size_t i = 0; i < 100; i++) {
+    StackableDB *db = blob_db_;
+    ColumnFamilyHandle *column_family = db->DefaultColumnFamily();
+    std::string key = "key" + std::to_string(i);
+    PinnableSlice pinnable_value;
+    ASSERT_OK(db->Get(ReadOptions(), column_family, key, &pinnable_value));
+    std::string string_value;
+    ASSERT_OK(db->Get(ReadOptions(), column_family, key, &string_value));
+    ASSERT_EQ(string_value, pinnable_value.ToString());
+    ASSERT_EQ(string_value, data[key]);
+  }
+}
+
+TEST_F(BlobDBTest, GetExpiration) {
+  Options options;
+  options.env = mock_env_.get();
+  BlobDBOptions bdb_options;
+  bdb_options.disable_background_tasks = true;
+  mock_clock_->SetCurrentTime(100);
+  Open(bdb_options, options);
+  ASSERT_OK(Put("key1", "value1"));
+  ASSERT_OK(PutWithTTL("key2", "value2", 200));
+  PinnableSlice value;
+  uint64_t expiration;
+  ASSERT_OK(blob_db_->Get(ReadOptions(), "key1", &value, &expiration));
+  ASSERT_EQ("value1", value.ToString());
+  ASSERT_EQ(kNoExpiration, expiration);
+  ASSERT_OK(blob_db_->Get(ReadOptions(), "key2", &value, &expiration));
+  ASSERT_EQ("value2", value.ToString());
+  ASSERT_EQ(300 /* = 100 + 200 */, expiration);
+}
+
+TEST_F(BlobDBTest, GetIOError) {
+  Options options;
+  options.env = fault_injection_env_.get();
+  BlobDBOptions bdb_options;
+  bdb_options.min_blob_size = 0;  // Make sure value write to blob file
+  bdb_options.disable_background_tasks = true;
+  Open(bdb_options, options);
+  ColumnFamilyHandle *column_family = blob_db_->DefaultColumnFamily();
+  PinnableSlice value;
+  ASSERT_OK(Put("foo", "bar"));
+  fault_injection_env_->SetFilesystemActive(false, Status::IOError());
+  Status s = blob_db_->Get(ReadOptions(), column_family, "foo", &value);
+  ASSERT_TRUE(s.IsIOError());
+  // Reactivate file system to allow test to close DB.
+  fault_injection_env_->SetFilesystemActive(true);
+}
+
+TEST_F(BlobDBTest, PutIOError) {
+  Options options;
+  options.env = fault_injection_env_.get();
+  BlobDBOptions bdb_options;
+  bdb_options.min_blob_size = 0;  // Make sure value write to blob file
+  bdb_options.disable_background_tasks = true;
+  Open(bdb_options, options);
+  fault_injection_env_->SetFilesystemActive(false, Status::IOError());
+  ASSERT_TRUE(Put("foo", "v1").IsIOError());
+  fault_injection_env_->SetFilesystemActive(true, Status::IOError());
+  ASSERT_OK(Put("bar", "v1"));
+}
+
+TEST_F(BlobDBTest, WriteBatch) {
+  Random rnd(301);
+  BlobDBOptions bdb_options;
+  bdb_options.min_blob_size = 0;
+  bdb_options.disable_background_tasks = true;
+  Open(bdb_options);
+  std::map<std::string, std::string> data;
+  for (size_t i = 0; i < 100; i++) {
+    WriteBatch batch;
+    for (size_t j = 0; j < 10; j++) {
+      PutRandomToWriteBatch("key" + std::to_string(j * 100 + i), &rnd, &batch,
+                            &data);
+    }
+
+    ASSERT_OK(blob_db_->Write(WriteOptions(), &batch));
+  }
+  VerifyDB(data);
+}
+
+TEST_F(BlobDBTest, Delete) {
+  Random rnd(301);
+  BlobDBOptions bdb_options;
+  bdb_options.min_blob_size = 0;
+  bdb_options.disable_background_tasks = true;
+  Open(bdb_options);
+  std::map<std::string, std::string> data;
+  for (size_t i = 0; i < 100; i++) {
+    PutRandom("key" + std::to_string(i), &rnd, &data);
+  }
+  for (size_t i = 0; i < 100; i += 5) {
+    Delete("key" + std::to_string(i), &data);
+  }
+  VerifyDB(data);
+}
+
+TEST_F(BlobDBTest, DeleteBatch) {
+  Random rnd(301);
+  BlobDBOptions bdb_options;
+  bdb_options.min_blob_size = 0;
+  bdb_options.disable_background_tasks = true;
+  Open(bdb_options);
+  for (size_t i = 0; i < 100; i++) {
+    PutRandom("key" + std::to_string(i), &rnd);
+  }
+  WriteBatch batch;
+  for (size_t i = 0; i < 100; i++) {
+    ASSERT_OK(batch.Delete("key" + std::to_string(i)));
+  }
+  ASSERT_OK(blob_db_->Write(WriteOptions(), &batch));
+  // DB should be empty.
+  VerifyDB({});
+}
+
+TEST_F(BlobDBTest, Override) {
+  Random rnd(301);
+  BlobDBOptions bdb_options;
+  bdb_options.min_blob_size = 0;
+  bdb_options.disable_background_tasks = true;
+  Open(bdb_options);
+  std::map<std::string, std::string> data;
+  for (int i = 0; i < 10000; i++) {
+    PutRandom("key" + std::to_string(i), &rnd, nullptr);
+  }
+  // override all the keys
+  for (int i = 0; i < 10000; i++) {
+    PutRandom("key" + std::to_string(i), &rnd, &data);
+  }
+  VerifyDB(data);
+}
+
+#ifdef SNAPPY
+TEST_F(BlobDBTest, Compression) {
+  Random rnd(301);
+  BlobDBOptions bdb_options;
+  bdb_options.min_blob_size = 0;
+  bdb_options.disable_background_tasks = true;
+  bdb_options.compression = CompressionType::kSnappyCompression;
+  Open(bdb_options);
+  std::map<std::string, std::string> data;
+  for (size_t i = 0; i < 100; i++) {
+    PutRandom("put-key" + std::to_string(i), &rnd, &data);
+  }
+  for (int i = 0; i < 100; i++) {
+    WriteBatch batch;
+    for (size_t j = 0; j < 10; j++) {
+      PutRandomToWriteBatch("write-batch-key" + std::to_string(j * 100 + i),
+                            &rnd, &batch, &data);
+    }
+    ASSERT_OK(blob_db_->Write(WriteOptions(), &batch));
+  }
+  VerifyDB(data);
+}
+
+TEST_F(BlobDBTest, DecompressAfterReopen) {
+  Random rnd(301);
+  BlobDBOptions bdb_options;
+  bdb_options.min_blob_size = 0;
+  bdb_options.disable_background_tasks = true;
+  bdb_options.compression = CompressionType::kSnappyCompression;
+  Open(bdb_options);
+  std::map<std::string, std::string> data;
+  for (size_t i = 0; i < 100; i++) {
+    PutRandom("put-key" + std::to_string(i), &rnd, &data);
+  }
+  VerifyDB(data);
+  bdb_options.compression = CompressionType::kNoCompression;
+  Reopen(bdb_options);
+  VerifyDB(data);
+}
+
+TEST_F(BlobDBTest, EnableDisableCompressionGC) {
+  Random rnd(301);
+  BlobDBOptions bdb_options;
+  bdb_options.min_blob_size = 0;
+  bdb_options.garbage_collection_cutoff = 1.0;
+  bdb_options.disable_background_tasks = true;
+  bdb_options.compression = kSnappyCompression;
+  Open(bdb_options);
+  std::map<std::string, std::string> data;
+  size_t data_idx = 0;
+  for (; data_idx < 100; data_idx++) {
+    PutRandom("put-key" + std::to_string(data_idx), &rnd, &data);
+  }
+  VerifyDB(data);
+  auto blob_files = blob_db_impl()->TEST_GetBlobFiles();
+  ASSERT_EQ(1, blob_files.size());
+  ASSERT_EQ(kSnappyCompression, blob_files[0]->GetCompressionType());
+
+  // disable compression
+  bdb_options.compression = kNoCompression;
+  Reopen(bdb_options);
+
+  // Add more data with new compression type
+  for (; data_idx < 200; data_idx++) {
+    PutRandom("put-key" + std::to_string(data_idx), &rnd, &data);
+  }
+  VerifyDB(data);
+
+  blob_files = blob_db_impl()->TEST_GetBlobFiles();
+  ASSERT_EQ(2, blob_files.size());
+  ASSERT_EQ(kNoCompression, blob_files[1]->GetCompressionType());
+
+  // Enable GC. If we do it earlier the snapshot release triggered compaction
+  // may compact files and trigger GC before we can verify there are two files.
+  bdb_options.enable_garbage_collection = true;
+  Reopen(bdb_options);
+
+  // Trigger compaction
+  ASSERT_OK(blob_db_->CompactRange(CompactRangeOptions(), nullptr, nullptr));
+  blob_db_impl()->TEST_DeleteObsoleteFiles();
+  VerifyDB(data);
+
+  blob_files = blob_db_impl()->TEST_GetBlobFiles();
+  for (auto bfile : blob_files) {
+    ASSERT_EQ(kNoCompression, bfile->GetCompressionType());
+  }
+
+  // enabling the compression again
+  bdb_options.compression = kSnappyCompression;
+  Reopen(bdb_options);
+
+  // Add more data with new compression type
+  for (; data_idx < 300; data_idx++) {
+    PutRandom("put-key" + std::to_string(data_idx), &rnd, &data);
+  }
+  VerifyDB(data);
+
+  // Trigger compaction
+  ASSERT_OK(blob_db_->CompactRange(CompactRangeOptions(), nullptr, nullptr));
+  blob_db_impl()->TEST_DeleteObsoleteFiles();
+  VerifyDB(data);
+
+  blob_files = blob_db_impl()->TEST_GetBlobFiles();
+  for (auto bfile : blob_files) {
+    ASSERT_EQ(kSnappyCompression, bfile->GetCompressionType());
+  }
+}
+
+#ifdef LZ4
+// Test switch compression types and run GC, it needs both Snappy and LZ4
+// support.
+TEST_F(BlobDBTest, ChangeCompressionGC) {
+  Random rnd(301);
+  BlobDBOptions bdb_options;
+  bdb_options.min_blob_size = 0;
+  bdb_options.garbage_collection_cutoff = 1.0;
+  bdb_options.disable_background_tasks = true;
+  bdb_options.compression = kLZ4Compression;
+  Open(bdb_options);
+  std::map<std::string, std::string> data;
+  size_t data_idx = 0;
+  for (; data_idx < 100; data_idx++) {
+    PutRandom("put-key" + std::to_string(data_idx), &rnd, &data);
+  }
+  VerifyDB(data);
+  auto blob_files = blob_db_impl()->TEST_GetBlobFiles();
+  ASSERT_EQ(1, blob_files.size());
+  ASSERT_EQ(kLZ4Compression, blob_files[0]->GetCompressionType());
+
+  // Change compression type
+  bdb_options.compression = kSnappyCompression;
+  Reopen(bdb_options);
+
+  // Add more data with Snappy compression type
+  for (; data_idx < 200; data_idx++) {
+    PutRandom("put-key" + std::to_string(data_idx), &rnd, &data);
+  }
+  VerifyDB(data);
+
+  // Verify blob file compression type
+  blob_files = blob_db_impl()->TEST_GetBlobFiles();
+  ASSERT_EQ(2, blob_files.size());
+  ASSERT_EQ(kSnappyCompression, blob_files[1]->GetCompressionType());
+
+  // Enable GC. If we do it earlier the snapshot release triggered compaction
+  // may compact files and trigger GC before we can verify there are two files.
+  bdb_options.enable_garbage_collection = true;
+  Reopen(bdb_options);
+
+  ASSERT_OK(blob_db_->CompactRange(CompactRangeOptions(), nullptr, nullptr));
+  VerifyDB(data);
+
+  blob_db_impl()->TEST_DeleteObsoleteFiles();
+  blob_files = blob_db_impl()->TEST_GetBlobFiles();
+  for (auto bfile : blob_files) {
+    ASSERT_EQ(kSnappyCompression, bfile->GetCompressionType());
+  }
+
+  // Disable compression
+  bdb_options.compression = kNoCompression;
+  Reopen(bdb_options);
+  for (; data_idx < 300; data_idx++) {
+    PutRandom("put-key" + std::to_string(data_idx), &rnd, &data);
+  }
+  VerifyDB(data);
+
+  ASSERT_OK(blob_db_->CompactRange(CompactRangeOptions(), nullptr, nullptr));
+  VerifyDB(data);
+
+  blob_db_impl()->TEST_DeleteObsoleteFiles();
+  blob_files = blob_db_impl()->TEST_GetBlobFiles();
+  for (auto bfile : blob_files) {
+    ASSERT_EQ(kNoCompression, bfile->GetCompressionType());
+  }
+
+  // switching different compression types to generate mixed compression types
+  bdb_options.compression = kSnappyCompression;
+  Reopen(bdb_options);
+  for (; data_idx < 400; data_idx++) {
+    PutRandom("put-key" + std::to_string(data_idx), &rnd, &data);
+  }
+  VerifyDB(data);
+
+  bdb_options.compression = kLZ4Compression;
+  Reopen(bdb_options);
+  for (; data_idx < 500; data_idx++) {
+    PutRandom("put-key" + std::to_string(data_idx), &rnd, &data);
+  }
+  VerifyDB(data);
+
+  ASSERT_OK(blob_db_->CompactRange(CompactRangeOptions(), nullptr, nullptr));
+  VerifyDB(data);
+
+  blob_db_impl()->TEST_DeleteObsoleteFiles();
+  blob_files = blob_db_impl()->TEST_GetBlobFiles();
+  for (auto bfile : blob_files) {
+    ASSERT_EQ(kLZ4Compression, bfile->GetCompressionType());
+  }
+}
+#endif  // LZ4
+#endif  // SNAPPY
+
+TEST_F(BlobDBTest, MultipleWriters) {
+  Open(BlobDBOptions());
+
+  std::vector<port::Thread> workers;
+  std::vector<std::map<std::string, std::string>> data_set(10);
+  for (uint32_t i = 0; i < 10; i++)
+    workers.push_back(port::Thread(
+        [&](uint32_t id) {
+          Random rnd(301 + id);
+          for (int j = 0; j < 100; j++) {
+            std::string key =
+                "key" + std::to_string(id) + "_" + std::to_string(j);
+            if (id < 5) {
+              PutRandom(key, &rnd, &data_set[id]);
+            } else {
+              WriteBatch batch;
+              PutRandomToWriteBatch(key, &rnd, &batch, &data_set[id]);
+              ASSERT_OK(blob_db_->Write(WriteOptions(), &batch));
+            }
+          }
+        },
+        i));
+  std::map<std::string, std::string> data;
+  for (size_t i = 0; i < 10; i++) {
+    workers[i].join();
+    data.insert(data_set[i].begin(), data_set[i].end());
+  }
+  VerifyDB(data);
+}
+
+TEST_F(BlobDBTest, SstFileManager) {
+  // run the same test for Get(), MultiGet() and Iterator each.
+  std::shared_ptr<SstFileManager> sst_file_manager(
+      NewSstFileManager(mock_env_.get()));
+  sst_file_manager->SetDeleteRateBytesPerSecond(1);
+  SstFileManagerImpl *sfm =
+      static_cast<SstFileManagerImpl *>(sst_file_manager.get());
+
+  BlobDBOptions bdb_options;
+  bdb_options.min_blob_size = 0;
+  bdb_options.enable_garbage_collection = true;
+  bdb_options.garbage_collection_cutoff = 1.0;
+  Options db_options;
+
+  int files_scheduled_to_delete = 0;
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
+      "SstFileManagerImpl::ScheduleFileDeletion", [&](void *arg) {
+        assert(arg);
+        const std::string *const file_path =
+            static_cast<const std::string *>(arg);
+        if (file_path->find(".blob") != std::string::npos) {
+          ++files_scheduled_to_delete;
+        }
+      });
+  SyncPoint::GetInstance()->EnableProcessing();
+  db_options.sst_file_manager = sst_file_manager;
+
+  Open(bdb_options, db_options);
+
+  // Create one obselete file and clean it.
+  ASSERT_OK(blob_db_->Put(WriteOptions(), "foo", "bar"));
+  auto blob_files = blob_db_impl()->TEST_GetBlobFiles();
+  ASSERT_EQ(1, blob_files.size());
+  std::shared_ptr<BlobFile> bfile = blob_files[0];
+  ASSERT_OK(blob_db_impl()->TEST_CloseBlobFile(bfile));
+  ASSERT_OK(blob_db_->CompactRange(CompactRangeOptions(), nullptr, nullptr));
+  blob_db_impl()->TEST_DeleteObsoleteFiles();
+
+  // Even if SSTFileManager is not set, DB is creating a dummy one.
+  ASSERT_EQ(1, files_scheduled_to_delete);
+  Destroy();
+  // Make sure that DestroyBlobDB() also goes through delete scheduler.
+  ASSERT_EQ(2, files_scheduled_to_delete);
+  SyncPoint::GetInstance()->DisableProcessing();
+  sfm->WaitForEmptyTrash();
+}
+
+TEST_F(BlobDBTest, SstFileManagerRestart) {
+  int files_scheduled_to_delete = 0;
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
+      "SstFileManagerImpl::ScheduleFileDeletion", [&](void *arg) {
+        assert(arg);
+        const std::string *const file_path =
+            static_cast<const std::string *>(arg);
+        if (file_path->find(".blob") != std::string::npos) {
+          ++files_scheduled_to_delete;
+        }
+      });
+
+  // run the same test for Get(), MultiGet() and Iterator each.
+  std::shared_ptr<SstFileManager> sst_file_manager(
+      NewSstFileManager(mock_env_.get()));
+  sst_file_manager->SetDeleteRateBytesPerSecond(1);
+  SstFileManagerImpl *sfm =
+      static_cast<SstFileManagerImpl *>(sst_file_manager.get());
+
+  BlobDBOptions bdb_options;
+  bdb_options.min_blob_size = 0;
+  Options db_options;
+
+  SyncPoint::GetInstance()->EnableProcessing();
+  db_options.sst_file_manager = sst_file_manager;
+
+  Open(bdb_options, db_options);
+  std::string blob_dir = blob_db_impl()->TEST_blob_dir();
+  ASSERT_OK(blob_db_->Put(WriteOptions(), "foo", "bar"));
+  Close();
+
+  // Create 3 dummy trash files under the blob_dir
+  const auto &fs = db_options.env->GetFileSystem();
+  ASSERT_OK(CreateFile(fs, blob_dir + "/000666.blob.trash", "", false));
+  ASSERT_OK(CreateFile(fs, blob_dir + "/000888.blob.trash", "", true));
+  ASSERT_OK(CreateFile(fs, blob_dir + "/something_not_match.trash", "", false));
+
+  // Make sure that reopening the DB rescan the existing trash files
+  Open(bdb_options, db_options);
+  ASSERT_EQ(files_scheduled_to_delete, 2);
+
+  sfm->WaitForEmptyTrash();
+
+  // There should be exact one file under the blob dir now.
+  std::vector<std::string> all_files;
+  ASSERT_OK(db_options.env->GetChildren(blob_dir, &all_files));
+  int nfiles = 0;
+  for (const auto &f : all_files) {
+    assert(!f.empty());
+    if (f[0] == '.') {
+      continue;
+    }
+    nfiles++;
+  }
+  ASSERT_EQ(nfiles, 1);
+
+  SyncPoint::GetInstance()->DisableProcessing();
+}
+
+TEST_F(BlobDBTest, SnapshotAndGarbageCollection) {
+  BlobDBOptions bdb_options;
+  bdb_options.min_blob_size = 0;
+  bdb_options.enable_garbage_collection = true;
+  bdb_options.garbage_collection_cutoff = 1.0;
+  bdb_options.disable_background_tasks = true;
+
+  Options options;
+  options.disable_auto_compactions = true;
+
+  // i = when to take snapshot
+  for (int i = 0; i < 4; i++) {
+    Destroy();
+    Open(bdb_options, options);
+
+    const Snapshot *snapshot = nullptr;
+
+    // First file
+    ASSERT_OK(Put("key1", "value"));
+    if (i == 0) {
+      snapshot = blob_db_->GetSnapshot();
+    }
+
+    auto blob_files = blob_db_impl()->TEST_GetBlobFiles();
+    ASSERT_EQ(1, blob_files.size());
+    ASSERT_OK(blob_db_impl()->TEST_CloseBlobFile(blob_files[0]));
+
+    // Second file
+    ASSERT_OK(Put("key2", "value"));
+    if (i == 1) {
+      snapshot = blob_db_->GetSnapshot();
+    }
+
+    blob_files = blob_db_impl()->TEST_GetBlobFiles();
+    ASSERT_EQ(2, blob_files.size());
+    auto bfile = blob_files[1];
+    ASSERT_FALSE(bfile->Immutable());
+    ASSERT_OK(blob_db_impl()->TEST_CloseBlobFile(bfile));
+
+    // Third file
+    ASSERT_OK(Put("key3", "value"));
+    if (i == 2) {
+      snapshot = blob_db_->GetSnapshot();
+    }
+
+    ASSERT_OK(blob_db_->CompactRange(CompactRangeOptions(), nullptr, nullptr));
+    ASSERT_TRUE(bfile->Obsolete());
+    ASSERT_EQ(blob_db_->GetLatestSequenceNumber(),
+              bfile->GetObsoleteSequence());
+
+    Delete("key2");
+    if (i == 3) {
+      snapshot = blob_db_->GetSnapshot();
+    }
+
+    ASSERT_EQ(4, blob_db_impl()->TEST_GetBlobFiles().size());
+    blob_db_impl()->TEST_DeleteObsoleteFiles();
+
+    if (i >= 2) {
+      // The snapshot shouldn't see data in bfile
+      ASSERT_EQ(2, blob_db_impl()->TEST_GetBlobFiles().size());
+      blob_db_->ReleaseSnapshot(snapshot);
+    } else {
+      // The snapshot will see data in bfile, so the file shouldn't be deleted
+      ASSERT_EQ(4, blob_db_impl()->TEST_GetBlobFiles().size());
+      blob_db_->ReleaseSnapshot(snapshot);
+      blob_db_impl()->TEST_DeleteObsoleteFiles();
+      ASSERT_EQ(2, blob_db_impl()->TEST_GetBlobFiles().size());
+    }
+  }
+}
+
+TEST_F(BlobDBTest, ColumnFamilyNotSupported) {
+  Options options;
+  options.env = mock_env_.get();
+  mock_clock_->SetCurrentTime(0);
+  Open(BlobDBOptions(), options);
+  ColumnFamilyHandle *default_handle = blob_db_->DefaultColumnFamily();
+  ColumnFamilyHandle *handle = nullptr;
+  std::string value;
+  std::vector<std::string> values;
+  // The call simply pass through to base db. It should succeed.
+  ASSERT_OK(
+      blob_db_->CreateColumnFamily(ColumnFamilyOptions(), "foo", &handle));
+  ASSERT_TRUE(blob_db_->Put(WriteOptions(), handle, "k", "v").IsNotSupported());
+  ASSERT_TRUE(blob_db_->PutWithTTL(WriteOptions(), handle, "k", "v", 60)
+                  .IsNotSupported());
+  ASSERT_TRUE(blob_db_->PutUntil(WriteOptions(), handle, "k", "v", 100)
+                  .IsNotSupported());
+  WriteBatch batch;
+  ASSERT_OK(batch.Put("k1", "v1"));
+  ASSERT_OK(batch.Put(handle, "k2", "v2"));
+  ASSERT_TRUE(blob_db_->Write(WriteOptions(), &batch).IsNotSupported());
+  ASSERT_TRUE(blob_db_->Get(ReadOptions(), "k1", &value).IsNotFound());
+  ASSERT_TRUE(
+      blob_db_->Get(ReadOptions(), handle, "k", &value).IsNotSupported());
+  auto statuses = blob_db_->MultiGet(ReadOptions(), {default_handle, handle},
+                                     {"k1", "k2"}, &values);
+  ASSERT_EQ(2, statuses.size());
+  ASSERT_TRUE(statuses[0].IsNotSupported());
+  ASSERT_TRUE(statuses[1].IsNotSupported());
+  ASSERT_EQ(nullptr, blob_db_->NewIterator(ReadOptions(), handle));
+  delete handle;
+}
+
+TEST_F(BlobDBTest, GetLiveFilesMetaData) {
+  Random rnd(301);
+
+  BlobDBOptions bdb_options;
+  bdb_options.blob_dir = "blob_dir";
+  bdb_options.path_relative = true;
+  bdb_options.ttl_range_secs = 10;
+  bdb_options.min_blob_size = 0;
+  bdb_options.disable_background_tasks = true;
+
+  Options options;
+  options.env = mock_env_.get();
+
+  Open(bdb_options, options);
+
+  std::map<std::string, std::string> data;
+  for (size_t i = 0; i < 100; i++) {
+    PutRandom("key" + std::to_string(i), &rnd, &data);
+  }
+
+  constexpr uint64_t expiration = 1000ULL;
+  PutRandomUntil("key100", expiration, &rnd, &data);
+
+  std::vector<LiveFileMetaData> metadata;
+  blob_db_->GetLiveFilesMetaData(&metadata);
+
+  ASSERT_EQ(2U, metadata.size());
+  // Path should be relative to db_name, but begin with slash.
+  const std::string filename1("/blob_dir/000001.blob");
+  ASSERT_EQ(filename1, metadata[0].name);
+  ASSERT_EQ(1, metadata[0].file_number);
+  ASSERT_EQ(0, metadata[0].oldest_ancester_time);
+  ASSERT_EQ(kDefaultColumnFamilyName, metadata[0].column_family_name);
+
+  const std::string filename2("/blob_dir/000002.blob");
+  ASSERT_EQ(filename2, metadata[1].name);
+  ASSERT_EQ(2, metadata[1].file_number);
+  ASSERT_EQ(expiration, metadata[1].oldest_ancester_time);
+  ASSERT_EQ(kDefaultColumnFamilyName, metadata[1].column_family_name);
+
+  std::vector<std::string> livefile;
+  uint64_t mfs;
+  ASSERT_OK(blob_db_->GetLiveFiles(livefile, &mfs, false));
+  ASSERT_EQ(5U, livefile.size());
+  ASSERT_EQ(filename1, livefile[3]);
+  ASSERT_EQ(filename2, livefile[4]);
+  VerifyDB(data);
+}
+
+TEST_F(BlobDBTest, MigrateFromPlainRocksDB) {
+  constexpr size_t kNumKey = 20;
+  constexpr size_t kNumIteration = 10;
+  Random rnd(301);
+  std::map<std::string, std::string> data;
+  std::vector<bool> is_blob(kNumKey, false);
+
+  // Write to plain rocksdb.
+  Options options;
+  options.create_if_missing = true;
+  DB *db = nullptr;
+  ASSERT_OK(DB::Open(options, dbname_, &db));
+  for (size_t i = 0; i < kNumIteration; i++) {
+    auto key_index = rnd.Next() % kNumKey;
+    std::string key = "key" + std::to_string(key_index);
+    PutRandom(db, key, &rnd, &data);
+  }
+  VerifyDB(db, data);
+  delete db;
+  db = nullptr;
+
+  // Open as blob db. Verify it can read existing data.
+  Open();
+  VerifyDB(blob_db_, data);
+  for (size_t i = 0; i < kNumIteration; i++) {
+    auto key_index = rnd.Next() % kNumKey;
+    std::string key = "key" + std::to_string(key_index);
+    is_blob[key_index] = true;
+    PutRandom(blob_db_, key, &rnd, &data);
+  }
+  VerifyDB(blob_db_, data);
+  delete blob_db_;
+  blob_db_ = nullptr;
+
+  // Verify plain db return error for keys written by blob db.
+  ASSERT_OK(DB::Open(options, dbname_, &db));
+  std::string value;
+  for (size_t i = 0; i < kNumKey; i++) {
+    std::string key = "key" + std::to_string(i);
+    Status s = db->Get(ReadOptions(), key, &value);
+    if (data.count(key) == 0) {
+      ASSERT_TRUE(s.IsNotFound());
+    } else if (is_blob[i]) {
+      ASSERT_TRUE(s.IsCorruption());
+    } else {
+      ASSERT_OK(s);
+      ASSERT_EQ(data[key], value);
+    }
+  }
+  delete db;
+}
+
+// Test to verify that a NoSpace IOError Status is returned on reaching
+// max_db_size limit.
+TEST_F(BlobDBTest, OutOfSpace) {
+  // Use mock env to stop wall clock.
+  Options options;
+  options.env = mock_env_.get();
+  BlobDBOptions bdb_options;
+  bdb_options.max_db_size = 200;
+  bdb_options.is_fifo = false;
+  bdb_options.disable_background_tasks = true;
+  Open(bdb_options);
+
+  // Each stored blob has an overhead of about 42 bytes currently.
+  // So a small key + a 100 byte blob should take up ~150 bytes in the db.
+  std::string value(100, 'v');
+  ASSERT_OK(blob_db_->PutWithTTL(WriteOptions(), "key1", value, 60));
+
+  // Putting another blob should fail as ading it would exceed the max_db_size
+  // limit.
+  Status s = blob_db_->PutWithTTL(WriteOptions(), "key2", value, 60);
+  ASSERT_TRUE(s.IsIOError());
+  ASSERT_TRUE(s.IsNoSpace());
+}
+
+TEST_F(BlobDBTest, FIFOEviction) {
+  BlobDBOptions bdb_options;
+  bdb_options.max_db_size = 200;
+  bdb_options.blob_file_size = 100;
+  bdb_options.is_fifo = true;
+  bdb_options.disable_background_tasks = true;
+  Open(bdb_options);
+
+  std::atomic<int> evict_count{0};
+  SyncPoint::GetInstance()->SetCallBack(
+      "BlobDBImpl::EvictOldestBlobFile:Evicted",
+      [&](void *) { evict_count++; });
+  SyncPoint::GetInstance()->EnableProcessing();
+
+  // Each stored blob has an overhead of 32 bytes currently.
+  // So a 100 byte blob should take up 132 bytes.
+  std::string value(100, 'v');
+  ASSERT_OK(blob_db_->PutWithTTL(WriteOptions(), "key1", value, 10));
+  VerifyDB({{"key1", value}});
+
+  ASSERT_EQ(1, blob_db_impl()->TEST_GetBlobFiles().size());
+
+  // Adding another 100 bytes blob would take the total size to 264 bytes
+  // (2*132). max_db_size will be exceeded
+  // than max_db_size and trigger FIFO eviction.
+  ASSERT_OK(blob_db_->PutWithTTL(WriteOptions(), "key2", value, 60));
+  ASSERT_EQ(1, evict_count);
+  // key1 will exist until corresponding file be deleted.
+  VerifyDB({{"key1", value}, {"key2", value}});
+
+  // Adding another 100 bytes blob without TTL.
+  ASSERT_OK(blob_db_->Put(WriteOptions(), "key3", value));
+  ASSERT_EQ(2, evict_count);
+  // key1 and key2 will exist until corresponding file be deleted.
+  VerifyDB({{"key1", value}, {"key2", value}, {"key3", value}});
+
+  // The fourth blob file, without TTL.
+  ASSERT_OK(blob_db_->Put(WriteOptions(), "key4", value));
+  ASSERT_EQ(3, evict_count);
+  VerifyDB(
+      {{"key1", value}, {"key2", value}, {"key3", value}, {"key4", value}});
+
+  auto blob_files = blob_db_impl()->TEST_GetBlobFiles();
+  ASSERT_EQ(4, blob_files.size());
+  ASSERT_TRUE(blob_files[0]->Obsolete());
+  ASSERT_TRUE(blob_files[1]->Obsolete());
+  ASSERT_TRUE(blob_files[2]->Obsolete());
+  ASSERT_FALSE(blob_files[3]->Obsolete());
+  auto obsolete_files = blob_db_impl()->TEST_GetObsoleteFiles();
+  ASSERT_EQ(3, obsolete_files.size());
+  ASSERT_EQ(blob_files[0], obsolete_files[0]);
+  ASSERT_EQ(blob_files[1], obsolete_files[1]);
+  ASSERT_EQ(blob_files[2], obsolete_files[2]);
+
+  blob_db_impl()->TEST_DeleteObsoleteFiles();
+  obsolete_files = blob_db_impl()->TEST_GetObsoleteFiles();
+  ASSERT_TRUE(obsolete_files.empty());
+  VerifyDB({{"key4", value}});
+}
+
+TEST_F(BlobDBTest, FIFOEviction_NoOldestFileToEvict) {
+  Options options;
+  BlobDBOptions bdb_options;
+  bdb_options.max_db_size = 1000;
+  bdb_options.blob_file_size = 5000;
+  bdb_options.is_fifo = true;
+  bdb_options.disable_background_tasks = true;
+  Open(bdb_options);
+
+  std::atomic<int> evict_count{0};
+  SyncPoint::GetInstance()->SetCallBack(
+      "BlobDBImpl::EvictOldestBlobFile:Evicted",
+      [&](void *) { evict_count++; });
+  SyncPoint::GetInstance()->EnableProcessing();
+
+  std::string value(2000, 'v');
+  ASSERT_TRUE(Put("foo", std::string(2000, 'v')).IsNoSpace());
+  ASSERT_EQ(0, evict_count);
+}
+
+TEST_F(BlobDBTest, FIFOEviction_NoEnoughBlobFilesToEvict) {
+  BlobDBOptions bdb_options;
+  bdb_options.is_fifo = true;
+  bdb_options.min_blob_size = 100;
+  bdb_options.disable_background_tasks = true;
+  Options options;
+  // Use mock env to stop wall clock.
+  options.env = mock_env_.get();
+  options.disable_auto_compactions = true;
+  auto statistics = CreateDBStatistics();
+  options.statistics = statistics;
+  Open(bdb_options, options);
+
+  ASSERT_EQ(0, blob_db_impl()->TEST_live_sst_size());
+  std::string small_value(50, 'v');
+  std::map<std::string, std::string> data;
+  // Insert some data into LSM tree to make sure FIFO eviction take SST
+  // file size into account.
+  for (int i = 0; i < 1000; i++) {
+    ASSERT_OK(Put("key" + std::to_string(i), small_value, &data));
+  }
+  ASSERT_OK(blob_db_->Flush(FlushOptions()));
+  uint64_t live_sst_size = 0;
+  ASSERT_TRUE(blob_db_->GetIntProperty(DB::Properties::kTotalSstFilesSize,
+                                       &live_sst_size));
+  ASSERT_TRUE(live_sst_size > 0);
+  ASSERT_EQ(live_sst_size, blob_db_impl()->TEST_live_sst_size());
+
+  bdb_options.max_db_size = live_sst_size + 2000;
+  Reopen(bdb_options, options);
+  ASSERT_EQ(live_sst_size, blob_db_impl()->TEST_live_sst_size());
+
+  std::string value_1k(1000, 'v');
+  ASSERT_OK(PutWithTTL("large_key1", value_1k, 60, &data));
+  ASSERT_EQ(0, statistics->getTickerCount(BLOB_DB_FIFO_NUM_FILES_EVICTED));
+  VerifyDB(data);
+  // large_key2 evicts large_key1
+  ASSERT_OK(PutWithTTL("large_key2", value_1k, 60, &data));
+  ASSERT_EQ(1, statistics->getTickerCount(BLOB_DB_FIFO_NUM_FILES_EVICTED));
+  blob_db_impl()->TEST_DeleteObsoleteFiles();
+  data.erase("large_key1");
+  VerifyDB(data);
+  // large_key3 get no enough space even after evicting large_key2, so it
+  // instead return no space error.
+  std::string value_2k(2000, 'v');
+  ASSERT_TRUE(PutWithTTL("large_key3", value_2k, 60).IsNoSpace());
+  ASSERT_EQ(1, statistics->getTickerCount(BLOB_DB_FIFO_NUM_FILES_EVICTED));
+  // Verify large_key2 still exists.
+  VerifyDB(data);
+}
+
+// Test flush or compaction will trigger FIFO eviction since they update
+// total SST file size.
+TEST_F(BlobDBTest, FIFOEviction_TriggerOnSSTSizeChange) {
+  BlobDBOptions bdb_options;
+  bdb_options.max_db_size = 1000;
+  bdb_options.is_fifo = true;
+  bdb_options.min_blob_size = 100;
+  bdb_options.disable_background_tasks = true;
+  Options options;
+  // Use mock env to stop wall clock.
+  options.env = mock_env_.get();
+  auto statistics = CreateDBStatistics();
+  options.statistics = statistics;
+  options.compression = kNoCompression;
+  Open(bdb_options, options);
+
+  std::string value(800, 'v');
+  ASSERT_OK(PutWithTTL("large_key", value, 60));
+  ASSERT_EQ(1, blob_db_impl()->TEST_GetBlobFiles().size());
+  ASSERT_EQ(0, statistics->getTickerCount(BLOB_DB_FIFO_NUM_FILES_EVICTED));
+  VerifyDB({{"large_key", value}});
+
+  // Insert some small keys and flush to bring DB out of space.
+  std::map<std::string, std::string> data;
+  for (int i = 0; i < 10; i++) {
+    ASSERT_OK(Put("key" + std::to_string(i), "v", &data));
+  }
+  ASSERT_OK(blob_db_->Flush(FlushOptions()));
+
+  // Verify large_key is deleted by FIFO eviction.
+  blob_db_impl()->TEST_DeleteObsoleteFiles();
+  ASSERT_EQ(0, blob_db_impl()->TEST_GetBlobFiles().size());
+  ASSERT_EQ(1, statistics->getTickerCount(BLOB_DB_FIFO_NUM_FILES_EVICTED));
+  VerifyDB(data);
+}
+
+TEST_F(BlobDBTest, InlineSmallValues) {
+  constexpr uint64_t kMaxExpiration = 1000;
+  Random rnd(301);
+  BlobDBOptions bdb_options;
+  bdb_options.ttl_range_secs = kMaxExpiration;
+  bdb_options.min_blob_size = 100;
+  bdb_options.blob_file_size = 256 * 1000 * 1000;
+  bdb_options.disable_background_tasks = true;
+  Options options;
+  options.env = mock_env_.get();
+  mock_clock_->SetCurrentTime(0);
+  Open(bdb_options, options);
+  std::map<std::string, std::string> data;
+  std::map<std::string, KeyVersion> versions;
+  for (size_t i = 0; i < 1000; i++) {
+    bool is_small_value = rnd.Next() % 2;
+    bool has_ttl = rnd.Next() % 2;
+    uint64_t expiration = rnd.Next() % kMaxExpiration;
+    int len = is_small_value ? 50 : 200;
+    std::string key = "key" + std::to_string(i);
+    std::string value = rnd.HumanReadableString(len);
+    std::string blob_index;
+    data[key] = value;
+    SequenceNumber sequence = blob_db_->GetLatestSequenceNumber() + 1;
+    if (!has_ttl) {
+      ASSERT_OK(blob_db_->Put(WriteOptions(), key, value));
+    } else {
+      ASSERT_OK(blob_db_->PutUntil(WriteOptions(), key, value, expiration));
+    }
+    ASSERT_EQ(blob_db_->GetLatestSequenceNumber(), sequence);
+    versions[key] =
+        KeyVersion(key, value, sequence,
+                   (is_small_value && !has_ttl) ? kTypeValue : kTypeBlobIndex);
+  }
+  VerifyDB(data);
+  VerifyBaseDB(versions);
+  auto *bdb_impl = static_cast<BlobDBImpl *>(blob_db_);
+  auto blob_files = bdb_impl->TEST_GetBlobFiles();
+  ASSERT_EQ(2, blob_files.size());
+  std::shared_ptr<BlobFile> non_ttl_file;
+  std::shared_ptr<BlobFile> ttl_file;
+  if (blob_files[0]->HasTTL()) {
+    ttl_file = blob_files[0];
+    non_ttl_file = blob_files[1];
+  } else {
+    non_ttl_file = blob_files[0];
+    ttl_file = blob_files[1];
+  }
+  ASSERT_FALSE(non_ttl_file->HasTTL());
+  ASSERT_TRUE(ttl_file->HasTTL());
+}
+
+TEST_F(BlobDBTest, UserCompactionFilter) {
+  class CustomerFilter : public CompactionFilter {
+   public:
+    bool Filter(int /*level*/, const Slice & /*key*/, const Slice &value,
+                std::string *new_value, bool *value_changed) const override {
+      *value_changed = false;
+      // changing value size to test value transitions between inlined data
+      // and stored-in-blob data
+      if (value.size() % 4 == 1) {
+        *new_value = value.ToString();
+        // double size by duplicating value
+        *new_value += *new_value;
+        *value_changed = true;
+        return false;
+      } else if (value.size() % 3 == 1) {
+        *new_value = value.ToString();
+        // trancate value size by half
+        *new_value = new_value->substr(0, new_value->size() / 2);
+        *value_changed = true;
+        return false;
+      } else if (value.size() % 2 == 1) {
+        return true;
+      }
+      return false;
+    }
+    bool IgnoreSnapshots() const override { return true; }
+    const char *Name() const override { return "CustomerFilter"; }
+  };
+  class CustomerFilterFactory : public CompactionFilterFactory {
+    const char *Name() const override { return "CustomerFilterFactory"; }
+    std::unique_ptr<CompactionFilter> CreateCompactionFilter(
+        const CompactionFilter::Context & /*context*/) override {
+      return std::unique_ptr<CompactionFilter>(new CustomerFilter());
+    }
+  };
+
+  constexpr size_t kNumPuts = 1 << 10;
+  // Generate both inlined and blob value
+  constexpr uint64_t kMinValueSize = 1 << 6;
+  constexpr uint64_t kMaxValueSize = 1 << 8;
+  constexpr uint64_t kMinBlobSize = 1 << 7;
+  static_assert(kMinValueSize < kMinBlobSize, "");
+  static_assert(kMaxValueSize > kMinBlobSize, "");
+
+  BlobDBOptions bdb_options;
+  bdb_options.min_blob_size = kMinBlobSize;
+  bdb_options.blob_file_size = kMaxValueSize * 10;
+  bdb_options.disable_background_tasks = true;
+  if (Snappy_Supported()) {
+    bdb_options.compression = CompressionType::kSnappyCompression;
+  }
+  // case_num == 0: Test user defined compaction filter
+  // case_num == 1: Test user defined compaction filter factory
+  for (int case_num = 0; case_num < 2; case_num++) {
+    Options options;
+    if (case_num == 0) {
+      options.compaction_filter = new CustomerFilter();
+    } else {
+      options.compaction_filter_factory.reset(new CustomerFilterFactory());
+    }
+    options.disable_auto_compactions = true;
+    options.env = mock_env_.get();
+    options.statistics = CreateDBStatistics();
+    Open(bdb_options, options);
+
+    std::map<std::string, std::string> data;
+    std::map<std::string, std::string> data_after_compact;
+    Random rnd(301);
+    uint64_t value_size = kMinValueSize;
+    int drop_record = 0;
+    for (size_t i = 0; i < kNumPuts; ++i) {
+      std::ostringstream oss;
+      oss << "key" << std::setw(4) << std::setfill('0') << i;
+
+      const std::string key(oss.str());
+      const std::string value = rnd.HumanReadableString((int)value_size);
+      const SequenceNumber sequence = blob_db_->GetLatestSequenceNumber() + 1;
+
+      ASSERT_OK(Put(key, value));
+      ASSERT_EQ(blob_db_->GetLatestSequenceNumber(), sequence);
+
+      data[key] = value;
+      if (value.length() % 4 == 1) {
+        data_after_compact[key] = value + value;
+      } else if (value.length() % 3 == 1) {
+        data_after_compact[key] = value.substr(0, value.size() / 2);
+      } else if (value.length() % 2 == 1) {
+        ++drop_record;
+      } else {
+        data_after_compact[key] = value;
+      }
+
+      if (++value_size > kMaxValueSize) {
+        value_size = kMinValueSize;
+      }
+    }
+    // Verify full data set
+    VerifyDB(data);
+    // Applying compaction filter for records
+    ASSERT_OK(blob_db_->CompactRange(CompactRangeOptions(), nullptr, nullptr));
+    // Verify data after compaction, only value with even length left.
+    VerifyDB(data_after_compact);
+    ASSERT_EQ(drop_record,
+              options.statistics->getTickerCount(COMPACTION_KEY_DROP_USER));
+    delete options.compaction_filter;
+    Destroy();
+  }
+}
+
+// Test user comapction filter when there is IO error on blob data.
+TEST_F(BlobDBTest, UserCompactionFilter_BlobIOError) {
+  class CustomerFilter : public CompactionFilter {
+   public:
+    bool Filter(int /*level*/, const Slice & /*key*/, const Slice &value,
+                std::string *new_value, bool *value_changed) const override {
+      *new_value = value.ToString() + "_new";
+      *value_changed = true;
+      return false;
+    }
+    bool IgnoreSnapshots() const override { return true; }
+    const char *Name() const override { return "CustomerFilter"; }
+  };
+
+  constexpr size_t kNumPuts = 100;
+  constexpr int kValueSize = 100;
+
+  BlobDBOptions bdb_options;
+  bdb_options.min_blob_size = 0;
+  bdb_options.blob_file_size = kValueSize * 10;
+  bdb_options.disable_background_tasks = true;
+  bdb_options.compression = CompressionType::kNoCompression;
+
+  std::vector<std::string> io_failure_cases = {
+      "BlobDBImpl::CreateBlobFileAndWriter",
+      "BlobIndexCompactionFilterBase::WriteBlobToNewFile",
+      "BlobDBImpl::CloseBlobFile"};
+
+  for (size_t case_num = 0; case_num < io_failure_cases.size(); case_num++) {
+    Options options;
+    options.compaction_filter = new CustomerFilter();
+    options.disable_auto_compactions = true;
+    options.env = fault_injection_env_.get();
+    options.statistics = CreateDBStatistics();
+    Open(bdb_options, options);
+
+    std::map<std::string, std::string> data;
+    Random rnd(301);
+    for (size_t i = 0; i < kNumPuts; ++i) {
+      std::ostringstream oss;
+      oss << "key" << std::setw(4) << std::setfill('0') << i;
+
+      const std::string key(oss.str());
+      const std::string value = rnd.HumanReadableString(kValueSize);
+      const SequenceNumber sequence = blob_db_->GetLatestSequenceNumber() + 1;
+
+      ASSERT_OK(Put(key, value));
+      ASSERT_EQ(blob_db_->GetLatestSequenceNumber(), sequence);
+      data[key] = value;
+    }
+
+    // Verify full data set
+    VerifyDB(data);
+
+    SyncPoint::GetInstance()->SetCallBack(
+        io_failure_cases[case_num], [&](void * /*arg*/) {
+          fault_injection_env_->SetFilesystemActive(false, Status::IOError());
+        });
+    SyncPoint::GetInstance()->EnableProcessing();
+    auto s = blob_db_->CompactRange(CompactRangeOptions(), nullptr, nullptr);
+    ASSERT_TRUE(s.IsIOError());
+
+    // Reactivate file system to allow test to verify and close DB.
+    fault_injection_env_->SetFilesystemActive(true);
+    SyncPoint::GetInstance()->DisableProcessing();
+    SyncPoint::GetInstance()->ClearAllCallBacks();
+
+    // Verify full data set after compaction failure
+    VerifyDB(data);
+
+    delete options.compaction_filter;
+    Destroy();
+  }
+}
+
+// Test comapction filter should remove any expired blob index.
+TEST_F(BlobDBTest, FilterExpiredBlobIndex) {
+  constexpr size_t kNumKeys = 100;
+  constexpr size_t kNumPuts = 1000;
+  constexpr uint64_t kMaxExpiration = 1000;
+  constexpr uint64_t kCompactTime = 500;
+  constexpr uint64_t kMinBlobSize = 100;
+  Random rnd(301);
+  mock_clock_->SetCurrentTime(0);
+  BlobDBOptions bdb_options;
+  bdb_options.min_blob_size = kMinBlobSize;
+  bdb_options.disable_background_tasks = true;
+  Options options;
+  options.env = mock_env_.get();
+  Open(bdb_options, options);
+
+  std::map<std::string, std::string> data;
+  std::map<std::string, std::string> data_after_compact;
+  for (size_t i = 0; i < kNumPuts; i++) {
+    bool is_small_value = rnd.Next() % 2;
+    bool has_ttl = rnd.Next() % 2;
+    uint64_t expiration = rnd.Next() % kMaxExpiration;
+    int len = is_small_value ? 10 : 200;
+    std::string key = "key" + std::to_string(rnd.Next() % kNumKeys);
+    std::string value = rnd.HumanReadableString(len);
+    if (!has_ttl) {
+      if (is_small_value) {
+        std::string blob_entry;
+        BlobIndex::EncodeInlinedTTL(&blob_entry, expiration, value);
+        // Fake blob index with TTL. See what it will do.
+        ASSERT_GT(kMinBlobSize, blob_entry.size());
+        value = blob_entry;
+      }
+      ASSERT_OK(Put(key, value));
+      data_after_compact[key] = value;
+    } else {
+      ASSERT_OK(PutUntil(key, value, expiration));
+      if (expiration <= kCompactTime) {
+        data_after_compact.erase(key);
+      } else {
+        data_after_compact[key] = value;
+      }
+    }
+    data[key] = value;
+  }
+  VerifyDB(data);
+
+  mock_clock_->SetCurrentTime(kCompactTime);
+  // Take a snapshot before compaction. Make sure expired blob indexes is
+  // filtered regardless of snapshot.
+  const Snapshot *snapshot = blob_db_->GetSnapshot();
+  // Issue manual compaction to trigger compaction filter.
+  ASSERT_OK(blob_db_->CompactRange(CompactRangeOptions(), nullptr, nullptr));
+  blob_db_->ReleaseSnapshot(snapshot);
+  // Verify expired blob index are filtered.
+  std::vector<KeyVersion> versions;
+  const size_t kMaxKeys = 10000;
+  ASSERT_OK(GetAllKeyVersions(blob_db_, "", "", kMaxKeys, &versions));
+  ASSERT_EQ(data_after_compact.size(), versions.size());
+  for (auto &version : versions) {
+    ASSERT_TRUE(data_after_compact.count(version.user_key) > 0);
+  }
+  VerifyDB(data_after_compact);
+}
+
+// Test compaction filter should remove any blob index where corresponding
+// blob file has been removed.
+TEST_F(BlobDBTest, FilterFileNotAvailable) {
+  BlobDBOptions bdb_options;
+  bdb_options.min_blob_size = 0;
+  bdb_options.disable_background_tasks = true;
+  Options options;
+  options.disable_auto_compactions = true;
+  Open(bdb_options, options);
+
+  ASSERT_OK(Put("foo", "v1"));
+  auto blob_files = blob_db_impl()->TEST_GetBlobFiles();
+  ASSERT_EQ(1, blob_files.size());
+  ASSERT_EQ(1, blob_files[0]->BlobFileNumber());
+  ASSERT_OK(blob_db_impl()->TEST_CloseBlobFile(blob_files[0]));
+
+  ASSERT_OK(Put("bar", "v2"));
+  blob_files = blob_db_impl()->TEST_GetBlobFiles();
+  ASSERT_EQ(2, blob_files.size());
+  ASSERT_EQ(2, blob_files[1]->BlobFileNumber());
+  ASSERT_OK(blob_db_impl()->TEST_CloseBlobFile(blob_files[1]));
+
+  const size_t kMaxKeys = 10000;
+
+  DB *base_db = blob_db_->GetRootDB();
+  std::vector<KeyVersion> versions;
+  ASSERT_OK(GetAllKeyVersions(base_db, "", "", kMaxKeys, &versions));
+  ASSERT_EQ(2, versions.size());
+  ASSERT_EQ("bar", versions[0].user_key);
+  ASSERT_EQ("foo", versions[1].user_key);
+  VerifyDB({{"bar", "v2"}, {"foo", "v1"}});
+
+  ASSERT_OK(blob_db_->CompactRange(CompactRangeOptions(), nullptr, nullptr));
+  ASSERT_OK(GetAllKeyVersions(base_db, "", "", kMaxKeys, &versions));
+  ASSERT_EQ(2, versions.size());
+  ASSERT_EQ("bar", versions[0].user_key);
+  ASSERT_EQ("foo", versions[1].user_key);
+  VerifyDB({{"bar", "v2"}, {"foo", "v1"}});
+
+  // Remove the first blob file and compact. foo should be remove from base db.
+  blob_db_impl()->TEST_ObsoleteBlobFile(blob_files[0]);
+  blob_db_impl()->TEST_DeleteObsoleteFiles();
+  ASSERT_OK(blob_db_->CompactRange(CompactRangeOptions(), nullptr, nullptr));
+  ASSERT_OK(GetAllKeyVersions(base_db, "", "", kMaxKeys, &versions));
+  ASSERT_EQ(1, versions.size());
+  ASSERT_EQ("bar", versions[0].user_key);
+  VerifyDB({{"bar", "v2"}});
+
+  // Remove the second blob file and compact. bar should be remove from base db.
+  blob_db_impl()->TEST_ObsoleteBlobFile(blob_files[1]);
+  blob_db_impl()->TEST_DeleteObsoleteFiles();
+  ASSERT_OK(blob_db_->CompactRange(CompactRangeOptions(), nullptr, nullptr));
+  ASSERT_OK(GetAllKeyVersions(base_db, "", "", kMaxKeys, &versions));
+  ASSERT_EQ(0, versions.size());
+  VerifyDB({});
+}
+
+// Test compaction filter should filter any inlined TTL keys that would have
+// been dropped by last FIFO eviction if they are store out-of-line.
+TEST_F(BlobDBTest, FilterForFIFOEviction) {
+  Random rnd(215);
+  BlobDBOptions bdb_options;
+  bdb_options.min_blob_size = 100;
+  bdb_options.ttl_range_secs = 60;
+  bdb_options.max_db_size = 0;
+  bdb_options.disable_background_tasks = true;
+  Options options;
+  // Use mock env to stop wall clock.
+  mock_clock_->SetCurrentTime(0);
+  options.env = mock_env_.get();
+  auto statistics = CreateDBStatistics();
+  options.statistics = statistics;
+  options.disable_auto_compactions = true;
+  Open(bdb_options, options);
+
+  std::map<std::string, std::string> data;
+  std::map<std::string, std::string> data_after_compact;
+  // Insert some small values that will be inlined.
+  for (int i = 0; i < 1000; i++) {
+    std::string key = "key" + std::to_string(i);
+    std::string value = rnd.HumanReadableString(50);
+    uint64_t ttl = rnd.Next() % 120 + 1;
+    ASSERT_OK(PutWithTTL(key, value, ttl, &data));
+    if (ttl >= 60) {
+      data_after_compact[key] = value;
+    }
+  }
+  uint64_t num_keys_to_evict = data.size() - data_after_compact.size();
+  ASSERT_OK(blob_db_->Flush(FlushOptions()));
+  uint64_t live_sst_size = blob_db_impl()->TEST_live_sst_size();
+  ASSERT_GT(live_sst_size, 0);
+  VerifyDB(data);
+
+  bdb_options.max_db_size = live_sst_size + 30000;
+  bdb_options.is_fifo = true;
+  Reopen(bdb_options, options);
+  VerifyDB(data);
+
+  // Put two large values, each on a different blob file.
+  std::string large_value(10000, 'v');
+  ASSERT_OK(PutWithTTL("large_key1", large_value, 90));
+  ASSERT_OK(PutWithTTL("large_key2", large_value, 150));
+  ASSERT_EQ(2, blob_db_impl()->TEST_GetBlobFiles().size());
+  ASSERT_EQ(0, statistics->getTickerCount(BLOB_DB_FIFO_NUM_FILES_EVICTED));
+  data["large_key1"] = large_value;
+  data["large_key2"] = large_value;
+  VerifyDB(data);
+
+  // Put a third large value which will bring the DB out of space.
+  // FIFO eviction will evict the file of large_key1.
+  ASSERT_OK(PutWithTTL("large_key3", large_value, 150));
+  ASSERT_EQ(1, statistics->getTickerCount(BLOB_DB_FIFO_NUM_FILES_EVICTED));
+  ASSERT_EQ(2, blob_db_impl()->TEST_GetBlobFiles().size());
+  blob_db_impl()->TEST_DeleteObsoleteFiles();
+  ASSERT_EQ(1, blob_db_impl()->TEST_GetBlobFiles().size());
+  data.erase("large_key1");
+  data["large_key3"] = large_value;
+  VerifyDB(data);
+
+  // Putting some more small values. These values shouldn't be evicted by
+  // compaction filter since they are inserted after FIFO eviction.
+  ASSERT_OK(PutWithTTL("foo", "v", 30, &data_after_compact));
+  ASSERT_OK(PutWithTTL("bar", "v", 30, &data_after_compact));
+
+  // FIFO eviction doesn't trigger again since there enough room for the flush.
+  ASSERT_OK(blob_db_->Flush(FlushOptions()));
+  ASSERT_EQ(1, statistics->getTickerCount(BLOB_DB_FIFO_NUM_FILES_EVICTED));
+
+  // Manual compact and check if compaction filter evict those keys with
+  // expiration < 60.
+  ASSERT_OK(blob_db_->CompactRange(CompactRangeOptions(), nullptr, nullptr));
+  // All keys with expiration < 60, plus large_key1 is filtered by
+  // compaction filter.
+  ASSERT_EQ(num_keys_to_evict + 1,
+            statistics->getTickerCount(BLOB_DB_BLOB_INDEX_EVICTED_COUNT));
+  ASSERT_EQ(1, statistics->getTickerCount(BLOB_DB_FIFO_NUM_FILES_EVICTED));
+  ASSERT_EQ(1, blob_db_impl()->TEST_GetBlobFiles().size());
+  data_after_compact["large_key2"] = large_value;
+  data_after_compact["large_key3"] = large_value;
+  VerifyDB(data_after_compact);
+}
+
+TEST_F(BlobDBTest, GarbageCollection) {
+  constexpr size_t kNumPuts = 1 << 10;
+
+  constexpr uint64_t kExpiration = 1000;
+  constexpr uint64_t kCompactTime = 500;
+
+  constexpr uint64_t kKeySize = 7;  // "key" + 4 digits
+
+  constexpr uint64_t kSmallValueSize = 1 << 6;
+  constexpr uint64_t kLargeValueSize = 1 << 8;
+  constexpr uint64_t kMinBlobSize = 1 << 7;
+  static_assert(kSmallValueSize < kMinBlobSize, "");
+  static_assert(kLargeValueSize > kMinBlobSize, "");
+
+  constexpr size_t kBlobsPerFile = 8;
+  constexpr size_t kNumBlobFiles = kNumPuts / kBlobsPerFile;
+  constexpr uint64_t kBlobFileSize =
+      BlobLogHeader::kSize +
+      (BlobLogRecord::kHeaderSize + kKeySize + kLargeValueSize) * kBlobsPerFile;
+
+  BlobDBOptions bdb_options;
+  bdb_options.min_blob_size = kMinBlobSize;
+  bdb_options.blob_file_size = kBlobFileSize;
+  bdb_options.enable_garbage_collection = true;
+  bdb_options.garbage_collection_cutoff = 0.25;
+  bdb_options.disable_background_tasks = true;
+
+  Options options;
+  options.env = mock_env_.get();
+  options.statistics = CreateDBStatistics();
+
+  Open(bdb_options, options);
+
+  std::map<std::string, std::string> data;
+  std::map<std::string, KeyVersion> blob_value_versions;
+  std::map<std::string, BlobIndexVersion> blob_index_versions;
+
+  Random rnd(301);
+
+  // Add a bunch of large non-TTL values. These will be written to non-TTL
+  // blob files and will be subject to GC.
+  for (size_t i = 0; i < kNumPuts; ++i) {
+    std::ostringstream oss;
+    oss << "key" << std::setw(4) << std::setfill('0') << i;
+
+    const std::string key(oss.str());
+    const std::string value = rnd.HumanReadableString(kLargeValueSize);
+    const SequenceNumber sequence = blob_db_->GetLatestSequenceNumber() + 1;
+
+    ASSERT_OK(Put(key, value));
+    ASSERT_EQ(blob_db_->GetLatestSequenceNumber(), sequence);
+
+    data[key] = value;
+    blob_value_versions[key] = KeyVersion(key, value, sequence, kTypeBlobIndex);
+    blob_index_versions[key] =
+        BlobIndexVersion(key, /* file_number */ (i >> 3) + 1, kNoExpiration,
+                         sequence, kTypeBlobIndex);
+  }
+
+  // Add some small and/or TTL values that will be ignored during GC.
+  // First, add a large TTL value will be written to its own TTL blob file.
+  {
+    const std::string key("key2000");
+    const std::string value = rnd.HumanReadableString(kLargeValueSize);
+    const SequenceNumber sequence = blob_db_->GetLatestSequenceNumber() + 1;
+
+    ASSERT_OK(PutUntil(key, value, kExpiration));
+    ASSERT_EQ(blob_db_->GetLatestSequenceNumber(), sequence);
+
+    data[key] = value;
+    blob_value_versions[key] = KeyVersion(key, value, sequence, kTypeBlobIndex);
+    blob_index_versions[key] =
+        BlobIndexVersion(key, /* file_number */ kNumBlobFiles + 1, kExpiration,
+                         sequence, kTypeBlobIndex);
+  }
+
+  // Now add a small TTL value (which will be inlined).
+  {
+    const std::string key("key3000");
+    const std::string value = rnd.HumanReadableString(kSmallValueSize);
+    const SequenceNumber sequence = blob_db_->GetLatestSequenceNumber() + 1;
+
+    ASSERT_OK(PutUntil(key, value, kExpiration));
+    ASSERT_EQ(blob_db_->GetLatestSequenceNumber(), sequence);
+
+    data[key] = value;
+    blob_value_versions[key] = KeyVersion(key, value, sequence, kTypeBlobIndex);
+    blob_index_versions[key] = BlobIndexVersion(
+        key, kInvalidBlobFileNumber, kExpiration, sequence, kTypeBlobIndex);
+  }
+
+  // Finally, add a small non-TTL value (which will be stored as a regular
+  // value).
+  {
+    const std::string key("key4000");
+    const std::string value = rnd.HumanReadableString(kSmallValueSize);
+    const SequenceNumber sequence = blob_db_->GetLatestSequenceNumber() + 1;
+
+    ASSERT_OK(Put(key, value));
+    ASSERT_EQ(blob_db_->GetLatestSequenceNumber(), sequence);
+
+    data[key] = value;
+    blob_value_versions[key] = KeyVersion(key, value, sequence, kTypeValue);
+    blob_index_versions[key] = BlobIndexVersion(
+        key, kInvalidBlobFileNumber, kNoExpiration, sequence, kTypeValue);
+  }
+
+  VerifyDB(data);
+  VerifyBaseDB(blob_value_versions);
+  VerifyBaseDBBlobIndex(blob_index_versions);
+
+  // At this point, we should have 128 immutable non-TTL files with file numbers
+  // 1..128.
+  {
+    auto live_imm_files = blob_db_impl()->TEST_GetLiveImmNonTTLFiles();
+    ASSERT_EQ(live_imm_files.size(), kNumBlobFiles);
+    for (size_t i = 0; i < kNumBlobFiles; ++i) {
+      ASSERT_EQ(live_imm_files[i]->BlobFileNumber(), i + 1);
+      ASSERT_EQ(live_imm_files[i]->GetFileSize(),
+                kBlobFileSize + BlobLogFooter::kSize);
+    }
+  }
+
+  mock_clock_->SetCurrentTime(kCompactTime);
+
+  ASSERT_OK(blob_db_->CompactRange(CompactRangeOptions(), nullptr, nullptr));
+
+  // We expect the data to remain the same and the blobs from the oldest N files
+  // to be moved to new files. Sequence numbers get zeroed out during the
+  // compaction.
+  VerifyDB(data);
+
+  for (auto &pair : blob_value_versions) {
+    KeyVersion &version = pair.second;
+    version.sequence = 0;
+  }
+
+  VerifyBaseDB(blob_value_versions);
+
+  const uint64_t cutoff = static_cast<uint64_t>(
+      bdb_options.garbage_collection_cutoff * kNumBlobFiles);
+  for (auto &pair : blob_index_versions) {
+    BlobIndexVersion &version = pair.second;
+
+    version.sequence = 0;
+
+    if (version.file_number == kInvalidBlobFileNumber) {
+      continue;
+    }
+
+    if (version.file_number > cutoff) {
+      continue;
+    }
+
+    version.file_number += kNumBlobFiles + 1;
+  }
+
+  VerifyBaseDBBlobIndex(blob_index_versions);
+
+  const Statistics *const statistics = options.statistics.get();
+  assert(statistics);
+
+  ASSERT_EQ(statistics->getTickerCount(BLOB_DB_GC_NUM_FILES), cutoff);
+  ASSERT_EQ(statistics->getTickerCount(BLOB_DB_GC_NUM_NEW_FILES), cutoff);
+  ASSERT_EQ(statistics->getTickerCount(BLOB_DB_GC_FAILURES), 0);
+  ASSERT_EQ(statistics->getTickerCount(BLOB_DB_GC_NUM_KEYS_RELOCATED),
+            cutoff * kBlobsPerFile);
+  ASSERT_EQ(statistics->getTickerCount(BLOB_DB_GC_BYTES_RELOCATED),
+            cutoff * kBlobsPerFile * kLargeValueSize);
+
+  // At this point, we should have 128 immutable non-TTL files with file numbers
+  // 33..128 and 130..161. (129 was taken by the TTL blob file.)
+  {
+    auto live_imm_files = blob_db_impl()->TEST_GetLiveImmNonTTLFiles();
+    ASSERT_EQ(live_imm_files.size(), kNumBlobFiles);
+    for (size_t i = 0; i < kNumBlobFiles; ++i) {
+      uint64_t expected_file_number = i + cutoff + 1;
+      if (expected_file_number > kNumBlobFiles) {
+        ++expected_file_number;
+      }
+
+      ASSERT_EQ(live_imm_files[i]->BlobFileNumber(), expected_file_number);
+      ASSERT_EQ(live_imm_files[i]->GetFileSize(),
+                kBlobFileSize + BlobLogFooter::kSize);
+    }
+  }
+}
+
+TEST_F(BlobDBTest, GarbageCollectionFailure) {
+  BlobDBOptions bdb_options;
+  bdb_options.min_blob_size = 0;
+  bdb_options.enable_garbage_collection = true;
+  bdb_options.garbage_collection_cutoff = 1.0;
+  bdb_options.disable_background_tasks = true;
+
+  Options db_options;
+  db_options.statistics = CreateDBStatistics();
+
+  Open(bdb_options, db_options);
+
+  // Write a couple of valid blobs.
+  ASSERT_OK(Put("foo", "bar"));
+  ASSERT_OK(Put("dead", "beef"));
+
+  // Write a fake blob reference into the base DB that points to a non-existing
+  // blob file.
+  std::string blob_index;
+  BlobIndex::EncodeBlob(&blob_index, /* file_number */ 1000, /* offset */ 1234,
+                        /* size */ 5678, kNoCompression);
+
+  WriteBatch batch;
+  ASSERT_OK(WriteBatchInternal::PutBlobIndex(
+      &batch, blob_db_->DefaultColumnFamily()->GetID(), "key", blob_index));
+  ASSERT_OK(blob_db_->GetRootDB()->Write(WriteOptions(), &batch));
+
+  auto blob_files = blob_db_impl()->TEST_GetBlobFiles();
+  ASSERT_EQ(blob_files.size(), 1);
+  auto blob_file = blob_files[0];
+  ASSERT_OK(blob_db_impl()->TEST_CloseBlobFile(blob_file));
+
+  ASSERT_TRUE(blob_db_->CompactRange(CompactRangeOptions(), nullptr, nullptr)
+                  .IsIOError());
+
+  const Statistics *const statistics = db_options.statistics.get();
+  assert(statistics);
+
+  ASSERT_EQ(statistics->getTickerCount(BLOB_DB_GC_NUM_FILES), 0);
+  ASSERT_EQ(statistics->getTickerCount(BLOB_DB_GC_NUM_NEW_FILES), 1);
+  ASSERT_EQ(statistics->getTickerCount(BLOB_DB_GC_FAILURES), 1);
+  ASSERT_EQ(statistics->getTickerCount(BLOB_DB_GC_NUM_KEYS_RELOCATED), 2);
+  ASSERT_EQ(statistics->getTickerCount(BLOB_DB_GC_BYTES_RELOCATED), 7);
+}
+
+// File should be evicted after expiration.
+TEST_F(BlobDBTest, EvictExpiredFile) {
+  BlobDBOptions bdb_options;
+  bdb_options.ttl_range_secs = 100;
+  bdb_options.min_blob_size = 0;
+  bdb_options.disable_background_tasks = true;
+  Options options;
+  options.env = mock_env_.get();
+  Open(bdb_options, options);
+  mock_clock_->SetCurrentTime(50);
+  std::map<std::string, std::string> data;
+  ASSERT_OK(PutWithTTL("foo", "bar", 100, &data));
+  auto blob_files = blob_db_impl()->TEST_GetBlobFiles();
+  ASSERT_EQ(1, blob_files.size());
+  auto blob_file = blob_files[0];
+  ASSERT_FALSE(blob_file->Immutable());
+  ASSERT_FALSE(blob_file->Obsolete());
+  VerifyDB(data);
+  mock_clock_->SetCurrentTime(250);
+  // The key should expired now.
+  blob_db_impl()->TEST_EvictExpiredFiles();
+  ASSERT_EQ(1, blob_db_impl()->TEST_GetBlobFiles().size());
+  ASSERT_EQ(1, blob_db_impl()->TEST_GetObsoleteFiles().size());
+  ASSERT_TRUE(blob_file->Immutable());
+  ASSERT_TRUE(blob_file->Obsolete());
+  blob_db_impl()->TEST_DeleteObsoleteFiles();
+  ASSERT_EQ(0, blob_db_impl()->TEST_GetBlobFiles().size());
+  ASSERT_EQ(0, blob_db_impl()->TEST_GetObsoleteFiles().size());
+  // Make sure we don't return garbage value after blob file being evicted,
+  // but the blob index still exists in the LSM tree.
+  std::string val = "";
+  ASSERT_TRUE(blob_db_->Get(ReadOptions(), "foo", &val).IsNotFound());
+  ASSERT_EQ("", val);
+}
+
+TEST_F(BlobDBTest, DisableFileDeletions) {
+  BlobDBOptions bdb_options;
+  bdb_options.disable_background_tasks = true;
+  Open(bdb_options);
+  std::map<std::string, std::string> data;
+  for (bool force : {true, false}) {
+    ASSERT_OK(Put("foo", "v", &data));
+    auto blob_files = blob_db_impl()->TEST_GetBlobFiles();
+    ASSERT_EQ(1, blob_files.size());
+    auto blob_file = blob_files[0];
+    ASSERT_OK(blob_db_impl()->TEST_CloseBlobFile(blob_file));
+    blob_db_impl()->TEST_ObsoleteBlobFile(blob_file);
+    ASSERT_EQ(1, blob_db_impl()->TEST_GetBlobFiles().size());
+    ASSERT_EQ(1, blob_db_impl()->TEST_GetObsoleteFiles().size());
+    // Call DisableFileDeletions twice.
+    ASSERT_OK(blob_db_->DisableFileDeletions());
+    ASSERT_OK(blob_db_->DisableFileDeletions());
+    // File deletions should be disabled.
+    blob_db_impl()->TEST_DeleteObsoleteFiles();
+    ASSERT_EQ(1, blob_db_impl()->TEST_GetBlobFiles().size());
+    ASSERT_EQ(1, blob_db_impl()->TEST_GetObsoleteFiles().size());
+    VerifyDB(data);
+    // Enable file deletions once. If force=true, file deletion is enabled.
+    // Otherwise it needs to enable it for a second time.
+    ASSERT_OK(blob_db_->EnableFileDeletions(force));
+    blob_db_impl()->TEST_DeleteObsoleteFiles();
+    if (!force) {
+      ASSERT_EQ(1, blob_db_impl()->TEST_GetBlobFiles().size());
+      ASSERT_EQ(1, blob_db_impl()->TEST_GetObsoleteFiles().size());
+      VerifyDB(data);
+      // Call EnableFileDeletions a second time.
+      ASSERT_OK(blob_db_->EnableFileDeletions(false));
+      blob_db_impl()->TEST_DeleteObsoleteFiles();
+    }
+    // Regardless of value of `force`, file should be deleted by now.
+    ASSERT_EQ(0, blob_db_impl()->TEST_GetBlobFiles().size());
+    ASSERT_EQ(0, blob_db_impl()->TEST_GetObsoleteFiles().size());
+    VerifyDB({});
+  }
+}
+
+TEST_F(BlobDBTest, MaintainBlobFileToSstMapping) {
+  BlobDBOptions bdb_options;
+  bdb_options.enable_garbage_collection = true;
+  bdb_options.disable_background_tasks = true;
+  Open(bdb_options);
+
+  // Register some dummy blob files.
+  blob_db_impl()->TEST_AddDummyBlobFile(1, /* immutable_sequence */ 200);
+  blob_db_impl()->TEST_AddDummyBlobFile(2, /* immutable_sequence */ 300);
+  blob_db_impl()->TEST_AddDummyBlobFile(3, /* immutable_sequence */ 400);
+  blob_db_impl()->TEST_AddDummyBlobFile(4, /* immutable_sequence */ 500);
+  blob_db_impl()->TEST_AddDummyBlobFile(5, /* immutable_sequence */ 600);
+
+  // Initialize the blob <-> SST file mapping. First, add some SST files with
+  // blob file references, then some without.
+  std::vector<LiveFileMetaData> live_files;
+
+  for (uint64_t i = 1; i <= 10; ++i) {
+    LiveFileMetaData live_file;
+    live_file.file_number = i;
+    live_file.oldest_blob_file_number = ((i - 1) % 5) + 1;
+
+    live_files.emplace_back(live_file);
+  }
+
+  for (uint64_t i = 11; i <= 20; ++i) {
+    LiveFileMetaData live_file;
+    live_file.file_number = i;
+
+    live_files.emplace_back(live_file);
+  }
+
+  blob_db_impl()->TEST_InitializeBlobFileToSstMapping(live_files);
+
+  // Check that the blob <-> SST mappings have been correctly initialized.
+  auto blob_files = blob_db_impl()->TEST_GetBlobFiles();
+
+  ASSERT_EQ(blob_files.size(), 5);
+
+  {
+    auto live_imm_files = blob_db_impl()->TEST_GetLiveImmNonTTLFiles();
+    ASSERT_EQ(live_imm_files.size(), 5);
+    for (size_t i = 0; i < 5; ++i) {
+      ASSERT_EQ(live_imm_files[i]->BlobFileNumber(), i + 1);
+    }
+
+    ASSERT_TRUE(blob_db_impl()->TEST_GetObsoleteFiles().empty());
+  }
+
+  {
+    const std::vector<std::unordered_set<uint64_t>> expected_sst_files{
+        {1, 6}, {2, 7}, {3, 8}, {4, 9}, {5, 10}};
+    const std::vector<bool> expected_obsolete{false, false, false, false,
+                                              false};
+    for (size_t i = 0; i < 5; ++i) {
+      const auto &blob_file = blob_files[i];
+      ASSERT_EQ(blob_file->GetLinkedSstFiles(), expected_sst_files[i]);
+      ASSERT_EQ(blob_file->Obsolete(), expected_obsolete[i]);
+    }
+
+    auto live_imm_files = blob_db_impl()->TEST_GetLiveImmNonTTLFiles();
+    ASSERT_EQ(live_imm_files.size(), 5);
+    for (size_t i = 0; i < 5; ++i) {
+      ASSERT_EQ(live_imm_files[i]->BlobFileNumber(), i + 1);
+    }
+
+    ASSERT_TRUE(blob_db_impl()->TEST_GetObsoleteFiles().empty());
+  }
+
+  // Simulate a flush where the SST does not reference any blob files.
+  {
+    FlushJobInfo info{};
+    info.file_number = 21;
+    info.smallest_seqno = 1;
+    info.largest_seqno = 100;
+
+    blob_db_impl()->TEST_ProcessFlushJobInfo(info);
+
+    const std::vector<std::unordered_set<uint64_t>> expected_sst_files{
+        {1, 6}, {2, 7}, {3, 8}, {4, 9}, {5, 10}};
+    const std::vector<bool> expected_obsolete{false, false, false, false,
+                                              false};
+    for (size_t i = 0; i < 5; ++i) {
+      const auto &blob_file = blob_files[i];
+      ASSERT_EQ(blob_file->GetLinkedSstFiles(), expected_sst_files[i]);
+      ASSERT_EQ(blob_file->Obsolete(), expected_obsolete[i]);
+    }
+
+    auto live_imm_files = blob_db_impl()->TEST_GetLiveImmNonTTLFiles();
+    ASSERT_EQ(live_imm_files.size(), 5);
+    for (size_t i = 0; i < 5; ++i) {
+      ASSERT_EQ(live_imm_files[i]->BlobFileNumber(), i + 1);
+    }
+
+    ASSERT_TRUE(blob_db_impl()->TEST_GetObsoleteFiles().empty());
+  }
+
+  // Simulate a flush where the SST references a blob file.
+  {
+    FlushJobInfo info{};
+    info.file_number = 22;
+    info.oldest_blob_file_number = 5;
+    info.smallest_seqno = 101;
+    info.largest_seqno = 200;
+
+    blob_db_impl()->TEST_ProcessFlushJobInfo(info);
+
+    const std::vector<std::unordered_set<uint64_t>> expected_sst_files{
+        {1, 6}, {2, 7}, {3, 8}, {4, 9}, {5, 10, 22}};
+    const std::vector<bool> expected_obsolete{false, false, false, false,
+                                              false};
+    for (size_t i = 0; i < 5; ++i) {
+      const auto &blob_file = blob_files[i];
+      ASSERT_EQ(blob_file->GetLinkedSstFiles(), expected_sst_files[i]);
+      ASSERT_EQ(blob_file->Obsolete(), expected_obsolete[i]);
+    }
+
+    auto live_imm_files = blob_db_impl()->TEST_GetLiveImmNonTTLFiles();
+    ASSERT_EQ(live_imm_files.size(), 5);
+    for (size_t i = 0; i < 5; ++i) {
+      ASSERT_EQ(live_imm_files[i]->BlobFileNumber(), i + 1);
+    }
+
+    ASSERT_TRUE(blob_db_impl()->TEST_GetObsoleteFiles().empty());
+  }
+
+  // Simulate a compaction. Some inputs and outputs have blob file references,
+  // some don't. There is also a trivial move (which means the SST appears on
+  // both the input and the output list). Blob file 1 loses all its linked SSTs,
+  // and since it got marked immutable at sequence number 200 which has already
+  // been flushed, it can be marked obsolete.
+  {
+    CompactionJobInfo info{};
+    info.input_file_infos.emplace_back(CompactionFileInfo{1, 1, 1});
+    info.input_file_infos.emplace_back(CompactionFileInfo{1, 2, 2});
+    info.input_file_infos.emplace_back(CompactionFileInfo{1, 6, 1});
+    info.input_file_infos.emplace_back(
+        CompactionFileInfo{1, 11, kInvalidBlobFileNumber});
+    info.input_file_infos.emplace_back(CompactionFileInfo{1, 22, 5});
+    info.output_file_infos.emplace_back(CompactionFileInfo{2, 22, 5});
+    info.output_file_infos.emplace_back(CompactionFileInfo{2, 23, 3});
+    info.output_file_infos.emplace_back(
+        CompactionFileInfo{2, 24, kInvalidBlobFileNumber});
+
+    blob_db_impl()->TEST_ProcessCompactionJobInfo(info);
+
+    const std::vector<std::unordered_set<uint64_t>> expected_sst_files{
+        {}, {7}, {3, 8, 23}, {4, 9}, {5, 10, 22}};
+    const std::vector<bool> expected_obsolete{true, false, false, false, false};
+    for (size_t i = 0; i < 5; ++i) {
+      const auto &blob_file = blob_files[i];
+      ASSERT_EQ(blob_file->GetLinkedSstFiles(), expected_sst_files[i]);
+      ASSERT_EQ(blob_file->Obsolete(), expected_obsolete[i]);
+    }
+
+    auto live_imm_files = blob_db_impl()->TEST_GetLiveImmNonTTLFiles();
+    ASSERT_EQ(live_imm_files.size(), 4);
+    for (size_t i = 0; i < 4; ++i) {
+      ASSERT_EQ(live_imm_files[i]->BlobFileNumber(), i + 2);
+    }
+
+    auto obsolete_files = blob_db_impl()->TEST_GetObsoleteFiles();
+    ASSERT_EQ(obsolete_files.size(), 1);
+    ASSERT_EQ(obsolete_files[0]->BlobFileNumber(), 1);
+  }
+
+  // Simulate a failed compaction. No mappings should be updated.
+  {
+    CompactionJobInfo info{};
+    info.input_file_infos.emplace_back(CompactionFileInfo{1, 7, 2});
+    info.input_file_infos.emplace_back(CompactionFileInfo{2, 22, 5});
+    info.output_file_infos.emplace_back(CompactionFileInfo{2, 25, 3});
+    info.status = Status::Corruption();
+
+    blob_db_impl()->TEST_ProcessCompactionJobInfo(info);
+
+    const std::vector<std::unordered_set<uint64_t>> expected_sst_files{
+        {}, {7}, {3, 8, 23}, {4, 9}, {5, 10, 22}};
+    const std::vector<bool> expected_obsolete{true, false, false, false, false};
+    for (size_t i = 0; i < 5; ++i) {
+      const auto &blob_file = blob_files[i];
+      ASSERT_EQ(blob_file->GetLinkedSstFiles(), expected_sst_files[i]);
+      ASSERT_EQ(blob_file->Obsolete(), expected_obsolete[i]);
+    }
+
+    auto live_imm_files = blob_db_impl()->TEST_GetLiveImmNonTTLFiles();
+    ASSERT_EQ(live_imm_files.size(), 4);
+    for (size_t i = 0; i < 4; ++i) {
+      ASSERT_EQ(live_imm_files[i]->BlobFileNumber(), i + 2);
+    }
+
+    auto obsolete_files = blob_db_impl()->TEST_GetObsoleteFiles();
+    ASSERT_EQ(obsolete_files.size(), 1);
+    ASSERT_EQ(obsolete_files[0]->BlobFileNumber(), 1);
+  }
+
+  // Simulate another compaction. Blob file 2 loses all its linked SSTs
+  // but since it got marked immutable at sequence number 300 which hasn't
+  // been flushed yet, it cannot be marked obsolete at this point.
+  {
+    CompactionJobInfo info{};
+    info.input_file_infos.emplace_back(CompactionFileInfo{1, 7, 2});
+    info.input_file_infos.emplace_back(CompactionFileInfo{2, 22, 5});
+    info.output_file_infos.emplace_back(CompactionFileInfo{2, 25, 3});
+
+    blob_db_impl()->TEST_ProcessCompactionJobInfo(info);
+
+    const std::vector<std::unordered_set<uint64_t>> expected_sst_files{
+        {}, {}, {3, 8, 23, 25}, {4, 9}, {5, 10}};
+    const std::vector<bool> expected_obsolete{true, false, false, false, false};
+    for (size_t i = 0; i < 5; ++i) {
+      const auto &blob_file = blob_files[i];
+      ASSERT_EQ(blob_file->GetLinkedSstFiles(), expected_sst_files[i]);
+      ASSERT_EQ(blob_file->Obsolete(), expected_obsolete[i]);
+    }
+
+    auto live_imm_files = blob_db_impl()->TEST_GetLiveImmNonTTLFiles();
+    ASSERT_EQ(live_imm_files.size(), 4);
+    for (size_t i = 0; i < 4; ++i) {
+      ASSERT_EQ(live_imm_files[i]->BlobFileNumber(), i + 2);
+    }
+
+    auto obsolete_files = blob_db_impl()->TEST_GetObsoleteFiles();
+    ASSERT_EQ(obsolete_files.size(), 1);
+    ASSERT_EQ(obsolete_files[0]->BlobFileNumber(), 1);
+  }
+
+  // Simulate a flush with largest sequence number 300. This will make it
+  // possible to mark blob file 2 obsolete.
+  {
+    FlushJobInfo info{};
+    info.file_number = 26;
+    info.smallest_seqno = 201;
+    info.largest_seqno = 300;
+
+    blob_db_impl()->TEST_ProcessFlushJobInfo(info);
+
+    const std::vector<std::unordered_set<uint64_t>> expected_sst_files{
+        {}, {}, {3, 8, 23, 25}, {4, 9}, {5, 10}};
+    const std::vector<bool> expected_obsolete{true, true, false, false, false};
+    for (size_t i = 0; i < 5; ++i) {
+      const auto &blob_file = blob_files[i];
+      ASSERT_EQ(blob_file->GetLinkedSstFiles(), expected_sst_files[i]);
+      ASSERT_EQ(blob_file->Obsolete(), expected_obsolete[i]);
+    }
+
+    auto live_imm_files = blob_db_impl()->TEST_GetLiveImmNonTTLFiles();
+    ASSERT_EQ(live_imm_files.size(), 3);
+    for (size_t i = 0; i < 3; ++i) {
+      ASSERT_EQ(live_imm_files[i]->BlobFileNumber(), i + 3);
+    }
+
+    auto obsolete_files = blob_db_impl()->TEST_GetObsoleteFiles();
+    ASSERT_EQ(obsolete_files.size(), 2);
+    ASSERT_EQ(obsolete_files[0]->BlobFileNumber(), 1);
+    ASSERT_EQ(obsolete_files[1]->BlobFileNumber(), 2);
+  }
+}
+
+TEST_F(BlobDBTest, ShutdownWait) {
+  BlobDBOptions bdb_options;
+  bdb_options.ttl_range_secs = 100;
+  bdb_options.min_blob_size = 0;
+  bdb_options.disable_background_tasks = false;
+  Options options;
+  options.env = mock_env_.get();
+
+  SyncPoint::GetInstance()->LoadDependency({
+      {"BlobDBImpl::EvictExpiredFiles:0", "BlobDBTest.ShutdownWait:0"},
+      {"BlobDBTest.ShutdownWait:1", "BlobDBImpl::EvictExpiredFiles:1"},
+      {"BlobDBImpl::EvictExpiredFiles:2", "BlobDBTest.ShutdownWait:2"},
+      {"BlobDBTest.ShutdownWait:3", "BlobDBImpl::EvictExpiredFiles:3"},
+  });
+  // Force all tasks to be scheduled immediately.
+  SyncPoint::GetInstance()->SetCallBack(
+      "TimeQueue::Add:item.end", [&](void *arg) {
+        std::chrono::steady_clock::time_point *tp =
+            static_cast<std::chrono::steady_clock::time_point *>(arg);
+        *tp =
+            std::chrono::steady_clock::now() - std::chrono::milliseconds(10000);
+      });
+
+  SyncPoint::GetInstance()->SetCallBack(
+      "BlobDBImpl::EvictExpiredFiles:cb", [&](void * /*arg*/) {
+        // Sleep 3 ms to increase the chance of data race.
+        // We've synced up the code so that EvictExpiredFiles()
+        // is called concurrently with ~BlobDBImpl().
+        // ~BlobDBImpl() is supposed to wait for all background
+        // task to shutdown before doing anything else. In order
+        // to use the same test to reproduce a bug of the waiting
+        // logic, we wait a little bit here, so that TSAN can
+        // catch the data race.
+        // We should improve the test if we find a better way.
+        Env::Default()->SleepForMicroseconds(3000);
+      });
+
+  SyncPoint::GetInstance()->EnableProcessing();
+
+  Open(bdb_options, options);
+  mock_clock_->SetCurrentTime(50);
+  std::map<std::string, std::string> data;
+  ASSERT_OK(PutWithTTL("foo", "bar", 100, &data));
+  auto blob_files = blob_db_impl()->TEST_GetBlobFiles();
+  ASSERT_EQ(1, blob_files.size());
+  auto blob_file = blob_files[0];
+  ASSERT_FALSE(blob_file->Immutable());
+  ASSERT_FALSE(blob_file->Obsolete());
+  VerifyDB(data);
+
+  TEST_SYNC_POINT("BlobDBTest.ShutdownWait:0");
+  mock_clock_->SetCurrentTime(250);
+  // The key should expired now.
+  TEST_SYNC_POINT("BlobDBTest.ShutdownWait:1");
+
+  TEST_SYNC_POINT("BlobDBTest.ShutdownWait:2");
+  TEST_SYNC_POINT("BlobDBTest.ShutdownWait:3");
+  Close();
+
+  SyncPoint::GetInstance()->DisableProcessing();
+}
+
+TEST_F(BlobDBTest, SyncBlobFileBeforeClose) {
+  Options options;
+  options.statistics = CreateDBStatistics();
+
+  BlobDBOptions blob_options;
+  blob_options.min_blob_size = 0;
+  blob_options.bytes_per_sync = 1 << 20;
+  blob_options.disable_background_tasks = true;
+
+  Open(blob_options, options);
+
+  ASSERT_OK(Put("foo", "bar"));
+
+  auto blob_files = blob_db_impl()->TEST_GetBlobFiles();
+  ASSERT_EQ(blob_files.size(), 1);
+
+  ASSERT_OK(blob_db_impl()->TEST_CloseBlobFile(blob_files[0]));
+  ASSERT_EQ(options.statistics->getTickerCount(BLOB_DB_BLOB_FILE_SYNCED), 1);
+}
+
+TEST_F(BlobDBTest, SyncBlobFileBeforeCloseIOError) {
+  Options options;
+  options.env = fault_injection_env_.get();
+
+  BlobDBOptions blob_options;
+  blob_options.min_blob_size = 0;
+  blob_options.bytes_per_sync = 1 << 20;
+  blob_options.disable_background_tasks = true;
+
+  Open(blob_options, options);
+
+  ASSERT_OK(Put("foo", "bar"));
+
+  auto blob_files = blob_db_impl()->TEST_GetBlobFiles();
+  ASSERT_EQ(blob_files.size(), 1);
+
+  SyncPoint::GetInstance()->SetCallBack(
+      "BlobLogWriter::Sync", [this](void * /* arg */) {
+        fault_injection_env_->SetFilesystemActive(false, Status::IOError());
+      });
+  SyncPoint::GetInstance()->EnableProcessing();
+
+  const Status s = blob_db_impl()->TEST_CloseBlobFile(blob_files[0]);
+
+  fault_injection_env_->SetFilesystemActive(true);
+  SyncPoint::GetInstance()->DisableProcessing();
+  SyncPoint::GetInstance()->ClearAllCallBacks();
+
+  ASSERT_TRUE(s.IsIOError());
+}
+
+}  //  namespace blob_db
+}  // namespace ROCKSDB_NAMESPACE
+
+// A black-box test for the ttl wrapper around rocksdb
+int main(int argc, char **argv) {
+  ROCKSDB_NAMESPACE::port::InstallStackTraceHandler();
+  ::testing::InitGoogleTest(&argc, argv);
+  return RUN_ALL_TESTS();
+}
+
+#else
+#include <stdio.h>
+
+int main(int /*argc*/, char** /*argv*/) {
+  fprintf(stderr, "SKIPPED as BlobDB is not supported in ROCKSDB_LITE\n");
+  return 0;
+}
+
+#endif  // !ROCKSDB_LITE
diff --git a/src/rocksdb/utilities/blob_db/blob_dump_tool.cc b/src/rocksdb/utilities/blob_db/blob_dump_tool.cc
new file mode 100644
index 000000000..1e0632990
--- /dev/null
+++ b/src/rocksdb/utilities/blob_db/blob_dump_tool.cc
@@ -0,0 +1,282 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+#ifndef ROCKSDB_LITE
+
+#include "utilities/blob_db/blob_dump_tool.h"
+
+#include <stdio.h>
+
+#include <cinttypes>
+#include <iostream>
+#include <memory>
+#include <string>
+
+#include "file/random_access_file_reader.h"
+#include "file/readahead_raf.h"
+#include "port/port.h"
+#include "rocksdb/convenience.h"
+#include "rocksdb/file_system.h"
+#include "table/format.h"
+#include "util/coding.h"
+#include "util/string_util.h"
+
+namespace ROCKSDB_NAMESPACE {
+namespace blob_db {
+
+BlobDumpTool::BlobDumpTool()
+    : reader_(nullptr), buffer_(nullptr), buffer_size_(0) {}
+
+Status BlobDumpTool::Run(const std::string& filename, DisplayType show_key,
+                         DisplayType show_blob,
+                         DisplayType show_uncompressed_blob,
+                         bool show_summary) {
+  constexpr size_t kReadaheadSize = 2 * 1024 * 1024;
+  Status s;
+  const auto fs = FileSystem::Default();
+  IOOptions io_opts;
+  s = fs->FileExists(filename, io_opts, nullptr);
+  if (!s.ok()) {
+    return s;
+  }
+  uint64_t file_size = 0;
+  s = fs->GetFileSize(filename, io_opts, &file_size, nullptr);
+  if (!s.ok()) {
+    return s;
+  }
+  std::unique_ptr<FSRandomAccessFile> file;
+  s = fs->NewRandomAccessFile(filename, FileOptions(), &file, nullptr);
+  if (!s.ok()) {
+    return s;
+  }
+  file = NewReadaheadRandomAccessFile(std::move(file), kReadaheadSize);
+  if (file_size == 0) {
+    return Status::Corruption("File is empty.");
+  }
+  reader_.reset(new RandomAccessFileReader(std::move(file), filename));
+  uint64_t offset = 0;
+  uint64_t footer_offset = 0;
+  CompressionType compression = kNoCompression;
+  s = DumpBlobLogHeader(&offset, &compression);
+  if (!s.ok()) {
+    return s;
+  }
+  s = DumpBlobLogFooter(file_size, &footer_offset);
+  if (!s.ok()) {
+    return s;
+  }
+  uint64_t total_records = 0;
+  uint64_t total_key_size = 0;
+  uint64_t total_blob_size = 0;
+  uint64_t total_uncompressed_blob_size = 0;
+  if (show_key != DisplayType::kNone || show_summary) {
+    while (offset < footer_offset) {
+      s = DumpRecord(show_key, show_blob, show_uncompressed_blob, show_summary,
+                     compression, &offset, &total_records, &total_key_size,
+                     &total_blob_size, &total_uncompressed_blob_size);
+      if (!s.ok()) {
+        break;
+      }
+    }
+  }
+  if (show_summary) {
+    fprintf(stdout, "Summary:\n");
+    fprintf(stdout, "  total records: %" PRIu64 "\n", total_records);
+    fprintf(stdout, "  total key size: %" PRIu64 "\n", total_key_size);
+    fprintf(stdout, "  total blob size: %" PRIu64 "\n", total_blob_size);
+    if (compression != kNoCompression) {
+      fprintf(stdout, "  total raw blob size: %" PRIu64 "\n",
+              total_uncompressed_blob_size);
+    }
+  }
+  return s;
+}
+
+Status BlobDumpTool::Read(uint64_t offset, size_t size, Slice* result) {
+  if (buffer_size_ < size) {
+    if (buffer_size_ == 0) {
+      buffer_size_ = 4096;
+    }
+    while (buffer_size_ < size) {
+      buffer_size_ *= 2;
+    }
+    buffer_.reset(new char[buffer_size_]);
+  }
+  Status s = reader_->Read(IOOptions(), offset, size, result, buffer_.get(),
+                           nullptr, Env::IO_TOTAL /* rate_limiter_priority */);
+  if (!s.ok()) {
+    return s;
+  }
+  if (result->size() != size) {
+    return Status::Corruption("Reach the end of the file unexpectedly.");
+  }
+  return s;
+}
+
+Status BlobDumpTool::DumpBlobLogHeader(uint64_t* offset,
+                                       CompressionType* compression) {
+  Slice slice;
+  Status s = Read(0, BlobLogHeader::kSize, &slice);
+  if (!s.ok()) {
+    return s;
+  }
+  BlobLogHeader header;
+  s = header.DecodeFrom(slice);
+  if (!s.ok()) {
+    return s;
+  }
+  fprintf(stdout, "Blob log header:\n");
+  fprintf(stdout, "  Version          : %" PRIu32 "\n", header.version);
+  fprintf(stdout, "  Column Family ID : %" PRIu32 "\n",
+          header.column_family_id);
+  std::string compression_str;
+  if (!GetStringFromCompressionType(&compression_str, header.compression)
+           .ok()) {
+    compression_str = "Unrecongnized compression type (" +
+                      std::to_string((int)header.compression) + ")";
+  }
+  fprintf(stdout, "  Compression      : %s\n", compression_str.c_str());
+  fprintf(stdout, "  Expiration range : %s\n",
+          GetString(header.expiration_range).c_str());
+  *offset = BlobLogHeader::kSize;
+  *compression = header.compression;
+  return s;
+}
+
+Status BlobDumpTool::DumpBlobLogFooter(uint64_t file_size,
+                                       uint64_t* footer_offset) {
+  auto no_footer = [&]() {
+    *footer_offset = file_size;
+    fprintf(stdout, "No blob log footer.\n");
+    return Status::OK();
+  };
+  if (file_size < BlobLogHeader::kSize + BlobLogFooter::kSize) {
+    return no_footer();
+  }
+  Slice slice;
+  *footer_offset = file_size - BlobLogFooter::kSize;
+  Status s = Read(*footer_offset, BlobLogFooter::kSize, &slice);
+  if (!s.ok()) {
+    return s;
+  }
+  BlobLogFooter footer;
+  s = footer.DecodeFrom(slice);
+  if (!s.ok()) {
+    return no_footer();
+  }
+  fprintf(stdout, "Blob log footer:\n");
+  fprintf(stdout, "  Blob count       : %" PRIu64 "\n", footer.blob_count);
+  fprintf(stdout, "  Expiration Range : %s\n",
+          GetString(footer.expiration_range).c_str());
+  return s;
+}
+
+Status BlobDumpTool::DumpRecord(DisplayType show_key, DisplayType show_blob,
+                                DisplayType show_uncompressed_blob,
+                                bool show_summary, CompressionType compression,
+                                uint64_t* offset, uint64_t* total_records,
+                                uint64_t* total_key_size,
+                                uint64_t* total_blob_size,
+                                uint64_t* total_uncompressed_blob_size) {
+  if (show_key != DisplayType::kNone) {
+    fprintf(stdout, "Read record with offset 0x%" PRIx64 " (%" PRIu64 "):\n",
+            *offset, *offset);
+  }
+  Slice slice;
+  Status s = Read(*offset, BlobLogRecord::kHeaderSize, &slice);
+  if (!s.ok()) {
+    return s;
+  }
+  BlobLogRecord record;
+  s = record.DecodeHeaderFrom(slice);
+  if (!s.ok()) {
+    return s;
+  }
+  uint64_t key_size = record.key_size;
+  uint64_t value_size = record.value_size;
+  if (show_key != DisplayType::kNone) {
+    fprintf(stdout, "  key size   : %" PRIu64 "\n", key_size);
+    fprintf(stdout, "  value size : %" PRIu64 "\n", value_size);
+    fprintf(stdout, "  expiration : %" PRIu64 "\n", record.expiration);
+  }
+  *offset += BlobLogRecord::kHeaderSize;
+  s = Read(*offset, static_cast<size_t>(key_size + value_size), &slice);
+  if (!s.ok()) {
+    return s;
+  }
+  // Decompress value
+  std::string uncompressed_value;
+  if (compression != kNoCompression &&
+      (show_uncompressed_blob != DisplayType::kNone || show_summary)) {
+    BlockContents contents;
+    UncompressionContext context(compression);
+    UncompressionInfo info(context, UncompressionDict::GetEmptyDict(),
+                           compression);
+    s = UncompressBlockData(
+        info, slice.data() + key_size, static_cast<size_t>(value_size),
+        &contents, 2 /*compress_format_version*/, ImmutableOptions(Options()));
+    if (!s.ok()) {
+      return s;
+    }
+    uncompressed_value = contents.data.ToString();
+  }
+  if (show_key != DisplayType::kNone) {
+    fprintf(stdout, "  key        : ");
+    DumpSlice(Slice(slice.data(), static_cast<size_t>(key_size)), show_key);
+    if (show_blob != DisplayType::kNone) {
+      fprintf(stdout, "  blob       : ");
+      DumpSlice(Slice(slice.data() + static_cast<size_t>(key_size),
+                      static_cast<size_t>(value_size)),
+                show_blob);
+    }
+    if (show_uncompressed_blob != DisplayType::kNone) {
+      fprintf(stdout, "  raw blob   : ");
+      DumpSlice(Slice(uncompressed_value), show_uncompressed_blob);
+    }
+  }
+  *offset += key_size + value_size;
+  *total_records += 1;
+  *total_key_size += key_size;
+  *total_blob_size += value_size;
+  *total_uncompressed_blob_size += uncompressed_value.size();
+  return s;
+}
+
+void BlobDumpTool::DumpSlice(const Slice s, DisplayType type) {
+  if (type == DisplayType::kRaw) {
+    fprintf(stdout, "%s\n", s.ToString().c_str());
+  } else if (type == DisplayType::kHex) {
+    fprintf(stdout, "%s\n", s.ToString(true /*hex*/).c_str());
+  } else if (type == DisplayType::kDetail) {
+    char buf[100];
+    for (size_t i = 0; i < s.size(); i += 16) {
+      memset(buf, 0, sizeof(buf));
+      for (size_t j = 0; j < 16 && i + j < s.size(); j++) {
+        unsigned char c = s[i + j];
+        snprintf(buf + j * 3 + 15, 2, "%x", c >> 4);
+        snprintf(buf + j * 3 + 16, 2, "%x", c & 0xf);
+        snprintf(buf + j + 65, 2, "%c", (0x20 <= c && c <= 0x7e) ? c : '.');
+      }
+      for (size_t p = 0; p + 1 < sizeof(buf); p++) {
+        if (buf[p] == 0) {
+          buf[p] = ' ';
+        }
+      }
+      fprintf(stdout, "%s\n", i == 0 ? buf + 15 : buf);
+    }
+  }
+}
+
+template <class T>
+std::string BlobDumpTool::GetString(std::pair<T, T> p) {
+  if (p.first == 0 && p.second == 0) {
+    return "nil";
+  }
+  return "(" + std::to_string(p.first) + ", " + std::to_string(p.second) + ")";
+}
+
+}  // namespace blob_db
+}  // namespace ROCKSDB_NAMESPACE
+
+#endif  // ROCKSDB_LITE
diff --git a/src/rocksdb/utilities/blob_db/blob_dump_tool.h b/src/rocksdb/utilities/blob_db/blob_dump_tool.h
new file mode 100644
index 000000000..bece564e1
--- /dev/null
+++ b/src/rocksdb/utilities/blob_db/blob_dump_tool.h
@@ -0,0 +1,58 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+#pragma once
+#ifndef ROCKSDB_LITE
+
+#include <memory>
+#include <string>
+#include <utility>
+
+#include "db/blob/blob_log_format.h"
+#include "file/random_access_file_reader.h"
+#include "rocksdb/slice.h"
+#include "rocksdb/status.h"
+
+namespace ROCKSDB_NAMESPACE {
+namespace blob_db {
+
+class BlobDumpTool {
+ public:
+  enum class DisplayType {
+    kNone,
+    kRaw,
+    kHex,
+    kDetail,
+  };
+
+  BlobDumpTool();
+
+  Status Run(const std::string& filename, DisplayType show_key,
+             DisplayType show_blob, DisplayType show_uncompressed_blob,
+             bool show_summary);
+
+ private:
+  std::unique_ptr<RandomAccessFileReader> reader_;
+  std::unique_ptr<char[]> buffer_;
+  size_t buffer_size_;
+
+  Status Read(uint64_t offset, size_t size, Slice* result);
+  Status DumpBlobLogHeader(uint64_t* offset, CompressionType* compression);
+  Status DumpBlobLogFooter(uint64_t file_size, uint64_t* footer_offset);
+  Status DumpRecord(DisplayType show_key, DisplayType show_blob,
+                    DisplayType show_uncompressed_blob, bool show_summary,
+                    CompressionType compression, uint64_t* offset,
+                    uint64_t* total_records, uint64_t* total_key_size,
+                    uint64_t* total_blob_size,
+                    uint64_t* total_uncompressed_blob_size);
+  void DumpSlice(const Slice s, DisplayType type);
+
+  template <class T>
+  std::string GetString(std::pair<T, T> p);
+};
+
+}  // namespace blob_db
+}  // namespace ROCKSDB_NAMESPACE
+
+#endif  // ROCKSDB_LITE
diff --git a/src/rocksdb/utilities/blob_db/blob_file.cc b/src/rocksdb/utilities/blob_db/blob_file.cc
new file mode 100644
index 000000000..c68e557c6
--- /dev/null
+++ b/src/rocksdb/utilities/blob_db/blob_file.cc
@@ -0,0 +1,318 @@
+
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+#ifndef ROCKSDB_LITE
+#include "utilities/blob_db/blob_file.h"
+
+#include <stdio.h>
+
+#include <algorithm>
+#include <cinttypes>
+#include <memory>
+
+#include "db/column_family.h"
+#include "db/db_impl/db_impl.h"
+#include "db/dbformat.h"
+#include "file/filename.h"
+#include "file/readahead_raf.h"
+#include "logging/logging.h"
+#include "utilities/blob_db/blob_db_impl.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+namespace blob_db {
+
+BlobFile::BlobFile(const BlobDBImpl* p, const std::string& bdir, uint64_t fn,
+                   Logger* info_log)
+    : parent_(p), path_to_dir_(bdir), file_number_(fn), info_log_(info_log) {}
+
+BlobFile::BlobFile(const BlobDBImpl* p, const std::string& bdir, uint64_t fn,
+                   Logger* info_log, uint32_t column_family_id,
+                   CompressionType compression, bool has_ttl,
+                   const ExpirationRange& expiration_range)
+    : parent_(p),
+      path_to_dir_(bdir),
+      file_number_(fn),
+      info_log_(info_log),
+      column_family_id_(column_family_id),
+      compression_(compression),
+      has_ttl_(has_ttl),
+      expiration_range_(expiration_range),
+      header_(column_family_id, compression, has_ttl, expiration_range),
+      header_valid_(true) {}
+
+BlobFile::~BlobFile() {
+  if (obsolete_) {
+    std::string pn(PathName());
+    Status s = Env::Default()->DeleteFile(PathName());
+    if (!s.ok()) {
+      // ROCKS_LOG_INFO(db_options_.info_log,
+      // "File could not be deleted %s", pn.c_str());
+    }
+  }
+}
+
+uint32_t BlobFile::GetColumnFamilyId() const { return column_family_id_; }
+
+std::string BlobFile::PathName() const {
+  return BlobFileName(path_to_dir_, file_number_);
+}
+
+std::string BlobFile::DumpState() const {
+  char str[1000];
+  snprintf(
+      str, sizeof(str),
+      "path: %s fn: %" PRIu64 " blob_count: %" PRIu64 " file_size: %" PRIu64
+      " closed: %d obsolete: %d expiration_range: (%" PRIu64 ", %" PRIu64
+      "), writer: %d reader: %d",
+      path_to_dir_.c_str(), file_number_, blob_count_.load(), file_size_.load(),
+      closed_.load(), obsolete_.load(), expiration_range_.first,
+      expiration_range_.second, (!!log_writer_), (!!ra_file_reader_));
+  return str;
+}
+
+void BlobFile::MarkObsolete(SequenceNumber sequence) {
+  assert(Immutable());
+  obsolete_sequence_ = sequence;
+  obsolete_.store(true);
+}
+
+Status BlobFile::WriteFooterAndCloseLocked(SequenceNumber sequence) {
+  BlobLogFooter footer;
+  footer.blob_count = blob_count_;
+  if (HasTTL()) {
+    footer.expiration_range = expiration_range_;
+  }
+
+  // this will close the file and reset the Writable File Pointer.
+  Status s = log_writer_->AppendFooter(footer, /* checksum_method */ nullptr,
+                                       /* checksum_value */ nullptr);
+  if (s.ok()) {
+    closed_ = true;
+    immutable_sequence_ = sequence;
+    file_size_ += BlobLogFooter::kSize;
+  }
+  // delete the sequential writer
+  log_writer_.reset();
+  return s;
+}
+
+Status BlobFile::ReadFooter(BlobLogFooter* bf) {
+  if (file_size_ < (BlobLogHeader::kSize + BlobLogFooter::kSize)) {
+    return Status::IOError("File does not have footer", PathName());
+  }
+
+  uint64_t footer_offset = file_size_ - BlobLogFooter::kSize;
+  // assume that ra_file_reader_ is valid before we enter this
+  assert(ra_file_reader_);
+
+  Slice result;
+  std::string buf;
+  AlignedBuf aligned_buf;
+  Status s;
+  // TODO: rate limit reading footers from blob files.
+  if (ra_file_reader_->use_direct_io()) {
+    s = ra_file_reader_->Read(IOOptions(), footer_offset, BlobLogFooter::kSize,
+                              &result, nullptr, &aligned_buf,
+                              Env::IO_TOTAL /* rate_limiter_priority */);
+  } else {
+    buf.reserve(BlobLogFooter::kSize + 10);
+    s = ra_file_reader_->Read(IOOptions(), footer_offset, BlobLogFooter::kSize,
+                              &result, &buf[0], nullptr,
+                              Env::IO_TOTAL /* rate_limiter_priority */);
+  }
+  if (!s.ok()) return s;
+  if (result.size() != BlobLogFooter::kSize) {
+    // should not happen
+    return Status::IOError("EOF reached before footer");
+  }
+
+  s = bf->DecodeFrom(result);
+  return s;
+}
+
+Status BlobFile::SetFromFooterLocked(const BlobLogFooter& footer) {
+  blob_count_ = footer.blob_count;
+  expiration_range_ = footer.expiration_range;
+  closed_ = true;
+  return Status::OK();
+}
+
+Status BlobFile::Fsync() {
+  Status s;
+  if (log_writer_.get()) {
+    s = log_writer_->Sync();
+  }
+  return s;
+}
+
+void BlobFile::CloseRandomAccessLocked() {
+  ra_file_reader_.reset();
+  last_access_ = -1;
+}
+
+Status BlobFile::GetReader(Env* env, const FileOptions& file_options,
+                           std::shared_ptr<RandomAccessFileReader>* reader,
+                           bool* fresh_open) {
+  assert(reader != nullptr);
+  assert(fresh_open != nullptr);
+  *fresh_open = false;
+  int64_t current_time = 0;
+  if (env->GetCurrentTime(&current_time).ok()) {
+    last_access_.store(current_time);
+  }
+  Status s;
+
+  {
+    ReadLock lockbfile_r(&mutex_);
+    if (ra_file_reader_) {
+      *reader = ra_file_reader_;
+      return s;
+    }
+  }
+
+  WriteLock lockbfile_w(&mutex_);
+  // Double check.
+  if (ra_file_reader_) {
+    *reader = ra_file_reader_;
+    return s;
+  }
+
+  std::unique_ptr<FSRandomAccessFile> rfile;
+  s = env->GetFileSystem()->NewRandomAccessFile(PathName(), file_options,
+                                                &rfile, nullptr);
+  if (!s.ok()) {
+    ROCKS_LOG_ERROR(info_log_,
+                    "Failed to open blob file for random-read: %s status: '%s'"
+                    " exists: '%s'",
+                    PathName().c_str(), s.ToString().c_str(),
+                    env->FileExists(PathName()).ToString().c_str());
+    return s;
+  }
+
+  ra_file_reader_ =
+      std::make_shared<RandomAccessFileReader>(std::move(rfile), PathName());
+  *reader = ra_file_reader_;
+  *fresh_open = true;
+  return s;
+}
+
+Status BlobFile::ReadMetadata(const std::shared_ptr<FileSystem>& fs,
+                              const FileOptions& file_options) {
+  assert(Immutable());
+  // Get file size.
+  uint64_t file_size = 0;
+  Status s =
+      fs->GetFileSize(PathName(), file_options.io_options, &file_size, nullptr);
+  if (s.ok()) {
+    file_size_ = file_size;
+  } else {
+    ROCKS_LOG_ERROR(info_log_,
+                    "Failed to get size of blob file %" PRIu64 ", status: %s",
+                    file_number_, s.ToString().c_str());
+    return s;
+  }
+  if (file_size < BlobLogHeader::kSize) {
+    ROCKS_LOG_ERROR(
+        info_log_, "Incomplete blob file blob file %" PRIu64 ", size: %" PRIu64,
+        file_number_, file_size);
+    return Status::Corruption("Incomplete blob file header.");
+  }
+
+  // Create file reader.
+  std::unique_ptr<RandomAccessFileReader> file_reader;
+  s = RandomAccessFileReader::Create(fs, PathName(), file_options, &file_reader,
+                                     nullptr);
+  if (!s.ok()) {
+    ROCKS_LOG_ERROR(info_log_,
+                    "Failed to open blob file %" PRIu64 ", status: %s",
+                    file_number_, s.ToString().c_str());
+    return s;
+  }
+
+  // Read file header.
+  std::string header_buf;
+  AlignedBuf aligned_buf;
+  Slice header_slice;
+  // TODO: rate limit reading headers from blob files.
+  if (file_reader->use_direct_io()) {
+    s = file_reader->Read(IOOptions(), 0, BlobLogHeader::kSize, &header_slice,
+                          nullptr, &aligned_buf,
+                          Env::IO_TOTAL /* rate_limiter_priority */);
+  } else {
+    header_buf.reserve(BlobLogHeader::kSize);
+    s = file_reader->Read(IOOptions(), 0, BlobLogHeader::kSize, &header_slice,
+                          &header_buf[0], nullptr,
+                          Env::IO_TOTAL /* rate_limiter_priority */);
+  }
+  if (!s.ok()) {
+    ROCKS_LOG_ERROR(
+        info_log_, "Failed to read header of blob file %" PRIu64 ", status: %s",
+        file_number_, s.ToString().c_str());
+    return s;
+  }
+  BlobLogHeader header;
+  s = header.DecodeFrom(header_slice);
+  if (!s.ok()) {
+    ROCKS_LOG_ERROR(info_log_,
+                    "Failed to decode header of blob file %" PRIu64
+                    ", status: %s",
+                    file_number_, s.ToString().c_str());
+    return s;
+  }
+  column_family_id_ = header.column_family_id;
+  compression_ = header.compression;
+  has_ttl_ = header.has_ttl;
+  if (has_ttl_) {
+    expiration_range_ = header.expiration_range;
+  }
+  header_valid_ = true;
+
+  // Read file footer.
+  if (file_size_ < BlobLogHeader::kSize + BlobLogFooter::kSize) {
+    // OK not to have footer.
+    assert(!footer_valid_);
+    return Status::OK();
+  }
+  std::string footer_buf;
+  Slice footer_slice;
+  // TODO: rate limit reading footers from blob files.
+  if (file_reader->use_direct_io()) {
+    s = file_reader->Read(IOOptions(), file_size - BlobLogFooter::kSize,
+                          BlobLogFooter::kSize, &footer_slice, nullptr,
+                          &aligned_buf,
+                          Env::IO_TOTAL /* rate_limiter_priority */);
+  } else {
+    footer_buf.reserve(BlobLogFooter::kSize);
+    s = file_reader->Read(IOOptions(), file_size - BlobLogFooter::kSize,
+                          BlobLogFooter::kSize, &footer_slice, &footer_buf[0],
+                          nullptr, Env::IO_TOTAL /* rate_limiter_priority */);
+  }
+  if (!s.ok()) {
+    ROCKS_LOG_ERROR(
+        info_log_, "Failed to read footer of blob file %" PRIu64 ", status: %s",
+        file_number_, s.ToString().c_str());
+    return s;
+  }
+  BlobLogFooter footer;
+  s = footer.DecodeFrom(footer_slice);
+  if (!s.ok()) {
+    // OK not to have footer.
+    assert(!footer_valid_);
+    return Status::OK();
+  }
+  blob_count_ = footer.blob_count;
+  if (has_ttl_) {
+    assert(header.expiration_range.first <= footer.expiration_range.first);
+    assert(header.expiration_range.second >= footer.expiration_range.second);
+    expiration_range_ = footer.expiration_range;
+  }
+  footer_valid_ = true;
+  return Status::OK();
+}
+
+}  // namespace blob_db
+}  // namespace ROCKSDB_NAMESPACE
+#endif  // ROCKSDB_LITE
diff --git a/src/rocksdb/utilities/blob_db/blob_file.h b/src/rocksdb/utilities/blob_db/blob_file.h
new file mode 100644
index 000000000..6f3f2bea7
--- /dev/null
+++ b/src/rocksdb/utilities/blob_db/blob_file.h
@@ -0,0 +1,246 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+#pragma once
+#ifndef ROCKSDB_LITE
+
+#include <atomic>
+#include <limits>
+#include <memory>
+#include <unordered_set>
+
+#include "db/blob/blob_log_format.h"
+#include "db/blob/blob_log_writer.h"
+#include "file/random_access_file_reader.h"
+#include "port/port.h"
+#include "rocksdb/env.h"
+#include "rocksdb/file_system.h"
+#include "rocksdb/options.h"
+
+namespace ROCKSDB_NAMESPACE {
+namespace blob_db {
+
+class BlobDBImpl;
+
+class BlobFile {
+  friend class BlobDBImpl;
+  friend struct BlobFileComparator;
+  friend struct BlobFileComparatorTTL;
+  friend class BlobIndexCompactionFilterBase;
+  friend class BlobIndexCompactionFilterGC;
+
+ private:
+  // access to parent
+  const BlobDBImpl* parent_{nullptr};
+
+  // path to blob directory
+  std::string path_to_dir_;
+
+  // the id of the file.
+  // the above 2 are created during file creation and never changed
+  // after that
+  uint64_t file_number_{0};
+
+  // The file numbers of the SST files whose oldest blob file reference
+  // points to this blob file.
+  std::unordered_set<uint64_t> linked_sst_files_;
+
+  // Info log.
+  Logger* info_log_{nullptr};
+
+  // Column family id.
+  uint32_t column_family_id_{std::numeric_limits<uint32_t>::max()};
+
+  // Compression type of blobs in the file
+  CompressionType compression_{kNoCompression};
+
+  // If true, the keys in this file all has TTL. Otherwise all keys don't
+  // have TTL.
+  bool has_ttl_{false};
+
+  // TTL range of blobs in the file.
+  ExpirationRange expiration_range_;
+
+  // number of blobs in the file
+  std::atomic<uint64_t> blob_count_{0};
+
+  // size of the file
+  std::atomic<uint64_t> file_size_{0};
+
+  BlobLogHeader header_;
+
+  // closed_ = true implies the file is no more mutable
+  // no more blobs will be appended and the footer has been written out
+  std::atomic<bool> closed_{false};
+
+  // The latest sequence number when the file was closed/made immutable.
+  SequenceNumber immutable_sequence_{0};
+
+  // Whether the file was marked obsolete (due to either TTL or GC).
+  // obsolete_ still needs to do iterator/snapshot checks
+  std::atomic<bool> obsolete_{false};
+
+  // The last sequence number by the time the file marked as obsolete.
+  // Data in this file is visible to a snapshot taken before the sequence.
+  SequenceNumber obsolete_sequence_{0};
+
+  // Sequential/Append writer for blobs
+  std::shared_ptr<BlobLogWriter> log_writer_;
+
+  // random access file reader for GET calls
+  std::shared_ptr<RandomAccessFileReader> ra_file_reader_;
+
+  // This Read-Write mutex is per file specific and protects
+  // all the datastructures
+  mutable port::RWMutex mutex_;
+
+  // time when the random access reader was last created.
+  std::atomic<std::int64_t> last_access_{-1};
+
+  bool header_valid_{false};
+
+  bool footer_valid_{false};
+
+ public:
+  BlobFile() = default;
+
+  BlobFile(const BlobDBImpl* parent, const std::string& bdir, uint64_t fnum,
+           Logger* info_log);
+
+  BlobFile(const BlobDBImpl* parent, const std::string& bdir, uint64_t fnum,
+           Logger* info_log, uint32_t column_family_id,
+           CompressionType compression, bool has_ttl,
+           const ExpirationRange& expiration_range);
+
+  ~BlobFile();
+
+  uint32_t GetColumnFamilyId() const;
+
+  // Returns log file's absolute pathname.
+  std::string PathName() const;
+
+  // Primary identifier for blob file.
+  // once the file is created, this never changes
+  uint64_t BlobFileNumber() const { return file_number_; }
+
+  // Get the set of SST files whose oldest blob file reference points to
+  // this file.
+  const std::unordered_set<uint64_t>& GetLinkedSstFiles() const {
+    return linked_sst_files_;
+  }
+
+  // Link an SST file whose oldest blob file reference points to this file.
+  void LinkSstFile(uint64_t sst_file_number) {
+    assert(linked_sst_files_.find(sst_file_number) == linked_sst_files_.end());
+    linked_sst_files_.insert(sst_file_number);
+  }
+
+  // Unlink an SST file whose oldest blob file reference points to this file.
+  void UnlinkSstFile(uint64_t sst_file_number) {
+    auto it = linked_sst_files_.find(sst_file_number);
+    assert(it != linked_sst_files_.end());
+    linked_sst_files_.erase(it);
+  }
+
+  // the following functions are atomic, and don't need
+  // read lock
+  uint64_t BlobCount() const {
+    return blob_count_.load(std::memory_order_acquire);
+  }
+
+  std::string DumpState() const;
+
+  // if the file is not taking any more appends.
+  bool Immutable() const { return closed_.load(); }
+
+  // Mark the file as immutable.
+  // REQUIRES: write lock held, or access from single thread (on DB open).
+  void MarkImmutable(SequenceNumber sequence) {
+    closed_ = true;
+    immutable_sequence_ = sequence;
+  }
+
+  SequenceNumber GetImmutableSequence() const {
+    assert(Immutable());
+    return immutable_sequence_;
+  }
+
+  // Whether the file was marked obsolete (due to either TTL or GC).
+  bool Obsolete() const {
+    assert(Immutable() || !obsolete_.load());
+    return obsolete_.load();
+  }
+
+  // Mark file as obsolete (due to either TTL or GC). The file is not visible to
+  // snapshots with sequence greater or equal to the given sequence.
+  void MarkObsolete(SequenceNumber sequence);
+
+  SequenceNumber GetObsoleteSequence() const {
+    assert(Obsolete());
+    return obsolete_sequence_;
+  }
+
+  Status Fsync();
+
+  uint64_t GetFileSize() const {
+    return file_size_.load(std::memory_order_acquire);
+  }
+
+  // All Get functions which are not atomic, will need ReadLock on the mutex
+
+  const ExpirationRange& GetExpirationRange() const {
+    return expiration_range_;
+  }
+
+  void ExtendExpirationRange(uint64_t expiration) {
+    expiration_range_.first = std::min(expiration_range_.first, expiration);
+    expiration_range_.second = std::max(expiration_range_.second, expiration);
+  }
+
+  bool HasTTL() const { return has_ttl_; }
+
+  void SetHasTTL(bool has_ttl) { has_ttl_ = has_ttl; }
+
+  CompressionType GetCompressionType() const { return compression_; }
+
+  std::shared_ptr<BlobLogWriter> GetWriter() const { return log_writer_; }
+
+  // Read blob file header and footer. Return corruption if file header is
+  // malform or incomplete. If footer is malform or incomplete, set
+  // footer_valid_ to false and return Status::OK.
+  Status ReadMetadata(const std::shared_ptr<FileSystem>& fs,
+                      const FileOptions& file_options);
+
+  Status GetReader(Env* env, const FileOptions& file_options,
+                   std::shared_ptr<RandomAccessFileReader>* reader,
+                   bool* fresh_open);
+
+ private:
+  Status ReadFooter(BlobLogFooter* footer);
+
+  Status WriteFooterAndCloseLocked(SequenceNumber sequence);
+
+  void CloseRandomAccessLocked();
+
+  // this is used, when you are reading only the footer of a
+  // previously closed file
+  Status SetFromFooterLocked(const BlobLogFooter& footer);
+
+  void set_expiration_range(const ExpirationRange& expiration_range) {
+    expiration_range_ = expiration_range;
+  }
+
+  // The following functions are atomic, and don't need locks
+  void SetFileSize(uint64_t fs) { file_size_ = fs; }
+
+  void SetBlobCount(uint64_t bc) { blob_count_ = bc; }
+
+  void BlobRecordAdded(uint64_t record_size) {
+    ++blob_count_;
+    file_size_ += record_size;
+  }
+};
+}  // namespace blob_db
+}  // namespace ROCKSDB_NAMESPACE
+#endif  // ROCKSDB_LITE
diff --git a/src/rocksdb/utilities/cache_dump_load.cc b/src/rocksdb/utilities/cache_dump_load.cc
new file mode 100644
index 000000000..9a7c76798
--- /dev/null
+++ b/src/rocksdb/utilities/cache_dump_load.cc
@@ -0,0 +1,69 @@
+//  Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#ifndef ROCKSDB_LITE
+
+#include "rocksdb/utilities/cache_dump_load.h"
+
+#include "file/writable_file_writer.h"
+#include "port/lang.h"
+#include "rocksdb/env.h"
+#include "rocksdb/file_system.h"
+#include "table/format.h"
+#include "util/crc32c.h"
+#include "utilities/cache_dump_load_impl.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+IOStatus NewToFileCacheDumpWriter(const std::shared_ptr<FileSystem>& fs,
+                                  const FileOptions& file_opts,
+                                  const std::string& file_name,
+                                  std::unique_ptr<CacheDumpWriter>* writer) {
+  std::unique_ptr<WritableFileWriter> file_writer;
+  IOStatus io_s = WritableFileWriter::Create(fs, file_name, file_opts,
+                                             &file_writer, nullptr);
+  if (!io_s.ok()) {
+    return io_s;
+  }
+  writer->reset(new ToFileCacheDumpWriter(std::move(file_writer)));
+  return io_s;
+}
+
+IOStatus NewFromFileCacheDumpReader(const std::shared_ptr<FileSystem>& fs,
+                                    const FileOptions& file_opts,
+                                    const std::string& file_name,
+                                    std::unique_ptr<CacheDumpReader>* reader) {
+  std::unique_ptr<RandomAccessFileReader> file_reader;
+  IOStatus io_s = RandomAccessFileReader::Create(fs, file_name, file_opts,
+                                                 &file_reader, nullptr);
+  if (!io_s.ok()) {
+    return io_s;
+  }
+  reader->reset(new FromFileCacheDumpReader(std::move(file_reader)));
+  return io_s;
+}
+
+Status NewDefaultCacheDumper(const CacheDumpOptions& dump_options,
+                             const std::shared_ptr<Cache>& cache,
+                             std::unique_ptr<CacheDumpWriter>&& writer,
+                             std::unique_ptr<CacheDumper>* cache_dumper) {
+  cache_dumper->reset(
+      new CacheDumperImpl(dump_options, cache, std::move(writer)));
+  return Status::OK();
+}
+
+Status NewDefaultCacheDumpedLoader(
+    const CacheDumpOptions& dump_options,
+    const BlockBasedTableOptions& toptions,
+    const std::shared_ptr<SecondaryCache>& secondary_cache,
+    std::unique_ptr<CacheDumpReader>&& reader,
+    std::unique_ptr<CacheDumpedLoader>* cache_dump_loader) {
+  cache_dump_loader->reset(new CacheDumpedLoaderImpl(
+      dump_options, toptions, secondary_cache, std::move(reader)));
+  return Status::OK();
+}
+
+}  // namespace ROCKSDB_NAMESPACE
+#endif  // ROCKSDB_LITE
diff --git a/src/rocksdb/utilities/cache_dump_load_impl.cc b/src/rocksdb/utilities/cache_dump_load_impl.cc
new file mode 100644
index 000000000..2b9f2a29d
--- /dev/null
+++ b/src/rocksdb/utilities/cache_dump_load_impl.cc
@@ -0,0 +1,393 @@
+//  Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#include "cache/cache_key.h"
+#include "table/block_based/block_based_table_reader.h"
+#ifndef ROCKSDB_LITE
+
+#include "cache/cache_entry_roles.h"
+#include "file/writable_file_writer.h"
+#include "port/lang.h"
+#include "rocksdb/env.h"
+#include "rocksdb/file_system.h"
+#include "rocksdb/utilities/ldb_cmd.h"
+#include "table/format.h"
+#include "util/crc32c.h"
+#include "utilities/cache_dump_load_impl.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+// Set the dump filter with a list of DBs. Block cache may be shared by multipe
+// DBs and we may only want to dump out the blocks belonging to certain DB(s).
+// Therefore, a filter is need to decide if the key of the block satisfy the
+// requirement.
+Status CacheDumperImpl::SetDumpFilter(std::vector<DB*> db_list) {
+  Status s = Status::OK();
+  for (size_t i = 0; i < db_list.size(); i++) {
+    assert(i < db_list.size());
+    TablePropertiesCollection ptc;
+    assert(db_list[i] != nullptr);
+    s = db_list[i]->GetPropertiesOfAllTables(&ptc);
+    if (!s.ok()) {
+      return s;
+    }
+    for (auto id = ptc.begin(); id != ptc.end(); id++) {
+      OffsetableCacheKey base;
+      // We only want to save cache entries that are portable to another
+      // DB::Open, so only save entries with stable keys.
+      bool is_stable;
+      BlockBasedTable::SetupBaseCacheKey(id->second.get(),
+                                         /*cur_db_session_id*/ "",
+                                         /*cur_file_num*/ 0, &base, &is_stable);
+      if (is_stable) {
+        Slice prefix_slice = base.CommonPrefixSlice();
+        assert(prefix_slice.size() == OffsetableCacheKey::kCommonPrefixSize);
+        prefix_filter_.insert(prefix_slice.ToString());
+      }
+    }
+  }
+  return s;
+}
+
+// This is the main function to dump out the cache block entries to the writer.
+// The writer may create a file or write to other systems. Currently, we will
+// iterate the whole block cache, get the blocks, and write them to the writer
+IOStatus CacheDumperImpl::DumpCacheEntriesToWriter() {
+  // Prepare stage, check the parameters.
+  if (cache_ == nullptr) {
+    return IOStatus::InvalidArgument("Cache is null");
+  }
+  if (writer_ == nullptr) {
+    return IOStatus::InvalidArgument("CacheDumpWriter is null");
+  }
+  // Set the system clock
+  if (options_.clock == nullptr) {
+    return IOStatus::InvalidArgument("System clock is null");
+  }
+  clock_ = options_.clock;
+  // We copy the Cache Deleter Role Map as its member.
+  role_map_ = CopyCacheDeleterRoleMap();
+  // Set the sequence number
+  sequence_num_ = 0;
+
+  // Dump stage, first, we write the hader
+  IOStatus io_s = WriteHeader();
+  if (!io_s.ok()) {
+    return io_s;
+  }
+
+  // Then, we iterate the block cache and dump out the blocks that are not
+  // filtered out.
+  cache_->ApplyToAllEntries(DumpOneBlockCallBack(), {});
+
+  // Finally, write the footer
+  io_s = WriteFooter();
+  if (!io_s.ok()) {
+    return io_s;
+  }
+  io_s = writer_->Close();
+  return io_s;
+}
+
+// Check if we need to filter out the block based on its key
+bool CacheDumperImpl::ShouldFilterOut(const Slice& key) {
+  if (key.size() < OffsetableCacheKey::kCommonPrefixSize) {
+    return /*filter out*/ true;
+  }
+  Slice key_prefix(key.data(), OffsetableCacheKey::kCommonPrefixSize);
+  std::string prefix = key_prefix.ToString();
+  // Filter out if not found
+  return prefix_filter_.find(prefix) == prefix_filter_.end();
+}
+
+// This is the callback function which will be applied to
+// Cache::ApplyToAllEntries. In this callback function, we will get the block
+// type, decide if the block needs to be dumped based on the filter, and write
+// the block through the provided writer.
+std::function<void(const Slice&, void*, size_t, Cache::DeleterFn)>
+CacheDumperImpl::DumpOneBlockCallBack() {
+  return [&](const Slice& key, void* value, size_t /*charge*/,
+             Cache::DeleterFn deleter) {
+    // Step 1: get the type of the block from role_map_
+    auto e = role_map_.find(deleter);
+    CacheEntryRole role;
+    CacheDumpUnitType type = CacheDumpUnitType::kBlockTypeMax;
+    if (e == role_map_.end()) {
+      role = CacheEntryRole::kMisc;
+    } else {
+      role = e->second;
+    }
+    bool filter_out = false;
+
+    // Step 2: based on the key prefix, check if the block should be filter out.
+    if (ShouldFilterOut(key)) {
+      filter_out = true;
+    }
+
+    // Step 3: based on the block type, get the block raw pointer and length.
+    const char* block_start = nullptr;
+    size_t block_len = 0;
+    switch (role) {
+      case CacheEntryRole::kDataBlock:
+        type = CacheDumpUnitType::kData;
+        block_start = (static_cast<Block*>(value))->data();
+        block_len = (static_cast<Block*>(value))->size();
+        break;
+      case CacheEntryRole::kFilterBlock:
+        type = CacheDumpUnitType::kFilter;
+        block_start = (static_cast<ParsedFullFilterBlock*>(value))
+                          ->GetBlockContentsData()
+                          .data();
+        block_len = (static_cast<ParsedFullFilterBlock*>(value))
+                        ->GetBlockContentsData()
+                        .size();
+        break;
+      case CacheEntryRole::kFilterMetaBlock:
+        type = CacheDumpUnitType::kFilterMetaBlock;
+        block_start = (static_cast<Block*>(value))->data();
+        block_len = (static_cast<Block*>(value))->size();
+        break;
+      case CacheEntryRole::kIndexBlock:
+        type = CacheDumpUnitType::kIndex;
+        block_start = (static_cast<Block*>(value))->data();
+        block_len = (static_cast<Block*>(value))->size();
+        break;
+      case CacheEntryRole::kDeprecatedFilterBlock:
+        // Obsolete
+        filter_out = true;
+        break;
+      case CacheEntryRole::kMisc:
+        filter_out = true;
+        break;
+      case CacheEntryRole::kOtherBlock:
+        filter_out = true;
+        break;
+      case CacheEntryRole::kWriteBuffer:
+        filter_out = true;
+        break;
+      default:
+        filter_out = true;
+    }
+
+    // Step 4: if the block should not be filter out, write the block to the
+    // CacheDumpWriter
+    if (!filter_out && block_start != nullptr) {
+      WriteBlock(type, key, Slice(block_start, block_len))
+          .PermitUncheckedError();
+    }
+  };
+}
+
+// Write the block to the writer. It takes the timestamp of the
+// block being copied from block cache, block type, key, block pointer,
+// block size and block checksum as the input. When writing the dumper raw
+// block, we first create the dump unit and encoude it to a string. Then,
+// we calculate the checksum of the whole dump unit string and store it in
+// the dump unit metadata.
+// First, we write the metadata first, which is a fixed size string. Then, we
+// Append the dump unit string to the writer.
+IOStatus CacheDumperImpl::WriteBlock(CacheDumpUnitType type, const Slice& key,
+                                     const Slice& value) {
+  uint64_t timestamp = clock_->NowMicros();
+  uint32_t value_checksum = crc32c::Value(value.data(), value.size());
+
+  // First, serialize the block information in a string
+  DumpUnit dump_unit;
+  dump_unit.timestamp = timestamp;
+  dump_unit.key = key;
+  dump_unit.type = type;
+  dump_unit.value_len = value.size();
+  dump_unit.value = const_cast<char*>(value.data());
+  dump_unit.value_checksum = value_checksum;
+  std::string encoded_data;
+  CacheDumperHelper::EncodeDumpUnit(dump_unit, &encoded_data);
+
+  // Second, create the metadata, which contains a sequence number, the dump
+  // unit string checksum and the string size. The sequence number monotonically
+  // increases from 0.
+  DumpUnitMeta unit_meta;
+  unit_meta.sequence_num = sequence_num_;
+  sequence_num_++;
+  unit_meta.dump_unit_checksum =
+      crc32c::Value(encoded_data.data(), encoded_data.size());
+  unit_meta.dump_unit_size = encoded_data.size();
+  std::string encoded_meta;
+  CacheDumperHelper::EncodeDumpUnitMeta(unit_meta, &encoded_meta);
+
+  // We write the metadata first.
+  assert(writer_ != nullptr);
+  IOStatus io_s = writer_->WriteMetadata(encoded_meta);
+  if (!io_s.ok()) {
+    return io_s;
+  }
+  // followed by the dump unit.
+  return writer_->WritePacket(encoded_data);
+}
+
+// Before we write any block, we write the header first to store the cache dump
+// format version, rocksdb version, and brief intro.
+IOStatus CacheDumperImpl::WriteHeader() {
+  std::string header_key = "header";
+  std::ostringstream s;
+  s << kTraceMagic << "\t"
+    << "Cache dump format version: " << kCacheDumpMajorVersion << "."
+    << kCacheDumpMinorVersion << "\t"
+    << "RocksDB Version: " << kMajorVersion << "." << kMinorVersion << "\t"
+    << "Format: dump_unit_metadata <sequence_number, dump_unit_checksum, "
+       "dump_unit_size>, dump_unit <timestamp, key, block_type, "
+       "block_size, block_data, block_checksum> cache_value\n";
+  std::string header_value(s.str());
+  CacheDumpUnitType type = CacheDumpUnitType::kHeader;
+  return WriteBlock(type, header_key, header_value);
+}
+
+// Write the footer after all the blocks are stored to indicate the ending.
+IOStatus CacheDumperImpl::WriteFooter() {
+  std::string footer_key = "footer";
+  std::string footer_value("cache dump completed");
+  CacheDumpUnitType type = CacheDumpUnitType::kFooter;
+  return WriteBlock(type, footer_key, footer_value);
+}
+
+// This is the main function to restore the cache entries to secondary cache.
+// First, we check if all the arguments are valid. Then, we read the block
+// sequentially from the reader and insert them to the secondary cache.
+IOStatus CacheDumpedLoaderImpl::RestoreCacheEntriesToSecondaryCache() {
+  // TODO: remove this line when options are used in the loader
+  (void)options_;
+  // Step 1: we check if all the arguments are valid
+  if (secondary_cache_ == nullptr) {
+    return IOStatus::InvalidArgument("Secondary Cache is null");
+  }
+  if (reader_ == nullptr) {
+    return IOStatus::InvalidArgument("CacheDumpReader is null");
+  }
+  // we copy the Cache Deleter Role Map as its member.
+  role_map_ = CopyCacheDeleterRoleMap();
+
+  // Step 2: read the header
+  // TODO: we need to check the cache dump format version and RocksDB version
+  // after the header is read out.
+  IOStatus io_s;
+  DumpUnit dump_unit;
+  std::string data;
+  io_s = ReadHeader(&data, &dump_unit);
+  if (!io_s.ok()) {
+    return io_s;
+  }
+
+  // Step 3: read out the rest of the blocks from the reader. The loop will stop
+  // either I/O status is not ok or we reach to the the end.
+  while (io_s.ok()) {
+    dump_unit.reset();
+    data.clear();
+    // read the content and store in the dump_unit
+    io_s = ReadCacheBlock(&data, &dump_unit);
+    if (!io_s.ok()) {
+      break;
+    }
+    if (dump_unit.type == CacheDumpUnitType::kFooter) {
+      break;
+    }
+    // Create the uncompressed_block based on the information in the dump_unit
+    // (There is no block trailer here compatible with block-based SST file.)
+    Slice content =
+        Slice(static_cast<char*>(dump_unit.value), dump_unit.value_len);
+    Status s = secondary_cache_->InsertSaved(dump_unit.key, content);
+    if (!s.ok()) {
+      io_s = status_to_io_status(std::move(s));
+    }
+  }
+  if (dump_unit.type == CacheDumpUnitType::kFooter) {
+    return IOStatus::OK();
+  } else {
+    return io_s;
+  }
+}
+
+// Read and copy the dump unit metadata to std::string data, decode and create
+// the unit metadata based on the string
+IOStatus CacheDumpedLoaderImpl::ReadDumpUnitMeta(std::string* data,
+                                                 DumpUnitMeta* unit_meta) {
+  assert(reader_ != nullptr);
+  assert(data != nullptr);
+  assert(unit_meta != nullptr);
+  IOStatus io_s = reader_->ReadMetadata(data);
+  if (!io_s.ok()) {
+    return io_s;
+  }
+  return status_to_io_status(
+      CacheDumperHelper::DecodeDumpUnitMeta(*data, unit_meta));
+}
+
+// Read and copy the dump unit to std::string data, decode and create the unit
+// based on the string
+IOStatus CacheDumpedLoaderImpl::ReadDumpUnit(size_t len, std::string* data,
+                                             DumpUnit* unit) {
+  assert(reader_ != nullptr);
+  assert(data != nullptr);
+  assert(unit != nullptr);
+  IOStatus io_s = reader_->ReadPacket(data);
+  if (!io_s.ok()) {
+    return io_s;
+  }
+  if (data->size() != len) {
+    return IOStatus::Corruption(
+        "The data being read out does not match the size stored in metadata!");
+  }
+  Slice block;
+  return status_to_io_status(CacheDumperHelper::DecodeDumpUnit(*data, unit));
+}
+
+// Read the header
+IOStatus CacheDumpedLoaderImpl::ReadHeader(std::string* data,
+                                           DumpUnit* dump_unit) {
+  DumpUnitMeta header_meta;
+  header_meta.reset();
+  std::string meta_string;
+  IOStatus io_s = ReadDumpUnitMeta(&meta_string, &header_meta);
+  if (!io_s.ok()) {
+    return io_s;
+  }
+
+  io_s = ReadDumpUnit(header_meta.dump_unit_size, data, dump_unit);
+  if (!io_s.ok()) {
+    return io_s;
+  }
+  uint32_t unit_checksum = crc32c::Value(data->data(), data->size());
+  if (unit_checksum != header_meta.dump_unit_checksum) {
+    return IOStatus::Corruption("Read header unit corrupted!");
+  }
+  return io_s;
+}
+
+// Read the blocks after header is read out
+IOStatus CacheDumpedLoaderImpl::ReadCacheBlock(std::string* data,
+                                               DumpUnit* dump_unit) {
+  // According to the write process, we read the dump_unit_metadata first
+  DumpUnitMeta unit_meta;
+  unit_meta.reset();
+  std::string unit_string;
+  IOStatus io_s = ReadDumpUnitMeta(&unit_string, &unit_meta);
+  if (!io_s.ok()) {
+    return io_s;
+  }
+
+  // Based on the information in the dump_unit_metadata, we read the dump_unit
+  // and verify if its content is correct.
+  io_s = ReadDumpUnit(unit_meta.dump_unit_size, data, dump_unit);
+  if (!io_s.ok()) {
+    return io_s;
+  }
+  uint32_t unit_checksum = crc32c::Value(data->data(), data->size());
+  if (unit_checksum != unit_meta.dump_unit_checksum) {
+    return IOStatus::Corruption(
+        "Checksum does not match! Read dumped unit corrupted!");
+  }
+  return io_s;
+}
+
+}  // namespace ROCKSDB_NAMESPACE
+#endif  // ROCKSDB_LITE
diff --git a/src/rocksdb/utilities/cache_dump_load_impl.h b/src/rocksdb/utilities/cache_dump_load_impl.h
new file mode 100644
index 000000000..9ca1ff45a
--- /dev/null
+++ b/src/rocksdb/utilities/cache_dump_load_impl.h
@@ -0,0 +1,359 @@
+//  Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#pragma once
+#ifndef ROCKSDB_LITE
+
+#include <unordered_map>
+
+#include "file/random_access_file_reader.h"
+#include "file/writable_file_writer.h"
+#include "rocksdb/utilities/cache_dump_load.h"
+#include "table/block_based/block.h"
+#include "table/block_based/block_like_traits.h"
+#include "table/block_based/block_type.h"
+#include "table/block_based/cachable_entry.h"
+#include "table/block_based/parsed_full_filter_block.h"
+#include "table/block_based/reader_common.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+// the read buffer size of for the default CacheDumpReader
+const unsigned int kDumpReaderBufferSize = 1024;  // 1KB
+static const unsigned int kSizePrefixLen = 4;
+
+enum CacheDumpUnitType : unsigned char {
+  kHeader = 1,
+  kFooter = 2,
+  kData = 3,
+  kFilter = 4,
+  kProperties = 5,
+  kCompressionDictionary = 6,
+  kRangeDeletion = 7,
+  kHashIndexPrefixes = 8,
+  kHashIndexMetadata = 9,
+  kMetaIndex = 10,
+  kIndex = 11,
+  kDeprecatedFilterBlock = 12,  // OBSOLETE / DEPRECATED
+  kFilterMetaBlock = 13,
+  kBlockTypeMax,
+};
+
+// The metadata of a dump unit. After it is serilized, its size is fixed 16
+// bytes.
+struct DumpUnitMeta {
+  // sequence number is a monotonically increasing number to indicate the order
+  // of the blocks being written. Header is 0.
+  uint32_t sequence_num;
+  // The Crc32c checksum of its dump unit.
+  uint32_t dump_unit_checksum;
+  // The dump unit size after the dump unit is serilized to a string.
+  uint64_t dump_unit_size;
+
+  void reset() {
+    sequence_num = 0;
+    dump_unit_checksum = 0;
+    dump_unit_size = 0;
+  }
+};
+
+// The data structure to hold a block and its information.
+struct DumpUnit {
+  // The timestamp when the block is identified, copied, and dumped from block
+  // cache
+  uint64_t timestamp;
+  // The type of the block
+  CacheDumpUnitType type;
+  // The key of this block when the block is referenced by this Cache
+  Slice key;
+  // The block size
+  size_t value_len;
+  // The Crc32c checksum of the block
+  uint32_t value_checksum;
+  // Pointer to the block. Note that, in the dump process, it points to a memory
+  // buffer copied from cache block. The buffer is freed when we process the
+  // next block. In the load process, we use an std::string to store the
+  // serialized dump_unit read from the reader. So it points to the memory
+  // address of the begin of the block in this string.
+  void* value;
+
+  DumpUnit() { reset(); }
+
+  void reset() {
+    timestamp = 0;
+    type = CacheDumpUnitType::kBlockTypeMax;
+    key.clear();
+    value_len = 0;
+    value_checksum = 0;
+    value = nullptr;
+  }
+};
+
+// The default implementation of the Cache Dumper
+class CacheDumperImpl : public CacheDumper {
+ public:
+  CacheDumperImpl(const CacheDumpOptions& dump_options,
+                  const std::shared_ptr<Cache>& cache,
+                  std::unique_ptr<CacheDumpWriter>&& writer)
+      : options_(dump_options), cache_(cache), writer_(std::move(writer)) {}
+  ~CacheDumperImpl() { writer_.reset(); }
+  Status SetDumpFilter(std::vector<DB*> db_list) override;
+  IOStatus DumpCacheEntriesToWriter() override;
+
+ private:
+  IOStatus WriteBlock(CacheDumpUnitType type, const Slice& key,
+                      const Slice& value);
+  IOStatus WriteHeader();
+  IOStatus WriteFooter();
+  bool ShouldFilterOut(const Slice& key);
+  std::function<void(const Slice&, void*, size_t, Cache::DeleterFn)>
+  DumpOneBlockCallBack();
+
+  CacheDumpOptions options_;
+  std::shared_ptr<Cache> cache_;
+  std::unique_ptr<CacheDumpWriter> writer_;
+  UnorderedMap<Cache::DeleterFn, CacheEntryRole> role_map_;
+  SystemClock* clock_;
+  uint32_t sequence_num_;
+  // The cache key prefix filter. Currently, we use db_session_id as the prefix,
+  // so using std::set to store the prefixes as filter is enough. Further
+  // improvement can be applied like BloomFilter or others to speedup the
+  // filtering.
+  std::set<std::string> prefix_filter_;
+};
+
+// The default implementation of CacheDumpedLoader
+class CacheDumpedLoaderImpl : public CacheDumpedLoader {
+ public:
+  CacheDumpedLoaderImpl(const CacheDumpOptions& dump_options,
+                        const BlockBasedTableOptions& /*toptions*/,
+                        const std::shared_ptr<SecondaryCache>& secondary_cache,
+                        std::unique_ptr<CacheDumpReader>&& reader)
+      : options_(dump_options),
+        secondary_cache_(secondary_cache),
+        reader_(std::move(reader)) {}
+  ~CacheDumpedLoaderImpl() {}
+  IOStatus RestoreCacheEntriesToSecondaryCache() override;
+
+ private:
+  IOStatus ReadDumpUnitMeta(std::string* data, DumpUnitMeta* unit_meta);
+  IOStatus ReadDumpUnit(size_t len, std::string* data, DumpUnit* unit);
+  IOStatus ReadHeader(std::string* data, DumpUnit* dump_unit);
+  IOStatus ReadCacheBlock(std::string* data, DumpUnit* dump_unit);
+
+  CacheDumpOptions options_;
+  std::shared_ptr<SecondaryCache> secondary_cache_;
+  std::unique_ptr<CacheDumpReader> reader_;
+  UnorderedMap<Cache::DeleterFn, CacheEntryRole> role_map_;
+};
+
+// The default implementation of CacheDumpWriter. We write the blocks to a file
+// sequentially.
+class ToFileCacheDumpWriter : public CacheDumpWriter {
+ public:
+  explicit ToFileCacheDumpWriter(
+      std::unique_ptr<WritableFileWriter>&& file_writer)
+      : file_writer_(std::move(file_writer)) {}
+
+  ~ToFileCacheDumpWriter() { Close().PermitUncheckedError(); }
+
+  // Write the serialized metadata to the file
+  virtual IOStatus WriteMetadata(const Slice& metadata) override {
+    assert(file_writer_ != nullptr);
+    std::string prefix;
+    PutFixed32(&prefix, static_cast<uint32_t>(metadata.size()));
+    IOStatus io_s = file_writer_->Append(Slice(prefix));
+    if (!io_s.ok()) {
+      return io_s;
+    }
+    io_s = file_writer_->Append(metadata);
+    return io_s;
+  }
+
+  // Write the serialized data to the file
+  virtual IOStatus WritePacket(const Slice& data) override {
+    assert(file_writer_ != nullptr);
+    std::string prefix;
+    PutFixed32(&prefix, static_cast<uint32_t>(data.size()));
+    IOStatus io_s = file_writer_->Append(Slice(prefix));
+    if (!io_s.ok()) {
+      return io_s;
+    }
+    io_s = file_writer_->Append(data);
+    return io_s;
+  }
+
+  // Reset the writer
+  virtual IOStatus Close() override {
+    file_writer_.reset();
+    return IOStatus::OK();
+  }
+
+ private:
+  std::unique_ptr<WritableFileWriter> file_writer_;
+};
+
+// The default implementation of CacheDumpReader. It is implemented based on
+// RandomAccessFileReader. Note that, we keep an internal variable to remember
+// the current offset.
+class FromFileCacheDumpReader : public CacheDumpReader {
+ public:
+  explicit FromFileCacheDumpReader(
+      std::unique_ptr<RandomAccessFileReader>&& reader)
+      : file_reader_(std::move(reader)),
+        offset_(0),
+        buffer_(new char[kDumpReaderBufferSize]) {}
+
+  ~FromFileCacheDumpReader() { delete[] buffer_; }
+
+  virtual IOStatus ReadMetadata(std::string* metadata) override {
+    uint32_t metadata_len = 0;
+    IOStatus io_s = ReadSizePrefix(&metadata_len);
+    if (!io_s.ok()) {
+      return io_s;
+    }
+    return Read(metadata_len, metadata);
+  }
+
+  virtual IOStatus ReadPacket(std::string* data) override {
+    uint32_t data_len = 0;
+    IOStatus io_s = ReadSizePrefix(&data_len);
+    if (!io_s.ok()) {
+      return io_s;
+    }
+    return Read(data_len, data);
+  }
+
+ private:
+  IOStatus ReadSizePrefix(uint32_t* len) {
+    std::string prefix;
+    IOStatus io_s = Read(kSizePrefixLen, &prefix);
+    if (!io_s.ok()) {
+      return io_s;
+    }
+    Slice encoded_slice(prefix);
+    if (!GetFixed32(&encoded_slice, len)) {
+      return IOStatus::Corruption("Decode size prefix string failed");
+    }
+    return IOStatus::OK();
+  }
+
+  IOStatus Read(size_t len, std::string* data) {
+    assert(file_reader_ != nullptr);
+    IOStatus io_s;
+
+    unsigned int bytes_to_read = static_cast<unsigned int>(len);
+    unsigned int to_read = bytes_to_read > kDumpReaderBufferSize
+                               ? kDumpReaderBufferSize
+                               : bytes_to_read;
+
+    while (to_read > 0) {
+      io_s = file_reader_->Read(IOOptions(), offset_, to_read, &result_,
+                                buffer_, nullptr,
+                                Env::IO_TOTAL /* rate_limiter_priority */);
+      if (!io_s.ok()) {
+        return io_s;
+      }
+      if (result_.size() < to_read) {
+        return IOStatus::Corruption("Corrupted cache dump file.");
+      }
+      data->append(result_.data(), result_.size());
+
+      offset_ += to_read;
+      bytes_to_read -= to_read;
+      to_read = bytes_to_read > kDumpReaderBufferSize ? kDumpReaderBufferSize
+                                                      : bytes_to_read;
+    }
+    return io_s;
+  }
+  std::unique_ptr<RandomAccessFileReader> file_reader_;
+  Slice result_;
+  size_t offset_;
+  char* buffer_;
+};
+
+// The cache dump and load helper class
+class CacheDumperHelper {
+ public:
+  // serialize the dump_unit_meta to a string, it is fixed 16 bytes size.
+  static void EncodeDumpUnitMeta(const DumpUnitMeta& meta, std::string* data) {
+    assert(data);
+    PutFixed32(data, static_cast<uint32_t>(meta.sequence_num));
+    PutFixed32(data, static_cast<uint32_t>(meta.dump_unit_checksum));
+    PutFixed64(data, meta.dump_unit_size);
+  }
+
+  // Serialize the dump_unit to a string.
+  static void EncodeDumpUnit(const DumpUnit& dump_unit, std::string* data) {
+    assert(data);
+    PutFixed64(data, dump_unit.timestamp);
+    data->push_back(dump_unit.type);
+    PutLengthPrefixedSlice(data, dump_unit.key);
+    PutFixed32(data, static_cast<uint32_t>(dump_unit.value_len));
+    PutFixed32(data, dump_unit.value_checksum);
+    PutLengthPrefixedSlice(data,
+                           Slice((char*)dump_unit.value, dump_unit.value_len));
+  }
+
+  // Deserialize the dump_unit_meta from a string
+  static Status DecodeDumpUnitMeta(const std::string& encoded_data,
+                                   DumpUnitMeta* unit_meta) {
+    assert(unit_meta != nullptr);
+    Slice encoded_slice = Slice(encoded_data);
+    if (!GetFixed32(&encoded_slice, &(unit_meta->sequence_num))) {
+      return Status::Incomplete("Decode dumped unit meta sequence_num failed");
+    }
+    if (!GetFixed32(&encoded_slice, &(unit_meta->dump_unit_checksum))) {
+      return Status::Incomplete(
+          "Decode dumped unit meta dump_unit_checksum failed");
+    }
+    if (!GetFixed64(&encoded_slice, &(unit_meta->dump_unit_size))) {
+      return Status::Incomplete(
+          "Decode dumped unit meta dump_unit_size failed");
+    }
+    return Status::OK();
+  }
+
+  // Deserialize the dump_unit from a string.
+  static Status DecodeDumpUnit(const std::string& encoded_data,
+                               DumpUnit* dump_unit) {
+    assert(dump_unit != nullptr);
+    Slice encoded_slice = Slice(encoded_data);
+
+    // Decode timestamp
+    if (!GetFixed64(&encoded_slice, &dump_unit->timestamp)) {
+      return Status::Incomplete("Decode dumped unit string failed");
+    }
+    // Decode the block type
+    dump_unit->type = static_cast<CacheDumpUnitType>(encoded_slice[0]);
+    encoded_slice.remove_prefix(1);
+    // Decode the key
+    if (!GetLengthPrefixedSlice(&encoded_slice, &(dump_unit->key))) {
+      return Status::Incomplete("Decode dumped unit string failed");
+    }
+    // Decode the value size
+    uint32_t value_len;
+    if (!GetFixed32(&encoded_slice, &value_len)) {
+      return Status::Incomplete("Decode dumped unit string failed");
+    }
+    dump_unit->value_len = static_cast<size_t>(value_len);
+    // Decode the value checksum
+    if (!GetFixed32(&encoded_slice, &(dump_unit->value_checksum))) {
+      return Status::Incomplete("Decode dumped unit string failed");
+    }
+    // Decode the block content and copy to the memory space whose pointer
+    // will be managed by the cache finally.
+    Slice block;
+    if (!GetLengthPrefixedSlice(&encoded_slice, &block)) {
+      return Status::Incomplete("Decode dumped unit string failed");
+    }
+    dump_unit->value = (void*)block.data();
+    assert(block.size() == dump_unit->value_len);
+    return Status::OK();
+  }
+};
+
+}  // namespace ROCKSDB_NAMESPACE
+#endif  // ROCKSDB_LITE
diff --git a/src/rocksdb/utilities/cassandra/cassandra_compaction_filter.cc b/src/rocksdb/utilities/cassandra/cassandra_compaction_filter.cc
new file mode 100644
index 000000000..4e48d63aa
--- /dev/null
+++ b/src/rocksdb/utilities/cassandra/cassandra_compaction_filter.cc
@@ -0,0 +1,110 @@
+// Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#include "utilities/cassandra/cassandra_compaction_filter.h"
+
+#include <string>
+
+#include "rocksdb/slice.h"
+#include "rocksdb/utilities/object_registry.h"
+#include "rocksdb/utilities/options_type.h"
+#include "utilities/cassandra/format.h"
+#include "utilities/cassandra/merge_operator.h"
+
+namespace ROCKSDB_NAMESPACE {
+namespace cassandra {
+static std::unordered_map<std::string, OptionTypeInfo>
+    cassandra_filter_type_info = {
+#ifndef ROCKSDB_LITE
+        {"purge_ttl_on_expiration",
+         {offsetof(struct CassandraOptions, purge_ttl_on_expiration),
+          OptionType::kBoolean, OptionVerificationType::kNormal,
+          OptionTypeFlags::kNone}},
+        {"gc_grace_period_in_seconds",
+         {offsetof(struct CassandraOptions, gc_grace_period_in_seconds),
+          OptionType::kUInt32T, OptionVerificationType::kNormal,
+          OptionTypeFlags::kNone}},
+#endif  // ROCKSDB_LITE
+};
+
+CassandraCompactionFilter::CassandraCompactionFilter(
+    bool purge_ttl_on_expiration, int32_t gc_grace_period_in_seconds)
+    : options_(gc_grace_period_in_seconds, 0, purge_ttl_on_expiration) {
+  RegisterOptions(&options_, &cassandra_filter_type_info);
+}
+
+CompactionFilter::Decision CassandraCompactionFilter::FilterV2(
+    int /*level*/, const Slice& /*key*/, ValueType value_type,
+    const Slice& existing_value, std::string* new_value,
+    std::string* /*skip_until*/) const {
+  bool value_changed = false;
+  RowValue row_value =
+      RowValue::Deserialize(existing_value.data(), existing_value.size());
+  RowValue compacted =
+      options_.purge_ttl_on_expiration
+          ? row_value.RemoveExpiredColumns(&value_changed)
+          : row_value.ConvertExpiredColumnsToTombstones(&value_changed);
+
+  if (value_type == ValueType::kValue) {
+    compacted = compacted.RemoveTombstones(options_.gc_grace_period_in_seconds);
+  }
+
+  if (compacted.Empty()) {
+    return Decision::kRemove;
+  }
+
+  if (value_changed) {
+    compacted.Serialize(new_value);
+    return Decision::kChangeValue;
+  }
+
+  return Decision::kKeep;
+}
+
+CassandraCompactionFilterFactory::CassandraCompactionFilterFactory(
+    bool purge_ttl_on_expiration, int32_t gc_grace_period_in_seconds)
+    : options_(gc_grace_period_in_seconds, 0, purge_ttl_on_expiration) {
+  RegisterOptions(&options_, &cassandra_filter_type_info);
+}
+
+std::unique_ptr<CompactionFilter>
+CassandraCompactionFilterFactory::CreateCompactionFilter(
+    const CompactionFilter::Context&) {
+  std::unique_ptr<CompactionFilter> result(new CassandraCompactionFilter(
+      options_.purge_ttl_on_expiration, options_.gc_grace_period_in_seconds));
+  return result;
+}
+
+#ifndef ROCKSDB_LITE
+int RegisterCassandraObjects(ObjectLibrary& library,
+                             const std::string& /*arg*/) {
+  library.AddFactory<MergeOperator>(
+      CassandraValueMergeOperator::kClassName(),
+      [](const std::string& /*uri*/, std::unique_ptr<MergeOperator>* guard,
+         std::string* /* errmsg */) {
+        guard->reset(new CassandraValueMergeOperator(0));
+        return guard->get();
+      });
+  library.AddFactory<CompactionFilter>(
+      CassandraCompactionFilter::kClassName(),
+      [](const std::string& /*uri*/,
+         std::unique_ptr<CompactionFilter>* /*guard */,
+         std::string* /* errmsg */) {
+        return new CassandraCompactionFilter(false, 0);
+      });
+  library.AddFactory<CompactionFilterFactory>(
+      CassandraCompactionFilterFactory::kClassName(),
+      [](const std::string& /*uri*/,
+         std::unique_ptr<CompactionFilterFactory>* guard,
+         std::string* /* errmsg */) {
+        guard->reset(new CassandraCompactionFilterFactory(false, 0));
+        return guard->get();
+      });
+  size_t num_types;
+  return static_cast<int>(library.GetFactoryCount(&num_types));
+}
+#endif  // ROCKSDB_LITE
+}  // namespace cassandra
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/utilities/cassandra/cassandra_compaction_filter.h b/src/rocksdb/utilities/cassandra/cassandra_compaction_filter.h
new file mode 100644
index 000000000..0325a4c39
--- /dev/null
+++ b/src/rocksdb/utilities/cassandra/cassandra_compaction_filter.h
@@ -0,0 +1,57 @@
+// Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#pragma once
+#include <string>
+
+#include "rocksdb/compaction_filter.h"
+#include "rocksdb/slice.h"
+#include "utilities/cassandra/cassandra_options.h"
+
+namespace ROCKSDB_NAMESPACE {
+namespace cassandra {
+
+/**
+ * Compaction filter for removing expired Cassandra data with ttl.
+ * If option `purge_ttl_on_expiration` is set to true, expired data
+ * will be directly purged. Otherwise expired data will be converted
+ * tombstones first, then be eventally removed after gc grace period.
+ * `purge_ttl_on_expiration` should only be on in the case all the
+ * writes have same ttl setting, otherwise it could bring old data back.
+ *
+ * Compaction filter is also in charge of removing tombstone that has been
+ * promoted to kValue type after serials of merging in compaction.
+ */
+class CassandraCompactionFilter : public CompactionFilter {
+ public:
+  explicit CassandraCompactionFilter(bool purge_ttl_on_expiration,
+                                     int32_t gc_grace_period_in_seconds);
+  static const char* kClassName() { return "CassandraCompactionFilter"; }
+  const char* Name() const override { return kClassName(); }
+
+  virtual Decision FilterV2(int level, const Slice& key, ValueType value_type,
+                            const Slice& existing_value, std::string* new_value,
+                            std::string* skip_until) const override;
+
+ private:
+  CassandraOptions options_;
+};
+
+class CassandraCompactionFilterFactory : public CompactionFilterFactory {
+ public:
+  explicit CassandraCompactionFilterFactory(bool purge_ttl_on_expiration,
+                                            int32_t gc_grace_period_in_seconds);
+  ~CassandraCompactionFilterFactory() override {}
+
+  std::unique_ptr<CompactionFilter> CreateCompactionFilter(
+      const CompactionFilter::Context& context) override;
+  static const char* kClassName() { return "CassandraCompactionFilterFactory"; }
+  const char* Name() const override { return kClassName(); }
+
+ private:
+  CassandraOptions options_;
+};
+}  // namespace cassandra
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/utilities/cassandra/cassandra_format_test.cc b/src/rocksdb/utilities/cassandra/cassandra_format_test.cc
new file mode 100644
index 000000000..4f12947ad
--- /dev/null
+++ b/src/rocksdb/utilities/cassandra/cassandra_format_test.cc
@@ -0,0 +1,377 @@
+// Copyright (c) 2017-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#include <cstring>
+#include <memory>
+
+#include "test_util/testharness.h"
+#include "utilities/cassandra/format.h"
+#include "utilities/cassandra/serialize.h"
+#include "utilities/cassandra/test_utils.h"
+
+namespace ROCKSDB_NAMESPACE {
+namespace cassandra {
+
+TEST(ColumnTest, Column) {
+  char data[4] = {'d', 'a', 't', 'a'};
+  int8_t mask = 0;
+  int8_t index = 1;
+  int64_t timestamp = 1494022807044;
+  Column c = Column(mask, index, timestamp, sizeof(data), data);
+
+  EXPECT_EQ(c.Index(), index);
+  EXPECT_EQ(c.Timestamp(), timestamp);
+  EXPECT_EQ(c.Size(), 14 + sizeof(data));
+
+  // Verify the serialization.
+  std::string dest;
+  dest.reserve(c.Size() * 2);
+  c.Serialize(&dest);
+
+  EXPECT_EQ(dest.size(), c.Size());
+  std::size_t offset = 0;
+  EXPECT_EQ(Deserialize<int8_t>(dest.c_str(), offset), mask);
+  offset += sizeof(int8_t);
+  EXPECT_EQ(Deserialize<int8_t>(dest.c_str(), offset), index);
+  offset += sizeof(int8_t);
+  EXPECT_EQ(Deserialize<int64_t>(dest.c_str(), offset), timestamp);
+  offset += sizeof(int64_t);
+  EXPECT_EQ(Deserialize<int32_t>(dest.c_str(), offset), sizeof(data));
+  offset += sizeof(int32_t);
+  EXPECT_TRUE(std::memcmp(data, dest.c_str() + offset, sizeof(data)) == 0);
+
+  // Verify the deserialization.
+  std::string saved_dest = dest;
+  std::shared_ptr<Column> c1 = Column::Deserialize(saved_dest.c_str(), 0);
+  EXPECT_EQ(c1->Index(), index);
+  EXPECT_EQ(c1->Timestamp(), timestamp);
+  EXPECT_EQ(c1->Size(), 14 + sizeof(data));
+
+  c1->Serialize(&dest);
+  EXPECT_EQ(dest.size(), 2 * c.Size());
+  EXPECT_TRUE(std::memcmp(dest.c_str(), dest.c_str() + c.Size(), c.Size()) ==
+              0);
+
+  // Verify the ColumnBase::Deserialization.
+  saved_dest = dest;
+  std::shared_ptr<ColumnBase> c2 =
+      ColumnBase::Deserialize(saved_dest.c_str(), c.Size());
+  c2->Serialize(&dest);
+  EXPECT_EQ(dest.size(), 3 * c.Size());
+  EXPECT_TRUE(std::memcmp(dest.c_str() + c.Size(), dest.c_str() + c.Size() * 2,
+                          c.Size()) == 0);
+}
+
+TEST(ExpiringColumnTest, ExpiringColumn) {
+  char data[4] = {'d', 'a', 't', 'a'};
+  int8_t mask = ColumnTypeMask::EXPIRATION_MASK;
+  int8_t index = 3;
+  int64_t timestamp = 1494022807044;
+  int32_t ttl = 3600;
+  ExpiringColumn c =
+      ExpiringColumn(mask, index, timestamp, sizeof(data), data, ttl);
+
+  EXPECT_EQ(c.Index(), index);
+  EXPECT_EQ(c.Timestamp(), timestamp);
+  EXPECT_EQ(c.Size(), 18 + sizeof(data));
+
+  // Verify the serialization.
+  std::string dest;
+  dest.reserve(c.Size() * 2);
+  c.Serialize(&dest);
+
+  EXPECT_EQ(dest.size(), c.Size());
+  std::size_t offset = 0;
+  EXPECT_EQ(Deserialize<int8_t>(dest.c_str(), offset), mask);
+  offset += sizeof(int8_t);
+  EXPECT_EQ(Deserialize<int8_t>(dest.c_str(), offset), index);
+  offset += sizeof(int8_t);
+  EXPECT_EQ(Deserialize<int64_t>(dest.c_str(), offset), timestamp);
+  offset += sizeof(int64_t);
+  EXPECT_EQ(Deserialize<int32_t>(dest.c_str(), offset), sizeof(data));
+  offset += sizeof(int32_t);
+  EXPECT_TRUE(std::memcmp(data, dest.c_str() + offset, sizeof(data)) == 0);
+  offset += sizeof(data);
+  EXPECT_EQ(Deserialize<int32_t>(dest.c_str(), offset), ttl);
+
+  // Verify the deserialization.
+  std::string saved_dest = dest;
+  std::shared_ptr<ExpiringColumn> c1 =
+      ExpiringColumn::Deserialize(saved_dest.c_str(), 0);
+  EXPECT_EQ(c1->Index(), index);
+  EXPECT_EQ(c1->Timestamp(), timestamp);
+  EXPECT_EQ(c1->Size(), 18 + sizeof(data));
+
+  c1->Serialize(&dest);
+  EXPECT_EQ(dest.size(), 2 * c.Size());
+  EXPECT_TRUE(std::memcmp(dest.c_str(), dest.c_str() + c.Size(), c.Size()) ==
+              0);
+
+  // Verify the ColumnBase::Deserialization.
+  saved_dest = dest;
+  std::shared_ptr<ColumnBase> c2 =
+      ColumnBase::Deserialize(saved_dest.c_str(), c.Size());
+  c2->Serialize(&dest);
+  EXPECT_EQ(dest.size(), 3 * c.Size());
+  EXPECT_TRUE(std::memcmp(dest.c_str() + c.Size(), dest.c_str() + c.Size() * 2,
+                          c.Size()) == 0);
+}
+
+TEST(TombstoneTest, TombstoneCollectable) {
+  int32_t now = (int32_t)time(nullptr);
+  int32_t gc_grace_seconds = 16440;
+  int32_t time_delta_seconds = 10;
+  EXPECT_TRUE(
+      Tombstone(ColumnTypeMask::DELETION_MASK, 0,
+                now - gc_grace_seconds - time_delta_seconds,
+                ToMicroSeconds(now - gc_grace_seconds - time_delta_seconds))
+          .Collectable(gc_grace_seconds));
+  EXPECT_FALSE(
+      Tombstone(ColumnTypeMask::DELETION_MASK, 0,
+                now - gc_grace_seconds + time_delta_seconds,
+                ToMicroSeconds(now - gc_grace_seconds + time_delta_seconds))
+          .Collectable(gc_grace_seconds));
+}
+
+TEST(TombstoneTest, Tombstone) {
+  int8_t mask = ColumnTypeMask::DELETION_MASK;
+  int8_t index = 2;
+  int32_t local_deletion_time = 1494022807;
+  int64_t marked_for_delete_at = 1494022807044;
+  Tombstone c =
+      Tombstone(mask, index, local_deletion_time, marked_for_delete_at);
+
+  EXPECT_EQ(c.Index(), index);
+  EXPECT_EQ(c.Timestamp(), marked_for_delete_at);
+  EXPECT_EQ(c.Size(), 14);
+
+  // Verify the serialization.
+  std::string dest;
+  dest.reserve(c.Size() * 2);
+  c.Serialize(&dest);
+
+  EXPECT_EQ(dest.size(), c.Size());
+  std::size_t offset = 0;
+  EXPECT_EQ(Deserialize<int8_t>(dest.c_str(), offset), mask);
+  offset += sizeof(int8_t);
+  EXPECT_EQ(Deserialize<int8_t>(dest.c_str(), offset), index);
+  offset += sizeof(int8_t);
+  EXPECT_EQ(Deserialize<int32_t>(dest.c_str(), offset), local_deletion_time);
+  offset += sizeof(int32_t);
+  EXPECT_EQ(Deserialize<int64_t>(dest.c_str(), offset), marked_for_delete_at);
+
+  // Verify the deserialization.
+  std::shared_ptr<Tombstone> c1 = Tombstone::Deserialize(dest.c_str(), 0);
+  EXPECT_EQ(c1->Index(), index);
+  EXPECT_EQ(c1->Timestamp(), marked_for_delete_at);
+  EXPECT_EQ(c1->Size(), 14);
+
+  c1->Serialize(&dest);
+  EXPECT_EQ(dest.size(), 2 * c.Size());
+  EXPECT_TRUE(std::memcmp(dest.c_str(), dest.c_str() + c.Size(), c.Size()) ==
+              0);
+
+  // Verify the ColumnBase::Deserialization.
+  std::shared_ptr<ColumnBase> c2 =
+      ColumnBase::Deserialize(dest.c_str(), c.Size());
+  c2->Serialize(&dest);
+  EXPECT_EQ(dest.size(), 3 * c.Size());
+  EXPECT_TRUE(std::memcmp(dest.c_str() + c.Size(), dest.c_str() + c.Size() * 2,
+                          c.Size()) == 0);
+}
+
+class RowValueTest : public testing::Test {};
+
+TEST(RowValueTest, RowTombstone) {
+  int32_t local_deletion_time = 1494022807;
+  int64_t marked_for_delete_at = 1494022807044;
+  RowValue r = RowValue(local_deletion_time, marked_for_delete_at);
+
+  EXPECT_EQ(r.Size(), 12);
+  EXPECT_EQ(r.IsTombstone(), true);
+  EXPECT_EQ(r.LastModifiedTime(), marked_for_delete_at);
+
+  // Verify the serialization.
+  std::string dest;
+  dest.reserve(r.Size() * 2);
+  r.Serialize(&dest);
+
+  EXPECT_EQ(dest.size(), r.Size());
+  std::size_t offset = 0;
+  EXPECT_EQ(Deserialize<int32_t>(dest.c_str(), offset), local_deletion_time);
+  offset += sizeof(int32_t);
+  EXPECT_EQ(Deserialize<int64_t>(dest.c_str(), offset), marked_for_delete_at);
+
+  // Verify the deserialization.
+  RowValue r1 = RowValue::Deserialize(dest.c_str(), r.Size());
+  EXPECT_EQ(r1.Size(), 12);
+  EXPECT_EQ(r1.IsTombstone(), true);
+  EXPECT_EQ(r1.LastModifiedTime(), marked_for_delete_at);
+
+  r1.Serialize(&dest);
+  EXPECT_EQ(dest.size(), 2 * r.Size());
+  EXPECT_TRUE(std::memcmp(dest.c_str(), dest.c_str() + r.Size(), r.Size()) ==
+              0);
+}
+
+TEST(RowValueTest, RowWithColumns) {
+  std::vector<std::shared_ptr<ColumnBase>> columns;
+  int64_t last_modified_time = 1494022807048;
+  std::size_t columns_data_size = 0;
+
+  char e_data[5] = {'e', 'd', 'a', 't', 'a'};
+  int8_t e_index = 0;
+  int64_t e_timestamp = 1494022807044;
+  int32_t e_ttl = 3600;
+  columns.push_back(std::shared_ptr<ExpiringColumn>(
+      new ExpiringColumn(ColumnTypeMask::EXPIRATION_MASK, e_index, e_timestamp,
+                         sizeof(e_data), e_data, e_ttl)));
+  columns_data_size += columns[0]->Size();
+
+  char c_data[4] = {'d', 'a', 't', 'a'};
+  int8_t c_index = 1;
+  int64_t c_timestamp = 1494022807048;
+  columns.push_back(std::shared_ptr<Column>(
+      new Column(0, c_index, c_timestamp, sizeof(c_data), c_data)));
+  columns_data_size += columns[1]->Size();
+
+  int8_t t_index = 2;
+  int32_t t_local_deletion_time = 1494022801;
+  int64_t t_marked_for_delete_at = 1494022807043;
+  columns.push_back(std::shared_ptr<Tombstone>(
+      new Tombstone(ColumnTypeMask::DELETION_MASK, t_index,
+                    t_local_deletion_time, t_marked_for_delete_at)));
+  columns_data_size += columns[2]->Size();
+
+  RowValue r = RowValue(std::move(columns), last_modified_time);
+
+  EXPECT_EQ(r.Size(), columns_data_size + 12);
+  EXPECT_EQ(r.IsTombstone(), false);
+  EXPECT_EQ(r.LastModifiedTime(), last_modified_time);
+
+  // Verify the serialization.
+  std::string dest;
+  dest.reserve(r.Size() * 2);
+  r.Serialize(&dest);
+
+  EXPECT_EQ(dest.size(), r.Size());
+  std::size_t offset = 0;
+  EXPECT_EQ(Deserialize<int32_t>(dest.c_str(), offset),
+            std::numeric_limits<int32_t>::max());
+  offset += sizeof(int32_t);
+  EXPECT_EQ(Deserialize<int64_t>(dest.c_str(), offset),
+            std::numeric_limits<int64_t>::min());
+  offset += sizeof(int64_t);
+
+  // Column0: ExpiringColumn
+  EXPECT_EQ(Deserialize<int8_t>(dest.c_str(), offset),
+            ColumnTypeMask::EXPIRATION_MASK);
+  offset += sizeof(int8_t);
+  EXPECT_EQ(Deserialize<int8_t>(dest.c_str(), offset), e_index);
+  offset += sizeof(int8_t);
+  EXPECT_EQ(Deserialize<int64_t>(dest.c_str(), offset), e_timestamp);
+  offset += sizeof(int64_t);
+  EXPECT_EQ(Deserialize<int32_t>(dest.c_str(), offset), sizeof(e_data));
+  offset += sizeof(int32_t);
+  EXPECT_TRUE(std::memcmp(e_data, dest.c_str() + offset, sizeof(e_data)) == 0);
+  offset += sizeof(e_data);
+  EXPECT_EQ(Deserialize<int32_t>(dest.c_str(), offset), e_ttl);
+  offset += sizeof(int32_t);
+
+  // Column1: Column
+  EXPECT_EQ(Deserialize<int8_t>(dest.c_str(), offset), 0);
+  offset += sizeof(int8_t);
+  EXPECT_EQ(Deserialize<int8_t>(dest.c_str(), offset), c_index);
+  offset += sizeof(int8_t);
+  EXPECT_EQ(Deserialize<int64_t>(dest.c_str(), offset), c_timestamp);
+  offset += sizeof(int64_t);
+  EXPECT_EQ(Deserialize<int32_t>(dest.c_str(), offset), sizeof(c_data));
+  offset += sizeof(int32_t);
+  EXPECT_TRUE(std::memcmp(c_data, dest.c_str() + offset, sizeof(c_data)) == 0);
+  offset += sizeof(c_data);
+
+  // Column2: Tombstone
+  EXPECT_EQ(Deserialize<int8_t>(dest.c_str(), offset),
+            ColumnTypeMask::DELETION_MASK);
+  offset += sizeof(int8_t);
+  EXPECT_EQ(Deserialize<int8_t>(dest.c_str(), offset), t_index);
+  offset += sizeof(int8_t);
+  EXPECT_EQ(Deserialize<int32_t>(dest.c_str(), offset), t_local_deletion_time);
+  offset += sizeof(int32_t);
+  EXPECT_EQ(Deserialize<int64_t>(dest.c_str(), offset), t_marked_for_delete_at);
+
+  // Verify the deserialization.
+  RowValue r1 = RowValue::Deserialize(dest.c_str(), r.Size());
+  EXPECT_EQ(r1.Size(), columns_data_size + 12);
+  EXPECT_EQ(r1.IsTombstone(), false);
+  EXPECT_EQ(r1.LastModifiedTime(), last_modified_time);
+
+  r1.Serialize(&dest);
+  EXPECT_EQ(dest.size(), 2 * r.Size());
+  EXPECT_TRUE(std::memcmp(dest.c_str(), dest.c_str() + r.Size(), r.Size()) ==
+              0);
+}
+
+TEST(RowValueTest, PurgeTtlShouldRemvoeAllColumnsExpired) {
+  int64_t now = time(nullptr);
+
+  auto row_value = CreateTestRowValue(
+      {CreateTestColumnSpec(kColumn, 0, ToMicroSeconds(now)),
+       CreateTestColumnSpec(kExpiringColumn, 1,
+                            ToMicroSeconds(now - kTtl - 10)),  // expired
+       CreateTestColumnSpec(kExpiringColumn, 2,
+                            ToMicroSeconds(now)),  // not expired
+       CreateTestColumnSpec(kTombstone, 3, ToMicroSeconds(now))});
+
+  bool changed = false;
+  auto purged = row_value.RemoveExpiredColumns(&changed);
+  EXPECT_TRUE(changed);
+  EXPECT_EQ(purged.get_columns().size(), 3);
+  VerifyRowValueColumns(purged.get_columns(), 0, kColumn, 0,
+                        ToMicroSeconds(now));
+  VerifyRowValueColumns(purged.get_columns(), 1, kExpiringColumn, 2,
+                        ToMicroSeconds(now));
+  VerifyRowValueColumns(purged.get_columns(), 2, kTombstone, 3,
+                        ToMicroSeconds(now));
+
+  purged.RemoveExpiredColumns(&changed);
+  EXPECT_FALSE(changed);
+}
+
+TEST(RowValueTest, ExpireTtlShouldConvertExpiredColumnsToTombstones) {
+  int64_t now = time(nullptr);
+
+  auto row_value = CreateTestRowValue(
+      {CreateTestColumnSpec(kColumn, 0, ToMicroSeconds(now)),
+       CreateTestColumnSpec(kExpiringColumn, 1,
+                            ToMicroSeconds(now - kTtl - 10)),  // expired
+       CreateTestColumnSpec(kExpiringColumn, 2,
+                            ToMicroSeconds(now)),  // not expired
+       CreateTestColumnSpec(kTombstone, 3, ToMicroSeconds(now))});
+
+  bool changed = false;
+  auto compacted = row_value.ConvertExpiredColumnsToTombstones(&changed);
+  EXPECT_TRUE(changed);
+  EXPECT_EQ(compacted.get_columns().size(), 4);
+  VerifyRowValueColumns(compacted.get_columns(), 0, kColumn, 0,
+                        ToMicroSeconds(now));
+  VerifyRowValueColumns(compacted.get_columns(), 1, kTombstone, 1,
+                        ToMicroSeconds(now - 10));
+  VerifyRowValueColumns(compacted.get_columns(), 2, kExpiringColumn, 2,
+                        ToMicroSeconds(now));
+  VerifyRowValueColumns(compacted.get_columns(), 3, kTombstone, 3,
+                        ToMicroSeconds(now));
+
+  compacted.ConvertExpiredColumnsToTombstones(&changed);
+  EXPECT_FALSE(changed);
+}
+}  // namespace cassandra
+}  // namespace ROCKSDB_NAMESPACE
+
+int main(int argc, char** argv) {
+  ROCKSDB_NAMESPACE::port::InstallStackTraceHandler();
+  ::testing::InitGoogleTest(&argc, argv);
+  return RUN_ALL_TESTS();
+}
diff --git a/src/rocksdb/utilities/cassandra/cassandra_functional_test.cc b/src/rocksdb/utilities/cassandra/cassandra_functional_test.cc
new file mode 100644
index 000000000..c5be836e8
--- /dev/null
+++ b/src/rocksdb/utilities/cassandra/cassandra_functional_test.cc
@@ -0,0 +1,446 @@
+// Copyright (c) 2017-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#include <iostream>
+
+#include "db/db_impl/db_impl.h"
+#include "rocksdb/convenience.h"
+#include "rocksdb/db.h"
+#include "rocksdb/merge_operator.h"
+#include "rocksdb/utilities/object_registry.h"
+#include "test_util/testharness.h"
+#include "util/cast_util.h"
+#include "util/random.h"
+#include "utilities/cassandra/cassandra_compaction_filter.h"
+#include "utilities/cassandra/merge_operator.h"
+#include "utilities/cassandra/test_utils.h"
+#include "utilities/merge_operators.h"
+
+namespace ROCKSDB_NAMESPACE {
+namespace cassandra {
+
+// Path to the database on file system
+const std::string kDbName = test::PerThreadDBPath("cassandra_functional_test");
+
+class CassandraStore {
+ public:
+  explicit CassandraStore(std::shared_ptr<DB> db)
+      : db_(db), write_option_(), get_option_() {
+    assert(db);
+  }
+
+  bool Append(const std::string& key, const RowValue& val) {
+    std::string result;
+    val.Serialize(&result);
+    Slice valSlice(result.data(), result.size());
+    auto s = db_->Merge(write_option_, key, valSlice);
+
+    if (s.ok()) {
+      return true;
+    } else {
+      std::cerr << "ERROR " << s.ToString() << std::endl;
+      return false;
+    }
+  }
+
+  bool Put(const std::string& key, const RowValue& val) {
+    std::string result;
+    val.Serialize(&result);
+    Slice valSlice(result.data(), result.size());
+    auto s = db_->Put(write_option_, key, valSlice);
+    if (s.ok()) {
+      return true;
+    } else {
+      std::cerr << "ERROR " << s.ToString() << std::endl;
+      return false;
+    }
+  }
+
+  Status Flush() {
+    Status s = dbfull()->TEST_FlushMemTable();
+    if (s.ok()) {
+      s = dbfull()->TEST_WaitForCompact();
+    }
+    return s;
+  }
+
+  Status Compact() {
+    return dbfull()->TEST_CompactRange(0, nullptr, nullptr,
+                                       db_->DefaultColumnFamily());
+  }
+
+  std::tuple<bool, RowValue> Get(const std::string& key) {
+    std::string result;
+    auto s = db_->Get(get_option_, key, &result);
+
+    if (s.ok()) {
+      return std::make_tuple(
+          true, RowValue::Deserialize(result.data(), result.size()));
+    }
+
+    if (!s.IsNotFound()) {
+      std::cerr << "ERROR " << s.ToString() << std::endl;
+    }
+
+    return std::make_tuple(false, RowValue(0, 0));
+  }
+
+ private:
+  std::shared_ptr<DB> db_;
+  WriteOptions write_option_;
+  ReadOptions get_option_;
+
+  DBImpl* dbfull() { return static_cast_with_check<DBImpl>(db_.get()); }
+};
+
+class TestCompactionFilterFactory : public CompactionFilterFactory {
+ public:
+  explicit TestCompactionFilterFactory(bool purge_ttl_on_expiration,
+                                       int32_t gc_grace_period_in_seconds)
+      : purge_ttl_on_expiration_(purge_ttl_on_expiration),
+        gc_grace_period_in_seconds_(gc_grace_period_in_seconds) {}
+
+  std::unique_ptr<CompactionFilter> CreateCompactionFilter(
+      const CompactionFilter::Context& /*context*/) override {
+    return std::unique_ptr<CompactionFilter>(new CassandraCompactionFilter(
+        purge_ttl_on_expiration_, gc_grace_period_in_seconds_));
+  }
+
+  const char* Name() const override { return "TestCompactionFilterFactory"; }
+
+ private:
+  bool purge_ttl_on_expiration_;
+  int32_t gc_grace_period_in_seconds_;
+};
+
+// The class for unit-testing
+class CassandraFunctionalTest : public testing::Test {
+ public:
+  CassandraFunctionalTest() {
+    EXPECT_OK(
+        DestroyDB(kDbName, Options()));  // Start each test with a fresh DB
+  }
+
+  std::shared_ptr<DB> OpenDb() {
+    DB* db;
+    Options options;
+    options.create_if_missing = true;
+    options.merge_operator.reset(
+        new CassandraValueMergeOperator(gc_grace_period_in_seconds_));
+    auto* cf_factory = new TestCompactionFilterFactory(
+        purge_ttl_on_expiration_, gc_grace_period_in_seconds_);
+    options.compaction_filter_factory.reset(cf_factory);
+    EXPECT_OK(DB::Open(options, kDbName, &db));
+    return std::shared_ptr<DB>(db);
+  }
+
+  bool purge_ttl_on_expiration_ = false;
+  int32_t gc_grace_period_in_seconds_ = 100;
+};
+
+// THE TEST CASES BEGIN HERE
+
+TEST_F(CassandraFunctionalTest, SimpleMergeTest) {
+  CassandraStore store(OpenDb());
+  int64_t now = time(nullptr);
+
+  store.Append(
+      "k1",
+      CreateTestRowValue({
+          CreateTestColumnSpec(kTombstone, 0, ToMicroSeconds(now + 5)),
+          CreateTestColumnSpec(kColumn, 1, ToMicroSeconds(now + 8)),
+          CreateTestColumnSpec(kExpiringColumn, 2, ToMicroSeconds(now + 5)),
+      }));
+  store.Append(
+      "k1",
+      CreateTestRowValue({
+          CreateTestColumnSpec(kColumn, 0, ToMicroSeconds(now + 2)),
+          CreateTestColumnSpec(kExpiringColumn, 1, ToMicroSeconds(now + 5)),
+          CreateTestColumnSpec(kTombstone, 2, ToMicroSeconds(now + 7)),
+          CreateTestColumnSpec(kExpiringColumn, 7, ToMicroSeconds(now + 17)),
+      }));
+  store.Append(
+      "k1",
+      CreateTestRowValue({
+          CreateTestColumnSpec(kExpiringColumn, 0, ToMicroSeconds(now + 6)),
+          CreateTestColumnSpec(kTombstone, 1, ToMicroSeconds(now + 5)),
+          CreateTestColumnSpec(kColumn, 2, ToMicroSeconds(now + 4)),
+          CreateTestColumnSpec(kTombstone, 11, ToMicroSeconds(now + 11)),
+      }));
+
+  auto ret = store.Get("k1");
+
+  ASSERT_TRUE(std::get<0>(ret));
+  RowValue& merged = std::get<1>(ret);
+  EXPECT_EQ(merged.get_columns().size(), 5);
+  VerifyRowValueColumns(merged.get_columns(), 0, kExpiringColumn, 0,
+                        ToMicroSeconds(now + 6));
+  VerifyRowValueColumns(merged.get_columns(), 1, kColumn, 1,
+                        ToMicroSeconds(now + 8));
+  VerifyRowValueColumns(merged.get_columns(), 2, kTombstone, 2,
+                        ToMicroSeconds(now + 7));
+  VerifyRowValueColumns(merged.get_columns(), 3, kExpiringColumn, 7,
+                        ToMicroSeconds(now + 17));
+  VerifyRowValueColumns(merged.get_columns(), 4, kTombstone, 11,
+                        ToMicroSeconds(now + 11));
+}
+
+constexpr int64_t kTestTimeoutSecs = 600;
+
+TEST_F(CassandraFunctionalTest,
+       CompactionShouldConvertExpiredColumnsToTombstone) {
+  CassandraStore store(OpenDb());
+  int64_t now = time(nullptr);
+
+  store.Append(
+      "k1",
+      CreateTestRowValue(
+          {CreateTestColumnSpec(kExpiringColumn, 0,
+                                ToMicroSeconds(now - kTtl - 20)),  // expired
+           CreateTestColumnSpec(
+               kExpiringColumn, 1,
+               ToMicroSeconds(now - kTtl + kTestTimeoutSecs)),  // not expired
+           CreateTestColumnSpec(kTombstone, 3, ToMicroSeconds(now))}));
+
+  ASSERT_OK(store.Flush());
+
+  store.Append(
+      "k1",
+      CreateTestRowValue(
+          {CreateTestColumnSpec(kExpiringColumn, 0,
+                                ToMicroSeconds(now - kTtl - 10)),  // expired
+           CreateTestColumnSpec(kColumn, 2, ToMicroSeconds(now))}));
+
+  ASSERT_OK(store.Flush());
+  ASSERT_OK(store.Compact());
+
+  auto ret = store.Get("k1");
+  ASSERT_TRUE(std::get<0>(ret));
+  RowValue& merged = std::get<1>(ret);
+  EXPECT_EQ(merged.get_columns().size(), 4);
+  VerifyRowValueColumns(merged.get_columns(), 0, kTombstone, 0,
+                        ToMicroSeconds(now - 10));
+  VerifyRowValueColumns(merged.get_columns(), 1, kExpiringColumn, 1,
+                        ToMicroSeconds(now - kTtl + kTestTimeoutSecs));
+  VerifyRowValueColumns(merged.get_columns(), 2, kColumn, 2,
+                        ToMicroSeconds(now));
+  VerifyRowValueColumns(merged.get_columns(), 3, kTombstone, 3,
+                        ToMicroSeconds(now));
+}
+
+TEST_F(CassandraFunctionalTest,
+       CompactionShouldPurgeExpiredColumnsIfPurgeTtlIsOn) {
+  purge_ttl_on_expiration_ = true;
+  CassandraStore store(OpenDb());
+  int64_t now = time(nullptr);
+
+  store.Append(
+      "k1",
+      CreateTestRowValue(
+          {CreateTestColumnSpec(kExpiringColumn, 0,
+                                ToMicroSeconds(now - kTtl - 20)),  // expired
+           CreateTestColumnSpec(kExpiringColumn, 1,
+                                ToMicroSeconds(now)),  // not expired
+           CreateTestColumnSpec(kTombstone, 3, ToMicroSeconds(now))}));
+
+  ASSERT_OK(store.Flush());
+
+  store.Append(
+      "k1",
+      CreateTestRowValue(
+          {CreateTestColumnSpec(kExpiringColumn, 0,
+                                ToMicroSeconds(now - kTtl - 10)),  // expired
+           CreateTestColumnSpec(kColumn, 2, ToMicroSeconds(now))}));
+
+  ASSERT_OK(store.Flush());
+  ASSERT_OK(store.Compact());
+
+  auto ret = store.Get("k1");
+  ASSERT_TRUE(std::get<0>(ret));
+  RowValue& merged = std::get<1>(ret);
+  EXPECT_EQ(merged.get_columns().size(), 3);
+  VerifyRowValueColumns(merged.get_columns(), 0, kExpiringColumn, 1,
+                        ToMicroSeconds(now));
+  VerifyRowValueColumns(merged.get_columns(), 1, kColumn, 2,
+                        ToMicroSeconds(now));
+  VerifyRowValueColumns(merged.get_columns(), 2, kTombstone, 3,
+                        ToMicroSeconds(now));
+}
+
+TEST_F(CassandraFunctionalTest,
+       CompactionShouldRemoveRowWhenAllColumnsExpiredIfPurgeTtlIsOn) {
+  purge_ttl_on_expiration_ = true;
+  CassandraStore store(OpenDb());
+  int64_t now = time(nullptr);
+
+  store.Append("k1", CreateTestRowValue({
+                         CreateTestColumnSpec(kExpiringColumn, 0,
+                                              ToMicroSeconds(now - kTtl - 20)),
+                         CreateTestColumnSpec(kExpiringColumn, 1,
+                                              ToMicroSeconds(now - kTtl - 20)),
+                     }));
+
+  ASSERT_OK(store.Flush());
+
+  store.Append("k1", CreateTestRowValue({
+                         CreateTestColumnSpec(kExpiringColumn, 0,
+                                              ToMicroSeconds(now - kTtl - 10)),
+                     }));
+
+  ASSERT_OK(store.Flush());
+  ASSERT_OK(store.Compact());
+  ASSERT_FALSE(std::get<0>(store.Get("k1")));
+}
+
+TEST_F(CassandraFunctionalTest,
+       CompactionShouldRemoveTombstoneExceedingGCGracePeriod) {
+  purge_ttl_on_expiration_ = true;
+  CassandraStore store(OpenDb());
+  int64_t now = time(nullptr);
+
+  store.Append("k1",
+               CreateTestRowValue(
+                   {CreateTestColumnSpec(
+                        kTombstone, 0,
+                        ToMicroSeconds(now - gc_grace_period_in_seconds_ - 1)),
+                    CreateTestColumnSpec(kColumn, 1, ToMicroSeconds(now))}));
+
+  store.Append("k2", CreateTestRowValue({CreateTestColumnSpec(
+                         kColumn, 0, ToMicroSeconds(now))}));
+
+  ASSERT_OK(store.Flush());
+
+  store.Append("k1", CreateTestRowValue({
+                         CreateTestColumnSpec(kColumn, 1, ToMicroSeconds(now)),
+                     }));
+
+  ASSERT_OK(store.Flush());
+  ASSERT_OK(store.Compact());
+
+  auto ret = store.Get("k1");
+  ASSERT_TRUE(std::get<0>(ret));
+  RowValue& gced = std::get<1>(ret);
+  EXPECT_EQ(gced.get_columns().size(), 1);
+  VerifyRowValueColumns(gced.get_columns(), 0, kColumn, 1, ToMicroSeconds(now));
+}
+
+TEST_F(CassandraFunctionalTest, CompactionShouldRemoveTombstoneFromPut) {
+  purge_ttl_on_expiration_ = true;
+  CassandraStore store(OpenDb());
+  int64_t now = time(nullptr);
+
+  store.Put("k1",
+            CreateTestRowValue({
+                CreateTestColumnSpec(
+                    kTombstone, 0,
+                    ToMicroSeconds(now - gc_grace_period_in_seconds_ - 1)),
+            }));
+
+  ASSERT_OK(store.Flush());
+  ASSERT_OK(store.Compact());
+  ASSERT_FALSE(std::get<0>(store.Get("k1")));
+}
+
+#ifndef ROCKSDB_LITE
+TEST_F(CassandraFunctionalTest, LoadMergeOperator) {
+  ConfigOptions config_options;
+  std::shared_ptr<MergeOperator> mo;
+  config_options.ignore_unsupported_options = false;
+
+  ASSERT_NOK(MergeOperator::CreateFromString(
+      config_options, CassandraValueMergeOperator::kClassName(), &mo));
+
+  config_options.registry->AddLibrary("cassandra", RegisterCassandraObjects,
+                                      "cassandra");
+
+  ASSERT_OK(MergeOperator::CreateFromString(
+      config_options, CassandraValueMergeOperator::kClassName(), &mo));
+  ASSERT_NE(mo, nullptr);
+  ASSERT_STREQ(mo->Name(), CassandraValueMergeOperator::kClassName());
+  mo.reset();
+  ASSERT_OK(MergeOperator::CreateFromString(
+      config_options,
+      std::string("operands_limit=20;gc_grace_period_in_seconds=42;id=") +
+          CassandraValueMergeOperator::kClassName(),
+      &mo));
+  ASSERT_NE(mo, nullptr);
+  ASSERT_STREQ(mo->Name(), CassandraValueMergeOperator::kClassName());
+  const auto* opts = mo->GetOptions<CassandraOptions>();
+  ASSERT_NE(opts, nullptr);
+  ASSERT_EQ(opts->gc_grace_period_in_seconds, 42);
+  ASSERT_EQ(opts->operands_limit, 20);
+}
+
+TEST_F(CassandraFunctionalTest, LoadCompactionFilter) {
+  ConfigOptions config_options;
+  const CompactionFilter* filter = nullptr;
+  config_options.ignore_unsupported_options = false;
+
+  ASSERT_NOK(CompactionFilter::CreateFromString(
+      config_options, CassandraCompactionFilter::kClassName(), &filter));
+  config_options.registry->AddLibrary("cassandra", RegisterCassandraObjects,
+                                      "cassandra");
+
+  ASSERT_OK(CompactionFilter::CreateFromString(
+      config_options, CassandraCompactionFilter::kClassName(), &filter));
+  ASSERT_NE(filter, nullptr);
+  ASSERT_STREQ(filter->Name(), CassandraCompactionFilter::kClassName());
+  delete filter;
+  filter = nullptr;
+  ASSERT_OK(CompactionFilter::CreateFromString(
+      config_options,
+      std::string(
+          "purge_ttl_on_expiration=true;gc_grace_period_in_seconds=42;id=") +
+          CassandraCompactionFilter::kClassName(),
+      &filter));
+  ASSERT_NE(filter, nullptr);
+  ASSERT_STREQ(filter->Name(), CassandraCompactionFilter::kClassName());
+  const auto* opts = filter->GetOptions<CassandraOptions>();
+  ASSERT_NE(opts, nullptr);
+  ASSERT_EQ(opts->gc_grace_period_in_seconds, 42);
+  ASSERT_TRUE(opts->purge_ttl_on_expiration);
+  delete filter;
+}
+
+TEST_F(CassandraFunctionalTest, LoadCompactionFilterFactory) {
+  ConfigOptions config_options;
+  std::shared_ptr<CompactionFilterFactory> factory;
+
+  config_options.ignore_unsupported_options = false;
+  ASSERT_NOK(CompactionFilterFactory::CreateFromString(
+      config_options, CassandraCompactionFilterFactory::kClassName(),
+      &factory));
+  config_options.registry->AddLibrary("cassandra", RegisterCassandraObjects,
+                                      "cassandra");
+
+  ASSERT_OK(CompactionFilterFactory::CreateFromString(
+      config_options, CassandraCompactionFilterFactory::kClassName(),
+      &factory));
+  ASSERT_NE(factory, nullptr);
+  ASSERT_STREQ(factory->Name(), CassandraCompactionFilterFactory::kClassName());
+  factory.reset();
+  ASSERT_OK(CompactionFilterFactory::CreateFromString(
+      config_options,
+      std::string(
+          "purge_ttl_on_expiration=true;gc_grace_period_in_seconds=42;id=") +
+          CassandraCompactionFilterFactory::kClassName(),
+      &factory));
+  ASSERT_NE(factory, nullptr);
+  ASSERT_STREQ(factory->Name(), CassandraCompactionFilterFactory::kClassName());
+  const auto* opts = factory->GetOptions<CassandraOptions>();
+  ASSERT_NE(opts, nullptr);
+  ASSERT_EQ(opts->gc_grace_period_in_seconds, 42);
+  ASSERT_TRUE(opts->purge_ttl_on_expiration);
+}
+#endif  // ROCKSDB_LITE
+
+}  // namespace cassandra
+}  // namespace ROCKSDB_NAMESPACE
+
+int main(int argc, char** argv) {
+  ROCKSDB_NAMESPACE::port::InstallStackTraceHandler();
+  ::testing::InitGoogleTest(&argc, argv);
+  return RUN_ALL_TESTS();
+}
diff --git a/src/rocksdb/utilities/cassandra/cassandra_options.h b/src/rocksdb/utilities/cassandra/cassandra_options.h
new file mode 100644
index 000000000..efa73a308
--- /dev/null
+++ b/src/rocksdb/utilities/cassandra/cassandra_options.h
@@ -0,0 +1,43 @@
+// Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#pragma once
+
+#include <cstddef>
+#include <cstdint>
+#include <string>
+
+#include "rocksdb/rocksdb_namespace.h"
+
+namespace ROCKSDB_NAMESPACE {
+class ObjectLibrary;
+namespace cassandra {
+struct CassandraOptions {
+  static const char* kName() { return "CassandraOptions"; }
+  CassandraOptions(int32_t _gc_grace_period_in_seconds, size_t _operands_limit,
+                   bool _purge_ttl_on_expiration = false)
+      : operands_limit(_operands_limit),
+        gc_grace_period_in_seconds(_gc_grace_period_in_seconds),
+        purge_ttl_on_expiration(_purge_ttl_on_expiration) {}
+  // Limit on the number of merge operands.
+  size_t operands_limit;
+
+  // How long (in seconds) tombstoned data remains before it is purged
+  int32_t gc_grace_period_in_seconds;
+
+  // If is set to true, expired data will be directly purged.
+  // Otherwise expired data will be converted tombstones first,
+  // then be eventually removed after gc grace period. This value should
+  // only true if all writes have same ttl setting, otherwise it could bring old
+  // data back.
+  bool purge_ttl_on_expiration;
+};
+#ifndef ROCKSDB_LITE
+extern "C" {
+int RegisterCassandraObjects(ObjectLibrary& library, const std::string& arg);
+}  // extern "C"
+#endif  // ROCKSDB_LITE
+}  // namespace cassandra
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/utilities/cassandra/cassandra_row_merge_test.cc b/src/rocksdb/utilities/cassandra/cassandra_row_merge_test.cc
new file mode 100644
index 000000000..0b4a89287
--- /dev/null
+++ b/src/rocksdb/utilities/cassandra/cassandra_row_merge_test.cc
@@ -0,0 +1,98 @@
+// Copyright (c) 2017-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#include <memory>
+
+#include "test_util/testharness.h"
+#include "utilities/cassandra/format.h"
+#include "utilities/cassandra/test_utils.h"
+
+namespace ROCKSDB_NAMESPACE {
+namespace cassandra {
+
+class RowValueMergeTest : public testing::Test {};
+
+TEST(RowValueMergeTest, Merge) {
+  std::vector<RowValue> row_values;
+  row_values.push_back(CreateTestRowValue({
+      CreateTestColumnSpec(kTombstone, 0, 5),
+      CreateTestColumnSpec(kColumn, 1, 8),
+      CreateTestColumnSpec(kExpiringColumn, 2, 5),
+  }));
+
+  row_values.push_back(CreateTestRowValue({
+      CreateTestColumnSpec(kColumn, 0, 2),
+      CreateTestColumnSpec(kExpiringColumn, 1, 5),
+      CreateTestColumnSpec(kTombstone, 2, 7),
+      CreateTestColumnSpec(kExpiringColumn, 7, 17),
+  }));
+
+  row_values.push_back(CreateTestRowValue({
+      CreateTestColumnSpec(kExpiringColumn, 0, 6),
+      CreateTestColumnSpec(kTombstone, 1, 5),
+      CreateTestColumnSpec(kColumn, 2, 4),
+      CreateTestColumnSpec(kTombstone, 11, 11),
+  }));
+
+  RowValue merged = RowValue::Merge(std::move(row_values));
+  EXPECT_FALSE(merged.IsTombstone());
+  EXPECT_EQ(merged.get_columns().size(), 5);
+  VerifyRowValueColumns(merged.get_columns(), 0, kExpiringColumn, 0, 6);
+  VerifyRowValueColumns(merged.get_columns(), 1, kColumn, 1, 8);
+  VerifyRowValueColumns(merged.get_columns(), 2, kTombstone, 2, 7);
+  VerifyRowValueColumns(merged.get_columns(), 3, kExpiringColumn, 7, 17);
+  VerifyRowValueColumns(merged.get_columns(), 4, kTombstone, 11, 11);
+}
+
+TEST(RowValueMergeTest, MergeWithRowTombstone) {
+  std::vector<RowValue> row_values;
+
+  // A row tombstone.
+  row_values.push_back(CreateRowTombstone(11));
+
+  // This row's timestamp is smaller than tombstone.
+  row_values.push_back(CreateTestRowValue({
+      CreateTestColumnSpec(kColumn, 0, 5),
+      CreateTestColumnSpec(kColumn, 1, 6),
+  }));
+
+  // Some of the column's row is smaller, some is larger.
+  row_values.push_back(CreateTestRowValue({
+      CreateTestColumnSpec(kColumn, 2, 10),
+      CreateTestColumnSpec(kColumn, 3, 12),
+  }));
+
+  // All of the column's rows are larger than tombstone.
+  row_values.push_back(CreateTestRowValue({
+      CreateTestColumnSpec(kColumn, 4, 13),
+      CreateTestColumnSpec(kColumn, 5, 14),
+  }));
+
+  RowValue merged = RowValue::Merge(std::move(row_values));
+  EXPECT_FALSE(merged.IsTombstone());
+  EXPECT_EQ(merged.get_columns().size(), 3);
+  VerifyRowValueColumns(merged.get_columns(), 0, kColumn, 3, 12);
+  VerifyRowValueColumns(merged.get_columns(), 1, kColumn, 4, 13);
+  VerifyRowValueColumns(merged.get_columns(), 2, kColumn, 5, 14);
+
+  // If the tombstone's timestamp is the latest, then it returns a
+  // row tombstone.
+  row_values.push_back(CreateRowTombstone(15));
+
+  row_values.push_back(CreateRowTombstone(17));
+
+  merged = RowValue::Merge(std::move(row_values));
+  EXPECT_TRUE(merged.IsTombstone());
+  EXPECT_EQ(merged.LastModifiedTime(), 17);
+}
+
+}  // namespace cassandra
+}  // namespace ROCKSDB_NAMESPACE
+
+int main(int argc, char** argv) {
+  ROCKSDB_NAMESPACE::port::InstallStackTraceHandler();
+  ::testing::InitGoogleTest(&argc, argv);
+  return RUN_ALL_TESTS();
+}
diff --git a/src/rocksdb/utilities/cassandra/cassandra_serialize_test.cc b/src/rocksdb/utilities/cassandra/cassandra_serialize_test.cc
new file mode 100644
index 000000000..c14d8fd80
--- /dev/null
+++ b/src/rocksdb/utilities/cassandra/cassandra_serialize_test.cc
@@ -0,0 +1,164 @@
+// Copyright (c) 2017-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#include "test_util/testharness.h"
+#include "utilities/cassandra/serialize.h"
+
+namespace ROCKSDB_NAMESPACE {
+namespace cassandra {
+
+TEST(SerializeTest, SerializeI64) {
+  std::string dest;
+  Serialize<int64_t>(0, &dest);
+  EXPECT_EQ(std::string({'\x00', '\x00', '\x00', '\x00', '\x00', '\x00', '\x00',
+                         '\x00'}),
+            dest);
+
+  dest.clear();
+  Serialize<int64_t>(1, &dest);
+  EXPECT_EQ(std::string({'\x00', '\x00', '\x00', '\x00', '\x00', '\x00', '\x00',
+                         '\x01'}),
+            dest);
+
+  dest.clear();
+  Serialize<int64_t>(-1, &dest);
+  EXPECT_EQ(std::string({'\xff', '\xff', '\xff', '\xff', '\xff', '\xff', '\xff',
+                         '\xff'}),
+            dest);
+
+  dest.clear();
+  Serialize<int64_t>(9223372036854775807, &dest);
+  EXPECT_EQ(std::string({'\x7f', '\xff', '\xff', '\xff', '\xff', '\xff', '\xff',
+                         '\xff'}),
+            dest);
+
+  dest.clear();
+  Serialize<int64_t>(-9223372036854775807, &dest);
+  EXPECT_EQ(std::string({'\x80', '\x00', '\x00', '\x00', '\x00', '\x00', '\x00',
+                         '\x01'}),
+            dest);
+}
+
+TEST(SerializeTest, DeserializeI64) {
+  std::string dest;
+  std::size_t offset = dest.size();
+  Serialize<int64_t>(0, &dest);
+  EXPECT_EQ(0, Deserialize<int64_t>(dest.c_str(), offset));
+
+  offset = dest.size();
+  Serialize<int64_t>(1, &dest);
+  EXPECT_EQ(1, Deserialize<int64_t>(dest.c_str(), offset));
+
+  offset = dest.size();
+  Serialize<int64_t>(-1, &dest);
+  EXPECT_EQ(-1, Deserialize<int64_t>(dest.c_str(), offset));
+
+  offset = dest.size();
+  Serialize<int64_t>(-9223372036854775807, &dest);
+  EXPECT_EQ(-9223372036854775807, Deserialize<int64_t>(dest.c_str(), offset));
+
+  offset = dest.size();
+  Serialize<int64_t>(9223372036854775807, &dest);
+  EXPECT_EQ(9223372036854775807, Deserialize<int64_t>(dest.c_str(), offset));
+}
+
+TEST(SerializeTest, SerializeI32) {
+  std::string dest;
+  Serialize<int32_t>(0, &dest);
+  EXPECT_EQ(std::string({'\x00', '\x00', '\x00', '\x00'}), dest);
+
+  dest.clear();
+  Serialize<int32_t>(1, &dest);
+  EXPECT_EQ(std::string({'\x00', '\x00', '\x00', '\x01'}), dest);
+
+  dest.clear();
+  Serialize<int32_t>(-1, &dest);
+  EXPECT_EQ(std::string({'\xff', '\xff', '\xff', '\xff'}), dest);
+
+  dest.clear();
+  Serialize<int32_t>(2147483647, &dest);
+  EXPECT_EQ(std::string({'\x7f', '\xff', '\xff', '\xff'}), dest);
+
+  dest.clear();
+  Serialize<int32_t>(-2147483648LL, &dest);
+  EXPECT_EQ(std::string({'\x80', '\x00', '\x00', '\x00'}), dest);
+}
+
+TEST(SerializeTest, DeserializeI32) {
+  std::string dest;
+  std::size_t offset = dest.size();
+  Serialize<int32_t>(0, &dest);
+  EXPECT_EQ(0, Deserialize<int32_t>(dest.c_str(), offset));
+
+  offset = dest.size();
+  Serialize<int32_t>(1, &dest);
+  EXPECT_EQ(1, Deserialize<int32_t>(dest.c_str(), offset));
+
+  offset = dest.size();
+  Serialize<int32_t>(-1, &dest);
+  EXPECT_EQ(-1, Deserialize<int32_t>(dest.c_str(), offset));
+
+  offset = dest.size();
+  Serialize<int32_t>(2147483647, &dest);
+  EXPECT_EQ(2147483647, Deserialize<int32_t>(dest.c_str(), offset));
+
+  offset = dest.size();
+  Serialize<int32_t>(-2147483648LL, &dest);
+  EXPECT_EQ(-2147483648LL, Deserialize<int32_t>(dest.c_str(), offset));
+}
+
+TEST(SerializeTest, SerializeI8) {
+  std::string dest;
+  Serialize<int8_t>(0, &dest);
+  EXPECT_EQ(std::string({'\x00'}), dest);
+
+  dest.clear();
+  Serialize<int8_t>(1, &dest);
+  EXPECT_EQ(std::string({'\x01'}), dest);
+
+  dest.clear();
+  Serialize<int8_t>(-1, &dest);
+  EXPECT_EQ(std::string({'\xff'}), dest);
+
+  dest.clear();
+  Serialize<int8_t>(127, &dest);
+  EXPECT_EQ(std::string({'\x7f'}), dest);
+
+  dest.clear();
+  Serialize<int8_t>(-128, &dest);
+  EXPECT_EQ(std::string({'\x80'}), dest);
+}
+
+TEST(SerializeTest, DeserializeI8) {
+  std::string dest;
+  std::size_t offset = dest.size();
+  Serialize<int8_t>(0, &dest);
+  EXPECT_EQ(0, Deserialize<int8_t>(dest.c_str(), offset));
+
+  offset = dest.size();
+  Serialize<int8_t>(1, &dest);
+  EXPECT_EQ(1, Deserialize<int8_t>(dest.c_str(), offset));
+
+  offset = dest.size();
+  Serialize<int8_t>(-1, &dest);
+  EXPECT_EQ(-1, Deserialize<int8_t>(dest.c_str(), offset));
+
+  offset = dest.size();
+  Serialize<int8_t>(127, &dest);
+  EXPECT_EQ(127, Deserialize<int8_t>(dest.c_str(), offset));
+
+  offset = dest.size();
+  Serialize<int8_t>(-128, &dest);
+  EXPECT_EQ(-128, Deserialize<int8_t>(dest.c_str(), offset));
+}
+
+}  // namespace cassandra
+}  // namespace ROCKSDB_NAMESPACE
+
+int main(int argc, char** argv) {
+  ROCKSDB_NAMESPACE::port::InstallStackTraceHandler();
+  ::testing::InitGoogleTest(&argc, argv);
+  return RUN_ALL_TESTS();
+}
diff --git a/src/rocksdb/utilities/cassandra/format.cc b/src/rocksdb/utilities/cassandra/format.cc
new file mode 100644
index 000000000..cc1dd2f28
--- /dev/null
+++ b/src/rocksdb/utilities/cassandra/format.cc
@@ -0,0 +1,367 @@
+// Copyright (c) 2017-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#include "format.h"
+
+#include <algorithm>
+#include <map>
+#include <memory>
+
+#include "utilities/cassandra/serialize.h"
+
+namespace ROCKSDB_NAMESPACE {
+namespace cassandra {
+namespace {
+const int32_t kDefaultLocalDeletionTime = std::numeric_limits<int32_t>::max();
+const int64_t kDefaultMarkedForDeleteAt = std::numeric_limits<int64_t>::min();
+}  // namespace
+
+ColumnBase::ColumnBase(int8_t mask, int8_t index)
+    : mask_(mask), index_(index) {}
+
+std::size_t ColumnBase::Size() const { return sizeof(mask_) + sizeof(index_); }
+
+int8_t ColumnBase::Mask() const { return mask_; }
+
+int8_t ColumnBase::Index() const { return index_; }
+
+void ColumnBase::Serialize(std::string* dest) const {
+  ROCKSDB_NAMESPACE::cassandra::Serialize<int8_t>(mask_, dest);
+  ROCKSDB_NAMESPACE::cassandra::Serialize<int8_t>(index_, dest);
+}
+
+std::shared_ptr<ColumnBase> ColumnBase::Deserialize(const char* src,
+                                                    std::size_t offset) {
+  int8_t mask = ROCKSDB_NAMESPACE::cassandra::Deserialize<int8_t>(src, offset);
+  if ((mask & ColumnTypeMask::DELETION_MASK) != 0) {
+    return Tombstone::Deserialize(src, offset);
+  } else if ((mask & ColumnTypeMask::EXPIRATION_MASK) != 0) {
+    return ExpiringColumn::Deserialize(src, offset);
+  } else {
+    return Column::Deserialize(src, offset);
+  }
+}
+
+Column::Column(int8_t mask, int8_t index, int64_t timestamp, int32_t value_size,
+               const char* value)
+    : ColumnBase(mask, index),
+      timestamp_(timestamp),
+      value_size_(value_size),
+      value_(value) {}
+
+int64_t Column::Timestamp() const { return timestamp_; }
+
+std::size_t Column::Size() const {
+  return ColumnBase::Size() + sizeof(timestamp_) + sizeof(value_size_) +
+         value_size_;
+}
+
+void Column::Serialize(std::string* dest) const {
+  ColumnBase::Serialize(dest);
+  ROCKSDB_NAMESPACE::cassandra::Serialize<int64_t>(timestamp_, dest);
+  ROCKSDB_NAMESPACE::cassandra::Serialize<int32_t>(value_size_, dest);
+  dest->append(value_, value_size_);
+}
+
+std::shared_ptr<Column> Column::Deserialize(const char* src,
+                                            std::size_t offset) {
+  int8_t mask = ROCKSDB_NAMESPACE::cassandra::Deserialize<int8_t>(src, offset);
+  offset += sizeof(mask);
+  int8_t index = ROCKSDB_NAMESPACE::cassandra::Deserialize<int8_t>(src, offset);
+  offset += sizeof(index);
+  int64_t timestamp =
+      ROCKSDB_NAMESPACE::cassandra::Deserialize<int64_t>(src, offset);
+  offset += sizeof(timestamp);
+  int32_t value_size =
+      ROCKSDB_NAMESPACE::cassandra::Deserialize<int32_t>(src, offset);
+  offset += sizeof(value_size);
+  return std::make_shared<Column>(mask, index, timestamp, value_size,
+                                  src + offset);
+}
+
+ExpiringColumn::ExpiringColumn(int8_t mask, int8_t index, int64_t timestamp,
+                               int32_t value_size, const char* value,
+                               int32_t ttl)
+    : Column(mask, index, timestamp, value_size, value), ttl_(ttl) {}
+
+std::size_t ExpiringColumn::Size() const {
+  return Column::Size() + sizeof(ttl_);
+}
+
+void ExpiringColumn::Serialize(std::string* dest) const {
+  Column::Serialize(dest);
+  ROCKSDB_NAMESPACE::cassandra::Serialize<int32_t>(ttl_, dest);
+}
+
+std::chrono::time_point<std::chrono::system_clock> ExpiringColumn::TimePoint()
+    const {
+  return std::chrono::time_point<std::chrono::system_clock>(
+      std::chrono::microseconds(Timestamp()));
+}
+
+std::chrono::seconds ExpiringColumn::Ttl() const {
+  return std::chrono::seconds(ttl_);
+}
+
+bool ExpiringColumn::Expired() const {
+  return TimePoint() + Ttl() < std::chrono::system_clock::now();
+}
+
+std::shared_ptr<Tombstone> ExpiringColumn::ToTombstone() const {
+  auto expired_at = (TimePoint() + Ttl()).time_since_epoch();
+  int32_t local_deletion_time = static_cast<int32_t>(
+      std::chrono::duration_cast<std::chrono::seconds>(expired_at).count());
+  int64_t marked_for_delete_at =
+      std::chrono::duration_cast<std::chrono::microseconds>(expired_at).count();
+  return std::make_shared<Tombstone>(
+      static_cast<int8_t>(ColumnTypeMask::DELETION_MASK), Index(),
+      local_deletion_time, marked_for_delete_at);
+}
+
+std::shared_ptr<ExpiringColumn> ExpiringColumn::Deserialize(
+    const char* src, std::size_t offset) {
+  int8_t mask = ROCKSDB_NAMESPACE::cassandra::Deserialize<int8_t>(src, offset);
+  offset += sizeof(mask);
+  int8_t index = ROCKSDB_NAMESPACE::cassandra::Deserialize<int8_t>(src, offset);
+  offset += sizeof(index);
+  int64_t timestamp =
+      ROCKSDB_NAMESPACE::cassandra::Deserialize<int64_t>(src, offset);
+  offset += sizeof(timestamp);
+  int32_t value_size =
+      ROCKSDB_NAMESPACE::cassandra::Deserialize<int32_t>(src, offset);
+  offset += sizeof(value_size);
+  const char* value = src + offset;
+  offset += value_size;
+  int32_t ttl = ROCKSDB_NAMESPACE::cassandra::Deserialize<int32_t>(src, offset);
+  return std::make_shared<ExpiringColumn>(mask, index, timestamp, value_size,
+                                          value, ttl);
+}
+
+Tombstone::Tombstone(int8_t mask, int8_t index, int32_t local_deletion_time,
+                     int64_t marked_for_delete_at)
+    : ColumnBase(mask, index),
+      local_deletion_time_(local_deletion_time),
+      marked_for_delete_at_(marked_for_delete_at) {}
+
+int64_t Tombstone::Timestamp() const { return marked_for_delete_at_; }
+
+std::size_t Tombstone::Size() const {
+  return ColumnBase::Size() + sizeof(local_deletion_time_) +
+         sizeof(marked_for_delete_at_);
+}
+
+void Tombstone::Serialize(std::string* dest) const {
+  ColumnBase::Serialize(dest);
+  ROCKSDB_NAMESPACE::cassandra::Serialize<int32_t>(local_deletion_time_, dest);
+  ROCKSDB_NAMESPACE::cassandra::Serialize<int64_t>(marked_for_delete_at_, dest);
+}
+
+bool Tombstone::Collectable(int32_t gc_grace_period_in_seconds) const {
+  auto local_deleted_at = std::chrono::time_point<std::chrono::system_clock>(
+      std::chrono::seconds(local_deletion_time_));
+  auto gc_grace_period = std::chrono::seconds(gc_grace_period_in_seconds);
+  return local_deleted_at + gc_grace_period < std::chrono::system_clock::now();
+}
+
+std::shared_ptr<Tombstone> Tombstone::Deserialize(const char* src,
+                                                  std::size_t offset) {
+  int8_t mask = ROCKSDB_NAMESPACE::cassandra::Deserialize<int8_t>(src, offset);
+  offset += sizeof(mask);
+  int8_t index = ROCKSDB_NAMESPACE::cassandra::Deserialize<int8_t>(src, offset);
+  offset += sizeof(index);
+  int32_t local_deletion_time =
+      ROCKSDB_NAMESPACE::cassandra::Deserialize<int32_t>(src, offset);
+  offset += sizeof(int32_t);
+  int64_t marked_for_delete_at =
+      ROCKSDB_NAMESPACE::cassandra::Deserialize<int64_t>(src, offset);
+  return std::make_shared<Tombstone>(mask, index, local_deletion_time,
+                                     marked_for_delete_at);
+}
+
+RowValue::RowValue(int32_t local_deletion_time, int64_t marked_for_delete_at)
+    : local_deletion_time_(local_deletion_time),
+      marked_for_delete_at_(marked_for_delete_at),
+      columns_(),
+      last_modified_time_(0) {}
+
+RowValue::RowValue(Columns columns, int64_t last_modified_time)
+    : local_deletion_time_(kDefaultLocalDeletionTime),
+      marked_for_delete_at_(kDefaultMarkedForDeleteAt),
+      columns_(std::move(columns)),
+      last_modified_time_(last_modified_time) {}
+
+std::size_t RowValue::Size() const {
+  std::size_t size =
+      sizeof(local_deletion_time_) + sizeof(marked_for_delete_at_);
+  for (const auto& column : columns_) {
+    size += column->Size();
+  }
+  return size;
+}
+
+int64_t RowValue::LastModifiedTime() const {
+  if (IsTombstone()) {
+    return marked_for_delete_at_;
+  } else {
+    return last_modified_time_;
+  }
+}
+
+bool RowValue::IsTombstone() const {
+  return marked_for_delete_at_ > kDefaultMarkedForDeleteAt;
+}
+
+void RowValue::Serialize(std::string* dest) const {
+  ROCKSDB_NAMESPACE::cassandra::Serialize<int32_t>(local_deletion_time_, dest);
+  ROCKSDB_NAMESPACE::cassandra::Serialize<int64_t>(marked_for_delete_at_, dest);
+  for (const auto& column : columns_) {
+    column->Serialize(dest);
+  }
+}
+
+RowValue RowValue::RemoveExpiredColumns(bool* changed) const {
+  *changed = false;
+  Columns new_columns;
+  for (auto& column : columns_) {
+    if (column->Mask() == ColumnTypeMask::EXPIRATION_MASK) {
+      std::shared_ptr<ExpiringColumn> expiring_column =
+          std::static_pointer_cast<ExpiringColumn>(column);
+
+      if (expiring_column->Expired()) {
+        *changed = true;
+        continue;
+      }
+    }
+
+    new_columns.push_back(column);
+  }
+  return RowValue(std::move(new_columns), last_modified_time_);
+}
+
+RowValue RowValue::ConvertExpiredColumnsToTombstones(bool* changed) const {
+  *changed = false;
+  Columns new_columns;
+  for (auto& column : columns_) {
+    if (column->Mask() == ColumnTypeMask::EXPIRATION_MASK) {
+      std::shared_ptr<ExpiringColumn> expiring_column =
+          std::static_pointer_cast<ExpiringColumn>(column);
+
+      if (expiring_column->Expired()) {
+        std::shared_ptr<Tombstone> tombstone = expiring_column->ToTombstone();
+        new_columns.push_back(tombstone);
+        *changed = true;
+        continue;
+      }
+    }
+    new_columns.push_back(column);
+  }
+  return RowValue(std::move(new_columns), last_modified_time_);
+}
+
+RowValue RowValue::RemoveTombstones(int32_t gc_grace_period) const {
+  Columns new_columns;
+  for (auto& column : columns_) {
+    if (column->Mask() == ColumnTypeMask::DELETION_MASK) {
+      std::shared_ptr<Tombstone> tombstone =
+          std::static_pointer_cast<Tombstone>(column);
+
+      if (tombstone->Collectable(gc_grace_period)) {
+        continue;
+      }
+    }
+
+    new_columns.push_back(column);
+  }
+  return RowValue(std::move(new_columns), last_modified_time_);
+}
+
+bool RowValue::Empty() const { return columns_.empty(); }
+
+RowValue RowValue::Deserialize(const char* src, std::size_t size) {
+  std::size_t offset = 0;
+  assert(size >= sizeof(local_deletion_time_) + sizeof(marked_for_delete_at_));
+  int32_t local_deletion_time =
+      ROCKSDB_NAMESPACE::cassandra::Deserialize<int32_t>(src, offset);
+  offset += sizeof(int32_t);
+  int64_t marked_for_delete_at =
+      ROCKSDB_NAMESPACE::cassandra::Deserialize<int64_t>(src, offset);
+  offset += sizeof(int64_t);
+  if (offset == size) {
+    return RowValue(local_deletion_time, marked_for_delete_at);
+  }
+
+  assert(local_deletion_time == kDefaultLocalDeletionTime);
+  assert(marked_for_delete_at == kDefaultMarkedForDeleteAt);
+  Columns columns;
+  int64_t last_modified_time = 0;
+  while (offset < size) {
+    auto c = ColumnBase::Deserialize(src, offset);
+    offset += c->Size();
+    assert(offset <= size);
+    last_modified_time = std::max(last_modified_time, c->Timestamp());
+    columns.push_back(std::move(c));
+  }
+
+  return RowValue(std::move(columns), last_modified_time);
+}
+
+// Merge multiple row values into one.
+// For each column in rows with same index, we pick the one with latest
+// timestamp. And we also take row tombstone into consideration, by iterating
+// each row from reverse timestamp order, and stop once we hit the first
+// row tombstone.
+RowValue RowValue::Merge(std::vector<RowValue>&& values) {
+  assert(values.size() > 0);
+  if (values.size() == 1) {
+    return std::move(values[0]);
+  }
+
+  // Merge columns by their last modified time, and skip once we hit
+  // a row tombstone.
+  std::sort(values.begin(), values.end(),
+            [](const RowValue& r1, const RowValue& r2) {
+              return r1.LastModifiedTime() > r2.LastModifiedTime();
+            });
+
+  std::map<int8_t, std::shared_ptr<ColumnBase>> merged_columns;
+  int64_t tombstone_timestamp = 0;
+
+  for (auto& value : values) {
+    if (value.IsTombstone()) {
+      if (merged_columns.size() == 0) {
+        return std::move(value);
+      }
+      tombstone_timestamp = value.LastModifiedTime();
+      break;
+    }
+    for (auto& column : value.columns_) {
+      int8_t index = column->Index();
+      if (merged_columns.find(index) == merged_columns.end()) {
+        merged_columns[index] = column;
+      } else {
+        if (column->Timestamp() > merged_columns[index]->Timestamp()) {
+          merged_columns[index] = column;
+        }
+      }
+    }
+  }
+
+  int64_t last_modified_time = 0;
+  Columns columns;
+  for (auto& pair : merged_columns) {
+    // For some row, its last_modified_time > row tombstone_timestamp, but
+    // it might have rows whose timestamp is ealier than tombstone, so we
+    // ned to filter these rows.
+    if (pair.second->Timestamp() <= tombstone_timestamp) {
+      continue;
+    }
+    last_modified_time = std::max(last_modified_time, pair.second->Timestamp());
+    columns.push_back(std::move(pair.second));
+  }
+  return RowValue(std::move(columns), last_modified_time);
+}
+
+}  // namespace cassandra
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/utilities/cassandra/format.h b/src/rocksdb/utilities/cassandra/format.h
new file mode 100644
index 000000000..1b2714735
--- /dev/null
+++ b/src/rocksdb/utilities/cassandra/format.h
@@ -0,0 +1,183 @@
+// Copyright (c) 2017-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+/**
+ * The encoding of Cassandra Row Value.
+ *
+ * A Cassandra Row Value could either be a row tombstone,
+ * or contains multiple columns, it has following fields:
+ *
+ * struct row_value {
+ *   int32_t local_deletion_time;  // Time in second when the row is deleted,
+ *                                 // only used for Cassandra tombstone gc.
+ *   int64_t marked_for_delete_at; // Ms that marked this row is deleted.
+ *   struct column_base columns[]; // For non tombstone row, all columns
+ *                                 // are stored here.
+ * }
+ *
+ * If the local_deletion_time and marked_for_delete_at is set, then this is
+ * a tombstone, otherwise it contains multiple columns.
+ *
+ * There are three type of Columns: Normal Column, Expiring Column and Column
+ * Tombstone, which have following fields:
+ *
+ * // Identify the type of the column.
+ * enum mask {
+ *   DELETION_MASK = 0x01,
+ *   EXPIRATION_MASK = 0x02,
+ * };
+ *
+ * struct column  {
+ *   int8_t mask = 0;
+ *   int8_t index;
+ *   int64_t timestamp;
+ *   int32_t value_length;
+ *   char value[value_length];
+ * }
+ *
+ * struct expiring_column  {
+ *   int8_t mask = mask.EXPIRATION_MASK;
+ *   int8_t index;
+ *   int64_t timestamp;
+ *   int32_t value_length;
+ *   char value[value_length];
+ *   int32_t ttl;
+ * }
+ *
+ * struct tombstone_column  {
+ *   int8_t mask = mask.DELETION_MASK;
+ *   int8_t index;
+ *   int32_t local_deletion_time; // Similar to row_value's field.
+ *   int64_t marked_for_delete_at;
+ *  }
+ */
+
+#pragma once
+#include <chrono>
+#include <memory>
+#include <vector>
+
+#include "rocksdb/merge_operator.h"
+#include "rocksdb/slice.h"
+
+namespace ROCKSDB_NAMESPACE {
+namespace cassandra {
+
+// Identify the type of the column.
+enum ColumnTypeMask {
+  DELETION_MASK = 0x01,
+  EXPIRATION_MASK = 0x02,
+};
+
+class ColumnBase {
+ public:
+  ColumnBase(int8_t mask, int8_t index);
+  virtual ~ColumnBase() = default;
+
+  virtual int64_t Timestamp() const = 0;
+  virtual int8_t Mask() const;
+  virtual int8_t Index() const;
+  virtual std::size_t Size() const;
+  virtual void Serialize(std::string* dest) const;
+  static std::shared_ptr<ColumnBase> Deserialize(const char* src,
+                                                 std::size_t offset);
+
+ private:
+  int8_t mask_;
+  int8_t index_;
+};
+
+class Column : public ColumnBase {
+ public:
+  Column(int8_t mask, int8_t index, int64_t timestamp, int32_t value_size,
+         const char* value);
+
+  virtual int64_t Timestamp() const override;
+  virtual std::size_t Size() const override;
+  virtual void Serialize(std::string* dest) const override;
+  static std::shared_ptr<Column> Deserialize(const char* src,
+                                             std::size_t offset);
+
+ private:
+  int64_t timestamp_;
+  int32_t value_size_;
+  const char* value_;
+};
+
+class Tombstone : public ColumnBase {
+ public:
+  Tombstone(int8_t mask, int8_t index, int32_t local_deletion_time,
+            int64_t marked_for_delete_at);
+
+  virtual int64_t Timestamp() const override;
+  virtual std::size_t Size() const override;
+  virtual void Serialize(std::string* dest) const override;
+  bool Collectable(int32_t gc_grace_period) const;
+  static std::shared_ptr<Tombstone> Deserialize(const char* src,
+                                                std::size_t offset);
+
+ private:
+  int32_t local_deletion_time_;
+  int64_t marked_for_delete_at_;
+};
+
+class ExpiringColumn : public Column {
+ public:
+  ExpiringColumn(int8_t mask, int8_t index, int64_t timestamp,
+                 int32_t value_size, const char* value, int32_t ttl);
+
+  virtual std::size_t Size() const override;
+  virtual void Serialize(std::string* dest) const override;
+  bool Expired() const;
+  std::shared_ptr<Tombstone> ToTombstone() const;
+
+  static std::shared_ptr<ExpiringColumn> Deserialize(const char* src,
+                                                     std::size_t offset);
+
+ private:
+  int32_t ttl_;
+  std::chrono::time_point<std::chrono::system_clock> TimePoint() const;
+  std::chrono::seconds Ttl() const;
+};
+
+using Columns = std::vector<std::shared_ptr<ColumnBase>>;
+
+class RowValue {
+ public:
+  // Create a Row Tombstone.
+  RowValue(int32_t local_deletion_time, int64_t marked_for_delete_at);
+  // Create a Row containing columns.
+  RowValue(Columns columns, int64_t last_modified_time);
+  RowValue(const RowValue& /*that*/) = delete;
+  RowValue(RowValue&& /*that*/) noexcept = default;
+  RowValue& operator=(const RowValue& /*that*/) = delete;
+  RowValue& operator=(RowValue&& /*that*/) = default;
+
+  std::size_t Size() const;
+  bool IsTombstone() const;
+  // For Tombstone this returns the marked_for_delete_at_,
+  // otherwise it returns the max timestamp of containing columns.
+  int64_t LastModifiedTime() const;
+  void Serialize(std::string* dest) const;
+  RowValue RemoveExpiredColumns(bool* changed) const;
+  RowValue ConvertExpiredColumnsToTombstones(bool* changed) const;
+  RowValue RemoveTombstones(int32_t gc_grace_period) const;
+  bool Empty() const;
+
+  static RowValue Deserialize(const char* src, std::size_t size);
+  // Merge multiple rows according to their timestamp.
+  static RowValue Merge(std::vector<RowValue>&& values);
+
+  const Columns& get_columns() { return columns_; }
+
+ private:
+  int32_t local_deletion_time_;
+  int64_t marked_for_delete_at_;
+  Columns columns_;
+  int64_t last_modified_time_;
+};
+
+}  // namespace cassandra
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/utilities/cassandra/merge_operator.cc b/src/rocksdb/utilities/cassandra/merge_operator.cc
new file mode 100644
index 000000000..bde5dcbad
--- /dev/null
+++ b/src/rocksdb/utilities/cassandra/merge_operator.cc
@@ -0,0 +1,82 @@
+//  Copyright (c) 2017-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#include "merge_operator.h"
+
+#include <assert.h>
+
+#include <memory>
+
+#include "rocksdb/merge_operator.h"
+#include "rocksdb/slice.h"
+#include "rocksdb/utilities/options_type.h"
+#include "utilities/cassandra/format.h"
+#include "utilities/merge_operators.h"
+
+namespace ROCKSDB_NAMESPACE {
+namespace cassandra {
+static std::unordered_map<std::string, OptionTypeInfo>
+    merge_operator_options_info = {
+#ifndef ROCKSDB_LITE
+        {"gc_grace_period_in_seconds",
+         {offsetof(struct CassandraOptions, gc_grace_period_in_seconds),
+          OptionType::kUInt32T, OptionVerificationType::kNormal,
+          OptionTypeFlags::kNone}},
+        {"operands_limit",
+         {offsetof(struct CassandraOptions, operands_limit), OptionType::kSizeT,
+          OptionVerificationType::kNormal, OptionTypeFlags::kNone}},
+#endif  // ROCKSDB_LITE
+};
+
+CassandraValueMergeOperator::CassandraValueMergeOperator(
+    int32_t gc_grace_period_in_seconds, size_t operands_limit)
+    : options_(gc_grace_period_in_seconds, operands_limit) {
+  RegisterOptions(&options_, &merge_operator_options_info);
+}
+
+// Implementation for the merge operation (merges two Cassandra values)
+bool CassandraValueMergeOperator::FullMergeV2(
+    const MergeOperationInput& merge_in,
+    MergeOperationOutput* merge_out) const {
+  // Clear the *new_value for writing.
+  merge_out->new_value.clear();
+  std::vector<RowValue> row_values;
+  if (merge_in.existing_value) {
+    row_values.push_back(RowValue::Deserialize(
+        merge_in.existing_value->data(), merge_in.existing_value->size()));
+  }
+
+  for (auto& operand : merge_in.operand_list) {
+    row_values.push_back(RowValue::Deserialize(operand.data(), operand.size()));
+  }
+
+  RowValue merged = RowValue::Merge(std::move(row_values));
+  merged = merged.RemoveTombstones(options_.gc_grace_period_in_seconds);
+  merge_out->new_value.reserve(merged.Size());
+  merged.Serialize(&(merge_out->new_value));
+
+  return true;
+}
+
+bool CassandraValueMergeOperator::PartialMergeMulti(
+    const Slice& /*key*/, const std::deque<Slice>& operand_list,
+    std::string* new_value, Logger* /*logger*/) const {
+  // Clear the *new_value for writing.
+  assert(new_value);
+  new_value->clear();
+
+  std::vector<RowValue> row_values;
+  for (auto& operand : operand_list) {
+    row_values.push_back(RowValue::Deserialize(operand.data(), operand.size()));
+  }
+  RowValue merged = RowValue::Merge(std::move(row_values));
+  new_value->reserve(merged.Size());
+  merged.Serialize(new_value);
+  return true;
+}
+
+}  // namespace cassandra
+
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/utilities/cassandra/merge_operator.h b/src/rocksdb/utilities/cassandra/merge_operator.h
new file mode 100644
index 000000000..af8725db7
--- /dev/null
+++ b/src/rocksdb/utilities/cassandra/merge_operator.h
@@ -0,0 +1,44 @@
+//  Copyright (c) 2017-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#pragma once
+#include "rocksdb/merge_operator.h"
+#include "rocksdb/slice.h"
+#include "utilities/cassandra/cassandra_options.h"
+
+namespace ROCKSDB_NAMESPACE {
+namespace cassandra {
+
+/**
+ * A MergeOperator for rocksdb that implements Cassandra row value merge.
+ */
+class CassandraValueMergeOperator : public MergeOperator {
+ public:
+  explicit CassandraValueMergeOperator(int32_t gc_grace_period_in_seconds,
+                                       size_t operands_limit = 0);
+
+  virtual bool FullMergeV2(const MergeOperationInput& merge_in,
+                           MergeOperationOutput* merge_out) const override;
+
+  virtual bool PartialMergeMulti(const Slice& key,
+                                 const std::deque<Slice>& operand_list,
+                                 std::string* new_value,
+                                 Logger* logger) const override;
+
+  const char* Name() const override { return kClassName(); }
+  static const char* kClassName() { return "CassandraValueMergeOperator"; }
+
+  virtual bool AllowSingleOperand() const override { return true; }
+
+  virtual bool ShouldMerge(const std::vector<Slice>& operands) const override {
+    return options_.operands_limit > 0 &&
+           operands.size() >= options_.operands_limit;
+  }
+
+ private:
+  CassandraOptions options_;
+};
+}  // namespace cassandra
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/utilities/cassandra/serialize.h b/src/rocksdb/utilities/cassandra/serialize.h
new file mode 100644
index 000000000..4bd552bfc
--- /dev/null
+++ b/src/rocksdb/utilities/cassandra/serialize.h
@@ -0,0 +1,81 @@
+//  Copyright (c) 2017-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+/**
+ * Helper functions which serialize and deserialize integers
+ * into bytes in big endian.
+ */
+
+#pragma once
+
+#include <cstdint>
+#include <string>
+
+#include "rocksdb/rocksdb_namespace.h"
+
+namespace ROCKSDB_NAMESPACE {
+namespace cassandra {
+namespace {
+const int64_t kCharMask = 0xFFLL;
+const int32_t kBitsPerByte = 8;
+}  // namespace
+
+template <typename T>
+void Serialize(T val, std::string* dest);
+
+template <typename T>
+T Deserialize(const char* src, std::size_t offset = 0);
+
+// Specializations
+template <>
+inline void Serialize<int8_t>(int8_t t, std::string* dest) {
+  dest->append(1, static_cast<char>(t & kCharMask));
+}
+
+template <>
+inline void Serialize<int32_t>(int32_t t, std::string* dest) {
+  for (unsigned long i = 0; i < sizeof(int32_t); i++) {
+    dest->append(
+        1, static_cast<char>((t >> (sizeof(int32_t) - 1 - i) * kBitsPerByte) &
+                             kCharMask));
+  }
+}
+
+template <>
+inline void Serialize<int64_t>(int64_t t, std::string* dest) {
+  for (unsigned long i = 0; i < sizeof(int64_t); i++) {
+    dest->append(
+        1, static_cast<char>((t >> (sizeof(int64_t) - 1 - i) * kBitsPerByte) &
+                             kCharMask));
+  }
+}
+
+template <>
+inline int8_t Deserialize<int8_t>(const char* src, std::size_t offset) {
+  return static_cast<int8_t>(src[offset]);
+}
+
+template <>
+inline int32_t Deserialize<int32_t>(const char* src, std::size_t offset) {
+  int32_t result = 0;
+  for (unsigned long i = 0; i < sizeof(int32_t); i++) {
+    result |= static_cast<int32_t>(static_cast<unsigned char>(src[offset + i]))
+              << ((sizeof(int32_t) - 1 - i) * kBitsPerByte);
+  }
+  return result;
+}
+
+template <>
+inline int64_t Deserialize<int64_t>(const char* src, std::size_t offset) {
+  int64_t result = 0;
+  for (unsigned long i = 0; i < sizeof(int64_t); i++) {
+    result |= static_cast<int64_t>(static_cast<unsigned char>(src[offset + i]))
+              << ((sizeof(int64_t) - 1 - i) * kBitsPerByte);
+  }
+  return result;
+}
+
+}  // namespace cassandra
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/utilities/cassandra/test_utils.cc b/src/rocksdb/utilities/cassandra/test_utils.cc
new file mode 100644
index 000000000..ec6e5752d
--- /dev/null
+++ b/src/rocksdb/utilities/cassandra/test_utils.cc
@@ -0,0 +1,69 @@
+//  Copyright (c) 2017-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#include "test_utils.h"
+
+namespace ROCKSDB_NAMESPACE {
+namespace cassandra {
+const char kData[] = {'d', 'a', 't', 'a'};
+const char kExpiringData[] = {'e', 'd', 'a', 't', 'a'};
+const int32_t kTtl = 86400;
+const int8_t kColumn = 0;
+const int8_t kTombstone = 1;
+const int8_t kExpiringColumn = 2;
+
+std::shared_ptr<ColumnBase> CreateTestColumn(int8_t mask, int8_t index,
+                                             int64_t timestamp) {
+  if ((mask & ColumnTypeMask::DELETION_MASK) != 0) {
+    return std::shared_ptr<Tombstone>(
+        new Tombstone(mask, index, ToSeconds(timestamp), timestamp));
+  } else if ((mask & ColumnTypeMask::EXPIRATION_MASK) != 0) {
+    return std::shared_ptr<ExpiringColumn>(new ExpiringColumn(
+        mask, index, timestamp, sizeof(kExpiringData), kExpiringData, kTtl));
+  } else {
+    return std::shared_ptr<Column>(
+        new Column(mask, index, timestamp, sizeof(kData), kData));
+  }
+}
+
+std::tuple<int8_t, int8_t, int64_t> CreateTestColumnSpec(int8_t mask,
+                                                         int8_t index,
+                                                         int64_t timestamp) {
+  return std::make_tuple(mask, index, timestamp);
+}
+
+RowValue CreateTestRowValue(
+    std::vector<std::tuple<int8_t, int8_t, int64_t>> column_specs) {
+  std::vector<std::shared_ptr<ColumnBase>> columns;
+  int64_t last_modified_time = 0;
+  for (auto spec : column_specs) {
+    auto c = CreateTestColumn(std::get<0>(spec), std::get<1>(spec),
+                              std::get<2>(spec));
+    last_modified_time = std::max(last_modified_time, c->Timestamp());
+    columns.push_back(std::move(c));
+  }
+  return RowValue(std::move(columns), last_modified_time);
+}
+
+RowValue CreateRowTombstone(int64_t timestamp) {
+  return RowValue(ToSeconds(timestamp), timestamp);
+}
+
+void VerifyRowValueColumns(
+    const std::vector<std::shared_ptr<ColumnBase>> &columns,
+    std::size_t index_of_vector, int8_t expected_mask, int8_t expected_index,
+    int64_t expected_timestamp) {
+  EXPECT_EQ(expected_timestamp, columns[index_of_vector]->Timestamp());
+  EXPECT_EQ(expected_mask, columns[index_of_vector]->Mask());
+  EXPECT_EQ(expected_index, columns[index_of_vector]->Index());
+}
+
+int64_t ToMicroSeconds(int64_t seconds) { return seconds * (int64_t)1000000; }
+
+int32_t ToSeconds(int64_t microseconds) {
+  return (int32_t)(microseconds / (int64_t)1000000);
+}
+}  // namespace cassandra
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/utilities/cassandra/test_utils.h b/src/rocksdb/utilities/cassandra/test_utils.h
new file mode 100644
index 000000000..be23f7076
--- /dev/null
+++ b/src/rocksdb/utilities/cassandra/test_utils.h
@@ -0,0 +1,42 @@
+//  Copyright (c) 2017-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#pragma once
+#include <memory>
+
+#include "test_util/testharness.h"
+#include "utilities/cassandra/format.h"
+#include "utilities/cassandra/serialize.h"
+
+namespace ROCKSDB_NAMESPACE {
+namespace cassandra {
+extern const char kData[];
+extern const char kExpiringData[];
+extern const int32_t kTtl;
+extern const int8_t kColumn;
+extern const int8_t kTombstone;
+extern const int8_t kExpiringColumn;
+
+std::shared_ptr<ColumnBase> CreateTestColumn(int8_t mask, int8_t index,
+                                             int64_t timestamp);
+
+std::tuple<int8_t, int8_t, int64_t> CreateTestColumnSpec(int8_t mask,
+                                                         int8_t index,
+                                                         int64_t timestamp);
+
+RowValue CreateTestRowValue(
+    std::vector<std::tuple<int8_t, int8_t, int64_t>> column_specs);
+
+RowValue CreateRowTombstone(int64_t timestamp);
+
+void VerifyRowValueColumns(
+    const std::vector<std::shared_ptr<ColumnBase>> &columns,
+    std::size_t index_of_vector, int8_t expected_mask, int8_t expected_index,
+    int64_t expected_timestamp);
+
+int64_t ToMicroSeconds(int64_t seconds);
+int32_t ToSeconds(int64_t microseconds);
+}  // namespace cassandra
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/utilities/checkpoint/checkpoint_impl.cc b/src/rocksdb/utilities/checkpoint/checkpoint_impl.cc
new file mode 100644
index 000000000..44ce70b1b
--- /dev/null
+++ b/src/rocksdb/utilities/checkpoint/checkpoint_impl.cc
@@ -0,0 +1,469 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+//
+// Copyright (c) 2012 Facebook.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file.
+
+#ifndef ROCKSDB_LITE
+
+#include "utilities/checkpoint/checkpoint_impl.h"
+
+#include <algorithm>
+#include <cinttypes>
+#include <string>
+#include <tuple>
+#include <unordered_set>
+#include <vector>
+
+#include "db/wal_manager.h"
+#include "file/file_util.h"
+#include "file/filename.h"
+#include "logging/logging.h"
+#include "port/port.h"
+#include "rocksdb/db.h"
+#include "rocksdb/env.h"
+#include "rocksdb/metadata.h"
+#include "rocksdb/options.h"
+#include "rocksdb/transaction_log.h"
+#include "rocksdb/types.h"
+#include "rocksdb/utilities/checkpoint.h"
+#include "test_util/sync_point.h"
+#include "util/cast_util.h"
+#include "util/file_checksum_helper.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+Status Checkpoint::Create(DB* db, Checkpoint** checkpoint_ptr) {
+  *checkpoint_ptr = new CheckpointImpl(db);
+  return Status::OK();
+}
+
+Status Checkpoint::CreateCheckpoint(const std::string& /*checkpoint_dir*/,
+                                    uint64_t /*log_size_for_flush*/,
+                                    uint64_t* /*sequence_number_ptr*/) {
+  return Status::NotSupported("");
+}
+
+void CheckpointImpl::CleanStagingDirectory(const std::string& full_private_path,
+                                           Logger* info_log) {
+  std::vector<std::string> subchildren;
+  Status s = db_->GetEnv()->FileExists(full_private_path);
+  if (s.IsNotFound()) {
+    return;
+  }
+  ROCKS_LOG_INFO(info_log, "File exists %s -- %s", full_private_path.c_str(),
+                 s.ToString().c_str());
+  s = db_->GetEnv()->GetChildren(full_private_path, &subchildren);
+  if (s.ok()) {
+    for (auto& subchild : subchildren) {
+      std::string subchild_path = full_private_path + "/" + subchild;
+      s = db_->GetEnv()->DeleteFile(subchild_path);
+      ROCKS_LOG_INFO(info_log, "Delete file %s -- %s", subchild_path.c_str(),
+                     s.ToString().c_str());
+    }
+  }
+  // finally delete the private dir
+  s = db_->GetEnv()->DeleteDir(full_private_path);
+  ROCKS_LOG_INFO(info_log, "Delete dir %s -- %s", full_private_path.c_str(),
+                 s.ToString().c_str());
+}
+
+Status Checkpoint::ExportColumnFamily(
+    ColumnFamilyHandle* /*handle*/, const std::string& /*export_dir*/,
+    ExportImportFilesMetaData** /*metadata*/) {
+  return Status::NotSupported("");
+}
+
+// Builds an openable snapshot of RocksDB
+Status CheckpointImpl::CreateCheckpoint(const std::string& checkpoint_dir,
+                                        uint64_t log_size_for_flush,
+                                        uint64_t* sequence_number_ptr) {
+  DBOptions db_options = db_->GetDBOptions();
+
+  Status s = db_->GetEnv()->FileExists(checkpoint_dir);
+  if (s.ok()) {
+    return Status::InvalidArgument("Directory exists");
+  } else if (!s.IsNotFound()) {
+    assert(s.IsIOError());
+    return s;
+  }
+
+  ROCKS_LOG_INFO(
+      db_options.info_log,
+      "Started the snapshot process -- creating snapshot in directory %s",
+      checkpoint_dir.c_str());
+
+  size_t final_nonslash_idx = checkpoint_dir.find_last_not_of('/');
+  if (final_nonslash_idx == std::string::npos) {
+    // npos means it's only slashes or empty. Non-empty means it's the root
+    // directory, but it shouldn't be because we verified above the directory
+    // doesn't exist.
+    assert(checkpoint_dir.empty());
+    return Status::InvalidArgument("invalid checkpoint directory name");
+  }
+
+  std::string full_private_path =
+      checkpoint_dir.substr(0, final_nonslash_idx + 1) + ".tmp";
+  ROCKS_LOG_INFO(db_options.info_log,
+                 "Snapshot process -- using temporary directory %s",
+                 full_private_path.c_str());
+  CleanStagingDirectory(full_private_path, db_options.info_log.get());
+  // create snapshot directory
+  s = db_->GetEnv()->CreateDir(full_private_path);
+  uint64_t sequence_number = 0;
+  if (s.ok()) {
+    // enable file deletions
+    s = db_->DisableFileDeletions();
+    const bool disabled_file_deletions = s.ok();
+
+    if (s.ok() || s.IsNotSupported()) {
+      s = CreateCustomCheckpoint(
+          [&](const std::string& src_dirname, const std::string& fname,
+              FileType) {
+            ROCKS_LOG_INFO(db_options.info_log, "Hard Linking %s",
+                           fname.c_str());
+            return db_->GetFileSystem()->LinkFile(
+                src_dirname + "/" + fname, full_private_path + "/" + fname,
+                IOOptions(), nullptr);
+          } /* link_file_cb */,
+          [&](const std::string& src_dirname, const std::string& fname,
+              uint64_t size_limit_bytes, FileType,
+              const std::string& /* checksum_func_name */,
+              const std::string& /* checksum_val */,
+              const Temperature temperature) {
+            ROCKS_LOG_INFO(db_options.info_log, "Copying %s", fname.c_str());
+            return CopyFile(db_->GetFileSystem(), src_dirname + "/" + fname,
+                            full_private_path + "/" + fname, size_limit_bytes,
+                            db_options.use_fsync, nullptr, temperature);
+          } /* copy_file_cb */,
+          [&](const std::string& fname, const std::string& contents, FileType) {
+            ROCKS_LOG_INFO(db_options.info_log, "Creating %s", fname.c_str());
+            return CreateFile(db_->GetFileSystem(),
+                              full_private_path + "/" + fname, contents,
+                              db_options.use_fsync);
+          } /* create_file_cb */,
+          &sequence_number, log_size_for_flush);
+
+      // we copied all the files, enable file deletions
+      if (disabled_file_deletions) {
+        Status ss = db_->EnableFileDeletions(false);
+        assert(ss.ok());
+        ss.PermitUncheckedError();
+      }
+    }
+  }
+
+  if (s.ok()) {
+    // move tmp private backup to real snapshot directory
+    s = db_->GetEnv()->RenameFile(full_private_path, checkpoint_dir);
+  }
+  if (s.ok()) {
+    std::unique_ptr<FSDirectory> checkpoint_directory;
+    s = db_->GetFileSystem()->NewDirectory(checkpoint_dir, IOOptions(),
+                                           &checkpoint_directory, nullptr);
+    if (s.ok() && checkpoint_directory != nullptr) {
+      s = checkpoint_directory->FsyncWithDirOptions(
+          IOOptions(), nullptr,
+          DirFsyncOptions(DirFsyncOptions::FsyncReason::kDirRenamed));
+    }
+  }
+
+  if (s.ok()) {
+    if (sequence_number_ptr != nullptr) {
+      *sequence_number_ptr = sequence_number;
+    }
+    // here we know that we succeeded and installed the new snapshot
+    ROCKS_LOG_INFO(db_options.info_log, "Snapshot DONE. All is good");
+    ROCKS_LOG_INFO(db_options.info_log, "Snapshot sequence number: %" PRIu64,
+                   sequence_number);
+  } else {
+    // clean all the files we might have created
+    ROCKS_LOG_INFO(db_options.info_log, "Snapshot failed -- %s",
+                   s.ToString().c_str());
+    CleanStagingDirectory(full_private_path, db_options.info_log.get());
+  }
+  return s;
+}
+
+Status CheckpointImpl::CreateCustomCheckpoint(
+    std::function<Status(const std::string& src_dirname,
+                         const std::string& src_fname, FileType type)>
+        link_file_cb,
+    std::function<
+        Status(const std::string& src_dirname, const std::string& src_fname,
+               uint64_t size_limit_bytes, FileType type,
+               const std::string& checksum_func_name,
+               const std::string& checksum_val, const Temperature temperature)>
+        copy_file_cb,
+    std::function<Status(const std::string& fname, const std::string& contents,
+                         FileType type)>
+        create_file_cb,
+    uint64_t* sequence_number, uint64_t log_size_for_flush,
+    bool get_live_table_checksum) {
+  *sequence_number = db_->GetLatestSequenceNumber();
+
+  LiveFilesStorageInfoOptions opts;
+  opts.include_checksum_info = get_live_table_checksum;
+  opts.wal_size_for_flush = log_size_for_flush;
+
+  std::vector<LiveFileStorageInfo> infos;
+  {
+    Status s = db_->GetLiveFilesStorageInfo(opts, &infos);
+    if (!s.ok()) {
+      return s;
+    }
+  }
+
+  // Verify that everything except WAL files are in same directory
+  // (db_paths / cf_paths not supported)
+  std::unordered_set<std::string> dirs;
+  for (auto& info : infos) {
+    if (info.file_type != kWalFile) {
+      dirs.insert(info.directory);
+    }
+  }
+  if (dirs.size() > 1) {
+    return Status::NotSupported(
+        "db_paths / cf_paths not supported for Checkpoint nor BackupEngine");
+  }
+
+  bool same_fs = true;
+
+  for (auto& info : infos) {
+    Status s;
+    if (!info.replacement_contents.empty()) {
+      // Currently should only be used for CURRENT file.
+      assert(info.file_type == kCurrentFile);
+
+      if (info.size != info.replacement_contents.size()) {
+        s = Status::Corruption("Inconsistent size metadata for " +
+                               info.relative_filename);
+      } else {
+        s = create_file_cb(info.relative_filename, info.replacement_contents,
+                           info.file_type);
+      }
+    } else {
+      if (same_fs && !info.trim_to_size) {
+        s = link_file_cb(info.directory, info.relative_filename,
+                         info.file_type);
+        if (s.IsNotSupported()) {
+          same_fs = false;
+          s = Status::OK();
+        }
+        s.MustCheck();
+      }
+      if (!same_fs || info.trim_to_size) {
+        assert(info.file_checksum_func_name.empty() ==
+               !opts.include_checksum_info);
+        // no assertion on file_checksum because empty is used for both "not
+        // set" and "unknown"
+        if (opts.include_checksum_info) {
+          s = copy_file_cb(info.directory, info.relative_filename, info.size,
+                           info.file_type, info.file_checksum_func_name,
+                           info.file_checksum, info.temperature);
+        } else {
+          s = copy_file_cb(info.directory, info.relative_filename, info.size,
+                           info.file_type, kUnknownFileChecksumFuncName,
+                           kUnknownFileChecksum, info.temperature);
+        }
+      }
+    }
+    if (!s.ok()) {
+      return s;
+    }
+  }
+
+  return Status::OK();
+}
+
+// Exports all live SST files of a specified Column Family onto export_dir,
+// returning SST files information in metadata.
+Status CheckpointImpl::ExportColumnFamily(
+    ColumnFamilyHandle* handle, const std::string& export_dir,
+    ExportImportFilesMetaData** metadata) {
+  auto cfh = static_cast_with_check<ColumnFamilyHandleImpl>(handle);
+  const auto cf_name = cfh->GetName();
+  const auto db_options = db_->GetDBOptions();
+
+  assert(metadata != nullptr);
+  assert(*metadata == nullptr);
+  auto s = db_->GetEnv()->FileExists(export_dir);
+  if (s.ok()) {
+    return Status::InvalidArgument("Specified export_dir exists");
+  } else if (!s.IsNotFound()) {
+    assert(s.IsIOError());
+    return s;
+  }
+
+  const auto final_nonslash_idx = export_dir.find_last_not_of('/');
+  if (final_nonslash_idx == std::string::npos) {
+    return Status::InvalidArgument("Specified export_dir invalid");
+  }
+  ROCKS_LOG_INFO(db_options.info_log,
+                 "[%s] export column family onto export directory %s",
+                 cf_name.c_str(), export_dir.c_str());
+
+  // Create a temporary export directory.
+  const auto tmp_export_dir =
+      export_dir.substr(0, final_nonslash_idx + 1) + ".tmp";
+  s = db_->GetEnv()->CreateDir(tmp_export_dir);
+
+  if (s.ok()) {
+    s = db_->Flush(ROCKSDB_NAMESPACE::FlushOptions(), handle);
+  }
+
+  ColumnFamilyMetaData db_metadata;
+  if (s.ok()) {
+    // Export live sst files with file deletions disabled.
+    s = db_->DisableFileDeletions();
+    if (s.ok()) {
+      db_->GetColumnFamilyMetaData(handle, &db_metadata);
+
+      s = ExportFilesInMetaData(
+          db_options, db_metadata,
+          [&](const std::string& src_dirname, const std::string& fname) {
+            ROCKS_LOG_INFO(db_options.info_log, "[%s] HardLinking %s",
+                           cf_name.c_str(), fname.c_str());
+            return db_->GetEnv()->LinkFile(src_dirname + fname,
+                                           tmp_export_dir + fname);
+          } /*link_file_cb*/,
+          [&](const std::string& src_dirname, const std::string& fname) {
+            ROCKS_LOG_INFO(db_options.info_log, "[%s] Copying %s",
+                           cf_name.c_str(), fname.c_str());
+            return CopyFile(db_->GetFileSystem(), src_dirname + fname,
+                            tmp_export_dir + fname, 0, db_options.use_fsync,
+                            nullptr, Temperature::kUnknown);
+          } /*copy_file_cb*/);
+
+      const auto enable_status = db_->EnableFileDeletions(false /*force*/);
+      if (s.ok()) {
+        s = enable_status;
+      }
+    }
+  }
+
+  auto moved_to_user_specified_dir = false;
+  if (s.ok()) {
+    // Move temporary export directory to the actual export directory.
+    s = db_->GetEnv()->RenameFile(tmp_export_dir, export_dir);
+  }
+
+  if (s.ok()) {
+    // Fsync export directory.
+    moved_to_user_specified_dir = true;
+    std::unique_ptr<FSDirectory> dir_ptr;
+    s = db_->GetFileSystem()->NewDirectory(export_dir, IOOptions(), &dir_ptr,
+                                           nullptr);
+    if (s.ok()) {
+      assert(dir_ptr != nullptr);
+      s = dir_ptr->FsyncWithDirOptions(
+          IOOptions(), nullptr,
+          DirFsyncOptions(DirFsyncOptions::FsyncReason::kDirRenamed));
+    }
+  }
+
+  if (s.ok()) {
+    // Export of files succeeded. Fill in the metadata information.
+    auto result_metadata = new ExportImportFilesMetaData();
+    result_metadata->db_comparator_name = handle->GetComparator()->Name();
+    for (const auto& level_metadata : db_metadata.levels) {
+      for (const auto& file_metadata : level_metadata.files) {
+        LiveFileMetaData live_file_metadata;
+        live_file_metadata.size = file_metadata.size;
+        live_file_metadata.name = std::move(file_metadata.name);
+        live_file_metadata.file_number = file_metadata.file_number;
+        live_file_metadata.db_path = export_dir;
+        live_file_metadata.smallest_seqno = file_metadata.smallest_seqno;
+        live_file_metadata.largest_seqno = file_metadata.largest_seqno;
+        live_file_metadata.smallestkey = std::move(file_metadata.smallestkey);
+        live_file_metadata.largestkey = std::move(file_metadata.largestkey);
+        live_file_metadata.oldest_blob_file_number =
+            file_metadata.oldest_blob_file_number;
+        live_file_metadata.level = level_metadata.level;
+        result_metadata->files.push_back(live_file_metadata);
+      }
+      *metadata = result_metadata;
+    }
+    ROCKS_LOG_INFO(db_options.info_log, "[%s] Export succeeded.",
+                   cf_name.c_str());
+  } else {
+    // Failure: Clean up all the files/directories created.
+    ROCKS_LOG_INFO(db_options.info_log, "[%s] Export failed. %s",
+                   cf_name.c_str(), s.ToString().c_str());
+    std::vector<std::string> subchildren;
+    const auto cleanup_dir =
+        moved_to_user_specified_dir ? export_dir : tmp_export_dir;
+    db_->GetEnv()->GetChildren(cleanup_dir, &subchildren);
+    for (const auto& subchild : subchildren) {
+      const auto subchild_path = cleanup_dir + "/" + subchild;
+      const auto status = db_->GetEnv()->DeleteFile(subchild_path);
+      if (!status.ok()) {
+        ROCKS_LOG_WARN(db_options.info_log, "Failed to cleanup file %s: %s",
+                       subchild_path.c_str(), status.ToString().c_str());
+      }
+    }
+    const auto status = db_->GetEnv()->DeleteDir(cleanup_dir);
+    if (!status.ok()) {
+      ROCKS_LOG_WARN(db_options.info_log, "Failed to cleanup dir %s: %s",
+                     cleanup_dir.c_str(), status.ToString().c_str());
+    }
+  }
+  return s;
+}
+
+Status CheckpointImpl::ExportFilesInMetaData(
+    const DBOptions& db_options, const ColumnFamilyMetaData& metadata,
+    std::function<Status(const std::string& src_dirname,
+                         const std::string& src_fname)>
+        link_file_cb,
+    std::function<Status(const std::string& src_dirname,
+                         const std::string& src_fname)>
+        copy_file_cb) {
+  Status s;
+  auto hardlink_file = true;
+
+  // Copy/hard link files in metadata.
+  size_t num_files = 0;
+  for (const auto& level_metadata : metadata.levels) {
+    for (const auto& file_metadata : level_metadata.files) {
+      uint64_t number;
+      FileType type;
+      const auto ok = ParseFileName(file_metadata.name, &number, &type);
+      if (!ok) {
+        s = Status::Corruption("Could not parse file name");
+        break;
+      }
+
+      // We should only get sst files here.
+      assert(type == kTableFile);
+      assert(file_metadata.size > 0 && file_metadata.name[0] == '/');
+      const auto src_fname = file_metadata.name;
+      ++num_files;
+
+      if (hardlink_file) {
+        s = link_file_cb(db_->GetName(), src_fname);
+        if (num_files == 1 && s.IsNotSupported()) {
+          // Fallback to copy if link failed due to cross-device directories.
+          hardlink_file = false;
+          s = Status::OK();
+        }
+      }
+      if (!hardlink_file) {
+        s = copy_file_cb(db_->GetName(), src_fname);
+      }
+      if (!s.ok()) {
+        break;
+      }
+    }
+  }
+  ROCKS_LOG_INFO(db_options.info_log, "Number of table files %" ROCKSDB_PRIszt,
+                 num_files);
+
+  return s;
+}
+}  // namespace ROCKSDB_NAMESPACE
+
+#endif  // ROCKSDB_LITE
diff --git a/src/rocksdb/utilities/checkpoint/checkpoint_impl.h b/src/rocksdb/utilities/checkpoint/checkpoint_impl.h
new file mode 100644
index 000000000..2947330cc
--- /dev/null
+++ b/src/rocksdb/utilities/checkpoint/checkpoint_impl.h
@@ -0,0 +1,66 @@
+//  Copyright (c) 2017-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#pragma once
+#ifndef ROCKSDB_LITE
+
+#include <string>
+
+#include "file/filename.h"
+#include "rocksdb/db.h"
+#include "rocksdb/utilities/checkpoint.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+class CheckpointImpl : public Checkpoint {
+ public:
+  explicit CheckpointImpl(DB* db) : db_(db) {}
+
+  Status CreateCheckpoint(const std::string& checkpoint_dir,
+                          uint64_t log_size_for_flush,
+                          uint64_t* sequence_number_ptr) override;
+
+  Status ExportColumnFamily(ColumnFamilyHandle* handle,
+                            const std::string& export_dir,
+                            ExportImportFilesMetaData** metadata) override;
+
+  // Checkpoint logic can be customized by providing callbacks for link, copy,
+  // or create.
+  Status CreateCustomCheckpoint(
+      std::function<Status(const std::string& src_dirname,
+                           const std::string& fname, FileType type)>
+          link_file_cb,
+      std::function<Status(const std::string& src_dirname,
+                           const std::string& fname, uint64_t size_limit_bytes,
+                           FileType type, const std::string& checksum_func_name,
+                           const std::string& checksum_val,
+                           const Temperature src_temperature)>
+          copy_file_cb,
+      std::function<Status(const std::string& fname,
+                           const std::string& contents, FileType type)>
+          create_file_cb,
+      uint64_t* sequence_number, uint64_t log_size_for_flush,
+      bool get_live_table_checksum = false);
+
+ private:
+  void CleanStagingDirectory(const std::string& path, Logger* info_log);
+
+  // Export logic customization by providing callbacks for link or copy.
+  Status ExportFilesInMetaData(
+      const DBOptions& db_options, const ColumnFamilyMetaData& metadata,
+      std::function<Status(const std::string& src_dirname,
+                           const std::string& fname)>
+          link_file_cb,
+      std::function<Status(const std::string& src_dirname,
+                           const std::string& fname)>
+          copy_file_cb);
+
+ private:
+  DB* db_;
+};
+
+}  // namespace ROCKSDB_NAMESPACE
+
+#endif  // ROCKSDB_LITE
diff --git a/src/rocksdb/utilities/checkpoint/checkpoint_test.cc b/src/rocksdb/utilities/checkpoint/checkpoint_test.cc
new file mode 100644
index 000000000..3da753d5f
--- /dev/null
+++ b/src/rocksdb/utilities/checkpoint/checkpoint_test.cc
@@ -0,0 +1,974 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+//
+// Copyright (c) 2011 The LevelDB Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file. See the AUTHORS file for names of contributors.
+
+// Syncpoint prevents us building and running tests in release
+#ifndef ROCKSDB_LITE
+#include "rocksdb/utilities/checkpoint.h"
+
+#ifndef OS_WIN
+#include <unistd.h>
+#endif
+#include <iostream>
+#include <thread>
+#include <utility>
+
+#include "db/db_impl/db_impl.h"
+#include "file/file_util.h"
+#include "port/port.h"
+#include "port/stack_trace.h"
+#include "rocksdb/db.h"
+#include "rocksdb/env.h"
+#include "rocksdb/utilities/transaction_db.h"
+#include "test_util/sync_point.h"
+#include "test_util/testharness.h"
+#include "test_util/testutil.h"
+#include "utilities/fault_injection_env.h"
+#include "utilities/fault_injection_fs.h"
+
+namespace ROCKSDB_NAMESPACE {
+class CheckpointTest : public testing::Test {
+ protected:
+  // Sequence of option configurations to try
+  enum OptionConfig {
+    kDefault = 0,
+  };
+  int option_config_;
+
+ public:
+  std::string dbname_;
+  std::string alternative_wal_dir_;
+  Env* env_;
+  DB* db_;
+  Options last_options_;
+  std::vector<ColumnFamilyHandle*> handles_;
+  std::string snapshot_name_;
+  std::string export_path_;
+  ColumnFamilyHandle* cfh_reverse_comp_;
+  ExportImportFilesMetaData* metadata_;
+
+  CheckpointTest() : env_(Env::Default()) {
+    env_->SetBackgroundThreads(1, Env::LOW);
+    env_->SetBackgroundThreads(1, Env::HIGH);
+    dbname_ = test::PerThreadDBPath(env_, "checkpoint_test");
+    alternative_wal_dir_ = dbname_ + "/wal";
+    auto options = CurrentOptions();
+    auto delete_options = options;
+    delete_options.wal_dir = alternative_wal_dir_;
+    EXPECT_OK(DestroyDB(dbname_, delete_options));
+    // Destroy it for not alternative WAL dir is used.
+    EXPECT_OK(DestroyDB(dbname_, options));
+    db_ = nullptr;
+    snapshot_name_ = test::PerThreadDBPath(env_, "snapshot");
+    std::string snapshot_tmp_name = snapshot_name_ + ".tmp";
+    EXPECT_OK(DestroyDB(snapshot_name_, options));
+    test::DeleteDir(env_, snapshot_name_);
+    EXPECT_OK(DestroyDB(snapshot_tmp_name, options));
+    test::DeleteDir(env_, snapshot_tmp_name);
+    Reopen(options);
+    export_path_ = test::PerThreadDBPath("/export");
+    DestroyDir(env_, export_path_).PermitUncheckedError();
+    cfh_reverse_comp_ = nullptr;
+    metadata_ = nullptr;
+  }
+
+  ~CheckpointTest() override {
+    ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->DisableProcessing();
+    ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->LoadDependency({});
+    ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->ClearAllCallBacks();
+    if (cfh_reverse_comp_) {
+      EXPECT_OK(db_->DestroyColumnFamilyHandle(cfh_reverse_comp_));
+      cfh_reverse_comp_ = nullptr;
+    }
+    if (metadata_) {
+      delete metadata_;
+      metadata_ = nullptr;
+    }
+    Close();
+    Options options;
+    options.db_paths.emplace_back(dbname_, 0);
+    options.db_paths.emplace_back(dbname_ + "_2", 0);
+    options.db_paths.emplace_back(dbname_ + "_3", 0);
+    options.db_paths.emplace_back(dbname_ + "_4", 0);
+    EXPECT_OK(DestroyDB(dbname_, options));
+    EXPECT_OK(DestroyDB(snapshot_name_, options));
+    DestroyDir(env_, export_path_).PermitUncheckedError();
+  }
+
+  // Return the current option configuration.
+  Options CurrentOptions() {
+    Options options;
+    options.env = env_;
+    options.create_if_missing = true;
+    return options;
+  }
+
+  void CreateColumnFamilies(const std::vector<std::string>& cfs,
+                            const Options& options) {
+    ColumnFamilyOptions cf_opts(options);
+    size_t cfi = handles_.size();
+    handles_.resize(cfi + cfs.size());
+    for (auto cf : cfs) {
+      ASSERT_OK(db_->CreateColumnFamily(cf_opts, cf, &handles_[cfi++]));
+    }
+  }
+
+  void CreateAndReopenWithCF(const std::vector<std::string>& cfs,
+                             const Options& options) {
+    CreateColumnFamilies(cfs, options);
+    std::vector<std::string> cfs_plus_default = cfs;
+    cfs_plus_default.insert(cfs_plus_default.begin(), kDefaultColumnFamilyName);
+    ReopenWithColumnFamilies(cfs_plus_default, options);
+  }
+
+  void ReopenWithColumnFamilies(const std::vector<std::string>& cfs,
+                                const std::vector<Options>& options) {
+    ASSERT_OK(TryReopenWithColumnFamilies(cfs, options));
+  }
+
+  void ReopenWithColumnFamilies(const std::vector<std::string>& cfs,
+                                const Options& options) {
+    ASSERT_OK(TryReopenWithColumnFamilies(cfs, options));
+  }
+
+  Status TryReopenWithColumnFamilies(const std::vector<std::string>& cfs,
+                                     const std::vector<Options>& options) {
+    Close();
+    EXPECT_EQ(cfs.size(), options.size());
+    std::vector<ColumnFamilyDescriptor> column_families;
+    for (size_t i = 0; i < cfs.size(); ++i) {
+      column_families.push_back(ColumnFamilyDescriptor(cfs[i], options[i]));
+    }
+    DBOptions db_opts = DBOptions(options[0]);
+    return DB::Open(db_opts, dbname_, column_families, &handles_, &db_);
+  }
+
+  Status TryReopenWithColumnFamilies(const std::vector<std::string>& cfs,
+                                     const Options& options) {
+    Close();
+    std::vector<Options> v_opts(cfs.size(), options);
+    return TryReopenWithColumnFamilies(cfs, v_opts);
+  }
+
+  void Reopen(const Options& options) { ASSERT_OK(TryReopen(options)); }
+
+  void CompactAll() {
+    for (auto h : handles_) {
+      ASSERT_OK(db_->CompactRange(CompactRangeOptions(), h, nullptr, nullptr));
+    }
+  }
+
+  void Close() {
+    for (auto h : handles_) {
+      delete h;
+    }
+    handles_.clear();
+    delete db_;
+    db_ = nullptr;
+  }
+
+  void DestroyAndReopen(const Options& options) {
+    // Destroy using last options
+    Destroy(last_options_);
+    ASSERT_OK(TryReopen(options));
+  }
+
+  void Destroy(const Options& options) {
+    Close();
+    ASSERT_OK(DestroyDB(dbname_, options));
+  }
+
+  Status ReadOnlyReopen(const Options& options) {
+    return DB::OpenForReadOnly(options, dbname_, &db_);
+  }
+
+  Status ReadOnlyReopenWithColumnFamilies(const std::vector<std::string>& cfs,
+                                          const Options& options) {
+    std::vector<ColumnFamilyDescriptor> column_families;
+    for (const auto& cf : cfs) {
+      column_families.emplace_back(cf, options);
+    }
+    return DB::OpenForReadOnly(options, dbname_, column_families, &handles_,
+                               &db_);
+  }
+
+  Status TryReopen(const Options& options) {
+    Close();
+    last_options_ = options;
+    return DB::Open(options, dbname_, &db_);
+  }
+
+  Status Flush(int cf = 0) {
+    if (cf == 0) {
+      return db_->Flush(FlushOptions());
+    } else {
+      return db_->Flush(FlushOptions(), handles_[cf]);
+    }
+  }
+
+  Status Put(const Slice& k, const Slice& v, WriteOptions wo = WriteOptions()) {
+    return db_->Put(wo, k, v);
+  }
+
+  Status Put(int cf, const Slice& k, const Slice& v,
+             WriteOptions wo = WriteOptions()) {
+    return db_->Put(wo, handles_[cf], k, v);
+  }
+
+  Status Delete(const std::string& k) { return db_->Delete(WriteOptions(), k); }
+
+  Status Delete(int cf, const std::string& k) {
+    return db_->Delete(WriteOptions(), handles_[cf], k);
+  }
+
+  std::string Get(const std::string& k, const Snapshot* snapshot = nullptr) {
+    ReadOptions options;
+    options.verify_checksums = true;
+    options.snapshot = snapshot;
+    std::string result;
+    Status s = db_->Get(options, k, &result);
+    if (s.IsNotFound()) {
+      result = "NOT_FOUND";
+    } else if (!s.ok()) {
+      result = s.ToString();
+    }
+    return result;
+  }
+
+  std::string Get(int cf, const std::string& k,
+                  const Snapshot* snapshot = nullptr) {
+    ReadOptions options;
+    options.verify_checksums = true;
+    options.snapshot = snapshot;
+    std::string result;
+    Status s = db_->Get(options, handles_[cf], k, &result);
+    if (s.IsNotFound()) {
+      result = "NOT_FOUND";
+    } else if (!s.ok()) {
+      result = s.ToString();
+    }
+    return result;
+  }
+};
+
+TEST_F(CheckpointTest, GetSnapshotLink) {
+  for (uint64_t log_size_for_flush : {0, 1000000}) {
+    Options options;
+    DB* snapshotDB;
+    ReadOptions roptions;
+    std::string result;
+    Checkpoint* checkpoint;
+
+    options = CurrentOptions();
+    delete db_;
+    db_ = nullptr;
+    ASSERT_OK(DestroyDB(dbname_, options));
+
+    // Create a database
+    options.create_if_missing = true;
+    ASSERT_OK(DB::Open(options, dbname_, &db_));
+    std::string key = std::string("foo");
+    ASSERT_OK(Put(key, "v1"));
+    // Take a snapshot
+    ASSERT_OK(Checkpoint::Create(db_, &checkpoint));
+    ASSERT_OK(checkpoint->CreateCheckpoint(snapshot_name_, log_size_for_flush));
+    ASSERT_OK(Put(key, "v2"));
+    ASSERT_EQ("v2", Get(key));
+    ASSERT_OK(Flush());
+    ASSERT_EQ("v2", Get(key));
+    // Open snapshot and verify contents while DB is running
+    options.create_if_missing = false;
+    ASSERT_OK(DB::Open(options, snapshot_name_, &snapshotDB));
+    ASSERT_OK(snapshotDB->Get(roptions, key, &result));
+    ASSERT_EQ("v1", result);
+    delete snapshotDB;
+    snapshotDB = nullptr;
+    delete db_;
+    db_ = nullptr;
+
+    // Destroy original DB
+    ASSERT_OK(DestroyDB(dbname_, options));
+
+    // Open snapshot and verify contents
+    options.create_if_missing = false;
+    dbname_ = snapshot_name_;
+    ASSERT_OK(DB::Open(options, dbname_, &db_));
+    ASSERT_EQ("v1", Get(key));
+    delete db_;
+    db_ = nullptr;
+    ASSERT_OK(DestroyDB(dbname_, options));
+    delete checkpoint;
+
+    // Restore DB name
+    dbname_ = test::PerThreadDBPath(env_, "db_test");
+  }
+}
+
+TEST_F(CheckpointTest, CheckpointWithBlob) {
+  // Create a database with a blob file
+  Options options = CurrentOptions();
+  options.create_if_missing = true;
+  options.enable_blob_files = true;
+  options.min_blob_size = 0;
+
+  Reopen(options);
+
+  constexpr char key[] = "key";
+  constexpr char blob[] = "blob";
+
+  ASSERT_OK(Put(key, blob));
+  ASSERT_OK(Flush());
+
+  // Create a checkpoint
+  Checkpoint* checkpoint = nullptr;
+  ASSERT_OK(Checkpoint::Create(db_, &checkpoint));
+
+  std::unique_ptr<Checkpoint> checkpoint_guard(checkpoint);
+
+  ASSERT_OK(checkpoint->CreateCheckpoint(snapshot_name_));
+
+  // Make sure it contains the blob file
+  std::vector<std::string> files;
+  ASSERT_OK(env_->GetChildren(snapshot_name_, &files));
+
+  bool blob_file_found = false;
+  for (const auto& file : files) {
+    uint64_t number = 0;
+    FileType type = kWalFile;
+
+    if (ParseFileName(file, &number, &type) && type == kBlobFile) {
+      blob_file_found = true;
+      break;
+    }
+  }
+
+  ASSERT_TRUE(blob_file_found);
+
+  // Make sure the checkpoint can be opened and the blob value read
+  options.create_if_missing = false;
+  DB* checkpoint_db = nullptr;
+  ASSERT_OK(DB::Open(options, snapshot_name_, &checkpoint_db));
+
+  std::unique_ptr<DB> checkpoint_db_guard(checkpoint_db);
+
+  PinnableSlice value;
+  ASSERT_OK(checkpoint_db->Get(
+      ReadOptions(), checkpoint_db->DefaultColumnFamily(), key, &value));
+
+  ASSERT_EQ(value, blob);
+}
+
+TEST_F(CheckpointTest, ExportColumnFamilyWithLinks) {
+  // Create a database
+  auto options = CurrentOptions();
+  options.create_if_missing = true;
+  CreateAndReopenWithCF({}, options);
+
+  // Helper to verify the number of files in metadata and export dir
+  auto verify_files_exported = [&](const ExportImportFilesMetaData& metadata,
+                                   int num_files_expected) {
+    ASSERT_EQ(metadata.files.size(), num_files_expected);
+    std::vector<std::string> subchildren;
+    ASSERT_OK(env_->GetChildren(export_path_, &subchildren));
+    ASSERT_EQ(subchildren.size(), num_files_expected);
+  };
+
+  // Test DefaultColumnFamily
+  {
+    const auto key = std::string("foo");
+    ASSERT_OK(Put(key, "v1"));
+
+    Checkpoint* checkpoint;
+    ASSERT_OK(Checkpoint::Create(db_, &checkpoint));
+
+    // Export the Tables and verify
+    ASSERT_OK(checkpoint->ExportColumnFamily(db_->DefaultColumnFamily(),
+                                             export_path_, &metadata_));
+    verify_files_exported(*metadata_, 1);
+    ASSERT_EQ(metadata_->db_comparator_name, options.comparator->Name());
+    ASSERT_OK(DestroyDir(env_, export_path_));
+    delete metadata_;
+    metadata_ = nullptr;
+
+    // Check again after compaction
+    CompactAll();
+    ASSERT_OK(Put(key, "v2"));
+    ASSERT_OK(checkpoint->ExportColumnFamily(db_->DefaultColumnFamily(),
+                                             export_path_, &metadata_));
+    verify_files_exported(*metadata_, 2);
+    ASSERT_EQ(metadata_->db_comparator_name, options.comparator->Name());
+    ASSERT_OK(DestroyDir(env_, export_path_));
+    delete metadata_;
+    metadata_ = nullptr;
+    delete checkpoint;
+  }
+
+  // Test non default column family with non default comparator
+  {
+    auto cf_options = CurrentOptions();
+    cf_options.comparator = ReverseBytewiseComparator();
+    ASSERT_OK(db_->CreateColumnFamily(cf_options, "yoyo", &cfh_reverse_comp_));
+
+    const auto key = std::string("foo");
+    ASSERT_OK(db_->Put(WriteOptions(), cfh_reverse_comp_, key, "v1"));
+
+    Checkpoint* checkpoint;
+    ASSERT_OK(Checkpoint::Create(db_, &checkpoint));
+
+    // Export the Tables and verify
+    ASSERT_OK(checkpoint->ExportColumnFamily(cfh_reverse_comp_, export_path_,
+                                             &metadata_));
+    verify_files_exported(*metadata_, 1);
+    ASSERT_EQ(metadata_->db_comparator_name,
+              ReverseBytewiseComparator()->Name());
+    delete checkpoint;
+  }
+}
+
+TEST_F(CheckpointTest, ExportColumnFamilyNegativeTest) {
+  // Create a database
+  auto options = CurrentOptions();
+  options.create_if_missing = true;
+  CreateAndReopenWithCF({}, options);
+
+  const auto key = std::string("foo");
+  ASSERT_OK(Put(key, "v1"));
+
+  Checkpoint* checkpoint;
+  ASSERT_OK(Checkpoint::Create(db_, &checkpoint));
+
+  // Export onto existing directory
+  ASSERT_OK(env_->CreateDirIfMissing(export_path_));
+  ASSERT_EQ(checkpoint->ExportColumnFamily(db_->DefaultColumnFamily(),
+                                           export_path_, &metadata_),
+            Status::InvalidArgument("Specified export_dir exists"));
+  ASSERT_OK(DestroyDir(env_, export_path_));
+
+  // Export with invalid directory specification
+  export_path_ = "";
+  ASSERT_EQ(checkpoint->ExportColumnFamily(db_->DefaultColumnFamily(),
+                                           export_path_, &metadata_),
+            Status::InvalidArgument("Specified export_dir invalid"));
+  delete checkpoint;
+}
+
+TEST_F(CheckpointTest, CheckpointCF) {
+  Options options = CurrentOptions();
+  CreateAndReopenWithCF({"one", "two", "three", "four", "five"}, options);
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->LoadDependency(
+      {{"CheckpointTest::CheckpointCF:2", "DBImpl::GetLiveFiles:2"},
+       {"DBImpl::GetLiveFiles:1", "CheckpointTest::CheckpointCF:1"}});
+
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->EnableProcessing();
+
+  ASSERT_OK(Put(0, "Default", "Default"));
+  ASSERT_OK(Put(1, "one", "one"));
+  ASSERT_OK(Put(2, "two", "two"));
+  ASSERT_OK(Put(3, "three", "three"));
+  ASSERT_OK(Put(4, "four", "four"));
+  ASSERT_OK(Put(5, "five", "five"));
+
+  DB* snapshotDB;
+  ReadOptions roptions;
+  std::string result;
+  std::vector<ColumnFamilyHandle*> cphandles;
+
+  // Take a snapshot
+  ROCKSDB_NAMESPACE::port::Thread t([&]() {
+    Checkpoint* checkpoint;
+    ASSERT_OK(Checkpoint::Create(db_, &checkpoint));
+    ASSERT_OK(checkpoint->CreateCheckpoint(snapshot_name_));
+    delete checkpoint;
+  });
+  TEST_SYNC_POINT("CheckpointTest::CheckpointCF:1");
+  ASSERT_OK(Put(0, "Default", "Default1"));
+  ASSERT_OK(Put(1, "one", "eleven"));
+  ASSERT_OK(Put(2, "two", "twelve"));
+  ASSERT_OK(Put(3, "three", "thirteen"));
+  ASSERT_OK(Put(4, "four", "fourteen"));
+  ASSERT_OK(Put(5, "five", "fifteen"));
+  TEST_SYNC_POINT("CheckpointTest::CheckpointCF:2");
+  t.join();
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->DisableProcessing();
+  ASSERT_OK(Put(1, "one", "twentyone"));
+  ASSERT_OK(Put(2, "two", "twentytwo"));
+  ASSERT_OK(Put(3, "three", "twentythree"));
+  ASSERT_OK(Put(4, "four", "twentyfour"));
+  ASSERT_OK(Put(5, "five", "twentyfive"));
+  ASSERT_OK(Flush());
+
+  // Open snapshot and verify contents while DB is running
+  options.create_if_missing = false;
+  std::vector<std::string> cfs;
+  cfs = {kDefaultColumnFamilyName, "one", "two", "three", "four", "five"};
+  std::vector<ColumnFamilyDescriptor> column_families;
+  for (size_t i = 0; i < cfs.size(); ++i) {
+    column_families.push_back(ColumnFamilyDescriptor(cfs[i], options));
+  }
+  ASSERT_OK(DB::Open(options, snapshot_name_, column_families, &cphandles,
+                     &snapshotDB));
+  ASSERT_OK(snapshotDB->Get(roptions, cphandles[0], "Default", &result));
+  ASSERT_EQ("Default1", result);
+  ASSERT_OK(snapshotDB->Get(roptions, cphandles[1], "one", &result));
+  ASSERT_EQ("eleven", result);
+  ASSERT_OK(snapshotDB->Get(roptions, cphandles[2], "two", &result));
+  for (auto h : cphandles) {
+    delete h;
+  }
+  cphandles.clear();
+  delete snapshotDB;
+  snapshotDB = nullptr;
+}
+
+TEST_F(CheckpointTest, CheckpointCFNoFlush) {
+  Options options = CurrentOptions();
+  CreateAndReopenWithCF({"one", "two", "three", "four", "five"}, options);
+
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->EnableProcessing();
+
+  ASSERT_OK(Put(0, "Default", "Default"));
+  ASSERT_OK(Put(1, "one", "one"));
+  ASSERT_OK(Flush());
+  ASSERT_OK(Put(2, "two", "two"));
+
+  DB* snapshotDB;
+  ReadOptions roptions;
+  std::string result;
+  std::vector<ColumnFamilyHandle*> cphandles;
+
+  // Take a snapshot
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
+      "DBImpl::BackgroundCallFlush:start", [&](void* /*arg*/) {
+        // Flush should never trigger.
+        FAIL();
+      });
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->EnableProcessing();
+  Checkpoint* checkpoint;
+  ASSERT_OK(Checkpoint::Create(db_, &checkpoint));
+  ASSERT_OK(checkpoint->CreateCheckpoint(snapshot_name_, 1000000));
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->DisableProcessing();
+
+  delete checkpoint;
+  ASSERT_OK(Put(1, "one", "two"));
+  ASSERT_OK(Flush(1));
+  ASSERT_OK(Put(2, "two", "twentytwo"));
+  Close();
+  EXPECT_OK(DestroyDB(dbname_, options));
+
+  // Open snapshot and verify contents while DB is running
+  options.create_if_missing = false;
+  std::vector<std::string> cfs;
+  cfs = {kDefaultColumnFamilyName, "one", "two", "three", "four", "five"};
+  std::vector<ColumnFamilyDescriptor> column_families;
+  for (size_t i = 0; i < cfs.size(); ++i) {
+    column_families.push_back(ColumnFamilyDescriptor(cfs[i], options));
+  }
+  ASSERT_OK(DB::Open(options, snapshot_name_, column_families, &cphandles,
+                     &snapshotDB));
+  ASSERT_OK(snapshotDB->Get(roptions, cphandles[0], "Default", &result));
+  ASSERT_EQ("Default", result);
+  ASSERT_OK(snapshotDB->Get(roptions, cphandles[1], "one", &result));
+  ASSERT_EQ("one", result);
+  ASSERT_OK(snapshotDB->Get(roptions, cphandles[2], "two", &result));
+  ASSERT_EQ("two", result);
+  for (auto h : cphandles) {
+    delete h;
+  }
+  cphandles.clear();
+  delete snapshotDB;
+  snapshotDB = nullptr;
+}
+
+TEST_F(CheckpointTest, CurrentFileModifiedWhileCheckpointing) {
+  Options options = CurrentOptions();
+  options.max_manifest_file_size = 0;  // always rollover manifest for file add
+  Reopen(options);
+
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->LoadDependency(
+      {// Get past the flush in the checkpoint thread before adding any keys to
+       // the db so the checkpoint thread won't hit the WriteManifest
+       // syncpoints.
+       {"CheckpointImpl::CreateCheckpoint:FlushDone",
+        "CheckpointTest::CurrentFileModifiedWhileCheckpointing:PrePut"},
+       // Roll the manifest during checkpointing right after live files are
+       // snapshotted.
+       {"CheckpointImpl::CreateCheckpoint:SavedLiveFiles1",
+        "VersionSet::LogAndApply:WriteManifest"},
+       {"VersionSet::LogAndApply:WriteManifestDone",
+        "CheckpointImpl::CreateCheckpoint:SavedLiveFiles2"}});
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->EnableProcessing();
+
+  ROCKSDB_NAMESPACE::port::Thread t([&]() {
+    Checkpoint* checkpoint;
+    ASSERT_OK(Checkpoint::Create(db_, &checkpoint));
+    ASSERT_OK(checkpoint->CreateCheckpoint(snapshot_name_));
+    delete checkpoint;
+  });
+  TEST_SYNC_POINT(
+      "CheckpointTest::CurrentFileModifiedWhileCheckpointing:PrePut");
+  ASSERT_OK(Put("Default", "Default1"));
+  ASSERT_OK(Flush());
+  t.join();
+
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->DisableProcessing();
+
+  DB* snapshotDB;
+  // Successful Open() implies that CURRENT pointed to the manifest in the
+  // checkpoint.
+  ASSERT_OK(DB::Open(options, snapshot_name_, &snapshotDB));
+  delete snapshotDB;
+  snapshotDB = nullptr;
+}
+
+TEST_F(CheckpointTest, CurrentFileModifiedWhileCheckpointing2PC) {
+  Close();
+  const std::string dbname = test::PerThreadDBPath("transaction_testdb");
+  ASSERT_OK(DestroyDB(dbname, CurrentOptions()));
+  test::DeleteDir(env_, dbname);
+
+  Options options = CurrentOptions();
+  options.allow_2pc = true;
+  // allow_2pc is implicitly set with tx prepare
+  // options.allow_2pc = true;
+  TransactionDBOptions txn_db_options;
+  TransactionDB* txdb;
+  Status s = TransactionDB::Open(options, txn_db_options, dbname, &txdb);
+  ASSERT_OK(s);
+  ColumnFamilyHandle* cfa;
+  ColumnFamilyHandle* cfb;
+  ColumnFamilyOptions cf_options;
+  ASSERT_OK(txdb->CreateColumnFamily(cf_options, "CFA", &cfa));
+
+  WriteOptions write_options;
+  // Insert something into CFB so lots of log files will be kept
+  // before creating the checkpoint.
+  ASSERT_OK(txdb->CreateColumnFamily(cf_options, "CFB", &cfb));
+  ASSERT_OK(txdb->Put(write_options, cfb, "", ""));
+
+  ReadOptions read_options;
+  std::string value;
+  TransactionOptions txn_options;
+  Transaction* txn = txdb->BeginTransaction(write_options, txn_options);
+  s = txn->SetName("xid");
+  ASSERT_OK(s);
+  ASSERT_EQ(txdb->GetTransactionByName("xid"), txn);
+
+  s = txn->Put(Slice("foo"), Slice("bar"));
+  ASSERT_OK(s);
+  s = txn->Put(cfa, Slice("foocfa"), Slice("barcfa"));
+  ASSERT_OK(s);
+  // Writing prepare into middle of first WAL, then flush WALs many times
+  for (int i = 1; i <= 100000; i++) {
+    Transaction* tx = txdb->BeginTransaction(write_options, txn_options);
+    ASSERT_OK(tx->SetName("x"));
+    ASSERT_OK(tx->Put(Slice(std::to_string(i)), Slice("val")));
+    ASSERT_OK(tx->Put(cfa, Slice("aaa"), Slice("111")));
+    ASSERT_OK(tx->Prepare());
+    ASSERT_OK(tx->Commit());
+    if (i % 10000 == 0) {
+      ASSERT_OK(txdb->Flush(FlushOptions()));
+    }
+    if (i == 88888) {
+      ASSERT_OK(txn->Prepare());
+    }
+    delete tx;
+  }
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->LoadDependency(
+      {{"CheckpointImpl::CreateCheckpoint:SavedLiveFiles1",
+        "CheckpointTest::CurrentFileModifiedWhileCheckpointing2PC:PreCommit"},
+       {"CheckpointTest::CurrentFileModifiedWhileCheckpointing2PC:PostCommit",
+        "CheckpointImpl::CreateCheckpoint:SavedLiveFiles2"}});
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->EnableProcessing();
+  ROCKSDB_NAMESPACE::port::Thread t([&]() {
+    Checkpoint* checkpoint;
+    ASSERT_OK(Checkpoint::Create(txdb, &checkpoint));
+    ASSERT_OK(checkpoint->CreateCheckpoint(snapshot_name_));
+    delete checkpoint;
+  });
+  TEST_SYNC_POINT(
+      "CheckpointTest::CurrentFileModifiedWhileCheckpointing2PC:PreCommit");
+  ASSERT_OK(txn->Commit());
+  delete txn;
+  TEST_SYNC_POINT(
+      "CheckpointTest::CurrentFileModifiedWhileCheckpointing2PC:PostCommit");
+  t.join();
+
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->DisableProcessing();
+
+  // No more than two logs files should exist.
+  std::vector<std::string> files;
+  ASSERT_OK(env_->GetChildren(snapshot_name_, &files));
+  int num_log_files = 0;
+  for (auto& file : files) {
+    uint64_t num;
+    FileType type;
+    WalFileType log_type;
+    if (ParseFileName(file, &num, &type, &log_type) && type == kWalFile) {
+      num_log_files++;
+    }
+  }
+  // One flush after preapare + one outstanding file before checkpoint + one log
+  // file generated after checkpoint.
+  ASSERT_LE(num_log_files, 3);
+
+  TransactionDB* snapshotDB;
+  std::vector<ColumnFamilyDescriptor> column_families;
+  column_families.push_back(
+      ColumnFamilyDescriptor(kDefaultColumnFamilyName, ColumnFamilyOptions()));
+  column_families.push_back(
+      ColumnFamilyDescriptor("CFA", ColumnFamilyOptions()));
+  column_families.push_back(
+      ColumnFamilyDescriptor("CFB", ColumnFamilyOptions()));
+  std::vector<ROCKSDB_NAMESPACE::ColumnFamilyHandle*> cf_handles;
+  ASSERT_OK(TransactionDB::Open(options, txn_db_options, snapshot_name_,
+                                column_families, &cf_handles, &snapshotDB));
+  ASSERT_OK(snapshotDB->Get(read_options, "foo", &value));
+  ASSERT_EQ(value, "bar");
+  ASSERT_OK(snapshotDB->Get(read_options, cf_handles[1], "foocfa", &value));
+  ASSERT_EQ(value, "barcfa");
+
+  delete cfa;
+  delete cfb;
+  delete cf_handles[0];
+  delete cf_handles[1];
+  delete cf_handles[2];
+  delete snapshotDB;
+  snapshotDB = nullptr;
+  delete txdb;
+}
+
+TEST_F(CheckpointTest, CheckpointInvalidDirectoryName) {
+  for (std::string checkpoint_dir : {"", "/", "////"}) {
+    Checkpoint* checkpoint;
+    ASSERT_OK(Checkpoint::Create(db_, &checkpoint));
+    ASSERT_TRUE(
+        checkpoint->CreateCheckpoint(checkpoint_dir).IsInvalidArgument());
+    delete checkpoint;
+  }
+}
+
+TEST_F(CheckpointTest, CheckpointWithParallelWrites) {
+  // When run with TSAN, this exposes the data race fixed in
+  // https://github.com/facebook/rocksdb/pull/3603
+  ASSERT_OK(Put("key1", "val1"));
+  port::Thread thread([this]() { ASSERT_OK(Put("key2", "val2")); });
+  Checkpoint* checkpoint;
+  ASSERT_OK(Checkpoint::Create(db_, &checkpoint));
+  ASSERT_OK(checkpoint->CreateCheckpoint(snapshot_name_));
+  delete checkpoint;
+  thread.join();
+}
+
+TEST_F(CheckpointTest, CheckpointWithUnsyncedDataDropped) {
+  Options options = CurrentOptions();
+  std::unique_ptr<FaultInjectionTestEnv> env(new FaultInjectionTestEnv(env_));
+  options.env = env.get();
+  Reopen(options);
+  ASSERT_OK(Put("key1", "val1"));
+  Checkpoint* checkpoint;
+  ASSERT_OK(Checkpoint::Create(db_, &checkpoint));
+  ASSERT_OK(checkpoint->CreateCheckpoint(snapshot_name_));
+  delete checkpoint;
+  ASSERT_OK(env->DropUnsyncedFileData());
+
+  // make sure it's openable even though whatever data that wasn't synced got
+  // dropped.
+  options.env = env_;
+  DB* snapshot_db;
+  ASSERT_OK(DB::Open(options, snapshot_name_, &snapshot_db));
+  ReadOptions read_opts;
+  std::string get_result;
+  ASSERT_OK(snapshot_db->Get(read_opts, "key1", &get_result));
+  ASSERT_EQ("val1", get_result);
+  delete snapshot_db;
+  delete db_;
+  db_ = nullptr;
+}
+
+TEST_F(CheckpointTest, CheckpointOptionsFileFailedToPersist) {
+  // Regression test for a bug where checkpoint failed on a DB where persisting
+  // OPTIONS file failed and the DB was opened with
+  // `fail_if_options_file_error == false`.
+  Options options = CurrentOptions();
+  options.fail_if_options_file_error = false;
+  auto fault_fs = std::make_shared<FaultInjectionTestFS>(FileSystem::Default());
+
+  // Setup `FaultInjectionTestFS` and `SyncPoint` callbacks to fail one
+  // operation when inside the OPTIONS file persisting code.
+  std::unique_ptr<Env> fault_fs_env(NewCompositeEnv(fault_fs));
+  fault_fs->SetRandomMetadataWriteError(1 /* one_in */);
+  SyncPoint::GetInstance()->SetCallBack(
+      "PersistRocksDBOptions:start", [fault_fs](void* /* arg */) {
+        fault_fs->EnableMetadataWriteErrorInjection();
+      });
+  SyncPoint::GetInstance()->SetCallBack(
+      "FaultInjectionTestFS::InjectMetadataWriteError:Injected",
+      [fault_fs](void* /* arg */) {
+        fault_fs->DisableMetadataWriteErrorInjection();
+      });
+  options.env = fault_fs_env.get();
+  SyncPoint::GetInstance()->EnableProcessing();
+
+  Reopen(options);
+  ASSERT_OK(Put("key1", "val1"));
+  Checkpoint* checkpoint;
+  ASSERT_OK(Checkpoint::Create(db_, &checkpoint));
+  ASSERT_OK(checkpoint->CreateCheckpoint(snapshot_name_));
+  delete checkpoint;
+
+  // Make sure it's usable.
+  options.env = env_;
+  DB* snapshot_db;
+  ASSERT_OK(DB::Open(options, snapshot_name_, &snapshot_db));
+  ReadOptions read_opts;
+  std::string get_result;
+  ASSERT_OK(snapshot_db->Get(read_opts, "key1", &get_result));
+  ASSERT_EQ("val1", get_result);
+  delete snapshot_db;
+  delete db_;
+  db_ = nullptr;
+}
+
+TEST_F(CheckpointTest, CheckpointReadOnlyDB) {
+  ASSERT_OK(Put("foo", "foo_value"));
+  ASSERT_OK(Flush());
+  Close();
+  Options options = CurrentOptions();
+  ASSERT_OK(ReadOnlyReopen(options));
+  Checkpoint* checkpoint = nullptr;
+  ASSERT_OK(Checkpoint::Create(db_, &checkpoint));
+  ASSERT_OK(checkpoint->CreateCheckpoint(snapshot_name_));
+  delete checkpoint;
+  checkpoint = nullptr;
+  Close();
+  DB* snapshot_db = nullptr;
+  ASSERT_OK(DB::Open(options, snapshot_name_, &snapshot_db));
+  ReadOptions read_opts;
+  std::string get_result;
+  ASSERT_OK(snapshot_db->Get(read_opts, "foo", &get_result));
+  ASSERT_EQ("foo_value", get_result);
+  delete snapshot_db;
+}
+
+TEST_F(CheckpointTest, CheckpointReadOnlyDBWithMultipleColumnFamilies) {
+  Options options = CurrentOptions();
+  CreateAndReopenWithCF({"pikachu", "eevee"}, options);
+  for (int i = 0; i != 3; ++i) {
+    ASSERT_OK(Put(i, "foo", "foo_value"));
+    ASSERT_OK(Flush(i));
+  }
+  Close();
+  Status s = ReadOnlyReopenWithColumnFamilies(
+      {kDefaultColumnFamilyName, "pikachu", "eevee"}, options);
+  ASSERT_OK(s);
+  Checkpoint* checkpoint = nullptr;
+  ASSERT_OK(Checkpoint::Create(db_, &checkpoint));
+  ASSERT_OK(checkpoint->CreateCheckpoint(snapshot_name_));
+  delete checkpoint;
+  checkpoint = nullptr;
+  Close();
+
+  std::vector<ColumnFamilyDescriptor> column_families{
+      {kDefaultColumnFamilyName, options},
+      {"pikachu", options},
+      {"eevee", options}};
+  DB* snapshot_db = nullptr;
+  std::vector<ColumnFamilyHandle*> snapshot_handles;
+  s = DB::Open(options, snapshot_name_, column_families, &snapshot_handles,
+               &snapshot_db);
+  ASSERT_OK(s);
+  ReadOptions read_opts;
+  for (int i = 0; i != 3; ++i) {
+    std::string get_result;
+    s = snapshot_db->Get(read_opts, snapshot_handles[i], "foo", &get_result);
+    ASSERT_OK(s);
+    ASSERT_EQ("foo_value", get_result);
+  }
+
+  for (auto snapshot_h : snapshot_handles) {
+    delete snapshot_h;
+  }
+  snapshot_handles.clear();
+  delete snapshot_db;
+}
+
+TEST_F(CheckpointTest, CheckpointWithDbPath) {
+  Options options = CurrentOptions();
+  options.db_paths.emplace_back(dbname_ + "_2", 0);
+  Reopen(options);
+  ASSERT_OK(Put("key1", "val1"));
+  Flush();
+  Checkpoint* checkpoint;
+  ASSERT_OK(Checkpoint::Create(db_, &checkpoint));
+  // Currently not supported
+  ASSERT_TRUE(checkpoint->CreateCheckpoint(snapshot_name_).IsNotSupported());
+  delete checkpoint;
+}
+
+TEST_F(CheckpointTest, PutRaceWithCheckpointTrackedWalSync) {
+  // Repro for a race condition where a user write comes in after the checkpoint
+  // syncs WAL for `track_and_verify_wals_in_manifest` but before the
+  // corresponding MANIFEST update. With the bug, that scenario resulted in an
+  // unopenable DB with error "Corruption: Size mismatch: WAL ...".
+  Options options = CurrentOptions();
+  std::unique_ptr<FaultInjectionTestEnv> fault_env(
+      new FaultInjectionTestEnv(env_));
+  options.env = fault_env.get();
+  options.track_and_verify_wals_in_manifest = true;
+  Reopen(options);
+
+  ASSERT_OK(Put("key1", "val1"));
+
+  SyncPoint::GetInstance()->SetCallBack(
+      "DBImpl::SyncWAL:BeforeMarkLogsSynced:1",
+      [this](void* /* arg */) { ASSERT_OK(Put("key2", "val2")); });
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->EnableProcessing();
+
+  std::unique_ptr<Checkpoint> checkpoint;
+  {
+    Checkpoint* checkpoint_ptr;
+    ASSERT_OK(Checkpoint::Create(db_, &checkpoint_ptr));
+    checkpoint.reset(checkpoint_ptr);
+  }
+
+  ASSERT_OK(checkpoint->CreateCheckpoint(snapshot_name_));
+
+  // Ensure callback ran.
+  ASSERT_EQ("val2", Get("key2"));
+
+  Close();
+
+  // Simulate full loss of unsynced data. This drops "key2" -> "val2" from the
+  // DB WAL.
+  fault_env->DropUnsyncedFileData();
+
+  // Before the bug fix, reopening the DB would fail because the MANIFEST's
+  // AddWal entry indicated the WAL should be synced through "key2" -> "val2".
+  Reopen(options);
+
+  // Need to close before `fault_env` goes out of scope.
+  Close();
+}
+
+}  // namespace ROCKSDB_NAMESPACE
+
+int main(int argc, char** argv) {
+  ROCKSDB_NAMESPACE::port::InstallStackTraceHandler();
+  ::testing::InitGoogleTest(&argc, argv);
+  return RUN_ALL_TESTS();
+}
+
+#else
+#include <stdio.h>
+
+int main(int /*argc*/, char** /*argv*/) {
+  fprintf(stderr, "SKIPPED as Checkpoint is not supported in ROCKSDB_LITE\n");
+  return 0;
+}
+
+#endif  // !ROCKSDB_LITE
diff --git a/src/rocksdb/utilities/compaction_filters.cc b/src/rocksdb/utilities/compaction_filters.cc
new file mode 100644
index 000000000..8763901c3
--- /dev/null
+++ b/src/rocksdb/utilities/compaction_filters.cc
@@ -0,0 +1,56 @@
+// Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.
+// Copyright (c) 2011 The LevelDB Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file. See the AUTHORS file for names of contributors.
+
+#include <memory>
+
+#include "rocksdb/compaction_filter.h"
+#include "rocksdb/options.h"
+#include "rocksdb/utilities/customizable_util.h"
+#include "rocksdb/utilities/options_type.h"
+#include "utilities/compaction_filters/layered_compaction_filter_base.h"
+#include "utilities/compaction_filters/remove_emptyvalue_compactionfilter.h"
+
+namespace ROCKSDB_NAMESPACE {
+#ifndef ROCKSDB_LITE
+static int RegisterBuiltinCompactionFilters(ObjectLibrary& library,
+                                            const std::string& /*arg*/) {
+  library.AddFactory<CompactionFilter>(
+      RemoveEmptyValueCompactionFilter::kClassName(),
+      [](const std::string& /*uri*/,
+         std::unique_ptr<CompactionFilter>* /*guard*/,
+         std::string* /*errmsg*/) {
+        return new RemoveEmptyValueCompactionFilter();
+      });
+  return 1;
+}
+#endif  // ROCKSDB_LITE
+Status CompactionFilter::CreateFromString(const ConfigOptions& config_options,
+                                          const std::string& value,
+                                          const CompactionFilter** result) {
+#ifndef ROCKSDB_LITE
+  static std::once_flag once;
+  std::call_once(once, [&]() {
+    RegisterBuiltinCompactionFilters(*(ObjectLibrary::Default().get()), "");
+  });
+#endif  // ROCKSDB_LITE
+  CompactionFilter* filter = const_cast<CompactionFilter*>(*result);
+  Status status = LoadStaticObject<CompactionFilter>(config_options, value,
+                                                     nullptr, &filter);
+  if (status.ok()) {
+    *result = const_cast<CompactionFilter*>(filter);
+  }
+  return status;
+}
+
+Status CompactionFilterFactory::CreateFromString(
+    const ConfigOptions& config_options, const std::string& value,
+    std::shared_ptr<CompactionFilterFactory>* result) {
+  // Currently there are no builtin CompactionFilterFactories.
+  // If any are introduced, they need to be registered here.
+  Status status = LoadSharedObject<CompactionFilterFactory>(
+      config_options, value, nullptr, result);
+  return status;
+}
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/utilities/compaction_filters/layered_compaction_filter_base.h b/src/rocksdb/utilities/compaction_filters/layered_compaction_filter_base.h
new file mode 100644
index 000000000..803fa94ae
--- /dev/null
+++ b/src/rocksdb/utilities/compaction_filters/layered_compaction_filter_base.h
@@ -0,0 +1,41 @@
+// Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.
+// Copyright (c) 2011 The LevelDB Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file. See the AUTHORS file for names of contributors.
+
+#pragma once
+#include <memory>
+
+#include "rocksdb/compaction_filter.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+// Abstract base class for building layered compaction filter on top of
+// user compaction filter.
+// See BlobIndexCompactionFilter or TtlCompactionFilter for a basic usage.
+class LayeredCompactionFilterBase : public CompactionFilter {
+ public:
+  LayeredCompactionFilterBase(
+      const CompactionFilter* _user_comp_filter,
+      std::unique_ptr<const CompactionFilter> _user_comp_filter_from_factory)
+      : user_comp_filter_(_user_comp_filter),
+        user_comp_filter_from_factory_(
+            std::move(_user_comp_filter_from_factory)) {
+    if (!user_comp_filter_) {
+      user_comp_filter_ = user_comp_filter_from_factory_.get();
+    }
+  }
+
+  // Return a pointer to user compaction filter
+  const CompactionFilter* user_comp_filter() const { return user_comp_filter_; }
+
+  const Customizable* Inner() const override { return user_comp_filter_; }
+
+ protected:
+  const CompactionFilter* user_comp_filter_;
+
+ private:
+  std::unique_ptr<const CompactionFilter> user_comp_filter_from_factory_;
+};
+
+}  //  namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/utilities/compaction_filters/remove_emptyvalue_compactionfilter.cc b/src/rocksdb/utilities/compaction_filters/remove_emptyvalue_compactionfilter.cc
new file mode 100644
index 000000000..b788dbf9b
--- /dev/null
+++ b/src/rocksdb/utilities/compaction_filters/remove_emptyvalue_compactionfilter.cc
@@ -0,0 +1,26 @@
+// Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#ifndef ROCKSDB_LITE
+
+#include "utilities/compaction_filters/remove_emptyvalue_compactionfilter.h"
+
+#include <string>
+
+#include "rocksdb/slice.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+bool RemoveEmptyValueCompactionFilter::Filter(int /*level*/,
+                                              const Slice& /*key*/,
+                                              const Slice& existing_value,
+                                              std::string* /*new_value*/,
+                                              bool* /*value_changed*/) const {
+  // remove kv pairs that have empty values
+  return existing_value.empty();
+}
+
+}  // namespace ROCKSDB_NAMESPACE
+#endif  // !ROCKSDB_LITE
diff --git a/src/rocksdb/utilities/compaction_filters/remove_emptyvalue_compactionfilter.h b/src/rocksdb/utilities/compaction_filters/remove_emptyvalue_compactionfilter.h
new file mode 100644
index 000000000..864ad15ff
--- /dev/null
+++ b/src/rocksdb/utilities/compaction_filters/remove_emptyvalue_compactionfilter.h
@@ -0,0 +1,28 @@
+// Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#ifndef ROCKSDB_LITE
+
+#pragma once
+
+#include <string>
+
+#include "rocksdb/compaction_filter.h"
+#include "rocksdb/slice.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+class RemoveEmptyValueCompactionFilter : public CompactionFilter {
+ public:
+  static const char* kClassName() { return "RemoveEmptyValueCompactionFilter"; }
+
+  const char* Name() const override { return kClassName(); }
+
+  bool Filter(int level, const Slice& key, const Slice& existing_value,
+              std::string* new_value, bool* value_changed) const override;
+};
+
+}  // namespace ROCKSDB_NAMESPACE
+#endif  // !ROCKSDB_LITE
diff --git a/src/rocksdb/utilities/convenience/info_log_finder.cc b/src/rocksdb/utilities/convenience/info_log_finder.cc
new file mode 100644
index 000000000..fe62fd561
--- /dev/null
+++ b/src/rocksdb/utilities/convenience/info_log_finder.cc
@@ -0,0 +1,26 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+//
+// Copyright (c) 2012 Facebook.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file.
+
+#include "rocksdb/utilities/info_log_finder.h"
+
+#include "file/filename.h"
+#include "rocksdb/env.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+Status GetInfoLogList(DB* db, std::vector<std::string>* info_log_list) {
+  if (!db) {
+    return Status::InvalidArgument("DB pointer is not valid");
+  }
+  std::string parent_path;
+  const Options& options = db->GetOptions();
+  return GetInfoLogFiles(options.env->GetFileSystem(), options.db_log_dir,
+                         db->GetName(), &parent_path, info_log_list);
+}
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/utilities/counted_fs.cc b/src/rocksdb/utilities/counted_fs.cc
new file mode 100644
index 000000000..e43f3a191
--- /dev/null
+++ b/src/rocksdb/utilities/counted_fs.cc
@@ -0,0 +1,379 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#include "utilities/counted_fs.h"
+
+#include <sstream>
+
+#include "rocksdb/file_system.h"
+#include "rocksdb/utilities/options_type.h"
+
+namespace ROCKSDB_NAMESPACE {
+namespace {
+class CountedSequentialFile : public FSSequentialFileOwnerWrapper {
+ private:
+  CountedFileSystem* fs_;
+
+ public:
+  CountedSequentialFile(std::unique_ptr<FSSequentialFile>&& f,
+                        CountedFileSystem* fs)
+      : FSSequentialFileOwnerWrapper(std::move(f)), fs_(fs) {}
+
+  ~CountedSequentialFile() override { fs_->counters()->closes++; }
+
+  IOStatus Read(size_t n, const IOOptions& options, Slice* result,
+                char* scratch, IODebugContext* dbg) override {
+    IOStatus rv = target()->Read(n, options, result, scratch, dbg);
+    fs_->counters()->reads.RecordOp(rv, result->size());
+    return rv;
+  }
+
+  IOStatus PositionedRead(uint64_t offset, size_t n, const IOOptions& options,
+                          Slice* result, char* scratch,
+                          IODebugContext* dbg) override {
+    IOStatus rv =
+        target()->PositionedRead(offset, n, options, result, scratch, dbg);
+    fs_->counters()->reads.RecordOp(rv, result->size());
+    return rv;
+  }
+};
+
+class CountedRandomAccessFile : public FSRandomAccessFileOwnerWrapper {
+ private:
+  CountedFileSystem* fs_;
+
+ public:
+  CountedRandomAccessFile(std::unique_ptr<FSRandomAccessFile>&& f,
+                          CountedFileSystem* fs)
+      : FSRandomAccessFileOwnerWrapper(std::move(f)), fs_(fs) {}
+
+  ~CountedRandomAccessFile() override { fs_->counters()->closes++; }
+
+  IOStatus Read(uint64_t offset, size_t n, const IOOptions& options,
+                Slice* result, char* scratch,
+                IODebugContext* dbg) const override {
+    IOStatus rv = target()->Read(offset, n, options, result, scratch, dbg);
+    fs_->counters()->reads.RecordOp(rv, result->size());
+    return rv;
+  }
+
+  IOStatus MultiRead(FSReadRequest* reqs, size_t num_reqs,
+                     const IOOptions& options, IODebugContext* dbg) override {
+    IOStatus rv = target()->MultiRead(reqs, num_reqs, options, dbg);
+    for (size_t r = 0; r < num_reqs; r++) {
+      fs_->counters()->reads.RecordOp(reqs[r].status, reqs[r].result.size());
+    }
+    return rv;
+  }
+};
+
+class CountedWritableFile : public FSWritableFileOwnerWrapper {
+ private:
+  CountedFileSystem* fs_;
+
+ public:
+  CountedWritableFile(std::unique_ptr<FSWritableFile>&& f,
+                      CountedFileSystem* fs)
+      : FSWritableFileOwnerWrapper(std::move(f)), fs_(fs) {}
+
+  IOStatus Append(const Slice& data, const IOOptions& options,
+                  IODebugContext* dbg) override {
+    IOStatus rv = target()->Append(data, options, dbg);
+    fs_->counters()->writes.RecordOp(rv, data.size());
+    return rv;
+  }
+
+  IOStatus Append(const Slice& data, const IOOptions& options,
+                  const DataVerificationInfo& info,
+                  IODebugContext* dbg) override {
+    IOStatus rv = target()->Append(data, options, info, dbg);
+    fs_->counters()->writes.RecordOp(rv, data.size());
+    return rv;
+  }
+
+  IOStatus PositionedAppend(const Slice& data, uint64_t offset,
+                            const IOOptions& options,
+                            IODebugContext* dbg) override {
+    IOStatus rv = target()->PositionedAppend(data, offset, options, dbg);
+    fs_->counters()->writes.RecordOp(rv, data.size());
+    return rv;
+  }
+
+  IOStatus PositionedAppend(const Slice& data, uint64_t offset,
+                            const IOOptions& options,
+                            const DataVerificationInfo& info,
+                            IODebugContext* dbg) override {
+    IOStatus rv = target()->PositionedAppend(data, offset, options, info, dbg);
+    fs_->counters()->writes.RecordOp(rv, data.size());
+    return rv;
+  }
+
+  IOStatus Close(const IOOptions& options, IODebugContext* dbg) override {
+    IOStatus rv = target()->Close(options, dbg);
+    if (rv.ok()) {
+      fs_->counters()->closes++;
+    }
+    return rv;
+  }
+
+  IOStatus Flush(const IOOptions& options, IODebugContext* dbg) override {
+    IOStatus rv = target()->Flush(options, dbg);
+    if (rv.ok()) {
+      fs_->counters()->flushes++;
+    }
+    return rv;
+  }
+
+  IOStatus Sync(const IOOptions& options, IODebugContext* dbg) override {
+    IOStatus rv = target()->Sync(options, dbg);
+    if (rv.ok()) {
+      fs_->counters()->syncs++;
+    }
+    return rv;
+  }
+
+  IOStatus Fsync(const IOOptions& options, IODebugContext* dbg) override {
+    IOStatus rv = target()->Fsync(options, dbg);
+    if (rv.ok()) {
+      fs_->counters()->fsyncs++;
+    }
+    return rv;
+  }
+
+  IOStatus RangeSync(uint64_t offset, uint64_t nbytes, const IOOptions& options,
+                     IODebugContext* dbg) override {
+    IOStatus rv = target()->RangeSync(offset, nbytes, options, dbg);
+    if (rv.ok()) {
+      fs_->counters()->syncs++;
+    }
+    return rv;
+  }
+};
+
+class CountedRandomRWFile : public FSRandomRWFileOwnerWrapper {
+ private:
+  mutable CountedFileSystem* fs_;
+
+ public:
+  CountedRandomRWFile(std::unique_ptr<FSRandomRWFile>&& f,
+                      CountedFileSystem* fs)
+      : FSRandomRWFileOwnerWrapper(std::move(f)), fs_(fs) {}
+  IOStatus Write(uint64_t offset, const Slice& data, const IOOptions& options,
+                 IODebugContext* dbg) override {
+    IOStatus rv = target()->Write(offset, data, options, dbg);
+    fs_->counters()->writes.RecordOp(rv, data.size());
+    return rv;
+  }
+
+  IOStatus Read(uint64_t offset, size_t n, const IOOptions& options,
+                Slice* result, char* scratch,
+                IODebugContext* dbg) const override {
+    IOStatus rv = target()->Read(offset, n, options, result, scratch, dbg);
+    fs_->counters()->reads.RecordOp(rv, result->size());
+    return rv;
+  }
+
+  IOStatus Flush(const IOOptions& options, IODebugContext* dbg) override {
+    IOStatus rv = target()->Flush(options, dbg);
+    if (rv.ok()) {
+      fs_->counters()->flushes++;
+    }
+    return rv;
+  }
+
+  IOStatus Sync(const IOOptions& options, IODebugContext* dbg) override {
+    IOStatus rv = target()->Sync(options, dbg);
+    if (rv.ok()) {
+      fs_->counters()->syncs++;
+    }
+    return rv;
+  }
+
+  IOStatus Fsync(const IOOptions& options, IODebugContext* dbg) override {
+    IOStatus rv = target()->Fsync(options, dbg);
+    if (rv.ok()) {
+      fs_->counters()->fsyncs++;
+    }
+    return rv;
+  }
+
+  IOStatus Close(const IOOptions& options, IODebugContext* dbg) override {
+    IOStatus rv = target()->Close(options, dbg);
+    if (rv.ok()) {
+      fs_->counters()->closes++;
+    }
+    return rv;
+  }
+};
+
+class CountedDirectory : public FSDirectoryWrapper {
+ private:
+  mutable CountedFileSystem* fs_;
+  bool closed_ = false;
+
+ public:
+  CountedDirectory(std::unique_ptr<FSDirectory>&& f, CountedFileSystem* fs)
+      : FSDirectoryWrapper(std::move(f)), fs_(fs) {}
+
+  IOStatus Fsync(const IOOptions& options, IODebugContext* dbg) override {
+    IOStatus rv = FSDirectoryWrapper::Fsync(options, dbg);
+    if (rv.ok()) {
+      fs_->counters()->dsyncs++;
+    }
+    return rv;
+  }
+
+  IOStatus Close(const IOOptions& options, IODebugContext* dbg) override {
+    IOStatus rv = FSDirectoryWrapper::Close(options, dbg);
+    if (rv.ok()) {
+      fs_->counters()->closes++;
+      fs_->counters()->dir_closes++;
+      closed_ = true;
+    }
+    return rv;
+  }
+
+  IOStatus FsyncWithDirOptions(const IOOptions& options, IODebugContext* dbg,
+                               const DirFsyncOptions& dir_options) override {
+    IOStatus rv =
+        FSDirectoryWrapper::FsyncWithDirOptions(options, dbg, dir_options);
+    if (rv.ok()) {
+      fs_->counters()->dsyncs++;
+    }
+    return rv;
+  }
+
+  ~CountedDirectory() {
+    if (!closed_) {
+      // TODO: fix DB+CF code to use explicit Close, not rely on destructor
+      fs_->counters()->closes++;
+      fs_->counters()->dir_closes++;
+    }
+  }
+};
+}  // anonymous namespace
+
+std::string FileOpCounters::PrintCounters() const {
+  std::stringstream ss;
+  ss << "Num files opened: " << opens.load(std::memory_order_relaxed)
+     << std::endl;
+  ss << "Num files deleted: " << deletes.load(std::memory_order_relaxed)
+     << std::endl;
+  ss << "Num files renamed: " << renames.load(std::memory_order_relaxed)
+     << std::endl;
+  ss << "Num Flush(): " << flushes.load(std::memory_order_relaxed) << std::endl;
+  ss << "Num Sync(): " << syncs.load(std::memory_order_relaxed) << std::endl;
+  ss << "Num Fsync(): " << fsyncs.load(std::memory_order_relaxed) << std::endl;
+  ss << "Num Dir Fsync(): " << dsyncs.load(std::memory_order_relaxed)
+     << std::endl;
+  ss << "Num Close(): " << closes.load(std::memory_order_relaxed) << std::endl;
+  ss << "Num Dir Open(): " << dir_opens.load(std::memory_order_relaxed)
+     << std::endl;
+  ss << "Num Dir Close(): " << dir_closes.load(std::memory_order_relaxed)
+     << std::endl;
+  ss << "Num Read(): " << reads.ops.load(std::memory_order_relaxed)
+     << std::endl;
+  ss << "Num Append(): " << writes.ops.load(std::memory_order_relaxed)
+     << std::endl;
+  ss << "Num bytes read: " << reads.bytes.load(std::memory_order_relaxed)
+     << std::endl;
+  ss << "Num bytes written: " << writes.bytes.load(std::memory_order_relaxed)
+     << std::endl;
+  return ss.str();
+}
+
+CountedFileSystem::CountedFileSystem(const std::shared_ptr<FileSystem>& base)
+    : FileSystemWrapper(base) {}
+
+IOStatus CountedFileSystem::NewSequentialFile(
+    const std::string& f, const FileOptions& options,
+    std::unique_ptr<FSSequentialFile>* r, IODebugContext* dbg) {
+  std::unique_ptr<FSSequentialFile> base;
+  IOStatus s = target()->NewSequentialFile(f, options, &base, dbg);
+  if (s.ok()) {
+    counters_.opens++;
+    r->reset(new CountedSequentialFile(std::move(base), this));
+  }
+  return s;
+}
+
+IOStatus CountedFileSystem::NewRandomAccessFile(
+    const std::string& f, const FileOptions& options,
+    std::unique_ptr<FSRandomAccessFile>* r, IODebugContext* dbg) {
+  std::unique_ptr<FSRandomAccessFile> base;
+  IOStatus s = target()->NewRandomAccessFile(f, options, &base, dbg);
+  if (s.ok()) {
+    counters_.opens++;
+    r->reset(new CountedRandomAccessFile(std::move(base), this));
+  }
+  return s;
+}
+
+IOStatus CountedFileSystem::NewWritableFile(const std::string& f,
+                                            const FileOptions& options,
+                                            std::unique_ptr<FSWritableFile>* r,
+                                            IODebugContext* dbg) {
+  std::unique_ptr<FSWritableFile> base;
+  IOStatus s = target()->NewWritableFile(f, options, &base, dbg);
+  if (s.ok()) {
+    counters_.opens++;
+    r->reset(new CountedWritableFile(std::move(base), this));
+  }
+  return s;
+}
+
+IOStatus CountedFileSystem::ReopenWritableFile(
+    const std::string& fname, const FileOptions& options,
+    std::unique_ptr<FSWritableFile>* result, IODebugContext* dbg) {
+  std::unique_ptr<FSWritableFile> base;
+  IOStatus s = target()->ReopenWritableFile(fname, options, &base, dbg);
+  if (s.ok()) {
+    counters_.opens++;
+    result->reset(new CountedWritableFile(std::move(base), this));
+  }
+  return s;
+}
+
+IOStatus CountedFileSystem::ReuseWritableFile(
+    const std::string& fname, const std::string& old_fname,
+    const FileOptions& options, std::unique_ptr<FSWritableFile>* result,
+    IODebugContext* dbg) {
+  std::unique_ptr<FSWritableFile> base;
+  IOStatus s =
+      target()->ReuseWritableFile(fname, old_fname, options, &base, dbg);
+  if (s.ok()) {
+    counters_.opens++;
+    result->reset(new CountedWritableFile(std::move(base), this));
+  }
+  return s;
+}
+
+IOStatus CountedFileSystem::NewRandomRWFile(
+    const std::string& name, const FileOptions& options,
+    std::unique_ptr<FSRandomRWFile>* result, IODebugContext* dbg) {
+  std::unique_ptr<FSRandomRWFile> base;
+  IOStatus s = target()->NewRandomRWFile(name, options, &base, dbg);
+  if (s.ok()) {
+    counters_.opens++;
+    result->reset(new CountedRandomRWFile(std::move(base), this));
+  }
+  return s;
+}
+
+IOStatus CountedFileSystem::NewDirectory(const std::string& name,
+                                         const IOOptions& options,
+                                         std::unique_ptr<FSDirectory>* result,
+                                         IODebugContext* dbg) {
+  std::unique_ptr<FSDirectory> base;
+  IOStatus s = target()->NewDirectory(name, options, &base, dbg);
+  if (s.ok()) {
+    counters_.opens++;
+    counters_.dir_opens++;
+    result->reset(new CountedDirectory(std::move(base), this));
+  }
+  return s;
+}
+
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/utilities/counted_fs.h b/src/rocksdb/utilities/counted_fs.h
new file mode 100644
index 000000000..cb8a8968f
--- /dev/null
+++ b/src/rocksdb/utilities/counted_fs.h
@@ -0,0 +1,158 @@
+//  Copyright (c) 2016-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#pragma once
+
+#include <atomic>
+#include <memory>
+
+#include "rocksdb/file_system.h"
+#include "rocksdb/io_status.h"
+#include "rocksdb/rocksdb_namespace.h"
+
+namespace ROCKSDB_NAMESPACE {
+class Logger;
+
+struct OpCounter {
+  std::atomic<int> ops;
+  std::atomic<uint64_t> bytes;
+
+  OpCounter() : ops(0), bytes(0) {}
+
+  void Reset() {
+    ops = 0;
+    bytes = 0;
+  }
+  void RecordOp(const IOStatus& io_s, size_t added_bytes) {
+    if (!io_s.IsNotSupported()) {
+      ops.fetch_add(1, std::memory_order_relaxed);
+    }
+    if (io_s.ok()) {
+      bytes.fetch_add(added_bytes, std::memory_order_relaxed);
+    }
+  }
+};
+
+struct FileOpCounters {
+  static const char* kName() { return "FileOpCounters"; }
+
+  std::atomic<int> opens;
+  std::atomic<int> closes;
+  std::atomic<int> deletes;
+  std::atomic<int> renames;
+  std::atomic<int> flushes;
+  std::atomic<int> syncs;
+  std::atomic<int> dsyncs;
+  std::atomic<int> fsyncs;
+  std::atomic<int> dir_opens;
+  std::atomic<int> dir_closes;
+  OpCounter reads;
+  OpCounter writes;
+
+  FileOpCounters()
+      : opens(0),
+        closes(0),
+        deletes(0),
+        renames(0),
+        flushes(0),
+        syncs(0),
+        dsyncs(0),
+        fsyncs(0),
+        dir_opens(0),
+        dir_closes(0) {}
+
+  void Reset() {
+    opens = 0;
+    closes = 0;
+    deletes = 0;
+    renames = 0;
+    flushes = 0;
+    syncs = 0;
+    dsyncs = 0;
+    fsyncs = 0;
+    dir_opens = 0;
+    dir_closes = 0;
+    reads.Reset();
+    writes.Reset();
+  }
+  std::string PrintCounters() const;
+};
+
+// A FileSystem class that counts operations (reads, writes, opens, closes, etc)
+class CountedFileSystem : public FileSystemWrapper {
+ public:
+ private:
+  FileOpCounters counters_;
+
+ public:
+  explicit CountedFileSystem(const std::shared_ptr<FileSystem>& base);
+  static const char* kClassName() { return "CountedFileSystem"; }
+  const char* Name() const override { return kClassName(); }
+
+  IOStatus NewSequentialFile(const std::string& f, const FileOptions& options,
+                             std::unique_ptr<FSSequentialFile>* r,
+                             IODebugContext* dbg) override;
+
+  IOStatus NewRandomAccessFile(const std::string& f,
+                               const FileOptions& file_opts,
+                               std::unique_ptr<FSRandomAccessFile>* r,
+                               IODebugContext* dbg) override;
+
+  IOStatus NewWritableFile(const std::string& f, const FileOptions& options,
+                           std::unique_ptr<FSWritableFile>* r,
+                           IODebugContext* dbg) override;
+  IOStatus ReopenWritableFile(const std::string& fname,
+                              const FileOptions& options,
+                              std::unique_ptr<FSWritableFile>* result,
+                              IODebugContext* dbg) override;
+
+  IOStatus ReuseWritableFile(const std::string& fname,
+                             const std::string& old_fname,
+                             const FileOptions& file_opts,
+                             std::unique_ptr<FSWritableFile>* result,
+                             IODebugContext* dbg) override;
+  IOStatus NewRandomRWFile(const std::string& name, const FileOptions& options,
+                           std::unique_ptr<FSRandomRWFile>* result,
+                           IODebugContext* dbg) override;
+
+  IOStatus NewDirectory(const std::string& name, const IOOptions& io_opts,
+                        std::unique_ptr<FSDirectory>* result,
+                        IODebugContext* dbg) override;
+
+  IOStatus DeleteFile(const std::string& fname, const IOOptions& options,
+                      IODebugContext* dbg) override {
+    IOStatus s = target()->DeleteFile(fname, options, dbg);
+    if (s.ok()) {
+      counters_.deletes++;
+    }
+    return s;
+  }
+
+  IOStatus RenameFile(const std::string& s, const std::string& t,
+                      const IOOptions& options, IODebugContext* dbg) override {
+    IOStatus st = target()->RenameFile(s, t, options, dbg);
+    if (st.ok()) {
+      counters_.renames++;
+    }
+    return st;
+  }
+
+  const FileOpCounters* counters() const { return &counters_; }
+
+  FileOpCounters* counters() { return &counters_; }
+
+  const void* GetOptionsPtr(const std::string& name) const override {
+    if (name == FileOpCounters::kName()) {
+      return counters();
+    } else {
+      return FileSystemWrapper::GetOptionsPtr(name);
+    }
+  }
+
+  // Prints the counters to a string
+  std::string PrintCounters() const { return counters_.PrintCounters(); }
+  void ResetCounters() { counters_.Reset(); }
+};
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/utilities/debug.cc b/src/rocksdb/utilities/debug.cc
new file mode 100644
index 000000000..f2c3bb513
--- /dev/null
+++ b/src/rocksdb/utilities/debug.cc
@@ -0,0 +1,120 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#ifndef ROCKSDB_LITE
+
+#include "rocksdb/utilities/debug.h"
+
+#include "db/db_impl/db_impl.h"
+#include "rocksdb/utilities/options_type.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+static std::unordered_map<std::string, ValueType> value_type_string_map = {
+    {"TypeDeletion", ValueType::kTypeDeletion},
+    {"TypeValue", ValueType::kTypeValue},
+    {"TypeMerge", ValueType::kTypeMerge},
+    {"TypeLogData", ValueType::kTypeLogData},
+    {"TypeColumnFamilyDeletion", ValueType::kTypeColumnFamilyDeletion},
+    {"TypeColumnFamilyValue", ValueType::kTypeColumnFamilyValue},
+    {"TypeColumnFamilyMerge", ValueType::kTypeColumnFamilyMerge},
+    {"TypeSingleDeletion", ValueType::kTypeSingleDeletion},
+    {"TypeColumnFamilySingleDeletion",
+     ValueType::kTypeColumnFamilySingleDeletion},
+    {"TypeBeginPrepareXID", ValueType::kTypeBeginPrepareXID},
+    {"TypeEndPrepareXID", ValueType::kTypeEndPrepareXID},
+    {"TypeCommitXID", ValueType::kTypeCommitXID},
+    {"TypeRollbackXID", ValueType::kTypeRollbackXID},
+    {"TypeNoop", ValueType::kTypeNoop},
+    {"TypeColumnFamilyRangeDeletion",
+     ValueType::kTypeColumnFamilyRangeDeletion},
+    {"TypeRangeDeletion", ValueType::kTypeRangeDeletion},
+    {"TypeColumnFamilyBlobIndex", ValueType::kTypeColumnFamilyBlobIndex},
+    {"TypeBlobIndex", ValueType::kTypeBlobIndex},
+    {"TypeBeginPersistedPrepareXID", ValueType::kTypeBeginPersistedPrepareXID},
+    {"TypeBeginUnprepareXID", ValueType::kTypeBeginUnprepareXID},
+    {"TypeDeletionWithTimestamp", ValueType::kTypeDeletionWithTimestamp},
+    {"TypeCommitXIDAndTimestamp", ValueType::kTypeCommitXIDAndTimestamp},
+    {"TypeWideColumnEntity", ValueType::kTypeWideColumnEntity},
+    {"TypeColumnFamilyWideColumnEntity",
+     ValueType::kTypeColumnFamilyWideColumnEntity}};
+
+std::string KeyVersion::GetTypeName() const {
+  std::string type_name;
+  if (SerializeEnum<ValueType>(value_type_string_map,
+                               static_cast<ValueType>(type), &type_name)) {
+    return type_name;
+  } else {
+    return "Invalid";
+  }
+}
+
+Status GetAllKeyVersions(DB* db, Slice begin_key, Slice end_key,
+                         size_t max_num_ikeys,
+                         std::vector<KeyVersion>* key_versions) {
+  if (nullptr == db) {
+    return Status::InvalidArgument("db cannot be null.");
+  }
+  return GetAllKeyVersions(db, db->DefaultColumnFamily(), begin_key, end_key,
+                           max_num_ikeys, key_versions);
+}
+
+Status GetAllKeyVersions(DB* db, ColumnFamilyHandle* cfh, Slice begin_key,
+                         Slice end_key, size_t max_num_ikeys,
+                         std::vector<KeyVersion>* key_versions) {
+  if (nullptr == db) {
+    return Status::InvalidArgument("db cannot be null.");
+  }
+  if (nullptr == cfh) {
+    return Status::InvalidArgument("Column family handle cannot be null.");
+  }
+  if (nullptr == key_versions) {
+    return Status::InvalidArgument("key_versions cannot be null.");
+  }
+  key_versions->clear();
+
+  DBImpl* idb = static_cast<DBImpl*>(db->GetRootDB());
+  auto icmp = InternalKeyComparator(idb->GetOptions(cfh).comparator);
+  ReadOptions read_options;
+  Arena arena;
+  ScopedArenaIterator iter(
+      idb->NewInternalIterator(read_options, &arena, kMaxSequenceNumber, cfh));
+
+  if (!begin_key.empty()) {
+    InternalKey ikey;
+    ikey.SetMinPossibleForUserKey(begin_key);
+    iter->Seek(ikey.Encode());
+  } else {
+    iter->SeekToFirst();
+  }
+
+  size_t num_keys = 0;
+  for (; iter->Valid(); iter->Next()) {
+    ParsedInternalKey ikey;
+    Status pik_status =
+        ParseInternalKey(iter->key(), &ikey, true /* log_err_key */);  // TODO
+    if (!pik_status.ok()) {
+      return pik_status;
+    }
+
+    if (!end_key.empty() &&
+        icmp.user_comparator()->Compare(ikey.user_key, end_key) > 0) {
+      break;
+    }
+
+    key_versions->emplace_back(ikey.user_key.ToString() /* _user_key */,
+                               iter->value().ToString() /* _value */,
+                               ikey.sequence /* _sequence */,
+                               static_cast<int>(ikey.type) /* _type */);
+    if (++num_keys >= max_num_ikeys) {
+      break;
+    }
+  }
+  return Status::OK();
+}
+
+}  // namespace ROCKSDB_NAMESPACE
+
+#endif  // ROCKSDB_LITE
diff --git a/src/rocksdb/utilities/env_mirror.cc b/src/rocksdb/utilities/env_mirror.cc
new file mode 100644
index 000000000..3ea323b42
--- /dev/null
+++ b/src/rocksdb/utilities/env_mirror.cc
@@ -0,0 +1,275 @@
+// Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.
+// Copyright (c) 2015, Red Hat, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+// Copyright (c) 2011 The LevelDB Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file. See the AUTHORS file for names of contributors.
+
+#ifndef ROCKSDB_LITE
+
+#include "rocksdb/utilities/env_mirror.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+// An implementation of Env that mirrors all work over two backend
+// Env's.  This is useful for debugging purposes.
+class SequentialFileMirror : public SequentialFile {
+ public:
+  std::unique_ptr<SequentialFile> a_, b_;
+  std::string fname;
+  explicit SequentialFileMirror(std::string f) : fname(f) {}
+
+  Status Read(size_t n, Slice* result, char* scratch) override {
+    Slice aslice;
+    Status as = a_->Read(n, &aslice, scratch);
+    if (as == Status::OK()) {
+      char* bscratch = new char[n];
+      Slice bslice;
+#ifndef NDEBUG
+      size_t off = 0;
+#endif
+      size_t left = aslice.size();
+      while (left) {
+        Status bs = b_->Read(left, &bslice, bscratch);
+#ifndef NDEBUG
+        assert(as == bs);
+        assert(memcmp(bscratch, scratch + off, bslice.size()) == 0);
+        off += bslice.size();
+#endif
+        left -= bslice.size();
+      }
+      delete[] bscratch;
+      *result = aslice;
+    } else {
+      Status bs = b_->Read(n, result, scratch);
+      assert(as == bs);
+    }
+    return as;
+  }
+
+  Status Skip(uint64_t n) override {
+    Status as = a_->Skip(n);
+    Status bs = b_->Skip(n);
+    assert(as == bs);
+    return as;
+  }
+  Status InvalidateCache(size_t offset, size_t length) override {
+    Status as = a_->InvalidateCache(offset, length);
+    Status bs = b_->InvalidateCache(offset, length);
+    assert(as == bs);
+    return as;
+  };
+};
+
+class RandomAccessFileMirror : public RandomAccessFile {
+ public:
+  std::unique_ptr<RandomAccessFile> a_, b_;
+  std::string fname;
+  explicit RandomAccessFileMirror(std::string f) : fname(f) {}
+
+  Status Read(uint64_t offset, size_t n, Slice* result,
+              char* scratch) const override {
+    Status as = a_->Read(offset, n, result, scratch);
+    if (as == Status::OK()) {
+      char* bscratch = new char[n];
+      Slice bslice;
+      size_t off = 0;
+      size_t left = result->size();
+      while (left) {
+        Status bs = b_->Read(offset + off, left, &bslice, bscratch);
+        assert(as == bs);
+        assert(memcmp(bscratch, scratch + off, bslice.size()) == 0);
+        off += bslice.size();
+        left -= bslice.size();
+      }
+      delete[] bscratch;
+    } else {
+      Status bs = b_->Read(offset, n, result, scratch);
+      assert(as == bs);
+    }
+    return as;
+  }
+
+  size_t GetUniqueId(char* id, size_t max_size) const override {
+    // NOTE: not verified
+    return a_->GetUniqueId(id, max_size);
+  }
+};
+
+class WritableFileMirror : public WritableFile {
+ public:
+  std::unique_ptr<WritableFile> a_, b_;
+  std::string fname;
+  explicit WritableFileMirror(std::string f, const EnvOptions& options)
+      : WritableFile(options), fname(f) {}
+
+  Status Append(const Slice& data) override {
+    Status as = a_->Append(data);
+    Status bs = b_->Append(data);
+    assert(as == bs);
+    return as;
+  }
+  Status Append(const Slice& data,
+                const DataVerificationInfo& /* verification_info */) override {
+    return Append(data);
+  }
+  Status PositionedAppend(const Slice& data, uint64_t offset) override {
+    Status as = a_->PositionedAppend(data, offset);
+    Status bs = b_->PositionedAppend(data, offset);
+    assert(as == bs);
+    return as;
+  }
+  Status PositionedAppend(
+      const Slice& data, uint64_t offset,
+      const DataVerificationInfo& /* verification_info */) override {
+    return PositionedAppend(data, offset);
+  }
+  Status Truncate(uint64_t size) override {
+    Status as = a_->Truncate(size);
+    Status bs = b_->Truncate(size);
+    assert(as == bs);
+    return as;
+  }
+  Status Close() override {
+    Status as = a_->Close();
+    Status bs = b_->Close();
+    assert(as == bs);
+    return as;
+  }
+  Status Flush() override {
+    Status as = a_->Flush();
+    Status bs = b_->Flush();
+    assert(as == bs);
+    return as;
+  }
+  Status Sync() override {
+    Status as = a_->Sync();
+    Status bs = b_->Sync();
+    assert(as == bs);
+    return as;
+  }
+  Status Fsync() override {
+    Status as = a_->Fsync();
+    Status bs = b_->Fsync();
+    assert(as == bs);
+    return as;
+  }
+  bool IsSyncThreadSafe() const override {
+    bool as = a_->IsSyncThreadSafe();
+    assert(as == b_->IsSyncThreadSafe());
+    return as;
+  }
+  void SetIOPriority(Env::IOPriority pri) override {
+    a_->SetIOPriority(pri);
+    b_->SetIOPriority(pri);
+  }
+  Env::IOPriority GetIOPriority() override {
+    // NOTE: we don't verify this one
+    return a_->GetIOPriority();
+  }
+  uint64_t GetFileSize() override {
+    uint64_t as = a_->GetFileSize();
+    assert(as == b_->GetFileSize());
+    return as;
+  }
+  void GetPreallocationStatus(size_t* block_size,
+                              size_t* last_allocated_block) override {
+    // NOTE: we don't verify this one
+    return a_->GetPreallocationStatus(block_size, last_allocated_block);
+  }
+  size_t GetUniqueId(char* id, size_t max_size) const override {
+    // NOTE: we don't verify this one
+    return a_->GetUniqueId(id, max_size);
+  }
+  Status InvalidateCache(size_t offset, size_t length) override {
+    Status as = a_->InvalidateCache(offset, length);
+    Status bs = b_->InvalidateCache(offset, length);
+    assert(as == bs);
+    return as;
+  }
+
+ protected:
+  Status Allocate(uint64_t offset, uint64_t length) override {
+    Status as = a_->Allocate(offset, length);
+    Status bs = b_->Allocate(offset, length);
+    assert(as == bs);
+    return as;
+  }
+  Status RangeSync(uint64_t offset, uint64_t nbytes) override {
+    Status as = a_->RangeSync(offset, nbytes);
+    Status bs = b_->RangeSync(offset, nbytes);
+    assert(as == bs);
+    return as;
+  }
+};
+
+Status EnvMirror::NewSequentialFile(const std::string& f,
+                                    std::unique_ptr<SequentialFile>* r,
+                                    const EnvOptions& options) {
+  if (f.find("/proc/") == 0) {
+    return a_->NewSequentialFile(f, r, options);
+  }
+  SequentialFileMirror* mf = new SequentialFileMirror(f);
+  Status as = a_->NewSequentialFile(f, &mf->a_, options);
+  Status bs = b_->NewSequentialFile(f, &mf->b_, options);
+  assert(as == bs);
+  if (as.ok())
+    r->reset(mf);
+  else
+    delete mf;
+  return as;
+}
+
+Status EnvMirror::NewRandomAccessFile(const std::string& f,
+                                      std::unique_ptr<RandomAccessFile>* r,
+                                      const EnvOptions& options) {
+  if (f.find("/proc/") == 0) {
+    return a_->NewRandomAccessFile(f, r, options);
+  }
+  RandomAccessFileMirror* mf = new RandomAccessFileMirror(f);
+  Status as = a_->NewRandomAccessFile(f, &mf->a_, options);
+  Status bs = b_->NewRandomAccessFile(f, &mf->b_, options);
+  assert(as == bs);
+  if (as.ok())
+    r->reset(mf);
+  else
+    delete mf;
+  return as;
+}
+
+Status EnvMirror::NewWritableFile(const std::string& f,
+                                  std::unique_ptr<WritableFile>* r,
+                                  const EnvOptions& options) {
+  if (f.find("/proc/") == 0) return a_->NewWritableFile(f, r, options);
+  WritableFileMirror* mf = new WritableFileMirror(f, options);
+  Status as = a_->NewWritableFile(f, &mf->a_, options);
+  Status bs = b_->NewWritableFile(f, &mf->b_, options);
+  assert(as == bs);
+  if (as.ok())
+    r->reset(mf);
+  else
+    delete mf;
+  return as;
+}
+
+Status EnvMirror::ReuseWritableFile(const std::string& fname,
+                                    const std::string& old_fname,
+                                    std::unique_ptr<WritableFile>* r,
+                                    const EnvOptions& options) {
+  if (fname.find("/proc/") == 0)
+    return a_->ReuseWritableFile(fname, old_fname, r, options);
+  WritableFileMirror* mf = new WritableFileMirror(fname, options);
+  Status as = a_->ReuseWritableFile(fname, old_fname, &mf->a_, options);
+  Status bs = b_->ReuseWritableFile(fname, old_fname, &mf->b_, options);
+  assert(as == bs);
+  if (as.ok())
+    r->reset(mf);
+  else
+    delete mf;
+  return as;
+}
+
+}  // namespace ROCKSDB_NAMESPACE
+#endif
diff --git a/src/rocksdb/utilities/env_mirror_test.cc b/src/rocksdb/utilities/env_mirror_test.cc
new file mode 100644
index 000000000..c372de1da
--- /dev/null
+++ b/src/rocksdb/utilities/env_mirror_test.cc
@@ -0,0 +1,226 @@
+// Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.
+//  Copyright (c) 2015, Red Hat, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#ifndef ROCKSDB_LITE
+
+#include "rocksdb/utilities/env_mirror.h"
+
+#include "env/mock_env.h"
+#include "test_util/testharness.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+class EnvMirrorTest : public testing::Test {
+ public:
+  Env* default_;
+  MockEnv *a_, *b_;
+  EnvMirror* env_;
+  const EnvOptions soptions_;
+
+  EnvMirrorTest()
+      : default_(Env::Default()),
+        a_(new MockEnv(default_)),
+        b_(new MockEnv(default_)),
+        env_(new EnvMirror(a_, b_)) {}
+  ~EnvMirrorTest() {
+    delete env_;
+    delete a_;
+    delete b_;
+  }
+};
+
+TEST_F(EnvMirrorTest, Basics) {
+  uint64_t file_size;
+  std::unique_ptr<WritableFile> writable_file;
+  std::vector<std::string> children;
+
+  ASSERT_OK(env_->CreateDir("/dir"));
+
+  // Check that the directory is empty.
+  ASSERT_EQ(Status::NotFound(), env_->FileExists("/dir/non_existent"));
+  ASSERT_TRUE(!env_->GetFileSize("/dir/non_existent", &file_size).ok());
+  ASSERT_OK(env_->GetChildren("/dir", &children));
+  ASSERT_EQ(0U, children.size());
+
+  // Create a file.
+  ASSERT_OK(env_->NewWritableFile("/dir/f", &writable_file, soptions_));
+  writable_file.reset();
+
+  // Check that the file exists.
+  ASSERT_OK(env_->FileExists("/dir/f"));
+  ASSERT_OK(a_->FileExists("/dir/f"));
+  ASSERT_OK(b_->FileExists("/dir/f"));
+  ASSERT_OK(env_->GetFileSize("/dir/f", &file_size));
+  ASSERT_EQ(0U, file_size);
+  ASSERT_OK(env_->GetChildren("/dir", &children));
+  ASSERT_EQ(1U, children.size());
+  ASSERT_EQ("f", children[0]);
+  ASSERT_OK(a_->GetChildren("/dir", &children));
+  ASSERT_EQ(1U, children.size());
+  ASSERT_EQ("f", children[0]);
+  ASSERT_OK(b_->GetChildren("/dir", &children));
+  ASSERT_EQ(1U, children.size());
+  ASSERT_EQ("f", children[0]);
+
+  // Write to the file.
+  ASSERT_OK(env_->NewWritableFile("/dir/f", &writable_file, soptions_));
+  ASSERT_OK(writable_file->Append("abc"));
+  writable_file.reset();
+
+  // Check for expected size.
+  ASSERT_OK(env_->GetFileSize("/dir/f", &file_size));
+  ASSERT_EQ(3U, file_size);
+  ASSERT_OK(a_->GetFileSize("/dir/f", &file_size));
+  ASSERT_EQ(3U, file_size);
+  ASSERT_OK(b_->GetFileSize("/dir/f", &file_size));
+  ASSERT_EQ(3U, file_size);
+
+  // Check that renaming works.
+  ASSERT_TRUE(!env_->RenameFile("/dir/non_existent", "/dir/g").ok());
+  ASSERT_OK(env_->RenameFile("/dir/f", "/dir/g"));
+  ASSERT_EQ(Status::NotFound(), env_->FileExists("/dir/f"));
+  ASSERT_OK(env_->FileExists("/dir/g"));
+  ASSERT_OK(env_->GetFileSize("/dir/g", &file_size));
+  ASSERT_EQ(3U, file_size);
+  ASSERT_OK(a_->FileExists("/dir/g"));
+  ASSERT_OK(a_->GetFileSize("/dir/g", &file_size));
+  ASSERT_EQ(3U, file_size);
+  ASSERT_OK(b_->FileExists("/dir/g"));
+  ASSERT_OK(b_->GetFileSize("/dir/g", &file_size));
+  ASSERT_EQ(3U, file_size);
+
+  // Check that opening non-existent file fails.
+  std::unique_ptr<SequentialFile> seq_file;
+  std::unique_ptr<RandomAccessFile> rand_file;
+  ASSERT_TRUE(
+      !env_->NewSequentialFile("/dir/non_existent", &seq_file, soptions_).ok());
+  ASSERT_TRUE(!seq_file);
+  ASSERT_TRUE(
+      !env_->NewRandomAccessFile("/dir/non_existent", &rand_file, soptions_)
+           .ok());
+  ASSERT_TRUE(!rand_file);
+
+  // Check that deleting works.
+  ASSERT_TRUE(!env_->DeleteFile("/dir/non_existent").ok());
+  ASSERT_OK(env_->DeleteFile("/dir/g"));
+  ASSERT_EQ(Status::NotFound(), env_->FileExists("/dir/g"));
+  ASSERT_OK(env_->GetChildren("/dir", &children));
+  ASSERT_EQ(0U, children.size());
+  ASSERT_OK(env_->DeleteDir("/dir"));
+}
+
+TEST_F(EnvMirrorTest, ReadWrite) {
+  std::unique_ptr<WritableFile> writable_file;
+  std::unique_ptr<SequentialFile> seq_file;
+  std::unique_ptr<RandomAccessFile> rand_file;
+  Slice result;
+  char scratch[100];
+
+  ASSERT_OK(env_->CreateDir("/dir"));
+
+  ASSERT_OK(env_->NewWritableFile("/dir/f", &writable_file, soptions_));
+  ASSERT_OK(writable_file->Append("hello "));
+  ASSERT_OK(writable_file->Append("world"));
+  writable_file.reset();
+
+  // Read sequentially.
+  ASSERT_OK(env_->NewSequentialFile("/dir/f", &seq_file, soptions_));
+  ASSERT_OK(seq_file->Read(5, &result, scratch));  // Read "hello".
+  ASSERT_EQ(0, result.compare("hello"));
+  ASSERT_OK(seq_file->Skip(1));
+  ASSERT_OK(seq_file->Read(1000, &result, scratch));  // Read "world".
+  ASSERT_EQ(0, result.compare("world"));
+  ASSERT_OK(seq_file->Read(1000, &result, scratch));  // Try reading past EOF.
+  ASSERT_EQ(0U, result.size());
+  ASSERT_OK(seq_file->Skip(100));  // Try to skip past end of file.
+  ASSERT_OK(seq_file->Read(1000, &result, scratch));
+  ASSERT_EQ(0U, result.size());
+
+  // Random reads.
+  ASSERT_OK(env_->NewRandomAccessFile("/dir/f", &rand_file, soptions_));
+  ASSERT_OK(rand_file->Read(6, 5, &result, scratch));  // Read "world".
+  ASSERT_EQ(0, result.compare("world"));
+  ASSERT_OK(rand_file->Read(0, 5, &result, scratch));  // Read "hello".
+  ASSERT_EQ(0, result.compare("hello"));
+  ASSERT_OK(rand_file->Read(10, 100, &result, scratch));  // Read "d".
+  ASSERT_EQ(0, result.compare("d"));
+
+  // Too high offset.
+  ASSERT_TRUE(!rand_file->Read(1000, 5, &result, scratch).ok());
+}
+
+TEST_F(EnvMirrorTest, Locks) {
+  FileLock* lock;
+
+  // These are no-ops, but we test they return success.
+  ASSERT_OK(env_->LockFile("some file", &lock));
+  ASSERT_OK(env_->UnlockFile(lock));
+}
+
+TEST_F(EnvMirrorTest, Misc) {
+  std::string test_dir;
+  ASSERT_OK(env_->GetTestDirectory(&test_dir));
+  ASSERT_TRUE(!test_dir.empty());
+
+  std::unique_ptr<WritableFile> writable_file;
+  ASSERT_OK(env_->NewWritableFile("/a/b", &writable_file, soptions_));
+
+  // These are no-ops, but we test they return success.
+  ASSERT_OK(writable_file->Sync());
+  ASSERT_OK(writable_file->Flush());
+  ASSERT_OK(writable_file->Close());
+  writable_file.reset();
+}
+
+TEST_F(EnvMirrorTest, LargeWrite) {
+  const size_t kWriteSize = 300 * 1024;
+  char* scratch = new char[kWriteSize * 2];
+
+  std::string write_data;
+  for (size_t i = 0; i < kWriteSize; ++i) {
+    write_data.append(1, static_cast<char>(i));
+  }
+
+  std::unique_ptr<WritableFile> writable_file;
+  ASSERT_OK(env_->NewWritableFile("/dir/f", &writable_file, soptions_));
+  ASSERT_OK(writable_file->Append("foo"));
+  ASSERT_OK(writable_file->Append(write_data));
+  writable_file.reset();
+
+  std::unique_ptr<SequentialFile> seq_file;
+  Slice result;
+  ASSERT_OK(env_->NewSequentialFile("/dir/f", &seq_file, soptions_));
+  ASSERT_OK(seq_file->Read(3, &result, scratch));  // Read "foo".
+  ASSERT_EQ(0, result.compare("foo"));
+
+  size_t read = 0;
+  std::string read_data;
+  while (read < kWriteSize) {
+    ASSERT_OK(seq_file->Read(kWriteSize - read, &result, scratch));
+    read_data.append(result.data(), result.size());
+    read += result.size();
+  }
+  ASSERT_TRUE(write_data == read_data);
+  delete[] scratch;
+}
+
+}  // namespace ROCKSDB_NAMESPACE
+
+int main(int argc, char** argv) {
+  ROCKSDB_NAMESPACE::port::InstallStackTraceHandler();
+  ::testing::InitGoogleTest(&argc, argv);
+  return RUN_ALL_TESTS();
+}
+
+#else
+#include <stdio.h>
+
+int main(int argc, char** argv) {
+  fprintf(stderr, "SKIPPED as EnvMirror is not supported in ROCKSDB_LITE\n");
+  return 0;
+}
+
+#endif  // !ROCKSDB_LITE
diff --git a/src/rocksdb/utilities/env_timed.cc b/src/rocksdb/utilities/env_timed.cc
new file mode 100644
index 000000000..1eb723146
--- /dev/null
+++ b/src/rocksdb/utilities/env_timed.cc
@@ -0,0 +1,187 @@
+// Copyright (c) 2017-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+#include "utilities/env_timed.h"
+
+#include "env/composite_env_wrapper.h"
+#include "monitoring/perf_context_imp.h"
+#include "rocksdb/env.h"
+#include "rocksdb/file_system.h"
+#include "rocksdb/status.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+#ifndef ROCKSDB_LITE
+TimedFileSystem::TimedFileSystem(const std::shared_ptr<FileSystem>& base)
+    : FileSystemWrapper(base) {}
+IOStatus TimedFileSystem::NewSequentialFile(
+    const std::string& fname, const FileOptions& options,
+    std::unique_ptr<FSSequentialFile>* result, IODebugContext* dbg) {
+  PERF_TIMER_GUARD(env_new_sequential_file_nanos);
+  return FileSystemWrapper::NewSequentialFile(fname, options, result, dbg);
+}
+
+IOStatus TimedFileSystem::NewRandomAccessFile(
+    const std::string& fname, const FileOptions& options,
+    std::unique_ptr<FSRandomAccessFile>* result, IODebugContext* dbg) {
+  PERF_TIMER_GUARD(env_new_random_access_file_nanos);
+  return FileSystemWrapper::NewRandomAccessFile(fname, options, result, dbg);
+}
+
+IOStatus TimedFileSystem::NewWritableFile(
+    const std::string& fname, const FileOptions& options,
+    std::unique_ptr<FSWritableFile>* result, IODebugContext* dbg) {
+  PERF_TIMER_GUARD(env_new_writable_file_nanos);
+  return FileSystemWrapper::NewWritableFile(fname, options, result, dbg);
+}
+
+IOStatus TimedFileSystem::ReuseWritableFile(
+    const std::string& fname, const std::string& old_fname,
+    const FileOptions& options, std::unique_ptr<FSWritableFile>* result,
+    IODebugContext* dbg) {
+  PERF_TIMER_GUARD(env_reuse_writable_file_nanos);
+  return FileSystemWrapper::ReuseWritableFile(fname, old_fname, options, result,
+                                              dbg);
+}
+
+IOStatus TimedFileSystem::NewRandomRWFile(
+    const std::string& fname, const FileOptions& options,
+    std::unique_ptr<FSRandomRWFile>* result, IODebugContext* dbg) {
+  PERF_TIMER_GUARD(env_new_random_rw_file_nanos);
+  return FileSystemWrapper::NewRandomRWFile(fname, options, result, dbg);
+}
+
+IOStatus TimedFileSystem::NewDirectory(const std::string& name,
+                                       const IOOptions& options,
+                                       std::unique_ptr<FSDirectory>* result,
+                                       IODebugContext* dbg) {
+  PERF_TIMER_GUARD(env_new_directory_nanos);
+  return FileSystemWrapper::NewDirectory(name, options, result, dbg);
+}
+
+IOStatus TimedFileSystem::FileExists(const std::string& fname,
+                                     const IOOptions& options,
+                                     IODebugContext* dbg) {
+  PERF_TIMER_GUARD(env_file_exists_nanos);
+  return FileSystemWrapper::FileExists(fname, options, dbg);
+}
+
+IOStatus TimedFileSystem::GetChildren(const std::string& dir,
+                                      const IOOptions& options,
+                                      std::vector<std::string>* result,
+                                      IODebugContext* dbg) {
+  PERF_TIMER_GUARD(env_get_children_nanos);
+  return FileSystemWrapper::GetChildren(dir, options, result, dbg);
+}
+
+IOStatus TimedFileSystem::GetChildrenFileAttributes(
+    const std::string& dir, const IOOptions& options,
+    std::vector<FileAttributes>* result, IODebugContext* dbg) {
+  PERF_TIMER_GUARD(env_get_children_file_attributes_nanos);
+  return FileSystemWrapper::GetChildrenFileAttributes(dir, options, result,
+                                                      dbg);
+}
+
+IOStatus TimedFileSystem::DeleteFile(const std::string& fname,
+                                     const IOOptions& options,
+                                     IODebugContext* dbg) {
+  PERF_TIMER_GUARD(env_delete_file_nanos);
+  return FileSystemWrapper::DeleteFile(fname, options, dbg);
+}
+
+IOStatus TimedFileSystem::CreateDir(const std::string& dirname,
+                                    const IOOptions& options,
+                                    IODebugContext* dbg) {
+  PERF_TIMER_GUARD(env_create_dir_nanos);
+  return FileSystemWrapper::CreateDir(dirname, options, dbg);
+}
+
+IOStatus TimedFileSystem::CreateDirIfMissing(const std::string& dirname,
+                                             const IOOptions& options,
+                                             IODebugContext* dbg) {
+  PERF_TIMER_GUARD(env_create_dir_if_missing_nanos);
+  return FileSystemWrapper::CreateDirIfMissing(dirname, options, dbg);
+}
+
+IOStatus TimedFileSystem::DeleteDir(const std::string& dirname,
+                                    const IOOptions& options,
+                                    IODebugContext* dbg) {
+  PERF_TIMER_GUARD(env_delete_dir_nanos);
+  return FileSystemWrapper::DeleteDir(dirname, options, dbg);
+}
+
+IOStatus TimedFileSystem::GetFileSize(const std::string& fname,
+                                      const IOOptions& options,
+                                      uint64_t* file_size,
+                                      IODebugContext* dbg) {
+  PERF_TIMER_GUARD(env_get_file_size_nanos);
+  return FileSystemWrapper::GetFileSize(fname, options, file_size, dbg);
+}
+
+IOStatus TimedFileSystem::GetFileModificationTime(const std::string& fname,
+                                                  const IOOptions& options,
+                                                  uint64_t* file_mtime,
+                                                  IODebugContext* dbg) {
+  PERF_TIMER_GUARD(env_get_file_modification_time_nanos);
+  return FileSystemWrapper::GetFileModificationTime(fname, options, file_mtime,
+                                                    dbg);
+}
+
+IOStatus TimedFileSystem::RenameFile(const std::string& src,
+                                     const std::string& dst,
+                                     const IOOptions& options,
+                                     IODebugContext* dbg) {
+  PERF_TIMER_GUARD(env_rename_file_nanos);
+  return FileSystemWrapper::RenameFile(src, dst, options, dbg);
+}
+
+IOStatus TimedFileSystem::LinkFile(const std::string& src,
+                                   const std::string& dst,
+                                   const IOOptions& options,
+                                   IODebugContext* dbg) {
+  PERF_TIMER_GUARD(env_link_file_nanos);
+  return FileSystemWrapper::LinkFile(src, dst, options, dbg);
+}
+
+IOStatus TimedFileSystem::LockFile(const std::string& fname,
+                                   const IOOptions& options, FileLock** lock,
+                                   IODebugContext* dbg) {
+  PERF_TIMER_GUARD(env_lock_file_nanos);
+  return FileSystemWrapper::LockFile(fname, options, lock, dbg);
+}
+
+IOStatus TimedFileSystem::UnlockFile(FileLock* lock, const IOOptions& options,
+                                     IODebugContext* dbg) {
+  PERF_TIMER_GUARD(env_unlock_file_nanos);
+  return FileSystemWrapper::UnlockFile(lock, options, dbg);
+}
+
+IOStatus TimedFileSystem::NewLogger(const std::string& fname,
+                                    const IOOptions& options,
+                                    std::shared_ptr<Logger>* result,
+                                    IODebugContext* dbg) {
+  PERF_TIMER_GUARD(env_new_logger_nanos);
+  return FileSystemWrapper::NewLogger(fname, options, result, dbg);
+}
+
+std::shared_ptr<FileSystem> NewTimedFileSystem(
+    const std::shared_ptr<FileSystem>& base) {
+  return std::make_shared<TimedFileSystem>(base);
+}
+
+// An environment that measures function call times for filesystem
+// operations, reporting results to variables in PerfContext.
+Env* NewTimedEnv(Env* base_env) {
+  std::shared_ptr<FileSystem> timed_fs =
+      NewTimedFileSystem(base_env->GetFileSystem());
+  return new CompositeEnvWrapper(base_env, timed_fs);
+}
+
+#else  // ROCKSDB_LITE
+
+Env* NewTimedEnv(Env* /*base_env*/) { return nullptr; }
+
+#endif  // !ROCKSDB_LITE
+
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/utilities/env_timed.h b/src/rocksdb/utilities/env_timed.h
new file mode 100644
index 000000000..2d34fd590
--- /dev/null
+++ b/src/rocksdb/utilities/env_timed.h
@@ -0,0 +1,97 @@
+// Copyright (c) 2019-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+//
+
+#pragma once
+#include "rocksdb/file_system.h"
+namespace ROCKSDB_NAMESPACE {
+#ifndef ROCKSDB_LITE
+class TimedFileSystem : public FileSystemWrapper {
+ public:
+  explicit TimedFileSystem(const std::shared_ptr<FileSystem>& base);
+
+  static const char* kClassName() { return "TimedFS"; }
+  const char* Name() const override { return kClassName(); }
+
+  IOStatus NewSequentialFile(const std::string& fname,
+                             const FileOptions& options,
+                             std::unique_ptr<FSSequentialFile>* result,
+                             IODebugContext* dbg) override;
+
+  IOStatus NewRandomAccessFile(const std::string& fname,
+                               const FileOptions& options,
+                               std::unique_ptr<FSRandomAccessFile>* result,
+                               IODebugContext* dbg) override;
+
+  IOStatus NewWritableFile(const std::string& fname, const FileOptions& options,
+                           std::unique_ptr<FSWritableFile>* result,
+                           IODebugContext* dbg) override;
+
+  IOStatus ReuseWritableFile(const std::string& fname,
+                             const std::string& old_fname,
+                             const FileOptions& options,
+                             std::unique_ptr<FSWritableFile>* result,
+                             IODebugContext* dbg) override;
+
+  IOStatus NewRandomRWFile(const std::string& fname, const FileOptions& options,
+                           std::unique_ptr<FSRandomRWFile>* result,
+                           IODebugContext* dbg) override;
+
+  IOStatus NewDirectory(const std::string& name, const IOOptions& options,
+                        std::unique_ptr<FSDirectory>* result,
+                        IODebugContext* dbg) override;
+
+  IOStatus FileExists(const std::string& fname, const IOOptions& options,
+                      IODebugContext* dbg) override;
+
+  IOStatus GetChildren(const std::string& dir, const IOOptions& options,
+                       std::vector<std::string>* result,
+                       IODebugContext* dbg) override;
+
+  IOStatus GetChildrenFileAttributes(const std::string& dir,
+                                     const IOOptions& options,
+                                     std::vector<FileAttributes>* result,
+                                     IODebugContext* dbg) override;
+
+  IOStatus DeleteFile(const std::string& fname, const IOOptions& options,
+                      IODebugContext* dbg) override;
+
+  IOStatus CreateDir(const std::string& dirname, const IOOptions& options,
+                     IODebugContext* dbg) override;
+
+  IOStatus CreateDirIfMissing(const std::string& dirname,
+                              const IOOptions& options,
+                              IODebugContext* dbg) override;
+
+  IOStatus DeleteDir(const std::string& dirname, const IOOptions& options,
+                     IODebugContext* dbg) override;
+
+  IOStatus GetFileSize(const std::string& fname, const IOOptions& options,
+                       uint64_t* file_size, IODebugContext* dbg) override;
+
+  IOStatus GetFileModificationTime(const std::string& fname,
+                                   const IOOptions& options,
+                                   uint64_t* file_mtime,
+                                   IODebugContext* dbg) override;
+
+  IOStatus RenameFile(const std::string& src, const std::string& dst,
+                      const IOOptions& options, IODebugContext* dbg) override;
+
+  IOStatus LinkFile(const std::string& src, const std::string& dst,
+                    const IOOptions& options, IODebugContext* dbg) override;
+
+  IOStatus LockFile(const std::string& fname, const IOOptions& options,
+                    FileLock** lock, IODebugContext* dbg) override;
+
+  IOStatus UnlockFile(FileLock* lock, const IOOptions& options,
+                      IODebugContext* dbg) override;
+
+  IOStatus NewLogger(const std::string& fname, const IOOptions& options,
+                     std::shared_ptr<Logger>* result,
+                     IODebugContext* dbg) override;
+};
+
+#endif  // ROCKSDB_LITE
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/utilities/env_timed_test.cc b/src/rocksdb/utilities/env_timed_test.cc
new file mode 100644
index 000000000..6e392579d
--- /dev/null
+++ b/src/rocksdb/utilities/env_timed_test.cc
@@ -0,0 +1,44 @@
+// Copyright (c) 2017-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#ifndef ROCKSDB_LITE
+
+#include "rocksdb/env.h"
+#include "rocksdb/perf_context.h"
+#include "test_util/testharness.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+class TimedEnvTest : public testing::Test {};
+
+TEST_F(TimedEnvTest, BasicTest) {
+  SetPerfLevel(PerfLevel::kEnableTime);
+  ASSERT_EQ(0, get_perf_context()->env_new_writable_file_nanos);
+
+  std::unique_ptr<Env> mem_env(NewMemEnv(Env::Default()));
+  std::unique_ptr<Env> timed_env(NewTimedEnv(mem_env.get()));
+  std::unique_ptr<WritableFile> writable_file;
+  ASSERT_OK(timed_env->NewWritableFile("f", &writable_file, EnvOptions()));
+
+  ASSERT_GT(get_perf_context()->env_new_writable_file_nanos, 0);
+}
+
+}  // namespace ROCKSDB_NAMESPACE
+
+int main(int argc, char** argv) {
+  ROCKSDB_NAMESPACE::port::InstallStackTraceHandler();
+  ::testing::InitGoogleTest(&argc, argv);
+  return RUN_ALL_TESTS();
+}
+
+#else  // ROCKSDB_LITE
+#include <stdio.h>
+
+int main(int /*argc*/, char** /*argv*/) {
+  fprintf(stderr, "SKIPPED as TimedEnv is not supported in ROCKSDB_LITE\n");
+  return 0;
+}
+
+#endif  // ROCKSDB_LITE
diff --git a/src/rocksdb/utilities/fault_injection_env.cc b/src/rocksdb/utilities/fault_injection_env.cc
new file mode 100644
index 000000000..b0495a8c1
--- /dev/null
+++ b/src/rocksdb/utilities/fault_injection_env.cc
@@ -0,0 +1,555 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+//
+// Copyright 2014 The LevelDB Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file. See the AUTHORS file for names of contributors.
+
+// This test uses a custom Env to keep track of the state of a filesystem as of
+// the last "sync". It then checks for data loss errors by purposely dropping
+// file data (or entire files) not protected by a "sync".
+
+#include "utilities/fault_injection_env.h"
+
+#include <functional>
+#include <utility>
+
+#include "util/random.h"
+namespace ROCKSDB_NAMESPACE {
+
+// Assume a filename, and not a directory name like "/foo/bar/"
+std::string GetDirName(const std::string filename) {
+  size_t found = filename.find_last_of("/\\");
+  if (found == std::string::npos) {
+    return "";
+  } else {
+    return filename.substr(0, found);
+  }
+}
+
+// A basic file truncation function suitable for this test.
+Status Truncate(Env* env, const std::string& filename, uint64_t length) {
+  std::unique_ptr<SequentialFile> orig_file;
+  const EnvOptions options;
+  Status s = env->NewSequentialFile(filename, &orig_file, options);
+  if (!s.ok()) {
+    fprintf(stderr, "Cannot open file %s for truncation: %s\n",
+            filename.c_str(), s.ToString().c_str());
+    return s;
+  }
+
+  std::unique_ptr<char[]> scratch(new char[length]);
+  ROCKSDB_NAMESPACE::Slice result;
+  s = orig_file->Read(length, &result, scratch.get());
+#ifdef OS_WIN
+  orig_file.reset();
+#endif
+  if (s.ok()) {
+    std::string tmp_name = GetDirName(filename) + "/truncate.tmp";
+    std::unique_ptr<WritableFile> tmp_file;
+    s = env->NewWritableFile(tmp_name, &tmp_file, options);
+    if (s.ok()) {
+      s = tmp_file->Append(result);
+      if (s.ok()) {
+        s = env->RenameFile(tmp_name, filename);
+      } else {
+        fprintf(stderr, "Cannot rename file %s to %s: %s\n", tmp_name.c_str(),
+                filename.c_str(), s.ToString().c_str());
+        env->DeleteFile(tmp_name);
+      }
+    }
+  }
+  if (!s.ok()) {
+    fprintf(stderr, "Cannot truncate file %s: %s\n", filename.c_str(),
+            s.ToString().c_str());
+  }
+
+  return s;
+}
+
+// Trim the tailing "/" in the end of `str`
+std::string TrimDirname(const std::string& str) {
+  size_t found = str.find_last_not_of("/");
+  if (found == std::string::npos) {
+    return str;
+  }
+  return str.substr(0, found + 1);
+}
+
+// Return pair <parent directory name, file name> of a full path.
+std::pair<std::string, std::string> GetDirAndName(const std::string& name) {
+  std::string dirname = GetDirName(name);
+  std::string fname = name.substr(dirname.size() + 1);
+  return std::make_pair(dirname, fname);
+}
+
+Status FileState::DropUnsyncedData(Env* env) const {
+  ssize_t sync_pos = pos_at_last_sync_ == -1 ? 0 : pos_at_last_sync_;
+  return Truncate(env, filename_, sync_pos);
+}
+
+Status FileState::DropRandomUnsyncedData(Env* env, Random* rand) const {
+  ssize_t sync_pos = pos_at_last_sync_ == -1 ? 0 : pos_at_last_sync_;
+  assert(pos_ >= sync_pos);
+  int range = static_cast<int>(pos_ - sync_pos);
+  uint64_t truncated_size =
+      static_cast<uint64_t>(sync_pos) + rand->Uniform(range);
+  return Truncate(env, filename_, truncated_size);
+}
+
+Status TestDirectory::Fsync() {
+  if (!env_->IsFilesystemActive()) {
+    return env_->GetError();
+  }
+  env_->SyncDir(dirname_);
+  return dir_->Fsync();
+}
+
+Status TestDirectory::Close() {
+  if (!env_->IsFilesystemActive()) {
+    return env_->GetError();
+  }
+  return dir_->Close();
+}
+
+TestRandomAccessFile::TestRandomAccessFile(
+    std::unique_ptr<RandomAccessFile>&& target, FaultInjectionTestEnv* env)
+    : target_(std::move(target)), env_(env) {
+  assert(target_);
+  assert(env_);
+}
+
+Status TestRandomAccessFile::Read(uint64_t offset, size_t n, Slice* result,
+                                  char* scratch) const {
+  assert(env_);
+  if (!env_->IsFilesystemActive()) {
+    return env_->GetError();
+  }
+
+  assert(target_);
+  return target_->Read(offset, n, result, scratch);
+}
+
+Status TestRandomAccessFile::Prefetch(uint64_t offset, size_t n) {
+  assert(env_);
+  if (!env_->IsFilesystemActive()) {
+    return env_->GetError();
+  }
+
+  assert(target_);
+  return target_->Prefetch(offset, n);
+}
+
+Status TestRandomAccessFile::MultiRead(ReadRequest* reqs, size_t num_reqs) {
+  assert(env_);
+  if (!env_->IsFilesystemActive()) {
+    const Status s = env_->GetError();
+
+    assert(reqs);
+    for (size_t i = 0; i < num_reqs; ++i) {
+      reqs[i].status = s;
+    }
+
+    return s;
+  }
+
+  assert(target_);
+  return target_->MultiRead(reqs, num_reqs);
+}
+
+TestWritableFile::TestWritableFile(const std::string& fname,
+                                   std::unique_ptr<WritableFile>&& f,
+                                   FaultInjectionTestEnv* env)
+    : state_(fname),
+      target_(std::move(f)),
+      writable_file_opened_(true),
+      env_(env) {
+  assert(target_ != nullptr);
+  state_.pos_ = 0;
+}
+
+TestWritableFile::~TestWritableFile() {
+  if (writable_file_opened_) {
+    Close().PermitUncheckedError();
+  }
+}
+
+Status TestWritableFile::Append(const Slice& data) {
+  if (!env_->IsFilesystemActive()) {
+    return env_->GetError();
+  }
+  Status s = target_->Append(data);
+  if (s.ok()) {
+    state_.pos_ += data.size();
+    env_->WritableFileAppended(state_);
+  }
+  return s;
+}
+
+Status TestWritableFile::Close() {
+  writable_file_opened_ = false;
+  Status s = target_->Close();
+  if (s.ok()) {
+    env_->WritableFileClosed(state_);
+  }
+  return s;
+}
+
+Status TestWritableFile::Flush() {
+  Status s = target_->Flush();
+  if (s.ok() && env_->IsFilesystemActive()) {
+    state_.pos_at_last_flush_ = state_.pos_;
+  }
+  return s;
+}
+
+Status TestWritableFile::Sync() {
+  if (!env_->IsFilesystemActive()) {
+    return Status::IOError("FaultInjectionTestEnv: not active");
+  }
+  // No need to actual sync.
+  state_.pos_at_last_sync_ = state_.pos_;
+  env_->WritableFileSynced(state_);
+  return Status::OK();
+}
+
+TestRandomRWFile::TestRandomRWFile(const std::string& /*fname*/,
+                                   std::unique_ptr<RandomRWFile>&& f,
+                                   FaultInjectionTestEnv* env)
+    : target_(std::move(f)), file_opened_(true), env_(env) {
+  assert(target_ != nullptr);
+}
+
+TestRandomRWFile::~TestRandomRWFile() {
+  if (file_opened_) {
+    Close().PermitUncheckedError();
+  }
+}
+
+Status TestRandomRWFile::Write(uint64_t offset, const Slice& data) {
+  if (!env_->IsFilesystemActive()) {
+    return env_->GetError();
+  }
+  return target_->Write(offset, data);
+}
+
+Status TestRandomRWFile::Read(uint64_t offset, size_t n, Slice* result,
+                              char* scratch) const {
+  if (!env_->IsFilesystemActive()) {
+    return env_->GetError();
+  }
+  return target_->Read(offset, n, result, scratch);
+}
+
+Status TestRandomRWFile::Close() {
+  file_opened_ = false;
+  return target_->Close();
+}
+
+Status TestRandomRWFile::Flush() {
+  if (!env_->IsFilesystemActive()) {
+    return env_->GetError();
+  }
+  return target_->Flush();
+}
+
+Status TestRandomRWFile::Sync() {
+  if (!env_->IsFilesystemActive()) {
+    return env_->GetError();
+  }
+  return target_->Sync();
+}
+
+Status FaultInjectionTestEnv::NewDirectory(const std::string& name,
+                                           std::unique_ptr<Directory>* result) {
+  std::unique_ptr<Directory> r;
+  Status s = target()->NewDirectory(name, &r);
+  assert(s.ok());
+  if (!s.ok()) {
+    return s;
+  }
+  result->reset(new TestDirectory(this, TrimDirname(name), r.release()));
+  return Status::OK();
+}
+
+Status FaultInjectionTestEnv::NewWritableFile(
+    const std::string& fname, std::unique_ptr<WritableFile>* result,
+    const EnvOptions& soptions) {
+  if (!IsFilesystemActive()) {
+    return GetError();
+  }
+  // Not allow overwriting files
+  Status s = target()->FileExists(fname);
+  if (s.ok()) {
+    return Status::Corruption("File already exists.");
+  } else if (!s.IsNotFound()) {
+    assert(s.IsIOError());
+    return s;
+  }
+  s = target()->NewWritableFile(fname, result, soptions);
+  if (s.ok()) {
+    result->reset(new TestWritableFile(fname, std::move(*result), this));
+    // WritableFileWriter* file is opened
+    // again then it will be truncated - so forget our saved state.
+    UntrackFile(fname);
+    MutexLock l(&mutex_);
+    open_managed_files_.insert(fname);
+    auto dir_and_name = GetDirAndName(fname);
+    auto& list = dir_to_new_files_since_last_sync_[dir_and_name.first];
+    list.insert(dir_and_name.second);
+  }
+  return s;
+}
+
+Status FaultInjectionTestEnv::ReopenWritableFile(
+    const std::string& fname, std::unique_ptr<WritableFile>* result,
+    const EnvOptions& soptions) {
+  if (!IsFilesystemActive()) {
+    return GetError();
+  }
+
+  bool exists;
+  Status s, exists_s = target()->FileExists(fname);
+  if (exists_s.IsNotFound()) {
+    exists = false;
+  } else if (exists_s.ok()) {
+    exists = true;
+  } else {
+    s = exists_s;
+    exists = false;
+  }
+
+  if (s.ok()) {
+    s = target()->ReopenWritableFile(fname, result, soptions);
+  }
+
+  // Only track files we created. Files created outside of this
+  // `FaultInjectionTestEnv` are not eligible for tracking/data dropping
+  // (for example, they may contain data a previous db_stress run expects to
+  // be recovered). This could be extended to track/drop data appended once
+  // the file is under `FaultInjectionTestEnv`'s control.
+  if (s.ok()) {
+    bool should_track;
+    {
+      MutexLock l(&mutex_);
+      if (db_file_state_.find(fname) != db_file_state_.end()) {
+        // It was written by this `Env` earlier.
+        assert(exists);
+        should_track = true;
+      } else if (!exists) {
+        // It was created by this `Env` just now.
+        should_track = true;
+        open_managed_files_.insert(fname);
+        auto dir_and_name = GetDirAndName(fname);
+        auto& list = dir_to_new_files_since_last_sync_[dir_and_name.first];
+        list.insert(dir_and_name.second);
+      } else {
+        should_track = false;
+      }
+    }
+    if (should_track) {
+      result->reset(new TestWritableFile(fname, std::move(*result), this));
+    }
+  }
+  return s;
+}
+
+Status FaultInjectionTestEnv::NewRandomRWFile(
+    const std::string& fname, std::unique_ptr<RandomRWFile>* result,
+    const EnvOptions& soptions) {
+  if (!IsFilesystemActive()) {
+    return GetError();
+  }
+  Status s = target()->NewRandomRWFile(fname, result, soptions);
+  if (s.ok()) {
+    result->reset(new TestRandomRWFile(fname, std::move(*result), this));
+    // WritableFileWriter* file is opened
+    // again then it will be truncated - so forget our saved state.
+    UntrackFile(fname);
+    MutexLock l(&mutex_);
+    open_managed_files_.insert(fname);
+    auto dir_and_name = GetDirAndName(fname);
+    auto& list = dir_to_new_files_since_last_sync_[dir_and_name.first];
+    list.insert(dir_and_name.second);
+  }
+  return s;
+}
+
+Status FaultInjectionTestEnv::NewRandomAccessFile(
+    const std::string& fname, std::unique_ptr<RandomAccessFile>* result,
+    const EnvOptions& soptions) {
+  if (!IsFilesystemActive()) {
+    return GetError();
+  }
+
+  assert(target());
+  const Status s = target()->NewRandomAccessFile(fname, result, soptions);
+  if (!s.ok()) {
+    return s;
+  }
+
+  assert(result);
+  result->reset(new TestRandomAccessFile(std::move(*result), this));
+
+  return Status::OK();
+}
+
+Status FaultInjectionTestEnv::DeleteFile(const std::string& f) {
+  if (!IsFilesystemActive()) {
+    return GetError();
+  }
+  Status s = EnvWrapper::DeleteFile(f);
+  if (s.ok()) {
+    UntrackFile(f);
+  }
+  return s;
+}
+
+Status FaultInjectionTestEnv::RenameFile(const std::string& s,
+                                         const std::string& t) {
+  if (!IsFilesystemActive()) {
+    return GetError();
+  }
+  Status ret = EnvWrapper::RenameFile(s, t);
+
+  if (ret.ok()) {
+    MutexLock l(&mutex_);
+    if (db_file_state_.find(s) != db_file_state_.end()) {
+      db_file_state_[t] = db_file_state_[s];
+      db_file_state_.erase(s);
+    }
+
+    auto sdn = GetDirAndName(s);
+    auto tdn = GetDirAndName(t);
+    if (dir_to_new_files_since_last_sync_[sdn.first].erase(sdn.second) != 0) {
+      auto& tlist = dir_to_new_files_since_last_sync_[tdn.first];
+      assert(tlist.find(tdn.second) == tlist.end());
+      tlist.insert(tdn.second);
+    }
+  }
+
+  return ret;
+}
+
+Status FaultInjectionTestEnv::LinkFile(const std::string& s,
+                                       const std::string& t) {
+  if (!IsFilesystemActive()) {
+    return GetError();
+  }
+  Status ret = EnvWrapper::LinkFile(s, t);
+
+  if (ret.ok()) {
+    MutexLock l(&mutex_);
+    if (db_file_state_.find(s) != db_file_state_.end()) {
+      db_file_state_[t] = db_file_state_[s];
+    }
+
+    auto sdn = GetDirAndName(s);
+    auto tdn = GetDirAndName(t);
+    if (dir_to_new_files_since_last_sync_[sdn.first].find(sdn.second) !=
+        dir_to_new_files_since_last_sync_[sdn.first].end()) {
+      auto& tlist = dir_to_new_files_since_last_sync_[tdn.first];
+      assert(tlist.find(tdn.second) == tlist.end());
+      tlist.insert(tdn.second);
+    }
+  }
+
+  return ret;
+}
+
+void FaultInjectionTestEnv::WritableFileClosed(const FileState& state) {
+  MutexLock l(&mutex_);
+  if (open_managed_files_.find(state.filename_) != open_managed_files_.end()) {
+    db_file_state_[state.filename_] = state;
+    open_managed_files_.erase(state.filename_);
+  }
+}
+
+void FaultInjectionTestEnv::WritableFileSynced(const FileState& state) {
+  MutexLock l(&mutex_);
+  if (open_managed_files_.find(state.filename_) != open_managed_files_.end()) {
+    if (db_file_state_.find(state.filename_) == db_file_state_.end()) {
+      db_file_state_.insert(std::make_pair(state.filename_, state));
+    } else {
+      db_file_state_[state.filename_] = state;
+    }
+  }
+}
+
+void FaultInjectionTestEnv::WritableFileAppended(const FileState& state) {
+  MutexLock l(&mutex_);
+  if (open_managed_files_.find(state.filename_) != open_managed_files_.end()) {
+    if (db_file_state_.find(state.filename_) == db_file_state_.end()) {
+      db_file_state_.insert(std::make_pair(state.filename_, state));
+    } else {
+      db_file_state_[state.filename_] = state;
+    }
+  }
+}
+
+// For every file that is not fully synced, make a call to `func` with
+// FileState of the file as the parameter.
+Status FaultInjectionTestEnv::DropFileData(
+    std::function<Status(Env*, FileState)> func) {
+  Status s;
+  MutexLock l(&mutex_);
+  for (std::map<std::string, FileState>::const_iterator it =
+           db_file_state_.begin();
+       s.ok() && it != db_file_state_.end(); ++it) {
+    const FileState& state = it->second;
+    if (!state.IsFullySynced()) {
+      s = func(target(), state);
+    }
+  }
+  return s;
+}
+
+Status FaultInjectionTestEnv::DropUnsyncedFileData() {
+  return DropFileData([&](Env* env, const FileState& state) {
+    return state.DropUnsyncedData(env);
+  });
+}
+
+Status FaultInjectionTestEnv::DropRandomUnsyncedFileData(Random* rnd) {
+  return DropFileData([&](Env* env, const FileState& state) {
+    return state.DropRandomUnsyncedData(env, rnd);
+  });
+}
+
+Status FaultInjectionTestEnv::DeleteFilesCreatedAfterLastDirSync() {
+  // Because DeleteFile access this container make a copy to avoid deadlock
+  std::map<std::string, std::set<std::string>> map_copy;
+  {
+    MutexLock l(&mutex_);
+    map_copy.insert(dir_to_new_files_since_last_sync_.begin(),
+                    dir_to_new_files_since_last_sync_.end());
+  }
+
+  for (auto& pair : map_copy) {
+    for (std::string name : pair.second) {
+      Status s = DeleteFile(pair.first + "/" + name);
+      if (!s.ok()) {
+        return s;
+      }
+    }
+  }
+  return Status::OK();
+}
+void FaultInjectionTestEnv::ResetState() {
+  MutexLock l(&mutex_);
+  db_file_state_.clear();
+  dir_to_new_files_since_last_sync_.clear();
+  SetFilesystemActiveNoLock(true);
+}
+
+void FaultInjectionTestEnv::UntrackFile(const std::string& f) {
+  MutexLock l(&mutex_);
+  auto dir_and_name = GetDirAndName(f);
+  dir_to_new_files_since_last_sync_[dir_and_name.first].erase(
+      dir_and_name.second);
+  db_file_state_.erase(f);
+  open_managed_files_.erase(f);
+}
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/utilities/fault_injection_env.h b/src/rocksdb/utilities/fault_injection_env.h
new file mode 100644
index 000000000..549bfe716
--- /dev/null
+++ b/src/rocksdb/utilities/fault_injection_env.h
@@ -0,0 +1,258 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+//
+// Copyright 2014 The LevelDB Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file. See the AUTHORS file for names of contributors.
+
+// This test uses a custom Env to keep track of the state of a filesystem as of
+// the last "sync". It then checks for data loss errors by purposely dropping
+// file data (or entire files) not protected by a "sync".
+
+#pragma once
+
+#include <map>
+#include <set>
+#include <string>
+
+#include "file/filename.h"
+#include "rocksdb/env.h"
+#include "util/mutexlock.h"
+
+namespace ROCKSDB_NAMESPACE {
+class Random;
+class TestWritableFile;
+class FaultInjectionTestEnv;
+
+struct FileState {
+  std::string filename_;
+  ssize_t pos_;
+  ssize_t pos_at_last_sync_;
+  ssize_t pos_at_last_flush_;
+
+  explicit FileState(const std::string& filename)
+      : filename_(filename),
+        pos_(-1),
+        pos_at_last_sync_(-1),
+        pos_at_last_flush_(-1) {}
+
+  FileState() : pos_(-1), pos_at_last_sync_(-1), pos_at_last_flush_(-1) {}
+
+  bool IsFullySynced() const { return pos_ <= 0 || pos_ == pos_at_last_sync_; }
+
+  Status DropUnsyncedData(Env* env) const;
+
+  Status DropRandomUnsyncedData(Env* env, Random* rand) const;
+};
+
+class TestRandomAccessFile : public RandomAccessFile {
+ public:
+  TestRandomAccessFile(std::unique_ptr<RandomAccessFile>&& target,
+                       FaultInjectionTestEnv* env);
+
+  Status Read(uint64_t offset, size_t n, Slice* result,
+              char* scratch) const override;
+
+  Status Prefetch(uint64_t offset, size_t n) override;
+
+  Status MultiRead(ReadRequest* reqs, size_t num_reqs) override;
+
+ private:
+  std::unique_ptr<RandomAccessFile> target_;
+  FaultInjectionTestEnv* env_;
+};
+
+// A wrapper around WritableFileWriter* file
+// is written to or sync'ed.
+class TestWritableFile : public WritableFile {
+ public:
+  explicit TestWritableFile(const std::string& fname,
+                            std::unique_ptr<WritableFile>&& f,
+                            FaultInjectionTestEnv* env);
+  virtual ~TestWritableFile();
+  virtual Status Append(const Slice& data) override;
+  virtual Status Append(
+      const Slice& data,
+      const DataVerificationInfo& /*verification_info*/) override {
+    return Append(data);
+  }
+  virtual Status Truncate(uint64_t size) override {
+    return target_->Truncate(size);
+  }
+  virtual Status Close() override;
+  virtual Status Flush() override;
+  virtual Status Sync() override;
+  virtual bool IsSyncThreadSafe() const override { return true; }
+  virtual Status PositionedAppend(const Slice& data, uint64_t offset) override {
+    return target_->PositionedAppend(data, offset);
+  }
+  virtual Status PositionedAppend(
+      const Slice& data, uint64_t offset,
+      const DataVerificationInfo& /*verification_info*/) override {
+    return PositionedAppend(data, offset);
+  }
+  virtual bool use_direct_io() const override {
+    return target_->use_direct_io();
+  };
+
+ private:
+  FileState state_;
+  std::unique_ptr<WritableFile> target_;
+  bool writable_file_opened_;
+  FaultInjectionTestEnv* env_;
+};
+
+// A wrapper around WritableFileWriter* file
+// is written to or sync'ed.
+class TestRandomRWFile : public RandomRWFile {
+ public:
+  explicit TestRandomRWFile(const std::string& fname,
+                            std::unique_ptr<RandomRWFile>&& f,
+                            FaultInjectionTestEnv* env);
+  virtual ~TestRandomRWFile();
+  Status Write(uint64_t offset, const Slice& data) override;
+  Status Read(uint64_t offset, size_t n, Slice* result,
+              char* scratch) const override;
+  Status Close() override;
+  Status Flush() override;
+  Status Sync() override;
+  size_t GetRequiredBufferAlignment() const override {
+    return target_->GetRequiredBufferAlignment();
+  }
+  bool use_direct_io() const override { return target_->use_direct_io(); };
+
+ private:
+  std::unique_ptr<RandomRWFile> target_;
+  bool file_opened_;
+  FaultInjectionTestEnv* env_;
+};
+
+class TestDirectory : public Directory {
+ public:
+  explicit TestDirectory(FaultInjectionTestEnv* env, std::string dirname,
+                         Directory* dir)
+      : env_(env), dirname_(dirname), dir_(dir) {}
+  ~TestDirectory() {}
+
+  virtual Status Fsync() override;
+  virtual Status Close() override;
+
+ private:
+  FaultInjectionTestEnv* env_;
+  std::string dirname_;
+  std::unique_ptr<Directory> dir_;
+};
+
+class FaultInjectionTestEnv : public EnvWrapper {
+ public:
+  explicit FaultInjectionTestEnv(Env* base)
+      : EnvWrapper(base), filesystem_active_(true) {}
+  virtual ~FaultInjectionTestEnv() { error_.PermitUncheckedError(); }
+
+  static const char* kClassName() { return "FaultInjectionTestEnv"; }
+  const char* Name() const override { return kClassName(); }
+
+  Status NewDirectory(const std::string& name,
+                      std::unique_ptr<Directory>* result) override;
+
+  Status NewWritableFile(const std::string& fname,
+                         std::unique_ptr<WritableFile>* result,
+                         const EnvOptions& soptions) override;
+
+  Status ReopenWritableFile(const std::string& fname,
+                            std::unique_ptr<WritableFile>* result,
+                            const EnvOptions& soptions) override;
+
+  Status NewRandomRWFile(const std::string& fname,
+                         std::unique_ptr<RandomRWFile>* result,
+                         const EnvOptions& soptions) override;
+
+  Status NewRandomAccessFile(const std::string& fname,
+                             std::unique_ptr<RandomAccessFile>* result,
+                             const EnvOptions& soptions) override;
+
+  virtual Status DeleteFile(const std::string& f) override;
+
+  virtual Status RenameFile(const std::string& s,
+                            const std::string& t) override;
+
+  virtual Status LinkFile(const std::string& s, const std::string& t) override;
+
+// Undef to eliminate clash on Windows
+#undef GetFreeSpace
+  virtual Status GetFreeSpace(const std::string& path,
+                              uint64_t* disk_free) override {
+    if (!IsFilesystemActive() &&
+        error_.subcode() == IOStatus::SubCode::kNoSpace) {
+      *disk_free = 0;
+      return Status::OK();
+    } else {
+      return target()->GetFreeSpace(path, disk_free);
+    }
+  }
+
+  void WritableFileClosed(const FileState& state);
+
+  void WritableFileSynced(const FileState& state);
+
+  void WritableFileAppended(const FileState& state);
+
+  // For every file that is not fully synced, make a call to `func` with
+  // FileState of the file as the parameter.
+  Status DropFileData(std::function<Status(Env*, FileState)> func);
+
+  Status DropUnsyncedFileData();
+
+  Status DropRandomUnsyncedFileData(Random* rnd);
+
+  Status DeleteFilesCreatedAfterLastDirSync();
+
+  void ResetState();
+
+  void UntrackFile(const std::string& f);
+
+  void SyncDir(const std::string& dirname) {
+    MutexLock l(&mutex_);
+    dir_to_new_files_since_last_sync_.erase(dirname);
+  }
+
+  // Setting the filesystem to inactive is the test equivalent to simulating a
+  // system reset. Setting to inactive will freeze our saved filesystem state so
+  // that it will stop being recorded. It can then be reset back to the state at
+  // the time of the reset.
+  bool IsFilesystemActive() {
+    MutexLock l(&mutex_);
+    return filesystem_active_;
+  }
+  void SetFilesystemActiveNoLock(
+      bool active, Status error = Status::Corruption("Not active")) {
+    error.PermitUncheckedError();
+    filesystem_active_ = active;
+    if (!active) {
+      error_ = error;
+    }
+    error.PermitUncheckedError();
+  }
+  void SetFilesystemActive(bool active,
+                           Status error = Status::Corruption("Not active")) {
+    error.PermitUncheckedError();
+    MutexLock l(&mutex_);
+    SetFilesystemActiveNoLock(active, error);
+    error.PermitUncheckedError();
+  }
+  void AssertNoOpenFile() { assert(open_managed_files_.empty()); }
+  Status GetError() { return error_; }
+
+ private:
+  port::Mutex mutex_;
+  std::map<std::string, FileState> db_file_state_;
+  std::set<std::string> open_managed_files_;
+  std::unordered_map<std::string, std::set<std::string>>
+      dir_to_new_files_since_last_sync_;
+  bool filesystem_active_;  // Record flushes, syncs, writes
+  Status error_;
+};
+
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/utilities/fault_injection_fs.cc b/src/rocksdb/utilities/fault_injection_fs.cc
new file mode 100644
index 000000000..549051856
--- /dev/null
+++ b/src/rocksdb/utilities/fault_injection_fs.cc
@@ -0,0 +1,1032 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+//
+// Copyright 2014 The LevelDB Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file. See the AUTHORS file for names of contributors.
+
+// This test uses a custom FileSystem to keep track of the state of a file
+// system the last "Sync". The data being written is cached in a "buffer".
+// Only when "Sync" is called, the data will be persistent. It can simulate
+// file data loss (or entire files) not protected by a "Sync". For any of the
+// FileSystem related operations, by specify the "IOStatus Error", a specific
+// error can be returned when file system is not activated.
+
+#include "utilities/fault_injection_fs.h"
+
+#include <algorithm>
+#include <functional>
+#include <utility>
+
+#include "env/composite_env_wrapper.h"
+#include "port/lang.h"
+#include "port/stack_trace.h"
+#include "test_util/sync_point.h"
+#include "util/coding.h"
+#include "util/crc32c.h"
+#include "util/random.h"
+#include "util/string_util.h"
+#include "util/xxhash.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+const std::string kNewFileNoOverwrite = "";
+
+// Assume a filename, and not a directory name like "/foo/bar/"
+std::string TestFSGetDirName(const std::string filename) {
+  size_t found = filename.find_last_of("/\\");
+  if (found == std::string::npos) {
+    return "";
+  } else {
+    return filename.substr(0, found);
+  }
+}
+
+// Trim the tailing "/" in the end of `str`
+std::string TestFSTrimDirname(const std::string& str) {
+  size_t found = str.find_last_not_of("/");
+  if (found == std::string::npos) {
+    return str;
+  }
+  return str.substr(0, found + 1);
+}
+
+// Return pair <parent directory name, file name> of a full path.
+std::pair<std::string, std::string> TestFSGetDirAndName(
+    const std::string& name) {
+  std::string dirname = TestFSGetDirName(name);
+  std::string fname = name.substr(dirname.size() + 1);
+  return std::make_pair(dirname, fname);
+}
+
+// Calculate the checksum of the data with corresponding checksum
+// type. If name does not match, no checksum is returned.
+void CalculateTypedChecksum(const ChecksumType& checksum_type, const char* data,
+                            size_t size, std::string* checksum) {
+  if (checksum_type == ChecksumType::kCRC32c) {
+    uint32_t v_crc32c = crc32c::Extend(0, data, size);
+    PutFixed32(checksum, v_crc32c);
+    return;
+  } else if (checksum_type == ChecksumType::kxxHash) {
+    uint32_t v = XXH32(data, size, 0);
+    PutFixed32(checksum, v);
+  }
+  return;
+}
+
+IOStatus FSFileState::DropUnsyncedData() {
+  buffer_.resize(0);
+  return IOStatus::OK();
+}
+
+IOStatus FSFileState::DropRandomUnsyncedData(Random* rand) {
+  int range = static_cast<int>(buffer_.size());
+  size_t truncated_size = static_cast<size_t>(rand->Uniform(range));
+  buffer_.resize(truncated_size);
+  return IOStatus::OK();
+}
+
+IOStatus TestFSDirectory::Fsync(const IOOptions& options, IODebugContext* dbg) {
+  if (!fs_->IsFilesystemActive()) {
+    return fs_->GetError();
+  }
+  {
+    IOStatus in_s = fs_->InjectMetadataWriteError();
+    if (!in_s.ok()) {
+      return in_s;
+    }
+  }
+  fs_->SyncDir(dirname_);
+  IOStatus s = dir_->Fsync(options, dbg);
+  {
+    IOStatus in_s = fs_->InjectMetadataWriteError();
+    if (!in_s.ok()) {
+      return in_s;
+    }
+  }
+  return s;
+}
+
+IOStatus TestFSDirectory::Close(const IOOptions& options, IODebugContext* dbg) {
+  if (!fs_->IsFilesystemActive()) {
+    return fs_->GetError();
+  }
+  IOStatus s = dir_->Close(options, dbg);
+  return s;
+}
+
+IOStatus TestFSDirectory::FsyncWithDirOptions(
+    const IOOptions& options, IODebugContext* dbg,
+    const DirFsyncOptions& dir_fsync_options) {
+  if (!fs_->IsFilesystemActive()) {
+    return fs_->GetError();
+  }
+  {
+    IOStatus in_s = fs_->InjectMetadataWriteError();
+    if (!in_s.ok()) {
+      return in_s;
+    }
+  }
+  fs_->SyncDir(dirname_);
+  IOStatus s = dir_->FsyncWithDirOptions(options, dbg, dir_fsync_options);
+  {
+    IOStatus in_s = fs_->InjectMetadataWriteError();
+    if (!in_s.ok()) {
+      return in_s;
+    }
+  }
+  return s;
+}
+
+TestFSWritableFile::TestFSWritableFile(const std::string& fname,
+                                       const FileOptions& file_opts,
+                                       std::unique_ptr<FSWritableFile>&& f,
+                                       FaultInjectionTestFS* fs)
+    : state_(fname),
+      file_opts_(file_opts),
+      target_(std::move(f)),
+      writable_file_opened_(true),
+      fs_(fs) {
+  assert(target_ != nullptr);
+  state_.pos_ = 0;
+}
+
+TestFSWritableFile::~TestFSWritableFile() {
+  if (writable_file_opened_) {
+    Close(IOOptions(), nullptr).PermitUncheckedError();
+  }
+}
+
+IOStatus TestFSWritableFile::Append(const Slice& data, const IOOptions& options,
+                                    IODebugContext* dbg) {
+  MutexLock l(&mutex_);
+  if (!fs_->IsFilesystemActive()) {
+    return fs_->GetError();
+  }
+  if (target_->use_direct_io()) {
+    target_->Append(data, options, dbg).PermitUncheckedError();
+  } else {
+    state_.buffer_.append(data.data(), data.size());
+    state_.pos_ += data.size();
+    fs_->WritableFileAppended(state_);
+  }
+  IOStatus io_s = fs_->InjectWriteError(state_.filename_);
+  return io_s;
+}
+
+// By setting the IngestDataCorruptionBeforeWrite(), the data corruption is
+// simulated.
+IOStatus TestFSWritableFile::Append(
+    const Slice& data, const IOOptions& options,
+    const DataVerificationInfo& verification_info, IODebugContext* dbg) {
+  MutexLock l(&mutex_);
+  if (!fs_->IsFilesystemActive()) {
+    return fs_->GetError();
+  }
+  if (fs_->ShouldDataCorruptionBeforeWrite()) {
+    return IOStatus::Corruption("Data is corrupted!");
+  }
+
+  // Calculate the checksum
+  std::string checksum;
+  CalculateTypedChecksum(fs_->GetChecksumHandoffFuncType(), data.data(),
+                         data.size(), &checksum);
+  if (fs_->GetChecksumHandoffFuncType() != ChecksumType::kNoChecksum &&
+      checksum != verification_info.checksum.ToString()) {
+    std::string msg = "Data is corrupted! Origin data checksum: " +
+                      verification_info.checksum.ToString() +
+                      "current data checksum: " + checksum;
+    return IOStatus::Corruption(msg);
+  }
+  if (target_->use_direct_io()) {
+    target_->Append(data, options, dbg).PermitUncheckedError();
+  } else {
+    state_.buffer_.append(data.data(), data.size());
+    state_.pos_ += data.size();
+    fs_->WritableFileAppended(state_);
+  }
+  IOStatus io_s = fs_->InjectWriteError(state_.filename_);
+  return io_s;
+}
+
+IOStatus TestFSWritableFile::PositionedAppend(
+    const Slice& data, uint64_t offset, const IOOptions& options,
+    const DataVerificationInfo& verification_info, IODebugContext* dbg) {
+  MutexLock l(&mutex_);
+  if (!fs_->IsFilesystemActive()) {
+    return fs_->GetError();
+  }
+  if (fs_->ShouldDataCorruptionBeforeWrite()) {
+    return IOStatus::Corruption("Data is corrupted!");
+  }
+
+  // Calculate the checksum
+  std::string checksum;
+  CalculateTypedChecksum(fs_->GetChecksumHandoffFuncType(), data.data(),
+                         data.size(), &checksum);
+  if (fs_->GetChecksumHandoffFuncType() != ChecksumType::kNoChecksum &&
+      checksum != verification_info.checksum.ToString()) {
+    std::string msg = "Data is corrupted! Origin data checksum: " +
+                      verification_info.checksum.ToString() +
+                      "current data checksum: " + checksum;
+    return IOStatus::Corruption(msg);
+  }
+  target_->PositionedAppend(data, offset, options, dbg);
+  IOStatus io_s = fs_->InjectWriteError(state_.filename_);
+  return io_s;
+}
+
+IOStatus TestFSWritableFile::Close(const IOOptions& options,
+                                   IODebugContext* dbg) {
+  MutexLock l(&mutex_);
+  if (!fs_->IsFilesystemActive()) {
+    return fs_->GetError();
+  }
+  {
+    IOStatus in_s = fs_->InjectMetadataWriteError();
+    if (!in_s.ok()) {
+      return in_s;
+    }
+  }
+  writable_file_opened_ = false;
+  IOStatus io_s;
+  if (!target_->use_direct_io()) {
+    io_s = target_->Append(state_.buffer_, options, dbg);
+  }
+  if (io_s.ok()) {
+    state_.buffer_.resize(0);
+    // Ignore sync errors
+    target_->Sync(options, dbg).PermitUncheckedError();
+    io_s = target_->Close(options, dbg);
+  }
+  if (io_s.ok()) {
+    fs_->WritableFileClosed(state_);
+    IOStatus in_s = fs_->InjectMetadataWriteError();
+    if (!in_s.ok()) {
+      return in_s;
+    }
+  }
+  return io_s;
+}
+
+IOStatus TestFSWritableFile::Flush(const IOOptions&, IODebugContext*) {
+  MutexLock l(&mutex_);
+  if (!fs_->IsFilesystemActive()) {
+    return fs_->GetError();
+  }
+  if (fs_->IsFilesystemActive()) {
+    state_.pos_at_last_flush_ = state_.pos_;
+  }
+  return IOStatus::OK();
+}
+
+IOStatus TestFSWritableFile::Sync(const IOOptions& options,
+                                  IODebugContext* dbg) {
+  MutexLock l(&mutex_);
+  if (!fs_->IsFilesystemActive()) {
+    return fs_->GetError();
+  }
+  if (target_->use_direct_io()) {
+    // For Direct IO mode, we don't buffer anything in TestFSWritableFile.
+    // So just return
+    return IOStatus::OK();
+  }
+  IOStatus io_s = target_->Append(state_.buffer_, options, dbg);
+  state_.buffer_.resize(0);
+  // Ignore sync errors
+  target_->Sync(options, dbg).PermitUncheckedError();
+  state_.pos_at_last_sync_ = state_.pos_;
+  fs_->WritableFileSynced(state_);
+  return io_s;
+}
+
+IOStatus TestFSWritableFile::RangeSync(uint64_t offset, uint64_t nbytes,
+                                       const IOOptions& options,
+                                       IODebugContext* dbg) {
+  MutexLock l(&mutex_);
+  if (!fs_->IsFilesystemActive()) {
+    return fs_->GetError();
+  }
+  // Assumes caller passes consecutive byte ranges.
+  uint64_t sync_limit = offset + nbytes;
+  uint64_t buf_begin =
+      state_.pos_at_last_sync_ < 0 ? 0 : state_.pos_at_last_sync_;
+
+  IOStatus io_s;
+  if (sync_limit < buf_begin) {
+    return io_s;
+  }
+  uint64_t num_to_sync = std::min(static_cast<uint64_t>(state_.buffer_.size()),
+                                  sync_limit - buf_begin);
+  Slice buf_to_sync(state_.buffer_.data(), num_to_sync);
+  io_s = target_->Append(buf_to_sync, options, dbg);
+  state_.buffer_ = state_.buffer_.substr(num_to_sync);
+  // Ignore sync errors
+  target_->RangeSync(offset, nbytes, options, dbg).PermitUncheckedError();
+  state_.pos_at_last_sync_ = offset + num_to_sync;
+  fs_->WritableFileSynced(state_);
+  return io_s;
+}
+
+TestFSRandomRWFile::TestFSRandomRWFile(const std::string& /*fname*/,
+                                       std::unique_ptr<FSRandomRWFile>&& f,
+                                       FaultInjectionTestFS* fs)
+    : target_(std::move(f)), file_opened_(true), fs_(fs) {
+  assert(target_ != nullptr);
+}
+
+TestFSRandomRWFile::~TestFSRandomRWFile() {
+  if (file_opened_) {
+    Close(IOOptions(), nullptr).PermitUncheckedError();
+  }
+}
+
+IOStatus TestFSRandomRWFile::Write(uint64_t offset, const Slice& data,
+                                   const IOOptions& options,
+                                   IODebugContext* dbg) {
+  if (!fs_->IsFilesystemActive()) {
+    return fs_->GetError();
+  }
+  return target_->Write(offset, data, options, dbg);
+}
+
+IOStatus TestFSRandomRWFile::Read(uint64_t offset, size_t n,
+                                  const IOOptions& options, Slice* result,
+                                  char* scratch, IODebugContext* dbg) const {
+  if (!fs_->IsFilesystemActive()) {
+    return fs_->GetError();
+  }
+  return target_->Read(offset, n, options, result, scratch, dbg);
+}
+
+IOStatus TestFSRandomRWFile::Close(const IOOptions& options,
+                                   IODebugContext* dbg) {
+  if (!fs_->IsFilesystemActive()) {
+    return fs_->GetError();
+  }
+  file_opened_ = false;
+  return target_->Close(options, dbg);
+}
+
+IOStatus TestFSRandomRWFile::Flush(const IOOptions& options,
+                                   IODebugContext* dbg) {
+  if (!fs_->IsFilesystemActive()) {
+    return fs_->GetError();
+  }
+  return target_->Flush(options, dbg);
+}
+
+IOStatus TestFSRandomRWFile::Sync(const IOOptions& options,
+                                  IODebugContext* dbg) {
+  if (!fs_->IsFilesystemActive()) {
+    return fs_->GetError();
+  }
+  return target_->Sync(options, dbg);
+}
+
+TestFSRandomAccessFile::TestFSRandomAccessFile(
+    const std::string& /*fname*/, std::unique_ptr<FSRandomAccessFile>&& f,
+    FaultInjectionTestFS* fs)
+    : target_(std::move(f)), fs_(fs) {
+  assert(target_ != nullptr);
+}
+
+IOStatus TestFSRandomAccessFile::Read(uint64_t offset, size_t n,
+                                      const IOOptions& options, Slice* result,
+                                      char* scratch,
+                                      IODebugContext* dbg) const {
+  if (!fs_->IsFilesystemActive()) {
+    return fs_->GetError();
+  }
+  IOStatus s = target_->Read(offset, n, options, result, scratch, dbg);
+  if (s.ok()) {
+    s = fs_->InjectThreadSpecificReadError(
+        FaultInjectionTestFS::ErrorOperation::kRead, result, use_direct_io(),
+        scratch, /*need_count_increase=*/true, /*fault_injected=*/nullptr);
+  }
+  if (s.ok() && fs_->ShouldInjectRandomReadError()) {
+    return IOStatus::IOError("Injected read error");
+  }
+  return s;
+}
+
+IOStatus TestFSRandomAccessFile::MultiRead(FSReadRequest* reqs, size_t num_reqs,
+                                           const IOOptions& options,
+                                           IODebugContext* dbg) {
+  if (!fs_->IsFilesystemActive()) {
+    return fs_->GetError();
+  }
+  IOStatus s = target_->MultiRead(reqs, num_reqs, options, dbg);
+  bool injected_error = false;
+  for (size_t i = 0; i < num_reqs; i++) {
+    if (!reqs[i].status.ok()) {
+      // Already seeing an error.
+      break;
+    }
+    bool this_injected_error;
+    reqs[i].status = fs_->InjectThreadSpecificReadError(
+        FaultInjectionTestFS::ErrorOperation::kMultiReadSingleReq,
+        &(reqs[i].result), use_direct_io(), reqs[i].scratch,
+        /*need_count_increase=*/true,
+        /*fault_injected=*/&this_injected_error);
+    injected_error |= this_injected_error;
+  }
+  if (s.ok()) {
+    s = fs_->InjectThreadSpecificReadError(
+        FaultInjectionTestFS::ErrorOperation::kMultiRead, nullptr,
+        use_direct_io(), nullptr, /*need_count_increase=*/!injected_error,
+        /*fault_injected=*/nullptr);
+  }
+  if (s.ok() && fs_->ShouldInjectRandomReadError()) {
+    return IOStatus::IOError("Injected read error");
+  }
+  return s;
+}
+
+size_t TestFSRandomAccessFile::GetUniqueId(char* id, size_t max_size) const {
+  if (fs_->ShouldFailGetUniqueId()) {
+    return 0;
+  } else {
+    return target_->GetUniqueId(id, max_size);
+  }
+}
+IOStatus TestFSSequentialFile::Read(size_t n, const IOOptions& options,
+                                    Slice* result, char* scratch,
+                                    IODebugContext* dbg) {
+  IOStatus s = target()->Read(n, options, result, scratch, dbg);
+  if (s.ok() && fs_->ShouldInjectRandomReadError()) {
+    return IOStatus::IOError("Injected seq read error");
+  }
+  return s;
+}
+
+IOStatus TestFSSequentialFile::PositionedRead(uint64_t offset, size_t n,
+                                              const IOOptions& options,
+                                              Slice* result, char* scratch,
+                                              IODebugContext* dbg) {
+  IOStatus s =
+      target()->PositionedRead(offset, n, options, result, scratch, dbg);
+  if (s.ok() && fs_->ShouldInjectRandomReadError()) {
+    return IOStatus::IOError("Injected seq positioned read error");
+  }
+  return s;
+}
+
+IOStatus FaultInjectionTestFS::NewDirectory(
+    const std::string& name, const IOOptions& options,
+    std::unique_ptr<FSDirectory>* result, IODebugContext* dbg) {
+  std::unique_ptr<FSDirectory> r;
+  IOStatus io_s = target()->NewDirectory(name, options, &r, dbg);
+  if (!io_s.ok()) {
+    return io_s;
+  }
+  result->reset(
+      new TestFSDirectory(this, TestFSTrimDirname(name), r.release()));
+  return IOStatus::OK();
+}
+
+IOStatus FaultInjectionTestFS::NewWritableFile(
+    const std::string& fname, const FileOptions& file_opts,
+    std::unique_ptr<FSWritableFile>* result, IODebugContext* dbg) {
+  if (!IsFilesystemActive()) {
+    return GetError();
+  }
+  {
+    IOStatus in_s = InjectMetadataWriteError();
+    if (!in_s.ok()) {
+      return in_s;
+    }
+  }
+
+  if (ShouldUseDiretWritable(fname)) {
+    return target()->NewWritableFile(fname, file_opts, result, dbg);
+  }
+
+  IOStatus io_s = target()->NewWritableFile(fname, file_opts, result, dbg);
+  if (io_s.ok()) {
+    result->reset(
+        new TestFSWritableFile(fname, file_opts, std::move(*result), this));
+    // WritableFileWriter* file is opened
+    // again then it will be truncated - so forget our saved state.
+    UntrackFile(fname);
+    {
+      MutexLock l(&mutex_);
+      open_managed_files_.insert(fname);
+      auto dir_and_name = TestFSGetDirAndName(fname);
+      auto& list = dir_to_new_files_since_last_sync_[dir_and_name.first];
+      // The new file could overwrite an old one. Here we simplify
+      // the implementation by assuming no file of this name after
+      // dropping unsynced files.
+      list[dir_and_name.second] = kNewFileNoOverwrite;
+    }
+    {
+      IOStatus in_s = InjectMetadataWriteError();
+      if (!in_s.ok()) {
+        return in_s;
+      }
+    }
+  }
+  return io_s;
+}
+
+IOStatus FaultInjectionTestFS::ReopenWritableFile(
+    const std::string& fname, const FileOptions& file_opts,
+    std::unique_ptr<FSWritableFile>* result, IODebugContext* dbg) {
+  if (!IsFilesystemActive()) {
+    return GetError();
+  }
+  if (ShouldUseDiretWritable(fname)) {
+    return target()->ReopenWritableFile(fname, file_opts, result, dbg);
+  }
+  {
+    IOStatus in_s = InjectMetadataWriteError();
+    if (!in_s.ok()) {
+      return in_s;
+    }
+  }
+
+  bool exists;
+  IOStatus io_s,
+      exists_s = target()->FileExists(fname, IOOptions(), nullptr /* dbg */);
+  if (exists_s.IsNotFound()) {
+    exists = false;
+  } else if (exists_s.ok()) {
+    exists = true;
+  } else {
+    io_s = exists_s;
+    exists = false;
+  }
+
+  if (io_s.ok()) {
+    io_s = target()->ReopenWritableFile(fname, file_opts, result, dbg);
+  }
+
+  // Only track files we created. Files created outside of this
+  // `FaultInjectionTestFS` are not eligible for tracking/data dropping
+  // (for example, they may contain data a previous db_stress run expects to
+  // be recovered). This could be extended to track/drop data appended once
+  // the file is under `FaultInjectionTestFS`'s control.
+  if (io_s.ok()) {
+    bool should_track;
+    {
+      MutexLock l(&mutex_);
+      if (db_file_state_.find(fname) != db_file_state_.end()) {
+        // It was written by this `FileSystem` earlier.
+        assert(exists);
+        should_track = true;
+      } else if (!exists) {
+        // It was created by this `FileSystem` just now.
+        should_track = true;
+        open_managed_files_.insert(fname);
+        auto dir_and_name = TestFSGetDirAndName(fname);
+        auto& list = dir_to_new_files_since_last_sync_[dir_and_name.first];
+        list[dir_and_name.second] = kNewFileNoOverwrite;
+      } else {
+        should_track = false;
+      }
+    }
+    if (should_track) {
+      result->reset(
+          new TestFSWritableFile(fname, file_opts, std::move(*result), this));
+    }
+    {
+      IOStatus in_s = InjectMetadataWriteError();
+      if (!in_s.ok()) {
+        return in_s;
+      }
+    }
+  }
+  return io_s;
+}
+
+IOStatus FaultInjectionTestFS::NewRandomRWFile(
+    const std::string& fname, const FileOptions& file_opts,
+    std::unique_ptr<FSRandomRWFile>* result, IODebugContext* dbg) {
+  if (!IsFilesystemActive()) {
+    return GetError();
+  }
+  if (ShouldUseDiretWritable(fname)) {
+    return target()->NewRandomRWFile(fname, file_opts, result, dbg);
+  }
+  {
+    IOStatus in_s = InjectMetadataWriteError();
+    if (!in_s.ok()) {
+      return in_s;
+    }
+  }
+  IOStatus io_s = target()->NewRandomRWFile(fname, file_opts, result, dbg);
+  if (io_s.ok()) {
+    result->reset(new TestFSRandomRWFile(fname, std::move(*result), this));
+    // WritableFileWriter* file is opened
+    // again then it will be truncated - so forget our saved state.
+    UntrackFile(fname);
+    {
+      MutexLock l(&mutex_);
+      open_managed_files_.insert(fname);
+      auto dir_and_name = TestFSGetDirAndName(fname);
+      auto& list = dir_to_new_files_since_last_sync_[dir_and_name.first];
+      // It could be overwriting an old file, but we simplify the
+      // implementation by ignoring it.
+      list[dir_and_name.second] = kNewFileNoOverwrite;
+    }
+    {
+      IOStatus in_s = InjectMetadataWriteError();
+      if (!in_s.ok()) {
+        return in_s;
+      }
+    }
+  }
+  return io_s;
+}
+
+IOStatus FaultInjectionTestFS::NewRandomAccessFile(
+    const std::string& fname, const FileOptions& file_opts,
+    std::unique_ptr<FSRandomAccessFile>* result, IODebugContext* dbg) {
+  if (!IsFilesystemActive()) {
+    return GetError();
+  }
+  if (ShouldInjectRandomReadError()) {
+    return IOStatus::IOError("Injected error when open random access file");
+  }
+  IOStatus io_s = InjectThreadSpecificReadError(ErrorOperation::kOpen, nullptr,
+                                                false, nullptr,
+                                                /*need_count_increase=*/true,
+                                                /*fault_injected=*/nullptr);
+  if (io_s.ok()) {
+    io_s = target()->NewRandomAccessFile(fname, file_opts, result, dbg);
+  }
+  if (io_s.ok()) {
+    result->reset(new TestFSRandomAccessFile(fname, std::move(*result), this));
+  }
+  return io_s;
+}
+
+IOStatus FaultInjectionTestFS::NewSequentialFile(
+    const std::string& fname, const FileOptions& file_opts,
+    std::unique_ptr<FSSequentialFile>* result, IODebugContext* dbg) {
+  if (!IsFilesystemActive()) {
+    return GetError();
+  }
+
+  if (ShouldInjectRandomReadError()) {
+    return IOStatus::IOError("Injected read error when creating seq file");
+  }
+  IOStatus io_s = target()->NewSequentialFile(fname, file_opts, result, dbg);
+  if (io_s.ok()) {
+    result->reset(new TestFSSequentialFile(std::move(*result), this));
+  }
+  return io_s;
+}
+
+IOStatus FaultInjectionTestFS::DeleteFile(const std::string& f,
+                                          const IOOptions& options,
+                                          IODebugContext* dbg) {
+  if (!IsFilesystemActive()) {
+    return GetError();
+  }
+  {
+    IOStatus in_s = InjectMetadataWriteError();
+    if (!in_s.ok()) {
+      return in_s;
+    }
+  }
+  IOStatus io_s = FileSystemWrapper::DeleteFile(f, options, dbg);
+  if (io_s.ok()) {
+    UntrackFile(f);
+    {
+      IOStatus in_s = InjectMetadataWriteError();
+      if (!in_s.ok()) {
+        return in_s;
+      }
+    }
+  }
+  return io_s;
+}
+
+IOStatus FaultInjectionTestFS::RenameFile(const std::string& s,
+                                          const std::string& t,
+                                          const IOOptions& options,
+                                          IODebugContext* dbg) {
+  if (!IsFilesystemActive()) {
+    return GetError();
+  }
+  {
+    IOStatus in_s = InjectMetadataWriteError();
+    if (!in_s.ok()) {
+      return in_s;
+    }
+  }
+
+  // We preserve contents of overwritten files up to a size threshold.
+  // We could keep previous file in another name, but we need to worry about
+  // garbage collect the those files. We do it if it is needed later.
+  // We ignore I/O errors here for simplicity.
+  std::string previous_contents = kNewFileNoOverwrite;
+  if (target()->FileExists(t, IOOptions(), nullptr).ok()) {
+    uint64_t file_size;
+    if (target()->GetFileSize(t, IOOptions(), &file_size, nullptr).ok() &&
+        file_size < 1024) {
+      ReadFileToString(target(), t, &previous_contents).PermitUncheckedError();
+    }
+  }
+  IOStatus io_s = FileSystemWrapper::RenameFile(s, t, options, dbg);
+
+  if (io_s.ok()) {
+    {
+      MutexLock l(&mutex_);
+      if (db_file_state_.find(s) != db_file_state_.end()) {
+        db_file_state_[t] = db_file_state_[s];
+        db_file_state_.erase(s);
+      }
+
+      auto sdn = TestFSGetDirAndName(s);
+      auto tdn = TestFSGetDirAndName(t);
+      if (dir_to_new_files_since_last_sync_[sdn.first].erase(sdn.second) != 0) {
+        auto& tlist = dir_to_new_files_since_last_sync_[tdn.first];
+        assert(tlist.find(tdn.second) == tlist.end());
+        tlist[tdn.second] = previous_contents;
+      }
+    }
+    IOStatus in_s = InjectMetadataWriteError();
+    if (!in_s.ok()) {
+      return in_s;
+    }
+  }
+
+  return io_s;
+}
+
+IOStatus FaultInjectionTestFS::LinkFile(const std::string& s,
+                                        const std::string& t,
+                                        const IOOptions& options,
+                                        IODebugContext* dbg) {
+  if (!IsFilesystemActive()) {
+    return GetError();
+  }
+  {
+    IOStatus in_s = InjectMetadataWriteError();
+    if (!in_s.ok()) {
+      return in_s;
+    }
+  }
+
+  // Using the value in `dir_to_new_files_since_last_sync_` for the source file
+  // may be a more reasonable choice.
+  std::string previous_contents = kNewFileNoOverwrite;
+
+  IOStatus io_s = FileSystemWrapper::LinkFile(s, t, options, dbg);
+
+  if (io_s.ok()) {
+    {
+      MutexLock l(&mutex_);
+      if (db_file_state_.find(s) != db_file_state_.end()) {
+        db_file_state_[t] = db_file_state_[s];
+      }
+
+      auto sdn = TestFSGetDirAndName(s);
+      auto tdn = TestFSGetDirAndName(t);
+      if (dir_to_new_files_since_last_sync_[sdn.first].find(sdn.second) !=
+          dir_to_new_files_since_last_sync_[sdn.first].end()) {
+        auto& tlist = dir_to_new_files_since_last_sync_[tdn.first];
+        assert(tlist.find(tdn.second) == tlist.end());
+        tlist[tdn.second] = previous_contents;
+      }
+    }
+    IOStatus in_s = InjectMetadataWriteError();
+    if (!in_s.ok()) {
+      return in_s;
+    }
+  }
+
+  return io_s;
+}
+
+void FaultInjectionTestFS::WritableFileClosed(const FSFileState& state) {
+  MutexLock l(&mutex_);
+  if (open_managed_files_.find(state.filename_) != open_managed_files_.end()) {
+    db_file_state_[state.filename_] = state;
+    open_managed_files_.erase(state.filename_);
+  }
+}
+
+void FaultInjectionTestFS::WritableFileSynced(const FSFileState& state) {
+  MutexLock l(&mutex_);
+  if (open_managed_files_.find(state.filename_) != open_managed_files_.end()) {
+    if (db_file_state_.find(state.filename_) == db_file_state_.end()) {
+      db_file_state_.insert(std::make_pair(state.filename_, state));
+    } else {
+      db_file_state_[state.filename_] = state;
+    }
+  }
+}
+
+void FaultInjectionTestFS::WritableFileAppended(const FSFileState& state) {
+  MutexLock l(&mutex_);
+  if (open_managed_files_.find(state.filename_) != open_managed_files_.end()) {
+    if (db_file_state_.find(state.filename_) == db_file_state_.end()) {
+      db_file_state_.insert(std::make_pair(state.filename_, state));
+    } else {
+      db_file_state_[state.filename_] = state;
+    }
+  }
+}
+
+IOStatus FaultInjectionTestFS::DropUnsyncedFileData() {
+  IOStatus io_s;
+  MutexLock l(&mutex_);
+  for (std::map<std::string, FSFileState>::iterator it = db_file_state_.begin();
+       io_s.ok() && it != db_file_state_.end(); ++it) {
+    FSFileState& fs_state = it->second;
+    if (!fs_state.IsFullySynced()) {
+      io_s = fs_state.DropUnsyncedData();
+    }
+  }
+  return io_s;
+}
+
+IOStatus FaultInjectionTestFS::DropRandomUnsyncedFileData(Random* rnd) {
+  IOStatus io_s;
+  MutexLock l(&mutex_);
+  for (std::map<std::string, FSFileState>::iterator it = db_file_state_.begin();
+       io_s.ok() && it != db_file_state_.end(); ++it) {
+    FSFileState& fs_state = it->second;
+    if (!fs_state.IsFullySynced()) {
+      io_s = fs_state.DropRandomUnsyncedData(rnd);
+    }
+  }
+  return io_s;
+}
+
+IOStatus FaultInjectionTestFS::DeleteFilesCreatedAfterLastDirSync(
+    const IOOptions& options, IODebugContext* dbg) {
+  // Because DeleteFile access this container make a copy to avoid deadlock
+  std::map<std::string, std::map<std::string, std::string>> map_copy;
+  {
+    MutexLock l(&mutex_);
+    map_copy.insert(dir_to_new_files_since_last_sync_.begin(),
+                    dir_to_new_files_since_last_sync_.end());
+  }
+
+  for (auto& pair : map_copy) {
+    for (auto& file_pair : pair.second) {
+      if (file_pair.second == kNewFileNoOverwrite) {
+        IOStatus io_s =
+            DeleteFile(pair.first + "/" + file_pair.first, options, dbg);
+        if (!io_s.ok()) {
+          return io_s;
+        }
+      } else {
+        IOStatus io_s =
+            WriteStringToFile(target(), file_pair.second,
+                              pair.first + "/" + file_pair.first, true);
+        if (!io_s.ok()) {
+          return io_s;
+        }
+      }
+    }
+  }
+  return IOStatus::OK();
+}
+
+void FaultInjectionTestFS::ResetState() {
+  MutexLock l(&mutex_);
+  db_file_state_.clear();
+  dir_to_new_files_since_last_sync_.clear();
+  SetFilesystemActiveNoLock(true);
+}
+
+void FaultInjectionTestFS::UntrackFile(const std::string& f) {
+  MutexLock l(&mutex_);
+  auto dir_and_name = TestFSGetDirAndName(f);
+  dir_to_new_files_since_last_sync_[dir_and_name.first].erase(
+      dir_and_name.second);
+  db_file_state_.erase(f);
+  open_managed_files_.erase(f);
+}
+
+IOStatus FaultInjectionTestFS::InjectThreadSpecificReadError(
+    ErrorOperation op, Slice* result, bool direct_io, char* scratch,
+    bool need_count_increase, bool* fault_injected) {
+  bool dummy_bool;
+  bool& ret_fault_injected = fault_injected ? *fault_injected : dummy_bool;
+  ret_fault_injected = false;
+  ErrorContext* ctx = static_cast<ErrorContext*>(thread_local_error_->Get());
+  if (ctx == nullptr || !ctx->enable_error_injection || !ctx->one_in) {
+    return IOStatus::OK();
+  }
+
+  if (ctx->rand.OneIn(ctx->one_in)) {
+    if (ctx->count == 0) {
+      ctx->message = "";
+    }
+    if (need_count_increase) {
+      ctx->count++;
+    }
+    if (ctx->callstack) {
+      free(ctx->callstack);
+    }
+    ctx->callstack = port::SaveStack(&ctx->frames);
+
+    if (op != ErrorOperation::kMultiReadSingleReq) {
+      // Likely non-per read status code for MultiRead
+      ctx->message += "error; ";
+      ret_fault_injected = true;
+      return IOStatus::IOError();
+    } else if (Random::GetTLSInstance()->OneIn(8)) {
+      assert(result);
+      // For a small chance, set the failure to status but turn the
+      // result to be empty, which is supposed to be caught for a check.
+      *result = Slice();
+      ctx->message += "inject empty result; ";
+      ret_fault_injected = true;
+    } else if (!direct_io && Random::GetTLSInstance()->OneIn(7) &&
+               scratch != nullptr && result->data() == scratch) {
+      assert(result);
+      // With direct I/O, many extra bytes might be read so corrupting
+      // one byte might not cause checksum mismatch. Skip checksum
+      // corruption injection.
+      // We only corrupt data if the result is filled to `scratch`. For other
+      // cases, the data might not be able to be modified (e.g mmaped files)
+      // or has unintended side effects.
+      // For a small chance, set the failure to status but corrupt the
+      // result in a way that checksum checking is supposed to fail.
+      // Corrupt the last byte, which is supposed to be a checksum byte
+      // It would work for CRC. Not 100% sure for xxhash and will adjust
+      // if it is not the case.
+      const_cast<char*>(result->data())[result->size() - 1]++;
+      ctx->message += "corrupt last byte; ";
+      ret_fault_injected = true;
+    } else {
+      ctx->message += "error result multiget single; ";
+      ret_fault_injected = true;
+      return IOStatus::IOError();
+    }
+  }
+  return IOStatus::OK();
+}
+
+bool FaultInjectionTestFS::TryParseFileName(const std::string& file_name,
+                                            uint64_t* number, FileType* type) {
+  std::size_t found = file_name.find_last_of("/");
+  std::string file = file_name.substr(found);
+  return ParseFileName(file, number, type);
+}
+
+IOStatus FaultInjectionTestFS::InjectWriteError(const std::string& file_name) {
+  MutexLock l(&mutex_);
+  if (!enable_write_error_injection_ || !write_error_one_in_) {
+    return IOStatus::OK();
+  }
+  bool allowed_type = false;
+
+  if (inject_for_all_file_types_) {
+    allowed_type = true;
+  } else {
+    uint64_t number;
+    FileType cur_type = kTempFile;
+    if (TryParseFileName(file_name, &number, &cur_type)) {
+      for (const auto& type : write_error_allowed_types_) {
+        if (cur_type == type) {
+          allowed_type = true;
+        }
+      }
+    }
+  }
+
+  if (allowed_type) {
+    if (write_error_rand_.OneIn(write_error_one_in_)) {
+      return GetError();
+    }
+  }
+  return IOStatus::OK();
+}
+
+IOStatus FaultInjectionTestFS::InjectMetadataWriteError() {
+  {
+    MutexLock l(&mutex_);
+    if (!enable_metadata_write_error_injection_ ||
+        !metadata_write_error_one_in_ ||
+        !write_error_rand_.OneIn(metadata_write_error_one_in_)) {
+      return IOStatus::OK();
+    }
+  }
+  TEST_SYNC_POINT("FaultInjectionTestFS::InjectMetadataWriteError:Injected");
+  return IOStatus::IOError();
+}
+
+void FaultInjectionTestFS::PrintFaultBacktrace() {
+#if defined(OS_LINUX)
+  ErrorContext* ctx = static_cast<ErrorContext*>(thread_local_error_->Get());
+  if (ctx == nullptr) {
+    return;
+  }
+  fprintf(stderr, "Injected error type = %d\n", ctx->type);
+  fprintf(stderr, "Message: %s\n", ctx->message.c_str());
+  port::PrintAndFreeStack(ctx->callstack, ctx->frames);
+  ctx->callstack = nullptr;
+#endif
+}
+
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/utilities/fault_injection_fs.h b/src/rocksdb/utilities/fault_injection_fs.h
new file mode 100644
index 000000000..53c9ccb6f
--- /dev/null
+++ b/src/rocksdb/utilities/fault_injection_fs.h
@@ -0,0 +1,584 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+//
+// Copyright 2014 The LevelDB Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file. See the AUTHORS file for names of contributors.
+
+// This test uses a custom FileSystem to keep track of the state of a file
+// system the last "Sync". The data being written is cached in a "buffer".
+// Only when "Sync" is called, the data will be persistent. It can similate
+// file data loss (or entire files) not protected by a "Sync". For any of the
+// FileSystem related operations, by specify the "IOStatus Error", a specific
+// error can be returned when file system is not activated.
+
+#pragma once
+
+#include <algorithm>
+#include <map>
+#include <set>
+#include <string>
+
+#include "file/filename.h"
+#include "rocksdb/file_system.h"
+#include "util/mutexlock.h"
+#include "util/random.h"
+#include "util/thread_local.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+class TestFSWritableFile;
+class FaultInjectionTestFS;
+
+struct FSFileState {
+  std::string filename_;
+  ssize_t pos_;
+  ssize_t pos_at_last_sync_;
+  ssize_t pos_at_last_flush_;
+  std::string buffer_;
+
+  explicit FSFileState(const std::string& filename)
+      : filename_(filename),
+        pos_(-1),
+        pos_at_last_sync_(-1),
+        pos_at_last_flush_(-1) {}
+
+  FSFileState() : pos_(-1), pos_at_last_sync_(-1), pos_at_last_flush_(-1) {}
+
+  bool IsFullySynced() const { return pos_ <= 0 || pos_ == pos_at_last_sync_; }
+
+  IOStatus DropUnsyncedData();
+
+  IOStatus DropRandomUnsyncedData(Random* rand);
+};
+
+// A wrapper around WritableFileWriter* file
+// is written to or sync'ed.
+class TestFSWritableFile : public FSWritableFile {
+ public:
+  explicit TestFSWritableFile(const std::string& fname,
+                              const FileOptions& file_opts,
+                              std::unique_ptr<FSWritableFile>&& f,
+                              FaultInjectionTestFS* fs);
+  virtual ~TestFSWritableFile();
+  virtual IOStatus Append(const Slice& data, const IOOptions&,
+                          IODebugContext*) override;
+  virtual IOStatus Append(const Slice& data, const IOOptions& options,
+                          const DataVerificationInfo& verification_info,
+                          IODebugContext* dbg) override;
+  virtual IOStatus Truncate(uint64_t size, const IOOptions& options,
+                            IODebugContext* dbg) override {
+    return target_->Truncate(size, options, dbg);
+  }
+  virtual IOStatus Close(const IOOptions& options,
+                         IODebugContext* dbg) override;
+  virtual IOStatus Flush(const IOOptions&, IODebugContext*) override;
+  virtual IOStatus Sync(const IOOptions& options, IODebugContext* dbg) override;
+  virtual IOStatus RangeSync(uint64_t /*offset*/, uint64_t /*nbytes*/,
+                             const IOOptions& options,
+                             IODebugContext* dbg) override;
+  virtual bool IsSyncThreadSafe() const override { return true; }
+  virtual IOStatus PositionedAppend(const Slice& data, uint64_t offset,
+                                    const IOOptions& options,
+                                    IODebugContext* dbg) override {
+    return target_->PositionedAppend(data, offset, options, dbg);
+  }
+  IOStatus PositionedAppend(const Slice& data, uint64_t offset,
+                            const IOOptions& options,
+                            const DataVerificationInfo& verification_info,
+                            IODebugContext* dbg) override;
+  virtual size_t GetRequiredBufferAlignment() const override {
+    return target_->GetRequiredBufferAlignment();
+  }
+  virtual bool use_direct_io() const override {
+    return target_->use_direct_io();
+  };
+
+ private:
+  FSFileState state_;  // Need protection by mutex_
+  FileOptions file_opts_;
+  std::unique_ptr<FSWritableFile> target_;
+  bool writable_file_opened_;
+  FaultInjectionTestFS* fs_;
+  port::Mutex mutex_;
+};
+
+// A wrapper around WritableFileWriter* file
+// is written to or sync'ed.
+class TestFSRandomRWFile : public FSRandomRWFile {
+ public:
+  explicit TestFSRandomRWFile(const std::string& fname,
+                              std::unique_ptr<FSRandomRWFile>&& f,
+                              FaultInjectionTestFS* fs);
+  virtual ~TestFSRandomRWFile();
+  IOStatus Write(uint64_t offset, const Slice& data, const IOOptions& options,
+                 IODebugContext* dbg) override;
+  IOStatus Read(uint64_t offset, size_t n, const IOOptions& options,
+                Slice* result, char* scratch,
+                IODebugContext* dbg) const override;
+  IOStatus Close(const IOOptions& options, IODebugContext* dbg) override;
+  IOStatus Flush(const IOOptions& options, IODebugContext* dbg) override;
+  IOStatus Sync(const IOOptions& options, IODebugContext* dbg) override;
+  size_t GetRequiredBufferAlignment() const override {
+    return target_->GetRequiredBufferAlignment();
+  }
+  bool use_direct_io() const override { return target_->use_direct_io(); };
+
+ private:
+  std::unique_ptr<FSRandomRWFile> target_;
+  bool file_opened_;
+  FaultInjectionTestFS* fs_;
+};
+
+class TestFSRandomAccessFile : public FSRandomAccessFile {
+ public:
+  explicit TestFSRandomAccessFile(const std::string& fname,
+                                  std::unique_ptr<FSRandomAccessFile>&& f,
+                                  FaultInjectionTestFS* fs);
+  ~TestFSRandomAccessFile() override {}
+  IOStatus Read(uint64_t offset, size_t n, const IOOptions& options,
+                Slice* result, char* scratch,
+                IODebugContext* dbg) const override;
+  IOStatus MultiRead(FSReadRequest* reqs, size_t num_reqs,
+                     const IOOptions& options, IODebugContext* dbg) override;
+  size_t GetRequiredBufferAlignment() const override {
+    return target_->GetRequiredBufferAlignment();
+  }
+  bool use_direct_io() const override { return target_->use_direct_io(); }
+
+  size_t GetUniqueId(char* id, size_t max_size) const override;
+
+ private:
+  std::unique_ptr<FSRandomAccessFile> target_;
+  FaultInjectionTestFS* fs_;
+};
+
+class TestFSSequentialFile : public FSSequentialFileOwnerWrapper {
+ public:
+  explicit TestFSSequentialFile(std::unique_ptr<FSSequentialFile>&& f,
+                                FaultInjectionTestFS* fs)
+      : FSSequentialFileOwnerWrapper(std::move(f)), fs_(fs) {}
+  IOStatus Read(size_t n, const IOOptions& options, Slice* result,
+                char* scratch, IODebugContext* dbg) override;
+  IOStatus PositionedRead(uint64_t offset, size_t n, const IOOptions& options,
+                          Slice* result, char* scratch,
+                          IODebugContext* dbg) override;
+
+ private:
+  FaultInjectionTestFS* fs_;
+};
+
+class TestFSDirectory : public FSDirectory {
+ public:
+  explicit TestFSDirectory(FaultInjectionTestFS* fs, std::string dirname,
+                           FSDirectory* dir)
+      : fs_(fs), dirname_(dirname), dir_(dir) {}
+  ~TestFSDirectory() {}
+
+  virtual IOStatus Fsync(const IOOptions& options,
+                         IODebugContext* dbg) override;
+
+  virtual IOStatus Close(const IOOptions& options,
+                         IODebugContext* dbg) override;
+
+  virtual IOStatus FsyncWithDirOptions(
+      const IOOptions& options, IODebugContext* dbg,
+      const DirFsyncOptions& dir_fsync_options) override;
+
+ private:
+  FaultInjectionTestFS* fs_;
+  std::string dirname_;
+  std::unique_ptr<FSDirectory> dir_;
+};
+
+class FaultInjectionTestFS : public FileSystemWrapper {
+ public:
+  explicit FaultInjectionTestFS(const std::shared_ptr<FileSystem>& base)
+      : FileSystemWrapper(base),
+        filesystem_active_(true),
+        filesystem_writable_(false),
+        thread_local_error_(new ThreadLocalPtr(DeleteThreadLocalErrorContext)),
+        enable_write_error_injection_(false),
+        enable_metadata_write_error_injection_(false),
+        write_error_rand_(0),
+        write_error_one_in_(0),
+        metadata_write_error_one_in_(0),
+        read_error_one_in_(0),
+        ingest_data_corruption_before_write_(false),
+        fail_get_file_unique_id_(false) {}
+  virtual ~FaultInjectionTestFS() { error_.PermitUncheckedError(); }
+
+  static const char* kClassName() { return "FaultInjectionTestFS"; }
+  const char* Name() const override { return kClassName(); }
+
+  IOStatus NewDirectory(const std::string& name, const IOOptions& options,
+                        std::unique_ptr<FSDirectory>* result,
+                        IODebugContext* dbg) override;
+
+  IOStatus NewWritableFile(const std::string& fname,
+                           const FileOptions& file_opts,
+                           std::unique_ptr<FSWritableFile>* result,
+                           IODebugContext* dbg) override;
+
+  IOStatus ReopenWritableFile(const std::string& fname,
+                              const FileOptions& file_opts,
+                              std::unique_ptr<FSWritableFile>* result,
+                              IODebugContext* dbg) override;
+
+  IOStatus NewRandomRWFile(const std::string& fname,
+                           const FileOptions& file_opts,
+                           std::unique_ptr<FSRandomRWFile>* result,
+                           IODebugContext* dbg) override;
+
+  IOStatus NewRandomAccessFile(const std::string& fname,
+                               const FileOptions& file_opts,
+                               std::unique_ptr<FSRandomAccessFile>* result,
+                               IODebugContext* dbg) override;
+  IOStatus NewSequentialFile(const std::string& f, const FileOptions& file_opts,
+                             std::unique_ptr<FSSequentialFile>* r,
+                             IODebugContext* dbg) override;
+
+  virtual IOStatus DeleteFile(const std::string& f, const IOOptions& options,
+                              IODebugContext* dbg) override;
+
+  virtual IOStatus RenameFile(const std::string& s, const std::string& t,
+                              const IOOptions& options,
+                              IODebugContext* dbg) override;
+
+  virtual IOStatus LinkFile(const std::string& src, const std::string& target,
+                            const IOOptions& options,
+                            IODebugContext* dbg) override;
+
+// Undef to eliminate clash on Windows
+#undef GetFreeSpace
+  virtual IOStatus GetFreeSpace(const std::string& path,
+                                const IOOptions& options, uint64_t* disk_free,
+                                IODebugContext* dbg) override {
+    IOStatus io_s;
+    if (!IsFilesystemActive() &&
+        error_.subcode() == IOStatus::SubCode::kNoSpace) {
+      *disk_free = 0;
+    } else {
+      io_s = target()->GetFreeSpace(path, options, disk_free, dbg);
+    }
+    return io_s;
+  }
+
+  void WritableFileClosed(const FSFileState& state);
+
+  void WritableFileSynced(const FSFileState& state);
+
+  void WritableFileAppended(const FSFileState& state);
+
+  IOStatus DropUnsyncedFileData();
+
+  IOStatus DropRandomUnsyncedFileData(Random* rnd);
+
+  IOStatus DeleteFilesCreatedAfterLastDirSync(const IOOptions& options,
+                                              IODebugContext* dbg);
+
+  void ResetState();
+
+  void UntrackFile(const std::string& f);
+
+  void SyncDir(const std::string& dirname) {
+    MutexLock l(&mutex_);
+    dir_to_new_files_since_last_sync_.erase(dirname);
+  }
+
+  // Setting the filesystem to inactive is the test equivalent to simulating a
+  // system reset. Setting to inactive will freeze our saved filesystem state so
+  // that it will stop being recorded. It can then be reset back to the state at
+  // the time of the reset.
+  bool IsFilesystemActive() {
+    MutexLock l(&mutex_);
+    return filesystem_active_;
+  }
+
+  // Setting filesystem_writable_ makes NewWritableFile. ReopenWritableFile,
+  // and NewRandomRWFile bypass FaultInjectionTestFS and go directly to the
+  // target FS
+  bool IsFilesystemDirectWritable() {
+    MutexLock l(&mutex_);
+    return filesystem_writable_;
+  }
+  bool ShouldUseDiretWritable(const std::string& file_name) {
+    MutexLock l(&mutex_);
+    if (filesystem_writable_) {
+      return true;
+    }
+    FileType file_type = kTempFile;
+    uint64_t file_number = 0;
+    if (!TryParseFileName(file_name, &file_number, &file_type)) {
+      return false;
+    }
+    return skip_direct_writable_types_.find(file_type) !=
+           skip_direct_writable_types_.end();
+  }
+  void SetFilesystemActiveNoLock(
+      bool active, IOStatus error = IOStatus::Corruption("Not active")) {
+    error.PermitUncheckedError();
+    filesystem_active_ = active;
+    if (!active) {
+      error_ = error;
+    }
+  }
+  void SetFilesystemActive(
+      bool active, IOStatus error = IOStatus::Corruption("Not active")) {
+    MutexLock l(&mutex_);
+    error.PermitUncheckedError();
+    SetFilesystemActiveNoLock(active, error);
+  }
+  void SetFilesystemDirectWritable(bool writable) {
+    MutexLock l(&mutex_);
+    filesystem_writable_ = writable;
+  }
+  void AssertNoOpenFile() { assert(open_managed_files_.empty()); }
+
+  IOStatus GetError() { return error_; }
+
+  void SetFileSystemIOError(IOStatus io_error) {
+    MutexLock l(&mutex_);
+    io_error.PermitUncheckedError();
+    error_ = io_error;
+  }
+
+  // To simulate the data corruption before data is written in FS
+  void IngestDataCorruptionBeforeWrite() {
+    MutexLock l(&mutex_);
+    ingest_data_corruption_before_write_ = true;
+  }
+
+  void NoDataCorruptionBeforeWrite() {
+    MutexLock l(&mutex_);
+    ingest_data_corruption_before_write_ = false;
+  }
+
+  bool ShouldDataCorruptionBeforeWrite() {
+    MutexLock l(&mutex_);
+    return ingest_data_corruption_before_write_;
+  }
+
+  void SetChecksumHandoffFuncType(const ChecksumType& func_type) {
+    MutexLock l(&mutex_);
+    checksum_handoff_func_tpye_ = func_type;
+  }
+
+  const ChecksumType& GetChecksumHandoffFuncType() {
+    MutexLock l(&mutex_);
+    return checksum_handoff_func_tpye_;
+  }
+
+  void SetFailGetUniqueId(bool flag) {
+    MutexLock l(&mutex_);
+    fail_get_file_unique_id_ = flag;
+  }
+
+  bool ShouldFailGetUniqueId() {
+    MutexLock l(&mutex_);
+    return fail_get_file_unique_id_;
+  }
+
+  // Specify what the operation, so we can inject the right type of error
+  enum ErrorOperation : char {
+    kRead = 0,
+    kMultiReadSingleReq = 1,
+    kMultiRead = 2,
+    kOpen,
+  };
+
+  // Set thread-local parameters for error injection. The first argument,
+  // seed is the seed for the random number generator, and one_in determines
+  // the probability of injecting error (i.e an error is injected with
+  // 1/one_in probability)
+  void SetThreadLocalReadErrorContext(uint32_t seed, int one_in) {
+    struct ErrorContext* ctx =
+        static_cast<struct ErrorContext*>(thread_local_error_->Get());
+    if (ctx == nullptr) {
+      ctx = new ErrorContext(seed);
+      thread_local_error_->Reset(ctx);
+    }
+    ctx->one_in = one_in;
+    ctx->count = 0;
+  }
+
+  static void DeleteThreadLocalErrorContext(void* p) {
+    ErrorContext* ctx = static_cast<ErrorContext*>(p);
+    delete ctx;
+  }
+
+  // This is to set the parameters for the write error injection.
+  // seed is the seed for the random number generator, and one_in determines
+  // the probability of injecting error (i.e an error is injected with
+  // 1/one_in probability). For write error, we can specify the error we
+  // want to inject. Types decides the file types we want to inject the
+  // error (e.g., Wal files, SST files), which is empty by default.
+  void SetRandomWriteError(uint32_t seed, int one_in, IOStatus error,
+                           bool inject_for_all_file_types,
+                           const std::vector<FileType>& types) {
+    MutexLock l(&mutex_);
+    Random tmp_rand(seed);
+    error.PermitUncheckedError();
+    error_ = error;
+    write_error_rand_ = tmp_rand;
+    write_error_one_in_ = one_in;
+    inject_for_all_file_types_ = inject_for_all_file_types;
+    write_error_allowed_types_ = types;
+  }
+
+  void SetSkipDirectWritableTypes(const std::set<FileType>& types) {
+    MutexLock l(&mutex_);
+    skip_direct_writable_types_ = types;
+  }
+
+  void SetRandomMetadataWriteError(int one_in) {
+    MutexLock l(&mutex_);
+    metadata_write_error_one_in_ = one_in;
+  }
+  // If the value is not 0, it is enabled. Otherwise, it is disabled.
+  void SetRandomReadError(int one_in) { read_error_one_in_ = one_in; }
+
+  bool ShouldInjectRandomReadError() {
+    return read_error_one_in() &&
+           Random::GetTLSInstance()->OneIn(read_error_one_in());
+  }
+
+  // Inject an write error with randomlized parameter and the predefined
+  // error type. Only the allowed file types will inject the write error
+  IOStatus InjectWriteError(const std::string& file_name);
+
+  // Ingest error to metadata operations.
+  IOStatus InjectMetadataWriteError();
+
+  // Inject an error. For a READ operation, a status of IOError(), a
+  // corruption in the contents of scratch, or truncation of slice
+  // are the types of error with equal probability. For OPEN,
+  // its always an IOError.
+  // fault_injected returns whether a fault is injected. It is needed
+  // because some fault is inected with IOStatus to be OK.
+  IOStatus InjectThreadSpecificReadError(ErrorOperation op, Slice* slice,
+                                         bool direct_io, char* scratch,
+                                         bool need_count_increase,
+                                         bool* fault_injected);
+
+  // Get the count of how many times we injected since the previous call
+  int GetAndResetErrorCount() {
+    ErrorContext* ctx = static_cast<ErrorContext*>(thread_local_error_->Get());
+    int count = 0;
+    if (ctx != nullptr) {
+      count = ctx->count;
+      ctx->count = 0;
+    }
+    return count;
+  }
+
+  void EnableErrorInjection() {
+    ErrorContext* ctx = static_cast<ErrorContext*>(thread_local_error_->Get());
+    if (ctx) {
+      ctx->enable_error_injection = true;
+    }
+  }
+
+  void EnableWriteErrorInjection() {
+    MutexLock l(&mutex_);
+    enable_write_error_injection_ = true;
+  }
+  void EnableMetadataWriteErrorInjection() {
+    MutexLock l(&mutex_);
+    enable_metadata_write_error_injection_ = true;
+  }
+
+  void DisableWriteErrorInjection() {
+    MutexLock l(&mutex_);
+    enable_write_error_injection_ = false;
+  }
+
+  void DisableErrorInjection() {
+    ErrorContext* ctx = static_cast<ErrorContext*>(thread_local_error_->Get());
+    if (ctx) {
+      ctx->enable_error_injection = false;
+    }
+  }
+
+  void DisableMetadataWriteErrorInjection() {
+    MutexLock l(&mutex_);
+    enable_metadata_write_error_injection_ = false;
+  }
+
+  int read_error_one_in() const { return read_error_one_in_.load(); }
+
+  int write_error_one_in() const { return write_error_one_in_; }
+
+  // We capture a backtrace every time a fault is injected, for debugging
+  // purposes. This call prints the backtrace to stderr and frees the
+  // saved callstack
+  void PrintFaultBacktrace();
+
+ private:
+  port::Mutex mutex_;
+  std::map<std::string, FSFileState> db_file_state_;
+  std::set<std::string> open_managed_files_;
+  // directory -> (file name -> file contents to recover)
+  // When data is recovered from unsyned parent directory, the files with
+  // empty file contents to recover is deleted. Those with non-empty ones
+  // will be recovered to content accordingly.
+  std::unordered_map<std::string, std::map<std::string, std::string>>
+      dir_to_new_files_since_last_sync_;
+  bool filesystem_active_;    // Record flushes, syncs, writes
+  bool filesystem_writable_;  // Bypass FaultInjectionTestFS and go directly
+                              // to underlying FS for writable files
+  IOStatus error_;
+
+  enum ErrorType : int {
+    kErrorTypeStatus = 0,
+    kErrorTypeCorruption,
+    kErrorTypeTruncated,
+    kErrorTypeMax
+  };
+
+  struct ErrorContext {
+    Random rand;
+    int one_in;
+    int count;
+    bool enable_error_injection;
+    void* callstack;
+    std::string message;
+    int frames;
+    ErrorType type;
+
+    explicit ErrorContext(uint32_t seed)
+        : rand(seed),
+          enable_error_injection(false),
+          callstack(nullptr),
+          frames(0) {}
+    ~ErrorContext() {
+      if (callstack) {
+        free(callstack);
+      }
+    }
+  };
+
+  std::unique_ptr<ThreadLocalPtr> thread_local_error_;
+  bool enable_write_error_injection_;
+  bool enable_metadata_write_error_injection_;
+  Random write_error_rand_;
+  int write_error_one_in_;
+  int metadata_write_error_one_in_;
+  std::atomic<int> read_error_one_in_;
+  bool inject_for_all_file_types_;
+  std::vector<FileType> write_error_allowed_types_;
+  // File types where direct writable is skipped.
+  std::set<FileType> skip_direct_writable_types_;
+  bool ingest_data_corruption_before_write_;
+  ChecksumType checksum_handoff_func_tpye_;
+  bool fail_get_file_unique_id_;
+
+  // Extract number of type from file name. Return false if failing to fine
+  // them.
+  bool TryParseFileName(const std::string& file_name, uint64_t* number,
+                        FileType* type);
+};
+
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/utilities/fault_injection_secondary_cache.cc b/src/rocksdb/utilities/fault_injection_secondary_cache.cc
new file mode 100644
index 000000000..2758c2a19
--- /dev/null
+++ b/src/rocksdb/utilities/fault_injection_secondary_cache.cc
@@ -0,0 +1,131 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+// This class implements a custom SecondaryCache that randomly injects an
+// error status into Inserts/Lookups based on a specified probability.
+
+#include "utilities/fault_injection_secondary_cache.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+void FaultInjectionSecondaryCache::ResultHandle::UpdateHandleValue(
+    FaultInjectionSecondaryCache::ResultHandle* handle) {
+  ErrorContext* ctx = handle->cache_->GetErrorContext();
+  if (!ctx->rand.OneIn(handle->cache_->prob_)) {
+    handle->value_ = handle->base_->Value();
+    handle->size_ = handle->base_->Size();
+  }
+  handle->base_.reset();
+}
+
+bool FaultInjectionSecondaryCache::ResultHandle::IsReady() {
+  bool ready = true;
+  if (base_) {
+    ready = base_->IsReady();
+    if (ready) {
+      UpdateHandleValue(this);
+    }
+  }
+  return ready;
+}
+
+void FaultInjectionSecondaryCache::ResultHandle::Wait() {
+  base_->Wait();
+  UpdateHandleValue(this);
+}
+
+void* FaultInjectionSecondaryCache::ResultHandle::Value() { return value_; }
+
+size_t FaultInjectionSecondaryCache::ResultHandle::Size() { return size_; }
+
+void FaultInjectionSecondaryCache::ResultHandle::WaitAll(
+    FaultInjectionSecondaryCache* cache,
+    std::vector<SecondaryCacheResultHandle*> handles) {
+  std::vector<SecondaryCacheResultHandle*> base_handles;
+  for (SecondaryCacheResultHandle* hdl : handles) {
+    FaultInjectionSecondaryCache::ResultHandle* handle =
+        static_cast<FaultInjectionSecondaryCache::ResultHandle*>(hdl);
+    if (!handle->base_) {
+      continue;
+    }
+    base_handles.emplace_back(handle->base_.get());
+  }
+
+  cache->base_->WaitAll(base_handles);
+  for (SecondaryCacheResultHandle* hdl : handles) {
+    FaultInjectionSecondaryCache::ResultHandle* handle =
+        static_cast<FaultInjectionSecondaryCache::ResultHandle*>(hdl);
+    if (handle->base_) {
+      UpdateHandleValue(handle);
+    }
+  }
+}
+
+FaultInjectionSecondaryCache::ErrorContext*
+FaultInjectionSecondaryCache::GetErrorContext() {
+  ErrorContext* ctx = static_cast<ErrorContext*>(thread_local_error_->Get());
+  if (!ctx) {
+    ctx = new ErrorContext(seed_);
+    thread_local_error_->Reset(ctx);
+  }
+
+  return ctx;
+}
+
+Status FaultInjectionSecondaryCache::Insert(
+    const Slice& key, void* value, const Cache::CacheItemHelper* helper) {
+  ErrorContext* ctx = GetErrorContext();
+  if (ctx->rand.OneIn(prob_)) {
+    return Status::IOError();
+  }
+
+  return base_->Insert(key, value, helper);
+}
+
+std::unique_ptr<SecondaryCacheResultHandle>
+FaultInjectionSecondaryCache::Lookup(const Slice& key,
+                                     const Cache::CreateCallback& create_cb,
+                                     bool wait, bool advise_erase,
+                                     bool& is_in_sec_cache) {
+  ErrorContext* ctx = GetErrorContext();
+  if (base_is_compressed_sec_cache_) {
+    if (ctx->rand.OneIn(prob_)) {
+      return nullptr;
+    } else {
+      return base_->Lookup(key, create_cb, wait, advise_erase, is_in_sec_cache);
+    }
+  } else {
+    std::unique_ptr<SecondaryCacheResultHandle> hdl =
+        base_->Lookup(key, create_cb, wait, advise_erase, is_in_sec_cache);
+    if (wait && ctx->rand.OneIn(prob_)) {
+      hdl.reset();
+    }
+    return std::unique_ptr<FaultInjectionSecondaryCache::ResultHandle>(
+        new FaultInjectionSecondaryCache::ResultHandle(this, std::move(hdl)));
+  }
+}
+
+void FaultInjectionSecondaryCache::Erase(const Slice& key) {
+  base_->Erase(key);
+}
+
+void FaultInjectionSecondaryCache::WaitAll(
+    std::vector<SecondaryCacheResultHandle*> handles) {
+  if (base_is_compressed_sec_cache_) {
+    ErrorContext* ctx = GetErrorContext();
+    std::vector<SecondaryCacheResultHandle*> base_handles;
+    for (SecondaryCacheResultHandle* hdl : handles) {
+      if (ctx->rand.OneIn(prob_)) {
+        continue;
+      }
+      base_handles.push_back(hdl);
+    }
+    base_->WaitAll(base_handles);
+  } else {
+    FaultInjectionSecondaryCache::ResultHandle::WaitAll(this, handles);
+  }
+}
+
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/utilities/fault_injection_secondary_cache.h b/src/rocksdb/utilities/fault_injection_secondary_cache.h
new file mode 100644
index 000000000..5321df626
--- /dev/null
+++ b/src/rocksdb/utilities/fault_injection_secondary_cache.h
@@ -0,0 +1,108 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#include "rocksdb/secondary_cache.h"
+#include "util/random.h"
+#include "util/thread_local.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+// This class implements a custom SecondaryCache that randomly injects an
+// error status into Inserts/Lookups based on a specified probability.
+// Its used by db_stress to verify correctness in the presence of
+// secondary cache errors.
+//
+class FaultInjectionSecondaryCache : public SecondaryCache {
+ public:
+  explicit FaultInjectionSecondaryCache(
+      const std::shared_ptr<SecondaryCache>& base, uint32_t seed, int prob)
+      : base_(base),
+        seed_(seed),
+        prob_(prob),
+        thread_local_error_(new ThreadLocalPtr(DeleteThreadLocalErrorContext)) {
+    if (std::strcmp(base_->Name(), "CompressedSecondaryCache") == 0) {
+      base_is_compressed_sec_cache_ = true;
+    }
+  }
+
+  virtual ~FaultInjectionSecondaryCache() override {}
+
+  const char* Name() const override { return "FaultInjectionSecondaryCache"; }
+
+  Status Insert(const Slice& key, void* value,
+                const Cache::CacheItemHelper* helper) override;
+
+  std::unique_ptr<SecondaryCacheResultHandle> Lookup(
+      const Slice& key, const Cache::CreateCallback& create_cb, bool wait,
+      bool advise_erase, bool& is_in_sec_cache) override;
+
+  bool SupportForceErase() const override { return base_->SupportForceErase(); }
+
+  void Erase(const Slice& key) override;
+
+  void WaitAll(std::vector<SecondaryCacheResultHandle*> handles) override;
+
+  Status SetCapacity(size_t capacity) override {
+    return base_->SetCapacity(capacity);
+  }
+
+  Status GetCapacity(size_t& capacity) override {
+    return base_->GetCapacity(capacity);
+  }
+
+  std::string GetPrintableOptions() const override {
+    return base_->GetPrintableOptions();
+  }
+
+ private:
+  class ResultHandle : public SecondaryCacheResultHandle {
+   public:
+    ResultHandle(FaultInjectionSecondaryCache* cache,
+                 std::unique_ptr<SecondaryCacheResultHandle>&& base)
+        : cache_(cache), base_(std::move(base)), value_(nullptr), size_(0) {}
+
+    ~ResultHandle() override {}
+
+    bool IsReady() override;
+
+    void Wait() override;
+
+    void* Value() override;
+
+    size_t Size() override;
+
+    static void WaitAll(FaultInjectionSecondaryCache* cache,
+                        std::vector<SecondaryCacheResultHandle*> handles);
+
+   private:
+    static void UpdateHandleValue(ResultHandle* handle);
+
+    FaultInjectionSecondaryCache* cache_;
+    std::unique_ptr<SecondaryCacheResultHandle> base_;
+    void* value_;
+    size_t size_;
+  };
+
+  static void DeleteThreadLocalErrorContext(void* p) {
+    ErrorContext* ctx = static_cast<ErrorContext*>(p);
+    delete ctx;
+  }
+
+  const std::shared_ptr<SecondaryCache> base_;
+  uint32_t seed_;
+  int prob_;
+  bool base_is_compressed_sec_cache_{false};
+
+  struct ErrorContext {
+    Random rand;
+
+    explicit ErrorContext(uint32_t seed) : rand(seed) {}
+  };
+  std::unique_ptr<ThreadLocalPtr> thread_local_error_;
+
+  ErrorContext* GetErrorContext();
+};
+
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/utilities/leveldb_options/leveldb_options.cc b/src/rocksdb/utilities/leveldb_options/leveldb_options.cc
new file mode 100644
index 000000000..125c3d956
--- /dev/null
+++ b/src/rocksdb/utilities/leveldb_options/leveldb_options.cc
@@ -0,0 +1,57 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+//
+// Copyright (c) 2011 The LevelDB Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file. See the AUTHORS file for names of contributors.
+
+#include "rocksdb/utilities/leveldb_options.h"
+
+#include "rocksdb/cache.h"
+#include "rocksdb/comparator.h"
+#include "rocksdb/env.h"
+#include "rocksdb/filter_policy.h"
+#include "rocksdb/options.h"
+#include "rocksdb/table.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+LevelDBOptions::LevelDBOptions()
+    : comparator(BytewiseComparator()),
+      create_if_missing(false),
+      error_if_exists(false),
+      paranoid_checks(false),
+      env(Env::Default()),
+      info_log(nullptr),
+      write_buffer_size(4 << 20),
+      max_open_files(1000),
+      block_cache(nullptr),
+      block_size(4096),
+      block_restart_interval(16),
+      compression(kSnappyCompression),
+      filter_policy(nullptr) {}
+
+Options ConvertOptions(const LevelDBOptions& leveldb_options) {
+  Options options = Options();
+  options.create_if_missing = leveldb_options.create_if_missing;
+  options.error_if_exists = leveldb_options.error_if_exists;
+  options.paranoid_checks = leveldb_options.paranoid_checks;
+  options.env = leveldb_options.env;
+  options.info_log.reset(leveldb_options.info_log);
+  options.write_buffer_size = leveldb_options.write_buffer_size;
+  options.max_open_files = leveldb_options.max_open_files;
+  options.compression = leveldb_options.compression;
+
+  BlockBasedTableOptions table_options;
+  table_options.block_cache.reset(leveldb_options.block_cache);
+  table_options.block_size = leveldb_options.block_size;
+  table_options.block_restart_interval = leveldb_options.block_restart_interval;
+  table_options.filter_policy.reset(leveldb_options.filter_policy);
+  options.table_factory.reset(NewBlockBasedTableFactory(table_options));
+
+  return options;
+}
+
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/utilities/memory/memory_test.cc b/src/rocksdb/utilities/memory/memory_test.cc
new file mode 100644
index 000000000..0b043af0e
--- /dev/null
+++ b/src/rocksdb/utilities/memory/memory_test.cc
@@ -0,0 +1,279 @@
+// Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#ifndef ROCKSDB_LITE
+
+#include "db/db_impl/db_impl.h"
+#include "rocksdb/cache.h"
+#include "rocksdb/table.h"
+#include "rocksdb/utilities/memory_util.h"
+#include "rocksdb/utilities/stackable_db.h"
+#include "table/block_based/block_based_table_factory.h"
+#include "test_util/testharness.h"
+#include "test_util/testutil.h"
+#include "util/random.h"
+#include "util/string_util.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+class MemoryTest : public testing::Test {
+ public:
+  MemoryTest() : kDbDir(test::PerThreadDBPath("memory_test")), rnd_(301) {
+    assert(Env::Default()->CreateDirIfMissing(kDbDir).ok());
+  }
+
+  std::string GetDBName(int id) { return kDbDir + "db_" + std::to_string(id); }
+
+  void UpdateUsagesHistory(const std::vector<DB*>& dbs) {
+    std::map<MemoryUtil::UsageType, uint64_t> usage_by_type;
+    ASSERT_OK(GetApproximateMemoryUsageByType(dbs, &usage_by_type));
+    for (int i = 0; i < MemoryUtil::kNumUsageTypes; ++i) {
+      usage_history_[i].push_back(
+          usage_by_type[static_cast<MemoryUtil::UsageType>(i)]);
+    }
+  }
+
+  void GetCachePointersFromTableFactory(
+      const TableFactory* factory,
+      std::unordered_set<const Cache*>* cache_set) {
+    const auto bbto = factory->GetOptions<BlockBasedTableOptions>();
+    if (bbto != nullptr) {
+      cache_set->insert(bbto->block_cache.get());
+      cache_set->insert(bbto->block_cache_compressed.get());
+    }
+  }
+
+  void GetCachePointers(const std::vector<DB*>& dbs,
+                        std::unordered_set<const Cache*>* cache_set) {
+    cache_set->clear();
+
+    for (auto* db : dbs) {
+      assert(db);
+
+      // Cache from DBImpl
+      StackableDB* sdb = dynamic_cast<StackableDB*>(db);
+      DBImpl* db_impl = dynamic_cast<DBImpl*>(sdb ? sdb->GetBaseDB() : db);
+      if (db_impl != nullptr) {
+        cache_set->insert(db_impl->TEST_table_cache());
+      }
+
+      // Cache from DBOptions
+      cache_set->insert(db->GetDBOptions().row_cache.get());
+
+      // Cache from table factories
+      std::unordered_map<std::string, const ImmutableCFOptions*> iopts_map;
+      if (db_impl != nullptr) {
+        ASSERT_OK(db_impl->TEST_GetAllImmutableCFOptions(&iopts_map));
+      }
+      for (auto pair : iopts_map) {
+        GetCachePointersFromTableFactory(pair.second->table_factory.get(),
+                                         cache_set);
+      }
+    }
+  }
+
+  Status GetApproximateMemoryUsageByType(
+      const std::vector<DB*>& dbs,
+      std::map<MemoryUtil::UsageType, uint64_t>* usage_by_type) {
+    std::unordered_set<const Cache*> cache_set;
+    GetCachePointers(dbs, &cache_set);
+
+    return MemoryUtil::GetApproximateMemoryUsageByType(dbs, cache_set,
+                                                       usage_by_type);
+  }
+
+  const std::string kDbDir;
+  Random rnd_;
+  std::vector<uint64_t> usage_history_[MemoryUtil::kNumUsageTypes];
+};
+
+TEST_F(MemoryTest, SharedBlockCacheTotal) {
+  std::vector<DB*> dbs;
+  std::vector<uint64_t> usage_by_type;
+  const int kNumDBs = 10;
+  const int kKeySize = 100;
+  const int kValueSize = 500;
+  Options opt;
+  opt.create_if_missing = true;
+  opt.write_buffer_size = kKeySize + kValueSize;
+  opt.max_write_buffer_number = 10;
+  opt.min_write_buffer_number_to_merge = 10;
+  opt.disable_auto_compactions = true;
+  BlockBasedTableOptions bbt_opts;
+  bbt_opts.block_cache = NewLRUCache(4096 * 1000 * 10);
+  for (int i = 0; i < kNumDBs; ++i) {
+    ASSERT_OK(DestroyDB(GetDBName(i), opt));
+    DB* db = nullptr;
+    ASSERT_OK(DB::Open(opt, GetDBName(i), &db));
+    dbs.push_back(db);
+  }
+
+  std::vector<std::string> keys_by_db[kNumDBs];
+
+  // Fill one memtable per Put to make memtable use more memory.
+  for (int p = 0; p < opt.min_write_buffer_number_to_merge / 2; ++p) {
+    for (int i = 0; i < kNumDBs; ++i) {
+      for (int j = 0; j < 100; ++j) {
+        keys_by_db[i].emplace_back(rnd_.RandomString(kKeySize));
+        ASSERT_OK(dbs[i]->Put(WriteOptions(), keys_by_db[i].back(),
+                              rnd_.RandomString(kValueSize)));
+      }
+      ASSERT_OK(dbs[i]->Flush(FlushOptions()));
+    }
+  }
+  for (int i = 0; i < kNumDBs; ++i) {
+    for (auto& key : keys_by_db[i]) {
+      std::string value;
+      ASSERT_OK(dbs[i]->Get(ReadOptions(), key, &value));
+    }
+    UpdateUsagesHistory(dbs);
+  }
+  for (size_t i = 1; i < usage_history_[MemoryUtil::kMemTableTotal].size();
+       ++i) {
+    // Expect EQ as we didn't flush more memtables.
+    ASSERT_EQ(usage_history_[MemoryUtil::kTableReadersTotal][i],
+              usage_history_[MemoryUtil::kTableReadersTotal][i - 1]);
+  }
+  for (int i = 0; i < kNumDBs; ++i) {
+    delete dbs[i];
+  }
+}
+
+TEST_F(MemoryTest, MemTableAndTableReadersTotal) {
+  std::vector<DB*> dbs;
+  std::vector<uint64_t> usage_by_type;
+  std::vector<std::vector<ColumnFamilyHandle*>> vec_handles;
+  const int kNumDBs = 10;
+  // These key/value sizes ensure each KV has its own memtable. Note that the
+  // minimum write_buffer_size allowed is 64 KB.
+  const int kKeySize = 100;
+  const int kValueSize = 1 << 16;
+  Options opt;
+  opt.create_if_missing = true;
+  opt.create_missing_column_families = true;
+  opt.write_buffer_size = kKeySize + kValueSize;
+  opt.max_write_buffer_number = 10;
+  opt.min_write_buffer_number_to_merge = 10;
+  opt.disable_auto_compactions = true;
+
+  std::vector<ColumnFamilyDescriptor> cf_descs = {
+      {kDefaultColumnFamilyName, ColumnFamilyOptions(opt)},
+      {"one", ColumnFamilyOptions(opt)},
+      {"two", ColumnFamilyOptions(opt)},
+  };
+
+  for (int i = 0; i < kNumDBs; ++i) {
+    ASSERT_OK(DestroyDB(GetDBName(i), opt));
+    std::vector<ColumnFamilyHandle*> handles;
+    dbs.emplace_back();
+    vec_handles.emplace_back();
+    ASSERT_OK(DB::Open(DBOptions(opt), GetDBName(i), cf_descs,
+                       &vec_handles.back(), &dbs.back()));
+  }
+
+  // Fill one memtable per Put to make memtable use more memory.
+  for (int p = 0; p < opt.min_write_buffer_number_to_merge / 2; ++p) {
+    for (int i = 0; i < kNumDBs; ++i) {
+      for (auto* handle : vec_handles[i]) {
+        ASSERT_OK(dbs[i]->Put(WriteOptions(), handle,
+                              rnd_.RandomString(kKeySize),
+                              rnd_.RandomString(kValueSize)));
+        UpdateUsagesHistory(dbs);
+      }
+    }
+  }
+  // Expect the usage history is monotonically increasing
+  for (size_t i = 1; i < usage_history_[MemoryUtil::kMemTableTotal].size();
+       ++i) {
+    ASSERT_GT(usage_history_[MemoryUtil::kMemTableTotal][i],
+              usage_history_[MemoryUtil::kMemTableTotal][i - 1]);
+    ASSERT_GT(usage_history_[MemoryUtil::kMemTableUnFlushed][i],
+              usage_history_[MemoryUtil::kMemTableUnFlushed][i - 1]);
+    ASSERT_EQ(usage_history_[MemoryUtil::kTableReadersTotal][i],
+              usage_history_[MemoryUtil::kTableReadersTotal][i - 1]);
+  }
+
+  size_t usage_check_point = usage_history_[MemoryUtil::kMemTableTotal].size();
+  std::vector<Iterator*> iters;
+
+  // Create an iterator and flush all memtables for each db
+  for (int i = 0; i < kNumDBs; ++i) {
+    iters.push_back(dbs[i]->NewIterator(ReadOptions()));
+    ASSERT_OK(dbs[i]->Flush(FlushOptions()));
+
+    for (int j = 0; j < 100; ++j) {
+      std::string value;
+      ASSERT_NOK(
+          dbs[i]->Get(ReadOptions(), rnd_.RandomString(kKeySize), &value));
+    }
+
+    UpdateUsagesHistory(dbs);
+  }
+  for (size_t i = usage_check_point;
+       i < usage_history_[MemoryUtil::kMemTableTotal].size(); ++i) {
+    // Since memtables are pinned by iterators, we don't expect the
+    // memory usage of all the memtables decreases as they are pinned
+    // by iterators.
+    ASSERT_GE(usage_history_[MemoryUtil::kMemTableTotal][i],
+              usage_history_[MemoryUtil::kMemTableTotal][i - 1]);
+    // Expect the usage history from the "usage_decay_point" is
+    // monotonically decreasing.
+    ASSERT_LT(usage_history_[MemoryUtil::kMemTableUnFlushed][i],
+              usage_history_[MemoryUtil::kMemTableUnFlushed][i - 1]);
+    // Expect the usage history of the table readers increases
+    // as we flush tables.
+    ASSERT_GT(usage_history_[MemoryUtil::kTableReadersTotal][i],
+              usage_history_[MemoryUtil::kTableReadersTotal][i - 1]);
+    ASSERT_GT(usage_history_[MemoryUtil::kCacheTotal][i],
+              usage_history_[MemoryUtil::kCacheTotal][i - 1]);
+  }
+  usage_check_point = usage_history_[MemoryUtil::kMemTableTotal].size();
+  for (int i = 0; i < kNumDBs; ++i) {
+    // iterator is not used.
+    ASSERT_OK(iters[i]->status());
+    delete iters[i];
+    UpdateUsagesHistory(dbs);
+  }
+  for (size_t i = usage_check_point;
+       i < usage_history_[MemoryUtil::kMemTableTotal].size(); ++i) {
+    // Expect the usage of all memtables decreasing as we delete iterators.
+    ASSERT_LT(usage_history_[MemoryUtil::kMemTableTotal][i],
+              usage_history_[MemoryUtil::kMemTableTotal][i - 1]);
+    // Since the memory usage of un-flushed memtables is only affected
+    // by Put and flush, we expect EQ here as we only delete iterators.
+    ASSERT_EQ(usage_history_[MemoryUtil::kMemTableUnFlushed][i],
+              usage_history_[MemoryUtil::kMemTableUnFlushed][i - 1]);
+    // Expect EQ as we didn't flush more memtables.
+    ASSERT_EQ(usage_history_[MemoryUtil::kTableReadersTotal][i],
+              usage_history_[MemoryUtil::kTableReadersTotal][i - 1]);
+  }
+
+  for (int i = 0; i < kNumDBs; ++i) {
+    for (auto* handle : vec_handles[i]) {
+      delete handle;
+    }
+    delete dbs[i];
+  }
+}
+}  // namespace ROCKSDB_NAMESPACE
+
+int main(int argc, char** argv) {
+#if !(defined NDEBUG) || !defined(OS_WIN)
+  ROCKSDB_NAMESPACE::port::InstallStackTraceHandler();
+  ::testing::InitGoogleTest(&argc, argv);
+  return RUN_ALL_TESTS();
+#else
+  return 0;
+#endif
+}
+
+#else
+#include <cstdio>
+
+int main(int /*argc*/, char** /*argv*/) {
+  printf("Skipped in RocksDBLite as utilities are not supported.\n");
+  return 0;
+}
+#endif  // !ROCKSDB_LITE
diff --git a/src/rocksdb/utilities/memory/memory_util.cc b/src/rocksdb/utilities/memory/memory_util.cc
new file mode 100644
index 000000000..13c81aec4
--- /dev/null
+++ b/src/rocksdb/utilities/memory/memory_util.cc
@@ -0,0 +1,52 @@
+// Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#ifndef ROCKSDB_LITE
+
+#include "rocksdb/utilities/memory_util.h"
+
+#include "db/db_impl/db_impl.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+Status MemoryUtil::GetApproximateMemoryUsageByType(
+    const std::vector<DB*>& dbs,
+    const std::unordered_set<const Cache*> cache_set,
+    std::map<MemoryUtil::UsageType, uint64_t>* usage_by_type) {
+  usage_by_type->clear();
+
+  // MemTable
+  for (auto* db : dbs) {
+    uint64_t usage = 0;
+    if (db->GetAggregatedIntProperty(DB::Properties::kSizeAllMemTables,
+                                     &usage)) {
+      (*usage_by_type)[MemoryUtil::kMemTableTotal] += usage;
+    }
+    if (db->GetAggregatedIntProperty(DB::Properties::kCurSizeAllMemTables,
+                                     &usage)) {
+      (*usage_by_type)[MemoryUtil::kMemTableUnFlushed] += usage;
+    }
+  }
+
+  // Table Readers
+  for (auto* db : dbs) {
+    uint64_t usage = 0;
+    if (db->GetAggregatedIntProperty(DB::Properties::kEstimateTableReadersMem,
+                                     &usage)) {
+      (*usage_by_type)[MemoryUtil::kTableReadersTotal] += usage;
+    }
+  }
+
+  // Cache
+  for (const auto* cache : cache_set) {
+    if (cache != nullptr) {
+      (*usage_by_type)[MemoryUtil::kCacheTotal] += cache->GetUsage();
+    }
+  }
+
+  return Status::OK();
+}
+}  // namespace ROCKSDB_NAMESPACE
+#endif  // !ROCKSDB_LITE
diff --git a/src/rocksdb/utilities/memory_allocators.h b/src/rocksdb/utilities/memory_allocators.h
new file mode 100644
index 000000000..c9e77a5b7
--- /dev/null
+++ b/src/rocksdb/utilities/memory_allocators.h
@@ -0,0 +1,104 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#pragma once
+
+#include <atomic>
+
+#include "rocksdb/memory_allocator.h"
+
+namespace ROCKSDB_NAMESPACE {
+// A memory allocator using new/delete
+class DefaultMemoryAllocator : public MemoryAllocator {
+ public:
+  static const char* kClassName() { return "DefaultMemoryAllocator"; }
+  const char* Name() const override { return kClassName(); }
+  void* Allocate(size_t size) override {
+    return static_cast<void*>(new char[size]);
+  }
+
+  void Deallocate(void* p) override { delete[] static_cast<char*>(p); }
+};
+
+// Base class for a MemoryAllocator.  This implementation does nothing
+// and implements the methods in failuse mode (assert if the methods are
+// invoked). Implementations can extend this class and override these methods
+// when they are enabled via compiler switches (e.g., the
+// JeMallocMemoryAllocator can define these methods if ROCKSDB_JEMALLOC is
+// defined at compile time.  If compiled in "disabled" mode, this class provides
+// default/failure implementations.  If compiled in "enabled" mode, the derived
+// class needs to provide the appopriate "enabled" methods for the "real"
+// implementation. Failure of the "real" implementation to implement ovreride
+// any of these methods will result in an assert failure.
+class BaseMemoryAllocator : public MemoryAllocator {
+ public:
+  void* Allocate(size_t /*size*/) override {
+    assert(false);
+    return nullptr;
+  }
+
+  void Deallocate(void* /*p*/) override { assert(false); }
+};
+
+// A Wrapped MemoryAllocator.  Delegates the memory allcator functions to the
+// wrapped one.
+class MemoryAllocatorWrapper : public MemoryAllocator {
+ public:
+  // Initialize an MemoryAllocatorWrapper that delegates all calls to *t
+  explicit MemoryAllocatorWrapper(const std::shared_ptr<MemoryAllocator>& t);
+  ~MemoryAllocatorWrapper() override {}
+
+  // Return the target to which to forward all calls
+  MemoryAllocator* target() const { return target_.get(); }
+  // Allocate a block of at least size. Has to be thread-safe.
+  void* Allocate(size_t size) override { return target_->Allocate(size); }
+
+  // Deallocate previously allocated block. Has to be thread-safe.
+  void Deallocate(void* p) override { return target_->Deallocate(p); }
+
+  // Returns the memory size of the block allocated at p. The default
+  // implementation that just returns the original allocation_size is fine.
+  size_t UsableSize(void* p, size_t allocation_size) const override {
+    return target_->UsableSize(p, allocation_size);
+  }
+
+  const Customizable* Inner() const override { return target_.get(); }
+
+ protected:
+  std::shared_ptr<MemoryAllocator> target_;
+};
+
+// A memory allocator that counts the number of allocations and deallocations
+// This class is useful if the number of memory allocations/dellocations is
+// important.
+class CountedMemoryAllocator : public MemoryAllocatorWrapper {
+ public:
+  CountedMemoryAllocator()
+      : MemoryAllocatorWrapper(std::make_shared<DefaultMemoryAllocator>()),
+        allocations_(0),
+        deallocations_(0) {}
+
+  explicit CountedMemoryAllocator(const std::shared_ptr<MemoryAllocator>& t)
+      : MemoryAllocatorWrapper(t), allocations_(0), deallocations_(0) {}
+  static const char* kClassName() { return "CountedMemoryAllocator"; }
+  const char* Name() const override { return kClassName(); }
+  std::string GetId() const override { return std::string(Name()); }
+  void* Allocate(size_t size) override {
+    allocations_++;
+    return MemoryAllocatorWrapper::Allocate(size);
+  }
+
+  void Deallocate(void* p) override {
+    deallocations_++;
+    MemoryAllocatorWrapper::Deallocate(p);
+  }
+  uint64_t GetNumAllocations() const { return allocations_; }
+  uint64_t GetNumDeallocations() const { return deallocations_; }
+
+ private:
+  std::atomic<uint64_t> allocations_;
+  std::atomic<uint64_t> deallocations_;
+};
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/utilities/merge_operators.cc b/src/rocksdb/utilities/merge_operators.cc
new file mode 100644
index 000000000..c97e9ce25
--- /dev/null
+++ b/src/rocksdb/utilities/merge_operators.cc
@@ -0,0 +1,120 @@
+// Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.
+// Copyright (c) 2011 The LevelDB Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file. See the AUTHORS file for names of contributors.
+
+#include "utilities/merge_operators.h"
+
+#include <memory>
+
+#include "rocksdb/merge_operator.h"
+#include "rocksdb/options.h"
+#include "rocksdb/utilities/customizable_util.h"
+#include "rocksdb/utilities/object_registry.h"
+#include "utilities/merge_operators/bytesxor.h"
+#include "utilities/merge_operators/sortlist.h"
+#include "utilities/merge_operators/string_append/stringappend.h"
+#include "utilities/merge_operators/string_append/stringappend2.h"
+
+namespace ROCKSDB_NAMESPACE {
+static bool LoadMergeOperator(const std::string& id,
+                              std::shared_ptr<MergeOperator>* result) {
+  bool success = true;
+  // TODO: Hook the "name" up to the actual Name() of the MergeOperators?
+  // Requires these classes be moved into a header file...
+  if (id == "put" || id == "PutOperator") {
+    *result = MergeOperators::CreatePutOperator();
+  } else if (id == "put_v1") {
+    *result = MergeOperators::CreateDeprecatedPutOperator();
+  } else if (id == "uint64add" || id == "UInt64AddOperator") {
+    *result = MergeOperators::CreateUInt64AddOperator();
+  } else if (id == "max" || id == "MaxOperator") {
+    *result = MergeOperators::CreateMaxOperator();
+#ifdef ROCKSDB_LITE
+    // The remainder of the classes are handled by the ObjectRegistry in
+    // non-LITE mode
+  } else if (id == StringAppendOperator::kNickName() ||
+             id == StringAppendOperator::kClassName()) {
+    *result = MergeOperators::CreateStringAppendOperator();
+  } else if (id == StringAppendTESTOperator::kNickName() ||
+             id == StringAppendTESTOperator::kClassName()) {
+    *result = MergeOperators::CreateStringAppendTESTOperator();
+  } else if (id == BytesXOROperator::kNickName() ||
+             id == BytesXOROperator::kClassName()) {
+    *result = MergeOperators::CreateBytesXOROperator();
+  } else if (id == SortList::kNickName() || id == SortList::kClassName()) {
+    *result = MergeOperators::CreateSortOperator();
+#endif  // ROCKSDB_LITE
+  } else {
+    success = false;
+  }
+  return success;
+}
+
+#ifndef ROCKSDB_LITE
+static int RegisterBuiltinMergeOperators(ObjectLibrary& library,
+                                         const std::string& /*arg*/) {
+  size_t num_types;
+  library.AddFactory<MergeOperator>(
+      ObjectLibrary::PatternEntry(StringAppendOperator::kClassName())
+          .AnotherName(StringAppendOperator::kNickName()),
+      [](const std::string& /*uri*/, std::unique_ptr<MergeOperator>* guard,
+         std::string* /*errmsg*/) {
+        guard->reset(new StringAppendOperator(","));
+        return guard->get();
+      });
+  library.AddFactory<MergeOperator>(
+      ObjectLibrary::PatternEntry(StringAppendTESTOperator::kClassName())
+          .AnotherName(StringAppendTESTOperator::kNickName()),
+      [](const std::string& /*uri*/, std::unique_ptr<MergeOperator>* guard,
+         std::string* /*errmsg*/) {
+        guard->reset(new StringAppendTESTOperator(","));
+        return guard->get();
+      });
+  library.AddFactory<MergeOperator>(
+      ObjectLibrary::PatternEntry(SortList::kClassName())
+          .AnotherName(SortList::kNickName()),
+      [](const std::string& /*uri*/, std::unique_ptr<MergeOperator>* guard,
+         std::string* /*errmsg*/) {
+        guard->reset(new SortList());
+        return guard->get();
+      });
+  library.AddFactory<MergeOperator>(
+      ObjectLibrary::PatternEntry(BytesXOROperator::kClassName())
+          .AnotherName(BytesXOROperator::kNickName()),
+      [](const std::string& /*uri*/, std::unique_ptr<MergeOperator>* guard,
+         std::string* /*errmsg*/) {
+        guard->reset(new BytesXOROperator());
+        return guard->get();
+      });
+
+  return static_cast<int>(library.GetFactoryCount(&num_types));
+}
+#endif  // ROCKSDB_LITE
+
+Status MergeOperator::CreateFromString(const ConfigOptions& config_options,
+                                       const std::string& value,
+                                       std::shared_ptr<MergeOperator>* result) {
+#ifndef ROCKSDB_LITE
+  static std::once_flag once;
+  std::call_once(once, [&]() {
+    RegisterBuiltinMergeOperators(*(ObjectLibrary::Default().get()), "");
+  });
+#endif  // ROCKSDB_LITE
+  return LoadSharedObject<MergeOperator>(config_options, value,
+                                         LoadMergeOperator, result);
+}
+
+std::shared_ptr<MergeOperator> MergeOperators::CreateFromStringId(
+    const std::string& id) {
+  std::shared_ptr<MergeOperator> result;
+  Status s = MergeOperator::CreateFromString(ConfigOptions(), id, &result);
+  if (s.ok()) {
+    return result;
+  } else {
+    // Empty or unknown, just return nullptr
+    return nullptr;
+  }
+}
+
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/utilities/merge_operators.h b/src/rocksdb/utilities/merge_operators.h
new file mode 100644
index 000000000..9b90107e3
--- /dev/null
+++ b/src/rocksdb/utilities/merge_operators.h
@@ -0,0 +1,36 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+//
+#pragma once
+#include <stdio.h>
+
+#include <memory>
+#include <string>
+
+#include "rocksdb/merge_operator.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+class MergeOperators {
+ public:
+  static std::shared_ptr<MergeOperator> CreatePutOperator();
+  static std::shared_ptr<MergeOperator> CreateDeprecatedPutOperator();
+  static std::shared_ptr<MergeOperator> CreateUInt64AddOperator();
+  static std::shared_ptr<MergeOperator> CreateStringAppendOperator();
+  static std::shared_ptr<MergeOperator> CreateStringAppendOperator(
+      char delim_char);
+  static std::shared_ptr<MergeOperator> CreateStringAppendOperator(
+      const std::string& delim);
+  static std::shared_ptr<MergeOperator> CreateStringAppendTESTOperator();
+  static std::shared_ptr<MergeOperator> CreateMaxOperator();
+  static std::shared_ptr<MergeOperator> CreateBytesXOROperator();
+  static std::shared_ptr<MergeOperator> CreateSortOperator();
+
+  // Will return a different merge operator depending on the string.
+  static std::shared_ptr<MergeOperator> CreateFromStringId(
+      const std::string& name);
+};
+
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/utilities/merge_operators/bytesxor.cc b/src/rocksdb/utilities/merge_operators/bytesxor.cc
new file mode 100644
index 000000000..fa09c18ea
--- /dev/null
+++ b/src/rocksdb/utilities/merge_operators/bytesxor.cc
@@ -0,0 +1,57 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#include "utilities/merge_operators/bytesxor.h"
+
+#include <algorithm>
+#include <string>
+
+namespace ROCKSDB_NAMESPACE {
+
+std::shared_ptr<MergeOperator> MergeOperators::CreateBytesXOROperator() {
+  return std::make_shared<BytesXOROperator>();
+}
+
+bool BytesXOROperator::Merge(const Slice& /*key*/, const Slice* existing_value,
+                             const Slice& value, std::string* new_value,
+                             Logger* /*logger*/) const {
+  XOR(existing_value, value, new_value);
+  return true;
+}
+
+void BytesXOROperator::XOR(const Slice* existing_value, const Slice& value,
+                           std::string* new_value) const {
+  if (!existing_value) {
+    new_value->clear();
+    new_value->assign(value.data(), value.size());
+    return;
+  }
+
+  size_t min_size = std::min(existing_value->size(), value.size());
+  size_t max_size = std::max(existing_value->size(), value.size());
+
+  new_value->clear();
+  new_value->reserve(max_size);
+
+  const char* existing_value_data = existing_value->data();
+  const char* value_data = value.data();
+
+  for (size_t i = 0; i < min_size; i++) {
+    new_value->push_back(existing_value_data[i] ^ value_data[i]);
+  }
+
+  if (existing_value->size() == max_size) {
+    for (size_t i = min_size; i < max_size; i++) {
+      new_value->push_back(existing_value_data[i]);
+    }
+  } else {
+    assert(value.size() == max_size);
+    for (size_t i = min_size; i < max_size; i++) {
+      new_value->push_back(value_data[i]);
+    }
+  }
+}
+
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/utilities/merge_operators/bytesxor.h b/src/rocksdb/utilities/merge_operators/bytesxor.h
new file mode 100644
index 000000000..3c7baacce
--- /dev/null
+++ b/src/rocksdb/utilities/merge_operators/bytesxor.h
@@ -0,0 +1,40 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#pragma once
+
+#include <algorithm>
+#include <memory>
+#include <string>
+
+#include "rocksdb/env.h"
+#include "rocksdb/merge_operator.h"
+#include "rocksdb/slice.h"
+#include "util/coding.h"
+#include "utilities/merge_operators.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+// A 'model' merge operator that XORs two (same sized) array of bytes.
+// Implemented as an AssociativeMergeOperator for simplicity and example.
+class BytesXOROperator : public AssociativeMergeOperator {
+ public:
+  // XORs the two array of bytes one byte at a time and stores the result
+  // in new_value. len is the number of xored bytes, and the length of new_value
+  virtual bool Merge(const Slice& key, const Slice* existing_value,
+                     const Slice& value, std::string* new_value,
+                     Logger* logger) const override;
+
+  static const char* kClassName() { return "BytesXOR"; }
+  static const char* kNickName() { return "bytesxor"; }
+
+  const char* NickName() const override { return kNickName(); }
+  const char* Name() const override { return kClassName(); }
+
+  void XOR(const Slice* existing_value, const Slice& value,
+           std::string* new_value) const;
+};
+
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/utilities/merge_operators/max.cc b/src/rocksdb/utilities/merge_operators/max.cc
new file mode 100644
index 000000000..de4abfa6f
--- /dev/null
+++ b/src/rocksdb/utilities/merge_operators/max.cc
@@ -0,0 +1,80 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#include <memory>
+
+#include "rocksdb/merge_operator.h"
+#include "rocksdb/slice.h"
+#include "utilities/merge_operators.h"
+
+using ROCKSDB_NAMESPACE::Logger;
+using ROCKSDB_NAMESPACE::MergeOperator;
+using ROCKSDB_NAMESPACE::Slice;
+
+namespace {  // anonymous namespace
+
+// Merge operator that picks the maximum operand, Comparison is based on
+// Slice::compare
+class MaxOperator : public MergeOperator {
+ public:
+  bool FullMergeV2(const MergeOperationInput& merge_in,
+                   MergeOperationOutput* merge_out) const override {
+    Slice& max = merge_out->existing_operand;
+    if (merge_in.existing_value) {
+      max = Slice(merge_in.existing_value->data(),
+                  merge_in.existing_value->size());
+    } else if (max.data() == nullptr) {
+      max = Slice();
+    }
+
+    for (const auto& op : merge_in.operand_list) {
+      if (max.compare(op) < 0) {
+        max = op;
+      }
+    }
+
+    return true;
+  }
+
+  bool PartialMerge(const Slice& /*key*/, const Slice& left_operand,
+                    const Slice& right_operand, std::string* new_value,
+                    Logger* /*logger*/) const override {
+    if (left_operand.compare(right_operand) >= 0) {
+      new_value->assign(left_operand.data(), left_operand.size());
+    } else {
+      new_value->assign(right_operand.data(), right_operand.size());
+    }
+    return true;
+  }
+
+  bool PartialMergeMulti(const Slice& /*key*/,
+                         const std::deque<Slice>& operand_list,
+                         std::string* new_value,
+                         Logger* /*logger*/) const override {
+    Slice max;
+    for (const auto& operand : operand_list) {
+      if (max.compare(operand) < 0) {
+        max = operand;
+      }
+    }
+
+    new_value->assign(max.data(), max.size());
+    return true;
+  }
+
+  static const char* kClassName() { return "MaxOperator"; }
+  static const char* kNickName() { return "max"; }
+  const char* Name() const override { return kClassName(); }
+  const char* NickName() const override { return kNickName(); }
+};
+
+}  // end of anonymous namespace
+
+namespace ROCKSDB_NAMESPACE {
+
+std::shared_ptr<MergeOperator> MergeOperators::CreateMaxOperator() {
+  return std::make_shared<MaxOperator>();
+}
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/utilities/merge_operators/put.cc b/src/rocksdb/utilities/merge_operators/put.cc
new file mode 100644
index 000000000..ccf9ff21f
--- /dev/null
+++ b/src/rocksdb/utilities/merge_operators/put.cc
@@ -0,0 +1,92 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#include <memory>
+
+#include "rocksdb/merge_operator.h"
+#include "rocksdb/slice.h"
+#include "utilities/merge_operators.h"
+
+namespace {  // anonymous namespace
+
+using ROCKSDB_NAMESPACE::Logger;
+using ROCKSDB_NAMESPACE::MergeOperator;
+using ROCKSDB_NAMESPACE::Slice;
+
+// A merge operator that mimics Put semantics
+// Since this merge-operator will not be used in production,
+// it is implemented as a non-associative merge operator to illustrate the
+// new interface and for testing purposes. (That is, we inherit from
+// the MergeOperator class rather than the AssociativeMergeOperator
+// which would be simpler in this case).
+//
+// From the client-perspective, semantics are the same.
+class PutOperator : public MergeOperator {
+ public:
+  bool FullMerge(const Slice& /*key*/, const Slice* /*existing_value*/,
+                 const std::deque<std::string>& operand_sequence,
+                 std::string* new_value, Logger* /*logger*/) const override {
+    // Put basically only looks at the current/latest value
+    assert(!operand_sequence.empty());
+    assert(new_value != nullptr);
+    new_value->assign(operand_sequence.back());
+    return true;
+  }
+
+  bool PartialMerge(const Slice& /*key*/, const Slice& /*left_operand*/,
+                    const Slice& right_operand, std::string* new_value,
+                    Logger* /*logger*/) const override {
+    new_value->assign(right_operand.data(), right_operand.size());
+    return true;
+  }
+
+  using MergeOperator::PartialMergeMulti;
+  bool PartialMergeMulti(const Slice& /*key*/,
+                         const std::deque<Slice>& operand_list,
+                         std::string* new_value,
+                         Logger* /*logger*/) const override {
+    new_value->assign(operand_list.back().data(), operand_list.back().size());
+    return true;
+  }
+
+  static const char* kClassName() { return "PutOperator"; }
+  static const char* kNickName() { return "put_v1"; }
+  const char* Name() const override { return kClassName(); }
+  const char* NickName() const override { return kNickName(); }
+};
+
+class PutOperatorV2 : public PutOperator {
+  bool FullMerge(const Slice& /*key*/, const Slice* /*existing_value*/,
+                 const std::deque<std::string>& /*operand_sequence*/,
+                 std::string* /*new_value*/,
+                 Logger* /*logger*/) const override {
+    assert(false);
+    return false;
+  }
+
+  bool FullMergeV2(const MergeOperationInput& merge_in,
+                   MergeOperationOutput* merge_out) const override {
+    // Put basically only looks at the current/latest value
+    assert(!merge_in.operand_list.empty());
+    merge_out->existing_operand = merge_in.operand_list.back();
+    return true;
+  }
+
+  static const char* kNickName() { return "put"; }
+  const char* NickName() const override { return kNickName(); }
+};
+
+}  // end of anonymous namespace
+
+namespace ROCKSDB_NAMESPACE {
+
+std::shared_ptr<MergeOperator> MergeOperators::CreateDeprecatedPutOperator() {
+  return std::make_shared<PutOperator>();
+}
+
+std::shared_ptr<MergeOperator> MergeOperators::CreatePutOperator() {
+  return std::make_shared<PutOperatorV2>();
+}
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/utilities/merge_operators/sortlist.cc b/src/rocksdb/utilities/merge_operators/sortlist.cc
new file mode 100644
index 000000000..67bfc7e5e
--- /dev/null
+++ b/src/rocksdb/utilities/merge_operators/sortlist.cc
@@ -0,0 +1,95 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+#include "utilities/merge_operators/sortlist.h"
+
+#include "rocksdb/merge_operator.h"
+#include "rocksdb/slice.h"
+#include "utilities/merge_operators.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+bool SortList::FullMergeV2(const MergeOperationInput& merge_in,
+                           MergeOperationOutput* merge_out) const {
+  std::vector<int> left;
+  for (Slice slice : merge_in.operand_list) {
+    std::vector<int> right;
+    MakeVector(right, slice);
+    left = Merge(left, right);
+  }
+  for (int i = 0; i < static_cast<int>(left.size()) - 1; i++) {
+    merge_out->new_value.append(std::to_string(left[i])).append(",");
+  }
+  merge_out->new_value.append(std::to_string(left.back()));
+  return true;
+}
+
+bool SortList::PartialMerge(const Slice& /*key*/, const Slice& left_operand,
+                            const Slice& right_operand, std::string* new_value,
+                            Logger* /*logger*/) const {
+  std::vector<int> left;
+  std::vector<int> right;
+  MakeVector(left, left_operand);
+  MakeVector(right, right_operand);
+  left = Merge(left, right);
+  for (int i = 0; i < static_cast<int>(left.size()) - 1; i++) {
+    new_value->append(std::to_string(left[i])).append(",");
+  }
+  new_value->append(std::to_string(left.back()));
+  return true;
+}
+
+bool SortList::PartialMergeMulti(const Slice& /*key*/,
+                                 const std::deque<Slice>& operand_list,
+                                 std::string* new_value,
+                                 Logger* /*logger*/) const {
+  (void)operand_list;
+  (void)new_value;
+  return true;
+}
+
+void SortList::MakeVector(std::vector<int>& operand, Slice slice) const {
+  do {
+    const char* begin = slice.data_;
+    while (*slice.data_ != ',' && *slice.data_) slice.data_++;
+    operand.push_back(std::stoi(std::string(begin, slice.data_)));
+  } while (0 != *slice.data_++);
+}
+
+std::vector<int> SortList::Merge(std::vector<int>& left,
+                                 std::vector<int>& right) const {
+  // Fill the resultant vector with sorted results from both vectors
+  std::vector<int> result;
+  unsigned left_it = 0, right_it = 0;
+
+  while (left_it < left.size() && right_it < right.size()) {
+    // If the left value is smaller than the right it goes next
+    // into the resultant vector
+    if (left[left_it] < right[right_it]) {
+      result.push_back(left[left_it]);
+      left_it++;
+    } else {
+      result.push_back(right[right_it]);
+      right_it++;
+    }
+  }
+
+  // Push the remaining data from both vectors onto the resultant
+  while (left_it < left.size()) {
+    result.push_back(left[left_it]);
+    left_it++;
+  }
+
+  while (right_it < right.size()) {
+    result.push_back(right[right_it]);
+    right_it++;
+  }
+
+  return result;
+}
+
+std::shared_ptr<MergeOperator> MergeOperators::CreateSortOperator() {
+  return std::make_shared<SortList>();
+}
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/utilities/merge_operators/sortlist.h b/src/rocksdb/utilities/merge_operators/sortlist.h
new file mode 100644
index 000000000..eaa4e76fb
--- /dev/null
+++ b/src/rocksdb/utilities/merge_operators/sortlist.h
@@ -0,0 +1,42 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+// A MergeOperator for RocksDB that implements Merge Sort.
+// It is built using the MergeOperator interface. The operator works by taking
+// an input which contains one or more merge operands where each operand is a
+// list of sorted ints and merges them to form a large sorted list.
+#pragma once
+
+#include "rocksdb/merge_operator.h"
+#include "rocksdb/slice.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+class SortList : public MergeOperator {
+ public:
+  bool FullMergeV2(const MergeOperationInput& merge_in,
+                   MergeOperationOutput* merge_out) const override;
+
+  bool PartialMerge(const Slice& /*key*/, const Slice& left_operand,
+                    const Slice& right_operand, std::string* new_value,
+                    Logger* /*logger*/) const override;
+
+  bool PartialMergeMulti(const Slice& key,
+                         const std::deque<Slice>& operand_list,
+                         std::string* new_value, Logger* logger) const override;
+
+  static const char* kClassName() { return "MergeSortOperator"; }
+  static const char* kNickName() { return "sortlist"; }
+
+  const char* Name() const override { return kClassName(); }
+  const char* NickName() const override { return kNickName(); }
+
+  void MakeVector(std::vector<int>& operand, Slice slice) const;
+
+ private:
+  std::vector<int> Merge(std::vector<int>& left, std::vector<int>& right) const;
+};
+
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/utilities/merge_operators/string_append/stringappend.cc b/src/rocksdb/utilities/merge_operators/string_append/stringappend.cc
new file mode 100644
index 000000000..5092cabcb
--- /dev/null
+++ b/src/rocksdb/utilities/merge_operators/string_append/stringappend.cc
@@ -0,0 +1,78 @@
+/**
+ * A MergeOperator for rocksdb that implements string append.
+ * @author Deon Nicholas (dnicholas@fb.com)
+ * Copyright 2013 Facebook
+ */
+
+#include "stringappend.h"
+
+#include <assert.h>
+
+#include <memory>
+
+#include "rocksdb/merge_operator.h"
+#include "rocksdb/slice.h"
+#include "rocksdb/utilities/options_type.h"
+#include "utilities/merge_operators.h"
+
+namespace ROCKSDB_NAMESPACE {
+namespace {
+static std::unordered_map<std::string, OptionTypeInfo>
+    stringappend_merge_type_info = {
+#ifndef ROCKSDB_LITE
+        {"delimiter",
+         {0, OptionType::kString, OptionVerificationType::kNormal,
+          OptionTypeFlags::kNone}},
+#endif  // ROCKSDB_LITE
+};
+}  // namespace
+// Constructor: also specify the delimiter character.
+StringAppendOperator::StringAppendOperator(char delim_char)
+    : delim_(1, delim_char) {
+  RegisterOptions("Delimiter", &delim_, &stringappend_merge_type_info);
+}
+
+StringAppendOperator::StringAppendOperator(const std::string& delim)
+    : delim_(delim) {
+  RegisterOptions("Delimiter", &delim_, &stringappend_merge_type_info);
+}
+
+// Implementation for the merge operation (concatenates two strings)
+bool StringAppendOperator::Merge(const Slice& /*key*/,
+                                 const Slice* existing_value,
+                                 const Slice& value, std::string* new_value,
+                                 Logger* /*logger*/) const {
+  // Clear the *new_value for writing.
+  assert(new_value);
+  new_value->clear();
+
+  if (!existing_value) {
+    // No existing_value. Set *new_value = value
+    new_value->assign(value.data(), value.size());
+  } else {
+    // Generic append (existing_value != null).
+    // Reserve *new_value to correct size, and apply concatenation.
+    new_value->reserve(existing_value->size() + delim_.size() + value.size());
+    new_value->assign(existing_value->data(), existing_value->size());
+    new_value->append(delim_);
+    new_value->append(value.data(), value.size());
+  }
+
+  return true;
+}
+
+std::shared_ptr<MergeOperator> MergeOperators::CreateStringAppendOperator() {
+  return std::make_shared<StringAppendOperator>(',');
+}
+
+std::shared_ptr<MergeOperator> MergeOperators::CreateStringAppendOperator(
+    char delim_char) {
+  return std::make_shared<StringAppendOperator>(delim_char);
+}
+
+std::shared_ptr<MergeOperator> MergeOperators::CreateStringAppendOperator(
+    const std::string& delim) {
+  return std::make_shared<StringAppendOperator>(delim);
+}
+
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/utilities/merge_operators/string_append/stringappend.h b/src/rocksdb/utilities/merge_operators/string_append/stringappend.h
new file mode 100644
index 000000000..153532382
--- /dev/null
+++ b/src/rocksdb/utilities/merge_operators/string_append/stringappend.h
@@ -0,0 +1,32 @@
+/**
+ * A MergeOperator for rocksdb that implements string append.
+ * @author Deon Nicholas (dnicholas@fb.com)
+ * Copyright 2013 Facebook
+ */
+
+#pragma once
+#include "rocksdb/merge_operator.h"
+#include "rocksdb/slice.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+class StringAppendOperator : public AssociativeMergeOperator {
+ public:
+  // Constructor: specify delimiter
+  explicit StringAppendOperator(char delim_char);
+  explicit StringAppendOperator(const std::string& delim);
+
+  virtual bool Merge(const Slice& key, const Slice* existing_value,
+                     const Slice& value, std::string* new_value,
+                     Logger* logger) const override;
+
+  static const char* kClassName() { return "StringAppendOperator"; }
+  static const char* kNickName() { return "stringappend"; }
+  virtual const char* Name() const override { return kClassName(); }
+  virtual const char* NickName() const override { return kNickName(); }
+
+ private:
+  std::string delim_;  // The delimiter is inserted between elements
+};
+
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/utilities/merge_operators/string_append/stringappend2.cc b/src/rocksdb/utilities/merge_operators/string_append/stringappend2.cc
new file mode 100644
index 000000000..36cb9ee34
--- /dev/null
+++ b/src/rocksdb/utilities/merge_operators/string_append/stringappend2.cc
@@ -0,0 +1,132 @@
+/**
+ * @author Deon Nicholas (dnicholas@fb.com)
+ * Copyright 2013 Facebook
+ */
+
+#include "stringappend2.h"
+
+#include <assert.h>
+
+#include <memory>
+#include <string>
+
+#include "rocksdb/merge_operator.h"
+#include "rocksdb/slice.h"
+#include "rocksdb/utilities/options_type.h"
+#include "utilities/merge_operators.h"
+
+namespace ROCKSDB_NAMESPACE {
+namespace {
+static std::unordered_map<std::string, OptionTypeInfo>
+    stringappend2_merge_type_info = {
+#ifndef ROCKSDB_LITE
+        {"delimiter",
+         {0, OptionType::kString, OptionVerificationType::kNormal,
+          OptionTypeFlags::kNone}},
+#endif  // ROCKSDB_LITE
+};
+}  // namespace
+
+// Constructor: also specify the delimiter character.
+StringAppendTESTOperator::StringAppendTESTOperator(char delim_char)
+    : delim_(1, delim_char) {
+  RegisterOptions("Delimiter", &delim_, &stringappend2_merge_type_info);
+}
+
+StringAppendTESTOperator::StringAppendTESTOperator(const std::string& delim)
+    : delim_(delim) {
+  RegisterOptions("Delimiter", &delim_, &stringappend2_merge_type_info);
+}
+
+// Implementation for the merge operation (concatenates two strings)
+bool StringAppendTESTOperator::FullMergeV2(
+    const MergeOperationInput& merge_in,
+    MergeOperationOutput* merge_out) const {
+  // Clear the *new_value for writing.
+  merge_out->new_value.clear();
+
+  if (merge_in.existing_value == nullptr && merge_in.operand_list.size() == 1) {
+    // Only one operand
+    merge_out->existing_operand = merge_in.operand_list.back();
+    return true;
+  }
+
+  // Compute the space needed for the final result.
+  size_t numBytes = 0;
+
+  for (auto it = merge_in.operand_list.begin();
+       it != merge_in.operand_list.end(); ++it) {
+    numBytes += it->size() + delim_.size();
+  }
+
+  // Only print the delimiter after the first entry has been printed
+  bool printDelim = false;
+
+  // Prepend the *existing_value if one exists.
+  if (merge_in.existing_value) {
+    merge_out->new_value.reserve(numBytes + merge_in.existing_value->size());
+    merge_out->new_value.append(merge_in.existing_value->data(),
+                                merge_in.existing_value->size());
+    printDelim = true;
+  } else if (numBytes) {
+    // Without the existing (initial) value, the delimiter before the first of
+    // subsequent operands becomes redundant.
+    merge_out->new_value.reserve(numBytes - delim_.size());
+  }
+
+  // Concatenate the sequence of strings (and add a delimiter between each)
+  for (auto it = merge_in.operand_list.begin();
+       it != merge_in.operand_list.end(); ++it) {
+    if (printDelim) {
+      merge_out->new_value.append(delim_);
+    }
+    merge_out->new_value.append(it->data(), it->size());
+    printDelim = true;
+  }
+
+  return true;
+}
+
+bool StringAppendTESTOperator::PartialMergeMulti(
+    const Slice& /*key*/, const std::deque<Slice>& /*operand_list*/,
+    std::string* /*new_value*/, Logger* /*logger*/) const {
+  return false;
+}
+
+// A version of PartialMerge that actually performs "partial merging".
+// Use this to simulate the exact behaviour of the StringAppendOperator.
+bool StringAppendTESTOperator::_AssocPartialMergeMulti(
+    const Slice& /*key*/, const std::deque<Slice>& operand_list,
+    std::string* new_value, Logger* /*logger*/) const {
+  // Clear the *new_value for writing
+  assert(new_value);
+  new_value->clear();
+  assert(operand_list.size() >= 2);
+
+  // Generic append
+  // Determine and reserve correct size for *new_value.
+  size_t size = 0;
+  for (const auto& operand : operand_list) {
+    size += operand.size();
+  }
+  size += (operand_list.size() - 1) * delim_.length();  // Delimiters
+  new_value->reserve(size);
+
+  // Apply concatenation
+  new_value->assign(operand_list.front().data(), operand_list.front().size());
+
+  for (std::deque<Slice>::const_iterator it = operand_list.begin() + 1;
+       it != operand_list.end(); ++it) {
+    new_value->append(delim_);
+    new_value->append(it->data(), it->size());
+  }
+
+  return true;
+}
+
+std::shared_ptr<MergeOperator>
+MergeOperators::CreateStringAppendTESTOperator() {
+  return std::make_shared<StringAppendTESTOperator>(',');
+}
+
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/utilities/merge_operators/string_append/stringappend2.h b/src/rocksdb/utilities/merge_operators/string_append/stringappend2.h
new file mode 100644
index 000000000..75389e4ae
--- /dev/null
+++ b/src/rocksdb/utilities/merge_operators/string_append/stringappend2.h
@@ -0,0 +1,52 @@
+// Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.
+/**
+ * A TEST MergeOperator for rocksdb that implements string append.
+ * It is built using the MergeOperator interface rather than the simpler
+ * AssociativeMergeOperator interface. This is useful for testing/benchmarking.
+ * While the two operators are semantically the same, all production code
+ * should use the StringAppendOperator defined in stringappend.{h,cc}. The
+ * operator defined in the present file is primarily for testing.
+ *
+ * @author Deon Nicholas (dnicholas@fb.com)
+ * Copyright 2013 Facebook
+ */
+
+#pragma once
+#include <deque>
+#include <string>
+
+#include "rocksdb/merge_operator.h"
+#include "rocksdb/slice.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+class StringAppendTESTOperator : public MergeOperator {
+ public:
+  // Constructor with delimiter
+  explicit StringAppendTESTOperator(char delim_char);
+  explicit StringAppendTESTOperator(const std::string& delim);
+
+  virtual bool FullMergeV2(const MergeOperationInput& merge_in,
+                           MergeOperationOutput* merge_out) const override;
+
+  virtual bool PartialMergeMulti(const Slice& key,
+                                 const std::deque<Slice>& operand_list,
+                                 std::string* new_value,
+                                 Logger* logger) const override;
+
+  static const char* kClassName() { return "StringAppendTESTOperator"; }
+  static const char* kNickName() { return "stringappendtest"; }
+  const char* Name() const override { return kClassName(); }
+  const char* NickName() const override { return kNickName(); }
+
+ private:
+  // A version of PartialMerge that actually performs "partial merging".
+  // Use this to simulate the exact behaviour of the StringAppendOperator.
+  bool _AssocPartialMergeMulti(const Slice& key,
+                               const std::deque<Slice>& operand_list,
+                               std::string* new_value, Logger* logger) const;
+
+  std::string delim_;  // The delimiter is inserted between elements
+};
+
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/utilities/merge_operators/string_append/stringappend_test.cc b/src/rocksdb/utilities/merge_operators/string_append/stringappend_test.cc
new file mode 100644
index 000000000..22b6144af
--- /dev/null
+++ b/src/rocksdb/utilities/merge_operators/string_append/stringappend_test.cc
@@ -0,0 +1,640 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+//
+
+/**
+ * An persistent map : key -> (list of strings), using rocksdb merge.
+ * This file is a test-harness / use-case for the StringAppendOperator.
+ *
+ * @author Deon Nicholas (dnicholas@fb.com)
+ * Copyright 2013 Facebook, Inc.
+ */
+
+#include "utilities/merge_operators/string_append/stringappend.h"
+
+#include <iostream>
+#include <map>
+#include <tuple>
+
+#include "port/stack_trace.h"
+#include "rocksdb/db.h"
+#include "rocksdb/merge_operator.h"
+#include "rocksdb/utilities/db_ttl.h"
+#include "test_util/testharness.h"
+#include "util/random.h"
+#include "utilities/merge_operators.h"
+#include "utilities/merge_operators/string_append/stringappend2.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+// Path to the database on file system
+const std::string kDbName = test::PerThreadDBPath("stringappend_test");
+
+namespace {
+// OpenDb opens a (possibly new) rocksdb database with a StringAppendOperator
+std::shared_ptr<DB> OpenNormalDb(const std::string& delim) {
+  DB* db;
+  Options options;
+  options.create_if_missing = true;
+  MergeOperator* mergeOperator;
+  if (delim.size() == 1) {
+    mergeOperator = new StringAppendOperator(delim[0]);
+  } else {
+    mergeOperator = new StringAppendOperator(delim);
+  }
+  options.merge_operator.reset(mergeOperator);
+  EXPECT_OK(DB::Open(options, kDbName, &db));
+  return std::shared_ptr<DB>(db);
+}
+
+#ifndef ROCKSDB_LITE  // TtlDb is not supported in Lite
+// Open a TtlDB with a non-associative StringAppendTESTOperator
+std::shared_ptr<DB> OpenTtlDb(const std::string& delim) {
+  DBWithTTL* db;
+  Options options;
+  options.create_if_missing = true;
+  MergeOperator* mergeOperator;
+  if (delim.size() == 1) {
+    mergeOperator = new StringAppendTESTOperator(delim[0]);
+  } else {
+    mergeOperator = new StringAppendTESTOperator(delim);
+  }
+  options.merge_operator.reset(mergeOperator);
+  EXPECT_OK(DBWithTTL::Open(options, kDbName, &db, 123456));
+  return std::shared_ptr<DB>(db);
+}
+#endif  // !ROCKSDB_LITE
+}  // namespace
+
+/// StringLists represents a set of string-lists, each with a key-index.
+/// Supports Append(list, string) and Get(list)
+class StringLists {
+ public:
+  // Constructor: specifies the rocksdb db
+  /* implicit */
+  StringLists(std::shared_ptr<DB> db)
+      : db_(db), merge_option_(), get_option_() {
+    assert(db);
+  }
+
+  // Append string val onto the list defined by key; return true on success
+  bool Append(const std::string& key, const std::string& val) {
+    Slice valSlice(val.data(), val.size());
+    auto s = db_->Merge(merge_option_, key, valSlice);
+
+    if (s.ok()) {
+      return true;
+    } else {
+      std::cerr << "ERROR " << s.ToString() << std::endl;
+      return false;
+    }
+  }
+
+  // Returns the list of strings associated with key (or "" if does not exist)
+  bool Get(const std::string& key, std::string* const result) {
+    assert(result != nullptr);  // we should have a place to store the result
+    auto s = db_->Get(get_option_, key, result);
+
+    if (s.ok()) {
+      return true;
+    }
+
+    // Either key does not exist, or there is some error.
+    *result = "";  // Always return empty string (just for convention)
+
+    // NotFound is okay; just return empty (similar to std::map)
+    // But network or db errors, etc, should fail the test (or at least yell)
+    if (!s.IsNotFound()) {
+      std::cerr << "ERROR " << s.ToString() << std::endl;
+    }
+
+    // Always return false if s.ok() was not true
+    return false;
+  }
+
+ private:
+  std::shared_ptr<DB> db_;
+  WriteOptions merge_option_;
+  ReadOptions get_option_;
+};
+
+// The class for unit-testing
+class StringAppendOperatorTest : public testing::Test,
+                                 public ::testing::WithParamInterface<bool> {
+ public:
+  StringAppendOperatorTest() {
+    EXPECT_OK(
+        DestroyDB(kDbName, Options()));  // Start each test with a fresh DB
+  }
+
+  void SetUp() override {
+#ifndef ROCKSDB_LITE  // TtlDb is not supported in Lite
+    bool if_use_ttl = GetParam();
+    if (if_use_ttl) {
+      fprintf(stderr, "Running tests with ttl db and generic operator.\n");
+      StringAppendOperatorTest::SetOpenDbFunction(&OpenTtlDb);
+      return;
+    }
+#endif  // !ROCKSDB_LITE
+    fprintf(stderr, "Running tests with regular db and operator.\n");
+    StringAppendOperatorTest::SetOpenDbFunction(&OpenNormalDb);
+  }
+
+  using OpenFuncPtr = std::shared_ptr<DB> (*)(const std::string&);
+
+  // Allows user to open databases with different configurations.
+  // e.g.: Can open a DB or a TtlDB, etc.
+  static void SetOpenDbFunction(OpenFuncPtr func) { OpenDb = func; }
+
+ protected:
+  static OpenFuncPtr OpenDb;
+};
+StringAppendOperatorTest::OpenFuncPtr StringAppendOperatorTest::OpenDb =
+    nullptr;
+
+// THE TEST CASES BEGIN HERE
+
+TEST_P(StringAppendOperatorTest, IteratorTest) {
+  auto db_ = OpenDb(",");
+  StringLists slists(db_);
+
+  slists.Append("k1", "v1");
+  slists.Append("k1", "v2");
+  slists.Append("k1", "v3");
+
+  slists.Append("k2", "a1");
+  slists.Append("k2", "a2");
+  slists.Append("k2", "a3");
+
+  std::string res;
+  std::unique_ptr<ROCKSDB_NAMESPACE::Iterator> it(
+      db_->NewIterator(ReadOptions()));
+  std::string k1("k1");
+  std::string k2("k2");
+  bool first = true;
+  for (it->Seek(k1); it->Valid(); it->Next()) {
+    res = it->value().ToString();
+    if (first) {
+      ASSERT_EQ(res, "v1,v2,v3");
+      first = false;
+    } else {
+      ASSERT_EQ(res, "a1,a2,a3");
+    }
+  }
+  slists.Append("k2", "a4");
+  slists.Append("k1", "v4");
+
+  // Snapshot should still be the same. Should ignore a4 and v4.
+  first = true;
+  for (it->Seek(k1); it->Valid(); it->Next()) {
+    res = it->value().ToString();
+    if (first) {
+      ASSERT_EQ(res, "v1,v2,v3");
+      first = false;
+    } else {
+      ASSERT_EQ(res, "a1,a2,a3");
+    }
+  }
+
+  // Should release the snapshot and be aware of the new stuff now
+  it.reset(db_->NewIterator(ReadOptions()));
+  first = true;
+  for (it->Seek(k1); it->Valid(); it->Next()) {
+    res = it->value().ToString();
+    if (first) {
+      ASSERT_EQ(res, "v1,v2,v3,v4");
+      first = false;
+    } else {
+      ASSERT_EQ(res, "a1,a2,a3,a4");
+    }
+  }
+
+  // start from k2 this time.
+  for (it->Seek(k2); it->Valid(); it->Next()) {
+    res = it->value().ToString();
+    if (first) {
+      ASSERT_EQ(res, "v1,v2,v3,v4");
+      first = false;
+    } else {
+      ASSERT_EQ(res, "a1,a2,a3,a4");
+    }
+  }
+
+  slists.Append("k3", "g1");
+
+  it.reset(db_->NewIterator(ReadOptions()));
+  first = true;
+  std::string k3("k3");
+  for (it->Seek(k2); it->Valid(); it->Next()) {
+    res = it->value().ToString();
+    if (first) {
+      ASSERT_EQ(res, "a1,a2,a3,a4");
+      first = false;
+    } else {
+      ASSERT_EQ(res, "g1");
+    }
+  }
+  for (it->Seek(k3); it->Valid(); it->Next()) {
+    res = it->value().ToString();
+    if (first) {
+      // should not be hit
+      ASSERT_EQ(res, "a1,a2,a3,a4");
+      first = false;
+    } else {
+      ASSERT_EQ(res, "g1");
+    }
+  }
+}
+
+TEST_P(StringAppendOperatorTest, SimpleTest) {
+  auto db = OpenDb(",");
+  StringLists slists(db);
+
+  slists.Append("k1", "v1");
+  slists.Append("k1", "v2");
+  slists.Append("k1", "v3");
+
+  std::string res;
+  ASSERT_TRUE(slists.Get("k1", &res));
+  ASSERT_EQ(res, "v1,v2,v3");
+}
+
+TEST_P(StringAppendOperatorTest, SimpleDelimiterTest) {
+  auto db = OpenDb("|");
+  StringLists slists(db);
+
+  slists.Append("k1", "v1");
+  slists.Append("k1", "v2");
+  slists.Append("k1", "v3");
+
+  std::string res;
+  ASSERT_TRUE(slists.Get("k1", &res));
+  ASSERT_EQ(res, "v1|v2|v3");
+}
+
+TEST_P(StringAppendOperatorTest, EmptyDelimiterTest) {
+  auto db = OpenDb("");
+  StringLists slists(db);
+
+  slists.Append("k1", "v1");
+  slists.Append("k1", "v2");
+  slists.Append("k1", "v3");
+
+  std::string res;
+  ASSERT_TRUE(slists.Get("k1", &res));
+  ASSERT_EQ(res, "v1v2v3");
+}
+
+TEST_P(StringAppendOperatorTest, MultiCharDelimiterTest) {
+  auto db = OpenDb("<>");
+  StringLists slists(db);
+
+  slists.Append("k1", "v1");
+  slists.Append("k1", "v2");
+  slists.Append("k1", "v3");
+
+  std::string res;
+  ASSERT_TRUE(slists.Get("k1", &res));
+  ASSERT_EQ(res, "v1<>v2<>v3");
+}
+
+TEST_P(StringAppendOperatorTest, DelimiterIsDefensivelyCopiedTest) {
+  std::string delimiter = "<>";
+  auto db = OpenDb(delimiter);
+  StringLists slists(db);
+
+  slists.Append("k1", "v1");
+  slists.Append("k1", "v2");
+  delimiter.clear();
+  slists.Append("k1", "v3");
+
+  std::string res;
+  ASSERT_TRUE(slists.Get("k1", &res));
+  ASSERT_EQ(res, "v1<>v2<>v3");
+}
+
+TEST_P(StringAppendOperatorTest, OneValueNoDelimiterTest) {
+  auto db = OpenDb("!");
+  StringLists slists(db);
+
+  slists.Append("random_key", "single_val");
+
+  std::string res;
+  ASSERT_TRUE(slists.Get("random_key", &res));
+  ASSERT_EQ(res, "single_val");
+}
+
+TEST_P(StringAppendOperatorTest, VariousKeys) {
+  auto db = OpenDb("\n");
+  StringLists slists(db);
+
+  slists.Append("c", "asdasd");
+  slists.Append("a", "x");
+  slists.Append("b", "y");
+  slists.Append("a", "t");
+  slists.Append("a", "r");
+  slists.Append("b", "2");
+  slists.Append("c", "asdasd");
+
+  std::string a, b, c;
+  bool sa, sb, sc;
+  sa = slists.Get("a", &a);
+  sb = slists.Get("b", &b);
+  sc = slists.Get("c", &c);
+
+  ASSERT_TRUE(sa && sb && sc);  // All three keys should have been found
+
+  ASSERT_EQ(a, "x\nt\nr");
+  ASSERT_EQ(b, "y\n2");
+  ASSERT_EQ(c, "asdasd\nasdasd");
+}
+
+// Generate semi random keys/words from a small distribution.
+TEST_P(StringAppendOperatorTest, RandomMixGetAppend) {
+  auto db = OpenDb(" ");
+  StringLists slists(db);
+
+  // Generate a list of random keys and values
+  const int kWordCount = 15;
+  std::string words[] = {"sdasd",     "triejf",       "fnjsdfn",  "dfjisdfsf",
+                         "342839",    "dsuha",        "mabuais",  "sadajsid",
+                         "jf9834hf",  "2d9j89",       "dj9823jd", "a",
+                         "dk02ed2dh", "$(jd4h984$(*", "mabz"};
+  const int kKeyCount = 6;
+  std::string keys[] = {"dhaiusdhu", "denidw", "daisda",
+                        "keykey",    "muki",   "shzassdianmd"};
+
+  // Will store a local copy of all data in order to verify correctness
+  std::map<std::string, std::string> parallel_copy;
+
+  // Generate a bunch of random queries (Append and Get)!
+  enum query_t { APPEND_OP, GET_OP, NUM_OPS };
+  Random randomGen(1337);  // deterministic seed; always get same results!
+
+  const int kNumQueries = 30;
+  for (int q = 0; q < kNumQueries; ++q) {
+    // Generate a random query (Append or Get) and random parameters
+    query_t query = (query_t)randomGen.Uniform((int)NUM_OPS);
+    std::string key = keys[randomGen.Uniform((int)kKeyCount)];
+    std::string word = words[randomGen.Uniform((int)kWordCount)];
+
+    // Apply the query and any checks.
+    if (query == APPEND_OP) {
+      // Apply the rocksdb test-harness Append defined above
+      slists.Append(key, word);  // apply the rocksdb append
+
+      // Apply the similar "Append" to the parallel copy
+      if (parallel_copy[key].size() > 0) {
+        parallel_copy[key] += " " + word;
+      } else {
+        parallel_copy[key] = word;
+      }
+
+    } else if (query == GET_OP) {
+      // Assumes that a non-existent key just returns <empty>
+      std::string res;
+      slists.Get(key, &res);
+      ASSERT_EQ(res, parallel_copy[key]);
+    }
+  }
+}
+
+TEST_P(StringAppendOperatorTest, BIGRandomMixGetAppend) {
+  auto db = OpenDb(" ");
+  StringLists slists(db);
+
+  // Generate a list of random keys and values
+  const int kWordCount = 15;
+  std::string words[] = {"sdasd",     "triejf",       "fnjsdfn",  "dfjisdfsf",
+                         "342839",    "dsuha",        "mabuais",  "sadajsid",
+                         "jf9834hf",  "2d9j89",       "dj9823jd", "a",
+                         "dk02ed2dh", "$(jd4h984$(*", "mabz"};
+  const int kKeyCount = 6;
+  std::string keys[] = {"dhaiusdhu", "denidw", "daisda",
+                        "keykey",    "muki",   "shzassdianmd"};
+
+  // Will store a local copy of all data in order to verify correctness
+  std::map<std::string, std::string> parallel_copy;
+
+  // Generate a bunch of random queries (Append and Get)!
+  enum query_t { APPEND_OP, GET_OP, NUM_OPS };
+  Random randomGen(9138204);  // deterministic seed
+
+  const int kNumQueries = 1000;
+  for (int q = 0; q < kNumQueries; ++q) {
+    // Generate a random query (Append or Get) and random parameters
+    query_t query = (query_t)randomGen.Uniform((int)NUM_OPS);
+    std::string key = keys[randomGen.Uniform((int)kKeyCount)];
+    std::string word = words[randomGen.Uniform((int)kWordCount)];
+
+    // Apply the query and any checks.
+    if (query == APPEND_OP) {
+      // Apply the rocksdb test-harness Append defined above
+      slists.Append(key, word);  // apply the rocksdb append
+
+      // Apply the similar "Append" to the parallel copy
+      if (parallel_copy[key].size() > 0) {
+        parallel_copy[key] += " " + word;
+      } else {
+        parallel_copy[key] = word;
+      }
+
+    } else if (query == GET_OP) {
+      // Assumes that a non-existent key just returns <empty>
+      std::string res;
+      slists.Get(key, &res);
+      ASSERT_EQ(res, parallel_copy[key]);
+    }
+  }
+}
+
+TEST_P(StringAppendOperatorTest, PersistentVariousKeys) {
+  // Perform the following operations in limited scope
+  {
+    auto db = OpenDb("\n");
+    StringLists slists(db);
+
+    slists.Append("c", "asdasd");
+    slists.Append("a", "x");
+    slists.Append("b", "y");
+    slists.Append("a", "t");
+    slists.Append("a", "r");
+    slists.Append("b", "2");
+    slists.Append("c", "asdasd");
+
+    std::string a, b, c;
+    ASSERT_TRUE(slists.Get("a", &a));
+    ASSERT_TRUE(slists.Get("b", &b));
+    ASSERT_TRUE(slists.Get("c", &c));
+
+    ASSERT_EQ(a, "x\nt\nr");
+    ASSERT_EQ(b, "y\n2");
+    ASSERT_EQ(c, "asdasd\nasdasd");
+  }
+
+  // Reopen the database (the previous changes should persist / be remembered)
+  {
+    auto db = OpenDb("\n");
+    StringLists slists(db);
+
+    slists.Append("c", "bbnagnagsx");
+    slists.Append("a", "sa");
+    slists.Append("b", "df");
+    slists.Append("a", "gh");
+    slists.Append("a", "jk");
+    slists.Append("b", "l;");
+    slists.Append("c", "rogosh");
+
+    // The previous changes should be on disk (L0)
+    // The most recent changes should be in memory (MemTable)
+    // Hence, this will test both Get() paths.
+    std::string a, b, c;
+    ASSERT_TRUE(slists.Get("a", &a));
+    ASSERT_TRUE(slists.Get("b", &b));
+    ASSERT_TRUE(slists.Get("c", &c));
+
+    ASSERT_EQ(a, "x\nt\nr\nsa\ngh\njk");
+    ASSERT_EQ(b, "y\n2\ndf\nl;");
+    ASSERT_EQ(c, "asdasd\nasdasd\nbbnagnagsx\nrogosh");
+  }
+
+  // Reopen the database (the previous changes should persist / be remembered)
+  {
+    auto db = OpenDb("\n");
+    StringLists slists(db);
+
+    // All changes should be on disk. This will test VersionSet Get()
+    std::string a, b, c;
+    ASSERT_TRUE(slists.Get("a", &a));
+    ASSERT_TRUE(slists.Get("b", &b));
+    ASSERT_TRUE(slists.Get("c", &c));
+
+    ASSERT_EQ(a, "x\nt\nr\nsa\ngh\njk");
+    ASSERT_EQ(b, "y\n2\ndf\nl;");
+    ASSERT_EQ(c, "asdasd\nasdasd\nbbnagnagsx\nrogosh");
+  }
+}
+
+TEST_P(StringAppendOperatorTest, PersistentFlushAndCompaction) {
+  // Perform the following operations in limited scope
+  {
+    auto db = OpenDb("\n");
+    StringLists slists(db);
+    std::string a, b, c;
+
+    // Append, Flush, Get
+    slists.Append("c", "asdasd");
+    ASSERT_OK(db->Flush(ROCKSDB_NAMESPACE::FlushOptions()));
+    ASSERT_TRUE(slists.Get("c", &c));
+    ASSERT_EQ(c, "asdasd");
+
+    // Append, Flush, Append, Get
+    slists.Append("a", "x");
+    slists.Append("b", "y");
+    ASSERT_OK(db->Flush(ROCKSDB_NAMESPACE::FlushOptions()));
+    slists.Append("a", "t");
+    slists.Append("a", "r");
+    slists.Append("b", "2");
+
+    ASSERT_TRUE(slists.Get("a", &a));
+    ASSERT_EQ(a, "x\nt\nr");
+
+    ASSERT_TRUE(slists.Get("b", &b));
+    ASSERT_EQ(b, "y\n2");
+
+    // Append, Get
+    ASSERT_TRUE(slists.Append("c", "asdasd"));
+    ASSERT_TRUE(slists.Append("b", "monkey"));
+
+    ASSERT_TRUE(slists.Get("a", &a));
+    ASSERT_TRUE(slists.Get("b", &b));
+    ASSERT_TRUE(slists.Get("c", &c));
+
+    ASSERT_EQ(a, "x\nt\nr");
+    ASSERT_EQ(b, "y\n2\nmonkey");
+    ASSERT_EQ(c, "asdasd\nasdasd");
+  }
+
+  // Reopen the database (the previous changes should persist / be remembered)
+  {
+    auto db = OpenDb("\n");
+    StringLists slists(db);
+    std::string a, b, c;
+
+    // Get (Quick check for persistence of previous database)
+    ASSERT_TRUE(slists.Get("a", &a));
+    ASSERT_EQ(a, "x\nt\nr");
+
+    // Append, Compact, Get
+    slists.Append("c", "bbnagnagsx");
+    slists.Append("a", "sa");
+    slists.Append("b", "df");
+    ASSERT_OK(db->CompactRange(CompactRangeOptions(), nullptr, nullptr));
+    ASSERT_TRUE(slists.Get("a", &a));
+    ASSERT_TRUE(slists.Get("b", &b));
+    ASSERT_TRUE(slists.Get("c", &c));
+    ASSERT_EQ(a, "x\nt\nr\nsa");
+    ASSERT_EQ(b, "y\n2\nmonkey\ndf");
+    ASSERT_EQ(c, "asdasd\nasdasd\nbbnagnagsx");
+
+    // Append, Get
+    slists.Append("a", "gh");
+    slists.Append("a", "jk");
+    slists.Append("b", "l;");
+    slists.Append("c", "rogosh");
+    ASSERT_TRUE(slists.Get("a", &a));
+    ASSERT_TRUE(slists.Get("b", &b));
+    ASSERT_TRUE(slists.Get("c", &c));
+    ASSERT_EQ(a, "x\nt\nr\nsa\ngh\njk");
+    ASSERT_EQ(b, "y\n2\nmonkey\ndf\nl;");
+    ASSERT_EQ(c, "asdasd\nasdasd\nbbnagnagsx\nrogosh");
+
+    // Compact, Get
+    ASSERT_OK(db->CompactRange(CompactRangeOptions(), nullptr, nullptr));
+    ASSERT_EQ(a, "x\nt\nr\nsa\ngh\njk");
+    ASSERT_EQ(b, "y\n2\nmonkey\ndf\nl;");
+    ASSERT_EQ(c, "asdasd\nasdasd\nbbnagnagsx\nrogosh");
+
+    // Append, Flush, Compact, Get
+    slists.Append("b", "afcg");
+    ASSERT_OK(db->Flush(ROCKSDB_NAMESPACE::FlushOptions()));
+    ASSERT_OK(db->CompactRange(CompactRangeOptions(), nullptr, nullptr));
+    ASSERT_TRUE(slists.Get("b", &b));
+    ASSERT_EQ(b, "y\n2\nmonkey\ndf\nl;\nafcg");
+  }
+}
+
+TEST_P(StringAppendOperatorTest, SimpleTestNullDelimiter) {
+  auto db = OpenDb(std::string(1, '\0'));
+  StringLists slists(db);
+
+  slists.Append("k1", "v1");
+  slists.Append("k1", "v2");
+  slists.Append("k1", "v3");
+
+  std::string res;
+  ASSERT_TRUE(slists.Get("k1", &res));
+
+  // Construct the desired string. Default constructor doesn't like '\0' chars.
+  std::string checker("v1,v2,v3");  // Verify that the string is right size.
+  checker[2] = '\0';                // Use null delimiter instead of comma.
+  checker[5] = '\0';
+  ASSERT_EQ(checker.size(), 8);  // Verify it is still the correct size
+
+  // Check that the rocksdb result string matches the desired string
+  ASSERT_EQ(res.size(), checker.size());
+  ASSERT_EQ(res, checker);
+}
+
+INSTANTIATE_TEST_CASE_P(StringAppendOperatorTest, StringAppendOperatorTest,
+                        testing::Bool());
+
+}  // namespace ROCKSDB_NAMESPACE
+
+int main(int argc, char** argv) {
+  ROCKSDB_NAMESPACE::port::InstallStackTraceHandler();
+  ::testing::InitGoogleTest(&argc, argv);
+  return RUN_ALL_TESTS();
+}
diff --git a/src/rocksdb/utilities/merge_operators/uint64add.cc b/src/rocksdb/utilities/merge_operators/uint64add.cc
new file mode 100644
index 000000000..5be2f5641
--- /dev/null
+++ b/src/rocksdb/utilities/merge_operators/uint64add.cc
@@ -0,0 +1,75 @@
+// Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#include <memory>
+
+#include "logging/logging.h"
+#include "rocksdb/env.h"
+#include "rocksdb/merge_operator.h"
+#include "rocksdb/slice.h"
+#include "util/coding.h"
+#include "utilities/merge_operators.h"
+
+namespace {  // anonymous namespace
+
+using ROCKSDB_NAMESPACE::AssociativeMergeOperator;
+using ROCKSDB_NAMESPACE::InfoLogLevel;
+using ROCKSDB_NAMESPACE::Logger;
+using ROCKSDB_NAMESPACE::Slice;
+
+// A 'model' merge operator with uint64 addition semantics
+// Implemented as an AssociativeMergeOperator for simplicity and example.
+class UInt64AddOperator : public AssociativeMergeOperator {
+ public:
+  bool Merge(const Slice& /*key*/, const Slice* existing_value,
+             const Slice& value, std::string* new_value,
+             Logger* logger) const override {
+    uint64_t orig_value = 0;
+    if (existing_value) {
+      orig_value = DecodeInteger(*existing_value, logger);
+    }
+    uint64_t operand = DecodeInteger(value, logger);
+
+    assert(new_value);
+    new_value->clear();
+    ROCKSDB_NAMESPACE::PutFixed64(new_value, orig_value + operand);
+
+    return true;  // Return true always since corruption will be treated as 0
+  }
+
+  static const char* kClassName() { return "UInt64AddOperator"; }
+  static const char* kNickName() { return "uint64add"; }
+  const char* Name() const override { return kClassName(); }
+  const char* NickName() const override { return kNickName(); }
+
+ private:
+  // Takes the string and decodes it into a uint64_t
+  // On error, prints a message and returns 0
+  uint64_t DecodeInteger(const Slice& value, Logger* logger) const {
+    uint64_t result = 0;
+
+    if (value.size() == sizeof(uint64_t)) {
+      result = ROCKSDB_NAMESPACE::DecodeFixed64(value.data());
+    } else if (logger != nullptr) {
+      // If value is corrupted, treat it as 0
+      ROCKS_LOG_ERROR(logger,
+                      "uint64 value corruption, size: %" ROCKSDB_PRIszt
+                      " > %" ROCKSDB_PRIszt,
+                      value.size(), sizeof(uint64_t));
+    }
+
+    return result;
+  }
+};
+
+}  // anonymous namespace
+
+namespace ROCKSDB_NAMESPACE {
+
+std::shared_ptr<MergeOperator> MergeOperators::CreateUInt64AddOperator() {
+  return std::make_shared<UInt64AddOperator>();
+}
+
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/utilities/object_registry.cc b/src/rocksdb/utilities/object_registry.cc
new file mode 100644
index 000000000..18834783d
--- /dev/null
+++ b/src/rocksdb/utilities/object_registry.cc
@@ -0,0 +1,383 @@
+// Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#include "rocksdb/utilities/object_registry.h"
+
+#include <ctype.h>
+
+#include "logging/logging.h"
+#include "port/lang.h"
+#include "rocksdb/customizable.h"
+#include "rocksdb/env.h"
+#include "util/string_util.h"
+
+namespace ROCKSDB_NAMESPACE {
+#ifndef ROCKSDB_LITE
+namespace {
+bool MatchesInteger(const std::string &target, size_t start, size_t pos) {
+  // If it is numeric, everything up to the match must be a number
+  int digits = 0;
+  if (target[start] == '-') {
+    start++;  // Allow negative numbers
+  }
+  while (start < pos) {
+    if (!isdigit(target[start++])) {
+      return false;
+    } else {
+      digits++;
+    }
+  }
+  return (digits > 0);
+}
+
+bool MatchesDecimal(const std::string &target, size_t start, size_t pos) {
+  int digits = 0;
+  if (target[start] == '-') {
+    start++;  // Allow negative numbers
+  }
+  for (bool point = false; start < pos; start++) {
+    if (target[start] == '.') {
+      if (point) {
+        return false;
+      } else {
+        point = true;
+      }
+    } else if (!isdigit(target[start])) {
+      return false;
+    } else {
+      digits++;
+    }
+  }
+  return (digits > 0);
+}
+}  // namespace
+
+size_t ObjectLibrary::PatternEntry::MatchSeparatorAt(
+    size_t start, Quantifier mode, const std::string &target, size_t tlen,
+    const std::string &separator) const {
+  size_t slen = separator.size();
+  // See if there is enough space.  If so, find the separator
+  if (tlen < start + slen) {
+    return std::string::npos;  // not enough space left
+  } else if (mode == kMatchExact) {
+    // Exact mode means the next thing we are looking for is the separator
+    if (target.compare(start, slen, separator) != 0) {
+      return std::string::npos;
+    } else {
+      return start + slen;  // Found the separator, return where we found it
+    }
+  } else {
+    auto pos = start + 1;
+    if (!separator.empty()) {
+      pos = target.find(separator, pos);
+    }
+    if (pos == std::string::npos) {
+      return pos;
+    } else if (mode == kMatchInteger) {
+      if (!MatchesInteger(target, start, pos)) {
+        return std::string::npos;
+      }
+    } else if (mode == kMatchDecimal) {
+      if (!MatchesDecimal(target, start, pos)) {
+        return std::string::npos;
+      }
+    }
+    return pos + slen;
+  }
+}
+
+bool ObjectLibrary::PatternEntry::MatchesTarget(const std::string &name,
+                                                size_t nlen,
+                                                const std::string &target,
+                                                size_t tlen) const {
+  if (separators_.empty()) {
+    assert(optional_);  // If there are no separators, it must be only a name
+    return nlen == tlen && name == target;
+  } else if (nlen == tlen) {  // The lengths are the same
+    return optional_ && name == target;
+  } else if (tlen < nlen + slength_) {
+    // The target is not long enough
+    return false;
+  } else if (target.compare(0, nlen, name) != 0) {
+    return false;  // Target does not start with name
+  } else {
+    // Loop through all of the separators one at a time matching them.
+    // Note that we first match the separator and then its quantifiers.
+    // Since we expect the separator first, we start with an exact match
+    // Subsequent matches will use the quantifier of the previous separator
+    size_t start = nlen;
+    auto mode = kMatchExact;
+    for (size_t idx = 0; idx < separators_.size(); ++idx) {
+      const auto &separator = separators_[idx];
+      start = MatchSeparatorAt(start, mode, target, tlen, separator.first);
+      if (start == std::string::npos) {
+        return false;
+      } else {
+        mode = separator.second;
+      }
+    }
+    // We have matched all of the separators.  Now check that what is left
+    // unmatched in the target is acceptable.
+    if (mode == kMatchExact) {
+      return (start == tlen);
+    } else if (start > tlen || (start == tlen && mode != kMatchZeroOrMore)) {
+      return false;
+    } else if (mode == kMatchInteger) {
+      return MatchesInteger(target, start, tlen);
+    } else if (mode == kMatchDecimal) {
+      return MatchesDecimal(target, start, tlen);
+    }
+  }
+  return true;
+}
+
+bool ObjectLibrary::PatternEntry::Matches(const std::string &target) const {
+  auto tlen = target.size();
+  if (MatchesTarget(name_, nlength_, target, tlen)) {
+    return true;
+  } else if (!names_.empty()) {
+    for (const auto &alt : names_) {
+      if (MatchesTarget(alt, alt.size(), target, tlen)) {
+        return true;
+      }
+    }
+  }
+  return false;
+}
+
+size_t ObjectLibrary::GetFactoryCount(size_t *types) const {
+  std::unique_lock<std::mutex> lock(mu_);
+  *types = factories_.size();
+  size_t factories = 0;
+  for (const auto &e : factories_) {
+    factories += e.second.size();
+  }
+  return factories;
+}
+
+size_t ObjectLibrary::GetFactoryCount(const std::string &type) const {
+  std::unique_lock<std::mutex> lock(mu_);
+  auto iter = factories_.find(type);
+  if (iter != factories_.end()) {
+    return iter->second.size();
+  } else {
+    return 0;
+  }
+}
+
+void ObjectLibrary::GetFactoryNames(const std::string &type,
+                                    std::vector<std::string> *names) const {
+  assert(names);
+  std::unique_lock<std::mutex> lock(mu_);
+  auto iter = factories_.find(type);
+  if (iter != factories_.end()) {
+    for (const auto &f : iter->second) {
+      names->push_back(f->Name());
+    }
+  }
+}
+
+void ObjectLibrary::GetFactoryTypes(
+    std::unordered_set<std::string> *types) const {
+  assert(types);
+  std::unique_lock<std::mutex> lock(mu_);
+  for (const auto &iter : factories_) {
+    types->insert(iter.first);
+  }
+}
+
+void ObjectLibrary::Dump(Logger *logger) const {
+  std::unique_lock<std::mutex> lock(mu_);
+  if (logger != nullptr && !factories_.empty()) {
+    ROCKS_LOG_HEADER(logger, "    Registered Library: %s\n", id_.c_str());
+    for (const auto &iter : factories_) {
+      ROCKS_LOG_HEADER(logger, "    Registered factories for type[%s] ",
+                       iter.first.c_str());
+      bool printed_one = false;
+      for (const auto &e : iter.second) {
+        ROCKS_LOG_HEADER(logger, "%c %s", (printed_one) ? ',' : ':', e->Name());
+        printed_one = true;
+      }
+    }
+  }
+}
+
+// Returns the Default singleton instance of the ObjectLibrary
+// This instance will contain most of the "standard" registered objects
+std::shared_ptr<ObjectLibrary> &ObjectLibrary::Default() {
+  // Use avoid destruction here so the default ObjectLibrary will not be
+  // statically destroyed and long-lived.
+  STATIC_AVOID_DESTRUCTION(std::shared_ptr<ObjectLibrary>, instance)
+  (std::make_shared<ObjectLibrary>("default"));
+  return instance;
+}
+
+ObjectRegistry::ObjectRegistry(const std::shared_ptr<ObjectLibrary> &library) {
+  libraries_.push_back(library);
+  for (const auto &b : builtins_) {
+    RegisterPlugin(b.first, b.second);
+  }
+}
+
+std::shared_ptr<ObjectRegistry> ObjectRegistry::Default() {
+  // Use avoid destruction here so the default ObjectRegistry will not be
+  // statically destroyed and long-lived.
+  STATIC_AVOID_DESTRUCTION(std::shared_ptr<ObjectRegistry>, instance)
+  (std::make_shared<ObjectRegistry>(ObjectLibrary::Default()));
+  return instance;
+}
+
+std::shared_ptr<ObjectRegistry> ObjectRegistry::NewInstance() {
+  return std::make_shared<ObjectRegistry>(Default());
+}
+
+std::shared_ptr<ObjectRegistry> ObjectRegistry::NewInstance(
+    const std::shared_ptr<ObjectRegistry> &parent) {
+  return std::make_shared<ObjectRegistry>(parent);
+}
+
+Status ObjectRegistry::SetManagedObject(
+    const std::string &type, const std::string &id,
+    const std::shared_ptr<Customizable> &object) {
+  std::string object_key = ToManagedObjectKey(type, id);
+  std::shared_ptr<Customizable> curr;
+  if (parent_ != nullptr) {
+    curr = parent_->GetManagedObject(type, id);
+  }
+  if (curr == nullptr) {
+    // We did not find the object in any parent.  Update in the current
+    std::unique_lock<std::mutex> lock(objects_mutex_);
+    auto iter = managed_objects_.find(object_key);
+    if (iter != managed_objects_.end()) {  // The object exists
+      curr = iter->second.lock();
+      if (curr != nullptr && curr != object) {
+        return Status::InvalidArgument("Object already exists: ", object_key);
+      } else {
+        iter->second = object;
+      }
+    } else {
+      // The object does not exist.  Add it
+      managed_objects_[object_key] = object;
+    }
+  } else if (curr != object) {
+    return Status::InvalidArgument("Object already exists: ", object_key);
+  }
+  return Status::OK();
+}
+
+std::shared_ptr<Customizable> ObjectRegistry::GetManagedObject(
+    const std::string &type, const std::string &id) const {
+  {
+    std::unique_lock<std::mutex> lock(objects_mutex_);
+    auto iter = managed_objects_.find(ToManagedObjectKey(type, id));
+    if (iter != managed_objects_.end()) {
+      return iter->second.lock();
+    }
+  }
+  if (parent_ != nullptr) {
+    return parent_->GetManagedObject(type, id);
+  } else {
+    return nullptr;
+  }
+}
+
+Status ObjectRegistry::ListManagedObjects(
+    const std::string &type, const std::string &name,
+    std::vector<std::shared_ptr<Customizable>> *results) const {
+  {
+    std::string key = ToManagedObjectKey(type, name);
+    std::unique_lock<std::mutex> lock(objects_mutex_);
+    for (auto iter = managed_objects_.lower_bound(key);
+         iter != managed_objects_.end() && StartsWith(iter->first, key);
+         ++iter) {
+      auto shared = iter->second.lock();
+      if (shared != nullptr) {
+        if (name.empty() || shared->IsInstanceOf(name)) {
+          results->emplace_back(shared);
+        }
+      }
+    }
+  }
+  if (parent_ != nullptr) {
+    return parent_->ListManagedObjects(type, name, results);
+  } else {
+    return Status::OK();
+  }
+}
+
+// Returns the number of registered types for this registry.
+// If specified (not-null), types is updated to include the names of the
+// registered types.
+size_t ObjectRegistry::GetFactoryCount(const std::string &type) const {
+  size_t count = 0;
+  if (parent_ != nullptr) {
+    count = parent_->GetFactoryCount(type);
+  }
+  std::unique_lock<std::mutex> lock(library_mutex_);
+  for (const auto &library : libraries_) {
+    count += library->GetFactoryCount(type);
+  }
+  return count;
+}
+
+void ObjectRegistry::GetFactoryNames(const std::string &type,
+                                     std::vector<std::string> *names) const {
+  assert(names);
+  names->clear();
+  if (parent_ != nullptr) {
+    parent_->GetFactoryNames(type, names);
+  }
+  std::unique_lock<std::mutex> lock(library_mutex_);
+  for (const auto &library : libraries_) {
+    library->GetFactoryNames(type, names);
+  }
+}
+
+void ObjectRegistry::GetFactoryTypes(
+    std::unordered_set<std::string> *types) const {
+  assert(types);
+  if (parent_ != nullptr) {
+    parent_->GetFactoryTypes(types);
+  }
+  std::unique_lock<std::mutex> lock(library_mutex_);
+  for (const auto &library : libraries_) {
+    library->GetFactoryTypes(types);
+  }
+}
+
+void ObjectRegistry::Dump(Logger *logger) const {
+  if (logger != nullptr) {
+    std::unique_lock<std::mutex> lock(library_mutex_);
+    if (!plugins_.empty()) {
+      ROCKS_LOG_HEADER(logger, "    Registered Plugins:");
+      bool printed_one = false;
+      for (const auto &plugin : plugins_) {
+        ROCKS_LOG_HEADER(logger, "%s%s", (printed_one) ? ", " : " ",
+                         plugin.c_str());
+        printed_one = true;
+      }
+      ROCKS_LOG_HEADER(logger, "\n");
+    }
+    for (auto iter = libraries_.crbegin(); iter != libraries_.crend(); ++iter) {
+      iter->get()->Dump(logger);
+    }
+  }
+  if (parent_ != nullptr) {
+    parent_->Dump(logger);
+  }
+}
+
+int ObjectRegistry::RegisterPlugin(const std::string &name,
+                                   const RegistrarFunc &func) {
+  if (!name.empty() && func != nullptr) {
+    plugins_.push_back(name);
+    return AddLibrary(name)->Register(func, name);
+  } else {
+    return -1;
+  }
+}
+
+#endif  // ROCKSDB_LITE
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/utilities/object_registry_test.cc b/src/rocksdb/utilities/object_registry_test.cc
new file mode 100644
index 000000000..90cd155ee
--- /dev/null
+++ b/src/rocksdb/utilities/object_registry_test.cc
@@ -0,0 +1,872 @@
+// Copyright (c) 2016-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#ifndef ROCKSDB_LITE
+
+#include "rocksdb/utilities/object_registry.h"
+
+#include "rocksdb/convenience.h"
+#include "rocksdb/customizable.h"
+#include "test_util/testharness.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+class ObjRegistryTest : public testing::Test {
+ public:
+  static int num_a, num_b;
+};
+
+int ObjRegistryTest::num_a = 0;
+int ObjRegistryTest::num_b = 0;
+static FactoryFunc<Env> test_reg_a = ObjectLibrary::Default()->AddFactory<Env>(
+    ObjectLibrary::PatternEntry("a", false).AddSeparator("://"),
+    [](const std::string& /*uri*/, std::unique_ptr<Env>* /*env_guard*/,
+       std::string* /* errmsg */) {
+      ++ObjRegistryTest::num_a;
+      return Env::Default();
+    });
+
+class WrappedEnv : public EnvWrapper {
+ private:
+  std::string id_;
+
+ public:
+  WrappedEnv(Env* t, const std::string& id) : EnvWrapper(t), id_(id) {}
+  const char* Name() const override { return id_.c_str(); }
+  std::string GetId() const override { return id_; }
+};
+static FactoryFunc<Env> test_reg_b = ObjectLibrary::Default()->AddFactory<Env>(
+    ObjectLibrary::PatternEntry("b", false).AddSeparator("://"),
+    [](const std::string& uri, std::unique_ptr<Env>* env_guard,
+       std::string* /* errmsg */) {
+      ++ObjRegistryTest::num_b;
+      // Env::Default() is a singleton so we can't grant ownership directly to
+      // the caller - we must wrap it first.
+      env_guard->reset(new WrappedEnv(Env::Default(), uri));
+      return env_guard->get();
+    });
+
+TEST_F(ObjRegistryTest, Basics) {
+  std::string msg;
+  std::unique_ptr<Env> guard;
+  Env* a_env = nullptr;
+
+  auto registry = ObjectRegistry::NewInstance();
+  ASSERT_NOK(registry->NewStaticObject<Env>("c://test", &a_env));
+  ASSERT_NOK(registry->NewUniqueObject<Env>("c://test", &guard));
+  ASSERT_EQ(a_env, nullptr);
+  ASSERT_EQ(guard, nullptr);
+  ASSERT_EQ(0, num_a);
+  ASSERT_EQ(0, num_b);
+
+  ASSERT_OK(registry->NewStaticObject<Env>("a://test", &a_env));
+  ASSERT_NE(a_env, nullptr);
+  ASSERT_EQ(1, num_a);
+  ASSERT_EQ(0, num_b);
+
+  ASSERT_OK(registry->NewUniqueObject<Env>("b://test", &guard));
+  ASSERT_NE(guard, nullptr);
+  ASSERT_EQ(1, num_a);
+  ASSERT_EQ(1, num_b);
+
+  Env* b_env = nullptr;
+  ASSERT_NOK(registry->NewStaticObject<Env>("b://test", &b_env));
+  ASSERT_EQ(b_env, nullptr);
+  ASSERT_EQ(1, num_a);
+  ASSERT_EQ(2, num_b);  // Created but rejected as not static
+
+  b_env = a_env;
+  ASSERT_NOK(registry->NewStaticObject<Env>("b://test", &b_env));
+  ASSERT_EQ(b_env, a_env);
+  ASSERT_EQ(1, num_a);
+  ASSERT_EQ(3, num_b);
+
+  b_env = guard.get();
+  ASSERT_NOK(registry->NewUniqueObject<Env>("a://test", &guard));
+  ASSERT_EQ(guard.get(), b_env);  // Unchanged
+  ASSERT_EQ(2, num_a);            // Created one but rejected it as not unique
+  ASSERT_EQ(3, num_b);
+}
+
+TEST_F(ObjRegistryTest, LocalRegistry) {
+  Env* env = nullptr;
+  auto registry = ObjectRegistry::NewInstance();
+  std::shared_ptr<ObjectLibrary> library =
+      std::make_shared<ObjectLibrary>("local");
+  registry->AddLibrary(library);
+  library->AddFactory<Env>(
+      "test-local",
+      [](const std::string& /*uri*/, std::unique_ptr<Env>* /*guard */,
+         std::string* /* errmsg */) { return Env::Default(); });
+
+  ObjectLibrary::Default()->AddFactory<Env>(
+      "test-global",
+      [](const std::string& /*uri*/, std::unique_ptr<Env>* /*guard */,
+         std::string* /* errmsg */) { return Env::Default(); });
+
+  ASSERT_NOK(
+      ObjectRegistry::NewInstance()->NewStaticObject<Env>("test-local", &env));
+  ASSERT_EQ(env, nullptr);
+  ASSERT_OK(
+      ObjectRegistry::NewInstance()->NewStaticObject<Env>("test-global", &env));
+  ASSERT_NE(env, nullptr);
+  ASSERT_OK(registry->NewStaticObject<Env>("test-local", &env));
+  ASSERT_NE(env, nullptr);
+  ASSERT_OK(registry->NewStaticObject<Env>("test-global", &env));
+  ASSERT_NE(env, nullptr);
+}
+
+static int RegisterTestUnguarded(ObjectLibrary& library,
+                                 const std::string& /*arg*/) {
+  library.AddFactory<Env>(
+      "unguarded",
+      [](const std::string& /*uri*/, std::unique_ptr<Env>* /*guard */,
+         std::string* /* errmsg */) { return Env::Default(); });
+  library.AddFactory<Env>(
+      "guarded", [](const std::string& uri, std::unique_ptr<Env>* guard,
+                    std::string* /* errmsg */) {
+        guard->reset(new WrappedEnv(Env::Default(), uri));
+        return guard->get();
+      });
+  return 2;
+}
+
+TEST_F(ObjRegistryTest, CheckShared) {
+  std::shared_ptr<Env> shared;
+  std::shared_ptr<ObjectRegistry> registry = ObjectRegistry::NewInstance();
+  registry->AddLibrary("shared", RegisterTestUnguarded, "");
+
+  ASSERT_OK(registry->NewSharedObject<Env>("guarded", &shared));
+  ASSERT_NE(shared, nullptr);
+  shared.reset();
+  ASSERT_NOK(registry->NewSharedObject<Env>("unguarded", &shared));
+  ASSERT_EQ(shared, nullptr);
+}
+
+TEST_F(ObjRegistryTest, CheckStatic) {
+  Env* env = nullptr;
+  std::shared_ptr<ObjectRegistry> registry = ObjectRegistry::NewInstance();
+  registry->AddLibrary("static", RegisterTestUnguarded, "");
+
+  ASSERT_NOK(registry->NewStaticObject<Env>("guarded", &env));
+  ASSERT_EQ(env, nullptr);
+  env = nullptr;
+  ASSERT_OK(registry->NewStaticObject<Env>("unguarded", &env));
+  ASSERT_NE(env, nullptr);
+}
+
+TEST_F(ObjRegistryTest, CheckUnique) {
+  std::unique_ptr<Env> unique;
+  std::shared_ptr<ObjectRegistry> registry = ObjectRegistry::NewInstance();
+  registry->AddLibrary("unique", RegisterTestUnguarded, "");
+
+  ASSERT_OK(registry->NewUniqueObject<Env>("guarded", &unique));
+  ASSERT_NE(unique, nullptr);
+  unique.reset();
+  ASSERT_NOK(registry->NewUniqueObject<Env>("unguarded", &unique));
+  ASSERT_EQ(unique, nullptr);
+}
+
+TEST_F(ObjRegistryTest, FailingFactory) {
+  std::shared_ptr<ObjectRegistry> registry = ObjectRegistry::NewInstance();
+  std::shared_ptr<ObjectLibrary> library =
+      std::make_shared<ObjectLibrary>("failing");
+  registry->AddLibrary(library);
+  library->AddFactory<Env>(
+      "failing", [](const std::string& /*uri*/,
+                    std::unique_ptr<Env>* /*guard */, std::string* errmsg) {
+        *errmsg = "Bad Factory";
+        return nullptr;
+      });
+  std::unique_ptr<Env> unique;
+  std::shared_ptr<Env> shared;
+  Env* pointer = nullptr;
+  Status s;
+  s = registry->NewUniqueObject<Env>("failing", &unique);
+  ASSERT_TRUE(s.IsInvalidArgument());
+  s = registry->NewSharedObject<Env>("failing", &shared);
+  ASSERT_TRUE(s.IsInvalidArgument());
+  s = registry->NewStaticObject<Env>("failing", &pointer);
+  ASSERT_TRUE(s.IsInvalidArgument());
+
+  s = registry->NewUniqueObject<Env>("missing", &unique);
+  ASSERT_TRUE(s.IsNotSupported());
+  s = registry->NewSharedObject<Env>("missing", &shared);
+  ASSERT_TRUE(s.IsNotSupported());
+  s = registry->NewStaticObject<Env>("missing", &pointer);
+  ASSERT_TRUE(s.IsNotSupported());
+}
+
+TEST_F(ObjRegistryTest, TestRegistryParents) {
+  auto grand = ObjectRegistry::Default();
+  auto parent = ObjectRegistry::NewInstance();  // parent with a grandparent
+  auto uncle = ObjectRegistry::NewInstance(grand);
+  auto child = ObjectRegistry::NewInstance(parent);
+  auto cousin = ObjectRegistry::NewInstance(uncle);
+
+  auto library = parent->AddLibrary("parent");
+  library->AddFactory<Env>(
+      "parent", [](const std::string& uri, std::unique_ptr<Env>* guard,
+                   std::string* /* errmsg */) {
+        guard->reset(new WrappedEnv(Env::Default(), uri));
+        return guard->get();
+      });
+  library = cousin->AddLibrary("cousin");
+  library->AddFactory<Env>(
+      "cousin", [](const std::string& uri, std::unique_ptr<Env>* guard,
+                   std::string* /* errmsg */) {
+        guard->reset(new WrappedEnv(Env::Default(), uri));
+        return guard->get();
+      });
+
+  Env* env = nullptr;
+  std::unique_ptr<Env> guard;
+  std::string msg;
+
+  // a:://* is registered in Default, so they should all work
+  ASSERT_OK(parent->NewStaticObject<Env>("a://test", &env));
+  ASSERT_OK(child->NewStaticObject<Env>("a://test", &env));
+  ASSERT_OK(uncle->NewStaticObject<Env>("a://test", &env));
+  ASSERT_OK(cousin->NewStaticObject<Env>("a://test", &env));
+
+  // The parent env is only registered for parent, not uncle,
+  // So parent and child should return success and uncle and cousin should fail
+  ASSERT_OK(parent->NewUniqueObject<Env>("parent", &guard));
+  ASSERT_OK(child->NewUniqueObject<Env>("parent", &guard));
+  ASSERT_NOK(uncle->NewUniqueObject<Env>("parent", &guard));
+  ASSERT_NOK(cousin->NewUniqueObject<Env>("parent", &guard));
+
+  // The cousin is only registered in the cousin, so all of the others should
+  // fail
+  ASSERT_OK(cousin->NewUniqueObject<Env>("cousin", &guard));
+  ASSERT_NOK(parent->NewUniqueObject<Env>("cousin", &guard));
+  ASSERT_NOK(child->NewUniqueObject<Env>("cousin", &guard));
+  ASSERT_NOK(uncle->NewUniqueObject<Env>("cousin", &guard));
+}
+
+class MyCustomizable : public Customizable {
+ public:
+  static const char* Type() { return "MyCustomizable"; }
+  MyCustomizable(const char* prefix, const std::string& id) : id_(id) {
+    name_ = id_.substr(0, strlen(prefix) - 1);
+  }
+  const char* Name() const override { return name_.c_str(); }
+  std::string GetId() const override { return id_; }
+
+ private:
+  std::string id_;
+  std::string name_;
+};
+
+TEST_F(ObjRegistryTest, TestFactoryCount) {
+  std::string msg;
+  auto grand = ObjectRegistry::Default();
+  auto local = ObjectRegistry::NewInstance();
+  std::unordered_set<std::string> grand_types, local_types;
+  std::vector<std::string> grand_names, local_names;
+
+  // Check how many types we have on startup.
+  // Grand should equal local
+  grand->GetFactoryTypes(&grand_types);
+  local->GetFactoryTypes(&local_types);
+  ASSERT_EQ(grand_types, local_types);
+  size_t grand_count = grand->GetFactoryCount(Env::Type());
+  size_t local_count = local->GetFactoryCount(Env::Type());
+
+  ASSERT_EQ(grand_count, local_count);
+  grand->GetFactoryNames(Env::Type(), &grand_names);
+  local->GetFactoryNames(Env::Type(), &local_names);
+  ASSERT_EQ(grand_names.size(), grand_count);
+  ASSERT_EQ(local_names.size(), local_count);
+  ASSERT_EQ(grand_names, local_names);
+
+  // Add an Env to the local registry.
+  // This will add one factory.
+  auto library = local->AddLibrary("local");
+  library->AddFactory<Env>(
+      "A", [](const std::string& /*uri*/, std::unique_ptr<Env>* /*guard */,
+              std::string* /* errmsg */) { return nullptr; });
+  ASSERT_EQ(local_count + 1, local->GetFactoryCount(Env::Type()));
+  ASSERT_EQ(grand_count, grand->GetFactoryCount(Env::Type()));
+  local->GetFactoryTypes(&local_types);
+  local->GetFactoryNames(Env::Type(), &local_names);
+  ASSERT_EQ(grand_names.size() + 1, local_names.size());
+  ASSERT_EQ(local_names.size(), local->GetFactoryCount(Env::Type()));
+
+  if (grand_count == 0) {
+    // There were no Env when we started.  Should have one more type
+    // than previously
+    ASSERT_NE(grand_types, local_types);
+    ASSERT_EQ(grand_types.size() + 1, local_types.size());
+  } else {
+    // There was an Env type when we started.  The types should match
+    ASSERT_EQ(grand_types, local_types);
+  }
+
+  // Add a MyCustomizable to the registry.  This should be a new type
+  library->AddFactory<MyCustomizable>(
+      "MY", [](const std::string& /*uri*/,
+               std::unique_ptr<MyCustomizable>* /*guard */,
+               std::string* /* errmsg */) { return nullptr; });
+  ASSERT_EQ(local_count + 1, local->GetFactoryCount(Env::Type()));
+  ASSERT_EQ(grand_count, grand->GetFactoryCount(Env::Type()));
+  ASSERT_EQ(0U, grand->GetFactoryCount(MyCustomizable::Type()));
+  ASSERT_EQ(1U, local->GetFactoryCount(MyCustomizable::Type()));
+
+  local->GetFactoryNames(MyCustomizable::Type(), &local_names);
+  ASSERT_EQ(1U, local_names.size());
+  ASSERT_EQ(local_names[0], "MY");
+
+  local->GetFactoryTypes(&local_types);
+  ASSERT_EQ(grand_count == 0 ? 2 : grand_types.size() + 1, local_types.size());
+
+  // Add the same name again.  We should now have 2 factories.
+  library->AddFactory<MyCustomizable>(
+      "MY", [](const std::string& /*uri*/,
+               std::unique_ptr<MyCustomizable>* /*guard */,
+               std::string* /* errmsg */) { return nullptr; });
+  local->GetFactoryNames(MyCustomizable::Type(), &local_names);
+  ASSERT_EQ(2U, local_names.size());
+}
+
+TEST_F(ObjRegistryTest, TestManagedObjects) {
+  auto registry = ObjectRegistry::NewInstance();
+  auto m_a1 = std::make_shared<MyCustomizable>("", "A");
+  auto m_a2 = std::make_shared<MyCustomizable>("", "A");
+
+  ASSERT_EQ(registry->GetManagedObject<MyCustomizable>("A"), nullptr);
+  ASSERT_OK(registry->SetManagedObject<MyCustomizable>(m_a1));
+  ASSERT_EQ(registry->GetManagedObject<MyCustomizable>("A"), m_a1);
+
+  ASSERT_NOK(registry->SetManagedObject<MyCustomizable>(m_a2));
+  ASSERT_OK(registry->SetManagedObject<MyCustomizable>(m_a1));
+  m_a1.reset();
+  ASSERT_EQ(registry->GetManagedObject<MyCustomizable>("A"), nullptr);
+  ASSERT_OK(registry->SetManagedObject<MyCustomizable>(m_a2));
+  ASSERT_EQ(registry->GetManagedObject<MyCustomizable>("A"), m_a2);
+}
+
+TEST_F(ObjRegistryTest, TestTwoManagedObjects) {
+  auto registry = ObjectRegistry::NewInstance();
+  auto m_a = std::make_shared<MyCustomizable>("", "A");
+  auto m_b = std::make_shared<MyCustomizable>("", "B");
+  std::vector<std::shared_ptr<MyCustomizable>> objects;
+
+  ASSERT_EQ(registry->GetManagedObject<MyCustomizable>("A"), nullptr);
+  ASSERT_EQ(registry->GetManagedObject<MyCustomizable>("B"), nullptr);
+  ASSERT_OK(registry->ListManagedObjects(&objects));
+  ASSERT_EQ(objects.size(), 0U);
+  ASSERT_OK(registry->SetManagedObject(m_a));
+  ASSERT_EQ(registry->GetManagedObject<MyCustomizable>("B"), nullptr);
+  ASSERT_EQ(registry->GetManagedObject<MyCustomizable>("A"), m_a);
+  ASSERT_OK(registry->ListManagedObjects(&objects));
+  ASSERT_EQ(objects.size(), 1U);
+  ASSERT_EQ(objects.front(), m_a);
+
+  ASSERT_OK(registry->SetManagedObject(m_b));
+  ASSERT_EQ(registry->GetManagedObject<MyCustomizable>("A"), m_a);
+  ASSERT_EQ(registry->GetManagedObject<MyCustomizable>("B"), m_b);
+  ASSERT_OK(registry->ListManagedObjects(&objects));
+  ASSERT_EQ(objects.size(), 2U);
+  ASSERT_OK(registry->ListManagedObjects("A", &objects));
+  ASSERT_EQ(objects.size(), 1U);
+  ASSERT_EQ(objects.front(), m_a);
+  ASSERT_OK(registry->ListManagedObjects("B", &objects));
+  ASSERT_EQ(objects.size(), 1U);
+  ASSERT_EQ(objects.front(), m_b);
+  ASSERT_OK(registry->ListManagedObjects("C", &objects));
+  ASSERT_EQ(objects.size(), 0U);
+
+  m_a.reset();
+  objects.clear();
+
+  ASSERT_EQ(registry->GetManagedObject<MyCustomizable>("B"), m_b);
+  ASSERT_EQ(registry->GetManagedObject<MyCustomizable>("A"), nullptr);
+  ASSERT_OK(registry->ListManagedObjects(&objects));
+  ASSERT_EQ(objects.size(), 1U);
+  ASSERT_EQ(objects.front(), m_b);
+
+  m_b.reset();
+  objects.clear();
+  ASSERT_EQ(registry->GetManagedObject<MyCustomizable>("A"), nullptr);
+  ASSERT_EQ(registry->GetManagedObject<MyCustomizable>("B"), nullptr);
+}
+
+TEST_F(ObjRegistryTest, TestAlternateNames) {
+  auto registry = ObjectRegistry::NewInstance();
+  auto m_a = std::make_shared<MyCustomizable>("", "A");
+  auto m_b = std::make_shared<MyCustomizable>("", "B");
+  std::vector<std::shared_ptr<MyCustomizable>> objects;
+  // Test no objects exist
+  ASSERT_EQ(registry->GetManagedObject<MyCustomizable>("A"), nullptr);
+  ASSERT_EQ(registry->GetManagedObject<MyCustomizable>("B"), nullptr);
+  ASSERT_EQ(registry->GetManagedObject<MyCustomizable>("TheOne"), nullptr);
+  ASSERT_OK(registry->ListManagedObjects(&objects));
+  ASSERT_EQ(objects.size(), 0U);
+
+  // Mark "TheOne" to be A
+  ASSERT_OK(registry->SetManagedObject("TheOne", m_a));
+  ASSERT_EQ(registry->GetManagedObject<MyCustomizable>("B"), nullptr);
+  ASSERT_EQ(registry->GetManagedObject<MyCustomizable>("A"), nullptr);
+  ASSERT_EQ(registry->GetManagedObject<MyCustomizable>("TheOne"), m_a);
+  ASSERT_OK(registry->ListManagedObjects(&objects));
+  ASSERT_EQ(objects.size(), 1U);
+  ASSERT_EQ(objects.front(), m_a);
+
+  // Try to mark "TheOne" again.
+  ASSERT_NOK(registry->SetManagedObject("TheOne", m_b));
+  ASSERT_OK(registry->SetManagedObject("TheOne", m_a));
+
+  // Add "A" as a managed object.  Registered 2x
+  ASSERT_OK(registry->SetManagedObject(m_a));
+  ASSERT_EQ(registry->GetManagedObject<MyCustomizable>("B"), nullptr);
+  ASSERT_EQ(registry->GetManagedObject<MyCustomizable>("A"), m_a);
+  ASSERT_EQ(registry->GetManagedObject<MyCustomizable>("TheOne"), m_a);
+  ASSERT_OK(registry->ListManagedObjects(&objects));
+  ASSERT_EQ(objects.size(), 2U);
+
+  // Delete "A".
+  m_a.reset();
+  objects.clear();
+
+  ASSERT_EQ(registry->GetManagedObject<MyCustomizable>("TheOne"), nullptr);
+  ASSERT_OK(registry->SetManagedObject("TheOne", m_b));
+  ASSERT_EQ(registry->GetManagedObject<MyCustomizable>("TheOne"), m_b);
+  ASSERT_OK(registry->ListManagedObjects(&objects));
+  ASSERT_EQ(objects.size(), 1U);
+  ASSERT_EQ(objects.front(), m_b);
+
+  m_b.reset();
+  objects.clear();
+  ASSERT_EQ(registry->GetManagedObject<MyCustomizable>("A"), nullptr);
+  ASSERT_EQ(registry->GetManagedObject<MyCustomizable>("A"), nullptr);
+  ASSERT_EQ(registry->GetManagedObject<MyCustomizable>("TheOne"), nullptr);
+  ASSERT_OK(registry->ListManagedObjects(&objects));
+  ASSERT_EQ(objects.size(), 0U);
+}
+
+TEST_F(ObjRegistryTest, TestTwoManagedClasses) {
+  class MyCustomizable2 : public MyCustomizable {
+   public:
+    static const char* Type() { return "MyCustomizable2"; }
+    MyCustomizable2(const char* prefix, const std::string& id)
+        : MyCustomizable(prefix, id) {}
+  };
+
+  auto registry = ObjectRegistry::NewInstance();
+  auto m_a1 = std::make_shared<MyCustomizable>("", "A");
+  auto m_a2 = std::make_shared<MyCustomizable2>("", "A");
+  std::vector<std::shared_ptr<MyCustomizable>> obj1s;
+  std::vector<std::shared_ptr<MyCustomizable2>> obj2s;
+
+  ASSERT_EQ(registry->GetManagedObject<MyCustomizable>("A"), nullptr);
+  ASSERT_EQ(registry->GetManagedObject<MyCustomizable2>("A"), nullptr);
+
+  ASSERT_OK(registry->SetManagedObject(m_a1));
+  ASSERT_EQ(registry->GetManagedObject<MyCustomizable>("A"), m_a1);
+  ASSERT_EQ(registry->GetManagedObject<MyCustomizable2>("A"), nullptr);
+
+  ASSERT_OK(registry->SetManagedObject(m_a2));
+  ASSERT_EQ(registry->GetManagedObject<MyCustomizable2>("A"), m_a2);
+  ASSERT_OK(registry->ListManagedObjects(&obj1s));
+  ASSERT_OK(registry->ListManagedObjects(&obj2s));
+  ASSERT_EQ(obj1s.size(), 1U);
+  ASSERT_EQ(obj2s.size(), 1U);
+  ASSERT_EQ(obj1s.front(), m_a1);
+  ASSERT_EQ(obj2s.front(), m_a2);
+  m_a1.reset();
+  obj1s.clear();
+  obj2s.clear();
+  ASSERT_EQ(registry->GetManagedObject<MyCustomizable>("A"), nullptr);
+  ASSERT_EQ(registry->GetManagedObject<MyCustomizable2>("A"), m_a2);
+
+  m_a2.reset();
+  ASSERT_EQ(registry->GetManagedObject<MyCustomizable>("A"), nullptr);
+  ASSERT_EQ(registry->GetManagedObject<MyCustomizable2>("A"), nullptr);
+}
+
+TEST_F(ObjRegistryTest, TestManagedObjectsWithParent) {
+  auto base = ObjectRegistry::NewInstance();
+  auto registry = ObjectRegistry::NewInstance(base);
+
+  auto m_a = std::make_shared<MyCustomizable>("", "A");
+  auto m_b = std::make_shared<MyCustomizable>("", "A");
+
+  ASSERT_EQ(registry->GetManagedObject<MyCustomizable>("A"), nullptr);
+  ASSERT_OK(base->SetManagedObject(m_a));
+  ASSERT_EQ(registry->GetManagedObject<MyCustomizable>("A"), m_a);
+
+  ASSERT_NOK(registry->SetManagedObject(m_b));
+  ASSERT_OK(registry->SetManagedObject(m_a));
+
+  m_a.reset();
+  ASSERT_EQ(registry->GetManagedObject<MyCustomizable>("A"), nullptr);
+  ASSERT_OK(registry->SetManagedObject(m_b));
+  ASSERT_EQ(registry->GetManagedObject<MyCustomizable>("A"), m_b);
+}
+
+TEST_F(ObjRegistryTest, TestGetOrCreateManagedObject) {
+  auto registry = ObjectRegistry::NewInstance();
+  registry->AddLibrary("test")->AddFactory<MyCustomizable>(
+      ObjectLibrary::PatternEntry::AsIndividualId("MC"),
+      [](const std::string& uri, std::unique_ptr<MyCustomizable>* guard,
+         std::string* /* errmsg */) {
+        guard->reset(new MyCustomizable("MC", uri));
+        return guard->get();
+      });
+  std::shared_ptr<MyCustomizable> m_a, m_b, obj;
+  std::vector<std::shared_ptr<MyCustomizable>> objs;
+
+  std::unordered_map<std::string, std::string> opt_map;
+
+  ASSERT_EQ(registry->GetManagedObject<MyCustomizable>("MC@A#1"), nullptr);
+  ASSERT_EQ(registry->GetManagedObject<MyCustomizable>("MC@B#1"), nullptr);
+  ASSERT_OK(registry->GetOrCreateManagedObject("MC@A#1", &m_a));
+  ASSERT_OK(registry->GetOrCreateManagedObject("MC@B#1", &m_b));
+  ASSERT_EQ(registry->GetManagedObject<MyCustomizable>("MC@A#1"), m_a);
+  ASSERT_OK(registry->GetOrCreateManagedObject("MC@A#1", &obj));
+  ASSERT_EQ(obj, m_a);
+  ASSERT_OK(registry->GetOrCreateManagedObject("MC@B#1", &obj));
+  ASSERT_EQ(obj, m_b);
+  ASSERT_OK(registry->ListManagedObjects(&objs));
+  ASSERT_EQ(objs.size(), 2U);
+
+  objs.clear();
+  m_a.reset();
+  obj.reset();
+  ASSERT_OK(registry->GetOrCreateManagedObject("MC@A#1", &m_a));
+  ASSERT_EQ(1, m_a.use_count());
+  ASSERT_OK(registry->GetOrCreateManagedObject("MC@B#1", &obj));
+  ASSERT_EQ(2, obj.use_count());
+}
+
+TEST_F(ObjRegistryTest, RegisterPlugin) {
+  std::shared_ptr<ObjectRegistry> registry = ObjectRegistry::NewInstance();
+  std::unique_ptr<Env> guard;
+  Env* env = nullptr;
+
+  ASSERT_NOK(registry->NewObject<Env>("unguarded", &env, &guard));
+  ASSERT_EQ(registry->RegisterPlugin("Missing", nullptr), -1);
+  ASSERT_EQ(registry->RegisterPlugin("", RegisterTestUnguarded), -1);
+  ASSERT_GT(registry->RegisterPlugin("Valid", RegisterTestUnguarded), 0);
+  ASSERT_OK(registry->NewObject<Env>("unguarded", &env, &guard));
+  ASSERT_NE(env, nullptr);
+}
+class PatternEntryTest : public testing::Test {};
+
+TEST_F(PatternEntryTest, TestSimpleEntry) {
+  ObjectLibrary::PatternEntry entry("ABC", true);
+
+  ASSERT_TRUE(entry.Matches("ABC"));
+  ASSERT_FALSE(entry.Matches("AABC"));
+  ASSERT_FALSE(entry.Matches("ABCA"));
+  ASSERT_FALSE(entry.Matches("AABCA"));
+  ASSERT_FALSE(entry.Matches("AB"));
+  ASSERT_FALSE(entry.Matches("BC"));
+  ASSERT_FALSE(entry.Matches("ABD"));
+  ASSERT_FALSE(entry.Matches("BCA"));
+}
+
+TEST_F(PatternEntryTest, TestPatternEntry) {
+  // Matches A:+
+  ObjectLibrary::PatternEntry entry("A", false);
+  entry.AddSeparator(":");
+  ASSERT_FALSE(entry.Matches("A"));
+  ASSERT_FALSE(entry.Matches("AA"));
+  ASSERT_FALSE(entry.Matches("AB"));
+  ASSERT_FALSE(entry.Matches("B"));
+  ASSERT_FALSE(entry.Matches("A:"));
+  ASSERT_FALSE(entry.Matches("AA:"));
+  ASSERT_FALSE(entry.Matches("AA:B"));
+  ASSERT_FALSE(entry.Matches("AA:BB"));
+  ASSERT_TRUE(entry.Matches("A:B"));
+  ASSERT_TRUE(entry.Matches("A:BB"));
+
+  entry.SetOptional(true);  // Now matches "A" or "A:+"
+  ASSERT_TRUE(entry.Matches("A"));
+  ASSERT_FALSE(entry.Matches("AA"));
+  ASSERT_FALSE(entry.Matches("AB"));
+  ASSERT_FALSE(entry.Matches("B"));
+  ASSERT_FALSE(entry.Matches("A:"));
+  ASSERT_FALSE(entry.Matches("AA:"));
+  ASSERT_FALSE(entry.Matches("AA:B"));
+  ASSERT_FALSE(entry.Matches("AA:BB"));
+  ASSERT_TRUE(entry.Matches("A:B"));
+  ASSERT_TRUE(entry.Matches("A:BB"));
+}
+
+TEST_F(PatternEntryTest, MatchZeroOrMore) {
+  // Matches A:*
+  ObjectLibrary::PatternEntry entry("A", false);
+  entry.AddSeparator(":", false);
+  ASSERT_FALSE(entry.Matches("A"));
+  ASSERT_FALSE(entry.Matches("AA"));
+  ASSERT_FALSE(entry.Matches("AB"));
+  ASSERT_FALSE(entry.Matches("B"));
+  ASSERT_TRUE(entry.Matches("A:"));
+  ASSERT_FALSE(entry.Matches("B:"));
+  ASSERT_FALSE(entry.Matches("B:A"));
+  ASSERT_FALSE(entry.Matches("AA:"));
+  ASSERT_FALSE(entry.Matches("AA:B"));
+  ASSERT_FALSE(entry.Matches("AA:BB"));
+  ASSERT_TRUE(entry.Matches("A:B"));
+  ASSERT_TRUE(entry.Matches("A:BB"));
+
+  entry.SetOptional(true);  // Now matches "A" or "A:*"
+  ASSERT_TRUE(entry.Matches("A"));
+  ASSERT_FALSE(entry.Matches("AA"));
+  ASSERT_FALSE(entry.Matches("AB"));
+  ASSERT_FALSE(entry.Matches("B"));
+  ASSERT_TRUE(entry.Matches("A:"));
+  ASSERT_FALSE(entry.Matches("B:"));
+  ASSERT_FALSE(entry.Matches("B:A"));
+  ASSERT_FALSE(entry.Matches("AA:"));
+  ASSERT_FALSE(entry.Matches("AA:B"));
+  ASSERT_FALSE(entry.Matches("AA:BB"));
+  ASSERT_TRUE(entry.Matches("A:B"));
+  ASSERT_TRUE(entry.Matches("A:BB"));
+}
+
+TEST_F(PatternEntryTest, TestSuffixEntry) {
+  ObjectLibrary::PatternEntry entry("AA", true);
+  entry.AddSuffix("BB");
+
+  ASSERT_TRUE(entry.Matches("AA"));
+  ASSERT_TRUE(entry.Matches("AABB"));
+
+  ASSERT_FALSE(entry.Matches("A"));
+  ASSERT_FALSE(entry.Matches("AB"));
+  ASSERT_FALSE(entry.Matches("B"));
+  ASSERT_FALSE(entry.Matches("BB"));
+  ASSERT_FALSE(entry.Matches("ABA"));
+  ASSERT_FALSE(entry.Matches("BBAA"));
+  ASSERT_FALSE(entry.Matches("AABBA"));
+  ASSERT_FALSE(entry.Matches("AABBB"));
+}
+
+TEST_F(PatternEntryTest, TestNumericEntry) {
+  ObjectLibrary::PatternEntry entry("A", false);
+  entry.AddNumber(":");
+  ASSERT_FALSE(entry.Matches("A"));
+  ASSERT_FALSE(entry.Matches("AA"));
+  ASSERT_FALSE(entry.Matches("A:"));
+  ASSERT_FALSE(entry.Matches("AA:"));
+  ASSERT_TRUE(entry.Matches("A:1"));
+  ASSERT_TRUE(entry.Matches("A:11"));
+  ASSERT_FALSE(entry.Matches("AA:1"));
+  ASSERT_FALSE(entry.Matches("AA:11"));
+  ASSERT_FALSE(entry.Matches("A:B"));
+  ASSERT_FALSE(entry.Matches("A:1B"));
+  ASSERT_FALSE(entry.Matches("A:B1"));
+
+  entry.AddSeparator(":", false);
+  ASSERT_FALSE(entry.Matches("A"));
+  ASSERT_FALSE(entry.Matches("AA"));
+  ASSERT_FALSE(entry.Matches("A:"));
+  ASSERT_FALSE(entry.Matches("AA:"));
+  ASSERT_TRUE(entry.Matches("A:1:"));
+  ASSERT_TRUE(entry.Matches("A:11:"));
+  ASSERT_FALSE(entry.Matches("A:1"));
+  ASSERT_FALSE(entry.Matches("A:B1:"));
+  ASSERT_FALSE(entry.Matches("A:1B:"));
+  ASSERT_FALSE(entry.Matches("A::"));
+}
+
+TEST_F(PatternEntryTest, TestDoubleEntry) {
+  ObjectLibrary::PatternEntry entry("A", false);
+  entry.AddNumber(":", false);
+  ASSERT_FALSE(entry.Matches("A"));
+  ASSERT_FALSE(entry.Matches("AA"));
+  ASSERT_FALSE(entry.Matches("A:"));
+  ASSERT_FALSE(entry.Matches("AA:"));
+  ASSERT_FALSE(entry.Matches("AA:1"));
+  ASSERT_FALSE(entry.Matches("AA:11"));
+  ASSERT_FALSE(entry.Matches("A:B"));
+  ASSERT_FALSE(entry.Matches("A:1B"));
+  ASSERT_FALSE(entry.Matches("A:B1"));
+  ASSERT_TRUE(entry.Matches("A:1"));
+  ASSERT_TRUE(entry.Matches("A:11"));
+  ASSERT_TRUE(entry.Matches("A:1.1"));
+  ASSERT_TRUE(entry.Matches("A:11.11"));
+  ASSERT_TRUE(entry.Matches("A:1."));
+  ASSERT_TRUE(entry.Matches("A:.1"));
+  ASSERT_TRUE(entry.Matches("A:0.1"));
+  ASSERT_TRUE(entry.Matches("A:1.0"));
+  ASSERT_TRUE(entry.Matches("A:1.0"));
+
+  ASSERT_FALSE(entry.Matches("A:1.0."));
+  ASSERT_FALSE(entry.Matches("A:1.0.2"));
+  ASSERT_FALSE(entry.Matches("A:.1.0"));
+  ASSERT_FALSE(entry.Matches("A:..10"));
+  ASSERT_FALSE(entry.Matches("A:10.."));
+  ASSERT_FALSE(entry.Matches("A:."));
+
+  entry.AddSeparator(":", false);
+  ASSERT_FALSE(entry.Matches("A:1"));
+  ASSERT_FALSE(entry.Matches("A:1.0"));
+
+  ASSERT_TRUE(entry.Matches("A:11:"));
+  ASSERT_TRUE(entry.Matches("A:1.1:"));
+  ASSERT_TRUE(entry.Matches("A:11.11:"));
+  ASSERT_TRUE(entry.Matches("A:1.:"));
+  ASSERT_TRUE(entry.Matches("A:.1:"));
+  ASSERT_TRUE(entry.Matches("A:0.1:"));
+  ASSERT_TRUE(entry.Matches("A:1.0:"));
+  ASSERT_TRUE(entry.Matches("A:1.0:"));
+
+  ASSERT_FALSE(entry.Matches("A:1.0.:"));
+  ASSERT_FALSE(entry.Matches("A:1.0.2:"));
+  ASSERT_FALSE(entry.Matches("A:.1.0:"));
+  ASSERT_FALSE(entry.Matches("A:..10:"));
+  ASSERT_FALSE(entry.Matches("A:10..:"));
+  ASSERT_FALSE(entry.Matches("A:.:"));
+  ASSERT_FALSE(entry.Matches("A::"));
+}
+
+TEST_F(PatternEntryTest, TestIndividualIdEntry) {
+  auto entry = ObjectLibrary::PatternEntry::AsIndividualId("AA");
+  ASSERT_TRUE(entry.Matches("AA"));
+  ASSERT_TRUE(entry.Matches("AA@123#456"));
+  ASSERT_TRUE(entry.Matches("AA@deadbeef#id"));
+
+  ASSERT_FALSE(entry.Matches("A"));
+  ASSERT_FALSE(entry.Matches("AAA"));
+  ASSERT_FALSE(entry.Matches("AA@123"));
+  ASSERT_FALSE(entry.Matches("AA@123#"));
+  ASSERT_FALSE(entry.Matches("AA@#123"));
+}
+
+TEST_F(PatternEntryTest, TestTwoNameEntry) {
+  ObjectLibrary::PatternEntry entry("A");
+  entry.AnotherName("B");
+  ASSERT_TRUE(entry.Matches("A"));
+  ASSERT_TRUE(entry.Matches("B"));
+  ASSERT_FALSE(entry.Matches("AA"));
+  ASSERT_FALSE(entry.Matches("BB"));
+  ASSERT_FALSE(entry.Matches("AA"));
+  ASSERT_FALSE(entry.Matches("BA"));
+  ASSERT_FALSE(entry.Matches("AB"));
+}
+
+TEST_F(PatternEntryTest, TestTwoPatternEntry) {
+  ObjectLibrary::PatternEntry entry("AA", false);
+  entry.AddSeparator(":");
+  entry.AddSeparator(":");
+  ASSERT_FALSE(entry.Matches("AA"));
+  ASSERT_FALSE(entry.Matches("AA:"));
+  ASSERT_FALSE(entry.Matches("AA::"));
+  ASSERT_FALSE(entry.Matches("AA::12"));
+  ASSERT_TRUE(entry.Matches("AA:1:2"));
+  ASSERT_TRUE(entry.Matches("AA:1:2:"));
+
+  ObjectLibrary::PatternEntry entry2("AA", false);
+  entry2.AddSeparator("::");
+  entry2.AddSeparator("##");
+  ASSERT_FALSE(entry2.Matches("AA"));
+  ASSERT_FALSE(entry2.Matches("AA:"));
+  ASSERT_FALSE(entry2.Matches("AA::"));
+  ASSERT_FALSE(entry2.Matches("AA::#"));
+  ASSERT_FALSE(entry2.Matches("AA::##"));
+  ASSERT_FALSE(entry2.Matches("AA##1::2"));
+  ASSERT_FALSE(entry2.Matches("AA::123##"));
+  ASSERT_TRUE(entry2.Matches("AA::1##2"));
+  ASSERT_TRUE(entry2.Matches("AA::12##34:"));
+  ASSERT_TRUE(entry2.Matches("AA::12::34##56"));
+  ASSERT_TRUE(entry2.Matches("AA::12##34::56"));
+}
+
+TEST_F(PatternEntryTest, TestTwoNumbersEntry) {
+  ObjectLibrary::PatternEntry entry("AA", false);
+  entry.AddNumber(":");
+  entry.AddNumber(":");
+  ASSERT_FALSE(entry.Matches("AA"));
+  ASSERT_FALSE(entry.Matches("AA:"));
+  ASSERT_FALSE(entry.Matches("AA::"));
+  ASSERT_FALSE(entry.Matches("AA::12"));
+  ASSERT_FALSE(entry.Matches("AA:1:2:"));
+  ASSERT_TRUE(entry.Matches("AA:1:2"));
+  ASSERT_TRUE(entry.Matches("AA:12:23456"));
+
+  ObjectLibrary::PatternEntry entry2("AA", false);
+  entry2.AddNumber(":");
+  entry2.AddNumber("#");
+  ASSERT_FALSE(entry2.Matches("AA"));
+  ASSERT_FALSE(entry2.Matches("AA:"));
+  ASSERT_FALSE(entry2.Matches("AA:#"));
+  ASSERT_FALSE(entry2.Matches("AA#:"));
+  ASSERT_FALSE(entry2.Matches("AA:123#"));
+  ASSERT_FALSE(entry2.Matches("AA:123#B"));
+  ASSERT_FALSE(entry2.Matches("AA:B#123"));
+  ASSERT_TRUE(entry2.Matches("AA:1#2"));
+  ASSERT_FALSE(entry2.Matches("AA:123#23:"));
+  ASSERT_FALSE(entry2.Matches("AA::12#234"));
+}
+
+TEST_F(PatternEntryTest, TestPatternAndSuffix) {
+  ObjectLibrary::PatternEntry entry("AA", false);
+  entry.AddSeparator("::");
+  entry.AddSuffix("##");
+  ASSERT_FALSE(entry.Matches("AA"));
+  ASSERT_FALSE(entry.Matches("AA::"));
+  ASSERT_FALSE(entry.Matches("AA::##"));
+  ASSERT_FALSE(entry.Matches("AB::1##"));
+  ASSERT_FALSE(entry.Matches("AB::1##2"));
+  ASSERT_FALSE(entry.Matches("AA##1::"));
+  ASSERT_TRUE(entry.Matches("AA::1##"));
+  ASSERT_FALSE(entry.Matches("AA::1###"));
+
+  ObjectLibrary::PatternEntry entry2("AA", false);
+  entry2.AddSuffix("::");
+  entry2.AddSeparator("##");
+  ASSERT_FALSE(entry2.Matches("AA"));
+  ASSERT_FALSE(entry2.Matches("AA::"));
+  ASSERT_FALSE(entry2.Matches("AA::##"));
+  ASSERT_FALSE(entry2.Matches("AB::1##"));
+  ASSERT_FALSE(entry2.Matches("AB::1##2"));
+  ASSERT_TRUE(entry2.Matches("AA::##12"));
+}
+
+TEST_F(PatternEntryTest, TestTwoNamesAndPattern) {
+  ObjectLibrary::PatternEntry entry("AA", true);
+  entry.AddSeparator("::");
+  entry.AnotherName("BBB");
+  ASSERT_TRUE(entry.Matches("AA"));
+  ASSERT_TRUE(entry.Matches("AA::1"));
+  ASSERT_TRUE(entry.Matches("BBB"));
+  ASSERT_TRUE(entry.Matches("BBB::2"));
+
+  ASSERT_FALSE(entry.Matches("AA::"));
+  ASSERT_FALSE(entry.Matches("AAA::"));
+  ASSERT_FALSE(entry.Matches("BBB::"));
+
+  entry.SetOptional(false);
+  ASSERT_FALSE(entry.Matches("AA"));
+  ASSERT_FALSE(entry.Matches("BBB"));
+
+  ASSERT_FALSE(entry.Matches("AA::"));
+  ASSERT_FALSE(entry.Matches("AAA::"));
+  ASSERT_FALSE(entry.Matches("BBB::"));
+
+  ASSERT_TRUE(entry.Matches("AA::1"));
+  ASSERT_TRUE(entry.Matches("BBB::2"));
+}
+
+}  // namespace ROCKSDB_NAMESPACE
+
+int main(int argc, char** argv) {
+  ROCKSDB_NAMESPACE::port::InstallStackTraceHandler();
+  ::testing::InitGoogleTest(&argc, argv);
+  return RUN_ALL_TESTS();
+}
+
+#else  // ROCKSDB_LITE
+#include <stdio.h>
+
+int main(int /*argc*/, char** /*argv*/) {
+  fprintf(stderr, "SKIPPED as ObjRegistry is not supported in ROCKSDB_LITE\n");
+  return 0;
+}
+
+#endif  // ROCKSDB_LITE
diff --git a/src/rocksdb/utilities/option_change_migration/option_change_migration.cc b/src/rocksdb/utilities/option_change_migration/option_change_migration.cc
new file mode 100644
index 000000000..e93d2152d
--- /dev/null
+++ b/src/rocksdb/utilities/option_change_migration/option_change_migration.cc
@@ -0,0 +1,186 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#include "rocksdb/utilities/option_change_migration.h"
+
+#ifndef ROCKSDB_LITE
+#include "rocksdb/db.h"
+
+namespace ROCKSDB_NAMESPACE {
+namespace {
+// Return a version of Options `opts` that allow us to open/write into a DB
+// without triggering an automatic compaction or stalling. This is guaranteed
+// by disabling automatic compactions and using huge values for stalling
+// triggers.
+Options GetNoCompactionOptions(const Options& opts) {
+  Options ret_opts = opts;
+  ret_opts.disable_auto_compactions = true;
+  ret_opts.level0_slowdown_writes_trigger = 999999;
+  ret_opts.level0_stop_writes_trigger = 999999;
+  ret_opts.soft_pending_compaction_bytes_limit = 0;
+  ret_opts.hard_pending_compaction_bytes_limit = 0;
+  return ret_opts;
+}
+
+Status OpenDb(const Options& options, const std::string& dbname,
+              std::unique_ptr<DB>* db) {
+  db->reset();
+  DB* tmpdb;
+  Status s = DB::Open(options, dbname, &tmpdb);
+  if (s.ok()) {
+    db->reset(tmpdb);
+  }
+  return s;
+}
+
+// l0_file_size specifies size of file on L0. Files will be range partitioned
+// after a full compaction so they are likely qualified to put on L0. If
+// left as 0, the files are compacted in a single file and put to L0. Otherwise,
+// will try to compact the files as size l0_file_size.
+Status CompactToLevel(const Options& options, const std::string& dbname,
+                      int dest_level, uint64_t l0_file_size, bool need_reopen) {
+  std::unique_ptr<DB> db;
+  Options no_compact_opts = GetNoCompactionOptions(options);
+  if (dest_level == 0) {
+    if (l0_file_size == 0) {
+      // Single file.
+      l0_file_size = 999999999999999;
+    }
+    // L0 has strict sequenceID requirements to files to it. It's safer
+    // to only put one compacted file to there.
+    // This is only used for converting to universal compaction with
+    // only one level. In this case, compacting to one file is also
+    // optimal.
+    no_compact_opts.target_file_size_base = l0_file_size;
+    no_compact_opts.max_compaction_bytes = l0_file_size;
+  }
+  Status s = OpenDb(no_compact_opts, dbname, &db);
+  if (!s.ok()) {
+    return s;
+  }
+  CompactRangeOptions cro;
+  cro.change_level = true;
+  cro.target_level = dest_level;
+  if (dest_level == 0) {
+    // cannot use kForceOptimized because the compaction is expected to
+    // generate one output file
+    cro.bottommost_level_compaction = BottommostLevelCompaction::kForce;
+  }
+  s = db->CompactRange(cro, nullptr, nullptr);
+
+  if (s.ok() && need_reopen) {
+    // Need to restart DB to rewrite the manifest file.
+    // In order to open a DB with specific num_levels, the manifest file should
+    // contain no record that mentiones any level beyond num_levels. Issuing a
+    // full compaction will move all the data to a level not exceeding
+    // num_levels, but the manifest may still contain previous record mentioning
+    // a higher level. Reopening the DB will force the manifest to be rewritten
+    // so that those records will be cleared.
+    db.reset();
+    s = OpenDb(no_compact_opts, dbname, &db);
+  }
+  return s;
+}
+
+Status MigrateToUniversal(std::string dbname, const Options& old_opts,
+                          const Options& new_opts) {
+  if (old_opts.num_levels <= new_opts.num_levels ||
+      old_opts.compaction_style == CompactionStyle::kCompactionStyleFIFO) {
+    return Status::OK();
+  } else {
+    bool need_compact = false;
+    {
+      std::unique_ptr<DB> db;
+      Options opts = GetNoCompactionOptions(old_opts);
+      Status s = OpenDb(opts, dbname, &db);
+      if (!s.ok()) {
+        return s;
+      }
+      ColumnFamilyMetaData metadata;
+      db->GetColumnFamilyMetaData(&metadata);
+      if (!metadata.levels.empty() &&
+          metadata.levels.back().level >= new_opts.num_levels) {
+        need_compact = true;
+      }
+    }
+    if (need_compact) {
+      return CompactToLevel(old_opts, dbname, new_opts.num_levels - 1,
+                            /*l0_file_size=*/0, true);
+    }
+    return Status::OK();
+  }
+}
+
+Status MigrateToLevelBase(std::string dbname, const Options& old_opts,
+                          const Options& new_opts) {
+  if (!new_opts.level_compaction_dynamic_level_bytes) {
+    if (old_opts.num_levels == 1) {
+      return Status::OK();
+    }
+    // Compact everything to level 1 to guarantee it can be safely opened.
+    Options opts = old_opts;
+    opts.target_file_size_base = new_opts.target_file_size_base;
+    // Although sometimes we can open the DB with the new option without error,
+    // We still want to compact the files to avoid the LSM tree to stuck
+    // in bad shape. For example, if the user changed the level size
+    // multiplier from 4 to 8, with the same data, we will have fewer
+    // levels. Unless we issue a full comaction, the LSM tree may stuck
+    // with more levels than needed and it won't recover automatically.
+    return CompactToLevel(opts, dbname, 1, /*l0_file_size=*/0, true);
+  } else {
+    // Compact everything to the last level to guarantee it can be safely
+    // opened.
+    if (old_opts.num_levels == 1) {
+      return Status::OK();
+    } else if (new_opts.num_levels > old_opts.num_levels) {
+      // Dynamic level mode requires data to be put in the last level first.
+      return CompactToLevel(new_opts, dbname, new_opts.num_levels - 1,
+                            /*l0_file_size=*/0, false);
+    } else {
+      Options opts = old_opts;
+      opts.target_file_size_base = new_opts.target_file_size_base;
+      return CompactToLevel(opts, dbname, new_opts.num_levels - 1,
+                            /*l0_file_size=*/0, true);
+    }
+  }
+}
+}  // namespace
+
+Status OptionChangeMigration(std::string dbname, const Options& old_opts,
+                             const Options& new_opts) {
+  if (old_opts.compaction_style == CompactionStyle::kCompactionStyleFIFO) {
+    // LSM generated by FIFO compaction can be opened by any compaction.
+    return Status::OK();
+  } else if (new_opts.compaction_style ==
+             CompactionStyle::kCompactionStyleUniversal) {
+    return MigrateToUniversal(dbname, old_opts, new_opts);
+  } else if (new_opts.compaction_style ==
+             CompactionStyle::kCompactionStyleLevel) {
+    return MigrateToLevelBase(dbname, old_opts, new_opts);
+  } else if (new_opts.compaction_style ==
+             CompactionStyle::kCompactionStyleFIFO) {
+    uint64_t l0_file_size = 0;
+    if (new_opts.compaction_options_fifo.max_table_files_size > 0) {
+      // Create at least 8 files when max_table_files_size hits, so that the DB
+      // doesn't just disappear. This in fact violates the FIFO condition, but
+      // otherwise, the migrated DB is unlikley to be usable.
+      l0_file_size = new_opts.compaction_options_fifo.max_table_files_size / 8;
+    }
+    return CompactToLevel(old_opts, dbname, 0, l0_file_size, true);
+  } else {
+    return Status::NotSupported(
+        "Do not how to migrate to this compaction style");
+  }
+}
+}  // namespace ROCKSDB_NAMESPACE
+#else
+namespace ROCKSDB_NAMESPACE {
+Status OptionChangeMigration(std::string /*dbname*/,
+                             const Options& /*old_opts*/,
+                             const Options& /*new_opts*/) {
+  return Status::NotSupported();
+}
+}  // namespace ROCKSDB_NAMESPACE
+#endif  // ROCKSDB_LITE
diff --git a/src/rocksdb/utilities/option_change_migration/option_change_migration_test.cc b/src/rocksdb/utilities/option_change_migration/option_change_migration_test.cc
new file mode 100644
index 000000000..71af45db1
--- /dev/null
+++ b/src/rocksdb/utilities/option_change_migration/option_change_migration_test.cc
@@ -0,0 +1,550 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+//
+// Copyright (c) 2011 The LevelDB Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file. See the AUTHORS file for names of contributors.
+
+#include "rocksdb/utilities/option_change_migration.h"
+
+#include <set>
+
+#include "db/db_test_util.h"
+#include "port/stack_trace.h"
+#include "util/random.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+class DBOptionChangeMigrationTests
+    : public DBTestBase,
+      public testing::WithParamInterface<
+          std::tuple<int, int, bool, int, int, bool, uint64_t>> {
+ public:
+  DBOptionChangeMigrationTests()
+      : DBTestBase("db_option_change_migration_test", /*env_do_fsync=*/true) {
+    level1_ = std::get<0>(GetParam());
+    compaction_style1_ = std::get<1>(GetParam());
+    is_dynamic1_ = std::get<2>(GetParam());
+
+    level2_ = std::get<3>(GetParam());
+    compaction_style2_ = std::get<4>(GetParam());
+    is_dynamic2_ = std::get<5>(GetParam());
+    fifo_max_table_files_size_ = std::get<6>(GetParam());
+  }
+
+  // Required if inheriting from testing::WithParamInterface<>
+  static void SetUpTestCase() {}
+  static void TearDownTestCase() {}
+
+  int level1_;
+  int compaction_style1_;
+  bool is_dynamic1_;
+
+  int level2_;
+  int compaction_style2_;
+  bool is_dynamic2_;
+
+  uint64_t fifo_max_table_files_size_;
+};
+
+#ifndef ROCKSDB_LITE
+TEST_P(DBOptionChangeMigrationTests, Migrate1) {
+  Options old_options = CurrentOptions();
+  old_options.compaction_style =
+      static_cast<CompactionStyle>(compaction_style1_);
+  if (old_options.compaction_style == CompactionStyle::kCompactionStyleLevel) {
+    old_options.level_compaction_dynamic_level_bytes = is_dynamic1_;
+  }
+  if (old_options.compaction_style == CompactionStyle::kCompactionStyleFIFO) {
+    old_options.max_open_files = -1;
+  }
+  old_options.level0_file_num_compaction_trigger = 3;
+  old_options.write_buffer_size = 64 * 1024;
+  old_options.target_file_size_base = 128 * 1024;
+  // Make level target of L1, L2 to be 200KB and 600KB
+  old_options.num_levels = level1_;
+  old_options.max_bytes_for_level_multiplier = 3;
+  old_options.max_bytes_for_level_base = 200 * 1024;
+
+  Reopen(old_options);
+
+  Random rnd(301);
+  int key_idx = 0;
+
+  // Generate at least 2MB of data
+  for (int num = 0; num < 20; num++) {
+    GenerateNewFile(&rnd, &key_idx);
+  }
+  ASSERT_OK(dbfull()->TEST_WaitForFlushMemTable());
+  ASSERT_OK(dbfull()->TEST_WaitForCompact());
+
+  // Will make sure exactly those keys are in the DB after migration.
+  std::set<std::string> keys;
+  {
+    std::unique_ptr<Iterator> it(db_->NewIterator(ReadOptions()));
+    it->SeekToFirst();
+    for (; it->Valid(); it->Next()) {
+      keys.insert(it->key().ToString());
+    }
+  }
+  Close();
+
+  Options new_options = old_options;
+  new_options.compaction_style =
+      static_cast<CompactionStyle>(compaction_style2_);
+  if (new_options.compaction_style == CompactionStyle::kCompactionStyleLevel) {
+    new_options.level_compaction_dynamic_level_bytes = is_dynamic2_;
+  }
+  if (new_options.compaction_style == CompactionStyle::kCompactionStyleFIFO) {
+    new_options.max_open_files = -1;
+  }
+  if (fifo_max_table_files_size_ != 0) {
+    new_options.compaction_options_fifo.max_table_files_size =
+        fifo_max_table_files_size_;
+  }
+  new_options.target_file_size_base = 256 * 1024;
+  new_options.num_levels = level2_;
+  new_options.max_bytes_for_level_base = 150 * 1024;
+  new_options.max_bytes_for_level_multiplier = 4;
+  ASSERT_OK(OptionChangeMigration(dbname_, old_options, new_options));
+  Reopen(new_options);
+
+  // Wait for compaction to finish and make sure it can reopen
+  ASSERT_OK(dbfull()->TEST_WaitForFlushMemTable());
+  ASSERT_OK(dbfull()->TEST_WaitForCompact());
+  Reopen(new_options);
+
+  {
+    std::unique_ptr<Iterator> it(db_->NewIterator(ReadOptions()));
+    it->SeekToFirst();
+    for (std::string key : keys) {
+      ASSERT_TRUE(it->Valid());
+      ASSERT_EQ(key, it->key().ToString());
+      it->Next();
+    }
+    ASSERT_TRUE(!it->Valid());
+  }
+}
+
+TEST_P(DBOptionChangeMigrationTests, Migrate2) {
+  Options old_options = CurrentOptions();
+  old_options.compaction_style =
+      static_cast<CompactionStyle>(compaction_style2_);
+  if (old_options.compaction_style == CompactionStyle::kCompactionStyleLevel) {
+    old_options.level_compaction_dynamic_level_bytes = is_dynamic2_;
+  }
+  if (old_options.compaction_style == CompactionStyle::kCompactionStyleFIFO) {
+    old_options.max_open_files = -1;
+  }
+  old_options.level0_file_num_compaction_trigger = 3;
+  old_options.write_buffer_size = 64 * 1024;
+  old_options.target_file_size_base = 128 * 1024;
+  // Make level target of L1, L2 to be 200KB and 600KB
+  old_options.num_levels = level2_;
+  old_options.max_bytes_for_level_multiplier = 3;
+  old_options.max_bytes_for_level_base = 200 * 1024;
+
+  Reopen(old_options);
+
+  Random rnd(301);
+  int key_idx = 0;
+
+  // Generate at least 2MB of data
+  for (int num = 0; num < 20; num++) {
+    GenerateNewFile(&rnd, &key_idx);
+  }
+  ASSERT_OK(dbfull()->TEST_WaitForFlushMemTable());
+  ASSERT_OK(dbfull()->TEST_WaitForCompact());
+
+  // Will make sure exactly those keys are in the DB after migration.
+  std::set<std::string> keys;
+  {
+    std::unique_ptr<Iterator> it(db_->NewIterator(ReadOptions()));
+    it->SeekToFirst();
+    for (; it->Valid(); it->Next()) {
+      keys.insert(it->key().ToString());
+    }
+  }
+
+  Close();
+
+  Options new_options = old_options;
+  new_options.compaction_style =
+      static_cast<CompactionStyle>(compaction_style1_);
+  if (new_options.compaction_style == CompactionStyle::kCompactionStyleLevel) {
+    new_options.level_compaction_dynamic_level_bytes = is_dynamic1_;
+  }
+  if (new_options.compaction_style == CompactionStyle::kCompactionStyleFIFO) {
+    new_options.max_open_files = -1;
+  }
+  if (fifo_max_table_files_size_ != 0) {
+    new_options.compaction_options_fifo.max_table_files_size =
+        fifo_max_table_files_size_;
+  }
+  new_options.target_file_size_base = 256 * 1024;
+  new_options.num_levels = level1_;
+  new_options.max_bytes_for_level_base = 150 * 1024;
+  new_options.max_bytes_for_level_multiplier = 4;
+  ASSERT_OK(OptionChangeMigration(dbname_, old_options, new_options));
+  Reopen(new_options);
+  // Wait for compaction to finish and make sure it can reopen
+  ASSERT_OK(dbfull()->TEST_WaitForFlushMemTable());
+  ASSERT_OK(dbfull()->TEST_WaitForCompact());
+  Reopen(new_options);
+
+  {
+    std::unique_ptr<Iterator> it(db_->NewIterator(ReadOptions()));
+    it->SeekToFirst();
+    for (std::string key : keys) {
+      ASSERT_TRUE(it->Valid());
+      ASSERT_EQ(key, it->key().ToString());
+      it->Next();
+    }
+    ASSERT_TRUE(!it->Valid());
+  }
+}
+
+TEST_P(DBOptionChangeMigrationTests, Migrate3) {
+  Options old_options = CurrentOptions();
+  old_options.compaction_style =
+      static_cast<CompactionStyle>(compaction_style1_);
+  if (old_options.compaction_style == CompactionStyle::kCompactionStyleLevel) {
+    old_options.level_compaction_dynamic_level_bytes = is_dynamic1_;
+  }
+  if (old_options.compaction_style == CompactionStyle::kCompactionStyleFIFO) {
+    old_options.max_open_files = -1;
+  }
+  old_options.level0_file_num_compaction_trigger = 3;
+  old_options.write_buffer_size = 64 * 1024;
+  old_options.target_file_size_base = 128 * 1024;
+  // Make level target of L1, L2 to be 200KB and 600KB
+  old_options.num_levels = level1_;
+  old_options.max_bytes_for_level_multiplier = 3;
+  old_options.max_bytes_for_level_base = 200 * 1024;
+
+  Reopen(old_options);
+  Random rnd(301);
+  for (int num = 0; num < 20; num++) {
+    for (int i = 0; i < 50; i++) {
+      ASSERT_OK(Put(Key(num * 100 + i), rnd.RandomString(900)));
+    }
+    Flush();
+    ASSERT_OK(dbfull()->TEST_WaitForCompact());
+    if (num == 9) {
+      // Issue a full compaction to generate some zero-out files
+      CompactRangeOptions cro;
+      cro.bottommost_level_compaction = BottommostLevelCompaction::kForce;
+      ASSERT_OK(dbfull()->CompactRange(cro, nullptr, nullptr));
+    }
+  }
+  ASSERT_OK(dbfull()->TEST_WaitForFlushMemTable());
+  ASSERT_OK(dbfull()->TEST_WaitForCompact());
+
+  // Will make sure exactly those keys are in the DB after migration.
+  std::set<std::string> keys;
+  {
+    std::unique_ptr<Iterator> it(db_->NewIterator(ReadOptions()));
+    it->SeekToFirst();
+    for (; it->Valid(); it->Next()) {
+      keys.insert(it->key().ToString());
+    }
+  }
+  Close();
+
+  Options new_options = old_options;
+  new_options.compaction_style =
+      static_cast<CompactionStyle>(compaction_style2_);
+  if (new_options.compaction_style == CompactionStyle::kCompactionStyleLevel) {
+    new_options.level_compaction_dynamic_level_bytes = is_dynamic2_;
+  }
+  if (new_options.compaction_style == CompactionStyle::kCompactionStyleFIFO) {
+    new_options.max_open_files = -1;
+  }
+  if (fifo_max_table_files_size_ != 0) {
+    new_options.compaction_options_fifo.max_table_files_size =
+        fifo_max_table_files_size_;
+  }
+  new_options.target_file_size_base = 256 * 1024;
+  new_options.num_levels = level2_;
+  new_options.max_bytes_for_level_base = 150 * 1024;
+  new_options.max_bytes_for_level_multiplier = 4;
+  ASSERT_OK(OptionChangeMigration(dbname_, old_options, new_options));
+  Reopen(new_options);
+
+  // Wait for compaction to finish and make sure it can reopen
+  ASSERT_OK(dbfull()->TEST_WaitForFlushMemTable());
+  ASSERT_OK(dbfull()->TEST_WaitForCompact());
+  Reopen(new_options);
+
+  {
+    std::unique_ptr<Iterator> it(db_->NewIterator(ReadOptions()));
+    it->SeekToFirst();
+    for (std::string key : keys) {
+      ASSERT_TRUE(it->Valid());
+      ASSERT_EQ(key, it->key().ToString());
+      it->Next();
+    }
+    ASSERT_TRUE(!it->Valid());
+  }
+}
+
+TEST_P(DBOptionChangeMigrationTests, Migrate4) {
+  Options old_options = CurrentOptions();
+  old_options.compaction_style =
+      static_cast<CompactionStyle>(compaction_style2_);
+  if (old_options.compaction_style == CompactionStyle::kCompactionStyleLevel) {
+    old_options.level_compaction_dynamic_level_bytes = is_dynamic2_;
+  }
+  if (old_options.compaction_style == CompactionStyle::kCompactionStyleFIFO) {
+    old_options.max_open_files = -1;
+  }
+  old_options.level0_file_num_compaction_trigger = 3;
+  old_options.write_buffer_size = 64 * 1024;
+  old_options.target_file_size_base = 128 * 1024;
+  // Make level target of L1, L2 to be 200KB and 600KB
+  old_options.num_levels = level2_;
+  old_options.max_bytes_for_level_multiplier = 3;
+  old_options.max_bytes_for_level_base = 200 * 1024;
+
+  Reopen(old_options);
+  Random rnd(301);
+  for (int num = 0; num < 20; num++) {
+    for (int i = 0; i < 50; i++) {
+      ASSERT_OK(Put(Key(num * 100 + i), rnd.RandomString(900)));
+    }
+    Flush();
+    ASSERT_OK(dbfull()->TEST_WaitForCompact());
+    if (num == 9) {
+      // Issue a full compaction to generate some zero-out files
+      CompactRangeOptions cro;
+      cro.bottommost_level_compaction = BottommostLevelCompaction::kForce;
+      ASSERT_OK(dbfull()->CompactRange(cro, nullptr, nullptr));
+    }
+  }
+  ASSERT_OK(dbfull()->TEST_WaitForFlushMemTable());
+  ASSERT_OK(dbfull()->TEST_WaitForCompact());
+
+  // Will make sure exactly those keys are in the DB after migration.
+  std::set<std::string> keys;
+  {
+    std::unique_ptr<Iterator> it(db_->NewIterator(ReadOptions()));
+    it->SeekToFirst();
+    for (; it->Valid(); it->Next()) {
+      keys.insert(it->key().ToString());
+    }
+  }
+
+  Close();
+
+  Options new_options = old_options;
+  new_options.compaction_style =
+      static_cast<CompactionStyle>(compaction_style1_);
+  if (new_options.compaction_style == CompactionStyle::kCompactionStyleLevel) {
+    new_options.level_compaction_dynamic_level_bytes = is_dynamic1_;
+  }
+  if (new_options.compaction_style == CompactionStyle::kCompactionStyleFIFO) {
+    new_options.max_open_files = -1;
+  }
+  if (fifo_max_table_files_size_ != 0) {
+    new_options.compaction_options_fifo.max_table_files_size =
+        fifo_max_table_files_size_;
+  }
+  new_options.target_file_size_base = 256 * 1024;
+  new_options.num_levels = level1_;
+  new_options.max_bytes_for_level_base = 150 * 1024;
+  new_options.max_bytes_for_level_multiplier = 4;
+  ASSERT_OK(OptionChangeMigration(dbname_, old_options, new_options));
+  Reopen(new_options);
+  // Wait for compaction to finish and make sure it can reopen
+  ASSERT_OK(dbfull()->TEST_WaitForFlushMemTable());
+  ASSERT_OK(dbfull()->TEST_WaitForCompact());
+  Reopen(new_options);
+
+  {
+    std::unique_ptr<Iterator> it(db_->NewIterator(ReadOptions()));
+    it->SeekToFirst();
+    for (std::string key : keys) {
+      ASSERT_TRUE(it->Valid());
+      ASSERT_EQ(key, it->key().ToString());
+      it->Next();
+    }
+    ASSERT_TRUE(!it->Valid());
+  }
+}
+
+INSTANTIATE_TEST_CASE_P(
+    DBOptionChangeMigrationTests, DBOptionChangeMigrationTests,
+    ::testing::Values(
+        std::make_tuple(3 /* old num_levels */, 0 /* old compaction style */,
+                        false /* is dynamic leveling in old option */,
+                        4 /* old num_levels */, 0 /* new compaction style */,
+                        false /* is dynamic leveling in new option */,
+                        0 /*fifo max_table_files_size*/),
+        std::make_tuple(3 /* old num_levels */, 0 /* old compaction style */,
+                        true /* is dynamic leveling in old option */,
+                        4 /* old num_levels */, 0 /* new compaction style */,
+                        true /* is dynamic leveling in new option */,
+                        0 /*fifo max_table_files_size*/),
+        std::make_tuple(3 /* old num_levels */, 0 /* old compaction style */,
+                        true /* is dynamic leveling in old option */,
+                        4 /* old num_levels */, 0 /* new compaction style */,
+                        false, 0 /*fifo max_table_files_size*/),
+        std::make_tuple(3 /* old num_levels */, 0 /* old compaction style */,
+                        false /* is dynamic leveling in old option */,
+                        4 /* old num_levels */, 0 /* new compaction style */,
+                        true /* is dynamic leveling in new option */,
+                        0 /*fifo max_table_files_size*/),
+        std::make_tuple(3 /* old num_levels */, 1 /* old compaction style */,
+                        false /* is dynamic leveling in old option */,
+                        4 /* old num_levels */, 1 /* new compaction style */,
+                        false /* is dynamic leveling in new option */,
+                        0 /*fifo max_table_files_size*/),
+        std::make_tuple(1 /* old num_levels */, 1 /* old compaction style */,
+                        false /* is dynamic leveling in old option */,
+                        4 /* old num_levels */, 1 /* new compaction style */,
+                        false /* is dynamic leveling in new option */,
+                        0 /*fifo max_table_files_size*/),
+        std::make_tuple(3 /* old num_levels */, 0 /* old compaction style */,
+                        false /* is dynamic leveling in old option */,
+                        4 /* old num_levels */, 1 /* new compaction style */,
+                        false /* is dynamic leveling in new option */,
+                        0 /*fifo max_table_files_size*/),
+        std::make_tuple(3 /* old num_levels */, 0 /* old compaction style */,
+                        false /* is dynamic leveling in old option */,
+                        1 /* old num_levels */, 1 /* new compaction style */,
+                        false /* is dynamic leveling in new option */,
+                        0 /*fifo max_table_files_size*/),
+        std::make_tuple(3 /* old num_levels */, 0 /* old compaction style */,
+                        true /* is dynamic leveling in old option */,
+                        4 /* old num_levels */, 1 /* new compaction style */,
+                        false /* is dynamic leveling in new option */,
+                        0 /*fifo max_table_files_size*/),
+        std::make_tuple(3 /* old num_levels */, 0 /* old compaction style */,
+                        true /* is dynamic leveling in old option */,
+                        1 /* old num_levels */, 1 /* new compaction style */,
+                        false /* is dynamic leveling in new option */,
+                        0 /*fifo max_table_files_size*/),
+        std::make_tuple(1 /* old num_levels */, 1 /* old compaction style */,
+                        false /* is dynamic leveling in old option */,
+                        4 /* old num_levels */, 0 /* new compaction style */,
+                        false /* is dynamic leveling in new option */,
+                        0 /*fifo max_table_files_size*/),
+        std::make_tuple(4 /* old num_levels */, 0 /* old compaction style */,
+                        false /* is dynamic leveling in old option */,
+                        1 /* old num_levels */, 2 /* new compaction style */,
+                        false /* is dynamic leveling in new option */,
+                        0 /*fifo max_table_files_size*/),
+        std::make_tuple(3 /* old num_levels */, 0 /* old compaction style */,
+                        true /* is dynamic leveling in old option */,
+                        2 /* old num_levels */, 2 /* new compaction style */,
+                        false /* is dynamic leveling in new option */,
+                        0 /*fifo max_table_files_size*/),
+        std::make_tuple(3 /* old num_levels */, 1 /* old compaction style */,
+                        false /* is dynamic leveling in old option */,
+                        3 /* old num_levels */, 2 /* new compaction style */,
+                        false /* is dynamic leveling in new option */,
+                        0 /*fifo max_table_files_size*/),
+        std::make_tuple(1 /* old num_levels */, 1 /* old compaction style */,
+                        false /* is dynamic leveling in old option */,
+                        4 /* old num_levels */, 2 /* new compaction style */,
+                        false /* is dynamic leveling in new option */, 0),
+        std::make_tuple(4 /* old num_levels */, 0 /* old compaction style */,
+                        false /* is dynamic leveling in old option */,
+                        1 /* old num_levels */, 2 /* new compaction style */,
+                        false /* is dynamic leveling in new option */,
+                        5 * 1024 * 1024 /*fifo max_table_files_size*/),
+        std::make_tuple(3 /* old num_levels */, 0 /* old compaction style */,
+                        true /* is dynamic leveling in old option */,
+                        2 /* old num_levels */, 2 /* new compaction style */,
+                        false /* is dynamic leveling in new option */,
+                        5 * 1024 * 1024 /*fifo max_table_files_size*/),
+        std::make_tuple(3 /* old num_levels */, 1 /* old compaction style */,
+                        false /* is dynamic leveling in old option */,
+                        3 /* old num_levels */, 2 /* new compaction style */,
+                        false /* is dynamic leveling in new option */,
+                        5 * 1024 * 1024 /*fifo max_table_files_size*/),
+        std::make_tuple(1 /* old num_levels */, 1 /* old compaction style */,
+                        false /* is dynamic leveling in old option */,
+                        4 /* old num_levels */, 2 /* new compaction style */,
+                        false /* is dynamic leveling in new option */,
+                        5 * 1024 * 1024 /*fifo max_table_files_size*/)));
+
+class DBOptionChangeMigrationTest : public DBTestBase {
+ public:
+  DBOptionChangeMigrationTest()
+      : DBTestBase("db_option_change_migration_test2", /*env_do_fsync=*/true) {}
+};
+
+TEST_F(DBOptionChangeMigrationTest, CompactedSrcToUniversal) {
+  Options old_options = CurrentOptions();
+  old_options.compaction_style = CompactionStyle::kCompactionStyleLevel;
+  old_options.max_compaction_bytes = 200 * 1024;
+  old_options.level_compaction_dynamic_level_bytes = false;
+  old_options.level0_file_num_compaction_trigger = 3;
+  old_options.write_buffer_size = 64 * 1024;
+  old_options.target_file_size_base = 128 * 1024;
+  // Make level target of L1, L2 to be 200KB and 600KB
+  old_options.num_levels = 4;
+  old_options.max_bytes_for_level_multiplier = 3;
+  old_options.max_bytes_for_level_base = 200 * 1024;
+
+  Reopen(old_options);
+  Random rnd(301);
+  for (int num = 0; num < 20; num++) {
+    for (int i = 0; i < 50; i++) {
+      ASSERT_OK(Put(Key(num * 100 + i), rnd.RandomString(900)));
+    }
+  }
+  Flush();
+  CompactRangeOptions cro;
+  cro.bottommost_level_compaction = BottommostLevelCompaction::kForce;
+  ASSERT_OK(dbfull()->CompactRange(cro, nullptr, nullptr));
+
+  // Will make sure exactly those keys are in the DB after migration.
+  std::set<std::string> keys;
+  {
+    std::unique_ptr<Iterator> it(db_->NewIterator(ReadOptions()));
+    it->SeekToFirst();
+    for (; it->Valid(); it->Next()) {
+      keys.insert(it->key().ToString());
+    }
+  }
+
+  Close();
+
+  Options new_options = old_options;
+  new_options.compaction_style = CompactionStyle::kCompactionStyleUniversal;
+  new_options.target_file_size_base = 256 * 1024;
+  new_options.num_levels = 1;
+  new_options.max_bytes_for_level_base = 150 * 1024;
+  new_options.max_bytes_for_level_multiplier = 4;
+  ASSERT_OK(OptionChangeMigration(dbname_, old_options, new_options));
+  Reopen(new_options);
+  // Wait for compaction to finish and make sure it can reopen
+  ASSERT_OK(dbfull()->TEST_WaitForFlushMemTable());
+  ASSERT_OK(dbfull()->TEST_WaitForCompact());
+  Reopen(new_options);
+
+  {
+    std::unique_ptr<Iterator> it(db_->NewIterator(ReadOptions()));
+    it->SeekToFirst();
+    for (std::string key : keys) {
+      ASSERT_TRUE(it->Valid());
+      ASSERT_EQ(key, it->key().ToString());
+      it->Next();
+    }
+    ASSERT_TRUE(!it->Valid());
+    ASSERT_OK(it->status());
+  }
+}
+
+#endif  // ROCKSDB_LITE
+}  // namespace ROCKSDB_NAMESPACE
+
+int main(int argc, char** argv) {
+  ROCKSDB_NAMESPACE::port::InstallStackTraceHandler();
+  ::testing::InitGoogleTest(&argc, argv);
+  return RUN_ALL_TESTS();
+}
diff --git a/src/rocksdb/utilities/options/options_util.cc b/src/rocksdb/utilities/options/options_util.cc
new file mode 100644
index 000000000..00c4b981a
--- /dev/null
+++ b/src/rocksdb/utilities/options/options_util.cc
@@ -0,0 +1,159 @@
+// Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#ifndef ROCKSDB_LITE
+
+#include "rocksdb/utilities/options_util.h"
+
+#include "file/filename.h"
+#include "options/options_parser.h"
+#include "rocksdb/convenience.h"
+#include "rocksdb/options.h"
+#include "table/block_based/block_based_table_factory.h"
+
+namespace ROCKSDB_NAMESPACE {
+Status LoadOptionsFromFile(const std::string& file_name, Env* env,
+                           DBOptions* db_options,
+                           std::vector<ColumnFamilyDescriptor>* cf_descs,
+                           bool ignore_unknown_options,
+                           std::shared_ptr<Cache>* cache) {
+  ConfigOptions config_options;
+  config_options.ignore_unknown_options = ignore_unknown_options;
+  config_options.input_strings_escaped = true;
+  config_options.env = env;
+
+  return LoadOptionsFromFile(config_options, file_name, db_options, cf_descs,
+                             cache);
+}
+
+Status LoadOptionsFromFile(const ConfigOptions& config_options,
+                           const std::string& file_name, DBOptions* db_options,
+                           std::vector<ColumnFamilyDescriptor>* cf_descs,
+                           std::shared_ptr<Cache>* cache) {
+  RocksDBOptionsParser parser;
+  const auto& fs = config_options.env->GetFileSystem();
+  Status s = parser.Parse(config_options, file_name, fs.get());
+  if (!s.ok()) {
+    return s;
+  }
+  *db_options = *parser.db_opt();
+  const std::vector<std::string>& cf_names = *parser.cf_names();
+  const std::vector<ColumnFamilyOptions>& cf_opts = *parser.cf_opts();
+  cf_descs->clear();
+  for (size_t i = 0; i < cf_opts.size(); ++i) {
+    cf_descs->push_back({cf_names[i], cf_opts[i]});
+    if (cache != nullptr) {
+      TableFactory* tf = cf_opts[i].table_factory.get();
+      if (tf != nullptr) {
+        auto* opts = tf->GetOptions<BlockBasedTableOptions>();
+        if (opts != nullptr) {
+          opts->block_cache = *cache;
+        }
+      }
+    }
+  }
+  return Status::OK();
+}
+
+Status GetLatestOptionsFileName(const std::string& dbpath, Env* env,
+                                std::string* options_file_name) {
+  Status s;
+  std::string latest_file_name;
+  uint64_t latest_time_stamp = 0;
+  std::vector<std::string> file_names;
+  s = env->GetChildren(dbpath, &file_names);
+  if (s.IsNotFound()) {
+    return Status::NotFound(Status::kPathNotFound,
+                            "No options files found in the DB directory.",
+                            dbpath);
+  } else if (!s.ok()) {
+    return s;
+  }
+  for (auto& file_name : file_names) {
+    uint64_t time_stamp;
+    FileType type;
+    if (ParseFileName(file_name, &time_stamp, &type) && type == kOptionsFile) {
+      if (time_stamp > latest_time_stamp) {
+        latest_time_stamp = time_stamp;
+        latest_file_name = file_name;
+      }
+    }
+  }
+  if (latest_file_name.size() == 0) {
+    return Status::NotFound(Status::kPathNotFound,
+                            "No options files found in the DB directory.",
+                            dbpath);
+  }
+  *options_file_name = latest_file_name;
+  return Status::OK();
+}
+
+Status LoadLatestOptions(const std::string& dbpath, Env* env,
+                         DBOptions* db_options,
+                         std::vector<ColumnFamilyDescriptor>* cf_descs,
+                         bool ignore_unknown_options,
+                         std::shared_ptr<Cache>* cache) {
+  ConfigOptions config_options;
+  config_options.ignore_unknown_options = ignore_unknown_options;
+  config_options.input_strings_escaped = true;
+  config_options.env = env;
+
+  return LoadLatestOptions(config_options, dbpath, db_options, cf_descs, cache);
+}
+
+Status LoadLatestOptions(const ConfigOptions& config_options,
+                         const std::string& dbpath, DBOptions* db_options,
+                         std::vector<ColumnFamilyDescriptor>* cf_descs,
+                         std::shared_ptr<Cache>* cache) {
+  std::string options_file_name;
+  Status s =
+      GetLatestOptionsFileName(dbpath, config_options.env, &options_file_name);
+  if (!s.ok()) {
+    return s;
+  }
+  return LoadOptionsFromFile(config_options, dbpath + "/" + options_file_name,
+                             db_options, cf_descs, cache);
+}
+
+Status CheckOptionsCompatibility(
+    const std::string& dbpath, Env* env, const DBOptions& db_options,
+    const std::vector<ColumnFamilyDescriptor>& cf_descs,
+    bool ignore_unknown_options) {
+  ConfigOptions config_options(db_options);
+  config_options.sanity_level = ConfigOptions::kSanityLevelLooselyCompatible;
+  config_options.ignore_unknown_options = ignore_unknown_options;
+  config_options.input_strings_escaped = true;
+  config_options.env = env;
+  return CheckOptionsCompatibility(config_options, dbpath, db_options,
+                                   cf_descs);
+}
+
+Status CheckOptionsCompatibility(
+    const ConfigOptions& config_options, const std::string& dbpath,
+    const DBOptions& db_options,
+    const std::vector<ColumnFamilyDescriptor>& cf_descs) {
+  std::string options_file_name;
+  Status s =
+      GetLatestOptionsFileName(dbpath, config_options.env, &options_file_name);
+  if (!s.ok()) {
+    return s;
+  }
+
+  std::vector<std::string> cf_names;
+  std::vector<ColumnFamilyOptions> cf_opts;
+  for (const auto& cf_desc : cf_descs) {
+    cf_names.push_back(cf_desc.name);
+    cf_opts.push_back(cf_desc.options);
+  }
+
+  const auto& fs = config_options.env->GetFileSystem();
+
+  return RocksDBOptionsParser::VerifyRocksDBOptionsFromFile(
+      config_options, db_options, cf_names, cf_opts,
+      dbpath + "/" + options_file_name, fs.get());
+}
+
+}  // namespace ROCKSDB_NAMESPACE
+#endif  // !ROCKSDB_LITE
diff --git a/src/rocksdb/utilities/options/options_util_test.cc b/src/rocksdb/utilities/options/options_util_test.cc
new file mode 100644
index 000000000..1c3b41ff2
--- /dev/null
+++ b/src/rocksdb/utilities/options/options_util_test.cc
@@ -0,0 +1,779 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#ifndef ROCKSDB_LITE
+
+#include "rocksdb/utilities/options_util.h"
+
+#include <cctype>
+#include <cinttypes>
+#include <unordered_map>
+
+#include "env/mock_env.h"
+#include "file/filename.h"
+#include "options/options_parser.h"
+#include "rocksdb/convenience.h"
+#include "rocksdb/db.h"
+#include "rocksdb/table.h"
+#include "test_util/testharness.h"
+#include "test_util/testutil.h"
+#include "util/random.h"
+
+#ifndef GFLAGS
+bool FLAGS_enable_print = false;
+#else
+#include "util/gflags_compat.h"
+using GFLAGS_NAMESPACE::ParseCommandLineFlags;
+DEFINE_bool(enable_print, false, "Print options generated to console.");
+#endif  // GFLAGS
+
+namespace ROCKSDB_NAMESPACE {
+class OptionsUtilTest : public testing::Test {
+ public:
+  OptionsUtilTest() : rnd_(0xFB) {
+    env_.reset(NewMemEnv(Env::Default()));
+    dbname_ = test::PerThreadDBPath("options_util_test");
+  }
+
+ protected:
+  std::unique_ptr<Env> env_;
+  std::string dbname_;
+  Random rnd_;
+};
+
+TEST_F(OptionsUtilTest, SaveAndLoad) {
+  const size_t kCFCount = 5;
+
+  DBOptions db_opt;
+  std::vector<std::string> cf_names;
+  std::vector<ColumnFamilyOptions> cf_opts;
+  test::RandomInitDBOptions(&db_opt, &rnd_);
+  for (size_t i = 0; i < kCFCount; ++i) {
+    cf_names.push_back(i == 0 ? kDefaultColumnFamilyName
+                              : test::RandomName(&rnd_, 10));
+    cf_opts.emplace_back();
+    test::RandomInitCFOptions(&cf_opts.back(), db_opt, &rnd_);
+  }
+
+  const std::string kFileName = "OPTIONS-123456";
+  ASSERT_OK(PersistRocksDBOptions(db_opt, cf_names, cf_opts, kFileName,
+                                  env_->GetFileSystem().get()));
+
+  DBOptions loaded_db_opt;
+  std::vector<ColumnFamilyDescriptor> loaded_cf_descs;
+  ASSERT_OK(LoadOptionsFromFile(kFileName, env_.get(), &loaded_db_opt,
+                                &loaded_cf_descs));
+  ConfigOptions exact;
+  exact.sanity_level = ConfigOptions::kSanityLevelExactMatch;
+  ASSERT_OK(
+      RocksDBOptionsParser::VerifyDBOptions(exact, db_opt, loaded_db_opt));
+  test::RandomInitDBOptions(&db_opt, &rnd_);
+  ASSERT_NOK(
+      RocksDBOptionsParser::VerifyDBOptions(exact, db_opt, loaded_db_opt));
+
+  for (size_t i = 0; i < kCFCount; ++i) {
+    ASSERT_EQ(cf_names[i], loaded_cf_descs[i].name);
+    ASSERT_OK(RocksDBOptionsParser::VerifyCFOptions(
+        exact, cf_opts[i], loaded_cf_descs[i].options));
+    ASSERT_OK(RocksDBOptionsParser::VerifyTableFactory(
+        exact, cf_opts[i].table_factory.get(),
+        loaded_cf_descs[i].options.table_factory.get()));
+    test::RandomInitCFOptions(&cf_opts[i], db_opt, &rnd_);
+    ASSERT_NOK(RocksDBOptionsParser::VerifyCFOptions(
+        exact, cf_opts[i], loaded_cf_descs[i].options));
+  }
+
+  ASSERT_OK(DestroyDB(dbname_, Options(db_opt, cf_opts[0])));
+  for (size_t i = 0; i < kCFCount; ++i) {
+    if (cf_opts[i].compaction_filter) {
+      delete cf_opts[i].compaction_filter;
+    }
+  }
+}
+
+TEST_F(OptionsUtilTest, SaveAndLoadWithCacheCheck) {
+  // creating db
+  DBOptions db_opt;
+  db_opt.create_if_missing = true;
+  // initialize BlockBasedTableOptions
+  std::shared_ptr<Cache> cache = NewLRUCache(1 * 1024);
+  BlockBasedTableOptions bbt_opts;
+  bbt_opts.block_size = 32 * 1024;
+  // saving cf options
+  std::vector<ColumnFamilyOptions> cf_opts;
+  ColumnFamilyOptions default_column_family_opt = ColumnFamilyOptions();
+  default_column_family_opt.table_factory.reset(
+      NewBlockBasedTableFactory(bbt_opts));
+  cf_opts.push_back(default_column_family_opt);
+
+  ColumnFamilyOptions cf_opt_sample = ColumnFamilyOptions();
+  cf_opt_sample.table_factory.reset(NewBlockBasedTableFactory(bbt_opts));
+  cf_opts.push_back(cf_opt_sample);
+
+  ColumnFamilyOptions cf_opt_plain_table_opt = ColumnFamilyOptions();
+  cf_opt_plain_table_opt.table_factory.reset(NewPlainTableFactory());
+  cf_opts.push_back(cf_opt_plain_table_opt);
+
+  std::vector<std::string> cf_names;
+  cf_names.push_back(kDefaultColumnFamilyName);
+  cf_names.push_back("cf_sample");
+  cf_names.push_back("cf_plain_table_sample");
+  // Saving DB in file
+  const std::string kFileName = "OPTIONS-LOAD_CACHE_123456";
+  ASSERT_OK(PersistRocksDBOptions(db_opt, cf_names, cf_opts, kFileName,
+                                  env_->GetFileSystem().get()));
+  DBOptions loaded_db_opt;
+  std::vector<ColumnFamilyDescriptor> loaded_cf_descs;
+
+  ConfigOptions config_options;
+  config_options.ignore_unknown_options = false;
+  config_options.input_strings_escaped = true;
+  config_options.env = env_.get();
+  ASSERT_OK(LoadOptionsFromFile(config_options, kFileName, &loaded_db_opt,
+                                &loaded_cf_descs, &cache));
+  for (size_t i = 0; i < loaded_cf_descs.size(); i++) {
+    auto* loaded_bbt_opt =
+        loaded_cf_descs[i]
+            .options.table_factory->GetOptions<BlockBasedTableOptions>();
+    // Expect the same cache will be loaded
+    if (loaded_bbt_opt != nullptr) {
+      ASSERT_EQ(loaded_bbt_opt->block_cache.get(), cache.get());
+    }
+  }
+
+  // Test the old interface
+  ASSERT_OK(LoadOptionsFromFile(kFileName, env_.get(), &loaded_db_opt,
+                                &loaded_cf_descs, false, &cache));
+  for (size_t i = 0; i < loaded_cf_descs.size(); i++) {
+    auto* loaded_bbt_opt =
+        loaded_cf_descs[i]
+            .options.table_factory->GetOptions<BlockBasedTableOptions>();
+    // Expect the same cache will be loaded
+    if (loaded_bbt_opt != nullptr) {
+      ASSERT_EQ(loaded_bbt_opt->block_cache.get(), cache.get());
+    }
+  }
+  ASSERT_OK(DestroyDB(dbname_, Options(loaded_db_opt, cf_opts[0])));
+}
+
+namespace {
+class DummyTableFactory : public TableFactory {
+ public:
+  DummyTableFactory() {}
+  ~DummyTableFactory() override {}
+
+  const char* Name() const override { return "DummyTableFactory"; }
+
+  using TableFactory::NewTableReader;
+  Status NewTableReader(
+      const ReadOptions& /*ro*/,
+      const TableReaderOptions& /*table_reader_options*/,
+      std::unique_ptr<RandomAccessFileReader>&& /*file*/,
+      uint64_t /*file_size*/, std::unique_ptr<TableReader>* /*table_reader*/,
+      bool /*prefetch_index_and_filter_in_cache*/) const override {
+    return Status::NotSupported();
+  }
+
+  TableBuilder* NewTableBuilder(
+      const TableBuilderOptions& /*table_builder_options*/,
+      WritableFileWriter* /*file*/) const override {
+    return nullptr;
+  }
+
+  Status ValidateOptions(
+      const DBOptions& /*db_opts*/,
+      const ColumnFamilyOptions& /*cf_opts*/) const override {
+    return Status::NotSupported();
+  }
+
+  std::string GetPrintableOptions() const override { return ""; }
+};
+
+class DummyMergeOperator : public MergeOperator {
+ public:
+  DummyMergeOperator() {}
+  ~DummyMergeOperator() override {}
+
+  bool FullMergeV2(const MergeOperationInput& /*merge_in*/,
+                   MergeOperationOutput* /*merge_out*/) const override {
+    return false;
+  }
+
+  bool PartialMergeMulti(const Slice& /*key*/,
+                         const std::deque<Slice>& /*operand_list*/,
+                         std::string* /*new_value*/,
+                         Logger* /*logger*/) const override {
+    return false;
+  }
+
+  const char* Name() const override { return "DummyMergeOperator"; }
+};
+
+class DummySliceTransform : public SliceTransform {
+ public:
+  DummySliceTransform() {}
+  ~DummySliceTransform() override {}
+
+  // Return the name of this transformation.
+  const char* Name() const override { return "DummySliceTransform"; }
+
+  // transform a src in domain to a dst in the range
+  Slice Transform(const Slice& src) const override { return src; }
+
+  // determine whether this is a valid src upon the function applies
+  bool InDomain(const Slice& /*src*/) const override { return false; }
+
+  // determine whether dst=Transform(src) for some src
+  bool InRange(const Slice& /*dst*/) const override { return false; }
+};
+
+}  // namespace
+
+TEST_F(OptionsUtilTest, SanityCheck) {
+  DBOptions db_opt;
+  std::vector<ColumnFamilyDescriptor> cf_descs;
+  const size_t kCFCount = 5;
+  for (size_t i = 0; i < kCFCount; ++i) {
+    cf_descs.emplace_back();
+    cf_descs.back().name =
+        (i == 0) ? kDefaultColumnFamilyName : test::RandomName(&rnd_, 10);
+
+    cf_descs.back().options.table_factory.reset(NewBlockBasedTableFactory());
+    // Assign non-null values to prefix_extractors except the first cf.
+    cf_descs.back().options.prefix_extractor.reset(
+        i != 0 ? test::RandomSliceTransform(&rnd_) : nullptr);
+    cf_descs.back().options.merge_operator.reset(
+        test::RandomMergeOperator(&rnd_));
+  }
+
+  db_opt.create_missing_column_families = true;
+  db_opt.create_if_missing = true;
+
+  ASSERT_OK(DestroyDB(dbname_, Options(db_opt, cf_descs[0].options)));
+  DB* db;
+  std::vector<ColumnFamilyHandle*> handles;
+  // open and persist the options
+  ASSERT_OK(DB::Open(db_opt, dbname_, cf_descs, &handles, &db));
+
+  // close the db
+  for (auto* handle : handles) {
+    delete handle;
+  }
+  delete db;
+
+  ConfigOptions config_options;
+  config_options.ignore_unknown_options = false;
+  config_options.input_strings_escaped = true;
+  config_options.sanity_level = ConfigOptions::kSanityLevelLooselyCompatible;
+  // perform sanity check
+  ASSERT_OK(
+      CheckOptionsCompatibility(config_options, dbname_, db_opt, cf_descs));
+
+  ASSERT_GE(kCFCount, 5);
+  // merge operator
+  {
+    std::shared_ptr<MergeOperator> merge_op =
+        cf_descs[0].options.merge_operator;
+
+    ASSERT_NE(merge_op.get(), nullptr);
+    cf_descs[0].options.merge_operator.reset();
+    ASSERT_NOK(
+        CheckOptionsCompatibility(config_options, dbname_, db_opt, cf_descs));
+
+    cf_descs[0].options.merge_operator.reset(new DummyMergeOperator());
+    ASSERT_NOK(
+        CheckOptionsCompatibility(config_options, dbname_, db_opt, cf_descs));
+
+    cf_descs[0].options.merge_operator = merge_op;
+    ASSERT_OK(
+        CheckOptionsCompatibility(config_options, dbname_, db_opt, cf_descs));
+  }
+
+  // prefix extractor
+  {
+    std::shared_ptr<const SliceTransform> prefix_extractor =
+        cf_descs[1].options.prefix_extractor;
+
+    // It's okay to set prefix_extractor to nullptr.
+    ASSERT_NE(prefix_extractor, nullptr);
+    cf_descs[1].options.prefix_extractor.reset();
+    ASSERT_OK(
+        CheckOptionsCompatibility(config_options, dbname_, db_opt, cf_descs));
+
+    cf_descs[1].options.prefix_extractor.reset(new DummySliceTransform());
+    ASSERT_OK(
+        CheckOptionsCompatibility(config_options, dbname_, db_opt, cf_descs));
+
+    cf_descs[1].options.prefix_extractor = prefix_extractor;
+    ASSERT_OK(
+        CheckOptionsCompatibility(config_options, dbname_, db_opt, cf_descs));
+  }
+
+  // prefix extractor nullptr case
+  {
+    std::shared_ptr<const SliceTransform> prefix_extractor =
+        cf_descs[0].options.prefix_extractor;
+
+    // It's okay to set prefix_extractor to nullptr.
+    ASSERT_EQ(prefix_extractor, nullptr);
+    cf_descs[0].options.prefix_extractor.reset();
+    ASSERT_OK(
+        CheckOptionsCompatibility(config_options, dbname_, db_opt, cf_descs));
+
+    // It's okay to change prefix_extractor from nullptr to non-nullptr
+    cf_descs[0].options.prefix_extractor.reset(new DummySliceTransform());
+    ASSERT_OK(
+        CheckOptionsCompatibility(config_options, dbname_, db_opt, cf_descs));
+
+    cf_descs[0].options.prefix_extractor = prefix_extractor;
+    ASSERT_OK(
+        CheckOptionsCompatibility(config_options, dbname_, db_opt, cf_descs));
+  }
+
+  // comparator
+  {
+    test::SimpleSuffixReverseComparator comparator;
+
+    auto* prev_comparator = cf_descs[2].options.comparator;
+    cf_descs[2].options.comparator = &comparator;
+    ASSERT_NOK(
+        CheckOptionsCompatibility(config_options, dbname_, db_opt, cf_descs));
+
+    cf_descs[2].options.comparator = prev_comparator;
+    ASSERT_OK(
+        CheckOptionsCompatibility(config_options, dbname_, db_opt, cf_descs));
+  }
+
+  // table factory
+  {
+    std::shared_ptr<TableFactory> table_factory =
+        cf_descs[3].options.table_factory;
+
+    ASSERT_NE(table_factory, nullptr);
+    cf_descs[3].options.table_factory.reset(new DummyTableFactory());
+    ASSERT_NOK(
+        CheckOptionsCompatibility(config_options, dbname_, db_opt, cf_descs));
+
+    cf_descs[3].options.table_factory = table_factory;
+    ASSERT_OK(
+        CheckOptionsCompatibility(config_options, dbname_, db_opt, cf_descs));
+  }
+  ASSERT_OK(DestroyDB(dbname_, Options(db_opt, cf_descs[0].options)));
+}
+
+TEST_F(OptionsUtilTest, LatestOptionsNotFound) {
+  std::unique_ptr<Env> env(NewMemEnv(Env::Default()));
+  Status s;
+  Options options;
+  ConfigOptions config_opts;
+  std::vector<ColumnFamilyDescriptor> cf_descs;
+
+  options.env = env.get();
+  options.create_if_missing = true;
+  config_opts.env = options.env;
+  config_opts.ignore_unknown_options = false;
+
+  std::vector<std::string> children;
+
+  std::string options_file_name;
+  ASSERT_OK(DestroyDB(dbname_, options));
+  // First, test where the db directory does not exist
+  ASSERT_NOK(options.env->GetChildren(dbname_, &children));
+
+  s = GetLatestOptionsFileName(dbname_, options.env, &options_file_name);
+  ASSERT_TRUE(s.IsNotFound());
+  ASSERT_TRUE(s.IsPathNotFound());
+
+  s = LoadLatestOptions(dbname_, options.env, &options, &cf_descs);
+  ASSERT_TRUE(s.IsNotFound());
+  ASSERT_TRUE(s.IsPathNotFound());
+
+  s = LoadLatestOptions(config_opts, dbname_, &options, &cf_descs);
+  ASSERT_TRUE(s.IsPathNotFound());
+
+  s = GetLatestOptionsFileName(dbname_, options.env, &options_file_name);
+  ASSERT_TRUE(s.IsNotFound());
+  ASSERT_TRUE(s.IsPathNotFound());
+
+  // Second, test where the db directory exists but is empty
+  ASSERT_OK(options.env->CreateDir(dbname_));
+
+  s = GetLatestOptionsFileName(dbname_, options.env, &options_file_name);
+  ASSERT_TRUE(s.IsNotFound());
+  ASSERT_TRUE(s.IsPathNotFound());
+
+  s = LoadLatestOptions(dbname_, options.env, &options, &cf_descs);
+  ASSERT_TRUE(s.IsNotFound());
+  ASSERT_TRUE(s.IsPathNotFound());
+
+  // Finally, test where a file exists but is not an "Options" file
+  std::unique_ptr<WritableFile> file;
+  ASSERT_OK(
+      options.env->NewWritableFile(dbname_ + "/temp.txt", &file, EnvOptions()));
+  ASSERT_OK(file->Close());
+  s = GetLatestOptionsFileName(dbname_, options.env, &options_file_name);
+  ASSERT_TRUE(s.IsNotFound());
+  ASSERT_TRUE(s.IsPathNotFound());
+
+  s = LoadLatestOptions(config_opts, dbname_, &options, &cf_descs);
+  ASSERT_TRUE(s.IsNotFound());
+  ASSERT_TRUE(s.IsPathNotFound());
+  ASSERT_OK(options.env->DeleteFile(dbname_ + "/temp.txt"));
+  ASSERT_OK(options.env->DeleteDir(dbname_));
+}
+
+TEST_F(OptionsUtilTest, LoadLatestOptions) {
+  Options options;
+  options.OptimizeForSmallDb();
+  ColumnFamilyDescriptor cf_desc;
+  ConfigOptions config_opts;
+  DBOptions db_opts;
+  std::vector<ColumnFamilyDescriptor> cf_descs;
+  std::vector<ColumnFamilyHandle*> handles;
+  DB* db;
+  options.create_if_missing = true;
+
+  ASSERT_OK(DestroyDB(dbname_, options));
+
+  cf_descs.emplace_back();
+  cf_descs.back().name = kDefaultColumnFamilyName;
+  cf_descs.back().options.table_factory.reset(NewBlockBasedTableFactory());
+  cf_descs.emplace_back();
+  cf_descs.back().name = "Plain";
+  cf_descs.back().options.table_factory.reset(NewPlainTableFactory());
+  db_opts.create_missing_column_families = true;
+  db_opts.create_if_missing = true;
+
+  // open and persist the options
+  ASSERT_OK(DB::Open(db_opts, dbname_, cf_descs, &handles, &db));
+
+  std::string options_file_name;
+  std::string new_options_file;
+
+  ASSERT_OK(GetLatestOptionsFileName(dbname_, options.env, &options_file_name));
+  ASSERT_OK(LoadLatestOptions(config_opts, dbname_, &db_opts, &cf_descs));
+  ASSERT_EQ(cf_descs.size(), 2U);
+  ASSERT_OK(RocksDBOptionsParser::VerifyDBOptions(config_opts,
+                                                  db->GetDBOptions(), db_opts));
+  ASSERT_OK(handles[0]->GetDescriptor(&cf_desc));
+  ASSERT_OK(RocksDBOptionsParser::VerifyCFOptions(config_opts, cf_desc.options,
+                                                  cf_descs[0].options));
+  ASSERT_OK(handles[1]->GetDescriptor(&cf_desc));
+  ASSERT_OK(RocksDBOptionsParser::VerifyCFOptions(config_opts, cf_desc.options,
+                                                  cf_descs[1].options));
+
+  // Now change some of the DBOptions
+  ASSERT_OK(db->SetDBOptions(
+      {{"delayed_write_rate", "1234"}, {"bytes_per_sync", "32768"}}));
+  ASSERT_OK(GetLatestOptionsFileName(dbname_, options.env, &new_options_file));
+  ASSERT_NE(options_file_name, new_options_file);
+  ASSERT_OK(LoadLatestOptions(config_opts, dbname_, &db_opts, &cf_descs));
+  ASSERT_OK(RocksDBOptionsParser::VerifyDBOptions(config_opts,
+                                                  db->GetDBOptions(), db_opts));
+  options_file_name = new_options_file;
+
+  // Now change some of the ColumnFamilyOptions
+  ASSERT_OK(db->SetOptions(handles[1], {{"write_buffer_size", "32768"}}));
+  ASSERT_OK(GetLatestOptionsFileName(dbname_, options.env, &new_options_file));
+  ASSERT_NE(options_file_name, new_options_file);
+  ASSERT_OK(LoadLatestOptions(config_opts, dbname_, &db_opts, &cf_descs));
+  ASSERT_OK(RocksDBOptionsParser::VerifyDBOptions(config_opts,
+                                                  db->GetDBOptions(), db_opts));
+  ASSERT_OK(handles[0]->GetDescriptor(&cf_desc));
+  ASSERT_OK(RocksDBOptionsParser::VerifyCFOptions(config_opts, cf_desc.options,
+                                                  cf_descs[0].options));
+  ASSERT_OK(handles[1]->GetDescriptor(&cf_desc));
+  ASSERT_OK(RocksDBOptionsParser::VerifyCFOptions(config_opts, cf_desc.options,
+                                                  cf_descs[1].options));
+
+  // close the db
+  for (auto* handle : handles) {
+    delete handle;
+  }
+  delete db;
+  ASSERT_OK(DestroyDB(dbname_, options, cf_descs));
+}
+
+static void WriteOptionsFile(Env* env, const std::string& path,
+                             const std::string& options_file, int major,
+                             int minor, const std::string& db_opts,
+                             const std::string& cf_opts,
+                             const std::string& bbt_opts = "") {
+  std::string options_file_header =
+      "\n"
+      "[Version]\n"
+      "  rocksdb_version=" +
+      std::to_string(major) + "." + std::to_string(minor) +
+      ".0\n"
+      "  options_file_version=1\n";
+
+  std::unique_ptr<WritableFile> wf;
+  ASSERT_OK(env->NewWritableFile(path + "/" + options_file, &wf, EnvOptions()));
+  ASSERT_OK(
+      wf->Append(options_file_header + "[ DBOptions ]\n" + db_opts + "\n"));
+  ASSERT_OK(wf->Append(
+      "[CFOptions   \"default\"]  # column family must be specified\n" +
+      cf_opts + "\n"));
+  ASSERT_OK(wf->Append("[TableOptions/BlockBasedTable   \"default\"]\n" +
+                       bbt_opts + "\n"));
+  ASSERT_OK(wf->Close());
+
+  std::string latest_options_file;
+  ASSERT_OK(GetLatestOptionsFileName(path, env, &latest_options_file));
+  ASSERT_EQ(latest_options_file, options_file);
+}
+
+TEST_F(OptionsUtilTest, BadLatestOptions) {
+  Status s;
+  ConfigOptions config_opts;
+  DBOptions db_opts;
+  std::vector<ColumnFamilyDescriptor> cf_descs;
+  Options options;
+  options.env = env_.get();
+  config_opts.env = env_.get();
+  config_opts.ignore_unknown_options = false;
+  config_opts.delimiter = "\n";
+
+  ConfigOptions ignore_opts = config_opts;
+  ignore_opts.ignore_unknown_options = true;
+
+  std::string options_file_name;
+
+  // Test where the db directory exists but is empty
+  ASSERT_OK(options.env->CreateDir(dbname_));
+  ASSERT_NOK(
+      GetLatestOptionsFileName(dbname_, options.env, &options_file_name));
+  ASSERT_NOK(LoadLatestOptions(config_opts, dbname_, &db_opts, &cf_descs));
+
+  // Write an options file for a previous major release with an unknown DB
+  // Option
+  WriteOptionsFile(options.env, dbname_, "OPTIONS-0001", ROCKSDB_MAJOR - 1,
+                   ROCKSDB_MINOR, "unknown_db_opt=true", "");
+  s = LoadLatestOptions(config_opts, dbname_, &db_opts, &cf_descs);
+  ASSERT_NOK(s);
+  ASSERT_TRUE(s.IsInvalidArgument());
+  // Even though ignore_unknown_options=true, we still return an error...
+  s = LoadLatestOptions(ignore_opts, dbname_, &db_opts, &cf_descs);
+  ASSERT_NOK(s);
+  ASSERT_TRUE(s.IsInvalidArgument());
+  // Write an options file for a previous minor release with an unknown CF
+  // Option
+  WriteOptionsFile(options.env, dbname_, "OPTIONS-0002", ROCKSDB_MAJOR,
+                   ROCKSDB_MINOR - 1, "", "unknown_cf_opt=true");
+  s = LoadLatestOptions(config_opts, dbname_, &db_opts, &cf_descs);
+  ASSERT_NOK(s);
+  ASSERT_TRUE(s.IsInvalidArgument());
+  // Even though ignore_unknown_options=true, we still return an error...
+  s = LoadLatestOptions(ignore_opts, dbname_, &db_opts, &cf_descs);
+  ASSERT_NOK(s);
+  ASSERT_TRUE(s.IsInvalidArgument());
+  // Write an options file for a previous minor release with an unknown BBT
+  // Option
+  WriteOptionsFile(options.env, dbname_, "OPTIONS-0003", ROCKSDB_MAJOR,
+                   ROCKSDB_MINOR - 1, "", "", "unknown_bbt_opt=true");
+  s = LoadLatestOptions(config_opts, dbname_, &db_opts, &cf_descs);
+  ASSERT_NOK(s);
+  ASSERT_TRUE(s.IsInvalidArgument());
+  // Even though ignore_unknown_options=true, we still return an error...
+  s = LoadLatestOptions(ignore_opts, dbname_, &db_opts, &cf_descs);
+  ASSERT_NOK(s);
+  ASSERT_TRUE(s.IsInvalidArgument());
+
+  // Write an options file for the current release with an unknown DB Option
+  WriteOptionsFile(options.env, dbname_, "OPTIONS-0004", ROCKSDB_MAJOR,
+                   ROCKSDB_MINOR, "unknown_db_opt=true", "");
+  s = LoadLatestOptions(config_opts, dbname_, &db_opts, &cf_descs);
+  ASSERT_NOK(s);
+  ASSERT_TRUE(s.IsInvalidArgument());
+  // Even though ignore_unknown_options=true, we still return an error...
+  s = LoadLatestOptions(ignore_opts, dbname_, &db_opts, &cf_descs);
+  ASSERT_NOK(s);
+  ASSERT_TRUE(s.IsInvalidArgument());
+
+  // Write an options file for the current release with an unknown CF Option
+  WriteOptionsFile(options.env, dbname_, "OPTIONS-0005", ROCKSDB_MAJOR,
+                   ROCKSDB_MINOR, "", "unknown_cf_opt=true");
+  s = LoadLatestOptions(config_opts, dbname_, &db_opts, &cf_descs);
+  ASSERT_NOK(s);
+  ASSERT_TRUE(s.IsInvalidArgument());
+  // Even though ignore_unknown_options=true, we still return an error...
+  s = LoadLatestOptions(ignore_opts, dbname_, &db_opts, &cf_descs);
+  ASSERT_NOK(s);
+  ASSERT_TRUE(s.IsInvalidArgument());
+
+  // Write an options file for the current release with an invalid DB Option
+  WriteOptionsFile(options.env, dbname_, "OPTIONS-0006", ROCKSDB_MAJOR,
+                   ROCKSDB_MINOR, "create_if_missing=hello", "");
+  s = LoadLatestOptions(config_opts, dbname_, &db_opts, &cf_descs);
+  ASSERT_NOK(s);
+  ASSERT_TRUE(s.IsInvalidArgument());
+  // Even though ignore_unknown_options=true, we still return an error...
+  s = LoadLatestOptions(ignore_opts, dbname_, &db_opts, &cf_descs);
+  ASSERT_NOK(s);
+  ASSERT_TRUE(s.IsInvalidArgument());
+
+  // Write an options file for the next release with an invalid DB Option
+  WriteOptionsFile(options.env, dbname_, "OPTIONS-0007", ROCKSDB_MAJOR,
+                   ROCKSDB_MINOR + 1, "create_if_missing=hello", "");
+  ASSERT_NOK(LoadLatestOptions(config_opts, dbname_, &db_opts, &cf_descs));
+  ASSERT_OK(LoadLatestOptions(ignore_opts, dbname_, &db_opts, &cf_descs));
+
+  // Write an options file for the next release with an unknown DB Option
+  WriteOptionsFile(options.env, dbname_, "OPTIONS-0008", ROCKSDB_MAJOR,
+                   ROCKSDB_MINOR + 1, "unknown_db_opt=true", "");
+  ASSERT_NOK(LoadLatestOptions(config_opts, dbname_, &db_opts, &cf_descs));
+  // Ignore the errors for future releases when ignore_unknown_options=true
+  ASSERT_OK(LoadLatestOptions(ignore_opts, dbname_, &db_opts, &cf_descs));
+
+  // Write an options file for the next major release with an unknown CF Option
+  WriteOptionsFile(options.env, dbname_, "OPTIONS-0009", ROCKSDB_MAJOR + 1,
+                   ROCKSDB_MINOR, "", "unknown_cf_opt=true");
+  ASSERT_NOK(LoadLatestOptions(config_opts, dbname_, &db_opts, &cf_descs));
+  // Ignore the errors for future releases when ignore_unknown_options=true
+  ASSERT_OK(LoadLatestOptions(ignore_opts, dbname_, &db_opts, &cf_descs));
+}
+
+TEST_F(OptionsUtilTest, RenameDatabaseDirectory) {
+  DB* db;
+  Options options;
+  DBOptions db_opts;
+  std::vector<ColumnFamilyDescriptor> cf_descs;
+  std::vector<ColumnFamilyHandle*> handles;
+
+  options.create_if_missing = true;
+
+  ASSERT_OK(DB::Open(options, dbname_, &db));
+  ASSERT_OK(db->Put(WriteOptions(), "foo", "value0"));
+  delete db;
+
+  auto new_dbname = dbname_ + "_2";
+
+  ASSERT_OK(options.env->RenameFile(dbname_, new_dbname));
+  ASSERT_OK(LoadLatestOptions(new_dbname, options.env, &db_opts, &cf_descs));
+  ASSERT_EQ(cf_descs.size(), 1U);
+
+  db_opts.create_if_missing = false;
+  ASSERT_OK(DB::Open(db_opts, new_dbname, cf_descs, &handles, &db));
+  std::string value;
+  ASSERT_OK(db->Get(ReadOptions(), "foo", &value));
+  ASSERT_EQ("value0", value);
+  // close the db
+  for (auto* handle : handles) {
+    delete handle;
+  }
+  delete db;
+  Options new_options(db_opts, cf_descs[0].options);
+  ASSERT_OK(DestroyDB(new_dbname, new_options, cf_descs));
+  ASSERT_OK(DestroyDB(dbname_, options));
+}
+
+TEST_F(OptionsUtilTest, WalDirSettings) {
+  DB* db;
+  Options options;
+  DBOptions db_opts;
+  std::vector<ColumnFamilyDescriptor> cf_descs;
+  std::vector<ColumnFamilyHandle*> handles;
+
+  options.create_if_missing = true;
+
+  // Open a DB with no wal dir set.  The wal_dir should stay empty
+  ASSERT_OK(DB::Open(options, dbname_, &db));
+  delete db;
+  ASSERT_OK(LoadLatestOptions(dbname_, options.env, &db_opts, &cf_descs));
+  ASSERT_EQ(db_opts.wal_dir, "");
+
+  // Open a DB with wal_dir == dbname.  The wal_dir should be set to empty
+  options.wal_dir = dbname_;
+  ASSERT_OK(DB::Open(options, dbname_, &db));
+  delete db;
+  ASSERT_OK(LoadLatestOptions(dbname_, options.env, &db_opts, &cf_descs));
+  ASSERT_EQ(db_opts.wal_dir, "");
+
+  // Open a DB with no wal_dir but a db_path==dbname_.  The wal_dir should be
+  // empty
+  options.wal_dir = "";
+  options.db_paths.emplace_back(dbname_, std::numeric_limits<uint64_t>::max());
+  ASSERT_OK(DB::Open(options, dbname_, &db));
+  delete db;
+  ASSERT_OK(LoadLatestOptions(dbname_, options.env, &db_opts, &cf_descs));
+  ASSERT_EQ(db_opts.wal_dir, "");
+
+  // Open a DB with no wal_dir==dbname_ and db_path==dbname_.  The wal_dir
+  // should be empty
+  options.wal_dir = dbname_ + "/";
+  options.db_paths.emplace_back(dbname_, std::numeric_limits<uint64_t>::max());
+  ASSERT_OK(DB::Open(options, dbname_, &db));
+  delete db;
+  ASSERT_OK(LoadLatestOptions(dbname_, options.env, &db_opts, &cf_descs));
+  ASSERT_EQ(db_opts.wal_dir, "");
+  ASSERT_OK(DestroyDB(dbname_, options));
+
+  // Open a DB with no wal_dir but db_path != db_name.  The wal_dir == dbname_
+  options.wal_dir = "";
+  options.db_paths.clear();
+  options.db_paths.emplace_back(dbname_ + "_0",
+                                std::numeric_limits<uint64_t>::max());
+  ASSERT_OK(DB::Open(options, dbname_, &db));
+  delete db;
+  ASSERT_OK(LoadLatestOptions(dbname_, options.env, &db_opts, &cf_descs));
+  ASSERT_EQ(db_opts.wal_dir, dbname_);
+  ASSERT_OK(DestroyDB(dbname_, options));
+
+  // Open a DB with wal_dir != db_name.  The wal_dir remains unchanged
+  options.wal_dir = dbname_ + "/wal";
+  options.db_paths.clear();
+  ASSERT_OK(DB::Open(options, dbname_, &db));
+  delete db;
+  ASSERT_OK(LoadLatestOptions(dbname_, options.env, &db_opts, &cf_descs));
+  ASSERT_EQ(db_opts.wal_dir, dbname_ + "/wal");
+  ASSERT_OK(DestroyDB(dbname_, options));
+}
+
+TEST_F(OptionsUtilTest, WalDirInOptins) {
+  DB* db;
+  Options options;
+  DBOptions db_opts;
+  std::vector<ColumnFamilyDescriptor> cf_descs;
+  std::vector<ColumnFamilyHandle*> handles;
+
+  // Store an options file with wal_dir=dbname_ and make sure it still loads
+  // when the input wal_dir is empty
+  options.create_if_missing = true;
+  options.wal_dir = "";
+  ASSERT_OK(DB::Open(options, dbname_, &db));
+  delete db;
+  options.wal_dir = dbname_;
+  std::string options_file;
+  ASSERT_OK(GetLatestOptionsFileName(dbname_, options.env, &options_file));
+  ASSERT_OK(PersistRocksDBOptions(options, {"default"}, {options},
+                                  dbname_ + "/" + options_file,
+                                  options.env->GetFileSystem().get()));
+  ASSERT_OK(LoadLatestOptions(dbname_, options.env, &db_opts, &cf_descs));
+  ASSERT_EQ(db_opts.wal_dir, dbname_);
+  options.wal_dir = "";
+  ASSERT_OK(DB::Open(options, dbname_, &db));
+  delete db;
+  ASSERT_OK(LoadLatestOptions(dbname_, options.env, &db_opts, &cf_descs));
+  ASSERT_EQ(db_opts.wal_dir, "");
+}
+}  // namespace ROCKSDB_NAMESPACE
+
+int main(int argc, char** argv) {
+  ROCKSDB_NAMESPACE::port::InstallStackTraceHandler();
+  ::testing::InitGoogleTest(&argc, argv);
+#ifdef GFLAGS
+  ParseCommandLineFlags(&argc, &argv, true);
+#endif  // GFLAGS
+  return RUN_ALL_TESTS();
+}
+
+#else
+#include <cstdio>
+
+int main(int /*argc*/, char** /*argv*/) {
+  printf("Skipped in RocksDBLite as utilities are not supported.\n");
+  return 0;
+}
+#endif  // !ROCKSDB_LITE
diff --git a/src/rocksdb/utilities/persistent_cache/block_cache_tier.cc b/src/rocksdb/utilities/persistent_cache/block_cache_tier.cc
new file mode 100644
index 000000000..8ad9bb1b1
--- /dev/null
+++ b/src/rocksdb/utilities/persistent_cache/block_cache_tier.cc
@@ -0,0 +1,422 @@
+//  Copyright (c) 2013, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+#ifndef ROCKSDB_LITE
+
+#include "utilities/persistent_cache/block_cache_tier.h"
+
+#include <utility>
+#include <vector>
+
+#include "logging/logging.h"
+#include "port/port.h"
+#include "test_util/sync_point.h"
+#include "util/stop_watch.h"
+#include "utilities/persistent_cache/block_cache_tier_file.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+//
+// BlockCacheImpl
+//
+Status BlockCacheTier::Open() {
+  Status status;
+
+  WriteLock _(&lock_);
+
+  assert(!size_);
+
+  // Check the validity of the options
+  status = opt_.ValidateSettings();
+  assert(status.ok());
+  if (!status.ok()) {
+    Error(opt_.log, "Invalid block cache options");
+    return status;
+  }
+
+  // Create base directory or cleanup existing directory
+  status = opt_.env->CreateDirIfMissing(opt_.path);
+  if (!status.ok()) {
+    Error(opt_.log, "Error creating directory %s. %s", opt_.path.c_str(),
+          status.ToString().c_str());
+    return status;
+  }
+
+  // Create base/<cache dir> directory
+  status = opt_.env->CreateDir(GetCachePath());
+  if (!status.ok()) {
+    // directory already exists, clean it up
+    status = CleanupCacheFolder(GetCachePath());
+    assert(status.ok());
+    if (!status.ok()) {
+      Error(opt_.log, "Error creating directory %s. %s", opt_.path.c_str(),
+            status.ToString().c_str());
+      return status;
+    }
+  }
+
+  // create a new file
+  assert(!cache_file_);
+  status = NewCacheFile();
+  if (!status.ok()) {
+    Error(opt_.log, "Error creating new file %s. %s", opt_.path.c_str(),
+          status.ToString().c_str());
+    return status;
+  }
+
+  assert(cache_file_);
+
+  if (opt_.pipeline_writes) {
+    assert(!insert_th_.joinable());
+    insert_th_ = port::Thread(&BlockCacheTier::InsertMain, this);
+  }
+
+  return Status::OK();
+}
+
+bool IsCacheFile(const std::string& file) {
+  // check if the file has .rc suffix
+  // Unfortunately regex support across compilers is not even, so we use simple
+  // string parsing
+  size_t pos = file.find(".");
+  if (pos == std::string::npos) {
+    return false;
+  }
+
+  std::string suffix = file.substr(pos);
+  return suffix == ".rc";
+}
+
+Status BlockCacheTier::CleanupCacheFolder(const std::string& folder) {
+  std::vector<std::string> files;
+  Status status = opt_.env->GetChildren(folder, &files);
+  if (!status.ok()) {
+    Error(opt_.log, "Error getting files for %s. %s", folder.c_str(),
+          status.ToString().c_str());
+    return status;
+  }
+
+  // cleanup files with the patter :digi:.rc
+  for (auto file : files) {
+    if (IsCacheFile(file)) {
+      // cache file
+      Info(opt_.log, "Removing file %s.", file.c_str());
+      status = opt_.env->DeleteFile(folder + "/" + file);
+      if (!status.ok()) {
+        Error(opt_.log, "Error deleting file %s. %s", file.c_str(),
+              status.ToString().c_str());
+        return status;
+      }
+    } else {
+      ROCKS_LOG_DEBUG(opt_.log, "Skipping file %s", file.c_str());
+    }
+  }
+  return Status::OK();
+}
+
+Status BlockCacheTier::Close() {
+  // stop the insert thread
+  if (opt_.pipeline_writes && insert_th_.joinable()) {
+    InsertOp op(/*quit=*/true);
+    insert_ops_.Push(std::move(op));
+    insert_th_.join();
+  }
+
+  // stop the writer before
+  writer_.Stop();
+
+  // clear all metadata
+  WriteLock _(&lock_);
+  metadata_.Clear();
+  return Status::OK();
+}
+
+template <class T>
+void Add(std::map<std::string, double>* stats, const std::string& key,
+         const T& t) {
+  stats->insert({key, static_cast<double>(t)});
+}
+
+PersistentCache::StatsType BlockCacheTier::Stats() {
+  std::map<std::string, double> stats;
+  Add(&stats, "persistentcache.blockcachetier.bytes_piplined",
+      stats_.bytes_pipelined_.Average());
+  Add(&stats, "persistentcache.blockcachetier.bytes_written",
+      stats_.bytes_written_.Average());
+  Add(&stats, "persistentcache.blockcachetier.bytes_read",
+      stats_.bytes_read_.Average());
+  Add(&stats, "persistentcache.blockcachetier.insert_dropped",
+      stats_.insert_dropped_);
+  Add(&stats, "persistentcache.blockcachetier.cache_hits", stats_.cache_hits_);
+  Add(&stats, "persistentcache.blockcachetier.cache_misses",
+      stats_.cache_misses_);
+  Add(&stats, "persistentcache.blockcachetier.cache_errors",
+      stats_.cache_errors_);
+  Add(&stats, "persistentcache.blockcachetier.cache_hits_pct",
+      stats_.CacheHitPct());
+  Add(&stats, "persistentcache.blockcachetier.cache_misses_pct",
+      stats_.CacheMissPct());
+  Add(&stats, "persistentcache.blockcachetier.read_hit_latency",
+      stats_.read_hit_latency_.Average());
+  Add(&stats, "persistentcache.blockcachetier.read_miss_latency",
+      stats_.read_miss_latency_.Average());
+  Add(&stats, "persistentcache.blockcachetier.write_latency",
+      stats_.write_latency_.Average());
+
+  auto out = PersistentCacheTier::Stats();
+  out.push_back(stats);
+  return out;
+}
+
+Status BlockCacheTier::Insert(const Slice& key, const char* data,
+                              const size_t size) {
+  // update stats
+  stats_.bytes_pipelined_.Add(size);
+
+  if (opt_.pipeline_writes) {
+    // off load the write to the write thread
+    insert_ops_.Push(
+        InsertOp(key.ToString(), std::move(std::string(data, size))));
+    return Status::OK();
+  }
+
+  assert(!opt_.pipeline_writes);
+  return InsertImpl(key, Slice(data, size));
+}
+
+void BlockCacheTier::InsertMain() {
+  while (true) {
+    InsertOp op(insert_ops_.Pop());
+
+    if (op.signal_) {
+      // that is a secret signal to exit
+      break;
+    }
+
+    size_t retry = 0;
+    Status s;
+    while ((s = InsertImpl(Slice(op.key_), Slice(op.data_))).IsTryAgain()) {
+      if (retry > kMaxRetry) {
+        break;
+      }
+
+      // this can happen when the buffers are full, we wait till some buffers
+      // are free. Why don't we wait inside the code. This is because we want
+      // to support both pipelined and non-pipelined mode
+      buffer_allocator_.WaitUntilUsable();
+      retry++;
+    }
+
+    if (!s.ok()) {
+      stats_.insert_dropped_++;
+    }
+  }
+}
+
+Status BlockCacheTier::InsertImpl(const Slice& key, const Slice& data) {
+  // pre-condition
+  assert(key.size());
+  assert(data.size());
+  assert(cache_file_);
+
+  StopWatchNano timer(opt_.clock, /*auto_start=*/true);
+
+  WriteLock _(&lock_);
+
+  LBA lba;
+  if (metadata_.Lookup(key, &lba)) {
+    // the key already exists, this is duplicate insert
+    return Status::OK();
+  }
+
+  while (!cache_file_->Append(key, data, &lba)) {
+    if (!cache_file_->Eof()) {
+      ROCKS_LOG_DEBUG(opt_.log, "Error inserting to cache file %d",
+                      cache_file_->cacheid());
+      stats_.write_latency_.Add(timer.ElapsedNanos() / 1000);
+      return Status::TryAgain();
+    }
+
+    assert(cache_file_->Eof());
+    Status status = NewCacheFile();
+    if (!status.ok()) {
+      return status;
+    }
+  }
+
+  // Insert into lookup index
+  BlockInfo* info = metadata_.Insert(key, lba);
+  assert(info);
+  if (!info) {
+    return Status::IOError("Unexpected error inserting to index");
+  }
+
+  // insert to cache file reverse mapping
+  cache_file_->Add(info);
+
+  // update stats
+  stats_.bytes_written_.Add(data.size());
+  stats_.write_latency_.Add(timer.ElapsedNanos() / 1000);
+  return Status::OK();
+}
+
+Status BlockCacheTier::Lookup(const Slice& key, std::unique_ptr<char[]>* val,
+                              size_t* size) {
+  StopWatchNano timer(opt_.clock, /*auto_start=*/true);
+
+  LBA lba;
+  bool status;
+  status = metadata_.Lookup(key, &lba);
+  if (!status) {
+    stats_.cache_misses_++;
+    stats_.read_miss_latency_.Add(timer.ElapsedNanos() / 1000);
+    return Status::NotFound("blockcache: key not found");
+  }
+
+  BlockCacheFile* const file = metadata_.Lookup(lba.cache_id_);
+  if (!file) {
+    // this can happen because the block index and cache file index are
+    // different, and the cache file might be removed between the two lookups
+    stats_.cache_misses_++;
+    stats_.read_miss_latency_.Add(timer.ElapsedNanos() / 1000);
+    return Status::NotFound("blockcache: cache file not found");
+  }
+
+  assert(file->refs_);
+
+  std::unique_ptr<char[]> scratch(new char[lba.size_]);
+  Slice blk_key;
+  Slice blk_val;
+
+  status = file->Read(lba, &blk_key, &blk_val, scratch.get());
+  --file->refs_;
+  if (!status) {
+    stats_.cache_misses_++;
+    stats_.cache_errors_++;
+    stats_.read_miss_latency_.Add(timer.ElapsedNanos() / 1000);
+    return Status::NotFound("blockcache: error reading data");
+  }
+
+  assert(blk_key == key);
+
+  val->reset(new char[blk_val.size()]);
+  memcpy(val->get(), blk_val.data(), blk_val.size());
+  *size = blk_val.size();
+
+  stats_.bytes_read_.Add(*size);
+  stats_.cache_hits_++;
+  stats_.read_hit_latency_.Add(timer.ElapsedNanos() / 1000);
+
+  return Status::OK();
+}
+
+bool BlockCacheTier::Erase(const Slice& key) {
+  WriteLock _(&lock_);
+  BlockInfo* info = metadata_.Remove(key);
+  assert(info);
+  delete info;
+  return true;
+}
+
+Status BlockCacheTier::NewCacheFile() {
+  lock_.AssertHeld();
+
+  TEST_SYNC_POINT_CALLBACK("BlockCacheTier::NewCacheFile:DeleteDir",
+                           (void*)(GetCachePath().c_str()));
+
+  std::unique_ptr<WriteableCacheFile> f(new WriteableCacheFile(
+      opt_.env, &buffer_allocator_, &writer_, GetCachePath(), writer_cache_id_,
+      opt_.cache_file_size, opt_.log));
+
+  bool status = f->Create(opt_.enable_direct_writes, opt_.enable_direct_reads);
+  if (!status) {
+    return Status::IOError("Error creating file");
+  }
+
+  Info(opt_.log, "Created cache file %d", writer_cache_id_);
+
+  writer_cache_id_++;
+  cache_file_ = f.release();
+
+  // insert to cache files tree
+  status = metadata_.Insert(cache_file_);
+  assert(status);
+  if (!status) {
+    Error(opt_.log, "Error inserting to metadata");
+    return Status::IOError("Error inserting to metadata");
+  }
+
+  return Status::OK();
+}
+
+bool BlockCacheTier::Reserve(const size_t size) {
+  WriteLock _(&lock_);
+  assert(size_ <= opt_.cache_size);
+
+  if (size + size_ <= opt_.cache_size) {
+    // there is enough space to write
+    size_ += size;
+    return true;
+  }
+
+  assert(size + size_ >= opt_.cache_size);
+  // there is not enough space to fit the requested data
+  // we can clear some space by evicting cold data
+
+  const double retain_fac = (100 - kEvictPct) / static_cast<double>(100);
+  while (size + size_ > opt_.cache_size * retain_fac) {
+    std::unique_ptr<BlockCacheFile> f(metadata_.Evict());
+    if (!f) {
+      // nothing is evictable
+      return false;
+    }
+    assert(!f->refs_);
+    uint64_t file_size;
+    if (!f->Delete(&file_size).ok()) {
+      // unable to delete file
+      return false;
+    }
+
+    assert(file_size <= size_);
+    size_ -= file_size;
+  }
+
+  size_ += size;
+  assert(size_ <= opt_.cache_size * 0.9);
+  return true;
+}
+
+Status NewPersistentCache(Env* const env, const std::string& path,
+                          const uint64_t size,
+                          const std::shared_ptr<Logger>& log,
+                          const bool optimized_for_nvm,
+                          std::shared_ptr<PersistentCache>* cache) {
+  if (!cache) {
+    return Status::IOError("invalid argument cache");
+  }
+
+  auto opt = PersistentCacheConfig(env, path, size, log);
+  if (optimized_for_nvm) {
+    // the default settings are optimized for SSD
+    // NVM devices are better accessed with 4K direct IO and written with
+    // parallelism
+    opt.enable_direct_writes = true;
+    opt.writer_qdepth = 4;
+    opt.writer_dispatch_size = 4 * 1024;
+  }
+
+  auto pcache = std::make_shared<BlockCacheTier>(opt);
+  Status s = pcache->Open();
+
+  if (!s.ok()) {
+    return s;
+  }
+
+  *cache = pcache;
+  return s;
+}
+
+}  // namespace ROCKSDB_NAMESPACE
+
+#endif  // ifndef ROCKSDB_LITE
diff --git a/src/rocksdb/utilities/persistent_cache/block_cache_tier.h b/src/rocksdb/utilities/persistent_cache/block_cache_tier.h
new file mode 100644
index 000000000..1aac287cc
--- /dev/null
+++ b/src/rocksdb/utilities/persistent_cache/block_cache_tier.h
@@ -0,0 +1,156 @@
+//  Copyright (c) 2013, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+#pragma once
+
+#ifndef ROCKSDB_LITE
+
+#ifndef OS_WIN
+#include <unistd.h>
+#endif  // ! OS_WIN
+
+#include <atomic>
+#include <list>
+#include <memory>
+#include <set>
+#include <sstream>
+#include <stdexcept>
+#include <string>
+#include <thread>
+
+#include "memory/arena.h"
+#include "memtable/skiplist.h"
+#include "monitoring/histogram.h"
+#include "port/port.h"
+#include "rocksdb/cache.h"
+#include "rocksdb/comparator.h"
+#include "rocksdb/persistent_cache.h"
+#include "rocksdb/system_clock.h"
+#include "util/coding.h"
+#include "util/crc32c.h"
+#include "util/mutexlock.h"
+#include "utilities/persistent_cache/block_cache_tier_file.h"
+#include "utilities/persistent_cache/block_cache_tier_metadata.h"
+#include "utilities/persistent_cache/persistent_cache_util.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+//
+// Block cache tier implementation
+//
+class BlockCacheTier : public PersistentCacheTier {
+ public:
+  explicit BlockCacheTier(const PersistentCacheConfig& opt)
+      : opt_(opt),
+        insert_ops_(static_cast<size_t>(opt_.max_write_pipeline_backlog_size)),
+        buffer_allocator_(opt.write_buffer_size, opt.write_buffer_count()),
+        writer_(this, opt_.writer_qdepth,
+                static_cast<size_t>(opt_.writer_dispatch_size)) {
+    Info(opt_.log, "Initializing allocator. size=%d B count=%" ROCKSDB_PRIszt,
+         opt_.write_buffer_size, opt_.write_buffer_count());
+  }
+
+  virtual ~BlockCacheTier() {
+    // Close is re-entrant so we can call close even if it is already closed
+    Close().PermitUncheckedError();
+    assert(!insert_th_.joinable());
+  }
+
+  Status Insert(const Slice& key, const char* data, const size_t size) override;
+  Status Lookup(const Slice& key, std::unique_ptr<char[]>* data,
+                size_t* size) override;
+  Status Open() override;
+  Status Close() override;
+  bool Erase(const Slice& key) override;
+  bool Reserve(const size_t size) override;
+
+  bool IsCompressed() override { return opt_.is_compressed; }
+
+  std::string GetPrintableOptions() const override { return opt_.ToString(); }
+
+  PersistentCache::StatsType Stats() override;
+
+  void TEST_Flush() override {
+    while (insert_ops_.Size()) {
+      /* sleep override */
+      SystemClock::Default()->SleepForMicroseconds(1000000);
+    }
+  }
+
+ private:
+  // Percentage of cache to be evicted when the cache is full
+  static const size_t kEvictPct = 10;
+  // Max attempts to insert key, value to cache in pipelined mode
+  static const size_t kMaxRetry = 3;
+
+  // Pipelined operation
+  struct InsertOp {
+    explicit InsertOp(const bool signal) : signal_(signal) {}
+    explicit InsertOp(std::string&& key, const std::string& data)
+        : key_(std::move(key)), data_(data) {}
+    ~InsertOp() {}
+
+    InsertOp() = delete;
+    InsertOp(InsertOp&& /*rhs*/) = default;
+    InsertOp& operator=(InsertOp&& rhs) = default;
+
+    // used for estimating size by bounded queue
+    size_t Size() { return data_.size() + key_.size(); }
+
+    std::string key_;
+    std::string data_;
+    bool signal_ = false;  // signal to request processing thread to exit
+  };
+
+  // entry point for insert thread
+  void InsertMain();
+  // insert implementation
+  Status InsertImpl(const Slice& key, const Slice& data);
+  // Create a new cache file
+  Status NewCacheFile();
+  // Get cache directory path
+  std::string GetCachePath() const { return opt_.path + "/cache"; }
+  // Cleanup folder
+  Status CleanupCacheFolder(const std::string& folder);
+
+  // Statistics
+  struct Statistics {
+    HistogramImpl bytes_pipelined_;
+    HistogramImpl bytes_written_;
+    HistogramImpl bytes_read_;
+    HistogramImpl read_hit_latency_;
+    HistogramImpl read_miss_latency_;
+    HistogramImpl write_latency_;
+    std::atomic<uint64_t> cache_hits_{0};
+    std::atomic<uint64_t> cache_misses_{0};
+    std::atomic<uint64_t> cache_errors_{0};
+    std::atomic<uint64_t> insert_dropped_{0};
+
+    double CacheHitPct() const {
+      const auto lookups = cache_hits_ + cache_misses_;
+      return lookups ? 100 * cache_hits_ / static_cast<double>(lookups) : 0.0;
+    }
+
+    double CacheMissPct() const {
+      const auto lookups = cache_hits_ + cache_misses_;
+      return lookups ? 100 * cache_misses_ / static_cast<double>(lookups) : 0.0;
+    }
+  };
+
+  port::RWMutex lock_;                          // Synchronization
+  const PersistentCacheConfig opt_;             // BlockCache options
+  BoundedQueue<InsertOp> insert_ops_;           // Ops waiting for insert
+  ROCKSDB_NAMESPACE::port::Thread insert_th_;   // Insert thread
+  uint32_t writer_cache_id_ = 0;                // Current cache file identifier
+  WriteableCacheFile* cache_file_ = nullptr;    // Current cache file reference
+  CacheWriteBufferAllocator buffer_allocator_;  // Buffer provider
+  ThreadedWriter writer_;                       // Writer threads
+  BlockCacheTierMetadata metadata_;             // Cache meta data manager
+  std::atomic<uint64_t> size_{0};               // Size of the cache
+  Statistics stats_;                            // Statistics
+};
+
+}  // namespace ROCKSDB_NAMESPACE
+
+#endif
diff --git a/src/rocksdb/utilities/persistent_cache/block_cache_tier_file.cc b/src/rocksdb/utilities/persistent_cache/block_cache_tier_file.cc
new file mode 100644
index 000000000..f4f8517ab
--- /dev/null
+++ b/src/rocksdb/utilities/persistent_cache/block_cache_tier_file.cc
@@ -0,0 +1,610 @@
+//  Copyright (c) 2013, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+#ifndef ROCKSDB_LITE
+
+#include "utilities/persistent_cache/block_cache_tier_file.h"
+
+#ifndef OS_WIN
+#include <unistd.h>
+#endif
+#include <functional>
+#include <memory>
+#include <vector>
+
+#include "env/composite_env_wrapper.h"
+#include "logging/logging.h"
+#include "port/port.h"
+#include "rocksdb/system_clock.h"
+#include "util/crc32c.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+//
+// File creation factories
+//
+Status NewWritableCacheFile(Env* const env, const std::string& filepath,
+                            std::unique_ptr<WritableFile>* file,
+                            const bool use_direct_writes = false) {
+  EnvOptions opt;
+  opt.use_direct_writes = use_direct_writes;
+  Status s = env->NewWritableFile(filepath, file, opt);
+  return s;
+}
+
+Status NewRandomAccessCacheFile(const std::shared_ptr<FileSystem>& fs,
+                                const std::string& filepath,
+                                std::unique_ptr<FSRandomAccessFile>* file,
+                                const bool use_direct_reads = true) {
+  assert(fs.get());
+
+  FileOptions opt;
+  opt.use_direct_reads = use_direct_reads;
+  return fs->NewRandomAccessFile(filepath, opt, file, nullptr);
+}
+
+//
+// BlockCacheFile
+//
+Status BlockCacheFile::Delete(uint64_t* size) {
+  assert(env_);
+
+  Status status = env_->GetFileSize(Path(), size);
+  if (!status.ok()) {
+    return status;
+  }
+  return env_->DeleteFile(Path());
+}
+
+//
+// CacheRecord
+//
+// Cache record represents the record on disk
+//
+// +--------+---------+----------+------------+---------------+-------------+
+// | magic  | crc     | key size | value size | key data      | value data  |
+// +--------+---------+----------+------------+---------------+-------------+
+// <-- 4 --><-- 4  --><-- 4   --><-- 4     --><-- key size  --><-- v-size -->
+//
+struct CacheRecordHeader {
+  CacheRecordHeader() : magic_(0), crc_(0), key_size_(0), val_size_(0) {}
+  CacheRecordHeader(const uint32_t magic, const uint32_t key_size,
+                    const uint32_t val_size)
+      : magic_(magic), crc_(0), key_size_(key_size), val_size_(val_size) {}
+
+  uint32_t magic_;
+  uint32_t crc_;
+  uint32_t key_size_;
+  uint32_t val_size_;
+};
+
+struct CacheRecord {
+  CacheRecord() {}
+  CacheRecord(const Slice& key, const Slice& val)
+      : hdr_(MAGIC, static_cast<uint32_t>(key.size()),
+             static_cast<uint32_t>(val.size())),
+        key_(key),
+        val_(val) {
+    hdr_.crc_ = ComputeCRC();
+  }
+
+  uint32_t ComputeCRC() const;
+  bool Serialize(std::vector<CacheWriteBuffer*>* bufs, size_t* woff);
+  bool Deserialize(const Slice& buf);
+
+  static uint32_t CalcSize(const Slice& key, const Slice& val) {
+    return static_cast<uint32_t>(sizeof(CacheRecordHeader) + key.size() +
+                                 val.size());
+  }
+
+  static const uint32_t MAGIC = 0xfefa;
+
+  bool Append(std::vector<CacheWriteBuffer*>* bufs, size_t* woff,
+              const char* data, const size_t size);
+
+  CacheRecordHeader hdr_;
+  Slice key_;
+  Slice val_;
+};
+
+static_assert(sizeof(CacheRecordHeader) == 16, "DataHeader is not aligned");
+
+uint32_t CacheRecord::ComputeCRC() const {
+  uint32_t crc = 0;
+  CacheRecordHeader tmp = hdr_;
+  tmp.crc_ = 0;
+  crc = crc32c::Extend(crc, reinterpret_cast<const char*>(&tmp), sizeof(tmp));
+  crc = crc32c::Extend(crc, reinterpret_cast<const char*>(key_.data()),
+                       key_.size());
+  crc = crc32c::Extend(crc, reinterpret_cast<const char*>(val_.data()),
+                       val_.size());
+  return crc;
+}
+
+bool CacheRecord::Serialize(std::vector<CacheWriteBuffer*>* bufs,
+                            size_t* woff) {
+  assert(bufs->size());
+  return Append(bufs, woff, reinterpret_cast<const char*>(&hdr_),
+                sizeof(hdr_)) &&
+         Append(bufs, woff, reinterpret_cast<const char*>(key_.data()),
+                key_.size()) &&
+         Append(bufs, woff, reinterpret_cast<const char*>(val_.data()),
+                val_.size());
+}
+
+bool CacheRecord::Append(std::vector<CacheWriteBuffer*>* bufs, size_t* woff,
+                         const char* data, const size_t data_size) {
+  assert(*woff < bufs->size());
+
+  const char* p = data;
+  size_t size = data_size;
+
+  while (size && *woff < bufs->size()) {
+    CacheWriteBuffer* buf = (*bufs)[*woff];
+    const size_t free = buf->Free();
+    if (size <= free) {
+      buf->Append(p, size);
+      size = 0;
+    } else {
+      buf->Append(p, free);
+      p += free;
+      size -= free;
+      assert(!buf->Free());
+      assert(buf->Used() == buf->Capacity());
+    }
+
+    if (!buf->Free()) {
+      *woff += 1;
+    }
+  }
+
+  assert(!size);
+
+  return !size;
+}
+
+bool CacheRecord::Deserialize(const Slice& data) {
+  assert(data.size() >= sizeof(CacheRecordHeader));
+  if (data.size() < sizeof(CacheRecordHeader)) {
+    return false;
+  }
+
+  memcpy(&hdr_, data.data(), sizeof(hdr_));
+
+  assert(hdr_.key_size_ + hdr_.val_size_ + sizeof(hdr_) == data.size());
+  if (hdr_.key_size_ + hdr_.val_size_ + sizeof(hdr_) != data.size()) {
+    return false;
+  }
+
+  key_ = Slice(data.data_ + sizeof(hdr_), hdr_.key_size_);
+  val_ = Slice(key_.data_ + hdr_.key_size_, hdr_.val_size_);
+
+  if (!(hdr_.magic_ == MAGIC && ComputeCRC() == hdr_.crc_)) {
+    fprintf(stderr, "** magic %d ** \n", hdr_.magic_);
+    fprintf(stderr, "** key_size %d ** \n", hdr_.key_size_);
+    fprintf(stderr, "** val_size %d ** \n", hdr_.val_size_);
+    fprintf(stderr, "** key %s ** \n", key_.ToString().c_str());
+    fprintf(stderr, "** val %s ** \n", val_.ToString().c_str());
+    for (size_t i = 0; i < hdr_.val_size_; ++i) {
+      fprintf(stderr, "%d.", (uint8_t)val_.data()[i]);
+    }
+    fprintf(stderr, "\n** cksum %d != %d **", hdr_.crc_, ComputeCRC());
+  }
+
+  assert(hdr_.magic_ == MAGIC && ComputeCRC() == hdr_.crc_);
+  return hdr_.magic_ == MAGIC && ComputeCRC() == hdr_.crc_;
+}
+
+//
+// RandomAccessFile
+//
+
+bool RandomAccessCacheFile::Open(const bool enable_direct_reads) {
+  WriteLock _(&rwlock_);
+  return OpenImpl(enable_direct_reads);
+}
+
+bool RandomAccessCacheFile::OpenImpl(const bool enable_direct_reads) {
+  rwlock_.AssertHeld();
+
+  ROCKS_LOG_DEBUG(log_, "Opening cache file %s", Path().c_str());
+  assert(env_);
+
+  std::unique_ptr<FSRandomAccessFile> file;
+  Status status = NewRandomAccessCacheFile(env_->GetFileSystem(), Path(), &file,
+                                           enable_direct_reads);
+  if (!status.ok()) {
+    Error(log_, "Error opening random access file %s. %s", Path().c_str(),
+          status.ToString().c_str());
+    return false;
+  }
+  freader_.reset(new RandomAccessFileReader(std::move(file), Path(),
+                                            env_->GetSystemClock().get()));
+
+  return true;
+}
+
+bool RandomAccessCacheFile::Read(const LBA& lba, Slice* key, Slice* val,
+                                 char* scratch) {
+  ReadLock _(&rwlock_);
+
+  assert(lba.cache_id_ == cache_id_);
+
+  if (!freader_) {
+    return false;
+  }
+
+  Slice result;
+  Status s = freader_->Read(IOOptions(), lba.off_, lba.size_, &result, scratch,
+                            nullptr, Env::IO_TOTAL /* rate_limiter_priority */);
+  if (!s.ok()) {
+    Error(log_, "Error reading from file %s. %s", Path().c_str(),
+          s.ToString().c_str());
+    return false;
+  }
+
+  assert(result.data() == scratch);
+
+  return ParseRec(lba, key, val, scratch);
+}
+
+bool RandomAccessCacheFile::ParseRec(const LBA& lba, Slice* key, Slice* val,
+                                     char* scratch) {
+  Slice data(scratch, lba.size_);
+
+  CacheRecord rec;
+  if (!rec.Deserialize(data)) {
+    assert(!"Error deserializing data");
+    Error(log_, "Error de-serializing record from file %s off %d",
+          Path().c_str(), lba.off_);
+    return false;
+  }
+
+  *key = Slice(rec.key_);
+  *val = Slice(rec.val_);
+
+  return true;
+}
+
+//
+// WriteableCacheFile
+//
+
+WriteableCacheFile::~WriteableCacheFile() {
+  WriteLock _(&rwlock_);
+  if (!eof_) {
+    // This file never flushed. We give priority to shutdown since this is a
+    // cache
+    // TODO(krad): Figure a way to flush the pending data
+    if (file_) {
+      assert(refs_ == 1);
+      --refs_;
+    }
+  }
+  assert(!refs_);
+  ClearBuffers();
+}
+
+bool WriteableCacheFile::Create(const bool /*enable_direct_writes*/,
+                                const bool enable_direct_reads) {
+  WriteLock _(&rwlock_);
+
+  enable_direct_reads_ = enable_direct_reads;
+
+  ROCKS_LOG_DEBUG(log_, "Creating new cache %s (max size is %d B)",
+                  Path().c_str(), max_size_);
+
+  assert(env_);
+
+  Status s = env_->FileExists(Path());
+  if (s.ok()) {
+    ROCKS_LOG_WARN(log_, "File %s already exists. %s", Path().c_str(),
+                   s.ToString().c_str());
+  }
+
+  s = NewWritableCacheFile(env_, Path(), &file_);
+  if (!s.ok()) {
+    ROCKS_LOG_WARN(log_, "Unable to create file %s. %s", Path().c_str(),
+                   s.ToString().c_str());
+    return false;
+  }
+
+  assert(!refs_);
+  ++refs_;
+
+  return true;
+}
+
+bool WriteableCacheFile::Append(const Slice& key, const Slice& val, LBA* lba) {
+  WriteLock _(&rwlock_);
+
+  if (eof_) {
+    // We can't append since the file is full
+    return false;
+  }
+
+  // estimate the space required to store the (key, val)
+  uint32_t rec_size = CacheRecord::CalcSize(key, val);
+
+  if (!ExpandBuffer(rec_size)) {
+    // unable to expand the buffer
+    ROCKS_LOG_DEBUG(log_, "Error expanding buffers. size=%d", rec_size);
+    return false;
+  }
+
+  lba->cache_id_ = cache_id_;
+  lba->off_ = disk_woff_;
+  lba->size_ = rec_size;
+
+  CacheRecord rec(key, val);
+  if (!rec.Serialize(&bufs_, &buf_woff_)) {
+    // unexpected error: unable to serialize the data
+    assert(!"Error serializing record");
+    return false;
+  }
+
+  disk_woff_ += rec_size;
+  eof_ = disk_woff_ >= max_size_;
+
+  // dispatch buffer for flush
+  DispatchBuffer();
+
+  return true;
+}
+
+bool WriteableCacheFile::ExpandBuffer(const size_t size) {
+  rwlock_.AssertHeld();
+  assert(!eof_);
+
+  // determine if there is enough space
+  size_t free = 0;  // compute the free space left in buffer
+  for (size_t i = buf_woff_; i < bufs_.size(); ++i) {
+    free += bufs_[i]->Free();
+    if (size <= free) {
+      // we have enough space in the buffer
+      return true;
+    }
+  }
+
+  // expand the buffer until there is enough space to write `size` bytes
+  assert(free < size);
+  assert(alloc_);
+
+  while (free < size) {
+    CacheWriteBuffer* const buf = alloc_->Allocate();
+    if (!buf) {
+      ROCKS_LOG_DEBUG(log_, "Unable to allocate buffers");
+      return false;
+    }
+
+    size_ += static_cast<uint32_t>(buf->Free());
+    free += buf->Free();
+    bufs_.push_back(buf);
+  }
+
+  assert(free >= size);
+  return true;
+}
+
+void WriteableCacheFile::DispatchBuffer() {
+  rwlock_.AssertHeld();
+
+  assert(bufs_.size());
+  assert(buf_doff_ <= buf_woff_);
+  assert(buf_woff_ <= bufs_.size());
+
+  if (pending_ios_) {
+    return;
+  }
+
+  if (!eof_ && buf_doff_ == buf_woff_) {
+    // dispatch buffer is pointing to write buffer and we haven't hit eof
+    return;
+  }
+
+  assert(eof_ || buf_doff_ < buf_woff_);
+  assert(buf_doff_ < bufs_.size());
+  assert(file_);
+  assert(alloc_);
+
+  auto* buf = bufs_[buf_doff_];
+  const uint64_t file_off = buf_doff_ * alloc_->BufferSize();
+
+  assert(!buf->Free() ||
+         (eof_ && buf_doff_ == buf_woff_ && buf_woff_ < bufs_.size()));
+  // we have reached end of file, and there is space in the last buffer
+  // pad it with zero for direct IO
+  buf->FillTrailingZeros();
+
+  assert(buf->Used() % kFileAlignmentSize == 0);
+
+  writer_->Write(file_.get(), buf, file_off,
+                 std::bind(&WriteableCacheFile::BufferWriteDone, this));
+  pending_ios_++;
+  buf_doff_++;
+}
+
+void WriteableCacheFile::BufferWriteDone() {
+  WriteLock _(&rwlock_);
+
+  assert(bufs_.size());
+
+  pending_ios_--;
+
+  if (buf_doff_ < bufs_.size()) {
+    DispatchBuffer();
+  }
+
+  if (eof_ && buf_doff_ >= bufs_.size() && !pending_ios_) {
+    // end-of-file reached, move to read mode
+    CloseAndOpenForReading();
+  }
+}
+
+void WriteableCacheFile::CloseAndOpenForReading() {
+  // Our env abstraction do not allow reading from a file opened for appending
+  // We need close the file and re-open it for reading
+  Close();
+  RandomAccessCacheFile::OpenImpl(enable_direct_reads_);
+}
+
+bool WriteableCacheFile::ReadBuffer(const LBA& lba, Slice* key, Slice* block,
+                                    char* scratch) {
+  rwlock_.AssertHeld();
+
+  if (!ReadBuffer(lba, scratch)) {
+    Error(log_, "Error reading from buffer. cache=%d off=%d", cache_id_,
+          lba.off_);
+    return false;
+  }
+
+  return ParseRec(lba, key, block, scratch);
+}
+
+bool WriteableCacheFile::ReadBuffer(const LBA& lba, char* data) {
+  rwlock_.AssertHeld();
+
+  assert(lba.off_ < disk_woff_);
+  assert(alloc_);
+
+  // we read from the buffers like reading from a flat file. The list of buffers
+  // are treated as contiguous stream of data
+
+  char* tmp = data;
+  size_t pending_nbytes = lba.size_;
+  // start buffer
+  size_t start_idx = lba.off_ / alloc_->BufferSize();
+  // offset into the start buffer
+  size_t start_off = lba.off_ % alloc_->BufferSize();
+
+  assert(start_idx <= buf_woff_);
+
+  for (size_t i = start_idx; pending_nbytes && i < bufs_.size(); ++i) {
+    assert(i <= buf_woff_);
+    auto* buf = bufs_[i];
+    assert(i == buf_woff_ || !buf->Free());
+    // bytes to write to the buffer
+    size_t nbytes = pending_nbytes > (buf->Used() - start_off)
+                        ? (buf->Used() - start_off)
+                        : pending_nbytes;
+    memcpy(tmp, buf->Data() + start_off, nbytes);
+
+    // left over to be written
+    pending_nbytes -= nbytes;
+    start_off = 0;
+    tmp += nbytes;
+  }
+
+  assert(!pending_nbytes);
+  if (pending_nbytes) {
+    return false;
+  }
+
+  assert(tmp == data + lba.size_);
+  return true;
+}
+
+void WriteableCacheFile::Close() {
+  rwlock_.AssertHeld();
+
+  assert(size_ >= max_size_);
+  assert(disk_woff_ >= max_size_);
+  assert(buf_doff_ == bufs_.size());
+  assert(bufs_.size() - buf_woff_ <= 1);
+  assert(!pending_ios_);
+
+  Info(log_, "Closing file %s. size=%d written=%d", Path().c_str(), size_,
+       disk_woff_);
+
+  ClearBuffers();
+  file_.reset();
+
+  assert(refs_);
+  --refs_;
+}
+
+void WriteableCacheFile::ClearBuffers() {
+  assert(alloc_);
+
+  for (size_t i = 0; i < bufs_.size(); ++i) {
+    alloc_->Deallocate(bufs_[i]);
+  }
+
+  bufs_.clear();
+}
+
+//
+// ThreadedFileWriter implementation
+//
+ThreadedWriter::ThreadedWriter(PersistentCacheTier* const cache,
+                               const size_t qdepth, const size_t io_size)
+    : Writer(cache), io_size_(io_size) {
+  for (size_t i = 0; i < qdepth; ++i) {
+    port::Thread th(&ThreadedWriter::ThreadMain, this);
+    threads_.push_back(std::move(th));
+  }
+}
+
+void ThreadedWriter::Stop() {
+  // notify all threads to exit
+  for (size_t i = 0; i < threads_.size(); ++i) {
+    q_.Push(IO(/*signal=*/true));
+  }
+
+  // wait for all threads to exit
+  for (auto& th : threads_) {
+    th.join();
+    assert(!th.joinable());
+  }
+  threads_.clear();
+}
+
+void ThreadedWriter::Write(WritableFile* const file, CacheWriteBuffer* buf,
+                           const uint64_t file_off,
+                           const std::function<void()> callback) {
+  q_.Push(IO(file, buf, file_off, callback));
+}
+
+void ThreadedWriter::ThreadMain() {
+  while (true) {
+    // Fetch the IO to process
+    IO io(q_.Pop());
+    if (io.signal_) {
+      // that's secret signal to exit
+      break;
+    }
+
+    // Reserve space for writing the buffer
+    while (!cache_->Reserve(io.buf_->Used())) {
+      // We can fail to reserve space if every file in the system
+      // is being currently accessed
+      /* sleep override */
+      SystemClock::Default()->SleepForMicroseconds(1000000);
+    }
+
+    DispatchIO(io);
+
+    io.callback_();
+  }
+}
+
+void ThreadedWriter::DispatchIO(const IO& io) {
+  size_t written = 0;
+  while (written < io.buf_->Used()) {
+    Slice data(io.buf_->Data() + written, io_size_);
+    Status s = io.file_->Append(data);
+    assert(s.ok());
+    if (!s.ok()) {
+      // That is definite IO error to device. There is not much we can
+      // do but ignore the failure. This can lead to corruption of data on
+      // disk, but the cache will skip while reading
+      fprintf(stderr, "Error writing data to file. %s\n", s.ToString().c_str());
+    }
+    written += io_size_;
+  }
+}
+
+}  // namespace ROCKSDB_NAMESPACE
+
+#endif
diff --git a/src/rocksdb/utilities/persistent_cache/block_cache_tier_file.h b/src/rocksdb/utilities/persistent_cache/block_cache_tier_file.h
new file mode 100644
index 000000000..1d265ab74
--- /dev/null
+++ b/src/rocksdb/utilities/persistent_cache/block_cache_tier_file.h
@@ -0,0 +1,293 @@
+//  Copyright (c) 2013, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+#pragma once
+
+#ifndef ROCKSDB_LITE
+
+#include <list>
+#include <memory>
+#include <string>
+#include <vector>
+
+#include "file/random_access_file_reader.h"
+#include "port/port.h"
+#include "rocksdb/comparator.h"
+#include "rocksdb/env.h"
+#include "util/crc32c.h"
+#include "util/mutexlock.h"
+#include "utilities/persistent_cache/block_cache_tier_file_buffer.h"
+#include "utilities/persistent_cache/lrulist.h"
+#include "utilities/persistent_cache/persistent_cache_tier.h"
+#include "utilities/persistent_cache/persistent_cache_util.h"
+
+// The io code path of persistent cache uses pipelined architecture
+//
+// client -> In Queue <-- BlockCacheTier --> Out Queue <-- Writer <--> Kernel
+//
+// This would enable the system to scale for GB/s of throughput which is
+// expected with modern devies like NVM.
+//
+// The file level operations are encapsulated in the following abstractions
+//
+// BlockCacheFile
+//       ^
+//       |
+//       |
+// RandomAccessCacheFile (For reading)
+//       ^
+//       |
+//       |
+// WriteableCacheFile (For writing)
+//
+// Write IO code path :
+//
+namespace ROCKSDB_NAMESPACE {
+
+class WriteableCacheFile;
+struct BlockInfo;
+
+// Represents a logical record on device
+//
+// (L)ogical (B)lock (Address = { cache-file-id, offset, size }
+struct LogicalBlockAddress {
+  LogicalBlockAddress() {}
+  explicit LogicalBlockAddress(const uint32_t cache_id, const uint32_t off,
+                               const uint16_t size)
+      : cache_id_(cache_id), off_(off), size_(size) {}
+
+  uint32_t cache_id_ = 0;
+  uint32_t off_ = 0;
+  uint32_t size_ = 0;
+};
+
+using LBA = LogicalBlockAddress;
+
+// class Writer
+//
+// Writer is the abstraction used for writing data to file. The component can be
+// multithreaded. It is the last step of write pipeline
+class Writer {
+ public:
+  explicit Writer(PersistentCacheTier* const cache) : cache_(cache) {}
+  virtual ~Writer() {}
+
+  // write buffer to file at the given offset
+  virtual void Write(WritableFile* const file, CacheWriteBuffer* buf,
+                     const uint64_t file_off,
+                     const std::function<void()> callback) = 0;
+  // stop the writer
+  virtual void Stop() = 0;
+
+  PersistentCacheTier* const cache_;
+};
+
+// class BlockCacheFile
+//
+// Generic interface to support building file specialized for read/writing
+class BlockCacheFile : public LRUElement<BlockCacheFile> {
+ public:
+  explicit BlockCacheFile(const uint32_t cache_id)
+      : LRUElement<BlockCacheFile>(), cache_id_(cache_id) {}
+
+  explicit BlockCacheFile(Env* const env, const std::string& dir,
+                          const uint32_t cache_id)
+      : LRUElement<BlockCacheFile>(),
+        env_(env),
+        dir_(dir),
+        cache_id_(cache_id) {}
+
+  virtual ~BlockCacheFile() {}
+
+  // append key/value to file and return LBA locator to user
+  virtual bool Append(const Slice& /*key*/, const Slice& /*val*/,
+                      LBA* const /*lba*/) {
+    assert(!"not implemented");
+    return false;
+  }
+
+  // read from the record locator (LBA) and return key, value and status
+  virtual bool Read(const LBA& /*lba*/, Slice* /*key*/, Slice* /*block*/,
+                    char* /*scratch*/) {
+    assert(!"not implemented");
+    return false;
+  }
+
+  // get file path
+  std::string Path() const {
+    return dir_ + "/" + std::to_string(cache_id_) + ".rc";
+  }
+  // get cache ID
+  uint32_t cacheid() const { return cache_id_; }
+  // Add block information to file data
+  // Block information is the list of index reference for this file
+  virtual void Add(BlockInfo* binfo) {
+    WriteLock _(&rwlock_);
+    block_infos_.push_back(binfo);
+  }
+  // get block information
+  std::list<BlockInfo*>& block_infos() { return block_infos_; }
+  // delete file and return the size of the file
+  virtual Status Delete(uint64_t* size);
+
+ protected:
+  port::RWMutex rwlock_;               // synchronization mutex
+  Env* const env_ = nullptr;           // Env for OS
+  const std::string dir_;              // Directory name
+  const uint32_t cache_id_;            // Cache id for the file
+  std::list<BlockInfo*> block_infos_;  // List of index entries mapping to the
+                                       // file content
+};
+
+// class RandomAccessFile
+//
+// Thread safe implementation for reading random data from file
+class RandomAccessCacheFile : public BlockCacheFile {
+ public:
+  explicit RandomAccessCacheFile(Env* const env, const std::string& dir,
+                                 const uint32_t cache_id,
+                                 const std::shared_ptr<Logger>& log)
+      : BlockCacheFile(env, dir, cache_id), log_(log) {}
+
+  virtual ~RandomAccessCacheFile() {}
+
+  // open file for reading
+  bool Open(const bool enable_direct_reads);
+  // read data from the disk
+  bool Read(const LBA& lba, Slice* key, Slice* block, char* scratch) override;
+
+ private:
+  std::unique_ptr<RandomAccessFileReader> freader_;
+
+ protected:
+  bool OpenImpl(const bool enable_direct_reads);
+  bool ParseRec(const LBA& lba, Slice* key, Slice* val, char* scratch);
+
+  std::shared_ptr<Logger> log_;  // log file
+};
+
+// class WriteableCacheFile
+//
+// All writes to the files are cached in buffers. The buffers are flushed to
+// disk as they get filled up. When file size reaches a certain size, a new file
+// will be created provided there is free space
+class WriteableCacheFile : public RandomAccessCacheFile {
+ public:
+  explicit WriteableCacheFile(Env* const env, CacheWriteBufferAllocator* alloc,
+                              Writer* writer, const std::string& dir,
+                              const uint32_t cache_id, const uint32_t max_size,
+                              const std::shared_ptr<Logger>& log)
+      : RandomAccessCacheFile(env, dir, cache_id, log),
+        alloc_(alloc),
+        writer_(writer),
+        max_size_(max_size) {}
+
+  virtual ~WriteableCacheFile();
+
+  // create file on disk
+  bool Create(const bool enable_direct_writes, const bool enable_direct_reads);
+
+  // read data from logical file
+  bool Read(const LBA& lba, Slice* key, Slice* block, char* scratch) override {
+    ReadLock _(&rwlock_);
+    const bool closed = eof_ && bufs_.empty();
+    if (closed) {
+      // the file is closed, read from disk
+      return RandomAccessCacheFile::Read(lba, key, block, scratch);
+    }
+    // file is still being written, read from buffers
+    return ReadBuffer(lba, key, block, scratch);
+  }
+
+  // append data to end of file
+  bool Append(const Slice&, const Slice&, LBA* const) override;
+  // End-of-file
+  bool Eof() const { return eof_; }
+
+ private:
+  friend class ThreadedWriter;
+
+  static const size_t kFileAlignmentSize = 4 * 1024;  // align file size
+
+  bool ReadBuffer(const LBA& lba, Slice* key, Slice* block, char* scratch);
+  bool ReadBuffer(const LBA& lba, char* data);
+  bool ExpandBuffer(const size_t size);
+  void DispatchBuffer();
+  void BufferWriteDone();
+  void CloseAndOpenForReading();
+  void ClearBuffers();
+  void Close();
+
+  // File layout in memory
+  //
+  // +------+------+------+------+------+------+
+  // | b0   | b1   | b2   | b3   | b4   | b5   |
+  // +------+------+------+------+------+------+
+  //        ^                           ^
+  //        |                           |
+  //      buf_doff_                   buf_woff_
+  //   (next buffer to           (next buffer to fill)
+  //   flush to disk)
+  //
+  //  The buffers are flushed to disk serially for a given file
+
+  CacheWriteBufferAllocator* const alloc_ = nullptr;  // Buffer provider
+  Writer* const writer_ = nullptr;                    // File writer thread
+  std::unique_ptr<WritableFile> file_;   // RocksDB Env file abstraction
+  std::vector<CacheWriteBuffer*> bufs_;  // Written buffers
+  uint32_t size_ = 0;                    // Size of the file
+  const uint32_t max_size_;              // Max size of the file
+  bool eof_ = false;                     // End of file
+  uint32_t disk_woff_ = 0;               // Offset to write on disk
+  size_t buf_woff_ = 0;                  // off into bufs_ to write
+  size_t buf_doff_ = 0;                  // off into bufs_ to dispatch
+  size_t pending_ios_ = 0;               // Number of ios to disk in-progress
+  bool enable_direct_reads_ = false;     // Should we enable direct reads
+                                         // when reading from disk
+};
+
+//
+// Abstraction to do writing to device. It is part of pipelined architecture.
+//
+class ThreadedWriter : public Writer {
+ public:
+  // Representation of IO to device
+  struct IO {
+    explicit IO(const bool signal) : signal_(signal) {}
+    explicit IO(WritableFile* const file, CacheWriteBuffer* const buf,
+                const uint64_t file_off, const std::function<void()> callback)
+        : file_(file), buf_(buf), file_off_(file_off), callback_(callback) {}
+
+    IO(const IO&) = default;
+    IO& operator=(const IO&) = default;
+    size_t Size() const { return sizeof(IO); }
+
+    WritableFile* file_ = nullptr;     // File to write to
+    CacheWriteBuffer* buf_ = nullptr;  // buffer to write
+    uint64_t file_off_ = 0;            // file offset
+    bool signal_ = false;              // signal to exit thread loop
+    std::function<void()> callback_;   // Callback on completion
+  };
+
+  explicit ThreadedWriter(PersistentCacheTier* const cache, const size_t qdepth,
+                          const size_t io_size);
+  virtual ~ThreadedWriter() { assert(threads_.empty()); }
+
+  void Stop() override;
+  void Write(WritableFile* const file, CacheWriteBuffer* buf,
+             const uint64_t file_off,
+             const std::function<void()> callback) override;
+
+ private:
+  void ThreadMain();
+  void DispatchIO(const IO& io);
+
+  const size_t io_size_ = 0;
+  BoundedQueue<IO> q_;
+  std::vector<port::Thread> threads_;
+};
+
+}  // namespace ROCKSDB_NAMESPACE
+
+#endif
diff --git a/src/rocksdb/utilities/persistent_cache/block_cache_tier_file_buffer.h b/src/rocksdb/utilities/persistent_cache/block_cache_tier_file_buffer.h
new file mode 100644
index 000000000..d4f02455a
--- /dev/null
+++ b/src/rocksdb/utilities/persistent_cache/block_cache_tier_file_buffer.h
@@ -0,0 +1,127 @@
+//  Copyright (c) 2013, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+#pragma once
+
+#include <list>
+#include <memory>
+#include <string>
+
+#include "memory/arena.h"
+#include "rocksdb/comparator.h"
+#include "util/mutexlock.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+//
+// CacheWriteBuffer
+//
+// Buffer abstraction that can be manipulated via append
+// (not thread safe)
+class CacheWriteBuffer {
+ public:
+  explicit CacheWriteBuffer(const size_t size) : size_(size), pos_(0) {
+    buf_.reset(new char[size_]);
+    assert(!pos_);
+    assert(size_);
+  }
+
+  virtual ~CacheWriteBuffer() {}
+
+  void Append(const char* buf, const size_t size) {
+    assert(pos_ + size <= size_);
+    memcpy(buf_.get() + pos_, buf, size);
+    pos_ += size;
+    assert(pos_ <= size_);
+  }
+
+  void FillTrailingZeros() {
+    assert(pos_ <= size_);
+    memset(buf_.get() + pos_, '0', size_ - pos_);
+    pos_ = size_;
+  }
+
+  void Reset() { pos_ = 0; }
+  size_t Free() const { return size_ - pos_; }
+  size_t Capacity() const { return size_; }
+  size_t Used() const { return pos_; }
+  char* Data() const { return buf_.get(); }
+
+ private:
+  std::unique_ptr<char[]> buf_;
+  const size_t size_;
+  size_t pos_;
+};
+
+//
+// CacheWriteBufferAllocator
+//
+// Buffer pool abstraction(not thread safe)
+//
+class CacheWriteBufferAllocator {
+ public:
+  explicit CacheWriteBufferAllocator(const size_t buffer_size,
+                                     const size_t buffer_count)
+      : cond_empty_(&lock_), buffer_size_(buffer_size) {
+    MutexLock _(&lock_);
+    buffer_size_ = buffer_size;
+    for (uint32_t i = 0; i < buffer_count; i++) {
+      auto* buf = new CacheWriteBuffer(buffer_size_);
+      assert(buf);
+      if (buf) {
+        bufs_.push_back(buf);
+        cond_empty_.Signal();
+      }
+    }
+  }
+
+  virtual ~CacheWriteBufferAllocator() {
+    MutexLock _(&lock_);
+    assert(bufs_.size() * buffer_size_ == Capacity());
+    for (auto* buf : bufs_) {
+      delete buf;
+    }
+    bufs_.clear();
+  }
+
+  CacheWriteBuffer* Allocate() {
+    MutexLock _(&lock_);
+    if (bufs_.empty()) {
+      return nullptr;
+    }
+
+    assert(!bufs_.empty());
+    CacheWriteBuffer* const buf = bufs_.front();
+    bufs_.pop_front();
+    return buf;
+  }
+
+  void Deallocate(CacheWriteBuffer* const buf) {
+    assert(buf);
+    MutexLock _(&lock_);
+    buf->Reset();
+    bufs_.push_back(buf);
+    cond_empty_.Signal();
+  }
+
+  void WaitUntilUsable() {
+    // We are asked to wait till we have buffers available
+    MutexLock _(&lock_);
+    while (bufs_.empty()) {
+      cond_empty_.Wait();
+    }
+  }
+
+  size_t Capacity() const { return bufs_.size() * buffer_size_; }
+  size_t Free() const { return bufs_.size() * buffer_size_; }
+  size_t BufferSize() const { return buffer_size_; }
+
+ private:
+  port::Mutex lock_;                   // Sync lock
+  port::CondVar cond_empty_;           // Condition var for empty buffers
+  size_t buffer_size_;                 // Size of each buffer
+  std::list<CacheWriteBuffer*> bufs_;  // Buffer stash
+};
+
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/utilities/persistent_cache/block_cache_tier_metadata.cc b/src/rocksdb/utilities/persistent_cache/block_cache_tier_metadata.cc
new file mode 100644
index 000000000..d73b5d0b4
--- /dev/null
+++ b/src/rocksdb/utilities/persistent_cache/block_cache_tier_metadata.cc
@@ -0,0 +1,86 @@
+//  Copyright (c) 2013, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+#ifndef ROCKSDB_LITE
+
+#include "utilities/persistent_cache/block_cache_tier_metadata.h"
+
+#include <functional>
+
+namespace ROCKSDB_NAMESPACE {
+
+bool BlockCacheTierMetadata::Insert(BlockCacheFile* file) {
+  return cache_file_index_.Insert(file);
+}
+
+BlockCacheFile* BlockCacheTierMetadata::Lookup(const uint32_t cache_id) {
+  BlockCacheFile* ret = nullptr;
+  BlockCacheFile lookup_key(cache_id);
+  bool ok = cache_file_index_.Find(&lookup_key, &ret);
+  if (ok) {
+    assert(ret->refs_);
+    return ret;
+  }
+  return nullptr;
+}
+
+BlockCacheFile* BlockCacheTierMetadata::Evict() {
+  using std::placeholders::_1;
+  auto fn = std::bind(&BlockCacheTierMetadata::RemoveAllKeys, this, _1);
+  return cache_file_index_.Evict(fn);
+}
+
+void BlockCacheTierMetadata::Clear() {
+  cache_file_index_.Clear([](BlockCacheFile* arg) { delete arg; });
+  block_index_.Clear([](BlockInfo* arg) { delete arg; });
+}
+
+BlockInfo* BlockCacheTierMetadata::Insert(const Slice& key, const LBA& lba) {
+  std::unique_ptr<BlockInfo> binfo(new BlockInfo(key, lba));
+  if (!block_index_.Insert(binfo.get())) {
+    return nullptr;
+  }
+  return binfo.release();
+}
+
+bool BlockCacheTierMetadata::Lookup(const Slice& key, LBA* lba) {
+  BlockInfo lookup_key(key);
+  BlockInfo* block;
+  port::RWMutex* rlock = nullptr;
+  if (!block_index_.Find(&lookup_key, &block, &rlock)) {
+    return false;
+  }
+
+  ReadUnlock _(rlock);
+  assert(block->key_ == key.ToString());
+  if (lba) {
+    *lba = block->lba_;
+  }
+  return true;
+}
+
+BlockInfo* BlockCacheTierMetadata::Remove(const Slice& key) {
+  BlockInfo lookup_key(key);
+  BlockInfo* binfo = nullptr;
+  bool ok __attribute__((__unused__));
+  ok = block_index_.Erase(&lookup_key, &binfo);
+  assert(ok);
+  return binfo;
+}
+
+void BlockCacheTierMetadata::RemoveAllKeys(BlockCacheFile* f) {
+  for (BlockInfo* binfo : f->block_infos()) {
+    BlockInfo* tmp = nullptr;
+    bool status = block_index_.Erase(binfo, &tmp);
+    (void)status;
+    assert(status);
+    assert(tmp == binfo);
+    delete binfo;
+  }
+  f->block_infos().clear();
+}
+
+}  // namespace ROCKSDB_NAMESPACE
+
+#endif
diff --git a/src/rocksdb/utilities/persistent_cache/block_cache_tier_metadata.h b/src/rocksdb/utilities/persistent_cache/block_cache_tier_metadata.h
new file mode 100644
index 000000000..2fcd50105
--- /dev/null
+++ b/src/rocksdb/utilities/persistent_cache/block_cache_tier_metadata.h
@@ -0,0 +1,124 @@
+//  Copyright (c) 2013, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+#pragma once
+
+#ifndef ROCKSDB_LITE
+
+#include <functional>
+#include <string>
+#include <unordered_map>
+
+#include "rocksdb/slice.h"
+#include "utilities/persistent_cache/block_cache_tier_file.h"
+#include "utilities/persistent_cache/hash_table.h"
+#include "utilities/persistent_cache/hash_table_evictable.h"
+#include "utilities/persistent_cache/lrulist.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+//
+// Block Cache Tier Metadata
+//
+// The BlockCacheTierMetadata holds all the metadata associated with block
+// cache. It
+// fundamentally contains 2 indexes and an LRU.
+//
+// Block Cache Index
+//
+// This is a forward index that maps a given key to a LBA (Logical Block
+// Address). LBA is a disk pointer that points to a record on the cache.
+//
+// LBA = { cache-id, offset, size }
+//
+// Cache File Index
+//
+// This is a forward index that maps a given cache-id to a cache file object.
+// Typically you would lookup using LBA and use the object to read or write
+struct BlockInfo {
+  explicit BlockInfo(const Slice& key, const LBA& lba = LBA())
+      : key_(key.ToString()), lba_(lba) {}
+
+  std::string key_;
+  LBA lba_;
+};
+
+class BlockCacheTierMetadata {
+ public:
+  explicit BlockCacheTierMetadata(const uint32_t blocks_capacity = 1024 * 1024,
+                                  const uint32_t cachefile_capacity = 10 * 1024)
+      : cache_file_index_(cachefile_capacity), block_index_(blocks_capacity) {}
+
+  virtual ~BlockCacheTierMetadata() {}
+
+  // Insert a given cache file
+  bool Insert(BlockCacheFile* file);
+
+  // Lookup cache file based on cache_id
+  BlockCacheFile* Lookup(const uint32_t cache_id);
+
+  // Insert block information to block index
+  BlockInfo* Insert(const Slice& key, const LBA& lba);
+  // bool Insert(BlockInfo* binfo);
+
+  // Lookup block information from block index
+  bool Lookup(const Slice& key, LBA* lba);
+
+  // Remove a given from the block index
+  BlockInfo* Remove(const Slice& key);
+
+  // Find and evict a cache file using LRU policy
+  BlockCacheFile* Evict();
+
+  // Clear the metadata contents
+  virtual void Clear();
+
+ protected:
+  // Remove all block information from a given file
+  virtual void RemoveAllKeys(BlockCacheFile* file);
+
+ private:
+  // Cache file index definition
+  //
+  // cache-id => BlockCacheFile
+  struct BlockCacheFileHash {
+    uint64_t operator()(const BlockCacheFile* rec) {
+      return std::hash<uint32_t>()(rec->cacheid());
+    }
+  };
+
+  struct BlockCacheFileEqual {
+    uint64_t operator()(const BlockCacheFile* lhs, const BlockCacheFile* rhs) {
+      return lhs->cacheid() == rhs->cacheid();
+    }
+  };
+
+  using CacheFileIndexType =
+      EvictableHashTable<BlockCacheFile, BlockCacheFileHash,
+                         BlockCacheFileEqual>;
+
+  // Block Lookup Index
+  //
+  // key => LBA
+  struct Hash {
+    size_t operator()(BlockInfo* node) const {
+      return std::hash<std::string>()(node->key_);
+    }
+  };
+
+  struct Equal {
+    size_t operator()(BlockInfo* lhs, BlockInfo* rhs) const {
+      return lhs->key_ == rhs->key_;
+    }
+  };
+
+  using BlockIndexType = HashTable<BlockInfo*, Hash, Equal>;
+
+  CacheFileIndexType cache_file_index_;
+  BlockIndexType block_index_;
+};
+
+}  // namespace ROCKSDB_NAMESPACE
+
+#endif
diff --git a/src/rocksdb/utilities/persistent_cache/hash_table.h b/src/rocksdb/utilities/persistent_cache/hash_table.h
new file mode 100644
index 000000000..b00b294ce
--- /dev/null
+++ b/src/rocksdb/utilities/persistent_cache/hash_table.h
@@ -0,0 +1,239 @@
+//  Copyright (c) 2013, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+//
+#pragma once
+
+#ifndef ROCKSDB_LITE
+
+#include <assert.h>
+
+#include <list>
+#include <vector>
+
+#ifdef OS_LINUX
+#include <sys/mman.h>
+#endif
+
+#include "rocksdb/env.h"
+#include "util/mutexlock.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+// HashTable<T, Hash, Equal>
+//
+// Traditional implementation of hash table with synchronization built on top
+// don't perform very well in multi-core scenarios. This is an implementation
+// designed for multi-core scenarios with high lock contention.
+//
+//                         |<-------- alpha ------------->|
+//               Buckets   Collision list
+//          ---- +----+    +---+---+--- ...... ---+---+---+
+//         /     |    |--->|   |   |              |   |   |
+//        /      +----+    +---+---+--- ...... ---+---+---+
+//       /       |    |
+// Locks/        +----+
+// +--+/         .    .
+// |  |          .    .
+// +--+          .    .
+// |  |          .    .
+// +--+          .    .
+// |  |          .    .
+// +--+          .    .
+//     \         +----+
+//      \        |    |
+//       \       +----+
+//        \      |    |
+//         \---- +----+
+//
+// The lock contention is spread over an array of locks. This helps improve
+// concurrent access. The spine is designed for a certain capacity and load
+// factor. When the capacity planning is done correctly we can expect
+// O(load_factor = 1) insert, access and remove time.
+//
+// Micro benchmark on debug build gives about .5 Million/sec rate of insert,
+// erase and lookup in parallel (total of about 1.5 Million ops/sec). If the
+// blocks were of 4K, the hash table can support  a virtual throughput of
+// 6 GB/s.
+//
+// T      Object type (contains both key and value)
+// Hash   Function that returns an hash from type T
+// Equal  Returns if two objects are equal
+//        (We need explicit equal for pointer type)
+//
+template <class T, class Hash, class Equal>
+class HashTable {
+ public:
+  explicit HashTable(const size_t capacity = 1024 * 1024,
+                     const float load_factor = 2.0, const uint32_t nlocks = 256)
+      : nbuckets_(
+            static_cast<uint32_t>(load_factor ? capacity / load_factor : 0)),
+        nlocks_(nlocks) {
+    // pre-conditions
+    assert(capacity);
+    assert(load_factor);
+    assert(nbuckets_);
+    assert(nlocks_);
+
+    buckets_.reset(new Bucket[nbuckets_]);
+#ifdef OS_LINUX
+    mlock(buckets_.get(), nbuckets_ * sizeof(Bucket));
+#endif
+
+    // initialize locks
+    locks_.reset(new port::RWMutex[nlocks_]);
+#ifdef OS_LINUX
+    mlock(locks_.get(), nlocks_ * sizeof(port::RWMutex));
+#endif
+
+    // post-conditions
+    assert(buckets_);
+    assert(locks_);
+  }
+
+  virtual ~HashTable() { AssertEmptyBuckets(); }
+
+  //
+  // Insert given record to hash table
+  //
+  bool Insert(const T& t) {
+    const uint64_t h = Hash()(t);
+    const uint32_t bucket_idx = h % nbuckets_;
+    const uint32_t lock_idx = bucket_idx % nlocks_;
+
+    WriteLock _(&locks_[lock_idx]);
+    auto& bucket = buckets_[bucket_idx];
+    return Insert(&bucket, t);
+  }
+
+  // Lookup hash table
+  //
+  // Please note that read lock should be held by the caller. This is because
+  // the caller owns the data, and should hold the read lock as long as he
+  // operates on the data.
+  bool Find(const T& t, T* ret, port::RWMutex** ret_lock) {
+    const uint64_t h = Hash()(t);
+    const uint32_t bucket_idx = h % nbuckets_;
+    const uint32_t lock_idx = bucket_idx % nlocks_;
+
+    port::RWMutex& lock = locks_[lock_idx];
+    lock.ReadLock();
+
+    auto& bucket = buckets_[bucket_idx];
+    if (Find(&bucket, t, ret)) {
+      *ret_lock = &lock;
+      return true;
+    }
+
+    lock.ReadUnlock();
+    return false;
+  }
+
+  //
+  // Erase a given key from the hash table
+  //
+  bool Erase(const T& t, T* ret) {
+    const uint64_t h = Hash()(t);
+    const uint32_t bucket_idx = h % nbuckets_;
+    const uint32_t lock_idx = bucket_idx % nlocks_;
+
+    WriteLock _(&locks_[lock_idx]);
+
+    auto& bucket = buckets_[bucket_idx];
+    return Erase(&bucket, t, ret);
+  }
+
+  // Fetch the mutex associated with a key
+  // This call is used to hold the lock for a given data for extended period of
+  // time.
+  port::RWMutex* GetMutex(const T& t) {
+    const uint64_t h = Hash()(t);
+    const uint32_t bucket_idx = h % nbuckets_;
+    const uint32_t lock_idx = bucket_idx % nlocks_;
+
+    return &locks_[lock_idx];
+  }
+
+  void Clear(void (*fn)(T)) {
+    for (uint32_t i = 0; i < nbuckets_; ++i) {
+      const uint32_t lock_idx = i % nlocks_;
+      WriteLock _(&locks_[lock_idx]);
+      for (auto& t : buckets_[i].list_) {
+        (*fn)(t);
+      }
+      buckets_[i].list_.clear();
+    }
+  }
+
+ protected:
+  // Models bucket of keys that hash to the same bucket number
+  struct Bucket {
+    std::list<T> list_;
+  };
+
+  // Substitute for std::find with custom comparator operator
+  typename std::list<T>::iterator Find(std::list<T>* list, const T& t) {
+    for (auto it = list->begin(); it != list->end(); ++it) {
+      if (Equal()(*it, t)) {
+        return it;
+      }
+    }
+    return list->end();
+  }
+
+  bool Insert(Bucket* bucket, const T& t) {
+    // Check if the key already exists
+    auto it = Find(&bucket->list_, t);
+    if (it != bucket->list_.end()) {
+      return false;
+    }
+
+    // insert to bucket
+    bucket->list_.push_back(t);
+    return true;
+  }
+
+  bool Find(Bucket* bucket, const T& t, T* ret) {
+    auto it = Find(&bucket->list_, t);
+    if (it != bucket->list_.end()) {
+      if (ret) {
+        *ret = *it;
+      }
+      return true;
+    }
+    return false;
+  }
+
+  bool Erase(Bucket* bucket, const T& t, T* ret) {
+    auto it = Find(&bucket->list_, t);
+    if (it != bucket->list_.end()) {
+      if (ret) {
+        *ret = *it;
+      }
+
+      bucket->list_.erase(it);
+      return true;
+    }
+    return false;
+  }
+
+  // assert that all buckets are empty
+  void AssertEmptyBuckets() {
+#ifndef NDEBUG
+    for (size_t i = 0; i < nbuckets_; ++i) {
+      WriteLock _(&locks_[i % nlocks_]);
+      assert(buckets_[i].list_.empty());
+    }
+#endif
+  }
+
+  const uint32_t nbuckets_;                 // No. of buckets in the spine
+  std::unique_ptr<Bucket[]> buckets_;       // Spine of the hash buckets
+  const uint32_t nlocks_;                   // No. of locks
+  std::unique_ptr<port::RWMutex[]> locks_;  // Granular locks
+};
+
+}  // namespace ROCKSDB_NAMESPACE
+
+#endif
diff --git a/src/rocksdb/utilities/persistent_cache/hash_table_bench.cc b/src/rocksdb/utilities/persistent_cache/hash_table_bench.cc
new file mode 100644
index 000000000..74d7e2edf
--- /dev/null
+++ b/src/rocksdb/utilities/persistent_cache/hash_table_bench.cc
@@ -0,0 +1,310 @@
+//  Copyright (c) 2013, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+//
+
+#if !defined(OS_WIN) && !defined(ROCKSDB_LITE)
+
+#ifndef GFLAGS
+#include <cstdio>
+int main() { fprintf(stderr, "Please install gflags to run tools\n"); }
+#else
+
+#include <sys/time.h>
+#include <unistd.h>
+
+#include <atomic>
+#include <functional>
+#include <string>
+#include <unordered_map>
+
+#include "port/port_posix.h"
+#include "port/sys_time.h"
+#include "rocksdb/env.h"
+#include "util/gflags_compat.h"
+#include "util/mutexlock.h"
+#include "util/random.h"
+#include "utilities/persistent_cache/hash_table.h"
+
+using std::string;
+
+DEFINE_int32(nsec, 10, "nsec");
+DEFINE_int32(nthread_write, 1, "insert %");
+DEFINE_int32(nthread_read, 1, "lookup %");
+DEFINE_int32(nthread_erase, 1, "erase %");
+
+namespace ROCKSDB_NAMESPACE {
+
+//
+// HashTableImpl interface
+//
+// Abstraction of a hash table implementation
+template <class Key, class Value>
+class HashTableImpl {
+ public:
+  virtual ~HashTableImpl() {}
+
+  virtual bool Insert(const Key& key, const Value& val) = 0;
+  virtual bool Erase(const Key& key) = 0;
+  virtual bool Lookup(const Key& key, Value* val) = 0;
+};
+
+// HashTableBenchmark
+//
+// Abstraction to test a given hash table implementation. The test mostly
+// focus on insert, lookup and erase. The test can operate in test mode and
+// benchmark mode.
+class HashTableBenchmark {
+ public:
+  explicit HashTableBenchmark(HashTableImpl<size_t, std::string>* impl,
+                              const size_t sec = 10,
+                              const size_t nthread_write = 1,
+                              const size_t nthread_read = 1,
+                              const size_t nthread_erase = 1)
+      : impl_(impl),
+        sec_(sec),
+        ninserts_(0),
+        nreads_(0),
+        nerases_(0),
+        nerases_failed_(0),
+        quit_(false) {
+    Prepop();
+
+    StartThreads(nthread_write, WriteMain);
+    StartThreads(nthread_read, ReadMain);
+    StartThreads(nthread_erase, EraseMain);
+
+    uint64_t start = NowInMillSec();
+    while (!quit_) {
+      quit_ = NowInMillSec() - start > sec_ * 1000;
+      /* sleep override */ sleep(1);
+    }
+
+    Env* env = Env::Default();
+    env->WaitForJoin();
+
+    if (sec_) {
+      printf("Result \n");
+      printf("====== \n");
+      printf("insert/sec = %f \n", ninserts_ / static_cast<double>(sec_));
+      printf("read/sec = %f \n", nreads_ / static_cast<double>(sec_));
+      printf("erases/sec = %f \n", nerases_ / static_cast<double>(sec_));
+      const uint64_t ops = ninserts_ + nreads_ + nerases_;
+      printf("ops/sec = %f \n", ops / static_cast<double>(sec_));
+      printf("erase fail = %d (%f%%)\n", static_cast<int>(nerases_failed_),
+             static_cast<float>(nerases_failed_ / nerases_ * 100));
+      printf("====== \n");
+    }
+  }
+
+  void RunWrite() {
+    while (!quit_) {
+      size_t k = insert_key_++;
+      std::string tmp(1000, k % 255);
+      bool status = impl_->Insert(k, tmp);
+      assert(status);
+      ninserts_++;
+    }
+  }
+
+  void RunRead() {
+    Random64 rgen(time(nullptr));
+    while (!quit_) {
+      std::string s;
+      size_t k = rgen.Next() % max_prepop_key;
+      bool status = impl_->Lookup(k, &s);
+      assert(status);
+      assert(s == std::string(1000, k % 255));
+      nreads_++;
+    }
+  }
+
+  void RunErase() {
+    while (!quit_) {
+      size_t k = erase_key_++;
+      bool status = impl_->Erase(k);
+      nerases_failed_ += !status;
+      nerases_++;
+    }
+  }
+
+ private:
+  // Start threads for a given function
+  void StartThreads(const size_t n, void (*fn)(void*)) {
+    Env* env = Env::Default();
+    for (size_t i = 0; i < n; ++i) {
+      env->StartThread(fn, this);
+    }
+  }
+
+  // Prepop the hash table with 1M keys
+  void Prepop() {
+    for (size_t i = 0; i < max_prepop_key; ++i) {
+      bool status = impl_->Insert(i, std::string(1000, i % 255));
+      assert(status);
+    }
+
+    erase_key_ = insert_key_ = max_prepop_key;
+
+    for (size_t i = 0; i < 10 * max_prepop_key; ++i) {
+      bool status = impl_->Insert(insert_key_++, std::string(1000, 'x'));
+      assert(status);
+    }
+  }
+
+  static uint64_t NowInMillSec() {
+    port::TimeVal tv;
+    port::GetTimeOfDay(&tv, /*tz=*/nullptr);
+    return tv.tv_sec * 1000 + tv.tv_usec / 1000;
+  }
+
+  //
+  //  Wrapper functions for thread entry
+  //
+  static void WriteMain(void* args) {
+    reinterpret_cast<HashTableBenchmark*>(args)->RunWrite();
+  }
+
+  static void ReadMain(void* args) {
+    reinterpret_cast<HashTableBenchmark*>(args)->RunRead();
+  }
+
+  static void EraseMain(void* args) {
+    reinterpret_cast<HashTableBenchmark*>(args)->RunErase();
+  }
+
+  HashTableImpl<size_t, std::string>* impl_;         // Implementation to test
+  const size_t sec_;                                 // Test time
+  const size_t max_prepop_key = 1ULL * 1024 * 1024;  // Max prepop key
+  std::atomic<size_t> insert_key_;                   // Last inserted key
+  std::atomic<size_t> erase_key_;                    // Erase key
+  std::atomic<size_t> ninserts_;                     // Number of inserts
+  std::atomic<size_t> nreads_;                       // Number of reads
+  std::atomic<size_t> nerases_;                      // Number of erases
+  std::atomic<size_t> nerases_failed_;               // Number of erases failed
+  bool quit_;  // Should the threads quit ?
+};
+
+//
+// SimpleImpl
+// Lock safe unordered_map implementation
+class SimpleImpl : public HashTableImpl<size_t, string> {
+ public:
+  bool Insert(const size_t& key, const string& val) override {
+    WriteLock _(&rwlock_);
+    map_.insert(make_pair(key, val));
+    return true;
+  }
+
+  bool Erase(const size_t& key) override {
+    WriteLock _(&rwlock_);
+    auto it = map_.find(key);
+    if (it == map_.end()) {
+      return false;
+    }
+    map_.erase(it);
+    return true;
+  }
+
+  bool Lookup(const size_t& key, string* val) override {
+    ReadLock _(&rwlock_);
+    auto it = map_.find(key);
+    if (it != map_.end()) {
+      *val = it->second;
+    }
+    return it != map_.end();
+  }
+
+ private:
+  port::RWMutex rwlock_;
+  std::unordered_map<size_t, string> map_;
+};
+
+//
+// GranularLockImpl
+// Thread safe custom RocksDB implementation of hash table with granular
+// locking
+class GranularLockImpl : public HashTableImpl<size_t, string> {
+ public:
+  bool Insert(const size_t& key, const string& val) override {
+    Node n(key, val);
+    return impl_.Insert(n);
+  }
+
+  bool Erase(const size_t& key) override {
+    Node n(key, string());
+    return impl_.Erase(n, nullptr);
+  }
+
+  bool Lookup(const size_t& key, string* val) override {
+    Node n(key, string());
+    port::RWMutex* rlock;
+    bool status = impl_.Find(n, &n, &rlock);
+    if (status) {
+      ReadUnlock _(rlock);
+      *val = n.val_;
+    }
+    return status;
+  }
+
+ private:
+  struct Node {
+    explicit Node(const size_t key, const string& val) : key_(key), val_(val) {}
+
+    size_t key_ = 0;
+    string val_;
+  };
+
+  struct Hash {
+    uint64_t operator()(const Node& node) {
+      return std::hash<uint64_t>()(node.key_);
+    }
+  };
+
+  struct Equal {
+    bool operator()(const Node& lhs, const Node& rhs) {
+      return lhs.key_ == rhs.key_;
+    }
+  };
+
+  HashTable<Node, Hash, Equal> impl_;
+};
+
+}  // namespace ROCKSDB_NAMESPACE
+
+//
+// main
+//
+int main(int argc, char** argv) {
+  GFLAGS_NAMESPACE::SetUsageMessage(std::string("\nUSAGE:\n") +
+                                    std::string(argv[0]) + " [OPTIONS]...");
+  GFLAGS_NAMESPACE::ParseCommandLineFlags(&argc, &argv, false);
+
+  //
+  // Micro benchmark unordered_map
+  //
+  printf("Micro benchmarking std::unordered_map \n");
+  {
+    ROCKSDB_NAMESPACE::SimpleImpl impl;
+    ROCKSDB_NAMESPACE::HashTableBenchmark _(
+        &impl, FLAGS_nsec, FLAGS_nthread_write, FLAGS_nthread_read,
+        FLAGS_nthread_erase);
+  }
+  //
+  // Micro benchmark scalable hash table
+  //
+  printf("Micro benchmarking scalable hash map \n");
+  {
+    ROCKSDB_NAMESPACE::GranularLockImpl impl;
+    ROCKSDB_NAMESPACE::HashTableBenchmark _(
+        &impl, FLAGS_nsec, FLAGS_nthread_write, FLAGS_nthread_read,
+        FLAGS_nthread_erase);
+  }
+
+  return 0;
+}
+#endif  // #ifndef GFLAGS
+#else
+int main(int /*argc*/, char** /*argv*/) { return 0; }
+#endif
diff --git a/src/rocksdb/utilities/persistent_cache/hash_table_evictable.h b/src/rocksdb/utilities/persistent_cache/hash_table_evictable.h
new file mode 100644
index 000000000..e10939b2f
--- /dev/null
+++ b/src/rocksdb/utilities/persistent_cache/hash_table_evictable.h
@@ -0,0 +1,168 @@
+//  Copyright (c) 2013, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+//
+#pragma once
+
+#ifndef ROCKSDB_LITE
+
+#include <functional>
+
+#include "util/random.h"
+#include "utilities/persistent_cache/hash_table.h"
+#include "utilities/persistent_cache/lrulist.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+// Evictable Hash Table
+//
+// Hash table index where least accessed (or one of the least accessed) elements
+// can be evicted.
+//
+// Please note EvictableHashTable can only be created for pointer type objects
+template <class T, class Hash, class Equal>
+class EvictableHashTable : private HashTable<T*, Hash, Equal> {
+ public:
+  using hash_table = HashTable<T*, Hash, Equal>;
+
+  explicit EvictableHashTable(const size_t capacity = 1024 * 1024,
+                              const float load_factor = 2.0,
+                              const uint32_t nlocks = 256)
+      : HashTable<T*, Hash, Equal>(capacity, load_factor, nlocks),
+        lru_lists_(new LRUList<T>[hash_table::nlocks_]) {
+    assert(lru_lists_);
+  }
+
+  virtual ~EvictableHashTable() { AssertEmptyLRU(); }
+
+  //
+  // Insert given record to hash table (and LRU list)
+  //
+  bool Insert(T* t) {
+    const uint64_t h = Hash()(t);
+    typename hash_table::Bucket& bucket = GetBucket(h);
+    LRUListType& lru = GetLRUList(h);
+    port::RWMutex& lock = GetMutex(h);
+
+    WriteLock _(&lock);
+    if (hash_table::Insert(&bucket, t)) {
+      lru.Push(t);
+      return true;
+    }
+    return false;
+  }
+
+  //
+  // Lookup hash table
+  //
+  // Please note that read lock should be held by the caller. This is because
+  // the caller owns the data, and should hold the read lock as long as he
+  // operates on the data.
+  bool Find(T* t, T** ret) {
+    const uint64_t h = Hash()(t);
+    typename hash_table::Bucket& bucket = GetBucket(h);
+    LRUListType& lru = GetLRUList(h);
+    port::RWMutex& lock = GetMutex(h);
+
+    ReadLock _(&lock);
+    if (hash_table::Find(&bucket, t, ret)) {
+      ++(*ret)->refs_;
+      lru.Touch(*ret);
+      return true;
+    }
+    return false;
+  }
+
+  //
+  // Evict one of the least recently used object
+  //
+  T* Evict(const std::function<void(T*)>& fn = nullptr) {
+    uint32_t random = Random::GetTLSInstance()->Next();
+    const size_t start_idx = random % hash_table::nlocks_;
+    T* t = nullptr;
+
+    // iterate from start_idx .. 0 .. start_idx
+    for (size_t i = 0; !t && i < hash_table::nlocks_; ++i) {
+      const size_t idx = (start_idx + i) % hash_table::nlocks_;
+
+      WriteLock _(&hash_table::locks_[idx]);
+      LRUListType& lru = lru_lists_[idx];
+      if (!lru.IsEmpty() && (t = lru.Pop()) != nullptr) {
+        assert(!t->refs_);
+        // We got an item to evict, erase from the bucket
+        const uint64_t h = Hash()(t);
+        typename hash_table::Bucket& bucket = GetBucket(h);
+        T* tmp = nullptr;
+        bool status = hash_table::Erase(&bucket, t, &tmp);
+        assert(t == tmp);
+        (void)status;
+        assert(status);
+        if (fn) {
+          fn(t);
+        }
+        break;
+      }
+      assert(!t);
+    }
+    return t;
+  }
+
+  void Clear(void (*fn)(T*)) {
+    for (uint32_t i = 0; i < hash_table::nbuckets_; ++i) {
+      const uint32_t lock_idx = i % hash_table::nlocks_;
+      WriteLock _(&hash_table::locks_[lock_idx]);
+      auto& lru_list = lru_lists_[lock_idx];
+      auto& bucket = hash_table::buckets_[i];
+      for (auto* t : bucket.list_) {
+        lru_list.Unlink(t);
+        (*fn)(t);
+      }
+      bucket.list_.clear();
+    }
+    // make sure that all LRU lists are emptied
+    AssertEmptyLRU();
+  }
+
+  void AssertEmptyLRU() {
+#ifndef NDEBUG
+    for (uint32_t i = 0; i < hash_table::nlocks_; ++i) {
+      WriteLock _(&hash_table::locks_[i]);
+      auto& lru_list = lru_lists_[i];
+      assert(lru_list.IsEmpty());
+    }
+#endif
+  }
+
+  //
+  // Fetch the mutex associated with a key
+  // This call is used to hold the lock for a given data for extended period of
+  // time.
+  port::RWMutex* GetMutex(T* t) { return hash_table::GetMutex(t); }
+
+ private:
+  using LRUListType = LRUList<T>;
+
+  typename hash_table::Bucket& GetBucket(const uint64_t h) {
+    const uint32_t bucket_idx = h % hash_table::nbuckets_;
+    return hash_table::buckets_[bucket_idx];
+  }
+
+  LRUListType& GetLRUList(const uint64_t h) {
+    const uint32_t bucket_idx = h % hash_table::nbuckets_;
+    const uint32_t lock_idx = bucket_idx % hash_table::nlocks_;
+    return lru_lists_[lock_idx];
+  }
+
+  port::RWMutex& GetMutex(const uint64_t h) {
+    const uint32_t bucket_idx = h % hash_table::nbuckets_;
+    const uint32_t lock_idx = bucket_idx % hash_table::nlocks_;
+    return hash_table::locks_[lock_idx];
+  }
+
+  std::unique_ptr<LRUListType[]> lru_lists_;
+};
+
+}  // namespace ROCKSDB_NAMESPACE
+
+#endif
diff --git a/src/rocksdb/utilities/persistent_cache/hash_table_test.cc b/src/rocksdb/utilities/persistent_cache/hash_table_test.cc
new file mode 100644
index 000000000..2f6387f5f
--- /dev/null
+++ b/src/rocksdb/utilities/persistent_cache/hash_table_test.cc
@@ -0,0 +1,163 @@
+//  Copyright (c) 2013, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+//
+#include "utilities/persistent_cache/hash_table.h"
+
+#include <stdlib.h>
+
+#include <iostream>
+#include <set>
+#include <string>
+
+#include "db/db_test_util.h"
+#include "memory/arena.h"
+#include "test_util/testharness.h"
+#include "util/random.h"
+#include "utilities/persistent_cache/hash_table_evictable.h"
+
+#ifndef ROCKSDB_LITE
+
+namespace ROCKSDB_NAMESPACE {
+
+struct HashTableTest : public testing::Test {
+  ~HashTableTest() override { map_.Clear(&HashTableTest::ClearNode); }
+
+  struct Node {
+    Node() {}
+    explicit Node(const uint64_t key, const std::string& val = std::string())
+        : key_(key), val_(val) {}
+
+    uint64_t key_ = 0;
+    std::string val_;
+  };
+
+  struct Equal {
+    bool operator()(const Node& lhs, const Node& rhs) {
+      return lhs.key_ == rhs.key_;
+    }
+  };
+
+  struct Hash {
+    uint64_t operator()(const Node& node) {
+      return std::hash<uint64_t>()(node.key_);
+    }
+  };
+
+  static void ClearNode(Node /*node*/) {}
+
+  HashTable<Node, Hash, Equal> map_;
+};
+
+struct EvictableHashTableTest : public testing::Test {
+  ~EvictableHashTableTest() override {
+    map_.Clear(&EvictableHashTableTest::ClearNode);
+  }
+
+  struct Node : LRUElement<Node> {
+    Node() {}
+    explicit Node(const uint64_t key, const std::string& val = std::string())
+        : key_(key), val_(val) {}
+
+    uint64_t key_ = 0;
+    std::string val_;
+    std::atomic<uint32_t> refs_{0};
+  };
+
+  struct Equal {
+    bool operator()(const Node* lhs, const Node* rhs) {
+      return lhs->key_ == rhs->key_;
+    }
+  };
+
+  struct Hash {
+    uint64_t operator()(const Node* node) {
+      return std::hash<uint64_t>()(node->key_);
+    }
+  };
+
+  static void ClearNode(Node* /*node*/) {}
+
+  EvictableHashTable<Node, Hash, Equal> map_;
+};
+
+TEST_F(HashTableTest, TestInsert) {
+  const uint64_t max_keys = 1024 * 1024;
+
+  // insert
+  for (uint64_t k = 0; k < max_keys; ++k) {
+    map_.Insert(Node(k, std::string(1000, k % 255)));
+  }
+
+  // verify
+  for (uint64_t k = 0; k < max_keys; ++k) {
+    Node val;
+    port::RWMutex* rlock = nullptr;
+    assert(map_.Find(Node(k), &val, &rlock));
+    rlock->ReadUnlock();
+    assert(val.val_ == std::string(1000, k % 255));
+  }
+}
+
+TEST_F(HashTableTest, TestErase) {
+  const uint64_t max_keys = 1024 * 1024;
+  // insert
+  for (uint64_t k = 0; k < max_keys; ++k) {
+    map_.Insert(Node(k, std::string(1000, k % 255)));
+  }
+
+  auto rand = Random64(time(nullptr));
+  // erase a few keys randomly
+  std::set<uint64_t> erased;
+  for (int i = 0; i < 1024; ++i) {
+    uint64_t k = rand.Next() % max_keys;
+    if (erased.find(k) != erased.end()) {
+      continue;
+    }
+    assert(map_.Erase(Node(k), /*ret=*/nullptr));
+    erased.insert(k);
+  }
+
+  // verify
+  for (uint64_t k = 0; k < max_keys; ++k) {
+    Node val;
+    port::RWMutex* rlock = nullptr;
+    bool status = map_.Find(Node(k), &val, &rlock);
+    if (erased.find(k) == erased.end()) {
+      assert(status);
+      rlock->ReadUnlock();
+      assert(val.val_ == std::string(1000, k % 255));
+    } else {
+      assert(!status);
+    }
+  }
+}
+
+TEST_F(EvictableHashTableTest, TestEvict) {
+  const uint64_t max_keys = 1024 * 1024;
+
+  // insert
+  for (uint64_t k = 0; k < max_keys; ++k) {
+    map_.Insert(new Node(k, std::string(1000, k % 255)));
+  }
+
+  // verify
+  for (uint64_t k = 0; k < max_keys; ++k) {
+    Node* val = map_.Evict();
+    // unfortunately we can't predict eviction value since it is from any one of
+    // the lock stripe
+    assert(val);
+    assert(val->val_ == std::string(1000, val->key_ % 255));
+    delete val;
+  }
+}
+
+}  // namespace ROCKSDB_NAMESPACE
+#endif
+
+int main(int argc, char** argv) {
+  ROCKSDB_NAMESPACE::port::InstallStackTraceHandler();
+  ::testing::InitGoogleTest(&argc, argv);
+  return RUN_ALL_TESTS();
+}
diff --git a/src/rocksdb/utilities/persistent_cache/lrulist.h b/src/rocksdb/utilities/persistent_cache/lrulist.h
new file mode 100644
index 000000000..a608890fc
--- /dev/null
+++ b/src/rocksdb/utilities/persistent_cache/lrulist.h
@@ -0,0 +1,174 @@
+//  Copyright (c) 2013, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+//
+#pragma once
+
+#ifndef ROCKSDB_LITE
+
+#include <atomic>
+
+#include "util/mutexlock.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+// LRU element definition
+//
+// Any object that needs to be part of the LRU algorithm should extend this
+// class
+template <class T>
+struct LRUElement {
+  explicit LRUElement() : next_(nullptr), prev_(nullptr), refs_(0) {}
+
+  virtual ~LRUElement() { assert(!refs_); }
+
+  T* next_;
+  T* prev_;
+  std::atomic<size_t> refs_;
+};
+
+// LRU implementation
+//
+// In place LRU implementation. There is no copy or allocation involved when
+// inserting or removing an element. This makes the data structure slim
+template <class T>
+class LRUList {
+ public:
+  virtual ~LRUList() {
+    MutexLock _(&lock_);
+    assert(!head_);
+    assert(!tail_);
+  }
+
+  // Push element into the LRU at the cold end
+  inline void Push(T* const t) {
+    assert(t);
+    assert(!t->next_);
+    assert(!t->prev_);
+
+    MutexLock _(&lock_);
+
+    assert((!head_ && !tail_) || (head_ && tail_));
+    assert(!head_ || !head_->prev_);
+    assert(!tail_ || !tail_->next_);
+
+    t->next_ = head_;
+    if (head_) {
+      head_->prev_ = t;
+    }
+
+    head_ = t;
+    if (!tail_) {
+      tail_ = t;
+    }
+  }
+
+  // Unlink the element from the LRU
+  inline void Unlink(T* const t) {
+    MutexLock _(&lock_);
+    UnlinkImpl(t);
+  }
+
+  // Evict an element from the LRU
+  inline T* Pop() {
+    MutexLock _(&lock_);
+
+    assert(tail_ && head_);
+    assert(!tail_->next_);
+    assert(!head_->prev_);
+
+    T* t = head_;
+    while (t && t->refs_) {
+      t = t->next_;
+    }
+
+    if (!t) {
+      // nothing can be evicted
+      return nullptr;
+    }
+
+    assert(!t->refs_);
+
+    // unlike the element
+    UnlinkImpl(t);
+    return t;
+  }
+
+  // Move the element from the front of the list to the back of the list
+  inline void Touch(T* const t) {
+    MutexLock _(&lock_);
+    UnlinkImpl(t);
+    PushBackImpl(t);
+  }
+
+  // Check if the LRU is empty
+  inline bool IsEmpty() const {
+    MutexLock _(&lock_);
+    return !head_ && !tail_;
+  }
+
+ private:
+  // Unlink an element from the LRU
+  void UnlinkImpl(T* const t) {
+    assert(t);
+
+    lock_.AssertHeld();
+
+    assert(head_ && tail_);
+    assert(t->prev_ || head_ == t);
+    assert(t->next_ || tail_ == t);
+
+    if (t->prev_) {
+      t->prev_->next_ = t->next_;
+    }
+    if (t->next_) {
+      t->next_->prev_ = t->prev_;
+    }
+
+    if (tail_ == t) {
+      tail_ = tail_->prev_;
+    }
+    if (head_ == t) {
+      head_ = head_->next_;
+    }
+
+    t->next_ = t->prev_ = nullptr;
+  }
+
+  // Insert an element at the hot end
+  inline void PushBack(T* const t) {
+    MutexLock _(&lock_);
+    PushBackImpl(t);
+  }
+
+  inline void PushBackImpl(T* const t) {
+    assert(t);
+    assert(!t->next_);
+    assert(!t->prev_);
+
+    lock_.AssertHeld();
+
+    assert((!head_ && !tail_) || (head_ && tail_));
+    assert(!head_ || !head_->prev_);
+    assert(!tail_ || !tail_->next_);
+
+    t->prev_ = tail_;
+    if (tail_) {
+      tail_->next_ = t;
+    }
+
+    tail_ = t;
+    if (!head_) {
+      head_ = tail_;
+    }
+  }
+
+  mutable port::Mutex lock_;  // synchronization primitive
+  T* head_ = nullptr;         // front (cold)
+  T* tail_ = nullptr;         // back (hot)
+};
+
+}  // namespace ROCKSDB_NAMESPACE
+
+#endif
diff --git a/src/rocksdb/utilities/persistent_cache/persistent_cache_bench.cc b/src/rocksdb/utilities/persistent_cache/persistent_cache_bench.cc
new file mode 100644
index 000000000..9d6e15d6b
--- /dev/null
+++ b/src/rocksdb/utilities/persistent_cache/persistent_cache_bench.cc
@@ -0,0 +1,359 @@
+//  Copyright (c) 2013, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+//
+#ifndef ROCKSDB_LITE
+
+#ifndef GFLAGS
+#include <cstdio>
+int main() { fprintf(stderr, "Please install gflags to run tools\n"); }
+#else
+#include <atomic>
+#include <functional>
+#include <memory>
+#include <sstream>
+#include <unordered_map>
+
+#include "monitoring/histogram.h"
+#include "port/port.h"
+#include "rocksdb/env.h"
+#include "rocksdb/system_clock.h"
+#include "table/block_based/block_builder.h"
+#include "util/gflags_compat.h"
+#include "util/mutexlock.h"
+#include "util/stop_watch.h"
+#include "utilities/persistent_cache/block_cache_tier.h"
+#include "utilities/persistent_cache/persistent_cache_tier.h"
+#include "utilities/persistent_cache/volatile_tier_impl.h"
+
+DEFINE_int32(nsec, 10, "nsec");
+DEFINE_int32(nthread_write, 1, "Insert threads");
+DEFINE_int32(nthread_read, 1, "Lookup threads");
+DEFINE_string(path, "/tmp/microbench/blkcache", "Path for cachefile");
+DEFINE_string(log_path, "/tmp/log", "Path for the log file");
+DEFINE_uint64(cache_size, std::numeric_limits<uint64_t>::max(), "Cache size");
+DEFINE_int32(iosize, 4 * 1024, "Read IO size");
+DEFINE_int32(writer_iosize, 4 * 1024, "File writer IO size");
+DEFINE_int32(writer_qdepth, 1, "File writer qdepth");
+DEFINE_bool(enable_pipelined_writes, false, "Enable async writes");
+DEFINE_string(cache_type, "block_cache",
+              "Cache type. (block_cache, volatile, tiered)");
+DEFINE_bool(benchmark, false, "Benchmark mode");
+DEFINE_int32(volatile_cache_pct, 10, "Percentage of cache in memory tier.");
+
+namespace ROCKSDB_NAMESPACE {
+
+std::unique_ptr<PersistentCacheTier> NewVolatileCache() {
+  assert(FLAGS_cache_size != std::numeric_limits<uint64_t>::max());
+  std::unique_ptr<PersistentCacheTier> pcache(
+      new VolatileCacheTier(FLAGS_cache_size));
+  return pcache;
+}
+
+std::unique_ptr<PersistentCacheTier> NewBlockCache() {
+  std::shared_ptr<Logger> log;
+  if (!Env::Default()->NewLogger(FLAGS_log_path, &log).ok()) {
+    fprintf(stderr, "Error creating log %s \n", FLAGS_log_path.c_str());
+    return nullptr;
+  }
+
+  PersistentCacheConfig opt(Env::Default(), FLAGS_path, FLAGS_cache_size, log);
+  opt.writer_dispatch_size = FLAGS_writer_iosize;
+  opt.writer_qdepth = FLAGS_writer_qdepth;
+  opt.pipeline_writes = FLAGS_enable_pipelined_writes;
+  opt.max_write_pipeline_backlog_size = std::numeric_limits<uint64_t>::max();
+  std::unique_ptr<PersistentCacheTier> cache(new BlockCacheTier(opt));
+  Status status = cache->Open();
+  return cache;
+}
+
+// create a new cache tier
+// construct a tiered RAM+Block cache
+std::unique_ptr<PersistentTieredCache> NewTieredCache(
+    const size_t mem_size, const PersistentCacheConfig& opt) {
+  std::unique_ptr<PersistentTieredCache> tcache(new PersistentTieredCache());
+  // create primary tier
+  assert(mem_size);
+  auto pcache =
+      std::shared_ptr<PersistentCacheTier>(new VolatileCacheTier(mem_size));
+  tcache->AddTier(pcache);
+  // create secondary tier
+  auto scache = std::shared_ptr<PersistentCacheTier>(new BlockCacheTier(opt));
+  tcache->AddTier(scache);
+
+  Status s = tcache->Open();
+  assert(s.ok());
+  return tcache;
+}
+
+std::unique_ptr<PersistentTieredCache> NewTieredCache() {
+  std::shared_ptr<Logger> log;
+  if (!Env::Default()->NewLogger(FLAGS_log_path, &log).ok()) {
+    fprintf(stderr, "Error creating log %s \n", FLAGS_log_path.c_str());
+    abort();
+  }
+
+  auto pct = FLAGS_volatile_cache_pct / static_cast<double>(100);
+  PersistentCacheConfig opt(Env::Default(), FLAGS_path,
+                            (1 - pct) * FLAGS_cache_size, log);
+  opt.writer_dispatch_size = FLAGS_writer_iosize;
+  opt.writer_qdepth = FLAGS_writer_qdepth;
+  opt.pipeline_writes = FLAGS_enable_pipelined_writes;
+  opt.max_write_pipeline_backlog_size = std::numeric_limits<uint64_t>::max();
+  return NewTieredCache(FLAGS_cache_size * pct, opt);
+}
+
+//
+// Benchmark driver
+//
+class CacheTierBenchmark {
+ public:
+  explicit CacheTierBenchmark(std::shared_ptr<PersistentCacheTier>&& cache)
+      : cache_(cache) {
+    if (FLAGS_nthread_read) {
+      fprintf(stdout, "Pre-populating\n");
+      Prepop();
+      fprintf(stdout, "Pre-population completed\n");
+    }
+
+    stats_.Clear();
+
+    // Start IO threads
+    std::list<port::Thread> threads;
+    Spawn(FLAGS_nthread_write, &threads,
+          std::bind(&CacheTierBenchmark::Write, this));
+    Spawn(FLAGS_nthread_read, &threads,
+          std::bind(&CacheTierBenchmark::Read, this));
+
+    // Wait till FLAGS_nsec and then signal to quit
+    StopWatchNano t(SystemClock::Default().get(), /*auto_start=*/true);
+    size_t sec = t.ElapsedNanos() / 1000000000ULL;
+    while (!quit_) {
+      sec = t.ElapsedNanos() / 1000000000ULL;
+      quit_ = sec > size_t(FLAGS_nsec);
+      /* sleep override */ sleep(1);
+    }
+
+    // Wait for threads to exit
+    Join(&threads);
+    // Print stats
+    PrintStats(sec);
+    // Close the cache
+    cache_->TEST_Flush();
+    cache_->Close();
+  }
+
+ private:
+  void PrintStats(const size_t sec) {
+    std::ostringstream msg;
+    msg << "Test stats" << std::endl
+        << "* Elapsed: " << sec << " s" << std::endl
+        << "* Write Latency:" << std::endl
+        << stats_.write_latency_.ToString() << std::endl
+        << "* Read Latency:" << std::endl
+        << stats_.read_latency_.ToString() << std::endl
+        << "* Bytes written:" << std::endl
+        << stats_.bytes_written_.ToString() << std::endl
+        << "* Bytes read:" << std::endl
+        << stats_.bytes_read_.ToString() << std::endl
+        << "Cache stats:" << std::endl
+        << cache_->PrintStats() << std::endl;
+    fprintf(stderr, "%s\n", msg.str().c_str());
+  }
+
+  //
+  // Insert implementation and corresponding helper functions
+  //
+  void Prepop() {
+    for (uint64_t i = 0; i < 1024 * 1024; ++i) {
+      InsertKey(i);
+      insert_key_limit_++;
+      read_key_limit_++;
+    }
+
+    // Wait until data is flushed
+    cache_->TEST_Flush();
+    // warmup the cache
+    for (uint64_t i = 0; i < 1024 * 1024; ReadKey(i++)) {
+    }
+  }
+
+  void Write() {
+    while (!quit_) {
+      InsertKey(insert_key_limit_++);
+    }
+  }
+
+  void InsertKey(const uint64_t key) {
+    // construct key
+    uint64_t k[3];
+    Slice block_key = FillKey(k, key);
+
+    // construct value
+    auto block = NewBlock(key);
+
+    // insert
+    StopWatchNano timer(SystemClock::Default().get(), /*auto_start=*/true);
+    while (true) {
+      Status status = cache_->Insert(block_key, block.get(), FLAGS_iosize);
+      if (status.ok()) {
+        break;
+      }
+
+      // transient error is possible if we run without pipelining
+      assert(!FLAGS_enable_pipelined_writes);
+    }
+
+    // adjust stats
+    const size_t elapsed_micro = timer.ElapsedNanos() / 1000;
+    stats_.write_latency_.Add(elapsed_micro);
+    stats_.bytes_written_.Add(FLAGS_iosize);
+  }
+
+  //
+  // Read implementation
+  //
+  void Read() {
+    while (!quit_) {
+      ReadKey(random() % read_key_limit_);
+    }
+  }
+
+  void ReadKey(const uint64_t val) {
+    // construct key
+    uint64_t k[3];
+    Slice key = FillKey(k, val);
+
+    // Lookup in cache
+    StopWatchNano timer(SystemClock::Default().get(), /*auto_start=*/true);
+    std::unique_ptr<char[]> block;
+    size_t size;
+    Status status = cache_->Lookup(key, &block, &size);
+    if (!status.ok()) {
+      fprintf(stderr, "%s\n", status.ToString().c_str());
+    }
+    assert(status.ok());
+    assert(size == (size_t)FLAGS_iosize);
+
+    // adjust stats
+    const size_t elapsed_micro = timer.ElapsedNanos() / 1000;
+    stats_.read_latency_.Add(elapsed_micro);
+    stats_.bytes_read_.Add(FLAGS_iosize);
+
+    // verify content
+    if (!FLAGS_benchmark) {
+      auto expected_block = NewBlock(val);
+      assert(memcmp(block.get(), expected_block.get(), FLAGS_iosize) == 0);
+    }
+  }
+
+  // create data for a key by filling with a certain pattern
+  std::unique_ptr<char[]> NewBlock(const uint64_t val) {
+    std::unique_ptr<char[]> data(new char[FLAGS_iosize]);
+    memset(data.get(), val % 255, FLAGS_iosize);
+    return data;
+  }
+
+  // spawn threads
+  void Spawn(const size_t n, std::list<port::Thread>* threads,
+             const std::function<void()>& fn) {
+    for (size_t i = 0; i < n; ++i) {
+      threads->emplace_back(fn);
+    }
+  }
+
+  // join threads
+  void Join(std::list<port::Thread>* threads) {
+    for (auto& th : *threads) {
+      th.join();
+    }
+  }
+
+  // construct key
+  Slice FillKey(uint64_t (&k)[3], const uint64_t val) {
+    k[0] = k[1] = 0;
+    k[2] = val;
+    void* p = static_cast<void*>(&k);
+    return Slice(static_cast<char*>(p), sizeof(k));
+  }
+
+  // benchmark stats
+  struct Stats {
+    void Clear() {
+      bytes_written_.Clear();
+      bytes_read_.Clear();
+      read_latency_.Clear();
+      write_latency_.Clear();
+    }
+
+    HistogramImpl bytes_written_;
+    HistogramImpl bytes_read_;
+    HistogramImpl read_latency_;
+    HistogramImpl write_latency_;
+  };
+
+  std::shared_ptr<PersistentCacheTier> cache_;  // cache implementation
+  std::atomic<uint64_t> insert_key_limit_{0};   // data inserted upto
+  std::atomic<uint64_t> read_key_limit_{0};     // data can be read safely upto
+  bool quit_ = false;                           // Quit thread ?
+  mutable Stats stats_;                         // Stats
+};
+
+}  // namespace ROCKSDB_NAMESPACE
+
+//
+// main
+//
+int main(int argc, char** argv) {
+  GFLAGS_NAMESPACE::SetUsageMessage(std::string("\nUSAGE:\n") +
+                                    std::string(argv[0]) + " [OPTIONS]...");
+  GFLAGS_NAMESPACE::ParseCommandLineFlags(&argc, &argv, false);
+
+  std::ostringstream msg;
+  msg << "Config" << std::endl
+      << "======" << std::endl
+      << "* nsec=" << FLAGS_nsec << std::endl
+      << "* nthread_write=" << FLAGS_nthread_write << std::endl
+      << "* path=" << FLAGS_path << std::endl
+      << "* cache_size=" << FLAGS_cache_size << std::endl
+      << "* iosize=" << FLAGS_iosize << std::endl
+      << "* writer_iosize=" << FLAGS_writer_iosize << std::endl
+      << "* writer_qdepth=" << FLAGS_writer_qdepth << std::endl
+      << "* enable_pipelined_writes=" << FLAGS_enable_pipelined_writes
+      << std::endl
+      << "* cache_type=" << FLAGS_cache_type << std::endl
+      << "* benchmark=" << FLAGS_benchmark << std::endl
+      << "* volatile_cache_pct=" << FLAGS_volatile_cache_pct << std::endl;
+
+  fprintf(stderr, "%s\n", msg.str().c_str());
+
+  std::shared_ptr<ROCKSDB_NAMESPACE::PersistentCacheTier> cache;
+  if (FLAGS_cache_type == "block_cache") {
+    fprintf(stderr, "Using block cache implementation\n");
+    cache = ROCKSDB_NAMESPACE::NewBlockCache();
+  } else if (FLAGS_cache_type == "volatile") {
+    fprintf(stderr, "Using volatile cache implementation\n");
+    cache = ROCKSDB_NAMESPACE::NewVolatileCache();
+  } else if (FLAGS_cache_type == "tiered") {
+    fprintf(stderr, "Using tiered cache implementation\n");
+    cache = ROCKSDB_NAMESPACE::NewTieredCache();
+  } else {
+    fprintf(stderr, "Unknown option for cache\n");
+  }
+
+  assert(cache);
+  if (!cache) {
+    fprintf(stderr, "Error creating cache\n");
+    abort();
+  }
+
+  std::unique_ptr<ROCKSDB_NAMESPACE::CacheTierBenchmark> benchmark(
+      new ROCKSDB_NAMESPACE::CacheTierBenchmark(std::move(cache)));
+
+  return 0;
+}
+#endif  // #ifndef GFLAGS
+#else
+int main(int, char**) { return 0; }
+#endif
diff --git a/src/rocksdb/utilities/persistent_cache/persistent_cache_test.cc b/src/rocksdb/utilities/persistent_cache/persistent_cache_test.cc
new file mode 100644
index 000000000..d1b18b68a
--- /dev/null
+++ b/src/rocksdb/utilities/persistent_cache/persistent_cache_test.cc
@@ -0,0 +1,462 @@
+//  Copyright (c) 2013, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+//
+// Copyright (c) 2011 The LevelDB Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file. See the AUTHORS file for names of contributors.
+#if !defined ROCKSDB_LITE
+
+#include "utilities/persistent_cache/persistent_cache_test.h"
+
+#include <functional>
+#include <memory>
+#include <thread>
+
+#include "file/file_util.h"
+#include "utilities/persistent_cache/block_cache_tier.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+static const double kStressFactor = .125;
+
+#ifdef OS_LINUX
+static void OnOpenForRead(void* arg) {
+  int* val = static_cast<int*>(arg);
+  *val &= ~O_DIRECT;
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
+      "NewRandomAccessFile:O_DIRECT",
+      std::bind(OnOpenForRead, std::placeholders::_1));
+}
+
+static void OnOpenForWrite(void* arg) {
+  int* val = static_cast<int*>(arg);
+  *val &= ~O_DIRECT;
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
+      "NewWritableFile:O_DIRECT",
+      std::bind(OnOpenForWrite, std::placeholders::_1));
+}
+#endif
+
+static void OnDeleteDir(void* arg) {
+  char* dir = static_cast<char*>(arg);
+  ASSERT_OK(DestroyDir(Env::Default(), std::string(dir)));
+}
+
+//
+// Simple logger that prints message on stdout
+//
+class ConsoleLogger : public Logger {
+ public:
+  using Logger::Logv;
+  ConsoleLogger() : Logger(InfoLogLevel::ERROR_LEVEL) {}
+
+  void Logv(const char* format, va_list ap) override {
+    MutexLock _(&lock_);
+    vprintf(format, ap);
+    printf("\n");
+  }
+
+  port::Mutex lock_;
+};
+
+// construct a tiered RAM+Block cache
+std::unique_ptr<PersistentTieredCache> NewTieredCache(
+    const size_t mem_size, const PersistentCacheConfig& opt) {
+  std::unique_ptr<PersistentTieredCache> tcache(new PersistentTieredCache());
+  // create primary tier
+  assert(mem_size);
+  auto pcache = std::shared_ptr<PersistentCacheTier>(new VolatileCacheTier(
+      /*is_compressed*/ true, mem_size));
+  tcache->AddTier(pcache);
+  // create secondary tier
+  auto scache = std::shared_ptr<PersistentCacheTier>(new BlockCacheTier(opt));
+  tcache->AddTier(scache);
+
+  Status s = tcache->Open();
+  assert(s.ok());
+  return tcache;
+}
+
+// create block cache
+std::unique_ptr<PersistentCacheTier> NewBlockCache(
+    Env* env, const std::string& path,
+    const uint64_t max_size = std::numeric_limits<uint64_t>::max(),
+    const bool enable_direct_writes = false) {
+  const uint32_t max_file_size =
+      static_cast<uint32_t>(12 * 1024 * 1024 * kStressFactor);
+  auto log = std::make_shared<ConsoleLogger>();
+  PersistentCacheConfig opt(env, path, max_size, log);
+  opt.cache_file_size = max_file_size;
+  opt.max_write_pipeline_backlog_size = std::numeric_limits<uint64_t>::max();
+  opt.enable_direct_writes = enable_direct_writes;
+  std::unique_ptr<PersistentCacheTier> scache(new BlockCacheTier(opt));
+  Status s = scache->Open();
+  assert(s.ok());
+  return scache;
+}
+
+// create a new cache tier
+std::unique_ptr<PersistentTieredCache> NewTieredCache(
+    Env* env, const std::string& path, const uint64_t max_volatile_cache_size,
+    const uint64_t max_block_cache_size =
+        std::numeric_limits<uint64_t>::max()) {
+  const uint32_t max_file_size =
+      static_cast<uint32_t>(12 * 1024 * 1024 * kStressFactor);
+  auto log = std::make_shared<ConsoleLogger>();
+  auto opt = PersistentCacheConfig(env, path, max_block_cache_size, log);
+  opt.cache_file_size = max_file_size;
+  opt.max_write_pipeline_backlog_size = std::numeric_limits<uint64_t>::max();
+  // create tier out of the two caches
+  auto cache = NewTieredCache(max_volatile_cache_size, opt);
+  return cache;
+}
+
+PersistentCacheTierTest::PersistentCacheTierTest()
+    : path_(test::PerThreadDBPath("cache_test")) {
+#ifdef OS_LINUX
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->EnableProcessing();
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
+      "NewRandomAccessFile:O_DIRECT", OnOpenForRead);
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
+      "NewWritableFile:O_DIRECT", OnOpenForWrite);
+#endif
+}
+
+// Block cache tests
+TEST_F(PersistentCacheTierTest, DISABLED_BlockCacheInsertWithFileCreateError) {
+  cache_ = NewBlockCache(Env::Default(), path_,
+                         /*size=*/std::numeric_limits<uint64_t>::max(),
+                         /*direct_writes=*/false);
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
+      "BlockCacheTier::NewCacheFile:DeleteDir", OnDeleteDir);
+
+  RunNegativeInsertTest(/*nthreads=*/1,
+                        /*max_keys*/
+                        static_cast<size_t>(10 * 1024 * kStressFactor));
+
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->ClearAllCallBacks();
+}
+
+// Travis is unable to handle the normal version of the tests running out of
+// fds, out of space and timeouts. This is an easier version of the test
+// specifically written for Travis
+TEST_F(PersistentCacheTierTest, DISABLED_BasicTest) {
+  cache_ = std::make_shared<VolatileCacheTier>();
+  RunInsertTest(/*nthreads=*/1, /*max_keys=*/1024);
+
+  cache_ = NewBlockCache(Env::Default(), path_,
+                         /*size=*/std::numeric_limits<uint64_t>::max(),
+                         /*direct_writes=*/true);
+  RunInsertTest(/*nthreads=*/1, /*max_keys=*/1024);
+
+  cache_ = NewTieredCache(Env::Default(), path_,
+                          /*memory_size=*/static_cast<size_t>(1 * 1024 * 1024));
+  RunInsertTest(/*nthreads=*/1, /*max_keys=*/1024);
+}
+
+// Volatile cache tests
+// DISABLED for now (somewhat expensive)
+TEST_F(PersistentCacheTierTest, DISABLED_VolatileCacheInsert) {
+  for (auto nthreads : {1, 5}) {
+    for (auto max_keys :
+         {10 * 1024 * kStressFactor, 1 * 1024 * 1024 * kStressFactor}) {
+      cache_ = std::make_shared<VolatileCacheTier>();
+      RunInsertTest(nthreads, static_cast<size_t>(max_keys));
+    }
+  }
+}
+
+// DISABLED for now (somewhat expensive)
+TEST_F(PersistentCacheTierTest, DISABLED_VolatileCacheInsertWithEviction) {
+  for (auto nthreads : {1, 5}) {
+    for (auto max_keys : {1 * 1024 * 1024 * kStressFactor}) {
+      cache_ = std::make_shared<VolatileCacheTier>(
+          /*compressed=*/true,
+          /*size=*/static_cast<size_t>(1 * 1024 * 1024 * kStressFactor));
+      RunInsertTestWithEviction(nthreads, static_cast<size_t>(max_keys));
+    }
+  }
+}
+
+// Block cache tests
+// DISABLED for now (expensive)
+TEST_F(PersistentCacheTierTest, DISABLED_BlockCacheInsert) {
+  for (auto direct_writes : {true, false}) {
+    for (auto nthreads : {1, 5}) {
+      for (auto max_keys :
+           {10 * 1024 * kStressFactor, 1 * 1024 * 1024 * kStressFactor}) {
+        cache_ = NewBlockCache(Env::Default(), path_,
+                               /*size=*/std::numeric_limits<uint64_t>::max(),
+                               direct_writes);
+        RunInsertTest(nthreads, static_cast<size_t>(max_keys));
+      }
+    }
+  }
+}
+
+// DISABLED for now (somewhat expensive)
+TEST_F(PersistentCacheTierTest, DISABLED_BlockCacheInsertWithEviction) {
+  for (auto nthreads : {1, 5}) {
+    for (auto max_keys : {1 * 1024 * 1024 * kStressFactor}) {
+      cache_ = NewBlockCache(
+          Env::Default(), path_,
+          /*max_size=*/static_cast<size_t>(200 * 1024 * 1024 * kStressFactor));
+      RunInsertTestWithEviction(nthreads, static_cast<size_t>(max_keys));
+    }
+  }
+}
+
+// Tiered cache tests
+// DISABLED for now (expensive)
+TEST_F(PersistentCacheTierTest, DISABLED_TieredCacheInsert) {
+  for (auto nthreads : {1, 5}) {
+    for (auto max_keys :
+         {10 * 1024 * kStressFactor, 1 * 1024 * 1024 * kStressFactor}) {
+      cache_ = NewTieredCache(
+          Env::Default(), path_,
+          /*memory_size=*/static_cast<size_t>(1 * 1024 * 1024 * kStressFactor));
+      RunInsertTest(nthreads, static_cast<size_t>(max_keys));
+    }
+  }
+}
+
+// the tests causes a lot of file deletions which Travis limited testing
+// environment cannot handle
+// DISABLED for now (somewhat expensive)
+TEST_F(PersistentCacheTierTest, DISABLED_TieredCacheInsertWithEviction) {
+  for (auto nthreads : {1, 5}) {
+    for (auto max_keys : {1 * 1024 * 1024 * kStressFactor}) {
+      cache_ = NewTieredCache(
+          Env::Default(), path_,
+          /*memory_size=*/static_cast<size_t>(1 * 1024 * 1024 * kStressFactor),
+          /*block_cache_size*/
+          static_cast<size_t>(200 * 1024 * 1024 * kStressFactor));
+      RunInsertTestWithEviction(nthreads, static_cast<size_t>(max_keys));
+    }
+  }
+}
+
+std::shared_ptr<PersistentCacheTier> MakeVolatileCache(
+    Env* /*env*/, const std::string& /*dbname*/) {
+  return std::make_shared<VolatileCacheTier>();
+}
+
+std::shared_ptr<PersistentCacheTier> MakeBlockCache(Env* env,
+                                                    const std::string& dbname) {
+  return NewBlockCache(env, dbname);
+}
+
+std::shared_ptr<PersistentCacheTier> MakeTieredCache(
+    Env* env, const std::string& dbname) {
+  const auto memory_size = 1 * 1024 * 1024 * kStressFactor;
+  return NewTieredCache(env, dbname, static_cast<size_t>(memory_size));
+}
+
+#ifdef OS_LINUX
+static void UniqueIdCallback(void* arg) {
+  int* result = reinterpret_cast<int*>(arg);
+  if (*result == -1) {
+    *result = 0;
+  }
+
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->ClearTrace();
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
+      "GetUniqueIdFromFile:FS_IOC_GETVERSION", UniqueIdCallback);
+}
+#endif
+
+TEST_F(PersistentCacheTierTest, FactoryTest) {
+  for (auto nvm_opt : {true, false}) {
+    ASSERT_FALSE(cache_);
+    auto log = std::make_shared<ConsoleLogger>();
+    std::shared_ptr<PersistentCache> cache;
+    ASSERT_OK(NewPersistentCache(Env::Default(), path_,
+                                 /*size=*/1 * 1024 * 1024 * 1024, log, nvm_opt,
+                                 &cache));
+    ASSERT_TRUE(cache);
+    ASSERT_EQ(cache->Stats().size(), 1);
+    ASSERT_TRUE(cache->Stats()[0].size());
+    cache.reset();
+  }
+}
+
+PersistentCacheDBTest::PersistentCacheDBTest()
+    : DBTestBase("cache_test", /*env_do_fsync=*/true) {
+#ifdef OS_LINUX
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->EnableProcessing();
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
+      "GetUniqueIdFromFile:FS_IOC_GETVERSION", UniqueIdCallback);
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
+      "NewRandomAccessFile:O_DIRECT", OnOpenForRead);
+#endif
+}
+
+// test template
+void PersistentCacheDBTest::RunTest(
+    const std::function<std::shared_ptr<PersistentCacheTier>(bool)>& new_pcache,
+    const size_t max_keys = 100 * 1024, const size_t max_usecase = 5) {
+  // number of insertion interations
+  int num_iter = static_cast<int>(max_keys * kStressFactor);
+
+  for (size_t iter = 0; iter < max_usecase; iter++) {
+    Options options;
+    options.write_buffer_size =
+        static_cast<size_t>(64 * 1024 * kStressFactor);  // small write buffer
+    options.statistics = ROCKSDB_NAMESPACE::CreateDBStatistics();
+    options = CurrentOptions(options);
+
+    // setup page cache
+    std::shared_ptr<PersistentCacheTier> pcache;
+    BlockBasedTableOptions table_options;
+    table_options.cache_index_and_filter_blocks = true;
+
+    const size_t size_max = std::numeric_limits<size_t>::max();
+
+    switch (iter) {
+      case 0:
+        // page cache, block cache, no-compressed cache
+        pcache = new_pcache(/*is_compressed=*/true);
+        table_options.persistent_cache = pcache;
+        table_options.block_cache = NewLRUCache(size_max);
+        table_options.block_cache_compressed = nullptr;
+        options.table_factory.reset(NewBlockBasedTableFactory(table_options));
+        break;
+      case 1:
+        // page cache, block cache, compressed cache
+        pcache = new_pcache(/*is_compressed=*/true);
+        table_options.persistent_cache = pcache;
+        table_options.block_cache = NewLRUCache(size_max);
+        table_options.block_cache_compressed = NewLRUCache(size_max);
+        options.table_factory.reset(NewBlockBasedTableFactory(table_options));
+        break;
+      case 2:
+        // page cache, block cache, compressed cache + KNoCompression
+        // both block cache and compressed cache, but DB is not compressed
+        // also, make block cache sizes bigger, to trigger block cache hits
+        pcache = new_pcache(/*is_compressed=*/true);
+        table_options.persistent_cache = pcache;
+        table_options.block_cache = NewLRUCache(size_max);
+        table_options.block_cache_compressed = NewLRUCache(size_max);
+        options.table_factory.reset(NewBlockBasedTableFactory(table_options));
+        options.compression = kNoCompression;
+        break;
+      case 3:
+        // page cache, no block cache, no compressed cache
+        pcache = new_pcache(/*is_compressed=*/false);
+        table_options.persistent_cache = pcache;
+        table_options.block_cache = nullptr;
+        table_options.block_cache_compressed = nullptr;
+        options.table_factory.reset(NewBlockBasedTableFactory(table_options));
+        break;
+      case 4:
+        // page cache, no block cache, no compressed cache
+        // Page cache caches compressed blocks
+        pcache = new_pcache(/*is_compressed=*/true);
+        table_options.persistent_cache = pcache;
+        table_options.block_cache = nullptr;
+        table_options.block_cache_compressed = nullptr;
+        options.table_factory.reset(NewBlockBasedTableFactory(table_options));
+        break;
+      default:
+        FAIL();
+    }
+
+    std::vector<std::string> values;
+    // insert data
+    Insert(options, table_options, num_iter, &values);
+    // flush all data in cache to device
+    pcache->TEST_Flush();
+    // verify data
+    Verify(num_iter, values);
+
+    auto block_miss = TestGetTickerCount(options, BLOCK_CACHE_MISS);
+    auto compressed_block_hit =
+        TestGetTickerCount(options, BLOCK_CACHE_COMPRESSED_HIT);
+    auto compressed_block_miss =
+        TestGetTickerCount(options, BLOCK_CACHE_COMPRESSED_MISS);
+    auto page_hit = TestGetTickerCount(options, PERSISTENT_CACHE_HIT);
+    auto page_miss = TestGetTickerCount(options, PERSISTENT_CACHE_MISS);
+
+    // check that we triggered the appropriate code paths in the cache
+    switch (iter) {
+      case 0:
+        // page cache, block cache, no-compressed cache
+        ASSERT_GT(page_miss, 0);
+        ASSERT_GT(page_hit, 0);
+        ASSERT_GT(block_miss, 0);
+        ASSERT_EQ(compressed_block_miss, 0);
+        ASSERT_EQ(compressed_block_hit, 0);
+        break;
+      case 1:
+        // page cache, block cache, compressed cache
+        ASSERT_GT(page_miss, 0);
+        ASSERT_GT(block_miss, 0);
+        ASSERT_GT(compressed_block_miss, 0);
+        break;
+      case 2:
+        // page cache, block cache, compressed cache + KNoCompression
+        ASSERT_GT(page_miss, 0);
+        ASSERT_GT(page_hit, 0);
+        ASSERT_GT(block_miss, 0);
+        ASSERT_GT(compressed_block_miss, 0);
+        // remember kNoCompression
+        ASSERT_EQ(compressed_block_hit, 0);
+        break;
+      case 3:
+      case 4:
+        // page cache, no block cache, no compressed cache
+        ASSERT_GT(page_miss, 0);
+        ASSERT_GT(page_hit, 0);
+        ASSERT_EQ(compressed_block_hit, 0);
+        ASSERT_EQ(compressed_block_miss, 0);
+        break;
+      default:
+        FAIL();
+    }
+
+    options.create_if_missing = true;
+    DestroyAndReopen(options);
+
+    ASSERT_OK(pcache->Close());
+  }
+}
+
+// Travis is unable to handle the normal version of the tests running out of
+// fds, out of space and timeouts. This is an easier version of the test
+// specifically written for Travis.
+// Now used generally because main tests are too expensive as unit tests.
+TEST_F(PersistentCacheDBTest, BasicTest) {
+  RunTest(std::bind(&MakeBlockCache, env_, dbname_), /*max_keys=*/1024,
+          /*max_usecase=*/1);
+}
+
+// test table with block page cache
+// DISABLED for now (very expensive, especially memory)
+TEST_F(PersistentCacheDBTest, DISABLED_BlockCacheTest) {
+  RunTest(std::bind(&MakeBlockCache, env_, dbname_));
+}
+
+// test table with volatile page cache
+// DISABLED for now (very expensive, especially memory)
+TEST_F(PersistentCacheDBTest, DISABLED_VolatileCacheTest) {
+  RunTest(std::bind(&MakeVolatileCache, env_, dbname_));
+}
+
+// test table with tiered page cache
+// DISABLED for now (very expensive, especially memory)
+TEST_F(PersistentCacheDBTest, DISABLED_TieredCacheTest) {
+  RunTest(std::bind(&MakeTieredCache, env_, dbname_));
+}
+
+}  // namespace ROCKSDB_NAMESPACE
+
+int main(int argc, char** argv) {
+  ROCKSDB_NAMESPACE::port::InstallStackTraceHandler();
+  ::testing::InitGoogleTest(&argc, argv);
+  return RUN_ALL_TESTS();
+}
+#else   // !defined ROCKSDB_LITE
+int main() { return 0; }
+#endif  // !defined ROCKSDB_LITE
diff --git a/src/rocksdb/utilities/persistent_cache/persistent_cache_test.h b/src/rocksdb/utilities/persistent_cache/persistent_cache_test.h
new file mode 100644
index 000000000..f13155ed6
--- /dev/null
+++ b/src/rocksdb/utilities/persistent_cache/persistent_cache_test.h
@@ -0,0 +1,286 @@
+//  Copyright (c) 2013, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+//
+// Copyright (c) 2011 The LevelDB Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file. See the AUTHORS file for names of contributors.
+#pragma once
+
+#ifndef ROCKSDB_LITE
+
+#include <functional>
+#include <limits>
+#include <list>
+#include <memory>
+#include <string>
+#include <thread>
+#include <vector>
+
+#include "db/db_test_util.h"
+#include "memory/arena.h"
+#include "port/port.h"
+#include "rocksdb/cache.h"
+#include "table/block_based/block_builder.h"
+#include "test_util/testharness.h"
+#include "util/random.h"
+#include "utilities/persistent_cache/volatile_tier_impl.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+//
+// Unit tests for testing PersistentCacheTier
+//
+class PersistentCacheTierTest : public testing::Test {
+ public:
+  PersistentCacheTierTest();
+  virtual ~PersistentCacheTierTest() {
+    if (cache_) {
+      Status s = cache_->Close();
+      assert(s.ok());
+    }
+  }
+
+ protected:
+  // Flush cache
+  void Flush() {
+    if (cache_) {
+      cache_->TEST_Flush();
+    }
+  }
+
+  // create threaded workload
+  template <class T>
+  std::list<port::Thread> SpawnThreads(const size_t n, const T& fn) {
+    std::list<port::Thread> threads;
+    for (size_t i = 0; i < n; i++) {
+      port::Thread th(fn);
+      threads.push_back(std::move(th));
+    }
+    return threads;
+  }
+
+  // Wait for threads to join
+  void Join(std::list<port::Thread>&& threads) {
+    for (auto& th : threads) {
+      th.join();
+    }
+    threads.clear();
+  }
+
+  // Run insert workload in threads
+  void Insert(const size_t nthreads, const size_t max_keys) {
+    key_ = 0;
+    max_keys_ = max_keys;
+    // spawn threads
+    auto fn = std::bind(&PersistentCacheTierTest::InsertImpl, this);
+    auto threads = SpawnThreads(nthreads, fn);
+    // join with threads
+    Join(std::move(threads));
+    // Flush cache
+    Flush();
+  }
+
+  // Run verification on the cache
+  void Verify(const size_t nthreads = 1, const bool eviction_enabled = false) {
+    stats_verify_hits_ = 0;
+    stats_verify_missed_ = 0;
+    key_ = 0;
+    // spawn threads
+    auto fn =
+        std::bind(&PersistentCacheTierTest::VerifyImpl, this, eviction_enabled);
+    auto threads = SpawnThreads(nthreads, fn);
+    // join with threads
+    Join(std::move(threads));
+  }
+
+  // pad 0 to numbers
+  std::string PaddedNumber(const size_t data, const size_t pad_size) {
+    assert(pad_size);
+    char* ret = new char[pad_size];
+    int pos = static_cast<int>(pad_size) - 1;
+    size_t count = 0;
+    size_t t = data;
+    // copy numbers
+    while (t) {
+      count++;
+      ret[pos--] = '0' + t % 10;
+      t = t / 10;
+    }
+    // copy 0s
+    while (pos >= 0) {
+      ret[pos--] = '0';
+    }
+    // post condition
+    assert(count <= pad_size);
+    assert(pos == -1);
+    std::string result(ret, pad_size);
+    delete[] ret;
+    return result;
+  }
+
+  // Insert workload implementation
+  void InsertImpl() {
+    const std::string prefix = "key_prefix_";
+
+    while (true) {
+      size_t i = key_++;
+      if (i >= max_keys_) {
+        break;
+      }
+
+      char data[4 * 1024];
+      memset(data, '0' + (i % 10), sizeof(data));
+      auto k = prefix + PaddedNumber(i, /*count=*/8);
+      Slice key(k);
+      while (true) {
+        Status status = cache_->Insert(key, data, sizeof(data));
+        if (status.ok()) {
+          break;
+        }
+        ASSERT_TRUE(status.IsTryAgain());
+        Env::Default()->SleepForMicroseconds(1 * 1000 * 1000);
+      }
+    }
+  }
+
+  // Verification implementation
+  void VerifyImpl(const bool eviction_enabled = false) {
+    const std::string prefix = "key_prefix_";
+    while (true) {
+      size_t i = key_++;
+      if (i >= max_keys_) {
+        break;
+      }
+
+      char edata[4 * 1024];
+      memset(edata, '0' + (i % 10), sizeof(edata));
+      auto k = prefix + PaddedNumber(i, /*count=*/8);
+      Slice key(k);
+      std::unique_ptr<char[]> block;
+      size_t block_size;
+
+      if (eviction_enabled) {
+        if (!cache_->Lookup(key, &block, &block_size).ok()) {
+          // assume that the key is evicted
+          stats_verify_missed_++;
+          continue;
+        }
+      }
+
+      ASSERT_OK(cache_->Lookup(key, &block, &block_size));
+      ASSERT_EQ(block_size, sizeof(edata));
+      ASSERT_EQ(memcmp(edata, block.get(), sizeof(edata)), 0);
+      stats_verify_hits_++;
+    }
+  }
+
+  // template for insert test
+  void RunInsertTest(const size_t nthreads, const size_t max_keys) {
+    Insert(nthreads, max_keys);
+    Verify(nthreads);
+    ASSERT_EQ(stats_verify_hits_, max_keys);
+    ASSERT_EQ(stats_verify_missed_, 0);
+
+    ASSERT_OK(cache_->Close());
+    cache_.reset();
+  }
+
+  // template for negative insert test
+  void RunNegativeInsertTest(const size_t nthreads, const size_t max_keys) {
+    Insert(nthreads, max_keys);
+    Verify(nthreads, /*eviction_enabled=*/true);
+    ASSERT_LT(stats_verify_hits_, max_keys);
+    ASSERT_GT(stats_verify_missed_, 0);
+
+    ASSERT_OK(cache_->Close());
+    cache_.reset();
+  }
+
+  // template for insert with eviction test
+  void RunInsertTestWithEviction(const size_t nthreads, const size_t max_keys) {
+    Insert(nthreads, max_keys);
+    Verify(nthreads, /*eviction_enabled=*/true);
+    ASSERT_EQ(stats_verify_hits_ + stats_verify_missed_, max_keys);
+    ASSERT_GT(stats_verify_hits_, 0);
+    ASSERT_GT(stats_verify_missed_, 0);
+
+    ASSERT_OK(cache_->Close());
+    cache_.reset();
+  }
+
+  const std::string path_;
+  std::shared_ptr<Logger> log_;
+  std::shared_ptr<PersistentCacheTier> cache_;
+  std::atomic<size_t> key_{0};
+  size_t max_keys_ = 0;
+  std::atomic<size_t> stats_verify_hits_{0};
+  std::atomic<size_t> stats_verify_missed_{0};
+};
+
+//
+// RocksDB tests
+//
+class PersistentCacheDBTest : public DBTestBase {
+ public:
+  PersistentCacheDBTest();
+
+  static uint64_t TestGetTickerCount(const Options& options,
+                                     Tickers ticker_type) {
+    return static_cast<uint32_t>(
+        options.statistics->getTickerCount(ticker_type));
+  }
+
+  // insert data to table
+  void Insert(const Options& options,
+              const BlockBasedTableOptions& /*table_options*/,
+              const int num_iter, std::vector<std::string>* values) {
+    CreateAndReopenWithCF({"pikachu"}, options);
+    // default column family doesn't have block cache
+    Options no_block_cache_opts;
+    no_block_cache_opts.statistics = options.statistics;
+    no_block_cache_opts = CurrentOptions(no_block_cache_opts);
+    BlockBasedTableOptions table_options_no_bc;
+    table_options_no_bc.no_block_cache = true;
+    no_block_cache_opts.table_factory.reset(
+        NewBlockBasedTableFactory(table_options_no_bc));
+    ReopenWithColumnFamilies(
+        {"default", "pikachu"},
+        std::vector<Options>({no_block_cache_opts, options}));
+
+    Random rnd(301);
+
+    // Write 8MB (80 values, each 100K)
+    ASSERT_EQ(NumTableFilesAtLevel(0, 1), 0);
+    std::string str;
+    for (int i = 0; i < num_iter; i++) {
+      if (i % 4 == 0) {  // high compression ratio
+        str = rnd.RandomString(1000);
+      }
+      values->push_back(str);
+      ASSERT_OK(Put(1, Key(i), (*values)[i]));
+    }
+
+    // flush all data from memtable so that reads are from block cache
+    ASSERT_OK(Flush(1));
+  }
+
+  // verify data
+  void Verify(const int num_iter, const std::vector<std::string>& values) {
+    for (int j = 0; j < 2; ++j) {
+      for (int i = 0; i < num_iter; i++) {
+        ASSERT_EQ(Get(1, Key(i)), values[i]);
+      }
+    }
+  }
+
+  // test template
+  void RunTest(const std::function<std::shared_ptr<PersistentCacheTier>(bool)>&
+                   new_pcache,
+               const size_t max_keys, const size_t max_usecase);
+};
+
+}  // namespace ROCKSDB_NAMESPACE
+
+#endif
diff --git a/src/rocksdb/utilities/persistent_cache/persistent_cache_tier.cc b/src/rocksdb/utilities/persistent_cache/persistent_cache_tier.cc
new file mode 100644
index 000000000..54cbce8f7
--- /dev/null
+++ b/src/rocksdb/utilities/persistent_cache/persistent_cache_tier.cc
@@ -0,0 +1,167 @@
+//  Copyright (c) 2013, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+//
+#ifndef ROCKSDB_LITE
+
+#include "utilities/persistent_cache/persistent_cache_tier.h"
+
+#include <cinttypes>
+#include <sstream>
+#include <string>
+
+namespace ROCKSDB_NAMESPACE {
+
+std::string PersistentCacheConfig::ToString() const {
+  std::string ret;
+  ret.reserve(20000);
+  const int kBufferSize = 200;
+  char buffer[kBufferSize];
+
+  snprintf(buffer, kBufferSize, "    path: %s\n", path.c_str());
+  ret.append(buffer);
+  snprintf(buffer, kBufferSize, "    enable_direct_reads: %d\n",
+           enable_direct_reads);
+  ret.append(buffer);
+  snprintf(buffer, kBufferSize, "    enable_direct_writes: %d\n",
+           enable_direct_writes);
+  ret.append(buffer);
+  snprintf(buffer, kBufferSize, "    cache_size: %" PRIu64 "\n", cache_size);
+  ret.append(buffer);
+  snprintf(buffer, kBufferSize, "    cache_file_size: %" PRIu32 "\n",
+           cache_file_size);
+  ret.append(buffer);
+  snprintf(buffer, kBufferSize, "    writer_qdepth: %" PRIu32 "\n",
+           writer_qdepth);
+  ret.append(buffer);
+  snprintf(buffer, kBufferSize, "    pipeline_writes: %d\n", pipeline_writes);
+  ret.append(buffer);
+  snprintf(buffer, kBufferSize,
+           "    max_write_pipeline_backlog_size: %" PRIu64 "\n",
+           max_write_pipeline_backlog_size);
+  ret.append(buffer);
+  snprintf(buffer, kBufferSize, "    write_buffer_size: %" PRIu32 "\n",
+           write_buffer_size);
+  ret.append(buffer);
+  snprintf(buffer, kBufferSize, "    writer_dispatch_size: %" PRIu64 "\n",
+           writer_dispatch_size);
+  ret.append(buffer);
+  snprintf(buffer, kBufferSize, "    is_compressed: %d\n", is_compressed);
+  ret.append(buffer);
+
+  return ret;
+}
+
+//
+// PersistentCacheTier implementation
+//
+Status PersistentCacheTier::Open() {
+  if (next_tier_) {
+    return next_tier_->Open();
+  }
+  return Status::OK();
+}
+
+Status PersistentCacheTier::Close() {
+  if (next_tier_) {
+    return next_tier_->Close();
+  }
+  return Status::OK();
+}
+
+bool PersistentCacheTier::Reserve(const size_t /*size*/) {
+  // default implementation is a pass through
+  return true;
+}
+
+bool PersistentCacheTier::Erase(const Slice& /*key*/) {
+  // default implementation is a pass through since not all cache tiers might
+  // support erase
+  return true;
+}
+
+std::string PersistentCacheTier::PrintStats() {
+  std::ostringstream os;
+  for (auto tier_stats : Stats()) {
+    os << "---- next tier -----" << std::endl;
+    for (auto stat : tier_stats) {
+      os << stat.first << ": " << stat.second << std::endl;
+    }
+  }
+  return os.str();
+}
+
+PersistentCache::StatsType PersistentCacheTier::Stats() {
+  if (next_tier_) {
+    return next_tier_->Stats();
+  }
+  return PersistentCache::StatsType{};
+}
+
+uint64_t PersistentCacheTier::NewId() {
+  return last_id_.fetch_add(1, std::memory_order_relaxed);
+}
+
+//
+// PersistentTieredCache implementation
+//
+PersistentTieredCache::~PersistentTieredCache() { assert(tiers_.empty()); }
+
+Status PersistentTieredCache::Open() {
+  assert(!tiers_.empty());
+  return tiers_.front()->Open();
+}
+
+Status PersistentTieredCache::Close() {
+  assert(!tiers_.empty());
+  Status status = tiers_.front()->Close();
+  if (status.ok()) {
+    tiers_.clear();
+  }
+  return status;
+}
+
+bool PersistentTieredCache::Erase(const Slice& key) {
+  assert(!tiers_.empty());
+  return tiers_.front()->Erase(key);
+}
+
+PersistentCache::StatsType PersistentTieredCache::Stats() {
+  assert(!tiers_.empty());
+  return tiers_.front()->Stats();
+}
+
+std::string PersistentTieredCache::PrintStats() {
+  assert(!tiers_.empty());
+  return tiers_.front()->PrintStats();
+}
+
+Status PersistentTieredCache::Insert(const Slice& page_key, const char* data,
+                                     const size_t size) {
+  assert(!tiers_.empty());
+  return tiers_.front()->Insert(page_key, data, size);
+}
+
+Status PersistentTieredCache::Lookup(const Slice& page_key,
+                                     std::unique_ptr<char[]>* data,
+                                     size_t* size) {
+  assert(!tiers_.empty());
+  return tiers_.front()->Lookup(page_key, data, size);
+}
+
+void PersistentTieredCache::AddTier(const Tier& tier) {
+  if (!tiers_.empty()) {
+    tiers_.back()->set_next_tier(tier);
+  }
+  tiers_.push_back(tier);
+}
+
+bool PersistentTieredCache::IsCompressed() {
+  assert(tiers_.size());
+  return tiers_.front()->IsCompressed();
+}
+
+}  // namespace ROCKSDB_NAMESPACE
+
+#endif
diff --git a/src/rocksdb/utilities/persistent_cache/persistent_cache_tier.h b/src/rocksdb/utilities/persistent_cache/persistent_cache_tier.h
new file mode 100644
index 000000000..65aadcd3f
--- /dev/null
+++ b/src/rocksdb/utilities/persistent_cache/persistent_cache_tier.h
@@ -0,0 +1,342 @@
+//  Copyright (c) 2013, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+//
+#pragma once
+
+#ifndef ROCKSDB_LITE
+
+#include <limits>
+#include <list>
+#include <map>
+#include <string>
+#include <vector>
+
+#include "monitoring/histogram.h"
+#include "rocksdb/env.h"
+#include "rocksdb/persistent_cache.h"
+#include "rocksdb/status.h"
+#include "rocksdb/system_clock.h"
+
+// Persistent Cache
+//
+// Persistent cache is tiered key-value cache that can use persistent medium. It
+// is a generic design and can leverage any storage medium -- disk/SSD/NVM/RAM.
+// The code has been kept generic but significant benchmark/design/development
+// time has been spent to make sure the cache performs appropriately for
+// respective storage medium.
+// The file defines
+// PersistentCacheTier    : Implementation that handles individual cache tier
+// PersistentTieresCache  : Implementation that handles all tiers as a logical
+//                          unit
+//
+// PersistentTieredCache architecture:
+// +--------------------------+ PersistentCacheTier that handles multiple tiers
+// | +----------------+       |
+// | | RAM            | PersistentCacheTier that handles RAM (VolatileCacheImpl)
+// | +----------------+       |
+// |   | next                 |
+// |   v                      |
+// | +----------------+       |
+// | | NVM            | PersistentCacheTier implementation that handles NVM
+// | +----------------+ (BlockCacheImpl)
+// |   | next                 |
+// |   V                      |
+// | +----------------+       |
+// | | LE-SSD         | PersistentCacheTier implementation that handles LE-SSD
+// | +----------------+ (BlockCacheImpl)
+// |   |                      |
+// |   V                      |
+// |  null                    |
+// +--------------------------+
+//               |
+//               V
+//              null
+namespace ROCKSDB_NAMESPACE {
+
+// Persistent Cache Config
+//
+// This struct captures all the options that are used to configure persistent
+// cache. Some of the terminologies used in naming the options are
+//
+// dispatch size :
+// This is the size in which IO is dispatched to the device
+//
+// write buffer size :
+// This is the size of an individual write buffer size. Write buffers are
+// grouped to form buffered file.
+//
+// cache size :
+// This is the logical maximum for the cache size
+//
+// qdepth :
+// This is the max number of IOs that can issues to the device in parallel
+//
+// pepeling :
+// The writer code path follows pipelined architecture, which means the
+// operations are handed off from one stage to another
+//
+// pipelining backlog size :
+// With the pipelined architecture, there can always be backlogging of ops in
+// pipeline queues. This is the maximum backlog size after which ops are dropped
+// from queue
+struct PersistentCacheConfig {
+  explicit PersistentCacheConfig(
+      Env* const _env, const std::string& _path, const uint64_t _cache_size,
+      const std::shared_ptr<Logger>& _log,
+      const uint32_t _write_buffer_size = 1 * 1024 * 1024 /*1MB*/) {
+    env = _env;
+    clock = (env != nullptr) ? env->GetSystemClock().get()
+                             : SystemClock::Default().get();
+    path = _path;
+    log = _log;
+    cache_size = _cache_size;
+    writer_dispatch_size = write_buffer_size = _write_buffer_size;
+  }
+
+  //
+  // Validate the settings. Our intentions are to catch erroneous settings ahead
+  // of time instead going violating invariants or causing dead locks.
+  //
+  Status ValidateSettings() const {
+    // (1) check pre-conditions for variables
+    if (!env || path.empty()) {
+      return Status::InvalidArgument("empty or null args");
+    }
+
+    // (2) assert size related invariants
+    // - cache size cannot be less than cache file size
+    // - individual write buffer size cannot be greater than cache file size
+    // - total write buffer size cannot be less than 2X cache file size
+    if (cache_size < cache_file_size || write_buffer_size >= cache_file_size ||
+        write_buffer_size * write_buffer_count() < 2 * cache_file_size) {
+      return Status::InvalidArgument("invalid cache size");
+    }
+
+    // (2) check writer settings
+    // - Queue depth cannot be 0
+    // - writer_dispatch_size cannot be greater than writer_buffer_size
+    // - dispatch size and buffer size need to be aligned
+    if (!writer_qdepth || writer_dispatch_size > write_buffer_size ||
+        write_buffer_size % writer_dispatch_size) {
+      return Status::InvalidArgument("invalid writer settings");
+    }
+
+    return Status::OK();
+  }
+
+  //
+  // Env abstraction to use for system level operations
+  //
+  Env* env;
+  SystemClock* clock;
+  //
+  // Path for the block cache where blocks are persisted
+  //
+  std::string path;
+
+  //
+  // Log handle for logging messages
+  //
+  std::shared_ptr<Logger> log;
+
+  //
+  // Enable direct IO for reading
+  //
+  bool enable_direct_reads = true;
+
+  //
+  // Enable direct IO for writing
+  //
+  bool enable_direct_writes = false;
+
+  //
+  // Logical cache size
+  //
+  uint64_t cache_size = std::numeric_limits<uint64_t>::max();
+
+  // cache-file-size
+  //
+  // Cache consists of multiples of small files. This parameter defines the
+  // size of an individual cache file
+  //
+  // default: 1M
+  uint32_t cache_file_size = 100ULL * 1024 * 1024;
+
+  // writer-qdepth
+  //
+  // The writers can issues IO to the devices in parallel. This parameter
+  // controls the max number if IOs that can issues in parallel to the block
+  // device
+  //
+  // default :1
+  uint32_t writer_qdepth = 1;
+
+  // pipeline-writes
+  //
+  // The write optionally follow pipelined architecture. This helps
+  // avoid regression in the eviction code path of the primary tier. This
+  // parameter defines if pipelining is enabled or disabled
+  //
+  // default: true
+  bool pipeline_writes = true;
+
+  // max-write-pipeline-backlog-size
+  //
+  // Max pipeline buffer size. This is the maximum backlog we can accumulate
+  // while waiting for writes. After the limit, new ops will be dropped.
+  //
+  // Default: 1GiB
+  uint64_t max_write_pipeline_backlog_size = 1ULL * 1024 * 1024 * 1024;
+
+  // write-buffer-size
+  //
+  // This is the size in which buffer slabs are allocated.
+  //
+  // Default: 1M
+  uint32_t write_buffer_size = 1ULL * 1024 * 1024;
+
+  // write-buffer-count
+  //
+  // This is the total number of buffer slabs. This is calculated as a factor of
+  // file size in order to avoid dead lock.
+  size_t write_buffer_count() const {
+    assert(write_buffer_size);
+    return static_cast<size_t>((writer_qdepth + 1.2) * cache_file_size /
+                               write_buffer_size);
+  }
+
+  // writer-dispatch-size
+  //
+  // The writer thread will dispatch the IO at the specified IO size
+  //
+  // default: 1M
+  uint64_t writer_dispatch_size = 1ULL * 1024 * 1024;
+
+  // is_compressed
+  //
+  // This option determines if the cache will run in compressed mode or
+  // uncompressed mode
+  bool is_compressed = true;
+
+  PersistentCacheConfig MakePersistentCacheConfig(
+      const std::string& path, const uint64_t size,
+      const std::shared_ptr<Logger>& log);
+
+  std::string ToString() const;
+};
+
+// Persistent Cache Tier
+//
+// This a logical abstraction that defines a tier of the persistent cache. Tiers
+// can be stacked over one another. PersistentCahe provides the basic definition
+// for accessing/storing in the cache. PersistentCacheTier extends the interface
+// to enable management and stacking of tiers.
+class PersistentCacheTier : public PersistentCache {
+ public:
+  using Tier = std::shared_ptr<PersistentCacheTier>;
+
+  virtual ~PersistentCacheTier() {}
+
+  // Open the persistent cache tier
+  virtual Status Open();
+
+  // Close the persistent cache tier
+  virtual Status Close();
+
+  // Reserve space up to 'size' bytes
+  virtual bool Reserve(const size_t size);
+
+  // Erase a key from the cache
+  virtual bool Erase(const Slice& key);
+
+  // Print stats to string recursively
+  virtual std::string PrintStats();
+
+  virtual PersistentCache::StatsType Stats() override;
+
+  // Insert to page cache
+  virtual Status Insert(const Slice& page_key, const char* data,
+                        const size_t size) override = 0;
+
+  // Lookup page cache by page identifier
+  virtual Status Lookup(const Slice& page_key, std::unique_ptr<char[]>* data,
+                        size_t* size) override = 0;
+
+  // Does it store compressed data ?
+  virtual bool IsCompressed() override = 0;
+
+  virtual std::string GetPrintableOptions() const override = 0;
+
+  virtual uint64_t NewId() override;
+
+  // Return a reference to next tier
+  virtual Tier& next_tier() { return next_tier_; }
+
+  // Set the value for next tier
+  virtual void set_next_tier(const Tier& tier) {
+    assert(!next_tier_);
+    next_tier_ = tier;
+  }
+
+  virtual void TEST_Flush() {
+    if (next_tier_) {
+      next_tier_->TEST_Flush();
+    }
+  }
+
+ private:
+  Tier next_tier_;  // next tier
+  std::atomic<uint64_t> last_id_{1};
+};
+
+// PersistentTieredCache
+//
+// Abstraction that helps you construct a tiers of persistent caches as a
+// unified cache. The tier(s) of cache will act a single tier for management
+// ease and support PersistentCache methods for accessing data.
+class PersistentTieredCache : public PersistentCacheTier {
+ public:
+  virtual ~PersistentTieredCache();
+
+  Status Open() override;
+  Status Close() override;
+  bool Erase(const Slice& key) override;
+  std::string PrintStats() override;
+  PersistentCache::StatsType Stats() override;
+  Status Insert(const Slice& page_key, const char* data,
+                const size_t size) override;
+  Status Lookup(const Slice& page_key, std::unique_ptr<char[]>* data,
+                size_t* size) override;
+  bool IsCompressed() override;
+
+  std::string GetPrintableOptions() const override {
+    return "PersistentTieredCache";
+  }
+
+  void AddTier(const Tier& tier);
+
+  Tier& next_tier() override {
+    auto it = tiers_.end();
+    return (*it)->next_tier();
+  }
+
+  void set_next_tier(const Tier& tier) override {
+    auto it = tiers_.end();
+    (*it)->set_next_tier(tier);
+  }
+
+  void TEST_Flush() override {
+    assert(!tiers_.empty());
+    tiers_.front()->TEST_Flush();
+    PersistentCacheTier::TEST_Flush();
+  }
+
+ protected:
+  std::list<Tier> tiers_;  // list of tiers top-down
+};
+
+}  // namespace ROCKSDB_NAMESPACE
+
+#endif
diff --git a/src/rocksdb/utilities/persistent_cache/persistent_cache_util.h b/src/rocksdb/utilities/persistent_cache/persistent_cache_util.h
new file mode 100644
index 000000000..2a769652d
--- /dev/null
+++ b/src/rocksdb/utilities/persistent_cache/persistent_cache_util.h
@@ -0,0 +1,67 @@
+//  Copyright (c) 2013, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+#pragma once
+
+#include <limits>
+#include <list>
+
+#include "util/mutexlock.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+//
+// Simple synchronized queue implementation with the option of
+// bounding the queue
+//
+// On overflow, the elements will be discarded
+//
+template <class T>
+class BoundedQueue {
+ public:
+  explicit BoundedQueue(
+      const size_t max_size = std::numeric_limits<size_t>::max())
+      : cond_empty_(&lock_), max_size_(max_size) {}
+
+  virtual ~BoundedQueue() {}
+
+  void Push(T&& t) {
+    MutexLock _(&lock_);
+    if (max_size_ != std::numeric_limits<size_t>::max() &&
+        size_ + t.Size() >= max_size_) {
+      // overflow
+      return;
+    }
+
+    size_ += t.Size();
+    q_.push_back(std::move(t));
+    cond_empty_.SignalAll();
+  }
+
+  T Pop() {
+    MutexLock _(&lock_);
+    while (q_.empty()) {
+      cond_empty_.Wait();
+    }
+
+    T t = std::move(q_.front());
+    size_ -= t.Size();
+    q_.pop_front();
+    return t;
+  }
+
+  size_t Size() const {
+    MutexLock _(&lock_);
+    return size_;
+  }
+
+ private:
+  mutable port::Mutex lock_;
+  port::CondVar cond_empty_;
+  std::list<T> q_;
+  size_t size_ = 0;
+  const size_t max_size_;
+};
+
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/utilities/persistent_cache/volatile_tier_impl.cc b/src/rocksdb/utilities/persistent_cache/volatile_tier_impl.cc
new file mode 100644
index 000000000..45d2830aa
--- /dev/null
+++ b/src/rocksdb/utilities/persistent_cache/volatile_tier_impl.cc
@@ -0,0 +1,140 @@
+//  Copyright (c) 2013, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+//
+#ifndef ROCKSDB_LITE
+
+#include "utilities/persistent_cache/volatile_tier_impl.h"
+
+#include <string>
+
+namespace ROCKSDB_NAMESPACE {
+
+void VolatileCacheTier::DeleteCacheData(VolatileCacheTier::CacheData* data) {
+  assert(data);
+  delete data;
+}
+
+VolatileCacheTier::~VolatileCacheTier() { index_.Clear(&DeleteCacheData); }
+
+PersistentCache::StatsType VolatileCacheTier::Stats() {
+  std::map<std::string, double> stat;
+  stat.insert({"persistent_cache.volatile_cache.hits",
+               static_cast<double>(stats_.cache_hits_)});
+  stat.insert({"persistent_cache.volatile_cache.misses",
+               static_cast<double>(stats_.cache_misses_)});
+  stat.insert({"persistent_cache.volatile_cache.inserts",
+               static_cast<double>(stats_.cache_inserts_)});
+  stat.insert({"persistent_cache.volatile_cache.evicts",
+               static_cast<double>(stats_.cache_evicts_)});
+  stat.insert({"persistent_cache.volatile_cache.hit_pct",
+               static_cast<double>(stats_.CacheHitPct())});
+  stat.insert({"persistent_cache.volatile_cache.miss_pct",
+               static_cast<double>(stats_.CacheMissPct())});
+
+  auto out = PersistentCacheTier::Stats();
+  out.push_back(stat);
+  return out;
+}
+
+Status VolatileCacheTier::Insert(const Slice& page_key, const char* data,
+                                 const size_t size) {
+  // precondition
+  assert(data);
+  assert(size);
+
+  // increment the size
+  size_ += size;
+
+  // check if we have overshot the limit, if so evict some space
+  while (size_ > max_size_) {
+    if (!Evict()) {
+      // unable to evict data, we give up so we don't spike read
+      // latency
+      assert(size_ >= size);
+      size_ -= size;
+      return Status::TryAgain("Unable to evict any data");
+    }
+  }
+
+  assert(size_ >= size);
+
+  // insert order: LRU, followed by index
+  std::string key(page_key.data(), page_key.size());
+  std::string value(data, size);
+  std::unique_ptr<CacheData> cache_data(
+      new CacheData(std::move(key), std::move(value)));
+  bool ok = index_.Insert(cache_data.get());
+  if (!ok) {
+    // decrement the size that we incremented ahead of time
+    assert(size_ >= size);
+    size_ -= size;
+    // failed to insert to cache, block already in cache
+    return Status::TryAgain("key already exists in volatile cache");
+  }
+
+  cache_data.release();
+  stats_.cache_inserts_++;
+  return Status::OK();
+}
+
+Status VolatileCacheTier::Lookup(const Slice& page_key,
+                                 std::unique_ptr<char[]>* result,
+                                 size_t* size) {
+  CacheData key(std::move(page_key.ToString()));
+  CacheData* kv;
+  bool ok = index_.Find(&key, &kv);
+  if (ok) {
+    // set return data
+    result->reset(new char[kv->value.size()]);
+    memcpy(result->get(), kv->value.c_str(), kv->value.size());
+    *size = kv->value.size();
+    // drop the reference on cache data
+    kv->refs_--;
+    // update stats
+    stats_.cache_hits_++;
+    return Status::OK();
+  }
+
+  stats_.cache_misses_++;
+
+  if (next_tier()) {
+    return next_tier()->Lookup(page_key, result, size);
+  }
+
+  return Status::NotFound("key not found in volatile cache");
+}
+
+bool VolatileCacheTier::Erase(const Slice& /*key*/) {
+  assert(!"not supported");
+  return true;
+}
+
+bool VolatileCacheTier::Evict() {
+  CacheData* edata = index_.Evict();
+  if (!edata) {
+    // not able to evict any object
+    return false;
+  }
+
+  stats_.cache_evicts_++;
+
+  // push the evicted object to the next level
+  if (next_tier()) {
+    // TODO: Should the insert error be ignored?
+    Status s = next_tier()->Insert(Slice(edata->key), edata->value.c_str(),
+                                   edata->value.size());
+    s.PermitUncheckedError();
+  }
+
+  // adjust size and destroy data
+  size_ -= edata->value.size();
+  delete edata;
+
+  return true;
+}
+
+}  // namespace ROCKSDB_NAMESPACE
+
+#endif
diff --git a/src/rocksdb/utilities/persistent_cache/volatile_tier_impl.h b/src/rocksdb/utilities/persistent_cache/volatile_tier_impl.h
new file mode 100644
index 000000000..09265e457
--- /dev/null
+++ b/src/rocksdb/utilities/persistent_cache/volatile_tier_impl.h
@@ -0,0 +1,141 @@
+//  Copyright (c) 2013, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+//
+#pragma once
+
+#ifndef ROCKSDB_LITE
+
+#include <atomic>
+#include <limits>
+#include <sstream>
+#include <string>
+#include <vector>
+
+#include "rocksdb/cache.h"
+#include "utilities/persistent_cache/hash_table.h"
+#include "utilities/persistent_cache/hash_table_evictable.h"
+#include "utilities/persistent_cache/persistent_cache_tier.h"
+
+// VolatileCacheTier
+//
+// This file provides persistent cache tier implementation for caching
+// key/values in RAM.
+//
+//        key/values
+//           |
+//           V
+// +-------------------+
+// | VolatileCacheTier | Store in an evictable hash table
+// +-------------------+
+//           |
+//           V
+//       on eviction
+//   pushed to next tier
+//
+// The implementation is designed to be concurrent. The evictable hash table
+// implementation is not concurrent at this point though.
+//
+// The eviction algorithm is LRU
+namespace ROCKSDB_NAMESPACE {
+
+class VolatileCacheTier : public PersistentCacheTier {
+ public:
+  explicit VolatileCacheTier(
+      const bool is_compressed = true,
+      const size_t max_size = std::numeric_limits<size_t>::max())
+      : is_compressed_(is_compressed), max_size_(max_size) {}
+
+  virtual ~VolatileCacheTier();
+
+  // insert to cache
+  Status Insert(const Slice& page_key, const char* data,
+                const size_t size) override;
+  // lookup key in cache
+  Status Lookup(const Slice& page_key, std::unique_ptr<char[]>* data,
+                size_t* size) override;
+
+  // is compressed cache ?
+  bool IsCompressed() override { return is_compressed_; }
+
+  // erase key from cache
+  bool Erase(const Slice& key) override;
+
+  std::string GetPrintableOptions() const override {
+    return "VolatileCacheTier";
+  }
+
+  // Expose stats as map
+  PersistentCache::StatsType Stats() override;
+
+ private:
+  //
+  // Cache data abstraction
+  //
+  struct CacheData : LRUElement<CacheData> {
+    explicit CacheData(CacheData&& rhs) noexcept
+        : key(std::move(rhs.key)), value(std::move(rhs.value)) {}
+
+    explicit CacheData(const std::string& _key, const std::string& _value = "")
+        : key(_key), value(_value) {}
+
+    virtual ~CacheData() {}
+
+    const std::string key;
+    const std::string value;
+  };
+
+  static void DeleteCacheData(CacheData* data);
+
+  //
+  // Index and LRU definition
+  //
+  struct CacheDataHash {
+    uint64_t operator()(const CacheData* obj) const {
+      assert(obj);
+      return std::hash<std::string>()(obj->key);
+    }
+  };
+
+  struct CacheDataEqual {
+    bool operator()(const CacheData* lhs, const CacheData* rhs) const {
+      assert(lhs);
+      assert(rhs);
+      return lhs->key == rhs->key;
+    }
+  };
+
+  struct Statistics {
+    std::atomic<uint64_t> cache_misses_{0};
+    std::atomic<uint64_t> cache_hits_{0};
+    std::atomic<uint64_t> cache_inserts_{0};
+    std::atomic<uint64_t> cache_evicts_{0};
+
+    double CacheHitPct() const {
+      auto lookups = cache_hits_ + cache_misses_;
+      return lookups ? 100 * cache_hits_ / static_cast<double>(lookups) : 0.0;
+    }
+
+    double CacheMissPct() const {
+      auto lookups = cache_hits_ + cache_misses_;
+      return lookups ? 100 * cache_misses_ / static_cast<double>(lookups) : 0.0;
+    }
+  };
+
+  using IndexType =
+      EvictableHashTable<CacheData, CacheDataHash, CacheDataEqual>;
+
+  // Evict LRU tail
+  bool Evict();
+
+  const bool is_compressed_ = true;    // does it store compressed data
+  IndexType index_;                    // in-memory cache
+  std::atomic<uint64_t> max_size_{0};  // Maximum size of the cache
+  std::atomic<uint64_t> size_{0};      // Size of the cache
+  Statistics stats_;
+};
+
+}  // namespace ROCKSDB_NAMESPACE
+
+#endif
diff --git a/src/rocksdb/utilities/simulator_cache/cache_simulator.cc b/src/rocksdb/utilities/simulator_cache/cache_simulator.cc
new file mode 100644
index 000000000..dc419e51a
--- /dev/null
+++ b/src/rocksdb/utilities/simulator_cache/cache_simulator.cc
@@ -0,0 +1,288 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#include "utilities/simulator_cache/cache_simulator.h"
+
+#include <algorithm>
+
+#include "db/dbformat.h"
+#include "rocksdb/trace_record.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+namespace {
+const std::string kGhostCachePrefix = "ghost_";
+}  // namespace
+
+GhostCache::GhostCache(std::shared_ptr<Cache> sim_cache)
+    : sim_cache_(sim_cache) {}
+
+bool GhostCache::Admit(const Slice& lookup_key) {
+  auto handle = sim_cache_->Lookup(lookup_key);
+  if (handle != nullptr) {
+    sim_cache_->Release(handle);
+    return true;
+  }
+  // TODO: Should we check for errors here?
+  auto s = sim_cache_->Insert(lookup_key, /*value=*/nullptr, lookup_key.size(),
+                              /*deleter=*/nullptr);
+  s.PermitUncheckedError();
+  return false;
+}
+
+CacheSimulator::CacheSimulator(std::unique_ptr<GhostCache>&& ghost_cache,
+                               std::shared_ptr<Cache> sim_cache)
+    : ghost_cache_(std::move(ghost_cache)), sim_cache_(sim_cache) {}
+
+void CacheSimulator::Access(const BlockCacheTraceRecord& access) {
+  bool admit = true;
+  const bool is_user_access =
+      BlockCacheTraceHelper::IsUserAccess(access.caller);
+  bool is_cache_miss = true;
+  if (ghost_cache_ && !access.no_insert) {
+    admit = ghost_cache_->Admit(access.block_key);
+  }
+  auto handle = sim_cache_->Lookup(access.block_key);
+  if (handle != nullptr) {
+    sim_cache_->Release(handle);
+    is_cache_miss = false;
+  } else {
+    if (!access.no_insert && admit && access.block_size > 0) {
+      // Ignore errors on insert
+      auto s = sim_cache_->Insert(access.block_key, /*value=*/nullptr,
+                                  access.block_size,
+                                  /*deleter=*/nullptr);
+      s.PermitUncheckedError();
+    }
+  }
+  miss_ratio_stats_.UpdateMetrics(access.access_timestamp, is_user_access,
+                                  is_cache_miss);
+}
+
+void MissRatioStats::UpdateMetrics(uint64_t timestamp_in_ms,
+                                   bool is_user_access, bool is_cache_miss) {
+  uint64_t timestamp_in_seconds = timestamp_in_ms / kMicrosInSecond;
+  num_accesses_timeline_[timestamp_in_seconds] += 1;
+  num_accesses_ += 1;
+  if (num_misses_timeline_.find(timestamp_in_seconds) ==
+      num_misses_timeline_.end()) {
+    num_misses_timeline_[timestamp_in_seconds] = 0;
+  }
+  if (is_cache_miss) {
+    num_misses_ += 1;
+    num_misses_timeline_[timestamp_in_seconds] += 1;
+  }
+  if (is_user_access) {
+    user_accesses_ += 1;
+    if (is_cache_miss) {
+      user_misses_ += 1;
+    }
+  }
+}
+
+Cache::Priority PrioritizedCacheSimulator::ComputeBlockPriority(
+    const BlockCacheTraceRecord& access) const {
+  if (access.block_type == TraceType::kBlockTraceFilterBlock ||
+      access.block_type == TraceType::kBlockTraceIndexBlock ||
+      access.block_type == TraceType::kBlockTraceUncompressionDictBlock) {
+    return Cache::Priority::HIGH;
+  }
+  return Cache::Priority::LOW;
+}
+
+void PrioritizedCacheSimulator::AccessKVPair(
+    const Slice& key, uint64_t value_size, Cache::Priority priority,
+    const BlockCacheTraceRecord& access, bool no_insert, bool is_user_access,
+    bool* is_cache_miss, bool* admitted, bool update_metrics) {
+  assert(is_cache_miss);
+  assert(admitted);
+  *is_cache_miss = true;
+  *admitted = true;
+  if (ghost_cache_ && !no_insert) {
+    *admitted = ghost_cache_->Admit(key);
+  }
+  auto handle = sim_cache_->Lookup(key);
+  if (handle != nullptr) {
+    sim_cache_->Release(handle);
+    *is_cache_miss = false;
+  } else if (!no_insert && *admitted && value_size > 0) {
+    // TODO: Should we check for an error here?
+    auto s = sim_cache_->Insert(key, /*value=*/nullptr, value_size,
+                                /*deleter=*/nullptr,
+                                /*handle=*/nullptr, priority);
+    s.PermitUncheckedError();
+  }
+  if (update_metrics) {
+    miss_ratio_stats_.UpdateMetrics(access.access_timestamp, is_user_access,
+                                    *is_cache_miss);
+  }
+}
+
+void PrioritizedCacheSimulator::Access(const BlockCacheTraceRecord& access) {
+  bool is_cache_miss = true;
+  bool admitted = true;
+  AccessKVPair(access.block_key, access.block_size,
+               ComputeBlockPriority(access), access, access.no_insert,
+               BlockCacheTraceHelper::IsUserAccess(access.caller),
+               &is_cache_miss, &admitted, /*update_metrics=*/true);
+}
+
+void HybridRowBlockCacheSimulator::Access(const BlockCacheTraceRecord& access) {
+  // TODO (haoyu): We only support Get for now. We need to extend the tracing
+  // for MultiGet, i.e., non-data block accesses must log all keys in a
+  // MultiGet.
+  bool is_cache_miss = true;
+  bool admitted = false;
+  if (access.caller == TableReaderCaller::kUserGet &&
+      access.get_id != BlockCacheTraceHelper::kReservedGetId) {
+    // This is a Get request.
+    const std::string& row_key = BlockCacheTraceHelper::ComputeRowKey(access);
+    GetRequestStatus& status = getid_status_map_[access.get_id];
+    if (status.is_complete) {
+      // This Get request completes.
+      // Skip future accesses to its index/filter/data
+      // blocks. These block lookups are unnecessary if we observe a hit for the
+      // referenced key-value pair already. Thus, we treat these lookups as
+      // hits. This is also to ensure the total number of accesses are the same
+      // when comparing to other policies.
+      miss_ratio_stats_.UpdateMetrics(access.access_timestamp,
+                                      /*is_user_access=*/true,
+                                      /*is_cache_miss=*/false);
+      return;
+    }
+    if (status.row_key_status.find(row_key) == status.row_key_status.end()) {
+      // This is the first time that this key is accessed. Look up the key-value
+      // pair first. Do not update the miss/accesses metrics here since it will
+      // be updated later.
+      AccessKVPair(row_key, access.referenced_data_size, Cache::Priority::HIGH,
+                   access,
+                   /*no_insert=*/false,
+                   /*is_user_access=*/true, &is_cache_miss, &admitted,
+                   /*update_metrics=*/false);
+      InsertResult result = InsertResult::NO_INSERT;
+      if (admitted && access.referenced_data_size > 0) {
+        result = InsertResult::INSERTED;
+      } else if (admitted) {
+        result = InsertResult::ADMITTED;
+      }
+      status.row_key_status[row_key] = result;
+    }
+    if (!is_cache_miss) {
+      // A cache hit.
+      status.is_complete = true;
+      miss_ratio_stats_.UpdateMetrics(access.access_timestamp,
+                                      /*is_user_access=*/true,
+                                      /*is_cache_miss=*/false);
+      return;
+    }
+    // The row key-value pair observes a cache miss. We need to access its
+    // index/filter/data blocks.
+    InsertResult inserted = status.row_key_status[row_key];
+    AccessKVPair(
+        access.block_key, access.block_size, ComputeBlockPriority(access),
+        access,
+        /*no_insert=*/!insert_blocks_upon_row_kvpair_miss_ || access.no_insert,
+        /*is_user_access=*/true, &is_cache_miss, &admitted,
+        /*update_metrics=*/true);
+    if (access.referenced_data_size > 0 && inserted == InsertResult::ADMITTED) {
+      // TODO: Should we check for an error here?
+      auto s = sim_cache_->Insert(row_key, /*value=*/nullptr,
+                                  access.referenced_data_size,
+                                  /*deleter=*/nullptr,
+                                  /*handle=*/nullptr, Cache::Priority::HIGH);
+      s.PermitUncheckedError();
+      status.row_key_status[row_key] = InsertResult::INSERTED;
+    }
+    return;
+  }
+  AccessKVPair(access.block_key, access.block_size,
+               ComputeBlockPriority(access), access, access.no_insert,
+               BlockCacheTraceHelper::IsUserAccess(access.caller),
+               &is_cache_miss, &admitted, /*update_metrics=*/true);
+}
+
+BlockCacheTraceSimulator::BlockCacheTraceSimulator(
+    uint64_t warmup_seconds, uint32_t downsample_ratio,
+    const std::vector<CacheConfiguration>& cache_configurations)
+    : warmup_seconds_(warmup_seconds),
+      downsample_ratio_(downsample_ratio),
+      cache_configurations_(cache_configurations) {}
+
+Status BlockCacheTraceSimulator::InitializeCaches() {
+  for (auto const& config : cache_configurations_) {
+    for (auto cache_capacity : config.cache_capacities) {
+      // Scale down the cache capacity since the trace contains accesses on
+      // 1/'downsample_ratio' blocks.
+      uint64_t simulate_cache_capacity = cache_capacity / downsample_ratio_;
+      std::shared_ptr<CacheSimulator> sim_cache;
+      std::unique_ptr<GhostCache> ghost_cache;
+      std::string cache_name = config.cache_name;
+      if (cache_name.find(kGhostCachePrefix) != std::string::npos) {
+        ghost_cache.reset(new GhostCache(
+            NewLRUCache(config.ghost_cache_capacity, /*num_shard_bits=*/1,
+                        /*strict_capacity_limit=*/false,
+                        /*high_pri_pool_ratio=*/0)));
+        cache_name = cache_name.substr(kGhostCachePrefix.size());
+      }
+      if (cache_name == "lru") {
+        sim_cache = std::make_shared<CacheSimulator>(
+            std::move(ghost_cache),
+            NewLRUCache(simulate_cache_capacity, config.num_shard_bits,
+                        /*strict_capacity_limit=*/false,
+                        /*high_pri_pool_ratio=*/0));
+      } else if (cache_name == "lru_priority") {
+        sim_cache = std::make_shared<PrioritizedCacheSimulator>(
+            std::move(ghost_cache),
+            NewLRUCache(simulate_cache_capacity, config.num_shard_bits,
+                        /*strict_capacity_limit=*/false,
+                        /*high_pri_pool_ratio=*/0.5));
+      } else if (cache_name == "lru_hybrid") {
+        sim_cache = std::make_shared<HybridRowBlockCacheSimulator>(
+            std::move(ghost_cache),
+            NewLRUCache(simulate_cache_capacity, config.num_shard_bits,
+                        /*strict_capacity_limit=*/false,
+                        /*high_pri_pool_ratio=*/0.5),
+            /*insert_blocks_upon_row_kvpair_miss=*/true);
+      } else if (cache_name == "lru_hybrid_no_insert_on_row_miss") {
+        sim_cache = std::make_shared<HybridRowBlockCacheSimulator>(
+            std::move(ghost_cache),
+            NewLRUCache(simulate_cache_capacity, config.num_shard_bits,
+                        /*strict_capacity_limit=*/false,
+                        /*high_pri_pool_ratio=*/0.5),
+            /*insert_blocks_upon_row_kvpair_miss=*/false);
+      } else {
+        // Not supported.
+        return Status::InvalidArgument("Unknown cache name " +
+                                       config.cache_name);
+      }
+      sim_caches_[config].push_back(sim_cache);
+    }
+  }
+  return Status::OK();
+}
+
+void BlockCacheTraceSimulator::Access(const BlockCacheTraceRecord& access) {
+  if (trace_start_time_ == 0) {
+    trace_start_time_ = access.access_timestamp;
+  }
+  // access.access_timestamp is in microseconds.
+  if (!warmup_complete_ &&
+      trace_start_time_ + warmup_seconds_ * kMicrosInSecond <=
+          access.access_timestamp) {
+    for (auto& config_caches : sim_caches_) {
+      for (auto& sim_cache : config_caches.second) {
+        sim_cache->reset_counter();
+      }
+    }
+    warmup_complete_ = true;
+  }
+  for (auto& config_caches : sim_caches_) {
+    for (auto& sim_cache : config_caches.second) {
+      sim_cache->Access(access);
+    }
+  }
+}
+
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/utilities/simulator_cache/cache_simulator.h b/src/rocksdb/utilities/simulator_cache/cache_simulator.h
new file mode 100644
index 000000000..6d4979013
--- /dev/null
+++ b/src/rocksdb/utilities/simulator_cache/cache_simulator.h
@@ -0,0 +1,231 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#pragma once
+
+#include <unordered_map>
+
+#include "cache/lru_cache.h"
+#include "trace_replay/block_cache_tracer.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+// A cache configuration provided by user.
+struct CacheConfiguration {
+  std::string cache_name;  // LRU.
+  uint32_t num_shard_bits;
+  uint64_t ghost_cache_capacity;  // ghost cache capacity in bytes.
+  std::vector<uint64_t>
+      cache_capacities;  // simulate cache capacities in bytes.
+
+  bool operator==(const CacheConfiguration& o) const {
+    return cache_name == o.cache_name && num_shard_bits == o.num_shard_bits &&
+           ghost_cache_capacity == o.ghost_cache_capacity;
+  }
+  bool operator<(const CacheConfiguration& o) const {
+    return cache_name < o.cache_name ||
+           (cache_name == o.cache_name && num_shard_bits < o.num_shard_bits) ||
+           (cache_name == o.cache_name && num_shard_bits == o.num_shard_bits &&
+            ghost_cache_capacity < o.ghost_cache_capacity);
+  }
+};
+
+class MissRatioStats {
+ public:
+  void reset_counter() {
+    num_misses_ = 0;
+    num_accesses_ = 0;
+    user_accesses_ = 0;
+    user_misses_ = 0;
+  }
+  double miss_ratio() const {
+    if (num_accesses_ == 0) {
+      return -1;
+    }
+    return static_cast<double>(num_misses_ * 100.0 / num_accesses_);
+  }
+  uint64_t total_accesses() const { return num_accesses_; }
+  uint64_t total_misses() const { return num_misses_; }
+
+  const std::map<uint64_t, uint64_t>& num_accesses_timeline() const {
+    return num_accesses_timeline_;
+  }
+
+  const std::map<uint64_t, uint64_t>& num_misses_timeline() const {
+    return num_misses_timeline_;
+  }
+
+  double user_miss_ratio() const {
+    if (user_accesses_ == 0) {
+      return -1;
+    }
+    return static_cast<double>(user_misses_ * 100.0 / user_accesses_);
+  }
+  uint64_t user_accesses() const { return user_accesses_; }
+  uint64_t user_misses() const { return user_misses_; }
+
+  void UpdateMetrics(uint64_t timestamp_in_ms, bool is_user_access,
+                     bool is_cache_miss);
+
+ private:
+  uint64_t num_accesses_ = 0;
+  uint64_t num_misses_ = 0;
+  uint64_t user_accesses_ = 0;
+  uint64_t user_misses_ = 0;
+
+  std::map<uint64_t, uint64_t> num_accesses_timeline_;
+  std::map<uint64_t, uint64_t> num_misses_timeline_;
+};
+
+// A ghost cache admits an entry on its second access.
+class GhostCache {
+ public:
+  explicit GhostCache(std::shared_ptr<Cache> sim_cache);
+  ~GhostCache() = default;
+  // No copy and move.
+  GhostCache(const GhostCache&) = delete;
+  GhostCache& operator=(const GhostCache&) = delete;
+  GhostCache(GhostCache&&) = delete;
+  GhostCache& operator=(GhostCache&&) = delete;
+
+  // Returns true if the lookup_key is in the ghost cache.
+  // Returns false otherwise.
+  bool Admit(const Slice& lookup_key);
+
+ private:
+  std::shared_ptr<Cache> sim_cache_;
+};
+
+// A cache simulator that runs against a block cache trace.
+class CacheSimulator {
+ public:
+  CacheSimulator(std::unique_ptr<GhostCache>&& ghost_cache,
+                 std::shared_ptr<Cache> sim_cache);
+  virtual ~CacheSimulator() = default;
+  // No copy and move.
+  CacheSimulator(const CacheSimulator&) = delete;
+  CacheSimulator& operator=(const CacheSimulator&) = delete;
+  CacheSimulator(CacheSimulator&&) = delete;
+  CacheSimulator& operator=(CacheSimulator&&) = delete;
+
+  virtual void Access(const BlockCacheTraceRecord& access);
+
+  void reset_counter() { miss_ratio_stats_.reset_counter(); }
+
+  const MissRatioStats& miss_ratio_stats() const { return miss_ratio_stats_; }
+
+ protected:
+  MissRatioStats miss_ratio_stats_;
+  std::unique_ptr<GhostCache> ghost_cache_;
+  std::shared_ptr<Cache> sim_cache_;
+};
+
+// A prioritized cache simulator that runs against a block cache trace.
+// It inserts missing index/filter/uncompression-dictionary blocks with high
+// priority in the cache.
+class PrioritizedCacheSimulator : public CacheSimulator {
+ public:
+  PrioritizedCacheSimulator(std::unique_ptr<GhostCache>&& ghost_cache,
+                            std::shared_ptr<Cache> sim_cache)
+      : CacheSimulator(std::move(ghost_cache), sim_cache) {}
+  void Access(const BlockCacheTraceRecord& access) override;
+
+ protected:
+  // Access the key-value pair and returns true upon a cache miss.
+  void AccessKVPair(const Slice& key, uint64_t value_size,
+                    Cache::Priority priority,
+                    const BlockCacheTraceRecord& access, bool no_insert,
+                    bool is_user_access, bool* is_cache_miss, bool* admitted,
+                    bool update_metrics);
+
+  Cache::Priority ComputeBlockPriority(
+      const BlockCacheTraceRecord& access) const;
+};
+
+// A hybrid row and block cache simulator. It looks up/inserts key-value pairs
+// referenced by Get/MultiGet requests, and not their accessed index/filter/data
+// blocks.
+//
+// Upon a Get/MultiGet request, it looks up the referenced key first.
+// If it observes a cache hit, future block accesses on this key-value pair is
+// skipped since the request is served already. Otherwise, it continues to look
+// up/insert its index/filter/data blocks. It also inserts the referenced
+// key-value pair in the cache for future lookups.
+class HybridRowBlockCacheSimulator : public PrioritizedCacheSimulator {
+ public:
+  HybridRowBlockCacheSimulator(std::unique_ptr<GhostCache>&& ghost_cache,
+                               std::shared_ptr<Cache> sim_cache,
+                               bool insert_blocks_upon_row_kvpair_miss)
+      : PrioritizedCacheSimulator(std::move(ghost_cache), sim_cache),
+        insert_blocks_upon_row_kvpair_miss_(
+            insert_blocks_upon_row_kvpair_miss) {}
+  void Access(const BlockCacheTraceRecord& access) override;
+
+ private:
+  enum InsertResult : char {
+    INSERTED,
+    ADMITTED,
+    NO_INSERT,
+  };
+
+  // We set is_complete to true when the referenced row-key of a get request
+  // hits the cache. If is_complete is true, we treat future accesses of this
+  // get request as hits.
+  //
+  // For each row key, it stores an enum. It is INSERTED when the
+  // kv-pair has been inserted into the cache, ADMITTED if it should be inserted
+  // but haven't been, NO_INSERT if it should not be inserted.
+  //
+  // A kv-pair is in ADMITTED state when we encounter this kv-pair but do not
+  // know its size. This may happen if the first access on the referenced key is
+  // an index/filter block.
+  struct GetRequestStatus {
+    bool is_complete = false;
+    std::map<std::string, InsertResult> row_key_status;
+  };
+
+  // A map stores get_id to a map of row keys.
+  std::map<uint64_t, GetRequestStatus> getid_status_map_;
+  bool insert_blocks_upon_row_kvpair_miss_;
+};
+
+// A block cache simulator that reports miss ratio curves given a set of cache
+// configurations.
+class BlockCacheTraceSimulator {
+ public:
+  // warmup_seconds: The number of seconds to warmup simulated caches. The
+  // hit/miss counters are reset after the warmup completes.
+  BlockCacheTraceSimulator(
+      uint64_t warmup_seconds, uint32_t downsample_ratio,
+      const std::vector<CacheConfiguration>& cache_configurations);
+  ~BlockCacheTraceSimulator() = default;
+  // No copy and move.
+  BlockCacheTraceSimulator(const BlockCacheTraceSimulator&) = delete;
+  BlockCacheTraceSimulator& operator=(const BlockCacheTraceSimulator&) = delete;
+  BlockCacheTraceSimulator(BlockCacheTraceSimulator&&) = delete;
+  BlockCacheTraceSimulator& operator=(BlockCacheTraceSimulator&&) = delete;
+
+  Status InitializeCaches();
+
+  void Access(const BlockCacheTraceRecord& access);
+
+  const std::map<CacheConfiguration,
+                 std::vector<std::shared_ptr<CacheSimulator>>>&
+  sim_caches() const {
+    return sim_caches_;
+  }
+
+ private:
+  const uint64_t warmup_seconds_;
+  const uint32_t downsample_ratio_;
+  const std::vector<CacheConfiguration> cache_configurations_;
+
+  bool warmup_complete_ = false;
+  std::map<CacheConfiguration, std::vector<std::shared_ptr<CacheSimulator>>>
+      sim_caches_;
+  uint64_t trace_start_time_ = 0;
+};
+
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/utilities/simulator_cache/cache_simulator_test.cc b/src/rocksdb/utilities/simulator_cache/cache_simulator_test.cc
new file mode 100644
index 000000000..2bc057c92
--- /dev/null
+++ b/src/rocksdb/utilities/simulator_cache/cache_simulator_test.cc
@@ -0,0 +1,497 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#include "utilities/simulator_cache/cache_simulator.h"
+
+#include <cstdlib>
+
+#include "rocksdb/env.h"
+#include "rocksdb/trace_record.h"
+#include "test_util/testharness.h"
+#include "test_util/testutil.h"
+
+namespace ROCKSDB_NAMESPACE {
+namespace {
+const std::string kBlockKeyPrefix = "test-block-";
+const std::string kRefKeyPrefix = "test-get-";
+const std::string kRefKeySequenceNumber = std::string(8, 'c');
+const uint64_t kGetId = 1;
+const uint64_t kGetBlockId = 100;
+const uint64_t kCompactionBlockId = 1000;
+const uint64_t kCacheSize = 1024 * 1024 * 1024;
+const uint64_t kGhostCacheSize = 1024 * 1024;
+}  // namespace
+
+class CacheSimulatorTest : public testing::Test {
+ public:
+  const size_t kNumBlocks = 5;
+  const size_t kValueSize = 1000;
+
+  CacheSimulatorTest() { env_ = ROCKSDB_NAMESPACE::Env::Default(); }
+
+  BlockCacheTraceRecord GenerateGetRecord(uint64_t getid) {
+    BlockCacheTraceRecord record;
+    record.block_type = TraceType::kBlockTraceDataBlock;
+    record.block_size = 4096;
+    record.block_key = kBlockKeyPrefix + std::to_string(kGetBlockId);
+    record.access_timestamp = env_->NowMicros();
+    record.cf_id = 0;
+    record.cf_name = "test";
+    record.caller = TableReaderCaller::kUserGet;
+    record.level = 6;
+    record.sst_fd_number = 0;
+    record.get_id = getid;
+    record.is_cache_hit = false;
+    record.no_insert = false;
+    record.referenced_key =
+        kRefKeyPrefix + std::to_string(kGetId) + kRefKeySequenceNumber;
+    record.referenced_key_exist_in_block = true;
+    record.referenced_data_size = 100;
+    record.num_keys_in_block = 300;
+    return record;
+  }
+
+  BlockCacheTraceRecord GenerateCompactionRecord() {
+    BlockCacheTraceRecord record;
+    record.block_type = TraceType::kBlockTraceDataBlock;
+    record.block_size = 4096;
+    record.block_key = kBlockKeyPrefix + std::to_string(kCompactionBlockId);
+    record.access_timestamp = env_->NowMicros();
+    record.cf_id = 0;
+    record.cf_name = "test";
+    record.caller = TableReaderCaller::kCompaction;
+    record.level = 6;
+    record.sst_fd_number = kCompactionBlockId;
+    record.is_cache_hit = false;
+    record.no_insert = true;
+    return record;
+  }
+
+  void AssertCache(std::shared_ptr<Cache> sim_cache,
+                   const MissRatioStats& miss_ratio_stats,
+                   uint64_t expected_usage, uint64_t expected_num_accesses,
+                   uint64_t expected_num_misses,
+                   std::vector<std::string> blocks,
+                   std::vector<std::string> keys) {
+    EXPECT_EQ(expected_usage, sim_cache->GetUsage());
+    EXPECT_EQ(expected_num_accesses, miss_ratio_stats.total_accesses());
+    EXPECT_EQ(expected_num_misses, miss_ratio_stats.total_misses());
+    for (auto const& block : blocks) {
+      auto handle = sim_cache->Lookup(block);
+      EXPECT_NE(nullptr, handle);
+      sim_cache->Release(handle);
+    }
+    for (auto const& key : keys) {
+      std::string row_key = kRefKeyPrefix + key + kRefKeySequenceNumber;
+      auto handle =
+          sim_cache->Lookup("0_" + ExtractUserKey(row_key).ToString());
+      EXPECT_NE(nullptr, handle);
+      sim_cache->Release(handle);
+    }
+  }
+
+  Env* env_;
+};
+
+TEST_F(CacheSimulatorTest, GhostCache) {
+  const std::string key1 = "test1";
+  const std::string key2 = "test2";
+  std::unique_ptr<GhostCache> ghost_cache(new GhostCache(
+      NewLRUCache(/*capacity=*/kGhostCacheSize, /*num_shard_bits=*/1,
+                  /*strict_capacity_limit=*/false,
+                  /*high_pri_pool_ratio=*/0)));
+  EXPECT_FALSE(ghost_cache->Admit(key1));
+  EXPECT_TRUE(ghost_cache->Admit(key1));
+  EXPECT_TRUE(ghost_cache->Admit(key1));
+  EXPECT_FALSE(ghost_cache->Admit(key2));
+  EXPECT_TRUE(ghost_cache->Admit(key2));
+}
+
+TEST_F(CacheSimulatorTest, CacheSimulator) {
+  const BlockCacheTraceRecord& access = GenerateGetRecord(kGetId);
+  const BlockCacheTraceRecord& compaction_access = GenerateCompactionRecord();
+  std::shared_ptr<Cache> sim_cache =
+      NewLRUCache(/*capacity=*/kCacheSize, /*num_shard_bits=*/1,
+                  /*strict_capacity_limit=*/false,
+                  /*high_pri_pool_ratio=*/0);
+  std::unique_ptr<CacheSimulator> cache_simulator(
+      new CacheSimulator(nullptr, sim_cache));
+  cache_simulator->Access(access);
+  cache_simulator->Access(access);
+  ASSERT_EQ(2, cache_simulator->miss_ratio_stats().total_accesses());
+  ASSERT_EQ(50, cache_simulator->miss_ratio_stats().miss_ratio());
+  ASSERT_EQ(2, cache_simulator->miss_ratio_stats().user_accesses());
+  ASSERT_EQ(50, cache_simulator->miss_ratio_stats().user_miss_ratio());
+
+  cache_simulator->Access(compaction_access);
+  cache_simulator->Access(compaction_access);
+  ASSERT_EQ(4, cache_simulator->miss_ratio_stats().total_accesses());
+  ASSERT_EQ(75, cache_simulator->miss_ratio_stats().miss_ratio());
+  ASSERT_EQ(2, cache_simulator->miss_ratio_stats().user_accesses());
+  ASSERT_EQ(50, cache_simulator->miss_ratio_stats().user_miss_ratio());
+
+  cache_simulator->reset_counter();
+  ASSERT_EQ(0, cache_simulator->miss_ratio_stats().total_accesses());
+  ASSERT_EQ(-1, cache_simulator->miss_ratio_stats().miss_ratio());
+  auto handle = sim_cache->Lookup(access.block_key);
+  ASSERT_NE(nullptr, handle);
+  sim_cache->Release(handle);
+  handle = sim_cache->Lookup(compaction_access.block_key);
+  ASSERT_EQ(nullptr, handle);
+}
+
+TEST_F(CacheSimulatorTest, GhostCacheSimulator) {
+  const BlockCacheTraceRecord& access = GenerateGetRecord(kGetId);
+  std::unique_ptr<GhostCache> ghost_cache(new GhostCache(
+      NewLRUCache(/*capacity=*/kGhostCacheSize, /*num_shard_bits=*/1,
+                  /*strict_capacity_limit=*/false,
+                  /*high_pri_pool_ratio=*/0)));
+  std::unique_ptr<CacheSimulator> cache_simulator(new CacheSimulator(
+      std::move(ghost_cache),
+      NewLRUCache(/*capacity=*/kCacheSize, /*num_shard_bits=*/1,
+                  /*strict_capacity_limit=*/false,
+                  /*high_pri_pool_ratio=*/0)));
+  cache_simulator->Access(access);
+  cache_simulator->Access(access);
+  ASSERT_EQ(2, cache_simulator->miss_ratio_stats().total_accesses());
+  // Both of them will be miss since we have a ghost cache.
+  ASSERT_EQ(100, cache_simulator->miss_ratio_stats().miss_ratio());
+}
+
+TEST_F(CacheSimulatorTest, PrioritizedCacheSimulator) {
+  const BlockCacheTraceRecord& access = GenerateGetRecord(kGetId);
+  std::shared_ptr<Cache> sim_cache =
+      NewLRUCache(/*capacity=*/kCacheSize, /*num_shard_bits=*/1,
+                  /*strict_capacity_limit=*/false,
+                  /*high_pri_pool_ratio=*/0);
+  std::unique_ptr<PrioritizedCacheSimulator> cache_simulator(
+      new PrioritizedCacheSimulator(nullptr, sim_cache));
+  cache_simulator->Access(access);
+  cache_simulator->Access(access);
+  ASSERT_EQ(2, cache_simulator->miss_ratio_stats().total_accesses());
+  ASSERT_EQ(50, cache_simulator->miss_ratio_stats().miss_ratio());
+
+  auto handle = sim_cache->Lookup(access.block_key);
+  ASSERT_NE(nullptr, handle);
+  sim_cache->Release(handle);
+}
+
+TEST_F(CacheSimulatorTest, GhostPrioritizedCacheSimulator) {
+  const BlockCacheTraceRecord& access = GenerateGetRecord(kGetId);
+  std::unique_ptr<GhostCache> ghost_cache(new GhostCache(
+      NewLRUCache(/*capacity=*/kGhostCacheSize, /*num_shard_bits=*/1,
+                  /*strict_capacity_limit=*/false,
+                  /*high_pri_pool_ratio=*/0)));
+  std::unique_ptr<PrioritizedCacheSimulator> cache_simulator(
+      new PrioritizedCacheSimulator(
+          std::move(ghost_cache),
+          NewLRUCache(/*capacity=*/kCacheSize, /*num_shard_bits=*/1,
+                      /*strict_capacity_limit=*/false,
+                      /*high_pri_pool_ratio=*/0)));
+  cache_simulator->Access(access);
+  cache_simulator->Access(access);
+  ASSERT_EQ(2, cache_simulator->miss_ratio_stats().total_accesses());
+  // Both of them will be miss since we have a ghost cache.
+  ASSERT_EQ(100, cache_simulator->miss_ratio_stats().miss_ratio());
+}
+
+TEST_F(CacheSimulatorTest, HybridRowBlockCacheSimulator) {
+  uint64_t block_id = 100;
+  BlockCacheTraceRecord first_get = GenerateGetRecord(kGetId);
+  first_get.get_from_user_specified_snapshot = true;
+  BlockCacheTraceRecord second_get = GenerateGetRecord(kGetId + 1);
+  second_get.referenced_data_size = 0;
+  second_get.referenced_key_exist_in_block = false;
+  second_get.get_from_user_specified_snapshot = true;
+  BlockCacheTraceRecord third_get = GenerateGetRecord(kGetId + 2);
+  third_get.referenced_data_size = 0;
+  third_get.referenced_key_exist_in_block = false;
+  third_get.referenced_key = kRefKeyPrefix + "third_get";
+  // We didn't find the referenced key in the third get.
+  third_get.referenced_key_exist_in_block = false;
+  third_get.referenced_data_size = 0;
+  std::shared_ptr<Cache> sim_cache =
+      NewLRUCache(/*capacity=*/kCacheSize, /*num_shard_bits=*/1,
+                  /*strict_capacity_limit=*/false,
+                  /*high_pri_pool_ratio=*/0);
+  std::unique_ptr<HybridRowBlockCacheSimulator> cache_simulator(
+      new HybridRowBlockCacheSimulator(
+          nullptr, sim_cache, /*insert_blocks_row_kvpair_misses=*/true));
+  // The first get request accesses 10 blocks. We should only report 10 accesses
+  // and 100% miss.
+  for (uint32_t i = 0; i < 10; i++) {
+    first_get.block_key = kBlockKeyPrefix + std::to_string(block_id);
+    cache_simulator->Access(first_get);
+    block_id++;
+  }
+
+  ASSERT_EQ(10, cache_simulator->miss_ratio_stats().total_accesses());
+  ASSERT_EQ(100, cache_simulator->miss_ratio_stats().miss_ratio());
+  ASSERT_EQ(10, cache_simulator->miss_ratio_stats().user_accesses());
+  ASSERT_EQ(100, cache_simulator->miss_ratio_stats().user_miss_ratio());
+  auto handle =
+      sim_cache->Lookup(std::to_string(first_get.sst_fd_number) + "_" +
+                        ExtractUserKey(first_get.referenced_key).ToString());
+  ASSERT_NE(nullptr, handle);
+  sim_cache->Release(handle);
+  for (uint32_t i = 100; i < block_id; i++) {
+    handle = sim_cache->Lookup(kBlockKeyPrefix + std::to_string(i));
+    ASSERT_NE(nullptr, handle);
+    sim_cache->Release(handle);
+  }
+
+  // The second get request accesses the same key. We should report 15
+  // access and 66% miss, 10 misses with 15 accesses.
+  // We do not consider these 5 block lookups as misses since the row hits the
+  // cache.
+  for (uint32_t i = 0; i < 5; i++) {
+    second_get.block_key = kBlockKeyPrefix + std::to_string(block_id);
+    cache_simulator->Access(second_get);
+    block_id++;
+  }
+  ASSERT_EQ(15, cache_simulator->miss_ratio_stats().total_accesses());
+  ASSERT_EQ(66, static_cast<uint64_t>(
+                    cache_simulator->miss_ratio_stats().miss_ratio()));
+  ASSERT_EQ(15, cache_simulator->miss_ratio_stats().user_accesses());
+  ASSERT_EQ(66, static_cast<uint64_t>(
+                    cache_simulator->miss_ratio_stats().user_miss_ratio()));
+  handle =
+      sim_cache->Lookup(std::to_string(second_get.sst_fd_number) + "_" +
+                        ExtractUserKey(second_get.referenced_key).ToString());
+  ASSERT_NE(nullptr, handle);
+  sim_cache->Release(handle);
+  for (uint32_t i = 100; i < block_id; i++) {
+    handle = sim_cache->Lookup(kBlockKeyPrefix + std::to_string(i));
+    if (i < 110) {
+      ASSERT_NE(nullptr, handle) << i;
+      sim_cache->Release(handle);
+    } else {
+      ASSERT_EQ(nullptr, handle) << i;
+    }
+  }
+
+  // The third get on a different key and does not have a size.
+  // This key should not be inserted into the cache.
+  for (uint32_t i = 0; i < 5; i++) {
+    third_get.block_key = kBlockKeyPrefix + std::to_string(block_id);
+    cache_simulator->Access(third_get);
+    block_id++;
+  }
+  ASSERT_EQ(20, cache_simulator->miss_ratio_stats().total_accesses());
+  ASSERT_EQ(75, static_cast<uint64_t>(
+                    cache_simulator->miss_ratio_stats().miss_ratio()));
+  ASSERT_EQ(20, cache_simulator->miss_ratio_stats().user_accesses());
+  ASSERT_EQ(75, static_cast<uint64_t>(
+                    cache_simulator->miss_ratio_stats().user_miss_ratio()));
+  // Assert that the third key is not inserted into the cache.
+  handle = sim_cache->Lookup(std::to_string(third_get.sst_fd_number) + "_" +
+                             third_get.referenced_key);
+  ASSERT_EQ(nullptr, handle);
+  for (uint32_t i = 100; i < block_id; i++) {
+    if (i < 110 || i >= 115) {
+      handle = sim_cache->Lookup(kBlockKeyPrefix + std::to_string(i));
+      ASSERT_NE(nullptr, handle) << i;
+      sim_cache->Release(handle);
+    } else {
+      handle = sim_cache->Lookup(kBlockKeyPrefix + std::to_string(i));
+      ASSERT_EQ(nullptr, handle) << i;
+    }
+  }
+}
+
+TEST_F(CacheSimulatorTest, HybridRowBlockCacheSimulatorGetTest) {
+  BlockCacheTraceRecord get = GenerateGetRecord(kGetId);
+  get.block_size = 1;
+  get.referenced_data_size = 0;
+  get.access_timestamp = 0;
+  get.block_key = "1";
+  get.get_id = 1;
+  get.get_from_user_specified_snapshot = false;
+  get.referenced_key =
+      kRefKeyPrefix + std::to_string(1) + kRefKeySequenceNumber;
+  get.no_insert = false;
+  get.sst_fd_number = 0;
+  get.get_from_user_specified_snapshot = false;
+
+  LRUCacheOptions co;
+  co.capacity = 16;
+  co.num_shard_bits = 1;
+  co.strict_capacity_limit = false;
+  co.high_pri_pool_ratio = 0;
+  co.metadata_charge_policy = kDontChargeCacheMetadata;
+  std::shared_ptr<Cache> sim_cache = NewLRUCache(co);
+  std::unique_ptr<HybridRowBlockCacheSimulator> cache_simulator(
+      new HybridRowBlockCacheSimulator(
+          nullptr, sim_cache, /*insert_blocks_row_kvpair_misses=*/true));
+  // Expect a miss and does not insert the row key-value pair since it does not
+  // have size.
+  cache_simulator->Access(get);
+  AssertCache(sim_cache, cache_simulator->miss_ratio_stats(), 1, 1, 1, {"1"},
+              {});
+  get.access_timestamp += 1;
+  get.referenced_data_size = 1;
+  get.block_key = "2";
+  cache_simulator->Access(get);
+  AssertCache(sim_cache, cache_simulator->miss_ratio_stats(), 3, 2, 2,
+              {"1", "2"}, {"1"});
+  get.access_timestamp += 1;
+  get.block_key = "3";
+  // K1 should not inserted again.
+  cache_simulator->Access(get);
+  AssertCache(sim_cache, cache_simulator->miss_ratio_stats(), 4, 3, 3,
+              {"1", "2", "3"}, {"1"});
+
+  // A second get request referencing the same key.
+  get.access_timestamp += 1;
+  get.get_id = 2;
+  get.block_key = "4";
+  get.referenced_data_size = 0;
+  cache_simulator->Access(get);
+  AssertCache(sim_cache, cache_simulator->miss_ratio_stats(), 4, 4, 3,
+              {"1", "2", "3"}, {"1"});
+
+  // A third get request searches three files, three different keys.
+  // And the second key observes a hit.
+  get.access_timestamp += 1;
+  get.referenced_data_size = 1;
+  get.get_id = 3;
+  get.block_key = "3";
+  get.referenced_key = kRefKeyPrefix + "2" + kRefKeySequenceNumber;
+  // K2 should observe a miss. Block 3 observes a hit.
+  cache_simulator->Access(get);
+  AssertCache(sim_cache, cache_simulator->miss_ratio_stats(), 5, 5, 3,
+              {"1", "2", "3"}, {"1", "2"});
+
+  get.access_timestamp += 1;
+  get.referenced_data_size = 1;
+  get.get_id = 3;
+  get.block_key = "4";
+  get.referenced_data_size = 1;
+  get.referenced_key = kRefKeyPrefix + "1" + kRefKeySequenceNumber;
+  // K1 should observe a hit.
+  cache_simulator->Access(get);
+  AssertCache(sim_cache, cache_simulator->miss_ratio_stats(), 5, 6, 3,
+              {"1", "2", "3"}, {"1", "2"});
+
+  get.access_timestamp += 1;
+  get.referenced_data_size = 1;
+  get.get_id = 3;
+  get.block_key = "4";
+  get.referenced_data_size = 1;
+  get.referenced_key = kRefKeyPrefix + "3" + kRefKeySequenceNumber;
+  // K3 should observe a miss.
+  // However, as the get already complete, we should not access k3 any more.
+  cache_simulator->Access(get);
+  AssertCache(sim_cache, cache_simulator->miss_ratio_stats(), 5, 7, 3,
+              {"1", "2", "3"}, {"1", "2"});
+
+  // A fourth get request searches one file and two blocks. One row key.
+  get.access_timestamp += 1;
+  get.get_id = 4;
+  get.block_key = "5";
+  get.referenced_key = kRefKeyPrefix + "4" + kRefKeySequenceNumber;
+  get.referenced_data_size = 1;
+  cache_simulator->Access(get);
+  AssertCache(sim_cache, cache_simulator->miss_ratio_stats(), 7, 8, 4,
+              {"1", "2", "3", "5"}, {"1", "2", "4"});
+  for (auto const& key : {"1", "2", "4"}) {
+    auto handle = sim_cache->Lookup("0_" + kRefKeyPrefix + key);
+    ASSERT_NE(nullptr, handle);
+    sim_cache->Release(handle);
+  }
+
+  // A bunch of insertions which evict cached row keys.
+  for (uint32_t i = 6; i < 100; i++) {
+    get.access_timestamp += 1;
+    get.get_id = 0;
+    get.block_key = std::to_string(i);
+    cache_simulator->Access(get);
+  }
+
+  get.get_id = 4;
+  // A different block.
+  get.block_key = "100";
+  // Same row key and should not be inserted again.
+  get.referenced_key = kRefKeyPrefix + "4" + kRefKeySequenceNumber;
+  get.referenced_data_size = 1;
+  cache_simulator->Access(get);
+  AssertCache(sim_cache, cache_simulator->miss_ratio_stats(), 16, 103, 99, {},
+              {});
+  for (auto const& key : {"1", "2", "4"}) {
+    auto handle = sim_cache->Lookup("0_" + kRefKeyPrefix + key);
+    ASSERT_EQ(nullptr, handle);
+  }
+}
+
+TEST_F(CacheSimulatorTest, HybridRowBlockNoInsertCacheSimulator) {
+  uint64_t block_id = 100;
+  BlockCacheTraceRecord first_get = GenerateGetRecord(kGetId);
+  std::shared_ptr<Cache> sim_cache =
+      NewLRUCache(/*capacity=*/kCacheSize, /*num_shard_bits=*/1,
+                  /*strict_capacity_limit=*/false,
+                  /*high_pri_pool_ratio=*/0);
+  std::unique_ptr<HybridRowBlockCacheSimulator> cache_simulator(
+      new HybridRowBlockCacheSimulator(
+          nullptr, sim_cache, /*insert_blocks_row_kvpair_misses=*/false));
+  for (uint32_t i = 0; i < 9; i++) {
+    first_get.block_key = kBlockKeyPrefix + std::to_string(block_id);
+    cache_simulator->Access(first_get);
+    block_id++;
+  }
+  auto handle =
+      sim_cache->Lookup(std::to_string(first_get.sst_fd_number) + "_" +
+                        ExtractUserKey(first_get.referenced_key).ToString());
+  ASSERT_NE(nullptr, handle);
+  sim_cache->Release(handle);
+  // All blocks are missing from the cache since insert_blocks_row_kvpair_misses
+  // is set to false.
+  for (uint32_t i = 100; i < block_id; i++) {
+    handle = sim_cache->Lookup(kBlockKeyPrefix + std::to_string(i));
+    ASSERT_EQ(nullptr, handle);
+  }
+}
+
+TEST_F(CacheSimulatorTest, GhostHybridRowBlockCacheSimulator) {
+  std::unique_ptr<GhostCache> ghost_cache(new GhostCache(
+      NewLRUCache(/*capacity=*/kGhostCacheSize, /*num_shard_bits=*/1,
+                  /*strict_capacity_limit=*/false,
+                  /*high_pri_pool_ratio=*/0)));
+  const BlockCacheTraceRecord& first_get = GenerateGetRecord(kGetId);
+  const BlockCacheTraceRecord& second_get = GenerateGetRecord(kGetId + 1);
+  const BlockCacheTraceRecord& third_get = GenerateGetRecord(kGetId + 2);
+  std::unique_ptr<HybridRowBlockCacheSimulator> cache_simulator(
+      new HybridRowBlockCacheSimulator(
+          std::move(ghost_cache),
+          NewLRUCache(/*capacity=*/kCacheSize, /*num_shard_bits=*/1,
+                      /*strict_capacity_limit=*/false,
+                      /*high_pri_pool_ratio=*/0),
+          /*insert_blocks_row_kvpair_misses=*/false));
+  // Two get requests access the same key.
+  cache_simulator->Access(first_get);
+  cache_simulator->Access(second_get);
+  ASSERT_EQ(2, cache_simulator->miss_ratio_stats().total_accesses());
+  ASSERT_EQ(100, cache_simulator->miss_ratio_stats().miss_ratio());
+  ASSERT_EQ(2, cache_simulator->miss_ratio_stats().user_accesses());
+  ASSERT_EQ(100, cache_simulator->miss_ratio_stats().user_miss_ratio());
+  // We insert the key-value pair upon the second get request. A third get
+  // request should observe a hit.
+  for (uint32_t i = 0; i < 10; i++) {
+    cache_simulator->Access(third_get);
+  }
+  ASSERT_EQ(12, cache_simulator->miss_ratio_stats().total_accesses());
+  ASSERT_EQ(16, static_cast<uint64_t>(
+                    cache_simulator->miss_ratio_stats().miss_ratio()));
+  ASSERT_EQ(12, cache_simulator->miss_ratio_stats().user_accesses());
+  ASSERT_EQ(16, static_cast<uint64_t>(
+                    cache_simulator->miss_ratio_stats().user_miss_ratio()));
+}
+
+}  // namespace ROCKSDB_NAMESPACE
+
+int main(int argc, char** argv) {
+  ROCKSDB_NAMESPACE::port::InstallStackTraceHandler();
+  ::testing::InitGoogleTest(&argc, argv);
+  return RUN_ALL_TESTS();
+}
diff --git a/src/rocksdb/utilities/simulator_cache/sim_cache.cc b/src/rocksdb/utilities/simulator_cache/sim_cache.cc
new file mode 100644
index 000000000..a883b52e7
--- /dev/null
+++ b/src/rocksdb/utilities/simulator_cache/sim_cache.cc
@@ -0,0 +1,364 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#include "rocksdb/utilities/sim_cache.h"
+
+#include <atomic>
+#include <iomanip>
+
+#include "file/writable_file_writer.h"
+#include "monitoring/statistics.h"
+#include "port/port.h"
+#include "rocksdb/env.h"
+#include "rocksdb/file_system.h"
+#include "util/mutexlock.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+namespace {
+
+class CacheActivityLogger {
+ public:
+  CacheActivityLogger()
+      : activity_logging_enabled_(false), max_logging_size_(0) {}
+
+  ~CacheActivityLogger() {
+    MutexLock l(&mutex_);
+
+    StopLoggingInternal();
+    bg_status_.PermitUncheckedError();
+  }
+
+  Status StartLogging(const std::string& activity_log_file, Env* env,
+                      uint64_t max_logging_size = 0) {
+    assert(activity_log_file != "");
+    assert(env != nullptr);
+
+    Status status;
+    FileOptions file_opts;
+
+    MutexLock l(&mutex_);
+
+    // Stop existing logging if any
+    StopLoggingInternal();
+
+    // Open log file
+    status = WritableFileWriter::Create(env->GetFileSystem(), activity_log_file,
+                                        file_opts, &file_writer_, nullptr);
+    if (!status.ok()) {
+      return status;
+    }
+
+    max_logging_size_ = max_logging_size;
+    activity_logging_enabled_.store(true);
+
+    return status;
+  }
+
+  void StopLogging() {
+    MutexLock l(&mutex_);
+
+    StopLoggingInternal();
+  }
+
+  void ReportLookup(const Slice& key) {
+    if (activity_logging_enabled_.load() == false) {
+      return;
+    }
+
+    std::ostringstream oss;
+    // line format: "LOOKUP - <KEY>"
+    oss << "LOOKUP - " << key.ToString(true) << std::endl;
+
+    MutexLock l(&mutex_);
+    Status s = file_writer_->Append(oss.str());
+    if (!s.ok() && bg_status_.ok()) {
+      bg_status_ = s;
+    }
+    if (MaxLoggingSizeReached() || !bg_status_.ok()) {
+      // Stop logging if we have reached the max file size or
+      // encountered an error
+      StopLoggingInternal();
+    }
+  }
+
+  void ReportAdd(const Slice& key, size_t size) {
+    if (activity_logging_enabled_.load() == false) {
+      return;
+    }
+
+    std::ostringstream oss;
+    // line format: "ADD - <KEY> - <KEY-SIZE>"
+    oss << "ADD - " << key.ToString(true) << " - " << size << std::endl;
+    MutexLock l(&mutex_);
+    Status s = file_writer_->Append(oss.str());
+    if (!s.ok() && bg_status_.ok()) {
+      bg_status_ = s;
+    }
+
+    if (MaxLoggingSizeReached() || !bg_status_.ok()) {
+      // Stop logging if we have reached the max file size or
+      // encountered an error
+      StopLoggingInternal();
+    }
+  }
+
+  Status& bg_status() {
+    MutexLock l(&mutex_);
+    return bg_status_;
+  }
+
+ private:
+  bool MaxLoggingSizeReached() {
+    mutex_.AssertHeld();
+
+    return (max_logging_size_ > 0 &&
+            file_writer_->GetFileSize() >= max_logging_size_);
+  }
+
+  void StopLoggingInternal() {
+    mutex_.AssertHeld();
+
+    if (!activity_logging_enabled_) {
+      return;
+    }
+
+    activity_logging_enabled_.store(false);
+    Status s = file_writer_->Close();
+    if (!s.ok() && bg_status_.ok()) {
+      bg_status_ = s;
+    }
+  }
+
+  // Mutex to sync writes to file_writer, and all following
+  // class data members
+  port::Mutex mutex_;
+  // Indicates if logging is currently enabled
+  // atomic to allow reads without mutex
+  std::atomic<bool> activity_logging_enabled_;
+  // When reached, we will stop logging and close the file
+  // Value of 0 means unlimited
+  uint64_t max_logging_size_;
+  std::unique_ptr<WritableFileWriter> file_writer_;
+  Status bg_status_;
+};
+
+// SimCacheImpl definition
+class SimCacheImpl : public SimCache {
+ public:
+  // capacity for real cache (ShardedLRUCache)
+  // test_capacity for key only cache
+  SimCacheImpl(std::shared_ptr<Cache> sim_cache, std::shared_ptr<Cache> cache)
+      : cache_(cache),
+        key_only_cache_(sim_cache),
+        miss_times_(0),
+        hit_times_(0),
+        stats_(nullptr) {}
+
+  ~SimCacheImpl() override {}
+  void SetCapacity(size_t capacity) override { cache_->SetCapacity(capacity); }
+
+  void SetStrictCapacityLimit(bool strict_capacity_limit) override {
+    cache_->SetStrictCapacityLimit(strict_capacity_limit);
+  }
+
+  using Cache::Insert;
+  Status Insert(const Slice& key, void* value, size_t charge,
+                void (*deleter)(const Slice& key, void* value), Handle** handle,
+                Priority priority) override {
+    // The handle and value passed in are for real cache, so we pass nullptr
+    // to key_only_cache_ for both instead. Also, the deleter function pointer
+    // will be called by user to perform some external operation which should
+    // be applied only once. Thus key_only_cache accepts an empty function.
+    // *Lambda function without capture can be assgined to a function pointer
+    Handle* h = key_only_cache_->Lookup(key);
+    if (h == nullptr) {
+      // TODO: Check for error here?
+      auto s = key_only_cache_->Insert(
+          key, nullptr, charge, [](const Slice& /*k*/, void* /*v*/) {}, nullptr,
+          priority);
+      s.PermitUncheckedError();
+    } else {
+      key_only_cache_->Release(h);
+    }
+
+    cache_activity_logger_.ReportAdd(key, charge);
+    if (!cache_) {
+      return Status::OK();
+    }
+    return cache_->Insert(key, value, charge, deleter, handle, priority);
+  }
+
+  using Cache::Lookup;
+  Handle* Lookup(const Slice& key, Statistics* stats) override {
+    Handle* h = key_only_cache_->Lookup(key);
+    if (h != nullptr) {
+      key_only_cache_->Release(h);
+      inc_hit_counter();
+      RecordTick(stats, SIM_BLOCK_CACHE_HIT);
+    } else {
+      inc_miss_counter();
+      RecordTick(stats, SIM_BLOCK_CACHE_MISS);
+    }
+
+    cache_activity_logger_.ReportLookup(key);
+    if (!cache_) {
+      return nullptr;
+    }
+    return cache_->Lookup(key, stats);
+  }
+
+  bool Ref(Handle* handle) override { return cache_->Ref(handle); }
+
+  using Cache::Release;
+  bool Release(Handle* handle, bool erase_if_last_ref = false) override {
+    return cache_->Release(handle, erase_if_last_ref);
+  }
+
+  void Erase(const Slice& key) override {
+    cache_->Erase(key);
+    key_only_cache_->Erase(key);
+  }
+
+  void* Value(Handle* handle) override { return cache_->Value(handle); }
+
+  uint64_t NewId() override { return cache_->NewId(); }
+
+  size_t GetCapacity() const override { return cache_->GetCapacity(); }
+
+  bool HasStrictCapacityLimit() const override {
+    return cache_->HasStrictCapacityLimit();
+  }
+
+  size_t GetUsage() const override { return cache_->GetUsage(); }
+
+  size_t GetUsage(Handle* handle) const override {
+    return cache_->GetUsage(handle);
+  }
+
+  size_t GetCharge(Handle* handle) const override {
+    return cache_->GetCharge(handle);
+  }
+
+  DeleterFn GetDeleter(Handle* handle) const override {
+    return cache_->GetDeleter(handle);
+  }
+
+  size_t GetPinnedUsage() const override { return cache_->GetPinnedUsage(); }
+
+  void DisownData() override {
+    cache_->DisownData();
+    key_only_cache_->DisownData();
+  }
+
+  void ApplyToAllCacheEntries(void (*callback)(void*, size_t),
+                              bool thread_safe) override {
+    // only apply to _cache since key_only_cache doesn't hold value
+    cache_->ApplyToAllCacheEntries(callback, thread_safe);
+  }
+
+  void ApplyToAllEntries(
+      const std::function<void(const Slice& key, void* value, size_t charge,
+                               DeleterFn deleter)>& callback,
+      const ApplyToAllEntriesOptions& opts) override {
+    cache_->ApplyToAllEntries(callback, opts);
+  }
+
+  void EraseUnRefEntries() override {
+    cache_->EraseUnRefEntries();
+    key_only_cache_->EraseUnRefEntries();
+  }
+
+  size_t GetSimCapacity() const override {
+    return key_only_cache_->GetCapacity();
+  }
+  size_t GetSimUsage() const override { return key_only_cache_->GetUsage(); }
+  void SetSimCapacity(size_t capacity) override {
+    key_only_cache_->SetCapacity(capacity);
+  }
+
+  uint64_t get_miss_counter() const override {
+    return miss_times_.load(std::memory_order_relaxed);
+  }
+
+  uint64_t get_hit_counter() const override {
+    return hit_times_.load(std::memory_order_relaxed);
+  }
+
+  void reset_counter() override {
+    miss_times_.store(0, std::memory_order_relaxed);
+    hit_times_.store(0, std::memory_order_relaxed);
+    SetTickerCount(stats_, SIM_BLOCK_CACHE_HIT, 0);
+    SetTickerCount(stats_, SIM_BLOCK_CACHE_MISS, 0);
+  }
+
+  std::string ToString() const override {
+    std::ostringstream oss;
+    oss << "SimCache MISSes:  " << get_miss_counter() << std::endl;
+    oss << "SimCache HITs:    " << get_hit_counter() << std::endl;
+    auto lookups = get_miss_counter() + get_hit_counter();
+    oss << "SimCache HITRATE: " << std::fixed << std::setprecision(2)
+        << (lookups == 0 ? 0 : get_hit_counter() * 100.0f / lookups)
+        << std::endl;
+    return oss.str();
+  }
+
+  std::string GetPrintableOptions() const override {
+    std::ostringstream oss;
+    oss << "    cache_options:" << std::endl;
+    oss << cache_->GetPrintableOptions();
+    oss << "    sim_cache_options:" << std::endl;
+    oss << key_only_cache_->GetPrintableOptions();
+    return oss.str();
+  }
+
+  Status StartActivityLogging(const std::string& activity_log_file, Env* env,
+                              uint64_t max_logging_size = 0) override {
+    return cache_activity_logger_.StartLogging(activity_log_file, env,
+                                               max_logging_size);
+  }
+
+  void StopActivityLogging() override { cache_activity_logger_.StopLogging(); }
+
+  Status GetActivityLoggingStatus() override {
+    return cache_activity_logger_.bg_status();
+  }
+
+ private:
+  std::shared_ptr<Cache> cache_;
+  std::shared_ptr<Cache> key_only_cache_;
+  std::atomic<uint64_t> miss_times_;
+  std::atomic<uint64_t> hit_times_;
+  Statistics* stats_;
+  CacheActivityLogger cache_activity_logger_;
+
+  void inc_miss_counter() {
+    miss_times_.fetch_add(1, std::memory_order_relaxed);
+  }
+  void inc_hit_counter() { hit_times_.fetch_add(1, std::memory_order_relaxed); }
+};
+
+}  // end anonymous namespace
+
+// For instrumentation purpose, use NewSimCache instead
+std::shared_ptr<SimCache> NewSimCache(std::shared_ptr<Cache> cache,
+                                      size_t sim_capacity, int num_shard_bits) {
+  LRUCacheOptions co;
+  co.capacity = sim_capacity;
+  co.num_shard_bits = num_shard_bits;
+  co.metadata_charge_policy = kDontChargeCacheMetadata;
+  return NewSimCache(NewLRUCache(co), cache, num_shard_bits);
+}
+
+std::shared_ptr<SimCache> NewSimCache(std::shared_ptr<Cache> sim_cache,
+                                      std::shared_ptr<Cache> cache,
+                                      int num_shard_bits) {
+  if (num_shard_bits >= 20) {
+    return nullptr;  // the cache cannot be sharded into too many fine pieces
+  }
+  return std::make_shared<SimCacheImpl>(sim_cache, cache);
+}
+
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/utilities/simulator_cache/sim_cache_test.cc b/src/rocksdb/utilities/simulator_cache/sim_cache_test.cc
new file mode 100644
index 000000000..2e37cd347
--- /dev/null
+++ b/src/rocksdb/utilities/simulator_cache/sim_cache_test.cc
@@ -0,0 +1,226 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#include "rocksdb/utilities/sim_cache.h"
+
+#include <cstdlib>
+
+#include "db/db_test_util.h"
+#include "port/stack_trace.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+class SimCacheTest : public DBTestBase {
+ private:
+  size_t miss_count_ = 0;
+  size_t hit_count_ = 0;
+  size_t insert_count_ = 0;
+  size_t failure_count_ = 0;
+
+ public:
+  const size_t kNumBlocks = 5;
+  const size_t kValueSize = 1000;
+
+  SimCacheTest() : DBTestBase("sim_cache_test", /*env_do_fsync=*/true) {}
+
+  BlockBasedTableOptions GetTableOptions() {
+    BlockBasedTableOptions table_options;
+    // Set a small enough block size so that each key-value get its own block.
+    table_options.block_size = 1;
+    return table_options;
+  }
+
+  Options GetOptions(const BlockBasedTableOptions& table_options) {
+    Options options = CurrentOptions();
+    options.create_if_missing = true;
+    // options.compression = kNoCompression;
+    options.statistics = ROCKSDB_NAMESPACE::CreateDBStatistics();
+    options.table_factory.reset(NewBlockBasedTableFactory(table_options));
+    return options;
+  }
+
+  void InitTable(const Options& /*options*/) {
+    std::string value(kValueSize, 'a');
+    for (size_t i = 0; i < kNumBlocks * 2; i++) {
+      ASSERT_OK(Put(std::to_string(i), value.c_str()));
+    }
+  }
+
+  void RecordCacheCounters(const Options& options) {
+    miss_count_ = TestGetTickerCount(options, BLOCK_CACHE_MISS);
+    hit_count_ = TestGetTickerCount(options, BLOCK_CACHE_HIT);
+    insert_count_ = TestGetTickerCount(options, BLOCK_CACHE_ADD);
+    failure_count_ = TestGetTickerCount(options, BLOCK_CACHE_ADD_FAILURES);
+  }
+
+  void CheckCacheCounters(const Options& options, size_t expected_misses,
+                          size_t expected_hits, size_t expected_inserts,
+                          size_t expected_failures) {
+    size_t new_miss_count = TestGetTickerCount(options, BLOCK_CACHE_MISS);
+    size_t new_hit_count = TestGetTickerCount(options, BLOCK_CACHE_HIT);
+    size_t new_insert_count = TestGetTickerCount(options, BLOCK_CACHE_ADD);
+    size_t new_failure_count =
+        TestGetTickerCount(options, BLOCK_CACHE_ADD_FAILURES);
+    ASSERT_EQ(miss_count_ + expected_misses, new_miss_count);
+    ASSERT_EQ(hit_count_ + expected_hits, new_hit_count);
+    ASSERT_EQ(insert_count_ + expected_inserts, new_insert_count);
+    ASSERT_EQ(failure_count_ + expected_failures, new_failure_count);
+    miss_count_ = new_miss_count;
+    hit_count_ = new_hit_count;
+    insert_count_ = new_insert_count;
+    failure_count_ = new_failure_count;
+  }
+};
+
+TEST_F(SimCacheTest, SimCache) {
+  ReadOptions read_options;
+  auto table_options = GetTableOptions();
+  auto options = GetOptions(table_options);
+  InitTable(options);
+  LRUCacheOptions co;
+  co.capacity = 0;
+  co.num_shard_bits = 0;
+  co.strict_capacity_limit = false;
+  co.metadata_charge_policy = kDontChargeCacheMetadata;
+  std::shared_ptr<SimCache> simCache = NewSimCache(NewLRUCache(co), 20000, 0);
+  table_options.block_cache = simCache;
+  options.table_factory.reset(NewBlockBasedTableFactory(table_options));
+  Reopen(options);
+  RecordCacheCounters(options);
+  // due to cache entry stats collector
+  uint64_t base_misses = simCache->get_miss_counter();
+
+  std::vector<std::unique_ptr<Iterator>> iterators(kNumBlocks);
+  Iterator* iter = nullptr;
+
+  // Load blocks into cache.
+  for (size_t i = 0; i < kNumBlocks; i++) {
+    iter = db_->NewIterator(read_options);
+    iter->Seek(std::to_string(i));
+    ASSERT_OK(iter->status());
+    CheckCacheCounters(options, 1, 0, 1, 0);
+    iterators[i].reset(iter);
+  }
+  ASSERT_EQ(kNumBlocks, simCache->get_hit_counter() +
+                            simCache->get_miss_counter() - base_misses);
+  ASSERT_EQ(0, simCache->get_hit_counter());
+  size_t usage = simCache->GetUsage();
+  ASSERT_LT(0, usage);
+  ASSERT_EQ(usage, simCache->GetSimUsage());
+  simCache->SetCapacity(usage);
+  ASSERT_EQ(usage, simCache->GetPinnedUsage());
+
+  // Test with strict capacity limit.
+  simCache->SetStrictCapacityLimit(true);
+  iter = db_->NewIterator(read_options);
+  iter->Seek(std::to_string(kNumBlocks * 2 - 1));
+  ASSERT_TRUE(iter->status().IsMemoryLimit());
+  CheckCacheCounters(options, 1, 0, 0, 1);
+  delete iter;
+  iter = nullptr;
+
+  // Release iterators and access cache again.
+  for (size_t i = 0; i < kNumBlocks; i++) {
+    iterators[i].reset();
+    CheckCacheCounters(options, 0, 0, 0, 0);
+  }
+  // Add kNumBlocks again
+  for (size_t i = 0; i < kNumBlocks; i++) {
+    std::unique_ptr<Iterator> it(db_->NewIterator(read_options));
+    it->Seek(std::to_string(i));
+    ASSERT_OK(it->status());
+    CheckCacheCounters(options, 0, 1, 0, 0);
+  }
+  ASSERT_EQ(5, simCache->get_hit_counter());
+  for (size_t i = kNumBlocks; i < kNumBlocks * 2; i++) {
+    std::unique_ptr<Iterator> it(db_->NewIterator(read_options));
+    it->Seek(std::to_string(i));
+    ASSERT_OK(it->status());
+    CheckCacheCounters(options, 1, 0, 1, 0);
+  }
+  ASSERT_EQ(0, simCache->GetPinnedUsage());
+  ASSERT_EQ(3 * kNumBlocks + 1, simCache->get_hit_counter() +
+                                    simCache->get_miss_counter() - base_misses);
+  ASSERT_EQ(6, simCache->get_hit_counter());
+}
+
+TEST_F(SimCacheTest, SimCacheLogging) {
+  auto table_options = GetTableOptions();
+  auto options = GetOptions(table_options);
+  options.disable_auto_compactions = true;
+  LRUCacheOptions co;
+  co.capacity = 1024 * 1024;
+  co.metadata_charge_policy = kDontChargeCacheMetadata;
+  std::shared_ptr<SimCache> sim_cache = NewSimCache(NewLRUCache(co), 20000, 0);
+  table_options.block_cache = sim_cache;
+  options.table_factory.reset(NewBlockBasedTableFactory(table_options));
+  Reopen(options);
+
+  int num_block_entries = 20;
+  for (int i = 0; i < num_block_entries; i++) {
+    ASSERT_OK(Put(Key(i), "val"));
+    ASSERT_OK(Flush());
+  }
+
+  std::string log_file = test::PerThreadDBPath(env_, "cache_log.txt");
+  ASSERT_OK(sim_cache->StartActivityLogging(log_file, env_));
+  for (int i = 0; i < num_block_entries; i++) {
+    ASSERT_EQ(Get(Key(i)), "val");
+  }
+  for (int i = 0; i < num_block_entries; i++) {
+    ASSERT_EQ(Get(Key(i)), "val");
+  }
+  sim_cache->StopActivityLogging();
+  ASSERT_OK(sim_cache->GetActivityLoggingStatus());
+
+  std::string file_contents = "";
+  ASSERT_OK(ReadFileToString(env_, log_file, &file_contents));
+  std::istringstream contents(file_contents);
+
+  int lookup_num = 0;
+  int add_num = 0;
+
+  std::string line;
+  // count number of lookups and additions
+  while (std::getline(contents, line)) {
+    // check if the line starts with LOOKUP or ADD
+    if (line.rfind("LOOKUP -", 0) == 0) {
+      ++lookup_num;
+    }
+    if (line.rfind("ADD -", 0) == 0) {
+      ++add_num;
+    }
+  }
+
+  // We asked for every block twice
+  ASSERT_EQ(lookup_num, num_block_entries * 2);
+
+  // We added every block only once, since the cache can hold all blocks
+  ASSERT_EQ(add_num, num_block_entries);
+
+  // Log things again but stop logging automatically after reaching 512 bytes
+  int max_size = 512;
+  ASSERT_OK(sim_cache->StartActivityLogging(log_file, env_, max_size));
+  for (int it = 0; it < 10; it++) {
+    for (int i = 0; i < num_block_entries; i++) {
+      ASSERT_EQ(Get(Key(i)), "val");
+    }
+  }
+  ASSERT_OK(sim_cache->GetActivityLoggingStatus());
+
+  uint64_t fsize = 0;
+  ASSERT_OK(env_->GetFileSize(log_file, &fsize));
+  // error margin of 100 bytes
+  ASSERT_LT(fsize, max_size + 100);
+  ASSERT_GT(fsize, max_size - 100);
+}
+
+}  // namespace ROCKSDB_NAMESPACE
+
+int main(int argc, char** argv) {
+  ROCKSDB_NAMESPACE::port::InstallStackTraceHandler();
+  ::testing::InitGoogleTest(&argc, argv);
+  return RUN_ALL_TESTS();
+}
diff --git a/src/rocksdb/utilities/table_properties_collectors/compact_on_deletion_collector.cc b/src/rocksdb/utilities/table_properties_collectors/compact_on_deletion_collector.cc
new file mode 100644
index 000000000..16f33934d
--- /dev/null
+++ b/src/rocksdb/utilities/table_properties_collectors/compact_on_deletion_collector.cc
@@ -0,0 +1,227 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#include "utilities/table_properties_collectors/compact_on_deletion_collector.h"
+
+#include <memory>
+#include <sstream>
+
+#include "rocksdb/utilities/customizable_util.h"
+#include "rocksdb/utilities/object_registry.h"
+#include "rocksdb/utilities/options_type.h"
+#include "rocksdb/utilities/table_properties_collectors.h"
+#include "util/string_util.h"
+
+namespace ROCKSDB_NAMESPACE {
+#ifndef ROCKSDB_LITE
+
+CompactOnDeletionCollector::CompactOnDeletionCollector(
+    size_t sliding_window_size, size_t deletion_trigger, double deletion_ratio)
+    : bucket_size_((sliding_window_size + kNumBuckets - 1) / kNumBuckets),
+      current_bucket_(0),
+      num_keys_in_current_bucket_(0),
+      num_deletions_in_observation_window_(0),
+      deletion_trigger_(deletion_trigger),
+      deletion_ratio_(deletion_ratio),
+      deletion_ratio_enabled_(deletion_ratio > 0 && deletion_ratio <= 1),
+      need_compaction_(false),
+      finished_(false) {
+  memset(num_deletions_in_buckets_, 0, sizeof(size_t) * kNumBuckets);
+}
+
+// AddUserKey() will be called when a new key/value pair is inserted into the
+// table.
+// @params key    the user key that is inserted into the table.
+// @params value  the value that is inserted into the table.
+// @params file_size  file size up to now
+Status CompactOnDeletionCollector::AddUserKey(const Slice& /*key*/,
+                                              const Slice& /*value*/,
+                                              EntryType type,
+                                              SequenceNumber /*seq*/,
+                                              uint64_t /*file_size*/) {
+  assert(!finished_);
+  if (!bucket_size_ && !deletion_ratio_enabled_) {
+    // This collector is effectively disabled
+    return Status::OK();
+  }
+
+  if (need_compaction_) {
+    // If the output file already needs to be compacted, skip the check.
+    return Status::OK();
+  }
+
+  if (deletion_ratio_enabled_) {
+    total_entries_++;
+    if (type == kEntryDelete) {
+      deletion_entries_++;
+    }
+  }
+
+  if (bucket_size_) {
+    if (num_keys_in_current_bucket_ == bucket_size_) {
+      // When the current bucket is full, advance the cursor of the
+      // ring buffer to the next bucket.
+      current_bucket_ = (current_bucket_ + 1) % kNumBuckets;
+
+      // Update the current count of observed deletion keys by excluding
+      // the number of deletion keys in the oldest bucket in the
+      // observation window.
+      assert(num_deletions_in_observation_window_ >=
+             num_deletions_in_buckets_[current_bucket_]);
+      num_deletions_in_observation_window_ -=
+          num_deletions_in_buckets_[current_bucket_];
+      num_deletions_in_buckets_[current_bucket_] = 0;
+      num_keys_in_current_bucket_ = 0;
+    }
+
+    num_keys_in_current_bucket_++;
+    if (type == kEntryDelete) {
+      num_deletions_in_observation_window_++;
+      num_deletions_in_buckets_[current_bucket_]++;
+      if (num_deletions_in_observation_window_ >= deletion_trigger_) {
+        need_compaction_ = true;
+      }
+    }
+  }
+
+  return Status::OK();
+}
+
+Status CompactOnDeletionCollector::Finish(
+    UserCollectedProperties* /*properties*/) {
+  if (!need_compaction_ && deletion_ratio_enabled_ && total_entries_ > 0) {
+    double ratio = static_cast<double>(deletion_entries_) / total_entries_;
+    need_compaction_ = ratio >= deletion_ratio_;
+  }
+  finished_ = true;
+  return Status::OK();
+}
+static std::unordered_map<std::string, OptionTypeInfo>
+    on_deletion_collector_type_info = {
+#ifndef ROCKSDB_LITE
+        {"window_size",
+         {0, OptionType::kUnknown, OptionVerificationType::kNormal,
+          OptionTypeFlags::kCompareNever | OptionTypeFlags::kMutable,
+          [](const ConfigOptions&, const std::string&, const std::string& value,
+             void* addr) {
+            auto* factory =
+                static_cast<CompactOnDeletionCollectorFactory*>(addr);
+            factory->SetWindowSize(ParseSizeT(value));
+            return Status::OK();
+          },
+          [](const ConfigOptions&, const std::string&, const void* addr,
+             std::string* value) {
+            const auto* factory =
+                static_cast<const CompactOnDeletionCollectorFactory*>(addr);
+            *value = std::to_string(factory->GetWindowSize());
+            return Status::OK();
+          },
+          nullptr}},
+        {"deletion_trigger",
+         {0, OptionType::kUnknown, OptionVerificationType::kNormal,
+          OptionTypeFlags::kCompareNever | OptionTypeFlags::kMutable,
+          [](const ConfigOptions&, const std::string&, const std::string& value,
+             void* addr) {
+            auto* factory =
+                static_cast<CompactOnDeletionCollectorFactory*>(addr);
+            factory->SetDeletionTrigger(ParseSizeT(value));
+            return Status::OK();
+          },
+          [](const ConfigOptions&, const std::string&, const void* addr,
+             std::string* value) {
+            const auto* factory =
+                static_cast<const CompactOnDeletionCollectorFactory*>(addr);
+            *value = std::to_string(factory->GetDeletionTrigger());
+            return Status::OK();
+          },
+          nullptr}},
+        {"deletion_ratio",
+         {0, OptionType::kUnknown, OptionVerificationType::kNormal,
+          OptionTypeFlags::kCompareNever | OptionTypeFlags::kMutable,
+          [](const ConfigOptions&, const std::string&, const std::string& value,
+             void* addr) {
+            auto* factory =
+                static_cast<CompactOnDeletionCollectorFactory*>(addr);
+            factory->SetDeletionRatio(ParseDouble(value));
+            return Status::OK();
+          },
+          [](const ConfigOptions&, const std::string&, const void* addr,
+             std::string* value) {
+            const auto* factory =
+                static_cast<const CompactOnDeletionCollectorFactory*>(addr);
+            *value = std::to_string(factory->GetDeletionRatio());
+            return Status::OK();
+          },
+          nullptr}},
+
+#endif  // ROCKSDB_LITE
+};
+
+CompactOnDeletionCollectorFactory::CompactOnDeletionCollectorFactory(
+    size_t sliding_window_size, size_t deletion_trigger, double deletion_ratio)
+    : sliding_window_size_(sliding_window_size),
+      deletion_trigger_(deletion_trigger),
+      deletion_ratio_(deletion_ratio) {
+  RegisterOptions("", this, &on_deletion_collector_type_info);
+}
+
+TablePropertiesCollector*
+CompactOnDeletionCollectorFactory::CreateTablePropertiesCollector(
+    TablePropertiesCollectorFactory::Context /*context*/) {
+  return new CompactOnDeletionCollector(sliding_window_size_.load(),
+                                        deletion_trigger_.load(),
+                                        deletion_ratio_.load());
+}
+
+std::string CompactOnDeletionCollectorFactory::ToString() const {
+  std::ostringstream cfg;
+  cfg << Name() << " (Sliding window size = " << sliding_window_size_.load()
+      << " Deletion trigger = " << deletion_trigger_.load()
+      << " Deletion ratio = " << deletion_ratio_.load() << ')';
+  return cfg.str();
+}
+
+std::shared_ptr<CompactOnDeletionCollectorFactory>
+NewCompactOnDeletionCollectorFactory(size_t sliding_window_size,
+                                     size_t deletion_trigger,
+                                     double deletion_ratio) {
+  return std::shared_ptr<CompactOnDeletionCollectorFactory>(
+      new CompactOnDeletionCollectorFactory(sliding_window_size,
+                                            deletion_trigger, deletion_ratio));
+}
+namespace {
+static int RegisterTablePropertiesCollectorFactories(
+    ObjectLibrary& library, const std::string& /*arg*/) {
+  library.AddFactory<TablePropertiesCollectorFactory>(
+      CompactOnDeletionCollectorFactory::kClassName(),
+      [](const std::string& /*uri*/,
+         std::unique_ptr<TablePropertiesCollectorFactory>* guard,
+         std::string* /* errmsg */) {
+        // By default, create a CompactionOnDeletionCollector that is disabled.
+        // Users will need to provide configuration parameters or call the
+        // corresponding Setter to enable the factory.
+        guard->reset(new CompactOnDeletionCollectorFactory(0, 0, 0));
+        return guard->get();
+      });
+  return 1;
+}
+}  // namespace
+#endif  // !ROCKSDB_LITE
+
+Status TablePropertiesCollectorFactory::CreateFromString(
+    const ConfigOptions& options, const std::string& value,
+    std::shared_ptr<TablePropertiesCollectorFactory>* result) {
+#ifndef ROCKSDB_LITE
+  static std::once_flag once;
+  std::call_once(once, [&]() {
+    RegisterTablePropertiesCollectorFactories(*(ObjectLibrary::Default().get()),
+                                              "");
+  });
+#endif  // ROCKSDB_LITE
+  return LoadSharedObject<TablePropertiesCollectorFactory>(options, value,
+                                                           nullptr, result);
+}
+
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/utilities/table_properties_collectors/compact_on_deletion_collector.h b/src/rocksdb/utilities/table_properties_collectors/compact_on_deletion_collector.h
new file mode 100644
index 000000000..2f7dc4f1b
--- /dev/null
+++ b/src/rocksdb/utilities/table_properties_collectors/compact_on_deletion_collector.h
@@ -0,0 +1,70 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#pragma once
+
+#ifndef ROCKSDB_LITE
+#include "rocksdb/utilities/table_properties_collectors.h"
+namespace ROCKSDB_NAMESPACE {
+
+class CompactOnDeletionCollector : public TablePropertiesCollector {
+ public:
+  CompactOnDeletionCollector(size_t sliding_window_size,
+                             size_t deletion_trigger, double deletion_raatio);
+
+  // AddUserKey() will be called when a new key/value pair is inserted into the
+  // table.
+  // @params key    the user key that is inserted into the table.
+  // @params value  the value that is inserted into the table.
+  // @params file_size  file size up to now
+  virtual Status AddUserKey(const Slice& key, const Slice& value,
+                            EntryType type, SequenceNumber seq,
+                            uint64_t file_size) override;
+
+  // Finish() will be called when a table has already been built and is ready
+  // for writing the properties block.
+  // @params properties  User will add their collected statistics to
+  // `properties`.
+  virtual Status Finish(UserCollectedProperties* /*properties*/) override;
+
+  // Return the human-readable properties, where the key is property name and
+  // the value is the human-readable form of value.
+  virtual UserCollectedProperties GetReadableProperties() const override {
+    return UserCollectedProperties();
+  }
+
+  // The name of the properties collector can be used for debugging purpose.
+  virtual const char* Name() const override {
+    return "CompactOnDeletionCollector";
+  }
+
+  // EXPERIMENTAL Return whether the output file should be further compacted
+  virtual bool NeedCompact() const override { return need_compaction_; }
+
+  static const int kNumBuckets = 128;
+
+ private:
+  void Reset();
+
+  // A ring buffer that used to count the number of deletion entries for every
+  // "bucket_size_" keys.
+  size_t num_deletions_in_buckets_[kNumBuckets];
+  // the number of keys in a bucket
+  size_t bucket_size_;
+
+  size_t current_bucket_;
+  size_t num_keys_in_current_bucket_;
+  size_t num_deletions_in_observation_window_;
+  size_t deletion_trigger_;
+  const double deletion_ratio_;
+  const bool deletion_ratio_enabled_;
+  size_t total_entries_ = 0;
+  size_t deletion_entries_ = 0;
+  // true if the current SST file needs to be compacted.
+  bool need_compaction_;
+  bool finished_;
+};
+}  // namespace ROCKSDB_NAMESPACE
+#endif  // !ROCKSDB_LITE
diff --git a/src/rocksdb/utilities/table_properties_collectors/compact_on_deletion_collector_test.cc b/src/rocksdb/utilities/table_properties_collectors/compact_on_deletion_collector_test.cc
new file mode 100644
index 000000000..88aeb8d5c
--- /dev/null
+++ b/src/rocksdb/utilities/table_properties_collectors/compact_on_deletion_collector_test.cc
@@ -0,0 +1,245 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+//
+// Copyright (c) 2011 The LevelDB Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file. See the AUTHORS file for names of contributors.
+
+#include <stdio.h>
+
+#ifndef ROCKSDB_LITE
+#include <algorithm>
+#include <cmath>
+#include <vector>
+
+#include "port/stack_trace.h"
+#include "rocksdb/table.h"
+#include "rocksdb/table_properties.h"
+#include "rocksdb/utilities/table_properties_collectors.h"
+#include "test_util/testharness.h"
+#include "util/random.h"
+#include "utilities/table_properties_collectors/compact_on_deletion_collector.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+TEST(CompactOnDeletionCollector, DeletionRatio) {
+  TablePropertiesCollectorFactory::Context context;
+  context.column_family_id =
+      TablePropertiesCollectorFactory::Context::kUnknownColumnFamily;
+  const size_t kTotalEntries = 100;
+
+  {
+    // Disable deletion ratio.
+    for (double deletion_ratio : {-1.5, -1.0, 0.0, 1.5, 2.0}) {
+      auto factory = NewCompactOnDeletionCollectorFactory(0, 0, deletion_ratio);
+      std::unique_ptr<TablePropertiesCollector> collector(
+          factory->CreateTablePropertiesCollector(context));
+      for (size_t i = 0; i < kTotalEntries; i++) {
+        // All entries are deletion entries.
+        ASSERT_OK(
+            collector->AddUserKey("hello", "rocksdb", kEntryDelete, 0, 0));
+        ASSERT_FALSE(collector->NeedCompact());
+      }
+      ASSERT_OK(collector->Finish(nullptr));
+      ASSERT_FALSE(collector->NeedCompact());
+    }
+  }
+
+  {
+    for (double deletion_ratio : {0.3, 0.5, 0.8, 1.0}) {
+      auto factory = NewCompactOnDeletionCollectorFactory(0, 0, deletion_ratio);
+      const size_t deletion_entries_trigger =
+          static_cast<size_t>(deletion_ratio * kTotalEntries);
+      for (int delta : {-1, 0, 1}) {
+        // Actual deletion entry ratio <, =, > deletion_ratio
+        size_t actual_deletion_entries = deletion_entries_trigger + delta;
+        std::unique_ptr<TablePropertiesCollector> collector(
+            factory->CreateTablePropertiesCollector(context));
+        for (size_t i = 0; i < kTotalEntries; i++) {
+          if (i < actual_deletion_entries) {
+            ASSERT_OK(
+                collector->AddUserKey("hello", "rocksdb", kEntryDelete, 0, 0));
+          } else {
+            ASSERT_OK(
+                collector->AddUserKey("hello", "rocksdb", kEntryPut, 0, 0));
+          }
+          ASSERT_FALSE(collector->NeedCompact());
+        }
+        ASSERT_OK(collector->Finish(nullptr));
+        if (delta >= 0) {
+          // >= deletion_ratio
+          ASSERT_TRUE(collector->NeedCompact());
+        } else {
+          ASSERT_FALSE(collector->NeedCompact());
+        }
+      }
+    }
+  }
+}
+
+TEST(CompactOnDeletionCollector, SlidingWindow) {
+  const int kWindowSizes[] = {1000, 10000, 10000, 127, 128,  129,
+                              255,  256,   257,   2,   10000};
+  const int kDeletionTriggers[] = {500, 9500, 4323, 47, 61, 128,
+                                   250, 250,  250,  2,  2};
+  TablePropertiesCollectorFactory::Context context;
+  context.column_family_id =
+      TablePropertiesCollectorFactory::Context::kUnknownColumnFamily;
+
+  std::vector<int> window_sizes;
+  std::vector<int> deletion_triggers;
+  // deterministic tests
+  for (int test = 0; test < 9; ++test) {
+    window_sizes.emplace_back(kWindowSizes[test]);
+    deletion_triggers.emplace_back(kDeletionTriggers[test]);
+  }
+
+  // randomize tests
+  Random rnd(301);
+  const int kMaxTestSize = 100000l;
+  for (int random_test = 0; random_test < 10; random_test++) {
+    int window_size = rnd.Uniform(kMaxTestSize) + 1;
+    int deletion_trigger = rnd.Uniform(window_size);
+    window_sizes.emplace_back(window_size);
+    deletion_triggers.emplace_back(deletion_trigger);
+  }
+
+  assert(window_sizes.size() == deletion_triggers.size());
+
+  for (size_t test = 0; test < window_sizes.size(); ++test) {
+    const int kBucketSize = 128;
+    const int kWindowSize = window_sizes[test];
+    const int kPaddedWindowSize =
+        kBucketSize * ((window_sizes[test] + kBucketSize - 1) / kBucketSize);
+    const int kNumDeletionTrigger = deletion_triggers[test];
+    const int kBias = (kNumDeletionTrigger + kBucketSize - 1) / kBucketSize;
+    // Simple test
+    {
+      auto factory = NewCompactOnDeletionCollectorFactory(kWindowSize,
+                                                          kNumDeletionTrigger);
+      const int kSample = 10;
+      for (int delete_rate = 0; delete_rate <= kSample; ++delete_rate) {
+        std::unique_ptr<TablePropertiesCollector> collector(
+            factory->CreateTablePropertiesCollector(context));
+        int deletions = 0;
+        for (int i = 0; i < kPaddedWindowSize; ++i) {
+          if (i % kSample < delete_rate) {
+            ASSERT_OK(
+                collector->AddUserKey("hello", "rocksdb", kEntryDelete, 0, 0));
+            deletions++;
+          } else {
+            ASSERT_OK(
+                collector->AddUserKey("hello", "rocksdb", kEntryPut, 0, 0));
+          }
+        }
+        if (collector->NeedCompact() != (deletions >= kNumDeletionTrigger) &&
+            std::abs(deletions - kNumDeletionTrigger) > kBias) {
+          fprintf(stderr,
+                  "[Error] collector->NeedCompact() != (%d >= %d)"
+                  " with kWindowSize = %d and kNumDeletionTrigger = %d\n",
+                  deletions, kNumDeletionTrigger, kWindowSize,
+                  kNumDeletionTrigger);
+          ASSERT_TRUE(false);
+        }
+        ASSERT_OK(collector->Finish(nullptr));
+      }
+    }
+
+    // Only one section of a file satisfies the compaction trigger
+    {
+      auto factory = NewCompactOnDeletionCollectorFactory(kWindowSize,
+                                                          kNumDeletionTrigger);
+      const int kSample = 10;
+      for (int delete_rate = 0; delete_rate <= kSample; ++delete_rate) {
+        std::unique_ptr<TablePropertiesCollector> collector(
+            factory->CreateTablePropertiesCollector(context));
+        int deletions = 0;
+        for (int section = 0; section < 5; ++section) {
+          int initial_entries = rnd.Uniform(kWindowSize) + kWindowSize;
+          for (int i = 0; i < initial_entries; ++i) {
+            ASSERT_OK(
+                collector->AddUserKey("hello", "rocksdb", kEntryPut, 0, 0));
+          }
+        }
+        for (int i = 0; i < kPaddedWindowSize; ++i) {
+          if (i % kSample < delete_rate) {
+            ASSERT_OK(
+                collector->AddUserKey("hello", "rocksdb", kEntryDelete, 0, 0));
+            deletions++;
+          } else {
+            ASSERT_OK(
+                collector->AddUserKey("hello", "rocksdb", kEntryPut, 0, 0));
+          }
+        }
+        for (int section = 0; section < 5; ++section) {
+          int ending_entries = rnd.Uniform(kWindowSize) + kWindowSize;
+          for (int i = 0; i < ending_entries; ++i) {
+            ASSERT_OK(
+                collector->AddUserKey("hello", "rocksdb", kEntryPut, 0, 0));
+          }
+        }
+        if (collector->NeedCompact() != (deletions >= kNumDeletionTrigger) &&
+            std::abs(deletions - kNumDeletionTrigger) > kBias) {
+          fprintf(stderr,
+                  "[Error] collector->NeedCompact() %d != (%d >= %d)"
+                  " with kWindowSize = %d, kNumDeletionTrigger = %d\n",
+                  collector->NeedCompact(), deletions, kNumDeletionTrigger,
+                  kWindowSize, kNumDeletionTrigger);
+          ASSERT_TRUE(false);
+        }
+        ASSERT_OK(collector->Finish(nullptr));
+      }
+    }
+
+    // TEST 3:  Issues a lots of deletes, but their density is not
+    // high enough to trigger compaction.
+    {
+      std::unique_ptr<TablePropertiesCollector> collector;
+      auto factory = NewCompactOnDeletionCollectorFactory(kWindowSize,
+                                                          kNumDeletionTrigger);
+      collector.reset(factory->CreateTablePropertiesCollector(context));
+      assert(collector->NeedCompact() == false);
+      // Insert "kNumDeletionTrigger * 0.95" deletions for every
+      // "kWindowSize" and verify compaction is not needed.
+      const int kDeletionsPerSection = kNumDeletionTrigger * 95 / 100;
+      if (kDeletionsPerSection >= 0) {
+        for (int section = 0; section < 200; ++section) {
+          for (int i = 0; i < kPaddedWindowSize; ++i) {
+            if (i < kDeletionsPerSection) {
+              ASSERT_OK(collector->AddUserKey("hello", "rocksdb", kEntryDelete,
+                                              0, 0));
+            } else {
+              ASSERT_OK(
+                  collector->AddUserKey("hello", "rocksdb", kEntryPut, 0, 0));
+            }
+          }
+        }
+        if (collector->NeedCompact() &&
+            std::abs(kDeletionsPerSection - kNumDeletionTrigger) > kBias) {
+          fprintf(stderr,
+                  "[Error] collector->NeedCompact() != false"
+                  " with kWindowSize = %d and kNumDeletionTrigger = %d\n",
+                  kWindowSize, kNumDeletionTrigger);
+          ASSERT_TRUE(false);
+        }
+        ASSERT_OK(collector->Finish(nullptr));
+      }
+    }
+  }
+}
+
+}  // namespace ROCKSDB_NAMESPACE
+
+int main(int argc, char** argv) {
+  ROCKSDB_NAMESPACE::port::InstallStackTraceHandler();
+  ::testing::InitGoogleTest(&argc, argv);
+  return RUN_ALL_TESTS();
+}
+#else
+int main(int /*argc*/, char** /*argv*/) {
+  fprintf(stderr, "SKIPPED as RocksDBLite does not include utilities.\n");
+  return 0;
+}
+#endif  // !ROCKSDB_LITE
diff --git a/src/rocksdb/utilities/trace/file_trace_reader_writer.cc b/src/rocksdb/utilities/trace/file_trace_reader_writer.cc
new file mode 100644
index 000000000..5886d3539
--- /dev/null
+++ b/src/rocksdb/utilities/trace/file_trace_reader_writer.cc
@@ -0,0 +1,133 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#include "utilities/trace/file_trace_reader_writer.h"
+
+#include "env/composite_env_wrapper.h"
+#include "file/random_access_file_reader.h"
+#include "file/writable_file_writer.h"
+#include "trace_replay/trace_replay.h"
+#include "util/coding.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+const unsigned int FileTraceReader::kBufferSize = 1024;  // 1KB
+
+FileTraceReader::FileTraceReader(
+    std::unique_ptr<RandomAccessFileReader>&& reader)
+    : file_reader_(std::move(reader)),
+      offset_(0),
+      buffer_(new char[kBufferSize]) {}
+
+FileTraceReader::~FileTraceReader() {
+  Close().PermitUncheckedError();
+  delete[] buffer_;
+}
+
+Status FileTraceReader::Close() {
+  file_reader_.reset();
+  return Status::OK();
+}
+
+Status FileTraceReader::Reset() {
+  if (file_reader_ == nullptr) {
+    return Status::IOError("TraceReader is closed.");
+  }
+  offset_ = 0;
+  return Status::OK();
+}
+
+Status FileTraceReader::Read(std::string* data) {
+  assert(file_reader_ != nullptr);
+  Status s = file_reader_->Read(IOOptions(), offset_, kTraceMetadataSize,
+                                &result_, buffer_, nullptr,
+                                Env::IO_TOTAL /* rate_limiter_priority */);
+  if (!s.ok()) {
+    return s;
+  }
+  if (result_.size() == 0) {
+    // No more data to read
+    // Todo: Come up with a better way to indicate end of data. May be this
+    // could be avoided once footer is introduced.
+    return Status::Incomplete();
+  }
+  if (result_.size() < kTraceMetadataSize) {
+    return Status::Corruption("Corrupted trace file.");
+  }
+  *data = result_.ToString();
+  offset_ += kTraceMetadataSize;
+
+  uint32_t payload_len =
+      DecodeFixed32(&buffer_[kTraceTimestampSize + kTraceTypeSize]);
+
+  // Read Payload
+  unsigned int bytes_to_read = payload_len;
+  unsigned int to_read =
+      bytes_to_read > kBufferSize ? kBufferSize : bytes_to_read;
+  while (to_read > 0) {
+    s = file_reader_->Read(IOOptions(), offset_, to_read, &result_, buffer_,
+                           nullptr, Env::IO_TOTAL /* rate_limiter_priority */);
+    if (!s.ok()) {
+      return s;
+    }
+    if (result_.size() < to_read) {
+      return Status::Corruption("Corrupted trace file.");
+    }
+    data->append(result_.data(), result_.size());
+
+    offset_ += to_read;
+    bytes_to_read -= to_read;
+    to_read = bytes_to_read > kBufferSize ? kBufferSize : bytes_to_read;
+  }
+
+  return s;
+}
+
+FileTraceWriter::FileTraceWriter(
+    std::unique_ptr<WritableFileWriter>&& file_writer)
+    : file_writer_(std::move(file_writer)) {}
+
+FileTraceWriter::~FileTraceWriter() { Close().PermitUncheckedError(); }
+
+Status FileTraceWriter::Close() {
+  file_writer_.reset();
+  return Status::OK();
+}
+
+Status FileTraceWriter::Write(const Slice& data) {
+  return file_writer_->Append(data);
+}
+
+uint64_t FileTraceWriter::GetFileSize() { return file_writer_->GetFileSize(); }
+
+Status NewFileTraceReader(Env* env, const EnvOptions& env_options,
+                          const std::string& trace_filename,
+                          std::unique_ptr<TraceReader>* trace_reader) {
+  std::unique_ptr<RandomAccessFileReader> file_reader;
+  Status s = RandomAccessFileReader::Create(
+      env->GetFileSystem(), trace_filename, FileOptions(env_options),
+      &file_reader, nullptr);
+  if (!s.ok()) {
+    return s;
+  }
+  trace_reader->reset(new FileTraceReader(std::move(file_reader)));
+  return s;
+}
+
+Status NewFileTraceWriter(Env* env, const EnvOptions& env_options,
+                          const std::string& trace_filename,
+                          std::unique_ptr<TraceWriter>* trace_writer) {
+  std::unique_ptr<WritableFileWriter> file_writer;
+  Status s = WritableFileWriter::Create(env->GetFileSystem(), trace_filename,
+                                        FileOptions(env_options), &file_writer,
+                                        nullptr);
+  if (!s.ok()) {
+    return s;
+  }
+  trace_writer->reset(new FileTraceWriter(std::move(file_writer)));
+  return s;
+}
+
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/utilities/trace/file_trace_reader_writer.h b/src/rocksdb/utilities/trace/file_trace_reader_writer.h
new file mode 100644
index 000000000..65d483108
--- /dev/null
+++ b/src/rocksdb/utilities/trace/file_trace_reader_writer.h
@@ -0,0 +1,48 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#pragma once
+
+#include "rocksdb/trace_reader_writer.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+class RandomAccessFileReader;
+class WritableFileWriter;
+
+// FileTraceReader allows reading RocksDB traces from a file.
+class FileTraceReader : public TraceReader {
+ public:
+  explicit FileTraceReader(std::unique_ptr<RandomAccessFileReader>&& reader);
+  ~FileTraceReader();
+
+  virtual Status Read(std::string* data) override;
+  virtual Status Close() override;
+  virtual Status Reset() override;
+
+ private:
+  std::unique_ptr<RandomAccessFileReader> file_reader_;
+  Slice result_;
+  size_t offset_;
+  char* const buffer_;
+
+  static const unsigned int kBufferSize;
+};
+
+// FileTraceWriter allows writing RocksDB traces to a file.
+class FileTraceWriter : public TraceWriter {
+ public:
+  explicit FileTraceWriter(std::unique_ptr<WritableFileWriter>&& file_writer);
+  ~FileTraceWriter();
+
+  virtual Status Write(const Slice& data) override;
+  virtual Status Close() override;
+  virtual uint64_t GetFileSize() override;
+
+ private:
+  std::unique_ptr<WritableFileWriter> file_writer_;
+};
+
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/utilities/trace/replayer_impl.cc b/src/rocksdb/utilities/trace/replayer_impl.cc
new file mode 100644
index 000000000..31023f1a2
--- /dev/null
+++ b/src/rocksdb/utilities/trace/replayer_impl.cc
@@ -0,0 +1,316 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#ifndef ROCKSDB_LITE
+
+#include "utilities/trace/replayer_impl.h"
+
+#include <cmath>
+#include <thread>
+
+#include "rocksdb/options.h"
+#include "rocksdb/slice.h"
+#include "rocksdb/system_clock.h"
+#include "util/threadpool_imp.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+ReplayerImpl::ReplayerImpl(DB* db,
+                           const std::vector<ColumnFamilyHandle*>& handles,
+                           std::unique_ptr<TraceReader>&& reader)
+    : Replayer(),
+      trace_reader_(std::move(reader)),
+      prepared_(false),
+      trace_end_(false),
+      header_ts_(0),
+      exec_handler_(TraceRecord::NewExecutionHandler(db, handles)),
+      env_(db->GetEnv()),
+      trace_file_version_(-1) {}
+
+ReplayerImpl::~ReplayerImpl() {
+  exec_handler_.reset();
+  trace_reader_.reset();
+}
+
+Status ReplayerImpl::Prepare() {
+  Trace header;
+  int db_version;
+  Status s = ReadHeader(&header);
+  if (!s.ok()) {
+    return s;
+  }
+  s = TracerHelper::ParseTraceHeader(header, &trace_file_version_, &db_version);
+  if (!s.ok()) {
+    return s;
+  }
+  header_ts_ = header.ts;
+  prepared_ = true;
+  trace_end_ = false;
+  return Status::OK();
+}
+
+Status ReplayerImpl::Next(std::unique_ptr<TraceRecord>* record) {
+  if (!prepared_) {
+    return Status::Incomplete("Not prepared!");
+  }
+  if (trace_end_) {
+    return Status::Incomplete("Trace end.");
+  }
+
+  Trace trace;
+  Status s = ReadTrace(&trace);  // ReadTrace is atomic
+  // Reached the trace end.
+  if (s.ok() && trace.type == kTraceEnd) {
+    trace_end_ = true;
+    return Status::Incomplete("Trace end.");
+  }
+  if (!s.ok() || record == nullptr) {
+    return s;
+  }
+
+  return TracerHelper::DecodeTraceRecord(&trace, trace_file_version_, record);
+}
+
+Status ReplayerImpl::Execute(const std::unique_ptr<TraceRecord>& record,
+                             std::unique_ptr<TraceRecordResult>* result) {
+  return record->Accept(exec_handler_.get(), result);
+}
+
+Status ReplayerImpl::Replay(
+    const ReplayOptions& options,
+    const std::function<void(Status, std::unique_ptr<TraceRecordResult>&&)>&
+        result_callback) {
+  if (options.fast_forward <= 0.0) {
+    return Status::InvalidArgument("Wrong fast forward speed!");
+  }
+
+  if (!prepared_) {
+    return Status::Incomplete("Not prepared!");
+  }
+  if (trace_end_) {
+    return Status::Incomplete("Trace end.");
+  }
+
+  Status s = Status::OK();
+
+  if (options.num_threads <= 1) {
+    // num_threads == 0 or num_threads == 1 uses single thread.
+    std::chrono::system_clock::time_point replay_epoch =
+        std::chrono::system_clock::now();
+
+    while (s.ok()) {
+      Trace trace;
+      s = ReadTrace(&trace);
+      // If already at trace end, ReadTrace should return Status::Incomplete().
+      if (!s.ok()) {
+        break;
+      }
+
+      // No need to sleep before breaking the loop if at the trace end.
+      if (trace.type == kTraceEnd) {
+        trace_end_ = true;
+        s = Status::Incomplete("Trace end.");
+        break;
+      }
+
+      // In single-threaded replay, decode first then sleep.
+      std::unique_ptr<TraceRecord> record;
+      s = TracerHelper::DecodeTraceRecord(&trace, trace_file_version_, &record);
+      if (!s.ok() && !s.IsNotSupported()) {
+        break;
+      }
+
+      std::chrono::system_clock::time_point sleep_to =
+          replay_epoch +
+          std::chrono::microseconds(static_cast<uint64_t>(std::llround(
+              1.0 * (trace.ts - header_ts_) / options.fast_forward)));
+      if (sleep_to > std::chrono::system_clock::now()) {
+        std::this_thread::sleep_until(sleep_to);
+      }
+
+      // Skip unsupported traces, stop for other errors.
+      if (s.IsNotSupported()) {
+        if (result_callback != nullptr) {
+          result_callback(s, nullptr);
+        }
+        s = Status::OK();
+        continue;
+      }
+
+      if (result_callback == nullptr) {
+        s = Execute(record, nullptr);
+      } else {
+        std::unique_ptr<TraceRecordResult> res;
+        s = Execute(record, &res);
+        result_callback(s, std::move(res));
+      }
+    }
+  } else {
+    // Multi-threaded replay.
+    ThreadPoolImpl thread_pool;
+    thread_pool.SetHostEnv(env_);
+    thread_pool.SetBackgroundThreads(static_cast<int>(options.num_threads));
+
+    std::mutex mtx;
+    // Background decoding and execution status.
+    Status bg_s = Status::OK();
+    uint64_t last_err_ts = static_cast<uint64_t>(-1);
+    // Callback function used in background work to update bg_s for the ealiest
+    // TraceRecord which has execution error. This is different from the
+    // timestamp of the first execution error (either start or end timestamp).
+    //
+    // Suppose TraceRecord R1, R2, with timestamps T1 < T2. Their execution
+    // timestamps are T1_start, T1_end, T2_start, T2_end.
+    // Single-thread: there must be T1_start < T1_end < T2_start < T2_end.
+    // Multi-thread: T1_start < T2_start may not be enforced. Orders of them are
+    // totally unknown.
+    // In order to report the same `first` error in both single-thread and
+    // multi-thread replay, we can only rely on the TraceRecords' timestamps,
+    // rather than their executin timestamps. Although in single-thread replay,
+    // the first error is also the last error, while in multi-thread replay, the
+    // first error may not be the first error in execution, and it may not be
+    // the last error in exeution as well.
+    auto error_cb = [&mtx, &bg_s, &last_err_ts](Status err, uint64_t err_ts) {
+      std::lock_guard<std::mutex> gd(mtx);
+      // Only record the first error.
+      if (!err.ok() && !err.IsNotSupported() && err_ts < last_err_ts) {
+        bg_s = err;
+        last_err_ts = err_ts;
+      }
+    };
+
+    std::chrono::system_clock::time_point replay_epoch =
+        std::chrono::system_clock::now();
+
+    while (bg_s.ok() && s.ok()) {
+      Trace trace;
+      s = ReadTrace(&trace);
+      // If already at trace end, ReadTrace should return Status::Incomplete().
+      if (!s.ok()) {
+        break;
+      }
+
+      TraceType trace_type = trace.type;
+
+      // No need to sleep before breaking the loop if at the trace end.
+      if (trace_type == kTraceEnd) {
+        trace_end_ = true;
+        s = Status::Incomplete("Trace end.");
+        break;
+      }
+
+      // In multi-threaded replay, sleep first then start decoding and
+      // execution in a thread.
+      std::chrono::system_clock::time_point sleep_to =
+          replay_epoch +
+          std::chrono::microseconds(static_cast<uint64_t>(std::llround(
+              1.0 * (trace.ts - header_ts_) / options.fast_forward)));
+      if (sleep_to > std::chrono::system_clock::now()) {
+        std::this_thread::sleep_until(sleep_to);
+      }
+
+      if (trace_type == kTraceWrite || trace_type == kTraceGet ||
+          trace_type == kTraceIteratorSeek ||
+          trace_type == kTraceIteratorSeekForPrev ||
+          trace_type == kTraceMultiGet) {
+        std::unique_ptr<ReplayerWorkerArg> ra(new ReplayerWorkerArg);
+        ra->trace_entry = std::move(trace);
+        ra->handler = exec_handler_.get();
+        ra->trace_file_version = trace_file_version_;
+        ra->error_cb = error_cb;
+        ra->result_cb = result_callback;
+        thread_pool.Schedule(&ReplayerImpl::BackgroundWork, ra.release(),
+                             nullptr, nullptr);
+      } else {
+        // Skip unsupported traces.
+        if (result_callback != nullptr) {
+          result_callback(Status::NotSupported("Unsupported trace type."),
+                          nullptr);
+        }
+      }
+    }
+
+    thread_pool.WaitForJobsAndJoinAllThreads();
+    if (!bg_s.ok()) {
+      s = bg_s;
+    }
+  }
+
+  if (s.IsIncomplete()) {
+    // Reaching eof returns Incomplete status at the moment.
+    // Could happen when killing a process without calling EndTrace() API.
+    // TODO: Add better error handling.
+    trace_end_ = true;
+    return Status::OK();
+  }
+  return s;
+}
+
+uint64_t ReplayerImpl::GetHeaderTimestamp() const { return header_ts_; }
+
+Status ReplayerImpl::ReadHeader(Trace* header) {
+  assert(header != nullptr);
+  Status s = trace_reader_->Reset();
+  if (!s.ok()) {
+    return s;
+  }
+  std::string encoded_trace;
+  // Read the trace head
+  s = trace_reader_->Read(&encoded_trace);
+  if (!s.ok()) {
+    return s;
+  }
+
+  return TracerHelper::DecodeHeader(encoded_trace, header);
+}
+
+Status ReplayerImpl::ReadTrace(Trace* trace) {
+  assert(trace != nullptr);
+  std::string encoded_trace;
+  // We don't know if TraceReader is implemented thread-safe, so we protect the
+  // reading trace part with a mutex. The decoding part does not need to be
+  // protected since it's local.
+  {
+    std::lock_guard<std::mutex> guard(mutex_);
+    Status s = trace_reader_->Read(&encoded_trace);
+    if (!s.ok()) {
+      return s;
+    }
+  }
+  return TracerHelper::DecodeTrace(encoded_trace, trace);
+}
+
+void ReplayerImpl::BackgroundWork(void* arg) {
+  std::unique_ptr<ReplayerWorkerArg> ra(
+      reinterpret_cast<ReplayerWorkerArg*>(arg));
+  assert(ra != nullptr);
+
+  std::unique_ptr<TraceRecord> record;
+  Status s = TracerHelper::DecodeTraceRecord(&(ra->trace_entry),
+                                             ra->trace_file_version, &record);
+  if (!s.ok()) {
+    // Stop the replay
+    if (ra->error_cb != nullptr) {
+      ra->error_cb(s, ra->trace_entry.ts);
+    }
+    // Report the result
+    if (ra->result_cb != nullptr) {
+      ra->result_cb(s, nullptr);
+    }
+    return;
+  }
+
+  if (ra->result_cb == nullptr) {
+    s = record->Accept(ra->handler, nullptr);
+  } else {
+    std::unique_ptr<TraceRecordResult> res;
+    s = record->Accept(ra->handler, &res);
+    ra->result_cb(s, std::move(res));
+  }
+  record.reset();
+}
+
+}  // namespace ROCKSDB_NAMESPACE
+#endif  // ROCKSDB_LITE
diff --git a/src/rocksdb/utilities/trace/replayer_impl.h b/src/rocksdb/utilities/trace/replayer_impl.h
new file mode 100644
index 000000000..367b0b51e
--- /dev/null
+++ b/src/rocksdb/utilities/trace/replayer_impl.h
@@ -0,0 +1,86 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#pragma once
+#ifndef ROCKSDB_LITE
+
+#include <atomic>
+#include <functional>
+#include <memory>
+#include <mutex>
+#include <unordered_map>
+
+#include "rocksdb/db.h"
+#include "rocksdb/env.h"
+#include "rocksdb/status.h"
+#include "rocksdb/trace_reader_writer.h"
+#include "rocksdb/trace_record.h"
+#include "rocksdb/trace_record_result.h"
+#include "rocksdb/utilities/replayer.h"
+#include "trace_replay/trace_replay.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+class ReplayerImpl : public Replayer {
+ public:
+  ReplayerImpl(DB* db, const std::vector<ColumnFamilyHandle*>& handles,
+               std::unique_ptr<TraceReader>&& reader);
+  ~ReplayerImpl() override;
+
+  using Replayer::Prepare;
+  Status Prepare() override;
+
+  using Replayer::Next;
+  Status Next(std::unique_ptr<TraceRecord>* record) override;
+
+  using Replayer::Execute;
+  Status Execute(const std::unique_ptr<TraceRecord>& record,
+                 std::unique_ptr<TraceRecordResult>* result) override;
+
+  using Replayer::Replay;
+  Status Replay(
+      const ReplayOptions& options,
+      const std::function<void(Status, std::unique_ptr<TraceRecordResult>&&)>&
+          result_callback) override;
+
+  using Replayer::GetHeaderTimestamp;
+  uint64_t GetHeaderTimestamp() const override;
+
+ private:
+  Status ReadHeader(Trace* header);
+  Status ReadTrace(Trace* trace);
+
+  // Generic function to execute a Trace in a thread pool.
+  static void BackgroundWork(void* arg);
+
+  std::unique_ptr<TraceReader> trace_reader_;
+  std::mutex mutex_;
+  std::atomic<bool> prepared_;
+  std::atomic<bool> trace_end_;
+  uint64_t header_ts_;
+  std::unique_ptr<TraceRecord::Handler> exec_handler_;
+  Env* env_;
+  // When reading the trace header, the trace file version can be parsed.
+  // Replayer will use different decode method to get the trace content based
+  // on different trace file version.
+  int trace_file_version_;
+};
+
+// Arguments passed to BackgroundWork() for replaying in a thread pool.
+struct ReplayerWorkerArg {
+  Trace trace_entry;
+  int trace_file_version;
+  // Handler to execute TraceRecord.
+  TraceRecord::Handler* handler;
+  // Callback function to report the error status and the timestamp of the
+  // TraceRecord (not the start/end timestamp of executing the TraceRecord).
+  std::function<void(Status, uint64_t)> error_cb;
+  // Callback function to report the trace execution status and operation
+  // execution status/result(s).
+  std::function<void(Status, std::unique_ptr<TraceRecordResult>&&)> result_cb;
+};
+
+}  // namespace ROCKSDB_NAMESPACE
+#endif  // ROCKSDB_LITE
diff --git a/src/rocksdb/utilities/transactions/lock/lock_manager.cc b/src/rocksdb/utilities/transactions/lock/lock_manager.cc
new file mode 100644
index 000000000..df16b32ad
--- /dev/null
+++ b/src/rocksdb/utilities/transactions/lock/lock_manager.cc
@@ -0,0 +1,29 @@
+// Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+// This source code is licensed under both the GPLv2 (found in the
+// COPYING file in the root directory) and Apache 2.0 License
+// (found in the LICENSE.Apache file in the root directory).
+
+#ifndef ROCKSDB_LITE
+
+#include "utilities/transactions/lock/lock_manager.h"
+
+#include "utilities/transactions/lock/point/point_lock_manager.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+std::shared_ptr<LockManager> NewLockManager(PessimisticTransactionDB* db,
+                                            const TransactionDBOptions& opt) {
+  assert(db);
+  if (opt.lock_mgr_handle) {
+    // A custom lock manager was provided in options
+    auto mgr = opt.lock_mgr_handle->getLockManager();
+    return std::shared_ptr<LockManager>(opt.lock_mgr_handle, mgr);
+  } else {
+    // Use a point lock manager by default
+    return std::shared_ptr<LockManager>(new PointLockManager(db, opt));
+  }
+}
+
+}  // namespace ROCKSDB_NAMESPACE
+
+#endif  // ROCKSDB_LITE
diff --git a/src/rocksdb/utilities/transactions/lock/lock_manager.h b/src/rocksdb/utilities/transactions/lock/lock_manager.h
new file mode 100644
index 000000000..a5ce1948c
--- /dev/null
+++ b/src/rocksdb/utilities/transactions/lock/lock_manager.h
@@ -0,0 +1,82 @@
+// Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+// This source code is licensed under both the GPLv2 (found in the
+// COPYING file in the root directory) and Apache 2.0 License
+// (found in the LICENSE.Apache file in the root directory).
+
+#pragma once
+#ifndef ROCKSDB_LITE
+
+#include "rocksdb/types.h"
+#include "rocksdb/utilities/transaction.h"
+#include "rocksdb/utilities/transaction_db.h"
+#include "utilities/transactions/lock/lock_tracker.h"
+#include "utilities/transactions/pessimistic_transaction.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+class PessimisticTransactionDB;
+
+class LockManager {
+ public:
+  virtual ~LockManager() {}
+
+  // Whether supports locking a specific key.
+  virtual bool IsPointLockSupported() const = 0;
+
+  // Whether supports locking a range of keys.
+  virtual bool IsRangeLockSupported() const = 0;
+
+  // Locks acquired through this LockManager should be tracked by
+  // the LockTrackers created through the returned factory.
+  virtual const LockTrackerFactory& GetLockTrackerFactory() const = 0;
+
+  // Enable locking for the specified column family.
+  // Caller should guarantee that this column family is not already enabled.
+  virtual void AddColumnFamily(const ColumnFamilyHandle* cf) = 0;
+
+  // Disable locking for the specified column family.
+  // Caller should guarantee that this column family is no longer used.
+  virtual void RemoveColumnFamily(const ColumnFamilyHandle* cf) = 0;
+
+  // Attempt to lock a key or a key range.  If OK status is returned, the caller
+  // is responsible for calling UnLock() on this key.
+  virtual Status TryLock(PessimisticTransaction* txn,
+                         ColumnFamilyId column_family_id,
+                         const std::string& key, Env* env, bool exclusive) = 0;
+  // The range [start, end] are inclusive at both sides.
+  virtual Status TryLock(PessimisticTransaction* txn,
+                         ColumnFamilyId column_family_id, const Endpoint& start,
+                         const Endpoint& end, Env* env, bool exclusive) = 0;
+
+  // Unlock a key or a range locked by TryLock().  txn must be the same
+  // Transaction that locked this key.
+  virtual void UnLock(PessimisticTransaction* txn, const LockTracker& tracker,
+                      Env* env) = 0;
+  virtual void UnLock(PessimisticTransaction* txn,
+                      ColumnFamilyId column_family_id, const std::string& key,
+                      Env* env) = 0;
+  virtual void UnLock(PessimisticTransaction* txn,
+                      ColumnFamilyId column_family_id, const Endpoint& start,
+                      const Endpoint& end, Env* env) = 0;
+
+  using PointLockStatus = std::unordered_multimap<ColumnFamilyId, KeyLockInfo>;
+  virtual PointLockStatus GetPointLockStatus() = 0;
+
+  using RangeLockStatus =
+      std::unordered_multimap<ColumnFamilyId, RangeLockInfo>;
+  virtual RangeLockStatus GetRangeLockStatus() = 0;
+
+  virtual std::vector<DeadlockPath> GetDeadlockInfoBuffer() = 0;
+
+  virtual void Resize(uint32_t new_size) = 0;
+};
+
+// LockManager should always be constructed through this factory method,
+// instead of constructing through concrete implementations' constructor.
+// Caller owns the returned pointer.
+std::shared_ptr<LockManager> NewLockManager(PessimisticTransactionDB* db,
+                                            const TransactionDBOptions& opt);
+
+}  // namespace ROCKSDB_NAMESPACE
+
+#endif  // ROCKSDB_LITE
diff --git a/src/rocksdb/utilities/transactions/lock/lock_tracker.h b/src/rocksdb/utilities/transactions/lock/lock_tracker.h
new file mode 100644
index 000000000..5fa228a82
--- /dev/null
+++ b/src/rocksdb/utilities/transactions/lock/lock_tracker.h
@@ -0,0 +1,209 @@
+// Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+// This source code is licensed under both the GPLv2 (found in the
+// COPYING file in the root directory) and Apache 2.0 License
+// (found in the LICENSE.Apache file in the root directory).
+
+#pragma once
+#ifndef ROCKSDB_LITE
+
+#include <memory>
+
+#include "rocksdb/rocksdb_namespace.h"
+#include "rocksdb/status.h"
+#include "rocksdb/types.h"
+#include "rocksdb/utilities/transaction_db.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+// Request for locking a single key.
+struct PointLockRequest {
+  // The id of the key's column family.
+  ColumnFamilyId column_family_id = 0;
+  // The key to lock.
+  std::string key;
+  // The sequence number from which there is no concurrent update to key.
+  SequenceNumber seq = 0;
+  // Whether the lock is acquired only for read.
+  bool read_only = false;
+  // Whether the lock is in exclusive mode.
+  bool exclusive = true;
+};
+
+// Request for locking a range of keys.
+struct RangeLockRequest {
+  // The id of the key's column family.
+  ColumnFamilyId column_family_id;
+
+  // The range to be locked
+  Endpoint start_endp;
+  Endpoint end_endp;
+};
+
+struct PointLockStatus {
+  // Whether the key is locked.
+  bool locked = false;
+  // Whether the key is locked in exclusive mode.
+  bool exclusive = true;
+  // The sequence number in the tracked PointLockRequest.
+  SequenceNumber seq = 0;
+};
+
+// Return status when calling LockTracker::Untrack.
+enum class UntrackStatus {
+  // The lock is not tracked at all, so no lock to untrack.
+  NOT_TRACKED,
+  // The lock is untracked but not removed from the tracker.
+  UNTRACKED,
+  // The lock is removed from the tracker.
+  REMOVED,
+};
+
+// Tracks the lock requests.
+// In PessimisticTransaction, it tracks the locks acquired through LockMgr;
+// In OptimisticTransaction, since there is no LockMgr, it tracks the lock
+// intention. Not thread-safe.
+class LockTracker {
+ public:
+  virtual ~LockTracker() {}
+
+  // Whether supports locking a specific key.
+  virtual bool IsPointLockSupported() const = 0;
+
+  // Whether supports locking a range of keys.
+  virtual bool IsRangeLockSupported() const = 0;
+
+  // Tracks the acquirement of a lock on key.
+  //
+  // If this method is not supported, leave it as a no-op.
+  virtual void Track(const PointLockRequest& /*lock_request*/) = 0;
+
+  // Untracks the lock on a key.
+  // seq and exclusive in lock_request are not used.
+  //
+  // If this method is not supported, leave it as a no-op and
+  // returns NOT_TRACKED.
+  virtual UntrackStatus Untrack(const PointLockRequest& /*lock_request*/) = 0;
+
+  // Counterpart of Track(const PointLockRequest&) for RangeLockRequest.
+  virtual void Track(const RangeLockRequest& /*lock_request*/) = 0;
+
+  // Counterpart of Untrack(const PointLockRequest&) for RangeLockRequest.
+  virtual UntrackStatus Untrack(const RangeLockRequest& /*lock_request*/) = 0;
+
+  // Merges lock requests tracked in the specified tracker into the current
+  // tracker.
+  //
+  // E.g. for point lock, if a key in tracker is not yet tracked,
+  // track this new key; otherwise, merge the tracked information of the key
+  // such as lock's exclusiveness, read/write statistics.
+  //
+  // If this method is not supported, leave it as a no-op.
+  //
+  // REQUIRED: the specified tracker must be of the same concrete class type as
+  // the current tracker.
+  virtual void Merge(const LockTracker& /*tracker*/) = 0;
+
+  // This is a reverse operation of Merge.
+  //
+  // E.g. for point lock, if a key exists in both current and the sepcified
+  // tracker, then subtract the information (such as read/write statistics) of
+  // the key in the specified tracker from the current tracker.
+  //
+  // If this method is not supported, leave it as a no-op.
+  //
+  // REQUIRED:
+  // The specified tracker must be of the same concrete class type as
+  // the current tracker.
+  // The tracked locks in the specified tracker must be a subset of those
+  // tracked by the current tracker.
+  virtual void Subtract(const LockTracker& /*tracker*/) = 0;
+
+  // Clears all tracked locks.
+  virtual void Clear() = 0;
+
+  // Gets the new locks (excluding the locks that have been tracked before the
+  // save point) tracked since the specified save point, the result is stored
+  // in an internally constructed LockTracker and returned.
+  //
+  // save_point_tracker is the tracker used by a SavePoint to track locks
+  // tracked after creating the SavePoint.
+  //
+  // The implementation should document whether point lock, or range lock, or
+  // both are considered in this method.
+  // If this method is not supported, returns nullptr.
+  //
+  // REQUIRED:
+  // The save_point_tracker must be of the same concrete class type as the
+  // current tracker.
+  // The tracked locks in the specified tracker must be a subset of those
+  // tracked by the current tracker.
+  virtual LockTracker* GetTrackedLocksSinceSavePoint(
+      const LockTracker& /*save_point_tracker*/) const = 0;
+
+  // Gets lock related information of the key.
+  //
+  // If point lock is not supported, always returns LockStatus with
+  // locked=false.
+  virtual PointLockStatus GetPointLockStatus(
+      ColumnFamilyId /*column_family_id*/,
+      const std::string& /*key*/) const = 0;
+
+  // Gets number of tracked point locks.
+  //
+  // If point lock is not supported, always returns 0.
+  virtual uint64_t GetNumPointLocks() const = 0;
+
+  class ColumnFamilyIterator {
+   public:
+    virtual ~ColumnFamilyIterator() {}
+
+    // Whether there are remaining column families.
+    virtual bool HasNext() const = 0;
+
+    // Gets next column family id.
+    //
+    // If HasNext is false, calling this method has undefined behavior.
+    virtual ColumnFamilyId Next() = 0;
+  };
+
+  // Gets an iterator for column families.
+  //
+  // Returned iterator must not be nullptr.
+  // If there is no column family to iterate,
+  // returns an empty non-null iterator.
+  // Caller owns the returned pointer.
+  virtual ColumnFamilyIterator* GetColumnFamilyIterator() const = 0;
+
+  class KeyIterator {
+   public:
+    virtual ~KeyIterator() {}
+
+    // Whether there are remaining keys.
+    virtual bool HasNext() const = 0;
+
+    // Gets the next key.
+    //
+    // If HasNext is false, calling this method has undefined behavior.
+    virtual const std::string& Next() = 0;
+  };
+
+  // Gets an iterator for keys with tracked point locks in the column family.
+  //
+  // The column family must exist.
+  // Returned iterator must not be nullptr.
+  // Caller owns the returned pointer.
+  virtual KeyIterator* GetKeyIterator(
+      ColumnFamilyId /*column_family_id*/) const = 0;
+};
+
+// LockTracker should always be constructed through this factory.
+// Each LockManager owns a LockTrackerFactory.
+class LockTrackerFactory {
+ public:
+  // Caller owns the returned pointer.
+  virtual LockTracker* Create() const = 0;
+  virtual ~LockTrackerFactory() {}
+};
+
+}  // namespace ROCKSDB_NAMESPACE
+#endif  // ROCKSDB_LITE
diff --git a/src/rocksdb/utilities/transactions/lock/point/point_lock_manager.cc b/src/rocksdb/utilities/transactions/lock/point/point_lock_manager.cc
new file mode 100644
index 000000000..b362a164d
--- /dev/null
+++ b/src/rocksdb/utilities/transactions/lock/point/point_lock_manager.cc
@@ -0,0 +1,721 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#ifndef ROCKSDB_LITE
+
+#include "utilities/transactions/lock/point/point_lock_manager.h"
+
+#include <algorithm>
+#include <cinttypes>
+#include <mutex>
+
+#include "monitoring/perf_context_imp.h"
+#include "rocksdb/slice.h"
+#include "rocksdb/utilities/transaction_db_mutex.h"
+#include "test_util/sync_point.h"
+#include "util/cast_util.h"
+#include "util/hash.h"
+#include "util/thread_local.h"
+#include "utilities/transactions/pessimistic_transaction_db.h"
+#include "utilities/transactions/transaction_db_mutex_impl.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+struct LockInfo {
+  bool exclusive;
+  autovector<TransactionID> txn_ids;
+
+  // Transaction locks are not valid after this time in us
+  uint64_t expiration_time;
+
+  LockInfo(TransactionID id, uint64_t time, bool ex)
+      : exclusive(ex), expiration_time(time) {
+    txn_ids.push_back(id);
+  }
+  LockInfo(const LockInfo& lock_info)
+      : exclusive(lock_info.exclusive),
+        txn_ids(lock_info.txn_ids),
+        expiration_time(lock_info.expiration_time) {}
+  void operator=(const LockInfo& lock_info) {
+    exclusive = lock_info.exclusive;
+    txn_ids = lock_info.txn_ids;
+    expiration_time = lock_info.expiration_time;
+  }
+  DECLARE_DEFAULT_MOVES(LockInfo);
+};
+
+struct LockMapStripe {
+  explicit LockMapStripe(std::shared_ptr<TransactionDBMutexFactory> factory) {
+    stripe_mutex = factory->AllocateMutex();
+    stripe_cv = factory->AllocateCondVar();
+    assert(stripe_mutex);
+    assert(stripe_cv);
+  }
+
+  // Mutex must be held before modifying keys map
+  std::shared_ptr<TransactionDBMutex> stripe_mutex;
+
+  // Condition Variable per stripe for waiting on a lock
+  std::shared_ptr<TransactionDBCondVar> stripe_cv;
+
+  // Locked keys mapped to the info about the transactions that locked them.
+  // TODO(agiardullo): Explore performance of other data structures.
+  UnorderedMap<std::string, LockInfo> keys;
+};
+
+// Map of #num_stripes LockMapStripes
+struct LockMap {
+  explicit LockMap(size_t num_stripes,
+                   std::shared_ptr<TransactionDBMutexFactory> factory)
+      : num_stripes_(num_stripes) {
+    lock_map_stripes_.reserve(num_stripes);
+    for (size_t i = 0; i < num_stripes; i++) {
+      LockMapStripe* stripe = new LockMapStripe(factory);
+      lock_map_stripes_.push_back(stripe);
+    }
+  }
+
+  ~LockMap() {
+    for (auto stripe : lock_map_stripes_) {
+      delete stripe;
+    }
+  }
+
+  // Number of sepearate LockMapStripes to create, each with their own Mutex
+  const size_t num_stripes_;
+
+  // Count of keys that are currently locked in this column family.
+  // (Only maintained if PointLockManager::max_num_locks_ is positive.)
+  std::atomic<int64_t> lock_cnt{0};
+
+  std::vector<LockMapStripe*> lock_map_stripes_;
+
+  size_t GetStripe(const std::string& key) const;
+};
+
+namespace {
+void UnrefLockMapsCache(void* ptr) {
+  // Called when a thread exits or a ThreadLocalPtr gets destroyed.
+  auto lock_maps_cache =
+      static_cast<UnorderedMap<uint32_t, std::shared_ptr<LockMap>>*>(ptr);
+  delete lock_maps_cache;
+}
+}  // anonymous namespace
+
+PointLockManager::PointLockManager(PessimisticTransactionDB* txn_db,
+                                   const TransactionDBOptions& opt)
+    : txn_db_impl_(txn_db),
+      default_num_stripes_(opt.num_stripes),
+      max_num_locks_(opt.max_num_locks),
+      lock_maps_cache_(new ThreadLocalPtr(&UnrefLockMapsCache)),
+      dlock_buffer_(opt.max_num_deadlocks),
+      mutex_factory_(opt.custom_mutex_factory
+                         ? opt.custom_mutex_factory
+                         : std::make_shared<TransactionDBMutexFactoryImpl>()) {}
+
+size_t LockMap::GetStripe(const std::string& key) const {
+  assert(num_stripes_ > 0);
+  return FastRange64(GetSliceNPHash64(key), num_stripes_);
+}
+
+void PointLockManager::AddColumnFamily(const ColumnFamilyHandle* cf) {
+  InstrumentedMutexLock l(&lock_map_mutex_);
+
+  if (lock_maps_.find(cf->GetID()) == lock_maps_.end()) {
+    lock_maps_.emplace(cf->GetID(), std::make_shared<LockMap>(
+                                        default_num_stripes_, mutex_factory_));
+  } else {
+    // column_family already exists in lock map
+    assert(false);
+  }
+}
+
+void PointLockManager::RemoveColumnFamily(const ColumnFamilyHandle* cf) {
+  // Remove lock_map for this column family.  Since the lock map is stored
+  // as a shared ptr, concurrent transactions can still keep using it
+  // until they release their references to it.
+  {
+    InstrumentedMutexLock l(&lock_map_mutex_);
+
+    auto lock_maps_iter = lock_maps_.find(cf->GetID());
+    if (lock_maps_iter == lock_maps_.end()) {
+      return;
+    }
+
+    lock_maps_.erase(lock_maps_iter);
+  }  // lock_map_mutex_
+
+  // Clear all thread-local caches
+  autovector<void*> local_caches;
+  lock_maps_cache_->Scrape(&local_caches, nullptr);
+  for (auto cache : local_caches) {
+    delete static_cast<LockMaps*>(cache);
+  }
+}
+
+// Look up the LockMap std::shared_ptr for a given column_family_id.
+// Note:  The LockMap is only valid as long as the caller is still holding on
+//   to the returned std::shared_ptr.
+std::shared_ptr<LockMap> PointLockManager::GetLockMap(
+    ColumnFamilyId column_family_id) {
+  // First check thread-local cache
+  if (lock_maps_cache_->Get() == nullptr) {
+    lock_maps_cache_->Reset(new LockMaps());
+  }
+
+  auto lock_maps_cache = static_cast<LockMaps*>(lock_maps_cache_->Get());
+
+  auto lock_map_iter = lock_maps_cache->find(column_family_id);
+  if (lock_map_iter != lock_maps_cache->end()) {
+    // Found lock map for this column family.
+    return lock_map_iter->second;
+  }
+
+  // Not found in local cache, grab mutex and check shared LockMaps
+  InstrumentedMutexLock l(&lock_map_mutex_);
+
+  lock_map_iter = lock_maps_.find(column_family_id);
+  if (lock_map_iter == lock_maps_.end()) {
+    return std::shared_ptr<LockMap>(nullptr);
+  } else {
+    // Found lock map.  Store in thread-local cache and return.
+    std::shared_ptr<LockMap>& lock_map = lock_map_iter->second;
+    lock_maps_cache->insert({column_family_id, lock_map});
+
+    return lock_map;
+  }
+}
+
+// Returns true if this lock has expired and can be acquired by another
+// transaction.
+// If false, sets *expire_time to the expiration time of the lock according
+// to Env->GetMicros() or 0 if no expiration.
+bool PointLockManager::IsLockExpired(TransactionID txn_id,
+                                     const LockInfo& lock_info, Env* env,
+                                     uint64_t* expire_time) {
+  if (lock_info.expiration_time == 0) {
+    *expire_time = 0;
+    return false;
+  }
+
+  auto now = env->NowMicros();
+  bool expired = lock_info.expiration_time <= now;
+  if (!expired) {
+    // return how many microseconds until lock will be expired
+    *expire_time = lock_info.expiration_time;
+  } else {
+    for (auto id : lock_info.txn_ids) {
+      if (txn_id == id) {
+        continue;
+      }
+
+      bool success = txn_db_impl_->TryStealingExpiredTransactionLocks(id);
+      if (!success) {
+        expired = false;
+        *expire_time = 0;
+        break;
+      }
+    }
+  }
+
+  return expired;
+}
+
+Status PointLockManager::TryLock(PessimisticTransaction* txn,
+                                 ColumnFamilyId column_family_id,
+                                 const std::string& key, Env* env,
+                                 bool exclusive) {
+  // Lookup lock map for this column family id
+  std::shared_ptr<LockMap> lock_map_ptr = GetLockMap(column_family_id);
+  LockMap* lock_map = lock_map_ptr.get();
+  if (lock_map == nullptr) {
+    char msg[255];
+    snprintf(msg, sizeof(msg), "Column family id not found: %" PRIu32,
+             column_family_id);
+
+    return Status::InvalidArgument(msg);
+  }
+
+  // Need to lock the mutex for the stripe that this key hashes to
+  size_t stripe_num = lock_map->GetStripe(key);
+  assert(lock_map->lock_map_stripes_.size() > stripe_num);
+  LockMapStripe* stripe = lock_map->lock_map_stripes_.at(stripe_num);
+
+  LockInfo lock_info(txn->GetID(), txn->GetExpirationTime(), exclusive);
+  int64_t timeout = txn->GetLockTimeout();
+
+  return AcquireWithTimeout(txn, lock_map, stripe, column_family_id, key, env,
+                            timeout, lock_info);
+}
+
+// Helper function for TryLock().
+Status PointLockManager::AcquireWithTimeout(
+    PessimisticTransaction* txn, LockMap* lock_map, LockMapStripe* stripe,
+    ColumnFamilyId column_family_id, const std::string& key, Env* env,
+    int64_t timeout, const LockInfo& lock_info) {
+  Status result;
+  uint64_t end_time = 0;
+
+  if (timeout > 0) {
+    uint64_t start_time = env->NowMicros();
+    end_time = start_time + timeout;
+  }
+
+  if (timeout < 0) {
+    // If timeout is negative, we wait indefinitely to acquire the lock
+    result = stripe->stripe_mutex->Lock();
+  } else {
+    result = stripe->stripe_mutex->TryLockFor(timeout);
+  }
+
+  if (!result.ok()) {
+    // failed to acquire mutex
+    return result;
+  }
+
+  // Acquire lock if we are able to
+  uint64_t expire_time_hint = 0;
+  autovector<TransactionID> wait_ids;
+  result = AcquireLocked(lock_map, stripe, key, env, lock_info,
+                         &expire_time_hint, &wait_ids);
+
+  if (!result.ok() && timeout != 0) {
+    PERF_TIMER_GUARD(key_lock_wait_time);
+    PERF_COUNTER_ADD(key_lock_wait_count, 1);
+    // If we weren't able to acquire the lock, we will keep retrying as long
+    // as the timeout allows.
+    bool timed_out = false;
+    do {
+      // Decide how long to wait
+      int64_t cv_end_time = -1;
+      if (expire_time_hint > 0 && end_time > 0) {
+        cv_end_time = std::min(expire_time_hint, end_time);
+      } else if (expire_time_hint > 0) {
+        cv_end_time = expire_time_hint;
+      } else if (end_time > 0) {
+        cv_end_time = end_time;
+      }
+
+      assert(result.IsBusy() || wait_ids.size() != 0);
+
+      // We are dependent on a transaction to finish, so perform deadlock
+      // detection.
+      if (wait_ids.size() != 0) {
+        if (txn->IsDeadlockDetect()) {
+          if (IncrementWaiters(txn, wait_ids, key, column_family_id,
+                               lock_info.exclusive, env)) {
+            result = Status::Busy(Status::SubCode::kDeadlock);
+            stripe->stripe_mutex->UnLock();
+            return result;
+          }
+        }
+        txn->SetWaitingTxn(wait_ids, column_family_id, &key);
+      }
+
+      TEST_SYNC_POINT("PointLockManager::AcquireWithTimeout:WaitingTxn");
+      if (cv_end_time < 0) {
+        // Wait indefinitely
+        result = stripe->stripe_cv->Wait(stripe->stripe_mutex);
+      } else {
+        uint64_t now = env->NowMicros();
+        if (static_cast<uint64_t>(cv_end_time) > now) {
+          result = stripe->stripe_cv->WaitFor(stripe->stripe_mutex,
+                                              cv_end_time - now);
+        }
+      }
+
+      if (wait_ids.size() != 0) {
+        txn->ClearWaitingTxn();
+        if (txn->IsDeadlockDetect()) {
+          DecrementWaiters(txn, wait_ids);
+        }
+      }
+
+      if (result.IsTimedOut()) {
+        timed_out = true;
+        // Even though we timed out, we will still make one more attempt to
+        // acquire lock below (it is possible the lock expired and we
+        // were never signaled).
+      }
+
+      if (result.ok() || result.IsTimedOut()) {
+        result = AcquireLocked(lock_map, stripe, key, env, lock_info,
+                               &expire_time_hint, &wait_ids);
+      }
+    } while (!result.ok() && !timed_out);
+  }
+
+  stripe->stripe_mutex->UnLock();
+
+  return result;
+}
+
+void PointLockManager::DecrementWaiters(
+    const PessimisticTransaction* txn,
+    const autovector<TransactionID>& wait_ids) {
+  std::lock_guard<std::mutex> lock(wait_txn_map_mutex_);
+  DecrementWaitersImpl(txn, wait_ids);
+}
+
+void PointLockManager::DecrementWaitersImpl(
+    const PessimisticTransaction* txn,
+    const autovector<TransactionID>& wait_ids) {
+  auto id = txn->GetID();
+  assert(wait_txn_map_.Contains(id));
+  wait_txn_map_.Delete(id);
+
+  for (auto wait_id : wait_ids) {
+    rev_wait_txn_map_.Get(wait_id)--;
+    if (rev_wait_txn_map_.Get(wait_id) == 0) {
+      rev_wait_txn_map_.Delete(wait_id);
+    }
+  }
+}
+
+bool PointLockManager::IncrementWaiters(
+    const PessimisticTransaction* txn,
+    const autovector<TransactionID>& wait_ids, const std::string& key,
+    const uint32_t& cf_id, const bool& exclusive, Env* const env) {
+  auto id = txn->GetID();
+  std::vector<int> queue_parents(
+      static_cast<size_t>(txn->GetDeadlockDetectDepth()));
+  std::vector<TransactionID> queue_values(
+      static_cast<size_t>(txn->GetDeadlockDetectDepth()));
+  std::lock_guard<std::mutex> lock(wait_txn_map_mutex_);
+  assert(!wait_txn_map_.Contains(id));
+
+  wait_txn_map_.Insert(id, {wait_ids, cf_id, exclusive, key});
+
+  for (auto wait_id : wait_ids) {
+    if (rev_wait_txn_map_.Contains(wait_id)) {
+      rev_wait_txn_map_.Get(wait_id)++;
+    } else {
+      rev_wait_txn_map_.Insert(wait_id, 1);
+    }
+  }
+
+  // No deadlock if nobody is waiting on self.
+  if (!rev_wait_txn_map_.Contains(id)) {
+    return false;
+  }
+
+  const auto* next_ids = &wait_ids;
+  int parent = -1;
+  int64_t deadlock_time = 0;
+  for (int tail = 0, head = 0; head < txn->GetDeadlockDetectDepth(); head++) {
+    int i = 0;
+    if (next_ids) {
+      for (; i < static_cast<int>(next_ids->size()) &&
+             tail + i < txn->GetDeadlockDetectDepth();
+           i++) {
+        queue_values[tail + i] = (*next_ids)[i];
+        queue_parents[tail + i] = parent;
+      }
+      tail += i;
+    }
+
+    // No more items in the list, meaning no deadlock.
+    if (tail == head) {
+      return false;
+    }
+
+    auto next = queue_values[head];
+    if (next == id) {
+      std::vector<DeadlockInfo> path;
+      while (head != -1) {
+        assert(wait_txn_map_.Contains(queue_values[head]));
+
+        auto extracted_info = wait_txn_map_.Get(queue_values[head]);
+        path.push_back({queue_values[head], extracted_info.m_cf_id,
+                        extracted_info.m_exclusive,
+                        extracted_info.m_waiting_key});
+        head = queue_parents[head];
+      }
+      if (!env->GetCurrentTime(&deadlock_time).ok()) {
+        /*
+          TODO(AR) this preserves the current behaviour whilst checking the
+          status of env->GetCurrentTime to ensure that ASSERT_STATUS_CHECKED
+          passes. Should we instead raise an error if !ok() ?
+        */
+        deadlock_time = 0;
+      }
+      std::reverse(path.begin(), path.end());
+      dlock_buffer_.AddNewPath(DeadlockPath(path, deadlock_time));
+      deadlock_time = 0;
+      DecrementWaitersImpl(txn, wait_ids);
+      return true;
+    } else if (!wait_txn_map_.Contains(next)) {
+      next_ids = nullptr;
+      continue;
+    } else {
+      parent = head;
+      next_ids = &(wait_txn_map_.Get(next).m_neighbors);
+    }
+  }
+
+  // Wait cycle too big, just assume deadlock.
+  if (!env->GetCurrentTime(&deadlock_time).ok()) {
+    /*
+      TODO(AR) this preserves the current behaviour whilst checking the status
+      of env->GetCurrentTime to ensure that ASSERT_STATUS_CHECKED passes.
+      Should we instead raise an error if !ok() ?
+    */
+    deadlock_time = 0;
+  }
+  dlock_buffer_.AddNewPath(DeadlockPath(deadlock_time, true));
+  DecrementWaitersImpl(txn, wait_ids);
+  return true;
+}
+
+// Try to lock this key after we have acquired the mutex.
+// Sets *expire_time to the expiration time in microseconds
+//  or 0 if no expiration.
+// REQUIRED:  Stripe mutex must be held.
+Status PointLockManager::AcquireLocked(LockMap* lock_map, LockMapStripe* stripe,
+                                       const std::string& key, Env* env,
+                                       const LockInfo& txn_lock_info,
+                                       uint64_t* expire_time,
+                                       autovector<TransactionID>* txn_ids) {
+  assert(txn_lock_info.txn_ids.size() == 1);
+
+  Status result;
+  // Check if this key is already locked
+  auto stripe_iter = stripe->keys.find(key);
+  if (stripe_iter != stripe->keys.end()) {
+    // Lock already held
+    LockInfo& lock_info = stripe_iter->second;
+    assert(lock_info.txn_ids.size() == 1 || !lock_info.exclusive);
+
+    if (lock_info.exclusive || txn_lock_info.exclusive) {
+      if (lock_info.txn_ids.size() == 1 &&
+          lock_info.txn_ids[0] == txn_lock_info.txn_ids[0]) {
+        // The list contains one txn and we're it, so just take it.
+        lock_info.exclusive = txn_lock_info.exclusive;
+        lock_info.expiration_time = txn_lock_info.expiration_time;
+      } else {
+        // Check if it's expired. Skips over txn_lock_info.txn_ids[0] in case
+        // it's there for a shared lock with multiple holders which was not
+        // caught in the first case.
+        if (IsLockExpired(txn_lock_info.txn_ids[0], lock_info, env,
+                          expire_time)) {
+          // lock is expired, can steal it
+          lock_info.txn_ids = txn_lock_info.txn_ids;
+          lock_info.exclusive = txn_lock_info.exclusive;
+          lock_info.expiration_time = txn_lock_info.expiration_time;
+          // lock_cnt does not change
+        } else {
+          result = Status::TimedOut(Status::SubCode::kLockTimeout);
+          *txn_ids = lock_info.txn_ids;
+        }
+      }
+    } else {
+      // We are requesting shared access to a shared lock, so just grant it.
+      lock_info.txn_ids.push_back(txn_lock_info.txn_ids[0]);
+      // Using std::max means that expiration time never goes down even when
+      // a transaction is removed from the list. The correct solution would be
+      // to track expiry for every transaction, but this would also work for
+      // now.
+      lock_info.expiration_time =
+          std::max(lock_info.expiration_time, txn_lock_info.expiration_time);
+    }
+  } else {  // Lock not held.
+    // Check lock limit
+    if (max_num_locks_ > 0 &&
+        lock_map->lock_cnt.load(std::memory_order_acquire) >= max_num_locks_) {
+      result = Status::Busy(Status::SubCode::kLockLimit);
+    } else {
+      // acquire lock
+      stripe->keys.emplace(key, txn_lock_info);
+
+      // Maintain lock count if there is a limit on the number of locks
+      if (max_num_locks_) {
+        lock_map->lock_cnt++;
+      }
+    }
+  }
+
+  return result;
+}
+
+void PointLockManager::UnLockKey(PessimisticTransaction* txn,
+                                 const std::string& key, LockMapStripe* stripe,
+                                 LockMap* lock_map, Env* env) {
+#ifdef NDEBUG
+  (void)env;
+#endif
+  TransactionID txn_id = txn->GetID();
+
+  auto stripe_iter = stripe->keys.find(key);
+  if (stripe_iter != stripe->keys.end()) {
+    auto& txns = stripe_iter->second.txn_ids;
+    auto txn_it = std::find(txns.begin(), txns.end(), txn_id);
+    // Found the key we locked.  unlock it.
+    if (txn_it != txns.end()) {
+      if (txns.size() == 1) {
+        stripe->keys.erase(stripe_iter);
+      } else {
+        auto last_it = txns.end() - 1;
+        if (txn_it != last_it) {
+          *txn_it = *last_it;
+        }
+        txns.pop_back();
+      }
+
+      if (max_num_locks_ > 0) {
+        // Maintain lock count if there is a limit on the number of locks.
+        assert(lock_map->lock_cnt.load(std::memory_order_relaxed) > 0);
+        lock_map->lock_cnt--;
+      }
+    }
+  } else {
+    // This key is either not locked or locked by someone else.  This should
+    // only happen if the unlocking transaction has expired.
+    assert(txn->GetExpirationTime() > 0 &&
+           txn->GetExpirationTime() < env->NowMicros());
+  }
+}
+
+void PointLockManager::UnLock(PessimisticTransaction* txn,
+                              ColumnFamilyId column_family_id,
+                              const std::string& key, Env* env) {
+  std::shared_ptr<LockMap> lock_map_ptr = GetLockMap(column_family_id);
+  LockMap* lock_map = lock_map_ptr.get();
+  if (lock_map == nullptr) {
+    // Column Family must have been dropped.
+    return;
+  }
+
+  // Lock the mutex for the stripe that this key hashes to
+  size_t stripe_num = lock_map->GetStripe(key);
+  assert(lock_map->lock_map_stripes_.size() > stripe_num);
+  LockMapStripe* stripe = lock_map->lock_map_stripes_.at(stripe_num);
+
+  stripe->stripe_mutex->Lock().PermitUncheckedError();
+  UnLockKey(txn, key, stripe, lock_map, env);
+  stripe->stripe_mutex->UnLock();
+
+  // Signal waiting threads to retry locking
+  stripe->stripe_cv->NotifyAll();
+}
+
+void PointLockManager::UnLock(PessimisticTransaction* txn,
+                              const LockTracker& tracker, Env* env) {
+  std::unique_ptr<LockTracker::ColumnFamilyIterator> cf_it(
+      tracker.GetColumnFamilyIterator());
+  assert(cf_it != nullptr);
+  while (cf_it->HasNext()) {
+    ColumnFamilyId cf = cf_it->Next();
+    std::shared_ptr<LockMap> lock_map_ptr = GetLockMap(cf);
+    LockMap* lock_map = lock_map_ptr.get();
+    if (!lock_map) {
+      // Column Family must have been dropped.
+      return;
+    }
+
+    // Bucket keys by lock_map_ stripe
+    UnorderedMap<size_t, std::vector<const std::string*>> keys_by_stripe(
+        lock_map->num_stripes_);
+    std::unique_ptr<LockTracker::KeyIterator> key_it(
+        tracker.GetKeyIterator(cf));
+    assert(key_it != nullptr);
+    while (key_it->HasNext()) {
+      const std::string& key = key_it->Next();
+      size_t stripe_num = lock_map->GetStripe(key);
+      keys_by_stripe[stripe_num].push_back(&key);
+    }
+
+    // For each stripe, grab the stripe mutex and unlock all keys in this stripe
+    for (auto& stripe_iter : keys_by_stripe) {
+      size_t stripe_num = stripe_iter.first;
+      auto& stripe_keys = stripe_iter.second;
+
+      assert(lock_map->lock_map_stripes_.size() > stripe_num);
+      LockMapStripe* stripe = lock_map->lock_map_stripes_.at(stripe_num);
+
+      stripe->stripe_mutex->Lock().PermitUncheckedError();
+
+      for (const std::string* key : stripe_keys) {
+        UnLockKey(txn, *key, stripe, lock_map, env);
+      }
+
+      stripe->stripe_mutex->UnLock();
+
+      // Signal waiting threads to retry locking
+      stripe->stripe_cv->NotifyAll();
+    }
+  }
+}
+
+PointLockManager::PointLockStatus PointLockManager::GetPointLockStatus() {
+  PointLockStatus data;
+  // Lock order here is important. The correct order is lock_map_mutex_, then
+  // for every column family ID in ascending order lock every stripe in
+  // ascending order.
+  InstrumentedMutexLock l(&lock_map_mutex_);
+
+  std::vector<uint32_t> cf_ids;
+  for (const auto& map : lock_maps_) {
+    cf_ids.push_back(map.first);
+  }
+  std::sort(cf_ids.begin(), cf_ids.end());
+
+  for (auto i : cf_ids) {
+    const auto& stripes = lock_maps_[i]->lock_map_stripes_;
+    // Iterate and lock all stripes in ascending order.
+    for (const auto& j : stripes) {
+      j->stripe_mutex->Lock().PermitUncheckedError();
+      for (const auto& it : j->keys) {
+        struct KeyLockInfo info;
+        info.exclusive = it.second.exclusive;
+        info.key = it.first;
+        for (const auto& id : it.second.txn_ids) {
+          info.ids.push_back(id);
+        }
+        data.insert({i, info});
+      }
+    }
+  }
+
+  // Unlock everything. Unlocking order is not important.
+  for (auto i : cf_ids) {
+    const auto& stripes = lock_maps_[i]->lock_map_stripes_;
+    for (const auto& j : stripes) {
+      j->stripe_mutex->UnLock();
+    }
+  }
+
+  return data;
+}
+
+std::vector<DeadlockPath> PointLockManager::GetDeadlockInfoBuffer() {
+  return dlock_buffer_.PrepareBuffer();
+}
+
+void PointLockManager::Resize(uint32_t target_size) {
+  dlock_buffer_.Resize(target_size);
+}
+
+PointLockManager::RangeLockStatus PointLockManager::GetRangeLockStatus() {
+  return {};
+}
+
+Status PointLockManager::TryLock(PessimisticTransaction* /* txn */,
+                                 ColumnFamilyId /* cf_id */,
+                                 const Endpoint& /* start */,
+                                 const Endpoint& /* end */, Env* /* env */,
+                                 bool /* exclusive */) {
+  return Status::NotSupported(
+      "PointLockManager does not support range locking");
+}
+
+void PointLockManager::UnLock(PessimisticTransaction* /* txn */,
+                              ColumnFamilyId /* cf_id */,
+                              const Endpoint& /* start */,
+                              const Endpoint& /* end */, Env* /* env */) {
+  // no-op
+}
+
+}  // namespace ROCKSDB_NAMESPACE
+#endif  // ROCKSDB_LITE
diff --git a/src/rocksdb/utilities/transactions/lock/point/point_lock_manager.h b/src/rocksdb/utilities/transactions/lock/point/point_lock_manager.h
new file mode 100644
index 000000000..eeb34f3be
--- /dev/null
+++ b/src/rocksdb/utilities/transactions/lock/point/point_lock_manager.h
@@ -0,0 +1,224 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#pragma once
+#ifndef ROCKSDB_LITE
+
+#include <memory>
+#include <string>
+#include <unordered_map>
+#include <utility>
+#include <vector>
+
+#include "monitoring/instrumented_mutex.h"
+#include "rocksdb/utilities/transaction.h"
+#include "util/autovector.h"
+#include "util/hash_containers.h"
+#include "util/hash_map.h"
+#include "util/thread_local.h"
+#include "utilities/transactions/lock/lock_manager.h"
+#include "utilities/transactions/lock/point/point_lock_tracker.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+class ColumnFamilyHandle;
+struct LockInfo;
+struct LockMap;
+struct LockMapStripe;
+
+template <class Path>
+class DeadlockInfoBufferTempl {
+ private:
+  std::vector<Path> paths_buffer_;
+  uint32_t buffer_idx_;
+  std::mutex paths_buffer_mutex_;
+
+  std::vector<Path> Normalize() {
+    auto working = paths_buffer_;
+
+    if (working.empty()) {
+      return working;
+    }
+
+    // Next write occurs at a nonexistent path's slot
+    if (paths_buffer_[buffer_idx_].empty()) {
+      working.resize(buffer_idx_);
+    } else {
+      std::rotate(working.begin(), working.begin() + buffer_idx_,
+                  working.end());
+    }
+
+    return working;
+  }
+
+ public:
+  explicit DeadlockInfoBufferTempl(uint32_t n_latest_dlocks)
+      : paths_buffer_(n_latest_dlocks), buffer_idx_(0) {}
+
+  void AddNewPath(Path path) {
+    std::lock_guard<std::mutex> lock(paths_buffer_mutex_);
+
+    if (paths_buffer_.empty()) {
+      return;
+    }
+
+    paths_buffer_[buffer_idx_] = std::move(path);
+    buffer_idx_ = (buffer_idx_ + 1) % paths_buffer_.size();
+  }
+
+  void Resize(uint32_t target_size) {
+    std::lock_guard<std::mutex> lock(paths_buffer_mutex_);
+
+    paths_buffer_ = Normalize();
+
+    // Drop the deadlocks that will no longer be needed ater the normalize
+    if (target_size < paths_buffer_.size()) {
+      paths_buffer_.erase(
+          paths_buffer_.begin(),
+          paths_buffer_.begin() + (paths_buffer_.size() - target_size));
+      buffer_idx_ = 0;
+    }
+    // Resize the buffer to the target size and restore the buffer's idx
+    else {
+      auto prev_size = paths_buffer_.size();
+      paths_buffer_.resize(target_size);
+      buffer_idx_ = (uint32_t)prev_size;
+    }
+  }
+
+  std::vector<Path> PrepareBuffer() {
+    std::lock_guard<std::mutex> lock(paths_buffer_mutex_);
+
+    // Reversing the normalized vector returns the latest deadlocks first
+    auto working = Normalize();
+    std::reverse(working.begin(), working.end());
+
+    return working;
+  }
+};
+
+using DeadlockInfoBuffer = DeadlockInfoBufferTempl<DeadlockPath>;
+
+struct TrackedTrxInfo {
+  autovector<TransactionID> m_neighbors;
+  uint32_t m_cf_id;
+  bool m_exclusive;
+  std::string m_waiting_key;
+};
+
+class PointLockManager : public LockManager {
+ public:
+  PointLockManager(PessimisticTransactionDB* db,
+                   const TransactionDBOptions& opt);
+  // No copying allowed
+  PointLockManager(const PointLockManager&) = delete;
+  PointLockManager& operator=(const PointLockManager&) = delete;
+
+  ~PointLockManager() override {}
+
+  bool IsPointLockSupported() const override { return true; }
+
+  bool IsRangeLockSupported() const override { return false; }
+
+  const LockTrackerFactory& GetLockTrackerFactory() const override {
+    return PointLockTrackerFactory::Get();
+  }
+
+  // Creates a new LockMap for this column family.  Caller should guarantee
+  // that this column family does not already exist.
+  void AddColumnFamily(const ColumnFamilyHandle* cf) override;
+  // Deletes the LockMap for this column family.  Caller should guarantee that
+  // this column family is no longer in use.
+  void RemoveColumnFamily(const ColumnFamilyHandle* cf) override;
+
+  Status TryLock(PessimisticTransaction* txn, ColumnFamilyId column_family_id,
+                 const std::string& key, Env* env, bool exclusive) override;
+  Status TryLock(PessimisticTransaction* txn, ColumnFamilyId column_family_id,
+                 const Endpoint& start, const Endpoint& end, Env* env,
+                 bool exclusive) override;
+
+  void UnLock(PessimisticTransaction* txn, const LockTracker& tracker,
+              Env* env) override;
+  void UnLock(PessimisticTransaction* txn, ColumnFamilyId column_family_id,
+              const std::string& key, Env* env) override;
+  void UnLock(PessimisticTransaction* txn, ColumnFamilyId column_family_id,
+              const Endpoint& start, const Endpoint& end, Env* env) override;
+
+  PointLockStatus GetPointLockStatus() override;
+
+  RangeLockStatus GetRangeLockStatus() override;
+
+  std::vector<DeadlockPath> GetDeadlockInfoBuffer() override;
+
+  void Resize(uint32_t new_size) override;
+
+ private:
+  PessimisticTransactionDB* txn_db_impl_;
+
+  // Default number of lock map stripes per column family
+  const size_t default_num_stripes_;
+
+  // Limit on number of keys locked per column family
+  const int64_t max_num_locks_;
+
+  // The following lock order must be satisfied in order to avoid deadlocking
+  // ourselves.
+  //   - lock_map_mutex_
+  //   - stripe mutexes in ascending cf id, ascending stripe order
+  //   - wait_txn_map_mutex_
+  //
+  // Must be held when accessing/modifying lock_maps_.
+  InstrumentedMutex lock_map_mutex_;
+
+  // Map of ColumnFamilyId to locked key info
+  using LockMaps = UnorderedMap<uint32_t, std::shared_ptr<LockMap>>;
+  LockMaps lock_maps_;
+
+  // Thread-local cache of entries in lock_maps_.  This is an optimization
+  // to avoid acquiring a mutex in order to look up a LockMap
+  std::unique_ptr<ThreadLocalPtr> lock_maps_cache_;
+
+  // Must be held when modifying wait_txn_map_ and rev_wait_txn_map_.
+  std::mutex wait_txn_map_mutex_;
+
+  // Maps from waitee -> number of waiters.
+  HashMap<TransactionID, int> rev_wait_txn_map_;
+  // Maps from waiter -> waitee.
+  HashMap<TransactionID, TrackedTrxInfo> wait_txn_map_;
+  DeadlockInfoBuffer dlock_buffer_;
+
+  // Used to allocate mutexes/condvars to use when locking keys
+  std::shared_ptr<TransactionDBMutexFactory> mutex_factory_;
+
+  bool IsLockExpired(TransactionID txn_id, const LockInfo& lock_info, Env* env,
+                     uint64_t* wait_time);
+
+  std::shared_ptr<LockMap> GetLockMap(uint32_t column_family_id);
+
+  Status AcquireWithTimeout(PessimisticTransaction* txn, LockMap* lock_map,
+                            LockMapStripe* stripe, uint32_t column_family_id,
+                            const std::string& key, Env* env, int64_t timeout,
+                            const LockInfo& lock_info);
+
+  Status AcquireLocked(LockMap* lock_map, LockMapStripe* stripe,
+                       const std::string& key, Env* env,
+                       const LockInfo& lock_info, uint64_t* wait_time,
+                       autovector<TransactionID>* txn_ids);
+
+  void UnLockKey(PessimisticTransaction* txn, const std::string& key,
+                 LockMapStripe* stripe, LockMap* lock_map, Env* env);
+
+  bool IncrementWaiters(const PessimisticTransaction* txn,
+                        const autovector<TransactionID>& wait_ids,
+                        const std::string& key, const uint32_t& cf_id,
+                        const bool& exclusive, Env* const env);
+  void DecrementWaiters(const PessimisticTransaction* txn,
+                        const autovector<TransactionID>& wait_ids);
+  void DecrementWaitersImpl(const PessimisticTransaction* txn,
+                            const autovector<TransactionID>& wait_ids);
+};
+
+}  // namespace ROCKSDB_NAMESPACE
+#endif  // ROCKSDB_LITE
diff --git a/src/rocksdb/utilities/transactions/lock/point/point_lock_manager_test.cc b/src/rocksdb/utilities/transactions/lock/point/point_lock_manager_test.cc
new file mode 100644
index 000000000..525fdea71
--- /dev/null
+++ b/src/rocksdb/utilities/transactions/lock/point/point_lock_manager_test.cc
@@ -0,0 +1,181 @@
+//  Copyright (c) 2020-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#ifndef ROCKSDB_LITE
+
+#include "utilities/transactions/lock/point/point_lock_manager_test.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+// This test is not applicable for Range Lock manager as Range Lock Manager
+// operates on Column Families, not their ids.
+TEST_F(PointLockManagerTest, LockNonExistingColumnFamily) {
+  MockColumnFamilyHandle cf(1024);
+  locker_->RemoveColumnFamily(&cf);
+  auto txn = NewTxn();
+  auto s = locker_->TryLock(txn, 1024, "k", env_, true);
+  ASSERT_TRUE(s.IsInvalidArgument());
+  ASSERT_STREQ(s.getState(), "Column family id not found: 1024");
+  delete txn;
+}
+
+TEST_F(PointLockManagerTest, LockStatus) {
+  MockColumnFamilyHandle cf1(1024), cf2(2048);
+  locker_->AddColumnFamily(&cf1);
+  locker_->AddColumnFamily(&cf2);
+
+  auto txn1 = NewTxn();
+  ASSERT_OK(locker_->TryLock(txn1, 1024, "k1", env_, true));
+  ASSERT_OK(locker_->TryLock(txn1, 2048, "k1", env_, true));
+
+  auto txn2 = NewTxn();
+  ASSERT_OK(locker_->TryLock(txn2, 1024, "k2", env_, false));
+  ASSERT_OK(locker_->TryLock(txn2, 2048, "k2", env_, false));
+
+  auto s = locker_->GetPointLockStatus();
+  ASSERT_EQ(s.size(), 4u);
+  for (uint32_t cf_id : {1024, 2048}) {
+    ASSERT_EQ(s.count(cf_id), 2u);
+    auto range = s.equal_range(cf_id);
+    for (auto it = range.first; it != range.second; it++) {
+      ASSERT_TRUE(it->second.key == "k1" || it->second.key == "k2");
+      if (it->second.key == "k1") {
+        ASSERT_EQ(it->second.exclusive, true);
+        ASSERT_EQ(it->second.ids.size(), 1u);
+        ASSERT_EQ(it->second.ids[0], txn1->GetID());
+      } else if (it->second.key == "k2") {
+        ASSERT_EQ(it->second.exclusive, false);
+        ASSERT_EQ(it->second.ids.size(), 1u);
+        ASSERT_EQ(it->second.ids[0], txn2->GetID());
+      }
+    }
+  }
+
+  // Cleanup
+  locker_->UnLock(txn1, 1024, "k1", env_);
+  locker_->UnLock(txn1, 2048, "k1", env_);
+  locker_->UnLock(txn2, 1024, "k2", env_);
+  locker_->UnLock(txn2, 2048, "k2", env_);
+
+  delete txn1;
+  delete txn2;
+}
+
+TEST_F(PointLockManagerTest, UnlockExclusive) {
+  MockColumnFamilyHandle cf(1);
+  locker_->AddColumnFamily(&cf);
+
+  auto txn1 = NewTxn();
+  ASSERT_OK(locker_->TryLock(txn1, 1, "k", env_, true));
+  locker_->UnLock(txn1, 1, "k", env_);
+
+  auto txn2 = NewTxn();
+  ASSERT_OK(locker_->TryLock(txn2, 1, "k", env_, true));
+
+  // Cleanup
+  locker_->UnLock(txn2, 1, "k", env_);
+
+  delete txn1;
+  delete txn2;
+}
+
+TEST_F(PointLockManagerTest, UnlockShared) {
+  MockColumnFamilyHandle cf(1);
+  locker_->AddColumnFamily(&cf);
+
+  auto txn1 = NewTxn();
+  ASSERT_OK(locker_->TryLock(txn1, 1, "k", env_, false));
+  locker_->UnLock(txn1, 1, "k", env_);
+
+  auto txn2 = NewTxn();
+  ASSERT_OK(locker_->TryLock(txn2, 1, "k", env_, true));
+
+  // Cleanup
+  locker_->UnLock(txn2, 1, "k", env_);
+
+  delete txn1;
+  delete txn2;
+}
+
+// This test doesn't work with Range Lock Manager, because Range Lock Manager
+// doesn't support deadlock_detect_depth.
+
+TEST_F(PointLockManagerTest, DeadlockDepthExceeded) {
+  // Tests that when detecting deadlock, if the detection depth is exceeded,
+  // it's also viewed as deadlock.
+  MockColumnFamilyHandle cf(1);
+  locker_->AddColumnFamily(&cf);
+  TransactionOptions txn_opt;
+  txn_opt.deadlock_detect = true;
+  txn_opt.deadlock_detect_depth = 1;
+  txn_opt.lock_timeout = 1000000;
+  auto txn1 = NewTxn(txn_opt);
+  auto txn2 = NewTxn(txn_opt);
+  auto txn3 = NewTxn(txn_opt);
+  auto txn4 = NewTxn(txn_opt);
+  // "a ->(k) b" means transaction a is waiting for transaction b to release
+  // the held lock on key k.
+  // txn4 ->(k3) -> txn3 ->(k2) txn2 ->(k1) txn1
+  // txn3's deadlock detection will exceed the detection depth 1,
+  // which will be viewed as a deadlock.
+  // NOTE:
+  // txn4 ->(k3) -> txn3 must be set up before
+  // txn3 ->(k2) -> txn2, because to trigger deadlock detection for txn3,
+  // it must have another txn waiting on it, which is txn4 in this case.
+  ASSERT_OK(locker_->TryLock(txn1, 1, "k1", env_, true));
+
+  port::Thread t1 = BlockUntilWaitingTxn(wait_sync_point_name_, [&]() {
+    ASSERT_OK(locker_->TryLock(txn2, 1, "k2", env_, true));
+    // block because txn1 is holding a lock on k1.
+    locker_->TryLock(txn2, 1, "k1", env_, true);
+  });
+
+  ASSERT_OK(locker_->TryLock(txn3, 1, "k3", env_, true));
+
+  port::Thread t2 = BlockUntilWaitingTxn(wait_sync_point_name_, [&]() {
+    // block because txn3 is holding a lock on k1.
+    locker_->TryLock(txn4, 1, "k3", env_, true);
+  });
+
+  auto s = locker_->TryLock(txn3, 1, "k2", env_, true);
+  ASSERT_TRUE(s.IsBusy());
+  ASSERT_EQ(s.subcode(), Status::SubCode::kDeadlock);
+
+  std::vector<DeadlockPath> deadlock_paths = locker_->GetDeadlockInfoBuffer();
+  ASSERT_EQ(deadlock_paths.size(), 1u);
+  ASSERT_TRUE(deadlock_paths[0].limit_exceeded);
+
+  locker_->UnLock(txn1, 1, "k1", env_);
+  locker_->UnLock(txn3, 1, "k3", env_);
+  t1.join();
+  t2.join();
+
+  delete txn4;
+  delete txn3;
+  delete txn2;
+  delete txn1;
+}
+
+INSTANTIATE_TEST_CASE_P(PointLockManager, AnyLockManagerTest,
+                        ::testing::Values(nullptr));
+
+}  // namespace ROCKSDB_NAMESPACE
+
+int main(int argc, char** argv) {
+  ROCKSDB_NAMESPACE::port::InstallStackTraceHandler();
+  ::testing::InitGoogleTest(&argc, argv);
+  return RUN_ALL_TESTS();
+}
+
+#else
+#include <stdio.h>
+
+int main(int /*argc*/, char** /*argv*/) {
+  fprintf(stderr,
+          "SKIPPED because Transactions are not supported in ROCKSDB_LITE\n");
+  return 0;
+}
+
+#endif  // ROCKSDB_LITE
diff --git a/src/rocksdb/utilities/transactions/lock/point/point_lock_manager_test.h b/src/rocksdb/utilities/transactions/lock/point/point_lock_manager_test.h
new file mode 100644
index 000000000..ca9f46bf9
--- /dev/null
+++ b/src/rocksdb/utilities/transactions/lock/point/point_lock_manager_test.h
@@ -0,0 +1,324 @@
+//  Copyright (c) Meta Platforms, Inc. and affiliates.
+//
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#include "file/file_util.h"
+#include "port/port.h"
+#include "port/stack_trace.h"
+#include "rocksdb/utilities/transaction_db.h"
+#include "test_util/testharness.h"
+#include "test_util/testutil.h"
+#include "utilities/transactions/lock/point/point_lock_manager.h"
+#include "utilities/transactions/pessimistic_transaction_db.h"
+#include "utilities/transactions/transaction_db_mutex_impl.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+class MockColumnFamilyHandle : public ColumnFamilyHandle {
+ public:
+  explicit MockColumnFamilyHandle(ColumnFamilyId cf_id) : cf_id_(cf_id) {}
+
+  ~MockColumnFamilyHandle() override {}
+
+  const std::string& GetName() const override { return name_; }
+
+  ColumnFamilyId GetID() const override { return cf_id_; }
+
+  Status GetDescriptor(ColumnFamilyDescriptor*) override {
+    return Status::OK();
+  }
+
+  const Comparator* GetComparator() const override {
+    return BytewiseComparator();
+  }
+
+ private:
+  ColumnFamilyId cf_id_;
+  std::string name_ = "MockCF";
+};
+
+class PointLockManagerTest : public testing::Test {
+ public:
+  void SetUp() override {
+    env_ = Env::Default();
+    db_dir_ = test::PerThreadDBPath("point_lock_manager_test");
+    ASSERT_OK(env_->CreateDir(db_dir_));
+
+    Options opt;
+    opt.create_if_missing = true;
+    TransactionDBOptions txn_opt;
+    txn_opt.transaction_lock_timeout = 0;
+
+    ASSERT_OK(TransactionDB::Open(opt, txn_opt, db_dir_, &db_));
+
+    // CAUTION: This test creates a separate lock manager object (right, NOT
+    // the one that the TransactionDB is using!), and runs tests on it.
+    locker_.reset(new PointLockManager(
+        static_cast<PessimisticTransactionDB*>(db_), txn_opt));
+
+    wait_sync_point_name_ = "PointLockManager::AcquireWithTimeout:WaitingTxn";
+  }
+
+  void TearDown() override {
+    delete db_;
+    EXPECT_OK(DestroyDir(env_, db_dir_));
+  }
+
+  PessimisticTransaction* NewTxn(
+      TransactionOptions txn_opt = TransactionOptions()) {
+    Transaction* txn = db_->BeginTransaction(WriteOptions(), txn_opt);
+    return reinterpret_cast<PessimisticTransaction*>(txn);
+  }
+
+ protected:
+  Env* env_;
+  std::shared_ptr<LockManager> locker_;
+  const char* wait_sync_point_name_;
+  friend void PointLockManagerTestExternalSetup(PointLockManagerTest*);
+
+ private:
+  std::string db_dir_;
+  TransactionDB* db_;
+};
+
+using init_func_t = void (*)(PointLockManagerTest*);
+
+class AnyLockManagerTest : public PointLockManagerTest,
+                           public testing::WithParamInterface<init_func_t> {
+ public:
+  void SetUp() override {
+    // If a custom setup function was provided, use it. Otherwise, use what we
+    // have inherited.
+    auto init_func = GetParam();
+    if (init_func)
+      (*init_func)(this);
+    else
+      PointLockManagerTest::SetUp();
+  }
+};
+
+TEST_P(AnyLockManagerTest, ReentrantExclusiveLock) {
+  // Tests that a txn can acquire exclusive lock on the same key repeatedly.
+  MockColumnFamilyHandle cf(1);
+  locker_->AddColumnFamily(&cf);
+  auto txn = NewTxn();
+  ASSERT_OK(locker_->TryLock(txn, 1, "k", env_, true));
+  ASSERT_OK(locker_->TryLock(txn, 1, "k", env_, true));
+
+  // Cleanup
+  locker_->UnLock(txn, 1, "k", env_);
+
+  delete txn;
+}
+
+TEST_P(AnyLockManagerTest, ReentrantSharedLock) {
+  // Tests that a txn can acquire shared lock on the same key repeatedly.
+  MockColumnFamilyHandle cf(1);
+  locker_->AddColumnFamily(&cf);
+  auto txn = NewTxn();
+  ASSERT_OK(locker_->TryLock(txn, 1, "k", env_, false));
+  ASSERT_OK(locker_->TryLock(txn, 1, "k", env_, false));
+
+  // Cleanup
+  locker_->UnLock(txn, 1, "k", env_);
+
+  delete txn;
+}
+
+TEST_P(AnyLockManagerTest, LockUpgrade) {
+  // Tests that a txn can upgrade from a shared lock to an exclusive lock.
+  MockColumnFamilyHandle cf(1);
+  locker_->AddColumnFamily(&cf);
+  auto txn = NewTxn();
+  ASSERT_OK(locker_->TryLock(txn, 1, "k", env_, false));
+  ASSERT_OK(locker_->TryLock(txn, 1, "k", env_, true));
+
+  // Cleanup
+  locker_->UnLock(txn, 1, "k", env_);
+  delete txn;
+}
+
+TEST_P(AnyLockManagerTest, LockDowngrade) {
+  // Tests that a txn can acquire a shared lock after acquiring an exclusive
+  // lock on the same key.
+  MockColumnFamilyHandle cf(1);
+  locker_->AddColumnFamily(&cf);
+  auto txn = NewTxn();
+  ASSERT_OK(locker_->TryLock(txn, 1, "k", env_, true));
+  ASSERT_OK(locker_->TryLock(txn, 1, "k", env_, false));
+
+  // Cleanup
+  locker_->UnLock(txn, 1, "k", env_);
+  delete txn;
+}
+
+TEST_P(AnyLockManagerTest, LockConflict) {
+  // Tests that lock conflicts lead to lock timeout.
+  MockColumnFamilyHandle cf(1);
+  locker_->AddColumnFamily(&cf);
+  auto txn1 = NewTxn();
+  auto txn2 = NewTxn();
+
+  {
+    // exclusive-exclusive conflict.
+    ASSERT_OK(locker_->TryLock(txn1, 1, "k1", env_, true));
+    auto s = locker_->TryLock(txn2, 1, "k1", env_, true);
+    ASSERT_TRUE(s.IsTimedOut());
+  }
+
+  {
+    // exclusive-shared conflict.
+    ASSERT_OK(locker_->TryLock(txn1, 1, "k2", env_, true));
+    auto s = locker_->TryLock(txn2, 1, "k2", env_, false);
+    ASSERT_TRUE(s.IsTimedOut());
+  }
+
+  {
+    // shared-exclusive conflict.
+    ASSERT_OK(locker_->TryLock(txn1, 1, "k2", env_, false));
+    auto s = locker_->TryLock(txn2, 1, "k2", env_, true);
+    ASSERT_TRUE(s.IsTimedOut());
+  }
+
+  // Cleanup
+  locker_->UnLock(txn1, 1, "k1", env_);
+  locker_->UnLock(txn1, 1, "k2", env_);
+
+  delete txn1;
+  delete txn2;
+}
+
+port::Thread BlockUntilWaitingTxn(const char* sync_point_name,
+                                  std::function<void()> f) {
+  std::atomic<bool> reached(false);
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
+      sync_point_name, [&](void* /*arg*/) { reached.store(true); });
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->EnableProcessing();
+
+  port::Thread t(f);
+
+  while (!reached.load()) {
+    std::this_thread::sleep_for(std::chrono::milliseconds(100));
+  }
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->DisableProcessing();
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->ClearAllCallBacks();
+
+  return t;
+}
+
+TEST_P(AnyLockManagerTest, SharedLocks) {
+  // Tests that shared locks can be concurrently held by multiple transactions.
+  MockColumnFamilyHandle cf(1);
+  locker_->AddColumnFamily(&cf);
+  auto txn1 = NewTxn();
+  auto txn2 = NewTxn();
+  ASSERT_OK(locker_->TryLock(txn1, 1, "k", env_, false));
+  ASSERT_OK(locker_->TryLock(txn2, 1, "k", env_, false));
+
+  // Cleanup
+  locker_->UnLock(txn1, 1, "k", env_);
+  locker_->UnLock(txn2, 1, "k", env_);
+
+  delete txn1;
+  delete txn2;
+}
+
+TEST_P(AnyLockManagerTest, Deadlock) {
+  // Tests that deadlock can be detected.
+  // Deadlock scenario:
+  // txn1 exclusively locks k1, and wants to lock k2;
+  // txn2 exclusively locks k2, and wants to lock k1.
+  MockColumnFamilyHandle cf(1);
+  locker_->AddColumnFamily(&cf);
+  TransactionOptions txn_opt;
+  txn_opt.deadlock_detect = true;
+  txn_opt.lock_timeout = 1000000;
+  auto txn1 = NewTxn(txn_opt);
+  auto txn2 = NewTxn(txn_opt);
+
+  ASSERT_OK(locker_->TryLock(txn1, 1, "k1", env_, true));
+  ASSERT_OK(locker_->TryLock(txn2, 1, "k2", env_, true));
+
+  // txn1 tries to lock k2, will block forever.
+  port::Thread t = BlockUntilWaitingTxn(wait_sync_point_name_, [&]() {
+    // block because txn2 is holding a lock on k2.
+    locker_->TryLock(txn1, 1, "k2", env_, true);
+  });
+
+  auto s = locker_->TryLock(txn2, 1, "k1", env_, true);
+  ASSERT_TRUE(s.IsBusy());
+  ASSERT_EQ(s.subcode(), Status::SubCode::kDeadlock);
+
+  std::vector<DeadlockPath> deadlock_paths = locker_->GetDeadlockInfoBuffer();
+  ASSERT_EQ(deadlock_paths.size(), 1u);
+  ASSERT_FALSE(deadlock_paths[0].limit_exceeded);
+
+  std::vector<DeadlockInfo> deadlocks = deadlock_paths[0].path;
+  ASSERT_EQ(deadlocks.size(), 2u);
+
+  ASSERT_EQ(deadlocks[0].m_txn_id, txn1->GetID());
+  ASSERT_EQ(deadlocks[0].m_cf_id, 1u);
+  ASSERT_TRUE(deadlocks[0].m_exclusive);
+  ASSERT_EQ(deadlocks[0].m_waiting_key, "k2");
+
+  ASSERT_EQ(deadlocks[1].m_txn_id, txn2->GetID());
+  ASSERT_EQ(deadlocks[1].m_cf_id, 1u);
+  ASSERT_TRUE(deadlocks[1].m_exclusive);
+  ASSERT_EQ(deadlocks[1].m_waiting_key, "k1");
+
+  locker_->UnLock(txn2, 1, "k2", env_);
+  t.join();
+
+  // Cleanup
+  locker_->UnLock(txn1, 1, "k1", env_);
+  locker_->UnLock(txn1, 1, "k2", env_);
+  delete txn2;
+  delete txn1;
+}
+
+TEST_P(AnyLockManagerTest, GetWaitingTxns_MultipleTxns) {
+  MockColumnFamilyHandle cf(1);
+  locker_->AddColumnFamily(&cf);
+
+  auto txn1 = NewTxn();
+  ASSERT_OK(locker_->TryLock(txn1, 1, "k", env_, false));
+
+  auto txn2 = NewTxn();
+  ASSERT_OK(locker_->TryLock(txn2, 1, "k", env_, false));
+
+  auto txn3 = NewTxn();
+  txn3->SetLockTimeout(10000);
+  port::Thread t1 = BlockUntilWaitingTxn(wait_sync_point_name_, [&]() {
+    ASSERT_OK(locker_->TryLock(txn3, 1, "k", env_, true));
+    locker_->UnLock(txn3, 1, "k", env_);
+  });
+
+  // Ok, now txn3 is waiting for lock on "k", which is owned by two
+  // transactions. Check that GetWaitingTxns reports this correctly
+  uint32_t wait_cf_id;
+  std::string wait_key;
+  auto waiters = txn3->GetWaitingTxns(&wait_cf_id, &wait_key);
+
+  ASSERT_EQ(wait_cf_id, 1u);
+  ASSERT_EQ(wait_key, "k");
+  ASSERT_EQ(waiters.size(), 2);
+  bool waits_correct =
+      (waiters[0] == txn1->GetID() && waiters[1] == txn2->GetID()) ||
+      (waiters[1] == txn1->GetID() && waiters[0] == txn2->GetID());
+  ASSERT_EQ(waits_correct, true);
+
+  // Release locks so txn3 can proceed with execution
+  locker_->UnLock(txn1, 1, "k", env_);
+  locker_->UnLock(txn2, 1, "k", env_);
+
+  // Wait until txn3 finishes
+  t1.join();
+
+  delete txn1;
+  delete txn2;
+  delete txn3;
+}
+
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/utilities/transactions/lock/point/point_lock_tracker.cc b/src/rocksdb/utilities/transactions/lock/point/point_lock_tracker.cc
new file mode 100644
index 000000000..6204a8f02
--- /dev/null
+++ b/src/rocksdb/utilities/transactions/lock/point/point_lock_tracker.cc
@@ -0,0 +1,257 @@
+// Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+// This source code is licensed under both the GPLv2 (found in the
+// COPYING file in the root directory) and Apache 2.0 License
+// (found in the LICENSE.Apache file in the root directory).
+
+#ifndef ROCKSDB_LITE
+
+#include "utilities/transactions/lock/point/point_lock_tracker.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+namespace {
+
+class TrackedKeysColumnFamilyIterator
+    : public LockTracker::ColumnFamilyIterator {
+ public:
+  explicit TrackedKeysColumnFamilyIterator(const TrackedKeys& keys)
+      : tracked_keys_(keys), it_(keys.begin()) {}
+
+  bool HasNext() const override { return it_ != tracked_keys_.end(); }
+
+  ColumnFamilyId Next() override { return (it_++)->first; }
+
+ private:
+  const TrackedKeys& tracked_keys_;
+  TrackedKeys::const_iterator it_;
+};
+
+class TrackedKeysIterator : public LockTracker::KeyIterator {
+ public:
+  TrackedKeysIterator(const TrackedKeys& keys, ColumnFamilyId id)
+      : key_infos_(keys.at(id)), it_(key_infos_.begin()) {}
+
+  bool HasNext() const override { return it_ != key_infos_.end(); }
+
+  const std::string& Next() override { return (it_++)->first; }
+
+ private:
+  const TrackedKeyInfos& key_infos_;
+  TrackedKeyInfos::const_iterator it_;
+};
+
+}  // namespace
+
+void PointLockTracker::Track(const PointLockRequest& r) {
+  auto& keys = tracked_keys_[r.column_family_id];
+  auto result = keys.try_emplace(r.key, r.seq);
+  auto it = result.first;
+  if (!result.second && r.seq < it->second.seq) {
+    // Now tracking this key with an earlier sequence number
+    it->second.seq = r.seq;
+  }
+  // else we do not update the seq. The smaller the tracked seq, the stronger it
+  // the guarantee since it implies from the seq onward there has not been a
+  // concurrent update to the key. So we update the seq if it implies stronger
+  // guarantees, i.e., if it is smaller than the existing tracked seq.
+
+  if (r.read_only) {
+    it->second.num_reads++;
+  } else {
+    it->second.num_writes++;
+  }
+
+  it->second.exclusive = it->second.exclusive || r.exclusive;
+}
+
+UntrackStatus PointLockTracker::Untrack(const PointLockRequest& r) {
+  auto cf_keys = tracked_keys_.find(r.column_family_id);
+  if (cf_keys == tracked_keys_.end()) {
+    return UntrackStatus::NOT_TRACKED;
+  }
+
+  auto& keys = cf_keys->second;
+  auto it = keys.find(r.key);
+  if (it == keys.end()) {
+    return UntrackStatus::NOT_TRACKED;
+  }
+
+  bool untracked = false;
+  auto& info = it->second;
+  if (r.read_only) {
+    if (info.num_reads > 0) {
+      info.num_reads--;
+      untracked = true;
+    }
+  } else {
+    if (info.num_writes > 0) {
+      info.num_writes--;
+      untracked = true;
+    }
+  }
+
+  bool removed = false;
+  if (info.num_reads == 0 && info.num_writes == 0) {
+    keys.erase(it);
+    if (keys.empty()) {
+      tracked_keys_.erase(cf_keys);
+    }
+    removed = true;
+  }
+
+  if (removed) {
+    return UntrackStatus::REMOVED;
+  }
+  if (untracked) {
+    return UntrackStatus::UNTRACKED;
+  }
+  return UntrackStatus::NOT_TRACKED;
+}
+
+void PointLockTracker::Merge(const LockTracker& tracker) {
+  const PointLockTracker& t = static_cast<const PointLockTracker&>(tracker);
+  for (const auto& cf_keys : t.tracked_keys_) {
+    ColumnFamilyId cf = cf_keys.first;
+    const auto& keys = cf_keys.second;
+
+    auto current_cf_keys = tracked_keys_.find(cf);
+    if (current_cf_keys == tracked_keys_.end()) {
+      tracked_keys_.emplace(cf_keys);
+    } else {
+      auto& current_keys = current_cf_keys->second;
+      for (const auto& key_info : keys) {
+        const std::string& key = key_info.first;
+        const TrackedKeyInfo& info = key_info.second;
+        // If key was not previously tracked, just copy the whole struct over.
+        // Otherwise, some merging needs to occur.
+        auto current_info = current_keys.find(key);
+        if (current_info == current_keys.end()) {
+          current_keys.emplace(key_info);
+        } else {
+          current_info->second.Merge(info);
+        }
+      }
+    }
+  }
+}
+
+void PointLockTracker::Subtract(const LockTracker& tracker) {
+  const PointLockTracker& t = static_cast<const PointLockTracker&>(tracker);
+  for (const auto& cf_keys : t.tracked_keys_) {
+    ColumnFamilyId cf = cf_keys.first;
+    const auto& keys = cf_keys.second;
+
+    auto& current_keys = tracked_keys_.at(cf);
+    for (const auto& key_info : keys) {
+      const std::string& key = key_info.first;
+      const TrackedKeyInfo& info = key_info.second;
+      uint32_t num_reads = info.num_reads;
+      uint32_t num_writes = info.num_writes;
+
+      auto current_key_info = current_keys.find(key);
+      assert(current_key_info != current_keys.end());
+
+      // Decrement the total reads/writes of this key by the number of
+      // reads/writes done since the last SavePoint.
+      if (num_reads > 0) {
+        assert(current_key_info->second.num_reads >= num_reads);
+        current_key_info->second.num_reads -= num_reads;
+      }
+      if (num_writes > 0) {
+        assert(current_key_info->second.num_writes >= num_writes);
+        current_key_info->second.num_writes -= num_writes;
+      }
+      if (current_key_info->second.num_reads == 0 &&
+          current_key_info->second.num_writes == 0) {
+        current_keys.erase(current_key_info);
+      }
+    }
+  }
+}
+
+LockTracker* PointLockTracker::GetTrackedLocksSinceSavePoint(
+    const LockTracker& save_point_tracker) const {
+  // Examine the number of reads/writes performed on all keys written
+  // since the last SavePoint and compare to the total number of reads/writes
+  // for each key.
+  LockTracker* t = new PointLockTracker();
+  const PointLockTracker& save_point_t =
+      static_cast<const PointLockTracker&>(save_point_tracker);
+  for (const auto& cf_keys : save_point_t.tracked_keys_) {
+    ColumnFamilyId cf = cf_keys.first;
+    const auto& keys = cf_keys.second;
+
+    auto& current_keys = tracked_keys_.at(cf);
+    for (const auto& key_info : keys) {
+      const std::string& key = key_info.first;
+      const TrackedKeyInfo& info = key_info.second;
+      uint32_t num_reads = info.num_reads;
+      uint32_t num_writes = info.num_writes;
+
+      auto current_key_info = current_keys.find(key);
+      assert(current_key_info != current_keys.end());
+      assert(current_key_info->second.num_reads >= num_reads);
+      assert(current_key_info->second.num_writes >= num_writes);
+
+      if (current_key_info->second.num_reads == num_reads &&
+          current_key_info->second.num_writes == num_writes) {
+        // All the reads/writes to this key were done in the last savepoint.
+        PointLockRequest r;
+        r.column_family_id = cf;
+        r.key = key;
+        r.seq = info.seq;
+        r.read_only = (num_writes == 0);
+        r.exclusive = info.exclusive;
+        t->Track(r);
+      }
+    }
+  }
+  return t;
+}
+
+PointLockStatus PointLockTracker::GetPointLockStatus(
+    ColumnFamilyId column_family_id, const std::string& key) const {
+  assert(IsPointLockSupported());
+  PointLockStatus status;
+  auto it = tracked_keys_.find(column_family_id);
+  if (it == tracked_keys_.end()) {
+    return status;
+  }
+
+  const auto& keys = it->second;
+  auto key_it = keys.find(key);
+  if (key_it == keys.end()) {
+    return status;
+  }
+
+  const TrackedKeyInfo& key_info = key_it->second;
+  status.locked = true;
+  status.exclusive = key_info.exclusive;
+  status.seq = key_info.seq;
+  return status;
+}
+
+uint64_t PointLockTracker::GetNumPointLocks() const {
+  uint64_t num_keys = 0;
+  for (const auto& cf_keys : tracked_keys_) {
+    num_keys += cf_keys.second.size();
+  }
+  return num_keys;
+}
+
+LockTracker::ColumnFamilyIterator* PointLockTracker::GetColumnFamilyIterator()
+    const {
+  return new TrackedKeysColumnFamilyIterator(tracked_keys_);
+}
+
+LockTracker::KeyIterator* PointLockTracker::GetKeyIterator(
+    ColumnFamilyId column_family_id) const {
+  assert(tracked_keys_.find(column_family_id) != tracked_keys_.end());
+  return new TrackedKeysIterator(tracked_keys_, column_family_id);
+}
+
+void PointLockTracker::Clear() { tracked_keys_.clear(); }
+
+}  // namespace ROCKSDB_NAMESPACE
+
+#endif  // ROCKSDB_LITE
diff --git a/src/rocksdb/utilities/transactions/lock/point/point_lock_tracker.h b/src/rocksdb/utilities/transactions/lock/point/point_lock_tracker.h
new file mode 100644
index 000000000..daf6f9aa2
--- /dev/null
+++ b/src/rocksdb/utilities/transactions/lock/point/point_lock_tracker.h
@@ -0,0 +1,99 @@
+// Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+// This source code is licensed under both the GPLv2 (found in the
+// COPYING file in the root directory) and Apache 2.0 License
+// (found in the LICENSE.Apache file in the root directory).
+
+#pragma once
+#ifndef ROCKSDB_LITE
+
+#include <memory>
+#include <string>
+#include <unordered_map>
+
+#include "utilities/transactions/lock/lock_tracker.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+struct TrackedKeyInfo {
+  // Earliest sequence number that is relevant to this transaction for this key
+  SequenceNumber seq;
+
+  uint32_t num_writes;
+  uint32_t num_reads;
+
+  bool exclusive;
+
+  explicit TrackedKeyInfo(SequenceNumber seq_no)
+      : seq(seq_no), num_writes(0), num_reads(0), exclusive(false) {}
+
+  void Merge(const TrackedKeyInfo& info) {
+    assert(seq <= info.seq);
+    num_reads += info.num_reads;
+    num_writes += info.num_writes;
+    exclusive = exclusive || info.exclusive;
+  }
+};
+
+using TrackedKeyInfos = std::unordered_map<std::string, TrackedKeyInfo>;
+
+using TrackedKeys = std::unordered_map<ColumnFamilyId, TrackedKeyInfos>;
+
+// Tracks point locks on single keys.
+class PointLockTracker : public LockTracker {
+ public:
+  PointLockTracker() = default;
+
+  PointLockTracker(const PointLockTracker&) = delete;
+  PointLockTracker& operator=(const PointLockTracker&) = delete;
+
+  bool IsPointLockSupported() const override { return true; }
+
+  bool IsRangeLockSupported() const override { return false; }
+
+  void Track(const PointLockRequest& lock_request) override;
+
+  UntrackStatus Untrack(const PointLockRequest& lock_request) override;
+
+  void Track(const RangeLockRequest& /*lock_request*/) override {}
+
+  UntrackStatus Untrack(const RangeLockRequest& /*lock_request*/) override {
+    return UntrackStatus::NOT_TRACKED;
+  }
+
+  void Merge(const LockTracker& tracker) override;
+
+  void Subtract(const LockTracker& tracker) override;
+
+  void Clear() override;
+
+  virtual LockTracker* GetTrackedLocksSinceSavePoint(
+      const LockTracker& save_point_tracker) const override;
+
+  PointLockStatus GetPointLockStatus(ColumnFamilyId column_family_id,
+                                     const std::string& key) const override;
+
+  uint64_t GetNumPointLocks() const override;
+
+  ColumnFamilyIterator* GetColumnFamilyIterator() const override;
+
+  KeyIterator* GetKeyIterator(ColumnFamilyId column_family_id) const override;
+
+ private:
+  TrackedKeys tracked_keys_;
+};
+
+class PointLockTrackerFactory : public LockTrackerFactory {
+ public:
+  static const PointLockTrackerFactory& Get() {
+    static const PointLockTrackerFactory instance;
+    return instance;
+  }
+
+  LockTracker* Create() const override { return new PointLockTracker(); }
+
+ private:
+  PointLockTrackerFactory() {}
+};
+
+}  // namespace ROCKSDB_NAMESPACE
+#endif  // ROCKSDB_LITE
diff --git a/src/rocksdb/utilities/transactions/lock/range/range_lock_manager.h b/src/rocksdb/utilities/transactions/lock/range/range_lock_manager.h
new file mode 100644
index 000000000..01899542e
--- /dev/null
+++ b/src/rocksdb/utilities/transactions/lock/range/range_lock_manager.h
@@ -0,0 +1,36 @@
+//  Copyright (c) Meta Platforms, Inc. and affiliates.
+//
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+//
+// Generic definitions for a Range-based Lock Manager
+//
+#pragma once
+#ifndef ROCKSDB_LITE
+
+#include "utilities/transactions/lock/lock_manager.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+/*
+  A base class for all Range-based lock managers
+
+  See also class RangeLockManagerHandle in
+  include/rocksdb/utilities/transaction_db.h
+*/
+class RangeLockManagerBase : public LockManager {
+ public:
+  // Geting a point lock is reduced to getting a range lock on a single-point
+  // range
+  using LockManager::TryLock;
+  Status TryLock(PessimisticTransaction* txn, ColumnFamilyId column_family_id,
+                 const std::string& key, Env* env, bool exclusive) override {
+    Endpoint endp(key.data(), key.size(), false);
+    return TryLock(txn, column_family_id, endp, endp, env, exclusive);
+  }
+};
+
+}  // namespace ROCKSDB_NAMESPACE
+#endif  // ROCKSDB_LITE
diff --git a/src/rocksdb/utilities/transactions/lock/range/range_locking_test.cc b/src/rocksdb/utilities/transactions/lock/range/range_locking_test.cc
new file mode 100644
index 000000000..bce66c1f3
--- /dev/null
+++ b/src/rocksdb/utilities/transactions/lock/range/range_locking_test.cc
@@ -0,0 +1,459 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#ifndef ROCKSDB_LITE
+#ifndef OS_WIN
+
+#include <algorithm>
+#include <functional>
+#include <string>
+#include <thread>
+
+#include "db/db_impl/db_impl.h"
+#include "port/port.h"
+#include "rocksdb/db.h"
+#include "rocksdb/options.h"
+#include "rocksdb/perf_context.h"
+#include "rocksdb/utilities/transaction.h"
+#include "rocksdb/utilities/transaction_db.h"
+#include "utilities/transactions/lock/point/point_lock_manager_test.h"
+#include "utilities/transactions/pessimistic_transaction_db.h"
+#include "utilities/transactions/transaction_test.h"
+
+using std::string;
+
+namespace ROCKSDB_NAMESPACE {
+
+class RangeLockingTest : public ::testing::Test {
+ public:
+  TransactionDB* db;
+  std::string dbname;
+  Options options;
+
+  std::shared_ptr<RangeLockManagerHandle> range_lock_mgr;
+  TransactionDBOptions txn_db_options;
+
+  RangeLockingTest() : db(nullptr) {
+    options.create_if_missing = true;
+    dbname = test::PerThreadDBPath("range_locking_testdb");
+
+    EXPECT_OK(DestroyDB(dbname, options));
+
+    range_lock_mgr.reset(NewRangeLockManager(nullptr));
+    txn_db_options.lock_mgr_handle = range_lock_mgr;
+
+    auto s = TransactionDB::Open(options, txn_db_options, dbname, &db);
+    assert(s.ok());
+  }
+
+  ~RangeLockingTest() {
+    delete db;
+    db = nullptr;
+    // This is to skip the assert statement in FaultInjectionTestEnv. There
+    // seems to be a bug in btrfs that the makes readdir return recently
+    // unlink-ed files. By using the default fs we simply ignore errors resulted
+    // from attempting to delete such files in DestroyDB.
+    EXPECT_OK(DestroyDB(dbname, options));
+  }
+
+  PessimisticTransaction* NewTxn(
+      TransactionOptions txn_opt = TransactionOptions()) {
+    Transaction* txn = db->BeginTransaction(WriteOptions(), txn_opt);
+    return reinterpret_cast<PessimisticTransaction*>(txn);
+  }
+};
+
+// TODO: set a smaller lock wait timeout so that the test runs faster.
+TEST_F(RangeLockingTest, BasicRangeLocking) {
+  WriteOptions write_options;
+  TransactionOptions txn_options;
+  std::string value;
+  ReadOptions read_options;
+  auto cf = db->DefaultColumnFamily();
+
+  Transaction* txn0 = db->BeginTransaction(write_options, txn_options);
+  Transaction* txn1 = db->BeginTransaction(write_options, txn_options);
+
+  // Get a range lock
+  ASSERT_OK(txn0->GetRangeLock(cf, Endpoint("a"), Endpoint("c")));
+
+  // Check that range Lock inhibits an overlapping range lock
+  {
+    auto s = txn1->GetRangeLock(cf, Endpoint("b"), Endpoint("z"));
+    ASSERT_TRUE(s.IsTimedOut());
+  }
+
+  // Check that range Lock inhibits an overlapping point lock
+  {
+    auto s = txn1->GetForUpdate(read_options, cf, Slice("b"), &value);
+    ASSERT_TRUE(s.IsTimedOut());
+  }
+
+  // Get a point lock, check that it inhibits range locks
+  ASSERT_OK(txn0->Put(cf, Slice("n"), Slice("value")));
+  {
+    auto s = txn1->GetRangeLock(cf, Endpoint("m"), Endpoint("p"));
+    ASSERT_TRUE(s.IsTimedOut());
+  }
+
+  ASSERT_OK(txn0->Commit());
+  txn1->Rollback();
+
+  delete txn0;
+  delete txn1;
+}
+
+TEST_F(RangeLockingTest, MyRocksLikeUpdate) {
+  WriteOptions write_options;
+  TransactionOptions txn_options;
+  Transaction* txn0 = db->BeginTransaction(write_options, txn_options);
+  auto cf = db->DefaultColumnFamily();
+  Status s;
+
+  // Get a range lock for the range we are about to update
+  ASSERT_OK(txn0->GetRangeLock(cf, Endpoint("a"), Endpoint("c")));
+
+  bool try_range_lock_called = false;
+
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
+      "RangeTreeLockManager::TryRangeLock:enter",
+      [&](void* /*arg*/) { try_range_lock_called = true; });
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->EnableProcessing();
+
+  // For performance reasons, the following must NOT call lock_mgr->TryLock():
+  // We verify that by checking the value of try_range_lock_called.
+  ASSERT_OK(txn0->Put(cf, Slice("b"), Slice("value"),
+                      /*assume_tracked=*/true));
+
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->DisableProcessing();
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->ClearAllCallBacks();
+  ASSERT_FALSE(try_range_lock_called);
+
+  txn0->Rollback();
+
+  delete txn0;
+}
+
+TEST_F(RangeLockingTest, UpgradeLockAndGetConflict) {
+  WriteOptions write_options;
+  TransactionOptions txn_options;
+  auto cf = db->DefaultColumnFamily();
+  Status s;
+  std::string value;
+  txn_options.lock_timeout = 10;
+
+  Transaction* txn0 = db->BeginTransaction(write_options, txn_options);
+  Transaction* txn1 = db->BeginTransaction(write_options, txn_options);
+
+  // Get the shared lock in txn0
+  s = txn0->GetForUpdate(ReadOptions(), cf, Slice("a"), &value,
+                         false /*exclusive*/);
+  ASSERT_TRUE(s.IsNotFound());
+
+  // Get the shared lock on the same key in txn1
+  s = txn1->GetForUpdate(ReadOptions(), cf, Slice("a"), &value,
+                         false /*exclusive*/);
+  ASSERT_TRUE(s.IsNotFound());
+
+  // Now, try getting an exclusive lock that overlaps with the above
+  s = txn0->GetRangeLock(cf, Endpoint("a"), Endpoint("b"));
+  ASSERT_TRUE(s.IsTimedOut());
+
+  txn0->Rollback();
+  txn1->Rollback();
+
+  delete txn0;
+  delete txn1;
+}
+
+TEST_F(RangeLockingTest, SnapshotValidation) {
+  Status s;
+  Slice key_slice = Slice("k");
+  ColumnFamilyHandle* cfh = db->DefaultColumnFamily();
+
+  auto txn0 = NewTxn();
+  txn0->Put(key_slice, Slice("initial"));
+  txn0->Commit();
+
+  // txn1
+  auto txn1 = NewTxn();
+  txn1->SetSnapshot();
+  std::string val1;
+  ASSERT_OK(txn1->Get(ReadOptions(), cfh, key_slice, &val1));
+  ASSERT_EQ(val1, "initial");
+  val1 = val1 + std::string("-txn1");
+
+  ASSERT_OK(txn1->Put(cfh, key_slice, Slice(val1)));
+
+  // txn2
+  auto txn2 = NewTxn();
+  txn2->SetSnapshot();
+  std::string val2;
+  // This will see the original value as nothing is committed
+  // This is also Get, so it is doesn't acquire any locks.
+  ASSERT_OK(txn2->Get(ReadOptions(), cfh, key_slice, &val2));
+  ASSERT_EQ(val2, "initial");
+
+  // txn1
+  ASSERT_OK(txn1->Commit());
+
+  // txn2
+  val2 = val2 + std::string("-txn2");
+  // Now, this call should do Snapshot Validation and fail:
+  s = txn2->Put(cfh, key_slice, Slice(val2));
+  ASSERT_TRUE(s.IsBusy());
+
+  ASSERT_OK(txn2->Commit());
+
+  delete txn0;
+  delete txn1;
+  delete txn2;
+}
+
+TEST_F(RangeLockingTest, MultipleTrxLockStatusData) {
+  WriteOptions write_options;
+  TransactionOptions txn_options;
+  auto cf = db->DefaultColumnFamily();
+
+  Transaction* txn0 = db->BeginTransaction(write_options, txn_options);
+  Transaction* txn1 = db->BeginTransaction(write_options, txn_options);
+
+  // Get a range lock
+  ASSERT_OK(txn0->GetRangeLock(cf, Endpoint("z"), Endpoint("z")));
+  ASSERT_OK(txn1->GetRangeLock(cf, Endpoint("b"), Endpoint("e")));
+
+  auto s = range_lock_mgr->GetRangeLockStatusData();
+  ASSERT_EQ(s.size(), 2);
+  for (auto it = s.begin(); it != s.end(); ++it) {
+    ASSERT_EQ(it->first, cf->GetID());
+    auto val = it->second;
+    ASSERT_FALSE(val.start.inf_suffix);
+    ASSERT_FALSE(val.end.inf_suffix);
+    ASSERT_TRUE(val.exclusive);
+    ASSERT_EQ(val.ids.size(), 1);
+    if (val.ids[0] == txn0->GetID()) {
+      ASSERT_EQ(val.start.slice, "z");
+      ASSERT_EQ(val.end.slice, "z");
+    } else if (val.ids[0] == txn1->GetID()) {
+      ASSERT_EQ(val.start.slice, "b");
+      ASSERT_EQ(val.end.slice, "e");
+    } else {
+      FAIL();  // Unknown transaction ID.
+    }
+  }
+
+  delete txn0;
+  delete txn1;
+}
+
+#if defined(__has_feature)
+#if __has_feature(thread_sanitizer)
+#define SKIP_LOCK_ESCALATION_TEST 1
+#endif
+#else
+#define SKIP_LOCK_ESCALATION_TEST 1
+#endif
+
+#ifndef SKIP_LOCK_ESCALATION_TEST
+TEST_F(RangeLockingTest, BasicLockEscalation) {
+  auto cf = db->DefaultColumnFamily();
+
+  auto counters = range_lock_mgr->GetStatus();
+
+  // Initially not using any lock memory
+  ASSERT_EQ(counters.current_lock_memory, 0);
+  ASSERT_EQ(counters.escalation_count, 0);
+
+  ASSERT_EQ(0, range_lock_mgr->SetMaxLockMemory(2000));
+
+  // Insert until we see lock escalations
+  auto txn = NewTxn();
+
+  // Get the locks until we hit an escalation
+  for (int i = 0; i < 2020; i++) {
+    std::ostringstream buf;
+    buf << std::setw(8) << std::setfill('0') << i;
+    std::string buf_str = buf.str();
+    ASSERT_OK(txn->GetRangeLock(cf, Endpoint(buf_str), Endpoint(buf_str)));
+  }
+  counters = range_lock_mgr->GetStatus();
+  ASSERT_GT(counters.escalation_count, 0);
+  ASSERT_LE(counters.current_lock_memory, 2000);
+
+  delete txn;
+}
+
+// An escalation barrier function. Allow escalation iff the first two bytes are
+// identical.
+static bool escalation_barrier(const Endpoint& a, const Endpoint& b) {
+  assert(a.slice.size() > 2);
+  assert(b.slice.size() > 2);
+  if (memcmp(a.slice.data(), b.slice.data(), 2)) {
+    return true;  // This is a barrier
+  } else {
+    return false;  // No barrier
+  }
+}
+
+TEST_F(RangeLockingTest, LockEscalationBarrier) {
+  auto cf = db->DefaultColumnFamily();
+
+  auto counters = range_lock_mgr->GetStatus();
+
+  // Initially not using any lock memory
+  ASSERT_EQ(counters.escalation_count, 0);
+
+  range_lock_mgr->SetMaxLockMemory(8000);
+  range_lock_mgr->SetEscalationBarrierFunc(escalation_barrier);
+
+  // Insert enough locks to cause lock escalations to happen
+  auto txn = NewTxn();
+  const int N = 2000;
+  for (int i = 0; i < N; i++) {
+    std::ostringstream buf;
+    buf << std::setw(4) << std::setfill('0') << i;
+    std::string buf_str = buf.str();
+    ASSERT_OK(txn->GetRangeLock(cf, Endpoint(buf_str), Endpoint(buf_str)));
+  }
+  counters = range_lock_mgr->GetStatus();
+  ASSERT_GT(counters.escalation_count, 0);
+
+  // Check that lock escalation was not performed across escalation barriers:
+  // Use another txn to acquire locks near the barriers.
+  auto txn2 = NewTxn();
+  range_lock_mgr->SetMaxLockMemory(500000);
+  for (int i = 100; i < N; i += 100) {
+    std::ostringstream buf;
+    buf << std::setw(4) << std::setfill('0') << i - 1 << "-a";
+    std::string buf_str = buf.str();
+    // Check that we CAN get a lock near the escalation barrier
+    ASSERT_OK(txn2->GetRangeLock(cf, Endpoint(buf_str), Endpoint(buf_str)));
+  }
+
+  txn->Rollback();
+  txn2->Rollback();
+  delete txn;
+  delete txn2;
+}
+
+#endif
+
+TEST_F(RangeLockingTest, LockWaitCount) {
+  TransactionOptions txn_options;
+  auto cf = db->DefaultColumnFamily();
+  txn_options.lock_timeout = 50;
+  Transaction* txn0 = db->BeginTransaction(WriteOptions(), txn_options);
+  Transaction* txn1 = db->BeginTransaction(WriteOptions(), txn_options);
+
+  // Get a range lock
+  ASSERT_OK(txn0->GetRangeLock(cf, Endpoint("a"), Endpoint("c")));
+
+  uint64_t lock_waits1 = range_lock_mgr->GetStatus().lock_wait_count;
+  // Attempt to get a conflicting lock
+  auto s = txn1->GetRangeLock(cf, Endpoint("b"), Endpoint("z"));
+  ASSERT_TRUE(s.IsTimedOut());
+
+  // Check that the counter was incremented
+  uint64_t lock_waits2 = range_lock_mgr->GetStatus().lock_wait_count;
+  ASSERT_EQ(lock_waits1 + 1, lock_waits2);
+
+  txn0->Rollback();
+  txn1->Rollback();
+
+  delete txn0;
+  delete txn1;
+}
+
+TEST_F(RangeLockingTest, LockWaiteeAccess) {
+  TransactionOptions txn_options;
+  auto cf = db->DefaultColumnFamily();
+  txn_options.lock_timeout = 60;
+  Transaction* txn0 = db->BeginTransaction(WriteOptions(), txn_options);
+  Transaction* txn1 = db->BeginTransaction(WriteOptions(), txn_options);
+
+  // Get a range lock
+  ASSERT_OK(txn0->GetRangeLock(cf, Endpoint("a"), Endpoint("c")));
+
+  std::atomic<bool> reached(false);
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
+      "RangeTreeLockManager::TryRangeLock:EnterWaitingTxn", [&](void* /*arg*/) {
+        reached.store(true);
+        std::this_thread::sleep_for(std::chrono::milliseconds(2000));
+      });
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->EnableProcessing();
+
+  port::Thread t([&]() {
+    // Attempt to get a conflicting lock
+    auto s = txn1->GetRangeLock(cf, Endpoint("b"), Endpoint("z"));
+    ASSERT_TRUE(s.ok());
+    txn1->Rollback();
+  });
+
+  while (!reached.load()) {
+    std::this_thread::sleep_for(std::chrono::milliseconds(100));
+  }
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->DisableProcessing();
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->ClearAllCallBacks();
+
+  // Release locks and free the transaction
+  txn0->Rollback();
+  delete txn0;
+
+  t.join();
+
+  delete txn1;
+}
+
+void PointLockManagerTestExternalSetup(PointLockManagerTest* self) {
+  self->env_ = Env::Default();
+  self->db_dir_ = test::PerThreadDBPath("point_lock_manager_test");
+  ASSERT_OK(self->env_->CreateDir(self->db_dir_));
+
+  Options opt;
+  opt.create_if_missing = true;
+  TransactionDBOptions txn_opt;
+  txn_opt.transaction_lock_timeout = 0;
+
+  auto mutex_factory = std::make_shared<TransactionDBMutexFactoryImpl>();
+  self->locker_.reset(NewRangeLockManager(mutex_factory)->getLockManager());
+  std::shared_ptr<RangeLockManagerHandle> range_lock_mgr =
+      std::dynamic_pointer_cast<RangeLockManagerHandle>(self->locker_);
+  txn_opt.lock_mgr_handle = range_lock_mgr;
+
+  ASSERT_OK(TransactionDB::Open(opt, txn_opt, self->db_dir_, &self->db_));
+  self->wait_sync_point_name_ = "RangeTreeLockManager::TryRangeLock:WaitingTxn";
+}
+
+INSTANTIATE_TEST_CASE_P(RangeLockManager, AnyLockManagerTest,
+                        ::testing::Values(PointLockManagerTestExternalSetup));
+
+}  // namespace ROCKSDB_NAMESPACE
+
+int main(int argc, char** argv) {
+  ROCKSDB_NAMESPACE::port::InstallStackTraceHandler();
+  ::testing::InitGoogleTest(&argc, argv);
+  return RUN_ALL_TESTS();
+}
+
+#else  // OS_WIN
+
+#include <stdio.h>
+int main(int /*argc*/, char** /*argv*/) {
+  fprintf(stderr, "skipped as Range Locking is not supported on Windows\n");
+  return 0;
+}
+
+#endif  // OS_WIN
+
+#else
+#include <stdio.h>
+
+int main(int /*argc*/, char** /*argv*/) {
+  fprintf(stderr,
+          "skipped as transactions are not supported in rocksdb_lite\n");
+  return 0;
+}
+
+#endif  // ROCKSDB_LITE
diff --git a/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/COPYING.AGPLv3 b/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/COPYING.AGPLv3
new file mode 100644
index 000000000..dba13ed2d
--- /dev/null
+++ b/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/COPYING.AGPLv3
@@ -0,0 +1,661 @@
+                    GNU AFFERO GENERAL PUBLIC LICENSE
+                       Version 3, 19 November 2007
+
+ Copyright (C) 2007 Free Software Foundation, Inc. <http://fsf.org/>
+ Everyone is permitted to copy and distribute verbatim copies
+ of this license document, but changing it is not allowed.
+
+                            Preamble
+
+  The GNU Affero General Public License is a free, copyleft license for
+software and other kinds of works, specifically designed to ensure
+cooperation with the community in the case of network server software.
+
+  The licenses for most software and other practical works are designed
+to take away your freedom to share and change the works.  By contrast,
+our General Public Licenses are intended to guarantee your freedom to
+share and change all versions of a program--to make sure it remains free
+software for all its users.
+
+  When we speak of free software, we are referring to freedom, not
+price.  Our General Public Licenses are designed to make sure that you
+have the freedom to distribute copies of free software (and charge for
+them if you wish), that you receive source code or can get it if you
+want it, that you can change the software or use pieces of it in new
+free programs, and that you know you can do these things.
+
+  Developers that use our General Public Licenses protect your rights
+with two steps: (1) assert copyright on the software, and (2) offer
+you this License which gives you legal permission to copy, distribute
+and/or modify the software.
+
+  A secondary benefit of defending all users' freedom is that
+improvements made in alternate versions of the program, if they
+receive widespread use, become available for other developers to
+incorporate.  Many developers of free software are heartened and
+encouraged by the resulting cooperation.  However, in the case of
+software used on network servers, this result may fail to come about.
+The GNU General Public License permits making a modified version and
+letting the public access it on a server without ever releasing its
+source code to the public.
+
+  The GNU Affero General Public License is designed specifically to
+ensure that, in such cases, the modified source code becomes available
+to the community.  It requires the operator of a network server to
+provide the source code of the modified version running there to the
+users of that server.  Therefore, public use of a modified version, on
+a publicly accessible server, gives the public access to the source
+code of the modified version.
+
+  An older license, called the Affero General Public License and
+published by Affero, was designed to accomplish similar goals.  This is
+a different license, not a version of the Affero GPL, but Affero has
+released a new version of the Affero GPL which permits relicensing under
+this license.
+
+  The precise terms and conditions for copying, distribution and
+modification follow.
+
+                       TERMS AND CONDITIONS
+
+  0. Definitions.
+
+  "This License" refers to version 3 of the GNU Affero General Public License.
+
+  "Copyright" also means copyright-like laws that apply to other kinds of
+works, such as semiconductor masks.
+
+  "The Program" refers to any copyrightable work licensed under this
+License.  Each licensee is addressed as "you".  "Licensees" and
+"recipients" may be individuals or organizations.
+
+  To "modify" a work means to copy from or adapt all or part of the work
+in a fashion requiring copyright permission, other than the making of an
+exact copy.  The resulting work is called a "modified version" of the
+earlier work or a work "based on" the earlier work.
+
+  A "covered work" means either the unmodified Program or a work based
+on the Program.
+
+  To "propagate" a work means to do anything with it that, without
+permission, would make you directly or secondarily liable for
+infringement under applicable copyright law, except executing it on a
+computer or modifying a private copy.  Propagation includes copying,
+distribution (with or without modification), making available to the
+public, and in some countries other activities as well.
+
+  To "convey" a work means any kind of propagation that enables other
+parties to make or receive copies.  Mere interaction with a user through
+a computer network, with no transfer of a copy, is not conveying.
+
+  An interactive user interface displays "Appropriate Legal Notices"
+to the extent that it includes a convenient and prominently visible
+feature that (1) displays an appropriate copyright notice, and (2)
+tells the user that there is no warranty for the work (except to the
+extent that warranties are provided), that licensees may convey the
+work under this License, and how to view a copy of this License.  If
+the interface presents a list of user commands or options, such as a
+menu, a prominent item in the list meets this criterion.
+
+  1. Source Code.
+
+  The "source code" for a work means the preferred form of the work
+for making modifications to it.  "Object code" means any non-source
+form of a work.
+
+  A "Standard Interface" means an interface that either is an official
+standard defined by a recognized standards body, or, in the case of
+interfaces specified for a particular programming language, one that
+is widely used among developers working in that language.
+
+  The "System Libraries" of an executable work include anything, other
+than the work as a whole, that (a) is included in the normal form of
+packaging a Major Component, but which is not part of that Major
+Component, and (b) serves only to enable use of the work with that
+Major Component, or to implement a Standard Interface for which an
+implementation is available to the public in source code form.  A
+"Major Component", in this context, means a major essential component
+(kernel, window system, and so on) of the specific operating system
+(if any) on which the executable work runs, or a compiler used to
+produce the work, or an object code interpreter used to run it.
+
+  The "Corresponding Source" for a work in object code form means all
+the source code needed to generate, install, and (for an executable
+work) run the object code and to modify the work, including scripts to
+control those activities.  However, it does not include the work's
+System Libraries, or general-purpose tools or generally available free
+programs which are used unmodified in performing those activities but
+which are not part of the work.  For example, Corresponding Source
+includes interface definition files associated with source files for
+the work, and the source code for shared libraries and dynamically
+linked subprograms that the work is specifically designed to require,
+such as by intimate data communication or control flow between those
+subprograms and other parts of the work.
+
+  The Corresponding Source need not include anything that users
+can regenerate automatically from other parts of the Corresponding
+Source.
+
+  The Corresponding Source for a work in source code form is that
+same work.
+
+  2. Basic Permissions.
+
+  All rights granted under this License are granted for the term of
+copyright on the Program, and are irrevocable provided the stated
+conditions are met.  This License explicitly affirms your unlimited
+permission to run the unmodified Program.  The output from running a
+covered work is covered by this License only if the output, given its
+content, constitutes a covered work.  This License acknowledges your
+rights of fair use or other equivalent, as provided by copyright law.
+
+  You may make, run and propagate covered works that you do not
+convey, without conditions so long as your license otherwise remains
+in force.  You may convey covered works to others for the sole purpose
+of having them make modifications exclusively for you, or provide you
+with facilities for running those works, provided that you comply with
+the terms of this License in conveying all material for which you do
+not control copyright.  Those thus making or running the covered works
+for you must do so exclusively on your behalf, under your direction
+and control, on terms that prohibit them from making any copies of
+your copyrighted material outside their relationship with you.
+
+  Conveying under any other circumstances is permitted solely under
+the conditions stated below.  Sublicensing is not allowed; section 10
+makes it unnecessary.
+
+  3. Protecting Users' Legal Rights From Anti-Circumvention Law.
+
+  No covered work shall be deemed part of an effective technological
+measure under any applicable law fulfilling obligations under article
+11 of the WIPO copyright treaty adopted on 20 December 1996, or
+similar laws prohibiting or restricting circumvention of such
+measures.
+
+  When you convey a covered work, you waive any legal power to forbid
+circumvention of technological measures to the extent such circumvention
+is effected by exercising rights under this License with respect to
+the covered work, and you disclaim any intention to limit operation or
+modification of the work as a means of enforcing, against the work's
+users, your or third parties' legal rights to forbid circumvention of
+technological measures.
+
+  4. Conveying Verbatim Copies.
+
+  You may convey verbatim copies of the Program's source code as you
+receive it, in any medium, provided that you conspicuously and
+appropriately publish on each copy an appropriate copyright notice;
+keep intact all notices stating that this License and any
+non-permissive terms added in accord with section 7 apply to the code;
+keep intact all notices of the absence of any warranty; and give all
+recipients a copy of this License along with the Program.
+
+  You may charge any price or no price for each copy that you convey,
+and you may offer support or warranty protection for a fee.
+
+  5. Conveying Modified Source Versions.
+
+  You may convey a work based on the Program, or the modifications to
+produce it from the Program, in the form of source code under the
+terms of section 4, provided that you also meet all of these conditions:
+
+    a) The work must carry prominent notices stating that you modified
+    it, and giving a relevant date.
+
+    b) The work must carry prominent notices stating that it is
+    released under this License and any conditions added under section
+    7.  This requirement modifies the requirement in section 4 to
+    "keep intact all notices".
+
+    c) You must license the entire work, as a whole, under this
+    License to anyone who comes into possession of a copy.  This
+    License will therefore apply, along with any applicable section 7
+    additional terms, to the whole of the work, and all its parts,
+    regardless of how they are packaged.  This License gives no
+    permission to license the work in any other way, but it does not
+    invalidate such permission if you have separately received it.
+
+    d) If the work has interactive user interfaces, each must display
+    Appropriate Legal Notices; however, if the Program has interactive
+    interfaces that do not display Appropriate Legal Notices, your
+    work need not make them do so.
+
+  A compilation of a covered work with other separate and independent
+works, which are not by their nature extensions of the covered work,
+and which are not combined with it such as to form a larger program,
+in or on a volume of a storage or distribution medium, is called an
+"aggregate" if the compilation and its resulting copyright are not
+used to limit the access or legal rights of the compilation's users
+beyond what the individual works permit.  Inclusion of a covered work
+in an aggregate does not cause this License to apply to the other
+parts of the aggregate.
+
+  6. Conveying Non-Source Forms.
+
+  You may convey a covered work in object code form under the terms
+of sections 4 and 5, provided that you also convey the
+machine-readable Corresponding Source under the terms of this License,
+in one of these ways:
+
+    a) Convey the object code in, or embodied in, a physical product
+    (including a physical distribution medium), accompanied by the
+    Corresponding Source fixed on a durable physical medium
+    customarily used for software interchange.
+
+    b) Convey the object code in, or embodied in, a physical product
+    (including a physical distribution medium), accompanied by a
+    written offer, valid for at least three years and valid for as
+    long as you offer spare parts or customer support for that product
+    model, to give anyone who possesses the object code either (1) a
+    copy of the Corresponding Source for all the software in the
+    product that is covered by this License, on a durable physical
+    medium customarily used for software interchange, for a price no
+    more than your reasonable cost of physically performing this
+    conveying of source, or (2) access to copy the
+    Corresponding Source from a network server at no charge.
+
+    c) Convey individual copies of the object code with a copy of the
+    written offer to provide the Corresponding Source.  This
+    alternative is allowed only occasionally and noncommercially, and
+    only if you received the object code with such an offer, in accord
+    with subsection 6b.
+
+    d) Convey the object code by offering access from a designated
+    place (gratis or for a charge), and offer equivalent access to the
+    Corresponding Source in the same way through the same place at no
+    further charge.  You need not require recipients to copy the
+    Corresponding Source along with the object code.  If the place to
+    copy the object code is a network server, the Corresponding Source
+    may be on a different server (operated by you or a third party)
+    that supports equivalent copying facilities, provided you maintain
+    clear directions next to the object code saying where to find the
+    Corresponding Source.  Regardless of what server hosts the
+    Corresponding Source, you remain obligated to ensure that it is
+    available for as long as needed to satisfy these requirements.
+
+    e) Convey the object code using peer-to-peer transmission, provided
+    you inform other peers where the object code and Corresponding
+    Source of the work are being offered to the general public at no
+    charge under subsection 6d.
+
+  A separable portion of the object code, whose source code is excluded
+from the Corresponding Source as a System Library, need not be
+included in conveying the object code work.
+
+  A "User Product" is either (1) a "consumer product", which means any
+tangible personal property which is normally used for personal, family,
+or household purposes, or (2) anything designed or sold for incorporation
+into a dwelling.  In determining whether a product is a consumer product,
+doubtful cases shall be resolved in favor of coverage.  For a particular
+product received by a particular user, "normally used" refers to a
+typical or common use of that class of product, regardless of the status
+of the particular user or of the way in which the particular user
+actually uses, or expects or is expected to use, the product.  A product
+is a consumer product regardless of whether the product has substantial
+commercial, industrial or non-consumer uses, unless such uses represent
+the only significant mode of use of the product.
+
+  "Installation Information" for a User Product means any methods,
+procedures, authorization keys, or other information required to install
+and execute modified versions of a covered work in that User Product from
+a modified version of its Corresponding Source.  The information must
+suffice to ensure that the continued functioning of the modified object
+code is in no case prevented or interfered with solely because
+modification has been made.
+
+  If you convey an object code work under this section in, or with, or
+specifically for use in, a User Product, and the conveying occurs as
+part of a transaction in which the right of possession and use of the
+User Product is transferred to the recipient in perpetuity or for a
+fixed term (regardless of how the transaction is characterized), the
+Corresponding Source conveyed under this section must be accompanied
+by the Installation Information.  But this requirement does not apply
+if neither you nor any third party retains the ability to install
+modified object code on the User Product (for example, the work has
+been installed in ROM).
+
+  The requirement to provide Installation Information does not include a
+requirement to continue to provide support service, warranty, or updates
+for a work that has been modified or installed by the recipient, or for
+the User Product in which it has been modified or installed.  Access to a
+network may be denied when the modification itself materially and
+adversely affects the operation of the network or violates the rules and
+protocols for communication across the network.
+
+  Corresponding Source conveyed, and Installation Information provided,
+in accord with this section must be in a format that is publicly
+documented (and with an implementation available to the public in
+source code form), and must require no special password or key for
+unpacking, reading or copying.
+
+  7. Additional Terms.
+
+  "Additional permissions" are terms that supplement the terms of this
+License by making exceptions from one or more of its conditions.
+Additional permissions that are applicable to the entire Program shall
+be treated as though they were included in this License, to the extent
+that they are valid under applicable law.  If additional permissions
+apply only to part of the Program, that part may be used separately
+under those permissions, but the entire Program remains governed by
+this License without regard to the additional permissions.
+
+  When you convey a copy of a covered work, you may at your option
+remove any additional permissions from that copy, or from any part of
+it.  (Additional permissions may be written to require their own
+removal in certain cases when you modify the work.)  You may place
+additional permissions on material, added by you to a covered work,
+for which you have or can give appropriate copyright permission.
+
+  Notwithstanding any other provision of this License, for material you
+add to a covered work, you may (if authorized by the copyright holders of
+that material) supplement the terms of this License with terms:
+
+    a) Disclaiming warranty or limiting liability differently from the
+    terms of sections 15 and 16 of this License; or
+
+    b) Requiring preservation of specified reasonable legal notices or
+    author attributions in that material or in the Appropriate Legal
+    Notices displayed by works containing it; or
+
+    c) Prohibiting misrepresentation of the origin of that material, or
+    requiring that modified versions of such material be marked in
+    reasonable ways as different from the original version; or
+
+    d) Limiting the use for publicity purposes of names of licensors or
+    authors of the material; or
+
+    e) Declining to grant rights under trademark law for use of some
+    trade names, trademarks, or service marks; or
+
+    f) Requiring indemnification of licensors and authors of that
+    material by anyone who conveys the material (or modified versions of
+    it) with contractual assumptions of liability to the recipient, for
+    any liability that these contractual assumptions directly impose on
+    those licensors and authors.
+
+  All other non-permissive additional terms are considered "further
+restrictions" within the meaning of section 10.  If the Program as you
+received it, or any part of it, contains a notice stating that it is
+governed by this License along with a term that is a further
+restriction, you may remove that term.  If a license document contains
+a further restriction but permits relicensing or conveying under this
+License, you may add to a covered work material governed by the terms
+of that license document, provided that the further restriction does
+not survive such relicensing or conveying.
+
+  If you add terms to a covered work in accord with this section, you
+must place, in the relevant source files, a statement of the
+additional terms that apply to those files, or a notice indicating
+where to find the applicable terms.
+
+  Additional terms, permissive or non-permissive, may be stated in the
+form of a separately written license, or stated as exceptions;
+the above requirements apply either way.
+
+  8. Termination.
+
+  You may not propagate or modify a covered work except as expressly
+provided under this License.  Any attempt otherwise to propagate or
+modify it is void, and will automatically terminate your rights under
+this License (including any patent licenses granted under the third
+paragraph of section 11).
+
+  However, if you cease all violation of this License, then your
+license from a particular copyright holder is reinstated (a)
+provisionally, unless and until the copyright holder explicitly and
+finally terminates your license, and (b) permanently, if the copyright
+holder fails to notify you of the violation by some reasonable means
+prior to 60 days after the cessation.
+
+  Moreover, your license from a particular copyright holder is
+reinstated permanently if the copyright holder notifies you of the
+violation by some reasonable means, this is the first time you have
+received notice of violation of this License (for any work) from that
+copyright holder, and you cure the violation prior to 30 days after
+your receipt of the notice.
+
+  Termination of your rights under this section does not terminate the
+licenses of parties who have received copies or rights from you under
+this License.  If your rights have been terminated and not permanently
+reinstated, you do not qualify to receive new licenses for the same
+material under section 10.
+
+  9. Acceptance Not Required for Having Copies.
+
+  You are not required to accept this License in order to receive or
+run a copy of the Program.  Ancillary propagation of a covered work
+occurring solely as a consequence of using peer-to-peer transmission
+to receive a copy likewise does not require acceptance.  However,
+nothing other than this License grants you permission to propagate or
+modify any covered work.  These actions infringe copyright if you do
+not accept this License.  Therefore, by modifying or propagating a
+covered work, you indicate your acceptance of this License to do so.
+
+  10. Automatic Licensing of Downstream Recipients.
+
+  Each time you convey a covered work, the recipient automatically
+receives a license from the original licensors, to run, modify and
+propagate that work, subject to this License.  You are not responsible
+for enforcing compliance by third parties with this License.
+
+  An "entity transaction" is a transaction transferring control of an
+organization, or substantially all assets of one, or subdividing an
+organization, or merging organizations.  If propagation of a covered
+work results from an entity transaction, each party to that
+transaction who receives a copy of the work also receives whatever
+licenses to the work the party's predecessor in interest had or could
+give under the previous paragraph, plus a right to possession of the
+Corresponding Source of the work from the predecessor in interest, if
+the predecessor has it or can get it with reasonable efforts.
+
+  You may not impose any further restrictions on the exercise of the
+rights granted or affirmed under this License.  For example, you may
+not impose a license fee, royalty, or other charge for exercise of
+rights granted under this License, and you may not initiate litigation
+(including a cross-claim or counterclaim in a lawsuit) alleging that
+any patent claim is infringed by making, using, selling, offering for
+sale, or importing the Program or any portion of it.
+
+  11. Patents.
+
+  A "contributor" is a copyright holder who authorizes use under this
+License of the Program or a work on which the Program is based.  The
+work thus licensed is called the contributor's "contributor version".
+
+  A contributor's "essential patent claims" are all patent claims
+owned or controlled by the contributor, whether already acquired or
+hereafter acquired, that would be infringed by some manner, permitted
+by this License, of making, using, or selling its contributor version,
+but do not include claims that would be infringed only as a
+consequence of further modification of the contributor version.  For
+purposes of this definition, "control" includes the right to grant
+patent sublicenses in a manner consistent with the requirements of
+this License.
+
+  Each contributor grants you a non-exclusive, worldwide, royalty-free
+patent license under the contributor's essential patent claims, to
+make, use, sell, offer for sale, import and otherwise run, modify and
+propagate the contents of its contributor version.
+
+  In the following three paragraphs, a "patent license" is any express
+agreement or commitment, however denominated, not to enforce a patent
+(such as an express permission to practice a patent or covenant not to
+sue for patent infringement).  To "grant" such a patent license to a
+party means to make such an agreement or commitment not to enforce a
+patent against the party.
+
+  If you convey a covered work, knowingly relying on a patent license,
+and the Corresponding Source of the work is not available for anyone
+to copy, free of charge and under the terms of this License, through a
+publicly available network server or other readily accessible means,
+then you must either (1) cause the Corresponding Source to be so
+available, or (2) arrange to deprive yourself of the benefit of the
+patent license for this particular work, or (3) arrange, in a manner
+consistent with the requirements of this License, to extend the patent
+license to downstream recipients.  "Knowingly relying" means you have
+actual knowledge that, but for the patent license, your conveying the
+covered work in a country, or your recipient's use of the covered work
+in a country, would infringe one or more identifiable patents in that
+country that you have reason to believe are valid.
+
+  If, pursuant to or in connection with a single transaction or
+arrangement, you convey, or propagate by procuring conveyance of, a
+covered work, and grant a patent license to some of the parties
+receiving the covered work authorizing them to use, propagate, modify
+or convey a specific copy of the covered work, then the patent license
+you grant is automatically extended to all recipients of the covered
+work and works based on it.
+
+  A patent license is "discriminatory" if it does not include within
+the scope of its coverage, prohibits the exercise of, or is
+conditioned on the non-exercise of one or more of the rights that are
+specifically granted under this License.  You may not convey a covered
+work if you are a party to an arrangement with a third party that is
+in the business of distributing software, under which you make payment
+to the third party based on the extent of your activity of conveying
+the work, and under which the third party grants, to any of the
+parties who would receive the covered work from you, a discriminatory
+patent license (a) in connection with copies of the covered work
+conveyed by you (or copies made from those copies), or (b) primarily
+for and in connection with specific products or compilations that
+contain the covered work, unless you entered into that arrangement,
+or that patent license was granted, prior to 28 March 2007.
+
+  Nothing in this License shall be construed as excluding or limiting
+any implied license or other defenses to infringement that may
+otherwise be available to you under applicable patent law.
+
+  12. No Surrender of Others' Freedom.
+
+  If conditions are imposed on you (whether by court order, agreement or
+otherwise) that contradict the conditions of this License, they do not
+excuse you from the conditions of this License.  If you cannot convey a
+covered work so as to satisfy simultaneously your obligations under this
+License and any other pertinent obligations, then as a consequence you may
+not convey it at all.  For example, if you agree to terms that obligate you
+to collect a royalty for further conveying from those to whom you convey
+the Program, the only way you could satisfy both those terms and this
+License would be to refrain entirely from conveying the Program.
+
+  13. Remote Network Interaction; Use with the GNU General Public License.
+
+  Notwithstanding any other provision of this License, if you modify the
+Program, your modified version must prominently offer all users
+interacting with it remotely through a computer network (if your version
+supports such interaction) an opportunity to receive the Corresponding
+Source of your version by providing access to the Corresponding Source
+from a network server at no charge, through some standard or customary
+means of facilitating copying of software.  This Corresponding Source
+shall include the Corresponding Source for any work covered by version 3
+of the GNU General Public License that is incorporated pursuant to the
+following paragraph.
+
+  Notwithstanding any other provision of this License, you have
+permission to link or combine any covered work with a work licensed
+under version 3 of the GNU General Public License into a single
+combined work, and to convey the resulting work.  The terms of this
+License will continue to apply to the part which is the covered work,
+but the work with which it is combined will remain governed by version
+3 of the GNU General Public License.
+
+  14. Revised Versions of this License.
+
+  The Free Software Foundation may publish revised and/or new versions of
+the GNU Affero General Public License from time to time.  Such new versions
+will be similar in spirit to the present version, but may differ in detail to
+address new problems or concerns.
+
+  Each version is given a distinguishing version number.  If the
+Program specifies that a certain numbered version of the GNU Affero General
+Public License "or any later version" applies to it, you have the
+option of following the terms and conditions either of that numbered
+version or of any later version published by the Free Software
+Foundation.  If the Program does not specify a version number of the
+GNU Affero General Public License, you may choose any version ever published
+by the Free Software Foundation.
+
+  If the Program specifies that a proxy can decide which future
+versions of the GNU Affero General Public License can be used, that proxy's
+public statement of acceptance of a version permanently authorizes you
+to choose that version for the Program.
+
+  Later license versions may give you additional or different
+permissions.  However, no additional obligations are imposed on any
+author or copyright holder as a result of your choosing to follow a
+later version.
+
+  15. Disclaimer of Warranty.
+
+  THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY
+APPLICABLE LAW.  EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT
+HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY
+OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO,
+THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+PURPOSE.  THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM
+IS WITH YOU.  SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF
+ALL NECESSARY SERVICING, REPAIR OR CORRECTION.
+
+  16. Limitation of Liability.
+
+  IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING
+WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS
+THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY
+GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE
+USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF
+DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD
+PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS),
+EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF
+SUCH DAMAGES.
+
+  17. Interpretation of Sections 15 and 16.
+
+  If the disclaimer of warranty and limitation of liability provided
+above cannot be given local legal effect according to their terms,
+reviewing courts shall apply local law that most closely approximates
+an absolute waiver of all civil liability in connection with the
+Program, unless a warranty or assumption of liability accompanies a
+copy of the Program in return for a fee.
+
+                     END OF TERMS AND CONDITIONS
+
+            How to Apply These Terms to Your New Programs
+
+  If you develop a new program, and you want it to be of the greatest
+possible use to the public, the best way to achieve this is to make it
+free software which everyone can redistribute and change under these terms.
+
+  To do so, attach the following notices to the program.  It is safest
+to attach them to the start of each source file to most effectively
+state the exclusion of warranty; and each file should have at least
+the "copyright" line and a pointer to where the full notice is found.
+
+    <one line to give the program's name and a brief idea of what it does.>
+    Copyright (C) <year>  <name of author>
+
+    This program is free software: you can redistribute it and/or modify
+    it under the terms of the GNU Affero General Public License as published by
+    the Free Software Foundation, either version 3 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU Affero General Public License for more details.
+
+    You should have received a copy of the GNU Affero General Public License
+    along with this program.  If not, see <http://www.gnu.org/licenses/>.
+
+Also add information on how to contact you by electronic and paper mail.
+
+  If your software can interact with users remotely through a computer
+network, you should also make sure that it provides a way for users to
+get its source.  For example, if your program is a web application, its
+interface could display a "Source" link that leads users to an archive
+of the code.  There are many ways you could offer source, and different
+solutions will be better for different programs; see section 13 for the
+specific requirements.
+
+  You should also get your employer (if you work as a programmer) or school,
+if any, to sign a "copyright disclaimer" for the program, if necessary.
+For more information on this, and how to apply and follow the GNU AGPL, see
+<http://www.gnu.org/licenses/>.
diff --git a/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/COPYING.APACHEv2 b/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/COPYING.APACHEv2
new file mode 100644
index 000000000..ecbfc770f
--- /dev/null
+++ b/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/COPYING.APACHEv2
@@ -0,0 +1,174 @@
+Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
diff --git a/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/COPYING.GPLv2 b/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/COPYING.GPLv2
new file mode 100644
index 000000000..d511905c1
--- /dev/null
+++ b/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/COPYING.GPLv2
@@ -0,0 +1,339 @@
+		    GNU GENERAL PUBLIC LICENSE
+		       Version 2, June 1991
+
+ Copyright (C) 1989, 1991 Free Software Foundation, Inc.,
+ 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
+ Everyone is permitted to copy and distribute verbatim copies
+ of this license document, but changing it is not allowed.
+
+			    Preamble
+
+  The licenses for most software are designed to take away your
+freedom to share and change it.  By contrast, the GNU General Public
+License is intended to guarantee your freedom to share and change free
+software--to make sure the software is free for all its users.  This
+General Public License applies to most of the Free Software
+Foundation's software and to any other program whose authors commit to
+using it.  (Some other Free Software Foundation software is covered by
+the GNU Lesser General Public License instead.)  You can apply it to
+your programs, too.
+
+  When we speak of free software, we are referring to freedom, not
+price.  Our General Public Licenses are designed to make sure that you
+have the freedom to distribute copies of free software (and charge for
+this service if you wish), that you receive source code or can get it
+if you want it, that you can change the software or use pieces of it
+in new free programs; and that you know you can do these things.
+
+  To protect your rights, we need to make restrictions that forbid
+anyone to deny you these rights or to ask you to surrender the rights.
+These restrictions translate to certain responsibilities for you if you
+distribute copies of the software, or if you modify it.
+
+  For example, if you distribute copies of such a program, whether
+gratis or for a fee, you must give the recipients all the rights that
+you have.  You must make sure that they, too, receive or can get the
+source code.  And you must show them these terms so they know their
+rights.
+
+  We protect your rights with two steps: (1) copyright the software, and
+(2) offer you this license which gives you legal permission to copy,
+distribute and/or modify the software.
+
+  Also, for each author's protection and ours, we want to make certain
+that everyone understands that there is no warranty for this free
+software.  If the software is modified by someone else and passed on, we
+want its recipients to know that what they have is not the original, so
+that any problems introduced by others will not reflect on the original
+authors' reputations.
+
+  Finally, any free program is threatened constantly by software
+patents.  We wish to avoid the danger that redistributors of a free
+program will individually obtain patent licenses, in effect making the
+program proprietary.  To prevent this, we have made it clear that any
+patent must be licensed for everyone's free use or not licensed at all.
+
+  The precise terms and conditions for copying, distribution and
+modification follow.
+
+		    GNU GENERAL PUBLIC LICENSE
+   TERMS AND CONDITIONS FOR COPYING, DISTRIBUTION AND MODIFICATION
+
+  0. This License applies to any program or other work which contains
+a notice placed by the copyright holder saying it may be distributed
+under the terms of this General Public License.  The "Program", below,
+refers to any such program or work, and a "work based on the Program"
+means either the Program or any derivative work under copyright law:
+that is to say, a work containing the Program or a portion of it,
+either verbatim or with modifications and/or translated into another
+language.  (Hereinafter, translation is included without limitation in
+the term "modification".)  Each licensee is addressed as "you".
+
+Activities other than copying, distribution and modification are not
+covered by this License; they are outside its scope.  The act of
+running the Program is not restricted, and the output from the Program
+is covered only if its contents constitute a work based on the
+Program (independent of having been made by running the Program).
+Whether that is true depends on what the Program does.
+
+  1. You may copy and distribute verbatim copies of the Program's
+source code as you receive it, in any medium, provided that you
+conspicuously and appropriately publish on each copy an appropriate
+copyright notice and disclaimer of warranty; keep intact all the
+notices that refer to this License and to the absence of any warranty;
+and give any other recipients of the Program a copy of this License
+along with the Program.
+
+You may charge a fee for the physical act of transferring a copy, and
+you may at your option offer warranty protection in exchange for a fee.
+
+  2. You may modify your copy or copies of the Program or any portion
+of it, thus forming a work based on the Program, and copy and
+distribute such modifications or work under the terms of Section 1
+above, provided that you also meet all of these conditions:
+
+    a) You must cause the modified files to carry prominent notices
+    stating that you changed the files and the date of any change.
+
+    b) You must cause any work that you distribute or publish, that in
+    whole or in part contains or is derived from the Program or any
+    part thereof, to be licensed as a whole at no charge to all third
+    parties under the terms of this License.
+
+    c) If the modified program normally reads commands interactively
+    when run, you must cause it, when started running for such
+    interactive use in the most ordinary way, to print or display an
+    announcement including an appropriate copyright notice and a
+    notice that there is no warranty (or else, saying that you provide
+    a warranty) and that users may redistribute the program under
+    these conditions, and telling the user how to view a copy of this
+    License.  (Exception: if the Program itself is interactive but
+    does not normally print such an announcement, your work based on
+    the Program is not required to print an announcement.)
+
+These requirements apply to the modified work as a whole.  If
+identifiable sections of that work are not derived from the Program,
+and can be reasonably considered independent and separate works in
+themselves, then this License, and its terms, do not apply to those
+sections when you distribute them as separate works.  But when you
+distribute the same sections as part of a whole which is a work based
+on the Program, the distribution of the whole must be on the terms of
+this License, whose permissions for other licensees extend to the
+entire whole, and thus to each and every part regardless of who wrote it.
+
+Thus, it is not the intent of this section to claim rights or contest
+your rights to work written entirely by you; rather, the intent is to
+exercise the right to control the distribution of derivative or
+collective works based on the Program.
+
+In addition, mere aggregation of another work not based on the Program
+with the Program (or with a work based on the Program) on a volume of
+a storage or distribution medium does not bring the other work under
+the scope of this License.
+
+  3. You may copy and distribute the Program (or a work based on it,
+under Section 2) in object code or executable form under the terms of
+Sections 1 and 2 above provided that you also do one of the following:
+
+    a) Accompany it with the complete corresponding machine-readable
+    source code, which must be distributed under the terms of Sections
+    1 and 2 above on a medium customarily used for software interchange; or,
+
+    b) Accompany it with a written offer, valid for at least three
+    years, to give any third party, for a charge no more than your
+    cost of physically performing source distribution, a complete
+    machine-readable copy of the corresponding source code, to be
+    distributed under the terms of Sections 1 and 2 above on a medium
+    customarily used for software interchange; or,
+
+    c) Accompany it with the information you received as to the offer
+    to distribute corresponding source code.  (This alternative is
+    allowed only for noncommercial distribution and only if you
+    received the program in object code or executable form with such
+    an offer, in accord with Subsection b above.)
+
+The source code for a work means the preferred form of the work for
+making modifications to it.  For an executable work, complete source
+code means all the source code for all modules it contains, plus any
+associated interface definition files, plus the scripts used to
+control compilation and installation of the executable.  However, as a
+special exception, the source code distributed need not include
+anything that is normally distributed (in either source or binary
+form) with the major components (compiler, kernel, and so on) of the
+operating system on which the executable runs, unless that component
+itself accompanies the executable.
+
+If distribution of executable or object code is made by offering
+access to copy from a designated place, then offering equivalent
+access to copy the source code from the same place counts as
+distribution of the source code, even though third parties are not
+compelled to copy the source along with the object code.
+
+  4. You may not copy, modify, sublicense, or distribute the Program
+except as expressly provided under this License.  Any attempt
+otherwise to copy, modify, sublicense or distribute the Program is
+void, and will automatically terminate your rights under this License.
+However, parties who have received copies, or rights, from you under
+this License will not have their licenses terminated so long as such
+parties remain in full compliance.
+
+  5. You are not required to accept this License, since you have not
+signed it.  However, nothing else grants you permission to modify or
+distribute the Program or its derivative works.  These actions are
+prohibited by law if you do not accept this License.  Therefore, by
+modifying or distributing the Program (or any work based on the
+Program), you indicate your acceptance of this License to do so, and
+all its terms and conditions for copying, distributing or modifying
+the Program or works based on it.
+
+  6. Each time you redistribute the Program (or any work based on the
+Program), the recipient automatically receives a license from the
+original licensor to copy, distribute or modify the Program subject to
+these terms and conditions.  You may not impose any further
+restrictions on the recipients' exercise of the rights granted herein.
+You are not responsible for enforcing compliance by third parties to
+this License.
+
+  7. If, as a consequence of a court judgment or allegation of patent
+infringement or for any other reason (not limited to patent issues),
+conditions are imposed on you (whether by court order, agreement or
+otherwise) that contradict the conditions of this License, they do not
+excuse you from the conditions of this License.  If you cannot
+distribute so as to satisfy simultaneously your obligations under this
+License and any other pertinent obligations, then as a consequence you
+may not distribute the Program at all.  For example, if a patent
+license would not permit royalty-free redistribution of the Program by
+all those who receive copies directly or indirectly through you, then
+the only way you could satisfy both it and this License would be to
+refrain entirely from distribution of the Program.
+
+If any portion of this section is held invalid or unenforceable under
+any particular circumstance, the balance of the section is intended to
+apply and the section as a whole is intended to apply in other
+circumstances.
+
+It is not the purpose of this section to induce you to infringe any
+patents or other property right claims or to contest validity of any
+such claims; this section has the sole purpose of protecting the
+integrity of the free software distribution system, which is
+implemented by public license practices.  Many people have made
+generous contributions to the wide range of software distributed
+through that system in reliance on consistent application of that
+system; it is up to the author/donor to decide if he or she is willing
+to distribute software through any other system and a licensee cannot
+impose that choice.
+
+This section is intended to make thoroughly clear what is believed to
+be a consequence of the rest of this License.
+
+  8. If the distribution and/or use of the Program is restricted in
+certain countries either by patents or by copyrighted interfaces, the
+original copyright holder who places the Program under this License
+may add an explicit geographical distribution limitation excluding
+those countries, so that distribution is permitted only in or among
+countries not thus excluded.  In such case, this License incorporates
+the limitation as if written in the body of this License.
+
+  9. The Free Software Foundation may publish revised and/or new versions
+of the General Public License from time to time.  Such new versions will
+be similar in spirit to the present version, but may differ in detail to
+address new problems or concerns.
+
+Each version is given a distinguishing version number.  If the Program
+specifies a version number of this License which applies to it and "any
+later version", you have the option of following the terms and conditions
+either of that version or of any later version published by the Free
+Software Foundation.  If the Program does not specify a version number of
+this License, you may choose any version ever published by the Free Software
+Foundation.
+
+  10. If you wish to incorporate parts of the Program into other free
+programs whose distribution conditions are different, write to the author
+to ask for permission.  For software which is copyrighted by the Free
+Software Foundation, write to the Free Software Foundation; we sometimes
+make exceptions for this.  Our decision will be guided by the two goals
+of preserving the free status of all derivatives of our free software and
+of promoting the sharing and reuse of software generally.
+
+			    NO WARRANTY
+
+  11. BECAUSE THE PROGRAM IS LICENSED FREE OF CHARGE, THERE IS NO WARRANTY
+FOR THE PROGRAM, TO THE EXTENT PERMITTED BY APPLICABLE LAW.  EXCEPT WHEN
+OTHERWISE STATED IN WRITING THE COPYRIGHT HOLDERS AND/OR OTHER PARTIES
+PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY OF ANY KIND, EITHER EXPRESSED
+OR IMPLIED, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF
+MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE.  THE ENTIRE RISK AS
+TO THE QUALITY AND PERFORMANCE OF THE PROGRAM IS WITH YOU.  SHOULD THE
+PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF ALL NECESSARY SERVICING,
+REPAIR OR CORRECTION.
+
+  12. IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING
+WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MAY MODIFY AND/OR
+REDISTRIBUTE THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES,
+INCLUDING ANY GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING
+OUT OF THE USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED
+TO LOSS OF DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY
+YOU OR THIRD PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER
+PROGRAMS), EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE
+POSSIBILITY OF SUCH DAMAGES.
+
+		     END OF TERMS AND CONDITIONS
+
+	    How to Apply These Terms to Your New Programs
+
+  If you develop a new program, and you want it to be of the greatest
+possible use to the public, the best way to achieve this is to make it
+free software which everyone can redistribute and change under these terms.
+
+  To do so, attach the following notices to the program.  It is safest
+to attach them to the start of each source file to most effectively
+convey the exclusion of warranty; and each file should have at least
+the "copyright" line and a pointer to where the full notice is found.
+
+    <one line to give the program's name and a brief idea of what it does.>
+    Copyright (C) <year>  <name of author>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+Also add information on how to contact you by electronic and paper mail.
+
+If the program is interactive, make it output a short notice like this
+when it starts in an interactive mode:
+
+    Gnomovision version 69, Copyright (C) year name of author
+    Gnomovision comes with ABSOLUTELY NO WARRANTY; for details type `show w'.
+    This is free software, and you are welcome to redistribute it
+    under certain conditions; type `show c' for details.
+
+The hypothetical commands `show w' and `show c' should show the appropriate
+parts of the General Public License.  Of course, the commands you use may
+be called something other than `show w' and `show c'; they could even be
+mouse-clicks or menu items--whatever suits your program.
+
+You should also get your employer (if you work as a programmer) or your
+school, if any, to sign a "copyright disclaimer" for the program, if
+necessary.  Here is a sample; alter the names:
+
+  Yoyodyne, Inc., hereby disclaims all copyright interest in the program
+  `Gnomovision' (which makes passes at compilers) written by James Hacker.
+
+  <signature of Ty Coon>, 1 April 1989
+  Ty Coon, President of Vice
+
+This General Public License does not permit incorporating your program into
+proprietary programs.  If your program is a subroutine library, you may
+consider it more useful to permit linking proprietary applications with the
+library.  If this is what you want to do, use the GNU Lesser General
+Public License instead of this License.
diff --git a/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/README b/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/README
new file mode 100644
index 000000000..2ea86bf46
--- /dev/null
+++ b/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/README
@@ -0,0 +1,13 @@
+The files in this directory originally come from
+https://github.com/percona/PerconaFT/.
+
+This directory only includes the "locktree" part of PerconaFT, and its
+dependencies.
+
+The following modifications were made:
+- Make locktree usable outside of PerconaFT library
+- Add shared read-only lock support
+
+The files named *_subst.* are substitutes of the PerconaFT's files, they
+contain replacements of PerconaFT's functionality.
+
diff --git a/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/db.h b/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/db.h
new file mode 100644
index 000000000..5aa826c8e
--- /dev/null
+++ b/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/db.h
@@ -0,0 +1,76 @@
+#ifndef _DB_H
+#define _DB_H
+
+#include <stdint.h>
+#include <sys/types.h>
+
+typedef struct __toku_dbt DBT;
+
+// port: this is currently not used
+struct simple_dbt {
+  uint32_t len;
+  void *data;
+};
+
+// engine status info
+// engine status is passed to handlerton as an array of
+// TOKU_ENGINE_STATUS_ROW_S[]
+typedef enum {
+  STATUS_FS_STATE = 0,  // interpret as file system state (redzone) enum
+  STATUS_UINT64,        // interpret as uint64_t
+  STATUS_CHARSTR,       // interpret as char *
+  STATUS_UNIXTIME,      // interpret as time_t
+  STATUS_TOKUTIME,      // interpret as tokutime_t
+  STATUS_PARCOUNT,      // interpret as PARTITIONED_COUNTER
+  STATUS_DOUBLE         // interpret as double
+} toku_engine_status_display_type;
+
+typedef enum {
+  TOKU_ENGINE_STATUS = (1ULL << 0),  // Include when asking for engine status
+  TOKU_GLOBAL_STATUS =
+      (1ULL << 1),  // Include when asking for information_schema.global_status
+} toku_engine_status_include_type;
+
+typedef struct __toku_engine_status_row {
+  const char *keyname;  // info schema key, should not change across revisions
+                        // without good reason
+  const char
+      *columnname;  // column for mysql, e.g. information_schema.global_status.
+                    // TOKUDB_ will automatically be prefixed.
+  const char *legend;  // the text that will appear at user interface
+  toku_engine_status_display_type type;  // how to interpret the value
+  toku_engine_status_include_type
+      include;  // which kinds of callers should get read this row?
+  union {
+    double dnum;
+    uint64_t num;
+    const char *str;
+    char datebuf[26];
+    struct partitioned_counter *parcount;
+  } value;
+} * TOKU_ENGINE_STATUS_ROW, TOKU_ENGINE_STATUS_ROW_S;
+
+#define DB_BUFFER_SMALL -30999
+#define DB_LOCK_DEADLOCK -30995
+#define DB_LOCK_NOTGRANTED -30994
+#define DB_NOTFOUND -30989
+#define DB_KEYEXIST -30996
+#define DB_DBT_MALLOC 8
+#define DB_DBT_REALLOC 64
+#define DB_DBT_USERMEM 256
+
+/* PerconaFT specific error codes */
+#define TOKUDB_OUT_OF_LOCKS -100000
+
+typedef void (*lock_wait_callback)(void *arg, uint64_t requesting_txnid,
+                                   uint64_t blocking_txnid);
+
+struct __toku_dbt {
+  void *data;
+  size_t size;
+  size_t ulen;
+  // One of DB_DBT_XXX flags
+  uint32_t flags;
+};
+
+#endif
diff --git a/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/ft/comparator.h b/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/ft/comparator.h
new file mode 100644
index 000000000..718efc623
--- /dev/null
+++ b/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/ft/comparator.h
@@ -0,0 +1,138 @@
+/* -*- mode: C; c-basic-offset: 4; indent-tabs-mode: nil -*- */
+// vim: ft=cpp:expandtab:ts=8:sw=4:softtabstop=4:
+/*======
+This file is part of PerconaFT.
+
+
+Copyright (c) 2006, 2015, Percona and/or its affiliates. All rights reserved.
+
+    PerconaFT is free software: you can redistribute it and/or modify
+    it under the terms of the GNU General Public License, version 2,
+    as published by the Free Software Foundation.
+
+    PerconaFT is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License
+    along with PerconaFT.  If not, see <http://www.gnu.org/licenses/>.
+
+----------------------------------------
+
+    PerconaFT is free software: you can redistribute it and/or modify
+    it under the terms of the GNU Affero General Public License, version 3,
+    as published by the Free Software Foundation.
+
+    PerconaFT is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU Affero General Public License for more details.
+
+    You should have received a copy of the GNU Affero General Public License
+    along with PerconaFT.  If not, see <http://www.gnu.org/licenses/>.
+
+----------------------------------------
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.
+======= */
+
+#ident \
+    "Copyright (c) 2006, 2015, Percona and/or its affiliates. All rights reserved."
+
+#pragma once
+
+#include <string.h>
+
+#include "../db.h"
+#include "../portability/memory.h"
+#include "../util/dbt.h"
+
+typedef int (*ft_compare_func)(void *arg, const DBT *a, const DBT *b);
+
+int toku_keycompare(const void *key1, size_t key1len, const void *key2,
+                    size_t key2len);
+
+int toku_builtin_compare_fun(const DBT *, const DBT *)
+    __attribute__((__visibility__("default")));
+
+namespace toku {
+
+// a comparator object encapsulates the data necessary for
+// comparing two keys in a fractal tree. it further understands
+// that points may be positive or negative infinity.
+
+class comparator {
+  void init(ft_compare_func cmp, void *cmp_arg, uint8_t memcmp_magic) {
+    _cmp = cmp;
+    _cmp_arg = cmp_arg;
+    _memcmp_magic = memcmp_magic;
+  }
+
+ public:
+  // This magic value is reserved to mean that the magic has not been set.
+  static const uint8_t MEMCMP_MAGIC_NONE = 0;
+
+  void create(ft_compare_func cmp, void *cmp_arg,
+              uint8_t memcmp_magic = MEMCMP_MAGIC_NONE) {
+    init(cmp, cmp_arg, memcmp_magic);
+  }
+
+  // inherit the attributes of another comparator, but keep our own
+  // copy of fake_db that is owned separately from the one given.
+  void inherit(const comparator &cmp) {
+    invariant_notnull(cmp._cmp);
+    init(cmp._cmp, cmp._cmp_arg, cmp._memcmp_magic);
+  }
+
+  // like inherit, but doesn't require that the this comparator
+  // was already created
+  void create_from(const comparator &cmp) { inherit(cmp); }
+
+  void destroy() {}
+
+  ft_compare_func get_compare_func() const { return _cmp; }
+
+  uint8_t get_memcmp_magic() const { return _memcmp_magic; }
+
+  bool valid() const { return _cmp != nullptr; }
+
+  inline bool dbt_has_memcmp_magic(const DBT *dbt) const {
+    return *reinterpret_cast<const char *>(dbt->data) == _memcmp_magic;
+  }
+
+  int operator()(const DBT *a, const DBT *b) const {
+    if (__builtin_expect(toku_dbt_is_infinite(a) || toku_dbt_is_infinite(b),
+                         0)) {
+      return toku_dbt_infinite_compare(a, b);
+    } else if (_memcmp_magic != MEMCMP_MAGIC_NONE
+               // If `a' has the memcmp magic..
+               && dbt_has_memcmp_magic(a)
+               // ..then we expect `b' to also have the memcmp magic
+               && __builtin_expect(dbt_has_memcmp_magic(b), 1)) {
+      assert(0);  // psergey: this branch should not be taken.
+      return toku_builtin_compare_fun(a, b);
+    } else {
+      // yikes, const sadness here
+      return _cmp(_cmp_arg, a, b);
+    }
+  }
+
+ private:
+  ft_compare_func _cmp;
+  void *_cmp_arg;
+
+  uint8_t _memcmp_magic;
+};
+
+} /* namespace toku */
diff --git a/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/ft/ft-status.h b/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/ft/ft-status.h
new file mode 100644
index 000000000..1b4511172
--- /dev/null
+++ b/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/ft/ft-status.h
@@ -0,0 +1,102 @@
+/* -*- mode: C++; c-basic-offset: 4; indent-tabs-mode: nil -*- */
+// vim: ft=cpp:expandtab:ts=8:sw=4:softtabstop=4:
+#ident "$Id$"
+/*======
+This file is part of PerconaFT.
+
+
+Copyright (c) 2006, 2015, Percona and/or its affiliates. All rights reserved.
+
+    PerconaFT is free software: you can redistribute it and/or modify
+    it under the terms of the GNU General Public License, version 2,
+    as published by the Free Software Foundation.
+
+    PerconaFT is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License
+    along with PerconaFT.  If not, see <http://www.gnu.org/licenses/>.
+
+----------------------------------------
+
+    PerconaFT is free software: you can redistribute it and/or modify
+    it under the terms of the GNU Affero General Public License, version 3,
+    as published by the Free Software Foundation.
+
+    PerconaFT is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU Affero General Public License for more details.
+
+    You should have received a copy of the GNU Affero General Public License
+    along with PerconaFT.  If not, see <http://www.gnu.org/licenses/>.
+
+----------------------------------------
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.
+======= */
+
+#ident \
+    "Copyright (c) 2006, 2015, Percona and/or its affiliates. All rights reserved."
+
+#pragma once
+
+#include "../db.h"
+#include "../portability/toku_race_tools.h"
+#include "../util/status.h"
+
+//
+// Lock Tree Manager statistics
+//
+class LTM_STATUS_S {
+ public:
+  enum {
+    LTM_SIZE_CURRENT = 0,
+    LTM_SIZE_LIMIT,
+    LTM_ESCALATION_COUNT,
+    LTM_ESCALATION_TIME,
+    LTM_ESCALATION_LATEST_RESULT,
+    LTM_NUM_LOCKTREES,
+    LTM_LOCK_REQUESTS_PENDING,
+    LTM_STO_NUM_ELIGIBLE,
+    LTM_STO_END_EARLY_COUNT,
+    LTM_STO_END_EARLY_TIME,
+    LTM_WAIT_COUNT,
+    LTM_WAIT_TIME,
+    LTM_LONG_WAIT_COUNT,
+    LTM_LONG_WAIT_TIME,
+    LTM_TIMEOUT_COUNT,
+    LTM_WAIT_ESCALATION_COUNT,
+    LTM_WAIT_ESCALATION_TIME,
+    LTM_LONG_WAIT_ESCALATION_COUNT,
+    LTM_LONG_WAIT_ESCALATION_TIME,
+    LTM_STATUS_NUM_ROWS  // must be last
+  };
+
+  void init(void);
+  void destroy(void);
+
+  TOKU_ENGINE_STATUS_ROW_S status[LTM_STATUS_NUM_ROWS];
+
+ private:
+  bool m_initialized = false;
+};
+typedef LTM_STATUS_S* LTM_STATUS;
+extern LTM_STATUS_S ltm_status;
+
+#define LTM_STATUS_VAL(x) ltm_status.status[LTM_STATUS_S::x].value.num
+
+void toku_status_init(void);     // just call ltm_status.init();
+void toku_status_destroy(void);  // just call ltm_status.destroy();
diff --git a/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/locktree/concurrent_tree.cc b/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/locktree/concurrent_tree.cc
new file mode 100644
index 000000000..5110cd482
--- /dev/null
+++ b/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/locktree/concurrent_tree.cc
@@ -0,0 +1,139 @@
+/* -*- mode: C++; c-basic-offset: 4; indent-tabs-mode: nil -*- */
+// vim: ft=cpp:expandtab:ts=8:sw=4:softtabstop=4:
+#ifndef ROCKSDB_LITE
+#ifndef OS_WIN
+#ident "$Id$"
+/*======
+This file is part of PerconaFT.
+
+
+Copyright (c) 2006, 2015, Percona and/or its affiliates. All rights reserved.
+
+    PerconaFT is free software: you can redistribute it and/or modify
+    it under the terms of the GNU General Public License, version 2,
+    as published by the Free Software Foundation.
+
+    PerconaFT is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License
+    along with PerconaFT.  If not, see <http://www.gnu.org/licenses/>.
+
+----------------------------------------
+
+    PerconaFT is free software: you can redistribute it and/or modify
+    it under the terms of the GNU Affero General Public License, version 3,
+    as published by the Free Software Foundation.
+
+    PerconaFT is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU Affero General Public License for more details.
+
+    You should have received a copy of the GNU Affero General Public License
+    along with PerconaFT.  If not, see <http://www.gnu.org/licenses/>.
+
+----------------------------------------
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.
+======= */
+
+#ident \
+    "Copyright (c) 2006, 2015, Percona and/or its affiliates. All rights reserved."
+
+#include "concurrent_tree.h"
+
+// PORT #include <toku_assert.h>
+namespace toku {
+
+void concurrent_tree::create(const comparator *cmp) {
+  // start with an empty root node. we do this instead of
+  // setting m_root to null so there's always a root to lock
+  m_root.create_root(cmp);
+}
+
+void concurrent_tree::destroy(void) { m_root.destroy_root(); }
+
+bool concurrent_tree::is_empty(void) { return m_root.is_empty(); }
+
+uint64_t concurrent_tree::get_insertion_memory_overhead(void) {
+  return sizeof(treenode);
+}
+
+void concurrent_tree::locked_keyrange::prepare(concurrent_tree *tree) {
+  // the first step in acquiring a locked keyrange is locking the root
+  treenode *const root = &tree->m_root;
+  m_tree = tree;
+  m_subtree = root;
+  m_range = keyrange::get_infinite_range();
+  root->mutex_lock();
+}
+
+void concurrent_tree::locked_keyrange::acquire(const keyrange &range) {
+  treenode *const root = &m_tree->m_root;
+
+  treenode *subtree;
+  if (root->is_empty() || root->range_overlaps(range)) {
+    subtree = root;
+  } else {
+    // we do not have a precomputed comparison hint, so pass null
+    const keyrange::comparison *cmp_hint = nullptr;
+    subtree = root->find_node_with_overlapping_child(range, cmp_hint);
+  }
+
+  // subtree is locked. it will be unlocked when this is release()'d
+  invariant_notnull(subtree);
+  m_range = range;
+  m_subtree = subtree;
+}
+
+bool concurrent_tree::locked_keyrange::add_shared_owner(const keyrange &range,
+                                                        TXNID new_owner) {
+  return m_subtree->insert(range, new_owner, /*is_shared*/ true);
+}
+
+void concurrent_tree::locked_keyrange::release(void) {
+  m_subtree->mutex_unlock();
+}
+
+void concurrent_tree::locked_keyrange::insert(const keyrange &range,
+                                              TXNID txnid, bool is_shared) {
+  // empty means no children, and only the root should ever be empty
+  if (m_subtree->is_empty()) {
+    m_subtree->set_range_and_txnid(range, txnid, is_shared);
+  } else {
+    m_subtree->insert(range, txnid, is_shared);
+  }
+}
+
+void concurrent_tree::locked_keyrange::remove(const keyrange &range,
+                                              TXNID txnid) {
+  invariant(!m_subtree->is_empty());
+  treenode *new_subtree = m_subtree->remove(range, txnid);
+  // if removing range changed the root of the subtree,
+  // then the subtree must be the root of the entire tree.
+  if (new_subtree == nullptr) {
+    invariant(m_subtree->is_root());
+    invariant(m_subtree->is_empty());
+  }
+}
+
+void concurrent_tree::locked_keyrange::remove_all(void) {
+  m_subtree->recursive_remove();
+}
+
+} /* namespace toku */
+#endif  // OS_WIN
+#endif  // ROCKSDB_LITE
diff --git a/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/locktree/concurrent_tree.h b/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/locktree/concurrent_tree.h
new file mode 100644
index 000000000..e1bfb86c5
--- /dev/null
+++ b/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/locktree/concurrent_tree.h
@@ -0,0 +1,174 @@
+/* -*- mode: C++; c-basic-offset: 4; indent-tabs-mode: nil -*- */
+// vim: ft=cpp:expandtab:ts=8:sw=2:softtabstop=2:
+#ident "$Id$"
+/*======
+This file is part of PerconaFT.
+
+
+Copyright (c) 2006, 2015, Percona and/or its affiliates. All rights reserved.
+
+    PerconaFT is free software: you can redistribute it and/or modify
+    it under the terms of the GNU General Public License, version 2,
+    as published by the Free Software Foundation.
+
+    PerconaFT is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License
+    along with PerconaFT.  If not, see <http://www.gnu.org/licenses/>.
+
+----------------------------------------
+
+    PerconaFT is free software: you can redistribute it and/or modify
+    it under the terms of the GNU Affero General Public License, version 3,
+    as published by the Free Software Foundation.
+
+    PerconaFT is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU Affero General Public License for more details.
+
+    You should have received a copy of the GNU Affero General Public License
+    along with PerconaFT.  If not, see <http://www.gnu.org/licenses/>.
+
+----------------------------------------
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.
+======= */
+
+#ident \
+    "Copyright (c) 2006, 2015, Percona and/or its affiliates. All rights reserved."
+
+#pragma once
+
+#include "../ft/comparator.h"
+#include "keyrange.h"
+#include "treenode.h"
+
+namespace toku {
+
+// A concurrent_tree stores non-overlapping ranges.
+// Access to disjoint parts of the tree usually occurs concurrently.
+
+class concurrent_tree {
+ public:
+  // A locked_keyrange gives you exclusive access to read and write
+  // operations that occur on any keys in that range. You only have
+  // the right to operate on keys in that range or keys that were read
+  // from the keyrange using iterate()
+  //
+  // Access model:
+  // - user prepares a locked keyrange. all threads serialize behind prepare().
+  // - user breaks the serialzation point by acquiring a range, or releasing.
+  // - one thread operates on a certain locked_keyrange object at a time.
+  // - when the thread is finished, it releases
+
+  class locked_keyrange {
+   public:
+    // effect: prepare to acquire a locked keyrange over the given
+    //         concurrent_tree, preventing other threads from preparing
+    //         until this thread either does acquire() or release().
+    // note: operations performed on a prepared keyrange are equivalent
+    //         to ones performed on an acquired keyrange over -inf, +inf.
+    // rationale: this provides the user with a serialization point for
+    // descending
+    //            or modifying the the tree. it also proives a convenient way of
+    //            doing serializable operations on the tree.
+    // There are two valid sequences of calls:
+    //  - prepare, acquire, [operations], release
+    //  - prepare, [operations],release
+    void prepare(concurrent_tree *tree);
+
+    // requires: the locked keyrange was prepare()'d
+    // effect: acquire a locked keyrange over the given concurrent_tree.
+    //         the locked keyrange represents the range of keys overlapped
+    //         by the given range
+    void acquire(const keyrange &range);
+
+    // effect: releases a locked keyrange and the mutex it holds
+    void release(void);
+
+    // effect: iterate over each range this locked_keyrange represents,
+    //         calling function->fn() on each node's keyrange and txnid
+    //         until there are no more or the function returns false
+    template <class F>
+    void iterate(F *function) const {
+      // if the subtree is non-empty, traverse it by calling the given
+      // function on each range, txnid pair found that overlaps.
+      if (!m_subtree->is_empty()) {
+        m_subtree->traverse_overlaps(m_range, function);
+      }
+    }
+
+    // Adds another owner to the lock on the specified keyrange.
+    // requires: the keyrange contains one treenode whose bounds are
+    //           exactly equal to the specifed range (no sub/supersets)
+    bool add_shared_owner(const keyrange &range, TXNID new_owner);
+
+    // inserts the given range into the tree, with an associated txnid.
+    // requires: range does not overlap with anything in this locked_keyrange
+    // rationale: caller is responsible for only inserting unique ranges
+    void insert(const keyrange &range, TXNID txnid, bool is_shared);
+
+    // effect: removes the given range from the tree.
+    //         - txnid=TXNID_ANY means remove the range no matter what its
+    //           owners are
+    //         - Other value means remove the specified txnid from
+    //           ownership (if the range has other owners, it will remain
+    //           in the tree)
+    // requires: range exists exactly in this locked_keyrange
+    // rationale: caller is responsible for only removing existing ranges
+    void remove(const keyrange &range, TXNID txnid);
+
+    // effect: removes all of the keys represented by this locked keyrange
+    // rationale: we'd like a fast way to empty out a tree
+    void remove_all(void);
+
+   private:
+    // the concurrent tree this locked keyrange is for
+    concurrent_tree *m_tree;
+
+    // the range of keys this locked keyrange represents
+    keyrange m_range;
+
+    // the subtree under which all overlapping ranges exist
+    treenode *m_subtree;
+
+    friend class concurrent_tree_unit_test;
+  };
+
+  // effect: initialize the tree to an empty state
+  void create(const comparator *cmp);
+
+  // effect: destroy the tree.
+  // requires: tree is empty
+  void destroy(void);
+
+  // returns: true iff the tree is empty
+  bool is_empty(void);
+
+  // returns: the memory overhead of a single insertion into the tree
+  static uint64_t get_insertion_memory_overhead(void);
+
+ private:
+  // the root needs to always exist so there's a lock to grab
+  // even if the tree is empty. that's why we store a treenode
+  // here and not a pointer to one.
+  treenode m_root;
+
+  friend class concurrent_tree_unit_test;
+};
+
+} /* namespace toku */
diff --git a/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/locktree/keyrange.cc b/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/locktree/keyrange.cc
new file mode 100644
index 000000000..e50ace5a9
--- /dev/null
+++ b/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/locktree/keyrange.cc
@@ -0,0 +1,222 @@
+/* -*- mode: C++; c-basic-offset: 4; indent-tabs-mode: nil -*- */
+// vim: ft=cpp:expandtab:ts=8:sw=4:softtabstop=4:
+#ifndef ROCKSDB_LITE
+#ifndef OS_WIN
+#ident "$Id$"
+/*======
+This file is part of PerconaFT.
+
+
+Copyright (c) 2006, 2015, Percona and/or its affiliates. All rights reserved.
+
+    PerconaFT is free software: you can redistribute it and/or modify
+    it under the terms of the GNU General Public License, version 2,
+    as published by the Free Software Foundation.
+
+    PerconaFT is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License
+    along with PerconaFT.  If not, see <http://www.gnu.org/licenses/>.
+
+----------------------------------------
+
+    PerconaFT is free software: you can redistribute it and/or modify
+    it under the terms of the GNU Affero General Public License, version 3,
+    as published by the Free Software Foundation.
+
+    PerconaFT is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU Affero General Public License for more details.
+
+    You should have received a copy of the GNU Affero General Public License
+    along with PerconaFT.  If not, see <http://www.gnu.org/licenses/>.
+
+----------------------------------------
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.
+======= */
+
+#ident \
+    "Copyright (c) 2006, 2015, Percona and/or its affiliates. All rights reserved."
+
+#include "keyrange.h"
+
+#include "../util/dbt.h"
+
+namespace toku {
+
+// create a keyrange by borrowing the left and right dbt
+// pointers. no memory is copied. no checks for infinity needed.
+void keyrange::create(const DBT *left, const DBT *right) {
+  init_empty();
+  m_left_key = left;
+  m_right_key = right;
+}
+
+// destroy the key copies. if they were never set, then destroy does nothing.
+void keyrange::destroy(void) {
+  toku_destroy_dbt(&m_left_key_copy);
+  toku_destroy_dbt(&m_right_key_copy);
+}
+
+// create a keyrange by copying the keys from the given range.
+void keyrange::create_copy(const keyrange &range) {
+  // start with an initialized, empty range
+  init_empty();
+
+  // optimize the case where the left and right keys are the same.
+  // we'd like to only have one copy of the data.
+  if (toku_dbt_equals(range.get_left_key(), range.get_right_key())) {
+    set_both_keys(range.get_left_key());
+  } else {
+    // replace our empty left and right keys with
+    // copies of the range's left and right keys
+    replace_left_key(range.get_left_key());
+    replace_right_key(range.get_right_key());
+  }
+}
+
+// extend this keyrange by choosing the leftmost and rightmost
+// endpoints between this range and the given. replaced keys
+// in this range are freed and inherited keys are copied.
+void keyrange::extend(const comparator &cmp, const keyrange &range) {
+  const DBT *range_left = range.get_left_key();
+  const DBT *range_right = range.get_right_key();
+  if (cmp(range_left, get_left_key()) < 0) {
+    replace_left_key(range_left);
+  }
+  if (cmp(range_right, get_right_key()) > 0) {
+    replace_right_key(range_right);
+  }
+}
+
+// how much memory does this keyrange take?
+// - the size of the left and right keys
+// --- ignore the fact that we may have optimized the point case.
+//     it complicates things for little gain.
+// - the size of the keyrange class itself
+uint64_t keyrange::get_memory_size(void) const {
+  const DBT *left_key = get_left_key();
+  const DBT *right_key = get_right_key();
+  return left_key->size + right_key->size + sizeof(keyrange);
+}
+
+// compare ranges.
+keyrange::comparison keyrange::compare(const comparator &cmp,
+                                       const keyrange &range) const {
+  if (cmp(get_right_key(), range.get_left_key()) < 0) {
+    return comparison::LESS_THAN;
+  } else if (cmp(get_left_key(), range.get_right_key()) > 0) {
+    return comparison::GREATER_THAN;
+  } else if (cmp(get_left_key(), range.get_left_key()) == 0 &&
+             cmp(get_right_key(), range.get_right_key()) == 0) {
+    return comparison::EQUALS;
+  } else {
+    return comparison::OVERLAPS;
+  }
+}
+
+bool keyrange::overlaps(const comparator &cmp, const keyrange &range) const {
+  // equality is a stronger form of overlapping.
+  // so two ranges "overlap" if they're either equal or just overlapping.
+  comparison c = compare(cmp, range);
+  return c == comparison::EQUALS || c == comparison::OVERLAPS;
+}
+
+keyrange keyrange::get_infinite_range(void) {
+  keyrange range;
+  range.create(toku_dbt_negative_infinity(), toku_dbt_positive_infinity());
+  return range;
+}
+
+void keyrange::init_empty(void) {
+  m_left_key = nullptr;
+  m_right_key = nullptr;
+  toku_init_dbt(&m_left_key_copy);
+  toku_init_dbt(&m_right_key_copy);
+  m_point_range = false;
+}
+
+const DBT *keyrange::get_left_key(void) const {
+  if (m_left_key) {
+    return m_left_key;
+  } else {
+    return &m_left_key_copy;
+  }
+}
+
+const DBT *keyrange::get_right_key(void) const {
+  if (m_right_key) {
+    return m_right_key;
+  } else {
+    return &m_right_key_copy;
+  }
+}
+
+// copy the given once and set both the left and right pointers.
+// optimization for point ranges, so the left and right ranges
+// are not copied twice.
+void keyrange::set_both_keys(const DBT *key) {
+  if (toku_dbt_is_infinite(key)) {
+    m_left_key = key;
+    m_right_key = key;
+  } else {
+    toku_clone_dbt(&m_left_key_copy, *key);
+    toku_copyref_dbt(&m_right_key_copy, m_left_key_copy);
+  }
+  m_point_range = true;
+}
+
+// destroy the current left key. set and possibly copy the new one
+void keyrange::replace_left_key(const DBT *key) {
+  // a little magic:
+  //
+  // if this is a point range, then the left and right keys share
+  // one copy of the data, and it lives in the left key copy. so
+  // if we're replacing the left key, move the real data to the
+  // right key copy instead of destroying it. now, the memory is
+  // owned by the right key and the left key may be replaced.
+  if (m_point_range) {
+    m_right_key_copy = m_left_key_copy;
+  } else {
+    toku_destroy_dbt(&m_left_key_copy);
+  }
+
+  if (toku_dbt_is_infinite(key)) {
+    m_left_key = key;
+  } else {
+    toku_clone_dbt(&m_left_key_copy, *key);
+    m_left_key = nullptr;
+  }
+  m_point_range = false;
+}
+
+// destroy the current right key. set and possibly copy the new one
+void keyrange::replace_right_key(const DBT *key) {
+  toku_destroy_dbt(&m_right_key_copy);
+  if (toku_dbt_is_infinite(key)) {
+    m_right_key = key;
+  } else {
+    toku_clone_dbt(&m_right_key_copy, *key);
+    m_right_key = nullptr;
+  }
+  m_point_range = false;
+}
+
+} /* namespace toku */
+#endif  // OS_WIN
+#endif  // ROCKSDB_LITE
diff --git a/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/locktree/keyrange.h b/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/locktree/keyrange.h
new file mode 100644
index 000000000..f9aeea0c4
--- /dev/null
+++ b/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/locktree/keyrange.h
@@ -0,0 +1,141 @@
+/* -*- mode: C++; c-basic-offset: 4; indent-tabs-mode: nil -*- */
+// vim: ft=cpp:expandtab:ts=8:sw=4:softtabstop=4:
+#ident "$Id$"
+/*======
+This file is part of PerconaFT.
+
+
+Copyright (c) 2006, 2015, Percona and/or its affiliates. All rights reserved.
+
+    PerconaFT is free software: you can redistribute it and/or modify
+    it under the terms of the GNU General Public License, version 2,
+    as published by the Free Software Foundation.
+
+    PerconaFT is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License
+    along with PerconaFT.  If not, see <http://www.gnu.org/licenses/>.
+
+----------------------------------------
+
+    PerconaFT is free software: you can redistribute it and/or modify
+    it under the terms of the GNU Affero General Public License, version 3,
+    as published by the Free Software Foundation.
+
+    PerconaFT is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU Affero General Public License for more details.
+
+    You should have received a copy of the GNU Affero General Public License
+    along with PerconaFT.  If not, see <http://www.gnu.org/licenses/>.
+
+----------------------------------------
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.
+======= */
+
+#ident \
+    "Copyright (c) 2006, 2015, Percona and/or its affiliates. All rights reserved."
+
+#pragma once
+
+#include "../ft/comparator.h"
+
+namespace toku {
+
+// A keyrange has a left and right key as endpoints.
+//
+// When a keyrange is created it owns no memory, but when it copies
+// or extends another keyrange, it copies memory as necessary. This
+// means it is cheap in the common case.
+
+class keyrange {
+ public:
+  // effect: constructor that borrows left and right key pointers.
+  //         no memory is allocated or copied.
+  void create(const DBT *left_key, const DBT *right_key);
+
+  // effect: constructor that allocates and copies another keyrange's points.
+  void create_copy(const keyrange &range);
+
+  // effect: destroys the keyrange, freeing any allocated memory
+  void destroy(void);
+
+  // effect: extends the keyrange by choosing the leftmost and rightmost
+  //         endpoints from this range and the given range.
+  //         replaced keys in this range are freed, new keys are copied.
+  void extend(const comparator &cmp, const keyrange &range);
+
+  // returns: the amount of memory this keyrange takes. does not account
+  //          for point optimizations or malloc overhead.
+  uint64_t get_memory_size(void) const;
+
+  // returns: pointer to the left key of this range
+  const DBT *get_left_key(void) const;
+
+  // returns: pointer to the right key of this range
+  const DBT *get_right_key(void) const;
+
+  // two ranges are either equal, lt, gt, or overlapping
+  enum comparison { EQUALS, LESS_THAN, GREATER_THAN, OVERLAPS };
+
+  // effect: compares this range to the given range
+  // returns: LESS_THAN    if given range is strictly to the left
+  //          GREATER_THAN if given range is strictly to the right
+  //          EQUALS       if given range has the same left and right endpoints
+  //          OVERLAPS     if at least one of the given range's endpoints falls
+  //                       between this range's endpoints
+  comparison compare(const comparator &cmp, const keyrange &range) const;
+
+  // returns: true if the range and the given range are equal or overlapping
+  bool overlaps(const comparator &cmp, const keyrange &range) const;
+
+  // returns: a keyrange representing -inf, +inf
+  static keyrange get_infinite_range(void);
+
+ private:
+  // some keys should be copied, some keys should not be.
+  //
+  // to support both, we use two DBTs for copies and two pointers
+  // for temporaries. the access rule is:
+  // - if a pointer is non-null, then it reprsents the key.
+  // - otherwise the pointer is null, and the key is in the copy.
+  DBT m_left_key_copy;
+  DBT m_right_key_copy;
+  const DBT *m_left_key;
+  const DBT *m_right_key;
+
+  // if this range is a point range, then m_left_key == m_right_key
+  // and the actual data is stored exactly once in m_left_key_copy.
+  bool m_point_range;
+
+  // effect: initializes a keyrange to be empty
+  void init_empty(void);
+
+  // effect: copies the given key once into the left key copy
+  //         and sets the right key copy to share the left.
+  // rationale: optimization for point ranges to only do one malloc
+  void set_both_keys(const DBT *key);
+
+  // effect: destroys the current left key. sets and copies the new one.
+  void replace_left_key(const DBT *key);
+
+  // effect: destroys the current right key. sets and copies the new one.
+  void replace_right_key(const DBT *key);
+};
+
+} /* namespace toku */
diff --git a/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/locktree/lock_request.cc b/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/locktree/lock_request.cc
new file mode 100644
index 000000000..3d217be70
--- /dev/null
+++ b/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/locktree/lock_request.cc
@@ -0,0 +1,527 @@
+/* -*- mode: C++; c-basic-offset: 4; indent-tabs-mode: nil -*- */
+// vim: ft=cpp:expandtab:ts=8:sw=2:softtabstop=2:
+#ifndef ROCKSDB_LITE
+#ifndef OS_WIN
+#ident "$Id$"
+/*======
+This file is part of PerconaFT.
+
+
+Copyright (c) 2006, 2015, Percona and/or its affiliates. All rights reserved.
+
+    PerconaFT is free software: you can redistribute it and/or modify
+    it under the terms of the GNU General Public License, version 2,
+    as published by the Free Software Foundation.
+
+    PerconaFT is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License
+    along with PerconaFT.  If not, see <http://www.gnu.org/licenses/>.
+
+----------------------------------------
+
+    PerconaFT is free software: you can redistribute it and/or modify
+    it under the terms of the GNU Affero General Public License, version 3,
+    as published by the Free Software Foundation.
+
+    PerconaFT is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU Affero General Public License for more details.
+
+    You should have received a copy of the GNU Affero General Public License
+    along with PerconaFT.  If not, see <http://www.gnu.org/licenses/>.
+
+----------------------------------------
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.
+======= */
+
+#ident \
+    "Copyright (c) 2006, 2015, Percona and/or its affiliates. All rights reserved."
+
+#include "lock_request.h"
+
+#include "../portability/toku_race_tools.h"
+#include "../portability/txn_subst.h"
+#include "../util/dbt.h"
+#include "locktree.h"
+
+namespace toku {
+
+// initialize a lock request's internals
+void lock_request::create(toku_external_mutex_factory_t mutex_factory) {
+  m_txnid = TXNID_NONE;
+  m_conflicting_txnid = TXNID_NONE;
+  m_start_time = 0;
+  m_left_key = nullptr;
+  m_right_key = nullptr;
+  toku_init_dbt(&m_left_key_copy);
+  toku_init_dbt(&m_right_key_copy);
+
+  m_type = type::UNKNOWN;
+  m_lt = nullptr;
+
+  m_complete_r = 0;
+  m_state = state::UNINITIALIZED;
+  m_info = nullptr;
+
+  // psergey-todo: this condition is for interruptible wait
+  // note: moved to here from lock_request::create:
+  toku_external_cond_init(mutex_factory, &m_wait_cond);
+
+  m_start_test_callback = nullptr;
+  m_start_before_pending_test_callback = nullptr;
+  m_retry_test_callback = nullptr;
+}
+
+// destroy a lock request.
+void lock_request::destroy(void) {
+  invariant(m_state != state::PENDING);
+  invariant(m_state != state::DESTROYED);
+  m_state = state::DESTROYED;
+  toku_destroy_dbt(&m_left_key_copy);
+  toku_destroy_dbt(&m_right_key_copy);
+  toku_external_cond_destroy(&m_wait_cond);
+}
+
+// set the lock request parameters. this API allows a lock request to be reused.
+void lock_request::set(locktree *lt, TXNID txnid, const DBT *left_key,
+                       const DBT *right_key, lock_request::type lock_type,
+                       bool big_txn, void *extra) {
+  invariant(m_state != state::PENDING);
+  m_lt = lt;
+
+  m_txnid = txnid;
+  m_left_key = left_key;
+  m_right_key = right_key;
+  toku_destroy_dbt(&m_left_key_copy);
+  toku_destroy_dbt(&m_right_key_copy);
+  m_type = lock_type;
+  m_state = state::INITIALIZED;
+  m_info = lt ? lt->get_lock_request_info() : nullptr;
+  m_big_txn = big_txn;
+  m_extra = extra;
+}
+
+// get rid of any stored left and right key copies and
+// replace them with copies of the given left and right key
+void lock_request::copy_keys() {
+  if (!toku_dbt_is_infinite(m_left_key)) {
+    toku_clone_dbt(&m_left_key_copy, *m_left_key);
+    m_left_key = &m_left_key_copy;
+  }
+  if (!toku_dbt_is_infinite(m_right_key)) {
+    toku_clone_dbt(&m_right_key_copy, *m_right_key);
+    m_right_key = &m_right_key_copy;
+  }
+}
+
+// what are the conflicts for this pending lock request?
+void lock_request::get_conflicts(txnid_set *conflicts) {
+  invariant(m_state == state::PENDING);
+  const bool is_write_request = m_type == type::WRITE;
+  m_lt->get_conflicts(is_write_request, m_txnid, m_left_key, m_right_key,
+                      conflicts);
+}
+
+// build a wait-for-graph for this lock request and the given conflict set
+// for each transaction B that blocks A's lock request
+//     if B is blocked then
+//         add (A,T) to the WFG and if B is new, fill in the WFG from B
+void lock_request::build_wait_graph(wfg *wait_graph,
+                                    const txnid_set &conflicts) {
+  uint32_t num_conflicts = conflicts.size();
+  for (uint32_t i = 0; i < num_conflicts; i++) {
+    TXNID conflicting_txnid = conflicts.get(i);
+    lock_request *conflicting_request = find_lock_request(conflicting_txnid);
+    invariant(conflicting_txnid != m_txnid);
+    invariant(conflicting_request != this);
+    if (conflicting_request) {
+      bool already_exists = wait_graph->node_exists(conflicting_txnid);
+      wait_graph->add_edge(m_txnid, conflicting_txnid);
+      if (!already_exists) {
+        // recursively build the wait for graph rooted at the conflicting
+        // request, given its set of lock conflicts.
+        txnid_set other_conflicts;
+        other_conflicts.create();
+        conflicting_request->get_conflicts(&other_conflicts);
+        conflicting_request->build_wait_graph(wait_graph, other_conflicts);
+        other_conflicts.destroy();
+      }
+    }
+  }
+}
+
+// returns: true if the current set of lock requests contains
+//          a deadlock, false otherwise.
+bool lock_request::deadlock_exists(const txnid_set &conflicts) {
+  wfg wait_graph;
+  wait_graph.create();
+
+  build_wait_graph(&wait_graph, conflicts);
+
+  std::function<void(TXNID)> reporter;
+  if (m_deadlock_cb) {
+    reporter = [this](TXNID a) {
+      lock_request *req = find_lock_request(a);
+      if (req) {
+        m_deadlock_cb(req->m_txnid, (req->m_type == lock_request::WRITE),
+                      req->m_left_key, req->m_right_key);
+      }
+    };
+  }
+
+  bool deadlock = wait_graph.cycle_exists_from_txnid(m_txnid, reporter);
+  wait_graph.destroy();
+  return deadlock;
+}
+
+// try to acquire a lock described by this lock request.
+int lock_request::start(void) {
+  int r;
+
+  txnid_set conflicts;
+  conflicts.create();
+  if (m_type == type::WRITE) {
+    r = m_lt->acquire_write_lock(m_txnid, m_left_key, m_right_key, &conflicts,
+                                 m_big_txn);
+  } else {
+    invariant(m_type == type::READ);
+    r = m_lt->acquire_read_lock(m_txnid, m_left_key, m_right_key, &conflicts,
+                                m_big_txn);
+  }
+
+  // if the lock is not granted, save it to the set of lock requests
+  // and check for a deadlock. if there is one, complete it as failed
+  if (r == DB_LOCK_NOTGRANTED) {
+    copy_keys();
+    m_state = state::PENDING;
+    m_start_time = toku_current_time_microsec() / 1000;
+    m_conflicting_txnid = conflicts.get(0);
+    if (m_start_before_pending_test_callback)
+      m_start_before_pending_test_callback();
+    toku_external_mutex_lock(&m_info->mutex);
+    insert_into_lock_requests();
+    if (deadlock_exists(conflicts)) {
+      remove_from_lock_requests();
+      r = DB_LOCK_DEADLOCK;
+    }
+    toku_external_mutex_unlock(&m_info->mutex);
+    if (m_start_test_callback) m_start_test_callback();  // test callback
+  }
+
+  if (r != DB_LOCK_NOTGRANTED) {
+    complete(r);
+  }
+
+  conflicts.destroy();
+  return r;
+}
+
+// sleep on the lock request until it becomes resolved or the wait time has
+// elapsed.
+int lock_request::wait(uint64_t wait_time_ms) {
+  return wait(wait_time_ms, 0, nullptr);
+}
+
+int lock_request::wait(uint64_t wait_time_ms, uint64_t killed_time_ms,
+                       int (*killed_callback)(void),
+                       void (*lock_wait_callback)(void *, lock_wait_infos *),
+                       void *callback_arg) {
+  uint64_t t_now = toku_current_time_microsec();
+  uint64_t t_start = t_now;
+  uint64_t t_end = t_start + wait_time_ms * 1000;
+
+  toku_external_mutex_lock(&m_info->mutex);
+
+  // check again, this time locking out other retry calls
+  if (m_state == state::PENDING) {
+    lock_wait_infos conflicts_collector;
+    retry(&conflicts_collector);
+    if (m_state == state::PENDING) {
+      report_waits(&conflicts_collector, lock_wait_callback, callback_arg);
+    }
+  }
+
+  while (m_state == state::PENDING) {
+    // check if this thread is killed
+    if (killed_callback && killed_callback()) {
+      remove_from_lock_requests();
+      complete(DB_LOCK_NOTGRANTED);
+      continue;
+    }
+
+    // compute the time until we should wait
+    uint64_t t_wait;
+    if (killed_time_ms == 0) {
+      t_wait = t_end;
+    } else {
+      t_wait = t_now + killed_time_ms * 1000;
+      if (t_wait > t_end) t_wait = t_end;
+    }
+
+    int r = toku_external_cond_timedwait(&m_wait_cond, &m_info->mutex,
+                                         (int64_t)(t_wait - t_now));
+    invariant(r == 0 || r == ETIMEDOUT);
+
+    t_now = toku_current_time_microsec();
+    if (m_state == state::PENDING && (t_now >= t_end)) {
+      m_info->counters.timeout_count += 1;
+
+      // if we're still pending and we timed out, then remove our
+      // request from the set of lock requests and fail.
+      remove_from_lock_requests();
+
+      // complete sets m_state to COMPLETE, breaking us out of the loop
+      complete(DB_LOCK_NOTGRANTED);
+    }
+  }
+
+  uint64_t t_real_end = toku_current_time_microsec();
+  uint64_t duration = t_real_end - t_start;
+  m_info->counters.wait_count += 1;
+  m_info->counters.wait_time += duration;
+  if (duration >= 1000000) {
+    m_info->counters.long_wait_count += 1;
+    m_info->counters.long_wait_time += duration;
+  }
+  toku_external_mutex_unlock(&m_info->mutex);
+
+  invariant(m_state == state::COMPLETE);
+  return m_complete_r;
+}
+
+// complete this lock request with the given return value
+void lock_request::complete(int complete_r) {
+  m_complete_r = complete_r;
+  m_state = state::COMPLETE;
+}
+
+const DBT *lock_request::get_left_key(void) const { return m_left_key; }
+
+const DBT *lock_request::get_right_key(void) const { return m_right_key; }
+
+TXNID lock_request::get_txnid(void) const { return m_txnid; }
+
+uint64_t lock_request::get_start_time(void) const { return m_start_time; }
+
+TXNID lock_request::get_conflicting_txnid(void) const {
+  return m_conflicting_txnid;
+}
+
+int lock_request::retry(lock_wait_infos *conflicts_collector) {
+  invariant(m_state == state::PENDING);
+  int r;
+  txnid_set conflicts;
+  conflicts.create();
+
+  if (m_type == type::WRITE) {
+    r = m_lt->acquire_write_lock(m_txnid, m_left_key, m_right_key, &conflicts,
+                                 m_big_txn);
+  } else {
+    r = m_lt->acquire_read_lock(m_txnid, m_left_key, m_right_key, &conflicts,
+                                m_big_txn);
+  }
+
+  // if the acquisition succeeded then remove ourselves from the
+  // set of lock requests, complete, and signal the waiting thread.
+  if (r == 0) {
+    remove_from_lock_requests();
+    complete(r);
+    if (m_retry_test_callback) m_retry_test_callback();  // test callback
+    toku_external_cond_broadcast(&m_wait_cond);
+  } else {
+    m_conflicting_txnid = conflicts.get(0);
+    add_conflicts_to_waits(&conflicts, conflicts_collector);
+  }
+  conflicts.destroy();
+
+  return r;
+}
+
+void lock_request::retry_all_lock_requests(
+    locktree *lt, void (*lock_wait_callback)(void *, lock_wait_infos *),
+    void *callback_arg, void (*after_retry_all_test_callback)(void)) {
+  lt_lock_request_info *info = lt->get_lock_request_info();
+
+  // if there are no pending lock requests than there is nothing to do
+  // the unlocked data race on pending_is_empty is OK since lock requests
+  // are retried after added to the pending set.
+  if (info->pending_is_empty) return;
+
+  // get my retry generation (post increment of retry_want)
+  unsigned long long my_retry_want = (info->retry_want += 1);
+
+  toku_mutex_lock(&info->retry_mutex);
+
+  // here is the group retry algorithm.
+  // get the latest retry_want count and use it as the generation number of
+  // this retry operation. if this retry generation is > the last retry
+  // generation, then do the lock retries.  otherwise, no lock retries
+  // are needed.
+  if ((my_retry_want - 1) == info->retry_done) {
+    for (;;) {
+      if (!info->running_retry) {
+        info->running_retry = true;
+        info->retry_done = info->retry_want;
+        toku_mutex_unlock(&info->retry_mutex);
+        retry_all_lock_requests_info(info, lock_wait_callback, callback_arg);
+        if (after_retry_all_test_callback) after_retry_all_test_callback();
+        toku_mutex_lock(&info->retry_mutex);
+        info->running_retry = false;
+        toku_cond_broadcast(&info->retry_cv);
+        break;
+      } else {
+        toku_cond_wait(&info->retry_cv, &info->retry_mutex);
+      }
+    }
+  }
+  toku_mutex_unlock(&info->retry_mutex);
+}
+
+void lock_request::retry_all_lock_requests_info(
+    lt_lock_request_info *info,
+    void (*lock_wait_callback)(void *, lock_wait_infos *), void *callback_arg) {
+  toku_external_mutex_lock(&info->mutex);
+  // retry all of the pending lock requests.
+  lock_wait_infos conflicts_collector;
+  for (uint32_t i = 0; i < info->pending_lock_requests.size();) {
+    lock_request *request;
+    int r = info->pending_lock_requests.fetch(i, &request);
+    invariant_zero(r);
+
+    // retry the lock request. if it didn't succeed,
+    // move on to the next lock request. otherwise
+    // the request is gone from the list so we may
+    // read the i'th entry for the next one.
+    r = request->retry(&conflicts_collector);
+    if (r != 0) {
+      i++;
+    }
+  }
+
+  // call report_waits while holding the pending queue lock since
+  // the waiter object is still valid while it's in the queue
+  report_waits(&conflicts_collector, lock_wait_callback, callback_arg);
+
+  // future threads should only retry lock requests if some still exist
+  info->should_retry_lock_requests = info->pending_lock_requests.size() > 0;
+  toku_external_mutex_unlock(&info->mutex);
+}
+
+void lock_request::add_conflicts_to_waits(txnid_set *conflicts,
+                                          lock_wait_infos *wait_conflicts) {
+  wait_conflicts->push_back({m_lt, get_txnid(), m_extra, {}});
+  uint32_t num_conflicts = conflicts->size();
+  for (uint32_t i = 0; i < num_conflicts; i++) {
+    wait_conflicts->back().waitees.push_back(conflicts->get(i));
+  }
+}
+
+void lock_request::report_waits(lock_wait_infos *wait_conflicts,
+                                void (*lock_wait_callback)(void *,
+                                                           lock_wait_infos *),
+                                void *callback_arg) {
+  if (lock_wait_callback) (*lock_wait_callback)(callback_arg, wait_conflicts);
+}
+
+void *lock_request::get_extra(void) const { return m_extra; }
+
+void lock_request::kill_waiter(void) {
+  remove_from_lock_requests();
+  complete(DB_LOCK_NOTGRANTED);
+  toku_external_cond_broadcast(&m_wait_cond);
+}
+
+void lock_request::kill_waiter(locktree *lt, void *extra) {
+  lt_lock_request_info *info = lt->get_lock_request_info();
+  toku_external_mutex_lock(&info->mutex);
+  for (uint32_t i = 0; i < info->pending_lock_requests.size(); i++) {
+    lock_request *request;
+    int r = info->pending_lock_requests.fetch(i, &request);
+    if (r == 0 && request->get_extra() == extra) {
+      request->kill_waiter();
+      break;
+    }
+  }
+  toku_external_mutex_unlock(&info->mutex);
+}
+
+// find another lock request by txnid. must hold the mutex.
+lock_request *lock_request::find_lock_request(const TXNID &txnid) {
+  lock_request *request;
+  int r = m_info->pending_lock_requests.find_zero<TXNID, find_by_txnid>(
+      txnid, &request, nullptr);
+  if (r != 0) {
+    request = nullptr;
+  }
+  return request;
+}
+
+// insert this lock request into the locktree's set. must hold the mutex.
+void lock_request::insert_into_lock_requests(void) {
+  uint32_t idx;
+  lock_request *request;
+  int r = m_info->pending_lock_requests.find_zero<TXNID, find_by_txnid>(
+      m_txnid, &request, &idx);
+  invariant(r == DB_NOTFOUND);
+  r = m_info->pending_lock_requests.insert_at(this, idx);
+  invariant_zero(r);
+  m_info->pending_is_empty = false;
+}
+
+// remove this lock request from the locktree's set. must hold the mutex.
+void lock_request::remove_from_lock_requests(void) {
+  uint32_t idx;
+  lock_request *request;
+  int r = m_info->pending_lock_requests.find_zero<TXNID, find_by_txnid>(
+      m_txnid, &request, &idx);
+  invariant_zero(r);
+  invariant(request == this);
+  r = m_info->pending_lock_requests.delete_at(idx);
+  invariant_zero(r);
+  if (m_info->pending_lock_requests.size() == 0)
+    m_info->pending_is_empty = true;
+}
+
+int lock_request::find_by_txnid(lock_request *const &request,
+                                const TXNID &txnid) {
+  TXNID request_txnid = request->m_txnid;
+  if (request_txnid < txnid) {
+    return -1;
+  } else if (request_txnid == txnid) {
+    return 0;
+  } else {
+    return 1;
+  }
+}
+
+void lock_request::set_start_test_callback(void (*f)(void)) {
+  m_start_test_callback = f;
+}
+
+void lock_request::set_start_before_pending_test_callback(void (*f)(void)) {
+  m_start_before_pending_test_callback = f;
+}
+
+void lock_request::set_retry_test_callback(void (*f)(void)) {
+  m_retry_test_callback = f;
+}
+
+} /* namespace toku */
+#endif  // OS_WIN
+#endif  // ROCKSDB_LITE
diff --git a/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/locktree/lock_request.h b/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/locktree/lock_request.h
new file mode 100644
index 000000000..d30e1e2ca
--- /dev/null
+++ b/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/locktree/lock_request.h
@@ -0,0 +1,255 @@
+/* -*- mode: C++; c-basic-offset: 4; indent-tabs-mode: nil -*- */
+// vim: ft=cpp:expandtab:ts=8:sw=2:softtabstop=2:
+#ident "$Id$"
+/*======
+This file is part of PerconaFT.
+
+
+Copyright (c) 2006, 2015, Percona and/or its affiliates. All rights reserved.
+
+    PerconaFT is free software: you can redistribute it and/or modify
+    it under the terms of the GNU General Public License, version 2,
+    as published by the Free Software Foundation.
+
+    PerconaFT is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License
+    along with PerconaFT.  If not, see <http://www.gnu.org/licenses/>.
+
+----------------------------------------
+
+    PerconaFT is free software: you can redistribute it and/or modify
+    it under the terms of the GNU Affero General Public License, version 3,
+    as published by the Free Software Foundation.
+
+    PerconaFT is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU Affero General Public License for more details.
+
+    You should have received a copy of the GNU Affero General Public License
+    along with PerconaFT.  If not, see <http://www.gnu.org/licenses/>.
+
+----------------------------------------
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.
+======= */
+
+#ident \
+    "Copyright (c) 2006, 2015, Percona and/or its affiliates. All rights reserved."
+
+#pragma once
+
+#include "../db.h"
+#include "../ft/comparator.h"
+#include "../portability/toku_pthread.h"
+#include "locktree.h"
+#include "txnid_set.h"
+#include "wfg.h"
+
+namespace toku {
+
+// Information about a lock wait
+struct lock_wait_info {
+  locktree *ltree;  // the tree where wait happens
+  TXNID waiter;     // the waiting transaction
+  void *m_extra;    // lock_request's m_extra
+
+  // The transactions that are waited for.
+  std::vector<TXNID> waitees;
+};
+
+typedef std::vector<lock_wait_info> lock_wait_infos;
+
+// A lock request contains the db, the key range, the lock type, and
+// the transaction id that describes a potential row range lock.
+//
+// the typical use case is:
+// - initialize a lock request
+// - start to try to acquire the lock
+// - do something else
+// - wait for the lock request to be resolved on a timed condition
+// - destroy the lock request
+// a lock request is resolved when its state is no longer pending, or
+// when it becomes granted, or timedout, or deadlocked. when resolved, the
+// state of the lock request is changed and any waiting threads are awakened.
+
+class lock_request {
+ public:
+  enum type { UNKNOWN, READ, WRITE };
+
+  // effect: Initializes a lock request.
+  void create(toku_external_mutex_factory_t mutex_factory);
+
+  // effect: Destroys a lock request.
+  void destroy(void);
+
+  // effect: Resets the lock request parameters, allowing it to be reused.
+  // requires: Lock request was already created at some point
+  void set(locktree *lt, TXNID txnid, const DBT *left_key, const DBT *right_key,
+           type lock_type, bool big_txn, void *extra = nullptr);
+
+  // effect: Tries to acquire a lock described by this lock request.
+  // returns: The return code of locktree::acquire_[write,read]_lock()
+  //          or DB_LOCK_DEADLOCK if this request would end up deadlocked.
+  int start(void);
+
+  // effect: Sleeps until either the request is granted or the wait time
+  // expires. returns: The return code of locktree::acquire_[write,read]_lock()
+  //          or simply DB_LOCK_NOTGRANTED if the wait time expired.
+  int wait(uint64_t wait_time_ms);
+  int wait(uint64_t wait_time_ms, uint64_t killed_time_ms,
+           int (*killed_callback)(void),
+           void (*lock_wait_callback)(void *, lock_wait_infos *) = nullptr,
+           void *callback_arg = nullptr);
+
+  // return: left end-point of the lock range
+  const DBT *get_left_key(void) const;
+
+  // return: right end-point of the lock range
+  const DBT *get_right_key(void) const;
+
+  // return: the txnid waiting for a lock
+  TXNID get_txnid(void) const;
+
+  // return: when this lock request started, as milliseconds from epoch
+  uint64_t get_start_time(void) const;
+
+  // return: which txnid is blocking this request (there may be more, though)
+  TXNID get_conflicting_txnid(void) const;
+
+  // effect: Retries all of the lock requests for the given locktree.
+  //         Any lock requests successfully restarted is completed and woken
+  //         up.
+  //         The rest remain pending.
+  static void retry_all_lock_requests(
+      locktree *lt,
+      void (*lock_wait_callback)(void *, lock_wait_infos *) = nullptr,
+      void *callback_arg = nullptr,
+      void (*after_retry_test_callback)(void) = nullptr);
+  static void retry_all_lock_requests_info(
+      lt_lock_request_info *info,
+      void (*lock_wait_callback)(void *, lock_wait_infos *),
+      void *callback_arg);
+
+  void set_start_test_callback(void (*f)(void));
+  void set_start_before_pending_test_callback(void (*f)(void));
+  void set_retry_test_callback(void (*f)(void));
+
+  void *get_extra(void) const;
+
+  void kill_waiter(void);
+  static void kill_waiter(locktree *lt, void *extra);
+
+ private:
+  enum state {
+    UNINITIALIZED,
+    INITIALIZED,
+    PENDING,
+    COMPLETE,
+    DESTROYED,
+  };
+
+  // The keys for a lock request are stored "unowned" in m_left_key
+  // and m_right_key. When the request is about to go to sleep, it
+  // copies these keys and stores them in m_left_key_copy etc and
+  // sets the temporary pointers to null.
+  TXNID m_txnid;
+  TXNID m_conflicting_txnid;
+  uint64_t m_start_time;
+  const DBT *m_left_key;
+  const DBT *m_right_key;
+  DBT m_left_key_copy;
+  DBT m_right_key_copy;
+
+  // The lock request type and associated locktree
+  type m_type;
+  locktree *m_lt;
+
+  // If the lock request is in the completed state, then its
+  // final return value is stored in m_complete_r
+  int m_complete_r;
+  state m_state;
+
+  toku_external_cond_t m_wait_cond;
+
+  bool m_big_txn;
+
+  // the lock request info state stored in the
+  // locktree that this lock request is for.
+  struct lt_lock_request_info *m_info;
+
+  void *m_extra;
+
+  // effect: tries again to acquire the lock described by this lock request
+  // returns: 0 if retrying the request succeeded and is now complete
+  int retry(lock_wait_infos *collector);
+
+  void complete(int complete_r);
+
+  // effect: Finds another lock request by txnid.
+  // requires: The lock request info mutex is held
+  lock_request *find_lock_request(const TXNID &txnid);
+
+  // effect: Insert this lock request into the locktree's set.
+  // requires: the locktree's mutex is held
+  void insert_into_lock_requests(void);
+
+  // effect: Removes this lock request from the locktree's set.
+  // requires: The lock request info mutex is held
+  void remove_from_lock_requests(void);
+
+  // effect: Asks this request's locktree which txnids are preventing
+  //         us from getting the lock described by this request.
+  // returns: conflicts is populated with the txnid's that this request
+  //          is blocked on
+  void get_conflicts(txnid_set *conflicts);
+
+  // effect: Builds a wait-for-graph for this lock request and the given
+  // conflict set
+  void build_wait_graph(wfg *wait_graph, const txnid_set &conflicts);
+
+  // returns: True if this lock request is in deadlock with the given conflicts
+  // set
+  bool deadlock_exists(const txnid_set &conflicts);
+
+  void copy_keys(void);
+
+  static int find_by_txnid(lock_request *const &request, const TXNID &txnid);
+
+  // Report list of conflicts to lock wait callback.
+  static void report_waits(lock_wait_infos *wait_conflicts,
+                           void (*lock_wait_callback)(void *,
+                                                      lock_wait_infos *),
+                           void *callback_arg);
+  void add_conflicts_to_waits(txnid_set *conflicts,
+                              lock_wait_infos *wait_conflicts);
+
+  void (*m_start_test_callback)(void);
+  void (*m_start_before_pending_test_callback)(void);
+  void (*m_retry_test_callback)(void);
+
+ public:
+  std::function<void(TXNID, bool, const DBT *, const DBT *)> m_deadlock_cb;
+
+  friend class lock_request_unit_test;
+};
+// PORT: lock_request is not a POD anymore due to use of toku_external_cond_t
+//  This is ok as the PODness is not really required: lock_request objects are
+//  not moved in memory or anything.
+// ENSURE_POD(lock_request);
+
+} /* namespace toku */
diff --git a/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/locktree/locktree.cc b/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/locktree/locktree.cc
new file mode 100644
index 000000000..3d6a590c7
--- /dev/null
+++ b/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/locktree/locktree.cc
@@ -0,0 +1,1023 @@
+/* -*- mode: C++; c-basic-offset: 4; indent-tabs-mode: nil -*- */
+// vim: ft=cpp:expandtab:ts=8:sw=2:softtabstop=2:
+#ifndef ROCKSDB_LITE
+#ifndef OS_WIN
+#ident "$Id$"
+/*======
+This file is part of PerconaFT.
+
+
+Copyright (c) 2006, 2015, Percona and/or its affiliates. All rights reserved.
+
+    PerconaFT is free software: you can redistribute it and/or modify
+    it under the terms of the GNU General Public License, version 2,
+    as published by the Free Software Foundation.
+
+    PerconaFT is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License
+    along with PerconaFT.  If not, see <http://www.gnu.org/licenses/>.
+
+----------------------------------------
+
+    PerconaFT is free software: you can redistribute it and/or modify
+    it under the terms of the GNU Affero General Public License, version 3,
+    as published by the Free Software Foundation.
+
+    PerconaFT is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU Affero General Public License for more details.
+
+    You should have received a copy of the GNU Affero General Public License
+    along with PerconaFT.  If not, see <http://www.gnu.org/licenses/>.
+
+----------------------------------------
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.
+======= */
+
+#ident \
+    "Copyright (c) 2006, 2015, Percona and/or its affiliates. All rights reserved."
+
+#include "locktree.h"
+
+#include <memory.h>
+
+#include "../portability/toku_pthread.h"
+#include "../portability/toku_time.h"
+#include "../util/growable_array.h"
+#include "range_buffer.h"
+
+// including the concurrent_tree here expands the templates
+// and "defines" the implementation, so we do it here in
+// the locktree source file instead of the header.
+#include "concurrent_tree.h"
+
+namespace toku {
+// A locktree represents the set of row locks owned by all transactions
+// over an open dictionary. Read and write ranges are represented as
+// a left and right key which are compared with the given descriptor
+// and comparison fn.
+//
+// Each locktree has a reference count which it manages
+// but does nothing based on the value of the reference count - it is
+// up to the user of the locktree to destroy it when it sees fit.
+
+void locktree::create(locktree_manager *mgr, DICTIONARY_ID dict_id,
+                      const comparator &cmp,
+                      toku_external_mutex_factory_t mutex_factory) {
+  m_mgr = mgr;
+  m_dict_id = dict_id;
+
+  m_cmp.create_from(cmp);
+  m_reference_count = 1;
+  m_userdata = nullptr;
+
+  XCALLOC(m_rangetree);
+  m_rangetree->create(&m_cmp);
+
+  m_sto_txnid = TXNID_NONE;
+  m_sto_buffer.create();
+  m_sto_score = STO_SCORE_THRESHOLD;
+  m_sto_end_early_count = 0;
+  m_sto_end_early_time = 0;
+
+  m_escalation_barrier = [](const DBT *, const DBT *, void *) -> bool {
+    return false;
+  };
+
+  m_lock_request_info.init(mutex_factory);
+}
+
+void locktree::set_escalation_barrier_func(
+    lt_escalation_barrier_check_func func, void *extra) {
+  m_escalation_barrier = func;
+  m_escalation_barrier_arg = extra;
+}
+
+void lt_lock_request_info::init(toku_external_mutex_factory_t mutex_factory) {
+  pending_lock_requests.create();
+  pending_is_empty = true;
+  toku_external_mutex_init(mutex_factory, &mutex);
+  retry_want = retry_done = 0;
+  ZERO_STRUCT(counters);
+  ZERO_STRUCT(retry_mutex);
+  toku_mutex_init(locktree_request_info_retry_mutex_key, &retry_mutex, nullptr);
+  toku_cond_init(locktree_request_info_retry_cv_key, &retry_cv, nullptr);
+  running_retry = false;
+
+  TOKU_VALGRIND_HG_DISABLE_CHECKING(&pending_is_empty,
+                                    sizeof(pending_is_empty));
+  TOKU_DRD_IGNORE_VAR(pending_is_empty);
+}
+
+void locktree::destroy(void) {
+  invariant(m_reference_count == 0);
+  invariant(m_lock_request_info.pending_lock_requests.size() == 0);
+  m_cmp.destroy();
+  m_rangetree->destroy();
+  toku_free(m_rangetree);
+  m_sto_buffer.destroy();
+  m_lock_request_info.destroy();
+}
+
+void lt_lock_request_info::destroy(void) {
+  pending_lock_requests.destroy();
+  toku_external_mutex_destroy(&mutex);
+  toku_mutex_destroy(&retry_mutex);
+  toku_cond_destroy(&retry_cv);
+}
+
+void locktree::add_reference(void) {
+  (void)toku_sync_add_and_fetch(&m_reference_count, 1);
+}
+
+uint32_t locktree::release_reference(void) {
+  return toku_sync_sub_and_fetch(&m_reference_count, 1);
+}
+
+uint32_t locktree::get_reference_count(void) { return m_reference_count; }
+
+// a container for a range/txnid pair
+struct row_lock {
+  keyrange range;
+  TXNID txnid;
+  bool is_shared;
+  TxnidVector *owners;
+};
+
+// iterate over a locked keyrange and copy out all of the data,
+// storing each row lock into the given growable array. the
+// caller does not own the range inside the returned row locks,
+// so remove from the tree with care using them as keys.
+static void iterate_and_get_overlapping_row_locks(
+    const concurrent_tree::locked_keyrange *lkr,
+    GrowableArray<row_lock> *row_locks) {
+  struct copy_fn_obj {
+    GrowableArray<row_lock> *row_locks;
+    bool fn(const keyrange &range, TXNID txnid, bool is_shared,
+            TxnidVector *owners) {
+      row_lock lock = {.range = range,
+                       .txnid = txnid,
+                       .is_shared = is_shared,
+                       .owners = owners};
+      row_locks->push(lock);
+      return true;
+    }
+  } copy_fn;
+  copy_fn.row_locks = row_locks;
+  lkr->iterate(&copy_fn);
+}
+
+// given a txnid and a set of overlapping row locks, determine
+// which txnids are conflicting, and store them in the conflicts
+// set, if given.
+static bool determine_conflicting_txnids(
+    const GrowableArray<row_lock> &row_locks, const TXNID &txnid,
+    txnid_set *conflicts) {
+  bool conflicts_exist = false;
+  const size_t num_overlaps = row_locks.get_size();
+  for (size_t i = 0; i < num_overlaps; i++) {
+    const row_lock lock = row_locks.fetch_unchecked(i);
+    const TXNID other_txnid = lock.txnid;
+    if (other_txnid != txnid) {
+      if (conflicts) {
+        if (other_txnid == TXNID_SHARED) {
+          // Add all shared lock owners, except this transaction.
+          for (TXNID shared_id : *lock.owners) {
+            if (shared_id != txnid) conflicts->add(shared_id);
+          }
+        } else {
+          conflicts->add(other_txnid);
+        }
+      }
+      conflicts_exist = true;
+    }
+  }
+  return conflicts_exist;
+}
+
+// how much memory does a row lock take up in a concurrent tree?
+static uint64_t row_lock_size_in_tree(const row_lock &lock) {
+  const uint64_t overhead = concurrent_tree::get_insertion_memory_overhead();
+  return lock.range.get_memory_size() + overhead;
+}
+
+// remove and destroy the given row lock from the locked keyrange,
+// then notify the memory tracker of the newly freed lock.
+static void remove_row_lock_from_tree(concurrent_tree::locked_keyrange *lkr,
+                                      const row_lock &lock, TXNID txnid,
+                                      locktree_manager *mgr) {
+  const uint64_t mem_released = row_lock_size_in_tree(lock);
+  lkr->remove(lock.range, txnid);
+  if (mgr != nullptr) {
+    mgr->note_mem_released(mem_released);
+  }
+}
+
+// insert a row lock into the locked keyrange, then notify
+// the memory tracker of this newly acquired lock.
+static void insert_row_lock_into_tree(concurrent_tree::locked_keyrange *lkr,
+                                      const row_lock &lock,
+                                      locktree_manager *mgr) {
+  uint64_t mem_used = row_lock_size_in_tree(lock);
+  lkr->insert(lock.range, lock.txnid, lock.is_shared);
+  if (mgr != nullptr) {
+    mgr->note_mem_used(mem_used);
+  }
+}
+
+void locktree::sto_begin(TXNID txnid) {
+  invariant(m_sto_txnid == TXNID_NONE);
+  invariant(m_sto_buffer.is_empty());
+  m_sto_txnid = txnid;
+}
+
+void locktree::sto_append(const DBT *left_key, const DBT *right_key,
+                          bool is_write_request) {
+  uint64_t buffer_mem, delta;
+
+  // psergey: the below two lines do not make any sense
+  // (and it's the same in upstream TokuDB)
+  keyrange range;
+  range.create(left_key, right_key);
+
+  buffer_mem = m_sto_buffer.total_memory_size();
+  m_sto_buffer.append(left_key, right_key, is_write_request);
+  delta = m_sto_buffer.total_memory_size() - buffer_mem;
+  if (m_mgr != nullptr) {
+    m_mgr->note_mem_used(delta);
+  }
+}
+
+void locktree::sto_end(void) {
+  uint64_t mem_size = m_sto_buffer.total_memory_size();
+  if (m_mgr != nullptr) {
+    m_mgr->note_mem_released(mem_size);
+  }
+  m_sto_buffer.destroy();
+  m_sto_buffer.create();
+  m_sto_txnid = TXNID_NONE;
+}
+
+void locktree::sto_end_early_no_accounting(void *prepared_lkr) {
+  sto_migrate_buffer_ranges_to_tree(prepared_lkr);
+  sto_end();
+  toku_unsafe_set(m_sto_score, 0);
+}
+
+void locktree::sto_end_early(void *prepared_lkr) {
+  m_sto_end_early_count++;
+
+  tokutime_t t0 = toku_time_now();
+  sto_end_early_no_accounting(prepared_lkr);
+  tokutime_t t1 = toku_time_now();
+
+  m_sto_end_early_time += (t1 - t0);
+}
+
+void locktree::sto_migrate_buffer_ranges_to_tree(void *prepared_lkr) {
+  // There should be something to migrate, and nothing in the rangetree.
+  invariant(!m_sto_buffer.is_empty());
+  invariant(m_rangetree->is_empty());
+
+  concurrent_tree sto_rangetree;
+  concurrent_tree::locked_keyrange sto_lkr;
+  sto_rangetree.create(&m_cmp);
+
+  // insert all of the ranges from the single txnid buffer into a new rangtree
+  range_buffer::iterator iter(&m_sto_buffer);
+  range_buffer::iterator::record rec;
+  while (iter.current(&rec)) {
+    sto_lkr.prepare(&sto_rangetree);
+    int r = acquire_lock_consolidated(&sto_lkr, m_sto_txnid, rec.get_left_key(),
+                                      rec.get_right_key(),
+                                      rec.get_exclusive_flag(), nullptr);
+    invariant_zero(r);
+    sto_lkr.release();
+    iter.next();
+  }
+
+  // Iterate the newly created rangetree and insert each range into the
+  // locktree's rangetree, on behalf of the old single txnid.
+  struct migrate_fn_obj {
+    concurrent_tree::locked_keyrange *dst_lkr;
+    bool fn(const keyrange &range, TXNID txnid, bool is_shared,
+            TxnidVector *owners) {
+      // There can't be multiple owners in STO mode
+      invariant_zero(owners);
+      dst_lkr->insert(range, txnid, is_shared);
+      return true;
+    }
+  } migrate_fn;
+  migrate_fn.dst_lkr =
+      static_cast<concurrent_tree::locked_keyrange *>(prepared_lkr);
+  sto_lkr.prepare(&sto_rangetree);
+  sto_lkr.iterate(&migrate_fn);
+  sto_lkr.remove_all();
+  sto_lkr.release();
+  sto_rangetree.destroy();
+  invariant(!m_rangetree->is_empty());
+}
+
+bool locktree::sto_try_acquire(void *prepared_lkr, TXNID txnid,
+                               const DBT *left_key, const DBT *right_key,
+                               bool is_write_request) {
+  if (m_rangetree->is_empty() && m_sto_buffer.is_empty() &&
+      toku_unsafe_fetch(m_sto_score) >= STO_SCORE_THRESHOLD) {
+    // We can do the optimization because the rangetree is empty, and
+    // we know its worth trying because the sto score is big enough.
+    sto_begin(txnid);
+  } else if (m_sto_txnid != TXNID_NONE) {
+    // We are currently doing the optimization. Check if we need to cancel
+    // it because a new txnid appeared, or if the current single txnid has
+    // taken too many locks already.
+    if (m_sto_txnid != txnid ||
+        m_sto_buffer.get_num_ranges() > STO_BUFFER_MAX_SIZE) {
+      sto_end_early(prepared_lkr);
+    }
+  }
+
+  // At this point the sto txnid is properly set. If it is valid, then
+  // this txnid can append its lock to the sto buffer successfully.
+  if (m_sto_txnid != TXNID_NONE) {
+    invariant(m_sto_txnid == txnid);
+    sto_append(left_key, right_key, is_write_request);
+    return true;
+  } else {
+    invariant(m_sto_buffer.is_empty());
+    return false;
+  }
+}
+
+/*
+  Do the same as iterate_and_get_overlapping_row_locks does, but also check for
+  this:
+    The set of overlapping rows locks consists of just one read-only shared
+    lock with the same endpoints as specified (in that case, we can just add
+    ourselves into that list)
+
+  @return true - One compatible shared lock
+         false - Otherwise
+*/
+static bool iterate_and_get_overlapping_row_locks2(
+    const concurrent_tree::locked_keyrange *lkr, const DBT *left_key,
+    const DBT *right_key, comparator *cmp, TXNID,
+    GrowableArray<row_lock> *row_locks) {
+  struct copy_fn_obj {
+    GrowableArray<row_lock> *row_locks;
+    bool first_call = true;
+    bool matching_lock_found = false;
+    const DBT *left_key, *right_key;
+    comparator *cmp;
+
+    bool fn(const keyrange &range, TXNID txnid, bool is_shared,
+            TxnidVector *owners) {
+      if (first_call) {
+        first_call = false;
+        if (is_shared && !(*cmp)(left_key, range.get_left_key()) &&
+            !(*cmp)(right_key, range.get_right_key())) {
+          matching_lock_found = true;
+        }
+      } else {
+        // if we see multiple matching locks, it doesn't matter whether
+        // the first one was matching.
+        matching_lock_found = false;
+      }
+      row_lock lock = {.range = range,
+                       .txnid = txnid,
+                       .is_shared = is_shared,
+                       .owners = owners};
+      row_locks->push(lock);
+      return true;
+    }
+  } copy_fn;
+  copy_fn.row_locks = row_locks;
+  copy_fn.left_key = left_key;
+  copy_fn.right_key = right_key;
+  copy_fn.cmp = cmp;
+  lkr->iterate(&copy_fn);
+  return copy_fn.matching_lock_found;
+}
+
+// try to acquire a lock and consolidate it with existing locks if possible
+// param: lkr, a prepared locked keyrange
+// return: 0 on success, DB_LOCK_NOTGRANTED if conflicting locks exist.
+int locktree::acquire_lock_consolidated(void *prepared_lkr, TXNID txnid,
+                                        const DBT *left_key,
+                                        const DBT *right_key,
+                                        bool is_write_request,
+                                        txnid_set *conflicts) {
+  int r = 0;
+  concurrent_tree::locked_keyrange *lkr;
+
+  keyrange requested_range;
+  requested_range.create(left_key, right_key);
+  lkr = static_cast<concurrent_tree::locked_keyrange *>(prepared_lkr);
+  lkr->acquire(requested_range);
+
+  // copy out the set of overlapping row locks.
+  GrowableArray<row_lock> overlapping_row_locks;
+  overlapping_row_locks.init();
+  bool matching_shared_lock_found = false;
+
+  if (is_write_request)
+    iterate_and_get_overlapping_row_locks(lkr, &overlapping_row_locks);
+  else {
+    matching_shared_lock_found = iterate_and_get_overlapping_row_locks2(
+        lkr, left_key, right_key, &m_cmp, txnid, &overlapping_row_locks);
+    // psergey-todo: what to do now? So, we have figured we have just one
+    // shareable lock. Need to add us into it as an owner but the lock
+    // pointer cannot be kept?
+    // A: use find_node_with_overlapping_child(key_range, nullptr);
+    //  then, add ourselves to the owner list.
+    // Dont' foreget to release the subtree after that.
+  }
+
+  if (matching_shared_lock_found) {
+    // there is just one non-confliting matching shared lock.
+    //  we are hilding a lock on it (see acquire() call above).
+    //  we need to modify it to indicate there is another locker...
+    if (lkr->add_shared_owner(requested_range, txnid)) {
+      // Pretend shared lock uses as much memory.
+      row_lock new_lock = {.range = requested_range,
+                           .txnid = txnid,
+                           .is_shared = false,
+                           .owners = nullptr};
+      uint64_t mem_used = row_lock_size_in_tree(new_lock);
+      if (m_mgr) {
+        m_mgr->note_mem_used(mem_used);
+      }
+    }
+    requested_range.destroy();
+    overlapping_row_locks.deinit();
+    return 0;
+  }
+
+  size_t num_overlapping_row_locks = overlapping_row_locks.get_size();
+
+  // if any overlapping row locks conflict with this request, bail out.
+
+  bool conflicts_exist =
+      determine_conflicting_txnids(overlapping_row_locks, txnid, conflicts);
+  if (!conflicts_exist) {
+    // there are no conflicts, so all of the overlaps are for the requesting
+    // txnid. so, we must consolidate all existing overlapping ranges and the
+    // requested range into one dominating range. then we insert the dominating
+    // range.
+    bool all_shared = !is_write_request;
+    for (size_t i = 0; i < num_overlapping_row_locks; i++) {
+      row_lock overlapping_lock = overlapping_row_locks.fetch_unchecked(i);
+      invariant(overlapping_lock.txnid == txnid);
+      requested_range.extend(m_cmp, overlapping_lock.range);
+      remove_row_lock_from_tree(lkr, overlapping_lock, TXNID_ANY, m_mgr);
+      all_shared = all_shared && overlapping_lock.is_shared;
+    }
+
+    row_lock new_lock = {.range = requested_range,
+                         .txnid = txnid,
+                         .is_shared = all_shared,
+                         .owners = nullptr};
+    insert_row_lock_into_tree(lkr, new_lock, m_mgr);
+  } else {
+    r = DB_LOCK_NOTGRANTED;
+  }
+
+  requested_range.destroy();
+  overlapping_row_locks.deinit();
+  return r;
+}
+
+// acquire a lock in the given key range, inclusive. if successful,
+// return 0. otherwise, populate the conflicts txnid_set with the set of
+// transactions that conflict with this request.
+int locktree::acquire_lock(bool is_write_request, TXNID txnid,
+                           const DBT *left_key, const DBT *right_key,
+                           txnid_set *conflicts) {
+  int r = 0;
+
+  // we are only supporting write locks for simplicity
+  // invariant(is_write_request);
+
+  // acquire and prepare a locked keyrange over the requested range.
+  // prepare is a serialzation point, so we take the opportunity to
+  // try the single txnid optimization first.
+  concurrent_tree::locked_keyrange lkr;
+  lkr.prepare(m_rangetree);
+
+  bool acquired =
+      sto_try_acquire(&lkr, txnid, left_key, right_key, is_write_request);
+  if (!acquired) {
+    r = acquire_lock_consolidated(&lkr, txnid, left_key, right_key,
+                                  is_write_request, conflicts);
+  }
+
+  lkr.release();
+  return r;
+}
+
+int locktree::try_acquire_lock(bool is_write_request, TXNID txnid,
+                               const DBT *left_key, const DBT *right_key,
+                               txnid_set *conflicts, bool big_txn) {
+  // All ranges in the locktree must have left endpoints <= right endpoints.
+  // Range comparisons rely on this fact, so we make a paranoid invariant here.
+  paranoid_invariant(m_cmp(left_key, right_key) <= 0);
+  int r = m_mgr == nullptr ? 0 : m_mgr->check_current_lock_constraints(big_txn);
+  if (r == 0) {
+    r = acquire_lock(is_write_request, txnid, left_key, right_key, conflicts);
+  }
+  return r;
+}
+
+// the locktree silently upgrades read locks to write locks for simplicity
+int locktree::acquire_read_lock(TXNID txnid, const DBT *left_key,
+                                const DBT *right_key, txnid_set *conflicts,
+                                bool big_txn) {
+  return try_acquire_lock(false, txnid, left_key, right_key, conflicts,
+                          big_txn);
+}
+
+int locktree::acquire_write_lock(TXNID txnid, const DBT *left_key,
+                                 const DBT *right_key, txnid_set *conflicts,
+                                 bool big_txn) {
+  return try_acquire_lock(true, txnid, left_key, right_key, conflicts, big_txn);
+}
+
+// typedef void (*dump_callback)(void *cdata, const DBT *left, const DBT *right,
+// TXNID txnid);
+void locktree::dump_locks(void *cdata, dump_callback cb) {
+  concurrent_tree::locked_keyrange lkr;
+  keyrange range;
+  range.create(toku_dbt_negative_infinity(), toku_dbt_positive_infinity());
+
+  lkr.prepare(m_rangetree);
+  lkr.acquire(range);
+
+  TXNID sto_txn;
+  if ((sto_txn = toku_unsafe_fetch(m_sto_txnid)) != TXNID_NONE) {
+    // insert all of the ranges from the single txnid buffer into a new rangtree
+    range_buffer::iterator iter(&m_sto_buffer);
+    range_buffer::iterator::record rec;
+    while (iter.current(&rec)) {
+      (*cb)(cdata, rec.get_left_key(), rec.get_right_key(), sto_txn,
+            !rec.get_exclusive_flag(), nullptr);
+      iter.next();
+    }
+  } else {
+    GrowableArray<row_lock> all_locks;
+    all_locks.init();
+    iterate_and_get_overlapping_row_locks(&lkr, &all_locks);
+
+    const size_t n_locks = all_locks.get_size();
+    for (size_t i = 0; i < n_locks; i++) {
+      const row_lock lock = all_locks.fetch_unchecked(i);
+      (*cb)(cdata, lock.range.get_left_key(), lock.range.get_right_key(),
+            lock.txnid, lock.is_shared, lock.owners);
+    }
+    all_locks.deinit();
+  }
+  lkr.release();
+  range.destroy();
+}
+
+void locktree::get_conflicts(bool is_write_request, TXNID txnid,
+                             const DBT *left_key, const DBT *right_key,
+                             txnid_set *conflicts) {
+  // because we only support write locks, ignore this bit for now.
+  (void)is_write_request;
+
+  // preparing and acquire a locked keyrange over the range
+  keyrange range;
+  range.create(left_key, right_key);
+  concurrent_tree::locked_keyrange lkr;
+  lkr.prepare(m_rangetree);
+  lkr.acquire(range);
+
+  // copy out the set of overlapping row locks and determine the conflicts
+  GrowableArray<row_lock> overlapping_row_locks;
+  overlapping_row_locks.init();
+  iterate_and_get_overlapping_row_locks(&lkr, &overlapping_row_locks);
+
+  // we don't care if conflicts exist. we just want the conflicts set populated.
+  (void)determine_conflicting_txnids(overlapping_row_locks, txnid, conflicts);
+
+  lkr.release();
+  overlapping_row_locks.deinit();
+  range.destroy();
+}
+
+// Effect:
+//  For each range in the lock tree that overlaps the given range and has
+//  the given txnid, remove it.
+// Rationale:
+//  In the common case, there is only the range [left_key, right_key] and
+//  it is associated with txnid, so this is a single tree delete.
+//
+//  However, consolidation and escalation change the objects in the tree
+//  without telling the txn anything.  In this case, the txn may own a
+//  large range lock that represents its ownership of many smaller range
+//  locks.  For example, the txn may think it owns point locks on keys 1,
+//  2, and 3, but due to escalation, only the object [1,3] exists in the
+//  tree.
+//
+//  The first call for a small lock will remove the large range lock, and
+//  the rest of the calls should do nothing.  After the first release,
+//  another thread can acquire one of the locks that the txn thinks it
+//  still owns.  That's ok, because the txn doesn't want it anymore (it
+//  unlocks everything at once), but it may find a lock that it does not
+//  own.
+//
+//  In our example, the txn unlocks key 1, which actually removes the
+//  whole lock [1,3].  Now, someone else can lock 2 before our txn gets
+//  around to unlocking 2, so we should not remove that lock.
+void locktree::remove_overlapping_locks_for_txnid(TXNID txnid,
+                                                  const DBT *left_key,
+                                                  const DBT *right_key) {
+  keyrange release_range;
+  release_range.create(left_key, right_key);
+
+  // acquire and prepare a locked keyrange over the release range
+  concurrent_tree::locked_keyrange lkr;
+  lkr.prepare(m_rangetree);
+  lkr.acquire(release_range);
+
+  // copy out the set of overlapping row locks.
+  GrowableArray<row_lock> overlapping_row_locks;
+  overlapping_row_locks.init();
+  iterate_and_get_overlapping_row_locks(&lkr, &overlapping_row_locks);
+  size_t num_overlapping_row_locks = overlapping_row_locks.get_size();
+
+  for (size_t i = 0; i < num_overlapping_row_locks; i++) {
+    row_lock lock = overlapping_row_locks.fetch_unchecked(i);
+    // If this isn't our lock, that's ok, just don't remove it.
+    // See rationale above.
+    // psergey-todo: for shared locks, just remove ourselves from the
+    //               owners.
+    if (lock.txnid == txnid || (lock.owners && lock.owners->contains(txnid))) {
+      remove_row_lock_from_tree(&lkr, lock, txnid, m_mgr);
+    }
+  }
+
+  lkr.release();
+  overlapping_row_locks.deinit();
+  release_range.destroy();
+}
+
+bool locktree::sto_txnid_is_valid_unsafe(void) const {
+  return toku_unsafe_fetch(m_sto_txnid) != TXNID_NONE;
+}
+
+int locktree::sto_get_score_unsafe(void) const {
+  return toku_unsafe_fetch(m_sto_score);
+}
+
+bool locktree::sto_try_release(TXNID txnid) {
+  bool released = false;
+  if (toku_unsafe_fetch(m_sto_txnid) != TXNID_NONE) {
+    // check the bit again with a prepared locked keyrange,
+    // which protects the optimization bits and rangetree data
+    concurrent_tree::locked_keyrange lkr;
+    lkr.prepare(m_rangetree);
+    if (m_sto_txnid != TXNID_NONE) {
+      // this txnid better be the single txnid on this locktree,
+      // or else we are in big trouble (meaning the logic is broken)
+      invariant(m_sto_txnid == txnid);
+      invariant(m_rangetree->is_empty());
+      sto_end();
+      released = true;
+    }
+    lkr.release();
+  }
+  return released;
+}
+
+// release all of the locks for a txnid whose endpoints are pairs
+// in the given range buffer.
+void locktree::release_locks(TXNID txnid, const range_buffer *ranges,
+                             bool all_trx_locks_hint) {
+  // try the single txn optimization. if it worked, then all of the
+  // locks are already released, otherwise we need to do it here.
+  bool released;
+  if (all_trx_locks_hint) {
+    // This will release all of the locks the transaction is holding
+    released = sto_try_release(txnid);
+  } else {
+    /*
+      psergey: we are asked to release *Some* of the locks the transaction
+      is holding.
+      We could try doing that without leaving the STO mode, but right now,
+      the easiest way is to exit the STO mode and let the non-STO code path
+      handle it.
+    */
+    if (toku_unsafe_fetch(m_sto_txnid) != TXNID_NONE) {
+      // check the bit again with a prepared locked keyrange,
+      // which protects the optimization bits and rangetree data
+      concurrent_tree::locked_keyrange lkr;
+      lkr.prepare(m_rangetree);
+      if (m_sto_txnid != TXNID_NONE) {
+        sto_end_early(&lkr);
+      }
+      lkr.release();
+    }
+    released = false;
+  }
+  if (!released) {
+    range_buffer::iterator iter(ranges);
+    range_buffer::iterator::record rec;
+    while (iter.current(&rec)) {
+      const DBT *left_key = rec.get_left_key();
+      const DBT *right_key = rec.get_right_key();
+      // All ranges in the locktree must have left endpoints <= right endpoints.
+      // Range comparisons rely on this fact, so we make a paranoid invariant
+      // here.
+      paranoid_invariant(m_cmp(left_key, right_key) <= 0);
+      remove_overlapping_locks_for_txnid(txnid, left_key, right_key);
+      iter.next();
+    }
+    // Increase the sto score slightly. Eventually it will hit
+    // the threshold and we'll try the optimization again. This
+    // is how a previously multithreaded system transitions into
+    // a single threaded system that benefits from the optimization.
+    if (toku_unsafe_fetch(m_sto_score) < STO_SCORE_THRESHOLD) {
+      toku_sync_fetch_and_add(&m_sto_score, 1);
+    }
+  }
+}
+
+// iterate over a locked keyrange and extract copies of the first N
+// row locks, storing each one into the given array of size N,
+// then removing each extracted lock from the locked keyrange.
+static int extract_first_n_row_locks(concurrent_tree::locked_keyrange *lkr,
+                                     locktree_manager *mgr, row_lock *row_locks,
+                                     int num_to_extract) {
+  struct extract_fn_obj {
+    int num_extracted;
+    int num_to_extract;
+    row_lock *row_locks;
+    bool fn(const keyrange &range, TXNID txnid, bool is_shared,
+            TxnidVector *owners) {
+      if (num_extracted < num_to_extract) {
+        row_lock lock;
+        lock.range.create_copy(range);
+        lock.txnid = txnid;
+        lock.is_shared = is_shared;
+        // deep-copy the set of owners:
+        if (owners)
+          lock.owners = new TxnidVector(*owners);
+        else
+          lock.owners = nullptr;
+        row_locks[num_extracted++] = lock;
+        return true;
+      } else {
+        return false;
+      }
+    }
+  } extract_fn;
+
+  extract_fn.row_locks = row_locks;
+  extract_fn.num_to_extract = num_to_extract;
+  extract_fn.num_extracted = 0;
+  lkr->iterate(&extract_fn);
+
+  // now that the ranges have been copied out, complete
+  // the extraction by removing the ranges from the tree.
+  // use remove_row_lock_from_tree() so we properly track the
+  // amount of memory and number of locks freed.
+  int num_extracted = extract_fn.num_extracted;
+  invariant(num_extracted <= num_to_extract);
+  for (int i = 0; i < num_extracted; i++) {
+    remove_row_lock_from_tree(lkr, row_locks[i], TXNID_ANY, mgr);
+  }
+
+  return num_extracted;
+}
+
+// Store each newly escalated lock in a range buffer for appropriate txnid.
+// We'll rebuild the locktree by iterating over these ranges, and then we
+// can pass back each txnid/buffer pair individually through a callback
+// to notify higher layers that locks have changed.
+struct txnid_range_buffer {
+  TXNID txnid;
+  range_buffer buffer;
+
+  static int find_by_txnid(struct txnid_range_buffer *const &other_buffer,
+                           const TXNID &txnid) {
+    if (txnid < other_buffer->txnid) {
+      return -1;
+    } else if (other_buffer->txnid == txnid) {
+      return 0;
+    } else {
+      return 1;
+    }
+  }
+};
+
+// escalate the locks in the locktree by merging adjacent
+// locks that have the same txnid into one larger lock.
+//
+// if there's only one txnid in the locktree then this
+// approach works well. if there are many txnids and each
+// has locks in a random/alternating order, then this does
+// not work so well.
+void locktree::escalate(lt_escalate_cb after_escalate_callback,
+                        void *after_escalate_callback_extra) {
+  omt<struct txnid_range_buffer *, struct txnid_range_buffer *> range_buffers;
+  range_buffers.create();
+
+  // prepare and acquire a locked keyrange on the entire locktree
+  concurrent_tree::locked_keyrange lkr;
+  keyrange infinite_range = keyrange::get_infinite_range();
+  lkr.prepare(m_rangetree);
+  lkr.acquire(infinite_range);
+
+  // if we're in the single txnid optimization, simply call it off.
+  // if you have to run escalation, you probably don't care about
+  // the optimization anyway, and this makes things easier.
+  if (m_sto_txnid != TXNID_NONE) {
+    // We are already accounting for this escalation time and
+    // count, so don't do it for sto_end_early too.
+    sto_end_early_no_accounting(&lkr);
+  }
+
+  // extract and remove batches of row locks from the locktree
+  int num_extracted;
+  const int num_row_locks_per_batch = 128;
+  row_lock *XCALLOC_N(num_row_locks_per_batch, extracted_buf);
+
+  // we always remove the "first" n because we are removing n
+  // each time we do an extraction. so this loops until its empty.
+  while ((num_extracted = extract_first_n_row_locks(
+              &lkr, m_mgr, extracted_buf, num_row_locks_per_batch)) > 0) {
+    int current_index = 0;
+    while (current_index < num_extracted) {
+      // every batch of extracted locks is in range-sorted order. search
+      // through them and merge adjacent locks with the same txnid into
+      // one dominating lock and save it to a set of escalated locks.
+      //
+      // first, find the index of the next row lock that
+      //  - belongs to a different txnid, or
+      //  - belongs to several txnids, or
+      //  - is a shared lock (we could potentially merge those but
+      //    currently we don't), or
+      //  - is across a lock escalation barrier.
+      int next_txnid_index = current_index + 1;
+
+      while (next_txnid_index < num_extracted &&
+             (extracted_buf[current_index].txnid ==
+              extracted_buf[next_txnid_index].txnid) &&
+             !extracted_buf[next_txnid_index].is_shared &&
+             !extracted_buf[next_txnid_index].owners &&
+             !m_escalation_barrier(
+                 extracted_buf[current_index].range.get_right_key(),
+                 extracted_buf[next_txnid_index].range.get_left_key(),
+                 m_escalation_barrier_arg)) {
+        next_txnid_index++;
+      }
+
+      // Create an escalated range for the current txnid that dominates
+      // each range between the current indext and the next txnid's index.
+      // const TXNID current_txnid = extracted_buf[current_index].txnid;
+      const DBT *escalated_left_key =
+          extracted_buf[current_index].range.get_left_key();
+      const DBT *escalated_right_key =
+          extracted_buf[next_txnid_index - 1].range.get_right_key();
+
+      // Try to find a range buffer for the current txnid. Create one if it
+      // doesn't exist. Then, append the new escalated range to the buffer. (If
+      // a lock is shared by multiple txnids, append it each of txnid's lists)
+      TxnidVector *owners_ptr;
+      TxnidVector singleton_owner;
+      if (extracted_buf[current_index].owners)
+        owners_ptr = extracted_buf[current_index].owners;
+      else {
+        singleton_owner.insert(extracted_buf[current_index].txnid);
+        owners_ptr = &singleton_owner;
+      }
+
+      for (auto cur_txnid : *owners_ptr) {
+        uint32_t idx;
+        struct txnid_range_buffer *existing_range_buffer;
+        int r =
+            range_buffers.find_zero<TXNID, txnid_range_buffer::find_by_txnid>(
+                cur_txnid, &existing_range_buffer, &idx);
+        if (r == DB_NOTFOUND) {
+          struct txnid_range_buffer *XMALLOC(new_range_buffer);
+          new_range_buffer->txnid = cur_txnid;
+          new_range_buffer->buffer.create();
+          new_range_buffer->buffer.append(
+              escalated_left_key, escalated_right_key,
+              !extracted_buf[current_index].is_shared);
+          range_buffers.insert_at(new_range_buffer, idx);
+        } else {
+          invariant_zero(r);
+          invariant(existing_range_buffer->txnid == cur_txnid);
+          existing_range_buffer->buffer.append(
+              escalated_left_key, escalated_right_key,
+              !extracted_buf[current_index].is_shared);
+        }
+      }
+
+      current_index = next_txnid_index;
+    }
+
+    // destroy the ranges copied during the extraction
+    for (int i = 0; i < num_extracted; i++) {
+      delete extracted_buf[i].owners;
+      extracted_buf[i].range.destroy();
+    }
+  }
+  toku_free(extracted_buf);
+
+  // Rebuild the locktree from each range in each range buffer,
+  // then notify higher layers that the txnid's locks have changed.
+  //
+  // (shared locks: if a lock was initially shared between transactions TRX1,
+  //  TRX2, etc, we will now try to acquire it acting on behalf on TRX1, on
+  //  TRX2, etc.  This will succeed and an identical shared lock will be
+  //  constructed)
+
+  invariant(m_rangetree->is_empty());
+  const uint32_t num_range_buffers = range_buffers.size();
+  for (uint32_t i = 0; i < num_range_buffers; i++) {
+    struct txnid_range_buffer *current_range_buffer;
+    int r = range_buffers.fetch(i, &current_range_buffer);
+    invariant_zero(r);
+    if (r == EINVAL)  // Shouldn't happen, avoid compiler warning
+      continue;
+
+    const TXNID current_txnid = current_range_buffer->txnid;
+    range_buffer::iterator iter(&current_range_buffer->buffer);
+    range_buffer::iterator::record rec;
+    while (iter.current(&rec)) {
+      keyrange range;
+      range.create(rec.get_left_key(), rec.get_right_key());
+      row_lock lock = {.range = range,
+                       .txnid = current_txnid,
+                       .is_shared = !rec.get_exclusive_flag(),
+                       .owners = nullptr};
+      insert_row_lock_into_tree(&lkr, lock, m_mgr);
+      iter.next();
+    }
+
+    // Notify higher layers that locks have changed for the current txnid
+    if (after_escalate_callback) {
+      after_escalate_callback(current_txnid, this, current_range_buffer->buffer,
+                              after_escalate_callback_extra);
+    }
+    current_range_buffer->buffer.destroy();
+  }
+
+  while (range_buffers.size() > 0) {
+    struct txnid_range_buffer *buffer;
+    int r = range_buffers.fetch(0, &buffer);
+    invariant_zero(r);
+    r = range_buffers.delete_at(0);
+    invariant_zero(r);
+    toku_free(buffer);
+  }
+  range_buffers.destroy();
+
+  lkr.release();
+}
+
+void *locktree::get_userdata(void) const { return m_userdata; }
+
+void locktree::set_userdata(void *userdata) { m_userdata = userdata; }
+
+struct lt_lock_request_info *locktree::get_lock_request_info(void) {
+  return &m_lock_request_info;
+}
+
+void locktree::set_comparator(const comparator &cmp) { m_cmp.inherit(cmp); }
+
+locktree_manager *locktree::get_manager(void) const { return m_mgr; }
+
+int locktree::compare(const locktree *lt) const {
+  if (m_dict_id.dictid < lt->m_dict_id.dictid) {
+    return -1;
+  } else if (m_dict_id.dictid == lt->m_dict_id.dictid) {
+    return 0;
+  } else {
+    return 1;
+  }
+}
+
+DICTIONARY_ID locktree::get_dict_id() const { return m_dict_id; }
+
+} /* namespace toku */
+#endif  // OS_WIN
+#endif  // ROCKSDB_LITE
diff --git a/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/locktree/locktree.h b/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/locktree/locktree.h
new file mode 100644
index 000000000..f0f4b042d
--- /dev/null
+++ b/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/locktree/locktree.h
@@ -0,0 +1,580 @@
+/* -*- mode: C++; c-basic-offset: 4; indent-tabs-mode: nil -*- */
+// vim: ft=cpp:expandtab:ts=8:sw=4:softtabstop=4:
+#ident "$Id$"
+/*======
+This file is part of PerconaFT.
+
+
+Copyright (c) 2006, 2015, Percona and/or its affiliates. All rights reserved.
+
+    PerconaFT is free software: you can redistribute it and/or modify
+    it under the terms of the GNU General Public License, version 2,
+    as published by the Free Software Foundation.
+
+    PerconaFT is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License
+    along with PerconaFT.  If not, see <http://www.gnu.org/licenses/>.
+
+----------------------------------------
+
+    PerconaFT is free software: you can redistribute it and/or modify
+    it under the terms of the GNU Affero General Public License, version 3,
+    as published by the Free Software Foundation.
+
+    PerconaFT is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU Affero General Public License for more details.
+
+    You should have received a copy of the GNU Affero General Public License
+    along with PerconaFT.  If not, see <http://www.gnu.org/licenses/>.
+
+----------------------------------------
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.
+======= */
+
+#ident \
+    "Copyright (c) 2006, 2015, Percona and/or its affiliates. All rights reserved."
+
+#pragma once
+
+#include <atomic>
+
+#include "../db.h"
+#include "../ft/comparator.h"
+#include "../portability/toku_external_pthread.h"
+#include "../portability/toku_pthread.h"
+#include "../portability/toku_time.h"
+// PORT #include <ft/ft-ops.h>  // just for DICTIONARY_ID..
+// PORT: ft-status for LTM_STATUS:
+#include "../ft/ft-status.h"
+
+struct DICTIONARY_ID {
+  uint64_t dictid;
+};
+
+#include "../util/omt.h"
+#include "range_buffer.h"
+#include "txnid_set.h"
+#include "wfg.h"
+
+namespace toku {
+
+class locktree;
+class locktree_manager;
+class lock_request;
+class concurrent_tree;
+
+typedef int (*lt_create_cb)(locktree *lt, void *extra);
+typedef void (*lt_destroy_cb)(locktree *lt);
+typedef void (*lt_escalate_cb)(TXNID txnid, const locktree *lt,
+                               const range_buffer &buffer, void *extra);
+
+typedef bool (*lt_escalation_barrier_check_func)(const DBT *a, const DBT *b,
+                                                 void *extra);
+
+struct lt_counters {
+  uint64_t wait_count, wait_time;
+  uint64_t long_wait_count, long_wait_time;
+  uint64_t timeout_count;
+
+  void add(const lt_counters &rhs) {
+    wait_count += rhs.wait_count;
+    wait_time += rhs.wait_time;
+    long_wait_count += rhs.long_wait_count;
+    long_wait_time += rhs.long_wait_time;
+    timeout_count += rhs.timeout_count;
+  }
+};
+
+// Lock request state for some locktree
+struct lt_lock_request_info {
+  omt<lock_request *> pending_lock_requests;
+  std::atomic_bool pending_is_empty;
+  toku_external_mutex_t mutex;
+  bool should_retry_lock_requests;
+  lt_counters counters;
+  std::atomic_ullong retry_want;
+  unsigned long long retry_done;
+  toku_mutex_t retry_mutex;
+  toku_cond_t retry_cv;
+  bool running_retry;
+
+  void init(toku_external_mutex_factory_t mutex_factory);
+  void destroy(void);
+};
+
+// The locktree manager manages a set of locktrees, one for each open
+// dictionary. Locktrees are retrieved from the manager. When they are no
+// longer needed, they are be released by the user.
+class locktree_manager {
+ public:
+  // param: create_cb, called just after a locktree is first created.
+  //        destroy_cb, called just before a locktree is destroyed.
+  //        escalate_cb, called after a locktree is escalated (with extra
+  //        param)
+  void create(lt_create_cb create_cb, lt_destroy_cb destroy_cb,
+              lt_escalate_cb escalate_cb, void *extra,
+              toku_external_mutex_factory_t mutex_factory_arg);
+
+  void destroy(void);
+
+  size_t get_max_lock_memory(void);
+
+  int set_max_lock_memory(size_t max_lock_memory);
+
+  // effect: Get a locktree from the manager. If a locktree exists with the
+  // given
+  //         dict_id, it is referenced and then returned. If one did not exist,
+  //         it is created. It will use the comparator for comparing keys. The
+  //         on_create callback (passed to locktree_manager::create()) will be
+  //         called with the given extra parameter.
+  locktree *get_lt(DICTIONARY_ID dict_id, const comparator &cmp,
+                   void *on_create_extra);
+
+  void reference_lt(locktree *lt);
+
+  // effect: Releases one reference on a locktree. If the reference count
+  // transitions
+  //         to zero, the on_destroy callback is called before it gets
+  //         destroyed.
+  void release_lt(locktree *lt);
+
+  void get_status(LTM_STATUS status);
+
+  // effect: calls the iterate function on each pending lock request
+  // note: holds the manager's mutex
+  typedef int (*lock_request_iterate_callback)(DICTIONARY_ID dict_id,
+                                               TXNID txnid, const DBT *left_key,
+                                               const DBT *right_key,
+                                               TXNID blocking_txnid,
+                                               uint64_t start_time,
+                                               void *extra);
+  int iterate_pending_lock_requests(lock_request_iterate_callback cb,
+                                    void *extra);
+
+  // effect: Determines if too many locks or too much memory is being used,
+  //         Runs escalation on the manager if so.
+  // param: big_txn, if the current transaction is 'big' (has spilled rollback
+  // logs) returns: 0 if there enough resources to create a new lock, or
+  // TOKUDB_OUT_OF_LOCKS
+  //          if there are not enough resources and lock escalation failed to
+  //          free up enough resources for a new lock.
+  int check_current_lock_constraints(bool big_txn);
+
+  bool over_big_threshold(void);
+
+  void note_mem_used(uint64_t mem_used);
+
+  void note_mem_released(uint64_t mem_freed);
+
+  bool out_of_locks(void) const;
+
+  // Escalate all locktrees
+  void escalate_all_locktrees(void);
+
+  // Escalate a set of locktrees
+  void escalate_locktrees(locktree **locktrees, int num_locktrees);
+
+  // effect: calls the private function run_escalation(), only ok to
+  //         do for tests.
+  // rationale: to get better stress test coverage, we want a way to
+  //            deterministicly trigger lock escalation.
+  void run_escalation_for_test(void);
+  void run_escalation(void);
+
+  // Add time t to the escalator's wait time statistics
+  void add_escalator_wait_time(uint64_t t);
+
+  void kill_waiter(void *extra);
+
+ private:
+  static const uint64_t DEFAULT_MAX_LOCK_MEMORY = 64L * 1024 * 1024;
+
+  // tracks the current number of locks and lock memory
+  uint64_t m_max_lock_memory;
+  uint64_t m_current_lock_memory;
+
+  struct lt_counters m_lt_counters;
+
+  // the create and destroy callbacks for the locktrees
+  lt_create_cb m_lt_create_callback;
+  lt_destroy_cb m_lt_destroy_callback;
+  lt_escalate_cb m_lt_escalate_callback;
+  void *m_lt_escalate_callback_extra;
+
+  omt<locktree *> m_locktree_map;
+
+  toku_external_mutex_factory_t mutex_factory;
+
+  // the manager's mutex protects the locktree map
+  toku_mutex_t m_mutex;
+
+  void mutex_lock(void);
+
+  void mutex_unlock(void);
+
+  // Manage the set of open locktrees
+  locktree *locktree_map_find(const DICTIONARY_ID &dict_id);
+  void locktree_map_put(locktree *lt);
+  void locktree_map_remove(locktree *lt);
+
+  static int find_by_dict_id(locktree *const &lt, const DICTIONARY_ID &dict_id);
+
+  void escalator_init(void);
+  void escalator_destroy(void);
+
+  // statistics about lock escalation.
+  toku_mutex_t m_escalation_mutex;
+  uint64_t m_escalation_count;
+  tokutime_t m_escalation_time;
+  uint64_t m_escalation_latest_result;
+  uint64_t m_wait_escalation_count;
+  uint64_t m_wait_escalation_time;
+  uint64_t m_long_wait_escalation_count;
+  uint64_t m_long_wait_escalation_time;
+
+  // the escalator coordinates escalation on a set of locktrees for a bunch of
+  // threads
+  class locktree_escalator {
+   public:
+    void create(void);
+    void destroy(void);
+    void run(locktree_manager *mgr, void (*escalate_locktrees_fun)(void *extra),
+             void *extra);
+
+   private:
+    toku_mutex_t m_escalator_mutex;
+    toku_cond_t m_escalator_done;
+    bool m_escalator_running;
+  };
+
+  locktree_escalator m_escalator;
+
+  friend class manager_unit_test;
+};
+
+// A locktree represents the set of row locks owned by all transactions
+// over an open dictionary. Read and write ranges are represented as
+// a left and right key which are compared with the given comparator
+//
+// Locktrees are not created and destroyed by the user. Instead, they are
+// referenced and released using the locktree manager.
+//
+// A sample workflow looks like this:
+// - Create a manager.
+// - Get a locktree by dictionaroy id from the manager.
+// - Perform read/write lock acquision on the locktree, add references to
+//   the locktree using the manager, release locks, release references, etc.
+// - ...
+// - Release the final reference to the locktree. It will be destroyed.
+// - Destroy the manager.
+class locktree {
+ public:
+  // effect: Creates a locktree
+  void create(locktree_manager *mgr, DICTIONARY_ID dict_id,
+              const comparator &cmp,
+              toku_external_mutex_factory_t mutex_factory);
+
+  void destroy(void);
+
+  // For thread-safe, external reference counting
+  void add_reference(void);
+
+  // requires: the reference count is > 0
+  // returns: the reference count, after decrementing it by one
+  uint32_t release_reference(void);
+
+  // returns: the current reference count
+  uint32_t get_reference_count(void);
+
+  // effect: Attempts to grant a read lock for the range of keys between
+  // [left_key, right_key]. returns: If the lock cannot be granted, return
+  // DB_LOCK_NOTGRANTED, and populate the
+  //          given conflicts set with the txnids that hold conflicting locks in
+  //          the range. If the locktree cannot create more locks, return
+  //          TOKUDB_OUT_OF_LOCKS.
+  // note: Read locks cannot be shared between txnids, as one would expect.
+  //       This is for simplicity since read locks are rare in MySQL.
+  int acquire_read_lock(TXNID txnid, const DBT *left_key, const DBT *right_key,
+                        txnid_set *conflicts, bool big_txn);
+
+  // effect: Attempts to grant a write lock for the range of keys between
+  // [left_key, right_key]. returns: If the lock cannot be granted, return
+  // DB_LOCK_NOTGRANTED, and populate the
+  //          given conflicts set with the txnids that hold conflicting locks in
+  //          the range. If the locktree cannot create more locks, return
+  //          TOKUDB_OUT_OF_LOCKS.
+  int acquire_write_lock(TXNID txnid, const DBT *left_key, const DBT *right_key,
+                         txnid_set *conflicts, bool big_txn);
+
+  // effect: populate the conflicts set with the txnids that would preventing
+  //         the given txnid from getting a lock on [left_key, right_key]
+  void get_conflicts(bool is_write_request, TXNID txnid, const DBT *left_key,
+                     const DBT *right_key, txnid_set *conflicts);
+
+  // effect: Release all of the lock ranges represented by the range buffer for
+  // a txnid.
+  void release_locks(TXNID txnid, const range_buffer *ranges,
+                     bool all_trx_locks_hint = false);
+
+  // effect: Runs escalation on this locktree
+  void escalate(lt_escalate_cb after_escalate_callback, void *extra);
+
+  // returns: The userdata associated with this locktree, or null if it has not
+  // been set.
+  void *get_userdata(void) const;
+
+  void set_userdata(void *userdata);
+
+  locktree_manager *get_manager(void) const;
+
+  void set_comparator(const comparator &cmp);
+
+  // Set the user-provided Lock Escalation Barrier check function and its
+  // argument
+  //
+  // Lock Escalation Barrier limits the scope of Lock Escalation.
+  // For two keys A and B (such that A < B),
+  // escalation_barrier_check_func(A, B)==true means that there's a lock
+  // escalation barrier between A and B, and lock escalation is not allowed to
+  // bridge the gap between A and B.
+  //
+  // This method sets the user-provided barrier check function and its
+  // parameter.
+  void set_escalation_barrier_func(lt_escalation_barrier_check_func func,
+                                   void *extra);
+
+  int compare(const locktree *lt) const;
+
+  DICTIONARY_ID get_dict_id() const;
+
+  // Private info struct for storing pending lock request state.
+  // Only to be used by lock requests. We store it here as
+  // something less opaque than usual to strike a tradeoff between
+  // abstraction and code complexity. It is still fairly abstract
+  // since the lock_request object is opaque
+  struct lt_lock_request_info *get_lock_request_info(void);
+
+  typedef void (*dump_callback)(void *cdata, const DBT *left, const DBT *right,
+                                TXNID txnid, bool is_shared,
+                                TxnidVector *owners);
+  void dump_locks(void *cdata, dump_callback cb);
+
+ private:
+  locktree_manager *m_mgr;
+  DICTIONARY_ID m_dict_id;
+  uint32_t m_reference_count;
+
+  // Since the memory referenced by this comparator is not owned by the
+  // locktree, the user must guarantee it will outlive the locktree.
+  //
+  // The ydb API accomplishes this by opening an ft_handle in the on_create
+  // callback, which will keep the underlying FT (and its descriptor) in memory
+  // for as long as the handle is open. The ft_handle is stored opaquely in the
+  // userdata pointer below. see locktree_manager::get_lt w/ on_create_extra
+  comparator m_cmp;
+
+  lt_escalation_barrier_check_func m_escalation_barrier;
+  void *m_escalation_barrier_arg;
+
+  concurrent_tree *m_rangetree;
+
+  void *m_userdata;
+  struct lt_lock_request_info m_lock_request_info;
+
+  // psergey-todo:
+  //  Each transaction also keeps a list of ranges it has locked.
+  //  So, when a transaction is running in STO mode, two identical
+  //  lists are kept: the STO lock list and transaction's owned locks
+  //  list. Why can't we do with just one list?
+
+  // The following fields and members prefixed with "sto_" are for
+  // the single txnid optimization, intended to speed up the case
+  // when only one transaction is using the locktree. If we know
+  // the locktree has only one transaction, then acquiring locks
+  // takes O(1) work and releasing all locks takes O(1) work.
+  //
+  // How do we know that the locktree only has a single txnid?
+  // What do we do if it does?
+  //
+  // When a txn with txnid T requests a lock:
+  // - If the tree is empty, the optimization is possible. Set the single
+  // txnid to T, and insert the lock range into the buffer.
+  // - If the tree is not empty, check if the single txnid is T. If so,
+  // append the lock range to the buffer. Otherwise, migrate all of
+  // the locks in the buffer into the rangetree on behalf of txnid T,
+  // and invalid the single txnid.
+  //
+  // When a txn with txnid T releases its locks:
+  // - If the single txnid is valid, it must be for T. Destroy the buffer.
+  // - If it's not valid, release locks the normal way in the rangetree.
+  //
+  // To carry out the optimization we need to record a single txnid
+  // and a range buffer for each locktree, each protected by the root
+  // lock of the locktree's rangetree. The root lock for a rangetree
+  // is grabbed by preparing a locked keyrange on the rangetree.
+  TXNID m_sto_txnid;
+  range_buffer m_sto_buffer;
+
+  // The single txnid optimization speeds up the case when only one
+  // transaction is using the locktree. But it has the potential to
+  // hurt the case when more than one txnid exists.
+  //
+  // There are two things we need to do to make the optimization only
+  // optimize the case we care about, and not hurt the general case.
+  //
+  // Bound the worst-case latency for lock migration when the
+  // optimization stops working:
+  // - Idea: Stop the optimization and migrate immediate if we notice
+  // the single txnid has takes many locks in the range buffer.
+  // - Implementation: Enforce a max size on the single txnid range buffer.
+  // - Analysis: Choosing the perfect max value, M, is difficult to do
+  // without some feedback from the field. Intuition tells us that M should
+  // not be so small that the optimization is worthless, and it should not
+  // be so big that it's unreasonable to have to wait behind a thread doing
+  // the work of converting M buffer locks into rangetree locks.
+  //
+  // Prevent concurrent-transaction workloads from trying the optimization
+  // in vain:
+  // - Idea: Don't even bother trying the optimization if we think the
+  // system is in a concurrent-transaction state.
+  // - Implementation: Do something even simpler than detecting whether the
+  // system is in a concurent-transaction state. Just keep a "score" value
+  // and some threshold. If at any time the locktree is eligible for the
+  // optimization, only do it if the score is at this threshold. When you
+  // actually do the optimization but someone has to migrate locks in the buffer
+  // (expensive), then reset the score back to zero. Each time a txn
+  // releases locks, the score is incremented by 1.
+  // - Analysis: If you let the threshold be "C", then at most 1 / C txns will
+  // do the optimization in a concurrent-transaction system. Similarly, it
+  // takes at most C txns to start using the single txnid optimzation, which
+  // is good when the system transitions from multithreaded to single threaded.
+  //
+  // STO_BUFFER_MAX_SIZE:
+  //
+  // We choose the max value to be 1 million since most transactions are smaller
+  // than 1 million and we can create a rangetree of 1 million elements in
+  // less than a second. So we can be pretty confident that this threshold
+  // enables the optimization almost always, and prevents super pathological
+  // latency issues for the first lock taken by a second thread.
+  //
+  // STO_SCORE_THRESHOLD:
+  //
+  // A simple first guess at a good value for the score threshold is 100.
+  // By our analysis, we'd end up doing the optimization in vain for
+  // around 1% of all transactions, which seems reasonable. Further,
+  // if the system goes single threaded, it ought to be pretty quick
+  // for 100 transactions to go by, so we won't have to wait long before
+  // we start doing the single txind optimzation again.
+  static const int STO_BUFFER_MAX_SIZE = 50 * 1024;
+  static const int STO_SCORE_THRESHOLD = 100;
+  int m_sto_score;
+
+  // statistics about time spent ending the STO early
+  uint64_t m_sto_end_early_count;
+  tokutime_t m_sto_end_early_time;
+
+  // effect: begins the single txnid optimizaiton, setting m_sto_txnid
+  //         to the given txnid.
+  // requires: m_sto_txnid is invalid
+  void sto_begin(TXNID txnid);
+
+  // effect: append a range to the sto buffer
+  // requires: m_sto_txnid is valid
+  void sto_append(const DBT *left_key, const DBT *right_key,
+                  bool is_write_request);
+
+  // effect: ends the single txnid optimization, releaseing any memory
+  //         stored in the sto buffer, notifying the tracker, and
+  //         invalidating m_sto_txnid.
+  // requires: m_sto_txnid is valid
+  void sto_end(void);
+
+  // params: prepared_lkr is a void * to a prepared locked keyrange. see below.
+  // effect: ends the single txnid optimization early, migrating buffer locks
+  //         into the rangetree, calling sto_end(), and then setting the
+  //         sto_score back to zero.
+  // requires: m_sto_txnid is valid
+  void sto_end_early(void *prepared_lkr);
+  void sto_end_early_no_accounting(void *prepared_lkr);
+
+  // params: prepared_lkr is a void * to a prepared locked keyrange. we can't
+  // use
+  //         the real type because the compiler won't allow us to forward
+  //         declare concurrent_tree::locked_keyrange without including
+  //         concurrent_tree.h, which we cannot do here because it is a template
+  //         implementation.
+  // requires: the prepared locked keyrange is for the locktree's rangetree
+  // requires: m_sto_txnid is valid
+  // effect: migrates each lock in the single txnid buffer into the locktree's
+  //         rangetree, notifying the memory tracker as necessary.
+  void sto_migrate_buffer_ranges_to_tree(void *prepared_lkr);
+
+  // effect: If m_sto_txnid is valid, then release the txnid's locks
+  //         by ending the optimization.
+  // requires: If m_sto_txnid is valid, it is equal to the given txnid
+  // returns: True if locks were released for this txnid
+  bool sto_try_release(TXNID txnid);
+
+  // params: prepared_lkr is a void * to a prepared locked keyrange. see above.
+  // requires: the prepared locked keyrange is for the locktree's rangetree
+  // effect: If m_sto_txnid is valid and equal to the given txnid, then
+  // append a range onto the buffer. Otherwise, if m_sto_txnid is valid
+  //        but not equal to this txnid, then migrate the buffer's locks
+  //        into the rangetree and end the optimization, setting the score
+  //        back to zero.
+  // returns: true if the lock was acquired for this txnid
+  bool sto_try_acquire(void *prepared_lkr, TXNID txnid, const DBT *left_key,
+                       const DBT *right_key, bool is_write_request);
+
+  // Effect:
+  //  Provides a hook for a helgrind suppression.
+  // Returns:
+  //  true if m_sto_txnid is not TXNID_NONE
+  bool sto_txnid_is_valid_unsafe(void) const;
+
+  // Effect:
+  //  Provides a hook for a helgrind suppression.
+  // Returns:
+  //  m_sto_score
+  int sto_get_score_unsafe(void) const;
+
+  void remove_overlapping_locks_for_txnid(TXNID txnid, const DBT *left_key,
+                                          const DBT *right_key);
+
+  int acquire_lock_consolidated(void *prepared_lkr, TXNID txnid,
+                                const DBT *left_key, const DBT *right_key,
+                                bool is_write_request, txnid_set *conflicts);
+
+  int acquire_lock(bool is_write_request, TXNID txnid, const DBT *left_key,
+                   const DBT *right_key, txnid_set *conflicts);
+
+  int try_acquire_lock(bool is_write_request, TXNID txnid, const DBT *left_key,
+                       const DBT *right_key, txnid_set *conflicts,
+                       bool big_txn);
+
+  friend class locktree_unit_test;
+  friend class manager_unit_test;
+  friend class lock_request_unit_test;
+
+  // engine status reaches into the locktree to read some stats
+  friend void locktree_manager::get_status(LTM_STATUS status);
+};
+
+} /* namespace toku */
diff --git a/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/locktree/manager.cc b/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/locktree/manager.cc
new file mode 100644
index 000000000..4186182be
--- /dev/null
+++ b/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/locktree/manager.cc
@@ -0,0 +1,527 @@
+/* -*- mode: C++; c-basic-offset: 4; indent-tabs-mode: nil -*- */
+// vim: ft=cpp:expandtab:ts=8:sw=4:softtabstop=4:
+#ifndef ROCKSDB_LITE
+#ifndef OS_WIN
+#ident "$Id$"
+/*======
+This file is part of PerconaFT.
+
+
+Copyright (c) 2006, 2015, Percona and/or its affiliates. All rights reserved.
+
+    PerconaFT is free software: you can redistribute it and/or modify
+    it under the terms of the GNU General Public License, version 2,
+    as published by the Free Software Foundation.
+
+    PerconaFT is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License
+    along with PerconaFT.  If not, see <http://www.gnu.org/licenses/>.
+
+----------------------------------------
+
+    PerconaFT is free software: you can redistribute it and/or modify
+    it under the terms of the GNU Affero General Public License, version 3,
+    as published by the Free Software Foundation.
+
+    PerconaFT is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU Affero General Public License for more details.
+
+    You should have received a copy of the GNU Affero General Public License
+    along with PerconaFT.  If not, see <http://www.gnu.org/licenses/>.
+
+----------------------------------------
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.
+======= */
+
+#ident \
+    "Copyright (c) 2006, 2015, Percona and/or its affiliates. All rights reserved."
+
+#include <stdlib.h>
+#include <string.h>
+
+#include "../portability/toku_pthread.h"
+#include "../util/status.h"
+#include "lock_request.h"
+#include "locktree.h"
+
+namespace toku {
+
+void locktree_manager::create(lt_create_cb create_cb, lt_destroy_cb destroy_cb,
+                              lt_escalate_cb escalate_cb, void *escalate_extra,
+                              toku_external_mutex_factory_t mutex_factory_arg) {
+  mutex_factory = mutex_factory_arg;
+  m_max_lock_memory = DEFAULT_MAX_LOCK_MEMORY;
+  m_current_lock_memory = 0;
+
+  m_locktree_map.create();
+  m_lt_create_callback = create_cb;
+  m_lt_destroy_callback = destroy_cb;
+  m_lt_escalate_callback = escalate_cb;
+  m_lt_escalate_callback_extra = escalate_extra;
+  ZERO_STRUCT(m_mutex);
+  toku_mutex_init(manager_mutex_key, &m_mutex, nullptr);
+
+  ZERO_STRUCT(m_lt_counters);
+
+  escalator_init();
+}
+
+void locktree_manager::destroy(void) {
+  escalator_destroy();
+  invariant(m_current_lock_memory == 0);
+  invariant(m_locktree_map.size() == 0);
+  m_locktree_map.destroy();
+  toku_mutex_destroy(&m_mutex);
+}
+
+void locktree_manager::mutex_lock(void) { toku_mutex_lock(&m_mutex); }
+
+void locktree_manager::mutex_unlock(void) { toku_mutex_unlock(&m_mutex); }
+
+size_t locktree_manager::get_max_lock_memory(void) { return m_max_lock_memory; }
+
+int locktree_manager::set_max_lock_memory(size_t max_lock_memory) {
+  int r = 0;
+  mutex_lock();
+  if (max_lock_memory < m_current_lock_memory) {
+    r = EDOM;
+  } else {
+    m_max_lock_memory = max_lock_memory;
+  }
+  mutex_unlock();
+  return r;
+}
+
+int locktree_manager::find_by_dict_id(locktree *const &lt,
+                                      const DICTIONARY_ID &dict_id) {
+  if (lt->get_dict_id().dictid < dict_id.dictid) {
+    return -1;
+  } else if (lt->get_dict_id().dictid == dict_id.dictid) {
+    return 0;
+  } else {
+    return 1;
+  }
+}
+
+locktree *locktree_manager::locktree_map_find(const DICTIONARY_ID &dict_id) {
+  locktree *lt;
+  int r = m_locktree_map.find_zero<DICTIONARY_ID, find_by_dict_id>(dict_id, &lt,
+                                                                   nullptr);
+  return r == 0 ? lt : nullptr;
+}
+
+void locktree_manager::locktree_map_put(locktree *lt) {
+  int r = m_locktree_map.insert<DICTIONARY_ID, find_by_dict_id>(
+      lt, lt->get_dict_id(), nullptr);
+  invariant_zero(r);
+}
+
+void locktree_manager::locktree_map_remove(locktree *lt) {
+  uint32_t idx;
+  locktree *found_lt;
+  int r = m_locktree_map.find_zero<DICTIONARY_ID, find_by_dict_id>(
+      lt->get_dict_id(), &found_lt, &idx);
+  invariant_zero(r);
+  invariant(found_lt == lt);
+  r = m_locktree_map.delete_at(idx);
+  invariant_zero(r);
+}
+
+locktree *locktree_manager::get_lt(DICTIONARY_ID dict_id, const comparator &cmp,
+                                   void *on_create_extra) {
+  // hold the mutex around searching and maybe
+  // inserting into the locktree map
+  mutex_lock();
+
+  locktree *lt = locktree_map_find(dict_id);
+  if (lt == nullptr) {
+    XCALLOC(lt);
+    lt->create(this, dict_id, cmp, mutex_factory);
+
+    // new locktree created - call the on_create callback
+    // and put it in the locktree map
+    if (m_lt_create_callback) {
+      int r = m_lt_create_callback(lt, on_create_extra);
+      if (r != 0) {
+        lt->release_reference();
+        lt->destroy();
+        toku_free(lt);
+        lt = nullptr;
+      }
+    }
+    if (lt) {
+      locktree_map_put(lt);
+    }
+  } else {
+    reference_lt(lt);
+  }
+
+  mutex_unlock();
+
+  return lt;
+}
+
+void locktree_manager::reference_lt(locktree *lt) {
+  // increment using a sync fetch and add.
+  // the caller guarantees that the lt won't be
+  // destroyed while we increment the count here.
+  //
+  // the caller can do this by already having an lt
+  // reference or by holding the manager mutex.
+  //
+  // if the manager's mutex is held, it is ok for the
+  // reference count to transition from 0 to 1 (no race),
+  // since we're serialized with other opens and closes.
+  lt->add_reference();
+}
+
+void locktree_manager::release_lt(locktree *lt) {
+  bool do_destroy = false;
+  DICTIONARY_ID dict_id = lt->get_dict_id();
+
+  // Release a reference on the locktree. If the count transitions to zero,
+  // then we *may* need to do the cleanup.
+  //
+  // Grab the manager's mutex and look for a locktree with this locktree's
+  // dictionary id. Since dictionary id's never get reused, any locktree
+  // found must be the one we just released a reference on.
+  //
+  // At least two things could have happened since we got the mutex:
+  // - Another thread gets a locktree with the same dict_id, increments
+  // the reference count. In this case, we shouldn't destroy it.
+  // - Another thread gets a locktree with the same dict_id and then
+  // releases it quickly, transitioning the reference count from zero to
+  // one and back to zero. In this case, only one of us should destroy it.
+  // It doesn't matter which. We originally missed this case, see #5776.
+  //
+  // After 5776, the high level rule for release is described below.
+  //
+  // If a thread releases a locktree and notices the reference count transition
+  // to zero, then that thread must immediately:
+  // - assume the locktree object is invalid
+  // - grab the manager's mutex
+  // - search the locktree map for a locktree with the same dict_id and remove
+  // it, if it exists. the destroy may be deferred.
+  // - release the manager's mutex
+  //
+  // This way, if many threads transition the same locktree's reference count
+  // from 1 to zero and wait behind the manager's mutex, only one of them will
+  // do the actual destroy and the others will happily do nothing.
+  uint32_t refs = lt->release_reference();
+  if (refs == 0) {
+    mutex_lock();
+    // lt may not have already been destroyed, so look it up.
+    locktree *find_lt = locktree_map_find(dict_id);
+    if (find_lt != nullptr) {
+      // A locktree is still in the map with that dict_id, so it must be
+      // equal to lt. This is true because dictionary ids are never reused.
+      // If the reference count is zero, it's our responsibility to remove
+      // it and do the destroy. Otherwise, someone still wants it.
+      // If the locktree is still valid then check if it should be deleted.
+      if (find_lt == lt) {
+        if (lt->get_reference_count() == 0) {
+          locktree_map_remove(lt);
+          do_destroy = true;
+        }
+        m_lt_counters.add(lt->get_lock_request_info()->counters);
+      }
+    }
+    mutex_unlock();
+  }
+
+  // if necessary, do the destroy without holding the mutex
+  if (do_destroy) {
+    if (m_lt_destroy_callback) {
+      m_lt_destroy_callback(lt);
+    }
+    lt->destroy();
+    toku_free(lt);
+  }
+}
+
+void locktree_manager::run_escalation(void) {
+  struct escalation_fn {
+    static void run(void *extra) {
+      locktree_manager *mgr = (locktree_manager *)extra;
+      mgr->escalate_all_locktrees();
+    };
+  };
+  m_escalator.run(this, escalation_fn::run, this);
+}
+
+// test-only version of lock escalation
+void locktree_manager::run_escalation_for_test(void) { run_escalation(); }
+
+void locktree_manager::escalate_all_locktrees(void) {
+  uint64_t t0 = toku_current_time_microsec();
+
+  // get all locktrees
+  mutex_lock();
+  int num_locktrees = m_locktree_map.size();
+  locktree **locktrees = new locktree *[num_locktrees];
+  for (int i = 0; i < num_locktrees; i++) {
+    int r = m_locktree_map.fetch(i, &locktrees[i]);
+    invariant_zero(r);
+    reference_lt(locktrees[i]);
+  }
+  mutex_unlock();
+
+  // escalate them
+  escalate_locktrees(locktrees, num_locktrees);
+
+  delete[] locktrees;
+
+  uint64_t t1 = toku_current_time_microsec();
+  add_escalator_wait_time(t1 - t0);
+}
+
+void locktree_manager::note_mem_used(uint64_t mem_used) {
+  (void)toku_sync_fetch_and_add(&m_current_lock_memory, mem_used);
+}
+
+void locktree_manager::note_mem_released(uint64_t mem_released) {
+  uint64_t old_mem_used =
+      toku_sync_fetch_and_sub(&m_current_lock_memory, mem_released);
+  invariant(old_mem_used >= mem_released);
+}
+
+bool locktree_manager::out_of_locks(void) const {
+  return m_current_lock_memory >= m_max_lock_memory;
+}
+
+bool locktree_manager::over_big_threshold(void) {
+  return m_current_lock_memory >= m_max_lock_memory / 2;
+}
+
+int locktree_manager::iterate_pending_lock_requests(
+    lock_request_iterate_callback callback, void *extra) {
+  mutex_lock();
+  int r = 0;
+  uint32_t num_locktrees = m_locktree_map.size();
+  for (uint32_t i = 0; i < num_locktrees && r == 0; i++) {
+    locktree *lt;
+    r = m_locktree_map.fetch(i, &lt);
+    invariant_zero(r);
+    if (r == EINVAL)  // Shouldn't happen, avoid compiler warning
+      continue;
+
+    struct lt_lock_request_info *info = lt->get_lock_request_info();
+    toku_external_mutex_lock(&info->mutex);
+
+    uint32_t num_requests = info->pending_lock_requests.size();
+    for (uint32_t k = 0; k < num_requests && r == 0; k++) {
+      lock_request *req;
+      r = info->pending_lock_requests.fetch(k, &req);
+      invariant_zero(r);
+      if (r == EINVAL) /* Shouldn't happen, avoid compiler warning */
+        continue;
+      r = callback(lt->get_dict_id(), req->get_txnid(), req->get_left_key(),
+                   req->get_right_key(), req->get_conflicting_txnid(),
+                   req->get_start_time(), extra);
+    }
+
+    toku_external_mutex_unlock(&info->mutex);
+  }
+  mutex_unlock();
+  return r;
+}
+
+int locktree_manager::check_current_lock_constraints(bool big_txn) {
+  int r = 0;
+  if (big_txn && over_big_threshold()) {
+    run_escalation();
+    if (over_big_threshold()) {
+      r = TOKUDB_OUT_OF_LOCKS;
+    }
+  }
+  if (r == 0 && out_of_locks()) {
+    run_escalation();
+    if (out_of_locks()) {
+      // return an error if we're still out of locks after escalation.
+      r = TOKUDB_OUT_OF_LOCKS;
+    }
+  }
+  return r;
+}
+
+void locktree_manager::escalator_init(void) {
+  ZERO_STRUCT(m_escalation_mutex);
+  toku_mutex_init(manager_escalation_mutex_key, &m_escalation_mutex, nullptr);
+  m_escalation_count = 0;
+  m_escalation_time = 0;
+  m_wait_escalation_count = 0;
+  m_wait_escalation_time = 0;
+  m_long_wait_escalation_count = 0;
+  m_long_wait_escalation_time = 0;
+  m_escalation_latest_result = 0;
+  m_escalator.create();
+}
+
+void locktree_manager::escalator_destroy(void) {
+  m_escalator.destroy();
+  toku_mutex_destroy(&m_escalation_mutex);
+}
+
+void locktree_manager::add_escalator_wait_time(uint64_t t) {
+  toku_mutex_lock(&m_escalation_mutex);
+  m_wait_escalation_count += 1;
+  m_wait_escalation_time += t;
+  if (t >= 1000000) {
+    m_long_wait_escalation_count += 1;
+    m_long_wait_escalation_time += t;
+  }
+  toku_mutex_unlock(&m_escalation_mutex);
+}
+
+void locktree_manager::escalate_locktrees(locktree **locktrees,
+                                          int num_locktrees) {
+  // there are too many row locks in the system and we need to tidy up.
+  //
+  // a simple implementation of escalation does not attempt
+  // to reduce the memory foot print of each txn's range buffer.
+  // doing so would require some layering hackery (or a callback)
+  // and more complicated locking. for now, just escalate each
+  // locktree individually, in-place.
+  tokutime_t t0 = toku_time_now();
+  for (int i = 0; i < num_locktrees; i++) {
+    locktrees[i]->escalate(m_lt_escalate_callback,
+                           m_lt_escalate_callback_extra);
+    release_lt(locktrees[i]);
+  }
+  tokutime_t t1 = toku_time_now();
+
+  toku_mutex_lock(&m_escalation_mutex);
+  m_escalation_count++;
+  m_escalation_time += (t1 - t0);
+  m_escalation_latest_result = m_current_lock_memory;
+  toku_mutex_unlock(&m_escalation_mutex);
+}
+
+struct escalate_args {
+  locktree_manager *mgr;
+  locktree **locktrees;
+  int num_locktrees;
+};
+
+void locktree_manager::locktree_escalator::create(void) {
+  ZERO_STRUCT(m_escalator_mutex);
+  toku_mutex_init(manager_escalator_mutex_key, &m_escalator_mutex, nullptr);
+  toku_cond_init(manager_m_escalator_done_key, &m_escalator_done, nullptr);
+  m_escalator_running = false;
+}
+
+void locktree_manager::locktree_escalator::destroy(void) {
+  toku_cond_destroy(&m_escalator_done);
+  toku_mutex_destroy(&m_escalator_mutex);
+}
+
+void locktree_manager::locktree_escalator::run(
+    locktree_manager *mgr, void (*escalate_locktrees_fun)(void *extra),
+    void *extra) {
+  uint64_t t0 = toku_current_time_microsec();
+  toku_mutex_lock(&m_escalator_mutex);
+  if (!m_escalator_running) {
+    // run escalation on this thread
+    m_escalator_running = true;
+    toku_mutex_unlock(&m_escalator_mutex);
+    escalate_locktrees_fun(extra);
+    toku_mutex_lock(&m_escalator_mutex);
+    m_escalator_running = false;
+    toku_cond_broadcast(&m_escalator_done);
+  } else {
+    toku_cond_wait(&m_escalator_done, &m_escalator_mutex);
+  }
+  toku_mutex_unlock(&m_escalator_mutex);
+  uint64_t t1 = toku_current_time_microsec();
+  mgr->add_escalator_wait_time(t1 - t0);
+}
+
+void locktree_manager::get_status(LTM_STATUS statp) {
+  ltm_status.init();
+  LTM_STATUS_VAL(LTM_SIZE_CURRENT) = m_current_lock_memory;
+  LTM_STATUS_VAL(LTM_SIZE_LIMIT) = m_max_lock_memory;
+  LTM_STATUS_VAL(LTM_ESCALATION_COUNT) = m_escalation_count;
+  LTM_STATUS_VAL(LTM_ESCALATION_TIME) = m_escalation_time;
+  LTM_STATUS_VAL(LTM_ESCALATION_LATEST_RESULT) = m_escalation_latest_result;
+  LTM_STATUS_VAL(LTM_WAIT_ESCALATION_COUNT) = m_wait_escalation_count;
+  LTM_STATUS_VAL(LTM_WAIT_ESCALATION_TIME) = m_wait_escalation_time;
+  LTM_STATUS_VAL(LTM_LONG_WAIT_ESCALATION_COUNT) = m_long_wait_escalation_count;
+  LTM_STATUS_VAL(LTM_LONG_WAIT_ESCALATION_TIME) = m_long_wait_escalation_time;
+
+  uint64_t lock_requests_pending = 0;
+  uint64_t sto_num_eligible = 0;
+  uint64_t sto_end_early_count = 0;
+  tokutime_t sto_end_early_time = 0;
+  uint32_t num_locktrees = 0;
+  struct lt_counters lt_counters;
+  ZERO_STRUCT(lt_counters);  // PORT: instead of ={}.
+
+  if (toku_mutex_trylock(&m_mutex) == 0) {
+    lt_counters = m_lt_counters;
+    num_locktrees = m_locktree_map.size();
+    for (uint32_t i = 0; i < num_locktrees; i++) {
+      locktree *lt;
+      int r = m_locktree_map.fetch(i, &lt);
+      invariant_zero(r);
+      if (r == EINVAL)  // Shouldn't happen, avoid compiler warning
+        continue;
+      if (toku_external_mutex_trylock(&lt->m_lock_request_info.mutex) == 0) {
+        lock_requests_pending +=
+            lt->m_lock_request_info.pending_lock_requests.size();
+        lt_counters.add(lt->get_lock_request_info()->counters);
+        toku_external_mutex_unlock(&lt->m_lock_request_info.mutex);
+      }
+      sto_num_eligible += lt->sto_txnid_is_valid_unsafe() ? 1 : 0;
+      sto_end_early_count += lt->m_sto_end_early_count;
+      sto_end_early_time += lt->m_sto_end_early_time;
+    }
+    mutex_unlock();
+  }
+
+  LTM_STATUS_VAL(LTM_NUM_LOCKTREES) = num_locktrees;
+  LTM_STATUS_VAL(LTM_LOCK_REQUESTS_PENDING) = lock_requests_pending;
+  LTM_STATUS_VAL(LTM_STO_NUM_ELIGIBLE) = sto_num_eligible;
+  LTM_STATUS_VAL(LTM_STO_END_EARLY_COUNT) = sto_end_early_count;
+  LTM_STATUS_VAL(LTM_STO_END_EARLY_TIME) = sto_end_early_time;
+  LTM_STATUS_VAL(LTM_WAIT_COUNT) = lt_counters.wait_count;
+  LTM_STATUS_VAL(LTM_WAIT_TIME) = lt_counters.wait_time;
+  LTM_STATUS_VAL(LTM_LONG_WAIT_COUNT) = lt_counters.long_wait_count;
+  LTM_STATUS_VAL(LTM_LONG_WAIT_TIME) = lt_counters.long_wait_time;
+  LTM_STATUS_VAL(LTM_TIMEOUT_COUNT) = lt_counters.timeout_count;
+  *statp = ltm_status;
+}
+
+void locktree_manager::kill_waiter(void *extra) {
+  mutex_lock();
+  int r = 0;
+  uint32_t num_locktrees = m_locktree_map.size();
+  for (uint32_t i = 0; i < num_locktrees; i++) {
+    locktree *lt;
+    r = m_locktree_map.fetch(i, &lt);
+    invariant_zero(r);
+    if (r) continue;  // Get rid of "may be used uninitialized" warning
+    lock_request::kill_waiter(lt, extra);
+  }
+  mutex_unlock();
+}
+
+} /* namespace toku */
+#endif  // OS_WIN
+#endif  // ROCKSDB_LITE
diff --git a/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/locktree/range_buffer.cc b/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/locktree/range_buffer.cc
new file mode 100644
index 000000000..1e1d23ef8
--- /dev/null
+++ b/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/locktree/range_buffer.cc
@@ -0,0 +1,265 @@
+/* -*- mode: C++; c-basic-offset: 4; indent-tabs-mode: nil -*- */
+// vim: ft=cpp:expandtab:ts=8:sw=4:softtabstop=4:
+#ifndef ROCKSDB_LITE
+#ifndef OS_WIN
+#ident "$Id$"
+/*======
+This file is part of PerconaFT.
+
+
+Copyright (c) 2006, 2015, Percona and/or its affiliates. All rights reserved.
+
+    PerconaFT is free software: you can redistribute it and/or modify
+    it under the terms of the GNU General Public License, version 2,
+    as published by the Free Software Foundation.
+
+    PerconaFT is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License
+    along with PerconaFT.  If not, see <http://www.gnu.org/licenses/>.
+
+----------------------------------------
+
+    PerconaFT is free software: you can redistribute it and/or modify
+    it under the terms of the GNU Affero General Public License, version 3,
+    as published by the Free Software Foundation.
+
+    PerconaFT is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU Affero General Public License for more details.
+
+    You should have received a copy of the GNU Affero General Public License
+    along with PerconaFT.  If not, see <http://www.gnu.org/licenses/>.
+
+----------------------------------------
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.
+======= */
+
+#ident \
+    "Copyright (c) 2006, 2015, Percona and/or its affiliates. All rights reserved."
+
+#include "range_buffer.h"
+
+#include <string.h>
+
+#include "../portability/memory.h"
+#include "../util/dbt.h"
+
+namespace toku {
+
+bool range_buffer::record_header::left_is_infinite(void) const {
+  return left_neg_inf || left_pos_inf;
+}
+
+bool range_buffer::record_header::right_is_infinite(void) const {
+  return right_neg_inf || right_pos_inf;
+}
+
+void range_buffer::record_header::init(const DBT *left_key,
+                                       const DBT *right_key,
+                                       bool is_exclusive) {
+  is_exclusive_lock = is_exclusive;
+  left_neg_inf = left_key == toku_dbt_negative_infinity();
+  left_pos_inf = left_key == toku_dbt_positive_infinity();
+  left_key_size = toku_dbt_is_infinite(left_key) ? 0 : left_key->size;
+  if (right_key) {
+    right_neg_inf = right_key == toku_dbt_negative_infinity();
+    right_pos_inf = right_key == toku_dbt_positive_infinity();
+    right_key_size = toku_dbt_is_infinite(right_key) ? 0 : right_key->size;
+  } else {
+    right_neg_inf = left_neg_inf;
+    right_pos_inf = left_pos_inf;
+    right_key_size = 0;
+  }
+}
+
+const DBT *range_buffer::iterator::record::get_left_key(void) const {
+  if (_header.left_neg_inf) {
+    return toku_dbt_negative_infinity();
+  } else if (_header.left_pos_inf) {
+    return toku_dbt_positive_infinity();
+  } else {
+    return &_left_key;
+  }
+}
+
+const DBT *range_buffer::iterator::record::get_right_key(void) const {
+  if (_header.right_neg_inf) {
+    return toku_dbt_negative_infinity();
+  } else if (_header.right_pos_inf) {
+    return toku_dbt_positive_infinity();
+  } else {
+    return &_right_key;
+  }
+}
+
+size_t range_buffer::iterator::record::size(void) const {
+  return sizeof(record_header) + _header.left_key_size + _header.right_key_size;
+}
+
+void range_buffer::iterator::record::deserialize(const char *buf) {
+  size_t current = 0;
+
+  // deserialize the header
+  memcpy(&_header, buf, sizeof(record_header));
+  current += sizeof(record_header);
+
+  // deserialize the left key if necessary
+  if (!_header.left_is_infinite()) {
+    // point the left DBT's buffer into ours
+    toku_fill_dbt(&_left_key, buf + current, _header.left_key_size);
+    current += _header.left_key_size;
+  }
+
+  // deserialize the right key if necessary
+  if (!_header.right_is_infinite()) {
+    if (_header.right_key_size == 0) {
+      toku_copyref_dbt(&_right_key, _left_key);
+    } else {
+      toku_fill_dbt(&_right_key, buf + current, _header.right_key_size);
+    }
+  }
+}
+
+toku::range_buffer::iterator::iterator()
+    : _ma_chunk_iterator(nullptr),
+      _current_chunk_base(nullptr),
+      _current_chunk_offset(0),
+      _current_chunk_max(0),
+      _current_rec_size(0) {}
+
+toku::range_buffer::iterator::iterator(const range_buffer *buffer)
+    : _ma_chunk_iterator(&buffer->_arena),
+      _current_chunk_base(nullptr),
+      _current_chunk_offset(0),
+      _current_chunk_max(0),
+      _current_rec_size(0) {
+  reset_current_chunk();
+}
+
+void range_buffer::iterator::reset_current_chunk() {
+  _current_chunk_base = _ma_chunk_iterator.current(&_current_chunk_max);
+  _current_chunk_offset = 0;
+}
+
+bool range_buffer::iterator::current(record *rec) {
+  if (_current_chunk_offset < _current_chunk_max) {
+    const char *buf = reinterpret_cast<const char *>(_current_chunk_base);
+    rec->deserialize(buf + _current_chunk_offset);
+    _current_rec_size = rec->size();
+    return true;
+  } else {
+    return false;
+  }
+}
+
+// move the iterator to the next record in the buffer
+void range_buffer::iterator::next(void) {
+  invariant(_current_chunk_offset < _current_chunk_max);
+  invariant(_current_rec_size > 0);
+
+  // the next record is _current_rec_size bytes forward
+  _current_chunk_offset += _current_rec_size;
+  // now, we don't know how big the current is, set it to 0.
+  _current_rec_size = 0;
+
+  if (_current_chunk_offset >= _current_chunk_max) {
+    // current chunk is exhausted, try moving to the next one
+    if (_ma_chunk_iterator.more()) {
+      _ma_chunk_iterator.next();
+      reset_current_chunk();
+    }
+  }
+}
+
+void range_buffer::create(void) {
+  // allocate buffer space lazily instead of on creation. this way,
+  // no malloc/free is done if the transaction ends up taking no locks.
+  _arena.create(0);
+  _num_ranges = 0;
+}
+
+void range_buffer::append(const DBT *left_key, const DBT *right_key,
+                          bool is_write_request) {
+  // if the keys are equal, then only one copy is stored.
+  if (toku_dbt_equals(left_key, right_key)) {
+    invariant(left_key->size <= MAX_KEY_SIZE);
+    append_point(left_key, is_write_request);
+  } else {
+    invariant(left_key->size <= MAX_KEY_SIZE);
+    invariant(right_key->size <= MAX_KEY_SIZE);
+    append_range(left_key, right_key, is_write_request);
+  }
+  _num_ranges++;
+}
+
+bool range_buffer::is_empty(void) const { return total_memory_size() == 0; }
+
+uint64_t range_buffer::total_memory_size(void) const {
+  return _arena.total_size_in_use();
+}
+
+int range_buffer::get_num_ranges(void) const { return _num_ranges; }
+
+void range_buffer::destroy(void) { _arena.destroy(); }
+
+void range_buffer::append_range(const DBT *left_key, const DBT *right_key,
+                                bool is_exclusive) {
+  size_t record_length =
+      sizeof(record_header) + left_key->size + right_key->size;
+  char *buf = reinterpret_cast<char *>(_arena.malloc_from_arena(record_length));
+
+  record_header h;
+  h.init(left_key, right_key, is_exclusive);
+
+  // serialize the header
+  memcpy(buf, &h, sizeof(record_header));
+  buf += sizeof(record_header);
+
+  // serialize the left key if necessary
+  if (!h.left_is_infinite()) {
+    memcpy(buf, left_key->data, left_key->size);
+    buf += left_key->size;
+  }
+
+  // serialize the right key if necessary
+  if (!h.right_is_infinite()) {
+    memcpy(buf, right_key->data, right_key->size);
+  }
+}
+
+void range_buffer::append_point(const DBT *key, bool is_exclusive) {
+  size_t record_length = sizeof(record_header) + key->size;
+  char *buf = reinterpret_cast<char *>(_arena.malloc_from_arena(record_length));
+
+  record_header h;
+  h.init(key, nullptr, is_exclusive);
+
+  // serialize the header
+  memcpy(buf, &h, sizeof(record_header));
+  buf += sizeof(record_header);
+
+  // serialize the key if necessary
+  if (!h.left_is_infinite()) {
+    memcpy(buf, key->data, key->size);
+  }
+}
+
+} /* namespace toku */
+#endif  // OS_WIN
+#endif  // ROCKSDB_LITE
diff --git a/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/locktree/range_buffer.h b/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/locktree/range_buffer.h
new file mode 100644
index 000000000..76e28d747
--- /dev/null
+++ b/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/locktree/range_buffer.h
@@ -0,0 +1,178 @@
+/* -*- mode: C++; c-basic-offset: 4; indent-tabs-mode: nil -*- */
+// vim: ft=cpp:expandtab:ts=8:sw=4:softtabstop=4:
+#ident "$Id$"
+/*======
+This file is part of PerconaFT.
+
+
+Copyright (c) 2006, 2015, Percona and/or its affiliates. All rights reserved.
+
+    PerconaFT is free software: you can redistribute it and/or modify
+    it under the terms of the GNU General Public License, version 2,
+    as published by the Free Software Foundation.
+
+    PerconaFT is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License
+    along with PerconaFT.  If not, see <http://www.gnu.org/licenses/>.
+
+----------------------------------------
+
+    PerconaFT is free software: you can redistribute it and/or modify
+    it under the terms of the GNU Affero General Public License, version 3,
+    as published by the Free Software Foundation.
+
+    PerconaFT is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU Affero General Public License for more details.
+
+    You should have received a copy of the GNU Affero General Public License
+    along with PerconaFT.  If not, see <http://www.gnu.org/licenses/>.
+
+----------------------------------------
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.
+======= */
+
+#ident \
+    "Copyright (c) 2006, 2015, Percona and/or its affiliates. All rights reserved."
+
+#pragma once
+
+#include <inttypes.h>
+#include <stdint.h>
+
+#include "../util/dbt.h"
+#include "../util/memarena.h"
+
+namespace toku {
+
+// a key range buffer represents a set of key ranges that can
+// be stored, iterated over, and then destroyed all at once.
+class range_buffer {
+ private:
+  // the key range buffer is a bunch of records in a row.
+  // each record has the following header, followed by the
+  // left key and right key data payload, if applicable.
+  // we limit keys to be 2^16, since we store lengths as 2 bytes.
+  static const size_t MAX_KEY_SIZE = 1 << 16;
+
+  struct record_header {
+    bool left_neg_inf;
+    bool left_pos_inf;
+    bool right_pos_inf;
+    bool right_neg_inf;
+    uint16_t left_key_size;
+    uint16_t right_key_size;
+    bool is_exclusive_lock;
+
+    bool left_is_infinite(void) const;
+
+    bool right_is_infinite(void) const;
+
+    void init(const DBT *left_key, const DBT *right_key, bool is_exclusive);
+  };
+  // PORT static_assert(sizeof(record_header) == 8, "record header format is
+  // off");
+
+ public:
+  // the iterator abstracts reading over a buffer of variable length
+  // records one by one until there are no more left.
+  class iterator {
+   public:
+    iterator();
+    iterator(const range_buffer *buffer);
+
+    // a record represents the user-view of a serialized key range.
+    // it handles positive and negative infinity and the optimized
+    // point range case, where left and right points share memory.
+    class record {
+     public:
+      // get a read-only pointer to the left key of this record's range
+      const DBT *get_left_key(void) const;
+
+      // get a read-only pointer to the right key of this record's range
+      const DBT *get_right_key(void) const;
+
+      // how big is this record? this tells us where the next record is
+      size_t size(void) const;
+
+      bool get_exclusive_flag() const { return _header.is_exclusive_lock; }
+
+      // populate a record header and point our DBT's
+      // buffers into ours if they are not infinite.
+      void deserialize(const char *buf);
+
+     private:
+      record_header _header;
+      DBT _left_key;
+      DBT _right_key;
+    };
+
+    // populate the given record object with the current
+    // the memory referred to by record is valid for only
+    // as long as the record exists.
+    bool current(record *rec);
+
+    // move the iterator to the next record in the buffer
+    void next(void);
+
+   private:
+    void reset_current_chunk();
+
+    // the key range buffer we are iterating over, the current
+    // offset in that buffer, and the size of the current record.
+    memarena::chunk_iterator _ma_chunk_iterator;
+    const void *_current_chunk_base;
+    size_t _current_chunk_offset;
+    size_t _current_chunk_max;
+    size_t _current_rec_size;
+  };
+
+  // allocate buffer space lazily instead of on creation. this way,
+  // no malloc/free is done if the transaction ends up taking no locks.
+  void create(void);
+
+  // append a left/right key range to the buffer.
+  // if the keys are equal, then only one copy is stored.
+  void append(const DBT *left_key, const DBT *right_key,
+              bool is_write_request = false);
+
+  // is this range buffer empty?
+  bool is_empty(void) const;
+
+  // how much memory is being used by this range buffer?
+  uint64_t total_memory_size(void) const;
+
+  // how many ranges are stored in this range buffer?
+  int get_num_ranges(void) const;
+
+  void destroy(void);
+
+ private:
+  memarena _arena;
+  int _num_ranges;
+
+  void append_range(const DBT *left_key, const DBT *right_key,
+                    bool is_write_request);
+
+  // append a point to the buffer. this is the space/time saving
+  // optimization for key ranges where left == right.
+  void append_point(const DBT *key, bool is_write_request);
+};
+
+} /* namespace toku */
diff --git a/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/locktree/treenode.cc b/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/locktree/treenode.cc
new file mode 100644
index 000000000..8997f634b
--- /dev/null
+++ b/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/locktree/treenode.cc
@@ -0,0 +1,520 @@
+/* -*- mode: C++; c-basic-offset: 4; indent-tabs-mode: nil -*- */
+// vim: ft=cpp:expandtab:ts=8:sw=4:softtabstop=4:
+#ifndef ROCKSDB_LITE
+#ifndef OS_WIN
+#ident "$Id$"
+/*======
+This file is part of PerconaFT.
+
+
+Copyright (c) 2006, 2015, Percona and/or its affiliates. All rights reserved.
+
+    PerconaFT is free software: you can redistribute it and/or modify
+    it under the terms of the GNU General Public License, version 2,
+    as published by the Free Software Foundation.
+
+    PerconaFT is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License
+    along with PerconaFT.  If not, see <http://www.gnu.org/licenses/>.
+
+----------------------------------------
+
+    PerconaFT is free software: you can redistribute it and/or modify
+    it under the terms of the GNU Affero General Public License, version 3,
+    as published by the Free Software Foundation.
+
+    PerconaFT is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU Affero General Public License for more details.
+
+    You should have received a copy of the GNU Affero General Public License
+    along with PerconaFT.  If not, see <http://www.gnu.org/licenses/>.
+
+----------------------------------------
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.
+======= */
+
+#ident \
+    "Copyright (c) 2006, 2015, Percona and/or its affiliates. All rights reserved."
+
+#include "treenode.h"
+
+#include "../portability/toku_race_tools.h"
+
+namespace toku {
+
+// TODO: source location info might have to be pulled up one caller
+// to be useful
+void treenode::mutex_lock(void) { toku_mutex_lock(&m_mutex); }
+
+void treenode::mutex_unlock(void) { toku_mutex_unlock(&m_mutex); }
+
+void treenode::init(const comparator *cmp) {
+  m_txnid = TXNID_NONE;
+  m_is_root = false;
+  m_is_empty = true;
+  m_cmp = cmp;
+
+  m_is_shared = false;
+  m_owners = nullptr;
+
+  // use an adaptive mutex at each node since we expect the time the
+  // lock is held to be relatively short compared to a context switch.
+  // indeed, this improves performance at high thread counts considerably.
+  memset(&m_mutex, 0, sizeof(toku_mutex_t));
+  toku_pthread_mutexattr_t attr;
+  toku_mutexattr_init(&attr);
+  toku_mutexattr_settype(&attr, TOKU_MUTEX_ADAPTIVE);
+  toku_mutex_init(treenode_mutex_key, &m_mutex, &attr);
+  toku_mutexattr_destroy(&attr);
+  m_left_child.set(nullptr);
+  m_right_child.set(nullptr);
+}
+
+void treenode::create_root(const comparator *cmp) {
+  init(cmp);
+  m_is_root = true;
+}
+
+void treenode::destroy_root(void) {
+  invariant(is_root());
+  invariant(is_empty());
+  toku_mutex_destroy(&m_mutex);
+  m_cmp = nullptr;
+}
+
+void treenode::set_range_and_txnid(const keyrange &range, TXNID txnid,
+                                   bool is_shared) {
+  // allocates a new copy of the range for this node
+  m_range.create_copy(range);
+  m_txnid = txnid;
+  m_is_shared = is_shared;
+  m_is_empty = false;
+}
+
+bool treenode::is_root(void) { return m_is_root; }
+
+bool treenode::is_empty(void) { return m_is_empty; }
+
+bool treenode::range_overlaps(const keyrange &range) {
+  return m_range.overlaps(*m_cmp, range);
+}
+
+treenode *treenode::alloc(const comparator *cmp, const keyrange &range,
+                          TXNID txnid, bool is_shared) {
+  treenode *XCALLOC(node);
+  node->init(cmp);
+  node->set_range_and_txnid(range, txnid, is_shared);
+  return node;
+}
+
+void treenode::swap_in_place(treenode *node1, treenode *node2) {
+  keyrange tmp_range = node1->m_range;
+  TXNID tmp_txnid = node1->m_txnid;
+  node1->m_range = node2->m_range;
+  node1->m_txnid = node2->m_txnid;
+  node2->m_range = tmp_range;
+  node2->m_txnid = tmp_txnid;
+
+  bool tmp_is_shared = node1->m_is_shared;
+  node1->m_is_shared = node2->m_is_shared;
+  node2->m_is_shared = tmp_is_shared;
+
+  auto tmp_m_owners = node1->m_owners;
+  node1->m_owners = node2->m_owners;
+  node2->m_owners = tmp_m_owners;
+}
+
+bool treenode::add_shared_owner(TXNID txnid) {
+  assert(m_is_shared);
+  if (txnid == m_txnid)
+    return false;  // acquiring a lock on the same range by the same trx
+
+  if (m_txnid != TXNID_SHARED) {
+    m_owners = new TxnidVector;
+    m_owners->insert(m_txnid);
+    m_txnid = TXNID_SHARED;
+  }
+  m_owners->insert(txnid);
+  return true;
+}
+
+void treenode::free(treenode *node) {
+  // destroy the range, freeing any copied keys
+  node->m_range.destroy();
+
+  if (node->m_owners) {
+    delete node->m_owners;
+    node->m_owners = nullptr;  // need this?
+  }
+
+  // the root is simply marked as empty.
+  if (node->is_root()) {
+    // PORT toku_mutex_assert_locked(&node->m_mutex);
+    node->m_is_empty = true;
+  } else {
+    // PORT toku_mutex_assert_unlocked(&node->m_mutex);
+    toku_mutex_destroy(&node->m_mutex);
+    toku_free(node);
+  }
+}
+
+uint32_t treenode::get_depth_estimate(void) const {
+  const uint32_t left_est = m_left_child.depth_est;
+  const uint32_t right_est = m_right_child.depth_est;
+  return (left_est > right_est ? left_est : right_est) + 1;
+}
+
+treenode *treenode::find_node_with_overlapping_child(
+    const keyrange &range, const keyrange::comparison *cmp_hint) {
+  // determine which child to look at based on a comparison. if we were
+  // given a comparison hint, use that. otherwise, compare them now.
+  keyrange::comparison c =
+      cmp_hint ? *cmp_hint : range.compare(*m_cmp, m_range);
+
+  treenode *child;
+  if (c == keyrange::comparison::LESS_THAN) {
+    child = lock_and_rebalance_left();
+  } else {
+    // The caller (locked_keyrange::acquire) handles the case where
+    // the root of the locked_keyrange is the node that overlaps.
+    // range is guaranteed not to overlap this node.
+    invariant(c == keyrange::comparison::GREATER_THAN);
+    child = lock_and_rebalance_right();
+  }
+
+  // if the search would lead us to an empty subtree (child == nullptr),
+  // or the child overlaps, then we know this node is the parent we want.
+  // otherwise we need to recur into that child.
+  if (child == nullptr) {
+    return this;
+  } else {
+    c = range.compare(*m_cmp, child->m_range);
+    if (c == keyrange::comparison::EQUALS ||
+        c == keyrange::comparison::OVERLAPS) {
+      child->mutex_unlock();
+      return this;
+    } else {
+      // unlock this node before recurring into the locked child,
+      // passing in a comparison hint since we just comapred range
+      // to the child's range.
+      mutex_unlock();
+      return child->find_node_with_overlapping_child(range, &c);
+    }
+  }
+}
+
+bool treenode::insert(const keyrange &range, TXNID txnid, bool is_shared) {
+  int rc = true;
+  // choose a child to check. if that child is null, then insert the new node
+  // there. otherwise recur down that child's subtree
+  keyrange::comparison c = range.compare(*m_cmp, m_range);
+  if (c == keyrange::comparison::LESS_THAN) {
+    treenode *left_child = lock_and_rebalance_left();
+    if (left_child == nullptr) {
+      left_child = treenode::alloc(m_cmp, range, txnid, is_shared);
+      m_left_child.set(left_child);
+    } else {
+      left_child->insert(range, txnid, is_shared);
+      left_child->mutex_unlock();
+    }
+  } else if (c == keyrange::comparison::GREATER_THAN) {
+    // invariant(c == keyrange::comparison::GREATER_THAN);
+    treenode *right_child = lock_and_rebalance_right();
+    if (right_child == nullptr) {
+      right_child = treenode::alloc(m_cmp, range, txnid, is_shared);
+      m_right_child.set(right_child);
+    } else {
+      right_child->insert(range, txnid, is_shared);
+      right_child->mutex_unlock();
+    }
+  } else if (c == keyrange::comparison::EQUALS) {
+    invariant(is_shared);
+    invariant(m_is_shared);
+    rc = add_shared_owner(txnid);
+  } else {
+    invariant(0);
+  }
+  return rc;
+}
+
+treenode *treenode::find_child_at_extreme(int direction, treenode **parent) {
+  treenode *child =
+      direction > 0 ? m_right_child.get_locked() : m_left_child.get_locked();
+
+  if (child) {
+    *parent = this;
+    treenode *child_extreme = child->find_child_at_extreme(direction, parent);
+    child->mutex_unlock();
+    return child_extreme;
+  } else {
+    return this;
+  }
+}
+
+treenode *treenode::find_leftmost_child(treenode **parent) {
+  return find_child_at_extreme(-1, parent);
+}
+
+treenode *treenode::find_rightmost_child(treenode **parent) {
+  return find_child_at_extreme(1, parent);
+}
+
+treenode *treenode::remove_root_of_subtree() {
+  // if this node has no children, just free it and return null
+  if (m_left_child.ptr == nullptr && m_right_child.ptr == nullptr) {
+    // treenode::free requires that non-root nodes are unlocked
+    if (!is_root()) {
+      mutex_unlock();
+    }
+    treenode::free(this);
+    return nullptr;
+  }
+
+  // we have a child, so get either the in-order successor or
+  // predecessor of this node to be our replacement.
+  // replacement_parent is updated by the find functions as
+  // they recur down the tree, so initialize it to this.
+  treenode *child, *replacement;
+  treenode *replacement_parent = this;
+  if (m_left_child.ptr != nullptr) {
+    child = m_left_child.get_locked();
+    replacement = child->find_rightmost_child(&replacement_parent);
+    invariant(replacement == child || replacement_parent != this);
+
+    // detach the replacement from its parent
+    if (replacement_parent == this) {
+      m_left_child = replacement->m_left_child;
+    } else {
+      replacement_parent->m_right_child = replacement->m_left_child;
+    }
+  } else {
+    child = m_right_child.get_locked();
+    replacement = child->find_leftmost_child(&replacement_parent);
+    invariant(replacement == child || replacement_parent != this);
+
+    // detach the replacement from its parent
+    if (replacement_parent == this) {
+      m_right_child = replacement->m_right_child;
+    } else {
+      replacement_parent->m_left_child = replacement->m_right_child;
+    }
+  }
+  child->mutex_unlock();
+
+  // swap in place with the detached replacement, then destroy it
+  treenode::swap_in_place(replacement, this);
+  treenode::free(replacement);
+
+  return this;
+}
+
+void treenode::recursive_remove(void) {
+  treenode *left = m_left_child.ptr;
+  if (left) {
+    left->recursive_remove();
+  }
+  m_left_child.set(nullptr);
+
+  treenode *right = m_right_child.ptr;
+  if (right) {
+    right->recursive_remove();
+  }
+  m_right_child.set(nullptr);
+
+  // we do not take locks on the way down, so we know non-root nodes
+  // are unlocked here and the caller is required to pass a locked
+  // root, so this free is correct.
+  treenode::free(this);
+}
+
+void treenode::remove_shared_owner(TXNID txnid) {
+  assert(m_owners->size() > 1);
+  m_owners->erase(txnid);
+  assert(m_owners->size() > 0);
+  /* if there is just one owner left, move it to m_txnid */
+  if (m_owners->size() == 1) {
+    m_txnid = *m_owners->begin();
+    delete m_owners;
+    m_owners = nullptr;
+  }
+}
+
+treenode *treenode::remove(const keyrange &range, TXNID txnid) {
+  treenode *child;
+  // if the range is equal to this node's range, then just remove
+  // the root of this subtree. otherwise search down the tree
+  // in either the left or right children.
+  keyrange::comparison c = range.compare(*m_cmp, m_range);
+  switch (c) {
+    case keyrange::comparison::EQUALS: {
+      // if we are the only owners, remove. Otherwise, just remove
+      // us from the owners list.
+      if (txnid != TXNID_ANY && has_multiple_owners()) {
+        remove_shared_owner(txnid);
+        return this;
+      } else {
+        return remove_root_of_subtree();
+      }
+    }
+    case keyrange::comparison::LESS_THAN:
+      child = m_left_child.get_locked();
+      invariant_notnull(child);
+      child = child->remove(range, txnid);
+
+      // unlock the child if there still is one.
+      // regardless, set the right child pointer
+      if (child) {
+        child->mutex_unlock();
+      }
+      m_left_child.set(child);
+      break;
+    case keyrange::comparison::GREATER_THAN:
+      child = m_right_child.get_locked();
+      invariant_notnull(child);
+      child = child->remove(range, txnid);
+
+      // unlock the child if there still is one.
+      // regardless, set the right child pointer
+      if (child) {
+        child->mutex_unlock();
+      }
+      m_right_child.set(child);
+      break;
+    case keyrange::comparison::OVERLAPS:
+      // shouldn't be overlapping, since the tree is
+      // non-overlapping and this range must exist
+      abort();
+  }
+
+  return this;
+}
+
+bool treenode::left_imbalanced(int threshold) const {
+  uint32_t left_depth = m_left_child.depth_est;
+  uint32_t right_depth = m_right_child.depth_est;
+  return m_left_child.ptr != nullptr && left_depth > threshold + right_depth;
+}
+
+bool treenode::right_imbalanced(int threshold) const {
+  uint32_t left_depth = m_left_child.depth_est;
+  uint32_t right_depth = m_right_child.depth_est;
+  return m_right_child.ptr != nullptr && right_depth > threshold + left_depth;
+}
+
+// effect: rebalances the subtree rooted at this node
+//         using AVL style O(1) rotations. unlocks this
+//         node if it is not the new root of the subtree.
+// requires: node is locked by this thread, children are not
+// returns: locked root node of the rebalanced tree
+treenode *treenode::maybe_rebalance(void) {
+  // if we end up not rotating at all, the new root is this
+  treenode *new_root = this;
+  treenode *child = nullptr;
+
+  if (left_imbalanced(IMBALANCE_THRESHOLD)) {
+    child = m_left_child.get_locked();
+    if (child->right_imbalanced(0)) {
+      treenode *grandchild = child->m_right_child.get_locked();
+
+      child->m_right_child = grandchild->m_left_child;
+      grandchild->m_left_child.set(child);
+
+      m_left_child = grandchild->m_right_child;
+      grandchild->m_right_child.set(this);
+
+      new_root = grandchild;
+    } else {
+      m_left_child = child->m_right_child;
+      child->m_right_child.set(this);
+      new_root = child;
+    }
+  } else if (right_imbalanced(IMBALANCE_THRESHOLD)) {
+    child = m_right_child.get_locked();
+    if (child->left_imbalanced(0)) {
+      treenode *grandchild = child->m_left_child.get_locked();
+
+      child->m_left_child = grandchild->m_right_child;
+      grandchild->m_right_child.set(child);
+
+      m_right_child = grandchild->m_left_child;
+      grandchild->m_left_child.set(this);
+
+      new_root = grandchild;
+    } else {
+      m_right_child = child->m_left_child;
+      child->m_left_child.set(this);
+      new_root = child;
+    }
+  }
+
+  // up to three nodes may be locked.
+  // - this
+  // - child
+  // - grandchild (but if it is locked, its the new root)
+  //
+  // one of them is the new root. we unlock everything except the new root.
+  if (child && child != new_root) {
+    TOKU_VALGRIND_RESET_MUTEX_ORDERING_INFO(&child->m_mutex);
+    child->mutex_unlock();
+  }
+  if (this != new_root) {
+    TOKU_VALGRIND_RESET_MUTEX_ORDERING_INFO(&m_mutex);
+    mutex_unlock();
+  }
+  TOKU_VALGRIND_RESET_MUTEX_ORDERING_INFO(&new_root->m_mutex);
+  return new_root;
+}
+
+treenode *treenode::lock_and_rebalance_left(void) {
+  treenode *child = m_left_child.get_locked();
+  if (child) {
+    treenode *new_root = child->maybe_rebalance();
+    m_left_child.set(new_root);
+    child = new_root;
+  }
+  return child;
+}
+
+treenode *treenode::lock_and_rebalance_right(void) {
+  treenode *child = m_right_child.get_locked();
+  if (child) {
+    treenode *new_root = child->maybe_rebalance();
+    m_right_child.set(new_root);
+    child = new_root;
+  }
+  return child;
+}
+
+void treenode::child_ptr::set(treenode *node) {
+  ptr = node;
+  depth_est = ptr ? ptr->get_depth_estimate() : 0;
+}
+
+treenode *treenode::child_ptr::get_locked(void) {
+  if (ptr) {
+    ptr->mutex_lock();
+    depth_est = ptr->get_depth_estimate();
+  }
+  return ptr;
+}
+
+} /* namespace toku */
+#endif  // OS_WIN
+#endif  // ROCKSDB_LITE
diff --git a/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/locktree/treenode.h b/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/locktree/treenode.h
new file mode 100644
index 000000000..ec25a8c58
--- /dev/null
+++ b/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/locktree/treenode.h
@@ -0,0 +1,302 @@
+/* -*- mode: C++; c-basic-offset: 4; indent-tabs-mode: nil -*- */
+// vim: ft=cpp:expandtab:ts=8:sw=2:softtabstop=2:
+#ident "$Id$"
+/*======
+This file is part of PerconaFT.
+
+
+Copyright (c) 2006, 2015, Percona and/or its affiliates. All rights reserved.
+
+    PerconaFT is free software: you can redistribute it and/or modify
+    it under the terms of the GNU General Public License, version 2,
+    as published by the Free Software Foundation.
+
+    PerconaFT is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License
+    along with PerconaFT.  If not, see <http://www.gnu.org/licenses/>.
+
+----------------------------------------
+
+    PerconaFT is free software: you can redistribute it and/or modify
+    it under the terms of the GNU Affero General Public License, version 3,
+    as published by the Free Software Foundation.
+
+    PerconaFT is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU Affero General Public License for more details.
+
+    You should have received a copy of the GNU Affero General Public License
+    along with PerconaFT.  If not, see <http://www.gnu.org/licenses/>.
+
+----------------------------------------
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.
+======= */
+
+#ident \
+    "Copyright (c) 2006, 2015, Percona and/or its affiliates. All rights reserved."
+
+#pragma once
+
+#include <string.h>
+
+#include "../ft/comparator.h"
+#include "../portability/memory.h"
+#include "../portability/toku_pthread.h"
+// PORT: we need LTM_STATUS
+#include "../ft/ft-status.h"
+#include "../portability/txn_subst.h"
+#include "keyrange.h"
+
+namespace toku {
+
+// a node in a tree with its own mutex
+// - range is the "key" of this node
+// - txnid is the single txnid associated with this node
+// - left and right children may be null
+//
+// to build a tree on top of this abstraction, the user:
+// - provides memory for a root node, initializes it via create_root()
+// - performs tree operations on the root node. memory management
+//   below the root node is handled by the abstraction, not the user.
+// this pattern:
+// - guaruntees a root node always exists.
+// - does not allow for rebalances on the root node
+
+class treenode {
+ public:
+  // every treenode function has some common requirements:
+  // - node is locked and children are never locked
+  // - node may be unlocked if no other thread has visibility
+
+  // effect: create the root node
+  void create_root(const comparator *cmp);
+
+  // effect: destroys the root node
+  void destroy_root(void);
+
+  // effect: sets the txnid and copies the given range for this node
+  void set_range_and_txnid(const keyrange &range, TXNID txnid, bool is_shared);
+
+  // returns: true iff this node is marked as empty
+  bool is_empty(void);
+
+  // returns: true if this is the root node, denoted by a null parent
+  bool is_root(void);
+
+  // returns: true if the given range overlaps with this node's range
+  bool range_overlaps(const keyrange &range);
+
+  // effect: locks the node
+  void mutex_lock(void);
+
+  // effect: unlocks the node
+  void mutex_unlock(void);
+
+  // return: node whose child overlaps, or a child that is empty
+  //         and would contain range if it existed
+  // given: if cmp_hint is non-null, then it is a precomputed
+  //        comparison of this node's range to the given range.
+  treenode *find_node_with_overlapping_child(
+      const keyrange &range, const keyrange::comparison *cmp_hint);
+
+  // effect: performs an in-order traversal of the ranges that overlap the
+  //         given range, calling function->fn() on each node that does
+  // requires: function signature is: bool fn(const keyrange &range, TXNID
+  // txnid) requires: fn returns true to keep iterating, false to stop iterating
+  // requires: fn does not attempt to use any ranges read out by value
+  //           after removing a node with an overlapping range from the tree.
+  template <class F>
+  void traverse_overlaps(const keyrange &range, F *function) {
+    keyrange::comparison c = range.compare(*m_cmp, m_range);
+    if (c == keyrange::comparison::EQUALS) {
+      // Doesn't matter if fn wants to keep going, there
+      // is nothing left, so return.
+      function->fn(m_range, m_txnid, m_is_shared, m_owners);
+      return;
+    }
+
+    treenode *left = m_left_child.get_locked();
+    if (left) {
+      if (c != keyrange::comparison::GREATER_THAN) {
+        // Target range is less than this node, or it overlaps this
+        // node.  There may be something on the left.
+        left->traverse_overlaps(range, function);
+      }
+      left->mutex_unlock();
+    }
+
+    if (c == keyrange::comparison::OVERLAPS) {
+      bool keep_going = function->fn(m_range, m_txnid, m_is_shared, m_owners);
+      if (!keep_going) {
+        return;
+      }
+    }
+
+    treenode *right = m_right_child.get_locked();
+    if (right) {
+      if (c != keyrange::comparison::LESS_THAN) {
+        // Target range is greater than this node, or it overlaps this
+        // node.  There may be something on the right.
+        right->traverse_overlaps(range, function);
+      }
+      right->mutex_unlock();
+    }
+  }
+
+  // effect: inserts the given range and txnid into a subtree, recursively
+  // requires: range does not overlap with any node below the subtree
+  bool insert(const keyrange &range, TXNID txnid, bool is_shared);
+
+  // effect: removes the given range from the subtree
+  // requires: range exists in the subtree
+  // returns: the root of the resulting subtree
+  treenode *remove(const keyrange &range, TXNID txnid);
+
+  // effect: removes this node and all of its children, recursively
+  // requires: every node at and below this node is unlocked
+  void recursive_remove(void);
+
+ private:
+  // the child_ptr is a light abstraction for the locking of
+  // a child and the maintenence of its depth estimate.
+
+  struct child_ptr {
+    // set the child pointer
+    void set(treenode *node);
+
+    // get and lock this child if it exists
+    treenode *get_locked(void);
+
+    treenode *ptr;
+    uint32_t depth_est;
+  };
+
+  // the balance factor at which a node is considered imbalanced
+  static const int32_t IMBALANCE_THRESHOLD = 2;
+
+  // node-level mutex
+  toku_mutex_t m_mutex;
+
+  // the range and txnid for this node. the range contains a copy
+  // of the keys originally inserted into the tree. nodes may
+  // swap ranges. but at the end of the day, when a node is
+  // destroyed, it frees the memory associated with whatever range
+  // it has at the time of destruction.
+  keyrange m_range;
+
+  void remove_shared_owner(TXNID txnid);
+
+  bool has_multiple_owners() { return (m_txnid == TXNID_SHARED); }
+
+ private:
+  // Owner transaction id.
+  // A value of TXNID_SHARED means this node has multiple owners
+  TXNID m_txnid;
+
+  // If true, this lock is a non-exclusive lock, and it can have either
+  // one or several owners.
+  bool m_is_shared;
+
+  // List of the owners, or nullptr if there's just one owner.
+  TxnidVector *m_owners;
+
+  // two child pointers
+  child_ptr m_left_child;
+  child_ptr m_right_child;
+
+  // comparator for ranges
+  // psergey-todo: Is there any sense to store the comparator in each tree
+  // node?
+  const comparator *m_cmp;
+
+  // marked for the root node. the root node is never free()'d
+  // when removed, but instead marked as empty.
+  bool m_is_root;
+
+  // marked for an empty node. only valid for the root.
+  bool m_is_empty;
+
+  // effect: initializes an empty node with the given comparator
+  void init(const comparator *cmp);
+
+  // requires: this is a shared node (m_is_shared==true)
+  // effect: another transaction is added as an owner.
+  // returns: true <=> added another owner
+  //          false <=> this transaction is already an owner
+  bool add_shared_owner(TXNID txnid);
+
+  // requires: *parent is initialized to something meaningful.
+  // requires: subtree is non-empty
+  // returns: the leftmost child of the given subtree
+  // returns: a pointer to the parent of said child in *parent, only
+  //          if this function recurred, otherwise it is untouched.
+  treenode *find_leftmost_child(treenode **parent);
+
+  // requires: *parent is initialized to something meaningful.
+  // requires: subtree is non-empty
+  // returns: the rightmost child of the given subtree
+  // returns: a pointer to the parent of said child in *parent, only
+  //          if this function recurred, otherwise it is untouched.
+  treenode *find_rightmost_child(treenode **parent);
+
+  // effect: remove the root of this subtree, destroying the old root
+  // returns: the new root of the subtree
+  treenode *remove_root_of_subtree(void);
+
+  // requires: subtree is non-empty, direction is not 0
+  // returns: the child of the subtree at either the left or rightmost extreme
+  treenode *find_child_at_extreme(int direction, treenode **parent);
+
+  // effect: retrieves and possibly rebalances the left child
+  // returns: a locked left child, if it exists
+  treenode *lock_and_rebalance_left(void);
+
+  // effect: retrieves and possibly rebalances the right child
+  // returns: a locked right child, if it exists
+  treenode *lock_and_rebalance_right(void);
+
+  // returns: the estimated depth of this subtree
+  uint32_t get_depth_estimate(void) const;
+
+  // returns: true iff left subtree depth is sufficiently less than the right
+  bool left_imbalanced(int threshold) const;
+
+  // returns: true iff right subtree depth is sufficiently greater than the left
+  bool right_imbalanced(int threshold) const;
+
+  // effect: performs an O(1) rebalance, which will "heal" an imbalance by at
+  // most 1. effect: if the new root is not this node, then this node is
+  // unlocked. returns: locked node representing the new root of the rebalanced
+  // subtree
+  treenode *maybe_rebalance(void);
+
+  // returns: allocated treenode populated with a copy of the range and txnid
+  static treenode *alloc(const comparator *cmp, const keyrange &range,
+                         TXNID txnid, bool is_shared);
+
+  // requires: node is a locked root node, or an unlocked non-root node
+  static void free(treenode *node);
+
+  // effect: swaps the range/txnid pairs for node1 and node2.
+  static void swap_in_place(treenode *node1, treenode *node2);
+
+  friend class concurrent_tree_unit_test;
+};
+
+} /* namespace toku */
diff --git a/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/locktree/txnid_set.cc b/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/locktree/txnid_set.cc
new file mode 100644
index 000000000..4caf1e26f
--- /dev/null
+++ b/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/locktree/txnid_set.cc
@@ -0,0 +1,120 @@
+/* -*- mode: C++; c-basic-offset: 4; indent-tabs-mode: nil -*- */
+// vim: ft=cpp:expandtab:ts=8:sw=4:softtabstop=4:
+#ifndef ROCKSDB_LITE
+#ifndef OS_WIN
+#ident "$Id$"
+/*======
+This file is part of PerconaFT.
+
+
+Copyright (c) 2006, 2015, Percona and/or its affiliates. All rights reserved.
+
+    PerconaFT is free software: you can redistribute it and/or modify
+    it under the terms of the GNU General Public License, version 2,
+    as published by the Free Software Foundation.
+
+    PerconaFT is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License
+    along with PerconaFT.  If not, see <http://www.gnu.org/licenses/>.
+
+----------------------------------------
+
+    PerconaFT is free software: you can redistribute it and/or modify
+    it under the terms of the GNU Affero General Public License, version 3,
+    as published by the Free Software Foundation.
+
+    PerconaFT is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU Affero General Public License for more details.
+
+    You should have received a copy of the GNU Affero General Public License
+    along with PerconaFT.  If not, see <http://www.gnu.org/licenses/>.
+
+----------------------------------------
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.
+======= */
+
+#ident \
+    "Copyright (c) 2006, 2015, Percona and/or its affiliates. All rights reserved."
+
+#include "txnid_set.h"
+
+#include "../db.h"
+
+namespace toku {
+
+int find_by_txnid(const TXNID &txnid_a, const TXNID &txnid_b);
+int find_by_txnid(const TXNID &txnid_a, const TXNID &txnid_b) {
+  if (txnid_a < txnid_b) {
+    return -1;
+  } else if (txnid_a == txnid_b) {
+    return 0;
+  } else {
+    return 1;
+  }
+}
+
+void txnid_set::create(void) {
+  // lazily allocate the underlying omt, since it is common
+  // to create a txnid set and never put anything in it.
+  m_txnids.create_no_array();
+}
+
+void txnid_set::destroy(void) { m_txnids.destroy(); }
+
+// Return true if the given transaction id is a member of the set.
+// Otherwise, return false.
+bool txnid_set::contains(TXNID txnid) const {
+  TXNID find_txnid;
+  int r = m_txnids.find_zero<TXNID, find_by_txnid>(txnid, &find_txnid, nullptr);
+  return r == 0 ? true : false;
+}
+
+// Add a given txnid to the set
+void txnid_set::add(TXNID txnid) {
+  int r = m_txnids.insert<TXNID, find_by_txnid>(txnid, txnid, nullptr);
+  invariant(r == 0 || r == DB_KEYEXIST);
+}
+
+// Delete a given txnid from the set.
+void txnid_set::remove(TXNID txnid) {
+  uint32_t idx;
+  int r = m_txnids.find_zero<TXNID, find_by_txnid>(txnid, nullptr, &idx);
+  if (r == 0) {
+    r = m_txnids.delete_at(idx);
+    invariant_zero(r);
+  }
+}
+
+// Return the size of the set
+uint32_t txnid_set::size(void) const { return m_txnids.size(); }
+
+// Get the ith id in the set, assuming that the set is sorted.
+TXNID txnid_set::get(uint32_t i) const {
+  TXNID txnid;
+  int r = m_txnids.fetch(i, &txnid);
+  if (r == EINVAL) /* Shouldn't happen, avoid compiler warning */
+    return TXNID_NONE;
+  invariant_zero(r);
+  return txnid;
+}
+
+} /* namespace toku */
+#endif  // OS_WIN
+#endif  // ROCKSDB_LITE
diff --git a/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/locktree/txnid_set.h b/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/locktree/txnid_set.h
new file mode 100644
index 000000000..d79c24fb0
--- /dev/null
+++ b/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/locktree/txnid_set.h
@@ -0,0 +1,92 @@
+/* -*- mode: C++; c-basic-offset: 4; indent-tabs-mode: nil -*- */
+// vim: ft=cpp:expandtab:ts=8:sw=4:softtabstop=4:
+#ident "$Id$"
+/*======
+This file is part of PerconaFT.
+
+
+Copyright (c) 2006, 2015, Percona and/or its affiliates. All rights reserved.
+
+    PerconaFT is free software: you can redistribute it and/or modify
+    it under the terms of the GNU General Public License, version 2,
+    as published by the Free Software Foundation.
+
+    PerconaFT is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License
+    along with PerconaFT.  If not, see <http://www.gnu.org/licenses/>.
+
+----------------------------------------
+
+    PerconaFT is free software: you can redistribute it and/or modify
+    it under the terms of the GNU Affero General Public License, version 3,
+    as published by the Free Software Foundation.
+
+    PerconaFT is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU Affero General Public License for more details.
+
+    You should have received a copy of the GNU Affero General Public License
+    along with PerconaFT.  If not, see <http://www.gnu.org/licenses/>.
+
+----------------------------------------
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.
+======= */
+
+#ident \
+    "Copyright (c) 2006, 2015, Percona and/or its affiliates. All rights reserved."
+
+#pragma once
+
+#include "../portability/txn_subst.h"
+#include "../util/omt.h"
+
+namespace toku {
+
+class txnid_set {
+ public:
+  // effect: Creates an empty set. Does not malloc space for
+  //         any entries yet. That is done lazily on add().
+  void create(void);
+
+  // effect: Destroy the set's internals.
+  void destroy(void);
+
+  // returns: True if the given txnid is a member of the set.
+  bool contains(TXNID id) const;
+
+  // effect: Adds a given txnid to the set if it did not exist
+  void add(TXNID txnid);
+
+  // effect: Deletes a txnid from the set if it exists.
+  void remove(TXNID txnid);
+
+  // returns: Size of the set
+  uint32_t size(void) const;
+
+  // returns: The "i'th" id in the set, as if it were sorted.
+  TXNID get(uint32_t i) const;
+
+ private:
+  toku::omt<TXNID> m_txnids;
+
+  friend class txnid_set_unit_test;
+};
+ENSURE_POD(txnid_set);
+
+} /* namespace toku */
diff --git a/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/locktree/wfg.cc b/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/locktree/wfg.cc
new file mode 100644
index 000000000..24536c88e
--- /dev/null
+++ b/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/locktree/wfg.cc
@@ -0,0 +1,213 @@
+/* -*- mode: C++; c-basic-offset: 4; indent-tabs-mode: nil -*- */
+// vim: ft=cpp:expandtab:ts=8:sw=4:softtabstop=4:
+#ifndef ROCKSDB_LITE
+#ifndef OS_WIN
+#ident "$Id$"
+/*======
+This file is part of PerconaFT.
+
+
+Copyright (c) 2006, 2015, Percona and/or its affiliates. All rights reserved.
+
+    PerconaFT is free software: you can redistribute it and/or modify
+    it under the terms of the GNU General Public License, version 2,
+    as published by the Free Software Foundation.
+
+    PerconaFT is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License
+    along with PerconaFT.  If not, see <http://www.gnu.org/licenses/>.
+
+----------------------------------------
+
+    PerconaFT is free software: you can redistribute it and/or modify
+    it under the terms of the GNU Affero General Public License, version 3,
+    as published by the Free Software Foundation.
+
+    PerconaFT is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU Affero General Public License for more details.
+
+    You should have received a copy of the GNU Affero General Public License
+    along with PerconaFT.  If not, see <http://www.gnu.org/licenses/>.
+
+----------------------------------------
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.
+======= */
+
+#ident \
+    "Copyright (c) 2006, 2015, Percona and/or its affiliates. All rights reserved."
+
+#include "../db.h"
+#include "../portability/memory.h"
+// PORT #include <toku_assert.h>
+#include <memory.h>
+#include <string.h>
+
+#include "txnid_set.h"
+#include "wfg.h"
+
+namespace toku {
+
+// Create a lock request graph
+void wfg::create(void) { m_nodes.create(); }
+
+// Destroy the internals of the lock request graph
+void wfg::destroy(void) {
+  uint32_t n_nodes = m_nodes.size();
+  for (uint32_t i = 0; i < n_nodes; i++) {
+    node *n;
+    int r = m_nodes.fetch(i, &n);
+    invariant_zero(r);
+    invariant_notnull(n);
+    if (r) continue;  // Get rid of "may be used uninitialized" warning
+    node::free(n);
+  }
+  m_nodes.destroy();
+}
+
+// Add an edge (a_id, b_id) to the graph
+void wfg::add_edge(TXNID a_txnid, TXNID b_txnid) {
+  node *a_node = find_create_node(a_txnid);
+  node *b_node = find_create_node(b_txnid);
+  a_node->edges.add(b_node->txnid);
+}
+
+// Return true if a node with the given transaction id exists in the graph.
+// Return false otherwise.
+bool wfg::node_exists(TXNID txnid) {
+  node *n = find_node(txnid);
+  return n != NULL;
+}
+
+bool wfg::cycle_exists_from_node(node *target, node *head,
+                                 std::function<void(TXNID)> reporter) {
+  bool cycle_found = false;
+  head->visited = true;
+  uint32_t n_edges = head->edges.size();
+  for (uint32_t i = 0; i < n_edges && !cycle_found; i++) {
+    TXNID edge_id = head->edges.get(i);
+    if (target->txnid == edge_id) {
+      cycle_found = true;
+      if (reporter) reporter(edge_id);
+    } else {
+      node *new_head = find_node(edge_id);
+      if (new_head && !new_head->visited) {
+        cycle_found = cycle_exists_from_node(target, new_head, reporter);
+        if (cycle_found && reporter) reporter(edge_id);
+      }
+    }
+  }
+  head->visited = false;
+  return cycle_found;
+}
+
+// Return true if there exists a cycle from a given transaction id in the graph.
+// Return false otherwise.
+bool wfg::cycle_exists_from_txnid(TXNID txnid,
+                                  std::function<void(TXNID)> reporter) {
+  node *a_node = find_node(txnid);
+  bool cycles_found = false;
+  if (a_node) {
+    cycles_found = cycle_exists_from_node(a_node, a_node, reporter);
+  }
+  return cycles_found;
+}
+
+// Apply a given function f to all of the nodes in the graph.  The apply
+// function returns when the function f is called for all of the nodes in the
+// graph, or the function f returns non-zero.
+void wfg::apply_nodes(int (*fn)(TXNID id, void *extra), void *extra) {
+  int r = 0;
+  uint32_t n_nodes = m_nodes.size();
+  for (uint32_t i = 0; i < n_nodes && r == 0; i++) {
+    node *n;
+    r = m_nodes.fetch(i, &n);
+    invariant_zero(r);
+    if (r) continue;  // Get rid of "may be used uninitialized" warning
+    r = fn(n->txnid, extra);
+  }
+}
+
+// Apply a given function f to all of the edges whose origin is a given node id.
+// The apply function returns when the function f is called for all edges in the
+// graph rooted at node id, or the function f returns non-zero.
+void wfg::apply_edges(TXNID txnid,
+                      int (*fn)(TXNID txnid, TXNID edge_txnid, void *extra),
+                      void *extra) {
+  node *n = find_node(txnid);
+  if (n) {
+    int r = 0;
+    uint32_t n_edges = n->edges.size();
+    for (uint32_t i = 0; i < n_edges && r == 0; i++) {
+      r = fn(txnid, n->edges.get(i), extra);
+    }
+  }
+}
+
+// find node by id
+wfg::node *wfg::find_node(TXNID txnid) {
+  node *n = nullptr;
+  int r = m_nodes.find_zero<TXNID, find_by_txnid>(txnid, &n, nullptr);
+  invariant(r == 0 || r == DB_NOTFOUND);
+  return n;
+}
+
+// this is the omt comparison function
+// nodes are compared by their txnid.
+int wfg::find_by_txnid(node *const &node_a, const TXNID &txnid_b) {
+  TXNID txnid_a = node_a->txnid;
+  if (txnid_a < txnid_b) {
+    return -1;
+  } else if (txnid_a == txnid_b) {
+    return 0;
+  } else {
+    return 1;
+  }
+}
+
+// insert a new node
+wfg::node *wfg::find_create_node(TXNID txnid) {
+  node *n;
+  uint32_t idx;
+  int r = m_nodes.find_zero<TXNID, find_by_txnid>(txnid, &n, &idx);
+  if (r == DB_NOTFOUND) {
+    n = node::alloc(txnid);
+    r = m_nodes.insert_at(n, idx);
+    invariant_zero(r);
+  }
+  invariant_notnull(n);
+  return n;
+}
+
+wfg::node *wfg::node::alloc(TXNID txnid) {
+  node *XCALLOC(n);
+  n->txnid = txnid;
+  n->visited = false;
+  n->edges.create();
+  return n;
+}
+
+void wfg::node::free(wfg::node *n) {
+  n->edges.destroy();
+  toku_free(n);
+}
+
+} /* namespace toku */
+#endif  // OS_WIN
+#endif  // ROCKSDB_LITE
diff --git a/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/locktree/wfg.h b/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/locktree/wfg.h
new file mode 100644
index 000000000..804202170
--- /dev/null
+++ b/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/locktree/wfg.h
@@ -0,0 +1,124 @@
+/* -*- mode: C++; c-basic-offset: 4; indent-tabs-mode: nil -*- */
+// vim: ft=cpp:expandtab:ts=8:sw=4:softtabstop=4:
+#ident "$Id$"
+/*======
+This file is part of PerconaFT.
+
+
+Copyright (c) 2006, 2015, Percona and/or its affiliates. All rights reserved.
+
+    PerconaFT is free software: you can redistribute it and/or modify
+    it under the terms of the GNU General Public License, version 2,
+    as published by the Free Software Foundation.
+
+    PerconaFT is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License
+    along with PerconaFT.  If not, see <http://www.gnu.org/licenses/>.
+
+----------------------------------------
+
+    PerconaFT is free software: you can redistribute it and/or modify
+    it under the terms of the GNU Affero General Public License, version 3,
+    as published by the Free Software Foundation.
+
+    PerconaFT is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU Affero General Public License for more details.
+
+    You should have received a copy of the GNU Affero General Public License
+    along with PerconaFT.  If not, see <http://www.gnu.org/licenses/>.
+
+----------------------------------------
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.
+======= */
+
+#ident \
+    "Copyright (c) 2006, 2015, Percona and/or its affiliates. All rights reserved."
+
+#pragma once
+
+#include <functional>
+
+#include "../util/omt.h"
+#include "txnid_set.h"
+
+namespace toku {
+
+// A wfg is a 'wait-for' graph. A directed edge in represents one
+// txn waiting for another to finish before it can acquire a lock.
+
+class wfg {
+ public:
+  // Create a lock request graph
+  void create(void);
+
+  // Destroy the internals of the lock request graph
+  void destroy(void);
+
+  // Add an edge (a_id, b_id) to the graph
+  void add_edge(TXNID a_txnid, TXNID b_txnid);
+
+  // Return true if a node with the given transaction id exists in the graph.
+  // Return false otherwise.
+  bool node_exists(TXNID txnid);
+
+  // Return true if there exists a cycle from a given transaction id in the
+  // graph. Return false otherwise.
+  bool cycle_exists_from_txnid(TXNID txnid,
+                               std::function<void(TXNID)> reporter);
+
+  // Apply a given function f to all of the nodes in the graph.  The apply
+  // function returns when the function f is called for all of the nodes in the
+  // graph, or the function f returns non-zero.
+  void apply_nodes(int (*fn)(TXNID txnid, void *extra), void *extra);
+
+  // Apply a given function f to all of the edges whose origin is a given node
+  // id. The apply function returns when the function f is called for all edges
+  // in the graph rooted at node id, or the function f returns non-zero.
+  void apply_edges(TXNID txnid,
+                   int (*fn)(TXNID txnid, TXNID edge_txnid, void *extra),
+                   void *extra);
+
+ private:
+  struct node {
+    // txnid for this node and the associated set of edges
+    TXNID txnid;
+    txnid_set edges;
+    bool visited;
+
+    static node *alloc(TXNID txnid);
+
+    static void free(node *n);
+  };
+  ENSURE_POD(node);
+
+  toku::omt<node *> m_nodes;
+
+  node *find_node(TXNID txnid);
+
+  node *find_create_node(TXNID txnid);
+
+  bool cycle_exists_from_node(node *target, node *head,
+                              std::function<void(TXNID)> reporter);
+
+  static int find_by_txnid(node *const &node_a, const TXNID &txnid_b);
+};
+ENSURE_POD(wfg);
+
+} /* namespace toku */
diff --git a/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/portability/memory.h b/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/portability/memory.h
new file mode 100644
index 000000000..0a621f8e0
--- /dev/null
+++ b/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/portability/memory.h
@@ -0,0 +1,215 @@
+/* -*- mode: C++; c-basic-offset: 4; indent-tabs-mode: nil -*- */
+// vim: ft=cpp:expandtab:ts=8:sw=4:softtabstop=4:
+#ident "$Id$"
+/*======
+This file is part of PerconaFT.
+
+
+Copyright (c) 2006, 2015, Percona and/or its affiliates. All rights reserved.
+
+    PerconaFT is free software: you can redistribute it and/or modify
+    it under the terms of the GNU General Public License, version 2,
+    as published by the Free Software Foundation.
+
+    PerconaFT is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License
+    along with PerconaFT.  If not, see <http://www.gnu.org/licenses/>.
+
+----------------------------------------
+
+    PerconaFT is free software: you can redistribute it and/or modify
+    it under the terms of the GNU Affero General Public License, version 3,
+    as published by the Free Software Foundation.
+
+    PerconaFT is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU Affero General Public License for more details.
+
+    You should have received a copy of the GNU Affero General Public License
+    along with PerconaFT.  If not, see <http://www.gnu.org/licenses/>.
+
+----------------------------------------
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.
+======= */
+
+#ident \
+    "Copyright (c) 2006, 2015, Percona and/or its affiliates. All rights reserved."
+
+#pragma once
+
+#include <stdlib.h>
+
+#include "toku_portability.h"
+
+/* Percona memory allocation functions and macros.
+ * These are functions for malloc and free */
+
+int toku_memory_startup(void) __attribute__((constructor));
+void toku_memory_shutdown(void) __attribute__((destructor));
+
+/* Generally: errno is set to 0 or a value to indicate problems. */
+
+// Everything should call toku_malloc() instead of malloc(), and toku_calloc()
+// instead of calloc() That way the tests can can, e.g.,  replace the malloc
+// function using toku_set_func_malloc().
+void *toku_calloc(size_t nmemb, size_t size)
+    __attribute__((__visibility__("default")));
+void *toku_xcalloc(size_t nmemb, size_t size)
+    __attribute__((__visibility__("default")));
+void *toku_malloc(size_t size) __attribute__((__visibility__("default")));
+void *toku_malloc_aligned(size_t alignment, size_t size)
+    __attribute__((__visibility__("default")));
+
+// xmalloc aborts instead of return NULL if we run out of memory
+void *toku_xmalloc(size_t size) __attribute__((__visibility__("default")));
+void *toku_xrealloc(void *, size_t size)
+    __attribute__((__visibility__("default")));
+void *toku_xmalloc_aligned(size_t alignment, size_t size)
+    __attribute__((__visibility__("default")));
+// Effect: Perform a os_malloc_aligned(size) with the additional property that
+// the returned pointer is a multiple of ALIGNMENT.
+//  Fail with a resource_assert if the allocation fails (don't return an error
+//  code). If the alloc_aligned function has been set then call it instead.
+// Requires: alignment is a power of two.
+
+void toku_free(void *) __attribute__((__visibility__("default")));
+
+size_t toku_malloc_usable_size(void *p)
+    __attribute__((__visibility__("default")));
+
+/* MALLOC is a macro that helps avoid a common error:
+ * Suppose I write
+ *    struct foo *x = malloc(sizeof(struct foo));
+ * That works fine.  But if I change it to this, I've probably made an mistake:
+ *    struct foo *x = malloc(sizeof(struct bar));
+ * It can get worse, since one might have something like
+ *    struct foo *x = malloc(sizeof(struct foo *))
+ * which looks reasonable, but it allocoates enough to hold a pointer instead of
+ * the amount needed for the struct. So instead, write struct foo *MALLOC(x);
+ * and you cannot go wrong.
+ */
+#define MALLOC(v) CAST_FROM_VOIDP(v, toku_malloc(sizeof(*v)))
+/* MALLOC_N is like calloc(Except no 0ing of data):  It makes an array.  Write
+ *   int *MALLOC_N(5,x);
+ * to make an array of 5 integers.
+ */
+#define MALLOC_N(n, v) CAST_FROM_VOIDP(v, toku_malloc((n) * sizeof(*v)))
+#define MALLOC_N_ALIGNED(align, n, v) \
+  CAST_FROM_VOIDP(v, toku_malloc_aligned((align), (n) * sizeof(*v)))
+
+// CALLOC_N is like calloc with auto-figuring out size of members
+#define CALLOC_N(n, v) CAST_FROM_VOIDP(v, toku_calloc((n), sizeof(*v)))
+
+#define CALLOC(v) CALLOC_N(1, v)
+
+// XMALLOC macros are like MALLOC except they abort if the operation fails
+#define XMALLOC(v) CAST_FROM_VOIDP(v, toku_xmalloc(sizeof(*v)))
+#define XMALLOC_N(n, v) CAST_FROM_VOIDP(v, toku_xmalloc((n) * sizeof(*v)))
+#define XCALLOC_N(n, v) CAST_FROM_VOIDP(v, toku_xcalloc((n), (sizeof(*v))))
+#define XCALLOC(v) XCALLOC_N(1, v)
+#define XREALLOC(v, s) CAST_FROM_VOIDP(v, toku_xrealloc(v, s))
+#define XREALLOC_N(n, v) CAST_FROM_VOIDP(v, toku_xrealloc(v, (n) * sizeof(*v)))
+
+#define XMALLOC_N_ALIGNED(align, n, v) \
+  CAST_FROM_VOIDP(v, toku_xmalloc_aligned((align), (n) * sizeof(*v)))
+
+#define XMEMDUP(dst, src) CAST_FROM_VOIDP(dst, toku_xmemdup(src, sizeof(*src)))
+#define XMEMDUP_N(dst, src, len) CAST_FROM_VOIDP(dst, toku_xmemdup(src, len))
+
+// ZERO_ARRAY writes zeroes to a stack-allocated array
+#define ZERO_ARRAY(o)          \
+  do {                         \
+    memset((o), 0, sizeof(o)); \
+  } while (0)
+// ZERO_STRUCT writes zeroes to a stack-allocated struct
+#define ZERO_STRUCT(o)          \
+  do {                          \
+    memset(&(o), 0, sizeof(o)); \
+  } while (0)
+
+/* Copy memory.  Analogous to strdup() */
+void *toku_memdup(const void *v, size_t len);
+/* Toku-version of strdup.  Use this so that it calls toku_malloc() */
+char *toku_strdup(const char *s) __attribute__((__visibility__("default")));
+/* Toku-version of strndup.  Use this so that it calls toku_malloc() */
+char *toku_strndup(const char *s, size_t n)
+    __attribute__((__visibility__("default")));
+/* Copy memory.  Analogous to strdup() Crashes instead of returning NULL */
+void *toku_xmemdup(const void *v, size_t len)
+    __attribute__((__visibility__("default")));
+/* Toku-version of strdup.  Use this so that it calls toku_xmalloc()  Crashes
+ * instead of returning NULL */
+char *toku_xstrdup(const char *s) __attribute__((__visibility__("default")));
+
+void toku_malloc_cleanup(
+    void); /* Before exiting, call this function to free up any internal data
+              structures from toku_malloc.  Otherwise valgrind will complain of
+              memory leaks. */
+
+/* Check to see if everything malloc'd was free.  Might be a no-op depending on
+ * how memory.c is configured. */
+void toku_memory_check_all_free(void);
+/* Check to see if memory is "sane".  Might be a no-op.  Probably better to
+ * simply use valgrind. */
+void toku_do_memory_check(void);
+
+typedef void *(*malloc_fun_t)(size_t);
+typedef void (*free_fun_t)(void *);
+typedef void *(*realloc_fun_t)(void *, size_t);
+typedef void *(*malloc_aligned_fun_t)(size_t /*alignment*/, size_t /*size*/);
+typedef void *(*realloc_aligned_fun_t)(size_t /*alignment*/, void * /*pointer*/,
+                                       size_t /*size*/);
+
+void toku_set_func_malloc(malloc_fun_t f);
+void toku_set_func_xmalloc_only(malloc_fun_t f);
+void toku_set_func_malloc_only(malloc_fun_t f);
+void toku_set_func_realloc(realloc_fun_t f);
+void toku_set_func_xrealloc_only(realloc_fun_t f);
+void toku_set_func_realloc_only(realloc_fun_t f);
+void toku_set_func_free(free_fun_t f);
+
+typedef struct memory_status {
+  uint64_t malloc_count;   // number of malloc operations
+  uint64_t free_count;     // number of free operations
+  uint64_t realloc_count;  // number of realloc operations
+  uint64_t malloc_fail;    // number of malloc operations that failed
+  uint64_t realloc_fail;   // number of realloc operations that failed
+  uint64_t requested;      // number of bytes requested
+  uint64_t used;   // number of bytes used (requested + overhead), obtained from
+                   // malloc_usable_size()
+  uint64_t freed;  // number of bytes freed;
+  uint64_t max_requested_size;  // largest attempted allocation size
+  uint64_t last_failed_size;    // size of the last failed allocation attempt
+  volatile uint64_t
+      max_in_use;  // maximum memory footprint (used - freed), approximate (not
+                   // worth threadsafety overhead for exact)
+  const char *mallocator_version;
+  uint64_t mmap_threshold;
+} LOCAL_MEMORY_STATUS_S, *LOCAL_MEMORY_STATUS;
+
+void toku_memory_get_status(LOCAL_MEMORY_STATUS s);
+
+// Effect: Like toku_memory_footprint, except instead of passing p,
+//   we pass toku_malloc_usable_size(p).
+size_t toku_memory_footprint_given_usable_size(size_t touched, size_t usable);
+
+// Effect: Return an estimate how how much space an object is using, possibly by
+//   using toku_malloc_usable_size(p).
+//   If p is NULL then returns 0.
+size_t toku_memory_footprint(void *p, size_t touched);
diff --git a/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/portability/toku_assert_subst.h b/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/portability/toku_assert_subst.h
new file mode 100644
index 000000000..af47800fb
--- /dev/null
+++ b/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/portability/toku_assert_subst.h
@@ -0,0 +1,39 @@
+//
+// A replacement for toku_assert.h
+//
+#pragma once
+
+#include <assert.h>
+#include <errno.h>
+
+#ifdef NDEBUG
+
+#define assert_zero(a) ((void)(a))
+#define invariant(a) ((void)(a))
+#define invariant_notnull(a) ((void)(a))
+#define invariant_zero(a) ((void)(a))
+
+#else
+
+#define assert_zero(a) assert((a) == 0)
+#define invariant(a) assert(a)
+#define invariant_notnull(a) assert(a)
+#define invariant_zero(a) assert_zero(a)
+
+#endif
+
+#define lazy_assert_zero(a) assert_zero(a)
+
+#define paranoid_invariant_zero(a) assert_zero(a)
+#define paranoid_invariant_notnull(a) assert(a)
+#define paranoid_invariant(a) assert(a)
+
+#define ENSURE_POD(type)                                                    \
+  static_assert(                                                            \
+      std::is_standard_layout<type>::value && std::is_trivial<type>::value, \
+      #type "isn't POD")
+
+inline int get_error_errno(void) {
+  invariant(errno);
+  return errno;
+}
diff --git a/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/portability/toku_atomic.h b/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/portability/toku_atomic.h
new file mode 100644
index 000000000..aaa2298fa
--- /dev/null
+++ b/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/portability/toku_atomic.h
@@ -0,0 +1,130 @@
+/* -*- mode: C++; c-basic-offset: 4; indent-tabs-mode: nil -*- */
+// vim: ft=cpp:expandtab:ts=8:sw=4:softtabstop=4:
+#ident "$Id$"
+/*======
+This file is part of PerconaFT.
+
+
+Copyright (c) 2006, 2015, Percona and/or its affiliates. All rights reserved.
+
+    PerconaFT is free software: you can redistribute it and/or modify
+    it under the terms of the GNU General Public License, version 2,
+    as published by the Free Software Foundation.
+
+    PerconaFT is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License
+    along with PerconaFT.  If not, see <http://www.gnu.org/licenses/>.
+
+----------------------------------------
+
+    PerconaFT is free software: you can redistribute it and/or modify
+    it under the terms of the GNU Affero General Public License, version 3,
+    as published by the Free Software Foundation.
+
+    PerconaFT is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU Affero General Public License for more details.
+
+    You should have received a copy of the GNU Affero General Public License
+    along with PerconaFT.  If not, see <http://www.gnu.org/licenses/>.
+
+----------------------------------------
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.
+======= */
+
+#ident \
+    "Copyright (c) 2006, 2015, Percona and/or its affiliates. All rights reserved."
+
+#pragma once
+
+// PORT2: #include <portability/toku_config.h>
+#include <stdbool.h>
+#include <stddef.h>
+#include <stdint.h>
+
+#include "toku_assert_subst.h"
+
+__attribute__((const, always_inline)) static inline intptr_t which_cache_line(
+    intptr_t addr) {
+  static const size_t assumed_cache_line_size = 64;
+  return addr / assumed_cache_line_size;
+}
+template <typename T>
+__attribute__((const, always_inline)) static inline bool crosses_boundary(
+    T *addr, size_t width) {
+  const intptr_t int_addr = reinterpret_cast<intptr_t>(addr);
+  const intptr_t last_byte = int_addr + width - 1;
+  return which_cache_line(int_addr) != which_cache_line(last_byte);
+}
+
+template <typename T, typename U>
+__attribute__((always_inline)) static inline T toku_sync_fetch_and_add(T *addr,
+                                                                       U diff) {
+  paranoid_invariant(!crosses_boundary(addr, sizeof *addr));
+  return __sync_fetch_and_add(addr, diff);
+}
+template <typename T, typename U>
+__attribute__((always_inline)) static inline T toku_sync_add_and_fetch(T *addr,
+                                                                       U diff) {
+  paranoid_invariant(!crosses_boundary(addr, sizeof *addr));
+  return __sync_add_and_fetch(addr, diff);
+}
+template <typename T, typename U>
+__attribute__((always_inline)) static inline T toku_sync_fetch_and_sub(T *addr,
+                                                                       U diff) {
+  paranoid_invariant(!crosses_boundary(addr, sizeof *addr));
+  return __sync_fetch_and_sub(addr, diff);
+}
+template <typename T, typename U>
+__attribute__((always_inline)) static inline T toku_sync_sub_and_fetch(T *addr,
+                                                                       U diff) {
+  paranoid_invariant(!crosses_boundary(addr, sizeof *addr));
+  return __sync_sub_and_fetch(addr, diff);
+}
+template <typename T, typename U, typename V>
+__attribute__((always_inline)) static inline T toku_sync_val_compare_and_swap(
+    T *addr, U oldval, V newval) {
+  paranoid_invariant(!crosses_boundary(addr, sizeof *addr));
+  return __sync_val_compare_and_swap(addr, oldval, newval);
+}
+template <typename T, typename U, typename V>
+__attribute__((always_inline)) static inline bool
+toku_sync_bool_compare_and_swap(T *addr, U oldval, V newval) {
+  paranoid_invariant(!crosses_boundary(addr, sizeof *addr));
+  return __sync_bool_compare_and_swap(addr, oldval, newval);
+}
+
+// in case you include this but not toku_portability.h
+#pragma GCC poison __sync_fetch_and_add
+#pragma GCC poison __sync_fetch_and_sub
+#pragma GCC poison __sync_fetch_and_or
+#pragma GCC poison __sync_fetch_and_and
+#pragma GCC poison __sync_fetch_and_xor
+#pragma GCC poison __sync_fetch_and_nand
+#pragma GCC poison __sync_add_and_fetch
+#pragma GCC poison __sync_sub_and_fetch
+#pragma GCC poison __sync_or_and_fetch
+#pragma GCC poison __sync_and_and_fetch
+#pragma GCC poison __sync_xor_and_fetch
+#pragma GCC poison __sync_nand_and_fetch
+#pragma GCC poison __sync_bool_compare_and_swap
+#pragma GCC poison __sync_val_compare_and_swap
+#pragma GCC poison __sync_synchronize
+#pragma GCC poison __sync_lock_test_and_set
+#pragma GCC poison __sync_release
diff --git a/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/portability/toku_external_pthread.h b/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/portability/toku_external_pthread.h
new file mode 100644
index 000000000..eb8291c1d
--- /dev/null
+++ b/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/portability/toku_external_pthread.h
@@ -0,0 +1,83 @@
+/*
+  A wrapper around ROCKSDB_NAMESPACE::TransactionDBMutexFactory-provided
+  condition and mutex that provides toku_pthread_*-like interface. The functions
+  are named
+
+    toku_external_{mutex|cond}_XXX
+
+  Lock Tree uses this mutex and condition for interruptible (long) lock waits.
+
+  (It also still uses toku_pthread_XXX calls for mutexes/conditions for
+   shorter waits on internal objects)
+*/
+
+#pragma once
+
+#include <pthread.h>
+#include <stdint.h>
+#include <time.h>
+
+#include "rocksdb/utilities/transaction_db.h"
+#include "rocksdb/utilities/transaction_db_mutex.h"
+#include "toku_portability.h"
+
+using ROCKSDB_NAMESPACE::TransactionDBCondVar;
+using ROCKSDB_NAMESPACE::TransactionDBMutex;
+
+typedef std::shared_ptr<ROCKSDB_NAMESPACE::TransactionDBMutexFactory>
+    toku_external_mutex_factory_t;
+
+typedef std::shared_ptr<TransactionDBMutex> toku_external_mutex_t;
+typedef std::shared_ptr<TransactionDBCondVar> toku_external_cond_t;
+
+static inline void toku_external_cond_init(
+    toku_external_mutex_factory_t mutex_factory, toku_external_cond_t *cond) {
+  *cond = mutex_factory->AllocateCondVar();
+}
+
+inline void toku_external_cond_destroy(toku_external_cond_t *cond) {
+  cond->reset();  // this will destroy the managed object
+}
+
+inline void toku_external_cond_signal(toku_external_cond_t *cond) {
+  (*cond)->Notify();
+}
+
+inline void toku_external_cond_broadcast(toku_external_cond_t *cond) {
+  (*cond)->NotifyAll();
+}
+
+inline int toku_external_cond_timedwait(toku_external_cond_t *cond,
+                                        toku_external_mutex_t *mutex,
+                                        int64_t timeout_microsec) {
+  auto res = (*cond)->WaitFor(*mutex, timeout_microsec);
+  if (res.ok())
+    return 0;
+  else
+    return ETIMEDOUT;
+}
+
+inline void toku_external_mutex_init(toku_external_mutex_factory_t factory,
+                                     toku_external_mutex_t *mutex) {
+  // Use placement new: the memory has been allocated but constructor wasn't
+  // called
+  new (mutex) toku_external_mutex_t;
+  *mutex = factory->AllocateMutex();
+}
+
+inline void toku_external_mutex_lock(toku_external_mutex_t *mutex) {
+  (*mutex)->Lock();
+}
+
+inline int toku_external_mutex_trylock(toku_external_mutex_t *mutex) {
+  (*mutex)->Lock();
+  return 0;
+}
+
+inline void toku_external_mutex_unlock(toku_external_mutex_t *mutex) {
+  (*mutex)->UnLock();
+}
+
+inline void toku_external_mutex_destroy(toku_external_mutex_t *mutex) {
+  mutex->reset();  // this will destroy the managed object
+}
diff --git a/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/portability/toku_instrumentation.h b/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/portability/toku_instrumentation.h
new file mode 100644
index 000000000..c967e7177
--- /dev/null
+++ b/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/portability/toku_instrumentation.h
@@ -0,0 +1,286 @@
+/*======
+This file is part of PerconaFT.
+
+Copyright (c) 2006, 2015, Percona and/or its affiliates. All rights reserved.
+
+    PerconaFT is free software: you can redistribute it and/or modify
+    it under the terms of the GNU General Public License, version 2,
+    as published by the Free Software Foundation.
+
+    PerconaFT is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License
+    along with PerconaFT.  If not, see <http://www.gnu.org/licenses/>.
+
+----------------------------------------
+
+    PerconaFT is free software: you can redistribute it and/or modify
+    it under the terms of the GNU Affero General Public License, version 3,
+    as published by the Free Software Foundation.
+
+    PerconaFT is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU Affero General Public License for more details.
+
+    You should have received a copy of the GNU Affero General Public License
+    along with PerconaFT.  If not, see <http://www.gnu.org/licenses/>.
+
+----------------------------------------
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.
+======= */
+
+#pragma once
+
+#include <stdio.h>  // FILE
+
+// Performance instrumentation object identifier type
+typedef unsigned int pfs_key_t;
+
+enum class toku_instr_object_type { mutex, rwlock, cond, thread, file };
+
+struct PSI_file;
+
+struct TOKU_FILE {
+  /** The real file. */
+  FILE *file;
+  struct PSI_file *key;
+  TOKU_FILE() : file(nullptr), key(nullptr) {}
+};
+
+struct PSI_mutex;
+struct PSI_cond;
+struct PSI_rwlock;
+
+struct toku_mutex_t;
+struct toku_cond_t;
+struct toku_pthread_rwlock_t;
+
+class toku_instr_key;
+
+class toku_instr_probe_empty {
+ public:
+  explicit toku_instr_probe_empty(UU(const toku_instr_key &key)) {}
+
+  void start_with_source_location(UU(const char *src_file), UU(int src_line)) {}
+
+  void stop() {}
+};
+
+#define TOKU_PROBE_START(p) p->start_with_source_location(__FILE__, __LINE__)
+#define TOKU_PROBE_STOP(p) p->stop
+
+extern toku_instr_key toku_uninstrumented;
+
+#ifndef MYSQL_TOKUDB_ENGINE
+
+#include <pthread.h>
+
+class toku_instr_key {
+ public:
+  toku_instr_key(UU(toku_instr_object_type type), UU(const char *group),
+                 UU(const char *name)) {}
+
+  explicit toku_instr_key(UU(pfs_key_t key_id)) {}
+  // No-instrumentation constructor:
+  toku_instr_key() {}
+  ~toku_instr_key() {}
+};
+
+typedef toku_instr_probe_empty toku_instr_probe;
+
+enum class toku_instr_file_op {
+  file_stream_open,
+  file_create,
+  file_open,
+  file_delete,
+  file_rename,
+  file_read,
+  file_write,
+  file_sync,
+  file_stream_close,
+  file_close,
+  file_stat
+};
+
+struct PSI_file {};
+struct PSI_mutex {};
+
+struct toku_io_instrumentation {};
+
+inline int toku_pthread_create(UU(const toku_instr_key &key), pthread_t *thread,
+                               const pthread_attr_t *attr,
+                               void *(*start_routine)(void *), void *arg) {
+  return pthread_create(thread, attr, start_routine, arg);
+}
+
+inline void toku_instr_register_current_thread() {}
+
+inline void toku_instr_delete_current_thread() {}
+
+// Instrument file creation, opening, closing, and renaming
+inline void toku_instr_file_open_begin(UU(toku_io_instrumentation &io_instr),
+                                       UU(const toku_instr_key &key),
+                                       UU(toku_instr_file_op op),
+                                       UU(const char *name),
+                                       UU(const char *src_file),
+                                       UU(int src_line)) {}
+
+inline void toku_instr_file_stream_open_end(
+    UU(toku_io_instrumentation &io_instr), UU(TOKU_FILE &file)) {}
+
+inline void toku_instr_file_open_end(UU(toku_io_instrumentation &io_instr),
+                                     UU(int fd)) {}
+
+inline void toku_instr_file_name_close_begin(
+    UU(toku_io_instrumentation &io_instr), UU(const toku_instr_key &key),
+    UU(toku_instr_file_op op), UU(const char *name), UU(const char *src_file),
+    UU(int src_line)) {}
+
+inline void toku_instr_file_stream_close_begin(
+    UU(toku_io_instrumentation &io_instr), UU(toku_instr_file_op op),
+    UU(TOKU_FILE &file), UU(const char *src_file), UU(int src_line)) {}
+
+inline void toku_instr_file_fd_close_begin(
+    UU(toku_io_instrumentation &io_instr), UU(toku_instr_file_op op),
+    UU(int fd), UU(const char *src_file), UU(int src_line)) {}
+
+inline void toku_instr_file_close_end(UU(toku_io_instrumentation &io_instr),
+                                      UU(int result)) {}
+
+inline void toku_instr_file_io_begin(UU(toku_io_instrumentation &io_instr),
+                                     UU(toku_instr_file_op op), UU(int fd),
+                                     UU(unsigned int count),
+                                     UU(const char *src_file),
+                                     UU(int src_line)) {}
+
+inline void toku_instr_file_name_io_begin(
+    UU(toku_io_instrumentation &io_instr), UU(const toku_instr_key &key),
+    UU(toku_instr_file_op op), UU(const char *name), UU(unsigned int count),
+    UU(const char *src_file), UU(int src_line)) {}
+
+inline void toku_instr_file_stream_io_begin(
+    UU(toku_io_instrumentation &io_instr), UU(toku_instr_file_op op),
+    UU(TOKU_FILE &file), UU(unsigned int count), UU(const char *src_file),
+    UU(int src_line)) {}
+
+inline void toku_instr_file_io_end(UU(toku_io_instrumentation &io_instr),
+                                   UU(unsigned int count)) {}
+
+struct toku_mutex_t;
+
+struct toku_mutex_instrumentation {};
+
+inline PSI_mutex *toku_instr_mutex_init(UU(const toku_instr_key &key),
+                                        UU(toku_mutex_t &mutex)) {
+  return nullptr;
+}
+
+inline void toku_instr_mutex_destroy(UU(PSI_mutex *&mutex_instr)) {}
+
+inline void toku_instr_mutex_lock_start(
+    UU(toku_mutex_instrumentation &mutex_instr), UU(toku_mutex_t &mutex),
+    UU(const char *src_file), UU(int src_line)) {}
+
+inline void toku_instr_mutex_trylock_start(
+    UU(toku_mutex_instrumentation &mutex_instr), UU(toku_mutex_t &mutex),
+    UU(const char *src_file), UU(int src_line)) {}
+
+inline void toku_instr_mutex_lock_end(
+    UU(toku_mutex_instrumentation &mutex_instr),
+    UU(int pthread_mutex_lock_result)) {}
+
+inline void toku_instr_mutex_unlock(UU(PSI_mutex *mutex_instr)) {}
+
+struct toku_cond_instrumentation {};
+
+enum class toku_instr_cond_op {
+  cond_wait,
+  cond_timedwait,
+};
+
+inline PSI_cond *toku_instr_cond_init(UU(const toku_instr_key &key),
+                                      UU(toku_cond_t &cond)) {
+  return nullptr;
+}
+
+inline void toku_instr_cond_destroy(UU(PSI_cond *&cond_instr)) {}
+
+inline void toku_instr_cond_wait_start(
+    UU(toku_cond_instrumentation &cond_instr), UU(toku_instr_cond_op op),
+    UU(toku_cond_t &cond), UU(toku_mutex_t &mutex), UU(const char *src_file),
+    UU(int src_line)) {}
+
+inline void toku_instr_cond_wait_end(UU(toku_cond_instrumentation &cond_instr),
+                                     UU(int pthread_cond_wait_result)) {}
+
+inline void toku_instr_cond_signal(UU(toku_cond_t &cond)) {}
+
+inline void toku_instr_cond_broadcast(UU(toku_cond_t &cond)) {}
+
+#if 0
+// rw locks are not used 
+// rwlock instrumentation
+struct toku_rwlock_instrumentation {};
+
+inline PSI_rwlock *toku_instr_rwlock_init(UU(const toku_instr_key &key),
+                                          UU(toku_pthread_rwlock_t &rwlock)) {
+    return nullptr;
+}
+
+inline void toku_instr_rwlock_destroy(UU(PSI_rwlock *&rwlock_instr)) {}
+
+inline void toku_instr_rwlock_rdlock_wait_start(
+    UU(toku_rwlock_instrumentation &rwlock_instr),
+    UU(toku_pthread_rwlock_t &rwlock),
+    UU(const char *src_file),
+    UU(int src_line)) {}
+
+inline void toku_instr_rwlock_wrlock_wait_start(
+    UU(toku_rwlock_instrumentation &rwlock_instr),
+    UU(toku_pthread_rwlock_t &rwlock),
+    UU(const char *src_file),
+    UU(int src_line)) {}
+
+inline void toku_instr_rwlock_rdlock_wait_end(
+    UU(toku_rwlock_instrumentation &rwlock_instr),
+    UU(int pthread_rwlock_wait_result)) {}
+
+inline void toku_instr_rwlock_wrlock_wait_end(
+    UU(toku_rwlock_instrumentation &rwlock_instr),
+    UU(int pthread_rwlock_wait_result)) {}
+
+inline void toku_instr_rwlock_unlock(UU(toku_pthread_rwlock_t &rwlock)) {}
+#endif
+
+#else  // MYSQL_TOKUDB_ENGINE
+// There can be not only mysql but also mongodb or any other PFS stuff
+#include <toku_instr_mysql.h>
+#endif  // MYSQL_TOKUDB_ENGINE
+
+// Mutexes
+extern toku_instr_key manager_escalation_mutex_key;
+extern toku_instr_key manager_escalator_mutex_key;
+extern toku_instr_key manager_mutex_key;
+extern toku_instr_key treenode_mutex_key;
+extern toku_instr_key locktree_request_info_mutex_key;
+extern toku_instr_key locktree_request_info_retry_mutex_key;
+
+// condition vars
+extern toku_instr_key lock_request_m_wait_cond_key;
+extern toku_instr_key locktree_request_info_retry_cv_key;
+extern toku_instr_key manager_m_escalator_done_key;  // unused
diff --git a/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/portability/toku_portability.h b/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/portability/toku_portability.h
new file mode 100644
index 000000000..9a95b38bd
--- /dev/null
+++ b/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/portability/toku_portability.h
@@ -0,0 +1,87 @@
+/* -*- mode: C++; c-basic-offset: 4; indent-tabs-mode: nil -*- */
+// vim: ft=cpp:expandtab:ts=8:sw=4:softtabstop=4:
+#ident "$Id$"
+/*======
+This file is part of PerconaFT.
+
+
+Copyright (c) 2006, 2015, Percona and/or its affiliates. All rights reserved.
+
+    PerconaFT is free software: you can redistribute it and/or modify
+    it under the terms of the GNU General Public License, version 2,
+    as published by the Free Software Foundation.
+
+    PerconaFT is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License
+    along with PerconaFT.  If not, see <http://www.gnu.org/licenses/>.
+
+----------------------------------------
+
+    PerconaFT is free software: you can redistribute it and/or modify
+    it under the terms of the GNU Affero General Public License, version 3,
+    as published by the Free Software Foundation.
+
+    PerconaFT is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU Affero General Public License for more details.
+
+    You should have received a copy of the GNU Affero General Public License
+    along with PerconaFT.  If not, see <http://www.gnu.org/licenses/>.
+
+----------------------------------------
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.
+======= */
+
+#ident \
+    "Copyright (c) 2006, 2015, Percona and/or its affiliates. All rights reserved."
+
+#pragma once
+
+#if defined(__clang__)
+#define constexpr_static_assert(a, b)
+#else
+#define constexpr_static_assert(a, b) static_assert(a, b)
+#endif
+
+// include here, before they get deprecated
+#include <inttypes.h>
+#include <stdint.h>
+#include <stdio.h>
+#include <sys/stat.h>
+#include <sys/time.h>
+#include <sys/types.h>
+#include <unistd.h>
+
+#include "toku_atomic.h"
+
+#if defined(__cplusplus)
+#include <type_traits>
+#endif
+
+#if defined(__cplusplus)
+// decltype() here gives a reference-to-pointer instead of just a pointer,
+// just use __typeof__
+#define CAST_FROM_VOIDP(name, value) name = static_cast<__typeof__(name)>(value)
+#else
+#define CAST_FROM_VOIDP(name, value) name = cast_to_typeof(name)(value)
+#endif
+
+#define UU(x) x __attribute__((__unused__))
+
+#include "toku_instrumentation.h"
diff --git a/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/portability/toku_pthread.h b/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/portability/toku_pthread.h
new file mode 100644
index 000000000..571b950e1
--- /dev/null
+++ b/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/portability/toku_pthread.h
@@ -0,0 +1,520 @@
+/* -*- mode: C++; c-basic-offset: 4; indent-tabs-mode: nil -*- */
+// vim: ft=cpp:expandtab:ts=8:sw=4:softtabstop=4:
+#ident "$Id$"
+/*======
+This file is part of PerconaFT.
+
+
+Copyright (c) 2006, 2015, Percona and/or its affiliates. All rights reserved.
+
+    PerconaFT is free software: you can redistribute it and/or modify
+    it under the terms of the GNU General Public License, version 2,
+    as published by the Free Software Foundation.
+
+    PerconaFT is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License
+    along with PerconaFT.  If not, see <http://www.gnu.org/licenses/>.
+
+----------------------------------------
+
+    PerconaFT is free software: you can redistribute it and/or modify
+    it under the terms of the GNU Affero General Public License, version 3,
+    as published by the Free Software Foundation.
+
+    PerconaFT is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU Affero General Public License for more details.
+
+    You should have received a copy of the GNU Affero General Public License
+    along with PerconaFT.  If not, see <http://www.gnu.org/licenses/>.
+
+----------------------------------------
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.
+======= */
+
+#ident \
+    "Copyright (c) 2006, 2015, Percona and/or its affiliates. All rights reserved."
+
+#pragma once
+
+#include <pthread.h>
+#include <stdint.h>
+#include <time.h>
+
+#include "toku_portability.h"
+// PORT2: #include "toku_assert.h"
+
+// TODO: some things moved toku_instrumentation.h, not necessarily the best
+// place
+typedef pthread_attr_t toku_pthread_attr_t;
+typedef pthread_t toku_pthread_t;
+typedef pthread_mutex_t toku_pthread_mutex_t;
+typedef pthread_condattr_t toku_pthread_condattr_t;
+typedef pthread_cond_t toku_pthread_cond_t;
+typedef pthread_rwlockattr_t toku_pthread_rwlockattr_t;
+typedef pthread_key_t toku_pthread_key_t;
+typedef struct timespec toku_timespec_t;
+
+// TODO: break this include loop
+#include <pthread.h>
+typedef pthread_mutexattr_t toku_pthread_mutexattr_t;
+
+struct toku_mutex_t {
+  pthread_mutex_t pmutex;
+  struct PSI_mutex *psi_mutex; /* The performance schema instrumentation hook */
+#if defined(TOKU_PTHREAD_DEBUG)
+  pthread_t owner;  // = pthread_self(); // for debugging
+  bool locked;
+  bool valid;
+  pfs_key_t instr_key_id;
+#endif  // defined(TOKU_PTHREAD_DEBUG)
+};
+
+struct toku_cond_t {
+  pthread_cond_t pcond;
+  struct PSI_cond *psi_cond;
+#if defined(TOKU_PTHREAD_DEBUG)
+  pfs_key_t instr_key_id;
+#endif  // defined(TOKU_PTHREAD_DEBUG)
+};
+
+#if defined(TOKU_PTHREAD_DEBUG)
+#define TOKU_COND_INITIALIZER \
+  { .pcond = PTHREAD_COND_INITIALIZER, .psi_cond = nullptr, .instr_key_id = 0 }
+#else
+#define TOKU_COND_INITIALIZER \
+  { .pcond = PTHREAD_COND_INITIALIZER, .psi_cond = nullptr }
+#endif  // defined(TOKU_PTHREAD_DEBUG)
+
+struct toku_pthread_rwlock_t {
+  pthread_rwlock_t rwlock;
+  struct PSI_rwlock *psi_rwlock;
+#if defined(TOKU_PTHREAD_DEBUG)
+  pfs_key_t instr_key_id;
+#endif  // defined(TOKU_PTHREAD_DEBUG)
+};
+
+typedef struct toku_mutex_aligned {
+  toku_mutex_t aligned_mutex __attribute__((__aligned__(64)));
+} toku_mutex_aligned_t;
+
+// Initializing with {} will fill in a struct with all zeros.
+// But you may also need a pragma to suppress the warnings, as follows
+//
+//   #pragma GCC diagnostic push
+//   #pragma GCC diagnostic ignored "-Wmissing-field-initializers"
+//   toku_mutex_t foo = ZERO_MUTEX_INITIALIZER;
+//   #pragma GCC diagnostic pop
+//
+// In general it will be a lot of busy work to make this codebase compile
+// cleanly with -Wmissing-field-initializers
+
+#define ZERO_MUTEX_INITIALIZER \
+  {}
+
+#if defined(TOKU_PTHREAD_DEBUG)
+#define TOKU_MUTEX_INITIALIZER                                             \
+  {                                                                        \
+    .pmutex = PTHREAD_MUTEX_INITIALIZER, .psi_mutex = nullptr, .owner = 0, \
+    .locked = false, .valid = true, .instr_key_id = 0                      \
+  }
+#else
+#define TOKU_MUTEX_INITIALIZER \
+  { .pmutex = PTHREAD_MUTEX_INITIALIZER, .psi_mutex = nullptr }
+#endif  // defined(TOKU_PTHREAD_DEBUG)
+
+// Darwin doesn't provide adaptive mutexes
+#if defined(__APPLE__)
+#define TOKU_MUTEX_ADAPTIVE PTHREAD_MUTEX_DEFAULT
+#if defined(TOKU_PTHREAD_DEBUG)
+#define TOKU_ADAPTIVE_MUTEX_INITIALIZER                                    \
+  {                                                                        \
+    .pmutex = PTHREAD_MUTEX_INITIALIZER, .psi_mutex = nullptr, .owner = 0, \
+    .locked = false, .valid = true, .instr_key_id = 0                      \
+  }
+#else
+#define TOKU_ADAPTIVE_MUTEX_INITIALIZER \
+  { .pmutex = PTHREAD_MUTEX_INITIALIZER, .psi_mutex = nullptr }
+#endif  // defined(TOKU_PTHREAD_DEBUG)
+#else   // __FreeBSD__, __linux__, at least
+#if defined(__GLIBC__)
+#define TOKU_MUTEX_ADAPTIVE PTHREAD_MUTEX_ADAPTIVE_NP
+#else
+// not all libc (e.g. musl) implement NP (Non-POSIX) attributes
+#define TOKU_MUTEX_ADAPTIVE PTHREAD_MUTEX_DEFAULT
+#endif
+#if defined(TOKU_PTHREAD_DEBUG)
+#define TOKU_ADAPTIVE_MUTEX_INITIALIZER                                    \
+  {                                                                        \
+    .pmutex = PTHREAD_ADAPTIVE_MUTEX_INITIALIZER_NP, .psi_mutex = nullptr, \
+    .owner = 0, .locked = false, .valid = true, .instr_key_id = 0          \
+  }
+#else
+#define TOKU_ADAPTIVE_MUTEX_INITIALIZER \
+  { .pmutex = PTHREAD_ADAPTIVE_MUTEX_INITIALIZER_NP, .psi_mutex = nullptr }
+#endif  // defined(TOKU_PTHREAD_DEBUG)
+#endif  // defined(__APPLE__)
+
+// Different OSes implement mutexes as different amounts of nested structs.
+// C++ will fill out all missing values with zeroes if you provide at least one
+// zero, but it needs the right amount of nesting.
+#if defined(__FreeBSD__)
+#define ZERO_COND_INITIALIZER \
+  { 0 }
+#elif defined(__APPLE__)
+#define ZERO_COND_INITIALIZER \
+  {                           \
+    { 0 }                     \
+  }
+#else  // __linux__, at least
+#define ZERO_COND_INITIALIZER \
+  {}
+#endif
+
+static inline void toku_mutexattr_init(toku_pthread_mutexattr_t *attr) {
+  int r = pthread_mutexattr_init(attr);
+  assert_zero(r);
+}
+
+static inline void toku_mutexattr_settype(toku_pthread_mutexattr_t *attr,
+                                          int type) {
+  int r = pthread_mutexattr_settype(attr, type);
+  assert_zero(r);
+}
+
+static inline void toku_mutexattr_destroy(toku_pthread_mutexattr_t *attr) {
+  int r = pthread_mutexattr_destroy(attr);
+  assert_zero(r);
+}
+
+#if defined(TOKU_PTHREAD_DEBUG)
+static inline void toku_mutex_assert_locked(const toku_mutex_t *mutex) {
+  invariant(mutex->locked);
+  invariant(mutex->owner == pthread_self());
+}
+#else
+static inline void toku_mutex_assert_locked(const toku_mutex_t *mutex
+                                            __attribute__((unused))) {}
+#endif  // defined(TOKU_PTHREAD_DEBUG)
+
+// asserting that a mutex is unlocked only makes sense
+// if the calling thread can guaruntee that no other threads
+// are trying to lock this mutex at the time of the assertion
+//
+// a good example of this is a tree with mutexes on each node.
+// when a node is locked the caller knows that no other threads
+// can be trying to lock its childrens' mutexes. the children
+// are in one of two fixed states: locked or unlocked.
+#if defined(TOKU_PTHREAD_DEBUG)
+static inline void toku_mutex_assert_unlocked(toku_mutex_t *mutex) {
+  invariant(mutex->owner == 0);
+  invariant(!mutex->locked);
+}
+#else
+static inline void toku_mutex_assert_unlocked(toku_mutex_t *mutex
+                                              __attribute__((unused))) {}
+#endif  // defined(TOKU_PTHREAD_DEBUG)
+
+#define toku_mutex_lock(M) \
+  toku_mutex_lock_with_source_location(M, __FILE__, __LINE__)
+
+static inline void toku_cond_init(toku_cond_t *cond,
+                                  const toku_pthread_condattr_t *attr) {
+  int r = pthread_cond_init(&cond->pcond, attr);
+  assert_zero(r);
+}
+
+#define toku_mutex_trylock(M) \
+  toku_mutex_trylock_with_source_location(M, __FILE__, __LINE__)
+
+inline void toku_mutex_unlock(toku_mutex_t *mutex) {
+#if defined(TOKU_PTHREAD_DEBUG)
+  invariant(mutex->owner == pthread_self());
+  invariant(mutex->valid);
+  invariant(mutex->locked);
+  mutex->locked = false;
+  mutex->owner = 0;
+#endif  // defined(TOKU_PTHREAD_DEBUG)
+  toku_instr_mutex_unlock(mutex->psi_mutex);
+  int r = pthread_mutex_unlock(&mutex->pmutex);
+  assert_zero(r);
+}
+
+inline void toku_mutex_lock_with_source_location(toku_mutex_t *mutex,
+                                                 const char *src_file,
+                                                 int src_line) {
+  toku_mutex_instrumentation mutex_instr;
+  toku_instr_mutex_lock_start(mutex_instr, *mutex, src_file, src_line);
+
+  const int r = pthread_mutex_lock(&mutex->pmutex);
+  toku_instr_mutex_lock_end(mutex_instr, r);
+
+  assert_zero(r);
+#if defined(TOKU_PTHREAD_DEBUG)
+  invariant(mutex->valid);
+  invariant(!mutex->locked);
+  invariant(mutex->owner == 0);
+  mutex->locked = true;
+  mutex->owner = pthread_self();
+#endif  // defined(TOKU_PTHREAD_DEBUG)
+}
+
+inline int toku_mutex_trylock_with_source_location(toku_mutex_t *mutex,
+                                                   const char *src_file,
+                                                   int src_line) {
+  toku_mutex_instrumentation mutex_instr;
+  toku_instr_mutex_trylock_start(mutex_instr, *mutex, src_file, src_line);
+
+  const int r = pthread_mutex_lock(&mutex->pmutex);
+  toku_instr_mutex_lock_end(mutex_instr, r);
+
+#if defined(TOKU_PTHREAD_DEBUG)
+  if (r == 0) {
+    invariant(mutex->valid);
+    invariant(!mutex->locked);
+    invariant(mutex->owner == 0);
+    mutex->locked = true;
+    mutex->owner = pthread_self();
+  }
+#endif  // defined(TOKU_PTHREAD_DEBUG)
+  return r;
+}
+
+#define toku_cond_wait(C, M) \
+  toku_cond_wait_with_source_location(C, M, __FILE__, __LINE__)
+
+#define toku_cond_timedwait(C, M, W) \
+  toku_cond_timedwait_with_source_location(C, M, W, __FILE__, __LINE__)
+
+inline void toku_cond_init(const toku_instr_key &key, toku_cond_t *cond,
+                           const pthread_condattr_t *attr) {
+  toku_instr_cond_init(key, *cond);
+  int r = pthread_cond_init(&cond->pcond, attr);
+  assert_zero(r);
+}
+
+inline void toku_cond_destroy(toku_cond_t *cond) {
+  toku_instr_cond_destroy(cond->psi_cond);
+  int r = pthread_cond_destroy(&cond->pcond);
+  assert_zero(r);
+}
+
+inline void toku_cond_wait_with_source_location(toku_cond_t *cond,
+                                                toku_mutex_t *mutex,
+                                                const char *src_file,
+                                                int src_line) {
+#if defined(TOKU_PTHREAD_DEBUG)
+  invariant(mutex->locked);
+  mutex->locked = false;
+  mutex->owner = 0;
+#endif  // defined(TOKU_PTHREAD_DEBUG)
+
+  /* Instrumentation start */
+  toku_cond_instrumentation cond_instr;
+  toku_instr_cond_wait_start(cond_instr, toku_instr_cond_op::cond_wait, *cond,
+                             *mutex, src_file, src_line);
+
+  /* Instrumented code */
+  const int r = pthread_cond_wait(&cond->pcond, &mutex->pmutex);
+
+  /* Instrumentation end */
+  toku_instr_cond_wait_end(cond_instr, r);
+
+  assert_zero(r);
+#if defined(TOKU_PTHREAD_DEBUG)
+  invariant(!mutex->locked);
+  mutex->locked = true;
+  mutex->owner = pthread_self();
+#endif  // defined(TOKU_PTHREAD_DEBUG)
+}
+
+inline int toku_cond_timedwait_with_source_location(toku_cond_t *cond,
+                                                    toku_mutex_t *mutex,
+                                                    toku_timespec_t *wakeup_at,
+                                                    const char *src_file,
+                                                    int src_line) {
+#if defined(TOKU_PTHREAD_DEBUG)
+  invariant(mutex->locked);
+  mutex->locked = false;
+  mutex->owner = 0;
+#endif  // defined(TOKU_PTHREAD_DEBUG)
+
+  /* Instrumentation start */
+  toku_cond_instrumentation cond_instr;
+  toku_instr_cond_wait_start(cond_instr, toku_instr_cond_op::cond_timedwait,
+                             *cond, *mutex, src_file, src_line);
+
+  /* Instrumented code */
+  const int r = pthread_cond_timedwait(&cond->pcond, &mutex->pmutex, wakeup_at);
+
+  /* Instrumentation end */
+  toku_instr_cond_wait_end(cond_instr, r);
+
+#if defined(TOKU_PTHREAD_DEBUG)
+  invariant(!mutex->locked);
+  mutex->locked = true;
+  mutex->owner = pthread_self();
+#endif  // defined(TOKU_PTHREAD_DEBUG)
+  return r;
+}
+
+inline void toku_cond_signal(toku_cond_t *cond) {
+  toku_instr_cond_signal(*cond);
+  const int r = pthread_cond_signal(&cond->pcond);
+  assert_zero(r);
+}
+
+inline void toku_cond_broadcast(toku_cond_t *cond) {
+  toku_instr_cond_broadcast(*cond);
+  const int r = pthread_cond_broadcast(&cond->pcond);
+  assert_zero(r);
+}
+
+inline void toku_mutex_init(const toku_instr_key &key, toku_mutex_t *mutex,
+                            const toku_pthread_mutexattr_t *attr) {
+#if defined(TOKU_PTHREAD_DEBUG)
+  mutex->valid = true;
+#endif  // defined(TOKU_PTHREAD_DEBUG)
+  toku_instr_mutex_init(key, *mutex);
+  const int r = pthread_mutex_init(&mutex->pmutex, attr);
+  assert_zero(r);
+#if defined(TOKU_PTHREAD_DEBUG)
+  mutex->locked = false;
+  invariant(mutex->valid);
+  mutex->valid = true;
+  mutex->owner = 0;
+#endif  // defined(TOKU_PTHREAD_DEBUG)
+}
+
+inline void toku_mutex_destroy(toku_mutex_t *mutex) {
+#if defined(TOKU_PTHREAD_DEBUG)
+  invariant(mutex->valid);
+  mutex->valid = false;
+  invariant(!mutex->locked);
+#endif  // defined(TOKU_PTHREAD_DEBUG)
+  toku_instr_mutex_destroy(mutex->psi_mutex);
+  int r = pthread_mutex_destroy(&mutex->pmutex);
+  assert_zero(r);
+}
+
+#define toku_pthread_rwlock_rdlock(RW) \
+  toku_pthread_rwlock_rdlock_with_source_location(RW, __FILE__, __LINE__)
+
+#define toku_pthread_rwlock_wrlock(RW) \
+  toku_pthread_rwlock_wrlock_with_source_location(RW, __FILE__, __LINE__)
+
+#if 0
+inline void toku_pthread_rwlock_init(
+    const toku_instr_key &key,
+    toku_pthread_rwlock_t *__restrict rwlock,
+    const toku_pthread_rwlockattr_t *__restrict attr) {
+    toku_instr_rwlock_init(key, *rwlock);
+    int r = pthread_rwlock_init(&rwlock->rwlock, attr);
+    assert_zero(r);
+}
+
+inline void toku_pthread_rwlock_destroy(toku_pthread_rwlock_t *rwlock) {
+    toku_instr_rwlock_destroy(rwlock->psi_rwlock);
+    int r = pthread_rwlock_destroy(&rwlock->rwlock);
+    assert_zero(r);
+}
+
+inline void toku_pthread_rwlock_rdlock_with_source_location(
+    toku_pthread_rwlock_t *rwlock,
+    const char *src_file,
+    uint src_line) {
+
+    /* Instrumentation start */
+    toku_rwlock_instrumentation rwlock_instr;
+    toku_instr_rwlock_rdlock_wait_start(
+        rwlock_instr, *rwlock, src_file, src_line);
+    /* Instrumented code */
+    const int r = pthread_rwlock_rdlock(&rwlock->rwlock);
+
+    /* Instrumentation end */
+    toku_instr_rwlock_rdlock_wait_end(rwlock_instr, r);
+
+    assert_zero(r);
+}
+
+inline void toku_pthread_rwlock_wrlock_with_source_location(
+    toku_pthread_rwlock_t *rwlock,
+    const char *src_file,
+    uint src_line) {
+
+    /* Instrumentation start */
+    toku_rwlock_instrumentation rwlock_instr;
+    toku_instr_rwlock_wrlock_wait_start(
+        rwlock_instr, *rwlock, src_file, src_line);
+    /* Instrumented code */
+    const int r = pthread_rwlock_wrlock(&rwlock->rwlock);
+
+    /* Instrumentation end */
+    toku_instr_rwlock_wrlock_wait_end(rwlock_instr, r);
+
+    assert_zero(r);
+}
+
+inline void toku_pthread_rwlock_rdunlock(toku_pthread_rwlock_t *rwlock) {
+    toku_instr_rwlock_unlock(*rwlock);
+    const int r = pthread_rwlock_unlock(&rwlock->rwlock);
+    assert_zero(r);
+}
+
+inline void toku_pthread_rwlock_wrunlock(toku_pthread_rwlock_t *rwlock) {
+    toku_instr_rwlock_unlock(*rwlock);
+    const int r = pthread_rwlock_unlock(&rwlock->rwlock);
+    assert_zero(r);
+}
+#endif
+
+static inline int toku_pthread_join(toku_pthread_t thread, void **value_ptr) {
+  return pthread_join(thread, value_ptr);
+}
+
+static inline int toku_pthread_detach(toku_pthread_t thread) {
+  return pthread_detach(thread);
+}
+
+static inline int toku_pthread_key_create(toku_pthread_key_t *key,
+                                          void (*destroyf)(void *)) {
+  return pthread_key_create(key, destroyf);
+}
+
+static inline int toku_pthread_key_delete(toku_pthread_key_t key) {
+  return pthread_key_delete(key);
+}
+
+static inline void *toku_pthread_getspecific(toku_pthread_key_t key) {
+  return pthread_getspecific(key);
+}
+
+static inline int toku_pthread_setspecific(toku_pthread_key_t key, void *data) {
+  return pthread_setspecific(key, data);
+}
+
+int toku_pthread_yield(void) __attribute__((__visibility__("default")));
+
+static inline toku_pthread_t toku_pthread_self(void) { return pthread_self(); }
+
+static inline void *toku_pthread_done(void *exit_value) {
+  toku_instr_delete_current_thread();
+  pthread_exit(exit_value);
+  return nullptr;  // Avoid compiler warning
+}
diff --git a/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/portability/toku_race_tools.h b/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/portability/toku_race_tools.h
new file mode 100644
index 000000000..3cb5b5790
--- /dev/null
+++ b/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/portability/toku_race_tools.h
@@ -0,0 +1,179 @@
+/* -*- mode: C++; c-basic-offset: 4; indent-tabs-mode: nil -*- */
+// vim: ft=cpp:expandtab:ts=8:sw=4:softtabstop=4:
+#ident "$Id$"
+/*======
+This file is part of PerconaFT.
+
+
+Copyright (c) 2006, 2015, Percona and/or its affiliates. All rights reserved.
+
+    PerconaFT is free software: you can redistribute it and/or modify
+    it under the terms of the GNU General Public License, version 2,
+    as published by the Free Software Foundation.
+
+    PerconaFT is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License
+    along with PerconaFT.  If not, see <http://www.gnu.org/licenses/>.
+
+----------------------------------------
+
+    PerconaFT is free software: you can redistribute it and/or modify
+    it under the terms of the GNU Affero General Public License, version 3,
+    as published by the Free Software Foundation.
+
+    PerconaFT is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU Affero General Public License for more details.
+
+    You should have received a copy of the GNU Affero General Public License
+    along with PerconaFT.  If not, see <http://www.gnu.org/licenses/>.
+
+----------------------------------------
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.
+======= */
+
+#ident \
+    "Copyright (c) 2006, 2015, Percona and/or its affiliates. All rights reserved."
+
+#pragma once
+
+// PORT2: #include <portability/toku_config.h>
+
+#ifdef HAVE_valgrind
+#undef USE_VALGRIND
+#define USE_VALGRIND 1
+#endif
+
+#if defined(__linux__) && USE_VALGRIND
+
+#include <valgrind/drd.h>
+#include <valgrind/helgrind.h>
+
+#define TOKU_ANNOTATE_NEW_MEMORY(p, size) ANNOTATE_NEW_MEMORY(p, size)
+#define TOKU_VALGRIND_HG_ENABLE_CHECKING(p, size) \
+  VALGRIND_HG_ENABLE_CHECKING(p, size)
+#define TOKU_VALGRIND_HG_DISABLE_CHECKING(p, size) \
+  VALGRIND_HG_DISABLE_CHECKING(p, size)
+#define TOKU_DRD_IGNORE_VAR(v) DRD_IGNORE_VAR(v)
+#define TOKU_DRD_STOP_IGNORING_VAR(v) DRD_STOP_IGNORING_VAR(v)
+#define TOKU_ANNOTATE_IGNORE_READS_BEGIN() ANNOTATE_IGNORE_READS_BEGIN()
+#define TOKU_ANNOTATE_IGNORE_READS_END() ANNOTATE_IGNORE_READS_END()
+#define TOKU_ANNOTATE_IGNORE_WRITES_BEGIN() ANNOTATE_IGNORE_WRITES_BEGIN()
+#define TOKU_ANNOTATE_IGNORE_WRITES_END() ANNOTATE_IGNORE_WRITES_END()
+
+/*
+ * How to make helgrind happy about tree rotations and new mutex orderings:
+ *
+ * // Tell helgrind that we unlocked it so that the next call doesn't get a
+ * "destroyed a locked mutex" error.
+ * // Tell helgrind that we destroyed the mutex.
+ * VALGRIND_HG_MUTEX_UNLOCK_PRE(&locka);
+ * VALGRIND_HG_MUTEX_DESTROY_PRE(&locka);
+ *
+ * // And recreate it.  It would be better to simply be able to say that the
+ * order on these two can now be reversed, because this code forgets all the
+ * ordering information for this mutex.
+ * // Then tell helgrind that we have locked it again.
+ * VALGRIND_HG_MUTEX_INIT_POST(&locka, 0);
+ * VALGRIND_HG_MUTEX_LOCK_POST(&locka);
+ *
+ * When the ordering of two locks changes, we don't need tell Helgrind about do
+ * both locks.  Just one is good enough.
+ */
+
+#define TOKU_VALGRIND_RESET_MUTEX_ORDERING_INFO(mutex) \
+  VALGRIND_HG_MUTEX_UNLOCK_PRE(mutex);                 \
+  VALGRIND_HG_MUTEX_DESTROY_PRE(mutex);                \
+  VALGRIND_HG_MUTEX_INIT_POST(mutex, 0);               \
+  VALGRIND_HG_MUTEX_LOCK_POST(mutex);
+
+#else  // !defined(__linux__) || !USE_VALGRIND
+
+#define NVALGRIND 1
+#define TOKU_ANNOTATE_NEW_MEMORY(p, size) ((void)0)
+#define TOKU_VALGRIND_HG_ENABLE_CHECKING(p, size) ((void)0)
+#define TOKU_VALGRIND_HG_DISABLE_CHECKING(p, size) ((void)0)
+#define TOKU_DRD_IGNORE_VAR(v)
+#define TOKU_DRD_STOP_IGNORING_VAR(v)
+#define TOKU_ANNOTATE_IGNORE_READS_BEGIN() ((void)0)
+#define TOKU_ANNOTATE_IGNORE_READS_END() ((void)0)
+#define TOKU_ANNOTATE_IGNORE_WRITES_BEGIN() ((void)0)
+#define TOKU_ANNOTATE_IGNORE_WRITES_END() ((void)0)
+#define TOKU_VALGRIND_RESET_MUTEX_ORDERING_INFO(mutex)
+#undef RUNNING_ON_VALGRIND
+#define RUNNING_ON_VALGRIND (0U)
+#endif
+
+// Valgrind 3.10.1 (and previous versions).
+// Problems with VALGRIND_HG_DISABLE_CHECKING and VALGRIND_HG_ENABLE_CHECKING.
+// Helgrind's implementation of disable and enable checking causes false races
+// to be reported.  In addition, the race report does not include ANY
+// information about the code that uses the helgrind disable and enable
+// functions.  Therefore, it is very difficult to figure out the cause of the
+// race. DRD does implement the disable and enable functions.
+
+// Problems with ANNOTATE_IGNORE_READS.
+// Helgrind does not implement ignore reads.
+// Annotate ignore reads is the way to inform DRD to ignore racy reads.
+
+// FT code uses unsafe reads in several places.  These unsafe reads have been
+// noted as valid since they use the toku_unsafe_fetch function. Unfortunately,
+// this causes helgrind to report erroneous data races which makes use of
+// helgrind problematic.
+
+// Unsafely fetch and return a `T' from src, telling drd to ignore
+// racey access to src for the next sizeof(*src) bytes
+template <typename T>
+T toku_unsafe_fetch(T *src) {
+  if (0)
+    TOKU_VALGRIND_HG_DISABLE_CHECKING(src,
+                                      sizeof *src);  // disabled, see comment
+  TOKU_ANNOTATE_IGNORE_READS_BEGIN();
+  T r = *src;
+  TOKU_ANNOTATE_IGNORE_READS_END();
+  if (0)
+    TOKU_VALGRIND_HG_ENABLE_CHECKING(src,
+                                     sizeof *src);  // disabled, see comment
+  return r;
+}
+
+template <typename T>
+T toku_unsafe_fetch(T &src) {
+  return toku_unsafe_fetch(&src);
+}
+
+// Unsafely set a `T' value into *dest from src, telling drd to ignore
+// racey access to dest for the next sizeof(*dest) bytes
+template <typename T>
+void toku_unsafe_set(T *dest, const T src) {
+  if (0)
+    TOKU_VALGRIND_HG_DISABLE_CHECKING(dest,
+                                      sizeof *dest);  // disabled, see comment
+  TOKU_ANNOTATE_IGNORE_WRITES_BEGIN();
+  *dest = src;
+  TOKU_ANNOTATE_IGNORE_WRITES_END();
+  if (0)
+    TOKU_VALGRIND_HG_ENABLE_CHECKING(dest,
+                                     sizeof *dest);  // disabled, see comment
+}
+
+template <typename T>
+void toku_unsafe_set(T &dest, const T src) {
+  toku_unsafe_set(&dest, src);
+}
diff --git a/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/portability/toku_time.h b/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/portability/toku_time.h
new file mode 100644
index 000000000..46111e7f0
--- /dev/null
+++ b/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/portability/toku_time.h
@@ -0,0 +1,193 @@
+/* -*- mode: C++; c-basic-offset: 4; indent-tabs-mode: nil -*- */
+// vim: ft=cpp:expandtab:ts=8:sw=4:softtabstop=4:
+#ident "$Id$"
+/*======
+This file is part of PerconaFT.
+
+
+Copyright (c) 2006, 2015, Percona and/or its affiliates. All rights reserved.
+
+    PerconaFT is free software: you can redistribute it and/or modify
+    it under the terms of the GNU General Public License, version 2,
+    as published by the Free Software Foundation.
+
+    PerconaFT is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License
+    along with PerconaFT.  If not, see <http://www.gnu.org/licenses/>.
+
+----------------------------------------
+
+    PerconaFT is free software: you can redistribute it and/or modify
+    it under the terms of the GNU Affero General Public License, version 3,
+    as published by the Free Software Foundation.
+
+    PerconaFT is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU Affero General Public License for more details.
+
+    You should have received a copy of the GNU Affero General Public License
+    along with PerconaFT.  If not, see <http://www.gnu.org/licenses/>.
+
+----------------------------------------
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.
+======= */
+
+#ident \
+    "Copyright (c) 2006, 2015, Percona and/or its affiliates. All rights reserved."
+
+#pragma once
+
+// PORT2: #include "toku_config.h"
+
+#include <stdint.h>
+#include <sys/time.h>
+#include <time.h>
+#if defined(__powerpc__)
+#include <sys/platform/ppc.h>
+#endif
+
+#if 0
+static inline float toku_tdiff (struct timeval *a, struct timeval *b) {
+    return (float)((a->tv_sec - b->tv_sec) + 1e-6 * (a->tv_usec - b->tv_usec));
+}
+// PORT2: temporary:
+#define HAVE_CLOCK_REALTIME
+#if !defined(HAVE_CLOCK_REALTIME)
+// OS X does not have clock_gettime, we fake clockid_t for the interface, and we'll implement it with clock_get_time.
+typedef int clockid_t;
+// just something bogus, it doesn't matter, we just want to make sure we're
+// only supporting this mode because we're not sure we can support other modes
+// without a real clock_gettime()
+#define CLOCK_REALTIME 0x01867234
+#endif
+int toku_clock_gettime(clockid_t clk_id, struct timespec *ts) __attribute__((__visibility__("default")));
+#endif
+
+// *************** Performance timers ************************
+// What do you really want from a performance timer:
+//  (1) Can determine actual time of day from the performance time.
+//  (2) Time goes forward, never backward.
+//  (3) Same time on different processors (or even different machines).
+//  (4) Time goes forward at a constant rate (doesn't get faster and slower)
+//  (5) Portable.
+//  (6) Getting the time is cheap.
+// Unfortuately it seems tough to get Properties 1-5.  So we go for Property 6,,
+// but we abstract it. We offer a type tokutime_t which can hold the time. This
+// type can be subtracted to get a time difference. We can get the present time
+// cheaply. We can convert this type to seconds (but that can be expensive). The
+// implementation is to use RDTSC (hence we lose property 3: not portable).
+// Recent machines have constant_tsc in which case we get property (4).
+// Recent OSs on recent machines (that have RDTSCP) fix the per-processor clock
+// skew, so we get property (3). We get property 2 with RDTSC (as long as
+// there's not any skew). We don't even try to get propety 1, since we don't
+// need it. The decision here is that these times are really accurate only on
+// modern machines with modern OSs.
+typedef uint64_t tokutime_t;  // Time type used in by tokutek timers.
+
+#if 0
+// The value of tokutime_t is not specified here. 
+// It might be microseconds since 1/1/1970 (if gettimeofday() is
+// used), or clock cycles since boot (if rdtsc is used).  Or something
+// else.
+// Two tokutime_t values can be subtracted to get a time difference.
+// Use tokutime_to_seconds to that convert difference  to seconds.
+// We want get_tokutime() to be fast, but don't care so much about tokutime_to_seconds();
+//
+// For accurate time calculations do the subtraction in the right order:
+//   Right:  tokutime_to_seconds(t1-t2);
+//   Wrong   tokutime_to_seconds(t1)-toku_time_to_seconds(t2);
+// Doing it the wrong way is likely to result in loss of precision.
+// A double can hold numbers up to about 53 bits.  RDTSC which uses about 33 bits every second, so that leaves
+// 2^20 seconds from booting (about 2 weeks) before the RDTSC value cannot be represented accurately as a double.
+//
+double tokutime_to_seconds(tokutime_t)  __attribute__((__visibility__("default"))); // Convert tokutime to seconds.
+
+#endif
+
+// Get the value of tokutime for right now.  We want this to be fast, so we
+// expose the implementation as RDTSC.
+static inline tokutime_t toku_time_now(void) {
+#if defined(__x86_64__) || defined(__i386__)
+  uint32_t lo, hi;
+  __asm__ __volatile__("rdtsc" : "=a"(lo), "=d"(hi));
+  return (uint64_t)hi << 32 | lo;
+#elif defined(__aarch64__)
+  uint64_t result;
+  __asm __volatile__("mrs %[rt], cntvct_el0" : [rt] "=r"(result));
+  return result;
+#elif defined(__powerpc__)
+  return __ppc_get_timebase();
+#elif defined(__s390x__)
+  uint64_t result;
+  asm volatile("stckf %0" : "=Q"(result) : : "cc");
+  return result;
+#elif defined(__riscv) && __riscv_xlen == 32
+  uint32_t cycles_lo, cycles_hi0, cycles_hi1;
+  // Implemented in assembly because Clang insisted on branching.
+  asm volatile(
+      "rdcycleh %0\n"
+      "rdcycle %1\n"
+      "rdcycleh %2\n"
+      "sub %0, %0, %2\n"
+      "seqz %0, %0\n"
+      "sub %0, zero, %0\n"
+      "and %1, %1, %0\n"
+      : "=r"(cycles_hi0), "=r"(cycles_lo), "=r"(cycles_hi1));
+  return (static_cast<uint64_t>(cycles_hi1) << 32) | cycles_lo;
+#elif defined(__riscv) && __riscv_xlen == 64
+  uint64_t cycles;
+  asm volatile("rdcycle %0" : "=r"(cycles));
+  return cycles;
+#else
+#error No timer implementation for this platform
+#endif
+}
+
+static inline uint64_t toku_current_time_microsec(void) {
+  struct timeval t;
+  gettimeofday(&t, NULL);
+  return t.tv_sec * (1UL * 1000 * 1000) + t.tv_usec;
+}
+
+#if 0
+// sleep microseconds
+static inline void toku_sleep_microsec(uint64_t ms) {
+    struct timeval  t;
+
+    t.tv_sec = ms / 1000000;
+    t.tv_usec = ms % 1000000;
+
+    select(0, NULL, NULL, NULL, &t);
+}
+#endif
+
+/*
+  PORT: Usage of this file:
+
+  uint64_t toku_current_time_microsec()   // uses gettimeoday
+      is used to track how much time various operations took (for example, lock
+      escalation). (TODO: it is not clear why these operations are tracked with
+      microsecond precision while others use nanoseconds)
+
+  tokutime_t toku_time_now() // uses rdtsc
+      seems to be used for a very similar purpose. This has greater precision
+
+  RocksDB environment provides Env::Default()->NowMicros() and NowNanos() which
+  should be adequate substitutes.
+*/
diff --git a/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/portability/txn_subst.h b/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/portability/txn_subst.h
new file mode 100644
index 000000000..803914862
--- /dev/null
+++ b/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/portability/txn_subst.h
@@ -0,0 +1,27 @@
+//
+// A substitute for ft/txn/txn.h
+//
+#pragma once
+
+#include <set>
+
+#include "../util/omt.h"
+
+typedef uint64_t TXNID;
+#define TXNID_NONE ((TXNID)0)
+
+// A set of transactions
+//  (TODO: consider using class toku::txnid_set. The reason for using STL
+//   container was that its API is easier)
+class TxnidVector : public std::set<TXNID> {
+ public:
+  bool contains(TXNID txnid) { return find(txnid) != end(); }
+};
+
+// A value for lock structures with a meaning "the lock is owned by multiple
+// transactions (and one has to check the TxnidVector to get their ids)
+#define TXNID_SHARED (TXNID(-1))
+
+// Auxiliary value meaning "any transaction id will do".  No real transaction
+// may have this is as id.
+#define TXNID_ANY (TXNID(-2))
diff --git a/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/standalone_port.cc b/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/standalone_port.cc
new file mode 100644
index 000000000..50dc879ce
--- /dev/null
+++ b/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/standalone_port.cc
@@ -0,0 +1,132 @@
+#ifndef ROCKSDB_LITE
+#ifndef OS_WIN
+/*
+  This is a dump ground to make Lock Tree work without the rest of TokuDB.
+*/
+#include <string.h>
+
+#include "db.h"
+#include "ft/ft-status.h"
+#include "portability/memory.h"
+#include "util/dbt.h"
+
+// portability/os_malloc.cc
+
+void toku_free(void *p) { free(p); }
+
+void *toku_xmalloc(size_t size) { return malloc(size); }
+
+void *toku_xrealloc(void *v, size_t size) { return realloc(v, size); }
+
+void *toku_xmemdup(const void *v, size_t len) {
+  void *p = toku_xmalloc(len);
+  memcpy(p, v, len);
+  return p;
+}
+
+// TODO: what are the X-functions? Xcalloc, Xrealloc?
+void *toku_xcalloc(size_t nmemb, size_t size) { return calloc(nmemb, size); }
+
+// ft-ft-opts.cc:
+
+// locktree
+toku_instr_key lock_request_m_wait_cond_key;
+toku_instr_key manager_m_escalator_done_key;
+toku_instr_key locktree_request_info_mutex_key;
+toku_instr_key locktree_request_info_retry_mutex_key;
+toku_instr_key locktree_request_info_retry_cv_key;
+
+toku_instr_key treenode_mutex_key;
+toku_instr_key manager_mutex_key;
+toku_instr_key manager_escalation_mutex_key;
+toku_instr_key manager_escalator_mutex_key;
+
+// portability/memory.cc
+size_t toku_memory_footprint(void *, size_t touched) { return touched; }
+
+// ft/ft-status.c
+// PORT2: note: the @c parameter to TOKUFT_STATUS_INIT must not start with
+//   "TOKU"
+LTM_STATUS_S ltm_status;
+void LTM_STATUS_S::init() {
+  if (m_initialized) return;
+#define LTM_STATUS_INIT(k, c, t, l)                    \
+  TOKUFT_STATUS_INIT((*this), k, c, t, "locktree: " l, \
+                     TOKU_ENGINE_STATUS | TOKU_GLOBAL_STATUS)
+  LTM_STATUS_INIT(LTM_SIZE_CURRENT, LOCKTREE_MEMORY_SIZE, STATUS_UINT64,
+                  "memory size");
+  LTM_STATUS_INIT(LTM_SIZE_LIMIT, LOCKTREE_MEMORY_SIZE_LIMIT, STATUS_UINT64,
+                  "memory size limit");
+  LTM_STATUS_INIT(LTM_ESCALATION_COUNT, LOCKTREE_ESCALATION_NUM, STATUS_UINT64,
+                  "number of times lock escalation ran");
+  LTM_STATUS_INIT(LTM_ESCALATION_TIME, LOCKTREE_ESCALATION_SECONDS,
+                  STATUS_TOKUTIME, "time spent running escalation (seconds)");
+  LTM_STATUS_INIT(LTM_ESCALATION_LATEST_RESULT,
+                  LOCKTREE_LATEST_POST_ESCALATION_MEMORY_SIZE, STATUS_UINT64,
+                  "latest post-escalation memory size");
+  LTM_STATUS_INIT(LTM_NUM_LOCKTREES, LOCKTREE_OPEN_CURRENT, STATUS_UINT64,
+                  "number of locktrees open now");
+  LTM_STATUS_INIT(LTM_LOCK_REQUESTS_PENDING, LOCKTREE_PENDING_LOCK_REQUESTS,
+                  STATUS_UINT64, "number of pending lock requests");
+  LTM_STATUS_INIT(LTM_STO_NUM_ELIGIBLE, LOCKTREE_STO_ELIGIBLE_NUM,
+                  STATUS_UINT64, "number of locktrees eligible for the STO");
+  LTM_STATUS_INIT(LTM_STO_END_EARLY_COUNT, LOCKTREE_STO_ENDED_NUM,
+                  STATUS_UINT64,
+                  "number of times a locktree ended the STO early");
+  LTM_STATUS_INIT(LTM_STO_END_EARLY_TIME, LOCKTREE_STO_ENDED_SECONDS,
+                  STATUS_TOKUTIME, "time spent ending the STO early (seconds)");
+  LTM_STATUS_INIT(LTM_WAIT_COUNT, LOCKTREE_WAIT_COUNT, STATUS_UINT64,
+                  "number of wait locks");
+  LTM_STATUS_INIT(LTM_WAIT_TIME, LOCKTREE_WAIT_TIME, STATUS_UINT64,
+                  "time waiting for locks");
+  LTM_STATUS_INIT(LTM_LONG_WAIT_COUNT, LOCKTREE_LONG_WAIT_COUNT, STATUS_UINT64,
+                  "number of long wait locks");
+  LTM_STATUS_INIT(LTM_LONG_WAIT_TIME, LOCKTREE_LONG_WAIT_TIME, STATUS_UINT64,
+                  "long time waiting for locks");
+  LTM_STATUS_INIT(LTM_TIMEOUT_COUNT, LOCKTREE_TIMEOUT_COUNT, STATUS_UINT64,
+                  "number of lock timeouts");
+  LTM_STATUS_INIT(LTM_WAIT_ESCALATION_COUNT, LOCKTREE_WAIT_ESCALATION_COUNT,
+                  STATUS_UINT64, "number of waits on lock escalation");
+  LTM_STATUS_INIT(LTM_WAIT_ESCALATION_TIME, LOCKTREE_WAIT_ESCALATION_TIME,
+                  STATUS_UINT64, "time waiting on lock escalation");
+  LTM_STATUS_INIT(LTM_LONG_WAIT_ESCALATION_COUNT,
+                  LOCKTREE_LONG_WAIT_ESCALATION_COUNT, STATUS_UINT64,
+                  "number of long waits on lock escalation");
+  LTM_STATUS_INIT(LTM_LONG_WAIT_ESCALATION_TIME,
+                  LOCKTREE_LONG_WAIT_ESCALATION_TIME, STATUS_UINT64,
+                  "long time waiting on lock escalation");
+
+  m_initialized = true;
+#undef LTM_STATUS_INIT
+}
+void LTM_STATUS_S::destroy() {
+  if (!m_initialized) return;
+  for (int i = 0; i < LTM_STATUS_NUM_ROWS; ++i) {
+    if (status[i].type == STATUS_PARCOUNT) {
+      // PORT: TODO?? destroy_partitioned_counter(status[i].value.parcount);
+    }
+  }
+}
+
+int toku_keycompare(const void *key1, size_t key1len, const void *key2,
+                    size_t key2len) {
+  size_t comparelen = key1len < key2len ? key1len : key2len;
+  int c = memcmp(key1, key2, comparelen);
+  if (__builtin_expect(c != 0, 1)) {
+    return c;
+  } else {
+    if (key1len < key2len) {
+      return -1;
+    } else if (key1len > key2len) {
+      return 1;
+    } else {
+      return 0;
+    }
+  }
+}
+
+int toku_builtin_compare_fun(const DBT *a, const DBT *b) {
+  return toku_keycompare(a->data, a->size, b->data, b->size);
+}
+#endif  // OS_WIN
+#endif  // ROCKSDB_LITE
diff --git a/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/util/dbt.cc b/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/util/dbt.cc
new file mode 100644
index 000000000..63cc3a267
--- /dev/null
+++ b/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/util/dbt.cc
@@ -0,0 +1,153 @@
+/* -*- mode: C++; c-basic-offset: 4; indent-tabs-mode: nil -*- */
+// vim: ft=cpp:expandtab:ts=8:sw=4:softtabstop=4:
+#ifndef ROCKSDB_LITE
+#ifndef OS_WIN
+#ident "$Id$"
+/*======
+This file is part of PerconaFT.
+
+
+Copyright (c) 2006, 2015, Percona and/or its affiliates. All rights reserved.
+
+    PerconaFT is free software: you can redistribute it and/or modify
+    it under the terms of the GNU General Public License, version 2,
+    as published by the Free Software Foundation.
+
+    PerconaFT is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License
+    along with PerconaFT.  If not, see <http://www.gnu.org/licenses/>.
+
+----------------------------------------
+
+    PerconaFT is free software: you can redistribute it and/or modify
+    it under the terms of the GNU Affero General Public License, version 3,
+    as published by the Free Software Foundation.
+
+    PerconaFT is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU Affero General Public License for more details.
+
+    You should have received a copy of the GNU Affero General Public License
+    along with PerconaFT.  If not, see <http://www.gnu.org/licenses/>.
+======= */
+
+#ident \
+    "Copyright (c) 2006, 2015, Percona and/or its affiliates. All rights reserved."
+
+#include "dbt.h"
+
+#include <string.h>
+
+#include "../db.h"
+#include "../portability/memory.h"
+
+DBT *toku_init_dbt(DBT *dbt) {
+  memset(dbt, 0, sizeof(*dbt));
+  return dbt;
+}
+
+DBT toku_empty_dbt(void) {
+  static const DBT empty_dbt = {.data = 0, .size = 0, .ulen = 0, .flags = 0};
+  return empty_dbt;
+}
+
+DBT *toku_init_dbt_flags(DBT *dbt, uint32_t flags) {
+  toku_init_dbt(dbt);
+  dbt->flags = flags;
+  return dbt;
+}
+
+void toku_destroy_dbt(DBT *dbt) {
+  switch (dbt->flags) {
+    case DB_DBT_MALLOC:
+    case DB_DBT_REALLOC:
+      toku_free(dbt->data);
+      toku_init_dbt(dbt);
+      break;
+  }
+}
+
+DBT *toku_fill_dbt(DBT *dbt, const void *k, size_t len) {
+  toku_init_dbt(dbt);
+  dbt->size = len;
+  dbt->data = (char *)k;
+  return dbt;
+}
+
+DBT *toku_memdup_dbt(DBT *dbt, const void *k, size_t len) {
+  toku_init_dbt_flags(dbt, DB_DBT_MALLOC);
+  dbt->size = len;
+  dbt->data = toku_xmemdup(k, len);
+  return dbt;
+}
+
+DBT *toku_copyref_dbt(DBT *dst, const DBT src) {
+  dst->flags = 0;
+  dst->ulen = 0;
+  dst->size = src.size;
+  dst->data = src.data;
+  return dst;
+}
+
+DBT *toku_clone_dbt(DBT *dst, const DBT &src) {
+  return toku_memdup_dbt(dst, src.data, src.size);
+}
+
+void toku_sdbt_cleanup(struct simple_dbt *sdbt) {
+  if (sdbt->data) toku_free(sdbt->data);
+  memset(sdbt, 0, sizeof(*sdbt));
+}
+
+const DBT *toku_dbt_positive_infinity(void) {
+  static DBT positive_infinity_dbt = {
+      .data = 0, .size = 0, .ulen = 0, .flags = 0};  // port
+  return &positive_infinity_dbt;
+}
+
+const DBT *toku_dbt_negative_infinity(void) {
+  static DBT negative_infinity_dbt = {
+      .data = 0, .size = 0, .ulen = 0, .flags = 0};  // port
+  return &negative_infinity_dbt;
+}
+
+bool toku_dbt_is_infinite(const DBT *dbt) {
+  return dbt == toku_dbt_positive_infinity() ||
+         dbt == toku_dbt_negative_infinity();
+}
+
+bool toku_dbt_is_empty(const DBT *dbt) {
+  // can't have a null data field with a non-zero size
+  paranoid_invariant(dbt->data != nullptr || dbt->size == 0);
+  return dbt->data == nullptr;
+}
+
+int toku_dbt_infinite_compare(const DBT *a, const DBT *b) {
+  if (a == b) {
+    return 0;
+  } else if (a == toku_dbt_positive_infinity()) {
+    return 1;
+  } else if (b == toku_dbt_positive_infinity()) {
+    return -1;
+  } else if (a == toku_dbt_negative_infinity()) {
+    return -1;
+  } else {
+    invariant(b == toku_dbt_negative_infinity());
+    return 1;
+  }
+}
+
+bool toku_dbt_equals(const DBT *a, const DBT *b) {
+  if (!toku_dbt_is_infinite(a) && !toku_dbt_is_infinite(b)) {
+    return a->data == b->data && a->size == b->size;
+  } else {
+    // a or b is infinite, so they're equal if they are the same infinite
+    return a == b ? true : false;
+  }
+}
+#endif  // OS_WIN
+#endif  // ROCKSDB_LITE
diff --git a/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/util/dbt.h b/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/util/dbt.h
new file mode 100644
index 000000000..d86c440f8
--- /dev/null
+++ b/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/util/dbt.h
@@ -0,0 +1,98 @@
+/* -*- mode: C++; c-basic-offset: 4; indent-tabs-mode: nil -*- */
+// vim: ft=cpp:expandtab:ts=8:sw=4:softtabstop=4:
+#ident "$Id$"
+/*======
+This file is part of PerconaFT.
+
+
+Copyright (c) 2006, 2015, Percona and/or its affiliates. All rights reserved.
+
+    PerconaFT is free software: you can redistribute it and/or modify
+    it under the terms of the GNU General Public License, version 2,
+    as published by the Free Software Foundation.
+
+    PerconaFT is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License
+    along with PerconaFT.  If not, see <http://www.gnu.org/licenses/>.
+
+----------------------------------------
+
+    PerconaFT is free software: you can redistribute it and/or modify
+    it under the terms of the GNU Affero General Public License, version 3,
+    as published by the Free Software Foundation.
+
+    PerconaFT is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU Affero General Public License for more details.
+
+    You should have received a copy of the GNU Affero General Public License
+    along with PerconaFT.  If not, see <http://www.gnu.org/licenses/>.
+
+----------------------------------------
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.
+======= */
+
+#ident \
+    "Copyright (c) 2006, 2015, Percona and/or its affiliates. All rights reserved."
+
+#pragma once
+
+#include "../db.h"
+
+// TODO: John
+// Document this API a little better so that DBT
+// memory management can be morm widely understood.
+
+DBT *toku_init_dbt(DBT *);
+
+// returns: an initialized but empty dbt (for which toku_dbt_is_empty() is true)
+DBT toku_empty_dbt(void);
+
+DBT *toku_init_dbt_flags(DBT *, uint32_t flags);
+
+void toku_destroy_dbt(DBT *);
+
+DBT *toku_fill_dbt(DBT *dbt, const void *k, size_t len);
+
+DBT *toku_memdup_dbt(DBT *dbt, const void *k, size_t len);
+
+DBT *toku_copyref_dbt(DBT *dst, const DBT src);
+
+DBT *toku_clone_dbt(DBT *dst, const DBT &src);
+
+void toku_sdbt_cleanup(struct simple_dbt *sdbt);
+
+// returns: special DBT pointer representing positive infinity
+const DBT *toku_dbt_positive_infinity(void);
+
+// returns: special DBT pointer representing negative infinity
+const DBT *toku_dbt_negative_infinity(void);
+
+// returns: true if the given dbt is either positive or negative infinity
+bool toku_dbt_is_infinite(const DBT *dbt);
+
+// returns: true if the given dbt has no data (ie: dbt->data == nullptr)
+bool toku_dbt_is_empty(const DBT *dbt);
+
+// effect: compares two potentially infinity-valued dbts
+// requires: at least one is infinite (assert otherwise)
+int toku_dbt_infinite_compare(const DBT *a, const DBT *b);
+
+// returns: true if the given dbts have the same data pointer and size
+bool toku_dbt_equals(const DBT *a, const DBT *b);
diff --git a/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/util/growable_array.h b/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/util/growable_array.h
new file mode 100644
index 000000000..158750fdb
--- /dev/null
+++ b/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/util/growable_array.h
@@ -0,0 +1,144 @@
+/* -*- mode: C++; c-basic-offset: 4; indent-tabs-mode: nil -*- */
+// vim: ft=cpp:expandtab:ts=8:sw=4:softtabstop=4:
+#ident "$Id$"
+/*======
+This file is part of PerconaFT.
+
+
+Copyright (c) 2006, 2015, Percona and/or its affiliates. All rights reserved.
+
+    PerconaFT is free software: you can redistribute it and/or modify
+    it under the terms of the GNU General Public License, version 2,
+    as published by the Free Software Foundation.
+
+    PerconaFT is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License
+    along with PerconaFT.  If not, see <http://www.gnu.org/licenses/>.
+
+----------------------------------------
+
+    PerconaFT is free software: you can redistribute it and/or modify
+    it under the terms of the GNU Affero General Public License, version 3,
+    as published by the Free Software Foundation.
+
+    PerconaFT is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU Affero General Public License for more details.
+
+    You should have received a copy of the GNU Affero General Public License
+    along with PerconaFT.  If not, see <http://www.gnu.org/licenses/>.
+
+----------------------------------------
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.
+======= */
+
+#ident \
+    "Copyright (c) 2006, 2015, Percona and/or its affiliates. All rights reserved."
+
+#pragma once
+
+#include <memory.h>
+
+//******************************************************************************
+//
+// Overview: A growable array is a little bit like std::vector except that
+//  it doesn't have constructors (hence can be used in static constructs, since
+//  the google style guide says no constructors), and it's a little simpler.
+// Operations:
+//   init and deinit (we don't have constructors and destructors).
+//   fetch_unchecked to get values out.
+//   store_unchecked to put values in.
+//   push to add an element at the end
+//   get_size to find out the size
+//   get_memory_size to find out how much memory the data stucture is using.
+//
+//******************************************************************************
+
+namespace toku {
+
+template <typename T>
+class GrowableArray {
+ public:
+  void init(void)
+  // Effect: Initialize the array to contain no elements.
+  {
+    m_array = NULL;
+    m_size = 0;
+    m_size_limit = 0;
+  }
+
+  void deinit(void)
+  // Effect: Deinitialize the array (freeing any memory it uses, for example).
+  {
+    toku_free(m_array);
+    m_array = NULL;
+    m_size = 0;
+    m_size_limit = 0;
+  }
+
+  T fetch_unchecked(size_t i) const
+  // Effect: Fetch the ith element.  If i is out of range, the system asserts.
+  {
+    return m_array[i];
+  }
+
+  void store_unchecked(size_t i, T v)
+  // Effect: Store v in the ith element.  If i is out of range, the system
+  // asserts.
+  {
+    paranoid_invariant(i < m_size);
+    m_array[i] = v;
+  }
+
+  void push(T v)
+  // Effect: Add v to the end of the array (increasing the size).  The amortized
+  // cost of this operation is constant. Implementation hint:  Double the size
+  // of the array when it gets too big so that the amortized cost stays
+  // constant.
+  {
+    if (m_size >= m_size_limit) {
+      if (m_array == NULL) {
+        m_size_limit = 1;
+      } else {
+        m_size_limit *= 2;
+      }
+      XREALLOC_N(m_size_limit, m_array);
+    }
+    m_array[m_size++] = v;
+  }
+
+  size_t get_size(void) const
+  // Effect: Return the number of elements in the array.
+  {
+    return m_size;
+  }
+  size_t memory_size(void) const
+  // Effect: Return the size (in bytes) that the array occupies in memory.  This
+  // is really only an estimate.
+  {
+    return sizeof(*this) + sizeof(T) * m_size_limit;
+  }
+
+ private:
+  T *m_array;
+  size_t m_size;
+  size_t m_size_limit;  // How much space is allocated in array.
+};
+
+}  // namespace toku
diff --git a/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/util/memarena.cc b/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/util/memarena.cc
new file mode 100644
index 000000000..0e7a9880b
--- /dev/null
+++ b/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/util/memarena.cc
@@ -0,0 +1,201 @@
+/* -*- mode: C++; c-basic-offset: 4; indent-tabs-mode: nil -*- */
+// vim: ft=cpp:expandtab:ts=8:sw=4:softtabstop=4:
+#ifndef ROCKSDB_LITE
+#ifndef OS_WIN
+#ident "$Id$"
+/*======
+This file is part of PerconaFT.
+
+
+Copyright (c) 2006, 2015, Percona and/or its affiliates. All rights reserved.
+
+    PerconaFT is free software: you can redistribute it and/or modify
+    it under the terms of the GNU General Public License, version 2,
+    as published by the Free Software Foundation.
+
+    PerconaFT is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License
+    along with PerconaFT.  If not, see <http://www.gnu.org/licenses/>.
+
+----------------------------------------
+
+    PerconaFT is free software: you can redistribute it and/or modify
+    it under the terms of the GNU Affero General Public License, version 3,
+    as published by the Free Software Foundation.
+
+    PerconaFT is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU Affero General Public License for more details.
+
+    You should have received a copy of the GNU Affero General Public License
+    along with PerconaFT.  If not, see <http://www.gnu.org/licenses/>.
+
+----------------------------------------
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.
+======= */
+
+#ident \
+    "Copyright (c) 2006, 2015, Percona and/or its affiliates. All rights reserved."
+
+#include "memarena.h"
+
+#include <string.h>
+
+#include <algorithm>
+
+#include "../portability/memory.h"
+
+void memarena::create(size_t initial_size) {
+  _current_chunk = arena_chunk();
+  _other_chunks = nullptr;
+  _size_of_other_chunks = 0;
+  _footprint_of_other_chunks = 0;
+  _n_other_chunks = 0;
+
+  _current_chunk.size = initial_size;
+  if (_current_chunk.size > 0) {
+    XMALLOC_N(_current_chunk.size, _current_chunk.buf);
+  }
+}
+
+void memarena::destroy(void) {
+  if (_current_chunk.buf) {
+    toku_free(_current_chunk.buf);
+  }
+  for (int i = 0; i < _n_other_chunks; i++) {
+    toku_free(_other_chunks[i].buf);
+  }
+  if (_other_chunks) {
+    toku_free(_other_chunks);
+  }
+  _current_chunk = arena_chunk();
+  _other_chunks = nullptr;
+  _n_other_chunks = 0;
+}
+
+static size_t round_to_page(size_t size) {
+  const size_t page_size = 4096;
+  const size_t r = page_size + ((size - 1) & ~(page_size - 1));
+  assert((r & (page_size - 1)) == 0);  // make sure it's aligned
+  assert(r >= size);                   // make sure it's not too small
+  assert(r <
+         size + page_size);  // make sure we didn't grow by more than a page.
+  return r;
+}
+
+static const size_t MEMARENA_MAX_CHUNK_SIZE = 64 * 1024 * 1024;
+
+void *memarena::malloc_from_arena(size_t size) {
+  if (_current_chunk.buf == nullptr ||
+      _current_chunk.size < _current_chunk.used + size) {
+    // The existing block isn't big enough.
+    // Add the block to the vector of blocks.
+    if (_current_chunk.buf) {
+      invariant(_current_chunk.size > 0);
+      int old_n = _n_other_chunks;
+      XREALLOC_N(old_n + 1, _other_chunks);
+      _other_chunks[old_n] = _current_chunk;
+      _n_other_chunks = old_n + 1;
+      _size_of_other_chunks += _current_chunk.size;
+      _footprint_of_other_chunks +=
+          toku_memory_footprint(_current_chunk.buf, _current_chunk.used);
+    }
+
+    // Make a new one. Grow the buffer size exponentially until we hit
+    // the max chunk size, but make it at least `size' bytes so the
+    // current allocation always fit.
+    size_t new_size =
+        std::min(MEMARENA_MAX_CHUNK_SIZE, 2 * _current_chunk.size);
+    if (new_size < size) {
+      new_size = size;
+    }
+    new_size = round_to_page(
+        new_size);  // at least size, but round to the next page size
+    XMALLOC_N(new_size, _current_chunk.buf);
+    _current_chunk.used = 0;
+    _current_chunk.size = new_size;
+  }
+  invariant(_current_chunk.buf != nullptr);
+
+  // allocate in the existing block.
+  char *p = _current_chunk.buf + _current_chunk.used;
+  _current_chunk.used += size;
+  return p;
+}
+
+void memarena::move_memory(memarena *dest) {
+  // Move memory to dest
+  XREALLOC_N(dest->_n_other_chunks + _n_other_chunks + 1, dest->_other_chunks);
+  dest->_size_of_other_chunks += _size_of_other_chunks + _current_chunk.size;
+  dest->_footprint_of_other_chunks +=
+      _footprint_of_other_chunks +
+      toku_memory_footprint(_current_chunk.buf, _current_chunk.used);
+  for (int i = 0; i < _n_other_chunks; i++) {
+    dest->_other_chunks[dest->_n_other_chunks++] = _other_chunks[i];
+  }
+  dest->_other_chunks[dest->_n_other_chunks++] = _current_chunk;
+
+  // Clear out this memarena's memory
+  toku_free(_other_chunks);
+  _current_chunk = arena_chunk();
+  _other_chunks = nullptr;
+  _size_of_other_chunks = 0;
+  _footprint_of_other_chunks = 0;
+  _n_other_chunks = 0;
+}
+
+size_t memarena::total_memory_size(void) const {
+  return sizeof(*this) + total_size_in_use() +
+         _n_other_chunks * sizeof(*_other_chunks);
+}
+
+size_t memarena::total_size_in_use(void) const {
+  return _size_of_other_chunks + _current_chunk.used;
+}
+
+size_t memarena::total_footprint(void) const {
+  return sizeof(*this) + _footprint_of_other_chunks +
+         toku_memory_footprint(_current_chunk.buf, _current_chunk.used) +
+         _n_other_chunks * sizeof(*_other_chunks);
+}
+
+////////////////////////////////////////////////////////////////////////////////
+
+const void *memarena::chunk_iterator::current(size_t *used) const {
+  if (_chunk_idx < 0) {
+    *used = _ma->_current_chunk.used;
+    return _ma->_current_chunk.buf;
+  } else if (_chunk_idx < _ma->_n_other_chunks) {
+    *used = _ma->_other_chunks[_chunk_idx].used;
+    return _ma->_other_chunks[_chunk_idx].buf;
+  }
+  *used = 0;
+  return nullptr;
+}
+
+void memarena::chunk_iterator::next() { _chunk_idx++; }
+
+bool memarena::chunk_iterator::more() const {
+  if (_chunk_idx < 0) {
+    return _ma->_current_chunk.buf != nullptr;
+  }
+  return _chunk_idx < _ma->_n_other_chunks;
+}
+#endif  // OS_WIN
+#endif  // ROCKSDB_LITE
diff --git a/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/util/memarena.h b/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/util/memarena.h
new file mode 100644
index 000000000..ddcc1144f
--- /dev/null
+++ b/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/util/memarena.h
@@ -0,0 +1,141 @@
+/* -*- mode: C++; c-basic-offset: 4; indent-tabs-mode: nil -*- */
+// vim: ft=cpp:expandtab:ts=8:sw=4:softtabstop=4:
+#ident "$Id$"
+/*======
+This file is part of PerconaFT.
+
+
+Copyright (c) 2006, 2015, Percona and/or its affiliates. All rights reserved.
+
+    PerconaFT is free software: you can redistribute it and/or modify
+    it under the terms of the GNU General Public License, version 2,
+    as published by the Free Software Foundation.
+
+    PerconaFT is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License
+    along with PerconaFT.  If not, see <http://www.gnu.org/licenses/>.
+
+----------------------------------------
+
+    PerconaFT is free software: you can redistribute it and/or modify
+    it under the terms of the GNU Affero General Public License, version 3,
+    as published by the Free Software Foundation.
+
+    PerconaFT is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU Affero General Public License for more details.
+
+    You should have received a copy of the GNU Affero General Public License
+    along with PerconaFT.  If not, see <http://www.gnu.org/licenses/>.
+
+----------------------------------------
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.
+======= */
+
+#ident \
+    "Copyright (c) 2006, 2015, Percona and/or its affiliates. All rights reserved."
+
+#pragma once
+
+#include <stdlib.h>
+
+/*
+ * A memarena is used to efficiently store a collection of objects that never
+ * move The pattern is allocate more and more stuff and free all of the items at
+ * once. The underlying memory will store 1 or more objects per chunk. Each
+ * chunk is contiguously laid out in memory but chunks are not necessarily
+ * contiguous with each other.
+ */
+class memarena {
+ public:
+  memarena()
+      : _current_chunk(arena_chunk()),
+        _other_chunks(nullptr),
+        _n_other_chunks(0),
+        _size_of_other_chunks(0),
+        _footprint_of_other_chunks(0) {}
+
+  // Effect: Create a memarena with the specified initial size
+  void create(size_t initial_size);
+
+  void destroy(void);
+
+  // Effect: Allocate some memory.  The returned value remains valid until the
+  // memarena is cleared or closed.
+  //  In case of ENOMEM, aborts.
+  void *malloc_from_arena(size_t size);
+
+  // Effect: Move all the memory from this memarena into DEST.
+  //         When SOURCE is closed the memory won't be freed.
+  //         When DEST is closed, the memory will be freed, unless DEST moves
+  //         its memory to another memarena...
+  void move_memory(memarena *dest);
+
+  // Effect: Calculate the amount of memory used by a memory arena.
+  size_t total_memory_size(void) const;
+
+  // Effect: Calculate the used space of the memory arena (ie: excludes unused
+  // space)
+  size_t total_size_in_use(void) const;
+
+  // Effect: Calculate the amount of memory used, according to
+  // toku_memory_footprint(),
+  //         which is a more expensive but more accurate count of memory used.
+  size_t total_footprint(void) const;
+
+  // iterator over the underlying chunks that store objects in the memarena.
+  // a chunk is represented by a pointer to const memory and a usable byte
+  // count.
+  class chunk_iterator {
+   public:
+    chunk_iterator(const memarena *ma) : _ma(ma), _chunk_idx(-1) {}
+
+    // returns: base pointer to the current chunk
+    //          *used set to the number of usable bytes
+    //          if more() is false, returns nullptr and *used = 0
+    const void *current(size_t *used) const;
+
+    // requires: more() is true
+    void next();
+
+    bool more() const;
+
+   private:
+    // -1 represents the 'initial' chunk in a memarena, ie: ma->_current_chunk
+    // >= 0 represents the i'th chunk in the ma->_other_chunks array
+    const memarena *_ma;
+    int _chunk_idx;
+  };
+
+ private:
+  struct arena_chunk {
+    arena_chunk() : buf(nullptr), used(0), size(0) {}
+    char *buf;
+    size_t used;
+    size_t size;
+  };
+
+  struct arena_chunk _current_chunk;
+  struct arena_chunk *_other_chunks;
+  int _n_other_chunks;
+  size_t _size_of_other_chunks;       // the buf_size of all the other chunks.
+  size_t _footprint_of_other_chunks;  // the footprint of all the other chunks.
+
+  friend class memarena_unit_test;
+};
diff --git a/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/util/omt.h b/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/util/omt.h
new file mode 100644
index 000000000..f208002d3
--- /dev/null
+++ b/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/util/omt.h
@@ -0,0 +1,794 @@
+/* -*- mode: C++; c-basic-offset: 4; indent-tabs-mode: nil -*- */
+// vim: ft=cpp:expandtab:ts=8:sw=4:softtabstop=4:
+#ident "$Id$"
+/*======
+This file is part of PerconaFT.
+
+
+Copyright (c) 2006, 2015, Percona and/or its affiliates. All rights reserved.
+
+    PerconaFT is free software: you can redistribute it and/or modify
+    it under the terms of the GNU General Public License, version 2,
+    as published by the Free Software Foundation.
+
+    PerconaFT is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License
+    along with PerconaFT.  If not, see <http://www.gnu.org/licenses/>.
+
+----------------------------------------
+
+    PerconaFT is free software: you can redistribute it and/or modify
+    it under the terms of the GNU Affero General Public License, version 3,
+    as published by the Free Software Foundation.
+
+    PerconaFT is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU Affero General Public License for more details.
+
+    You should have received a copy of the GNU Affero General Public License
+    along with PerconaFT.  If not, see <http://www.gnu.org/licenses/>.
+
+----------------------------------------
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.
+======= */
+
+#ident \
+    "Copyright (c) 2006, 2015, Percona and/or its affiliates. All rights reserved."
+
+#pragma once
+
+#include <memory.h>
+#include <stdint.h>
+
+#include "../portability/toku_portability.h"
+#include "../portability/toku_race_tools.h"
+#include "growable_array.h"
+
+namespace toku {
+
+/**
+ * Order Maintenance Tree (OMT)
+ *
+ * Maintains a collection of totally ordered values, where each value has an
+ * integer weight. The OMT is a mutable datatype.
+ *
+ * The Abstraction:
+ *
+ * An OMT is a vector of values, $V$, where $|V|$ is the length of the vector.
+ * The vector is numbered from $0$ to $|V|-1$.
+ * Each value has a weight.  The weight of the $i$th element is denoted
+ * $w(V_i)$.
+ *
+ * We can create a new OMT, which is the empty vector.
+ *
+ * We can insert a new element $x$ into slot $i$, changing $V$ into $V'$ where
+ *  $|V'|=1+|V|$       and
+ *
+ *   V'_j = V_j       if $j<i$
+ *          x         if $j=i$
+ *          V_{j-1}   if $j>i$.
+ *
+ * We can specify $i$ using a kind of function instead of as an integer.
+ * Let $b$ be a function mapping from values to nonzero integers, such that
+ * the signum of $b$ is monotically increasing.
+ * We can specify $i$ as the minimum integer such that $b(V_i)>0$.
+ *
+ * We look up a value using its index, or using a Heaviside function.
+ * For lookups, we allow $b$ to be zero for some values, and again the signum of
+ * $b$ must be monotonically increasing. When lookup up values, we can look up
+ *  $V_i$ where $i$ is the minimum integer such that $b(V_i)=0$.   (With a
+ * special return code if no such value exists.) (Rationale:  Ordinarily we want
+ * $i$ to be unique.  But for various reasons we want to allow multiple zeros,
+ * and we want the smallest $i$ in that case.) $V_i$ where $i$ is the minimum
+ * integer such that $b(V_i)>0$.   (Or an indication that no such value exists.)
+ *  $V_i$ where $i$ is the maximum integer such that $b(V_i)<0$.   (Or an
+ * indication that no such value exists.)
+ *
+ * When looking up a value using a Heaviside function, we get the value and its
+ * index.
+ *
+ * We can also split an OMT into two OMTs, splitting the weight of the values
+ * evenly. Find a value $j$ such that the values to the left of $j$ have about
+ * the same total weight as the values to the right of $j$. The resulting two
+ * OMTs contain the values to the left of $j$ and the values to the right of $j$
+ * respectively. All of the values from the original OMT go into one of the new
+ * OMTs. If the weights of the values don't split exactly evenly, then the
+ * implementation has the freedom to choose whether the new left OMT or the new
+ * right OMT is larger.
+ *
+ * Performance:
+ *  Insertion and deletion should run with $O(\log |V|)$ time and $O(\log |V|)$
+ * calls to the Heaviside function. The memory required is O(|V|).
+ *
+ * Usage:
+ *  The omt is templated by two parameters:
+ *   - omtdata_t is what will be stored within the omt.  These could be pointers
+ * or real data types (ints, structs).
+ *   - omtdataout_t is what will be returned by find and related functions.  By
+ * default, it is the same as omtdata_t, but you can set it to (omtdata_t *). To
+ * create an omt which will store "TXNID"s, for example, it is a good idea to
+ * typedef the template: typedef omt<TXNID> txnid_omt_t; If you are storing
+ * structs, you may want to be able to get a pointer to the data actually stored
+ * in the omt (see find_zero).  To do this, use the second template parameter:
+ *   typedef omt<struct foo, struct foo *> foo_omt_t;
+ */
+
+namespace omt_internal {
+
+template <bool subtree_supports_marks>
+class subtree_templated {
+ private:
+  uint32_t m_index;
+
+ public:
+  static const uint32_t NODE_NULL = UINT32_MAX;
+  inline void set_to_null(void) { m_index = NODE_NULL; }
+
+  inline bool is_null(void) const { return NODE_NULL == this->get_index(); }
+
+  inline uint32_t get_index(void) const { return m_index; }
+
+  inline void set_index(uint32_t index) {
+    paranoid_invariant(index != NODE_NULL);
+    m_index = index;
+  }
+} __attribute__((__packed__, aligned(4)));
+
+template <>
+class subtree_templated<true> {
+ private:
+  uint32_t m_bitfield;
+  static const uint32_t MASK_INDEX = ~(((uint32_t)1) << 31);
+  static const uint32_t MASK_BIT = ((uint32_t)1) << 31;
+
+  inline void set_index_internal(uint32_t new_index) {
+    m_bitfield = (m_bitfield & MASK_BIT) | new_index;
+  }
+
+ public:
+  static const uint32_t NODE_NULL = INT32_MAX;
+  inline void set_to_null(void) { this->set_index_internal(NODE_NULL); }
+
+  inline bool is_null(void) const { return NODE_NULL == this->get_index(); }
+
+  inline uint32_t get_index(void) const {
+    TOKU_DRD_IGNORE_VAR(m_bitfield);
+    const uint32_t bits = m_bitfield;
+    TOKU_DRD_STOP_IGNORING_VAR(m_bitfield);
+    return bits & MASK_INDEX;
+  }
+
+  inline void set_index(uint32_t index) {
+    paranoid_invariant(index < NODE_NULL);
+    this->set_index_internal(index);
+  }
+
+  inline bool get_bit(void) const {
+    TOKU_DRD_IGNORE_VAR(m_bitfield);
+    const uint32_t bits = m_bitfield;
+    TOKU_DRD_STOP_IGNORING_VAR(m_bitfield);
+    return (bits & MASK_BIT) != 0;
+  }
+
+  inline void enable_bit(void) {
+    // These bits may be set by a thread with a write lock on some
+    // leaf, and the index can be read by another thread with a (read
+    // or write) lock on another thread.  Also, the has_marks_below
+    // bit can be set by two threads simultaneously.  Neither of these
+    // are real races, so if we are using DRD we should tell it to
+    // ignore these bits just while we set this bit.  If there were a
+    // race in setting the index, that would be a real race.
+    TOKU_DRD_IGNORE_VAR(m_bitfield);
+    m_bitfield |= MASK_BIT;
+    TOKU_DRD_STOP_IGNORING_VAR(m_bitfield);
+  }
+
+  inline void disable_bit(void) { m_bitfield &= MASK_INDEX; }
+} __attribute__((__packed__));
+
+template <typename omtdata_t, bool subtree_supports_marks>
+class omt_node_templated {
+ public:
+  omtdata_t value;
+  uint32_t weight;
+  subtree_templated<subtree_supports_marks> left;
+  subtree_templated<subtree_supports_marks> right;
+
+  // this needs to be in both implementations because we don't have
+  // a "static if" the caller can use
+  inline void clear_stolen_bits(void) {}
+};  // note: originally this class had __attribute__((__packed__, aligned(4)))
+
+template <typename omtdata_t>
+class omt_node_templated<omtdata_t, true> {
+ public:
+  omtdata_t value;
+  uint32_t weight;
+  subtree_templated<true> left;
+  subtree_templated<true> right;
+  inline bool get_marked(void) const { return left.get_bit(); }
+  inline void set_marked_bit(void) { return left.enable_bit(); }
+  inline void unset_marked_bit(void) { return left.disable_bit(); }
+
+  inline bool get_marks_below(void) const { return right.get_bit(); }
+  inline void set_marks_below_bit(void) {
+    // This function can be called by multiple threads.
+    // Checking first reduces cache invalidation.
+    if (!this->get_marks_below()) {
+      right.enable_bit();
+    }
+  }
+  inline void unset_marks_below_bit(void) { right.disable_bit(); }
+
+  inline void clear_stolen_bits(void) {
+    this->unset_marked_bit();
+    this->unset_marks_below_bit();
+  }
+};  // note: originally this class had __attribute__((__packed__, aligned(4)))
+
+}  // namespace omt_internal
+
+template <typename omtdata_t, typename omtdataout_t = omtdata_t,
+          bool supports_marks = false>
+class omt {
+ public:
+  /**
+   * Effect: Create an empty OMT.
+   * Performance: constant time.
+   */
+  void create(void);
+
+  /**
+   * Effect: Create an empty OMT with no internal allocated space.
+   * Performance: constant time.
+   * Rationale: In some cases we need a valid omt but don't want to malloc.
+   */
+  void create_no_array(void);
+
+  /**
+   * Effect: Create a OMT containing values.  The number of values is in
+   * numvalues. Stores the new OMT in *omtp. Requires: this has not been created
+   * yet Requires: values != NULL Requires: values is sorted Performance:
+   * time=O(numvalues) Rationale:    Normally to insert N values takes O(N lg N)
+   * amortized time. If the N values are known in advance, are sorted, and the
+   * structure is empty, we can batch insert them much faster.
+   */
+  __attribute__((nonnull)) void create_from_sorted_array(
+      const omtdata_t *const values, const uint32_t numvalues);
+
+  /**
+   * Effect: Create an OMT containing values.  The number of values is in
+   * numvalues. On success the OMT takes ownership of *values array, and sets
+   * values=NULL. Requires: this has not been created yet Requires: values !=
+   * NULL Requires: *values is sorted Requires: *values was allocated with
+   * toku_malloc Requires: Capacity of the *values array is <= new_capacity
+   * Requires: On success, *values may not be accessed again by the caller.
+   * Performance:  time=O(1)
+   * Rational:     create_from_sorted_array takes O(numvalues) time.
+   *               By taking ownership of the array, we save a malloc and
+   * memcpy, and possibly a free (if the caller is done with the array).
+   */
+  void create_steal_sorted_array(omtdata_t **const values,
+                                 const uint32_t numvalues,
+                                 const uint32_t new_capacity);
+
+  /**
+   * Effect: Create a new OMT, storing it in *newomt.
+   *  The values to the right of index (starting at index) are moved to *newomt.
+   * Requires: newomt != NULL
+   * Returns
+   *    0             success,
+   *    EINVAL        if index > toku_omt_size(omt)
+   * On nonzero return, omt and *newomt are unmodified.
+   * Performance: time=O(n)
+   * Rationale:  We don't need a split-evenly operation.  We need to split items
+   * so that their total sizes are even, and other similar splitting criteria.
+   * It's easy to split evenly by calling size(), and dividing by two.
+   */
+  __attribute__((nonnull)) int split_at(omt *const newomt, const uint32_t idx);
+
+  /**
+   * Effect: Appends leftomt and rightomt to produce a new omt.
+   *  Creates this as the new omt.
+   *  leftomt and rightomt are destroyed.
+   * Performance: time=O(n) is acceptable, but one can imagine implementations
+   * that are O(\log n) worst-case.
+   */
+  __attribute__((nonnull)) void merge(omt *const leftomt, omt *const rightomt);
+
+  /**
+   * Effect: Creates a copy of an omt.
+   *  Creates this as the clone.
+   *  Each element is copied directly.  If they are pointers, the underlying
+   * data is not duplicated. Performance: O(n) or the running time of
+   * fill_array_with_subtree_values()
+   */
+  void clone(const omt &src);
+
+  /**
+   * Effect: Set the tree to be empty.
+   *  Note: Will not reallocate or resize any memory.
+   * Performance: time=O(1)
+   */
+  void clear(void);
+
+  /**
+   * Effect:  Destroy an OMT, freeing all its memory.
+   *   If the values being stored are pointers, their underlying data is not
+   * freed.  See free_items() Those values may be freed before or after calling
+   * toku_omt_destroy. Rationale: Returns no values since free() cannot fail.
+   * Rationale: Does not free the underlying pointers to reduce complexity.
+   * Performance:  time=O(1)
+   */
+  void destroy(void);
+
+  /**
+   * Effect: return |this|.
+   * Performance:  time=O(1)
+   */
+  uint32_t size(void) const;
+
+  /**
+   * Effect:  Insert value into the OMT.
+   *   If there is some i such that $h(V_i, v)=0$ then returns DB_KEYEXIST.
+   *   Otherwise, let i be the minimum value such that $h(V_i, v)>0$.
+   *      If no such i exists, then let i be |V|
+   *   Then this has the same effect as
+   *    insert_at(tree, value, i);
+   *   If idx!=NULL then i is stored in *idx
+   * Requires:  The signum of h must be monotonically increasing.
+   * Returns:
+   *    0            success
+   *    DB_KEYEXIST  the key is present (h was equal to zero for some value)
+   * On nonzero return, omt is unchanged.
+   * Performance: time=O(\log N) amortized.
+   * Rationale: Some future implementation may be O(\log N) worst-case time, but
+   * O(\log N) amortized is good enough for now.
+   */
+  template <typename omtcmp_t, int (*h)(const omtdata_t &, const omtcmp_t &)>
+  int insert(const omtdata_t &value, const omtcmp_t &v, uint32_t *const idx);
+
+  /**
+   * Effect: Increases indexes of all items at slot >= idx by 1.
+   *         Insert value into the position at idx.
+   * Returns:
+   *   0         success
+   *   EINVAL    if idx > this->size()
+   * On error, omt is unchanged.
+   * Performance: time=O(\log N) amortized time.
+   * Rationale: Some future implementation may be O(\log N) worst-case time, but
+   * O(\log N) amortized is good enough for now.
+   */
+  int insert_at(const omtdata_t &value, const uint32_t idx);
+
+  /**
+   * Effect:  Replaces the item at idx with value.
+   * Returns:
+   *   0       success
+   *   EINVAL  if idx>=this->size()
+   * On error, omt is unchanged.
+   * Performance: time=O(\log N)
+   * Rationale: The FT needs to be able to replace a value with another copy of
+   * the same value (allocated in a different location)
+   *
+   */
+  int set_at(const omtdata_t &value, const uint32_t idx);
+
+  /**
+   * Effect: Delete the item in slot idx.
+   *         Decreases indexes of all items at slot > idx by 1.
+   * Returns
+   *     0            success
+   *     EINVAL       if idx>=this->size()
+   * On error, omt is unchanged.
+   * Rationale: To delete an item, first find its index using find or find_zero,
+   * then delete it. Performance: time=O(\log N) amortized.
+   */
+  int delete_at(const uint32_t idx);
+
+  /**
+   * Effect:  Iterate over the values of the omt, from left to right, calling f
+   * on each value. The first argument passed to f is a ref-to-const of the
+   * value stored in the omt. The second argument passed to f is the index of
+   * the value. The third argument passed to f is iterate_extra. The indices run
+   * from 0 (inclusive) to this->size() (exclusive). Requires: f != NULL
+   * Returns:
+   *  If f ever returns nonzero, then the iteration stops, and the value
+   * returned by f is returned by iterate. If f always returns zero, then
+   * iterate returns 0. Requires:  Don't modify the omt while running.  (E.g., f
+   * may not insert or delete values from the omt.) Performance: time=O(i+\log
+   * N) where i is the number of times f is called, and N is the number of
+   * elements in the omt. Rationale: Although the functional iterator requires
+   * defining another function (as opposed to C++ style iterator), it is much
+   * easier to read. Rationale: We may at some point use functors, but for now
+   * this is a smaller change from the old OMT.
+   */
+  template <typename iterate_extra_t,
+            int (*f)(const omtdata_t &, const uint32_t, iterate_extra_t *const)>
+  int iterate(iterate_extra_t *const iterate_extra) const;
+
+  /**
+   * Effect:  Iterate over the values of the omt, from left to right, calling f
+   * on each value. The first argument passed to f is a ref-to-const of the
+   * value stored in the omt. The second argument passed to f is the index of
+   * the value. The third argument passed to f is iterate_extra. The indices run
+   * from 0 (inclusive) to this->size() (exclusive). We will iterate only over
+   * [left,right)
+   *
+   * Requires: left <= right
+   * Requires: f != NULL
+   * Returns:
+   *  EINVAL  if right > this->size()
+   *  If f ever returns nonzero, then the iteration stops, and the value
+   * returned by f is returned by iterate_on_range. If f always returns zero,
+   * then iterate_on_range returns 0. Requires:  Don't modify the omt while
+   * running.  (E.g., f may not insert or delete values from the omt.)
+   * Performance: time=O(i+\log N) where i is the number of times f is called,
+   * and N is the number of elements in the omt. Rational: Although the
+   * functional iterator requires defining another function (as opposed to C++
+   * style iterator), it is much easier to read.
+   */
+  template <typename iterate_extra_t,
+            int (*f)(const omtdata_t &, const uint32_t, iterate_extra_t *const)>
+  int iterate_on_range(const uint32_t left, const uint32_t right,
+                       iterate_extra_t *const iterate_extra) const;
+
+  /**
+   * Effect: Iterate over the values of the omt, and mark the nodes that are
+   * visited. Other than the marks, this behaves the same as iterate_on_range.
+   * Requires: supports_marks == true
+   * Performance: time=O(i+\log N) where i is the number of times f is called,
+   * and N is the number of elements in the omt. Notes: This function MAY be
+   * called concurrently by multiple threads, but not concurrently with any
+   * other non-const function.
+   */
+  template <typename iterate_extra_t,
+            int (*f)(const omtdata_t &, const uint32_t, iterate_extra_t *const)>
+  int iterate_and_mark_range(const uint32_t left, const uint32_t right,
+                             iterate_extra_t *const iterate_extra);
+
+  /**
+   * Effect: Iterate over the values of the omt, from left to right, calling f
+   * on each value whose node has been marked. Other than the marks, this
+   * behaves the same as iterate. Requires: supports_marks == true Performance:
+   * time=O(i+\log N) where i is the number of times f is called, and N is the
+   * number of elements in the omt.
+   */
+  template <typename iterate_extra_t,
+            int (*f)(const omtdata_t &, const uint32_t, iterate_extra_t *const)>
+  int iterate_over_marked(iterate_extra_t *const iterate_extra) const;
+
+  /**
+   * Effect: Delete all elements from the omt, whose nodes have been marked.
+   * Requires: supports_marks == true
+   * Performance: time=O(N + i\log N) where i is the number of marked elements,
+   * {c,sh}ould be faster
+   */
+  void delete_all_marked(void);
+
+  /**
+   * Effect: Verify that the internal state of the marks in the tree are
+   * self-consistent. Crashes the system if the marks are in a bad state.
+   * Requires: supports_marks == true
+   * Performance: time=O(N)
+   * Notes:
+   *  Even though this is a const function, it requires exclusive access.
+   * Rationale:
+   *  The current implementation of the marks relies on a sort of
+   *  "cache" bit representing the state of bits below it in the tree.
+   *  This allows glass-box testing that these bits are correct.
+   */
+  void verify_marks_consistent(void) const;
+
+  /**
+   * Effect: None
+   * Returns whether there are any marks in the tree.
+   */
+  bool has_marks(void) const;
+
+  /**
+   * Effect:  Iterate over the values of the omt, from left to right, calling f
+   * on each value. The first argument passed to f is a pointer to the value
+   * stored in the omt. The second argument passed to f is the index of the
+   * value. The third argument passed to f is iterate_extra. The indices run
+   * from 0 (inclusive) to this->size() (exclusive). Requires: same as for
+   * iterate() Returns: same as for iterate() Performance: same as for iterate()
+   * Rationale: In general, most iterators should use iterate() since they
+   * should not modify the data stored in the omt.  This function is for
+   * iterators which need to modify values (for example, free_items). Rationale:
+   * We assume if you are transforming the data in place, you want to do it to
+   * everything at once, so there is not yet an iterate_on_range_ptr (but there
+   * could be).
+   */
+  template <typename iterate_extra_t,
+            int (*f)(omtdata_t *, const uint32_t, iterate_extra_t *const)>
+  void iterate_ptr(iterate_extra_t *const iterate_extra);
+
+  /**
+   * Effect: Set *value=V_idx
+   * Returns
+   *    0             success
+   *    EINVAL        if index>=toku_omt_size(omt)
+   * On nonzero return, *value is unchanged
+   * Performance: time=O(\log N)
+   */
+  int fetch(const uint32_t idx, omtdataout_t *const value) const;
+
+  /**
+   * Effect:  Find the smallest i such that h(V_i, extra)>=0
+   *  If there is such an i and h(V_i,extra)==0 then set *idxp=i, set *value =
+   * V_i, and return 0. If there is such an i and h(V_i,extra)>0  then set
+   * *idxp=i and return DB_NOTFOUND. If there is no such i then set
+   * *idx=this->size() and return DB_NOTFOUND. Note: value is of type
+   * omtdataout_t, which may be of type (omtdata_t) or (omtdata_t *) but is
+   * fixed by the instantiation. If it is the value type, then the value is
+   * copied out (even if the value type is a pointer to something else) If it is
+   * the pointer type, then *value is set to a pointer to the data within the
+   * omt. This is determined by the type of the omt as initially declared. If
+   * the omt is declared as omt<foo_t>, then foo_t's will be stored and foo_t's
+   * will be returned by find and related functions. If the omt is declared as
+   * omt<foo_t, foo_t *>, then foo_t's will be stored, and pointers to the
+   * stored items will be returned by find and related functions. Rationale:
+   *  Structs too small for malloc should be stored directly in the omt.
+   *  These structs may need to be edited as they exist inside the omt, so we
+   * need a way to get a pointer within the omt. Using separate functions for
+   * returning pointers and values increases code duplication and reduces
+   * type-checking. That also reduces the ability of the creator of a data
+   * structure to give advice to its future users. Slight overloading in this
+   * case seemed to provide a better API and better type checking.
+   */
+  template <typename omtcmp_t, int (*h)(const omtdata_t &, const omtcmp_t &)>
+  int find_zero(const omtcmp_t &extra, omtdataout_t *const value,
+                uint32_t *const idxp) const;
+
+  /**
+   *   Effect:
+   *    If direction >0 then find the smallest i such that h(V_i,extra)>0.
+   *    If direction <0 then find the largest  i such that h(V_i,extra)<0.
+   *    (Direction may not be equal to zero.)
+   *    If value!=NULL then store V_i in *value
+   *    If idxp!=NULL then store i in *idxp.
+   *   Requires: The signum of h is monotically increasing.
+   *   Returns
+   *      0             success
+   *      DB_NOTFOUND   no such value is found.
+   *   On nonzero return, *value and *idxp are unchanged
+   *   Performance: time=O(\log N)
+   *   Rationale:
+   *     Here's how to use the find function to find various things
+   *       Cases for find:
+   *        find first value:         ( h(v)=+1, direction=+1 )
+   *        find last value           ( h(v)=-1, direction=-1 )
+   *        find first X              ( h(v)=(v< x) ? -1 : 1    direction=+1 )
+   *        find last X               ( h(v)=(v<=x) ? -1 : 1    direction=-1 )
+   *        find X or successor to X  ( same as find first X. )
+   *
+   *   Rationale: To help understand heaviside functions and behavor of find:
+   *    There are 7 kinds of heaviside functions.
+   *    The signus of the h must be monotonically increasing.
+   *    Given a function of the following form, A is the element
+   *    returned for direction>0, B is the element returned
+   *    for direction<0, C is the element returned for
+   *    direction==0 (see find_zero) (with a return of 0), and D is the element
+   *    returned for direction==0 (see find_zero) with a return of DB_NOTFOUND.
+   *    If any of A, B, or C are not found, then asking for the
+   *    associated direction will return DB_NOTFOUND.
+   *    See find_zero for more information.
+   *
+   *    Let the following represent the signus of the heaviside function.
+   *
+   *    -...-
+   *        A
+   *         D
+   *
+   *    +...+
+   *    B
+   *    D
+   *
+   *    0...0
+   *    C
+   *
+   *    -...-0...0
+   *        AC
+   *
+   *    0...0+...+
+   *    C    B
+   *
+   *    -...-+...+
+   *        AB
+   *         D
+   *
+   *    -...-0...0+...+
+   *        AC    B
+   */
+  template <typename omtcmp_t, int (*h)(const omtdata_t &, const omtcmp_t &)>
+  int find(const omtcmp_t &extra, int direction, omtdataout_t *const value,
+           uint32_t *const idxp) const;
+
+  /**
+   * Effect: Return the size (in bytes) of the omt, as it resides in main
+   * memory.  If the data stored are pointers, don't include the size of what
+   * they all point to.
+   */
+  size_t memory_size(void);
+
+ private:
+  typedef uint32_t node_idx;
+  typedef omt_internal::subtree_templated<supports_marks> subtree;
+  typedef omt_internal::omt_node_templated<omtdata_t, supports_marks> omt_node;
+  ENSURE_POD(subtree);
+
+  struct omt_array {
+    uint32_t start_idx;
+    uint32_t num_values;
+    omtdata_t *values;
+  };
+
+  struct omt_tree {
+    subtree root;
+    uint32_t free_idx;
+    omt_node *nodes;
+  };
+
+  bool is_array;
+  uint32_t capacity;
+  union {
+    struct omt_array a;
+    struct omt_tree t;
+  } d;
+
+  __attribute__((nonnull)) void unmark(const subtree &subtree,
+                                       const uint32_t index,
+                                       GrowableArray<node_idx> *const indexes);
+
+  void create_internal_no_array(const uint32_t new_capacity);
+
+  void create_internal(const uint32_t new_capacity);
+
+  uint32_t nweight(const subtree &subtree) const;
+
+  node_idx node_malloc(void);
+
+  void node_free(const node_idx idx);
+
+  void maybe_resize_array(const uint32_t n);
+
+  __attribute__((nonnull)) void fill_array_with_subtree_values(
+      omtdata_t *const array, const subtree &subtree) const;
+
+  void convert_to_array(void);
+
+  __attribute__((nonnull)) void rebuild_from_sorted_array(
+      subtree *const subtree, const omtdata_t *const values,
+      const uint32_t numvalues);
+
+  void convert_to_tree(void);
+
+  void maybe_resize_or_convert(const uint32_t n);
+
+  bool will_need_rebalance(const subtree &subtree, const int leftmod,
+                           const int rightmod) const;
+
+  __attribute__((nonnull)) void insert_internal(
+      subtree *const subtreep, const omtdata_t &value, const uint32_t idx,
+      subtree **const rebalance_subtree);
+
+  void set_at_internal_array(const omtdata_t &value, const uint32_t idx);
+
+  void set_at_internal(const subtree &subtree, const omtdata_t &value,
+                       const uint32_t idx);
+
+  void delete_internal(subtree *const subtreep, const uint32_t idx,
+                       omt_node *const copyn,
+                       subtree **const rebalance_subtree);
+
+  template <typename iterate_extra_t,
+            int (*f)(const omtdata_t &, const uint32_t, iterate_extra_t *const)>
+  int iterate_internal_array(const uint32_t left, const uint32_t right,
+                             iterate_extra_t *const iterate_extra) const;
+
+  template <typename iterate_extra_t,
+            int (*f)(omtdata_t *, const uint32_t, iterate_extra_t *const)>
+  void iterate_ptr_internal(const uint32_t left, const uint32_t right,
+                            const subtree &subtree, const uint32_t idx,
+                            iterate_extra_t *const iterate_extra);
+
+  template <typename iterate_extra_t,
+            int (*f)(omtdata_t *, const uint32_t, iterate_extra_t *const)>
+  void iterate_ptr_internal_array(const uint32_t left, const uint32_t right,
+                                  iterate_extra_t *const iterate_extra);
+
+  template <typename iterate_extra_t,
+            int (*f)(const omtdata_t &, const uint32_t, iterate_extra_t *const)>
+  int iterate_internal(const uint32_t left, const uint32_t right,
+                       const subtree &subtree, const uint32_t idx,
+                       iterate_extra_t *const iterate_extra) const;
+
+  template <typename iterate_extra_t,
+            int (*f)(const omtdata_t &, const uint32_t, iterate_extra_t *const)>
+  int iterate_and_mark_range_internal(const uint32_t left, const uint32_t right,
+                                      const subtree &subtree,
+                                      const uint32_t idx,
+                                      iterate_extra_t *const iterate_extra);
+
+  template <typename iterate_extra_t,
+            int (*f)(const omtdata_t &, const uint32_t, iterate_extra_t *const)>
+  int iterate_over_marked_internal(const subtree &subtree, const uint32_t idx,
+                                   iterate_extra_t *const iterate_extra) const;
+
+  uint32_t verify_marks_consistent_internal(const subtree &subtree,
+                                            const bool allow_marks) const;
+
+  void fetch_internal_array(const uint32_t i, omtdataout_t *const value) const;
+
+  void fetch_internal(const subtree &subtree, const uint32_t i,
+                      omtdataout_t *const value) const;
+
+  __attribute__((nonnull)) void fill_array_with_subtree_idxs(
+      node_idx *const array, const subtree &subtree) const;
+
+  __attribute__((nonnull)) void rebuild_subtree_from_idxs(
+      subtree *const subtree, const node_idx *const idxs,
+      const uint32_t numvalues);
+
+  __attribute__((nonnull)) void rebalance(subtree *const subtree);
+
+  __attribute__((nonnull)) static void copyout(omtdata_t *const out,
+                                               const omt_node *const n);
+
+  __attribute__((nonnull)) static void copyout(omtdata_t **const out,
+                                               omt_node *const n);
+
+  __attribute__((nonnull)) static void copyout(
+      omtdata_t *const out, const omtdata_t *const stored_value_ptr);
+
+  __attribute__((nonnull)) static void copyout(
+      omtdata_t **const out, omtdata_t *const stored_value_ptr);
+
+  template <typename omtcmp_t, int (*h)(const omtdata_t &, const omtcmp_t &)>
+  int find_internal_zero_array(const omtcmp_t &extra, omtdataout_t *const value,
+                               uint32_t *const idxp) const;
+
+  template <typename omtcmp_t, int (*h)(const omtdata_t &, const omtcmp_t &)>
+  int find_internal_zero(const subtree &subtree, const omtcmp_t &extra,
+                         omtdataout_t *const value, uint32_t *const idxp) const;
+
+  template <typename omtcmp_t, int (*h)(const omtdata_t &, const omtcmp_t &)>
+  int find_internal_plus_array(const omtcmp_t &extra, omtdataout_t *const value,
+                               uint32_t *const idxp) const;
+
+  template <typename omtcmp_t, int (*h)(const omtdata_t &, const omtcmp_t &)>
+  int find_internal_plus(const subtree &subtree, const omtcmp_t &extra,
+                         omtdataout_t *const value, uint32_t *const idxp) const;
+
+  template <typename omtcmp_t, int (*h)(const omtdata_t &, const omtcmp_t &)>
+  int find_internal_minus_array(const omtcmp_t &extra,
+                                omtdataout_t *const value,
+                                uint32_t *const idxp) const;
+
+  template <typename omtcmp_t, int (*h)(const omtdata_t &, const omtcmp_t &)>
+  int find_internal_minus(const subtree &subtree, const omtcmp_t &extra,
+                          omtdataout_t *const value,
+                          uint32_t *const idxp) const;
+};
+
+}  // namespace toku
+
+// include the implementation here
+#include "omt_impl.h"
diff --git a/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/util/omt_impl.h b/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/util/omt_impl.h
new file mode 100644
index 000000000..e77986716
--- /dev/null
+++ b/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/util/omt_impl.h
@@ -0,0 +1,1295 @@
+/* -*- mode: C++; c-basic-offset: 4; indent-tabs-mode: nil -*- */
+// vim: ft=cpp:expandtab:ts=8:sw=4:softtabstop=4:
+#ident "$Id$"
+/*======
+This file is part of PerconaFT.
+
+
+Copyright (c) 2006, 2015, Percona and/or its affiliates. All rights reserved.
+
+    PerconaFT is free software: you can redistribute it and/or modify
+    it under the terms of the GNU General Public License, version 2,
+    as published by the Free Software Foundation.
+
+    PerconaFT is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License
+    along with PerconaFT.  If not, see <http://www.gnu.org/licenses/>.
+
+----------------------------------------
+
+    PerconaFT is free software: you can redistribute it and/or modify
+    it under the terms of the GNU Affero General Public License, version 3,
+    as published by the Free Software Foundation.
+
+    PerconaFT is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU Affero General Public License for more details.
+
+    You should have received a copy of the GNU Affero General Public License
+    along with PerconaFT.  If not, see <http://www.gnu.org/licenses/>.
+
+----------------------------------------
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.
+======= */
+
+#ident \
+    "Copyright (c) 2006, 2015, Percona and/or its affiliates. All rights reserved."
+
+#include <string.h>
+
+#include "../db.h"
+#include "../portability/memory.h"
+
+namespace toku {
+
+template <typename omtdata_t, typename omtdataout_t, bool supports_marks>
+void omt<omtdata_t, omtdataout_t, supports_marks>::create(void) {
+  this->create_internal(2);
+  if (supports_marks) {
+    this->convert_to_tree();
+  }
+}
+
+template <typename omtdata_t, typename omtdataout_t, bool supports_marks>
+void omt<omtdata_t, omtdataout_t, supports_marks>::create_no_array(void) {
+  if (!supports_marks) {
+    this->create_internal_no_array(0);
+  } else {
+    this->is_array = false;
+    this->capacity = 0;
+    this->d.t.nodes = nullptr;
+    this->d.t.root.set_to_null();
+    this->d.t.free_idx = 0;
+  }
+}
+
+template <typename omtdata_t, typename omtdataout_t, bool supports_marks>
+void omt<omtdata_t, omtdataout_t, supports_marks>::create_from_sorted_array(
+    const omtdata_t *const values, const uint32_t numvalues) {
+  this->create_internal(numvalues);
+  memcpy(this->d.a.values, values, numvalues * (sizeof values[0]));
+  this->d.a.num_values = numvalues;
+  if (supports_marks) {
+    this->convert_to_tree();
+  }
+}
+
+template <typename omtdata_t, typename omtdataout_t, bool supports_marks>
+void omt<omtdata_t, omtdataout_t, supports_marks>::create_steal_sorted_array(
+    omtdata_t **const values, const uint32_t numvalues,
+    const uint32_t new_capacity) {
+  paranoid_invariant_notnull(values);
+  this->create_internal_no_array(new_capacity);
+  this->d.a.num_values = numvalues;
+  this->d.a.values = *values;
+  *values = nullptr;
+  if (supports_marks) {
+    this->convert_to_tree();
+  }
+}
+
+template <typename omtdata_t, typename omtdataout_t, bool supports_marks>
+int omt<omtdata_t, omtdataout_t, supports_marks>::split_at(omt *const newomt,
+                                                           const uint32_t idx) {
+  barf_if_marked(*this);
+  paranoid_invariant_notnull(newomt);
+  if (idx > this->size()) {
+    return EINVAL;
+  }
+  this->convert_to_array();
+  const uint32_t newsize = this->size() - idx;
+  newomt->create_from_sorted_array(&this->d.a.values[this->d.a.start_idx + idx],
+                                   newsize);
+  this->d.a.num_values = idx;
+  this->maybe_resize_array(idx);
+  if (supports_marks) {
+    this->convert_to_tree();
+  }
+  return 0;
+}
+
+template <typename omtdata_t, typename omtdataout_t, bool supports_marks>
+void omt<omtdata_t, omtdataout_t, supports_marks>::merge(omt *const leftomt,
+                                                         omt *const rightomt) {
+  barf_if_marked(*this);
+  paranoid_invariant_notnull(leftomt);
+  paranoid_invariant_notnull(rightomt);
+  const uint32_t leftsize = leftomt->size();
+  const uint32_t rightsize = rightomt->size();
+  const uint32_t newsize = leftsize + rightsize;
+
+  if (leftomt->is_array) {
+    if (leftomt->capacity -
+            (leftomt->d.a.start_idx + leftomt->d.a.num_values) >=
+        rightsize) {
+      this->create_steal_sorted_array(
+          &leftomt->d.a.values, leftomt->d.a.num_values, leftomt->capacity);
+      this->d.a.start_idx = leftomt->d.a.start_idx;
+    } else {
+      this->create_internal(newsize);
+      memcpy(&this->d.a.values[0], &leftomt->d.a.values[leftomt->d.a.start_idx],
+             leftomt->d.a.num_values * (sizeof this->d.a.values[0]));
+    }
+  } else {
+    this->create_internal(newsize);
+    leftomt->fill_array_with_subtree_values(&this->d.a.values[0],
+                                            leftomt->d.t.root);
+  }
+  leftomt->destroy();
+  this->d.a.num_values = leftsize;
+
+  if (rightomt->is_array) {
+    memcpy(&this->d.a.values[this->d.a.start_idx + this->d.a.num_values],
+           &rightomt->d.a.values[rightomt->d.a.start_idx],
+           rightomt->d.a.num_values * (sizeof this->d.a.values[0]));
+  } else {
+    rightomt->fill_array_with_subtree_values(
+        &this->d.a.values[this->d.a.start_idx + this->d.a.num_values],
+        rightomt->d.t.root);
+  }
+  rightomt->destroy();
+  this->d.a.num_values += rightsize;
+  paranoid_invariant(this->size() == newsize);
+  if (supports_marks) {
+    this->convert_to_tree();
+  }
+}
+
+template <typename omtdata_t, typename omtdataout_t, bool supports_marks>
+void omt<omtdata_t, omtdataout_t, supports_marks>::clone(const omt &src) {
+  barf_if_marked(*this);
+  this->create_internal(src.size());
+  if (src.is_array) {
+    memcpy(&this->d.a.values[0], &src.d.a.values[src.d.a.start_idx],
+           src.d.a.num_values * (sizeof this->d.a.values[0]));
+  } else {
+    src.fill_array_with_subtree_values(&this->d.a.values[0], src.d.t.root);
+  }
+  this->d.a.num_values = src.size();
+  if (supports_marks) {
+    this->convert_to_tree();
+  }
+}
+
+template <typename omtdata_t, typename omtdataout_t, bool supports_marks>
+void omt<omtdata_t, omtdataout_t, supports_marks>::clear(void) {
+  if (this->is_array) {
+    this->d.a.start_idx = 0;
+    this->d.a.num_values = 0;
+  } else {
+    this->d.t.root.set_to_null();
+    this->d.t.free_idx = 0;
+  }
+}
+
+template <typename omtdata_t, typename omtdataout_t, bool supports_marks>
+void omt<omtdata_t, omtdataout_t, supports_marks>::destroy(void) {
+  this->clear();
+  this->capacity = 0;
+  if (this->is_array) {
+    if (this->d.a.values != nullptr) {
+      toku_free(this->d.a.values);
+    }
+    this->d.a.values = nullptr;
+  } else {
+    if (this->d.t.nodes != nullptr) {
+      toku_free(this->d.t.nodes);
+    }
+    this->d.t.nodes = nullptr;
+  }
+}
+
+template <typename omtdata_t, typename omtdataout_t, bool supports_marks>
+uint32_t omt<omtdata_t, omtdataout_t, supports_marks>::size(void) const {
+  if (this->is_array) {
+    return this->d.a.num_values;
+  } else {
+    return this->nweight(this->d.t.root);
+  }
+}
+
+template <typename omtdata_t, typename omtdataout_t, bool supports_marks>
+template <typename omtcmp_t, int (*h)(const omtdata_t &, const omtcmp_t &)>
+int omt<omtdata_t, omtdataout_t, supports_marks>::insert(const omtdata_t &value,
+                                                         const omtcmp_t &v,
+                                                         uint32_t *const idx) {
+  int r;
+  uint32_t insert_idx;
+
+  r = this->find_zero<omtcmp_t, h>(v, nullptr, &insert_idx);
+  if (r == 0) {
+    if (idx) *idx = insert_idx;
+    return DB_KEYEXIST;
+  }
+  if (r != DB_NOTFOUND) return r;
+
+  if ((r = this->insert_at(value, insert_idx))) return r;
+  if (idx) *idx = insert_idx;
+
+  return 0;
+}
+
+// The following 3 functions implement a static if for us.
+template <typename omtdata_t, typename omtdataout_t>
+static void barf_if_marked(const omt<omtdata_t, omtdataout_t, false> &UU(omt)) {
+}
+
+template <typename omtdata_t, typename omtdataout_t>
+static void barf_if_marked(const omt<omtdata_t, omtdataout_t, true> &omt) {
+  invariant(!omt.has_marks());
+}
+
+template <typename omtdata_t, typename omtdataout_t, bool supports_marks>
+bool omt<omtdata_t, omtdataout_t, supports_marks>::has_marks(void) const {
+  static_assert(supports_marks, "Does not support marks");
+  if (this->d.t.root.is_null()) {
+    return false;
+  }
+  const omt_node &node = this->d.t.nodes[this->d.t.root.get_index()];
+  return node.get_marks_below() || node.get_marked();
+}
+
+template <typename omtdata_t, typename omtdataout_t, bool supports_marks>
+int omt<omtdata_t, omtdataout_t, supports_marks>::insert_at(
+    const omtdata_t &value, const uint32_t idx) {
+  barf_if_marked(*this);
+  if (idx > this->size()) {
+    return EINVAL;
+  }
+
+  this->maybe_resize_or_convert(this->size() + 1);
+  if (this->is_array && idx != this->d.a.num_values &&
+      (idx != 0 || this->d.a.start_idx == 0)) {
+    this->convert_to_tree();
+  }
+  if (this->is_array) {
+    if (idx == this->d.a.num_values) {
+      this->d.a.values[this->d.a.start_idx + this->d.a.num_values] = value;
+    } else {
+      this->d.a.values[--this->d.a.start_idx] = value;
+    }
+    this->d.a.num_values++;
+  } else {
+    subtree *rebalance_subtree = nullptr;
+    this->insert_internal(&this->d.t.root, value, idx, &rebalance_subtree);
+    if (rebalance_subtree != nullptr) {
+      this->rebalance(rebalance_subtree);
+    }
+  }
+  return 0;
+}
+
+template <typename omtdata_t, typename omtdataout_t, bool supports_marks>
+int omt<omtdata_t, omtdataout_t, supports_marks>::set_at(const omtdata_t &value,
+                                                         const uint32_t idx) {
+  barf_if_marked(*this);
+  if (idx >= this->size()) {
+    return EINVAL;
+  }
+
+  if (this->is_array) {
+    this->set_at_internal_array(value, idx);
+  } else {
+    this->set_at_internal(this->d.t.root, value, idx);
+  }
+  return 0;
+}
+
+template <typename omtdata_t, typename omtdataout_t, bool supports_marks>
+int omt<omtdata_t, omtdataout_t, supports_marks>::delete_at(
+    const uint32_t idx) {
+  barf_if_marked(*this);
+  if (idx >= this->size()) {
+    return EINVAL;
+  }
+
+  this->maybe_resize_or_convert(this->size() - 1);
+  if (this->is_array && idx != 0 && idx != this->d.a.num_values - 1) {
+    this->convert_to_tree();
+  }
+  if (this->is_array) {
+    // Testing for 0 does not rule out it being the last entry.
+    // Test explicitly for num_values-1
+    if (idx != this->d.a.num_values - 1) {
+      this->d.a.start_idx++;
+    }
+    this->d.a.num_values--;
+  } else {
+    subtree *rebalance_subtree = nullptr;
+    this->delete_internal(&this->d.t.root, idx, nullptr, &rebalance_subtree);
+    if (rebalance_subtree != nullptr) {
+      this->rebalance(rebalance_subtree);
+    }
+  }
+  return 0;
+}
+
+template <typename omtdata_t, typename omtdataout_t, bool supports_marks>
+template <typename iterate_extra_t,
+          int (*f)(const omtdata_t &, const uint32_t, iterate_extra_t *const)>
+int omt<omtdata_t, omtdataout_t, supports_marks>::iterate(
+    iterate_extra_t *const iterate_extra) const {
+  return this->iterate_on_range<iterate_extra_t, f>(0, this->size(),
+                                                    iterate_extra);
+}
+
+template <typename omtdata_t, typename omtdataout_t, bool supports_marks>
+template <typename iterate_extra_t,
+          int (*f)(const omtdata_t &, const uint32_t, iterate_extra_t *const)>
+int omt<omtdata_t, omtdataout_t, supports_marks>::iterate_on_range(
+    const uint32_t left, const uint32_t right,
+    iterate_extra_t *const iterate_extra) const {
+  if (right > this->size()) {
+    return EINVAL;
+  }
+  if (left == right) {
+    return 0;
+  }
+  if (this->is_array) {
+    return this->iterate_internal_array<iterate_extra_t, f>(left, right,
+                                                            iterate_extra);
+  }
+  return this->iterate_internal<iterate_extra_t, f>(left, right, this->d.t.root,
+                                                    0, iterate_extra);
+}
+
+template <typename omtdata_t, typename omtdataout_t, bool supports_marks>
+template <typename iterate_extra_t,
+          int (*f)(const omtdata_t &, const uint32_t, iterate_extra_t *const)>
+int omt<omtdata_t, omtdataout_t, supports_marks>::iterate_and_mark_range(
+    const uint32_t left, const uint32_t right,
+    iterate_extra_t *const iterate_extra) {
+  static_assert(supports_marks, "does not support marks");
+  if (right > this->size()) {
+    return EINVAL;
+  }
+  if (left == right) {
+    return 0;
+  }
+  paranoid_invariant(!this->is_array);
+  return this->iterate_and_mark_range_internal<iterate_extra_t, f>(
+      left, right, this->d.t.root, 0, iterate_extra);
+}
+
+// TODO: We can optimize this if we steal 3 bits.  1 bit: this node is
+// marked.  1 bit: left subtree has marks. 1 bit: right subtree has marks.
+template <typename omtdata_t, typename omtdataout_t, bool supports_marks>
+template <typename iterate_extra_t,
+          int (*f)(const omtdata_t &, const uint32_t, iterate_extra_t *const)>
+int omt<omtdata_t, omtdataout_t, supports_marks>::iterate_over_marked(
+    iterate_extra_t *const iterate_extra) const {
+  static_assert(supports_marks, "does not support marks");
+  paranoid_invariant(!this->is_array);
+  return this->iterate_over_marked_internal<iterate_extra_t, f>(
+      this->d.t.root, 0, iterate_extra);
+}
+
+template <typename omtdata_t, typename omtdataout_t, bool supports_marks>
+void omt<omtdata_t, omtdataout_t, supports_marks>::unmark(
+    const subtree &st, const uint32_t index,
+    GrowableArray<node_idx> *const indexes) {
+  if (st.is_null()) {
+    return;
+  }
+  omt_node &n = this->d.t.nodes[st.get_index()];
+  const uint32_t index_root = index + this->nweight(n.left);
+
+  const bool below = n.get_marks_below();
+  if (below) {
+    this->unmark(n.left, index, indexes);
+  }
+  if (n.get_marked()) {
+    indexes->push(index_root);
+  }
+  n.clear_stolen_bits();
+  if (below) {
+    this->unmark(n.right, index_root + 1, indexes);
+  }
+}
+
+template <typename omtdata_t, typename omtdataout_t, bool supports_marks>
+void omt<omtdata_t, omtdataout_t, supports_marks>::delete_all_marked(void) {
+  static_assert(supports_marks, "does not support marks");
+  if (!this->has_marks()) {
+    return;
+  }
+  paranoid_invariant(!this->is_array);
+  GrowableArray<node_idx> marked_indexes;
+  marked_indexes.init();
+
+  // Remove all marks.
+  // We need to delete all the stolen bits before calling delete_at to
+  // prevent barfing.
+  this->unmark(this->d.t.root, 0, &marked_indexes);
+
+  for (uint32_t i = 0; i < marked_indexes.get_size(); i++) {
+    // Delete from left to right, shift by number already deleted.
+    // Alternative is delete from right to left.
+    int r = this->delete_at(marked_indexes.fetch_unchecked(i) - i);
+    lazy_assert_zero(r);
+  }
+  marked_indexes.deinit();
+  barf_if_marked(*this);
+}
+
+template <typename omtdata_t, typename omtdataout_t, bool supports_marks>
+uint32_t
+omt<omtdata_t, omtdataout_t, supports_marks>::verify_marks_consistent_internal(
+    const subtree &st, const bool UU(allow_marks)) const {
+  if (st.is_null()) {
+    return 0;
+  }
+  const omt_node &node = this->d.t.nodes[st.get_index()];
+  uint32_t num_marks =
+      verify_marks_consistent_internal(node.left, node.get_marks_below());
+  num_marks +=
+      verify_marks_consistent_internal(node.right, node.get_marks_below());
+  if (node.get_marks_below()) {
+    paranoid_invariant(allow_marks);
+    paranoid_invariant(num_marks > 0);
+  } else {
+    // redundant with invariant below, but nice to have explicitly
+    paranoid_invariant(num_marks == 0);
+  }
+  if (node.get_marked()) {
+    paranoid_invariant(allow_marks);
+    ++num_marks;
+  }
+  return num_marks;
+}
+
+template <typename omtdata_t, typename omtdataout_t, bool supports_marks>
+void omt<omtdata_t, omtdataout_t, supports_marks>::verify_marks_consistent(
+    void) const {
+  static_assert(supports_marks, "does not support marks");
+  paranoid_invariant(!this->is_array);
+  this->verify_marks_consistent_internal(this->d.t.root, true);
+}
+
+template <typename omtdata_t, typename omtdataout_t, bool supports_marks>
+template <typename iterate_extra_t,
+          int (*f)(omtdata_t *, const uint32_t, iterate_extra_t *const)>
+void omt<omtdata_t, omtdataout_t, supports_marks>::iterate_ptr(
+    iterate_extra_t *const iterate_extra) {
+  if (this->is_array) {
+    this->iterate_ptr_internal_array<iterate_extra_t, f>(0, this->size(),
+                                                         iterate_extra);
+  } else {
+    this->iterate_ptr_internal<iterate_extra_t, f>(
+        0, this->size(), this->d.t.root, 0, iterate_extra);
+  }
+}
+
+template <typename omtdata_t, typename omtdataout_t, bool supports_marks>
+int omt<omtdata_t, omtdataout_t, supports_marks>::fetch(
+    const uint32_t idx, omtdataout_t *const value) const {
+  if (idx >= this->size()) {
+    return EINVAL;
+  }
+  if (this->is_array) {
+    this->fetch_internal_array(idx, value);
+  } else {
+    this->fetch_internal(this->d.t.root, idx, value);
+  }
+  return 0;
+}
+
+template <typename omtdata_t, typename omtdataout_t, bool supports_marks>
+template <typename omtcmp_t, int (*h)(const omtdata_t &, const omtcmp_t &)>
+int omt<omtdata_t, omtdataout_t, supports_marks>::find_zero(
+    const omtcmp_t &extra, omtdataout_t *const value,
+    uint32_t *const idxp) const {
+  uint32_t tmp_index;
+  uint32_t *const child_idxp = (idxp != nullptr) ? idxp : &tmp_index;
+  int r;
+  if (this->is_array) {
+    r = this->find_internal_zero_array<omtcmp_t, h>(extra, value, child_idxp);
+  } else {
+    r = this->find_internal_zero<omtcmp_t, h>(this->d.t.root, extra, value,
+                                              child_idxp);
+  }
+  return r;
+}
+
+template <typename omtdata_t, typename omtdataout_t, bool supports_marks>
+template <typename omtcmp_t, int (*h)(const omtdata_t &, const omtcmp_t &)>
+int omt<omtdata_t, omtdataout_t, supports_marks>::find(
+    const omtcmp_t &extra, int direction, omtdataout_t *const value,
+    uint32_t *const idxp) const {
+  uint32_t tmp_index;
+  uint32_t *const child_idxp = (idxp != nullptr) ? idxp : &tmp_index;
+  paranoid_invariant(direction != 0);
+  if (direction < 0) {
+    if (this->is_array) {
+      return this->find_internal_minus_array<omtcmp_t, h>(extra, value,
+                                                          child_idxp);
+    } else {
+      return this->find_internal_minus<omtcmp_t, h>(this->d.t.root, extra,
+                                                    value, child_idxp);
+    }
+  } else {
+    if (this->is_array) {
+      return this->find_internal_plus_array<omtcmp_t, h>(extra, value,
+                                                         child_idxp);
+    } else {
+      return this->find_internal_plus<omtcmp_t, h>(this->d.t.root, extra, value,
+                                                   child_idxp);
+    }
+  }
+}
+
+template <typename omtdata_t, typename omtdataout_t, bool supports_marks>
+size_t omt<omtdata_t, omtdataout_t, supports_marks>::memory_size(void) {
+  if (this->is_array) {
+    return (sizeof *this) + this->capacity * (sizeof this->d.a.values[0]);
+  }
+  return (sizeof *this) + this->capacity * (sizeof this->d.t.nodes[0]);
+}
+
+template <typename omtdata_t, typename omtdataout_t, bool supports_marks>
+void omt<omtdata_t, omtdataout_t, supports_marks>::create_internal_no_array(
+    const uint32_t new_capacity) {
+  this->is_array = true;
+  this->d.a.start_idx = 0;
+  this->d.a.num_values = 0;
+  this->d.a.values = nullptr;
+  this->capacity = new_capacity;
+}
+
+template <typename omtdata_t, typename omtdataout_t, bool supports_marks>
+void omt<omtdata_t, omtdataout_t, supports_marks>::create_internal(
+    const uint32_t new_capacity) {
+  this->create_internal_no_array(new_capacity);
+  XMALLOC_N(this->capacity, this->d.a.values);
+}
+
+template <typename omtdata_t, typename omtdataout_t, bool supports_marks>
+uint32_t omt<omtdata_t, omtdataout_t, supports_marks>::nweight(
+    const subtree &st) const {
+  if (st.is_null()) {
+    return 0;
+  } else {
+    return this->d.t.nodes[st.get_index()].weight;
+  }
+}
+
+template <typename omtdata_t, typename omtdataout_t, bool supports_marks>
+typename omt<omtdata_t, omtdataout_t, supports_marks>::node_idx
+omt<omtdata_t, omtdataout_t, supports_marks>::node_malloc(void) {
+  paranoid_invariant(this->d.t.free_idx < this->capacity);
+  omt_node &n = this->d.t.nodes[this->d.t.free_idx];
+  n.clear_stolen_bits();
+  return this->d.t.free_idx++;
+}
+
+template <typename omtdata_t, typename omtdataout_t, bool supports_marks>
+void omt<omtdata_t, omtdataout_t, supports_marks>::node_free(
+    const node_idx UU(idx)) {
+  paranoid_invariant(idx < this->capacity);
+}
+
+template <typename omtdata_t, typename omtdataout_t, bool supports_marks>
+void omt<omtdata_t, omtdataout_t, supports_marks>::maybe_resize_array(
+    const uint32_t n) {
+  const uint32_t new_size = n <= 2 ? 4 : 2 * n;
+  const uint32_t room = this->capacity - this->d.a.start_idx;
+
+  if (room < n || this->capacity / 2 >= new_size) {
+    omtdata_t *XMALLOC_N(new_size, tmp_values);
+    if (this->d.a.num_values) {
+      memcpy(tmp_values, &this->d.a.values[this->d.a.start_idx],
+             this->d.a.num_values * (sizeof tmp_values[0]));
+    }
+    this->d.a.start_idx = 0;
+    this->capacity = new_size;
+    toku_free(this->d.a.values);
+    this->d.a.values = tmp_values;
+  }
+}
+
+template <typename omtdata_t, typename omtdataout_t, bool supports_marks>
+void omt<omtdata_t, omtdataout_t,
+         supports_marks>::fill_array_with_subtree_values(omtdata_t *const array,
+                                                         const subtree &st)
+    const {
+  if (st.is_null()) return;
+  const omt_node &tree = this->d.t.nodes[st.get_index()];
+  this->fill_array_with_subtree_values(&array[0], tree.left);
+  array[this->nweight(tree.left)] = tree.value;
+  this->fill_array_with_subtree_values(&array[this->nweight(tree.left) + 1],
+                                       tree.right);
+}
+
+template <typename omtdata_t, typename omtdataout_t, bool supports_marks>
+void omt<omtdata_t, omtdataout_t, supports_marks>::convert_to_array(void) {
+  if (!this->is_array) {
+    const uint32_t num_values = this->size();
+    uint32_t new_size = 2 * num_values;
+    new_size = new_size < 4 ? 4 : new_size;
+
+    omtdata_t *XMALLOC_N(new_size, tmp_values);
+    this->fill_array_with_subtree_values(tmp_values, this->d.t.root);
+    toku_free(this->d.t.nodes);
+    this->is_array = true;
+    this->capacity = new_size;
+    this->d.a.num_values = num_values;
+    this->d.a.values = tmp_values;
+    this->d.a.start_idx = 0;
+  }
+}
+
+template <typename omtdata_t, typename omtdataout_t, bool supports_marks>
+void omt<omtdata_t, omtdataout_t, supports_marks>::rebuild_from_sorted_array(
+    subtree *const st, const omtdata_t *const values,
+    const uint32_t numvalues) {
+  if (numvalues == 0) {
+    st->set_to_null();
+  } else {
+    const uint32_t halfway = numvalues / 2;
+    const node_idx newidx = this->node_malloc();
+    omt_node *const newnode = &this->d.t.nodes[newidx];
+    newnode->weight = numvalues;
+    newnode->value = values[halfway];
+    st->set_index(newidx);
+    // update everything before the recursive calls so the second call
+    // can be a tail call.
+    this->rebuild_from_sorted_array(&newnode->left, &values[0], halfway);
+    this->rebuild_from_sorted_array(&newnode->right, &values[halfway + 1],
+                                    numvalues - (halfway + 1));
+  }
+}
+
+template <typename omtdata_t, typename omtdataout_t, bool supports_marks>
+void omt<omtdata_t, omtdataout_t, supports_marks>::convert_to_tree(void) {
+  if (this->is_array) {
+    const uint32_t num_nodes = this->size();
+    uint32_t new_size = num_nodes * 2;
+    new_size = new_size < 4 ? 4 : new_size;
+
+    omt_node *XMALLOC_N(new_size, new_nodes);
+    omtdata_t *const values = this->d.a.values;
+    omtdata_t *const tmp_values = &values[this->d.a.start_idx];
+    this->is_array = false;
+    this->d.t.nodes = new_nodes;
+    this->capacity = new_size;
+    this->d.t.free_idx = 0;
+    this->d.t.root.set_to_null();
+    this->rebuild_from_sorted_array(&this->d.t.root, tmp_values, num_nodes);
+    toku_free(values);
+  }
+}
+
+template <typename omtdata_t, typename omtdataout_t, bool supports_marks>
+void omt<omtdata_t, omtdataout_t, supports_marks>::maybe_resize_or_convert(
+    const uint32_t n) {
+  if (this->is_array) {
+    this->maybe_resize_array(n);
+  } else {
+    const uint32_t new_size = n <= 2 ? 4 : 2 * n;
+    const uint32_t num_nodes = this->nweight(this->d.t.root);
+    if ((this->capacity / 2 >= new_size) ||
+        (this->d.t.free_idx >= this->capacity && num_nodes < n) ||
+        (this->capacity < n)) {
+      this->convert_to_array();
+      // if we had a free list, the "supports_marks" version could
+      // just resize, as it is now, we have to convert to and back
+      // from an array.
+      if (supports_marks) {
+        this->convert_to_tree();
+      }
+    }
+  }
+}
+
+template <typename omtdata_t, typename omtdataout_t, bool supports_marks>
+bool omt<omtdata_t, omtdataout_t, supports_marks>::will_need_rebalance(
+    const subtree &st, const int leftmod, const int rightmod) const {
+  if (st.is_null()) {
+    return false;
+  }
+  const omt_node &n = this->d.t.nodes[st.get_index()];
+  // one of the 1's is for the root.
+  // the other is to take ceil(n/2)
+  const uint32_t weight_left = this->nweight(n.left) + leftmod;
+  const uint32_t weight_right = this->nweight(n.right) + rightmod;
+  return ((1 + weight_left < (1 + 1 + weight_right) / 2) ||
+          (1 + weight_right < (1 + 1 + weight_left) / 2));
+}
+
+template <typename omtdata_t, typename omtdataout_t, bool supports_marks>
+void omt<omtdata_t, omtdataout_t, supports_marks>::insert_internal(
+    subtree *const subtreep, const omtdata_t &value, const uint32_t idx,
+    subtree **const rebalance_subtree) {
+  if (subtreep->is_null()) {
+    paranoid_invariant_zero(idx);
+    const node_idx newidx = this->node_malloc();
+    omt_node *const newnode = &this->d.t.nodes[newidx];
+    newnode->weight = 1;
+    newnode->left.set_to_null();
+    newnode->right.set_to_null();
+    newnode->value = value;
+    subtreep->set_index(newidx);
+  } else {
+    omt_node &n = this->d.t.nodes[subtreep->get_index()];
+    n.weight++;
+    if (idx <= this->nweight(n.left)) {
+      if (*rebalance_subtree == nullptr &&
+          this->will_need_rebalance(*subtreep, 1, 0)) {
+        *rebalance_subtree = subtreep;
+      }
+      this->insert_internal(&n.left, value, idx, rebalance_subtree);
+    } else {
+      if (*rebalance_subtree == nullptr &&
+          this->will_need_rebalance(*subtreep, 0, 1)) {
+        *rebalance_subtree = subtreep;
+      }
+      const uint32_t sub_index = idx - this->nweight(n.left) - 1;
+      this->insert_internal(&n.right, value, sub_index, rebalance_subtree);
+    }
+  }
+}
+
+template <typename omtdata_t, typename omtdataout_t, bool supports_marks>
+void omt<omtdata_t, omtdataout_t, supports_marks>::set_at_internal_array(
+    const omtdata_t &value, const uint32_t idx) {
+  this->d.a.values[this->d.a.start_idx + idx] = value;
+}
+
+template <typename omtdata_t, typename omtdataout_t, bool supports_marks>
+void omt<omtdata_t, omtdataout_t, supports_marks>::set_at_internal(
+    const subtree &st, const omtdata_t &value, const uint32_t idx) {
+  paranoid_invariant(!st.is_null());
+  omt_node &n = this->d.t.nodes[st.get_index()];
+  const uint32_t leftweight = this->nweight(n.left);
+  if (idx < leftweight) {
+    this->set_at_internal(n.left, value, idx);
+  } else if (idx == leftweight) {
+    n.value = value;
+  } else {
+    this->set_at_internal(n.right, value, idx - leftweight - 1);
+  }
+}
+
+template <typename omtdata_t, typename omtdataout_t, bool supports_marks>
+void omt<omtdata_t, omtdataout_t, supports_marks>::delete_internal(
+    subtree *const subtreep, const uint32_t idx, omt_node *const copyn,
+    subtree **const rebalance_subtree) {
+  paranoid_invariant_notnull(subtreep);
+  paranoid_invariant_notnull(rebalance_subtree);
+  paranoid_invariant(!subtreep->is_null());
+  omt_node &n = this->d.t.nodes[subtreep->get_index()];
+  const uint32_t leftweight = this->nweight(n.left);
+  if (idx < leftweight) {
+    n.weight--;
+    if (*rebalance_subtree == nullptr &&
+        this->will_need_rebalance(*subtreep, -1, 0)) {
+      *rebalance_subtree = subtreep;
+    }
+    this->delete_internal(&n.left, idx, copyn, rebalance_subtree);
+  } else if (idx == leftweight) {
+    if (n.left.is_null()) {
+      const uint32_t oldidx = subtreep->get_index();
+      *subtreep = n.right;
+      if (copyn != nullptr) {
+        copyn->value = n.value;
+      }
+      this->node_free(oldidx);
+    } else if (n.right.is_null()) {
+      const uint32_t oldidx = subtreep->get_index();
+      *subtreep = n.left;
+      if (copyn != nullptr) {
+        copyn->value = n.value;
+      }
+      this->node_free(oldidx);
+    } else {
+      if (*rebalance_subtree == nullptr &&
+          this->will_need_rebalance(*subtreep, 0, -1)) {
+        *rebalance_subtree = subtreep;
+      }
+      // don't need to copy up value, it's only used by this
+      // next call, and when that gets to the bottom there
+      // won't be any more recursion
+      n.weight--;
+      this->delete_internal(&n.right, 0, &n, rebalance_subtree);
+    }
+  } else {
+    n.weight--;
+    if (*rebalance_subtree == nullptr &&
+        this->will_need_rebalance(*subtreep, 0, -1)) {
+      *rebalance_subtree = subtreep;
+    }
+    this->delete_internal(&n.right, idx - leftweight - 1, copyn,
+                          rebalance_subtree);
+  }
+}
+
+template <typename omtdata_t, typename omtdataout_t, bool supports_marks>
+template <typename iterate_extra_t,
+          int (*f)(const omtdata_t &, const uint32_t, iterate_extra_t *const)>
+int omt<omtdata_t, omtdataout_t, supports_marks>::iterate_internal_array(
+    const uint32_t left, const uint32_t right,
+    iterate_extra_t *const iterate_extra) const {
+  int r;
+  for (uint32_t i = left; i < right; ++i) {
+    r = f(this->d.a.values[this->d.a.start_idx + i], i, iterate_extra);
+    if (r != 0) {
+      return r;
+    }
+  }
+  return 0;
+}
+
+template <typename omtdata_t, typename omtdataout_t, bool supports_marks>
+template <typename iterate_extra_t,
+          int (*f)(omtdata_t *, const uint32_t, iterate_extra_t *const)>
+void omt<omtdata_t, omtdataout_t, supports_marks>::iterate_ptr_internal(
+    const uint32_t left, const uint32_t right, const subtree &st,
+    const uint32_t idx, iterate_extra_t *const iterate_extra) {
+  if (!st.is_null()) {
+    omt_node &n = this->d.t.nodes[st.get_index()];
+    const uint32_t idx_root = idx + this->nweight(n.left);
+    if (left < idx_root) {
+      this->iterate_ptr_internal<iterate_extra_t, f>(left, right, n.left, idx,
+                                                     iterate_extra);
+    }
+    if (left <= idx_root && idx_root < right) {
+      int r = f(&n.value, idx_root, iterate_extra);
+      lazy_assert_zero(r);
+    }
+    if (idx_root + 1 < right) {
+      this->iterate_ptr_internal<iterate_extra_t, f>(
+          left, right, n.right, idx_root + 1, iterate_extra);
+    }
+  }
+}
+
+template <typename omtdata_t, typename omtdataout_t, bool supports_marks>
+template <typename iterate_extra_t,
+          int (*f)(omtdata_t *, const uint32_t, iterate_extra_t *const)>
+void omt<omtdata_t, omtdataout_t, supports_marks>::iterate_ptr_internal_array(
+    const uint32_t left, const uint32_t right,
+    iterate_extra_t *const iterate_extra) {
+  for (uint32_t i = left; i < right; ++i) {
+    int r = f(&this->d.a.values[this->d.a.start_idx + i], i, iterate_extra);
+    lazy_assert_zero(r);
+  }
+}
+
+template <typename omtdata_t, typename omtdataout_t, bool supports_marks>
+template <typename iterate_extra_t,
+          int (*f)(const omtdata_t &, const uint32_t, iterate_extra_t *const)>
+int omt<omtdata_t, omtdataout_t, supports_marks>::iterate_internal(
+    const uint32_t left, const uint32_t right, const subtree &st,
+    const uint32_t idx, iterate_extra_t *const iterate_extra) const {
+  if (st.is_null()) {
+    return 0;
+  }
+  int r;
+  const omt_node &n = this->d.t.nodes[st.get_index()];
+  const uint32_t idx_root = idx + this->nweight(n.left);
+  if (left < idx_root) {
+    r = this->iterate_internal<iterate_extra_t, f>(left, right, n.left, idx,
+                                                   iterate_extra);
+    if (r != 0) {
+      return r;
+    }
+  }
+  if (left <= idx_root && idx_root < right) {
+    r = f(n.value, idx_root, iterate_extra);
+    if (r != 0) {
+      return r;
+    }
+  }
+  if (idx_root + 1 < right) {
+    return this->iterate_internal<iterate_extra_t, f>(
+        left, right, n.right, idx_root + 1, iterate_extra);
+  }
+  return 0;
+}
+
+template <typename omtdata_t, typename omtdataout_t, bool supports_marks>
+template <typename iterate_extra_t,
+          int (*f)(const omtdata_t &, const uint32_t, iterate_extra_t *const)>
+int omt<omtdata_t, omtdataout_t, supports_marks>::
+    iterate_and_mark_range_internal(const uint32_t left, const uint32_t right,
+                                    const subtree &st, const uint32_t idx,
+                                    iterate_extra_t *const iterate_extra) {
+  paranoid_invariant(!st.is_null());
+  int r;
+  omt_node &n = this->d.t.nodes[st.get_index()];
+  const uint32_t idx_root = idx + this->nweight(n.left);
+  if (left < idx_root && !n.left.is_null()) {
+    n.set_marks_below_bit();
+    r = this->iterate_and_mark_range_internal<iterate_extra_t, f>(
+        left, right, n.left, idx, iterate_extra);
+    if (r != 0) {
+      return r;
+    }
+  }
+  if (left <= idx_root && idx_root < right) {
+    n.set_marked_bit();
+    r = f(n.value, idx_root, iterate_extra);
+    if (r != 0) {
+      return r;
+    }
+  }
+  if (idx_root + 1 < right && !n.right.is_null()) {
+    n.set_marks_below_bit();
+    return this->iterate_and_mark_range_internal<iterate_extra_t, f>(
+        left, right, n.right, idx_root + 1, iterate_extra);
+  }
+  return 0;
+}
+
+template <typename omtdata_t, typename omtdataout_t, bool supports_marks>
+template <typename iterate_extra_t,
+          int (*f)(const omtdata_t &, const uint32_t, iterate_extra_t *const)>
+int omt<omtdata_t, omtdataout_t, supports_marks>::iterate_over_marked_internal(
+    const subtree &st, const uint32_t idx,
+    iterate_extra_t *const iterate_extra) const {
+  if (st.is_null()) {
+    return 0;
+  }
+  int r;
+  const omt_node &n = this->d.t.nodes[st.get_index()];
+  const uint32_t idx_root = idx + this->nweight(n.left);
+  if (n.get_marks_below()) {
+    r = this->iterate_over_marked_internal<iterate_extra_t, f>(n.left, idx,
+                                                               iterate_extra);
+    if (r != 0) {
+      return r;
+    }
+  }
+  if (n.get_marked()) {
+    r = f(n.value, idx_root, iterate_extra);
+    if (r != 0) {
+      return r;
+    }
+  }
+  if (n.get_marks_below()) {
+    return this->iterate_over_marked_internal<iterate_extra_t, f>(
+        n.right, idx_root + 1, iterate_extra);
+  }
+  return 0;
+}
+
+template <typename omtdata_t, typename omtdataout_t, bool supports_marks>
+void omt<omtdata_t, omtdataout_t, supports_marks>::fetch_internal_array(
+    const uint32_t i, omtdataout_t *const value) const {
+  if (value != nullptr) {
+    copyout(value, &this->d.a.values[this->d.a.start_idx + i]);
+  }
+}
+
+template <typename omtdata_t, typename omtdataout_t, bool supports_marks>
+void omt<omtdata_t, omtdataout_t, supports_marks>::fetch_internal(
+    const subtree &st, const uint32_t i, omtdataout_t *const value) const {
+  omt_node &n = this->d.t.nodes[st.get_index()];
+  const uint32_t leftweight = this->nweight(n.left);
+  if (i < leftweight) {
+    this->fetch_internal(n.left, i, value);
+  } else if (i == leftweight) {
+    if (value != nullptr) {
+      copyout(value, &n);
+    }
+  } else {
+    this->fetch_internal(n.right, i - leftweight - 1, value);
+  }
+}
+
+template <typename omtdata_t, typename omtdataout_t, bool supports_marks>
+void omt<omtdata_t, omtdataout_t, supports_marks>::fill_array_with_subtree_idxs(
+    node_idx *const array, const subtree &st) const {
+  if (!st.is_null()) {
+    const omt_node &tree = this->d.t.nodes[st.get_index()];
+    this->fill_array_with_subtree_idxs(&array[0], tree.left);
+    array[this->nweight(tree.left)] = st.get_index();
+    this->fill_array_with_subtree_idxs(&array[this->nweight(tree.left) + 1],
+                                       tree.right);
+  }
+}
+
+template <typename omtdata_t, typename omtdataout_t, bool supports_marks>
+void omt<omtdata_t, omtdataout_t, supports_marks>::rebuild_subtree_from_idxs(
+    subtree *const st, const node_idx *const idxs, const uint32_t numvalues) {
+  if (numvalues == 0) {
+    st->set_to_null();
+  } else {
+    uint32_t halfway = numvalues / 2;
+    st->set_index(idxs[halfway]);
+    // node_idx newidx = idxs[halfway];
+    omt_node &newnode = this->d.t.nodes[st->get_index()];
+    newnode.weight = numvalues;
+    // value is already in there.
+    this->rebuild_subtree_from_idxs(&newnode.left, &idxs[0], halfway);
+    this->rebuild_subtree_from_idxs(&newnode.right, &idxs[halfway + 1],
+                                    numvalues - (halfway + 1));
+    // n_idx = newidx;
+  }
+}
+
+template <typename omtdata_t, typename omtdataout_t, bool supports_marks>
+void omt<omtdata_t, omtdataout_t, supports_marks>::rebalance(
+    subtree *const st) {
+  node_idx idx = st->get_index();
+  if (idx == this->d.t.root.get_index()) {
+    // Try to convert to an array.
+    // If this fails, (malloc) nothing will have changed.
+    // In the failure case we continue on to the standard rebalance
+    // algorithm.
+    this->convert_to_array();
+    if (supports_marks) {
+      this->convert_to_tree();
+    }
+  } else {
+    const omt_node &n = this->d.t.nodes[idx];
+    node_idx *tmp_array;
+    size_t mem_needed = n.weight * (sizeof tmp_array[0]);
+    size_t mem_free =
+        (this->capacity - this->d.t.free_idx) * (sizeof this->d.t.nodes[0]);
+    bool malloced;
+    if (mem_needed <= mem_free) {
+      // There is sufficient free space at the end of the nodes array
+      // to hold enough node indexes to rebalance.
+      malloced = false;
+      tmp_array =
+          reinterpret_cast<node_idx *>(&this->d.t.nodes[this->d.t.free_idx]);
+    } else {
+      malloced = true;
+      XMALLOC_N(n.weight, tmp_array);
+    }
+    this->fill_array_with_subtree_idxs(tmp_array, *st);
+    this->rebuild_subtree_from_idxs(st, tmp_array, n.weight);
+    if (malloced) toku_free(tmp_array);
+  }
+}
+
+template <typename omtdata_t, typename omtdataout_t, bool supports_marks>
+void omt<omtdata_t, omtdataout_t, supports_marks>::copyout(
+    omtdata_t *const out, const omt_node *const n) {
+  *out = n->value;
+}
+
+template <typename omtdata_t, typename omtdataout_t, bool supports_marks>
+void omt<omtdata_t, omtdataout_t, supports_marks>::copyout(
+    omtdata_t **const out, omt_node *const n) {
+  *out = &n->value;
+}
+
+template <typename omtdata_t, typename omtdataout_t, bool supports_marks>
+void omt<omtdata_t, omtdataout_t, supports_marks>::copyout(
+    omtdata_t *const out, const omtdata_t *const stored_value_ptr) {
+  *out = *stored_value_ptr;
+}
+
+template <typename omtdata_t, typename omtdataout_t, bool supports_marks>
+void omt<omtdata_t, omtdataout_t, supports_marks>::copyout(
+    omtdata_t **const out, omtdata_t *const stored_value_ptr) {
+  *out = stored_value_ptr;
+}
+
+template <typename omtdata_t, typename omtdataout_t, bool supports_marks>
+template <typename omtcmp_t, int (*h)(const omtdata_t &, const omtcmp_t &)>
+int omt<omtdata_t, omtdataout_t, supports_marks>::find_internal_zero_array(
+    const omtcmp_t &extra, omtdataout_t *const value,
+    uint32_t *const idxp) const {
+  paranoid_invariant_notnull(idxp);
+  uint32_t min = this->d.a.start_idx;
+  uint32_t limit = this->d.a.start_idx + this->d.a.num_values;
+  uint32_t best_pos = subtree::NODE_NULL;
+  uint32_t best_zero = subtree::NODE_NULL;
+
+  while (min != limit) {
+    uint32_t mid = (min + limit) / 2;
+    int hv = h(this->d.a.values[mid], extra);
+    if (hv < 0) {
+      min = mid + 1;
+    } else if (hv > 0) {
+      best_pos = mid;
+      limit = mid;
+    } else {
+      best_zero = mid;
+      limit = mid;
+    }
+  }
+  if (best_zero != subtree::NODE_NULL) {
+    // Found a zero
+    if (value != nullptr) {
+      copyout(value, &this->d.a.values[best_zero]);
+    }
+    *idxp = best_zero - this->d.a.start_idx;
+    return 0;
+  }
+  if (best_pos != subtree::NODE_NULL)
+    *idxp = best_pos - this->d.a.start_idx;
+  else
+    *idxp = this->d.a.num_values;
+  return DB_NOTFOUND;
+}
+
+template <typename omtdata_t, typename omtdataout_t, bool supports_marks>
+template <typename omtcmp_t, int (*h)(const omtdata_t &, const omtcmp_t &)>
+int omt<omtdata_t, omtdataout_t, supports_marks>::find_internal_zero(
+    const subtree &st, const omtcmp_t &extra, omtdataout_t *const value,
+    uint32_t *const idxp) const {
+  paranoid_invariant_notnull(idxp);
+  if (st.is_null()) {
+    *idxp = 0;
+    return DB_NOTFOUND;
+  }
+  omt_node &n = this->d.t.nodes[st.get_index()];
+  int hv = h(n.value, extra);
+  if (hv < 0) {
+    int r = this->find_internal_zero<omtcmp_t, h>(n.right, extra, value, idxp);
+    *idxp += this->nweight(n.left) + 1;
+    return r;
+  } else if (hv > 0) {
+    return this->find_internal_zero<omtcmp_t, h>(n.left, extra, value, idxp);
+  } else {
+    int r = this->find_internal_zero<omtcmp_t, h>(n.left, extra, value, idxp);
+    if (r == DB_NOTFOUND) {
+      *idxp = this->nweight(n.left);
+      if (value != nullptr) {
+        copyout(value, &n);
+      }
+      r = 0;
+    }
+    return r;
+  }
+}
+
+template <typename omtdata_t, typename omtdataout_t, bool supports_marks>
+template <typename omtcmp_t, int (*h)(const omtdata_t &, const omtcmp_t &)>
+int omt<omtdata_t, omtdataout_t, supports_marks>::find_internal_plus_array(
+    const omtcmp_t &extra, omtdataout_t *const value,
+    uint32_t *const idxp) const {
+  paranoid_invariant_notnull(idxp);
+  uint32_t min = this->d.a.start_idx;
+  uint32_t limit = this->d.a.start_idx + this->d.a.num_values;
+  uint32_t best = subtree::NODE_NULL;
+
+  while (min != limit) {
+    const uint32_t mid = (min + limit) / 2;
+    const int hv = h(this->d.a.values[mid], extra);
+    if (hv > 0) {
+      best = mid;
+      limit = mid;
+    } else {
+      min = mid + 1;
+    }
+  }
+  if (best == subtree::NODE_NULL) {
+    return DB_NOTFOUND;
+  }
+  if (value != nullptr) {
+    copyout(value, &this->d.a.values[best]);
+  }
+  *idxp = best - this->d.a.start_idx;
+  return 0;
+}
+
+template <typename omtdata_t, typename omtdataout_t, bool supports_marks>
+template <typename omtcmp_t, int (*h)(const omtdata_t &, const omtcmp_t &)>
+int omt<omtdata_t, omtdataout_t, supports_marks>::find_internal_plus(
+    const subtree &st, const omtcmp_t &extra, omtdataout_t *const value,
+    uint32_t *const idxp) const {
+  paranoid_invariant_notnull(idxp);
+  if (st.is_null()) {
+    return DB_NOTFOUND;
+  }
+  omt_node *const n = &this->d.t.nodes[st.get_index()];
+  int hv = h(n->value, extra);
+  int r;
+  if (hv > 0) {
+    r = this->find_internal_plus<omtcmp_t, h>(n->left, extra, value, idxp);
+    if (r == DB_NOTFOUND) {
+      *idxp = this->nweight(n->left);
+      if (value != nullptr) {
+        copyout(value, n);
+      }
+      r = 0;
+    }
+  } else {
+    r = this->find_internal_plus<omtcmp_t, h>(n->right, extra, value, idxp);
+    if (r == 0) {
+      *idxp += this->nweight(n->left) + 1;
+    }
+  }
+  return r;
+}
+
+template <typename omtdata_t, typename omtdataout_t, bool supports_marks>
+template <typename omtcmp_t, int (*h)(const omtdata_t &, const omtcmp_t &)>
+int omt<omtdata_t, omtdataout_t, supports_marks>::find_internal_minus_array(
+    const omtcmp_t &extra, omtdataout_t *const value,
+    uint32_t *const idxp) const {
+  paranoid_invariant_notnull(idxp);
+  uint32_t min = this->d.a.start_idx;
+  uint32_t limit = this->d.a.start_idx + this->d.a.num_values;
+  uint32_t best = subtree::NODE_NULL;
+
+  while (min != limit) {
+    const uint32_t mid = (min + limit) / 2;
+    const int hv = h(this->d.a.values[mid], extra);
+    if (hv < 0) {
+      best = mid;
+      min = mid + 1;
+    } else {
+      limit = mid;
+    }
+  }
+  if (best == subtree::NODE_NULL) {
+    return DB_NOTFOUND;
+  }
+  if (value != nullptr) {
+    copyout(value, &this->d.a.values[best]);
+  }
+  *idxp = best - this->d.a.start_idx;
+  return 0;
+}
+
+template <typename omtdata_t, typename omtdataout_t, bool supports_marks>
+template <typename omtcmp_t, int (*h)(const omtdata_t &, const omtcmp_t &)>
+int omt<omtdata_t, omtdataout_t, supports_marks>::find_internal_minus(
+    const subtree &st, const omtcmp_t &extra, omtdataout_t *const value,
+    uint32_t *const idxp) const {
+  paranoid_invariant_notnull(idxp);
+  if (st.is_null()) {
+    return DB_NOTFOUND;
+  }
+  omt_node *const n = &this->d.t.nodes[st.get_index()];
+  int hv = h(n->value, extra);
+  if (hv < 0) {
+    int r =
+        this->find_internal_minus<omtcmp_t, h>(n->right, extra, value, idxp);
+    if (r == 0) {
+      *idxp += this->nweight(n->left) + 1;
+    } else if (r == DB_NOTFOUND) {
+      *idxp = this->nweight(n->left);
+      if (value != nullptr) {
+        copyout(value, n);
+      }
+      r = 0;
+    }
+    return r;
+  } else {
+    return this->find_internal_minus<omtcmp_t, h>(n->left, extra, value, idxp);
+  }
+}
+}  // namespace toku
diff --git a/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/util/partitioned_counter.h b/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/util/partitioned_counter.h
new file mode 100644
index 000000000..f20eeedf2
--- /dev/null
+++ b/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/util/partitioned_counter.h
@@ -0,0 +1,165 @@
+/* -*- mode: C++; c-basic-offset: 4; indent-tabs-mode: nil -*- */
+// vim: ft=cpp:expandtab:ts=8:sw=4:softtabstop=4:
+#ident "$Id$"
+/*======
+This file is part of PerconaFT.
+
+
+Copyright (c) 2006, 2015, Percona and/or its affiliates. All rights reserved.
+
+    PerconaFT is free software: you can redistribute it and/or modify
+    it under the terms of the GNU General Public License, version 2,
+    as published by the Free Software Foundation.
+
+    PerconaFT is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License
+    along with PerconaFT.  If not, see <http://www.gnu.org/licenses/>.
+
+----------------------------------------
+
+    PerconaFT is free software: you can redistribute it and/or modify
+    it under the terms of the GNU Affero General Public License, version 3,
+    as published by the Free Software Foundation.
+
+    PerconaFT is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU Affero General Public License for more details.
+
+    You should have received a copy of the GNU Affero General Public License
+    along with PerconaFT.  If not, see <http://www.gnu.org/licenses/>.
+
+----------------------------------------
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.
+======= */
+
+#ident \
+    "Copyright (c) 2006, 2015, Percona and/or its affiliates. All rights reserved."
+
+#pragma once
+
+// Overview: A partitioned_counter provides a counter that can be incremented
+// and the running sum can be read at any time.
+//  We assume that increments are frequent, whereas reading is infrequent.
+// Implementation hint: Use thread-local storage so each thread increments its
+// own data.  The increment does not require a lock or atomic operation.
+//  Reading the data can be performed by iterating over the thread-local
+//  versions, summing them up. The data structure also includes a sum for all
+//  the threads that have died. Use a pthread_key to create the thread-local
+//  versions.  When a thread finishes, the system calls pthread_key destructor
+//  which can add that thread's copy into the sum_of_dead counter.
+// Rationale: For statistics such as are found in engine status, we need a
+// counter that requires no cache misses to increment.  We've seen significant
+//  performance speedups by removing certain counters.  Rather than removing
+//  those statistics, we would like to just make the counter fast. We generally
+//  increment the counters frequently, and want to fetch the values
+//  infrequently. The counters are monotonic. The counters can be split into
+//  many counters, which can be summed up at the end. We don't care if we get
+//  slightly out-of-date counter sums when we read the counter.  We don't care
+//  if there is a race on reading the a counter
+//   variable and incrementing.
+//  See tests/test_partitioned_counter.c for some performance measurements.
+// Operations:
+//   create_partitioned_counter    Create a counter initialized to zero.
+//   destroy_partitioned_counter   Destroy it.
+//   increment_partitioned_counter Increment it.  This is the frequent
+//   operation. read_partitioned_counter      Get the current value.  This is
+//   infrequent.
+// See partitioned_counter.cc for the abstraction function and representation
+// invariant.
+//
+// The google style guide says to avoid using constructors, and it appears that
+// constructors may have broken all the tests, because they called
+// pthread_key_create before the key was actually created.  So the google style
+// guide may have some wisdom there...
+//
+// This version does not use constructors, essentially reverrting to the google
+// C++ style guide.
+//
+
+// The old C interface.  This required a bunch of explicit
+// ___attribute__((__destructor__)) functions to remember to destroy counters at
+// the end.
+#if defined(__cplusplus)
+extern "C" {
+#endif
+
+typedef struct partitioned_counter *PARTITIONED_COUNTER;
+PARTITIONED_COUNTER create_partitioned_counter(void);
+// Effect: Create a counter, initialized to zero.
+
+void destroy_partitioned_counter(PARTITIONED_COUNTER);
+// Effect: Destroy the counter.  No operations on that counter are permitted
+// after this.
+
+void increment_partitioned_counter(PARTITIONED_COUNTER, uint64_t amount);
+// Effect: Increment the counter by amount.
+// Requires: No overflows.  This is a 64-bit unsigned counter.
+
+uint64_t read_partitioned_counter(PARTITIONED_COUNTER)
+    __attribute__((__visibility__("default")));
+// Effect: Return the current value of the counter.
+
+void partitioned_counters_init(void);
+// Effect: Initialize any partitioned counters data structures that must be set
+// up before any partitioned counters run.
+
+void partitioned_counters_destroy(void);
+// Effect: Destroy any partitioned counters data structures.
+
+#if defined(__cplusplus)
+};
+#endif
+
+#if 0
+#include <pthread.h>
+
+#include "fttypes.h"
+
+// Used inside the PARTITIONED_COUNTER.
+struct linked_list_head {
+    struct linked_list_element *first;
+};
+
+
+class PARTITIONED_COUNTER {
+public:
+    PARTITIONED_COUNTER(void);
+    // Effect: Construct a counter, initialized to zero.
+
+    ~PARTITIONED_COUNTER(void);
+    // Effect: Destruct the counter.
+
+    void increment(uint64_t amount);
+    // Effect: Increment the counter by amount.  This is a 64-bit unsigned counter, and if you overflow it, you will get overflowed results (that is mod 2^64).
+    // Requires: Don't use this from a static constructor or destructor.
+
+    uint64_t read(void);
+    // Effect: Read the sum.
+    // Requires: Don't use this from a static constructor or destructor.
+
+private:
+    uint64_t       _sum_of_dead;             // The sum of all thread-local counts from threads that have terminated.
+    pthread_key_t   _key;                     // The pthread_key which gives us the hook to construct and destruct thread-local storage.
+    struct linked_list_head _ll_counter_head; // A linked list of all the thread-local information for this counter.
+    
+    // This function is used to destroy the thread-local part of the state when a thread terminates.
+    // But it's not the destructor for the local part of the counter, it's a destructor on a "dummy" key just so that we get a notification when a thread ends.
+    friend void destroy_thread_local_part_of_partitioned_counters (void *);
+};
+#endif
diff --git a/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/util/status.h b/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/util/status.h
new file mode 100644
index 000000000..3fd0095d0
--- /dev/null
+++ b/src/rocksdb/utilities/transactions/lock/range/range_tree/lib/util/status.h
@@ -0,0 +1,76 @@
+/* -*- mode: C++; c-basic-offset: 4; indent-tabs-mode: nil -*- */
+// vim: ft=cpp:expandtab:ts=8:sw=4:softtabstop=4:
+#ident "$Id$"
+/*======
+This file is part of PerconaFT.
+
+
+Copyright (c) 2006, 2015, Percona and/or its affiliates. All rights reserved.
+
+    PerconaFT is free software: you can redistribute it and/or modify
+    it under the terms of the GNU General Public License, version 2,
+    as published by the Free Software Foundation.
+
+    PerconaFT is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License
+    along with PerconaFT.  If not, see <http://www.gnu.org/licenses/>.
+
+----------------------------------------
+
+    PerconaFT is free software: you can redistribute it and/or modify
+    it under the terms of the GNU Affero General Public License, version 3,
+    as published by the Free Software Foundation.
+
+    PerconaFT is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU Affero General Public License for more details.
+
+    You should have received a copy of the GNU Affero General Public License
+    along with PerconaFT.  If not, see <http://www.gnu.org/licenses/>.
+
+----------------------------------------
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.
+======= */
+
+#ident \
+    "Copyright (c) 2006, 2015, Percona and/or its affiliates. All rights reserved."
+
+#pragma once
+
+#include "partitioned_counter.h"
+// PORT2: #include <util/constexpr.h>
+
+#define TOKUFT_STATUS_INIT(array, k, c, t, l, inc)                    \
+  do {                                                                \
+    array.status[k].keyname = #k;                                     \
+    array.status[k].columnname = #c;                                  \
+    array.status[k].type = t;                                         \
+    array.status[k].legend = l;                                       \
+    constexpr_static_assert(                                          \
+        strcmp(#c, "NULL") && strcmp(#c, "0"),                        \
+        "Use nullptr for no column name instead of NULL, 0, etc..."); \
+    constexpr_static_assert(                                          \
+        (inc) == TOKU_ENGINE_STATUS || strcmp(#c, "nullptr"),         \
+        "Missing column name.");                                      \
+    array.status[k].include =                                         \
+        static_cast<toku_engine_status_include_type>(inc);            \
+    if (t == STATUS_PARCOUNT) {                                       \
+      array.status[k].value.parcount = create_partitioned_counter();  \
+    }                                                                 \
+  } while (0)
diff --git a/src/rocksdb/utilities/transactions/lock/range/range_tree/range_tree_lock_manager.cc b/src/rocksdb/utilities/transactions/lock/range/range_tree/range_tree_lock_manager.cc
new file mode 100644
index 000000000..531165dea
--- /dev/null
+++ b/src/rocksdb/utilities/transactions/lock/range/range_tree/range_tree_lock_manager.cc
@@ -0,0 +1,503 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#ifndef ROCKSDB_LITE
+#ifndef OS_WIN
+
+#include "utilities/transactions/lock/range/range_tree/range_tree_lock_manager.h"
+
+#include <algorithm>
+#include <cinttypes>
+#include <mutex>
+
+#include "monitoring/perf_context_imp.h"
+#include "rocksdb/slice.h"
+#include "rocksdb/utilities/transaction_db_mutex.h"
+#include "test_util/sync_point.h"
+#include "util/cast_util.h"
+#include "util/hash.h"
+#include "util/thread_local.h"
+#include "utilities/transactions/lock/range/range_tree/range_tree_lock_tracker.h"
+#include "utilities/transactions/pessimistic_transaction_db.h"
+#include "utilities/transactions/transaction_db_mutex_impl.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+RangeLockManagerHandle* NewRangeLockManager(
+    std::shared_ptr<TransactionDBMutexFactory> mutex_factory) {
+  std::shared_ptr<TransactionDBMutexFactory> use_factory;
+
+  if (mutex_factory) {
+    use_factory = mutex_factory;
+  } else {
+    use_factory.reset(new TransactionDBMutexFactoryImpl());
+  }
+  return new RangeTreeLockManager(use_factory);
+}
+
+static const char SUFFIX_INFIMUM = 0x0;
+static const char SUFFIX_SUPREMUM = 0x1;
+
+// Convert Endpoint into an internal format used for storing it in locktree
+// (DBT structure is used for passing endpoints to locktree and getting back)
+void serialize_endpoint(const Endpoint& endp, std::string* buf) {
+  buf->push_back(endp.inf_suffix ? SUFFIX_SUPREMUM : SUFFIX_INFIMUM);
+  buf->append(endp.slice.data(), endp.slice.size());
+}
+
+// Decode the endpoint from the format it is stored in the locktree (DBT) to
+// the one used outside: either Endpoint or EndpointWithString
+template <typename EndpointStruct>
+void deserialize_endpoint(const DBT* dbt, EndpointStruct* endp) {
+  assert(dbt->size >= 1);
+  const char* dbt_data = (const char*)dbt->data;
+  char suffix = dbt_data[0];
+  assert(suffix == SUFFIX_INFIMUM || suffix == SUFFIX_SUPREMUM);
+  endp->inf_suffix = (suffix == SUFFIX_SUPREMUM);
+  endp->slice = decltype(EndpointStruct::slice)(dbt_data + 1, dbt->size - 1);
+}
+
+// Get a range lock on [start_key; end_key] range
+Status RangeTreeLockManager::TryLock(PessimisticTransaction* txn,
+                                     uint32_t column_family_id,
+                                     const Endpoint& start_endp,
+                                     const Endpoint& end_endp, Env*,
+                                     bool exclusive) {
+  toku::lock_request request;
+  request.create(mutex_factory_);
+  DBT start_key_dbt, end_key_dbt;
+
+  TEST_SYNC_POINT("RangeTreeLockManager::TryRangeLock:enter");
+  std::string start_key;
+  std::string end_key;
+  serialize_endpoint(start_endp, &start_key);
+  serialize_endpoint(end_endp, &end_key);
+
+  toku_fill_dbt(&start_key_dbt, start_key.data(), start_key.size());
+  toku_fill_dbt(&end_key_dbt, end_key.data(), end_key.size());
+
+  auto lt = GetLockTreeForCF(column_family_id);
+
+  // Put the key waited on into request's m_extra. See
+  // wait_callback_for_locktree for details.
+  std::string wait_key(start_endp.slice.data(), start_endp.slice.size());
+
+  request.set(lt.get(), (TXNID)txn, &start_key_dbt, &end_key_dbt,
+              exclusive ? toku::lock_request::WRITE : toku::lock_request::READ,
+              false /* not a big txn */, &wait_key);
+
+  // This is for "periodically wake up and check if the wait is killed" feature
+  // which we are not using.
+  uint64_t killed_time_msec = 0;
+  uint64_t wait_time_msec = txn->GetLockTimeout();
+
+  if (wait_time_msec == static_cast<uint64_t>(-1)) {
+    // The transaction has no wait timeout. lock_request::wait doesn't support
+    // this, it needs a number of milliseconds to wait. Pass it one year to
+    // be safe.
+    wait_time_msec = uint64_t(1000) * 60 * 60 * 24 * 365;
+  } else {
+    // convert microseconds to milliseconds
+    wait_time_msec = (wait_time_msec + 500) / 1000;
+  }
+
+  std::vector<RangeDeadlockInfo> di_path;
+  request.m_deadlock_cb = [&](TXNID txnid, bool is_exclusive,
+                              const DBT* start_dbt, const DBT* end_dbt) {
+    EndpointWithString start;
+    EndpointWithString end;
+    deserialize_endpoint(start_dbt, &start);
+    deserialize_endpoint(end_dbt, &end);
+
+    di_path.push_back({txnid, column_family_id, is_exclusive, std::move(start),
+                       std::move(end)});
+  };
+
+  request.start();
+
+  const int r = request.wait(wait_time_msec, killed_time_msec,
+                             nullptr,  // killed_callback
+                             wait_callback_for_locktree, nullptr);
+
+  // Inform the txn that we are no longer waiting:
+  txn->ClearWaitingTxn();
+
+  request.destroy();
+  switch (r) {
+    case 0:
+      break;  // fall through
+    case DB_LOCK_NOTGRANTED:
+      return Status::TimedOut(Status::SubCode::kLockTimeout);
+    case TOKUDB_OUT_OF_LOCKS:
+      return Status::Busy(Status::SubCode::kLockLimit);
+    case DB_LOCK_DEADLOCK: {
+      std::reverse(di_path.begin(), di_path.end());
+      dlock_buffer_.AddNewPath(
+          RangeDeadlockPath(di_path, request.get_start_time()));
+      return Status::Busy(Status::SubCode::kDeadlock);
+    }
+    default:
+      assert(0);
+      return Status::Busy(Status::SubCode::kLockLimit);
+  }
+
+  return Status::OK();
+}
+
+// Wait callback that locktree library will call to inform us about
+// the lock waits that are in progress.
+void wait_callback_for_locktree(void*, toku::lock_wait_infos* infos) {
+  TEST_SYNC_POINT("RangeTreeLockManager::TryRangeLock:EnterWaitingTxn");
+  for (auto wait_info : *infos) {
+    // As long as we hold the lock on the locktree's pending request queue
+    // this should be safe.
+    auto txn = (PessimisticTransaction*)wait_info.waiter;
+    auto cf_id = (ColumnFamilyId)wait_info.ltree->get_dict_id().dictid;
+
+    autovector<TransactionID> waitee_ids;
+    for (auto waitee : wait_info.waitees) {
+      waitee_ids.push_back(waitee);
+    }
+    txn->SetWaitingTxn(waitee_ids, cf_id, (std::string*)wait_info.m_extra);
+  }
+
+  // Here we can assume that the locktree code will now wait for some lock
+  TEST_SYNC_POINT("RangeTreeLockManager::TryRangeLock:WaitingTxn");
+}
+
+void RangeTreeLockManager::UnLock(PessimisticTransaction* txn,
+                                  ColumnFamilyId column_family_id,
+                                  const std::string& key, Env*) {
+  auto locktree = GetLockTreeForCF(column_family_id);
+  std::string endp_image;
+  serialize_endpoint({key.data(), key.size(), false}, &endp_image);
+
+  DBT key_dbt;
+  toku_fill_dbt(&key_dbt, endp_image.data(), endp_image.size());
+
+  toku::range_buffer range_buf;
+  range_buf.create();
+  range_buf.append(&key_dbt, &key_dbt);
+
+  locktree->release_locks((TXNID)txn, &range_buf);
+  range_buf.destroy();
+
+  toku::lock_request::retry_all_lock_requests(
+      locktree.get(), wait_callback_for_locktree, nullptr);
+}
+
+void RangeTreeLockManager::UnLock(PessimisticTransaction* txn,
+                                  const LockTracker& tracker, Env*) {
+  const RangeTreeLockTracker* range_tracker =
+      static_cast<const RangeTreeLockTracker*>(&tracker);
+
+  RangeTreeLockTracker* range_trx_tracker =
+      static_cast<RangeTreeLockTracker*>(&txn->GetTrackedLocks());
+  bool all_keys = (range_trx_tracker == range_tracker);
+
+  // tracked_locks_->range_list may hold nullptr if the transaction has never
+  // acquired any locks.
+  ((RangeTreeLockTracker*)range_tracker)->ReleaseLocks(this, txn, all_keys);
+}
+
+int RangeTreeLockManager::CompareDbtEndpoints(void* arg, const DBT* a_key,
+                                              const DBT* b_key) {
+  const char* a = (const char*)a_key->data;
+  const char* b = (const char*)b_key->data;
+
+  size_t a_len = a_key->size;
+  size_t b_len = b_key->size;
+
+  size_t min_len = std::min(a_len, b_len);
+
+  // Compare the values. The first byte encodes the endpoint type, its value
+  // is either SUFFIX_INFIMUM or SUFFIX_SUPREMUM.
+  Comparator* cmp = (Comparator*)arg;
+  int res = cmp->Compare(Slice(a + 1, min_len - 1), Slice(b + 1, min_len - 1));
+  if (!res) {
+    if (b_len > min_len) {
+      // a is shorter;
+      if (a[0] == SUFFIX_INFIMUM) {
+        return -1;  //"a is smaller"
+      } else {
+        // a is considered padded with 0xFF:FF:FF:FF...
+        return 1;  // "a" is bigger
+      }
+    } else if (a_len > min_len) {
+      // the opposite of the above: b is shorter.
+      if (b[0] == SUFFIX_INFIMUM) {
+        return 1;  //"b is smaller"
+      } else {
+        // b is considered padded with 0xFF:FF:FF:FF...
+        return -1;  // "b" is bigger
+      }
+    } else {
+      // the lengths are equal (and the key values, too)
+      if (a[0] < b[0]) {
+        return -1;
+      } else if (a[0] > b[0]) {
+        return 1;
+      } else {
+        return 0;
+      }
+    }
+  } else {
+    return res;
+  }
+}
+
+namespace {
+void UnrefLockTreeMapsCache(void* ptr) {
+  // Called when a thread exits or a ThreadLocalPtr gets destroyed.
+  auto lock_tree_map_cache = static_cast<
+      std::unordered_map<ColumnFamilyId, std::shared_ptr<toku::locktree>>*>(
+      ptr);
+  delete lock_tree_map_cache;
+}
+}  // anonymous namespace
+
+RangeTreeLockManager::RangeTreeLockManager(
+    std::shared_ptr<TransactionDBMutexFactory> mutex_factory)
+    : mutex_factory_(mutex_factory),
+      ltree_lookup_cache_(new ThreadLocalPtr(&UnrefLockTreeMapsCache)),
+      dlock_buffer_(10) {
+  ltm_.create(on_create, on_destroy, on_escalate, nullptr, mutex_factory_);
+}
+
+int RangeTreeLockManager::on_create(toku::locktree* lt, void* arg) {
+  // arg is a pointer to RangeTreeLockManager
+  lt->set_escalation_barrier_func(&OnEscalationBarrierCheck, arg);
+  return 0;
+}
+
+bool RangeTreeLockManager::OnEscalationBarrierCheck(const DBT* a, const DBT* b,
+                                                    void* extra) {
+  Endpoint a_endp, b_endp;
+  deserialize_endpoint(a, &a_endp);
+  deserialize_endpoint(b, &b_endp);
+  auto self = static_cast<RangeTreeLockManager*>(extra);
+  return self->barrier_func_(a_endp, b_endp);
+}
+
+void RangeTreeLockManager::SetRangeDeadlockInfoBufferSize(
+    uint32_t target_size) {
+  dlock_buffer_.Resize(target_size);
+}
+
+void RangeTreeLockManager::Resize(uint32_t target_size) {
+  SetRangeDeadlockInfoBufferSize(target_size);
+}
+
+std::vector<RangeDeadlockPath>
+RangeTreeLockManager::GetRangeDeadlockInfoBuffer() {
+  return dlock_buffer_.PrepareBuffer();
+}
+
+std::vector<DeadlockPath> RangeTreeLockManager::GetDeadlockInfoBuffer() {
+  std::vector<DeadlockPath> res;
+  std::vector<RangeDeadlockPath> data = GetRangeDeadlockInfoBuffer();
+  // report left endpoints
+  for (auto it = data.begin(); it != data.end(); ++it) {
+    std::vector<DeadlockInfo> path;
+
+    for (auto it2 = it->path.begin(); it2 != it->path.end(); ++it2) {
+      path.push_back(
+          {it2->m_txn_id, it2->m_cf_id, it2->m_exclusive, it2->m_start.slice});
+    }
+    res.push_back(DeadlockPath(path, it->deadlock_time));
+  }
+  return res;
+}
+
+// @brief  Lock Escalation Callback function
+//
+// @param txnid   Transaction whose locks got escalated
+// @param lt      Lock Tree where escalation is happening
+// @param buffer  Escalation result: list of locks that this transaction now
+//                owns in this lock tree.
+// @param void*   Callback context
+void RangeTreeLockManager::on_escalate(TXNID txnid, const toku::locktree* lt,
+                                       const toku::range_buffer& buffer,
+                                       void*) {
+  auto txn = (PessimisticTransaction*)txnid;
+  ((RangeTreeLockTracker*)&txn->GetTrackedLocks())->ReplaceLocks(lt, buffer);
+}
+
+RangeTreeLockManager::~RangeTreeLockManager() {
+  autovector<void*> local_caches;
+  ltree_lookup_cache_->Scrape(&local_caches, nullptr);
+  for (auto cache : local_caches) {
+    delete static_cast<LockTreeMap*>(cache);
+  }
+  ltree_map_.clear();  // this will call release_lt() for all locktrees
+  ltm_.destroy();
+}
+
+RangeLockManagerHandle::Counters RangeTreeLockManager::GetStatus() {
+  LTM_STATUS_S ltm_status_test;
+  ltm_.get_status(&ltm_status_test);
+  Counters res;
+
+  // Searching status variable by its string name is how Toku's unit tests
+  // do it (why didn't they make LTM_ESCALATION_COUNT constant visible?)
+  // lookup keyname in status
+  for (int i = 0; i < LTM_STATUS_S::LTM_STATUS_NUM_ROWS; i++) {
+    TOKU_ENGINE_STATUS_ROW status = &ltm_status_test.status[i];
+    if (strcmp(status->keyname, "LTM_ESCALATION_COUNT") == 0) {
+      res.escalation_count = status->value.num;
+      continue;
+    }
+    if (strcmp(status->keyname, "LTM_WAIT_COUNT") == 0) {
+      res.lock_wait_count = status->value.num;
+      continue;
+    }
+    if (strcmp(status->keyname, "LTM_SIZE_CURRENT") == 0) {
+      res.current_lock_memory = status->value.num;
+    }
+  }
+  return res;
+}
+
+std::shared_ptr<toku::locktree> RangeTreeLockManager::MakeLockTreePtr(
+    toku::locktree* lt) {
+  toku::locktree_manager* ltm = &ltm_;
+  return std::shared_ptr<toku::locktree>(
+      lt, [ltm](toku::locktree* p) { ltm->release_lt(p); });
+}
+
+void RangeTreeLockManager::AddColumnFamily(const ColumnFamilyHandle* cfh) {
+  uint32_t column_family_id = cfh->GetID();
+
+  InstrumentedMutexLock l(&ltree_map_mutex_);
+  if (ltree_map_.find(column_family_id) == ltree_map_.end()) {
+    DICTIONARY_ID dict_id = {.dictid = column_family_id};
+    toku::comparator cmp;
+    cmp.create(CompareDbtEndpoints, (void*)cfh->GetComparator());
+    toku::locktree* ltree =
+        ltm_.get_lt(dict_id, cmp,
+                    /* on_create_extra*/ static_cast<void*>(this));
+    // This is ok to because get_lt has copied the comparator:
+    cmp.destroy();
+
+    ltree_map_.insert({column_family_id, MakeLockTreePtr(ltree)});
+  }
+}
+
+void RangeTreeLockManager::RemoveColumnFamily(const ColumnFamilyHandle* cfh) {
+  uint32_t column_family_id = cfh->GetID();
+  // Remove lock_map for this column family.  Since the lock map is stored
+  // as a shared ptr, concurrent transactions can still keep using it
+  // until they release their references to it.
+
+  // TODO what if one drops a column family while transaction(s) still have
+  // locks in it?
+  // locktree uses column family'c Comparator* as the criteria to do tree
+  // ordering. If the comparator is gone, we won't even be able to remove the
+  // elements from the locktree.
+  // A possible solution might be to remove everything right now:
+  //  - wait until everyone traversing the locktree are gone
+  //  - remove everything from the locktree.
+  //  - some transactions may have acquired locks in their LockTracker objects.
+  //    Arrange something so we don't blow up when they try to release them.
+  //  - ...
+  // This use case (drop column family while somebody is using it) doesn't seem
+  // the priority, though.
+
+  {
+    InstrumentedMutexLock l(&ltree_map_mutex_);
+
+    auto lock_maps_iter = ltree_map_.find(column_family_id);
+    assert(lock_maps_iter != ltree_map_.end());
+    ltree_map_.erase(lock_maps_iter);
+  }  // lock_map_mutex_
+
+  autovector<void*> local_caches;
+  ltree_lookup_cache_->Scrape(&local_caches, nullptr);
+  for (auto cache : local_caches) {
+    delete static_cast<LockTreeMap*>(cache);
+  }
+}
+
+std::shared_ptr<toku::locktree> RangeTreeLockManager::GetLockTreeForCF(
+    ColumnFamilyId column_family_id) {
+  // First check thread-local cache
+  if (ltree_lookup_cache_->Get() == nullptr) {
+    ltree_lookup_cache_->Reset(new LockTreeMap());
+  }
+
+  auto ltree_map_cache = static_cast<LockTreeMap*>(ltree_lookup_cache_->Get());
+
+  auto it = ltree_map_cache->find(column_family_id);
+  if (it != ltree_map_cache->end()) {
+    // Found lock map for this column family.
+    return it->second;
+  }
+
+  // Not found in local cache, grab mutex and check shared LockMaps
+  InstrumentedMutexLock l(&ltree_map_mutex_);
+
+  it = ltree_map_.find(column_family_id);
+  if (it == ltree_map_.end()) {
+    return nullptr;
+  } else {
+    // Found lock map.  Store in thread-local cache and return.
+    ltree_map_cache->insert({column_family_id, it->second});
+    return it->second;
+  }
+}
+
+struct LOCK_PRINT_CONTEXT {
+  RangeLockManagerHandle::RangeLockStatus* data;  // Save locks here
+  uint32_t cfh_id;  // Column Family whose tree we are traversing
+};
+
+// Report left endpoints of the acquired locks
+LockManager::PointLockStatus RangeTreeLockManager::GetPointLockStatus() {
+  PointLockStatus res;
+  LockManager::RangeLockStatus data = GetRangeLockStatus();
+  // report left endpoints
+  for (auto it = data.begin(); it != data.end(); ++it) {
+    auto& val = it->second;
+    res.insert({it->first, {val.start.slice, val.ids, val.exclusive}});
+  }
+  return res;
+}
+
+static void push_into_lock_status_data(void* param, const DBT* left,
+                                       const DBT* right, TXNID txnid_arg,
+                                       bool is_shared, TxnidVector* owners) {
+  struct LOCK_PRINT_CONTEXT* ctx = (LOCK_PRINT_CONTEXT*)param;
+  struct RangeLockInfo info;
+
+  info.exclusive = !is_shared;
+
+  deserialize_endpoint(left, &info.start);
+  deserialize_endpoint(right, &info.end);
+
+  if (txnid_arg != TXNID_SHARED) {
+    info.ids.push_back(txnid_arg);
+  } else {
+    for (auto it : *owners) {
+      info.ids.push_back(it);
+    }
+  }
+  ctx->data->insert({ctx->cfh_id, info});
+}
+
+LockManager::RangeLockStatus RangeTreeLockManager::GetRangeLockStatus() {
+  LockManager::RangeLockStatus data;
+  {
+    InstrumentedMutexLock l(&ltree_map_mutex_);
+    for (auto it : ltree_map_) {
+      LOCK_PRINT_CONTEXT ctx = {&data, it.first};
+      it.second->dump_locks((void*)&ctx, push_into_lock_status_data);
+    }
+  }
+  return data;
+}
+
+}  // namespace ROCKSDB_NAMESPACE
+#endif  // OS_WIN
+#endif  // ROCKSDB_LITE
diff --git a/src/rocksdb/utilities/transactions/lock/range/range_tree/range_tree_lock_manager.h b/src/rocksdb/utilities/transactions/lock/range/range_tree/range_tree_lock_manager.h
new file mode 100644
index 000000000..e4236d600
--- /dev/null
+++ b/src/rocksdb/utilities/transactions/lock/range/range_tree/range_tree_lock_manager.h
@@ -0,0 +1,137 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#pragma once
+#ifndef ROCKSDB_LITE
+#ifndef OS_WIN
+
+// For DeadlockInfoBuffer:
+#include "util/thread_local.h"
+#include "utilities/transactions/lock/point/point_lock_manager.h"
+#include "utilities/transactions/lock/range/range_lock_manager.h"
+
+// Lock Tree library:
+#include "utilities/transactions/lock/range/range_tree/lib/locktree/lock_request.h"
+#include "utilities/transactions/lock/range/range_tree/lib/locktree/locktree.h"
+#include "utilities/transactions/lock/range/range_tree/range_tree_lock_tracker.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+typedef DeadlockInfoBufferTempl<RangeDeadlockPath> RangeDeadlockInfoBuffer;
+
+// A Range Lock Manager that uses PerconaFT's locktree library
+class RangeTreeLockManager : public RangeLockManagerBase,
+                             public RangeLockManagerHandle {
+ public:
+  LockManager* getLockManager() override { return this; }
+
+  void AddColumnFamily(const ColumnFamilyHandle* cfh) override;
+  void RemoveColumnFamily(const ColumnFamilyHandle* cfh) override;
+
+  void Resize(uint32_t) override;
+  std::vector<DeadlockPath> GetDeadlockInfoBuffer() override;
+
+  std::vector<RangeDeadlockPath> GetRangeDeadlockInfoBuffer() override;
+  void SetRangeDeadlockInfoBufferSize(uint32_t target_size) override;
+
+  // Get a lock on a range
+  //  @note only exclusive locks are currently supported (requesting a
+  //  non-exclusive lock will get an exclusive one)
+  using LockManager::TryLock;
+  Status TryLock(PessimisticTransaction* txn, ColumnFamilyId column_family_id,
+                 const Endpoint& start_endp, const Endpoint& end_endp, Env* env,
+                 bool exclusive) override;
+
+  void UnLock(PessimisticTransaction* txn, const LockTracker& tracker,
+              Env* env) override;
+  void UnLock(PessimisticTransaction* txn, ColumnFamilyId column_family_id,
+              const std::string& key, Env* env) override;
+  void UnLock(PessimisticTransaction*, ColumnFamilyId, const Endpoint&,
+              const Endpoint&, Env*) override {
+    // TODO: range unlock does nothing...
+  }
+
+  explicit RangeTreeLockManager(
+      std::shared_ptr<TransactionDBMutexFactory> mutex_factory);
+
+  ~RangeTreeLockManager() override;
+
+  int SetMaxLockMemory(size_t max_lock_memory) override {
+    return ltm_.set_max_lock_memory(max_lock_memory);
+  }
+
+  size_t GetMaxLockMemory() override { return ltm_.get_max_lock_memory(); }
+
+  Counters GetStatus() override;
+
+  bool IsPointLockSupported() const override {
+    // One could have acquired a point lock (it is reduced to range lock)
+    return true;
+  }
+
+  PointLockStatus GetPointLockStatus() override;
+
+  // This is from LockManager
+  LockManager::RangeLockStatus GetRangeLockStatus() override;
+
+  // This has the same meaning as GetRangeLockStatus but is from
+  // RangeLockManagerHandle
+  RangeLockManagerHandle::RangeLockStatus GetRangeLockStatusData() override {
+    return GetRangeLockStatus();
+  }
+
+  bool IsRangeLockSupported() const override { return true; }
+
+  const LockTrackerFactory& GetLockTrackerFactory() const override {
+    return RangeTreeLockTrackerFactory::Get();
+  }
+
+  // Get the locktree which stores locks for the Column Family with given cf_id
+  std::shared_ptr<toku::locktree> GetLockTreeForCF(ColumnFamilyId cf_id);
+
+  void SetEscalationBarrierFunc(EscalationBarrierFunc func) override {
+    barrier_func_ = func;
+  }
+
+ private:
+  toku::locktree_manager ltm_;
+
+  EscalationBarrierFunc barrier_func_ =
+      [](const Endpoint&, const Endpoint&) -> bool { return false; };
+
+  std::shared_ptr<TransactionDBMutexFactory> mutex_factory_;
+
+  // Map from cf_id to locktree*. Can only be accessed while holding the
+  // ltree_map_mutex_. Must use a custom deleter that calls ltm_.release_lt
+  using LockTreeMap =
+      std::unordered_map<ColumnFamilyId, std::shared_ptr<toku::locktree>>;
+  LockTreeMap ltree_map_;
+
+  InstrumentedMutex ltree_map_mutex_;
+
+  // Per-thread cache of ltree_map_.
+  // (uses the same approach as TransactionLockMgr::lock_maps_cache_)
+  std::unique_ptr<ThreadLocalPtr> ltree_lookup_cache_;
+
+  RangeDeadlockInfoBuffer dlock_buffer_;
+
+  std::shared_ptr<toku::locktree> MakeLockTreePtr(toku::locktree* lt);
+  static int CompareDbtEndpoints(void* arg, const DBT* a_key, const DBT* b_key);
+
+  // Callbacks
+  static int on_create(toku::locktree*, void*);
+  static void on_destroy(toku::locktree*) {}
+  static void on_escalate(TXNID txnid, const toku::locktree* lt,
+                          const toku::range_buffer& buffer, void* extra);
+
+  static bool OnEscalationBarrierCheck(const DBT* a, const DBT* b, void* extra);
+};
+
+void serialize_endpoint(const Endpoint& endp, std::string* buf);
+void wait_callback_for_locktree(void* cdata, toku::lock_wait_infos* infos);
+
+}  // namespace ROCKSDB_NAMESPACE
+#endif  // OS_WIN
+#endif  // ROCKSDB_LITE
diff --git a/src/rocksdb/utilities/transactions/lock/range/range_tree/range_tree_lock_tracker.cc b/src/rocksdb/utilities/transactions/lock/range/range_tree/range_tree_lock_tracker.cc
new file mode 100644
index 000000000..be1e1478b
--- /dev/null
+++ b/src/rocksdb/utilities/transactions/lock/range/range_tree/range_tree_lock_tracker.cc
@@ -0,0 +1,156 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#ifndef ROCKSDB_LITE
+#ifndef OS_WIN
+
+#include "utilities/transactions/lock/range/range_tree/range_tree_lock_tracker.h"
+
+#include "utilities/transactions/lock/range/range_tree/range_tree_lock_manager.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+RangeLockList *RangeTreeLockTracker::getOrCreateList() {
+  if (range_list_) return range_list_.get();
+
+  // Doesn't exist, create
+  range_list_.reset(new RangeLockList());
+  return range_list_.get();
+}
+
+void RangeTreeLockTracker::Track(const PointLockRequest &lock_req) {
+  DBT key_dbt;
+  std::string key;
+  serialize_endpoint(Endpoint(lock_req.key, false), &key);
+  toku_fill_dbt(&key_dbt, key.data(), key.size());
+  RangeLockList *rl = getOrCreateList();
+  rl->Append(lock_req.column_family_id, &key_dbt, &key_dbt);
+}
+
+void RangeTreeLockTracker::Track(const RangeLockRequest &lock_req) {
+  DBT start_dbt, end_dbt;
+  std::string start_key, end_key;
+
+  serialize_endpoint(lock_req.start_endp, &start_key);
+  serialize_endpoint(lock_req.end_endp, &end_key);
+
+  toku_fill_dbt(&start_dbt, start_key.data(), start_key.size());
+  toku_fill_dbt(&end_dbt, end_key.data(), end_key.size());
+
+  RangeLockList *rl = getOrCreateList();
+  rl->Append(lock_req.column_family_id, &start_dbt, &end_dbt);
+}
+
+PointLockStatus RangeTreeLockTracker::GetPointLockStatus(
+    ColumnFamilyId /*cf_id*/, const std::string & /*key*/) const {
+  // This function is not expected to be called as RangeTreeLockTracker::
+  // IsPointLockSupported() returns false. Return the status which indicates
+  // the point is not locked.
+  PointLockStatus p;
+  p.locked = false;
+  p.exclusive = true;
+  p.seq = 0;
+  return p;
+}
+
+void RangeTreeLockTracker::Clear() { range_list_.reset(); }
+
+void RangeLockList::Append(ColumnFamilyId cf_id, const DBT *left_key,
+                           const DBT *right_key) {
+  MutexLock l(&mutex_);
+  // Only the transaction owner thread calls this function.
+  // The same thread does the lock release, so we can be certain nobody is
+  // releasing the locks concurrently.
+  assert(!releasing_locks_.load());
+  auto it = buffers_.find(cf_id);
+  if (it == buffers_.end()) {
+    // create a new one
+    it = buffers_.emplace(cf_id, std::make_shared<toku::range_buffer>()).first;
+    it->second->create();
+  }
+  it->second->append(left_key, right_key);
+}
+
+void RangeLockList::ReleaseLocks(RangeTreeLockManager *mgr,
+                                 PessimisticTransaction *txn,
+                                 bool all_trx_locks) {
+  {
+    MutexLock l(&mutex_);
+    // The lt->release_locks() call below will walk range_list->buffer_. We
+    // need to prevent lock escalation callback from replacing
+    // range_list->buffer_ while we are doing that.
+    //
+    // Additional complication here is internal mutex(es) in the locktree
+    // (let's call them latches):
+    // - Lock escalation first obtains latches on the lock tree
+    // - Then, it calls RangeTreeLockManager::on_escalate to replace
+    // transaction's range_list->buffer_. = Access to that buffer must be
+    // synchronized, so it will want to acquire the range_list->mutex_.
+    //
+    // While in this function we would want to do the reverse:
+    // - Acquire range_list->mutex_ to prevent access to the range_list.
+    // - Then, lt->release_locks() call will walk through the range_list
+    // - and acquire latches on parts of the lock tree to remove locks from
+    //   it.
+    //
+    // How do we avoid the deadlock? The idea is that here we set
+    // releasing_locks_=true, and release the mutex.
+    // All other users of the range_list must:
+    // - Acquire the mutex, then check that releasing_locks_=false.
+    //   (the code in this function doesnt do that as there's only one thread
+    //    that releases transaction's locks)
+    releasing_locks_.store(true);
+  }
+
+  for (auto it : buffers_) {
+    // Don't try to call release_locks() if the buffer is empty! if we are
+    //  not holding any locks, the lock tree might be in the STO-mode with
+    //  another transaction, and our attempt to release an empty set of locks
+    //  will cause an assertion failure.
+    if (it.second->get_num_ranges()) {
+      auto lt_ptr = mgr->GetLockTreeForCF(it.first);
+      toku::locktree *lt = lt_ptr.get();
+
+      lt->release_locks((TXNID)txn, it.second.get(), all_trx_locks);
+
+      it.second->destroy();
+      it.second->create();
+
+      toku::lock_request::retry_all_lock_requests(lt,
+                                                  wait_callback_for_locktree);
+    }
+  }
+
+  Clear();
+  releasing_locks_.store(false);
+}
+
+void RangeLockList::ReplaceLocks(const toku::locktree *lt,
+                                 const toku::range_buffer &buffer) {
+  MutexLock l(&mutex_);
+  if (releasing_locks_.load()) {
+    // Do nothing. The transaction is releasing its locks, so it will not care
+    // about having a correct list of ranges. (In TokuDB,
+    // toku_db_txn_escalate_callback() makes use of this property, too)
+    return;
+  }
+
+  ColumnFamilyId cf_id = (ColumnFamilyId)lt->get_dict_id().dictid;
+
+  auto it = buffers_.find(cf_id);
+  it->second->destroy();
+  it->second->create();
+
+  toku::range_buffer::iterator iter(&buffer);
+  toku::range_buffer::iterator::record rec;
+  while (iter.current(&rec)) {
+    it->second->append(rec.get_left_key(), rec.get_right_key());
+    iter.next();
+  }
+}
+
+}  // namespace ROCKSDB_NAMESPACE
+#endif  // OS_WIN
+#endif  // ROCKSDB_LITE
diff --git a/src/rocksdb/utilities/transactions/lock/range/range_tree/range_tree_lock_tracker.h b/src/rocksdb/utilities/transactions/lock/range/range_tree/range_tree_lock_tracker.h
new file mode 100644
index 000000000..4ef48d252
--- /dev/null
+++ b/src/rocksdb/utilities/transactions/lock/range/range_tree/range_tree_lock_tracker.h
@@ -0,0 +1,146 @@
+// Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+// This source code is licensed under both the GPLv2 (found in the
+// COPYING file in the root directory) and Apache 2.0 License
+// (found in the LICENSE.Apache file in the root directory).
+
+#pragma once
+
+#include <memory>
+#include <string>
+#include <unordered_map>
+
+#include "util/mutexlock.h"
+#include "utilities/transactions/lock/lock_tracker.h"
+#include "utilities/transactions/pessimistic_transaction.h"
+
+// Range Locking:
+#include "lib/locktree/lock_request.h"
+#include "lib/locktree/locktree.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+class RangeTreeLockManager;
+
+// Storage for locks that are currently held by a transaction.
+//
+// Locks are kept in toku::range_buffer because toku::locktree::release_locks()
+// accepts that as an argument.
+//
+// Note: the list of locks may differ slighly from the contents of the lock
+// tree, due to concurrency between lock acquisition, lock release, and lock
+// escalation. See MDEV-18227 and RangeTreeLockManager::UnLock for details.
+// This property is currently harmless.
+//
+// Append() and ReleaseLocks() are not thread-safe, as they are expected to be
+// called only by the owner transaction. ReplaceLocks() is safe to call from
+// other threads.
+class RangeLockList {
+ public:
+  ~RangeLockList() { Clear(); }
+
+  RangeLockList() : releasing_locks_(false) {}
+
+  void Append(ColumnFamilyId cf_id, const DBT* left_key, const DBT* right_key);
+  void ReleaseLocks(RangeTreeLockManager* mgr, PessimisticTransaction* txn,
+                    bool all_trx_locks);
+  void ReplaceLocks(const toku::locktree* lt, const toku::range_buffer& buffer);
+
+ private:
+  void Clear() {
+    for (auto it : buffers_) {
+      it.second->destroy();
+    }
+    buffers_.clear();
+  }
+
+  std::unordered_map<ColumnFamilyId, std::shared_ptr<toku::range_buffer>>
+      buffers_;
+  port::Mutex mutex_;
+  std::atomic<bool> releasing_locks_;
+};
+
+// A LockTracker-based object that is used together with RangeTreeLockManager.
+class RangeTreeLockTracker : public LockTracker {
+ public:
+  RangeTreeLockTracker() : range_list_(nullptr) {}
+
+  RangeTreeLockTracker(const RangeTreeLockTracker&) = delete;
+  RangeTreeLockTracker& operator=(const RangeTreeLockTracker&) = delete;
+
+  void Track(const PointLockRequest&) override;
+  void Track(const RangeLockRequest&) override;
+
+  bool IsPointLockSupported() const override {
+    // This indicates that we don't implement GetPointLockStatus()
+    return false;
+  }
+  bool IsRangeLockSupported() const override { return true; }
+
+  // a Not-supported dummy implementation.
+  UntrackStatus Untrack(const RangeLockRequest& /*lock_request*/) override {
+    return UntrackStatus::NOT_TRACKED;
+  }
+
+  UntrackStatus Untrack(const PointLockRequest& /*lock_request*/) override {
+    return UntrackStatus::NOT_TRACKED;
+  }
+
+  // "If this method is not supported, leave it as a no-op."
+  void Merge(const LockTracker&) override {}
+
+  // "If this method is not supported, leave it as a no-op."
+  void Subtract(const LockTracker&) override {}
+
+  void Clear() override;
+
+  // "If this method is not supported, returns nullptr."
+  virtual LockTracker* GetTrackedLocksSinceSavePoint(
+      const LockTracker&) const override {
+    return nullptr;
+  }
+
+  PointLockStatus GetPointLockStatus(ColumnFamilyId column_family_id,
+                                     const std::string& key) const override;
+
+  // The return value is only used for tests
+  uint64_t GetNumPointLocks() const override { return 0; }
+
+  ColumnFamilyIterator* GetColumnFamilyIterator() const override {
+    return nullptr;
+  }
+
+  KeyIterator* GetKeyIterator(
+      ColumnFamilyId /*column_family_id*/) const override {
+    return nullptr;
+  }
+
+  void ReleaseLocks(RangeTreeLockManager* mgr, PessimisticTransaction* txn,
+                    bool all_trx_locks) {
+    if (range_list_) range_list_->ReleaseLocks(mgr, txn, all_trx_locks);
+  }
+
+  void ReplaceLocks(const toku::locktree* lt,
+                    const toku::range_buffer& buffer) {
+    // range_list_ cannot be NULL here
+    range_list_->ReplaceLocks(lt, buffer);
+  }
+
+ private:
+  RangeLockList* getOrCreateList();
+  std::unique_ptr<RangeLockList> range_list_;
+};
+
+class RangeTreeLockTrackerFactory : public LockTrackerFactory {
+ public:
+  static const RangeTreeLockTrackerFactory& Get() {
+    static const RangeTreeLockTrackerFactory instance;
+    return instance;
+  }
+
+  LockTracker* Create() const override { return new RangeTreeLockTracker(); }
+
+ private:
+  RangeTreeLockTrackerFactory() {}
+};
+
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/utilities/transactions/optimistic_transaction.cc b/src/rocksdb/utilities/transactions/optimistic_transaction.cc
new file mode 100644
index 000000000..0ee0f28b6
--- /dev/null
+++ b/src/rocksdb/utilities/transactions/optimistic_transaction.cc
@@ -0,0 +1,196 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#ifndef ROCKSDB_LITE
+
+#include "utilities/transactions/optimistic_transaction.h"
+
+#include <string>
+
+#include "db/column_family.h"
+#include "db/db_impl/db_impl.h"
+#include "rocksdb/comparator.h"
+#include "rocksdb/db.h"
+#include "rocksdb/status.h"
+#include "rocksdb/utilities/optimistic_transaction_db.h"
+#include "util/cast_util.h"
+#include "util/string_util.h"
+#include "utilities/transactions/lock/point/point_lock_tracker.h"
+#include "utilities/transactions/optimistic_transaction.h"
+#include "utilities/transactions/optimistic_transaction_db_impl.h"
+#include "utilities/transactions/transaction_util.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+struct WriteOptions;
+
+OptimisticTransaction::OptimisticTransaction(
+    OptimisticTransactionDB* txn_db, const WriteOptions& write_options,
+    const OptimisticTransactionOptions& txn_options)
+    : TransactionBaseImpl(txn_db->GetBaseDB(), write_options,
+                          PointLockTrackerFactory::Get()),
+      txn_db_(txn_db) {
+  Initialize(txn_options);
+}
+
+void OptimisticTransaction::Initialize(
+    const OptimisticTransactionOptions& txn_options) {
+  if (txn_options.set_snapshot) {
+    SetSnapshot();
+  }
+}
+
+void OptimisticTransaction::Reinitialize(
+    OptimisticTransactionDB* txn_db, const WriteOptions& write_options,
+    const OptimisticTransactionOptions& txn_options) {
+  TransactionBaseImpl::Reinitialize(txn_db->GetBaseDB(), write_options);
+  Initialize(txn_options);
+}
+
+OptimisticTransaction::~OptimisticTransaction() {}
+
+void OptimisticTransaction::Clear() { TransactionBaseImpl::Clear(); }
+
+Status OptimisticTransaction::Prepare() {
+  return Status::InvalidArgument(
+      "Two phase commit not supported for optimistic transactions.");
+}
+
+Status OptimisticTransaction::Commit() {
+  auto txn_db_impl = static_cast_with_check<OptimisticTransactionDBImpl,
+                                            OptimisticTransactionDB>(txn_db_);
+  assert(txn_db_impl);
+  switch (txn_db_impl->GetValidatePolicy()) {
+    case OccValidationPolicy::kValidateParallel:
+      return CommitWithParallelValidate();
+    case OccValidationPolicy::kValidateSerial:
+      return CommitWithSerialValidate();
+    default:
+      assert(0);
+  }
+  // unreachable, just void compiler complain
+  return Status::OK();
+}
+
+Status OptimisticTransaction::CommitWithSerialValidate() {
+  // Set up callback which will call CheckTransactionForConflicts() to
+  // check whether this transaction is safe to be committed.
+  OptimisticTransactionCallback callback(this);
+
+  DBImpl* db_impl = static_cast_with_check<DBImpl>(db_->GetRootDB());
+
+  Status s = db_impl->WriteWithCallback(
+      write_options_, GetWriteBatch()->GetWriteBatch(), &callback);
+
+  if (s.ok()) {
+    Clear();
+  }
+
+  return s;
+}
+
+Status OptimisticTransaction::CommitWithParallelValidate() {
+  auto txn_db_impl = static_cast_with_check<OptimisticTransactionDBImpl,
+                                            OptimisticTransactionDB>(txn_db_);
+  assert(txn_db_impl);
+  DBImpl* db_impl = static_cast_with_check<DBImpl>(db_->GetRootDB());
+  assert(db_impl);
+  const size_t space = txn_db_impl->GetLockBucketsSize();
+  std::set<size_t> lk_idxes;
+  std::vector<std::unique_lock<std::mutex>> lks;
+  std::unique_ptr<LockTracker::ColumnFamilyIterator> cf_it(
+      tracked_locks_->GetColumnFamilyIterator());
+  assert(cf_it != nullptr);
+  while (cf_it->HasNext()) {
+    ColumnFamilyId cf = cf_it->Next();
+    std::unique_ptr<LockTracker::KeyIterator> key_it(
+        tracked_locks_->GetKeyIterator(cf));
+    assert(key_it != nullptr);
+    while (key_it->HasNext()) {
+      const std::string& key = key_it->Next();
+      lk_idxes.insert(FastRange64(GetSliceNPHash64(key), space));
+    }
+  }
+  // NOTE: in a single txn, all bucket-locks are taken in ascending order.
+  // In this way, txns from different threads all obey this rule so that
+  // deadlock can be avoided.
+  for (auto v : lk_idxes) {
+    lks.emplace_back(txn_db_impl->LockBucket(v));
+  }
+
+  Status s = TransactionUtil::CheckKeysForConflicts(db_impl, *tracked_locks_,
+                                                    true /* cache_only */);
+  if (!s.ok()) {
+    return s;
+  }
+
+  s = db_impl->Write(write_options_, GetWriteBatch()->GetWriteBatch());
+  if (s.ok()) {
+    Clear();
+  }
+
+  return s;
+}
+
+Status OptimisticTransaction::Rollback() {
+  Clear();
+  return Status::OK();
+}
+
+// Record this key so that we can check it for conflicts at commit time.
+//
+// 'exclusive' is unused for OptimisticTransaction.
+Status OptimisticTransaction::TryLock(ColumnFamilyHandle* column_family,
+                                      const Slice& key, bool read_only,
+                                      bool exclusive, const bool do_validate,
+                                      const bool assume_tracked) {
+  assert(!assume_tracked);  // not supported
+  (void)assume_tracked;
+  if (!do_validate) {
+    return Status::OK();
+  }
+  uint32_t cfh_id = GetColumnFamilyID(column_family);
+
+  SetSnapshotIfNeeded();
+
+  SequenceNumber seq;
+  if (snapshot_) {
+    seq = snapshot_->GetSequenceNumber();
+  } else {
+    seq = db_->GetLatestSequenceNumber();
+  }
+
+  std::string key_str = key.ToString();
+
+  TrackKey(cfh_id, key_str, seq, read_only, exclusive);
+
+  // Always return OK. Confilct checking will happen at commit time.
+  return Status::OK();
+}
+
+// Returns OK if it is safe to commit this transaction.  Returns Status::Busy
+// if there are read or write conflicts that would prevent us from committing OR
+// if we can not determine whether there would be any such conflicts.
+//
+// Should only be called on writer thread in order to avoid any race conditions
+// in detecting write conflicts.
+Status OptimisticTransaction::CheckTransactionForConflicts(DB* db) {
+  auto db_impl = static_cast_with_check<DBImpl>(db);
+
+  // Since we are on the write thread and do not want to block other writers,
+  // we will do a cache-only conflict check.  This can result in TryAgain
+  // getting returned if there is not sufficient memtable history to check
+  // for conflicts.
+  return TransactionUtil::CheckKeysForConflicts(db_impl, *tracked_locks_,
+                                                true /* cache_only */);
+}
+
+Status OptimisticTransaction::SetName(const TransactionName& /* unused */) {
+  return Status::InvalidArgument("Optimistic transactions cannot be named.");
+}
+
+}  // namespace ROCKSDB_NAMESPACE
+
+#endif  // ROCKSDB_LITE
diff --git a/src/rocksdb/utilities/transactions/optimistic_transaction.h b/src/rocksdb/utilities/transactions/optimistic_transaction.h
new file mode 100644
index 000000000..de23233d5
--- /dev/null
+++ b/src/rocksdb/utilities/transactions/optimistic_transaction.h
@@ -0,0 +1,101 @@
+// Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#pragma once
+
+#ifndef ROCKSDB_LITE
+
+#include <stack>
+#include <string>
+#include <unordered_map>
+#include <vector>
+
+#include "db/write_callback.h"
+#include "rocksdb/db.h"
+#include "rocksdb/slice.h"
+#include "rocksdb/snapshot.h"
+#include "rocksdb/status.h"
+#include "rocksdb/types.h"
+#include "rocksdb/utilities/optimistic_transaction_db.h"
+#include "rocksdb/utilities/transaction.h"
+#include "rocksdb/utilities/write_batch_with_index.h"
+#include "utilities/transactions/transaction_base.h"
+#include "utilities/transactions/transaction_util.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+class OptimisticTransaction : public TransactionBaseImpl {
+ public:
+  OptimisticTransaction(OptimisticTransactionDB* db,
+                        const WriteOptions& write_options,
+                        const OptimisticTransactionOptions& txn_options);
+  // No copying allowed
+  OptimisticTransaction(const OptimisticTransaction&) = delete;
+  void operator=(const OptimisticTransaction&) = delete;
+
+  virtual ~OptimisticTransaction();
+
+  void Reinitialize(OptimisticTransactionDB* txn_db,
+                    const WriteOptions& write_options,
+                    const OptimisticTransactionOptions& txn_options);
+
+  Status Prepare() override;
+
+  Status Commit() override;
+
+  Status Rollback() override;
+
+  Status SetName(const TransactionName& name) override;
+
+ protected:
+  Status TryLock(ColumnFamilyHandle* column_family, const Slice& key,
+                 bool read_only, bool exclusive, const bool do_validate = true,
+                 const bool assume_tracked = false) override;
+
+ private:
+  ROCKSDB_FIELD_UNUSED OptimisticTransactionDB* const txn_db_;
+
+  friend class OptimisticTransactionCallback;
+
+  void Initialize(const OptimisticTransactionOptions& txn_options);
+
+  // Returns OK if it is safe to commit this transaction.  Returns Status::Busy
+  // if there are read or write conflicts that would prevent us from committing
+  // OR if we can not determine whether there would be any such conflicts.
+  //
+  // Should only be called on writer thread.
+  Status CheckTransactionForConflicts(DB* db);
+
+  void Clear() override;
+
+  void UnlockGetForUpdate(ColumnFamilyHandle* /* unused */,
+                          const Slice& /* unused */) override {
+    // Nothing to unlock.
+  }
+
+  Status CommitWithSerialValidate();
+
+  Status CommitWithParallelValidate();
+};
+
+// Used at commit time to trigger transaction validation
+class OptimisticTransactionCallback : public WriteCallback {
+ public:
+  explicit OptimisticTransactionCallback(OptimisticTransaction* txn)
+      : txn_(txn) {}
+
+  Status Callback(DB* db) override {
+    return txn_->CheckTransactionForConflicts(db);
+  }
+
+  bool AllowWriteBatching() override { return false; }
+
+ private:
+  OptimisticTransaction* txn_;
+};
+
+}  // namespace ROCKSDB_NAMESPACE
+
+#endif  // ROCKSDB_LITE
diff --git a/src/rocksdb/utilities/transactions/optimistic_transaction_db_impl.cc b/src/rocksdb/utilities/transactions/optimistic_transaction_db_impl.cc
new file mode 100644
index 000000000..bffb3d5ed
--- /dev/null
+++ b/src/rocksdb/utilities/transactions/optimistic_transaction_db_impl.cc
@@ -0,0 +1,111 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#ifndef ROCKSDB_LITE
+
+#include "utilities/transactions/optimistic_transaction_db_impl.h"
+
+#include <string>
+#include <vector>
+
+#include "db/db_impl/db_impl.h"
+#include "rocksdb/db.h"
+#include "rocksdb/options.h"
+#include "rocksdb/utilities/optimistic_transaction_db.h"
+#include "utilities/transactions/optimistic_transaction.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+Transaction* OptimisticTransactionDBImpl::BeginTransaction(
+    const WriteOptions& write_options,
+    const OptimisticTransactionOptions& txn_options, Transaction* old_txn) {
+  if (old_txn != nullptr) {
+    ReinitializeTransaction(old_txn, write_options, txn_options);
+    return old_txn;
+  } else {
+    return new OptimisticTransaction(this, write_options, txn_options);
+  }
+}
+
+std::unique_lock<std::mutex> OptimisticTransactionDBImpl::LockBucket(
+    size_t idx) {
+  assert(idx < bucketed_locks_.size());
+  return std::unique_lock<std::mutex>(*bucketed_locks_[idx]);
+}
+
+Status OptimisticTransactionDB::Open(const Options& options,
+                                     const std::string& dbname,
+                                     OptimisticTransactionDB** dbptr) {
+  DBOptions db_options(options);
+  ColumnFamilyOptions cf_options(options);
+  std::vector<ColumnFamilyDescriptor> column_families;
+  column_families.push_back(
+      ColumnFamilyDescriptor(kDefaultColumnFamilyName, cf_options));
+  std::vector<ColumnFamilyHandle*> handles;
+  Status s = Open(db_options, dbname, column_families, &handles, dbptr);
+  if (s.ok()) {
+    assert(handles.size() == 1);
+    // i can delete the handle since DBImpl is always holding a reference to
+    // default column family
+    delete handles[0];
+  }
+
+  return s;
+}
+
+Status OptimisticTransactionDB::Open(
+    const DBOptions& db_options, const std::string& dbname,
+    const std::vector<ColumnFamilyDescriptor>& column_families,
+    std::vector<ColumnFamilyHandle*>* handles,
+    OptimisticTransactionDB** dbptr) {
+  return OptimisticTransactionDB::Open(db_options,
+                                       OptimisticTransactionDBOptions(), dbname,
+                                       column_families, handles, dbptr);
+}
+
+Status OptimisticTransactionDB::Open(
+    const DBOptions& db_options,
+    const OptimisticTransactionDBOptions& occ_options,
+    const std::string& dbname,
+    const std::vector<ColumnFamilyDescriptor>& column_families,
+    std::vector<ColumnFamilyHandle*>* handles,
+    OptimisticTransactionDB** dbptr) {
+  Status s;
+  DB* db;
+
+  std::vector<ColumnFamilyDescriptor> column_families_copy = column_families;
+
+  // Enable MemTable History if not already enabled
+  for (auto& column_family : column_families_copy) {
+    ColumnFamilyOptions* options = &column_family.options;
+
+    if (options->max_write_buffer_size_to_maintain == 0 &&
+        options->max_write_buffer_number_to_maintain == 0) {
+      // Setting to -1 will set the History size to
+      // max_write_buffer_number * write_buffer_size.
+      options->max_write_buffer_size_to_maintain = -1;
+    }
+  }
+
+  s = DB::Open(db_options, dbname, column_families_copy, handles, &db);
+
+  if (s.ok()) {
+    *dbptr = new OptimisticTransactionDBImpl(db, occ_options);
+  }
+
+  return s;
+}
+
+void OptimisticTransactionDBImpl::ReinitializeTransaction(
+    Transaction* txn, const WriteOptions& write_options,
+    const OptimisticTransactionOptions& txn_options) {
+  assert(dynamic_cast<OptimisticTransaction*>(txn) != nullptr);
+  auto txn_impl = reinterpret_cast<OptimisticTransaction*>(txn);
+
+  txn_impl->Reinitialize(this, write_options, txn_options);
+}
+
+}  // namespace ROCKSDB_NAMESPACE
+#endif  // ROCKSDB_LITE
diff --git a/src/rocksdb/utilities/transactions/optimistic_transaction_db_impl.h b/src/rocksdb/utilities/transactions/optimistic_transaction_db_impl.h
new file mode 100644
index 000000000..88e86ea4a
--- /dev/null
+++ b/src/rocksdb/utilities/transactions/optimistic_transaction_db_impl.h
@@ -0,0 +1,88 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#pragma once
+#ifndef ROCKSDB_LITE
+
+#include <algorithm>
+#include <mutex>
+#include <vector>
+
+#include "rocksdb/db.h"
+#include "rocksdb/options.h"
+#include "rocksdb/utilities/optimistic_transaction_db.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+class OptimisticTransactionDBImpl : public OptimisticTransactionDB {
+ public:
+  explicit OptimisticTransactionDBImpl(
+      DB* db, const OptimisticTransactionDBOptions& occ_options,
+      bool take_ownership = true)
+      : OptimisticTransactionDB(db),
+        db_owner_(take_ownership),
+        validate_policy_(occ_options.validate_policy) {
+    if (validate_policy_ == OccValidationPolicy::kValidateParallel) {
+      uint32_t bucket_size = std::max(16u, occ_options.occ_lock_buckets);
+      bucketed_locks_.reserve(bucket_size);
+      for (size_t i = 0; i < bucket_size; ++i) {
+        bucketed_locks_.emplace_back(
+            std::unique_ptr<std::mutex>(new std::mutex));
+      }
+    }
+  }
+
+  ~OptimisticTransactionDBImpl() {
+    // Prevent this stackable from destroying
+    // base db
+    if (!db_owner_) {
+      db_ = nullptr;
+    }
+  }
+
+  Transaction* BeginTransaction(const WriteOptions& write_options,
+                                const OptimisticTransactionOptions& txn_options,
+                                Transaction* old_txn) override;
+
+  // Transactional `DeleteRange()` is not yet supported.
+  using StackableDB::DeleteRange;
+  virtual Status DeleteRange(const WriteOptions&, ColumnFamilyHandle*,
+                             const Slice&, const Slice&) override {
+    return Status::NotSupported();
+  }
+
+  // Range deletions also must not be snuck into `WriteBatch`es as they are
+  // incompatible with `OptimisticTransactionDB`.
+  virtual Status Write(const WriteOptions& write_opts,
+                       WriteBatch* batch) override {
+    if (batch->HasDeleteRange()) {
+      return Status::NotSupported();
+    }
+    return OptimisticTransactionDB::Write(write_opts, batch);
+  }
+
+  size_t GetLockBucketsSize() const { return bucketed_locks_.size(); }
+
+  OccValidationPolicy GetValidatePolicy() const { return validate_policy_; }
+
+  std::unique_lock<std::mutex> LockBucket(size_t idx);
+
+ private:
+  // NOTE: used in validation phase. Each key is hashed into some
+  // bucket. We then take the lock in the hash value order to avoid deadlock.
+  std::vector<std::unique_ptr<std::mutex>> bucketed_locks_;
+
+  bool db_owner_;
+
+  const OccValidationPolicy validate_policy_;
+
+  void ReinitializeTransaction(Transaction* txn,
+                               const WriteOptions& write_options,
+                               const OptimisticTransactionOptions& txn_options =
+                                   OptimisticTransactionOptions());
+};
+
+}  // namespace ROCKSDB_NAMESPACE
+#endif  // ROCKSDB_LITE
diff --git a/src/rocksdb/utilities/transactions/optimistic_transaction_test.cc b/src/rocksdb/utilities/transactions/optimistic_transaction_test.cc
new file mode 100644
index 000000000..aa8192c32
--- /dev/null
+++ b/src/rocksdb/utilities/transactions/optimistic_transaction_test.cc
@@ -0,0 +1,1491 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#ifndef ROCKSDB_LITE
+
+#include <functional>
+#include <string>
+#include <thread>
+
+#include "db/db_impl/db_impl.h"
+#include "db/db_test_util.h"
+#include "port/port.h"
+#include "rocksdb/db.h"
+#include "rocksdb/perf_context.h"
+#include "rocksdb/utilities/optimistic_transaction_db.h"
+#include "rocksdb/utilities/transaction.h"
+#include "test_util/sync_point.h"
+#include "test_util/testharness.h"
+#include "test_util/transaction_test_util.h"
+#include "util/crc32c.h"
+#include "util/random.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+class OptimisticTransactionTest
+    : public testing::Test,
+      public testing::WithParamInterface<OccValidationPolicy> {
+ public:
+  OptimisticTransactionDB* txn_db;
+  std::string dbname;
+  Options options;
+
+  OptimisticTransactionTest() {
+    options.create_if_missing = true;
+    options.max_write_buffer_number = 2;
+    options.max_write_buffer_size_to_maintain = 2 * Arena::kInlineSize;
+    options.merge_operator.reset(new TestPutOperator());
+    dbname = test::PerThreadDBPath("optimistic_transaction_testdb");
+
+    EXPECT_OK(DestroyDB(dbname, options));
+    Open();
+  }
+  ~OptimisticTransactionTest() override {
+    delete txn_db;
+    EXPECT_OK(DestroyDB(dbname, options));
+  }
+
+  void Reopen() {
+    delete txn_db;
+    txn_db = nullptr;
+    Open();
+  }
+
+ private:
+  void Open() {
+    ColumnFamilyOptions cf_options(options);
+    OptimisticTransactionDBOptions occ_opts;
+    occ_opts.validate_policy = GetParam();
+    std::vector<ColumnFamilyDescriptor> column_families;
+    std::vector<ColumnFamilyHandle*> handles;
+    column_families.push_back(
+        ColumnFamilyDescriptor(kDefaultColumnFamilyName, cf_options));
+    Status s =
+        OptimisticTransactionDB::Open(DBOptions(options), occ_opts, dbname,
+                                      column_families, &handles, &txn_db);
+
+    ASSERT_OK(s);
+    ASSERT_NE(txn_db, nullptr);
+    ASSERT_EQ(handles.size(), 1);
+    delete handles[0];
+  }
+};
+
+TEST_P(OptimisticTransactionTest, SuccessTest) {
+  WriteOptions write_options;
+  ReadOptions read_options;
+  std::string value;
+
+  ASSERT_OK(txn_db->Put(write_options, Slice("foo"), Slice("bar")));
+  ASSERT_OK(txn_db->Put(write_options, Slice("foo2"), Slice("bar")));
+
+  Transaction* txn = txn_db->BeginTransaction(write_options);
+  ASSERT_NE(txn, nullptr);
+
+  ASSERT_OK(txn->GetForUpdate(read_options, "foo", &value));
+  ASSERT_EQ(value, "bar");
+
+  ASSERT_OK(txn->Put(Slice("foo"), Slice("bar2")));
+
+  ASSERT_OK(txn->GetForUpdate(read_options, "foo", &value));
+  ASSERT_EQ(value, "bar2");
+
+  ASSERT_OK(txn->Commit());
+
+  ASSERT_OK(txn_db->Get(read_options, "foo", &value));
+  ASSERT_EQ(value, "bar2");
+
+  delete txn;
+}
+
+TEST_P(OptimisticTransactionTest, WriteConflictTest) {
+  WriteOptions write_options;
+  ReadOptions read_options;
+  std::string value;
+
+  ASSERT_OK(txn_db->Put(write_options, "foo", "bar"));
+  ASSERT_OK(txn_db->Put(write_options, "foo2", "bar"));
+
+  Transaction* txn = txn_db->BeginTransaction(write_options);
+  ASSERT_NE(txn, nullptr);
+
+  ASSERT_OK(txn->Put("foo", "bar2"));
+
+  // This Put outside of a transaction will conflict with the previous write
+  ASSERT_OK(txn_db->Put(write_options, "foo", "barz"));
+
+  ASSERT_OK(txn_db->Get(read_options, "foo", &value));
+  ASSERT_EQ(value, "barz");
+  ASSERT_EQ(1, txn->GetNumKeys());
+
+  Status s = txn->Commit();
+  ASSERT_TRUE(s.IsBusy());  // Txn should not commit
+
+  // Verify that transaction did not write anything
+  ASSERT_OK(txn_db->Get(read_options, "foo", &value));
+  ASSERT_EQ(value, "barz");
+  ASSERT_OK(txn_db->Get(read_options, "foo2", &value));
+  ASSERT_EQ(value, "bar");
+
+  delete txn;
+}
+
+TEST_P(OptimisticTransactionTest, WriteConflictTest2) {
+  WriteOptions write_options;
+  ReadOptions read_options;
+  OptimisticTransactionOptions txn_options;
+  std::string value;
+
+  ASSERT_OK(txn_db->Put(write_options, "foo", "bar"));
+  ASSERT_OK(txn_db->Put(write_options, "foo2", "bar"));
+
+  txn_options.set_snapshot = true;
+  Transaction* txn = txn_db->BeginTransaction(write_options, txn_options);
+  ASSERT_NE(txn, nullptr);
+
+  // This Put outside of a transaction will conflict with a later write
+  ASSERT_OK(txn_db->Put(write_options, "foo", "barz"));
+
+  ASSERT_OK(txn->Put(
+      "foo", "bar2"));  // Conflicts with write done after snapshot taken
+
+  ASSERT_OK(txn_db->Get(read_options, "foo", &value));
+  ASSERT_EQ(value, "barz");
+
+  Status s = txn->Commit();
+  ASSERT_TRUE(s.IsBusy());  // Txn should not commit
+
+  // Verify that transaction did not write anything
+  ASSERT_OK(txn_db->Get(read_options, "foo", &value));
+  ASSERT_EQ(value, "barz");
+  ASSERT_OK(txn_db->Get(read_options, "foo2", &value));
+  ASSERT_EQ(value, "bar");
+
+  delete txn;
+}
+
+TEST_P(OptimisticTransactionTest, WriteConflictTest3) {
+  ASSERT_OK(txn_db->Put(WriteOptions(), "foo", "bar"));
+
+  Transaction* txn = txn_db->BeginTransaction(WriteOptions());
+  ASSERT_NE(txn, nullptr);
+
+  std::string value;
+  ASSERT_OK(txn->GetForUpdate(ReadOptions(), "foo", &value));
+  ASSERT_EQ(value, "bar");
+  ASSERT_OK(txn->Merge("foo", "bar3"));
+
+  // Merge outside of a transaction should conflict with the previous merge
+  ASSERT_OK(txn_db->Merge(WriteOptions(), "foo", "bar2"));
+  ASSERT_OK(txn_db->Get(ReadOptions(), "foo", &value));
+  ASSERT_EQ(value, "bar2");
+
+  ASSERT_EQ(1, txn->GetNumKeys());
+
+  Status s = txn->Commit();
+  EXPECT_TRUE(s.IsBusy());  // Txn should not commit
+
+  // Verify that transaction did not write anything
+  ASSERT_OK(txn_db->Get(ReadOptions(), "foo", &value));
+  ASSERT_EQ(value, "bar2");
+
+  delete txn;
+}
+
+TEST_P(OptimisticTransactionTest, WriteConflict4) {
+  ASSERT_OK(txn_db->Put(WriteOptions(), "foo", "bar"));
+
+  Transaction* txn = txn_db->BeginTransaction(WriteOptions());
+  ASSERT_NE(txn, nullptr);
+
+  std::string value;
+  ASSERT_OK(txn->GetForUpdate(ReadOptions(), "foo", &value));
+  ASSERT_EQ(value, "bar");
+  ASSERT_OK(txn->Merge("foo", "bar3"));
+
+  // Range delete outside of a transaction should conflict with the previous
+  // merge inside txn
+  auto* dbimpl = static_cast_with_check<DBImpl>(txn_db->GetRootDB());
+  ColumnFamilyHandle* default_cf = dbimpl->DefaultColumnFamily();
+  ASSERT_OK(dbimpl->DeleteRange(WriteOptions(), default_cf, "foo", "foo1"));
+  Status s = txn_db->Get(ReadOptions(), "foo", &value);
+  ASSERT_TRUE(s.IsNotFound());
+
+  ASSERT_EQ(1, txn->GetNumKeys());
+
+  s = txn->Commit();
+  EXPECT_TRUE(s.IsBusy());  // Txn should not commit
+
+  // Verify that transaction did not write anything
+  s = txn_db->Get(ReadOptions(), "foo", &value);
+  ASSERT_TRUE(s.IsNotFound());
+
+  delete txn;
+}
+
+TEST_P(OptimisticTransactionTest, ReadConflictTest) {
+  WriteOptions write_options;
+  ReadOptions read_options, snapshot_read_options;
+  OptimisticTransactionOptions txn_options;
+  std::string value;
+
+  ASSERT_OK(txn_db->Put(write_options, "foo", "bar"));
+  ASSERT_OK(txn_db->Put(write_options, "foo2", "bar"));
+
+  txn_options.set_snapshot = true;
+  Transaction* txn = txn_db->BeginTransaction(write_options, txn_options);
+  ASSERT_NE(txn, nullptr);
+
+  txn->SetSnapshot();
+  snapshot_read_options.snapshot = txn->GetSnapshot();
+
+  ASSERT_OK(txn->GetForUpdate(snapshot_read_options, "foo", &value));
+  ASSERT_EQ(value, "bar");
+
+  // This Put outside of a transaction will conflict with the previous read
+  ASSERT_OK(txn_db->Put(write_options, "foo", "barz"));
+
+  ASSERT_OK(txn_db->Get(read_options, "foo", &value));
+  ASSERT_EQ(value, "barz");
+
+  Status s = txn->Commit();
+  ASSERT_TRUE(s.IsBusy());  // Txn should not commit
+
+  // Verify that transaction did not write anything
+  ASSERT_OK(txn->GetForUpdate(read_options, "foo", &value));
+  ASSERT_EQ(value, "barz");
+  ASSERT_OK(txn->GetForUpdate(read_options, "foo2", &value));
+  ASSERT_EQ(value, "bar");
+
+  delete txn;
+}
+
+TEST_P(OptimisticTransactionTest, TxnOnlyTest) {
+  // Test to make sure transactions work when there are no other writes in an
+  // empty db.
+
+  WriteOptions write_options;
+  ReadOptions read_options;
+  std::string value;
+
+  Transaction* txn = txn_db->BeginTransaction(write_options);
+  ASSERT_NE(txn, nullptr);
+
+  ASSERT_OK(txn->Put("x", "y"));
+
+  ASSERT_OK(txn->Commit());
+
+  delete txn;
+}
+
+TEST_P(OptimisticTransactionTest, FlushTest) {
+  WriteOptions write_options;
+  ReadOptions read_options, snapshot_read_options;
+  std::string value;
+
+  ASSERT_OK(txn_db->Put(write_options, Slice("foo"), Slice("bar")));
+  ASSERT_OK(txn_db->Put(write_options, Slice("foo2"), Slice("bar")));
+
+  Transaction* txn = txn_db->BeginTransaction(write_options);
+  ASSERT_NE(txn, nullptr);
+
+  snapshot_read_options.snapshot = txn->GetSnapshot();
+
+  ASSERT_OK(txn->GetForUpdate(snapshot_read_options, "foo", &value));
+  ASSERT_EQ(value, "bar");
+
+  ASSERT_OK(txn->Put(Slice("foo"), Slice("bar2")));
+
+  ASSERT_OK(txn->GetForUpdate(snapshot_read_options, "foo", &value));
+  ASSERT_EQ(value, "bar2");
+
+  // Put a random key so we have a memtable to flush
+  ASSERT_OK(txn_db->Put(write_options, "dummy", "dummy"));
+
+  // force a memtable flush
+  FlushOptions flush_ops;
+  ASSERT_OK(txn_db->Flush(flush_ops));
+
+  // txn should commit since the flushed table is still in MemtableList History
+  ASSERT_OK(txn->Commit());
+
+  ASSERT_OK(txn_db->Get(read_options, "foo", &value));
+  ASSERT_EQ(value, "bar2");
+
+  delete txn;
+}
+
+TEST_P(OptimisticTransactionTest, FlushTest2) {
+  WriteOptions write_options;
+  ReadOptions read_options, snapshot_read_options;
+  std::string value;
+
+  ASSERT_OK(txn_db->Put(write_options, Slice("foo"), Slice("bar")));
+  ASSERT_OK(txn_db->Put(write_options, Slice("foo2"), Slice("bar")));
+
+  Transaction* txn = txn_db->BeginTransaction(write_options);
+  ASSERT_NE(txn, nullptr);
+
+  snapshot_read_options.snapshot = txn->GetSnapshot();
+
+  ASSERT_OK(txn->GetForUpdate(snapshot_read_options, "foo", &value));
+  ASSERT_EQ(value, "bar");
+
+  ASSERT_OK(txn->Put(Slice("foo"), Slice("bar2")));
+
+  ASSERT_OK(txn->GetForUpdate(snapshot_read_options, "foo", &value));
+  ASSERT_EQ(value, "bar2");
+
+  // Put a random key so we have a MemTable to flush
+  ASSERT_OK(txn_db->Put(write_options, "dummy", "dummy"));
+
+  // force a memtable flush
+  FlushOptions flush_ops;
+  ASSERT_OK(txn_db->Flush(flush_ops));
+
+  // Put a random key so we have a MemTable to flush
+  ASSERT_OK(txn_db->Put(write_options, "dummy", "dummy2"));
+
+  // force a memtable flush
+  ASSERT_OK(txn_db->Flush(flush_ops));
+
+  ASSERT_OK(txn_db->Put(write_options, "dummy", "dummy3"));
+
+  // force a memtable flush
+  // Since our test db has max_write_buffer_number=2, this flush will cause
+  // the first memtable to get purged from the MemtableList history.
+  ASSERT_OK(txn_db->Flush(flush_ops));
+
+  Status s = txn->Commit();
+  // txn should not commit since MemTableList History is not large enough
+  ASSERT_TRUE(s.IsTryAgain());
+
+  ASSERT_OK(txn_db->Get(read_options, "foo", &value));
+  ASSERT_EQ(value, "bar");
+
+  delete txn;
+}
+
+// Trigger the condition where some old memtables are skipped when doing
+// TransactionUtil::CheckKey(), and make sure the result is still correct.
+TEST_P(OptimisticTransactionTest, CheckKeySkipOldMemtable) {
+  const int kAttemptHistoryMemtable = 0;
+  const int kAttemptImmMemTable = 1;
+  for (int attempt = kAttemptHistoryMemtable; attempt <= kAttemptImmMemTable;
+       attempt++) {
+    Reopen();
+
+    WriteOptions write_options;
+    ReadOptions read_options;
+    ReadOptions snapshot_read_options;
+    ReadOptions snapshot_read_options2;
+    std::string value;
+
+    ASSERT_OK(txn_db->Put(write_options, Slice("foo"), Slice("bar")));
+    ASSERT_OK(txn_db->Put(write_options, Slice("foo2"), Slice("bar")));
+
+    Transaction* txn = txn_db->BeginTransaction(write_options);
+    ASSERT_TRUE(txn != nullptr);
+
+    Transaction* txn2 = txn_db->BeginTransaction(write_options);
+    ASSERT_TRUE(txn2 != nullptr);
+
+    snapshot_read_options.snapshot = txn->GetSnapshot();
+    ASSERT_OK(txn->GetForUpdate(snapshot_read_options, "foo", &value));
+    ASSERT_EQ(value, "bar");
+    ASSERT_OK(txn->Put(Slice("foo"), Slice("bar2")));
+
+    snapshot_read_options2.snapshot = txn2->GetSnapshot();
+    ASSERT_OK(txn2->GetForUpdate(snapshot_read_options2, "foo2", &value));
+    ASSERT_EQ(value, "bar");
+    ASSERT_OK(txn2->Put(Slice("foo2"), Slice("bar2")));
+
+    // txn updates "foo" and txn2 updates "foo2", and now a write is
+    // issued for "foo", which conflicts with txn but not txn2
+    ASSERT_OK(txn_db->Put(write_options, "foo", "bar"));
+
+    if (attempt == kAttemptImmMemTable) {
+      // For the second attempt, hold flush from beginning. The memtable
+      // will be switched to immutable after calling TEST_SwitchMemtable()
+      // while CheckKey() is called.
+      ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->LoadDependency(
+          {{"OptimisticTransactionTest.CheckKeySkipOldMemtable",
+            "FlushJob::Start"}});
+      ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->EnableProcessing();
+    }
+
+    // force a memtable flush. The memtable should still be kept
+    FlushOptions flush_ops;
+    if (attempt == kAttemptHistoryMemtable) {
+      ASSERT_OK(txn_db->Flush(flush_ops));
+    } else {
+      ASSERT_EQ(attempt, kAttemptImmMemTable);
+      DBImpl* db_impl = static_cast<DBImpl*>(txn_db->GetRootDB());
+      ASSERT_OK(db_impl->TEST_SwitchMemtable());
+    }
+    uint64_t num_imm_mems;
+    ASSERT_TRUE(txn_db->GetIntProperty(DB::Properties::kNumImmutableMemTable,
+                                       &num_imm_mems));
+    if (attempt == kAttemptHistoryMemtable) {
+      ASSERT_EQ(0, num_imm_mems);
+    } else {
+      ASSERT_EQ(attempt, kAttemptImmMemTable);
+      ASSERT_EQ(1, num_imm_mems);
+    }
+
+    // Put something in active memtable
+    ASSERT_OK(txn_db->Put(write_options, Slice("foo3"), Slice("bar")));
+
+    // Create txn3 after flushing, when this transaction is commited,
+    // only need to check the active memtable
+    Transaction* txn3 = txn_db->BeginTransaction(write_options);
+    ASSERT_TRUE(txn3 != nullptr);
+
+    // Commit both of txn and txn2. txn will conflict but txn2 will
+    // pass. In both ways, both memtables are queried.
+    SetPerfLevel(PerfLevel::kEnableCount);
+
+    get_perf_context()->Reset();
+    Status s = txn->Commit();
+    // We should have checked two memtables
+    ASSERT_EQ(2, get_perf_context()->get_from_memtable_count);
+    // txn should fail because of conflict, even if the memtable
+    // has flushed, because it is still preserved in history.
+    ASSERT_TRUE(s.IsBusy());
+
+    get_perf_context()->Reset();
+    s = txn2->Commit();
+    // We should have checked two memtables
+    ASSERT_EQ(2, get_perf_context()->get_from_memtable_count);
+    ASSERT_TRUE(s.ok());
+
+    ASSERT_OK(txn3->Put(Slice("foo2"), Slice("bar2")));
+    get_perf_context()->Reset();
+    s = txn3->Commit();
+    // txn3 is created after the active memtable is created, so that is the only
+    // memtable to check.
+    ASSERT_EQ(1, get_perf_context()->get_from_memtable_count);
+    ASSERT_TRUE(s.ok());
+
+    TEST_SYNC_POINT("OptimisticTransactionTest.CheckKeySkipOldMemtable");
+    ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->DisableProcessing();
+
+    SetPerfLevel(PerfLevel::kDisable);
+
+    delete txn;
+    delete txn2;
+    delete txn3;
+  }
+}
+
+TEST_P(OptimisticTransactionTest, NoSnapshotTest) {
+  WriteOptions write_options;
+  ReadOptions read_options;
+  std::string value;
+
+  ASSERT_OK(txn_db->Put(write_options, "AAA", "bar"));
+
+  Transaction* txn = txn_db->BeginTransaction(write_options);
+  ASSERT_NE(txn, nullptr);
+
+  // Modify key after transaction start
+  ASSERT_OK(txn_db->Put(write_options, "AAA", "bar1"));
+
+  // Read and write without a snapshot
+  ASSERT_OK(txn->GetForUpdate(read_options, "AAA", &value));
+  ASSERT_EQ(value, "bar1");
+  ASSERT_OK(txn->Put("AAA", "bar2"));
+
+  // Should commit since read/write was done after data changed
+  ASSERT_OK(txn->Commit());
+
+  ASSERT_OK(txn->GetForUpdate(read_options, "AAA", &value));
+  ASSERT_EQ(value, "bar2");
+
+  delete txn;
+}
+
+TEST_P(OptimisticTransactionTest, MultipleSnapshotTest) {
+  WriteOptions write_options;
+  ReadOptions read_options, snapshot_read_options;
+  std::string value;
+
+  ASSERT_OK(txn_db->Put(write_options, "AAA", "bar"));
+  ASSERT_OK(txn_db->Put(write_options, "BBB", "bar"));
+  ASSERT_OK(txn_db->Put(write_options, "CCC", "bar"));
+
+  Transaction* txn = txn_db->BeginTransaction(write_options);
+  ASSERT_NE(txn, nullptr);
+
+  ASSERT_OK(txn_db->Put(write_options, "AAA", "bar1"));
+
+  // Read and write without a snapshot
+  ASSERT_OK(txn->GetForUpdate(read_options, "AAA", &value));
+  ASSERT_EQ(value, "bar1");
+  ASSERT_OK(txn->Put("AAA", "bar2"));
+
+  // Modify BBB before snapshot is taken
+  ASSERT_OK(txn_db->Put(write_options, "BBB", "bar1"));
+
+  txn->SetSnapshot();
+  snapshot_read_options.snapshot = txn->GetSnapshot();
+
+  // Read and write with snapshot
+  ASSERT_OK(txn->GetForUpdate(snapshot_read_options, "BBB", &value));
+  ASSERT_EQ(value, "bar1");
+  ASSERT_OK(txn->Put("BBB", "bar2"));
+
+  ASSERT_OK(txn_db->Put(write_options, "CCC", "bar1"));
+
+  // Set a new snapshot
+  txn->SetSnapshot();
+  snapshot_read_options.snapshot = txn->GetSnapshot();
+
+  // Read and write with snapshot
+  ASSERT_OK(txn->GetForUpdate(snapshot_read_options, "CCC", &value));
+  ASSERT_EQ(value, "bar1");
+  ASSERT_OK(txn->Put("CCC", "bar2"));
+
+  ASSERT_OK(txn->GetForUpdate(read_options, "AAA", &value));
+  ASSERT_EQ(value, "bar2");
+  ASSERT_OK(txn->GetForUpdate(read_options, "BBB", &value));
+  ASSERT_EQ(value, "bar2");
+  ASSERT_OK(txn->GetForUpdate(read_options, "CCC", &value));
+  ASSERT_EQ(value, "bar2");
+
+  ASSERT_OK(txn_db->Get(read_options, "AAA", &value));
+  ASSERT_EQ(value, "bar1");
+  ASSERT_OK(txn_db->Get(read_options, "BBB", &value));
+  ASSERT_EQ(value, "bar1");
+  ASSERT_OK(txn_db->Get(read_options, "CCC", &value));
+  ASSERT_EQ(value, "bar1");
+
+  ASSERT_OK(txn->Commit());
+
+  ASSERT_OK(txn_db->Get(read_options, "AAA", &value));
+  ASSERT_EQ(value, "bar2");
+  ASSERT_OK(txn_db->Get(read_options, "BBB", &value));
+  ASSERT_EQ(value, "bar2");
+  ASSERT_OK(txn_db->Get(read_options, "CCC", &value));
+  ASSERT_EQ(value, "bar2");
+
+  // verify that we track multiple writes to the same key at different snapshots
+  delete txn;
+  txn = txn_db->BeginTransaction(write_options);
+
+  // Potentially conflicting writes
+  ASSERT_OK(txn_db->Put(write_options, "ZZZ", "zzz"));
+  ASSERT_OK(txn_db->Put(write_options, "XXX", "xxx"));
+
+  txn->SetSnapshot();
+
+  OptimisticTransactionOptions txn_options;
+  txn_options.set_snapshot = true;
+  Transaction* txn2 = txn_db->BeginTransaction(write_options, txn_options);
+  txn2->SetSnapshot();
+
+  // This should not conflict in txn since the snapshot is later than the
+  // previous write (spoiler alert:  it will later conflict with txn2).
+  ASSERT_OK(txn->Put("ZZZ", "zzzz"));
+  ASSERT_OK(txn->Commit());
+
+  delete txn;
+
+  // This will conflict since the snapshot is earlier than another write to ZZZ
+  ASSERT_OK(txn2->Put("ZZZ", "xxxxx"));
+
+  Status s = txn2->Commit();
+  ASSERT_TRUE(s.IsBusy());
+
+  delete txn2;
+}
+
+TEST_P(OptimisticTransactionTest, ColumnFamiliesTest) {
+  WriteOptions write_options;
+  ReadOptions read_options, snapshot_read_options;
+  OptimisticTransactionOptions txn_options;
+  std::string value;
+
+  ColumnFamilyHandle *cfa, *cfb;
+  ColumnFamilyOptions cf_options;
+
+  // Create 2 new column families
+  ASSERT_OK(txn_db->CreateColumnFamily(cf_options, "CFA", &cfa));
+  ASSERT_OK(txn_db->CreateColumnFamily(cf_options, "CFB", &cfb));
+
+  delete cfa;
+  delete cfb;
+  delete txn_db;
+  txn_db = nullptr;
+
+  // open DB with three column families
+  std::vector<ColumnFamilyDescriptor> column_families;
+  // have to open default column family
+  column_families.push_back(
+      ColumnFamilyDescriptor(kDefaultColumnFamilyName, ColumnFamilyOptions()));
+  // open the new column families
+  column_families.push_back(
+      ColumnFamilyDescriptor("CFA", ColumnFamilyOptions()));
+  column_families.push_back(
+      ColumnFamilyDescriptor("CFB", ColumnFamilyOptions()));
+  std::vector<ColumnFamilyHandle*> handles;
+  ASSERT_OK(OptimisticTransactionDB::Open(options, dbname, column_families,
+                                          &handles, &txn_db));
+  assert(txn_db != nullptr);
+  ASSERT_NE(txn_db, nullptr);
+
+  Transaction* txn = txn_db->BeginTransaction(write_options);
+  ASSERT_NE(txn, nullptr);
+
+  txn->SetSnapshot();
+  snapshot_read_options.snapshot = txn->GetSnapshot();
+
+  txn_options.set_snapshot = true;
+  Transaction* txn2 = txn_db->BeginTransaction(write_options, txn_options);
+  ASSERT_TRUE(txn2);
+
+  // Write some data to the db
+  WriteBatch batch;
+  ASSERT_OK(batch.Put("foo", "foo"));
+  ASSERT_OK(batch.Put(handles[1], "AAA", "bar"));
+  ASSERT_OK(batch.Put(handles[1], "AAAZZZ", "bar"));
+  ASSERT_OK(txn_db->Write(write_options, &batch));
+  ASSERT_OK(txn_db->Delete(write_options, handles[1], "AAAZZZ"));
+
+  // These keys do no conflict with existing writes since they're in
+  // different column families
+  ASSERT_OK(txn->Delete("AAA"));
+  Status s =
+      txn->GetForUpdate(snapshot_read_options, handles[1], "foo", &value);
+  ASSERT_TRUE(s.IsNotFound());
+  Slice key_slice("AAAZZZ");
+  Slice value_slices[2] = {Slice("bar"), Slice("bar")};
+  ASSERT_OK(txn->Put(handles[2], SliceParts(&key_slice, 1),
+                     SliceParts(value_slices, 2)));
+
+  ASSERT_EQ(3, txn->GetNumKeys());
+
+  // Txn should commit
+  ASSERT_OK(txn->Commit());
+  s = txn_db->Get(read_options, "AAA", &value);
+  ASSERT_TRUE(s.IsNotFound());
+  s = txn_db->Get(read_options, handles[2], "AAAZZZ", &value);
+  ASSERT_EQ(value, "barbar");
+
+  Slice key_slices[3] = {Slice("AAA"), Slice("ZZ"), Slice("Z")};
+  Slice value_slice("barbarbar");
+  // This write will cause a conflict with the earlier batch write
+  ASSERT_OK(txn2->Put(handles[1], SliceParts(key_slices, 3),
+                      SliceParts(&value_slice, 1)));
+
+  ASSERT_OK(txn2->Delete(handles[2], "XXX"));
+  ASSERT_OK(txn2->Delete(handles[1], "XXX"));
+  s = txn2->GetForUpdate(snapshot_read_options, handles[1], "AAA", &value);
+  ASSERT_TRUE(s.IsNotFound());
+
+  // Verify txn did not commit
+  s = txn2->Commit();
+  ASSERT_TRUE(s.IsBusy());
+  s = txn_db->Get(read_options, handles[1], "AAAZZZ", &value);
+  ASSERT_TRUE(s.IsNotFound());
+  ASSERT_EQ(value, "barbar");
+
+  delete txn;
+  delete txn2;
+
+  txn = txn_db->BeginTransaction(write_options, txn_options);
+  snapshot_read_options.snapshot = txn->GetSnapshot();
+
+  txn2 = txn_db->BeginTransaction(write_options, txn_options);
+  ASSERT_NE(txn, nullptr);
+
+  std::vector<ColumnFamilyHandle*> multiget_cfh = {handles[1], handles[2],
+                                                   handles[0], handles[2]};
+  std::vector<Slice> multiget_keys = {"AAA", "AAAZZZ", "foo", "foo"};
+  std::vector<std::string> values(4);
+
+  std::vector<Status> results = txn->MultiGetForUpdate(
+      snapshot_read_options, multiget_cfh, multiget_keys, &values);
+  ASSERT_OK(results[0]);
+  ASSERT_OK(results[1]);
+  ASSERT_OK(results[2]);
+  ASSERT_TRUE(results[3].IsNotFound());
+  ASSERT_EQ(values[0], "bar");
+  ASSERT_EQ(values[1], "barbar");
+  ASSERT_EQ(values[2], "foo");
+
+  ASSERT_OK(txn->Delete(handles[2], "ZZZ"));
+  ASSERT_OK(txn->Put(handles[2], "ZZZ", "YYY"));
+  ASSERT_OK(txn->Put(handles[2], "ZZZ", "YYYY"));
+  ASSERT_OK(txn->Delete(handles[2], "ZZZ"));
+  ASSERT_OK(txn->Put(handles[2], "AAAZZZ", "barbarbar"));
+
+  ASSERT_EQ(5, txn->GetNumKeys());
+
+  // Txn should commit
+  ASSERT_OK(txn->Commit());
+  s = txn_db->Get(read_options, handles[2], "ZZZ", &value);
+  ASSERT_TRUE(s.IsNotFound());
+
+  // Put a key which will conflict with the next txn using the previous snapshot
+  ASSERT_OK(txn_db->Put(write_options, handles[2], "foo", "000"));
+
+  results = txn2->MultiGetForUpdate(snapshot_read_options, multiget_cfh,
+                                    multiget_keys, &values);
+  ASSERT_OK(results[0]);
+  ASSERT_OK(results[1]);
+  ASSERT_OK(results[2]);
+  ASSERT_TRUE(results[3].IsNotFound());
+  ASSERT_EQ(values[0], "bar");
+  ASSERT_EQ(values[1], "barbar");
+  ASSERT_EQ(values[2], "foo");
+
+  // Verify Txn Did not Commit
+  s = txn2->Commit();
+  ASSERT_TRUE(s.IsBusy());
+
+  s = txn_db->DropColumnFamily(handles[1]);
+  ASSERT_OK(s);
+  s = txn_db->DropColumnFamily(handles[2]);
+  ASSERT_OK(s);
+
+  delete txn;
+  delete txn2;
+
+  for (auto handle : handles) {
+    delete handle;
+  }
+}
+
+TEST_P(OptimisticTransactionTest, EmptyTest) {
+  WriteOptions write_options;
+  ReadOptions read_options;
+  std::string value;
+
+  ASSERT_OK(txn_db->Put(write_options, "aaa", "aaa"));
+
+  Transaction* txn = txn_db->BeginTransaction(write_options);
+  ASSERT_OK(txn->Commit());
+  delete txn;
+
+  txn = txn_db->BeginTransaction(write_options);
+  ASSERT_OK(txn->Rollback());
+  delete txn;
+
+  txn = txn_db->BeginTransaction(write_options);
+  ASSERT_OK(txn->GetForUpdate(read_options, "aaa", &value));
+  ASSERT_EQ(value, "aaa");
+
+  ASSERT_OK(txn->Commit());
+  delete txn;
+
+  txn = txn_db->BeginTransaction(write_options);
+  txn->SetSnapshot();
+  ASSERT_OK(txn->GetForUpdate(read_options, "aaa", &value));
+  ASSERT_EQ(value, "aaa");
+
+  ASSERT_OK(txn_db->Put(write_options, "aaa", "xxx"));
+  Status s = txn->Commit();
+  ASSERT_TRUE(s.IsBusy());
+  delete txn;
+}
+
+TEST_P(OptimisticTransactionTest, PredicateManyPreceders) {
+  WriteOptions write_options;
+  ReadOptions read_options1, read_options2;
+  OptimisticTransactionOptions txn_options;
+  std::string value;
+
+  txn_options.set_snapshot = true;
+  Transaction* txn1 = txn_db->BeginTransaction(write_options, txn_options);
+  read_options1.snapshot = txn1->GetSnapshot();
+
+  Transaction* txn2 = txn_db->BeginTransaction(write_options);
+  txn2->SetSnapshot();
+  read_options2.snapshot = txn2->GetSnapshot();
+
+  std::vector<Slice> multiget_keys = {"1", "2", "3"};
+  std::vector<std::string> multiget_values;
+
+  std::vector<Status> results =
+      txn1->MultiGetForUpdate(read_options1, multiget_keys, &multiget_values);
+  ASSERT_TRUE(results[0].IsNotFound());
+  ASSERT_TRUE(results[1].IsNotFound());
+  ASSERT_TRUE(results[2].IsNotFound());
+
+  ASSERT_OK(txn2->Put("2", "x"));
+
+  ASSERT_OK(txn2->Commit());
+
+  multiget_values.clear();
+  results =
+      txn1->MultiGetForUpdate(read_options1, multiget_keys, &multiget_values);
+  ASSERT_TRUE(results[0].IsNotFound());
+  ASSERT_TRUE(results[1].IsNotFound());
+  ASSERT_TRUE(results[2].IsNotFound());
+
+  // should not commit since txn2 wrote a key txn has read
+  Status s = txn1->Commit();
+  ASSERT_TRUE(s.IsBusy());
+
+  delete txn1;
+  delete txn2;
+
+  txn1 = txn_db->BeginTransaction(write_options, txn_options);
+  read_options1.snapshot = txn1->GetSnapshot();
+
+  txn2 = txn_db->BeginTransaction(write_options, txn_options);
+  read_options2.snapshot = txn2->GetSnapshot();
+
+  ASSERT_OK(txn1->Put("4", "x"));
+
+  ASSERT_OK(txn2->Delete("4"));
+
+  // txn1 can commit since txn2's delete hasn't happened yet (it's just batched)
+  ASSERT_OK(txn1->Commit());
+
+  s = txn2->GetForUpdate(read_options2, "4", &value);
+  ASSERT_TRUE(s.IsNotFound());
+
+  // txn2 cannot commit since txn1 changed "4"
+  s = txn2->Commit();
+  ASSERT_TRUE(s.IsBusy());
+
+  delete txn1;
+  delete txn2;
+}
+
+TEST_P(OptimisticTransactionTest, LostUpdate) {
+  WriteOptions write_options;
+  ReadOptions read_options, read_options1, read_options2;
+  OptimisticTransactionOptions txn_options;
+  std::string value;
+
+  // Test 2 transactions writing to the same key in multiple orders and
+  // with/without snapshots
+
+  Transaction* txn1 = txn_db->BeginTransaction(write_options);
+  Transaction* txn2 = txn_db->BeginTransaction(write_options);
+
+  ASSERT_OK(txn1->Put("1", "1"));
+  ASSERT_OK(txn2->Put("1", "2"));
+
+  ASSERT_OK(txn1->Commit());
+
+  Status s = txn2->Commit();
+  ASSERT_TRUE(s.IsBusy());
+
+  delete txn1;
+  delete txn2;
+
+  txn_options.set_snapshot = true;
+  txn1 = txn_db->BeginTransaction(write_options, txn_options);
+  read_options1.snapshot = txn1->GetSnapshot();
+
+  txn2 = txn_db->BeginTransaction(write_options, txn_options);
+  read_options2.snapshot = txn2->GetSnapshot();
+
+  ASSERT_OK(txn1->Put("1", "3"));
+  ASSERT_OK(txn2->Put("1", "4"));
+
+  ASSERT_OK(txn1->Commit());
+
+  s = txn2->Commit();
+  ASSERT_TRUE(s.IsBusy());
+
+  delete txn1;
+  delete txn2;
+
+  txn1 = txn_db->BeginTransaction(write_options, txn_options);
+  read_options1.snapshot = txn1->GetSnapshot();
+
+  txn2 = txn_db->BeginTransaction(write_options, txn_options);
+  read_options2.snapshot = txn2->GetSnapshot();
+
+  ASSERT_OK(txn1->Put("1", "5"));
+  ASSERT_OK(txn1->Commit());
+
+  ASSERT_OK(txn2->Put("1", "6"));
+  s = txn2->Commit();
+  ASSERT_TRUE(s.IsBusy());
+
+  delete txn1;
+  delete txn2;
+
+  txn1 = txn_db->BeginTransaction(write_options, txn_options);
+  read_options1.snapshot = txn1->GetSnapshot();
+
+  txn2 = txn_db->BeginTransaction(write_options, txn_options);
+  read_options2.snapshot = txn2->GetSnapshot();
+
+  ASSERT_OK(txn1->Put("1", "5"));
+  ASSERT_OK(txn1->Commit());
+
+  txn2->SetSnapshot();
+  ASSERT_OK(txn2->Put("1", "6"));
+  ASSERT_OK(txn2->Commit());
+
+  delete txn1;
+  delete txn2;
+
+  txn1 = txn_db->BeginTransaction(write_options);
+  txn2 = txn_db->BeginTransaction(write_options);
+
+  ASSERT_OK(txn1->Put("1", "7"));
+  ASSERT_OK(txn1->Commit());
+
+  ASSERT_OK(txn2->Put("1", "8"));
+  ASSERT_OK(txn2->Commit());
+
+  delete txn1;
+  delete txn2;
+
+  ASSERT_OK(txn_db->Get(read_options, "1", &value));
+  ASSERT_EQ(value, "8");
+}
+
+TEST_P(OptimisticTransactionTest, UntrackedWrites) {
+  WriteOptions write_options;
+  ReadOptions read_options;
+  std::string value;
+  Status s;
+
+  // Verify transaction rollback works for untracked keys.
+  Transaction* txn = txn_db->BeginTransaction(write_options);
+  ASSERT_OK(txn->PutUntracked("untracked", "0"));
+  ASSERT_OK(txn->Rollback());
+  s = txn_db->Get(read_options, "untracked", &value);
+  ASSERT_TRUE(s.IsNotFound());
+
+  delete txn;
+  txn = txn_db->BeginTransaction(write_options);
+
+  ASSERT_OK(txn->Put("tracked", "1"));
+  ASSERT_OK(txn->PutUntracked("untracked", "1"));
+  ASSERT_OK(txn->MergeUntracked("untracked", "2"));
+  ASSERT_OK(txn->DeleteUntracked("untracked"));
+
+  // Write to the untracked key outside of the transaction and verify
+  // it doesn't prevent the transaction from committing.
+  ASSERT_OK(txn_db->Put(write_options, "untracked", "x"));
+
+  ASSERT_OK(txn->Commit());
+
+  s = txn_db->Get(read_options, "untracked", &value);
+  ASSERT_TRUE(s.IsNotFound());
+
+  delete txn;
+  txn = txn_db->BeginTransaction(write_options);
+
+  ASSERT_OK(txn->Put("tracked", "10"));
+  ASSERT_OK(txn->PutUntracked("untracked", "A"));
+
+  // Write to tracked key outside of the transaction and verify that the
+  // untracked keys are not written when the commit fails.
+  ASSERT_OK(txn_db->Delete(write_options, "tracked"));
+
+  s = txn->Commit();
+  ASSERT_TRUE(s.IsBusy());
+
+  s = txn_db->Get(read_options, "untracked", &value);
+  ASSERT_TRUE(s.IsNotFound());
+
+  delete txn;
+}
+
+TEST_P(OptimisticTransactionTest, IteratorTest) {
+  WriteOptions write_options;
+  ReadOptions read_options, snapshot_read_options;
+  OptimisticTransactionOptions txn_options;
+  std::string value;
+
+  // Write some keys to the db
+  ASSERT_OK(txn_db->Put(write_options, "A", "a"));
+  ASSERT_OK(txn_db->Put(write_options, "G", "g"));
+  ASSERT_OK(txn_db->Put(write_options, "F", "f"));
+  ASSERT_OK(txn_db->Put(write_options, "C", "c"));
+  ASSERT_OK(txn_db->Put(write_options, "D", "d"));
+
+  Transaction* txn = txn_db->BeginTransaction(write_options);
+  ASSERT_NE(txn, nullptr);
+
+  // Write some keys in a txn
+  ASSERT_OK(txn->Put("B", "b"));
+  ASSERT_OK(txn->Put("H", "h"));
+  ASSERT_OK(txn->Delete("D"));
+  ASSERT_OK(txn->Put("E", "e"));
+
+  txn->SetSnapshot();
+  const Snapshot* snapshot = txn->GetSnapshot();
+
+  // Write some keys to the db after the snapshot
+  ASSERT_OK(txn_db->Put(write_options, "BB", "xx"));
+  ASSERT_OK(txn_db->Put(write_options, "C", "xx"));
+
+  read_options.snapshot = snapshot;
+  Iterator* iter = txn->GetIterator(read_options);
+  ASSERT_OK(iter->status());
+  iter->SeekToFirst();
+
+  // Read all keys via iter and lock them all
+  std::string results[] = {"a", "b", "c", "e", "f", "g", "h"};
+  for (int i = 0; i < 7; i++) {
+    ASSERT_OK(iter->status());
+    ASSERT_TRUE(iter->Valid());
+    ASSERT_EQ(results[i], iter->value().ToString());
+
+    ASSERT_OK(txn->GetForUpdate(read_options, iter->key(), nullptr));
+
+    iter->Next();
+  }
+  ASSERT_FALSE(iter->Valid());
+
+  iter->Seek("G");
+  ASSERT_OK(iter->status());
+  ASSERT_TRUE(iter->Valid());
+  ASSERT_EQ("g", iter->value().ToString());
+
+  iter->Prev();
+  ASSERT_OK(iter->status());
+  ASSERT_TRUE(iter->Valid());
+  ASSERT_EQ("f", iter->value().ToString());
+
+  iter->Seek("D");
+  ASSERT_OK(iter->status());
+  ASSERT_TRUE(iter->Valid());
+  ASSERT_EQ("e", iter->value().ToString());
+
+  iter->Seek("C");
+  ASSERT_OK(iter->status());
+  ASSERT_TRUE(iter->Valid());
+  ASSERT_EQ("c", iter->value().ToString());
+
+  iter->Next();
+  ASSERT_OK(iter->status());
+  ASSERT_TRUE(iter->Valid());
+  ASSERT_EQ("e", iter->value().ToString());
+
+  iter->Seek("");
+  ASSERT_OK(iter->status());
+  ASSERT_TRUE(iter->Valid());
+  ASSERT_EQ("a", iter->value().ToString());
+
+  iter->Seek("X");
+  ASSERT_OK(iter->status());
+  ASSERT_FALSE(iter->Valid());
+
+  iter->SeekToLast();
+  ASSERT_OK(iter->status());
+  ASSERT_TRUE(iter->Valid());
+  ASSERT_EQ("h", iter->value().ToString());
+
+  // key "C" was modified in the db after txn's snapshot.  txn will not commit.
+  Status s = txn->Commit();
+  ASSERT_TRUE(s.IsBusy());
+
+  delete iter;
+  delete txn;
+}
+
+TEST_P(OptimisticTransactionTest, DeleteRangeSupportTest) {
+  // `OptimisticTransactionDB` does not allow range deletion in any API.
+  ASSERT_TRUE(
+      txn_db
+          ->DeleteRange(WriteOptions(), txn_db->DefaultColumnFamily(), "a", "b")
+          .IsNotSupported());
+  WriteBatch wb;
+  ASSERT_OK(wb.DeleteRange("a", "b"));
+  ASSERT_NOK(txn_db->Write(WriteOptions(), &wb));
+}
+
+TEST_P(OptimisticTransactionTest, SavepointTest) {
+  WriteOptions write_options;
+  ReadOptions read_options, snapshot_read_options;
+  OptimisticTransactionOptions txn_options;
+  std::string value;
+
+  Transaction* txn = txn_db->BeginTransaction(write_options);
+  ASSERT_NE(txn, nullptr);
+
+  Status s = txn->RollbackToSavePoint();
+  ASSERT_TRUE(s.IsNotFound());
+
+  txn->SetSavePoint();  // 1
+
+  ASSERT_OK(txn->RollbackToSavePoint());  // Rollback to beginning of txn
+  s = txn->RollbackToSavePoint();
+  ASSERT_TRUE(s.IsNotFound());
+
+  ASSERT_OK(txn->Put("B", "b"));
+
+  ASSERT_OK(txn->Commit());
+
+  ASSERT_OK(txn_db->Get(read_options, "B", &value));
+  ASSERT_EQ("b", value);
+
+  delete txn;
+  txn = txn_db->BeginTransaction(write_options);
+  ASSERT_NE(txn, nullptr);
+
+  ASSERT_OK(txn->Put("A", "a"));
+  ASSERT_OK(txn->Put("B", "bb"));
+  ASSERT_OK(txn->Put("C", "c"));
+
+  txn->SetSavePoint();  // 2
+
+  ASSERT_OK(txn->Delete("B"));
+  ASSERT_OK(txn->Put("C", "cc"));
+  ASSERT_OK(txn->Put("D", "d"));
+
+  ASSERT_OK(txn->RollbackToSavePoint());  // Rollback to 2
+
+  ASSERT_OK(txn->Get(read_options, "A", &value));
+  ASSERT_EQ("a", value);
+  ASSERT_OK(txn->Get(read_options, "B", &value));
+  ASSERT_EQ("bb", value);
+  ASSERT_OK(txn->Get(read_options, "C", &value));
+  ASSERT_EQ("c", value);
+  s = txn->Get(read_options, "D", &value);
+  ASSERT_TRUE(s.IsNotFound());
+
+  ASSERT_OK(txn->Put("A", "a"));
+  ASSERT_OK(txn->Put("E", "e"));
+
+  // Rollback to beginning of txn
+  s = txn->RollbackToSavePoint();
+  ASSERT_TRUE(s.IsNotFound());
+  ASSERT_OK(txn->Rollback());
+
+  s = txn->Get(read_options, "A", &value);
+  ASSERT_TRUE(s.IsNotFound());
+  ASSERT_OK(txn->Get(read_options, "B", &value));
+  ASSERT_EQ("b", value);
+  s = txn->Get(read_options, "D", &value);
+  ASSERT_TRUE(s.IsNotFound());
+  s = txn->Get(read_options, "D", &value);
+  ASSERT_TRUE(s.IsNotFound());
+  s = txn->Get(read_options, "E", &value);
+  ASSERT_TRUE(s.IsNotFound());
+
+  ASSERT_OK(txn->Put("A", "aa"));
+  ASSERT_OK(txn->Put("F", "f"));
+
+  txn->SetSavePoint();  // 3
+  txn->SetSavePoint();  // 4
+
+  ASSERT_OK(txn->Put("G", "g"));
+  ASSERT_OK(txn->Delete("F"));
+  ASSERT_OK(txn->Delete("B"));
+
+  ASSERT_OK(txn->Get(read_options, "A", &value));
+  ASSERT_EQ("aa", value);
+
+  s = txn->Get(read_options, "F", &value);
+  ASSERT_TRUE(s.IsNotFound());
+
+  s = txn->Get(read_options, "B", &value);
+  ASSERT_TRUE(s.IsNotFound());
+
+  ASSERT_OK(txn->RollbackToSavePoint());  // Rollback to 3
+
+  ASSERT_OK(txn->Get(read_options, "F", &value));
+  ASSERT_EQ("f", value);
+
+  s = txn->Get(read_options, "G", &value);
+  ASSERT_TRUE(s.IsNotFound());
+
+  ASSERT_OK(txn->Commit());
+
+  ASSERT_OK(txn_db->Get(read_options, "F", &value));
+  ASSERT_EQ("f", value);
+
+  s = txn_db->Get(read_options, "G", &value);
+  ASSERT_TRUE(s.IsNotFound());
+
+  ASSERT_OK(txn_db->Get(read_options, "A", &value));
+  ASSERT_EQ("aa", value);
+
+  ASSERT_OK(txn_db->Get(read_options, "B", &value));
+  ASSERT_EQ("b", value);
+
+  s = txn_db->Get(read_options, "C", &value);
+  ASSERT_TRUE(s.IsNotFound());
+
+  s = txn_db->Get(read_options, "D", &value);
+  ASSERT_TRUE(s.IsNotFound());
+
+  s = txn_db->Get(read_options, "E", &value);
+  ASSERT_TRUE(s.IsNotFound());
+
+  delete txn;
+}
+
+TEST_P(OptimisticTransactionTest, UndoGetForUpdateTest) {
+  WriteOptions write_options;
+  ReadOptions read_options, snapshot_read_options;
+  OptimisticTransactionOptions txn_options;
+  std::string value;
+
+  ASSERT_OK(txn_db->Put(write_options, "A", ""));
+
+  Transaction* txn1 = txn_db->BeginTransaction(write_options);
+  ASSERT_TRUE(txn1);
+
+  ASSERT_OK(txn1->GetForUpdate(read_options, "A", &value));
+
+  txn1->UndoGetForUpdate("A");
+
+  Transaction* txn2 = txn_db->BeginTransaction(write_options);
+  txn2->Put("A", "x");
+  ASSERT_OK(txn2->Commit());
+  delete txn2;
+
+  // Verify that txn1 can commit since A isn't conflict checked
+  ASSERT_OK(txn1->Commit());
+  delete txn1;
+
+  txn1 = txn_db->BeginTransaction(write_options);
+  ASSERT_OK(txn1->Put("A", "a"));
+
+  ASSERT_OK(txn1->GetForUpdate(read_options, "A", &value));
+
+  txn1->UndoGetForUpdate("A");
+
+  txn2 = txn_db->BeginTransaction(write_options);
+  ASSERT_OK(txn2->Put("A", "x"));
+  ASSERT_OK(txn2->Commit());
+  delete txn2;
+
+  // Verify that txn1 cannot commit since A will still be conflict checked
+  Status s = txn1->Commit();
+  ASSERT_TRUE(s.IsBusy());
+  delete txn1;
+
+  txn1 = txn_db->BeginTransaction(write_options);
+
+  ASSERT_OK(txn1->GetForUpdate(read_options, "A", &value));
+  ASSERT_OK(txn1->GetForUpdate(read_options, "A", &value));
+
+  txn1->UndoGetForUpdate("A");
+
+  txn2 = txn_db->BeginTransaction(write_options);
+  ASSERT_OK(txn2->Put("A", "x"));
+  ASSERT_OK(txn2->Commit());
+  delete txn2;
+
+  // Verify that txn1 cannot commit since A will still be conflict checked
+  s = txn1->Commit();
+  ASSERT_TRUE(s.IsBusy());
+  delete txn1;
+
+  txn1 = txn_db->BeginTransaction(write_options);
+
+  ASSERT_OK(txn1->GetForUpdate(read_options, "A", &value));
+  ASSERT_OK(txn1->GetForUpdate(read_options, "A", &value));
+
+  txn1->UndoGetForUpdate("A");
+  txn1->UndoGetForUpdate("A");
+
+  txn2 = txn_db->BeginTransaction(write_options);
+  ASSERT_OK(txn2->Put("A", "x"));
+  ASSERT_OK(txn2->Commit());
+  delete txn2;
+
+  // Verify that txn1 can commit since A isn't conflict checked
+  ASSERT_OK(txn1->Commit());
+  delete txn1;
+
+  txn1 = txn_db->BeginTransaction(write_options);
+
+  ASSERT_OK(txn1->GetForUpdate(read_options, "A", &value));
+
+  txn1->SetSavePoint();
+  txn1->UndoGetForUpdate("A");
+
+  txn2 = txn_db->BeginTransaction(write_options);
+  ASSERT_OK(txn2->Put("A", "x"));
+  ASSERT_OK(txn2->Commit());
+  delete txn2;
+
+  // Verify that txn1 cannot commit since A will still be conflict checked
+  s = txn1->Commit();
+  ASSERT_TRUE(s.IsBusy());
+  delete txn1;
+
+  txn1 = txn_db->BeginTransaction(write_options);
+
+  ASSERT_OK(txn1->GetForUpdate(read_options, "A", &value));
+
+  txn1->SetSavePoint();
+  ASSERT_OK(txn1->GetForUpdate(read_options, "A", &value));
+  txn1->UndoGetForUpdate("A");
+
+  txn2 = txn_db->BeginTransaction(write_options);
+  ASSERT_OK(txn2->Put("A", "x"));
+  ASSERT_OK(txn2->Commit());
+  delete txn2;
+
+  // Verify that txn1 cannot commit since A will still be conflict checked
+  s = txn1->Commit();
+  ASSERT_TRUE(s.IsBusy());
+  delete txn1;
+
+  txn1 = txn_db->BeginTransaction(write_options);
+
+  ASSERT_OK(txn1->GetForUpdate(read_options, "A", &value));
+
+  txn1->SetSavePoint();
+  ASSERT_OK(txn1->GetForUpdate(read_options, "A", &value));
+  txn1->UndoGetForUpdate("A");
+
+  ASSERT_OK(txn1->RollbackToSavePoint());
+  txn1->UndoGetForUpdate("A");
+
+  txn2 = txn_db->BeginTransaction(write_options);
+  ASSERT_OK(txn2->Put("A", "x"));
+  ASSERT_OK(txn2->Commit());
+  delete txn2;
+
+  // Verify that txn1 can commit since A isn't conflict checked
+  ASSERT_OK(txn1->Commit());
+  delete txn1;
+}
+
+namespace {
+Status OptimisticTransactionStressTestInserter(OptimisticTransactionDB* db,
+                                               const size_t num_transactions,
+                                               const size_t num_sets,
+                                               const size_t num_keys_per_set) {
+  size_t seed = std::hash<std::thread::id>()(std::this_thread::get_id());
+  Random64 _rand(seed);
+  WriteOptions write_options;
+  ReadOptions read_options;
+  OptimisticTransactionOptions txn_options;
+  txn_options.set_snapshot = true;
+
+  RandomTransactionInserter inserter(&_rand, write_options, read_options,
+                                     num_keys_per_set,
+                                     static_cast<uint16_t>(num_sets));
+
+  for (size_t t = 0; t < num_transactions; t++) {
+    bool success = inserter.OptimisticTransactionDBInsert(db, txn_options);
+    if (!success) {
+      // unexpected failure
+      return inserter.GetLastStatus();
+    }
+  }
+
+  inserter.GetLastStatus().PermitUncheckedError();
+
+  // Make sure at least some of the transactions succeeded.  It's ok if
+  // some failed due to write-conflicts.
+  if (inserter.GetFailureCount() > num_transactions / 2) {
+    return Status::TryAgain("Too many transactions failed! " +
+                            std::to_string(inserter.GetFailureCount()) + " / " +
+                            std::to_string(num_transactions));
+  }
+
+  return Status::OK();
+}
+}  // namespace
+
+TEST_P(OptimisticTransactionTest, OptimisticTransactionStressTest) {
+  const size_t num_threads = 4;
+  const size_t num_transactions_per_thread = 10000;
+  const size_t num_sets = 3;
+  const size_t num_keys_per_set = 100;
+  // Setting the key-space to be 100 keys should cause enough write-conflicts
+  // to make this test interesting.
+
+  std::vector<port::Thread> threads;
+
+  std::function<void()> call_inserter = [&] {
+    ASSERT_OK(OptimisticTransactionStressTestInserter(
+        txn_db, num_transactions_per_thread, num_sets, num_keys_per_set));
+  };
+
+  // Create N threads that use RandomTransactionInserter to write
+  // many transactions.
+  for (uint32_t i = 0; i < num_threads; i++) {
+    threads.emplace_back(call_inserter);
+  }
+
+  // Wait for all threads to run
+  for (auto& t : threads) {
+    t.join();
+  }
+
+  // Verify that data is consistent
+  Status s = RandomTransactionInserter::Verify(txn_db, num_sets);
+  ASSERT_OK(s);
+}
+
+TEST_P(OptimisticTransactionTest, SequenceNumberAfterRecoverTest) {
+  WriteOptions write_options;
+  OptimisticTransactionOptions transaction_options;
+
+  Transaction* transaction(
+      txn_db->BeginTransaction(write_options, transaction_options));
+  Status s = transaction->Put("foo", "val");
+  ASSERT_OK(s);
+  s = transaction->Put("foo2", "val");
+  ASSERT_OK(s);
+  s = transaction->Put("foo3", "val");
+  ASSERT_OK(s);
+  s = transaction->Commit();
+  ASSERT_OK(s);
+  delete transaction;
+
+  Reopen();
+  transaction = txn_db->BeginTransaction(write_options, transaction_options);
+  s = transaction->Put("bar", "val");
+  ASSERT_OK(s);
+  s = transaction->Put("bar2", "val");
+  ASSERT_OK(s);
+  s = transaction->Commit();
+  ASSERT_OK(s);
+
+  delete transaction;
+}
+
+TEST_P(OptimisticTransactionTest, TimestampedSnapshotMissingCommitTs) {
+  std::unique_ptr<Transaction> txn(txn_db->BeginTransaction(WriteOptions()));
+  ASSERT_OK(txn->Put("a", "v"));
+  Status s = txn->CommitAndTryCreateSnapshot();
+  ASSERT_TRUE(s.IsInvalidArgument());
+}
+
+TEST_P(OptimisticTransactionTest, TimestampedSnapshotSetCommitTs) {
+  std::unique_ptr<Transaction> txn(txn_db->BeginTransaction(WriteOptions()));
+  ASSERT_OK(txn->Put("a", "v"));
+  std::shared_ptr<const Snapshot> snapshot;
+  Status s = txn->CommitAndTryCreateSnapshot(nullptr, /*ts=*/100, &snapshot);
+  ASSERT_TRUE(s.IsNotSupported());
+}
+
+INSTANTIATE_TEST_CASE_P(
+    InstanceOccGroup, OptimisticTransactionTest,
+    testing::Values(OccValidationPolicy::kValidateSerial,
+                    OccValidationPolicy::kValidateParallel));
+
+}  // namespace ROCKSDB_NAMESPACE
+
+int main(int argc, char** argv) {
+  ROCKSDB_NAMESPACE::port::InstallStackTraceHandler();
+  ::testing::InitGoogleTest(&argc, argv);
+  return RUN_ALL_TESTS();
+}
+
+#else
+#include <stdio.h>
+
+int main(int /*argc*/, char** /*argv*/) {
+  fprintf(
+      stderr,
+      "SKIPPED as optimistic_transaction is not supported in ROCKSDB_LITE\n");
+  return 0;
+}
+
+#endif  // !ROCKSDB_LITE
diff --git a/src/rocksdb/utilities/transactions/pessimistic_transaction.cc b/src/rocksdb/utilities/transactions/pessimistic_transaction.cc
new file mode 100644
index 000000000..cb8fd3bb6
--- /dev/null
+++ b/src/rocksdb/utilities/transactions/pessimistic_transaction.cc
@@ -0,0 +1,1175 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#ifndef ROCKSDB_LITE
+
+#include "utilities/transactions/pessimistic_transaction.h"
+
+#include <map>
+#include <set>
+#include <string>
+#include <vector>
+
+#include "db/column_family.h"
+#include "db/db_impl/db_impl.h"
+#include "logging/logging.h"
+#include "rocksdb/comparator.h"
+#include "rocksdb/db.h"
+#include "rocksdb/snapshot.h"
+#include "rocksdb/status.h"
+#include "rocksdb/utilities/transaction_db.h"
+#include "test_util/sync_point.h"
+#include "util/cast_util.h"
+#include "util/string_util.h"
+#include "utilities/transactions/pessimistic_transaction_db.h"
+#include "utilities/transactions/transaction_util.h"
+#include "utilities/write_batch_with_index/write_batch_with_index_internal.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+struct WriteOptions;
+
+std::atomic<TransactionID> PessimisticTransaction::txn_id_counter_(1);
+
+TransactionID PessimisticTransaction::GenTxnID() {
+  return txn_id_counter_.fetch_add(1);
+}
+
+PessimisticTransaction::PessimisticTransaction(
+    TransactionDB* txn_db, const WriteOptions& write_options,
+    const TransactionOptions& txn_options, const bool init)
+    : TransactionBaseImpl(
+          txn_db->GetRootDB(), write_options,
+          static_cast_with_check<PessimisticTransactionDB>(txn_db)
+              ->GetLockTrackerFactory()),
+      txn_db_impl_(nullptr),
+      expiration_time_(0),
+      txn_id_(0),
+      waiting_cf_id_(0),
+      waiting_key_(nullptr),
+      lock_timeout_(0),
+      deadlock_detect_(false),
+      deadlock_detect_depth_(0),
+      skip_concurrency_control_(false) {
+  txn_db_impl_ = static_cast_with_check<PessimisticTransactionDB>(txn_db);
+  db_impl_ = static_cast_with_check<DBImpl>(db_);
+  if (init) {
+    Initialize(txn_options);
+  }
+}
+
+void PessimisticTransaction::Initialize(const TransactionOptions& txn_options) {
+  // Range lock manager uses address of transaction object as TXNID
+  const TransactionDBOptions& db_options = txn_db_impl_->GetTxnDBOptions();
+  if (db_options.lock_mgr_handle &&
+      db_options.lock_mgr_handle->getLockManager()->IsRangeLockSupported()) {
+    txn_id_ = reinterpret_cast<TransactionID>(this);
+  } else {
+    txn_id_ = GenTxnID();
+  }
+
+  txn_state_ = STARTED;
+
+  deadlock_detect_ = txn_options.deadlock_detect;
+  deadlock_detect_depth_ = txn_options.deadlock_detect_depth;
+  write_batch_.SetMaxBytes(txn_options.max_write_batch_size);
+  skip_concurrency_control_ = txn_options.skip_concurrency_control;
+
+  lock_timeout_ = txn_options.lock_timeout * 1000;
+  if (lock_timeout_ < 0) {
+    // Lock timeout not set, use default
+    lock_timeout_ =
+        txn_db_impl_->GetTxnDBOptions().transaction_lock_timeout * 1000;
+  }
+
+  if (txn_options.expiration >= 0) {
+    expiration_time_ = start_time_ + txn_options.expiration * 1000;
+  } else {
+    expiration_time_ = 0;
+  }
+
+  if (txn_options.set_snapshot) {
+    SetSnapshot();
+  }
+
+  if (expiration_time_ > 0) {
+    txn_db_impl_->InsertExpirableTransaction(txn_id_, this);
+  }
+  use_only_the_last_commit_time_batch_for_recovery_ =
+      txn_options.use_only_the_last_commit_time_batch_for_recovery;
+  skip_prepare_ = txn_options.skip_prepare;
+
+  read_timestamp_ = kMaxTxnTimestamp;
+  commit_timestamp_ = kMaxTxnTimestamp;
+}
+
+PessimisticTransaction::~PessimisticTransaction() {
+  txn_db_impl_->UnLock(this, *tracked_locks_);
+  if (expiration_time_ > 0) {
+    txn_db_impl_->RemoveExpirableTransaction(txn_id_);
+  }
+  if (!name_.empty() && txn_state_ != COMMITTED) {
+    txn_db_impl_->UnregisterTransaction(this);
+  }
+}
+
+void PessimisticTransaction::Clear() {
+  txn_db_impl_->UnLock(this, *tracked_locks_);
+  TransactionBaseImpl::Clear();
+}
+
+void PessimisticTransaction::Reinitialize(
+    TransactionDB* txn_db, const WriteOptions& write_options,
+    const TransactionOptions& txn_options) {
+  if (!name_.empty() && txn_state_ != COMMITTED) {
+    txn_db_impl_->UnregisterTransaction(this);
+  }
+  TransactionBaseImpl::Reinitialize(txn_db->GetRootDB(), write_options);
+  Initialize(txn_options);
+}
+
+bool PessimisticTransaction::IsExpired() const {
+  if (expiration_time_ > 0) {
+    if (dbimpl_->GetSystemClock()->NowMicros() >= expiration_time_) {
+      // Transaction is expired.
+      return true;
+    }
+  }
+
+  return false;
+}
+
+WriteCommittedTxn::WriteCommittedTxn(TransactionDB* txn_db,
+                                     const WriteOptions& write_options,
+                                     const TransactionOptions& txn_options)
+    : PessimisticTransaction(txn_db, write_options, txn_options) {}
+
+Status WriteCommittedTxn::GetForUpdate(const ReadOptions& read_options,
+                                       ColumnFamilyHandle* column_family,
+                                       const Slice& key, std::string* value,
+                                       bool exclusive, const bool do_validate) {
+  return GetForUpdateImpl(read_options, column_family, key, value, exclusive,
+                          do_validate);
+}
+
+Status WriteCommittedTxn::GetForUpdate(const ReadOptions& read_options,
+                                       ColumnFamilyHandle* column_family,
+                                       const Slice& key,
+                                       PinnableSlice* pinnable_val,
+                                       bool exclusive, const bool do_validate) {
+  return GetForUpdateImpl(read_options, column_family, key, pinnable_val,
+                          exclusive, do_validate);
+}
+
+template <typename TValue>
+inline Status WriteCommittedTxn::GetForUpdateImpl(
+    const ReadOptions& read_options, ColumnFamilyHandle* column_family,
+    const Slice& key, TValue* value, bool exclusive, const bool do_validate) {
+  column_family =
+      column_family ? column_family : db_impl_->DefaultColumnFamily();
+  assert(column_family);
+  if (!read_options.timestamp) {
+    const Comparator* const ucmp = column_family->GetComparator();
+    assert(ucmp);
+    size_t ts_sz = ucmp->timestamp_size();
+    if (0 == ts_sz) {
+      return TransactionBaseImpl::GetForUpdate(read_options, column_family, key,
+                                               value, exclusive, do_validate);
+    }
+  } else {
+    Status s = db_impl_->FailIfTsMismatchCf(
+        column_family, *(read_options.timestamp), /*ts_for_read=*/true);
+    if (!s.ok()) {
+      return s;
+    }
+  }
+
+  if (!do_validate) {
+    return Status::InvalidArgument(
+        "If do_validate is false then GetForUpdate with read_timestamp is not "
+        "defined.");
+  } else if (kMaxTxnTimestamp == read_timestamp_) {
+    return Status::InvalidArgument("read_timestamp must be set for validation");
+  }
+
+  if (!read_options.timestamp) {
+    ReadOptions read_opts_copy = read_options;
+    char ts_buf[sizeof(kMaxTxnTimestamp)];
+    EncodeFixed64(ts_buf, read_timestamp_);
+    Slice ts(ts_buf, sizeof(ts_buf));
+    read_opts_copy.timestamp = &ts;
+    return TransactionBaseImpl::GetForUpdate(read_opts_copy, column_family, key,
+                                             value, exclusive, do_validate);
+  }
+  assert(read_options.timestamp);
+  const char* const ts_buf = read_options.timestamp->data();
+  assert(read_options.timestamp->size() == sizeof(kMaxTxnTimestamp));
+  TxnTimestamp ts = DecodeFixed64(ts_buf);
+  if (ts != read_timestamp_) {
+    return Status::InvalidArgument("Must read from the same read_timestamp");
+  }
+  return TransactionBaseImpl::GetForUpdate(read_options, column_family, key,
+                                           value, exclusive, do_validate);
+}
+
+Status WriteCommittedTxn::Put(ColumnFamilyHandle* column_family,
+                              const Slice& key, const Slice& value,
+                              const bool assume_tracked) {
+  const bool do_validate = !assume_tracked;
+  return Operate(column_family, key, do_validate, assume_tracked,
+                 [column_family, &key, &value, this]() {
+                   Status s =
+                       GetBatchForWrite()->Put(column_family, key, value);
+                   if (s.ok()) {
+                     ++num_puts_;
+                   }
+                   return s;
+                 });
+}
+
+Status WriteCommittedTxn::Put(ColumnFamilyHandle* column_family,
+                              const SliceParts& key, const SliceParts& value,
+                              const bool assume_tracked) {
+  const bool do_validate = !assume_tracked;
+  return Operate(column_family, key, do_validate, assume_tracked,
+                 [column_family, &key, &value, this]() {
+                   Status s =
+                       GetBatchForWrite()->Put(column_family, key, value);
+                   if (s.ok()) {
+                     ++num_puts_;
+                   }
+                   return s;
+                 });
+}
+
+Status WriteCommittedTxn::PutUntracked(ColumnFamilyHandle* column_family,
+                                       const Slice& key, const Slice& value) {
+  return Operate(
+      column_family, key, /*do_validate=*/false,
+      /*assume_tracked=*/false, [column_family, &key, &value, this]() {
+        Status s = GetBatchForWrite()->Put(column_family, key, value);
+        if (s.ok()) {
+          ++num_puts_;
+        }
+        return s;
+      });
+}
+
+Status WriteCommittedTxn::PutUntracked(ColumnFamilyHandle* column_family,
+                                       const SliceParts& key,
+                                       const SliceParts& value) {
+  return Operate(
+      column_family, key, /*do_validate=*/false,
+      /*assume_tracked=*/false, [column_family, &key, &value, this]() {
+        Status s = GetBatchForWrite()->Put(column_family, key, value);
+        if (s.ok()) {
+          ++num_puts_;
+        }
+        return s;
+      });
+}
+
+Status WriteCommittedTxn::Delete(ColumnFamilyHandle* column_family,
+                                 const Slice& key, const bool assume_tracked) {
+  const bool do_validate = !assume_tracked;
+  return Operate(column_family, key, do_validate, assume_tracked,
+                 [column_family, &key, this]() {
+                   Status s = GetBatchForWrite()->Delete(column_family, key);
+                   if (s.ok()) {
+                     ++num_deletes_;
+                   }
+                   return s;
+                 });
+}
+
+Status WriteCommittedTxn::Delete(ColumnFamilyHandle* column_family,
+                                 const SliceParts& key,
+                                 const bool assume_tracked) {
+  const bool do_validate = !assume_tracked;
+  return Operate(column_family, key, do_validate, assume_tracked,
+                 [column_family, &key, this]() {
+                   Status s = GetBatchForWrite()->Delete(column_family, key);
+                   if (s.ok()) {
+                     ++num_deletes_;
+                   }
+                   return s;
+                 });
+}
+
+Status WriteCommittedTxn::DeleteUntracked(ColumnFamilyHandle* column_family,
+                                          const Slice& key) {
+  return Operate(column_family, key, /*do_validate=*/false,
+                 /*assume_tracked=*/false, [column_family, &key, this]() {
+                   Status s = GetBatchForWrite()->Delete(column_family, key);
+                   if (s.ok()) {
+                     ++num_deletes_;
+                   }
+                   return s;
+                 });
+}
+
+Status WriteCommittedTxn::DeleteUntracked(ColumnFamilyHandle* column_family,
+                                          const SliceParts& key) {
+  return Operate(column_family, key, /*do_validate=*/false,
+                 /*assume_tracked=*/false, [column_family, &key, this]() {
+                   Status s = GetBatchForWrite()->Delete(column_family, key);
+                   if (s.ok()) {
+                     ++num_deletes_;
+                   }
+                   return s;
+                 });
+}
+
+Status WriteCommittedTxn::SingleDelete(ColumnFamilyHandle* column_family,
+                                       const Slice& key,
+                                       const bool assume_tracked) {
+  const bool do_validate = !assume_tracked;
+  return Operate(column_family, key, do_validate, assume_tracked,
+                 [column_family, &key, this]() {
+                   Status s =
+                       GetBatchForWrite()->SingleDelete(column_family, key);
+                   if (s.ok()) {
+                     ++num_deletes_;
+                   }
+                   return s;
+                 });
+}
+
+Status WriteCommittedTxn::SingleDelete(ColumnFamilyHandle* column_family,
+                                       const SliceParts& key,
+                                       const bool assume_tracked) {
+  const bool do_validate = !assume_tracked;
+  return Operate(column_family, key, do_validate, assume_tracked,
+                 [column_family, &key, this]() {
+                   Status s =
+                       GetBatchForWrite()->SingleDelete(column_family, key);
+                   if (s.ok()) {
+                     ++num_deletes_;
+                   }
+                   return s;
+                 });
+}
+
+Status WriteCommittedTxn::SingleDeleteUntracked(
+    ColumnFamilyHandle* column_family, const Slice& key) {
+  return Operate(column_family, key, /*do_validate=*/false,
+                 /*assume_tracked=*/false, [column_family, &key, this]() {
+                   Status s =
+                       GetBatchForWrite()->SingleDelete(column_family, key);
+                   if (s.ok()) {
+                     ++num_deletes_;
+                   }
+                   return s;
+                 });
+}
+
+Status WriteCommittedTxn::Merge(ColumnFamilyHandle* column_family,
+                                const Slice& key, const Slice& value,
+                                const bool assume_tracked) {
+  const bool do_validate = !assume_tracked;
+  return Operate(column_family, key, do_validate, assume_tracked,
+                 [column_family, &key, &value, this]() {
+                   Status s =
+                       GetBatchForWrite()->Merge(column_family, key, value);
+                   if (s.ok()) {
+                     ++num_merges_;
+                   }
+                   return s;
+                 });
+}
+
+template <typename TKey, typename TOperation>
+Status WriteCommittedTxn::Operate(ColumnFamilyHandle* column_family,
+                                  const TKey& key, const bool do_validate,
+                                  const bool assume_tracked,
+                                  TOperation&& operation) {
+  Status s;
+  if constexpr (std::is_same_v<Slice, TKey>) {
+    s = TryLock(column_family, key, /*read_only=*/false, /*exclusive=*/true,
+                do_validate, assume_tracked);
+  } else if constexpr (std::is_same_v<SliceParts, TKey>) {
+    std::string key_buf;
+    Slice contiguous_key(key, &key_buf);
+    s = TryLock(column_family, contiguous_key, /*read_only=*/false,
+                /*exclusive=*/true, do_validate, assume_tracked);
+  }
+  if (!s.ok()) {
+    return s;
+  }
+  column_family =
+      column_family ? column_family : db_impl_->DefaultColumnFamily();
+  assert(column_family);
+  const Comparator* const ucmp = column_family->GetComparator();
+  assert(ucmp);
+  size_t ts_sz = ucmp->timestamp_size();
+  if (ts_sz > 0) {
+    assert(ts_sz == sizeof(TxnTimestamp));
+    if (!IndexingEnabled()) {
+      cfs_with_ts_tracked_when_indexing_disabled_.insert(
+          column_family->GetID());
+    }
+  }
+  return operation();
+}
+
+Status WriteCommittedTxn::SetReadTimestampForValidation(TxnTimestamp ts) {
+  if (read_timestamp_ < kMaxTxnTimestamp && ts < read_timestamp_) {
+    return Status::InvalidArgument(
+        "Cannot decrease read timestamp for validation");
+  }
+  read_timestamp_ = ts;
+  return Status::OK();
+}
+
+Status WriteCommittedTxn::SetCommitTimestamp(TxnTimestamp ts) {
+  if (read_timestamp_ < kMaxTxnTimestamp && ts <= read_timestamp_) {
+    return Status::InvalidArgument(
+        "Cannot commit at timestamp smaller than or equal to read timestamp");
+  }
+  commit_timestamp_ = ts;
+  return Status::OK();
+}
+
+Status PessimisticTransaction::CommitBatch(WriteBatch* batch) {
+  if (batch && WriteBatchInternal::HasKeyWithTimestamp(*batch)) {
+    // CommitBatch() needs to lock the keys in the batch.
+    // However, the application also needs to specify the timestamp for the
+    // keys in batch before calling this API.
+    // This means timestamp order may violate the order of locking, thus
+    // violate the sequence number order for the same user key.
+    // Therefore, we disallow this operation for now.
+    return Status::NotSupported(
+        "Batch to commit includes timestamp assigned before locking");
+  }
+
+  std::unique_ptr<LockTracker> keys_to_unlock(lock_tracker_factory_.Create());
+  Status s = LockBatch(batch, keys_to_unlock.get());
+
+  if (!s.ok()) {
+    return s;
+  }
+
+  bool can_commit = false;
+
+  if (IsExpired()) {
+    s = Status::Expired();
+  } else if (expiration_time_ > 0) {
+    TransactionState expected = STARTED;
+    can_commit = std::atomic_compare_exchange_strong(&txn_state_, &expected,
+                                                     AWAITING_COMMIT);
+  } else if (txn_state_ == STARTED) {
+    // lock stealing is not a concern
+    can_commit = true;
+  }
+
+  if (can_commit) {
+    txn_state_.store(AWAITING_COMMIT);
+    s = CommitBatchInternal(batch);
+    if (s.ok()) {
+      txn_state_.store(COMMITTED);
+    }
+  } else if (txn_state_ == LOCKS_STOLEN) {
+    s = Status::Expired();
+  } else {
+    s = Status::InvalidArgument("Transaction is not in state for commit.");
+  }
+
+  txn_db_impl_->UnLock(this, *keys_to_unlock);
+
+  return s;
+}
+
+Status PessimisticTransaction::Prepare() {
+  if (name_.empty()) {
+    return Status::InvalidArgument(
+        "Cannot prepare a transaction that has not been named.");
+  }
+
+  if (IsExpired()) {
+    return Status::Expired();
+  }
+
+  Status s;
+  bool can_prepare = false;
+
+  if (expiration_time_ > 0) {
+    // must concern ourselves with expiraton and/or lock stealing
+    // need to compare/exchange bc locks could be stolen under us here
+    TransactionState expected = STARTED;
+    can_prepare = std::atomic_compare_exchange_strong(&txn_state_, &expected,
+                                                      AWAITING_PREPARE);
+  } else if (txn_state_ == STARTED) {
+    // expiration and lock stealing is not possible
+    txn_state_.store(AWAITING_PREPARE);
+    can_prepare = true;
+  }
+
+  if (can_prepare) {
+    // transaction can't expire after preparation
+    expiration_time_ = 0;
+    assert(log_number_ == 0 ||
+           txn_db_impl_->GetTxnDBOptions().write_policy == WRITE_UNPREPARED);
+
+    s = PrepareInternal();
+    if (s.ok()) {
+      txn_state_.store(PREPARED);
+    }
+  } else if (txn_state_ == LOCKS_STOLEN) {
+    s = Status::Expired();
+  } else if (txn_state_ == PREPARED) {
+    s = Status::InvalidArgument("Transaction has already been prepared.");
+  } else if (txn_state_ == COMMITTED) {
+    s = Status::InvalidArgument("Transaction has already been committed.");
+  } else if (txn_state_ == ROLLEDBACK) {
+    s = Status::InvalidArgument("Transaction has already been rolledback.");
+  } else {
+    s = Status::InvalidArgument("Transaction is not in state for commit.");
+  }
+
+  return s;
+}
+
+Status WriteCommittedTxn::PrepareInternal() {
+  WriteOptions write_options = write_options_;
+  write_options.disableWAL = false;
+  auto s = WriteBatchInternal::MarkEndPrepare(GetWriteBatch()->GetWriteBatch(),
+                                              name_);
+  assert(s.ok());
+  class MarkLogCallback : public PreReleaseCallback {
+   public:
+    MarkLogCallback(DBImpl* db, bool two_write_queues)
+        : db_(db), two_write_queues_(two_write_queues) {
+      (void)two_write_queues_;  // to silence unused private field warning
+    }
+    virtual Status Callback(SequenceNumber, bool is_mem_disabled,
+                            uint64_t log_number, size_t /*index*/,
+                            size_t /*total*/) override {
+#ifdef NDEBUG
+      (void)is_mem_disabled;
+#endif
+      assert(log_number != 0);
+      assert(!two_write_queues_ || is_mem_disabled);  // implies the 2nd queue
+      db_->logs_with_prep_tracker()->MarkLogAsContainingPrepSection(log_number);
+      return Status::OK();
+    }
+
+   private:
+    DBImpl* db_;
+    bool two_write_queues_;
+  } mark_log_callback(db_impl_,
+                      db_impl_->immutable_db_options().two_write_queues);
+
+  WriteCallback* const kNoWriteCallback = nullptr;
+  const uint64_t kRefNoLog = 0;
+  const bool kDisableMemtable = true;
+  SequenceNumber* const KIgnoreSeqUsed = nullptr;
+  const size_t kNoBatchCount = 0;
+  s = db_impl_->WriteImpl(write_options, GetWriteBatch()->GetWriteBatch(),
+                          kNoWriteCallback, &log_number_, kRefNoLog,
+                          kDisableMemtable, KIgnoreSeqUsed, kNoBatchCount,
+                          &mark_log_callback);
+  return s;
+}
+
+Status PessimisticTransaction::Commit() {
+  bool commit_without_prepare = false;
+  bool commit_prepared = false;
+
+  if (IsExpired()) {
+    return Status::Expired();
+  }
+
+  if (expiration_time_ > 0) {
+    // we must atomicaly compare and exchange the state here because at
+    // this state in the transaction it is possible for another thread
+    // to change our state out from under us in the even that we expire and have
+    // our locks stolen. In this case the only valid state is STARTED because
+    // a state of PREPARED would have a cleared expiration_time_.
+    TransactionState expected = STARTED;
+    commit_without_prepare = std::atomic_compare_exchange_strong(
+        &txn_state_, &expected, AWAITING_COMMIT);
+    TEST_SYNC_POINT("TransactionTest::ExpirableTransactionDataRace:1");
+  } else if (txn_state_ == PREPARED) {
+    // expiration and lock stealing is not a concern
+    commit_prepared = true;
+  } else if (txn_state_ == STARTED) {
+    // expiration and lock stealing is not a concern
+    if (skip_prepare_) {
+      commit_without_prepare = true;
+    } else {
+      return Status::TxnNotPrepared();
+    }
+  }
+
+  Status s;
+  if (commit_without_prepare) {
+    assert(!commit_prepared);
+    if (WriteBatchInternal::Count(GetCommitTimeWriteBatch()) > 0) {
+      s = Status::InvalidArgument(
+          "Commit-time batch contains values that will not be committed.");
+    } else {
+      txn_state_.store(AWAITING_COMMIT);
+      if (log_number_ > 0) {
+        dbimpl_->logs_with_prep_tracker()->MarkLogAsHavingPrepSectionFlushed(
+            log_number_);
+      }
+      s = CommitWithoutPrepareInternal();
+      if (!name_.empty()) {
+        txn_db_impl_->UnregisterTransaction(this);
+      }
+      Clear();
+      if (s.ok()) {
+        txn_state_.store(COMMITTED);
+      }
+    }
+  } else if (commit_prepared) {
+    txn_state_.store(AWAITING_COMMIT);
+
+    s = CommitInternal();
+
+    if (!s.ok()) {
+      ROCKS_LOG_WARN(db_impl_->immutable_db_options().info_log,
+                     "Commit write failed");
+      return s;
+    }
+
+    // FindObsoleteFiles must now look to the memtables
+    // to determine what prep logs must be kept around,
+    // not the prep section heap.
+    assert(log_number_ > 0);
+    dbimpl_->logs_with_prep_tracker()->MarkLogAsHavingPrepSectionFlushed(
+        log_number_);
+    txn_db_impl_->UnregisterTransaction(this);
+
+    Clear();
+    txn_state_.store(COMMITTED);
+  } else if (txn_state_ == LOCKS_STOLEN) {
+    s = Status::Expired();
+  } else if (txn_state_ == COMMITTED) {
+    s = Status::InvalidArgument("Transaction has already been committed.");
+  } else if (txn_state_ == ROLLEDBACK) {
+    s = Status::InvalidArgument("Transaction has already been rolledback.");
+  } else {
+    s = Status::InvalidArgument("Transaction is not in state for commit.");
+  }
+
+  return s;
+}
+
+Status WriteCommittedTxn::CommitWithoutPrepareInternal() {
+  WriteBatchWithIndex* wbwi = GetWriteBatch();
+  assert(wbwi);
+  WriteBatch* wb = wbwi->GetWriteBatch();
+  assert(wb);
+
+  const bool needs_ts = WriteBatchInternal::HasKeyWithTimestamp(*wb);
+  if (needs_ts && commit_timestamp_ == kMaxTxnTimestamp) {
+    return Status::InvalidArgument("Must assign a commit timestamp");
+  }
+
+  if (needs_ts) {
+    assert(commit_timestamp_ != kMaxTxnTimestamp);
+    char commit_ts_buf[sizeof(kMaxTxnTimestamp)];
+    EncodeFixed64(commit_ts_buf, commit_timestamp_);
+    Slice commit_ts(commit_ts_buf, sizeof(commit_ts_buf));
+
+    Status s =
+        wb->UpdateTimestamps(commit_ts, [wbwi, this](uint32_t cf) -> size_t {
+          auto cf_iter = cfs_with_ts_tracked_when_indexing_disabled_.find(cf);
+          if (cf_iter != cfs_with_ts_tracked_when_indexing_disabled_.end()) {
+            return sizeof(kMaxTxnTimestamp);
+          }
+          const Comparator* ucmp =
+              WriteBatchWithIndexInternal::GetUserComparator(*wbwi, cf);
+          return ucmp ? ucmp->timestamp_size()
+                      : std::numeric_limits<uint64_t>::max();
+        });
+    if (!s.ok()) {
+      return s;
+    }
+  }
+
+  uint64_t seq_used = kMaxSequenceNumber;
+  SnapshotCreationCallback snapshot_creation_cb(db_impl_, commit_timestamp_,
+                                                snapshot_notifier_, snapshot_);
+  PostMemTableCallback* post_mem_cb = nullptr;
+  if (snapshot_needed_) {
+    if (commit_timestamp_ == kMaxTxnTimestamp) {
+      return Status::InvalidArgument("Must set transaction commit timestamp");
+    } else {
+      post_mem_cb = &snapshot_creation_cb;
+    }
+  }
+  auto s = db_impl_->WriteImpl(write_options_, wb,
+                               /*callback*/ nullptr, /*log_used*/ nullptr,
+                               /*log_ref*/ 0, /*disable_memtable*/ false,
+                               &seq_used, /*batch_cnt=*/0,
+                               /*pre_release_callback=*/nullptr, post_mem_cb);
+  assert(!s.ok() || seq_used != kMaxSequenceNumber);
+  if (s.ok()) {
+    SetId(seq_used);
+  }
+  return s;
+}
+
+Status WriteCommittedTxn::CommitBatchInternal(WriteBatch* batch, size_t) {
+  uint64_t seq_used = kMaxSequenceNumber;
+  auto s = db_impl_->WriteImpl(write_options_, batch, /*callback*/ nullptr,
+                               /*log_used*/ nullptr, /*log_ref*/ 0,
+                               /*disable_memtable*/ false, &seq_used);
+  assert(!s.ok() || seq_used != kMaxSequenceNumber);
+  if (s.ok()) {
+    SetId(seq_used);
+  }
+  return s;
+}
+
+Status WriteCommittedTxn::CommitInternal() {
+  WriteBatchWithIndex* wbwi = GetWriteBatch();
+  assert(wbwi);
+  WriteBatch* wb = wbwi->GetWriteBatch();
+  assert(wb);
+
+  const bool needs_ts = WriteBatchInternal::HasKeyWithTimestamp(*wb);
+  if (needs_ts && commit_timestamp_ == kMaxTxnTimestamp) {
+    return Status::InvalidArgument("Must assign a commit timestamp");
+  }
+  // We take the commit-time batch and append the Commit marker.
+  // The Memtable will ignore the Commit marker in non-recovery mode
+  WriteBatch* working_batch = GetCommitTimeWriteBatch();
+
+  Status s;
+  if (!needs_ts) {
+    s = WriteBatchInternal::MarkCommit(working_batch, name_);
+  } else {
+    assert(commit_timestamp_ != kMaxTxnTimestamp);
+    char commit_ts_buf[sizeof(kMaxTxnTimestamp)];
+    EncodeFixed64(commit_ts_buf, commit_timestamp_);
+    Slice commit_ts(commit_ts_buf, sizeof(commit_ts_buf));
+    s = WriteBatchInternal::MarkCommitWithTimestamp(working_batch, name_,
+                                                    commit_ts);
+    if (s.ok()) {
+      s = wb->UpdateTimestamps(commit_ts, [wbwi, this](uint32_t cf) -> size_t {
+        if (cfs_with_ts_tracked_when_indexing_disabled_.find(cf) !=
+            cfs_with_ts_tracked_when_indexing_disabled_.end()) {
+          return sizeof(kMaxTxnTimestamp);
+        }
+        const Comparator* ucmp =
+            WriteBatchWithIndexInternal::GetUserComparator(*wbwi, cf);
+        return ucmp ? ucmp->timestamp_size()
+                    : std::numeric_limits<uint64_t>::max();
+      });
+    }
+  }
+
+  if (!s.ok()) {
+    return s;
+  }
+
+  // any operations appended to this working_batch will be ignored from WAL
+  working_batch->MarkWalTerminationPoint();
+
+  // insert prepared batch into Memtable only skipping WAL.
+  // Memtable will ignore BeginPrepare/EndPrepare markers
+  // in non recovery mode and simply insert the values
+  s = WriteBatchInternal::Append(working_batch, wb);
+  assert(s.ok());
+
+  uint64_t seq_used = kMaxSequenceNumber;
+  SnapshotCreationCallback snapshot_creation_cb(db_impl_, commit_timestamp_,
+                                                snapshot_notifier_, snapshot_);
+  PostMemTableCallback* post_mem_cb = nullptr;
+  if (snapshot_needed_) {
+    if (commit_timestamp_ == kMaxTxnTimestamp) {
+      s = Status::InvalidArgument("Must set transaction commit timestamp");
+      return s;
+    } else {
+      post_mem_cb = &snapshot_creation_cb;
+    }
+  }
+  s = db_impl_->WriteImpl(write_options_, working_batch, /*callback*/ nullptr,
+                          /*log_used*/ nullptr, /*log_ref*/ log_number_,
+                          /*disable_memtable*/ false, &seq_used,
+                          /*batch_cnt=*/0, /*pre_release_callback=*/nullptr,
+                          post_mem_cb);
+  assert(!s.ok() || seq_used != kMaxSequenceNumber);
+  if (s.ok()) {
+    SetId(seq_used);
+  }
+  return s;
+}
+
+Status PessimisticTransaction::Rollback() {
+  Status s;
+  if (txn_state_ == PREPARED) {
+    txn_state_.store(AWAITING_ROLLBACK);
+
+    s = RollbackInternal();
+
+    if (s.ok()) {
+      // we do not need to keep our prepared section around
+      assert(log_number_ > 0);
+      dbimpl_->logs_with_prep_tracker()->MarkLogAsHavingPrepSectionFlushed(
+          log_number_);
+      Clear();
+      txn_state_.store(ROLLEDBACK);
+    }
+  } else if (txn_state_ == STARTED) {
+    if (log_number_ > 0) {
+      assert(txn_db_impl_->GetTxnDBOptions().write_policy == WRITE_UNPREPARED);
+      assert(GetId() > 0);
+      s = RollbackInternal();
+
+      if (s.ok()) {
+        dbimpl_->logs_with_prep_tracker()->MarkLogAsHavingPrepSectionFlushed(
+            log_number_);
+      }
+    }
+    // prepare couldn't have taken place
+    Clear();
+  } else if (txn_state_ == COMMITTED) {
+    s = Status::InvalidArgument("This transaction has already been committed.");
+  } else {
+    s = Status::InvalidArgument(
+        "Two phase transaction is not in state for rollback.");
+  }
+
+  return s;
+}
+
+Status WriteCommittedTxn::RollbackInternal() {
+  WriteBatch rollback_marker;
+  auto s = WriteBatchInternal::MarkRollback(&rollback_marker, name_);
+  assert(s.ok());
+  s = db_impl_->WriteImpl(write_options_, &rollback_marker);
+  return s;
+}
+
+Status PessimisticTransaction::RollbackToSavePoint() {
+  if (txn_state_ != STARTED) {
+    return Status::InvalidArgument("Transaction is beyond state for rollback.");
+  }
+
+  if (save_points_ != nullptr && !save_points_->empty()) {
+    // Unlock any keys locked since last transaction
+    auto& save_point_tracker = *save_points_->top().new_locks_;
+    std::unique_ptr<LockTracker> t(
+        tracked_locks_->GetTrackedLocksSinceSavePoint(save_point_tracker));
+    if (t) {
+      txn_db_impl_->UnLock(this, *t);
+    }
+  }
+
+  return TransactionBaseImpl::RollbackToSavePoint();
+}
+
+// Lock all keys in this batch.
+// On success, caller should unlock keys_to_unlock
+Status PessimisticTransaction::LockBatch(WriteBatch* batch,
+                                         LockTracker* keys_to_unlock) {
+  if (!batch) {
+    return Status::InvalidArgument("batch is nullptr");
+  }
+
+  class Handler : public WriteBatch::Handler {
+   public:
+    // Sorted map of column_family_id to sorted set of keys.
+    // Since LockBatch() always locks keys in sorted order, it cannot deadlock
+    // with itself.  We're not using a comparator here since it doesn't matter
+    // what the sorting is as long as it's consistent.
+    std::map<uint32_t, std::set<std::string>> keys_;
+
+    Handler() {}
+
+    void RecordKey(uint32_t column_family_id, const Slice& key) {
+      std::string key_str = key.ToString();
+
+      auto& cfh_keys = keys_[column_family_id];
+      auto iter = cfh_keys.find(key_str);
+      if (iter == cfh_keys.end()) {
+        // key not yet seen, store it.
+        cfh_keys.insert({std::move(key_str)});
+      }
+    }
+
+    Status PutCF(uint32_t column_family_id, const Slice& key,
+                 const Slice& /* unused */) override {
+      RecordKey(column_family_id, key);
+      return Status::OK();
+    }
+    Status MergeCF(uint32_t column_family_id, const Slice& key,
+                   const Slice& /* unused */) override {
+      RecordKey(column_family_id, key);
+      return Status::OK();
+    }
+    Status DeleteCF(uint32_t column_family_id, const Slice& key) override {
+      RecordKey(column_family_id, key);
+      return Status::OK();
+    }
+  };
+
+  // Iterating on this handler will add all keys in this batch into keys
+  Handler handler;
+  Status s = batch->Iterate(&handler);
+  if (!s.ok()) {
+    return s;
+  }
+
+  // Attempt to lock all keys
+  for (const auto& cf_iter : handler.keys_) {
+    uint32_t cfh_id = cf_iter.first;
+    auto& cfh_keys = cf_iter.second;
+
+    for (const auto& key_iter : cfh_keys) {
+      const std::string& key = key_iter;
+
+      s = txn_db_impl_->TryLock(this, cfh_id, key, true /* exclusive */);
+      if (!s.ok()) {
+        break;
+      }
+      PointLockRequest r;
+      r.column_family_id = cfh_id;
+      r.key = key;
+      r.seq = kMaxSequenceNumber;
+      r.read_only = false;
+      r.exclusive = true;
+      keys_to_unlock->Track(r);
+    }
+
+    if (!s.ok()) {
+      break;
+    }
+  }
+
+  if (!s.ok()) {
+    txn_db_impl_->UnLock(this, *keys_to_unlock);
+  }
+
+  return s;
+}
+
+// Attempt to lock this key.
+// Returns OK if the key has been successfully locked.  Non-ok, otherwise.
+// If check_shapshot is true and this transaction has a snapshot set,
+// this key will only be locked if there have been no writes to this key since
+// the snapshot time.
+Status PessimisticTransaction::TryLock(ColumnFamilyHandle* column_family,
+                                       const Slice& key, bool read_only,
+                                       bool exclusive, const bool do_validate,
+                                       const bool assume_tracked) {
+  assert(!assume_tracked || !do_validate);
+  Status s;
+  if (UNLIKELY(skip_concurrency_control_)) {
+    return s;
+  }
+  uint32_t cfh_id = GetColumnFamilyID(column_family);
+  std::string key_str = key.ToString();
+
+  PointLockStatus status;
+  bool lock_upgrade;
+  bool previously_locked;
+  if (tracked_locks_->IsPointLockSupported()) {
+    status = tracked_locks_->GetPointLockStatus(cfh_id, key_str);
+    previously_locked = status.locked;
+    lock_upgrade = previously_locked && exclusive && !status.exclusive;
+  } else {
+    // If the record is tracked, we can assume it was locked, too.
+    previously_locked = assume_tracked;
+    status.locked = false;
+    lock_upgrade = false;
+  }
+
+  // Lock this key if this transactions hasn't already locked it or we require
+  // an upgrade.
+  if (!previously_locked || lock_upgrade) {
+    s = txn_db_impl_->TryLock(this, cfh_id, key_str, exclusive);
+  }
+
+  const ColumnFamilyHandle* const cfh =
+      column_family ? column_family : db_impl_->DefaultColumnFamily();
+  assert(cfh);
+  const Comparator* const ucmp = cfh->GetComparator();
+  assert(ucmp);
+  size_t ts_sz = ucmp->timestamp_size();
+
+  SetSnapshotIfNeeded();
+
+  // Even though we do not care about doing conflict checking for this write,
+  // we still need to take a lock to make sure we do not cause a conflict with
+  // some other write.  However, we do not need to check if there have been
+  // any writes since this transaction's snapshot.
+  // TODO(agiardullo): could optimize by supporting shared txn locks in the
+  // future.
+  SequenceNumber tracked_at_seq =
+      status.locked ? status.seq : kMaxSequenceNumber;
+  if (!do_validate || (snapshot_ == nullptr &&
+                       (0 == ts_sz || kMaxTxnTimestamp == read_timestamp_))) {
+    if (assume_tracked && !previously_locked &&
+        tracked_locks_->IsPointLockSupported()) {
+      s = Status::InvalidArgument(
+          "assume_tracked is set but it is not tracked yet");
+    }
+    // Need to remember the earliest sequence number that we know that this
+    // key has not been modified after.  This is useful if this same
+    // transaction later tries to lock this key again.
+    if (tracked_at_seq == kMaxSequenceNumber) {
+      // Since we haven't checked a snapshot, we only know this key has not
+      // been modified since after we locked it.
+      // Note: when last_seq_same_as_publish_seq_==false this is less than the
+      // latest allocated seq but it is ok since i) this is just a heuristic
+      // used only as a hint to avoid actual check for conflicts, ii) this would
+      // cause a false positive only if the snapthot is taken right after the
+      // lock, which would be an unusual sequence.
+      tracked_at_seq = db_->GetLatestSequenceNumber();
+    }
+  } else if (s.ok()) {
+    // If a snapshot is set, we need to make sure the key hasn't been modified
+    // since the snapshot.  This must be done after we locked the key.
+    // If we already have validated an earilier snapshot it must has been
+    // reflected in tracked_at_seq and ValidateSnapshot will return OK.
+    s = ValidateSnapshot(column_family, key, &tracked_at_seq);
+
+    if (!s.ok()) {
+      // Failed to validate key
+      // Unlock key we just locked
+      if (lock_upgrade) {
+        s = txn_db_impl_->TryLock(this, cfh_id, key_str, false /* exclusive */);
+        assert(s.ok());
+      } else if (!previously_locked) {
+        txn_db_impl_->UnLock(this, cfh_id, key.ToString());
+      }
+    }
+  }
+
+  if (s.ok()) {
+    // We must track all the locked keys so that we can unlock them later. If
+    // the key is already locked, this func will update some stats on the
+    // tracked key. It could also update the tracked_at_seq if it is lower
+    // than the existing tracked key seq. These stats are necessary for
+    // RollbackToSavePoint to determine whether a key can be safely removed
+    // from tracked_keys_. Removal can only be done if a key was only locked
+    // during the current savepoint.
+    //
+    // Recall that if assume_tracked is true, we assume that TrackKey has been
+    // called previously since the last savepoint, with the same exclusive
+    // setting, and at a lower sequence number, so skipping here should be
+    // safe.
+    if (!assume_tracked) {
+      TrackKey(cfh_id, key_str, tracked_at_seq, read_only, exclusive);
+    } else {
+#ifndef NDEBUG
+      if (tracked_locks_->IsPointLockSupported()) {
+        PointLockStatus lock_status =
+            tracked_locks_->GetPointLockStatus(cfh_id, key_str);
+        assert(lock_status.locked);
+        assert(lock_status.seq <= tracked_at_seq);
+        assert(lock_status.exclusive == exclusive);
+      }
+#endif
+    }
+  }
+
+  return s;
+}
+
+Status PessimisticTransaction::GetRangeLock(ColumnFamilyHandle* column_family,
+                                            const Endpoint& start_endp,
+                                            const Endpoint& end_endp) {
+  ColumnFamilyHandle* cfh =
+      column_family ? column_family : db_impl_->DefaultColumnFamily();
+  uint32_t cfh_id = GetColumnFamilyID(cfh);
+
+  Status s = txn_db_impl_->TryRangeLock(this, cfh_id, start_endp, end_endp);
+
+  if (s.ok()) {
+    RangeLockRequest req{cfh_id, start_endp, end_endp};
+    tracked_locks_->Track(req);
+  }
+  return s;
+}
+
+// Return OK() if this key has not been modified more recently than the
+// transaction snapshot_.
+// tracked_at_seq is the global seq at which we either locked the key or already
+// have done ValidateSnapshot.
+Status PessimisticTransaction::ValidateSnapshot(
+    ColumnFamilyHandle* column_family, const Slice& key,
+    SequenceNumber* tracked_at_seq) {
+  assert(snapshot_ || read_timestamp_ < kMaxTxnTimestamp);
+
+  SequenceNumber snap_seq = 0;
+  if (snapshot_) {
+    snap_seq = snapshot_->GetSequenceNumber();
+    if (*tracked_at_seq <= snap_seq) {
+      // If the key has been previous validated (or locked) at a sequence number
+      // earlier than the current snapshot's sequence number, we already know it
+      // has not been modified aftter snap_seq either.
+      return Status::OK();
+    }
+  } else {
+    snap_seq = db_impl_->GetLatestSequenceNumber();
+  }
+
+  // Otherwise we have either
+  // 1: tracked_at_seq == kMaxSequenceNumber, i.e., first time tracking the key
+  // 2: snap_seq < tracked_at_seq: last time we lock the key was via
+  // do_validate=false which means we had skipped ValidateSnapshot. In both
+  // cases we should do ValidateSnapshot now.
+
+  *tracked_at_seq = snap_seq;
+
+  ColumnFamilyHandle* cfh =
+      column_family ? column_family : db_impl_->DefaultColumnFamily();
+
+  assert(cfh);
+  const Comparator* const ucmp = cfh->GetComparator();
+  assert(ucmp);
+  size_t ts_sz = ucmp->timestamp_size();
+  std::string ts_buf;
+  if (ts_sz > 0 && read_timestamp_ < kMaxTxnTimestamp) {
+    assert(ts_sz == sizeof(read_timestamp_));
+    PutFixed64(&ts_buf, read_timestamp_);
+  }
+
+  return TransactionUtil::CheckKeyForConflicts(
+      db_impl_, cfh, key.ToString(), snap_seq, ts_sz == 0 ? nullptr : &ts_buf,
+      false /* cache_only */);
+}
+
+bool PessimisticTransaction::TryStealingLocks() {
+  assert(IsExpired());
+  TransactionState expected = STARTED;
+  return std::atomic_compare_exchange_strong(&txn_state_, &expected,
+                                             LOCKS_STOLEN);
+}
+
+void PessimisticTransaction::UnlockGetForUpdate(
+    ColumnFamilyHandle* column_family, const Slice& key) {
+  txn_db_impl_->UnLock(this, GetColumnFamilyID(column_family), key.ToString());
+}
+
+Status PessimisticTransaction::SetName(const TransactionName& name) {
+  Status s;
+  if (txn_state_ == STARTED) {
+    if (name_.length()) {
+      s = Status::InvalidArgument("Transaction has already been named.");
+    } else if (txn_db_impl_->GetTransactionByName(name) != nullptr) {
+      s = Status::InvalidArgument("Transaction name must be unique.");
+    } else if (name.length() < 1 || name.length() > 512) {
+      s = Status::InvalidArgument(
+          "Transaction name length must be between 1 and 512 chars.");
+    } else {
+      name_ = name;
+      txn_db_impl_->RegisterTransaction(this);
+    }
+  } else {
+    s = Status::InvalidArgument("Transaction is beyond state for naming.");
+  }
+  return s;
+}
+
+}  // namespace ROCKSDB_NAMESPACE
+
+#endif  // ROCKSDB_LITE
diff --git a/src/rocksdb/utilities/transactions/pessimistic_transaction.h b/src/rocksdb/utilities/transactions/pessimistic_transaction.h
new file mode 100644
index 000000000..d43d1d3ac
--- /dev/null
+++ b/src/rocksdb/utilities/transactions/pessimistic_transaction.h
@@ -0,0 +1,313 @@
+// Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#pragma once
+
+#ifndef ROCKSDB_LITE
+
+#include <algorithm>
+#include <atomic>
+#include <mutex>
+#include <stack>
+#include <string>
+#include <unordered_map>
+#include <vector>
+
+#include "db/write_callback.h"
+#include "rocksdb/db.h"
+#include "rocksdb/slice.h"
+#include "rocksdb/snapshot.h"
+#include "rocksdb/status.h"
+#include "rocksdb/types.h"
+#include "rocksdb/utilities/transaction.h"
+#include "rocksdb/utilities/transaction_db.h"
+#include "rocksdb/utilities/write_batch_with_index.h"
+#include "util/autovector.h"
+#include "utilities/transactions/transaction_base.h"
+#include "utilities/transactions/transaction_util.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+class PessimisticTransactionDB;
+
+// A transaction under pessimistic concurrency control. This class implements
+// the locking API and interfaces with the lock manager as well as the
+// pessimistic transactional db.
+class PessimisticTransaction : public TransactionBaseImpl {
+ public:
+  PessimisticTransaction(TransactionDB* db, const WriteOptions& write_options,
+                         const TransactionOptions& txn_options,
+                         const bool init = true);
+  // No copying allowed
+  PessimisticTransaction(const PessimisticTransaction&) = delete;
+  void operator=(const PessimisticTransaction&) = delete;
+
+  ~PessimisticTransaction() override;
+
+  void Reinitialize(TransactionDB* txn_db, const WriteOptions& write_options,
+                    const TransactionOptions& txn_options);
+
+  Status Prepare() override;
+
+  Status Commit() override;
+
+  // It is basically Commit without going through Prepare phase. The write batch
+  // is also directly provided instead of expecting txn to gradually batch the
+  // transactions writes to an internal write batch.
+  Status CommitBatch(WriteBatch* batch);
+
+  Status Rollback() override;
+
+  Status RollbackToSavePoint() override;
+
+  Status SetName(const TransactionName& name) override;
+
+  // Generate a new unique transaction identifier
+  static TransactionID GenTxnID();
+
+  TransactionID GetID() const override { return txn_id_; }
+
+  std::vector<TransactionID> GetWaitingTxns(uint32_t* column_family_id,
+                                            std::string* key) const override {
+    std::lock_guard<std::mutex> lock(wait_mutex_);
+    std::vector<TransactionID> ids(waiting_txn_ids_.size());
+    if (key) *key = waiting_key_ ? *waiting_key_ : "";
+    if (column_family_id) *column_family_id = waiting_cf_id_;
+    std::copy(waiting_txn_ids_.begin(), waiting_txn_ids_.end(), ids.begin());
+    return ids;
+  }
+
+  void SetWaitingTxn(autovector<TransactionID> ids, uint32_t column_family_id,
+                     const std::string* key) {
+    std::lock_guard<std::mutex> lock(wait_mutex_);
+    waiting_txn_ids_ = ids;
+    waiting_cf_id_ = column_family_id;
+    waiting_key_ = key;
+  }
+
+  void ClearWaitingTxn() {
+    std::lock_guard<std::mutex> lock(wait_mutex_);
+    waiting_txn_ids_.clear();
+    waiting_cf_id_ = 0;
+    waiting_key_ = nullptr;
+  }
+
+  // Returns the time (in microseconds according to Env->GetMicros())
+  // that this transaction will be expired.  Returns 0 if this transaction does
+  // not expire.
+  uint64_t GetExpirationTime() const { return expiration_time_; }
+
+  // returns true if this transaction has an expiration_time and has expired.
+  bool IsExpired() const;
+
+  // Returns the number of microseconds a transaction can wait on acquiring a
+  // lock or -1 if there is no timeout.
+  int64_t GetLockTimeout() const { return lock_timeout_; }
+  void SetLockTimeout(int64_t timeout) override {
+    lock_timeout_ = timeout * 1000;
+  }
+
+  // Returns true if locks were stolen successfully, false otherwise.
+  bool TryStealingLocks();
+
+  bool IsDeadlockDetect() const override { return deadlock_detect_; }
+
+  int64_t GetDeadlockDetectDepth() const { return deadlock_detect_depth_; }
+
+  virtual Status GetRangeLock(ColumnFamilyHandle* column_family,
+                              const Endpoint& start_key,
+                              const Endpoint& end_key) override;
+
+ protected:
+  // Refer to
+  // TransactionOptions::use_only_the_last_commit_time_batch_for_recovery
+  bool use_only_the_last_commit_time_batch_for_recovery_ = false;
+  // Refer to
+  // TransactionOptions::skip_prepare
+  bool skip_prepare_ = false;
+
+  virtual Status PrepareInternal() = 0;
+
+  virtual Status CommitWithoutPrepareInternal() = 0;
+
+  // batch_cnt if non-zero is the number of sub-batches. A sub-batch is a batch
+  // with no duplicate keys. If zero, then the number of sub-batches is unknown.
+  virtual Status CommitBatchInternal(WriteBatch* batch,
+                                     size_t batch_cnt = 0) = 0;
+
+  virtual Status CommitInternal() = 0;
+
+  virtual Status RollbackInternal() = 0;
+
+  virtual void Initialize(const TransactionOptions& txn_options);
+
+  Status LockBatch(WriteBatch* batch, LockTracker* keys_to_unlock);
+
+  Status TryLock(ColumnFamilyHandle* column_family, const Slice& key,
+                 bool read_only, bool exclusive, const bool do_validate = true,
+                 const bool assume_tracked = false) override;
+
+  void Clear() override;
+
+  PessimisticTransactionDB* txn_db_impl_;
+  DBImpl* db_impl_;
+
+  // If non-zero, this transaction should not be committed after this time (in
+  // microseconds according to Env->NowMicros())
+  uint64_t expiration_time_;
+
+  // Timestamp used by the transaction to perform all GetForUpdate.
+  // Use this timestamp for conflict checking.
+  // read_timestamp_ == kMaxTxnTimestamp means this transaction has not
+  // performed any GetForUpdate. It is possible that the transaction has
+  // performed blind writes or Get, though.
+  TxnTimestamp read_timestamp_{kMaxTxnTimestamp};
+  TxnTimestamp commit_timestamp_{kMaxTxnTimestamp};
+
+ private:
+  friend class TransactionTest_ValidateSnapshotTest_Test;
+  // Used to create unique ids for transactions.
+  static std::atomic<TransactionID> txn_id_counter_;
+
+  // Unique ID for this transaction
+  TransactionID txn_id_;
+
+  // IDs for the transactions that are blocking the current transaction.
+  //
+  // empty if current transaction is not waiting.
+  autovector<TransactionID> waiting_txn_ids_;
+
+  // The following two represents the (cf, key) that a transaction is waiting
+  // on.
+  //
+  // If waiting_key_ is not null, then the pointer should always point to
+  // a valid string object. The reason is that it is only non-null when the
+  // transaction is blocked in the PointLockManager::AcquireWithTimeout
+  // function. At that point, the key string object is one of the function
+  // parameters.
+  uint32_t waiting_cf_id_;
+  const std::string* waiting_key_;
+
+  // Mutex protecting waiting_txn_ids_, waiting_cf_id_ and waiting_key_.
+  mutable std::mutex wait_mutex_;
+
+  // Timeout in microseconds when locking a key or -1 if there is no timeout.
+  int64_t lock_timeout_;
+
+  // Whether to perform deadlock detection or not.
+  bool deadlock_detect_;
+
+  // Whether to perform deadlock detection or not.
+  int64_t deadlock_detect_depth_;
+
+  // Refer to TransactionOptions::skip_concurrency_control
+  bool skip_concurrency_control_;
+
+  virtual Status ValidateSnapshot(ColumnFamilyHandle* column_family,
+                                  const Slice& key,
+                                  SequenceNumber* tracked_at_seq);
+
+  void UnlockGetForUpdate(ColumnFamilyHandle* column_family,
+                          const Slice& key) override;
+};
+
+class WriteCommittedTxn : public PessimisticTransaction {
+ public:
+  WriteCommittedTxn(TransactionDB* db, const WriteOptions& write_options,
+                    const TransactionOptions& txn_options);
+  // No copying allowed
+  WriteCommittedTxn(const WriteCommittedTxn&) = delete;
+  void operator=(const WriteCommittedTxn&) = delete;
+
+  ~WriteCommittedTxn() override {}
+
+  using TransactionBaseImpl::GetForUpdate;
+  Status GetForUpdate(const ReadOptions& read_options,
+                      ColumnFamilyHandle* column_family, const Slice& key,
+                      std::string* value, bool exclusive,
+                      const bool do_validate) override;
+  Status GetForUpdate(const ReadOptions& read_options,
+                      ColumnFamilyHandle* column_family, const Slice& key,
+                      PinnableSlice* pinnable_val, bool exclusive,
+                      const bool do_validate) override;
+
+  using TransactionBaseImpl::Put;
+  // `key` does NOT include timestamp even when it's enabled.
+  Status Put(ColumnFamilyHandle* column_family, const Slice& key,
+             const Slice& value, const bool assume_tracked = false) override;
+  Status Put(ColumnFamilyHandle* column_family, const SliceParts& key,
+             const SliceParts& value,
+             const bool assume_tracked = false) override;
+
+  using TransactionBaseImpl::PutUntracked;
+  Status PutUntracked(ColumnFamilyHandle* column_family, const Slice& key,
+                      const Slice& value) override;
+  Status PutUntracked(ColumnFamilyHandle* column_family, const SliceParts& key,
+                      const SliceParts& value) override;
+
+  using TransactionBaseImpl::Delete;
+  // `key` does NOT include timestamp even when it's enabled.
+  Status Delete(ColumnFamilyHandle* column_family, const Slice& key,
+                const bool assume_tracked = false) override;
+  Status Delete(ColumnFamilyHandle* column_family, const SliceParts& key,
+                const bool assume_tracked = false) override;
+
+  using TransactionBaseImpl::DeleteUntracked;
+  Status DeleteUntracked(ColumnFamilyHandle* column_family,
+                         const Slice& key) override;
+  Status DeleteUntracked(ColumnFamilyHandle* column_family,
+                         const SliceParts& key) override;
+
+  using TransactionBaseImpl::SingleDelete;
+  // `key` does NOT include timestamp even when it's enabled.
+  Status SingleDelete(ColumnFamilyHandle* column_family, const Slice& key,
+                      const bool assume_tracked = false) override;
+  Status SingleDelete(ColumnFamilyHandle* column_family, const SliceParts& key,
+                      const bool assume_tracked = false) override;
+
+  using TransactionBaseImpl::SingleDeleteUntracked;
+  Status SingleDeleteUntracked(ColumnFamilyHandle* column_family,
+                               const Slice& key) override;
+
+  using TransactionBaseImpl::Merge;
+  Status Merge(ColumnFamilyHandle* column_family, const Slice& key,
+               const Slice& value, const bool assume_tracked = false) override;
+
+  Status SetReadTimestampForValidation(TxnTimestamp ts) override;
+  Status SetCommitTimestamp(TxnTimestamp ts) override;
+  TxnTimestamp GetCommitTimestamp() const override { return commit_timestamp_; }
+
+ private:
+  template <typename TValue>
+  Status GetForUpdateImpl(const ReadOptions& read_options,
+                          ColumnFamilyHandle* column_family, const Slice& key,
+                          TValue* value, bool exclusive,
+                          const bool do_validate);
+
+  template <typename TKey, typename TOperation>
+  Status Operate(ColumnFamilyHandle* column_family, const TKey& key,
+                 const bool do_validate, const bool assume_tracked,
+                 TOperation&& operation);
+
+  Status PrepareInternal() override;
+
+  Status CommitWithoutPrepareInternal() override;
+
+  Status CommitBatchInternal(WriteBatch* batch, size_t batch_cnt) override;
+
+  Status CommitInternal() override;
+
+  Status RollbackInternal() override;
+
+  // Column families that enable timestamps and whose data are written when
+  // indexing_enabled_ is false. If a key is written when indexing_enabled_ is
+  // true, then the corresponding column family is not added to cfs_with_ts
+  // even if it enables timestamp.
+  std::unordered_set<uint32_t> cfs_with_ts_tracked_when_indexing_disabled_;
+};
+
+}  // namespace ROCKSDB_NAMESPACE
+
+#endif  // ROCKSDB_LITE
diff --git a/src/rocksdb/utilities/transactions/pessimistic_transaction_db.cc b/src/rocksdb/utilities/transactions/pessimistic_transaction_db.cc
new file mode 100644
index 000000000..950ef8042
--- /dev/null
+++ b/src/rocksdb/utilities/transactions/pessimistic_transaction_db.cc
@@ -0,0 +1,782 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#ifndef ROCKSDB_LITE
+
+#include "utilities/transactions/pessimistic_transaction_db.h"
+
+#include <cinttypes>
+#include <sstream>
+#include <string>
+#include <unordered_set>
+#include <vector>
+
+#include "db/db_impl/db_impl.h"
+#include "logging/logging.h"
+#include "rocksdb/db.h"
+#include "rocksdb/options.h"
+#include "rocksdb/utilities/transaction_db.h"
+#include "test_util/sync_point.h"
+#include "util/cast_util.h"
+#include "util/mutexlock.h"
+#include "utilities/transactions/pessimistic_transaction.h"
+#include "utilities/transactions/transaction_db_mutex_impl.h"
+#include "utilities/transactions/write_prepared_txn_db.h"
+#include "utilities/transactions/write_unprepared_txn_db.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+PessimisticTransactionDB::PessimisticTransactionDB(
+    DB* db, const TransactionDBOptions& txn_db_options)
+    : TransactionDB(db),
+      db_impl_(static_cast_with_check<DBImpl>(db)),
+      txn_db_options_(txn_db_options),
+      lock_manager_(NewLockManager(this, txn_db_options)) {
+  assert(db_impl_ != nullptr);
+  info_log_ = db_impl_->GetDBOptions().info_log;
+}
+
+// Support initiliazing PessimisticTransactionDB from a stackable db
+//
+//    PessimisticTransactionDB
+//     ^        ^
+//     |        |
+//     |        +
+//     |   StackableDB
+//     |   ^
+//     |   |
+//     +   +
+//     DBImpl
+//       ^
+//       |(inherit)
+//       +
+//       DB
+//
+PessimisticTransactionDB::PessimisticTransactionDB(
+    StackableDB* db, const TransactionDBOptions& txn_db_options)
+    : TransactionDB(db),
+      db_impl_(static_cast_with_check<DBImpl>(db->GetRootDB())),
+      txn_db_options_(txn_db_options),
+      lock_manager_(NewLockManager(this, txn_db_options)) {
+  assert(db_impl_ != nullptr);
+}
+
+PessimisticTransactionDB::~PessimisticTransactionDB() {
+  while (!transactions_.empty()) {
+    delete transactions_.begin()->second;
+    // TODO(myabandeh): this seems to be an unsafe approach as it is not quite
+    // clear whether delete would also remove the entry from transactions_.
+  }
+}
+
+Status PessimisticTransactionDB::VerifyCFOptions(
+    const ColumnFamilyOptions& cf_options) {
+  const Comparator* const ucmp = cf_options.comparator;
+  assert(ucmp);
+  size_t ts_sz = ucmp->timestamp_size();
+  if (0 == ts_sz) {
+    return Status::OK();
+  }
+  if (ts_sz != sizeof(TxnTimestamp)) {
+    std::ostringstream oss;
+    oss << "Timestamp of transaction must have " << sizeof(TxnTimestamp)
+        << " bytes. CF comparator " << std::string(ucmp->Name())
+        << " timestamp size is " << ts_sz << " bytes";
+    return Status::InvalidArgument(oss.str());
+  }
+  if (txn_db_options_.write_policy != WRITE_COMMITTED) {
+    return Status::NotSupported("Only WriteCommittedTxn supports timestamp");
+  }
+  return Status::OK();
+}
+
+Status PessimisticTransactionDB::Initialize(
+    const std::vector<size_t>& compaction_enabled_cf_indices,
+    const std::vector<ColumnFamilyHandle*>& handles) {
+  for (auto cf_ptr : handles) {
+    AddColumnFamily(cf_ptr);
+  }
+  // Verify cf options
+  for (auto handle : handles) {
+    ColumnFamilyDescriptor cfd;
+    Status s = handle->GetDescriptor(&cfd);
+    if (!s.ok()) {
+      return s;
+    }
+    s = VerifyCFOptions(cfd.options);
+    if (!s.ok()) {
+      return s;
+    }
+  }
+
+  // Re-enable compaction for the column families that initially had
+  // compaction enabled.
+  std::vector<ColumnFamilyHandle*> compaction_enabled_cf_handles;
+  compaction_enabled_cf_handles.reserve(compaction_enabled_cf_indices.size());
+  for (auto index : compaction_enabled_cf_indices) {
+    compaction_enabled_cf_handles.push_back(handles[index]);
+  }
+
+  Status s = EnableAutoCompaction(compaction_enabled_cf_handles);
+
+  // create 'real' transactions from recovered shell transactions
+  auto dbimpl = static_cast_with_check<DBImpl>(GetRootDB());
+  assert(dbimpl != nullptr);
+  auto rtrxs = dbimpl->recovered_transactions();
+
+  for (auto it = rtrxs.begin(); it != rtrxs.end(); ++it) {
+    auto recovered_trx = it->second;
+    assert(recovered_trx);
+    assert(recovered_trx->batches_.size() == 1);
+    const auto& seq = recovered_trx->batches_.begin()->first;
+    const auto& batch_info = recovered_trx->batches_.begin()->second;
+    assert(batch_info.log_number_);
+    assert(recovered_trx->name_.length());
+
+    WriteOptions w_options;
+    w_options.sync = true;
+    TransactionOptions t_options;
+    // This would help avoiding deadlock for keys that although exist in the WAL
+    // did not go through concurrency control. This includes the merge that
+    // MyRocks uses for auto-inc columns. It is safe to do so, since (i) if
+    // there is a conflict between the keys of two transactions that must be
+    // avoided, it is already avoided by the application, MyRocks, before the
+    // restart (ii) application, MyRocks, guarntees to rollback/commit the
+    // recovered transactions before new transactions start.
+    t_options.skip_concurrency_control = true;
+
+    Transaction* real_trx = BeginTransaction(w_options, t_options, nullptr);
+    assert(real_trx);
+    real_trx->SetLogNumber(batch_info.log_number_);
+    assert(seq != kMaxSequenceNumber);
+    if (GetTxnDBOptions().write_policy != WRITE_COMMITTED) {
+      real_trx->SetId(seq);
+    }
+
+    s = real_trx->SetName(recovered_trx->name_);
+    if (!s.ok()) {
+      break;
+    }
+
+    s = real_trx->RebuildFromWriteBatch(batch_info.batch_);
+    // WriteCommitted set this to to disable this check that is specific to
+    // WritePrepared txns
+    assert(batch_info.batch_cnt_ == 0 ||
+           real_trx->GetWriteBatch()->SubBatchCnt() == batch_info.batch_cnt_);
+    real_trx->SetState(Transaction::PREPARED);
+    if (!s.ok()) {
+      break;
+    }
+  }
+  if (s.ok()) {
+    dbimpl->DeleteAllRecoveredTransactions();
+  }
+  return s;
+}
+
+Transaction* WriteCommittedTxnDB::BeginTransaction(
+    const WriteOptions& write_options, const TransactionOptions& txn_options,
+    Transaction* old_txn) {
+  if (old_txn != nullptr) {
+    ReinitializeTransaction(old_txn, write_options, txn_options);
+    return old_txn;
+  } else {
+    return new WriteCommittedTxn(this, write_options, txn_options);
+  }
+}
+
+TransactionDBOptions PessimisticTransactionDB::ValidateTxnDBOptions(
+    const TransactionDBOptions& txn_db_options) {
+  TransactionDBOptions validated = txn_db_options;
+
+  if (txn_db_options.num_stripes == 0) {
+    validated.num_stripes = 1;
+  }
+
+  return validated;
+}
+
+Status TransactionDB::Open(const Options& options,
+                           const TransactionDBOptions& txn_db_options,
+                           const std::string& dbname, TransactionDB** dbptr) {
+  DBOptions db_options(options);
+  ColumnFamilyOptions cf_options(options);
+  std::vector<ColumnFamilyDescriptor> column_families;
+  column_families.push_back(
+      ColumnFamilyDescriptor(kDefaultColumnFamilyName, cf_options));
+  std::vector<ColumnFamilyHandle*> handles;
+  Status s = TransactionDB::Open(db_options, txn_db_options, dbname,
+                                 column_families, &handles, dbptr);
+  if (s.ok()) {
+    assert(handles.size() == 1);
+    // i can delete the handle since DBImpl is always holding a reference to
+    // default column family
+    delete handles[0];
+  }
+
+  return s;
+}
+
+Status TransactionDB::Open(
+    const DBOptions& db_options, const TransactionDBOptions& txn_db_options,
+    const std::string& dbname,
+    const std::vector<ColumnFamilyDescriptor>& column_families,
+    std::vector<ColumnFamilyHandle*>* handles, TransactionDB** dbptr) {
+  Status s;
+  DB* db = nullptr;
+  if (txn_db_options.write_policy == WRITE_COMMITTED &&
+      db_options.unordered_write) {
+    return Status::NotSupported(
+        "WRITE_COMMITTED is incompatible with unordered_writes");
+  }
+  if (txn_db_options.write_policy == WRITE_UNPREPARED &&
+      db_options.unordered_write) {
+    // TODO(lth): support it
+    return Status::NotSupported(
+        "WRITE_UNPREPARED is currently incompatible with unordered_writes");
+  }
+  if (txn_db_options.write_policy == WRITE_PREPARED &&
+      db_options.unordered_write && !db_options.two_write_queues) {
+    return Status::NotSupported(
+        "WRITE_PREPARED is incompatible with unordered_writes if "
+        "two_write_queues is not enabled.");
+  }
+
+  std::vector<ColumnFamilyDescriptor> column_families_copy = column_families;
+  std::vector<size_t> compaction_enabled_cf_indices;
+  DBOptions db_options_2pc = db_options;
+  PrepareWrap(&db_options_2pc, &column_families_copy,
+              &compaction_enabled_cf_indices);
+  const bool use_seq_per_batch =
+      txn_db_options.write_policy == WRITE_PREPARED ||
+      txn_db_options.write_policy == WRITE_UNPREPARED;
+  const bool use_batch_per_txn =
+      txn_db_options.write_policy == WRITE_COMMITTED ||
+      txn_db_options.write_policy == WRITE_PREPARED;
+  s = DBImpl::Open(db_options_2pc, dbname, column_families_copy, handles, &db,
+                   use_seq_per_batch, use_batch_per_txn);
+  if (s.ok()) {
+    ROCKS_LOG_WARN(db->GetDBOptions().info_log,
+                   "Transaction write_policy is %" PRId32,
+                   static_cast<int>(txn_db_options.write_policy));
+    // if WrapDB return non-ok, db will be deleted in WrapDB() via
+    // ~StackableDB().
+    s = WrapDB(db, txn_db_options, compaction_enabled_cf_indices, *handles,
+               dbptr);
+  }
+  return s;
+}
+
+void TransactionDB::PrepareWrap(
+    DBOptions* db_options, std::vector<ColumnFamilyDescriptor>* column_families,
+    std::vector<size_t>* compaction_enabled_cf_indices) {
+  compaction_enabled_cf_indices->clear();
+
+  // Enable MemTable History if not already enabled
+  for (size_t i = 0; i < column_families->size(); i++) {
+    ColumnFamilyOptions* cf_options = &(*column_families)[i].options;
+
+    if (cf_options->max_write_buffer_size_to_maintain == 0 &&
+        cf_options->max_write_buffer_number_to_maintain == 0) {
+      // Setting to -1 will set the History size to
+      // max_write_buffer_number * write_buffer_size.
+      cf_options->max_write_buffer_size_to_maintain = -1;
+    }
+    if (!cf_options->disable_auto_compactions) {
+      // Disable compactions momentarily to prevent race with DB::Open
+      cf_options->disable_auto_compactions = true;
+      compaction_enabled_cf_indices->push_back(i);
+    }
+  }
+  db_options->allow_2pc = true;
+}
+
+namespace {
+template <typename DBType>
+Status WrapAnotherDBInternal(
+    DBType* db, const TransactionDBOptions& txn_db_options,
+    const std::vector<size_t>& compaction_enabled_cf_indices,
+    const std::vector<ColumnFamilyHandle*>& handles, TransactionDB** dbptr) {
+  assert(db != nullptr);
+  assert(dbptr != nullptr);
+  *dbptr = nullptr;
+  std::unique_ptr<PessimisticTransactionDB> txn_db;
+  // txn_db owns object pointed to by the raw db pointer.
+  switch (txn_db_options.write_policy) {
+    case WRITE_UNPREPARED:
+      txn_db.reset(new WriteUnpreparedTxnDB(
+          db, PessimisticTransactionDB::ValidateTxnDBOptions(txn_db_options)));
+      break;
+    case WRITE_PREPARED:
+      txn_db.reset(new WritePreparedTxnDB(
+          db, PessimisticTransactionDB::ValidateTxnDBOptions(txn_db_options)));
+      break;
+    case WRITE_COMMITTED:
+    default:
+      txn_db.reset(new WriteCommittedTxnDB(
+          db, PessimisticTransactionDB::ValidateTxnDBOptions(txn_db_options)));
+  }
+  txn_db->UpdateCFComparatorMap(handles);
+  Status s = txn_db->Initialize(compaction_enabled_cf_indices, handles);
+  // In case of a failure at this point, db is deleted via the txn_db destructor
+  // and set to nullptr.
+  if (s.ok()) {
+    *dbptr = txn_db.release();
+  } else {
+    for (auto* h : handles) {
+      delete h;
+    }
+    // txn_db still owns db, and ~StackableDB() will be called when txn_db goes
+    // out of scope, deleting the input db pointer.
+    ROCKS_LOG_FATAL(db->GetDBOptions().info_log,
+                    "Failed to initialize txn_db: %s", s.ToString().c_str());
+  }
+  return s;
+}
+}  // namespace
+
+Status TransactionDB::WrapDB(
+    // make sure this db is already opened with memtable history enabled,
+    // auto compaction distabled and 2 phase commit enabled
+    DB* db, const TransactionDBOptions& txn_db_options,
+    const std::vector<size_t>& compaction_enabled_cf_indices,
+    const std::vector<ColumnFamilyHandle*>& handles, TransactionDB** dbptr) {
+  return WrapAnotherDBInternal(db, txn_db_options,
+                               compaction_enabled_cf_indices, handles, dbptr);
+}
+
+Status TransactionDB::WrapStackableDB(
+    // make sure this stackable_db is already opened with memtable history
+    // enabled, auto compaction distabled and 2 phase commit enabled
+    StackableDB* db, const TransactionDBOptions& txn_db_options,
+    const std::vector<size_t>& compaction_enabled_cf_indices,
+    const std::vector<ColumnFamilyHandle*>& handles, TransactionDB** dbptr) {
+  return WrapAnotherDBInternal(db, txn_db_options,
+                               compaction_enabled_cf_indices, handles, dbptr);
+}
+
+// Let LockManager know that this column family exists so it can
+// allocate a LockMap for it.
+void PessimisticTransactionDB::AddColumnFamily(
+    const ColumnFamilyHandle* handle) {
+  lock_manager_->AddColumnFamily(handle);
+}
+
+Status PessimisticTransactionDB::CreateColumnFamily(
+    const ColumnFamilyOptions& options, const std::string& column_family_name,
+    ColumnFamilyHandle** handle) {
+  InstrumentedMutexLock l(&column_family_mutex_);
+  Status s = VerifyCFOptions(options);
+  if (!s.ok()) {
+    return s;
+  }
+
+  s = db_->CreateColumnFamily(options, column_family_name, handle);
+  if (s.ok()) {
+    lock_manager_->AddColumnFamily(*handle);
+    UpdateCFComparatorMap(*handle);
+  }
+
+  return s;
+}
+
+Status PessimisticTransactionDB::CreateColumnFamilies(
+    const ColumnFamilyOptions& options,
+    const std::vector<std::string>& column_family_names,
+    std::vector<ColumnFamilyHandle*>* handles) {
+  InstrumentedMutexLock l(&column_family_mutex_);
+
+  Status s = VerifyCFOptions(options);
+  if (!s.ok()) {
+    return s;
+  }
+
+  s = db_->CreateColumnFamilies(options, column_family_names, handles);
+  if (s.ok()) {
+    for (auto* handle : *handles) {
+      lock_manager_->AddColumnFamily(handle);
+      UpdateCFComparatorMap(handle);
+    }
+  }
+
+  return s;
+}
+
+Status PessimisticTransactionDB::CreateColumnFamilies(
+    const std::vector<ColumnFamilyDescriptor>& column_families,
+    std::vector<ColumnFamilyHandle*>* handles) {
+  InstrumentedMutexLock l(&column_family_mutex_);
+
+  for (auto& cf_desc : column_families) {
+    Status s = VerifyCFOptions(cf_desc.options);
+    if (!s.ok()) {
+      return s;
+    }
+  }
+
+  Status s = db_->CreateColumnFamilies(column_families, handles);
+  if (s.ok()) {
+    for (auto* handle : *handles) {
+      lock_manager_->AddColumnFamily(handle);
+      UpdateCFComparatorMap(handle);
+    }
+  }
+
+  return s;
+}
+
+// Let LockManager know that it can deallocate the LockMap for this
+// column family.
+Status PessimisticTransactionDB::DropColumnFamily(
+    ColumnFamilyHandle* column_family) {
+  InstrumentedMutexLock l(&column_family_mutex_);
+
+  Status s = db_->DropColumnFamily(column_family);
+  if (s.ok()) {
+    lock_manager_->RemoveColumnFamily(column_family);
+  }
+
+  return s;
+}
+
+Status PessimisticTransactionDB::DropColumnFamilies(
+    const std::vector<ColumnFamilyHandle*>& column_families) {
+  InstrumentedMutexLock l(&column_family_mutex_);
+
+  Status s = db_->DropColumnFamilies(column_families);
+  if (s.ok()) {
+    for (auto* handle : column_families) {
+      lock_manager_->RemoveColumnFamily(handle);
+    }
+  }
+
+  return s;
+}
+
+Status PessimisticTransactionDB::TryLock(PessimisticTransaction* txn,
+                                         uint32_t cfh_id,
+                                         const std::string& key,
+                                         bool exclusive) {
+  return lock_manager_->TryLock(txn, cfh_id, key, GetEnv(), exclusive);
+}
+
+Status PessimisticTransactionDB::TryRangeLock(PessimisticTransaction* txn,
+                                              uint32_t cfh_id,
+                                              const Endpoint& start_endp,
+                                              const Endpoint& end_endp) {
+  return lock_manager_->TryLock(txn, cfh_id, start_endp, end_endp, GetEnv(),
+                                /*exclusive=*/true);
+}
+
+void PessimisticTransactionDB::UnLock(PessimisticTransaction* txn,
+                                      const LockTracker& keys) {
+  lock_manager_->UnLock(txn, keys, GetEnv());
+}
+
+void PessimisticTransactionDB::UnLock(PessimisticTransaction* txn,
+                                      uint32_t cfh_id, const std::string& key) {
+  lock_manager_->UnLock(txn, cfh_id, key, GetEnv());
+}
+
+// Used when wrapping DB write operations in a transaction
+Transaction* PessimisticTransactionDB::BeginInternalTransaction(
+    const WriteOptions& options) {
+  TransactionOptions txn_options;
+  Transaction* txn = BeginTransaction(options, txn_options, nullptr);
+
+  // Use default timeout for non-transactional writes
+  txn->SetLockTimeout(txn_db_options_.default_lock_timeout);
+  return txn;
+}
+
+// All user Put, Merge, Delete, and Write requests must be intercepted to make
+// sure that they lock all keys that they are writing to avoid causing conflicts
+// with any concurrent transactions. The easiest way to do this is to wrap all
+// write operations in a transaction.
+//
+// Put(), Merge(), and Delete() only lock a single key per call.  Write() will
+// sort its keys before locking them.  This guarantees that TransactionDB write
+// methods cannot deadlock with each other (but still could deadlock with a
+// Transaction).
+Status PessimisticTransactionDB::Put(const WriteOptions& options,
+                                     ColumnFamilyHandle* column_family,
+                                     const Slice& key, const Slice& val) {
+  Status s = FailIfCfEnablesTs(this, column_family);
+  if (!s.ok()) {
+    return s;
+  }
+
+  Transaction* txn = BeginInternalTransaction(options);
+  txn->DisableIndexing();
+
+  // Since the client didn't create a transaction, they don't care about
+  // conflict checking for this write.  So we just need to do PutUntracked().
+  s = txn->PutUntracked(column_family, key, val);
+
+  if (s.ok()) {
+    s = txn->Commit();
+  }
+
+  delete txn;
+
+  return s;
+}
+
+Status PessimisticTransactionDB::Delete(const WriteOptions& wopts,
+                                        ColumnFamilyHandle* column_family,
+                                        const Slice& key) {
+  Status s = FailIfCfEnablesTs(this, column_family);
+  if (!s.ok()) {
+    return s;
+  }
+
+  Transaction* txn = BeginInternalTransaction(wopts);
+  txn->DisableIndexing();
+
+  // Since the client didn't create a transaction, they don't care about
+  // conflict checking for this write.  So we just need to do
+  // DeleteUntracked().
+  s = txn->DeleteUntracked(column_family, key);
+
+  if (s.ok()) {
+    s = txn->Commit();
+  }
+
+  delete txn;
+
+  return s;
+}
+
+Status PessimisticTransactionDB::SingleDelete(const WriteOptions& wopts,
+                                              ColumnFamilyHandle* column_family,
+                                              const Slice& key) {
+  Status s = FailIfCfEnablesTs(this, column_family);
+  if (!s.ok()) {
+    return s;
+  }
+
+  Transaction* txn = BeginInternalTransaction(wopts);
+  txn->DisableIndexing();
+
+  // Since the client didn't create a transaction, they don't care about
+  // conflict checking for this write.  So we just need to do
+  // SingleDeleteUntracked().
+  s = txn->SingleDeleteUntracked(column_family, key);
+
+  if (s.ok()) {
+    s = txn->Commit();
+  }
+
+  delete txn;
+
+  return s;
+}
+
+Status PessimisticTransactionDB::Merge(const WriteOptions& options,
+                                       ColumnFamilyHandle* column_family,
+                                       const Slice& key, const Slice& value) {
+  Status s = FailIfCfEnablesTs(this, column_family);
+  if (!s.ok()) {
+    return s;
+  }
+
+  Transaction* txn = BeginInternalTransaction(options);
+  txn->DisableIndexing();
+
+  // Since the client didn't create a transaction, they don't care about
+  // conflict checking for this write.  So we just need to do
+  // MergeUntracked().
+  s = txn->MergeUntracked(column_family, key, value);
+
+  if (s.ok()) {
+    s = txn->Commit();
+  }
+
+  delete txn;
+
+  return s;
+}
+
+Status PessimisticTransactionDB::Write(const WriteOptions& opts,
+                                       WriteBatch* updates) {
+  return WriteWithConcurrencyControl(opts, updates);
+}
+
+Status WriteCommittedTxnDB::Write(const WriteOptions& opts,
+                                  WriteBatch* updates) {
+  Status s = FailIfBatchHasTs(updates);
+  if (!s.ok()) {
+    return s;
+  }
+  if (txn_db_options_.skip_concurrency_control) {
+    return db_impl_->Write(opts, updates);
+  } else {
+    return WriteWithConcurrencyControl(opts, updates);
+  }
+}
+
+Status WriteCommittedTxnDB::Write(
+    const WriteOptions& opts,
+    const TransactionDBWriteOptimizations& optimizations, WriteBatch* updates) {
+  Status s = FailIfBatchHasTs(updates);
+  if (!s.ok()) {
+    return s;
+  }
+  if (optimizations.skip_concurrency_control) {
+    return db_impl_->Write(opts, updates);
+  } else {
+    return WriteWithConcurrencyControl(opts, updates);
+  }
+}
+
+void PessimisticTransactionDB::InsertExpirableTransaction(
+    TransactionID tx_id, PessimisticTransaction* tx) {
+  assert(tx->GetExpirationTime() > 0);
+  std::lock_guard<std::mutex> lock(map_mutex_);
+  expirable_transactions_map_.insert({tx_id, tx});
+}
+
+void PessimisticTransactionDB::RemoveExpirableTransaction(TransactionID tx_id) {
+  std::lock_guard<std::mutex> lock(map_mutex_);
+  expirable_transactions_map_.erase(tx_id);
+}
+
+bool PessimisticTransactionDB::TryStealingExpiredTransactionLocks(
+    TransactionID tx_id) {
+  std::lock_guard<std::mutex> lock(map_mutex_);
+
+  auto tx_it = expirable_transactions_map_.find(tx_id);
+  if (tx_it == expirable_transactions_map_.end()) {
+    return true;
+  }
+  PessimisticTransaction& tx = *(tx_it->second);
+  return tx.TryStealingLocks();
+}
+
+void PessimisticTransactionDB::ReinitializeTransaction(
+    Transaction* txn, const WriteOptions& write_options,
+    const TransactionOptions& txn_options) {
+  auto txn_impl = static_cast_with_check<PessimisticTransaction>(txn);
+
+  txn_impl->Reinitialize(this, write_options, txn_options);
+}
+
+Transaction* PessimisticTransactionDB::GetTransactionByName(
+    const TransactionName& name) {
+  std::lock_guard<std::mutex> lock(name_map_mutex_);
+  auto it = transactions_.find(name);
+  if (it == transactions_.end()) {
+    return nullptr;
+  } else {
+    return it->second;
+  }
+}
+
+void PessimisticTransactionDB::GetAllPreparedTransactions(
+    std::vector<Transaction*>* transv) {
+  assert(transv);
+  transv->clear();
+  std::lock_guard<std::mutex> lock(name_map_mutex_);
+  for (auto it = transactions_.begin(); it != transactions_.end(); ++it) {
+    if (it->second->GetState() == Transaction::PREPARED) {
+      transv->push_back(it->second);
+    }
+  }
+}
+
+LockManager::PointLockStatus PessimisticTransactionDB::GetLockStatusData() {
+  return lock_manager_->GetPointLockStatus();
+}
+
+std::vector<DeadlockPath> PessimisticTransactionDB::GetDeadlockInfoBuffer() {
+  return lock_manager_->GetDeadlockInfoBuffer();
+}
+
+void PessimisticTransactionDB::SetDeadlockInfoBufferSize(uint32_t target_size) {
+  lock_manager_->Resize(target_size);
+}
+
+void PessimisticTransactionDB::RegisterTransaction(Transaction* txn) {
+  assert(txn);
+  assert(txn->GetName().length() > 0);
+  assert(GetTransactionByName(txn->GetName()) == nullptr);
+  assert(txn->GetState() == Transaction::STARTED);
+  std::lock_guard<std::mutex> lock(name_map_mutex_);
+  transactions_[txn->GetName()] = txn;
+}
+
+void PessimisticTransactionDB::UnregisterTransaction(Transaction* txn) {
+  assert(txn);
+  std::lock_guard<std::mutex> lock(name_map_mutex_);
+  auto it = transactions_.find(txn->GetName());
+  assert(it != transactions_.end());
+  transactions_.erase(it);
+}
+
+std::pair<Status, std::shared_ptr<const Snapshot>>
+PessimisticTransactionDB::CreateTimestampedSnapshot(TxnTimestamp ts) {
+  if (kMaxTxnTimestamp == ts) {
+    return std::make_pair(Status::InvalidArgument("invalid ts"), nullptr);
+  }
+  assert(db_impl_);
+  return db_impl_->CreateTimestampedSnapshot(kMaxSequenceNumber, ts);
+}
+
+std::shared_ptr<const Snapshot>
+PessimisticTransactionDB::GetTimestampedSnapshot(TxnTimestamp ts) const {
+  assert(db_impl_);
+  return db_impl_->GetTimestampedSnapshot(ts);
+}
+
+void PessimisticTransactionDB::ReleaseTimestampedSnapshotsOlderThan(
+    TxnTimestamp ts) {
+  assert(db_impl_);
+  db_impl_->ReleaseTimestampedSnapshotsOlderThan(ts);
+}
+
+Status PessimisticTransactionDB::GetTimestampedSnapshots(
+    TxnTimestamp ts_lb, TxnTimestamp ts_ub,
+    std::vector<std::shared_ptr<const Snapshot>>& timestamped_snapshots) const {
+  assert(db_impl_);
+  return db_impl_->GetTimestampedSnapshots(ts_lb, ts_ub, timestamped_snapshots);
+}
+
+Status SnapshotCreationCallback::operator()(SequenceNumber seq,
+                                            bool disable_memtable) {
+  assert(db_impl_);
+  assert(commit_ts_ != kMaxTxnTimestamp);
+
+  const bool two_write_queues =
+      db_impl_->immutable_db_options().two_write_queues;
+  assert(!two_write_queues || !disable_memtable);
+#ifdef NDEBUG
+  (void)two_write_queues;
+  (void)disable_memtable;
+#endif
+
+  const bool seq_per_batch = db_impl_->seq_per_batch();
+  if (!seq_per_batch) {
+    assert(db_impl_->GetLastPublishedSequence() <= seq);
+  } else {
+    assert(db_impl_->GetLastPublishedSequence() < seq);
+  }
+
+  // Create a snapshot which can also be used for write conflict checking.
+  auto ret = db_impl_->CreateTimestampedSnapshot(seq, commit_ts_);
+  snapshot_creation_status_ = ret.first;
+  snapshot_ = ret.second;
+  if (snapshot_creation_status_.ok()) {
+    assert(snapshot_);
+  } else {
+    assert(!snapshot_);
+  }
+  if (snapshot_ && snapshot_notifier_) {
+    snapshot_notifier_->SnapshotCreated(snapshot_.get());
+  }
+  return Status::OK();
+}
+
+}  // namespace ROCKSDB_NAMESPACE
+#endif  // ROCKSDB_LITE
diff --git a/src/rocksdb/utilities/transactions/pessimistic_transaction_db.h b/src/rocksdb/utilities/transactions/pessimistic_transaction_db.h
new file mode 100644
index 000000000..25cd11054
--- /dev/null
+++ b/src/rocksdb/utilities/transactions/pessimistic_transaction_db.h
@@ -0,0 +1,318 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#pragma once
+#ifndef ROCKSDB_LITE
+
+#include <mutex>
+#include <queue>
+#include <set>
+#include <string>
+#include <unordered_map>
+#include <vector>
+
+#include "db/db_iter.h"
+#include "db/read_callback.h"
+#include "db/snapshot_checker.h"
+#include "rocksdb/db.h"
+#include "rocksdb/options.h"
+#include "rocksdb/utilities/transaction_db.h"
+#include "util/cast_util.h"
+#include "utilities/transactions/lock/lock_manager.h"
+#include "utilities/transactions/lock/range/range_lock_manager.h"
+#include "utilities/transactions/pessimistic_transaction.h"
+#include "utilities/transactions/write_prepared_txn.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+class PessimisticTransactionDB : public TransactionDB {
+ public:
+  explicit PessimisticTransactionDB(DB* db,
+                                    const TransactionDBOptions& txn_db_options);
+
+  explicit PessimisticTransactionDB(StackableDB* db,
+                                    const TransactionDBOptions& txn_db_options);
+
+  virtual ~PessimisticTransactionDB();
+
+  virtual const Snapshot* GetSnapshot() override { return db_->GetSnapshot(); }
+
+  virtual Status Initialize(
+      const std::vector<size_t>& compaction_enabled_cf_indices,
+      const std::vector<ColumnFamilyHandle*>& handles);
+
+  Transaction* BeginTransaction(const WriteOptions& write_options,
+                                const TransactionOptions& txn_options,
+                                Transaction* old_txn) override = 0;
+
+  using StackableDB::Put;
+  virtual Status Put(const WriteOptions& options,
+                     ColumnFamilyHandle* column_family, const Slice& key,
+                     const Slice& val) override;
+
+  using StackableDB::Delete;
+  virtual Status Delete(const WriteOptions& wopts,
+                        ColumnFamilyHandle* column_family,
+                        const Slice& key) override;
+
+  using StackableDB::SingleDelete;
+  virtual Status SingleDelete(const WriteOptions& wopts,
+                              ColumnFamilyHandle* column_family,
+                              const Slice& key) override;
+
+  using StackableDB::Merge;
+  virtual Status Merge(const WriteOptions& options,
+                       ColumnFamilyHandle* column_family, const Slice& key,
+                       const Slice& value) override;
+
+  using TransactionDB::Write;
+  virtual Status Write(const WriteOptions& opts, WriteBatch* updates) override;
+  inline Status WriteWithConcurrencyControl(const WriteOptions& opts,
+                                            WriteBatch* updates) {
+    Status s;
+    if (opts.protection_bytes_per_key > 0) {
+      s = WriteBatchInternal::UpdateProtectionInfo(
+          updates, opts.protection_bytes_per_key);
+    }
+    if (s.ok()) {
+      // Need to lock all keys in this batch to prevent write conflicts with
+      // concurrent transactions.
+      Transaction* txn = BeginInternalTransaction(opts);
+      txn->DisableIndexing();
+
+      auto txn_impl = static_cast_with_check<PessimisticTransaction>(txn);
+
+      // Since commitBatch sorts the keys before locking, concurrent Write()
+      // operations will not cause a deadlock.
+      // In order to avoid a deadlock with a concurrent Transaction,
+      // Transactions should use a lock timeout.
+      s = txn_impl->CommitBatch(updates);
+
+      delete txn;
+    }
+
+    return s;
+  }
+
+  using StackableDB::CreateColumnFamily;
+  virtual Status CreateColumnFamily(const ColumnFamilyOptions& options,
+                                    const std::string& column_family_name,
+                                    ColumnFamilyHandle** handle) override;
+
+  Status CreateColumnFamilies(
+      const ColumnFamilyOptions& options,
+      const std::vector<std::string>& column_family_names,
+      std::vector<ColumnFamilyHandle*>* handles) override;
+
+  Status CreateColumnFamilies(
+      const std::vector<ColumnFamilyDescriptor>& column_families,
+      std::vector<ColumnFamilyHandle*>* handles) override;
+
+  using StackableDB::DropColumnFamily;
+  virtual Status DropColumnFamily(ColumnFamilyHandle* column_family) override;
+
+  Status DropColumnFamilies(
+      const std::vector<ColumnFamilyHandle*>& column_families) override;
+
+  Status TryLock(PessimisticTransaction* txn, uint32_t cfh_id,
+                 const std::string& key, bool exclusive);
+  Status TryRangeLock(PessimisticTransaction* txn, uint32_t cfh_id,
+                      const Endpoint& start_endp, const Endpoint& end_endp);
+
+  void UnLock(PessimisticTransaction* txn, const LockTracker& keys);
+  void UnLock(PessimisticTransaction* txn, uint32_t cfh_id,
+              const std::string& key);
+
+  void AddColumnFamily(const ColumnFamilyHandle* handle);
+
+  static TransactionDBOptions ValidateTxnDBOptions(
+      const TransactionDBOptions& txn_db_options);
+
+  const TransactionDBOptions& GetTxnDBOptions() const {
+    return txn_db_options_;
+  }
+
+  void InsertExpirableTransaction(TransactionID tx_id,
+                                  PessimisticTransaction* tx);
+  void RemoveExpirableTransaction(TransactionID tx_id);
+
+  // If transaction is no longer available, locks can be stolen
+  // If transaction is available, try stealing locks directly from transaction
+  // It is the caller's responsibility to ensure that the referred transaction
+  // is expirable (GetExpirationTime() > 0) and that it is expired.
+  bool TryStealingExpiredTransactionLocks(TransactionID tx_id);
+
+  Transaction* GetTransactionByName(const TransactionName& name) override;
+
+  void RegisterTransaction(Transaction* txn);
+  void UnregisterTransaction(Transaction* txn);
+
+  // not thread safe. current use case is during recovery (single thread)
+  void GetAllPreparedTransactions(std::vector<Transaction*>* trans) override;
+
+  LockManager::PointLockStatus GetLockStatusData() override;
+
+  std::vector<DeadlockPath> GetDeadlockInfoBuffer() override;
+  void SetDeadlockInfoBufferSize(uint32_t target_size) override;
+
+  // The default implementation does nothing. The actual implementation is moved
+  // to the child classes that actually need this information. This was due to
+  // an odd performance drop we observed when the added std::atomic member to
+  // the base class even when the subclass do not read it in the fast path.
+  virtual void UpdateCFComparatorMap(const std::vector<ColumnFamilyHandle*>&) {}
+  virtual void UpdateCFComparatorMap(ColumnFamilyHandle*) {}
+
+  // Use the returned factory to create LockTrackers in transactions.
+  const LockTrackerFactory& GetLockTrackerFactory() const {
+    return lock_manager_->GetLockTrackerFactory();
+  }
+
+  std::pair<Status, std::shared_ptr<const Snapshot>> CreateTimestampedSnapshot(
+      TxnTimestamp ts) override;
+
+  std::shared_ptr<const Snapshot> GetTimestampedSnapshot(
+      TxnTimestamp ts) const override;
+
+  void ReleaseTimestampedSnapshotsOlderThan(TxnTimestamp ts) override;
+
+  Status GetTimestampedSnapshots(TxnTimestamp ts_lb, TxnTimestamp ts_ub,
+                                 std::vector<std::shared_ptr<const Snapshot>>&
+                                     timestamped_snapshots) const override;
+
+ protected:
+  DBImpl* db_impl_;
+  std::shared_ptr<Logger> info_log_;
+  const TransactionDBOptions txn_db_options_;
+
+  static Status FailIfBatchHasTs(const WriteBatch* wb);
+
+  static Status FailIfCfEnablesTs(const DB* db,
+                                  const ColumnFamilyHandle* column_family);
+
+  void ReinitializeTransaction(
+      Transaction* txn, const WriteOptions& write_options,
+      const TransactionOptions& txn_options = TransactionOptions());
+
+  virtual Status VerifyCFOptions(const ColumnFamilyOptions& cf_options);
+
+ private:
+  friend class WritePreparedTxnDB;
+  friend class WritePreparedTxnDBMock;
+  friend class WriteUnpreparedTxn;
+  friend class TransactionTest_DoubleCrashInRecovery_Test;
+  friend class TransactionTest_DoubleEmptyWrite_Test;
+  friend class TransactionTest_DuplicateKeys_Test;
+  friend class TransactionTest_PersistentTwoPhaseTransactionTest_Test;
+  friend class TransactionTest_TwoPhaseDoubleRecoveryTest_Test;
+  friend class TransactionTest_TwoPhaseOutOfOrderDelete_Test;
+  friend class TransactionStressTest_TwoPhaseLongPrepareTest_Test;
+  friend class WriteUnpreparedTransactionTest_RecoveryTest_Test;
+  friend class WriteUnpreparedTransactionTest_MarkLogWithPrepSection_Test;
+
+  Transaction* BeginInternalTransaction(const WriteOptions& options);
+
+  std::shared_ptr<LockManager> lock_manager_;
+
+  // Must be held when adding/dropping column families.
+  InstrumentedMutex column_family_mutex_;
+
+  // Used to ensure that no locks are stolen from an expirable transaction
+  // that has started a commit. Only transactions with an expiration time
+  // should be in this map.
+  std::mutex map_mutex_;
+  std::unordered_map<TransactionID, PessimisticTransaction*>
+      expirable_transactions_map_;
+
+  // map from name to two phase transaction instance
+  std::mutex name_map_mutex_;
+  std::unordered_map<TransactionName, Transaction*> transactions_;
+
+  // Signal that we are testing a crash scenario. Some asserts could be relaxed
+  // in such cases.
+  virtual void TEST_Crash() {}
+};
+
+// A PessimisticTransactionDB that writes the data to the DB after the commit.
+// In this way the DB only contains the committed data.
+class WriteCommittedTxnDB : public PessimisticTransactionDB {
+ public:
+  explicit WriteCommittedTxnDB(DB* db,
+                               const TransactionDBOptions& txn_db_options)
+      : PessimisticTransactionDB(db, txn_db_options) {}
+
+  explicit WriteCommittedTxnDB(StackableDB* db,
+                               const TransactionDBOptions& txn_db_options)
+      : PessimisticTransactionDB(db, txn_db_options) {}
+
+  virtual ~WriteCommittedTxnDB() {}
+
+  Transaction* BeginTransaction(const WriteOptions& write_options,
+                                const TransactionOptions& txn_options,
+                                Transaction* old_txn) override;
+
+  // Optimized version of ::Write that makes use of skip_concurrency_control
+  // hint
+  using TransactionDB::Write;
+  virtual Status Write(const WriteOptions& opts,
+                       const TransactionDBWriteOptimizations& optimizations,
+                       WriteBatch* updates) override;
+  virtual Status Write(const WriteOptions& opts, WriteBatch* updates) override;
+};
+
+inline Status PessimisticTransactionDB::FailIfBatchHasTs(
+    const WriteBatch* batch) {
+  if (batch != nullptr && WriteBatchInternal::HasKeyWithTimestamp(*batch)) {
+    return Status::NotSupported(
+        "Writes with timestamp must go through transaction API instead of "
+        "TransactionDB.");
+  }
+  return Status::OK();
+}
+
+inline Status PessimisticTransactionDB::FailIfCfEnablesTs(
+    const DB* db, const ColumnFamilyHandle* column_family) {
+  assert(db);
+  column_family = column_family ? column_family : db->DefaultColumnFamily();
+  assert(column_family);
+  const Comparator* const ucmp = column_family->GetComparator();
+  assert(ucmp);
+  if (ucmp->timestamp_size() > 0) {
+    return Status::NotSupported(
+        "Write operation with user timestamp must go through the transaction "
+        "API instead of TransactionDB.");
+  }
+  return Status::OK();
+}
+
+class SnapshotCreationCallback : public PostMemTableCallback {
+ public:
+  explicit SnapshotCreationCallback(
+      DBImpl* dbi, TxnTimestamp commit_ts,
+      const std::shared_ptr<TransactionNotifier>& notifier,
+      std::shared_ptr<const Snapshot>& snapshot)
+      : db_impl_(dbi),
+        commit_ts_(commit_ts),
+        snapshot_notifier_(notifier),
+        snapshot_(snapshot) {
+    assert(db_impl_);
+  }
+
+  ~SnapshotCreationCallback() override {
+    snapshot_creation_status_.PermitUncheckedError();
+  }
+
+  Status operator()(SequenceNumber seq, bool disable_memtable) override;
+
+ private:
+  DBImpl* const db_impl_;
+  const TxnTimestamp commit_ts_;
+  std::shared_ptr<TransactionNotifier> snapshot_notifier_;
+  std::shared_ptr<const Snapshot>& snapshot_;
+
+  Status snapshot_creation_status_;
+};
+
+}  // namespace ROCKSDB_NAMESPACE
+#endif  // ROCKSDB_LITE
diff --git a/src/rocksdb/utilities/transactions/snapshot_checker.cc b/src/rocksdb/utilities/transactions/snapshot_checker.cc
new file mode 100644
index 000000000..76d16681a
--- /dev/null
+++ b/src/rocksdb/utilities/transactions/snapshot_checker.cc
@@ -0,0 +1,53 @@
+// Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#include "db/snapshot_checker.h"
+
+#ifdef ROCKSDB_LITE
+#include <assert.h>
+#endif  // ROCKSDB_LITE
+
+#include "port/lang.h"
+#include "utilities/transactions/write_prepared_txn_db.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+#ifdef ROCKSDB_LITE
+WritePreparedSnapshotChecker::WritePreparedSnapshotChecker(
+    WritePreparedTxnDB* /*txn_db*/) {}
+
+SnapshotCheckerResult WritePreparedSnapshotChecker::CheckInSnapshot(
+    SequenceNumber /*sequence*/, SequenceNumber /*snapshot_sequence*/) const {
+  // Should never be called in LITE mode.
+  assert(false);
+  return SnapshotCheckerResult::kInSnapshot;
+}
+
+#else
+
+WritePreparedSnapshotChecker::WritePreparedSnapshotChecker(
+    WritePreparedTxnDB* txn_db)
+    : txn_db_(txn_db){};
+
+SnapshotCheckerResult WritePreparedSnapshotChecker::CheckInSnapshot(
+    SequenceNumber sequence, SequenceNumber snapshot_sequence) const {
+  bool snapshot_released = false;
+  // TODO(myabandeh): set min_uncommitted
+  bool in_snapshot = txn_db_->IsInSnapshot(
+      sequence, snapshot_sequence, kMinUnCommittedSeq, &snapshot_released);
+  if (snapshot_released) {
+    return SnapshotCheckerResult::kSnapshotReleased;
+  }
+  return in_snapshot ? SnapshotCheckerResult::kInSnapshot
+                     : SnapshotCheckerResult::kNotInSnapshot;
+}
+
+#endif  // ROCKSDB_LITE
+
+DisableGCSnapshotChecker* DisableGCSnapshotChecker::Instance() {
+  STATIC_AVOID_DESTRUCTION(DisableGCSnapshotChecker, instance);
+  return &instance;
+}
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/utilities/transactions/timestamped_snapshot_test.cc b/src/rocksdb/utilities/transactions/timestamped_snapshot_test.cc
new file mode 100644
index 000000000..e9b474415
--- /dev/null
+++ b/src/rocksdb/utilities/transactions/timestamped_snapshot_test.cc
@@ -0,0 +1,466 @@
+//  Copyright (c) Meta Platforms, Inc. and affiliates. All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#ifdef ROCKSDB_LITE
+#include <cstdio>
+
+int main(int /*argc*/, char** /*argv*/) {
+  fprintf(stderr, "SKIPPED as Transactions are not supported in LITE mode\n");
+  return 0;
+}
+#else  // ROCKSDB_LITE
+#include <cassert>
+
+#include "util/cast_util.h"
+#include "utilities/transactions/transaction_test.h"
+
+namespace ROCKSDB_NAMESPACE {
+INSTANTIATE_TEST_CASE_P(
+    Unsupported, TimestampedSnapshotWithTsSanityCheck,
+    ::testing::Values(
+        std::make_tuple(false, false, WRITE_PREPARED, kOrderedWrite),
+        std::make_tuple(false, true, WRITE_PREPARED, kUnorderedWrite),
+        std::make_tuple(false, false, WRITE_UNPREPARED, kOrderedWrite)));
+
+INSTANTIATE_TEST_CASE_P(WriteCommitted, TransactionTest,
+                        ::testing::Combine(::testing::Bool(), ::testing::Bool(),
+                                           ::testing::Values(WRITE_COMMITTED),
+                                           ::testing::Values(kOrderedWrite)));
+
+namespace {
+// Not thread-safe. Caller needs to provide external synchronization.
+class TsCheckingTxnNotifier : public TransactionNotifier {
+ public:
+  explicit TsCheckingTxnNotifier() = default;
+
+  ~TsCheckingTxnNotifier() override {}
+
+  void SnapshotCreated(const Snapshot* new_snapshot) override {
+    assert(new_snapshot);
+    if (prev_snapshot_seq_ != kMaxSequenceNumber) {
+      assert(prev_snapshot_seq_ <= new_snapshot->GetSequenceNumber());
+    }
+    prev_snapshot_seq_ = new_snapshot->GetSequenceNumber();
+    if (prev_snapshot_ts_ != kMaxTxnTimestamp) {
+      assert(prev_snapshot_ts_ <= new_snapshot->GetTimestamp());
+    }
+    prev_snapshot_ts_ = new_snapshot->GetTimestamp();
+  }
+
+  TxnTimestamp prev_snapshot_ts() const { return prev_snapshot_ts_; }
+
+ private:
+  SequenceNumber prev_snapshot_seq_ = kMaxSequenceNumber;
+  TxnTimestamp prev_snapshot_ts_ = kMaxTxnTimestamp;
+};
+}  // anonymous namespace
+
+TEST_P(TimestampedSnapshotWithTsSanityCheck, WithoutCommitTs) {
+  std::unique_ptr<Transaction> txn(
+      db->BeginTransaction(WriteOptions(), TransactionOptions()));
+  assert(txn);
+  ASSERT_OK(txn->SetName("txn0"));
+  ASSERT_OK(txn->Put("a", "v"));
+  ASSERT_OK(txn->Prepare());
+  Status s = txn->CommitAndTryCreateSnapshot();
+  ASSERT_TRUE(s.IsInvalidArgument());
+  ASSERT_OK(txn->Rollback());
+
+  txn.reset(db->BeginTransaction(WriteOptions(), TransactionOptions()));
+  assert(txn);
+  ASSERT_OK(txn->SetName("txn0"));
+  ASSERT_OK(txn->Put("a", "v"));
+  s = txn->CommitAndTryCreateSnapshot();
+  ASSERT_TRUE(s.IsInvalidArgument());
+}
+
+TEST_P(TimestampedSnapshotWithTsSanityCheck, SetCommitTs) {
+  std::unique_ptr<Transaction> txn(
+      db->BeginTransaction(WriteOptions(), TransactionOptions()));
+  assert(txn);
+  ASSERT_OK(txn->SetName("txn0"));
+  ASSERT_OK(txn->Put("a", "v"));
+  ASSERT_OK(txn->Prepare());
+  std::shared_ptr<const Snapshot> snapshot;
+  Status s = txn->CommitAndTryCreateSnapshot(nullptr, 10, &snapshot);
+  ASSERT_TRUE(s.IsNotSupported());
+  ASSERT_OK(txn->Rollback());
+
+  txn.reset(db->BeginTransaction(WriteOptions(), TransactionOptions()));
+  assert(txn);
+  ASSERT_OK(txn->SetName("txn0"));
+  ASSERT_OK(txn->Put("a", "v"));
+  s = txn->CommitAndTryCreateSnapshot(nullptr, 10, &snapshot);
+  ASSERT_TRUE(s.IsNotSupported());
+}
+
+TEST_P(TransactionTest, WithoutCommitTs) {
+  std::unique_ptr<Transaction> txn(
+      db->BeginTransaction(WriteOptions(), TransactionOptions()));
+  assert(txn);
+  ASSERT_OK(txn->SetName("txn0"));
+  ASSERT_OK(txn->Put("a", "v"));
+  ASSERT_OK(txn->Prepare());
+  Status s = txn->CommitAndTryCreateSnapshot();
+  ASSERT_TRUE(s.IsInvalidArgument());
+  ASSERT_OK(txn->Rollback());
+
+  txn.reset(db->BeginTransaction(WriteOptions(), TransactionOptions()));
+  assert(txn);
+  ASSERT_OK(txn->SetName("txn0"));
+  ASSERT_OK(txn->Put("a", "v"));
+  s = txn->CommitAndTryCreateSnapshot();
+  ASSERT_TRUE(s.IsInvalidArgument());
+}
+
+TEST_P(TransactionTest, ReuseExistingTxn) {
+  Transaction* txn = db->BeginTransaction(WriteOptions(), TransactionOptions());
+  assert(txn);
+  ASSERT_OK(txn->SetName("txn0"));
+  ASSERT_OK(txn->Put("a", "v1"));
+  ASSERT_OK(txn->Prepare());
+
+  auto notifier = std::make_shared<TsCheckingTxnNotifier>();
+  std::shared_ptr<const Snapshot> snapshot1;
+  Status s =
+      txn->CommitAndTryCreateSnapshot(notifier, /*commit_ts=*/100, &snapshot1);
+  ASSERT_OK(s);
+  ASSERT_EQ(100, snapshot1->GetTimestamp());
+
+  Transaction* txn1 =
+      db->BeginTransaction(WriteOptions(), TransactionOptions(), txn);
+  assert(txn1 == txn);
+  ASSERT_OK(txn1->SetName("txn1"));
+  ASSERT_OK(txn->Put("a", "v2"));
+  ASSERT_OK(txn->Prepare());
+  std::shared_ptr<const Snapshot> snapshot2;
+  s = txn->CommitAndTryCreateSnapshot(notifier, /*commit_ts=*/110, &snapshot2);
+  ASSERT_OK(s);
+  ASSERT_EQ(110, snapshot2->GetTimestamp());
+  delete txn;
+
+  {
+    std::string value;
+    ReadOptions read_opts;
+    read_opts.snapshot = snapshot1.get();
+    ASSERT_OK(db->Get(read_opts, "a", &value));
+    ASSERT_EQ("v1", value);
+
+    read_opts.snapshot = snapshot2.get();
+    ASSERT_OK(db->Get(read_opts, "a", &value));
+    ASSERT_EQ("v2", value);
+  }
+}
+
+TEST_P(TransactionTest, CreateSnapshotWhenCommit) {
+  std::unique_ptr<Transaction> txn(
+      db->BeginTransaction(WriteOptions(), TransactionOptions()));
+  assert(txn);
+
+  constexpr int batch_size = 10;
+  for (int i = 0; i < batch_size; ++i) {
+    ASSERT_OK(db->Put(WriteOptions(), "k" + std::to_string(i), "v0"));
+  }
+  const SequenceNumber seq0 = db->GetLatestSequenceNumber();
+  ASSERT_EQ(static_cast<SequenceNumber>(batch_size), seq0);
+
+  txn->SetSnapshot();
+  {
+    const Snapshot* const snapshot = txn->GetSnapshot();
+    assert(snapshot);
+    ASSERT_EQ(seq0, snapshot->GetSequenceNumber());
+  }
+
+  for (int i = 0; i < batch_size; ++i) {
+    ASSERT_OK(txn->Put("k" + std::to_string(i), "v1"));
+  }
+  ASSERT_OK(txn->SetName("txn0"));
+  ASSERT_OK(txn->Prepare());
+
+  std::shared_ptr<const Snapshot> snapshot;
+  constexpr TxnTimestamp timestamp = 1;
+  auto notifier = std::make_shared<TsCheckingTxnNotifier>();
+  Status s = txn->CommitAndTryCreateSnapshot(notifier, timestamp, &snapshot);
+  ASSERT_OK(s);
+  ASSERT_LT(notifier->prev_snapshot_ts(), kMaxTxnTimestamp);
+  assert(snapshot);
+  ASSERT_EQ(timestamp, snapshot->GetTimestamp());
+  ASSERT_EQ(seq0 + batch_size, snapshot->GetSequenceNumber());
+  const Snapshot* const raw_snapshot_ptr = txn->GetSnapshot();
+  ASSERT_EQ(raw_snapshot_ptr, snapshot.get());
+  ASSERT_EQ(snapshot, txn->GetTimestampedSnapshot());
+
+  {
+    std::shared_ptr<const Snapshot> snapshot1 =
+        db->GetLatestTimestampedSnapshot();
+    ASSERT_EQ(snapshot, snapshot1);
+  }
+  {
+    std::shared_ptr<const Snapshot> snapshot1 =
+        db->GetTimestampedSnapshot(timestamp);
+    ASSERT_EQ(snapshot, snapshot1);
+  }
+  {
+    std::vector<std::shared_ptr<const Snapshot> > snapshots;
+    s = db->GetAllTimestampedSnapshots(snapshots);
+    ASSERT_OK(s);
+    ASSERT_EQ(std::vector<std::shared_ptr<const Snapshot> >{snapshot},
+              snapshots);
+  }
+}
+
+TEST_P(TransactionTest, CreateSnapshot) {
+  // First create a non-timestamped snapshot
+  ManagedSnapshot snapshot_guard(db);
+  for (int i = 0; i < 10; ++i) {
+    ASSERT_OK(db->Put(WriteOptions(), "k" + std::to_string(i),
+                      "v0_" + std::to_string(i)));
+  }
+  {
+    auto ret = db->CreateTimestampedSnapshot(kMaxTxnTimestamp);
+    ASSERT_TRUE(ret.first.IsInvalidArgument());
+    auto snapshot = ret.second;
+    ASSERT_EQ(nullptr, snapshot.get());
+  }
+  constexpr TxnTimestamp timestamp = 100;
+  Status s;
+  std::shared_ptr<const Snapshot> ts_snap0;
+  std::tie(s, ts_snap0) = db->CreateTimestampedSnapshot(timestamp);
+  ASSERT_OK(s);
+  assert(ts_snap0);
+  ASSERT_EQ(timestamp, ts_snap0->GetTimestamp());
+  for (int i = 0; i < 10; ++i) {
+    ASSERT_OK(db->Delete(WriteOptions(), "k" + std::to_string(i)));
+  }
+  {
+    ReadOptions read_opts;
+    read_opts.snapshot = ts_snap0.get();
+    for (int i = 0; i < 10; ++i) {
+      std::string value;
+      s = db->Get(read_opts, "k" + std::to_string(i), &value);
+      ASSERT_OK(s);
+      ASSERT_EQ("v0_" + std::to_string(i), value);
+    }
+  }
+  {
+    std::shared_ptr<const Snapshot> snapshot =
+        db->GetLatestTimestampedSnapshot();
+    ASSERT_EQ(ts_snap0, snapshot);
+  }
+  {
+    std::shared_ptr<const Snapshot> snapshot =
+        db->GetTimestampedSnapshot(timestamp);
+    ASSERT_OK(s);
+    ASSERT_EQ(ts_snap0, snapshot);
+  }
+  {
+    std::vector<std::shared_ptr<const Snapshot> > snapshots;
+    s = db->GetAllTimestampedSnapshots(snapshots);
+    ASSERT_OK(s);
+    ASSERT_EQ(std::vector<std::shared_ptr<const Snapshot> >{ts_snap0},
+              snapshots);
+  }
+}
+
+TEST_P(TransactionTest, SequenceAndTsOrder) {
+  Status s;
+  std::shared_ptr<const Snapshot> snapshot;
+  std::tie(s, snapshot) = db->CreateTimestampedSnapshot(100);
+  ASSERT_OK(s);
+  assert(snapshot);
+  {
+    // Cannot request smaller timestamp for the new timestamped snapshot.
+    std::shared_ptr<const Snapshot> tmp_snapshot;
+    std::tie(s, tmp_snapshot) = db->CreateTimestampedSnapshot(50);
+    ASSERT_TRUE(s.IsInvalidArgument());
+    ASSERT_EQ(nullptr, tmp_snapshot.get());
+  }
+
+  // If requesting a new timestamped snapshot with the same timestamp and
+  // sequence number, we avoid creating new snapshot object but reuse
+  // exisisting one.
+  std::shared_ptr<const Snapshot> snapshot1;
+  std::tie(s, snapshot1) = db->CreateTimestampedSnapshot(100);
+  ASSERT_OK(s);
+  ASSERT_EQ(snapshot.get(), snapshot1.get());
+
+  // If there is no write, but we request a larger timestamp, we still create
+  // a new snapshot object.
+  std::shared_ptr<const Snapshot> snapshot2;
+  std::tie(s, snapshot2) = db->CreateTimestampedSnapshot(200);
+  ASSERT_OK(s);
+  assert(snapshot2);
+  ASSERT_NE(snapshot.get(), snapshot2.get());
+  ASSERT_EQ(snapshot2->GetSequenceNumber(), snapshot->GetSequenceNumber());
+  ASSERT_EQ(200, snapshot2->GetTimestamp());
+
+  // Increase sequence number.
+  ASSERT_OK(db->Put(WriteOptions(), "foo", "v0"));
+  {
+    // We are requesting the same timestamp for a larger sequence number, thus
+    // we cannot create timestamped snapshot.
+    std::shared_ptr<const Snapshot> tmp_snapshot;
+    std::tie(s, tmp_snapshot) = db->CreateTimestampedSnapshot(200);
+    ASSERT_TRUE(s.IsInvalidArgument());
+    ASSERT_EQ(nullptr, tmp_snapshot.get());
+  }
+  {
+    std::unique_ptr<Transaction> txn1(
+        db->BeginTransaction(WriteOptions(), TransactionOptions()));
+    ASSERT_OK(txn1->Put("bar", "v0"));
+    std::shared_ptr<const Snapshot> ss;
+    ASSERT_OK(txn1->CommitAndTryCreateSnapshot(nullptr, 200, &ss));
+    // Cannot create snapshot because requested timestamp is the same as the
+    // latest timestamped snapshot while sequence number is strictly higher.
+    ASSERT_EQ(nullptr, ss);
+  }
+  {
+    std::unique_ptr<Transaction> txn2(
+        db->BeginTransaction(WriteOptions(), TransactionOptions()));
+    ASSERT_OK(txn2->Put("bar", "v0"));
+    std::shared_ptr<const Snapshot> ss;
+    // Application should never do this. This is just to demonstrate error
+    // handling.
+    ASSERT_OK(txn2->CommitAndTryCreateSnapshot(nullptr, 100, &ss));
+    // Cannot create snapshot because requested timestamp is smaller than
+    // latest timestamped snapshot.
+    ASSERT_EQ(nullptr, ss);
+  }
+}
+
+TEST_P(TransactionTest, CloseDbWithSnapshots) {
+  std::unique_ptr<Transaction> txn(
+      db->BeginTransaction(WriteOptions(), TransactionOptions()));
+  ASSERT_OK(txn->SetName("txn0"));
+  ASSERT_OK(txn->Put("foo", "v"));
+  ASSERT_OK(txn->Prepare());
+  std::shared_ptr<const Snapshot> snapshot;
+  constexpr TxnTimestamp timestamp = 121;
+  auto notifier = std::make_shared<TsCheckingTxnNotifier>();
+  ASSERT_OK(txn->CommitAndTryCreateSnapshot(notifier, timestamp, &snapshot));
+  assert(snapshot);
+  ASSERT_LT(notifier->prev_snapshot_ts(), kMaxTxnTimestamp);
+  ASSERT_EQ(timestamp, snapshot->GetTimestamp());
+  ASSERT_TRUE(db->Close().IsAborted());
+}
+
+TEST_P(TransactionTest, MultipleTimestampedSnapshots) {
+  auto* dbimpl = static_cast_with_check<DBImpl>(db->GetRootDB());
+  assert(dbimpl);
+  const bool seq_per_batch = dbimpl->seq_per_batch();
+  // TODO: remove the following assert(!seq_per_batch) once timestamped snapshot
+  // is supported in write-prepared/write-unprepared transactions.
+  assert(!seq_per_batch);
+  constexpr size_t txn_size = 10;
+  constexpr TxnTimestamp ts_delta = 10;
+  constexpr size_t num_txns = 100;
+  std::vector<std::shared_ptr<const Snapshot> > snapshots(num_txns);
+  constexpr TxnTimestamp start_ts = 10000;
+  auto notifier = std::make_shared<TsCheckingTxnNotifier>();
+  for (size_t i = 0; i < num_txns; ++i) {
+    std::unique_ptr<Transaction> txn(
+        db->BeginTransaction(WriteOptions(), TransactionOptions()));
+    ASSERT_OK(txn->SetName("txn" + std::to_string(i)));
+    for (size_t j = 0; j < txn_size; ++j) {
+      ASSERT_OK(txn->Put("k" + std::to_string(j),
+                         "v" + std::to_string(j) + "_" + std::to_string(i)));
+    }
+    if (0 == (i % 2)) {
+      ASSERT_OK(txn->Prepare());
+    }
+    ASSERT_OK(txn->CommitAndTryCreateSnapshot(notifier, start_ts + i * ts_delta,
+                                              &snapshots[i]));
+    assert(snapshots[i]);
+    ASSERT_LT(notifier->prev_snapshot_ts(), kMaxTxnTimestamp);
+    ASSERT_EQ(start_ts + i * ts_delta, snapshots[i]->GetTimestamp());
+  }
+
+  {
+    auto snapshot = db->GetTimestampedSnapshot(start_ts + 1);
+    ASSERT_EQ(nullptr, snapshot);
+  }
+
+  constexpr TxnTimestamp max_ts = start_ts + num_txns * ts_delta;
+  for (size_t i = 0; i < num_txns; ++i) {
+    auto snapshot = db->GetTimestampedSnapshot(start_ts + i * ts_delta);
+    ASSERT_EQ(snapshots[i], snapshot);
+
+    std::vector<std::shared_ptr<const Snapshot> > tmp_snapshots;
+    Status s = db->GetTimestampedSnapshots(max_ts, start_ts + i * ts_delta,
+                                           tmp_snapshots);
+    ASSERT_TRUE(s.IsInvalidArgument());
+    ASSERT_TRUE(tmp_snapshots.empty());
+
+    for (size_t j = i; j < num_txns; ++j) {
+      std::vector<std::shared_ptr<const Snapshot> > expected_snapshots(
+          snapshots.begin() + i, snapshots.begin() + j);
+      tmp_snapshots.clear();
+      s = db->GetTimestampedSnapshots(start_ts + i * ts_delta,
+                                      start_ts + j * ts_delta, tmp_snapshots);
+      if (i < j) {
+        ASSERT_OK(s);
+      } else {
+        ASSERT_TRUE(s.IsInvalidArgument());
+      }
+      ASSERT_EQ(expected_snapshots, tmp_snapshots);
+    }
+  }
+
+  {
+    std::vector<std::shared_ptr<const Snapshot> > tmp_snapshots;
+    const Status s = db->GetAllTimestampedSnapshots(tmp_snapshots);
+    ASSERT_OK(s);
+    ASSERT_EQ(snapshots, tmp_snapshots);
+
+    const std::shared_ptr<const Snapshot> latest_snapshot =
+        db->GetLatestTimestampedSnapshot();
+    ASSERT_EQ(snapshots.back(), latest_snapshot);
+  }
+
+  for (size_t i = 0; i <= num_txns; ++i) {
+    std::vector<std::shared_ptr<const Snapshot> > snapshots1(
+        snapshots.begin() + i, snapshots.end());
+    if (i > 0) {
+      auto snapshot1 =
+          db->GetTimestampedSnapshot(start_ts + (i - 1) * ts_delta);
+      assert(snapshot1);
+      ASSERT_EQ(start_ts + (i - 1) * ts_delta, snapshot1->GetTimestamp());
+    }
+
+    db->ReleaseTimestampedSnapshotsOlderThan(start_ts + i * ts_delta);
+
+    if (i > 0) {
+      auto snapshot1 =
+          db->GetTimestampedSnapshot(start_ts + (i - 1) * ts_delta);
+      ASSERT_EQ(nullptr, snapshot1);
+    }
+
+    std::vector<std::shared_ptr<const Snapshot> > tmp_snapshots;
+    const Status s = db->GetAllTimestampedSnapshots(tmp_snapshots);
+    ASSERT_OK(s);
+    ASSERT_EQ(snapshots1, tmp_snapshots);
+  }
+
+  // Even after released by db, the applications still hold reference to shared
+  // snapshots.
+  for (size_t i = 0; i < num_txns; ++i) {
+    assert(snapshots[i]);
+    ASSERT_EQ(start_ts + i * ts_delta, snapshots[i]->GetTimestamp());
+  }
+
+  snapshots.clear();
+  ASSERT_OK(db->Close());
+  delete db;
+  db = nullptr;
+}
+
+}  // namespace ROCKSDB_NAMESPACE
+
+int main(int argc, char** argv) {
+  ROCKSDB_NAMESPACE::port::InstallStackTraceHandler();
+  ::testing::InitGoogleTest(&argc, argv);
+  return RUN_ALL_TESTS();
+}
+#endif  // !ROCKSDB_LITE
diff --git a/src/rocksdb/utilities/transactions/transaction_base.cc b/src/rocksdb/utilities/transactions/transaction_base.cc
new file mode 100644
index 000000000..83fd94ac8
--- /dev/null
+++ b/src/rocksdb/utilities/transactions/transaction_base.cc
@@ -0,0 +1,731 @@
+// Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#ifndef ROCKSDB_LITE
+
+#include "utilities/transactions/transaction_base.h"
+
+#include <cinttypes>
+
+#include "db/column_family.h"
+#include "db/db_impl/db_impl.h"
+#include "logging/logging.h"
+#include "rocksdb/comparator.h"
+#include "rocksdb/db.h"
+#include "rocksdb/status.h"
+#include "util/cast_util.h"
+#include "util/string_util.h"
+#include "utilities/transactions/lock/lock_tracker.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+Status Transaction::CommitAndTryCreateSnapshot(
+    std::shared_ptr<TransactionNotifier> notifier, TxnTimestamp ts,
+    std::shared_ptr<const Snapshot>* snapshot) {
+  if (snapshot) {
+    snapshot->reset();
+  }
+  TxnTimestamp commit_ts = GetCommitTimestamp();
+  if (commit_ts == kMaxTxnTimestamp) {
+    if (ts == kMaxTxnTimestamp) {
+      return Status::InvalidArgument("Commit timestamp unset");
+    } else {
+      const Status s = SetCommitTimestamp(ts);
+      if (!s.ok()) {
+        return s;
+      }
+    }
+  } else if (ts != kMaxTxnTimestamp) {
+    if (ts != commit_ts) {
+      // For now we treat this as error.
+      return Status::InvalidArgument("Different commit ts specified");
+    }
+  }
+  SetSnapshotOnNextOperation(notifier);
+  Status s = Commit();
+  if (!s.ok()) {
+    return s;
+  }
+  assert(s.ok());
+  // If we reach here, we must return ok status for this function.
+  std::shared_ptr<const Snapshot> new_snapshot = GetTimestampedSnapshot();
+
+  if (snapshot) {
+    *snapshot = new_snapshot;
+  }
+  return Status::OK();
+}
+
+TransactionBaseImpl::TransactionBaseImpl(
+    DB* db, const WriteOptions& write_options,
+    const LockTrackerFactory& lock_tracker_factory)
+    : db_(db),
+      dbimpl_(static_cast_with_check<DBImpl>(db)),
+      write_options_(write_options),
+      cmp_(GetColumnFamilyUserComparator(db->DefaultColumnFamily())),
+      lock_tracker_factory_(lock_tracker_factory),
+      start_time_(dbimpl_->GetSystemClock()->NowMicros()),
+      write_batch_(cmp_, 0, true, 0, write_options.protection_bytes_per_key),
+      tracked_locks_(lock_tracker_factory_.Create()),
+      commit_time_batch_(0 /* reserved_bytes */, 0 /* max_bytes */,
+                         write_options.protection_bytes_per_key,
+                         0 /* default_cf_ts_sz */),
+      indexing_enabled_(true) {
+  assert(dynamic_cast<DBImpl*>(db_) != nullptr);
+  log_number_ = 0;
+  if (dbimpl_->allow_2pc()) {
+    InitWriteBatch();
+  }
+}
+
+TransactionBaseImpl::~TransactionBaseImpl() {
+  // Release snapshot if snapshot is set
+  SetSnapshotInternal(nullptr);
+}
+
+void TransactionBaseImpl::Clear() {
+  save_points_.reset(nullptr);
+  write_batch_.Clear();
+  commit_time_batch_.Clear();
+  tracked_locks_->Clear();
+  num_puts_ = 0;
+  num_deletes_ = 0;
+  num_merges_ = 0;
+
+  if (dbimpl_->allow_2pc()) {
+    InitWriteBatch();
+  }
+}
+
+void TransactionBaseImpl::Reinitialize(DB* db,
+                                       const WriteOptions& write_options) {
+  Clear();
+  ClearSnapshot();
+  id_ = 0;
+  db_ = db;
+  name_.clear();
+  log_number_ = 0;
+  write_options_ = write_options;
+  start_time_ = dbimpl_->GetSystemClock()->NowMicros();
+  indexing_enabled_ = true;
+  cmp_ = GetColumnFamilyUserComparator(db_->DefaultColumnFamily());
+  WriteBatchInternal::UpdateProtectionInfo(
+      write_batch_.GetWriteBatch(), write_options_.protection_bytes_per_key)
+      .PermitUncheckedError();
+  WriteBatchInternal::UpdateProtectionInfo(
+      &commit_time_batch_, write_options_.protection_bytes_per_key)
+      .PermitUncheckedError();
+}
+
+void TransactionBaseImpl::SetSnapshot() {
+  const Snapshot* snapshot = dbimpl_->GetSnapshotForWriteConflictBoundary();
+  SetSnapshotInternal(snapshot);
+}
+
+void TransactionBaseImpl::SetSnapshotInternal(const Snapshot* snapshot) {
+  // Set a custom deleter for the snapshot_ SharedPtr as the snapshot needs to
+  // be released, not deleted when it is no longer referenced.
+  snapshot_.reset(snapshot, std::bind(&TransactionBaseImpl::ReleaseSnapshot,
+                                      this, std::placeholders::_1, db_));
+  snapshot_needed_ = false;
+  snapshot_notifier_ = nullptr;
+}
+
+void TransactionBaseImpl::SetSnapshotOnNextOperation(
+    std::shared_ptr<TransactionNotifier> notifier) {
+  snapshot_needed_ = true;
+  snapshot_notifier_ = notifier;
+}
+
+void TransactionBaseImpl::SetSnapshotIfNeeded() {
+  if (snapshot_needed_) {
+    std::shared_ptr<TransactionNotifier> notifier = snapshot_notifier_;
+    SetSnapshot();
+    if (notifier != nullptr) {
+      notifier->SnapshotCreated(GetSnapshot());
+    }
+  }
+}
+
+Status TransactionBaseImpl::TryLock(ColumnFamilyHandle* column_family,
+                                    const SliceParts& key, bool read_only,
+                                    bool exclusive, const bool do_validate,
+                                    const bool assume_tracked) {
+  size_t key_size = 0;
+  for (int i = 0; i < key.num_parts; ++i) {
+    key_size += key.parts[i].size();
+  }
+
+  std::string str;
+  str.reserve(key_size);
+
+  for (int i = 0; i < key.num_parts; ++i) {
+    str.append(key.parts[i].data(), key.parts[i].size());
+  }
+
+  return TryLock(column_family, str, read_only, exclusive, do_validate,
+                 assume_tracked);
+}
+
+void TransactionBaseImpl::SetSavePoint() {
+  if (save_points_ == nullptr) {
+    save_points_.reset(
+        new std::stack<TransactionBaseImpl::SavePoint,
+                       autovector<TransactionBaseImpl::SavePoint>>());
+  }
+  save_points_->emplace(snapshot_, snapshot_needed_, snapshot_notifier_,
+                        num_puts_, num_deletes_, num_merges_,
+                        lock_tracker_factory_);
+  write_batch_.SetSavePoint();
+}
+
+Status TransactionBaseImpl::RollbackToSavePoint() {
+  if (save_points_ != nullptr && save_points_->size() > 0) {
+    // Restore saved SavePoint
+    TransactionBaseImpl::SavePoint& save_point = save_points_->top();
+    snapshot_ = save_point.snapshot_;
+    snapshot_needed_ = save_point.snapshot_needed_;
+    snapshot_notifier_ = save_point.snapshot_notifier_;
+    num_puts_ = save_point.num_puts_;
+    num_deletes_ = save_point.num_deletes_;
+    num_merges_ = save_point.num_merges_;
+
+    // Rollback batch
+    Status s = write_batch_.RollbackToSavePoint();
+    assert(s.ok());
+
+    // Rollback any keys that were tracked since the last savepoint
+    tracked_locks_->Subtract(*save_point.new_locks_);
+
+    save_points_->pop();
+
+    return s;
+  } else {
+    assert(write_batch_.RollbackToSavePoint().IsNotFound());
+    return Status::NotFound();
+  }
+}
+
+Status TransactionBaseImpl::PopSavePoint() {
+  if (save_points_ == nullptr || save_points_->empty()) {
+    // No SavePoint yet.
+    assert(write_batch_.PopSavePoint().IsNotFound());
+    return Status::NotFound();
+  }
+
+  assert(!save_points_->empty());
+  // If there is another savepoint A below the current savepoint B, then A needs
+  // to inherit tracked_keys in B so that if we rollback to savepoint A, we
+  // remember to unlock keys in B. If there is no other savepoint below, then we
+  // can safely discard savepoint info.
+  if (save_points_->size() == 1) {
+    save_points_->pop();
+  } else {
+    TransactionBaseImpl::SavePoint top(lock_tracker_factory_);
+    std::swap(top, save_points_->top());
+    save_points_->pop();
+
+    save_points_->top().new_locks_->Merge(*top.new_locks_);
+  }
+
+  return write_batch_.PopSavePoint();
+}
+
+Status TransactionBaseImpl::Get(const ReadOptions& read_options,
+                                ColumnFamilyHandle* column_family,
+                                const Slice& key, std::string* value) {
+  assert(value != nullptr);
+  PinnableSlice pinnable_val(value);
+  assert(!pinnable_val.IsPinned());
+  auto s = Get(read_options, column_family, key, &pinnable_val);
+  if (s.ok() && pinnable_val.IsPinned()) {
+    value->assign(pinnable_val.data(), pinnable_val.size());
+  }  // else value is already assigned
+  return s;
+}
+
+Status TransactionBaseImpl::Get(const ReadOptions& read_options,
+                                ColumnFamilyHandle* column_family,
+                                const Slice& key, PinnableSlice* pinnable_val) {
+  return write_batch_.GetFromBatchAndDB(db_, read_options, column_family, key,
+                                        pinnable_val);
+}
+
+Status TransactionBaseImpl::GetForUpdate(const ReadOptions& read_options,
+                                         ColumnFamilyHandle* column_family,
+                                         const Slice& key, std::string* value,
+                                         bool exclusive,
+                                         const bool do_validate) {
+  if (!do_validate && read_options.snapshot != nullptr) {
+    return Status::InvalidArgument(
+        "If do_validate is false then GetForUpdate with snapshot is not "
+        "defined.");
+  }
+  Status s =
+      TryLock(column_family, key, true /* read_only */, exclusive, do_validate);
+
+  if (s.ok() && value != nullptr) {
+    assert(value != nullptr);
+    PinnableSlice pinnable_val(value);
+    assert(!pinnable_val.IsPinned());
+    s = Get(read_options, column_family, key, &pinnable_val);
+    if (s.ok() && pinnable_val.IsPinned()) {
+      value->assign(pinnable_val.data(), pinnable_val.size());
+    }  // else value is already assigned
+  }
+  return s;
+}
+
+Status TransactionBaseImpl::GetForUpdate(const ReadOptions& read_options,
+                                         ColumnFamilyHandle* column_family,
+                                         const Slice& key,
+                                         PinnableSlice* pinnable_val,
+                                         bool exclusive,
+                                         const bool do_validate) {
+  if (!do_validate && read_options.snapshot != nullptr) {
+    return Status::InvalidArgument(
+        "If do_validate is false then GetForUpdate with snapshot is not "
+        "defined.");
+  }
+  Status s =
+      TryLock(column_family, key, true /* read_only */, exclusive, do_validate);
+
+  if (s.ok() && pinnable_val != nullptr) {
+    s = Get(read_options, column_family, key, pinnable_val);
+  }
+  return s;
+}
+
+std::vector<Status> TransactionBaseImpl::MultiGet(
+    const ReadOptions& read_options,
+    const std::vector<ColumnFamilyHandle*>& column_family,
+    const std::vector<Slice>& keys, std::vector<std::string>* values) {
+  size_t num_keys = keys.size();
+  values->resize(num_keys);
+
+  std::vector<Status> stat_list(num_keys);
+  for (size_t i = 0; i < num_keys; ++i) {
+    stat_list[i] = Get(read_options, column_family[i], keys[i], &(*values)[i]);
+  }
+
+  return stat_list;
+}
+
+void TransactionBaseImpl::MultiGet(const ReadOptions& read_options,
+                                   ColumnFamilyHandle* column_family,
+                                   const size_t num_keys, const Slice* keys,
+                                   PinnableSlice* values, Status* statuses,
+                                   const bool sorted_input) {
+  write_batch_.MultiGetFromBatchAndDB(db_, read_options, column_family,
+                                      num_keys, keys, values, statuses,
+                                      sorted_input);
+}
+
+std::vector<Status> TransactionBaseImpl::MultiGetForUpdate(
+    const ReadOptions& read_options,
+    const std::vector<ColumnFamilyHandle*>& column_family,
+    const std::vector<Slice>& keys, std::vector<std::string>* values) {
+  // Regardless of whether the MultiGet succeeded, track these keys.
+  size_t num_keys = keys.size();
+  values->resize(num_keys);
+
+  // Lock all keys
+  for (size_t i = 0; i < num_keys; ++i) {
+    Status s = TryLock(column_family[i], keys[i], true /* read_only */,
+                       true /* exclusive */);
+    if (!s.ok()) {
+      // Fail entire multiget if we cannot lock all keys
+      return std::vector<Status>(num_keys, s);
+    }
+  }
+
+  // TODO(agiardullo): optimize multiget?
+  std::vector<Status> stat_list(num_keys);
+  for (size_t i = 0; i < num_keys; ++i) {
+    stat_list[i] = Get(read_options, column_family[i], keys[i], &(*values)[i]);
+  }
+
+  return stat_list;
+}
+
+Iterator* TransactionBaseImpl::GetIterator(const ReadOptions& read_options) {
+  Iterator* db_iter = db_->NewIterator(read_options);
+  assert(db_iter);
+
+  return write_batch_.NewIteratorWithBase(db_->DefaultColumnFamily(), db_iter,
+                                          &read_options);
+}
+
+Iterator* TransactionBaseImpl::GetIterator(const ReadOptions& read_options,
+                                           ColumnFamilyHandle* column_family) {
+  Iterator* db_iter = db_->NewIterator(read_options, column_family);
+  assert(db_iter);
+
+  return write_batch_.NewIteratorWithBase(column_family, db_iter,
+                                          &read_options);
+}
+
+Status TransactionBaseImpl::Put(ColumnFamilyHandle* column_family,
+                                const Slice& key, const Slice& value,
+                                const bool assume_tracked) {
+  const bool do_validate = !assume_tracked;
+  Status s = TryLock(column_family, key, false /* read_only */,
+                     true /* exclusive */, do_validate, assume_tracked);
+
+  if (s.ok()) {
+    s = GetBatchForWrite()->Put(column_family, key, value);
+    if (s.ok()) {
+      num_puts_++;
+    }
+  }
+
+  return s;
+}
+
+Status TransactionBaseImpl::Put(ColumnFamilyHandle* column_family,
+                                const SliceParts& key, const SliceParts& value,
+                                const bool assume_tracked) {
+  const bool do_validate = !assume_tracked;
+  Status s = TryLock(column_family, key, false /* read_only */,
+                     true /* exclusive */, do_validate, assume_tracked);
+
+  if (s.ok()) {
+    s = GetBatchForWrite()->Put(column_family, key, value);
+    if (s.ok()) {
+      num_puts_++;
+    }
+  }
+
+  return s;
+}
+
+Status TransactionBaseImpl::Merge(ColumnFamilyHandle* column_family,
+                                  const Slice& key, const Slice& value,
+                                  const bool assume_tracked) {
+  const bool do_validate = !assume_tracked;
+  Status s = TryLock(column_family, key, false /* read_only */,
+                     true /* exclusive */, do_validate, assume_tracked);
+
+  if (s.ok()) {
+    s = GetBatchForWrite()->Merge(column_family, key, value);
+    if (s.ok()) {
+      num_merges_++;
+    }
+  }
+
+  return s;
+}
+
+Status TransactionBaseImpl::Delete(ColumnFamilyHandle* column_family,
+                                   const Slice& key,
+                                   const bool assume_tracked) {
+  const bool do_validate = !assume_tracked;
+  Status s = TryLock(column_family, key, false /* read_only */,
+                     true /* exclusive */, do_validate, assume_tracked);
+
+  if (s.ok()) {
+    s = GetBatchForWrite()->Delete(column_family, key);
+    if (s.ok()) {
+      num_deletes_++;
+    }
+  }
+
+  return s;
+}
+
+Status TransactionBaseImpl::Delete(ColumnFamilyHandle* column_family,
+                                   const SliceParts& key,
+                                   const bool assume_tracked) {
+  const bool do_validate = !assume_tracked;
+  Status s = TryLock(column_family, key, false /* read_only */,
+                     true /* exclusive */, do_validate, assume_tracked);
+
+  if (s.ok()) {
+    s = GetBatchForWrite()->Delete(column_family, key);
+    if (s.ok()) {
+      num_deletes_++;
+    }
+  }
+
+  return s;
+}
+
+Status TransactionBaseImpl::SingleDelete(ColumnFamilyHandle* column_family,
+                                         const Slice& key,
+                                         const bool assume_tracked) {
+  const bool do_validate = !assume_tracked;
+  Status s = TryLock(column_family, key, false /* read_only */,
+                     true /* exclusive */, do_validate, assume_tracked);
+
+  if (s.ok()) {
+    s = GetBatchForWrite()->SingleDelete(column_family, key);
+    if (s.ok()) {
+      num_deletes_++;
+    }
+  }
+
+  return s;
+}
+
+Status TransactionBaseImpl::SingleDelete(ColumnFamilyHandle* column_family,
+                                         const SliceParts& key,
+                                         const bool assume_tracked) {
+  const bool do_validate = !assume_tracked;
+  Status s = TryLock(column_family, key, false /* read_only */,
+                     true /* exclusive */, do_validate, assume_tracked);
+
+  if (s.ok()) {
+    s = GetBatchForWrite()->SingleDelete(column_family, key);
+    if (s.ok()) {
+      num_deletes_++;
+    }
+  }
+
+  return s;
+}
+
+Status TransactionBaseImpl::PutUntracked(ColumnFamilyHandle* column_family,
+                                         const Slice& key, const Slice& value) {
+  Status s = TryLock(column_family, key, false /* read_only */,
+                     true /* exclusive */, false /* do_validate */);
+
+  if (s.ok()) {
+    s = GetBatchForWrite()->Put(column_family, key, value);
+    if (s.ok()) {
+      num_puts_++;
+    }
+  }
+
+  return s;
+}
+
+Status TransactionBaseImpl::PutUntracked(ColumnFamilyHandle* column_family,
+                                         const SliceParts& key,
+                                         const SliceParts& value) {
+  Status s = TryLock(column_family, key, false /* read_only */,
+                     true /* exclusive */, false /* do_validate */);
+
+  if (s.ok()) {
+    s = GetBatchForWrite()->Put(column_family, key, value);
+    if (s.ok()) {
+      num_puts_++;
+    }
+  }
+
+  return s;
+}
+
+Status TransactionBaseImpl::MergeUntracked(ColumnFamilyHandle* column_family,
+                                           const Slice& key,
+                                           const Slice& value) {
+  Status s = TryLock(column_family, key, false /* read_only */,
+                     true /* exclusive */, false /* do_validate */);
+
+  if (s.ok()) {
+    s = GetBatchForWrite()->Merge(column_family, key, value);
+    if (s.ok()) {
+      num_merges_++;
+    }
+  }
+
+  return s;
+}
+
+Status TransactionBaseImpl::DeleteUntracked(ColumnFamilyHandle* column_family,
+                                            const Slice& key) {
+  Status s = TryLock(column_family, key, false /* read_only */,
+                     true /* exclusive */, false /* do_validate */);
+
+  if (s.ok()) {
+    s = GetBatchForWrite()->Delete(column_family, key);
+    if (s.ok()) {
+      num_deletes_++;
+    }
+  }
+
+  return s;
+}
+
+Status TransactionBaseImpl::DeleteUntracked(ColumnFamilyHandle* column_family,
+                                            const SliceParts& key) {
+  Status s = TryLock(column_family, key, false /* read_only */,
+                     true /* exclusive */, false /* do_validate */);
+
+  if (s.ok()) {
+    s = GetBatchForWrite()->Delete(column_family, key);
+    if (s.ok()) {
+      num_deletes_++;
+    }
+  }
+
+  return s;
+}
+
+Status TransactionBaseImpl::SingleDeleteUntracked(
+    ColumnFamilyHandle* column_family, const Slice& key) {
+  Status s = TryLock(column_family, key, false /* read_only */,
+                     true /* exclusive */, false /* do_validate */);
+
+  if (s.ok()) {
+    s = GetBatchForWrite()->SingleDelete(column_family, key);
+    if (s.ok()) {
+      num_deletes_++;
+    }
+  }
+
+  return s;
+}
+
+void TransactionBaseImpl::PutLogData(const Slice& blob) {
+  auto s = write_batch_.PutLogData(blob);
+  (void)s;
+  assert(s.ok());
+}
+
+WriteBatchWithIndex* TransactionBaseImpl::GetWriteBatch() {
+  return &write_batch_;
+}
+
+uint64_t TransactionBaseImpl::GetElapsedTime() const {
+  return (dbimpl_->GetSystemClock()->NowMicros() - start_time_) / 1000;
+}
+
+uint64_t TransactionBaseImpl::GetNumPuts() const { return num_puts_; }
+
+uint64_t TransactionBaseImpl::GetNumDeletes() const { return num_deletes_; }
+
+uint64_t TransactionBaseImpl::GetNumMerges() const { return num_merges_; }
+
+uint64_t TransactionBaseImpl::GetNumKeys() const {
+  return tracked_locks_->GetNumPointLocks();
+}
+
+void TransactionBaseImpl::TrackKey(uint32_t cfh_id, const std::string& key,
+                                   SequenceNumber seq, bool read_only,
+                                   bool exclusive) {
+  PointLockRequest r;
+  r.column_family_id = cfh_id;
+  r.key = key;
+  r.seq = seq;
+  r.read_only = read_only;
+  r.exclusive = exclusive;
+
+  // Update map of all tracked keys for this transaction
+  tracked_locks_->Track(r);
+
+  if (save_points_ != nullptr && !save_points_->empty()) {
+    // Update map of tracked keys in this SavePoint
+    save_points_->top().new_locks_->Track(r);
+  }
+}
+
+// Gets the write batch that should be used for Put/Merge/Deletes.
+//
+// Returns either a WriteBatch or WriteBatchWithIndex depending on whether
+// DisableIndexing() has been called.
+WriteBatchBase* TransactionBaseImpl::GetBatchForWrite() {
+  if (indexing_enabled_) {
+    // Use WriteBatchWithIndex
+    return &write_batch_;
+  } else {
+    // Don't use WriteBatchWithIndex. Return base WriteBatch.
+    return write_batch_.GetWriteBatch();
+  }
+}
+
+void TransactionBaseImpl::ReleaseSnapshot(const Snapshot* snapshot, DB* db) {
+  if (snapshot != nullptr) {
+    ROCKS_LOG_DETAILS(dbimpl_->immutable_db_options().info_log,
+                      "ReleaseSnapshot %" PRIu64 " Set",
+                      snapshot->GetSequenceNumber());
+    db->ReleaseSnapshot(snapshot);
+  }
+}
+
+void TransactionBaseImpl::UndoGetForUpdate(ColumnFamilyHandle* column_family,
+                                           const Slice& key) {
+  PointLockRequest r;
+  r.column_family_id = GetColumnFamilyID(column_family);
+  r.key = key.ToString();
+  r.read_only = true;
+
+  bool can_untrack = false;
+  if (save_points_ != nullptr && !save_points_->empty()) {
+    // If there is no GetForUpdate of the key in this save point,
+    // then cannot untrack from the global lock tracker.
+    UntrackStatus s = save_points_->top().new_locks_->Untrack(r);
+    can_untrack = (s != UntrackStatus::NOT_TRACKED);
+  } else {
+    // No save point, so can untrack from the global lock tracker.
+    can_untrack = true;
+  }
+
+  if (can_untrack) {
+    // If erased from the global tracker, then can unlock the key.
+    UntrackStatus s = tracked_locks_->Untrack(r);
+    bool can_unlock = (s == UntrackStatus::REMOVED);
+    if (can_unlock) {
+      UnlockGetForUpdate(column_family, key);
+    }
+  }
+}
+
+Status TransactionBaseImpl::RebuildFromWriteBatch(WriteBatch* src_batch) {
+  struct IndexedWriteBatchBuilder : public WriteBatch::Handler {
+    Transaction* txn_;
+    DBImpl* db_;
+    IndexedWriteBatchBuilder(Transaction* txn, DBImpl* db)
+        : txn_(txn), db_(db) {
+      assert(dynamic_cast<TransactionBaseImpl*>(txn_) != nullptr);
+    }
+
+    Status PutCF(uint32_t cf, const Slice& key, const Slice& val) override {
+      return txn_->Put(db_->GetColumnFamilyHandle(cf), key, val);
+    }
+
+    Status DeleteCF(uint32_t cf, const Slice& key) override {
+      return txn_->Delete(db_->GetColumnFamilyHandle(cf), key);
+    }
+
+    Status SingleDeleteCF(uint32_t cf, const Slice& key) override {
+      return txn_->SingleDelete(db_->GetColumnFamilyHandle(cf), key);
+    }
+
+    Status MergeCF(uint32_t cf, const Slice& key, const Slice& val) override {
+      return txn_->Merge(db_->GetColumnFamilyHandle(cf), key, val);
+    }
+
+    // this is used for reconstructing prepared transactions upon
+    // recovery. there should not be any meta markers in the batches
+    // we are processing.
+    Status MarkBeginPrepare(bool) override { return Status::InvalidArgument(); }
+
+    Status MarkEndPrepare(const Slice&) override {
+      return Status::InvalidArgument();
+    }
+
+    Status MarkCommit(const Slice&) override {
+      return Status::InvalidArgument();
+    }
+
+    Status MarkCommitWithTimestamp(const Slice&, const Slice&) override {
+      return Status::InvalidArgument();
+    }
+
+    Status MarkRollback(const Slice&) override {
+      return Status::InvalidArgument();
+    }
+  };
+
+  IndexedWriteBatchBuilder copycat(this, dbimpl_);
+  return src_batch->Iterate(&copycat);
+}
+
+WriteBatch* TransactionBaseImpl::GetCommitTimeWriteBatch() {
+  return &commit_time_batch_;
+}
+}  // namespace ROCKSDB_NAMESPACE
+
+#endif  // ROCKSDB_LITE
diff --git a/src/rocksdb/utilities/transactions/transaction_base.h b/src/rocksdb/utilities/transactions/transaction_base.h
new file mode 100644
index 000000000..1bcb20ca9
--- /dev/null
+++ b/src/rocksdb/utilities/transactions/transaction_base.h
@@ -0,0 +1,384 @@
+// Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+// This source code is licensed under both the GPLv2 (found in the
+// COPYING file in the root directory) and Apache 2.0 License
+// (found in the LICENSE.Apache file in the root directory).
+
+#pragma once
+
+#ifndef ROCKSDB_LITE
+
+#include <stack>
+#include <string>
+#include <vector>
+
+#include "db/write_batch_internal.h"
+#include "rocksdb/db.h"
+#include "rocksdb/slice.h"
+#include "rocksdb/snapshot.h"
+#include "rocksdb/status.h"
+#include "rocksdb/types.h"
+#include "rocksdb/utilities/transaction.h"
+#include "rocksdb/utilities/transaction_db.h"
+#include "rocksdb/utilities/write_batch_with_index.h"
+#include "util/autovector.h"
+#include "utilities/transactions/lock/lock_tracker.h"
+#include "utilities/transactions/transaction_util.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+class TransactionBaseImpl : public Transaction {
+ public:
+  TransactionBaseImpl(DB* db, const WriteOptions& write_options,
+                      const LockTrackerFactory& lock_tracker_factory);
+
+  ~TransactionBaseImpl() override;
+
+  // Remove pending operations queued in this transaction.
+  virtual void Clear();
+
+  void Reinitialize(DB* db, const WriteOptions& write_options);
+
+  // Called before executing Put, Merge, Delete, and GetForUpdate.  If TryLock
+  // returns non-OK, the Put/Merge/Delete/GetForUpdate will be failed.
+  // do_validate will be false if called from PutUntracked, DeleteUntracked,
+  // MergeUntracked, or GetForUpdate(do_validate=false)
+  virtual Status TryLock(ColumnFamilyHandle* column_family, const Slice& key,
+                         bool read_only, bool exclusive,
+                         const bool do_validate = true,
+                         const bool assume_tracked = false) = 0;
+
+  void SetSavePoint() override;
+
+  Status RollbackToSavePoint() override;
+
+  Status PopSavePoint() override;
+
+  using Transaction::Get;
+  Status Get(const ReadOptions& options, ColumnFamilyHandle* column_family,
+             const Slice& key, std::string* value) override;
+
+  Status Get(const ReadOptions& options, ColumnFamilyHandle* column_family,
+             const Slice& key, PinnableSlice* value) override;
+
+  Status Get(const ReadOptions& options, const Slice& key,
+             std::string* value) override {
+    return Get(options, db_->DefaultColumnFamily(), key, value);
+  }
+
+  using Transaction::GetForUpdate;
+  Status GetForUpdate(const ReadOptions& options,
+                      ColumnFamilyHandle* column_family, const Slice& key,
+                      std::string* value, bool exclusive,
+                      const bool do_validate) override;
+
+  Status GetForUpdate(const ReadOptions& options,
+                      ColumnFamilyHandle* column_family, const Slice& key,
+                      PinnableSlice* pinnable_val, bool exclusive,
+                      const bool do_validate) override;
+
+  Status GetForUpdate(const ReadOptions& options, const Slice& key,
+                      std::string* value, bool exclusive,
+                      const bool do_validate) override {
+    return GetForUpdate(options, db_->DefaultColumnFamily(), key, value,
+                        exclusive, do_validate);
+  }
+
+  using Transaction::MultiGet;
+  std::vector<Status> MultiGet(
+      const ReadOptions& options,
+      const std::vector<ColumnFamilyHandle*>& column_family,
+      const std::vector<Slice>& keys,
+      std::vector<std::string>* values) override;
+
+  std::vector<Status> MultiGet(const ReadOptions& options,
+                               const std::vector<Slice>& keys,
+                               std::vector<std::string>* values) override {
+    return MultiGet(options,
+                    std::vector<ColumnFamilyHandle*>(
+                        keys.size(), db_->DefaultColumnFamily()),
+                    keys, values);
+  }
+
+  void MultiGet(const ReadOptions& options, ColumnFamilyHandle* column_family,
+                const size_t num_keys, const Slice* keys, PinnableSlice* values,
+                Status* statuses, const bool sorted_input = false) override;
+
+  using Transaction::MultiGetForUpdate;
+  std::vector<Status> MultiGetForUpdate(
+      const ReadOptions& options,
+      const std::vector<ColumnFamilyHandle*>& column_family,
+      const std::vector<Slice>& keys,
+      std::vector<std::string>* values) override;
+
+  std::vector<Status> MultiGetForUpdate(
+      const ReadOptions& options, const std::vector<Slice>& keys,
+      std::vector<std::string>* values) override {
+    return MultiGetForUpdate(options,
+                             std::vector<ColumnFamilyHandle*>(
+                                 keys.size(), db_->DefaultColumnFamily()),
+                             keys, values);
+  }
+
+  Iterator* GetIterator(const ReadOptions& read_options) override;
+  Iterator* GetIterator(const ReadOptions& read_options,
+                        ColumnFamilyHandle* column_family) override;
+
+  Status Put(ColumnFamilyHandle* column_family, const Slice& key,
+             const Slice& value, const bool assume_tracked = false) override;
+  Status Put(const Slice& key, const Slice& value) override {
+    return Put(nullptr, key, value);
+  }
+
+  Status Put(ColumnFamilyHandle* column_family, const SliceParts& key,
+             const SliceParts& value,
+             const bool assume_tracked = false) override;
+  Status Put(const SliceParts& key, const SliceParts& value) override {
+    return Put(nullptr, key, value);
+  }
+
+  Status Merge(ColumnFamilyHandle* column_family, const Slice& key,
+               const Slice& value, const bool assume_tracked = false) override;
+  Status Merge(const Slice& key, const Slice& value) override {
+    return Merge(nullptr, key, value);
+  }
+
+  Status Delete(ColumnFamilyHandle* column_family, const Slice& key,
+                const bool assume_tracked = false) override;
+  Status Delete(const Slice& key) override { return Delete(nullptr, key); }
+  Status Delete(ColumnFamilyHandle* column_family, const SliceParts& key,
+                const bool assume_tracked = false) override;
+  Status Delete(const SliceParts& key) override { return Delete(nullptr, key); }
+
+  Status SingleDelete(ColumnFamilyHandle* column_family, const Slice& key,
+                      const bool assume_tracked = false) override;
+  Status SingleDelete(const Slice& key) override {
+    return SingleDelete(nullptr, key);
+  }
+  Status SingleDelete(ColumnFamilyHandle* column_family, const SliceParts& key,
+                      const bool assume_tracked = false) override;
+  Status SingleDelete(const SliceParts& key) override {
+    return SingleDelete(nullptr, key);
+  }
+
+  Status PutUntracked(ColumnFamilyHandle* column_family, const Slice& key,
+                      const Slice& value) override;
+  Status PutUntracked(const Slice& key, const Slice& value) override {
+    return PutUntracked(nullptr, key, value);
+  }
+
+  Status PutUntracked(ColumnFamilyHandle* column_family, const SliceParts& key,
+                      const SliceParts& value) override;
+  Status PutUntracked(const SliceParts& key, const SliceParts& value) override {
+    return PutUntracked(nullptr, key, value);
+  }
+
+  Status MergeUntracked(ColumnFamilyHandle* column_family, const Slice& key,
+                        const Slice& value) override;
+  Status MergeUntracked(const Slice& key, const Slice& value) override {
+    return MergeUntracked(nullptr, key, value);
+  }
+
+  Status DeleteUntracked(ColumnFamilyHandle* column_family,
+                         const Slice& key) override;
+  Status DeleteUntracked(const Slice& key) override {
+    return DeleteUntracked(nullptr, key);
+  }
+  Status DeleteUntracked(ColumnFamilyHandle* column_family,
+                         const SliceParts& key) override;
+  Status DeleteUntracked(const SliceParts& key) override {
+    return DeleteUntracked(nullptr, key);
+  }
+
+  Status SingleDeleteUntracked(ColumnFamilyHandle* column_family,
+                               const Slice& key) override;
+  Status SingleDeleteUntracked(const Slice& key) override {
+    return SingleDeleteUntracked(nullptr, key);
+  }
+
+  void PutLogData(const Slice& blob) override;
+
+  WriteBatchWithIndex* GetWriteBatch() override;
+
+  virtual void SetLockTimeout(int64_t /*timeout*/) override { /* Do nothing */
+  }
+
+  const Snapshot* GetSnapshot() const override {
+    // will return nullptr when there is no snapshot
+    return snapshot_.get();
+  }
+
+  std::shared_ptr<const Snapshot> GetTimestampedSnapshot() const override {
+    return snapshot_;
+  }
+
+  virtual void SetSnapshot() override;
+  void SetSnapshotOnNextOperation(
+      std::shared_ptr<TransactionNotifier> notifier = nullptr) override;
+
+  void ClearSnapshot() override {
+    snapshot_.reset();
+    snapshot_needed_ = false;
+    snapshot_notifier_ = nullptr;
+  }
+
+  void DisableIndexing() override { indexing_enabled_ = false; }
+
+  void EnableIndexing() override { indexing_enabled_ = true; }
+
+  bool IndexingEnabled() const { return indexing_enabled_; }
+
+  uint64_t GetElapsedTime() const override;
+
+  uint64_t GetNumPuts() const override;
+
+  uint64_t GetNumDeletes() const override;
+
+  uint64_t GetNumMerges() const override;
+
+  uint64_t GetNumKeys() const override;
+
+  void UndoGetForUpdate(ColumnFamilyHandle* column_family,
+                        const Slice& key) override;
+  void UndoGetForUpdate(const Slice& key) override {
+    return UndoGetForUpdate(nullptr, key);
+  };
+
+  WriteOptions* GetWriteOptions() override { return &write_options_; }
+
+  void SetWriteOptions(const WriteOptions& write_options) override {
+    write_options_ = write_options;
+  }
+
+  // Used for memory management for snapshot_
+  void ReleaseSnapshot(const Snapshot* snapshot, DB* db);
+
+  // iterates over the given batch and makes the appropriate inserts.
+  // used for rebuilding prepared transactions after recovery.
+  virtual Status RebuildFromWriteBatch(WriteBatch* src_batch) override;
+
+  WriteBatch* GetCommitTimeWriteBatch() override;
+
+  LockTracker& GetTrackedLocks() { return *tracked_locks_; }
+
+ protected:
+  // Add a key to the list of tracked keys.
+  //
+  // seqno is the earliest seqno this key was involved with this transaction.
+  // readonly should be set to true if no data was written for this key
+  void TrackKey(uint32_t cfh_id, const std::string& key, SequenceNumber seqno,
+                bool readonly, bool exclusive);
+
+  // Called when UndoGetForUpdate determines that this key can be unlocked.
+  virtual void UnlockGetForUpdate(ColumnFamilyHandle* column_family,
+                                  const Slice& key) = 0;
+
+  // Sets a snapshot if SetSnapshotOnNextOperation() has been called.
+  void SetSnapshotIfNeeded();
+
+  // Initialize write_batch_ for 2PC by inserting Noop.
+  inline void InitWriteBatch(bool clear = false) {
+    if (clear) {
+      write_batch_.Clear();
+    }
+    assert(write_batch_.GetDataSize() == WriteBatchInternal::kHeader);
+    auto s = WriteBatchInternal::InsertNoop(write_batch_.GetWriteBatch());
+    assert(s.ok());
+  }
+
+  WriteBatchBase* GetBatchForWrite();
+
+  DB* db_;
+  DBImpl* dbimpl_;
+
+  WriteOptions write_options_;
+
+  const Comparator* cmp_;
+
+  const LockTrackerFactory& lock_tracker_factory_;
+
+  // Stores that time the txn was constructed, in microseconds.
+  uint64_t start_time_;
+
+  // Stores the current snapshot that was set by SetSnapshot or null if
+  // no snapshot is currently set.
+  std::shared_ptr<const Snapshot> snapshot_;
+
+  // Count of various operations pending in this transaction
+  uint64_t num_puts_ = 0;
+  uint64_t num_deletes_ = 0;
+  uint64_t num_merges_ = 0;
+
+  struct SavePoint {
+    std::shared_ptr<const Snapshot> snapshot_;
+    bool snapshot_needed_ = false;
+    std::shared_ptr<TransactionNotifier> snapshot_notifier_;
+    uint64_t num_puts_ = 0;
+    uint64_t num_deletes_ = 0;
+    uint64_t num_merges_ = 0;
+
+    // Record all locks tracked since the last savepoint
+    std::shared_ptr<LockTracker> new_locks_;
+
+    SavePoint(std::shared_ptr<const Snapshot> snapshot, bool snapshot_needed,
+              std::shared_ptr<TransactionNotifier> snapshot_notifier,
+              uint64_t num_puts, uint64_t num_deletes, uint64_t num_merges,
+              const LockTrackerFactory& lock_tracker_factory)
+        : snapshot_(snapshot),
+          snapshot_needed_(snapshot_needed),
+          snapshot_notifier_(snapshot_notifier),
+          num_puts_(num_puts),
+          num_deletes_(num_deletes),
+          num_merges_(num_merges),
+          new_locks_(lock_tracker_factory.Create()) {}
+
+    explicit SavePoint(const LockTrackerFactory& lock_tracker_factory)
+        : new_locks_(lock_tracker_factory.Create()) {}
+  };
+
+  // Records writes pending in this transaction
+  WriteBatchWithIndex write_batch_;
+
+  // For Pessimistic Transactions this is the set of acquired locks.
+  // Optimistic Transactions will keep note the requested locks (not actually
+  // locked), and do conflict checking until commit time based on the tracked
+  // lock requests.
+  std::unique_ptr<LockTracker> tracked_locks_;
+
+  // Stack of the Snapshot saved at each save point. Saved snapshots may be
+  // nullptr if there was no snapshot at the time SetSavePoint() was called.
+  std::unique_ptr<std::stack<TransactionBaseImpl::SavePoint,
+                             autovector<TransactionBaseImpl::SavePoint>>>
+      save_points_;
+
+ private:
+  friend class WriteCommittedTxn;
+  friend class WritePreparedTxn;
+
+  // Extra data to be persisted with the commit. Note this is only used when
+  // prepare phase is not skipped.
+  WriteBatch commit_time_batch_;
+
+  // If true, future Put/Merge/Deletes will be indexed in the
+  // WriteBatchWithIndex.
+  // If false, future Put/Merge/Deletes will be inserted directly into the
+  // underlying WriteBatch and not indexed in the WriteBatchWithIndex.
+  bool indexing_enabled_;
+
+  // SetSnapshotOnNextOperation() has been called and the snapshot has not yet
+  // been reset.
+  bool snapshot_needed_ = false;
+
+  // SetSnapshotOnNextOperation() has been called and the caller would like
+  // a notification through the TransactionNotifier interface
+  std::shared_ptr<TransactionNotifier> snapshot_notifier_ = nullptr;
+
+  Status TryLock(ColumnFamilyHandle* column_family, const SliceParts& key,
+                 bool read_only, bool exclusive, const bool do_validate = true,
+                 const bool assume_tracked = false);
+
+  void SetSnapshotInternal(const Snapshot* snapshot);
+};
+
+}  // namespace ROCKSDB_NAMESPACE
+
+#endif  // ROCKSDB_LITE
diff --git a/src/rocksdb/utilities/transactions/transaction_db_mutex_impl.cc b/src/rocksdb/utilities/transactions/transaction_db_mutex_impl.cc
new file mode 100644
index 000000000..345c4be90
--- /dev/null
+++ b/src/rocksdb/utilities/transactions/transaction_db_mutex_impl.cc
@@ -0,0 +1,135 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#ifndef ROCKSDB_LITE
+
+#include "utilities/transactions/transaction_db_mutex_impl.h"
+
+#include <chrono>
+#include <condition_variable>
+#include <functional>
+#include <mutex>
+
+#include "rocksdb/utilities/transaction_db_mutex.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+class TransactionDBMutexImpl : public TransactionDBMutex {
+ public:
+  TransactionDBMutexImpl() {}
+  ~TransactionDBMutexImpl() override {}
+
+  Status Lock() override;
+
+  Status TryLockFor(int64_t timeout_time) override;
+
+  void UnLock() override { mutex_.unlock(); }
+
+  friend class TransactionDBCondVarImpl;
+
+ private:
+  std::mutex mutex_;
+};
+
+class TransactionDBCondVarImpl : public TransactionDBCondVar {
+ public:
+  TransactionDBCondVarImpl() {}
+  ~TransactionDBCondVarImpl() override {}
+
+  Status Wait(std::shared_ptr<TransactionDBMutex> mutex) override;
+
+  Status WaitFor(std::shared_ptr<TransactionDBMutex> mutex,
+                 int64_t timeout_time) override;
+
+  void Notify() override { cv_.notify_one(); }
+
+  void NotifyAll() override { cv_.notify_all(); }
+
+ private:
+  std::condition_variable cv_;
+};
+
+std::shared_ptr<TransactionDBMutex>
+TransactionDBMutexFactoryImpl::AllocateMutex() {
+  return std::shared_ptr<TransactionDBMutex>(new TransactionDBMutexImpl());
+}
+
+std::shared_ptr<TransactionDBCondVar>
+TransactionDBMutexFactoryImpl::AllocateCondVar() {
+  return std::shared_ptr<TransactionDBCondVar>(new TransactionDBCondVarImpl());
+}
+
+Status TransactionDBMutexImpl::Lock() {
+  mutex_.lock();
+  return Status::OK();
+}
+
+Status TransactionDBMutexImpl::TryLockFor(int64_t timeout_time) {
+  bool locked = true;
+
+  if (timeout_time == 0) {
+    locked = mutex_.try_lock();
+  } else {
+    // Previously, this code used a std::timed_mutex.  However, this was changed
+    // due to known bugs in gcc versions < 4.9.
+    // https://gcc.gnu.org/bugzilla/show_bug.cgi?id=54562
+    //
+    // Since this mutex isn't held for long and only a single mutex is ever
+    // held at a time, it is reasonable to ignore the lock timeout_time here
+    // and only check it when waiting on the condition_variable.
+    mutex_.lock();
+  }
+
+  if (!locked) {
+    // timeout acquiring mutex
+    return Status::TimedOut(Status::SubCode::kMutexTimeout);
+  }
+
+  return Status::OK();
+}
+
+Status TransactionDBCondVarImpl::Wait(
+    std::shared_ptr<TransactionDBMutex> mutex) {
+  auto mutex_impl = reinterpret_cast<TransactionDBMutexImpl*>(mutex.get());
+
+  std::unique_lock<std::mutex> lock(mutex_impl->mutex_, std::adopt_lock);
+  cv_.wait(lock);
+
+  // Make sure unique_lock doesn't unlock mutex when it destructs
+  lock.release();
+
+  return Status::OK();
+}
+
+Status TransactionDBCondVarImpl::WaitFor(
+    std::shared_ptr<TransactionDBMutex> mutex, int64_t timeout_time) {
+  Status s;
+
+  auto mutex_impl = reinterpret_cast<TransactionDBMutexImpl*>(mutex.get());
+  std::unique_lock<std::mutex> lock(mutex_impl->mutex_, std::adopt_lock);
+
+  if (timeout_time < 0) {
+    // If timeout is negative, do not use a timeout
+    cv_.wait(lock);
+  } else {
+    auto duration = std::chrono::microseconds(timeout_time);
+    auto cv_status = cv_.wait_for(lock, duration);
+
+    // Check if the wait stopped due to timing out.
+    if (cv_status == std::cv_status::timeout) {
+      s = Status::TimedOut(Status::SubCode::kMutexTimeout);
+    }
+  }
+
+  // Make sure unique_lock doesn't unlock mutex when it destructs
+  lock.release();
+
+  // CV was signaled, or we spuriously woke up (but didn't time out)
+  return s;
+}
+
+}  // namespace ROCKSDB_NAMESPACE
+
+#endif  // ROCKSDB_LITE
diff --git a/src/rocksdb/utilities/transactions/transaction_db_mutex_impl.h b/src/rocksdb/utilities/transactions/transaction_db_mutex_impl.h
new file mode 100644
index 000000000..fbee92832
--- /dev/null
+++ b/src/rocksdb/utilities/transactions/transaction_db_mutex_impl.h
@@ -0,0 +1,26 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#pragma once
+#ifndef ROCKSDB_LITE
+
+#include "rocksdb/utilities/transaction_db_mutex.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+class TransactionDBMutex;
+class TransactionDBCondVar;
+
+// Default implementation of TransactionDBMutexFactory.  May be overridden
+// by TransactionDBOptions.custom_mutex_factory.
+class TransactionDBMutexFactoryImpl : public TransactionDBMutexFactory {
+ public:
+  std::shared_ptr<TransactionDBMutex> AllocateMutex() override;
+  std::shared_ptr<TransactionDBCondVar> AllocateCondVar() override;
+};
+
+}  // namespace ROCKSDB_NAMESPACE
+
+#endif  // ROCKSDB_LITE
diff --git a/src/rocksdb/utilities/transactions/transaction_test.cc b/src/rocksdb/utilities/transactions/transaction_test.cc
new file mode 100644
index 000000000..caf1566b9
--- /dev/null
+++ b/src/rocksdb/utilities/transactions/transaction_test.cc
@@ -0,0 +1,6550 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#ifndef ROCKSDB_LITE
+
+#include "utilities/transactions/transaction_test.h"
+
+#include <algorithm>
+#include <functional>
+#include <string>
+#include <thread>
+
+#include "db/db_impl/db_impl.h"
+#include "port/port.h"
+#include "rocksdb/db.h"
+#include "rocksdb/options.h"
+#include "rocksdb/perf_context.h"
+#include "rocksdb/utilities/transaction.h"
+#include "rocksdb/utilities/transaction_db.h"
+#include "table/mock_table.h"
+#include "test_util/sync_point.h"
+#include "test_util/testharness.h"
+#include "test_util/testutil.h"
+#include "test_util/transaction_test_util.h"
+#include "util/random.h"
+#include "util/string_util.h"
+#include "utilities/fault_injection_env.h"
+#include "utilities/merge_operators.h"
+#include "utilities/merge_operators/string_append/stringappend.h"
+#include "utilities/transactions/pessimistic_transaction_db.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+INSTANTIATE_TEST_CASE_P(
+    DBAsBaseDB, TransactionTest,
+    ::testing::Values(
+        std::make_tuple(false, false, WRITE_COMMITTED, kOrderedWrite),
+        std::make_tuple(false, true, WRITE_COMMITTED, kOrderedWrite),
+        std::make_tuple(false, false, WRITE_PREPARED, kOrderedWrite),
+        std::make_tuple(false, true, WRITE_PREPARED, kOrderedWrite),
+        std::make_tuple(false, true, WRITE_PREPARED, kUnorderedWrite),
+        std::make_tuple(false, false, WRITE_UNPREPARED, kOrderedWrite),
+        std::make_tuple(false, true, WRITE_UNPREPARED, kOrderedWrite)));
+INSTANTIATE_TEST_CASE_P(
+    DBAsBaseDB, TransactionStressTest,
+    ::testing::Values(
+        std::make_tuple(false, false, WRITE_COMMITTED, kOrderedWrite),
+        std::make_tuple(false, true, WRITE_COMMITTED, kOrderedWrite),
+        std::make_tuple(false, false, WRITE_PREPARED, kOrderedWrite),
+        std::make_tuple(false, true, WRITE_PREPARED, kOrderedWrite),
+        std::make_tuple(false, true, WRITE_PREPARED, kUnorderedWrite),
+        std::make_tuple(false, false, WRITE_UNPREPARED, kOrderedWrite),
+        std::make_tuple(false, true, WRITE_UNPREPARED, kOrderedWrite)));
+INSTANTIATE_TEST_CASE_P(
+    StackableDBAsBaseDB, TransactionTest,
+    ::testing::Values(
+        std::make_tuple(true, true, WRITE_COMMITTED, kOrderedWrite),
+        std::make_tuple(true, true, WRITE_PREPARED, kOrderedWrite),
+        std::make_tuple(true, true, WRITE_UNPREPARED, kOrderedWrite)));
+
+// MySQLStyleTransactionTest takes far too long for valgrind to run. Only do it
+// in full mode (`ROCKSDB_FULL_VALGRIND_RUN` compiler flag is set).
+#if !defined(ROCKSDB_VALGRIND_RUN) || defined(ROCKSDB_FULL_VALGRIND_RUN)
+INSTANTIATE_TEST_CASE_P(
+    MySQLStyleTransactionTest, MySQLStyleTransactionTest,
+    ::testing::Values(
+        std::make_tuple(false, false, WRITE_COMMITTED, kOrderedWrite, false),
+        std::make_tuple(false, true, WRITE_COMMITTED, kOrderedWrite, false),
+        std::make_tuple(false, false, WRITE_PREPARED, kOrderedWrite, false),
+        std::make_tuple(false, false, WRITE_PREPARED, kOrderedWrite, true),
+        std::make_tuple(false, true, WRITE_PREPARED, kOrderedWrite, false),
+        std::make_tuple(false, true, WRITE_PREPARED, kOrderedWrite, true),
+        std::make_tuple(false, false, WRITE_UNPREPARED, kOrderedWrite, false),
+        std::make_tuple(false, false, WRITE_UNPREPARED, kOrderedWrite, true),
+        std::make_tuple(false, true, WRITE_UNPREPARED, kOrderedWrite, false),
+        std::make_tuple(false, true, WRITE_UNPREPARED, kOrderedWrite, true),
+        std::make_tuple(false, true, WRITE_PREPARED, kUnorderedWrite, false),
+        std::make_tuple(false, true, WRITE_PREPARED, kUnorderedWrite, true)));
+#endif  // !defined(ROCKSDB_VALGRIND_RUN) || defined(ROCKSDB_FULL_VALGRIND_RUN)
+
+TEST_P(TransactionTest, DoubleEmptyWrite) {
+  WriteOptions write_options;
+  write_options.sync = true;
+  write_options.disableWAL = false;
+
+  WriteBatch batch;
+
+  ASSERT_OK(db->Write(write_options, &batch));
+  ASSERT_OK(db->Write(write_options, &batch));
+
+  // Also test committing empty transactions in 2PC
+  TransactionOptions txn_options;
+  Transaction* txn0 = db->BeginTransaction(write_options, txn_options);
+  ASSERT_OK(txn0->SetName("xid"));
+  ASSERT_OK(txn0->Prepare());
+  ASSERT_OK(txn0->Commit());
+  delete txn0;
+
+  // Also test that it works during recovery
+  txn0 = db->BeginTransaction(write_options, txn_options);
+  ASSERT_OK(txn0->SetName("xid2"));
+  ASSERT_OK(txn0->Put(Slice("foo0"), Slice("bar0a")));
+  ASSERT_OK(txn0->Prepare());
+  delete txn0;
+  reinterpret_cast<PessimisticTransactionDB*>(db)->TEST_Crash();
+  ASSERT_OK(ReOpenNoDelete());
+  assert(db != nullptr);
+  txn0 = db->GetTransactionByName("xid2");
+  ASSERT_OK(txn0->Commit());
+  delete txn0;
+}
+
+TEST_P(TransactionTest, SuccessTest) {
+  ASSERT_OK(db->ResetStats());
+
+  WriteOptions write_options;
+  ReadOptions read_options;
+  std::string value;
+
+  ASSERT_OK(db->Put(write_options, Slice("foo"), Slice("bar")));
+  ASSERT_OK(db->Put(write_options, Slice("foo2"), Slice("bar")));
+
+  Transaction* txn = db->BeginTransaction(write_options, TransactionOptions());
+  ASSERT_TRUE(txn);
+
+  ASSERT_EQ(0, txn->GetNumPuts());
+  ASSERT_LE(0, txn->GetID());
+
+  ASSERT_OK(txn->GetForUpdate(read_options, "foo", &value));
+  ASSERT_EQ(value, "bar");
+
+  ASSERT_OK(txn->Put(Slice("foo"), Slice("bar2")));
+
+  ASSERT_EQ(1, txn->GetNumPuts());
+
+  ASSERT_OK(txn->GetForUpdate(read_options, "foo", &value));
+  ASSERT_EQ(value, "bar2");
+
+  ASSERT_OK(txn->Commit());
+
+  ASSERT_OK(db->Get(read_options, "foo", &value));
+  ASSERT_EQ(value, "bar2");
+
+  delete txn;
+}
+
+TEST_P(TransactionTest, SwitchMemtableDuringPrepareAndCommit_WC) {
+  const TxnDBWritePolicy write_policy = std::get<2>(GetParam());
+
+  if (write_policy != TxnDBWritePolicy::WRITE_COMMITTED) {
+    ROCKSDB_GTEST_BYPASS("Test applies to write-committed only");
+    return;
+  }
+
+  ASSERT_OK(db->Put(WriteOptions(), "key0", "value"));
+
+  TransactionOptions txn_opts;
+  txn_opts.use_only_the_last_commit_time_batch_for_recovery = true;
+  Transaction* txn = db->BeginTransaction(WriteOptions(), txn_opts);
+  assert(txn);
+
+  SyncPoint::GetInstance()->DisableProcessing();
+  SyncPoint::GetInstance()->ClearAllCallBacks();
+  SyncPoint::GetInstance()->SetCallBack(
+      "FlushJob::WriteLevel0Table", [&](void* arg) {
+        // db mutex not held.
+        auto* mems = reinterpret_cast<autovector<MemTable*>*>(arg);
+        assert(mems);
+        ASSERT_EQ(1, mems->size());
+        auto* ctwb = txn->GetCommitTimeWriteBatch();
+        ASSERT_OK(ctwb->Put("gtid", "123"));
+        ASSERT_OK(txn->Commit());
+        delete txn;
+      });
+  SyncPoint::GetInstance()->EnableProcessing();
+
+  ASSERT_OK(txn->Put("key1", "value"));
+  ASSERT_OK(txn->SetName("txn1"));
+
+  ASSERT_OK(txn->Prepare());
+
+  auto dbimpl = static_cast_with_check<DBImpl>(db->GetRootDB());
+  ASSERT_OK(dbimpl->TEST_SwitchMemtable(nullptr));
+  ASSERT_OK(dbimpl->TEST_FlushMemTable(
+      /*wait=*/false, /*allow_write_stall=*/true, /*cfh=*/nullptr));
+
+  ASSERT_OK(dbimpl->TEST_WaitForFlushMemTable());
+
+  {
+    std::string value;
+    ASSERT_OK(db->Get(ReadOptions(), "key1", &value));
+    ASSERT_EQ("value", value);
+  }
+
+  delete db;
+  db = nullptr;
+  Status s;
+  if (use_stackable_db_ == false) {
+    s = TransactionDB::Open(options, txn_db_options, dbname, &db);
+  } else {
+    s = OpenWithStackableDB();
+  }
+  ASSERT_OK(s);
+  assert(db);
+
+  {
+    std::string value;
+    ASSERT_OK(db->Get(ReadOptions(), "gtid", &value));
+    ASSERT_EQ("123", value);
+
+    ASSERT_OK(db->Get(ReadOptions(), "key1", &value));
+    ASSERT_EQ("value", value);
+  }
+
+  SyncPoint::GetInstance()->DisableProcessing();
+  SyncPoint::GetInstance()->ClearAllCallBacks();
+}
+
+// The test clarifies the contract of do_validate and assume_tracked
+// in GetForUpdate and Put/Merge/Delete
+TEST_P(TransactionTest, AssumeExclusiveTracked) {
+  WriteOptions write_options;
+  ReadOptions read_options;
+  std::string value;
+  Status s;
+  TransactionOptions txn_options;
+  txn_options.lock_timeout = 1;
+  const bool EXCLUSIVE = true;
+  const bool DO_VALIDATE = true;
+  const bool ASSUME_LOCKED = true;
+
+  Transaction* txn = db->BeginTransaction(write_options, txn_options);
+  ASSERT_TRUE(txn);
+  txn->SetSnapshot();
+
+  // commit a value after the snapshot is taken
+  ASSERT_OK(db->Put(write_options, Slice("foo"), Slice("bar")));
+
+  // By default write should fail to the commit after our snapshot
+  s = txn->GetForUpdate(read_options, "foo", &value, EXCLUSIVE);
+  ASSERT_TRUE(s.IsBusy());
+  // But the user could direct the db to skip validating the snapshot. The read
+  // value then should be the most recently committed
+  ASSERT_OK(
+      txn->GetForUpdate(read_options, "foo", &value, EXCLUSIVE, !DO_VALIDATE));
+  ASSERT_EQ(value, "bar");
+
+  // Although ValidateSnapshot is skipped the key must have still got locked
+  s = db->Put(write_options, Slice("foo"), Slice("bar"));
+  ASSERT_TRUE(s.IsTimedOut());
+
+  // By default the write operations should fail due to the commit after the
+  // snapshot
+  s = txn->Put(Slice("foo"), Slice("bar1"));
+  ASSERT_TRUE(s.IsBusy());
+  s = txn->Put(db->DefaultColumnFamily(), Slice("foo"), Slice("bar1"),
+               !ASSUME_LOCKED);
+  ASSERT_TRUE(s.IsBusy());
+  // But the user could direct the db that it already assumes exclusive lock on
+  // the key due to the previous GetForUpdate call.
+  ASSERT_OK(txn->Put(db->DefaultColumnFamily(), Slice("foo"), Slice("bar1"),
+                     ASSUME_LOCKED));
+  ASSERT_OK(txn->Merge(db->DefaultColumnFamily(), Slice("foo"), Slice("bar2"),
+                       ASSUME_LOCKED));
+  ASSERT_OK(
+      txn->Delete(db->DefaultColumnFamily(), Slice("foo"), ASSUME_LOCKED));
+  ASSERT_OK(txn->SingleDelete(db->DefaultColumnFamily(), Slice("foo"),
+                              ASSUME_LOCKED));
+
+  ASSERT_OK(txn->Rollback());
+  delete txn;
+}
+
+// This test clarifies the contract of ValidateSnapshot
+TEST_P(TransactionTest, ValidateSnapshotTest) {
+  for (bool with_flush : {true}) {
+    for (bool with_2pc : {true}) {
+      ASSERT_OK(ReOpen());
+      WriteOptions write_options;
+      ReadOptions read_options;
+      std::string value;
+
+      assert(db != nullptr);
+      Transaction* txn1 =
+          db->BeginTransaction(write_options, TransactionOptions());
+      ASSERT_TRUE(txn1);
+      ASSERT_OK(txn1->Put(Slice("foo"), Slice("bar1")));
+      if (with_2pc) {
+        ASSERT_OK(txn1->SetName("xid1"));
+        ASSERT_OK(txn1->Prepare());
+      }
+
+      if (with_flush) {
+        auto db_impl = static_cast_with_check<DBImpl>(db->GetRootDB());
+        ASSERT_OK(db_impl->TEST_FlushMemTable(true));
+        // Make sure the flushed memtable is not kept in memory
+        int max_memtable_in_history =
+            std::max(
+                options.max_write_buffer_number,
+                static_cast<int>(options.max_write_buffer_size_to_maintain) /
+                    static_cast<int>(options.write_buffer_size)) +
+            1;
+        for (int i = 0; i < max_memtable_in_history; i++) {
+          ASSERT_OK(db->Put(write_options, Slice("key"), Slice("value")));
+          ASSERT_OK(db_impl->TEST_FlushMemTable(true));
+        }
+      }
+
+      Transaction* txn2 =
+          db->BeginTransaction(write_options, TransactionOptions());
+      ASSERT_TRUE(txn2);
+      txn2->SetSnapshot();
+
+      ASSERT_OK(txn1->Commit());
+      delete txn1;
+
+      auto pes_txn2 = dynamic_cast<PessimisticTransaction*>(txn2);
+      // Test the simple case where the key is not tracked yet
+      auto trakced_seq = kMaxSequenceNumber;
+      auto s = pes_txn2->ValidateSnapshot(db->DefaultColumnFamily(), "foo",
+                                          &trakced_seq);
+      ASSERT_TRUE(s.IsBusy());
+      delete txn2;
+    }
+  }
+}
+
+TEST_P(TransactionTest, WaitingTxn) {
+  WriteOptions write_options;
+  ReadOptions read_options;
+  TransactionOptions txn_options;
+  std::string value;
+  Status s;
+
+  txn_options.lock_timeout = 1;
+  s = db->Put(write_options, Slice("foo"), Slice("bar"));
+  ASSERT_OK(s);
+
+  /* create second cf */
+  ColumnFamilyHandle* cfa;
+  ColumnFamilyOptions cf_options;
+  s = db->CreateColumnFamily(cf_options, "CFA", &cfa);
+  ASSERT_OK(s);
+  s = db->Put(write_options, cfa, Slice("foo"), Slice("bar"));
+  ASSERT_OK(s);
+
+  Transaction* txn1 = db->BeginTransaction(write_options, txn_options);
+  Transaction* txn2 = db->BeginTransaction(write_options, txn_options);
+  TransactionID id1 = txn1->GetID();
+  ASSERT_TRUE(txn1);
+  ASSERT_TRUE(txn2);
+
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
+      "PointLockManager::AcquireWithTimeout:WaitingTxn", [&](void* /*arg*/) {
+        std::string key;
+        uint32_t cf_id;
+        std::vector<TransactionID> wait = txn2->GetWaitingTxns(&cf_id, &key);
+        ASSERT_EQ(key, "foo");
+        ASSERT_EQ(wait.size(), 1);
+        ASSERT_EQ(wait[0], id1);
+        ASSERT_EQ(cf_id, 0U);
+      });
+
+  get_perf_context()->Reset();
+  // lock key in default cf
+  s = txn1->GetForUpdate(read_options, "foo", &value);
+  ASSERT_OK(s);
+  ASSERT_EQ(value, "bar");
+  ASSERT_EQ(get_perf_context()->key_lock_wait_count, 0);
+
+  // lock key in cfa
+  s = txn1->GetForUpdate(read_options, cfa, "foo", &value);
+  ASSERT_OK(s);
+  ASSERT_EQ(value, "bar");
+  ASSERT_EQ(get_perf_context()->key_lock_wait_count, 0);
+
+  auto lock_data = db->GetLockStatusData();
+  // Locked keys exist in both column family.
+  ASSERT_EQ(lock_data.size(), 2);
+
+  auto cf_iterator = lock_data.begin();
+
+  // The iterator points to an unordered_multimap
+  // thus the test can not assume any particular order.
+
+  // Column family is 1 or 0 (cfa).
+  if (cf_iterator->first != 1 && cf_iterator->first != 0) {
+    FAIL();
+  }
+  // The locked key is "foo" and is locked by txn1
+  ASSERT_EQ(cf_iterator->second.key, "foo");
+  ASSERT_EQ(cf_iterator->second.ids.size(), 1);
+  ASSERT_EQ(cf_iterator->second.ids[0], txn1->GetID());
+
+  cf_iterator++;
+
+  // Column family is 0 (default) or 1.
+  if (cf_iterator->first != 1 && cf_iterator->first != 0) {
+    FAIL();
+  }
+  // The locked key is "foo" and is locked by txn1
+  ASSERT_EQ(cf_iterator->second.key, "foo");
+  ASSERT_EQ(cf_iterator->second.ids.size(), 1);
+  ASSERT_EQ(cf_iterator->second.ids[0], txn1->GetID());
+
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->EnableProcessing();
+
+  s = txn2->GetForUpdate(read_options, "foo", &value);
+  ASSERT_TRUE(s.IsTimedOut());
+  ASSERT_EQ(s.ToString(), "Operation timed out: Timeout waiting to lock key");
+  ASSERT_EQ(get_perf_context()->key_lock_wait_count, 1);
+  ASSERT_GE(get_perf_context()->key_lock_wait_time, 0);
+
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->DisableProcessing();
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->ClearAllCallBacks();
+
+  delete cfa;
+  delete txn1;
+  delete txn2;
+}
+
+TEST_P(TransactionTest, SharedLocks) {
+  WriteOptions write_options;
+  ReadOptions read_options;
+  TransactionOptions txn_options;
+  Status s;
+
+  txn_options.lock_timeout = 1;
+  s = db->Put(write_options, Slice("foo"), Slice("bar"));
+  ASSERT_OK(s);
+
+  Transaction* txn1 = db->BeginTransaction(write_options, txn_options);
+  Transaction* txn2 = db->BeginTransaction(write_options, txn_options);
+  Transaction* txn3 = db->BeginTransaction(write_options, txn_options);
+  ASSERT_TRUE(txn1);
+  ASSERT_TRUE(txn2);
+  ASSERT_TRUE(txn3);
+
+  // Test shared access between txns
+  s = txn1->GetForUpdate(read_options, "foo", nullptr, false /* exclusive */);
+  ASSERT_OK(s);
+
+  s = txn2->GetForUpdate(read_options, "foo", nullptr, false /* exclusive */);
+  ASSERT_OK(s);
+
+  s = txn3->GetForUpdate(read_options, "foo", nullptr, false /* exclusive */);
+  ASSERT_OK(s);
+
+  auto lock_data = db->GetLockStatusData();
+  ASSERT_EQ(lock_data.size(), 1);
+
+  auto cf_iterator = lock_data.begin();
+  ASSERT_EQ(cf_iterator->second.key, "foo");
+
+  // We compare whether the set of txns locking this key is the same. To do
+  // this, we need to sort both vectors so that the comparison is done
+  // correctly.
+  std::vector<TransactionID> expected_txns = {txn1->GetID(), txn2->GetID(),
+                                              txn3->GetID()};
+  std::vector<TransactionID> lock_txns = cf_iterator->second.ids;
+  ASSERT_EQ(expected_txns, lock_txns);
+  ASSERT_FALSE(cf_iterator->second.exclusive);
+
+  ASSERT_OK(txn1->Rollback());
+  ASSERT_OK(txn2->Rollback());
+  ASSERT_OK(txn3->Rollback());
+
+  // Test txn1 and txn2 sharing a lock and txn3 trying to obtain it.
+  s = txn1->GetForUpdate(read_options, "foo", nullptr, false /* exclusive */);
+  ASSERT_OK(s);
+
+  s = txn2->GetForUpdate(read_options, "foo", nullptr, false /* exclusive */);
+  ASSERT_OK(s);
+
+  s = txn3->GetForUpdate(read_options, "foo", nullptr);
+  ASSERT_TRUE(s.IsTimedOut());
+  ASSERT_EQ(s.ToString(), "Operation timed out: Timeout waiting to lock key");
+
+  txn1->UndoGetForUpdate("foo");
+  s = txn3->GetForUpdate(read_options, "foo", nullptr);
+  ASSERT_TRUE(s.IsTimedOut());
+  ASSERT_EQ(s.ToString(), "Operation timed out: Timeout waiting to lock key");
+
+  txn2->UndoGetForUpdate("foo");
+  s = txn3->GetForUpdate(read_options, "foo", nullptr);
+  ASSERT_OK(s);
+
+  ASSERT_OK(txn1->Rollback());
+  ASSERT_OK(txn2->Rollback());
+  ASSERT_OK(txn3->Rollback());
+
+  // Test txn1 and txn2 sharing a lock and txn2 trying to upgrade lock.
+  s = txn1->GetForUpdate(read_options, "foo", nullptr, false /* exclusive */);
+  ASSERT_OK(s);
+
+  s = txn2->GetForUpdate(read_options, "foo", nullptr, false /* exclusive */);
+  ASSERT_OK(s);
+
+  s = txn2->GetForUpdate(read_options, "foo", nullptr);
+  ASSERT_TRUE(s.IsTimedOut());
+  ASSERT_EQ(s.ToString(), "Operation timed out: Timeout waiting to lock key");
+
+  txn1->UndoGetForUpdate("foo");
+  s = txn2->GetForUpdate(read_options, "foo", nullptr);
+  ASSERT_OK(s);
+
+  ASSERT_OK(txn1->Rollback());
+  ASSERT_OK(txn2->Rollback());
+
+  // Test txn1 trying to downgrade its lock.
+  s = txn1->GetForUpdate(read_options, "foo", nullptr, true /* exclusive */);
+  ASSERT_OK(s);
+
+  s = txn2->GetForUpdate(read_options, "foo", nullptr, false /* exclusive */);
+  ASSERT_TRUE(s.IsTimedOut());
+  ASSERT_EQ(s.ToString(), "Operation timed out: Timeout waiting to lock key");
+
+  // Should still fail after "downgrading".
+  s = txn1->GetForUpdate(read_options, "foo", nullptr, false /* exclusive */);
+  ASSERT_OK(s);
+
+  s = txn2->GetForUpdate(read_options, "foo", nullptr, false /* exclusive */);
+  ASSERT_TRUE(s.IsTimedOut());
+  ASSERT_EQ(s.ToString(), "Operation timed out: Timeout waiting to lock key");
+
+  ASSERT_OK(txn1->Rollback());
+  ASSERT_OK(txn2->Rollback());
+
+  // Test txn1 holding an exclusive lock and txn2 trying to obtain shared
+  // access.
+  s = txn1->GetForUpdate(read_options, "foo", nullptr);
+  ASSERT_OK(s);
+
+  s = txn2->GetForUpdate(read_options, "foo", nullptr, false /* exclusive */);
+  ASSERT_TRUE(s.IsTimedOut());
+  ASSERT_EQ(s.ToString(), "Operation timed out: Timeout waiting to lock key");
+
+  txn1->UndoGetForUpdate("foo");
+  s = txn2->GetForUpdate(read_options, "foo", nullptr, false /* exclusive */);
+  ASSERT_OK(s);
+
+  delete txn1;
+  delete txn2;
+  delete txn3;
+}
+
+TEST_P(TransactionTest, DeadlockCycleShared) {
+  WriteOptions write_options;
+  ReadOptions read_options;
+  TransactionOptions txn_options;
+
+  txn_options.lock_timeout = 1000000;
+  txn_options.deadlock_detect = true;
+
+  // Set up a wait for chain like this:
+  //
+  // Tn -> T(n*2)
+  // Tn -> T(n*2 + 1)
+  //
+  // So we have:
+  // T1 -> T2 -> T4 ...
+  //    |     |> T5 ...
+  //    |> T3 -> T6 ...
+  //          |> T7 ...
+  // up to T31, then T[16 - 31] -> T1.
+  // Note that Tn holds lock on floor(n / 2).
+
+  std::vector<Transaction*> txns(31);
+
+  for (uint32_t i = 0; i < 31; i++) {
+    txns[i] = db->BeginTransaction(write_options, txn_options);
+    ASSERT_TRUE(txns[i]);
+    auto s = txns[i]->GetForUpdate(read_options, std::to_string((i + 1) / 2),
+                                   nullptr, false /* exclusive */);
+    ASSERT_OK(s);
+  }
+
+  std::atomic<uint32_t> checkpoints(0);
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
+      "PointLockManager::AcquireWithTimeout:WaitingTxn",
+      [&](void* /*arg*/) { checkpoints.fetch_add(1); });
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->EnableProcessing();
+
+  // We want the leaf transactions to block and hold everyone back.
+  std::vector<port::Thread> threads;
+  for (uint32_t i = 0; i < 15; i++) {
+    std::function<void()> blocking_thread = [&, i] {
+      auto s = txns[i]->GetForUpdate(read_options, std::to_string(i + 1),
+                                     nullptr, true /* exclusive */);
+      ASSERT_OK(s);
+      ASSERT_OK(txns[i]->Rollback());
+      delete txns[i];
+    };
+    threads.emplace_back(blocking_thread);
+  }
+
+  // Wait until all threads are waiting on each other.
+  while (checkpoints.load() != 15) {
+    /* sleep override */
+    std::this_thread::sleep_for(std::chrono::milliseconds(100));
+  }
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->DisableProcessing();
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->ClearAllCallBacks();
+
+  // Complete the cycle T[16 - 31] -> T1
+  for (uint32_t i = 15; i < 31; i++) {
+    auto s =
+        txns[i]->GetForUpdate(read_options, "0", nullptr, true /* exclusive */);
+    ASSERT_TRUE(s.IsDeadlock());
+
+    // Calculate next buffer len, plateau at 5 when 5 records are inserted.
+    const uint32_t curr_dlock_buffer_len_ =
+        (i - 14 > kInitialMaxDeadlocks) ? kInitialMaxDeadlocks : (i - 14);
+
+    auto dlock_buffer = db->GetDeadlockInfoBuffer();
+    ASSERT_EQ(dlock_buffer.size(), curr_dlock_buffer_len_);
+    auto dlock_entry = dlock_buffer[0].path;
+    ASSERT_EQ(dlock_entry.size(), kInitialMaxDeadlocks);
+    int64_t pre_deadlock_time = dlock_buffer[0].deadlock_time;
+    int64_t cur_deadlock_time = 0;
+    for (auto const& dl_path_rec : dlock_buffer) {
+      cur_deadlock_time = dl_path_rec.deadlock_time;
+      ASSERT_NE(cur_deadlock_time, 0);
+      ASSERT_TRUE(cur_deadlock_time <= pre_deadlock_time);
+      pre_deadlock_time = cur_deadlock_time;
+    }
+
+    int64_t curr_waiting_key = 0;
+
+    // Offset of each txn id from the root of the shared dlock tree's txn id.
+    int64_t offset_root = dlock_entry[0].m_txn_id - 1;
+    // Offset of the final entry in the dlock path from the root's txn id.
+    TransactionID leaf_id =
+        dlock_entry[dlock_entry.size() - 1].m_txn_id - offset_root;
+
+    for (auto it = dlock_entry.rbegin(); it != dlock_entry.rend(); ++it) {
+      auto dl_node = *it;
+      ASSERT_EQ(dl_node.m_txn_id, offset_root + leaf_id);
+      ASSERT_EQ(dl_node.m_cf_id, 0U);
+      ASSERT_EQ(dl_node.m_waiting_key, std::to_string(curr_waiting_key));
+      ASSERT_EQ(dl_node.m_exclusive, true);
+
+      if (curr_waiting_key == 0) {
+        curr_waiting_key = leaf_id;
+      }
+      curr_waiting_key /= 2;
+      leaf_id /= 2;
+    }
+  }
+
+  // Rollback the leaf transaction.
+  for (uint32_t i = 15; i < 31; i++) {
+    ASSERT_OK(txns[i]->Rollback());
+    delete txns[i];
+  }
+
+  for (auto& t : threads) {
+    t.join();
+  }
+
+  // Downsize the buffer and verify the 3 latest deadlocks are preserved.
+  auto dlock_buffer_before_resize = db->GetDeadlockInfoBuffer();
+  db->SetDeadlockInfoBufferSize(3);
+  auto dlock_buffer_after_resize = db->GetDeadlockInfoBuffer();
+  ASSERT_EQ(dlock_buffer_after_resize.size(), 3);
+
+  for (uint32_t i = 0; i < dlock_buffer_after_resize.size(); i++) {
+    for (uint32_t j = 0; j < dlock_buffer_after_resize[i].path.size(); j++) {
+      ASSERT_EQ(dlock_buffer_after_resize[i].path[j].m_txn_id,
+                dlock_buffer_before_resize[i].path[j].m_txn_id);
+    }
+  }
+
+  // Upsize the buffer and verify the 3 latest dealocks are preserved.
+  dlock_buffer_before_resize = db->GetDeadlockInfoBuffer();
+  db->SetDeadlockInfoBufferSize(5);
+  dlock_buffer_after_resize = db->GetDeadlockInfoBuffer();
+  ASSERT_EQ(dlock_buffer_after_resize.size(), 3);
+
+  for (uint32_t i = 0; i < dlock_buffer_before_resize.size(); i++) {
+    for (uint32_t j = 0; j < dlock_buffer_before_resize[i].path.size(); j++) {
+      ASSERT_EQ(dlock_buffer_after_resize[i].path[j].m_txn_id,
+                dlock_buffer_before_resize[i].path[j].m_txn_id);
+    }
+  }
+
+  // Downsize to 0 and verify the size is consistent.
+  dlock_buffer_before_resize = db->GetDeadlockInfoBuffer();
+  db->SetDeadlockInfoBufferSize(0);
+  dlock_buffer_after_resize = db->GetDeadlockInfoBuffer();
+  ASSERT_EQ(dlock_buffer_after_resize.size(), 0);
+
+  // Upsize from 0 to verify the size is persistent.
+  dlock_buffer_before_resize = db->GetDeadlockInfoBuffer();
+  db->SetDeadlockInfoBufferSize(3);
+  dlock_buffer_after_resize = db->GetDeadlockInfoBuffer();
+  ASSERT_EQ(dlock_buffer_after_resize.size(), 0);
+
+  // Contrived case of shared lock of cycle size 2 to verify that a shared
+  // lock causing a deadlock is correctly reported as "shared" in the buffer.
+  std::vector<Transaction*> txns_shared(2);
+
+  // Create a cycle of size 2.
+  for (uint32_t i = 0; i < 2; i++) {
+    txns_shared[i] = db->BeginTransaction(write_options, txn_options);
+    ASSERT_TRUE(txns_shared[i]);
+    auto s =
+        txns_shared[i]->GetForUpdate(read_options, std::to_string(i), nullptr);
+    ASSERT_OK(s);
+  }
+
+  std::atomic<uint32_t> checkpoints_shared(0);
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
+      "PointLockManager::AcquireWithTimeout:WaitingTxn",
+      [&](void* /*arg*/) { checkpoints_shared.fetch_add(1); });
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->EnableProcessing();
+
+  std::vector<port::Thread> threads_shared;
+  for (uint32_t i = 0; i < 1; i++) {
+    std::function<void()> blocking_thread = [&, i] {
+      auto s = txns_shared[i]->GetForUpdate(read_options, std::to_string(i + 1),
+                                            nullptr);
+      ASSERT_OK(s);
+      ASSERT_OK(txns_shared[i]->Rollback());
+      delete txns_shared[i];
+    };
+    threads_shared.emplace_back(blocking_thread);
+  }
+
+  // Wait until all threads are waiting on each other.
+  while (checkpoints_shared.load() != 1) {
+    /* sleep override */
+    std::this_thread::sleep_for(std::chrono::milliseconds(100));
+  }
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->DisableProcessing();
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->ClearAllCallBacks();
+
+  // Complete the cycle T2 -> T1 with a shared lock.
+  auto s = txns_shared[1]->GetForUpdate(read_options, "0", nullptr, false);
+  ASSERT_TRUE(s.IsDeadlock());
+
+  auto dlock_buffer = db->GetDeadlockInfoBuffer();
+
+  // Verify the size of the buffer and the single path.
+  ASSERT_EQ(dlock_buffer.size(), 1);
+  ASSERT_EQ(dlock_buffer[0].path.size(), 2);
+
+  // Verify the exclusivity field of the transactions in the deadlock path.
+  ASSERT_TRUE(dlock_buffer[0].path[0].m_exclusive);
+  ASSERT_FALSE(dlock_buffer[0].path[1].m_exclusive);
+  ASSERT_OK(txns_shared[1]->Rollback());
+  delete txns_shared[1];
+
+  for (auto& t : threads_shared) {
+    t.join();
+  }
+}
+
+#if !defined(ROCKSDB_VALGRIND_RUN) || defined(ROCKSDB_FULL_VALGRIND_RUN)
+TEST_P(TransactionStressTest, DeadlockCycle) {
+  WriteOptions write_options;
+  ReadOptions read_options;
+  TransactionOptions txn_options;
+
+  // offset by 2 from the max depth to test edge case
+  const uint32_t kMaxCycleLength = 52;
+
+  txn_options.lock_timeout = 1000000;
+  txn_options.deadlock_detect = true;
+
+  for (uint32_t len = 2; len < kMaxCycleLength; len++) {
+    // Set up a long wait for chain like this:
+    //
+    // T1 -> T2 -> T3 -> ... -> Tlen
+
+    std::vector<Transaction*> txns(len);
+
+    for (uint32_t i = 0; i < len; i++) {
+      txns[i] = db->BeginTransaction(write_options, txn_options);
+      ASSERT_TRUE(txns[i]);
+      auto s = txns[i]->GetForUpdate(read_options, std::to_string(i), nullptr);
+      ASSERT_OK(s);
+    }
+
+    std::atomic<uint32_t> checkpoints(0);
+    ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
+        "PointLockManager::AcquireWithTimeout:WaitingTxn",
+        [&](void* /*arg*/) { checkpoints.fetch_add(1); });
+    ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->EnableProcessing();
+
+    // We want the last transaction in the chain to block and hold everyone
+    // back.
+    std::vector<port::Thread> threads;
+    for (uint32_t i = 0; i + 1 < len; i++) {
+      std::function<void()> blocking_thread = [&, i] {
+        auto s =
+            txns[i]->GetForUpdate(read_options, std::to_string(i + 1), nullptr);
+        ASSERT_OK(s);
+        ASSERT_OK(txns[i]->Rollback());
+        delete txns[i];
+      };
+      threads.emplace_back(blocking_thread);
+    }
+
+    // Wait until all threads are waiting on each other.
+    while (checkpoints.load() != len - 1) {
+      /* sleep override */
+      std::this_thread::sleep_for(std::chrono::milliseconds(100));
+    }
+    ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->DisableProcessing();
+    ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->ClearAllCallBacks();
+
+    // Complete the cycle Tlen -> T1
+    auto s = txns[len - 1]->GetForUpdate(read_options, "0", nullptr);
+    ASSERT_TRUE(s.IsDeadlock());
+
+    const uint32_t dlock_buffer_size_ = (len - 1 > 5) ? 5 : (len - 1);
+    uint32_t curr_waiting_key = 0;
+    TransactionID curr_txn_id = txns[0]->GetID();
+
+    auto dlock_buffer = db->GetDeadlockInfoBuffer();
+    ASSERT_EQ(dlock_buffer.size(), dlock_buffer_size_);
+    uint32_t check_len = len;
+    bool check_limit_flag = false;
+
+    // Special case for a deadlock path that exceeds the maximum depth.
+    if (len > 50) {
+      check_len = 0;
+      check_limit_flag = true;
+    }
+    auto dlock_entry = dlock_buffer[0].path;
+    ASSERT_EQ(dlock_entry.size(), check_len);
+    ASSERT_EQ(dlock_buffer[0].limit_exceeded, check_limit_flag);
+
+    int64_t pre_deadlock_time = dlock_buffer[0].deadlock_time;
+    int64_t cur_deadlock_time = 0;
+    for (auto const& dl_path_rec : dlock_buffer) {
+      cur_deadlock_time = dl_path_rec.deadlock_time;
+      ASSERT_NE(cur_deadlock_time, 0);
+      ASSERT_TRUE(cur_deadlock_time <= pre_deadlock_time);
+      pre_deadlock_time = cur_deadlock_time;
+    }
+
+    // Iterates backwards over path verifying decreasing txn_ids.
+    for (auto it = dlock_entry.rbegin(); it != dlock_entry.rend(); ++it) {
+      auto dl_node = *it;
+      ASSERT_EQ(dl_node.m_txn_id, len + curr_txn_id - 1);
+      ASSERT_EQ(dl_node.m_cf_id, 0u);
+      ASSERT_EQ(dl_node.m_waiting_key, std::to_string(curr_waiting_key));
+      ASSERT_EQ(dl_node.m_exclusive, true);
+
+      curr_txn_id--;
+      if (curr_waiting_key == 0) {
+        curr_waiting_key = len;
+      }
+      curr_waiting_key--;
+    }
+
+    // Rollback the last transaction.
+    ASSERT_OK(txns[len - 1]->Rollback());
+    delete txns[len - 1];
+
+    for (auto& t : threads) {
+      t.join();
+    }
+  }
+}
+
+TEST_P(TransactionStressTest, DeadlockStress) {
+  const uint32_t NUM_TXN_THREADS = 10;
+  const uint32_t NUM_KEYS = 100;
+  const uint32_t NUM_ITERS = 1000;
+
+  WriteOptions write_options;
+  ReadOptions read_options;
+  TransactionOptions txn_options;
+
+  txn_options.lock_timeout = 1000000;
+  txn_options.deadlock_detect = true;
+  std::vector<std::string> keys;
+
+  for (uint32_t i = 0; i < NUM_KEYS; i++) {
+    ASSERT_OK(db->Put(write_options, Slice(std::to_string(i)), Slice("")));
+    keys.push_back(std::to_string(i));
+  }
+
+  size_t tid = std::hash<std::thread::id>()(std::this_thread::get_id());
+  Random rnd(static_cast<uint32_t>(tid));
+  std::function<void(uint32_t)> stress_thread = [&](uint32_t seed) {
+    std::default_random_engine g(seed);
+
+    Transaction* txn;
+    for (uint32_t i = 0; i < NUM_ITERS; i++) {
+      txn = db->BeginTransaction(write_options, txn_options);
+      auto random_keys = keys;
+      std::shuffle(random_keys.begin(), random_keys.end(), g);
+
+      // Lock keys in random order.
+      for (const auto& k : random_keys) {
+        // Lock mostly for shared access, but exclusive 1/4 of the time.
+        auto s =
+            txn->GetForUpdate(read_options, k, nullptr, txn->GetID() % 4 == 0);
+        if (!s.ok()) {
+          ASSERT_TRUE(s.IsDeadlock());
+          ASSERT_OK(txn->Rollback());
+          break;
+        }
+      }
+
+      delete txn;
+    }
+  };
+
+  std::vector<port::Thread> threads;
+  for (uint32_t i = 0; i < NUM_TXN_THREADS; i++) {
+    threads.emplace_back(stress_thread, rnd.Next());
+  }
+
+  for (auto& t : threads) {
+    t.join();
+  }
+}
+#endif  // !defined(ROCKSDB_VALGRIND_RUN) || defined(ROCKSDB_FULL_VALGRIND_RUN)
+
+TEST_P(TransactionTest, CommitTimeBatchFailTest) {
+  WriteOptions write_options;
+  TransactionOptions txn_options;
+
+  std::string value;
+  Status s;
+
+  Transaction* txn1 = db->BeginTransaction(write_options, txn_options);
+  ASSERT_TRUE(txn1);
+
+  ASSERT_OK(txn1->GetCommitTimeWriteBatch()->Put("cat", "dog"));
+
+  s = txn1->Put("foo", "bar");
+  ASSERT_OK(s);
+
+  // fails due to non-empty commit-time batch
+  s = txn1->Commit();
+  ASSERT_EQ(s, Status::InvalidArgument());
+
+  delete txn1;
+}
+
+TEST_P(TransactionTest, LogMarkLeakTest) {
+  TransactionOptions txn_options;
+  WriteOptions write_options;
+  options.write_buffer_size = 1024;
+  ASSERT_OK(ReOpenNoDelete());
+  assert(db != nullptr);
+  Random rnd(47);
+  std::vector<Transaction*> txns;
+  DBImpl* db_impl = static_cast_with_check<DBImpl>(db->GetRootDB());
+  // At the beginning there should be no log containing prepare data
+  ASSERT_EQ(db_impl->TEST_FindMinLogContainingOutstandingPrep(), 0);
+  for (size_t i = 0; i < 100; i++) {
+    Transaction* txn = db->BeginTransaction(write_options, txn_options);
+    ASSERT_OK(txn->SetName("xid" + std::to_string(i)));
+    ASSERT_OK(txn->Put(Slice("foo" + std::to_string(i)), Slice("bar")));
+    ASSERT_OK(txn->Prepare());
+    ASSERT_GT(db_impl->TEST_FindMinLogContainingOutstandingPrep(), 0);
+    if (rnd.OneIn(5)) {
+      txns.push_back(txn);
+    } else {
+      ASSERT_OK(txn->Commit());
+      delete txn;
+    }
+    ASSERT_OK(db_impl->TEST_FlushMemTable(true));
+  }
+  for (auto txn : txns) {
+    ASSERT_OK(txn->Commit());
+    delete txn;
+  }
+  // At the end there should be no log left containing prepare data
+  ASSERT_EQ(db_impl->TEST_FindMinLogContainingOutstandingPrep(), 0);
+  // Make sure that the underlying data structures are properly truncated and
+  // cause not leak
+  ASSERT_EQ(db_impl->TEST_PreparedSectionCompletedSize(), 0);
+  ASSERT_EQ(db_impl->TEST_LogsWithPrepSize(), 0);
+}
+
+TEST_P(TransactionTest, SimpleTwoPhaseTransactionTest) {
+  for (bool cwb4recovery : {true, false}) {
+    ASSERT_OK(ReOpen());
+    WriteOptions write_options;
+    ReadOptions read_options;
+
+    TransactionOptions txn_options;
+    txn_options.use_only_the_last_commit_time_batch_for_recovery = cwb4recovery;
+
+    std::string value;
+    Status s;
+
+    DBImpl* db_impl = static_cast_with_check<DBImpl>(db->GetRootDB());
+
+    Transaction* txn = db->BeginTransaction(write_options, txn_options);
+    s = txn->SetName("xid");
+    ASSERT_OK(s);
+
+    ASSERT_EQ(db->GetTransactionByName("xid"), txn);
+
+    // transaction put
+    s = txn->Put(Slice("foo"), Slice("bar"));
+    ASSERT_OK(s);
+    ASSERT_EQ(1, txn->GetNumPuts());
+
+    // regular db put
+    s = db->Put(write_options, Slice("foo2"), Slice("bar2"));
+    ASSERT_OK(s);
+    ASSERT_EQ(1, txn->GetNumPuts());
+
+    // regular db read
+    ASSERT_OK(db->Get(read_options, "foo2", &value));
+    ASSERT_EQ(value, "bar2");
+
+    // commit time put
+    if (cwb4recovery) {
+      ASSERT_OK(
+          txn->GetCommitTimeWriteBatch()->Put(Slice("gtid"), Slice("dogs")));
+      ASSERT_OK(
+          txn->GetCommitTimeWriteBatch()->Put(Slice("gtid2"), Slice("cats")));
+    }
+
+    // nothing has been prepped yet
+    ASSERT_EQ(db_impl->TEST_FindMinLogContainingOutstandingPrep(), 0);
+
+    s = txn->Prepare();
+    ASSERT_OK(s);
+
+    // data not im mem yet
+    s = db->Get(read_options, Slice("foo"), &value);
+    ASSERT_TRUE(s.IsNotFound());
+    s = db->Get(read_options, Slice("gtid"), &value);
+    ASSERT_TRUE(s.IsNotFound());
+
+    // find trans in list of prepared transactions
+    std::vector<Transaction*> prepared_trans;
+    db->GetAllPreparedTransactions(&prepared_trans);
+    ASSERT_EQ(prepared_trans.size(), 1);
+    ASSERT_EQ(prepared_trans.front()->GetName(), "xid");
+
+    auto log_containing_prep =
+        db_impl->TEST_FindMinLogContainingOutstandingPrep();
+    ASSERT_GT(log_containing_prep, 0);
+
+    // make commit
+    s = txn->Commit();
+    ASSERT_OK(s);
+
+    // value is now available
+    s = db->Get(read_options, "foo", &value);
+    ASSERT_OK(s);
+    ASSERT_EQ(value, "bar");
+
+    // we already committed
+    s = txn->Commit();
+    ASSERT_EQ(s, Status::InvalidArgument());
+
+    // no longer is prepared results
+    db->GetAllPreparedTransactions(&prepared_trans);
+    ASSERT_EQ(prepared_trans.size(), 0);
+    ASSERT_EQ(db->GetTransactionByName("xid"), nullptr);
+
+    // heap should not care about prepared section anymore
+    ASSERT_EQ(db_impl->TEST_FindMinLogContainingOutstandingPrep(), 0);
+
+    switch (txn_db_options.write_policy) {
+      case WRITE_COMMITTED:
+        // but now our memtable should be referencing the prep section
+        ASSERT_GE(log_containing_prep, db_impl->MinLogNumberToKeep());
+        ASSERT_EQ(log_containing_prep,
+                  db_impl->TEST_FindMinPrepLogReferencedByMemTable());
+        break;
+      case WRITE_PREPARED:
+      case WRITE_UNPREPARED:
+        // In these modes memtable do not ref the prep sections
+        ASSERT_EQ(0, db_impl->TEST_FindMinPrepLogReferencedByMemTable());
+        break;
+      default:
+        assert(false);
+    }
+
+    ASSERT_OK(db_impl->TEST_FlushMemTable(true));
+    // After flush the recoverable state must be visible
+    if (cwb4recovery) {
+      s = db->Get(read_options, "gtid", &value);
+      ASSERT_OK(s);
+      ASSERT_EQ(value, "dogs");
+
+      s = db->Get(read_options, "gtid2", &value);
+      ASSERT_OK(s);
+      ASSERT_EQ(value, "cats");
+    }
+
+    // after memtable flush we can now relese the log
+    ASSERT_GT(db_impl->MinLogNumberToKeep(), log_containing_prep);
+    ASSERT_EQ(0, db_impl->TEST_FindMinPrepLogReferencedByMemTable());
+
+    delete txn;
+
+    if (cwb4recovery) {
+      // kill and reopen to trigger recovery
+      s = ReOpenNoDelete();
+      ASSERT_OK(s);
+      assert(db != nullptr);
+      s = db->Get(read_options, "gtid", &value);
+      ASSERT_OK(s);
+      ASSERT_EQ(value, "dogs");
+
+      s = db->Get(read_options, "gtid2", &value);
+      ASSERT_OK(s);
+      ASSERT_EQ(value, "cats");
+    }
+  }
+}
+
+TEST_P(TransactionTest, TwoPhaseNameTest) {
+  Status s;
+
+  WriteOptions write_options;
+  TransactionOptions txn_options;
+  Transaction* txn1 = db->BeginTransaction(write_options, txn_options);
+  Transaction* txn2 = db->BeginTransaction(write_options, txn_options);
+  Transaction* txn3 = db->BeginTransaction(write_options, txn_options);
+  ASSERT_TRUE(txn3);
+  delete txn3;
+
+  // cant prepare txn without name
+  s = txn1->Prepare();
+  ASSERT_EQ(s, Status::InvalidArgument());
+
+  // name too short
+  s = txn1->SetName("");
+  ASSERT_EQ(s, Status::InvalidArgument());
+
+  // name too long
+  s = txn1->SetName(std::string(513, 'x'));
+  ASSERT_EQ(s, Status::InvalidArgument());
+
+  // valid set name
+  s = txn1->SetName("name1");
+  ASSERT_OK(s);
+
+  // cant have duplicate name
+  s = txn2->SetName("name1");
+  ASSERT_EQ(s, Status::InvalidArgument());
+
+  // shouldn't be able to prepare
+  s = txn2->Prepare();
+  ASSERT_EQ(s, Status::InvalidArgument());
+
+  // valid name set
+  s = txn2->SetName("name2");
+  ASSERT_OK(s);
+
+  // cant reset name
+  s = txn2->SetName("name3");
+  ASSERT_EQ(s, Status::InvalidArgument());
+
+  ASSERT_EQ(txn1->GetName(), "name1");
+  ASSERT_EQ(txn2->GetName(), "name2");
+
+  s = txn1->Prepare();
+  ASSERT_OK(s);
+
+  // can't rename after prepare
+  s = txn1->SetName("name4");
+  ASSERT_EQ(s, Status::InvalidArgument());
+
+  ASSERT_OK(txn1->Rollback());
+  ASSERT_OK(txn2->Rollback());
+  delete txn1;
+  delete txn2;
+}
+
+TEST_P(TransactionTest, TwoPhaseEmptyWriteTest) {
+  for (bool cwb4recovery : {true, false}) {
+    for (bool test_with_empty_wal : {true, false}) {
+      if (!cwb4recovery && test_with_empty_wal) {
+        continue;
+      }
+      ASSERT_OK(ReOpen());
+      Status s;
+      std::string value;
+
+      WriteOptions write_options;
+      ReadOptions read_options;
+      TransactionOptions txn_options;
+      txn_options.use_only_the_last_commit_time_batch_for_recovery =
+          cwb4recovery;
+      Transaction* txn1 = db->BeginTransaction(write_options, txn_options);
+      ASSERT_TRUE(txn1);
+      Transaction* txn2 = db->BeginTransaction(write_options, txn_options);
+      ASSERT_TRUE(txn2);
+
+      s = txn1->SetName("joe");
+      ASSERT_OK(s);
+
+      s = txn2->SetName("bob");
+      ASSERT_OK(s);
+
+      s = txn1->Prepare();
+      ASSERT_OK(s);
+
+      s = txn1->Commit();
+      ASSERT_OK(s);
+
+      delete txn1;
+
+      if (cwb4recovery) {
+        ASSERT_OK(
+            txn2->GetCommitTimeWriteBatch()->Put(Slice("foo"), Slice("bar")));
+      }
+
+      s = txn2->Prepare();
+      ASSERT_OK(s);
+
+      s = txn2->Commit();
+      ASSERT_OK(s);
+
+      delete txn2;
+      if (cwb4recovery) {
+        if (test_with_empty_wal) {
+          DBImpl* db_impl = static_cast_with_check<DBImpl>(db->GetRootDB());
+          ASSERT_OK(db_impl->TEST_FlushMemTable(true));
+          // After flush the state must be visible
+          s = db->Get(read_options, "foo", &value);
+          ASSERT_OK(s);
+          ASSERT_EQ(value, "bar");
+        }
+        ASSERT_OK(db->FlushWAL(true));
+        // kill and reopen to trigger recovery
+        s = ReOpenNoDelete();
+        ASSERT_OK(s);
+        assert(db != nullptr);
+        s = db->Get(read_options, "foo", &value);
+        ASSERT_OK(s);
+        ASSERT_EQ(value, "bar");
+      }
+    }
+  }
+}
+
+#if !defined(ROCKSDB_VALGRIND_RUN) || defined(ROCKSDB_FULL_VALGRIND_RUN)
+TEST_P(TransactionStressTest, TwoPhaseExpirationTest) {
+  Status s;
+
+  WriteOptions write_options;
+  TransactionOptions txn_options;
+  txn_options.expiration = 500;  // 500ms
+  Transaction* txn1 = db->BeginTransaction(write_options, txn_options);
+  Transaction* txn2 = db->BeginTransaction(write_options, txn_options);
+  ASSERT_TRUE(txn1);
+  ASSERT_TRUE(txn1);
+
+  s = txn1->SetName("joe");
+  ASSERT_OK(s);
+  s = txn2->SetName("bob");
+  ASSERT_OK(s);
+
+  s = txn1->Prepare();
+  ASSERT_OK(s);
+
+  /* sleep override */
+  std::this_thread::sleep_for(std::chrono::milliseconds(1000));
+
+  s = txn1->Commit();
+  ASSERT_OK(s);
+
+  s = txn2->Prepare();
+  ASSERT_EQ(s, Status::Expired());
+
+  delete txn1;
+  delete txn2;
+}
+
+TEST_P(TransactionTest, TwoPhaseRollbackTest) {
+  WriteOptions write_options;
+  ReadOptions read_options;
+
+  TransactionOptions txn_options;
+
+  std::string value;
+  Status s;
+
+  DBImpl* db_impl = static_cast_with_check<DBImpl>(db->GetRootDB());
+  Transaction* txn = db->BeginTransaction(write_options, txn_options);
+  s = txn->SetName("xid");
+  ASSERT_OK(s);
+
+  // transaction put
+  s = txn->Put(Slice("tfoo"), Slice("tbar"));
+  ASSERT_OK(s);
+
+  // value is readable form txn
+  s = txn->Get(read_options, Slice("tfoo"), &value);
+  ASSERT_OK(s);
+  ASSERT_EQ(value, "tbar");
+
+  // issue rollback
+  s = txn->Rollback();
+  ASSERT_OK(s);
+
+  // value is nolonger readable
+  s = txn->Get(read_options, Slice("tfoo"), &value);
+  ASSERT_TRUE(s.IsNotFound());
+  ASSERT_EQ(txn->GetNumPuts(), 0);
+
+  // put new txn values
+  s = txn->Put(Slice("tfoo2"), Slice("tbar2"));
+  ASSERT_OK(s);
+
+  // new value is readable from txn
+  s = txn->Get(read_options, Slice("tfoo2"), &value);
+  ASSERT_OK(s);
+  ASSERT_EQ(value, "tbar2");
+
+  s = txn->Prepare();
+  ASSERT_OK(s);
+
+  // flush to next wal
+  s = db->Put(write_options, Slice("foo"), Slice("bar"));
+  ASSERT_OK(s);
+  ASSERT_OK(db_impl->TEST_FlushMemTable(true));
+
+  // issue rollback (marker written to WAL)
+  s = txn->Rollback();
+  ASSERT_OK(s);
+
+  // value is nolonger readable
+  s = txn->Get(read_options, Slice("tfoo2"), &value);
+  ASSERT_TRUE(s.IsNotFound());
+  ASSERT_EQ(txn->GetNumPuts(), 0);
+
+  // make commit
+  s = txn->Commit();
+  ASSERT_EQ(s, Status::InvalidArgument());
+
+  // try rollback again
+  s = txn->Rollback();
+  ASSERT_EQ(s, Status::InvalidArgument());
+
+  delete txn;
+}
+
+TEST_P(TransactionTest, PersistentTwoPhaseTransactionTest) {
+  WriteOptions write_options;
+  write_options.sync = true;
+  write_options.disableWAL = false;
+  ReadOptions read_options;
+
+  TransactionOptions txn_options;
+
+  std::string value;
+  Status s;
+
+  DBImpl* db_impl = static_cast_with_check<DBImpl>(db->GetRootDB());
+
+  Transaction* txn = db->BeginTransaction(write_options, txn_options);
+  s = txn->SetName("xid");
+  ASSERT_OK(s);
+
+  ASSERT_EQ(db->GetTransactionByName("xid"), txn);
+
+  // transaction put
+  s = txn->Put(Slice("foo"), Slice("bar"));
+  ASSERT_OK(s);
+  ASSERT_EQ(1, txn->GetNumPuts());
+
+  // txn read
+  s = txn->Get(read_options, "foo", &value);
+  ASSERT_OK(s);
+  ASSERT_EQ(value, "bar");
+
+  // regular db put
+  s = db->Put(write_options, Slice("foo2"), Slice("bar2"));
+  ASSERT_OK(s);
+  ASSERT_EQ(1, txn->GetNumPuts());
+
+  ASSERT_OK(db_impl->TEST_FlushMemTable(true));
+
+  // regular db read
+  db->Get(read_options, "foo2", &value);
+  ASSERT_EQ(value, "bar2");
+
+  // nothing has been prepped yet
+  ASSERT_EQ(db_impl->TEST_FindMinLogContainingOutstandingPrep(), 0);
+
+  // prepare
+  s = txn->Prepare();
+  ASSERT_OK(s);
+
+  // still not available to db
+  s = db->Get(read_options, Slice("foo"), &value);
+  ASSERT_TRUE(s.IsNotFound());
+
+  ASSERT_OK(db->FlushWAL(false));
+  delete txn;
+  // kill and reopen
+  reinterpret_cast<PessimisticTransactionDB*>(db)->TEST_Crash();
+  s = ReOpenNoDelete();
+  ASSERT_OK(s);
+  assert(db != nullptr);
+  db_impl = static_cast_with_check<DBImpl>(db->GetRootDB());
+
+  // find trans in list of prepared transactions
+  std::vector<Transaction*> prepared_trans;
+  db->GetAllPreparedTransactions(&prepared_trans);
+  ASSERT_EQ(prepared_trans.size(), 1);
+
+  txn = prepared_trans.front();
+  ASSERT_TRUE(txn);
+  ASSERT_EQ(txn->GetName(), "xid");
+  ASSERT_EQ(db->GetTransactionByName("xid"), txn);
+
+  // log has been marked
+  auto log_containing_prep =
+      db_impl->TEST_FindMinLogContainingOutstandingPrep();
+  ASSERT_GT(log_containing_prep, 0);
+
+  // value is readable from txn
+  s = txn->Get(read_options, "foo", &value);
+  ASSERT_OK(s);
+  ASSERT_EQ(value, "bar");
+
+  // make commit
+  s = txn->Commit();
+  ASSERT_OK(s);
+
+  // value is now available
+  db->Get(read_options, "foo", &value);
+  ASSERT_EQ(value, "bar");
+
+  // we already committed
+  s = txn->Commit();
+  ASSERT_EQ(s, Status::InvalidArgument());
+
+  // no longer is prepared results
+  prepared_trans.clear();
+  db->GetAllPreparedTransactions(&prepared_trans);
+  ASSERT_EQ(prepared_trans.size(), 0);
+
+  // transaction should no longer be visible
+  ASSERT_EQ(db->GetTransactionByName("xid"), nullptr);
+
+  // heap should not care about prepared section anymore
+  ASSERT_EQ(db_impl->TEST_FindMinLogContainingOutstandingPrep(), 0);
+
+  switch (txn_db_options.write_policy) {
+    case WRITE_COMMITTED:
+      // but now our memtable should be referencing the prep section
+      ASSERT_EQ(log_containing_prep,
+                db_impl->TEST_FindMinPrepLogReferencedByMemTable());
+      ASSERT_GE(log_containing_prep, db_impl->MinLogNumberToKeep());
+
+      break;
+    case WRITE_PREPARED:
+    case WRITE_UNPREPARED:
+      // In these modes memtable do not ref the prep sections
+      ASSERT_EQ(0, db_impl->TEST_FindMinPrepLogReferencedByMemTable());
+      break;
+    default:
+      assert(false);
+  }
+
+  // Add a dummy record to memtable before a flush. Otherwise, the
+  // memtable will be empty and flush will be skipped.
+  s = db->Put(write_options, Slice("foo3"), Slice("bar3"));
+  ASSERT_OK(s);
+
+  ASSERT_OK(db_impl->TEST_FlushMemTable(true));
+
+  // after memtable flush we can now release the log
+  ASSERT_GT(db_impl->MinLogNumberToKeep(), log_containing_prep);
+  ASSERT_EQ(0, db_impl->TEST_FindMinPrepLogReferencedByMemTable());
+
+  delete txn;
+
+  // deleting transaction should unregister transaction
+  ASSERT_EQ(db->GetTransactionByName("xid"), nullptr);
+}
+#endif  // !defined(ROCKSDB_VALGRIND_RUN) || defined(ROCKSDB_FULL_VALGRIND_RUN)
+
+// TODO this test needs to be updated with serial commits
+TEST_P(TransactionTest, DISABLED_TwoPhaseMultiThreadTest) {
+  // mix transaction writes and regular writes
+  const uint32_t NUM_TXN_THREADS = 50;
+  std::atomic<uint32_t> txn_thread_num(0);
+
+  std::function<void()> txn_write_thread = [&]() {
+    uint32_t id = txn_thread_num.fetch_add(1);
+
+    WriteOptions write_options;
+    write_options.sync = true;
+    write_options.disableWAL = false;
+    TransactionOptions txn_options;
+    txn_options.lock_timeout = 1000000;
+    if (id % 2 == 0) {
+      txn_options.expiration = 1000000;
+    }
+    TransactionName name("xid_" + std::string(1, 'A' + static_cast<char>(id)));
+    Transaction* txn = db->BeginTransaction(write_options, txn_options);
+    ASSERT_OK(txn->SetName(name));
+    for (int i = 0; i < 10; i++) {
+      std::string key(name + "_" + std::string(1, static_cast<char>('A' + i)));
+      ASSERT_OK(txn->Put(key, "val"));
+    }
+    ASSERT_OK(txn->Prepare());
+    ASSERT_OK(txn->Commit());
+    delete txn;
+  };
+
+  // assure that all thread are in the same write group
+  std::atomic<uint32_t> t_wait_on_prepare(0);
+  std::atomic<uint32_t> t_wait_on_commit(0);
+
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
+      "WriteThread::JoinBatchGroup:Wait", [&](void* arg) {
+        auto* writer = reinterpret_cast<WriteThread::Writer*>(arg);
+
+        if (writer->ShouldWriteToWAL()) {
+          t_wait_on_prepare.fetch_add(1);
+          // wait for friends
+          while (t_wait_on_prepare.load() < NUM_TXN_THREADS) {
+            env->SleepForMicroseconds(10);
+          }
+        } else if (writer->ShouldWriteToMemtable()) {
+          t_wait_on_commit.fetch_add(1);
+          // wait for friends
+          while (t_wait_on_commit.load() < NUM_TXN_THREADS) {
+            env->SleepForMicroseconds(10);
+          }
+        } else {
+          FAIL();
+        }
+      });
+
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->EnableProcessing();
+
+  // do all the writes
+  std::vector<port::Thread> threads;
+  for (uint32_t i = 0; i < NUM_TXN_THREADS; i++) {
+    threads.emplace_back(txn_write_thread);
+  }
+  for (auto& t : threads) {
+    t.join();
+  }
+
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->DisableProcessing();
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->ClearAllCallBacks();
+
+  ReadOptions read_options;
+  std::string value;
+  Status s;
+  for (uint32_t t = 0; t < NUM_TXN_THREADS; t++) {
+    TransactionName name("xid_" + std::string(1, 'A' + static_cast<char>(t)));
+    for (int i = 0; i < 10; i++) {
+      std::string key(name + "_" + std::string(1, static_cast<char>('A' + i)));
+      s = db->Get(read_options, key, &value);
+      ASSERT_OK(s);
+      ASSERT_EQ(value, "val");
+    }
+  }
+}
+
+#if !defined(ROCKSDB_VALGRIND_RUN) || defined(ROCKSDB_FULL_VALGRIND_RUN)
+TEST_P(TransactionStressTest, TwoPhaseLongPrepareTest) {
+  WriteOptions write_options;
+  write_options.sync = true;
+  write_options.disableWAL = false;
+  ReadOptions read_options;
+  TransactionOptions txn_options;
+
+  std::string value;
+  Status s;
+
+  Transaction* txn = db->BeginTransaction(write_options, txn_options);
+  s = txn->SetName("bob");
+  ASSERT_OK(s);
+
+  // transaction put
+  s = txn->Put(Slice("foo"), Slice("bar"));
+  ASSERT_OK(s);
+
+  // prepare
+  s = txn->Prepare();
+  ASSERT_OK(s);
+
+  delete txn;
+
+  for (int i = 0; i < 1000; i++) {
+    std::string key(i, 'k');
+    std::string val(1000, 'v');
+    assert(db != nullptr);
+    s = db->Put(write_options, key, val);
+    ASSERT_OK(s);
+
+    if (i % 29 == 0) {
+      // crash
+      env->SetFilesystemActive(false);
+      reinterpret_cast<PessimisticTransactionDB*>(db)->TEST_Crash();
+      ReOpenNoDelete();
+    } else if (i % 37 == 0) {
+      // close
+      ReOpenNoDelete();
+    }
+  }
+
+  // commit old txn
+  txn = db->GetTransactionByName("bob");
+  ASSERT_TRUE(txn);
+  s = txn->Commit();
+  ASSERT_OK(s);
+
+  // verify data txn data
+  s = db->Get(read_options, "foo", &value);
+  ASSERT_EQ(s, Status::OK());
+  ASSERT_EQ(value, "bar");
+
+  // verify non txn data
+  for (int i = 0; i < 1000; i++) {
+    std::string key(i, 'k');
+    std::string val(1000, 'v');
+    s = db->Get(read_options, key, &value);
+    ASSERT_EQ(s, Status::OK());
+    ASSERT_EQ(value, val);
+  }
+
+  delete txn;
+}
+
+TEST_P(TransactionTest, TwoPhaseSequenceTest) {
+  WriteOptions write_options;
+  write_options.sync = true;
+  write_options.disableWAL = false;
+  ReadOptions read_options;
+
+  TransactionOptions txn_options;
+
+  std::string value;
+  Status s;
+
+  Transaction* txn = db->BeginTransaction(write_options, txn_options);
+  s = txn->SetName("xid");
+  ASSERT_OK(s);
+
+  // transaction put
+  s = txn->Put(Slice("foo"), Slice("bar"));
+  ASSERT_OK(s);
+  s = txn->Put(Slice("foo2"), Slice("bar2"));
+  ASSERT_OK(s);
+  s = txn->Put(Slice("foo3"), Slice("bar3"));
+  ASSERT_OK(s);
+  s = txn->Put(Slice("foo4"), Slice("bar4"));
+  ASSERT_OK(s);
+
+  // prepare
+  s = txn->Prepare();
+  ASSERT_OK(s);
+
+  // make commit
+  s = txn->Commit();
+  ASSERT_OK(s);
+
+  delete txn;
+
+  // kill and reopen
+  env->SetFilesystemActive(false);
+  ReOpenNoDelete();
+  assert(db != nullptr);
+
+  // value is now available
+  s = db->Get(read_options, "foo4", &value);
+  ASSERT_EQ(s, Status::OK());
+  ASSERT_EQ(value, "bar4");
+}
+#endif  // !defined(ROCKSDB_VALGRIND_RUN) || defined(ROCKSDB_FULL_VALGRIND_RUN)
+
+TEST_P(TransactionTest, TwoPhaseDoubleRecoveryTest) {
+  WriteOptions write_options;
+  write_options.sync = true;
+  write_options.disableWAL = false;
+  ReadOptions read_options;
+
+  TransactionOptions txn_options;
+
+  std::string value;
+  Status s;
+
+  Transaction* txn = db->BeginTransaction(write_options, txn_options);
+  s = txn->SetName("a");
+  ASSERT_OK(s);
+
+  // transaction put
+  s = txn->Put(Slice("foo"), Slice("bar"));
+  ASSERT_OK(s);
+
+  // prepare
+  s = txn->Prepare();
+  ASSERT_OK(s);
+
+  delete txn;
+
+  // kill and reopen
+  env->SetFilesystemActive(false);
+  reinterpret_cast<PessimisticTransactionDB*>(db)->TEST_Crash();
+  ReOpenNoDelete();
+
+  // commit old txn
+  assert(db != nullptr);  // Make clang analyze happy.
+  txn = db->GetTransactionByName("a");
+  assert(txn != nullptr);
+  s = txn->Commit();
+  ASSERT_OK(s);
+
+  s = db->Get(read_options, "foo", &value);
+  ASSERT_EQ(s, Status::OK());
+  ASSERT_EQ(value, "bar");
+
+  delete txn;
+
+  txn = db->BeginTransaction(write_options, txn_options);
+  s = txn->SetName("b");
+  ASSERT_OK(s);
+
+  s = txn->Put(Slice("foo2"), Slice("bar2"));
+  ASSERT_OK(s);
+
+  s = txn->Prepare();
+  ASSERT_OK(s);
+
+  s = txn->Commit();
+  ASSERT_OK(s);
+
+  delete txn;
+
+  // kill and reopen
+  env->SetFilesystemActive(false);
+  ASSERT_OK(ReOpenNoDelete());
+  assert(db != nullptr);
+
+  // value is now available
+  s = db->Get(read_options, "foo", &value);
+  ASSERT_EQ(s, Status::OK());
+  ASSERT_EQ(value, "bar");
+
+  s = db->Get(read_options, "foo2", &value);
+  ASSERT_EQ(s, Status::OK());
+  ASSERT_EQ(value, "bar2");
+}
+
+TEST_P(TransactionTest, TwoPhaseLogRollingTest) {
+  DBImpl* db_impl = static_cast_with_check<DBImpl>(db->GetRootDB());
+
+  Status s;
+  std::string v;
+  ColumnFamilyHandle *cfa, *cfb;
+
+  // Create 2 new column families
+  ColumnFamilyOptions cf_options;
+  s = db->CreateColumnFamily(cf_options, "CFA", &cfa);
+  ASSERT_OK(s);
+  s = db->CreateColumnFamily(cf_options, "CFB", &cfb);
+  ASSERT_OK(s);
+
+  WriteOptions wopts;
+  wopts.disableWAL = false;
+  wopts.sync = true;
+
+  TransactionOptions topts1;
+  Transaction* txn1 = db->BeginTransaction(wopts, topts1);
+  s = txn1->SetName("xid1");
+  ASSERT_OK(s);
+
+  TransactionOptions topts2;
+  Transaction* txn2 = db->BeginTransaction(wopts, topts2);
+  s = txn2->SetName("xid2");
+  ASSERT_OK(s);
+
+  // transaction put in two column families
+  s = txn1->Put(cfa, "ka1", "va1");
+  ASSERT_OK(s);
+
+  // transaction put in two column families
+  s = txn2->Put(cfa, "ka2", "va2");
+  ASSERT_OK(s);
+  s = txn2->Put(cfb, "kb2", "vb2");
+  ASSERT_OK(s);
+
+  // write prep section to wal
+  s = txn1->Prepare();
+  ASSERT_OK(s);
+
+  // our log should be in the heap
+  ASSERT_EQ(db_impl->TEST_FindMinLogContainingOutstandingPrep(),
+            txn1->GetLogNumber());
+  ASSERT_EQ(db_impl->TEST_LogfileNumber(), txn1->GetLastLogNumber());
+
+  // flush default cf to crate new log
+  s = db->Put(wopts, "foo", "bar");
+  ASSERT_OK(s);
+  s = db_impl->TEST_FlushMemTable(true);
+  ASSERT_OK(s);
+
+  // make sure we are on a new log
+  ASSERT_GT(db_impl->TEST_LogfileNumber(), txn1->GetLastLogNumber());
+
+  // put txn2 prep section in this log
+  s = txn2->Prepare();
+  ASSERT_OK(s);
+  ASSERT_EQ(db_impl->TEST_LogfileNumber(), txn2->GetLastLogNumber());
+
+  // heap should still see first log
+  ASSERT_EQ(db_impl->TEST_FindMinLogContainingOutstandingPrep(),
+            txn1->GetLogNumber());
+
+  // commit txn1
+  s = txn1->Commit();
+  ASSERT_OK(s);
+
+  // heap should now show txn2s log
+  ASSERT_EQ(db_impl->TEST_FindMinLogContainingOutstandingPrep(),
+            txn2->GetLogNumber());
+
+  switch (txn_db_options.write_policy) {
+    case WRITE_COMMITTED:
+      // we should see txn1s log refernced by the memtables
+      ASSERT_EQ(txn1->GetLogNumber(),
+                db_impl->TEST_FindMinPrepLogReferencedByMemTable());
+      break;
+    case WRITE_PREPARED:
+    case WRITE_UNPREPARED:
+      // In these modes memtable do not ref the prep sections
+      ASSERT_EQ(0, db_impl->TEST_FindMinPrepLogReferencedByMemTable());
+      break;
+    default:
+      assert(false);
+  }
+
+  // flush default cf to crate new log
+  s = db->Put(wopts, "foo", "bar2");
+  ASSERT_OK(s);
+  s = db_impl->TEST_FlushMemTable(true);
+  ASSERT_OK(s);
+
+  // make sure we are on a new log
+  ASSERT_GT(db_impl->TEST_LogfileNumber(), txn2->GetLastLogNumber());
+
+  // commit txn2
+  s = txn2->Commit();
+  ASSERT_OK(s);
+
+  // heap should not show any logs
+  ASSERT_EQ(db_impl->TEST_FindMinLogContainingOutstandingPrep(), 0);
+
+  switch (txn_db_options.write_policy) {
+    case WRITE_COMMITTED:
+      // should show the first txn log
+      ASSERT_EQ(txn1->GetLogNumber(),
+                db_impl->TEST_FindMinPrepLogReferencedByMemTable());
+      break;
+    case WRITE_PREPARED:
+    case WRITE_UNPREPARED:
+      // In these modes memtable do not ref the prep sections
+      ASSERT_EQ(0, db_impl->TEST_FindMinPrepLogReferencedByMemTable());
+      break;
+    default:
+      assert(false);
+  }
+
+  // flush only cfa memtable
+  s = db_impl->TEST_FlushMemTable(true, false, cfa);
+  ASSERT_OK(s);
+
+  switch (txn_db_options.write_policy) {
+    case WRITE_COMMITTED:
+      // should show the first txn log
+      ASSERT_EQ(txn2->GetLogNumber(),
+                db_impl->TEST_FindMinPrepLogReferencedByMemTable());
+      break;
+    case WRITE_PREPARED:
+    case WRITE_UNPREPARED:
+      // In these modes memtable do not ref the prep sections
+      ASSERT_EQ(0, db_impl->TEST_FindMinPrepLogReferencedByMemTable());
+      break;
+    default:
+      assert(false);
+  }
+
+  // flush only cfb memtable
+  s = db_impl->TEST_FlushMemTable(true, false, cfb);
+  ASSERT_OK(s);
+
+  // should show not dependency on logs
+  ASSERT_EQ(db_impl->TEST_FindMinPrepLogReferencedByMemTable(), 0);
+  ASSERT_EQ(db_impl->TEST_FindMinLogContainingOutstandingPrep(), 0);
+
+  delete txn1;
+  delete txn2;
+  delete cfa;
+  delete cfb;
+}
+
+TEST_P(TransactionTest, TwoPhaseLogRollingTest2) {
+  DBImpl* db_impl = static_cast_with_check<DBImpl>(db->GetRootDB());
+
+  Status s;
+  ColumnFamilyHandle *cfa, *cfb;
+
+  ColumnFamilyOptions cf_options;
+  s = db->CreateColumnFamily(cf_options, "CFA", &cfa);
+  ASSERT_OK(s);
+  s = db->CreateColumnFamily(cf_options, "CFB", &cfb);
+  ASSERT_OK(s);
+
+  WriteOptions wopts;
+  wopts.disableWAL = false;
+  wopts.sync = true;
+
+  auto cfh_a = static_cast_with_check<ColumnFamilyHandleImpl>(cfa);
+  auto cfh_b = static_cast_with_check<ColumnFamilyHandleImpl>(cfb);
+
+  TransactionOptions topts1;
+  Transaction* txn1 = db->BeginTransaction(wopts, topts1);
+  s = txn1->SetName("xid1");
+  ASSERT_OK(s);
+  s = txn1->Put(cfa, "boys", "girls1");
+  ASSERT_OK(s);
+
+  Transaction* txn2 = db->BeginTransaction(wopts, topts1);
+  s = txn2->SetName("xid2");
+  ASSERT_OK(s);
+  s = txn2->Put(cfb, "up", "down1");
+  ASSERT_OK(s);
+
+  // prepre transaction in LOG A
+  s = txn1->Prepare();
+  ASSERT_OK(s);
+
+  // prepre transaction in LOG A
+  s = txn2->Prepare();
+  ASSERT_OK(s);
+
+  // regular put so that mem table can actually be flushed for log rolling
+  s = db->Put(wopts, "cats", "dogs1");
+  ASSERT_OK(s);
+
+  auto prepare_log_no = txn1->GetLastLogNumber();
+
+  // roll to LOG B
+  s = db_impl->TEST_FlushMemTable(true);
+  ASSERT_OK(s);
+
+  // now we pause background work so that
+  // imm()s are not flushed before we can check their status
+  s = db_impl->PauseBackgroundWork();
+  ASSERT_OK(s);
+
+  ASSERT_GT(db_impl->TEST_LogfileNumber(), prepare_log_no);
+  switch (txn_db_options.write_policy) {
+    case WRITE_COMMITTED:
+      // This cf is empty and should ref the latest log
+      ASSERT_GT(cfh_a->cfd()->GetLogNumber(), prepare_log_no);
+      ASSERT_EQ(cfh_a->cfd()->GetLogNumber(), db_impl->TEST_LogfileNumber());
+      break;
+    case WRITE_PREPARED:
+    case WRITE_UNPREPARED:
+      // This cf is not flushed yet and should ref the log that has its data
+      ASSERT_EQ(cfh_a->cfd()->GetLogNumber(), prepare_log_no);
+      break;
+    default:
+      assert(false);
+  }
+  ASSERT_EQ(db_impl->TEST_FindMinLogContainingOutstandingPrep(),
+            txn1->GetLogNumber());
+  ASSERT_EQ(db_impl->TEST_FindMinPrepLogReferencedByMemTable(), 0);
+
+  // commit in LOG B
+  s = txn1->Commit();
+  ASSERT_OK(s);
+
+  switch (txn_db_options.write_policy) {
+    case WRITE_COMMITTED:
+      ASSERT_EQ(db_impl->TEST_FindMinPrepLogReferencedByMemTable(),
+                prepare_log_no);
+      break;
+    case WRITE_PREPARED:
+    case WRITE_UNPREPARED:
+      // In these modes memtable do not ref the prep sections
+      ASSERT_EQ(db_impl->TEST_FindMinPrepLogReferencedByMemTable(), 0);
+      break;
+    default:
+      assert(false);
+  }
+
+  ASSERT_TRUE(!db_impl->TEST_UnableToReleaseOldestLog());
+
+  // request a flush for all column families such that the earliest
+  // alive log file can be killed
+  ASSERT_OK(db_impl->TEST_SwitchWAL());
+  // log cannot be flushed because txn2 has not been commited
+  ASSERT_TRUE(!db_impl->TEST_IsLogGettingFlushed());
+  ASSERT_TRUE(db_impl->TEST_UnableToReleaseOldestLog());
+
+  // assert that cfa has a flush requested
+  ASSERT_TRUE(cfh_a->cfd()->imm()->HasFlushRequested());
+
+  switch (txn_db_options.write_policy) {
+    case WRITE_COMMITTED:
+      // cfb should not be flushed becuse it has no data from LOG A
+      ASSERT_TRUE(!cfh_b->cfd()->imm()->HasFlushRequested());
+      break;
+    case WRITE_PREPARED:
+    case WRITE_UNPREPARED:
+      // cfb should be flushed becuse it has prepared data from LOG A
+      ASSERT_TRUE(cfh_b->cfd()->imm()->HasFlushRequested());
+      break;
+    default:
+      assert(false);
+  }
+
+  // cfb now has data from LOG A
+  s = txn2->Commit();
+  ASSERT_OK(s);
+
+  ASSERT_OK(db_impl->TEST_SwitchWAL());
+  ASSERT_TRUE(!db_impl->TEST_UnableToReleaseOldestLog());
+
+  // we should see that cfb now has a flush requested
+  ASSERT_TRUE(cfh_b->cfd()->imm()->HasFlushRequested());
+
+  // all data in LOG A resides in a memtable that has been
+  // requested for a flush
+  ASSERT_TRUE(db_impl->TEST_IsLogGettingFlushed());
+
+  delete txn1;
+  delete txn2;
+  delete cfa;
+  delete cfb;
+}
+/*
+ * 1) use prepare to keep first log around to determine starting sequence
+ * during recovery.
+ * 2) insert many values, skipping wal, to increase seqid.
+ * 3) insert final value into wal
+ * 4) recover and see that final value was properly recovered - not
+ * hidden behind improperly summed sequence ids
+ */
+TEST_P(TransactionTest, TwoPhaseOutOfOrderDelete) {
+  DBImpl* db_impl = static_cast_with_check<DBImpl>(db->GetRootDB());
+  WriteOptions wal_on, wal_off;
+  wal_on.sync = true;
+  wal_on.disableWAL = false;
+  wal_off.disableWAL = true;
+  ReadOptions read_options;
+  TransactionOptions txn_options;
+
+  std::string value;
+  Status s;
+
+  Transaction* txn1 = db->BeginTransaction(wal_on, txn_options);
+
+  s = txn1->SetName("1");
+  ASSERT_OK(s);
+
+  s = db->Put(wal_on, "first", "first");
+  ASSERT_OK(s);
+
+  s = txn1->Put(Slice("dummy"), Slice("dummy"));
+  ASSERT_OK(s);
+  s = txn1->Prepare();
+  ASSERT_OK(s);
+
+  s = db->Put(wal_off, "cats", "dogs1");
+  ASSERT_OK(s);
+  s = db->Put(wal_off, "cats", "dogs2");
+  ASSERT_OK(s);
+  s = db->Put(wal_off, "cats", "dogs3");
+  ASSERT_OK(s);
+
+  s = db_impl->TEST_FlushMemTable(true);
+  ASSERT_OK(s);
+
+  s = db->Put(wal_on, "cats", "dogs4");
+  ASSERT_OK(s);
+
+  ASSERT_OK(db->FlushWAL(false));
+
+  // kill and reopen
+  env->SetFilesystemActive(false);
+  reinterpret_cast<PessimisticTransactionDB*>(db)->TEST_Crash();
+  ASSERT_OK(ReOpenNoDelete());
+  assert(db != nullptr);
+
+  s = db->Get(read_options, "first", &value);
+  ASSERT_OK(s);
+  ASSERT_EQ(value, "first");
+
+  s = db->Get(read_options, "cats", &value);
+  ASSERT_OK(s);
+  ASSERT_EQ(value, "dogs4");
+}
+
+TEST_P(TransactionTest, FirstWriteTest) {
+  WriteOptions write_options;
+
+  // Test conflict checking against the very first write to a db.
+  // The transaction's snapshot will have seq 1 and the following write
+  // will have sequence 1.
+  Status s = db->Put(write_options, "A", "a");
+
+  Transaction* txn = db->BeginTransaction(write_options);
+  txn->SetSnapshot();
+
+  ASSERT_OK(s);
+
+  s = txn->Put("A", "b");
+  ASSERT_OK(s);
+
+  delete txn;
+}
+
+TEST_P(TransactionTest, FirstWriteTest2) {
+  WriteOptions write_options;
+
+  Transaction* txn = db->BeginTransaction(write_options);
+  txn->SetSnapshot();
+
+  // Test conflict checking against the very first write to a db.
+  // The transaction's snapshot is a seq 0 while the following write
+  // will have sequence 1.
+  Status s = db->Put(write_options, "A", "a");
+  ASSERT_OK(s);
+
+  s = txn->Put("A", "b");
+  ASSERT_TRUE(s.IsBusy());
+
+  delete txn;
+}
+
+TEST_P(TransactionTest, WriteOptionsTest) {
+  WriteOptions write_options;
+  write_options.sync = true;
+  write_options.disableWAL = true;
+
+  Transaction* txn = db->BeginTransaction(write_options);
+  ASSERT_TRUE(txn);
+
+  ASSERT_TRUE(txn->GetWriteOptions()->sync);
+
+  write_options.sync = false;
+  txn->SetWriteOptions(write_options);
+  ASSERT_FALSE(txn->GetWriteOptions()->sync);
+  ASSERT_TRUE(txn->GetWriteOptions()->disableWAL);
+
+  delete txn;
+}
+
+TEST_P(TransactionTest, WriteConflictTest) {
+  WriteOptions write_options;
+  ReadOptions read_options;
+  std::string value;
+  Status s;
+
+  ASSERT_OK(db->Put(write_options, "foo", "A"));
+  ASSERT_OK(db->Put(write_options, "foo2", "B"));
+
+  Transaction* txn = db->BeginTransaction(write_options);
+  ASSERT_TRUE(txn);
+
+  s = txn->Put("foo", "A2");
+  ASSERT_OK(s);
+
+  s = txn->Put("foo2", "B2");
+  ASSERT_OK(s);
+
+  // This Put outside of a transaction will conflict with the previous write
+  s = db->Put(write_options, "foo", "xxx");
+  ASSERT_TRUE(s.IsTimedOut());
+
+  s = db->Get(read_options, "foo", &value);
+  ASSERT_EQ(value, "A");
+
+  s = txn->Commit();
+  ASSERT_OK(s);
+
+  db->Get(read_options, "foo", &value);
+  ASSERT_EQ(value, "A2");
+  db->Get(read_options, "foo2", &value);
+  ASSERT_EQ(value, "B2");
+
+  delete txn;
+}
+
+TEST_P(TransactionTest, WriteConflictTest2) {
+  WriteOptions write_options;
+  ReadOptions read_options;
+  TransactionOptions txn_options;
+  std::string value;
+  Status s;
+
+  ASSERT_OK(db->Put(write_options, "foo", "bar"));
+
+  txn_options.set_snapshot = true;
+  Transaction* txn = db->BeginTransaction(write_options, txn_options);
+  ASSERT_TRUE(txn);
+
+  // This Put outside of a transaction will conflict with a later write
+  s = db->Put(write_options, "foo", "barz");
+  ASSERT_OK(s);
+
+  s = txn->Put("foo2", "X");
+  ASSERT_OK(s);
+
+  s = txn->Put("foo",
+               "bar2");  // Conflicts with write done after snapshot taken
+  ASSERT_TRUE(s.IsBusy());
+
+  s = txn->Put("foo3", "Y");
+  ASSERT_OK(s);
+
+  s = db->Get(read_options, "foo", &value);
+  ASSERT_EQ(value, "barz");
+
+  ASSERT_EQ(2, txn->GetNumKeys());
+
+  s = txn->Commit();
+  ASSERT_OK(s);  // Txn should commit, but only write foo2 and foo3
+
+  // Verify that transaction wrote foo2 and foo3 but not foo
+  db->Get(read_options, "foo", &value);
+  ASSERT_EQ(value, "barz");
+
+  db->Get(read_options, "foo2", &value);
+  ASSERT_EQ(value, "X");
+
+  db->Get(read_options, "foo3", &value);
+  ASSERT_EQ(value, "Y");
+
+  delete txn;
+}
+
+TEST_P(TransactionTest, ReadConflictTest) {
+  WriteOptions write_options;
+  ReadOptions read_options, snapshot_read_options;
+  TransactionOptions txn_options;
+  std::string value;
+  Status s;
+
+  ASSERT_OK(db->Put(write_options, "foo", "bar"));
+  ASSERT_OK(db->Put(write_options, "foo2", "bar"));
+
+  txn_options.set_snapshot = true;
+  Transaction* txn = db->BeginTransaction(write_options, txn_options);
+  ASSERT_TRUE(txn);
+
+  txn->SetSnapshot();
+  snapshot_read_options.snapshot = txn->GetSnapshot();
+
+  ASSERT_OK(txn->GetForUpdate(snapshot_read_options, "foo", &value));
+  ASSERT_EQ(value, "bar");
+
+  // This Put outside of a transaction will conflict with the previous read
+  s = db->Put(write_options, "foo", "barz");
+  ASSERT_TRUE(s.IsTimedOut());
+
+  s = db->Get(read_options, "foo", &value);
+  ASSERT_EQ(value, "bar");
+
+  s = txn->Get(read_options, "foo", &value);
+  ASSERT_EQ(value, "bar");
+
+  s = txn->Commit();
+  ASSERT_OK(s);
+
+  delete txn;
+}
+
+TEST_P(TransactionTest, TxnOnlyTest) {
+  // Test to make sure transactions work when there are no other writes in an
+  // empty db.
+
+  WriteOptions write_options;
+  ReadOptions read_options;
+  std::string value;
+  Status s;
+
+  Transaction* txn = db->BeginTransaction(write_options);
+  ASSERT_TRUE(txn);
+
+  s = txn->Put("x", "y");
+  ASSERT_OK(s);
+
+  s = txn->Commit();
+  ASSERT_OK(s);
+
+  delete txn;
+}
+
+TEST_P(TransactionTest, FlushTest) {
+  WriteOptions write_options;
+  ReadOptions read_options, snapshot_read_options;
+  std::string value;
+  Status s;
+
+  ASSERT_OK(db->Put(write_options, Slice("foo"), Slice("bar")));
+  ASSERT_OK(db->Put(write_options, Slice("foo2"), Slice("bar")));
+
+  Transaction* txn = db->BeginTransaction(write_options);
+  ASSERT_TRUE(txn);
+
+  snapshot_read_options.snapshot = txn->GetSnapshot();
+
+  ASSERT_OK(txn->GetForUpdate(snapshot_read_options, "foo", &value));
+  ASSERT_EQ(value, "bar");
+
+  s = txn->Put(Slice("foo"), Slice("bar2"));
+  ASSERT_OK(s);
+
+  ASSERT_OK(txn->GetForUpdate(snapshot_read_options, "foo", &value));
+  ASSERT_EQ(value, "bar2");
+
+  // Put a random key so we have a memtable to flush
+  s = db->Put(write_options, "dummy", "dummy");
+  ASSERT_OK(s);
+
+  // force a memtable flush
+  FlushOptions flush_ops;
+  db->Flush(flush_ops);
+
+  s = txn->Commit();
+  // txn should commit since the flushed table is still in MemtableList History
+  ASSERT_OK(s);
+
+  db->Get(read_options, "foo", &value);
+  ASSERT_EQ(value, "bar2");
+
+  delete txn;
+}
+
+TEST_P(TransactionTest, FlushTest2) {
+  const size_t num_tests = 3;
+
+  for (size_t n = 0; n < num_tests; n++) {
+    // Test different table factories
+    switch (n) {
+      case 0:
+        break;
+      case 1:
+        options.table_factory.reset(new mock::MockTableFactory());
+        break;
+      case 2: {
+        PlainTableOptions pt_opts;
+        pt_opts.hash_table_ratio = 0;
+        options.table_factory.reset(NewPlainTableFactory(pt_opts));
+        break;
+      }
+    }
+
+    Status s = ReOpen();
+    ASSERT_OK(s);
+    assert(db != nullptr);
+
+    WriteOptions write_options;
+    ReadOptions read_options, snapshot_read_options;
+    TransactionOptions txn_options;
+    std::string value;
+
+    DBImpl* db_impl = static_cast_with_check<DBImpl>(db->GetRootDB());
+
+    ASSERT_OK(db->Put(write_options, Slice("foo"), Slice("bar")));
+    ASSERT_OK(db->Put(write_options, Slice("foo2"), Slice("bar2")));
+    ASSERT_OK(db->Put(write_options, Slice("foo3"), Slice("bar3")));
+
+    txn_options.set_snapshot = true;
+    Transaction* txn = db->BeginTransaction(write_options, txn_options);
+    ASSERT_TRUE(txn);
+
+    snapshot_read_options.snapshot = txn->GetSnapshot();
+
+    ASSERT_OK(txn->GetForUpdate(snapshot_read_options, "foo", &value));
+    ASSERT_EQ(value, "bar");
+
+    s = txn->Put(Slice("foo"), Slice("bar2"));
+    ASSERT_OK(s);
+
+    ASSERT_OK(txn->GetForUpdate(snapshot_read_options, "foo", &value));
+    ASSERT_EQ(value, "bar2");
+    // verify foo is locked by txn
+    s = db->Delete(write_options, "foo");
+    ASSERT_TRUE(s.IsTimedOut());
+
+    s = db->Put(write_options, "Z", "z");
+    ASSERT_OK(s);
+    s = db->Put(write_options, "dummy", "dummy");
+    ASSERT_OK(s);
+
+    s = db->Put(write_options, "S", "s");
+    ASSERT_OK(s);
+    s = db->SingleDelete(write_options, "S");
+    ASSERT_OK(s);
+
+    s = txn->Delete("S");
+    // Should fail after encountering a write to S in memtable
+    ASSERT_TRUE(s.IsBusy());
+
+    // force a memtable flush
+    s = db_impl->TEST_FlushMemTable(true);
+    ASSERT_OK(s);
+
+    // Put a random key so we have a MemTable to flush
+    s = db->Put(write_options, "dummy", "dummy2");
+    ASSERT_OK(s);
+
+    // force a memtable flush
+    ASSERT_OK(db_impl->TEST_FlushMemTable(true));
+
+    s = db->Put(write_options, "dummy", "dummy3");
+    ASSERT_OK(s);
+
+    // force a memtable flush
+    // Since our test db has max_write_buffer_number=2, this flush will cause
+    // the first memtable to get purged from the MemtableList history.
+    ASSERT_OK(db_impl->TEST_FlushMemTable(true));
+
+    s = txn->Put("X", "Y");
+    // Should succeed after verifying there is no write to X in SST file
+    ASSERT_OK(s);
+
+    s = txn->Put("Z", "zz");
+    // Should fail after encountering a write to Z in SST file
+    ASSERT_TRUE(s.IsBusy());
+
+    s = txn->GetForUpdate(read_options, "foo2", &value);
+    // should succeed since key was written before txn started
+    ASSERT_OK(s);
+    // verify foo2 is locked by txn
+    s = db->Delete(write_options, "foo2");
+    ASSERT_TRUE(s.IsTimedOut());
+
+    s = txn->Delete("S");
+    // Should fail after encountering a write to S in SST file
+    ASSERT_TRUE(s.IsBusy());
+
+    // Write a bunch of keys to db to force a compaction
+    Random rnd(47);
+    for (int i = 0; i < 1000; i++) {
+      s = db->Put(write_options, std::to_string(i),
+                  test::CompressibleString(&rnd, 0.8, 100, &value));
+      ASSERT_OK(s);
+    }
+
+    s = txn->Put("X", "yy");
+    // Should succeed after verifying there is no write to X in SST file
+    ASSERT_OK(s);
+
+    s = txn->Put("Z", "zzz");
+    // Should fail after encountering a write to Z in SST file
+    ASSERT_TRUE(s.IsBusy());
+
+    s = txn->Delete("S");
+    // Should fail after encountering a write to S in SST file
+    ASSERT_TRUE(s.IsBusy());
+
+    s = txn->GetForUpdate(read_options, "foo3", &value);
+    // should succeed since key was written before txn started
+    ASSERT_OK(s);
+    // verify foo3 is locked by txn
+    s = db->Delete(write_options, "foo3");
+    ASSERT_TRUE(s.IsTimedOut());
+
+    ASSERT_OK(db_impl->TEST_WaitForCompact());
+
+    s = txn->Commit();
+    ASSERT_OK(s);
+
+    // Transaction should only write the keys that succeeded.
+    s = db->Get(read_options, "foo", &value);
+    ASSERT_EQ(value, "bar2");
+
+    s = db->Get(read_options, "X", &value);
+    ASSERT_OK(s);
+    ASSERT_EQ("yy", value);
+
+    s = db->Get(read_options, "Z", &value);
+    ASSERT_OK(s);
+    ASSERT_EQ("z", value);
+
+    delete txn;
+  }
+}
+
+TEST_P(TransactionTest, NoSnapshotTest) {
+  WriteOptions write_options;
+  ReadOptions read_options;
+  std::string value;
+  Status s;
+
+  ASSERT_OK(db->Put(write_options, "AAA", "bar"));
+
+  Transaction* txn = db->BeginTransaction(write_options);
+  ASSERT_TRUE(txn);
+
+  // Modify key after transaction start
+  ASSERT_OK(db->Put(write_options, "AAA", "bar1"));
+
+  // Read and write without a snap
+  ASSERT_OK(txn->GetForUpdate(read_options, "AAA", &value));
+  ASSERT_EQ(value, "bar1");
+  s = txn->Put("AAA", "bar2");
+  ASSERT_OK(s);
+
+  // Should commit since read/write was done after data changed
+  s = txn->Commit();
+  ASSERT_OK(s);
+
+  ASSERT_OK(txn->GetForUpdate(read_options, "AAA", &value));
+  ASSERT_EQ(value, "bar2");
+
+  delete txn;
+}
+
+TEST_P(TransactionTest, MultipleSnapshotTest) {
+  WriteOptions write_options;
+  ReadOptions read_options, snapshot_read_options;
+  std::string value;
+  Status s;
+
+  ASSERT_OK(db->Put(write_options, "AAA", "bar"));
+  ASSERT_OK(db->Put(write_options, "BBB", "bar"));
+  ASSERT_OK(db->Put(write_options, "CCC", "bar"));
+
+  Transaction* txn = db->BeginTransaction(write_options);
+  ASSERT_TRUE(txn);
+
+  ASSERT_OK(db->Put(write_options, "AAA", "bar1"));
+
+  // Read and write without a snapshot
+  ASSERT_OK(txn->GetForUpdate(read_options, "AAA", &value));
+  ASSERT_EQ(value, "bar1");
+  s = txn->Put("AAA", "bar2");
+  ASSERT_OK(s);
+
+  // Modify BBB before snapshot is taken
+  ASSERT_OK(db->Put(write_options, "BBB", "bar1"));
+
+  txn->SetSnapshot();
+  snapshot_read_options.snapshot = txn->GetSnapshot();
+
+  // Read and write with snapshot
+  ASSERT_OK(txn->GetForUpdate(snapshot_read_options, "BBB", &value));
+  ASSERT_EQ(value, "bar1");
+  s = txn->Put("BBB", "bar2");
+  ASSERT_OK(s);
+
+  ASSERT_OK(db->Put(write_options, "CCC", "bar1"));
+
+  // Set a new snapshot
+  txn->SetSnapshot();
+  snapshot_read_options.snapshot = txn->GetSnapshot();
+
+  // Read and write with snapshot
+  ASSERT_OK(txn->GetForUpdate(snapshot_read_options, "CCC", &value));
+  ASSERT_EQ(value, "bar1");
+  s = txn->Put("CCC", "bar2");
+  ASSERT_OK(s);
+
+  s = txn->GetForUpdate(read_options, "AAA", &value);
+  ASSERT_OK(s);
+  ASSERT_EQ(value, "bar2");
+  s = txn->GetForUpdate(read_options, "BBB", &value);
+  ASSERT_OK(s);
+  ASSERT_EQ(value, "bar2");
+  s = txn->GetForUpdate(read_options, "CCC", &value);
+  ASSERT_OK(s);
+  ASSERT_EQ(value, "bar2");
+
+  s = db->Get(read_options, "AAA", &value);
+  ASSERT_OK(s);
+  ASSERT_EQ(value, "bar1");
+  s = db->Get(read_options, "BBB", &value);
+  ASSERT_OK(s);
+  ASSERT_EQ(value, "bar1");
+  s = db->Get(read_options, "CCC", &value);
+  ASSERT_OK(s);
+  ASSERT_EQ(value, "bar1");
+
+  s = txn->Commit();
+  ASSERT_OK(s);
+
+  s = db->Get(read_options, "AAA", &value);
+  ASSERT_OK(s);
+  ASSERT_EQ(value, "bar2");
+  s = db->Get(read_options, "BBB", &value);
+  ASSERT_OK(s);
+  ASSERT_EQ(value, "bar2");
+  s = db->Get(read_options, "CCC", &value);
+  ASSERT_OK(s);
+  ASSERT_EQ(value, "bar2");
+
+  // verify that we track multiple writes to the same key at different snapshots
+  delete txn;
+  txn = db->BeginTransaction(write_options);
+
+  // Potentially conflicting writes
+  ASSERT_OK(db->Put(write_options, "ZZZ", "zzz"));
+  ASSERT_OK(db->Put(write_options, "XXX", "xxx"));
+
+  txn->SetSnapshot();
+
+  TransactionOptions txn_options;
+  txn_options.set_snapshot = true;
+  Transaction* txn2 = db->BeginTransaction(write_options, txn_options);
+  txn2->SetSnapshot();
+
+  // This should not conflict in txn since the snapshot is later than the
+  // previous write (spoiler alert:  it will later conflict with txn2).
+  s = txn->Put("ZZZ", "zzzz");
+  ASSERT_OK(s);
+
+  s = txn->Commit();
+  ASSERT_OK(s);
+
+  delete txn;
+
+  // This will conflict since the snapshot is earlier than another write to ZZZ
+  s = txn2->Put("ZZZ", "xxxxx");
+  ASSERT_TRUE(s.IsBusy());
+
+  s = txn2->Commit();
+  ASSERT_OK(s);
+
+  s = db->Get(read_options, "ZZZ", &value);
+  ASSERT_OK(s);
+  ASSERT_EQ(value, "zzzz");
+
+  delete txn2;
+}
+
+TEST_P(TransactionTest, ColumnFamiliesTest) {
+  WriteOptions write_options;
+  ReadOptions read_options, snapshot_read_options;
+  TransactionOptions txn_options;
+  std::string value;
+  Status s;
+
+  ColumnFamilyHandle *cfa, *cfb;
+  ColumnFamilyOptions cf_options;
+
+  // Create 2 new column families
+  s = db->CreateColumnFamily(cf_options, "CFA", &cfa);
+  ASSERT_OK(s);
+  s = db->CreateColumnFamily(cf_options, "CFB", &cfb);
+  ASSERT_OK(s);
+
+  delete cfa;
+  delete cfb;
+  delete db;
+  db = nullptr;
+
+  // open DB with three column families
+  std::vector<ColumnFamilyDescriptor> column_families;
+  // have to open default column family
+  column_families.push_back(
+      ColumnFamilyDescriptor(kDefaultColumnFamilyName, ColumnFamilyOptions()));
+  // open the new column families
+  column_families.push_back(
+      ColumnFamilyDescriptor("CFA", ColumnFamilyOptions()));
+  column_families.push_back(
+      ColumnFamilyDescriptor("CFB", ColumnFamilyOptions()));
+
+  std::vector<ColumnFamilyHandle*> handles;
+
+  ASSERT_OK(ReOpenNoDelete(column_families, &handles));
+  assert(db != nullptr);
+
+  Transaction* txn = db->BeginTransaction(write_options);
+  ASSERT_TRUE(txn);
+
+  txn->SetSnapshot();
+  snapshot_read_options.snapshot = txn->GetSnapshot();
+
+  txn_options.set_snapshot = true;
+  Transaction* txn2 = db->BeginTransaction(write_options, txn_options);
+  ASSERT_TRUE(txn2);
+
+  // Write some data to the db
+  WriteBatch batch;
+  ASSERT_OK(batch.Put("foo", "foo"));
+  ASSERT_OK(batch.Put(handles[1], "AAA", "bar"));
+  ASSERT_OK(batch.Put(handles[1], "AAAZZZ", "bar"));
+  s = db->Write(write_options, &batch);
+  ASSERT_OK(s);
+  ASSERT_OK(db->Delete(write_options, handles[1], "AAAZZZ"));
+
+  // These keys do not conflict with existing writes since they're in
+  // different column families
+  s = txn->Delete("AAA");
+  ASSERT_OK(s);
+  s = txn->GetForUpdate(snapshot_read_options, handles[1], "foo", &value);
+  ASSERT_TRUE(s.IsNotFound());
+  Slice key_slice("AAAZZZ");
+  Slice value_slices[2] = {Slice("bar"), Slice("bar")};
+  s = txn->Put(handles[2], SliceParts(&key_slice, 1),
+               SliceParts(value_slices, 2));
+  ASSERT_OK(s);
+  ASSERT_EQ(3, txn->GetNumKeys());
+
+  s = txn->Commit();
+  ASSERT_OK(s);
+  s = db->Get(read_options, "AAA", &value);
+  ASSERT_TRUE(s.IsNotFound());
+  s = db->Get(read_options, handles[2], "AAAZZZ", &value);
+  ASSERT_EQ(value, "barbar");
+
+  Slice key_slices[3] = {Slice("AAA"), Slice("ZZ"), Slice("Z")};
+  Slice value_slice("barbarbar");
+
+  s = txn2->Delete(handles[2], "XXX");
+  ASSERT_OK(s);
+  s = txn2->Delete(handles[1], "XXX");
+  ASSERT_OK(s);
+
+  // This write will cause a conflict with the earlier batch write
+  s = txn2->Put(handles[1], SliceParts(key_slices, 3),
+                SliceParts(&value_slice, 1));
+  ASSERT_TRUE(s.IsBusy());
+
+  s = txn2->Commit();
+  ASSERT_OK(s);
+  // In the above the latest change to AAAZZZ in handles[1] is delete.
+  s = db->Get(read_options, handles[1], "AAAZZZ", &value);
+  ASSERT_TRUE(s.IsNotFound());
+
+  delete txn;
+  delete txn2;
+
+  txn = db->BeginTransaction(write_options, txn_options);
+  snapshot_read_options.snapshot = txn->GetSnapshot();
+
+  txn2 = db->BeginTransaction(write_options, txn_options);
+  ASSERT_TRUE(txn);
+
+  std::vector<ColumnFamilyHandle*> multiget_cfh = {handles[1], handles[2],
+                                                   handles[0], handles[2]};
+  std::vector<Slice> multiget_keys = {"AAA", "AAAZZZ", "foo", "foo"};
+  std::vector<std::string> values(4);
+  std::vector<Status> results = txn->MultiGetForUpdate(
+      snapshot_read_options, multiget_cfh, multiget_keys, &values);
+  ASSERT_OK(results[0]);
+  ASSERT_OK(results[1]);
+  ASSERT_OK(results[2]);
+  ASSERT_TRUE(results[3].IsNotFound());
+  ASSERT_EQ(values[0], "bar");
+  ASSERT_EQ(values[1], "barbar");
+  ASSERT_EQ(values[2], "foo");
+
+  s = txn->SingleDelete(handles[2], "ZZZ");
+  ASSERT_OK(s);
+  s = txn->Put(handles[2], "ZZZ", "YYY");
+  ASSERT_OK(s);
+  s = txn->Put(handles[2], "ZZZ", "YYYY");
+  ASSERT_OK(s);
+  s = txn->Delete(handles[2], "ZZZ");
+  ASSERT_OK(s);
+  s = txn->Put(handles[2], "AAAZZZ", "barbarbar");
+  ASSERT_OK(s);
+
+  ASSERT_EQ(5, txn->GetNumKeys());
+
+  // Txn should commit
+  s = txn->Commit();
+  ASSERT_OK(s);
+  s = db->Get(read_options, handles[2], "ZZZ", &value);
+  ASSERT_TRUE(s.IsNotFound());
+
+  // Put a key which will conflict with the next txn using the previous snapshot
+  ASSERT_OK(db->Put(write_options, handles[2], "foo", "000"));
+
+  results = txn2->MultiGetForUpdate(snapshot_read_options, multiget_cfh,
+                                    multiget_keys, &values);
+  // All results should fail since there was a conflict
+  ASSERT_TRUE(results[0].IsBusy());
+  ASSERT_TRUE(results[1].IsBusy());
+  ASSERT_TRUE(results[2].IsBusy());
+  ASSERT_TRUE(results[3].IsBusy());
+
+  s = db->Get(read_options, handles[2], "foo", &value);
+  ASSERT_EQ(value, "000");
+
+  s = txn2->Commit();
+  ASSERT_OK(s);
+
+  s = db->DropColumnFamily(handles[1]);
+  ASSERT_OK(s);
+  s = db->DropColumnFamily(handles[2]);
+  ASSERT_OK(s);
+
+  delete txn;
+  delete txn2;
+
+  for (auto handle : handles) {
+    delete handle;
+  }
+}
+
+TEST_P(TransactionTest, MultiGetBatchedTest) {
+  WriteOptions write_options;
+  ReadOptions read_options, snapshot_read_options;
+  TransactionOptions txn_options;
+  std::string value;
+  Status s;
+
+  ColumnFamilyHandle* cf;
+  ColumnFamilyOptions cf_options;
+
+  // Create a new column families
+  s = db->CreateColumnFamily(cf_options, "CF", &cf);
+  ASSERT_OK(s);
+
+  delete cf;
+  delete db;
+  db = nullptr;
+
+  // open DB with three column families
+  std::vector<ColumnFamilyDescriptor> column_families;
+  // have to open default column family
+  column_families.push_back(
+      ColumnFamilyDescriptor(kDefaultColumnFamilyName, ColumnFamilyOptions()));
+  // open the new column families
+  cf_options.merge_operator = MergeOperators::CreateStringAppendOperator();
+  column_families.push_back(ColumnFamilyDescriptor("CF", cf_options));
+
+  std::vector<ColumnFamilyHandle*> handles;
+
+  options.merge_operator = MergeOperators::CreateStringAppendOperator();
+  ASSERT_OK(ReOpenNoDelete(column_families, &handles));
+  assert(db != nullptr);
+
+  // Write some data to the db
+  WriteBatch batch;
+  ASSERT_OK(batch.Put(handles[1], "aaa", "val1"));
+  ASSERT_OK(batch.Put(handles[1], "bbb", "val2"));
+  ASSERT_OK(batch.Put(handles[1], "ccc", "val3"));
+  ASSERT_OK(batch.Put(handles[1], "ddd", "foo"));
+  ASSERT_OK(batch.Put(handles[1], "eee", "val5"));
+  ASSERT_OK(batch.Put(handles[1], "fff", "val6"));
+  ASSERT_OK(batch.Merge(handles[1], "ggg", "foo"));
+  s = db->Write(write_options, &batch);
+  ASSERT_OK(s);
+
+  Transaction* txn = db->BeginTransaction(write_options);
+  ASSERT_TRUE(txn);
+
+  txn->SetSnapshot();
+  snapshot_read_options.snapshot = txn->GetSnapshot();
+
+  txn_options.set_snapshot = true;
+  // Write some data to the db
+  s = txn->Delete(handles[1], "bbb");
+  ASSERT_OK(s);
+  s = txn->Put(handles[1], "ccc", "val3_new");
+  ASSERT_OK(s);
+  s = txn->Merge(handles[1], "ddd", "bar");
+  ASSERT_OK(s);
+
+  std::vector<Slice> keys = {"aaa", "bbb", "ccc", "ddd", "eee", "fff", "ggg"};
+  std::vector<PinnableSlice> values(keys.size());
+  std::vector<Status> statuses(keys.size());
+
+  txn->MultiGet(snapshot_read_options, handles[1], keys.size(), keys.data(),
+                values.data(), statuses.data());
+  ASSERT_TRUE(statuses[0].ok());
+  ASSERT_EQ(values[0], "val1");
+  ASSERT_TRUE(statuses[1].IsNotFound());
+  ASSERT_TRUE(statuses[2].ok());
+  ASSERT_EQ(values[2], "val3_new");
+  ASSERT_TRUE(statuses[3].ok());
+  ASSERT_EQ(values[3], "foo,bar");
+  ASSERT_TRUE(statuses[4].ok());
+  ASSERT_EQ(values[4], "val5");
+  ASSERT_TRUE(statuses[5].ok());
+  ASSERT_EQ(values[5], "val6");
+  ASSERT_TRUE(statuses[6].ok());
+  ASSERT_EQ(values[6], "foo");
+  delete txn;
+  for (auto handle : handles) {
+    delete handle;
+  }
+}
+
+// This test calls WriteBatchWithIndex::MultiGetFromBatchAndDB with a large
+// number of keys, i.e greater than MultiGetContext::MAX_BATCH_SIZE, which is
+// is 32. This forces autovector allocations in the MultiGet code paths
+// to use std::vector in addition to stack allocations. The MultiGet keys
+// includes Merges, which are handled specially in MultiGetFromBatchAndDB by
+// allocating an autovector of MergeContexts
+TEST_P(TransactionTest, MultiGetLargeBatchedTest) {
+  WriteOptions write_options;
+  ReadOptions read_options, snapshot_read_options;
+  std::string value;
+  Status s;
+
+  ColumnFamilyHandle* cf;
+  ColumnFamilyOptions cf_options;
+
+  std::vector<std::string> key_str;
+  for (int i = 0; i < 100; ++i) {
+    key_str.emplace_back(std::to_string(i));
+  }
+  // Create a new column families
+  s = db->CreateColumnFamily(cf_options, "CF", &cf);
+  ASSERT_OK(s);
+
+  delete cf;
+  delete db;
+  db = nullptr;
+
+  // open DB with three column families
+  std::vector<ColumnFamilyDescriptor> column_families;
+  // have to open default column family
+  column_families.push_back(
+      ColumnFamilyDescriptor(kDefaultColumnFamilyName, ColumnFamilyOptions()));
+  // open the new column families
+  cf_options.merge_operator = MergeOperators::CreateStringAppendOperator();
+  column_families.push_back(ColumnFamilyDescriptor("CF", cf_options));
+
+  std::vector<ColumnFamilyHandle*> handles;
+
+  options.merge_operator = MergeOperators::CreateStringAppendOperator();
+  ASSERT_OK(ReOpenNoDelete(column_families, &handles));
+  assert(db != nullptr);
+
+  // Write some data to the db
+  WriteBatch batch;
+  for (int i = 0; i < 3 * MultiGetContext::MAX_BATCH_SIZE; ++i) {
+    std::string val = "val" + std::to_string(i);
+    ASSERT_OK(batch.Put(handles[1], key_str[i], val));
+  }
+  s = db->Write(write_options, &batch);
+  ASSERT_OK(s);
+
+  WriteBatchWithIndex wb;
+  // Write some data to the db
+  s = wb.Delete(handles[1], std::to_string(1));
+  ASSERT_OK(s);
+  s = wb.Put(handles[1], std::to_string(2), "new_val" + std::to_string(2));
+  ASSERT_OK(s);
+  // Write a lot of merges so when we call MultiGetFromBatchAndDB later on,
+  // it is forced to use std::vector in ROCKSDB_NAMESPACE::autovector to
+  // allocate MergeContexts. The number of merges needs to be >
+  // MultiGetContext::MAX_BATCH_SIZE
+  for (int i = 8; i < MultiGetContext::MAX_BATCH_SIZE + 24; ++i) {
+    s = wb.Merge(handles[1], std::to_string(i), "merge");
+    ASSERT_OK(s);
+  }
+
+  // MultiGet a lot of keys in order to force std::vector reallocations
+  std::vector<Slice> keys;
+  for (int i = 0; i < MultiGetContext::MAX_BATCH_SIZE + 32; ++i) {
+    keys.emplace_back(key_str[i]);
+  }
+  std::vector<PinnableSlice> values(keys.size());
+  std::vector<Status> statuses(keys.size());
+
+  wb.MultiGetFromBatchAndDB(db, snapshot_read_options, handles[1], keys.size(),
+                            keys.data(), values.data(), statuses.data(), false);
+  for (size_t i = 0; i < keys.size(); ++i) {
+    if (i == 1) {
+      ASSERT_TRUE(statuses[1].IsNotFound());
+    } else if (i == 2) {
+      ASSERT_TRUE(statuses[2].ok());
+      ASSERT_EQ(values[2], "new_val" + std::to_string(2));
+    } else if (i >= 8 && i < 56) {
+      ASSERT_TRUE(statuses[i].ok());
+      ASSERT_EQ(values[i], "val" + std::to_string(i) + ",merge");
+    } else {
+      ASSERT_TRUE(statuses[i].ok());
+      if (values[i] != "val" + std::to_string(i)) {
+        ASSERT_EQ(values[i], "val" + std::to_string(i));
+      }
+    }
+  }
+
+  for (auto handle : handles) {
+    delete handle;
+  }
+}
+
+TEST_P(TransactionTest, MultiGetSnapshot) {
+  WriteOptions write_options;
+  TransactionOptions transaction_options;
+  Transaction* txn1 = db->BeginTransaction(write_options, transaction_options);
+
+  Slice key = "foo";
+
+  Status s = txn1->Put(key, "bar");
+  ASSERT_OK(s);
+
+  s = txn1->SetName("test");
+  ASSERT_OK(s);
+
+  s = txn1->Prepare();
+  ASSERT_OK(s);
+
+  // Get snapshot between prepare and commit
+  // Un-committed data should be invisible to other transactions
+  const Snapshot* s1 = db->GetSnapshot();
+
+  s = txn1->Commit();
+  ASSERT_OK(s);
+  delete txn1;
+
+  Transaction* txn2 = db->BeginTransaction(write_options, transaction_options);
+  ReadOptions read_options;
+  read_options.snapshot = s1;
+
+  std::vector<Slice> keys;
+  std::vector<PinnableSlice> values(1);
+  std::vector<Status> statuses(1);
+  keys.push_back(key);
+  auto cfd = db->DefaultColumnFamily();
+  txn2->MultiGet(read_options, cfd, 1, keys.data(), values.data(),
+                 statuses.data());
+  ASSERT_TRUE(statuses[0].IsNotFound());
+  delete txn2;
+
+  db->ReleaseSnapshot(s1);
+}
+
+TEST_P(TransactionTest, ColumnFamiliesTest2) {
+  WriteOptions write_options;
+  ReadOptions read_options, snapshot_read_options;
+  std::string value;
+  Status s;
+
+  ColumnFamilyHandle *one, *two;
+  ColumnFamilyOptions cf_options;
+
+  // Create 2 new column families
+  s = db->CreateColumnFamily(cf_options, "ONE", &one);
+  ASSERT_OK(s);
+  s = db->CreateColumnFamily(cf_options, "TWO", &two);
+  ASSERT_OK(s);
+
+  Transaction* txn1 = db->BeginTransaction(write_options);
+  ASSERT_TRUE(txn1);
+  Transaction* txn2 = db->BeginTransaction(write_options);
+  ASSERT_TRUE(txn2);
+
+  s = txn1->Put(one, "X", "1");
+  ASSERT_OK(s);
+  s = txn1->Put(two, "X", "2");
+  ASSERT_OK(s);
+  s = txn1->Put("X", "0");
+  ASSERT_OK(s);
+
+  s = txn2->Put(one, "X", "11");
+  ASSERT_TRUE(s.IsTimedOut());
+
+  s = txn1->Commit();
+  ASSERT_OK(s);
+
+  // Drop first column family
+  s = db->DropColumnFamily(one);
+  ASSERT_OK(s);
+
+  // Should fail since column family was dropped.
+  s = txn2->Commit();
+  ASSERT_OK(s);
+
+  delete txn1;
+  txn1 = db->BeginTransaction(write_options);
+  ASSERT_TRUE(txn1);
+
+  // Should fail since column family was dropped
+  s = txn1->Put(one, "X", "111");
+  ASSERT_TRUE(s.IsInvalidArgument());
+
+  s = txn1->Put(two, "X", "222");
+  ASSERT_OK(s);
+
+  s = txn1->Put("X", "000");
+  ASSERT_OK(s);
+
+  s = txn1->Commit();
+  ASSERT_OK(s);
+
+  s = db->Get(read_options, two, "X", &value);
+  ASSERT_OK(s);
+  ASSERT_EQ("222", value);
+
+  s = db->Get(read_options, "X", &value);
+  ASSERT_OK(s);
+  ASSERT_EQ("000", value);
+
+  s = db->DropColumnFamily(two);
+  ASSERT_OK(s);
+
+  delete txn1;
+  delete txn2;
+
+  delete one;
+  delete two;
+}
+
+TEST_P(TransactionTest, EmptyTest) {
+  WriteOptions write_options;
+  ReadOptions read_options;
+  std::string value;
+  Status s;
+
+  s = db->Put(write_options, "aaa", "aaa");
+  ASSERT_OK(s);
+
+  Transaction* txn = db->BeginTransaction(write_options);
+  s = txn->Commit();
+  ASSERT_OK(s);
+  delete txn;
+
+  txn = db->BeginTransaction(write_options);
+  ASSERT_OK(txn->Rollback());
+  delete txn;
+
+  txn = db->BeginTransaction(write_options);
+  s = txn->GetForUpdate(read_options, "aaa", &value);
+  ASSERT_EQ(value, "aaa");
+
+  s = txn->Commit();
+  ASSERT_OK(s);
+  delete txn;
+
+  txn = db->BeginTransaction(write_options);
+  txn->SetSnapshot();
+
+  s = txn->GetForUpdate(read_options, "aaa", &value);
+  ASSERT_EQ(value, "aaa");
+
+  // Conflicts with previous GetForUpdate
+  s = db->Put(write_options, "aaa", "xxx");
+  ASSERT_TRUE(s.IsTimedOut());
+
+  // transaction expired!
+  s = txn->Commit();
+  ASSERT_OK(s);
+  delete txn;
+}
+
+TEST_P(TransactionTest, PredicateManyPreceders) {
+  WriteOptions write_options;
+  ReadOptions read_options1, read_options2;
+  TransactionOptions txn_options;
+  std::string value;
+  Status s;
+
+  txn_options.set_snapshot = true;
+  Transaction* txn1 = db->BeginTransaction(write_options, txn_options);
+  read_options1.snapshot = txn1->GetSnapshot();
+
+  Transaction* txn2 = db->BeginTransaction(write_options);
+  txn2->SetSnapshot();
+  read_options2.snapshot = txn2->GetSnapshot();
+
+  std::vector<Slice> multiget_keys = {"1", "2", "3"};
+  std::vector<std::string> multiget_values;
+
+  std::vector<Status> results =
+      txn1->MultiGetForUpdate(read_options1, multiget_keys, &multiget_values);
+  ASSERT_EQ(results.size(), 3);
+  ASSERT_TRUE(results[0].IsNotFound());
+  ASSERT_TRUE(results[1].IsNotFound());
+  ASSERT_TRUE(results[2].IsNotFound());
+
+  s = txn2->Put("2", "x");  // Conflict's with txn1's MultiGetForUpdate
+  ASSERT_TRUE(s.IsTimedOut());
+
+  ASSERT_OK(txn2->Rollback());
+
+  multiget_values.clear();
+  results =
+      txn1->MultiGetForUpdate(read_options1, multiget_keys, &multiget_values);
+  ASSERT_EQ(results.size(), 3);
+  ASSERT_TRUE(results[0].IsNotFound());
+  ASSERT_TRUE(results[1].IsNotFound());
+  ASSERT_TRUE(results[2].IsNotFound());
+
+  s = txn1->Commit();
+  ASSERT_OK(s);
+
+  delete txn1;
+  delete txn2;
+
+  txn1 = db->BeginTransaction(write_options, txn_options);
+  read_options1.snapshot = txn1->GetSnapshot();
+
+  txn2 = db->BeginTransaction(write_options, txn_options);
+  read_options2.snapshot = txn2->GetSnapshot();
+
+  s = txn1->Put("4", "x");
+  ASSERT_OK(s);
+
+  s = txn2->Delete("4");  // conflict
+  ASSERT_TRUE(s.IsTimedOut());
+
+  s = txn1->Commit();
+  ASSERT_OK(s);
+
+  s = txn2->GetForUpdate(read_options2, "4", &value);
+  ASSERT_TRUE(s.IsBusy());
+
+  ASSERT_OK(txn2->Rollback());
+
+  delete txn1;
+  delete txn2;
+}
+
+TEST_P(TransactionTest, LostUpdate) {
+  WriteOptions write_options;
+  ReadOptions read_options, read_options1, read_options2;
+  TransactionOptions txn_options;
+  std::string value;
+  Status s;
+
+  // Test 2 transactions writing to the same key in multiple orders and
+  // with/without snapshots
+
+  Transaction* txn1 = db->BeginTransaction(write_options);
+  Transaction* txn2 = db->BeginTransaction(write_options);
+
+  s = txn1->Put("1", "1");
+  ASSERT_OK(s);
+
+  s = txn2->Put("1", "2");  // conflict
+  ASSERT_TRUE(s.IsTimedOut());
+
+  s = txn2->Commit();
+  ASSERT_OK(s);
+
+  s = txn1->Commit();
+  ASSERT_OK(s);
+
+  s = db->Get(read_options, "1", &value);
+  ASSERT_OK(s);
+  ASSERT_EQ("1", value);
+
+  delete txn1;
+  delete txn2;
+
+  txn_options.set_snapshot = true;
+  txn1 = db->BeginTransaction(write_options, txn_options);
+  read_options1.snapshot = txn1->GetSnapshot();
+
+  txn2 = db->BeginTransaction(write_options, txn_options);
+  read_options2.snapshot = txn2->GetSnapshot();
+
+  s = txn1->Put("1", "3");
+  ASSERT_OK(s);
+  s = txn2->Put("1", "4");  // conflict
+  ASSERT_TRUE(s.IsTimedOut());
+
+  s = txn1->Commit();
+  ASSERT_OK(s);
+
+  s = txn2->Commit();
+  ASSERT_OK(s);
+
+  s = db->Get(read_options, "1", &value);
+  ASSERT_OK(s);
+  ASSERT_EQ("3", value);
+
+  delete txn1;
+  delete txn2;
+
+  txn1 = db->BeginTransaction(write_options, txn_options);
+  read_options1.snapshot = txn1->GetSnapshot();
+
+  txn2 = db->BeginTransaction(write_options, txn_options);
+  read_options2.snapshot = txn2->GetSnapshot();
+
+  s = txn1->Put("1", "5");
+  ASSERT_OK(s);
+
+  s = txn1->Commit();
+  ASSERT_OK(s);
+
+  s = txn2->Put("1", "6");
+  ASSERT_TRUE(s.IsBusy());
+  s = txn2->Commit();
+  ASSERT_OK(s);
+
+  s = db->Get(read_options, "1", &value);
+  ASSERT_OK(s);
+  ASSERT_EQ("5", value);
+
+  delete txn1;
+  delete txn2;
+
+  txn1 = db->BeginTransaction(write_options, txn_options);
+  read_options1.snapshot = txn1->GetSnapshot();
+
+  txn2 = db->BeginTransaction(write_options, txn_options);
+  read_options2.snapshot = txn2->GetSnapshot();
+
+  s = txn1->Put("1", "7");
+  ASSERT_OK(s);
+  s = txn1->Commit();
+  ASSERT_OK(s);
+
+  txn2->SetSnapshot();
+  s = txn2->Put("1", "8");
+  ASSERT_OK(s);
+  s = txn2->Commit();
+  ASSERT_OK(s);
+
+  s = db->Get(read_options, "1", &value);
+  ASSERT_OK(s);
+  ASSERT_EQ("8", value);
+
+  delete txn1;
+  delete txn2;
+
+  txn1 = db->BeginTransaction(write_options);
+  txn2 = db->BeginTransaction(write_options);
+
+  s = txn1->Put("1", "9");
+  ASSERT_OK(s);
+  s = txn1->Commit();
+  ASSERT_OK(s);
+
+  s = txn2->Put("1", "10");
+  ASSERT_OK(s);
+  s = txn2->Commit();
+  ASSERT_OK(s);
+
+  delete txn1;
+  delete txn2;
+
+  s = db->Get(read_options, "1", &value);
+  ASSERT_OK(s);
+  ASSERT_EQ(value, "10");
+}
+
+TEST_P(TransactionTest, UntrackedWrites) {
+  if (txn_db_options.write_policy == WRITE_UNPREPARED) {
+    // TODO(lth): For WriteUnprepared, validate that untracked writes are
+    // not supported.
+    return;
+  }
+
+  WriteOptions write_options;
+  ReadOptions read_options;
+  std::string value;
+  Status s;
+
+  // Verify transaction rollback works for untracked keys.
+  Transaction* txn = db->BeginTransaction(write_options);
+  txn->SetSnapshot();
+
+  s = txn->PutUntracked("untracked", "0");
+  ASSERT_OK(s);
+  ASSERT_OK(txn->Rollback());
+  s = db->Get(read_options, "untracked", &value);
+  ASSERT_TRUE(s.IsNotFound());
+
+  delete txn;
+  txn = db->BeginTransaction(write_options);
+  txn->SetSnapshot();
+
+  s = db->Put(write_options, "untracked", "x");
+  ASSERT_OK(s);
+
+  // Untracked writes should succeed even though key was written after snapshot
+  s = txn->PutUntracked("untracked", "1");
+  ASSERT_OK(s);
+  s = txn->MergeUntracked("untracked", "2");
+  ASSERT_OK(s);
+  s = txn->DeleteUntracked("untracked");
+  ASSERT_OK(s);
+
+  // Conflict
+  s = txn->Put("untracked", "3");
+  ASSERT_TRUE(s.IsBusy());
+
+  s = txn->Commit();
+  ASSERT_OK(s);
+
+  s = db->Get(read_options, "untracked", &value);
+  ASSERT_TRUE(s.IsNotFound());
+
+  delete txn;
+}
+
+TEST_P(TransactionTest, ExpiredTransaction) {
+  WriteOptions write_options;
+  ReadOptions read_options;
+  TransactionOptions txn_options;
+  std::string value;
+  Status s;
+
+  // Set txn expiration timeout to 0 microseconds (expires instantly)
+  txn_options.expiration = 0;
+  Transaction* txn1 = db->BeginTransaction(write_options, txn_options);
+
+  s = txn1->Put("X", "1");
+  ASSERT_OK(s);
+
+  s = txn1->Put("Y", "1");
+  ASSERT_OK(s);
+
+  Transaction* txn2 = db->BeginTransaction(write_options);
+
+  // txn2 should be able to write to X since txn1 has expired
+  s = txn2->Put("X", "2");
+  ASSERT_OK(s);
+
+  s = txn2->Commit();
+  ASSERT_OK(s);
+  s = db->Get(read_options, "X", &value);
+  ASSERT_OK(s);
+  ASSERT_EQ("2", value);
+
+  s = txn1->Put("Z", "1");
+  ASSERT_OK(s);
+
+  // txn1 should fail to commit since it is expired
+  s = txn1->Commit();
+  ASSERT_TRUE(s.IsExpired());
+
+  s = db->Get(read_options, "Y", &value);
+  ASSERT_TRUE(s.IsNotFound());
+
+  s = db->Get(read_options, "Z", &value);
+  ASSERT_TRUE(s.IsNotFound());
+
+  delete txn1;
+  delete txn2;
+}
+
+TEST_P(TransactionTest, ReinitializeTest) {
+  WriteOptions write_options;
+  ReadOptions read_options;
+  TransactionOptions txn_options;
+  std::string value;
+  Status s;
+
+  // Set txn expiration timeout to 0 microseconds (expires instantly)
+  txn_options.expiration = 0;
+  Transaction* txn1 = db->BeginTransaction(write_options, txn_options);
+
+  // Reinitialize transaction to no long expire
+  txn_options.expiration = -1;
+  txn1 = db->BeginTransaction(write_options, txn_options, txn1);
+
+  s = txn1->Put("Z", "z");
+  ASSERT_OK(s);
+
+  // Should commit since not expired
+  s = txn1->Commit();
+  ASSERT_OK(s);
+
+  txn1 = db->BeginTransaction(write_options, txn_options, txn1);
+
+  s = txn1->Put("Z", "zz");
+  ASSERT_OK(s);
+
+  // Reinitilize txn1 and verify that Z gets unlocked
+  txn1 = db->BeginTransaction(write_options, txn_options, txn1);
+
+  Transaction* txn2 = db->BeginTransaction(write_options, txn_options, nullptr);
+  s = txn2->Put("Z", "zzz");
+  ASSERT_OK(s);
+  s = txn2->Commit();
+  ASSERT_OK(s);
+  delete txn2;
+
+  s = db->Get(read_options, "Z", &value);
+  ASSERT_OK(s);
+  ASSERT_EQ(value, "zzz");
+
+  // Verify snapshots get reinitialized correctly
+  txn1->SetSnapshot();
+  s = txn1->Put("Z", "zzzz");
+  ASSERT_OK(s);
+
+  s = txn1->Commit();
+  ASSERT_OK(s);
+
+  s = db->Get(read_options, "Z", &value);
+  ASSERT_OK(s);
+  ASSERT_EQ(value, "zzzz");
+
+  txn1 = db->BeginTransaction(write_options, txn_options, txn1);
+  const Snapshot* snapshot = txn1->GetSnapshot();
+  ASSERT_FALSE(snapshot);
+
+  txn_options.set_snapshot = true;
+  txn1 = db->BeginTransaction(write_options, txn_options, txn1);
+  snapshot = txn1->GetSnapshot();
+  ASSERT_TRUE(snapshot);
+
+  s = txn1->Put("Z", "a");
+  ASSERT_OK(s);
+
+  ASSERT_OK(txn1->Rollback());
+
+  s = txn1->Put("Y", "y");
+  ASSERT_OK(s);
+
+  txn_options.set_snapshot = false;
+  txn1 = db->BeginTransaction(write_options, txn_options, txn1);
+  snapshot = txn1->GetSnapshot();
+  ASSERT_FALSE(snapshot);
+
+  s = txn1->Put("X", "x");
+  ASSERT_OK(s);
+
+  s = txn1->Commit();
+  ASSERT_OK(s);
+
+  s = db->Get(read_options, "Z", &value);
+  ASSERT_OK(s);
+  ASSERT_EQ(value, "zzzz");
+
+  s = db->Get(read_options, "Y", &value);
+  ASSERT_TRUE(s.IsNotFound());
+
+  txn1 = db->BeginTransaction(write_options, txn_options, txn1);
+
+  s = txn1->SetName("name");
+  ASSERT_OK(s);
+
+  s = txn1->Prepare();
+  ASSERT_OK(s);
+  s = txn1->Commit();
+  ASSERT_OK(s);
+
+  txn1 = db->BeginTransaction(write_options, txn_options, txn1);
+
+  s = txn1->SetName("name");
+  ASSERT_OK(s);
+
+  delete txn1;
+}
+
+TEST_P(TransactionTest, Rollback) {
+  WriteOptions write_options;
+  ReadOptions read_options;
+  TransactionOptions txn_options;
+  std::string value;
+  Status s;
+
+  Transaction* txn1 = db->BeginTransaction(write_options, txn_options);
+
+  ASSERT_OK(s);
+
+  s = txn1->Put("X", "1");
+  ASSERT_OK(s);
+
+  Transaction* txn2 = db->BeginTransaction(write_options);
+
+  // txn2 should not be able to write to X since txn1 has it locked
+  s = txn2->Put("X", "2");
+  ASSERT_TRUE(s.IsTimedOut());
+
+  ASSERT_OK(txn1->Rollback());
+  delete txn1;
+
+  // txn2 should now be able to write to X
+  s = txn2->Put("X", "3");
+  ASSERT_OK(s);
+
+  s = txn2->Commit();
+  ASSERT_OK(s);
+
+  s = db->Get(read_options, "X", &value);
+  ASSERT_OK(s);
+  ASSERT_EQ("3", value);
+
+  delete txn2;
+}
+
+TEST_P(TransactionTest, LockLimitTest) {
+  WriteOptions write_options;
+  ReadOptions read_options, snapshot_read_options;
+  TransactionOptions txn_options;
+  std::string value;
+  Status s;
+
+  delete db;
+  db = nullptr;
+
+  // Open DB with a lock limit of 3
+  txn_db_options.max_num_locks = 3;
+  ASSERT_OK(ReOpen());
+  assert(db != nullptr);
+  ASSERT_OK(s);
+
+  // Create a txn and verify we can only lock up to 3 keys
+  Transaction* txn = db->BeginTransaction(write_options, txn_options);
+  ASSERT_TRUE(txn);
+
+  s = txn->Put("X", "x");
+  ASSERT_OK(s);
+
+  s = txn->Put("Y", "y");
+  ASSERT_OK(s);
+
+  s = txn->Put("Z", "z");
+  ASSERT_OK(s);
+
+  // lock limit reached
+  s = txn->Put("W", "w");
+  ASSERT_TRUE(s.IsBusy());
+
+  // re-locking same key shouldn't put us over the limit
+  s = txn->Put("X", "xx");
+  ASSERT_OK(s);
+
+  s = txn->GetForUpdate(read_options, "W", &value);
+  ASSERT_TRUE(s.IsBusy());
+  s = txn->GetForUpdate(read_options, "V", &value);
+  ASSERT_TRUE(s.IsBusy());
+
+  // re-locking same key shouldn't put us over the limit
+  s = txn->GetForUpdate(read_options, "Y", &value);
+  ASSERT_OK(s);
+  ASSERT_EQ("y", value);
+
+  s = txn->Get(read_options, "W", &value);
+  ASSERT_TRUE(s.IsNotFound());
+
+  Transaction* txn2 = db->BeginTransaction(write_options, txn_options);
+  ASSERT_TRUE(txn2);
+
+  // "X" currently locked
+  s = txn2->Put("X", "x");
+  ASSERT_TRUE(s.IsTimedOut());
+
+  // lock limit reached
+  s = txn2->Put("M", "m");
+  ASSERT_TRUE(s.IsBusy());
+
+  s = txn->Commit();
+  ASSERT_OK(s);
+
+  s = db->Get(read_options, "X", &value);
+  ASSERT_OK(s);
+  ASSERT_EQ("xx", value);
+
+  s = db->Get(read_options, "W", &value);
+  ASSERT_TRUE(s.IsNotFound());
+
+  // Committing txn should release its locks and allow txn2 to proceed
+  s = txn2->Put("X", "x2");
+  ASSERT_OK(s);
+
+  s = txn2->Delete("X");
+  ASSERT_OK(s);
+
+  s = txn2->Put("M", "m");
+  ASSERT_OK(s);
+
+  s = txn2->Put("Z", "z2");
+  ASSERT_OK(s);
+
+  // lock limit reached
+  s = txn2->Delete("Y");
+  ASSERT_TRUE(s.IsBusy());
+
+  s = txn2->Commit();
+  ASSERT_OK(s);
+
+  s = db->Get(read_options, "Z", &value);
+  ASSERT_OK(s);
+  ASSERT_EQ("z2", value);
+
+  s = db->Get(read_options, "Y", &value);
+  ASSERT_OK(s);
+  ASSERT_EQ("y", value);
+
+  s = db->Get(read_options, "X", &value);
+  ASSERT_TRUE(s.IsNotFound());
+
+  delete txn;
+  delete txn2;
+}
+
+TEST_P(TransactionTest, IteratorTest) {
+  // This test does writes without snapshot validation, and then tries to create
+  // iterator later, which is unsupported in write unprepared.
+  if (txn_db_options.write_policy == WRITE_UNPREPARED) {
+    return;
+  }
+
+  WriteOptions write_options;
+  ReadOptions read_options, snapshot_read_options;
+  std::string value;
+  Status s;
+
+  // Write some keys to the db
+  s = db->Put(write_options, "A", "a");
+  ASSERT_OK(s);
+
+  s = db->Put(write_options, "G", "g");
+  ASSERT_OK(s);
+
+  s = db->Put(write_options, "F", "f");
+  ASSERT_OK(s);
+
+  s = db->Put(write_options, "C", "c");
+  ASSERT_OK(s);
+
+  s = db->Put(write_options, "D", "d");
+  ASSERT_OK(s);
+
+  Transaction* txn = db->BeginTransaction(write_options);
+  ASSERT_TRUE(txn);
+
+  // Write some keys in a txn
+  s = txn->Put("B", "b");
+  ASSERT_OK(s);
+
+  s = txn->Put("H", "h");
+  ASSERT_OK(s);
+
+  s = txn->Delete("D");
+  ASSERT_OK(s);
+
+  s = txn->Put("E", "e");
+  ASSERT_OK(s);
+
+  txn->SetSnapshot();
+  const Snapshot* snapshot = txn->GetSnapshot();
+
+  // Write some keys to the db after the snapshot
+  s = db->Put(write_options, "BB", "xx");
+  ASSERT_OK(s);
+
+  s = db->Put(write_options, "C", "xx");
+  ASSERT_OK(s);
+
+  read_options.snapshot = snapshot;
+  Iterator* iter = txn->GetIterator(read_options);
+  ASSERT_OK(iter->status());
+  iter->SeekToFirst();
+
+  // Read all keys via iter and lock them all
+  std::string results[] = {"a", "b", "c", "e", "f", "g", "h"};
+  for (int i = 0; i < 7; i++) {
+    ASSERT_OK(iter->status());
+    ASSERT_TRUE(iter->Valid());
+    ASSERT_EQ(results[i], iter->value().ToString());
+
+    s = txn->GetForUpdate(read_options, iter->key(), nullptr);
+    if (i == 2) {
+      // "C" was modified after txn's snapshot
+      ASSERT_TRUE(s.IsBusy());
+    } else {
+      ASSERT_OK(s);
+    }
+
+    iter->Next();
+  }
+  ASSERT_FALSE(iter->Valid());
+
+  iter->Seek("G");
+  ASSERT_OK(iter->status());
+  ASSERT_TRUE(iter->Valid());
+  ASSERT_EQ("g", iter->value().ToString());
+
+  iter->Prev();
+  ASSERT_OK(iter->status());
+  ASSERT_TRUE(iter->Valid());
+  ASSERT_EQ("f", iter->value().ToString());
+
+  iter->Seek("D");
+  ASSERT_OK(iter->status());
+  ASSERT_TRUE(iter->Valid());
+  ASSERT_EQ("e", iter->value().ToString());
+
+  iter->Seek("C");
+  ASSERT_OK(iter->status());
+  ASSERT_TRUE(iter->Valid());
+  ASSERT_EQ("c", iter->value().ToString());
+
+  iter->Next();
+  ASSERT_OK(iter->status());
+  ASSERT_TRUE(iter->Valid());
+  ASSERT_EQ("e", iter->value().ToString());
+
+  iter->Seek("");
+  ASSERT_OK(iter->status());
+  ASSERT_TRUE(iter->Valid());
+  ASSERT_EQ("a", iter->value().ToString());
+
+  iter->Seek("X");
+  ASSERT_OK(iter->status());
+  ASSERT_FALSE(iter->Valid());
+
+  iter->SeekToLast();
+  ASSERT_OK(iter->status());
+  ASSERT_TRUE(iter->Valid());
+  ASSERT_EQ("h", iter->value().ToString());
+
+  s = txn->Commit();
+  ASSERT_OK(s);
+
+  delete iter;
+  delete txn;
+}
+
+TEST_P(TransactionTest, DisableIndexingTest) {
+  // Skip this test for write unprepared. It does not solely rely on WBWI for
+  // read your own writes, so depending on whether batches are flushed or not,
+  // only some writes will be visible.
+  //
+  // Also, write unprepared does not support creating iterators if there has
+  // been txn->Put() without snapshot validation.
+  if (txn_db_options.write_policy == WRITE_UNPREPARED) {
+    return;
+  }
+
+  WriteOptions write_options;
+  ReadOptions read_options;
+  std::string value;
+  Status s;
+
+  Transaction* txn = db->BeginTransaction(write_options);
+  ASSERT_TRUE(txn);
+
+  s = txn->Put("A", "a");
+  ASSERT_OK(s);
+
+  s = txn->Get(read_options, "A", &value);
+  ASSERT_OK(s);
+  ASSERT_EQ("a", value);
+
+  txn->DisableIndexing();
+
+  s = txn->Put("B", "b");
+  ASSERT_OK(s);
+
+  s = txn->Get(read_options, "B", &value);
+  ASSERT_TRUE(s.IsNotFound());
+
+  Iterator* iter = txn->GetIterator(read_options);
+  ASSERT_OK(iter->status());
+
+  iter->Seek("B");
+  ASSERT_OK(iter->status());
+  ASSERT_FALSE(iter->Valid());
+
+  s = txn->Delete("A");
+
+  s = txn->Get(read_options, "A", &value);
+  ASSERT_OK(s);
+  ASSERT_EQ("a", value);
+
+  txn->EnableIndexing();
+
+  s = txn->Put("B", "bb");
+  ASSERT_OK(s);
+
+  iter->Seek("B");
+  ASSERT_OK(iter->status());
+  ASSERT_TRUE(iter->Valid());
+  ASSERT_EQ("bb", iter->value().ToString());
+
+  s = txn->Get(read_options, "B", &value);
+  ASSERT_OK(s);
+  ASSERT_EQ("bb", value);
+
+  s = txn->Put("A", "aa");
+  ASSERT_OK(s);
+
+  s = txn->Get(read_options, "A", &value);
+  ASSERT_OK(s);
+  ASSERT_EQ("aa", value);
+
+  delete iter;
+  delete txn;
+}
+
+TEST_P(TransactionTest, SavepointTest) {
+  WriteOptions write_options;
+  ReadOptions read_options, snapshot_read_options;
+  std::string value;
+  Status s;
+
+  Transaction* txn = db->BeginTransaction(write_options);
+  ASSERT_TRUE(txn);
+
+  ASSERT_EQ(0, txn->GetNumPuts());
+
+  s = txn->RollbackToSavePoint();
+  ASSERT_TRUE(s.IsNotFound());
+
+  txn->SetSavePoint();  // 1
+
+  ASSERT_OK(txn->RollbackToSavePoint());  // Rollback to beginning of txn
+  s = txn->RollbackToSavePoint();
+  ASSERT_TRUE(s.IsNotFound());
+
+  s = txn->Put("B", "b");
+  ASSERT_OK(s);
+
+  ASSERT_EQ(1, txn->GetNumPuts());
+  ASSERT_EQ(0, txn->GetNumDeletes());
+
+  s = txn->Commit();
+  ASSERT_OK(s);
+
+  s = db->Get(read_options, "B", &value);
+  ASSERT_OK(s);
+  ASSERT_EQ("b", value);
+
+  delete txn;
+  txn = db->BeginTransaction(write_options);
+  ASSERT_TRUE(txn);
+
+  s = txn->Put("A", "a");
+  ASSERT_OK(s);
+
+  s = txn->Put("B", "bb");
+  ASSERT_OK(s);
+
+  s = txn->Put("C", "c");
+  ASSERT_OK(s);
+
+  txn->SetSavePoint();  // 2
+
+  s = txn->Delete("B");
+  ASSERT_OK(s);
+
+  s = txn->Put("C", "cc");
+  ASSERT_OK(s);
+
+  s = txn->Put("D", "d");
+  ASSERT_OK(s);
+
+  ASSERT_EQ(5, txn->GetNumPuts());
+  ASSERT_EQ(1, txn->GetNumDeletes());
+
+  ASSERT_OK(txn->RollbackToSavePoint());  // Rollback to 2
+
+  ASSERT_EQ(3, txn->GetNumPuts());
+  ASSERT_EQ(0, txn->GetNumDeletes());
+
+  s = txn->Get(read_options, "A", &value);
+  ASSERT_OK(s);
+  ASSERT_EQ("a", value);
+
+  s = txn->Get(read_options, "B", &value);
+  ASSERT_OK(s);
+  ASSERT_EQ("bb", value);
+
+  s = txn->Get(read_options, "C", &value);
+  ASSERT_OK(s);
+  ASSERT_EQ("c", value);
+
+  s = txn->Get(read_options, "D", &value);
+  ASSERT_TRUE(s.IsNotFound());
+
+  s = txn->Put("A", "a");
+  ASSERT_OK(s);
+
+  s = txn->Put("E", "e");
+  ASSERT_OK(s);
+
+  ASSERT_EQ(5, txn->GetNumPuts());
+  ASSERT_EQ(0, txn->GetNumDeletes());
+
+  // Rollback to beginning of txn
+  s = txn->RollbackToSavePoint();
+  ASSERT_TRUE(s.IsNotFound());
+  ASSERT_OK(txn->Rollback());
+
+  ASSERT_EQ(0, txn->GetNumPuts());
+  ASSERT_EQ(0, txn->GetNumDeletes());
+
+  s = txn->Get(read_options, "A", &value);
+  ASSERT_TRUE(s.IsNotFound());
+
+  s = txn->Get(read_options, "B", &value);
+  ASSERT_OK(s);
+  ASSERT_EQ("b", value);
+
+  s = txn->Get(read_options, "D", &value);
+  ASSERT_TRUE(s.IsNotFound());
+
+  s = txn->Get(read_options, "D", &value);
+  ASSERT_TRUE(s.IsNotFound());
+
+  s = txn->Get(read_options, "E", &value);
+  ASSERT_TRUE(s.IsNotFound());
+
+  s = txn->Put("A", "aa");
+  ASSERT_OK(s);
+
+  s = txn->Put("F", "f");
+  ASSERT_OK(s);
+
+  ASSERT_EQ(2, txn->GetNumPuts());
+  ASSERT_EQ(0, txn->GetNumDeletes());
+
+  txn->SetSavePoint();  // 3
+  txn->SetSavePoint();  // 4
+
+  s = txn->Put("G", "g");
+  ASSERT_OK(s);
+
+  s = txn->SingleDelete("F");
+  ASSERT_OK(s);
+
+  s = txn->Delete("B");
+  ASSERT_OK(s);
+
+  s = txn->Get(read_options, "A", &value);
+  ASSERT_OK(s);
+  ASSERT_EQ("aa", value);
+
+  s = txn->Get(read_options, "F", &value);
+  // According to db.h, doing a SingleDelete on a key that has been
+  // overwritten will have undefinied behavior.  So it is unclear what the
+  // result of fetching "F" should be. The current implementation will
+  // return NotFound in this case.
+  ASSERT_TRUE(s.IsNotFound());
+
+  s = txn->Get(read_options, "B", &value);
+  ASSERT_TRUE(s.IsNotFound());
+
+  ASSERT_EQ(3, txn->GetNumPuts());
+  ASSERT_EQ(2, txn->GetNumDeletes());
+
+  ASSERT_OK(txn->RollbackToSavePoint());  // Rollback to 3
+
+  ASSERT_EQ(2, txn->GetNumPuts());
+  ASSERT_EQ(0, txn->GetNumDeletes());
+
+  s = txn->Get(read_options, "F", &value);
+  ASSERT_OK(s);
+  ASSERT_EQ("f", value);
+
+  s = txn->Get(read_options, "G", &value);
+  ASSERT_TRUE(s.IsNotFound());
+
+  s = txn->Commit();
+  ASSERT_OK(s);
+
+  s = db->Get(read_options, "F", &value);
+  ASSERT_OK(s);
+  ASSERT_EQ("f", value);
+
+  s = db->Get(read_options, "G", &value);
+  ASSERT_TRUE(s.IsNotFound());
+
+  s = db->Get(read_options, "A", &value);
+  ASSERT_OK(s);
+  ASSERT_EQ("aa", value);
+
+  s = db->Get(read_options, "B", &value);
+  ASSERT_OK(s);
+  ASSERT_EQ("b", value);
+
+  s = db->Get(read_options, "C", &value);
+  ASSERT_TRUE(s.IsNotFound());
+
+  s = db->Get(read_options, "D", &value);
+  ASSERT_TRUE(s.IsNotFound());
+
+  s = db->Get(read_options, "E", &value);
+  ASSERT_TRUE(s.IsNotFound());
+
+  delete txn;
+}
+
+TEST_P(TransactionTest, SavepointTest2) {
+  WriteOptions write_options;
+  ReadOptions read_options;
+  TransactionOptions txn_options;
+  Status s;
+
+  txn_options.lock_timeout = 1;  // 1 ms
+  Transaction* txn1 = db->BeginTransaction(write_options, txn_options);
+  ASSERT_TRUE(txn1);
+
+  s = txn1->Put("A", "");
+  ASSERT_OK(s);
+
+  txn1->SetSavePoint();  // 1
+
+  s = txn1->Put("A", "a");
+  ASSERT_OK(s);
+
+  s = txn1->Put("C", "c");
+  ASSERT_OK(s);
+
+  txn1->SetSavePoint();  // 2
+
+  s = txn1->Put("A", "a");
+  ASSERT_OK(s);
+  s = txn1->Put("B", "b");
+  ASSERT_OK(s);
+
+  ASSERT_OK(txn1->RollbackToSavePoint());  // Rollback to 2
+
+  // Verify that "A" and "C" is still locked while "B" is not
+  Transaction* txn2 = db->BeginTransaction(write_options, txn_options);
+  ASSERT_TRUE(txn2);
+
+  s = txn2->Put("A", "a2");
+  ASSERT_TRUE(s.IsTimedOut());
+  s = txn2->Put("C", "c2");
+  ASSERT_TRUE(s.IsTimedOut());
+  s = txn2->Put("B", "b2");
+  ASSERT_OK(s);
+
+  s = txn1->Put("A", "aa");
+  ASSERT_OK(s);
+  s = txn1->Put("B", "bb");
+  ASSERT_TRUE(s.IsTimedOut());
+
+  s = txn2->Commit();
+  ASSERT_OK(s);
+  delete txn2;
+
+  s = txn1->Put("A", "aaa");
+  ASSERT_OK(s);
+  s = txn1->Put("B", "bbb");
+  ASSERT_OK(s);
+  s = txn1->Put("C", "ccc");
+  ASSERT_OK(s);
+
+  txn1->SetSavePoint();                    // 3
+  ASSERT_OK(txn1->RollbackToSavePoint());  // Rollback to 3
+
+  // Verify that "A", "B", "C" are still locked
+  txn2 = db->BeginTransaction(write_options, txn_options);
+  ASSERT_TRUE(txn2);
+
+  s = txn2->Put("A", "a2");
+  ASSERT_TRUE(s.IsTimedOut());
+  s = txn2->Put("B", "b2");
+  ASSERT_TRUE(s.IsTimedOut());
+  s = txn2->Put("C", "c2");
+  ASSERT_TRUE(s.IsTimedOut());
+
+  ASSERT_OK(txn1->RollbackToSavePoint());  // Rollback to 1
+
+  // Verify that only "A" is locked
+  s = txn2->Put("A", "a3");
+  ASSERT_TRUE(s.IsTimedOut());
+  s = txn2->Put("B", "b3");
+  ASSERT_OK(s);
+  s = txn2->Put("C", "c3po");
+  ASSERT_OK(s);
+
+  s = txn1->Commit();
+  ASSERT_OK(s);
+  delete txn1;
+
+  // Verify "A" "C" "B" are no longer locked
+  s = txn2->Put("A", "a4");
+  ASSERT_OK(s);
+  s = txn2->Put("B", "b4");
+  ASSERT_OK(s);
+  s = txn2->Put("C", "c4");
+  ASSERT_OK(s);
+
+  s = txn2->Commit();
+  ASSERT_OK(s);
+  delete txn2;
+}
+
+TEST_P(TransactionTest, SavepointTest3) {
+  WriteOptions write_options;
+  ReadOptions read_options;
+  TransactionOptions txn_options;
+  Status s;
+
+  txn_options.lock_timeout = 1;  // 1 ms
+  Transaction* txn1 = db->BeginTransaction(write_options, txn_options);
+  ASSERT_TRUE(txn1);
+
+  s = txn1->PopSavePoint();  // No SavePoint present
+  ASSERT_TRUE(s.IsNotFound());
+
+  s = txn1->Put("A", "");
+  ASSERT_OK(s);
+
+  s = txn1->PopSavePoint();  // Still no SavePoint present
+  ASSERT_TRUE(s.IsNotFound());
+
+  txn1->SetSavePoint();  // 1
+
+  s = txn1->Put("A", "a");
+  ASSERT_OK(s);
+
+  s = txn1->PopSavePoint();  // Remove 1
+  ASSERT_TRUE(txn1->RollbackToSavePoint().IsNotFound());
+
+  // Verify that "A" is still locked
+  Transaction* txn2 = db->BeginTransaction(write_options, txn_options);
+  ASSERT_TRUE(txn2);
+
+  s = txn2->Put("A", "a2");
+  ASSERT_TRUE(s.IsTimedOut());
+  delete txn2;
+
+  txn1->SetSavePoint();  // 2
+
+  s = txn1->Put("B", "b");
+  ASSERT_OK(s);
+
+  txn1->SetSavePoint();  // 3
+
+  s = txn1->Put("B", "b2");
+  ASSERT_OK(s);
+
+  ASSERT_OK(txn1->RollbackToSavePoint());  // Roll back to 2
+
+  s = txn1->PopSavePoint();
+  ASSERT_OK(s);
+
+  s = txn1->PopSavePoint();
+  ASSERT_TRUE(s.IsNotFound());
+
+  s = txn1->Commit();
+  ASSERT_OK(s);
+  delete txn1;
+
+  std::string value;
+
+  // tnx1 should have modified "A" to "a"
+  s = db->Get(read_options, "A", &value);
+  ASSERT_OK(s);
+  ASSERT_EQ("a", value);
+
+  // tnx1 should have set "B" to just "b"
+  s = db->Get(read_options, "B", &value);
+  ASSERT_OK(s);
+  ASSERT_EQ("b", value);
+
+  s = db->Get(read_options, "C", &value);
+  ASSERT_TRUE(s.IsNotFound());
+}
+
+TEST_P(TransactionTest, SavepointTest4) {
+  WriteOptions write_options;
+  ReadOptions read_options;
+  TransactionOptions txn_options;
+  Status s;
+
+  txn_options.lock_timeout = 1;  // 1 ms
+  Transaction* txn1 = db->BeginTransaction(write_options, txn_options);
+  ASSERT_TRUE(txn1);
+
+  txn1->SetSavePoint();  // 1
+  s = txn1->Put("A", "a");
+  ASSERT_OK(s);
+
+  txn1->SetSavePoint();  // 2
+  s = txn1->Put("B", "b");
+  ASSERT_OK(s);
+
+  s = txn1->PopSavePoint();  // Remove 2
+  ASSERT_OK(s);
+
+  // Verify that A/B still exists.
+  std::string value;
+  ASSERT_OK(txn1->Get(read_options, "A", &value));
+  ASSERT_EQ("a", value);
+
+  ASSERT_OK(txn1->Get(read_options, "B", &value));
+  ASSERT_EQ("b", value);
+
+  ASSERT_OK(txn1->RollbackToSavePoint());  // Rollback to 1
+
+  // Verify that everything was rolled back.
+  s = txn1->Get(read_options, "A", &value);
+  ASSERT_TRUE(s.IsNotFound());
+
+  s = txn1->Get(read_options, "B", &value);
+  ASSERT_TRUE(s.IsNotFound());
+
+  // Nothing should be locked
+  Transaction* txn2 = db->BeginTransaction(write_options, txn_options);
+  ASSERT_TRUE(txn2);
+
+  s = txn2->Put("A", "");
+  ASSERT_OK(s);
+
+  s = txn2->Put("B", "");
+  ASSERT_OK(s);
+
+  delete txn2;
+  delete txn1;
+}
+
+TEST_P(TransactionTest, UndoGetForUpdateTest) {
+  WriteOptions write_options;
+  ReadOptions read_options;
+  TransactionOptions txn_options;
+  std::string value;
+  Status s;
+
+  txn_options.lock_timeout = 1;  // 1 ms
+  Transaction* txn1 = db->BeginTransaction(write_options, txn_options);
+  ASSERT_TRUE(txn1);
+
+  txn1->UndoGetForUpdate("A");
+
+  s = txn1->Commit();
+  ASSERT_OK(s);
+  delete txn1;
+
+  txn1 = db->BeginTransaction(write_options, txn_options);
+
+  txn1->UndoGetForUpdate("A");
+  s = txn1->GetForUpdate(read_options, "A", &value);
+  ASSERT_TRUE(s.IsNotFound());
+
+  // Verify that A is locked
+  Transaction* txn2 = db->BeginTransaction(write_options, txn_options);
+  s = txn2->Put("A", "a");
+  ASSERT_TRUE(s.IsTimedOut());
+
+  txn1->UndoGetForUpdate("A");
+
+  // Verify that A is now unlocked
+  s = txn2->Put("A", "a2");
+  ASSERT_OK(s);
+  ASSERT_OK(txn2->Commit());
+  delete txn2;
+  s = db->Get(read_options, "A", &value);
+  ASSERT_OK(s);
+  ASSERT_EQ("a2", value);
+
+  s = txn1->Delete("A");
+  ASSERT_OK(s);
+  s = txn1->GetForUpdate(read_options, "A", &value);
+  ASSERT_TRUE(s.IsNotFound());
+
+  s = txn1->Put("B", "b3");
+  ASSERT_OK(s);
+  s = txn1->GetForUpdate(read_options, "B", &value);
+  ASSERT_OK(s);
+
+  txn1->UndoGetForUpdate("A");
+  txn1->UndoGetForUpdate("B");
+
+  // Verify that A and B are still locked
+  txn2 = db->BeginTransaction(write_options, txn_options);
+  s = txn2->Put("A", "a4");
+  ASSERT_TRUE(s.IsTimedOut());
+  s = txn2->Put("B", "b4");
+  ASSERT_TRUE(s.IsTimedOut());
+
+  ASSERT_OK(txn1->Rollback());
+  delete txn1;
+
+  // Verify that A and B are no longer locked
+  s = txn2->Put("A", "a5");
+  ASSERT_OK(s);
+  s = txn2->Put("B", "b5");
+  ASSERT_OK(s);
+  s = txn2->Commit();
+  delete txn2;
+  ASSERT_OK(s);
+
+  txn1 = db->BeginTransaction(write_options, txn_options);
+
+  s = txn1->GetForUpdate(read_options, "A", &value);
+  ASSERT_OK(s);
+  s = txn1->GetForUpdate(read_options, "A", &value);
+  ASSERT_OK(s);
+  s = txn1->GetForUpdate(read_options, "C", &value);
+  ASSERT_TRUE(s.IsNotFound());
+  s = txn1->GetForUpdate(read_options, "A", &value);
+  ASSERT_OK(s);
+  s = txn1->GetForUpdate(read_options, "C", &value);
+  ASSERT_TRUE(s.IsNotFound());
+  s = txn1->GetForUpdate(read_options, "B", &value);
+  ASSERT_OK(s);
+  s = txn1->Put("B", "b5");
+  s = txn1->GetForUpdate(read_options, "B", &value);
+  ASSERT_OK(s);
+
+  txn1->UndoGetForUpdate("A");
+  txn1->UndoGetForUpdate("B");
+  txn1->UndoGetForUpdate("C");
+  txn1->UndoGetForUpdate("X");
+
+  // Verify A,B,C are locked
+  txn2 = db->BeginTransaction(write_options, txn_options);
+  s = txn2->Put("A", "a6");
+  ASSERT_TRUE(s.IsTimedOut());
+  s = txn2->Delete("B");
+  ASSERT_TRUE(s.IsTimedOut());
+  s = txn2->Put("C", "c6");
+  ASSERT_TRUE(s.IsTimedOut());
+  s = txn2->Put("X", "x6");
+  ASSERT_OK(s);
+
+  txn1->UndoGetForUpdate("A");
+  txn1->UndoGetForUpdate("B");
+  txn1->UndoGetForUpdate("C");
+  txn1->UndoGetForUpdate("X");
+
+  // Verify A,B are locked and C is not
+  s = txn2->Put("A", "a6");
+  ASSERT_TRUE(s.IsTimedOut());
+  s = txn2->Delete("B");
+  ASSERT_TRUE(s.IsTimedOut());
+  s = txn2->Put("C", "c6");
+  ASSERT_OK(s);
+  s = txn2->Put("X", "x6");
+  ASSERT_OK(s);
+
+  txn1->UndoGetForUpdate("A");
+  txn1->UndoGetForUpdate("B");
+  txn1->UndoGetForUpdate("C");
+  txn1->UndoGetForUpdate("X");
+
+  // Verify B is locked and A and C are not
+  s = txn2->Put("A", "a7");
+  ASSERT_OK(s);
+  s = txn2->Delete("B");
+  ASSERT_TRUE(s.IsTimedOut());
+  s = txn2->Put("C", "c7");
+  ASSERT_OK(s);
+  s = txn2->Put("X", "x7");
+  ASSERT_OK(s);
+
+  s = txn2->Commit();
+  ASSERT_OK(s);
+  delete txn2;
+
+  s = txn1->Commit();
+  ASSERT_OK(s);
+  delete txn1;
+}
+
+TEST_P(TransactionTest, UndoGetForUpdateTest2) {
+  WriteOptions write_options;
+  ReadOptions read_options;
+  TransactionOptions txn_options;
+  std::string value;
+  Status s;
+
+  s = db->Put(write_options, "A", "");
+  ASSERT_OK(s);
+
+  txn_options.lock_timeout = 1;  // 1 ms
+  Transaction* txn1 = db->BeginTransaction(write_options, txn_options);
+  ASSERT_TRUE(txn1);
+
+  s = txn1->GetForUpdate(read_options, "A", &value);
+  ASSERT_OK(s);
+  s = txn1->GetForUpdate(read_options, "B", &value);
+  ASSERT_TRUE(s.IsNotFound());
+
+  s = txn1->Put("F", "f");
+  ASSERT_OK(s);
+
+  txn1->SetSavePoint();  // 1
+
+  txn1->UndoGetForUpdate("A");
+
+  s = txn1->GetForUpdate(read_options, "C", &value);
+  ASSERT_TRUE(s.IsNotFound());
+  s = txn1->GetForUpdate(read_options, "D", &value);
+  ASSERT_TRUE(s.IsNotFound());
+
+  s = txn1->Put("E", "e");
+  ASSERT_OK(s);
+  s = txn1->GetForUpdate(read_options, "E", &value);
+  ASSERT_OK(s);
+
+  s = txn1->GetForUpdate(read_options, "F", &value);
+  ASSERT_OK(s);
+
+  // Verify A,B,C,D,E,F are still locked
+  Transaction* txn2 = db->BeginTransaction(write_options, txn_options);
+  s = txn2->Put("A", "a1");
+  ASSERT_TRUE(s.IsTimedOut());
+  s = txn2->Put("B", "b1");
+  ASSERT_TRUE(s.IsTimedOut());
+  s = txn2->Put("C", "c1");
+  ASSERT_TRUE(s.IsTimedOut());
+  s = txn2->Put("D", "d1");
+  ASSERT_TRUE(s.IsTimedOut());
+  s = txn2->Put("E", "e1");
+  ASSERT_TRUE(s.IsTimedOut());
+  s = txn2->Put("F", "f1");
+  ASSERT_TRUE(s.IsTimedOut());
+
+  txn1->UndoGetForUpdate("C");
+  txn1->UndoGetForUpdate("E");
+
+  // Verify A,B,D,E,F are still locked and C is not.
+  s = txn2->Put("A", "a2");
+  ASSERT_TRUE(s.IsTimedOut());
+  s = txn2->Put("B", "b2");
+  ASSERT_TRUE(s.IsTimedOut());
+  s = txn2->Put("D", "d2");
+  ASSERT_TRUE(s.IsTimedOut());
+  s = txn2->Put("E", "e2");
+  ASSERT_TRUE(s.IsTimedOut());
+  s = txn2->Put("F", "f2");
+  ASSERT_TRUE(s.IsTimedOut());
+  s = txn2->Put("C", "c2");
+  ASSERT_OK(s);
+
+  txn1->SetSavePoint();  // 2
+
+  s = txn1->Put("H", "h");
+  ASSERT_OK(s);
+
+  txn1->UndoGetForUpdate("A");
+  txn1->UndoGetForUpdate("B");
+  txn1->UndoGetForUpdate("C");
+  txn1->UndoGetForUpdate("D");
+  txn1->UndoGetForUpdate("E");
+  txn1->UndoGetForUpdate("F");
+  txn1->UndoGetForUpdate("G");
+  txn1->UndoGetForUpdate("H");
+
+  // Verify A,B,D,E,F,H are still locked and C,G are not.
+  s = txn2->Put("A", "a3");
+  ASSERT_TRUE(s.IsTimedOut());
+  s = txn2->Put("B", "b3");
+  ASSERT_TRUE(s.IsTimedOut());
+  s = txn2->Put("D", "d3");
+  ASSERT_TRUE(s.IsTimedOut());
+  s = txn2->Put("E", "e3");
+  ASSERT_TRUE(s.IsTimedOut());
+  s = txn2->Put("F", "f3");
+  ASSERT_TRUE(s.IsTimedOut());
+  s = txn2->Put("H", "h3");
+  ASSERT_TRUE(s.IsTimedOut());
+  s = txn2->Put("C", "c3");
+  ASSERT_OK(s);
+  s = txn2->Put("G", "g3");
+  ASSERT_OK(s);
+
+  ASSERT_OK(txn1->RollbackToSavePoint());  // rollback to 2
+
+  // Verify A,B,D,E,F are still locked and C,G,H are not.
+  s = txn2->Put("A", "a3");
+  ASSERT_TRUE(s.IsTimedOut());
+  s = txn2->Put("B", "b3");
+  ASSERT_TRUE(s.IsTimedOut());
+  s = txn2->Put("D", "d3");
+  ASSERT_TRUE(s.IsTimedOut());
+  s = txn2->Put("E", "e3");
+  ASSERT_TRUE(s.IsTimedOut());
+  s = txn2->Put("F", "f3");
+  ASSERT_TRUE(s.IsTimedOut());
+  s = txn2->Put("C", "c3");
+  ASSERT_OK(s);
+  s = txn2->Put("G", "g3");
+  ASSERT_OK(s);
+  s = txn2->Put("H", "h3");
+  ASSERT_OK(s);
+
+  txn1->UndoGetForUpdate("A");
+  txn1->UndoGetForUpdate("B");
+  txn1->UndoGetForUpdate("C");
+  txn1->UndoGetForUpdate("D");
+  txn1->UndoGetForUpdate("E");
+  txn1->UndoGetForUpdate("F");
+  txn1->UndoGetForUpdate("G");
+  txn1->UndoGetForUpdate("H");
+
+  // Verify A,B,E,F are still locked and C,D,G,H are not.
+  s = txn2->Put("A", "a3");
+  ASSERT_TRUE(s.IsTimedOut());
+  s = txn2->Put("B", "b3");
+  ASSERT_TRUE(s.IsTimedOut());
+  s = txn2->Put("E", "e3");
+  ASSERT_TRUE(s.IsTimedOut());
+  s = txn2->Put("F", "f3");
+  ASSERT_TRUE(s.IsTimedOut());
+  s = txn2->Put("C", "c3");
+  ASSERT_OK(s);
+  s = txn2->Put("D", "d3");
+  ASSERT_OK(s);
+  s = txn2->Put("G", "g3");
+  ASSERT_OK(s);
+  s = txn2->Put("H", "h3");
+  ASSERT_OK(s);
+
+  ASSERT_OK(txn1->RollbackToSavePoint());  // rollback to 1
+
+  // Verify A,B,F are still locked and C,D,E,G,H are not.
+  s = txn2->Put("A", "a3");
+  ASSERT_TRUE(s.IsTimedOut());
+  s = txn2->Put("B", "b3");
+  ASSERT_TRUE(s.IsTimedOut());
+  s = txn2->Put("F", "f3");
+  ASSERT_TRUE(s.IsTimedOut());
+  s = txn2->Put("C", "c3");
+  ASSERT_OK(s);
+  s = txn2->Put("D", "d3");
+  ASSERT_OK(s);
+  s = txn2->Put("E", "e3");
+  ASSERT_OK(s);
+  s = txn2->Put("G", "g3");
+  ASSERT_OK(s);
+  s = txn2->Put("H", "h3");
+  ASSERT_OK(s);
+
+  txn1->UndoGetForUpdate("A");
+  txn1->UndoGetForUpdate("B");
+  txn1->UndoGetForUpdate("C");
+  txn1->UndoGetForUpdate("D");
+  txn1->UndoGetForUpdate("E");
+  txn1->UndoGetForUpdate("F");
+  txn1->UndoGetForUpdate("G");
+  txn1->UndoGetForUpdate("H");
+
+  // Verify F is still locked and A,B,C,D,E,G,H are not.
+  s = txn2->Put("F", "f3");
+  ASSERT_TRUE(s.IsTimedOut());
+  s = txn2->Put("A", "a3");
+  ASSERT_OK(s);
+  s = txn2->Put("B", "b3");
+  ASSERT_OK(s);
+  s = txn2->Put("C", "c3");
+  ASSERT_OK(s);
+  s = txn2->Put("D", "d3");
+  ASSERT_OK(s);
+  s = txn2->Put("E", "e3");
+  ASSERT_OK(s);
+  s = txn2->Put("G", "g3");
+  ASSERT_OK(s);
+  s = txn2->Put("H", "h3");
+  ASSERT_OK(s);
+
+  s = txn1->Commit();
+  ASSERT_OK(s);
+  s = txn2->Commit();
+  ASSERT_OK(s);
+
+  delete txn1;
+  delete txn2;
+}
+
+TEST_P(TransactionTest, TimeoutTest) {
+  WriteOptions write_options;
+  ReadOptions read_options;
+  std::string value;
+  Status s;
+
+  delete db;
+  db = nullptr;
+
+  // transaction writes have an infinite timeout,
+  // but we will override this when we start a txn
+  // db writes have infinite timeout
+  txn_db_options.transaction_lock_timeout = -1;
+  txn_db_options.default_lock_timeout = -1;
+
+  s = TransactionDB::Open(options, txn_db_options, dbname, &db);
+  assert(db != nullptr);
+  ASSERT_OK(s);
+
+  s = db->Put(write_options, "aaa", "aaa");
+  ASSERT_OK(s);
+
+  TransactionOptions txn_options0;
+  txn_options0.expiration = 100;   // 100ms
+  txn_options0.lock_timeout = 50;  // txn timeout no longer infinite
+  Transaction* txn1 = db->BeginTransaction(write_options, txn_options0);
+
+  s = txn1->GetForUpdate(read_options, "aaa", nullptr);
+  ASSERT_OK(s);
+
+  // Conflicts with previous GetForUpdate.
+  // Since db writes do not have a timeout, this should eventually succeed when
+  // the transaction expires.
+  s = db->Put(write_options, "aaa", "xxx");
+  ASSERT_OK(s);
+
+  ASSERT_GE(txn1->GetElapsedTime(),
+            static_cast<uint64_t>(txn_options0.expiration));
+
+  s = txn1->Commit();
+  ASSERT_TRUE(s.IsExpired());  // expired!
+
+  s = db->Get(read_options, "aaa", &value);
+  ASSERT_OK(s);
+  ASSERT_EQ("xxx", value);
+
+  delete txn1;
+  delete db;
+
+  // transaction writes have 10ms timeout,
+  // db writes have infinite timeout
+  txn_db_options.transaction_lock_timeout = 50;
+  txn_db_options.default_lock_timeout = -1;
+
+  s = TransactionDB::Open(options, txn_db_options, dbname, &db);
+  ASSERT_OK(s);
+
+  s = db->Put(write_options, "aaa", "aaa");
+  ASSERT_OK(s);
+
+  TransactionOptions txn_options;
+  txn_options.expiration = 100;  // 100ms
+  txn1 = db->BeginTransaction(write_options, txn_options);
+
+  s = txn1->GetForUpdate(read_options, "aaa", nullptr);
+  ASSERT_OK(s);
+
+  // Conflicts with previous GetForUpdate.
+  // Since db writes do not have a timeout, this should eventually succeed when
+  // the transaction expires.
+  s = db->Put(write_options, "aaa", "xxx");
+  ASSERT_OK(s);
+
+  s = txn1->Commit();
+  ASSERT_NOK(s);  // expired!
+
+  s = db->Get(read_options, "aaa", &value);
+  ASSERT_OK(s);
+  ASSERT_EQ("xxx", value);
+
+  delete txn1;
+  txn_options.expiration = 6000000;  // 100 minutes
+  txn_options.lock_timeout = 1;      // 1ms
+  txn1 = db->BeginTransaction(write_options, txn_options);
+  txn1->SetLockTimeout(100);
+
+  TransactionOptions txn_options2;
+  txn_options2.expiration = 10;  // 10ms
+  Transaction* txn2 = db->BeginTransaction(write_options, txn_options2);
+  ASSERT_OK(s);
+
+  s = txn2->Put("a", "2");
+  ASSERT_OK(s);
+
+  // txn1 has a lock timeout longer than txn2's expiration, so it will win
+  s = txn1->Delete("a");
+  ASSERT_OK(s);
+
+  s = txn1->Commit();
+  ASSERT_OK(s);
+
+  // txn2 should be expired out since txn1 waiting until its timeout expired.
+  s = txn2->Commit();
+  ASSERT_TRUE(s.IsExpired());
+
+  delete txn1;
+  delete txn2;
+  txn_options.expiration = 6000000;  // 100 minutes
+  txn1 = db->BeginTransaction(write_options, txn_options);
+  txn_options2.expiration = 100000000;
+  txn2 = db->BeginTransaction(write_options, txn_options2);
+
+  s = txn1->Delete("asdf");
+  ASSERT_OK(s);
+
+  // txn2 has a smaller lock timeout than txn1's expiration, so it will time out
+  s = txn2->Delete("asdf");
+  ASSERT_TRUE(s.IsTimedOut());
+  ASSERT_EQ(s.ToString(), "Operation timed out: Timeout waiting to lock key");
+
+  s = txn1->Commit();
+  ASSERT_OK(s);
+
+  s = txn2->Put("asdf", "asdf");
+  ASSERT_OK(s);
+
+  s = txn2->Commit();
+  ASSERT_OK(s);
+
+  s = db->Get(read_options, "asdf", &value);
+  ASSERT_OK(s);
+  ASSERT_EQ("asdf", value);
+
+  delete txn1;
+  delete txn2;
+}
+
+TEST_P(TransactionTest, SingleDeleteTest) {
+  WriteOptions write_options;
+  ReadOptions read_options;
+  std::string value;
+  Status s;
+
+  Transaction* txn = db->BeginTransaction(write_options);
+  ASSERT_TRUE(txn);
+
+  s = txn->SingleDelete("A");
+  ASSERT_OK(s);
+
+  s = txn->Get(read_options, "A", &value);
+  ASSERT_TRUE(s.IsNotFound());
+
+  s = txn->Commit();
+  ASSERT_OK(s);
+  delete txn;
+
+  txn = db->BeginTransaction(write_options);
+
+  s = txn->SingleDelete("A");
+  ASSERT_OK(s);
+
+  s = txn->Put("A", "a");
+  ASSERT_OK(s);
+
+  s = txn->Get(read_options, "A", &value);
+  ASSERT_OK(s);
+  ASSERT_EQ("a", value);
+
+  s = txn->Commit();
+  ASSERT_OK(s);
+  delete txn;
+
+  s = db->Get(read_options, "A", &value);
+  ASSERT_OK(s);
+  ASSERT_EQ("a", value);
+
+  txn = db->BeginTransaction(write_options);
+
+  s = txn->SingleDelete("A");
+  ASSERT_OK(s);
+
+  s = txn->Get(read_options, "A", &value);
+  ASSERT_TRUE(s.IsNotFound());
+
+  s = txn->Commit();
+  ASSERT_OK(s);
+  delete txn;
+
+  s = db->Get(read_options, "A", &value);
+  ASSERT_TRUE(s.IsNotFound());
+
+  txn = db->BeginTransaction(write_options);
+  Transaction* txn2 = db->BeginTransaction(write_options);
+  txn2->SetSnapshot();
+
+  s = txn->Put("A", "a");
+  ASSERT_OK(s);
+
+  s = txn->Put("A", "a2");
+  ASSERT_OK(s);
+
+  s = txn->SingleDelete("A");
+  ASSERT_OK(s);
+
+  s = txn->SingleDelete("B");
+  ASSERT_OK(s);
+
+  // According to db.h, doing a SingleDelete on a key that has been
+  // overwritten will have undefinied behavior.  So it is unclear what the
+  // result of fetching "A" should be. The current implementation will
+  // return NotFound in this case.
+  s = txn->Get(read_options, "A", &value);
+  ASSERT_TRUE(s.IsNotFound());
+
+  s = txn2->Put("B", "b");
+  ASSERT_TRUE(s.IsTimedOut());
+  s = txn2->Commit();
+  ASSERT_OK(s);
+  delete txn2;
+
+  s = txn->Commit();
+  ASSERT_OK(s);
+  delete txn;
+
+  // According to db.h, doing a SingleDelete on a key that has been
+  // overwritten will have undefinied behavior.  So it is unclear what the
+  // result of fetching "A" should be. The current implementation will
+  // return NotFound in this case.
+  s = db->Get(read_options, "A", &value);
+  ASSERT_TRUE(s.IsNotFound());
+
+  s = db->Get(read_options, "B", &value);
+  ASSERT_TRUE(s.IsNotFound());
+}
+
+TEST_P(TransactionTest, MergeTest) {
+  WriteOptions write_options;
+  ReadOptions read_options;
+  std::string value;
+  Status s;
+
+  Transaction* txn = db->BeginTransaction(write_options, TransactionOptions());
+  ASSERT_TRUE(txn);
+
+  s = db->Put(write_options, "A", "a0");
+  ASSERT_OK(s);
+
+  s = txn->Merge("A", "1");
+  ASSERT_OK(s);
+
+  s = txn->Merge("A", "2");
+  ASSERT_OK(s);
+
+  s = txn->Get(read_options, "A", &value);
+  ASSERT_OK(s);
+  ASSERT_EQ("a0,1,2", value);
+
+  s = txn->Put("A", "a");
+  ASSERT_OK(s);
+
+  s = txn->Get(read_options, "A", &value);
+  ASSERT_OK(s);
+  ASSERT_EQ("a", value);
+
+  s = txn->Merge("A", "3");
+  ASSERT_OK(s);
+
+  s = txn->Get(read_options, "A", &value);
+  ASSERT_OK(s);
+  ASSERT_EQ("a,3", value);
+
+  TransactionOptions txn_options;
+  txn_options.lock_timeout = 1;  // 1 ms
+  Transaction* txn2 = db->BeginTransaction(write_options, txn_options);
+  ASSERT_TRUE(txn2);
+
+  // verify that txn has "A" locked
+  s = txn2->Merge("A", "4");
+  ASSERT_TRUE(s.IsTimedOut());
+
+  s = txn2->Commit();
+  ASSERT_OK(s);
+  delete txn2;
+
+  s = txn->Commit();
+  ASSERT_OK(s);
+  delete txn;
+
+  s = db->Get(read_options, "A", &value);
+  ASSERT_OK(s);
+  ASSERT_EQ("a,3", value);
+}
+
+TEST_P(TransactionTest, DeleteRangeSupportTest) {
+  // The `DeleteRange()` API is banned everywhere.
+  ASSERT_TRUE(
+      db->DeleteRange(WriteOptions(), db->DefaultColumnFamily(), "a", "b")
+          .IsNotSupported());
+
+  // But range deletions can be added via the `Write()` API by specifying the
+  // proper flags to promise there are no conflicts according to the DB type
+  // (see `TransactionDB::DeleteRange()` API doc for details).
+  for (bool skip_concurrency_control : {false, true}) {
+    for (bool skip_duplicate_key_check : {false, true}) {
+      ASSERT_OK(db->Put(WriteOptions(), "a", "val"));
+      WriteBatch wb;
+      ASSERT_OK(wb.DeleteRange("a", "b"));
+      TransactionDBWriteOptimizations flags;
+      flags.skip_concurrency_control = skip_concurrency_control;
+      flags.skip_duplicate_key_check = skip_duplicate_key_check;
+      Status s = db->Write(WriteOptions(), flags, &wb);
+      std::string value;
+      switch (txn_db_options.write_policy) {
+        case WRITE_COMMITTED:
+          if (skip_concurrency_control) {
+            ASSERT_OK(s);
+            ASSERT_TRUE(db->Get(ReadOptions(), "a", &value).IsNotFound());
+          } else {
+            ASSERT_NOK(s);
+            ASSERT_OK(db->Get(ReadOptions(), "a", &value));
+          }
+          break;
+        case WRITE_PREPARED:
+          // Intentional fall-through
+        case WRITE_UNPREPARED:
+          if (skip_concurrency_control && skip_duplicate_key_check) {
+            ASSERT_OK(s);
+            ASSERT_TRUE(db->Get(ReadOptions(), "a", &value).IsNotFound());
+          } else {
+            ASSERT_NOK(s);
+            ASSERT_OK(db->Get(ReadOptions(), "a", &value));
+          }
+          break;
+      }
+      // Without any promises from the user, range deletion via other `Write()`
+      // APIs are still banned.
+      ASSERT_OK(db->Put(WriteOptions(), "a", "val"));
+      ASSERT_NOK(db->Write(WriteOptions(), &wb));
+      ASSERT_OK(db->Get(ReadOptions(), "a", &value));
+    }
+  }
+}
+
+TEST_P(TransactionTest, DeferSnapshotTest) {
+  WriteOptions write_options;
+  ReadOptions read_options;
+  std::string value;
+  Status s;
+
+  s = db->Put(write_options, "A", "a0");
+  ASSERT_OK(s);
+
+  Transaction* txn1 = db->BeginTransaction(write_options);
+  Transaction* txn2 = db->BeginTransaction(write_options);
+
+  txn1->SetSnapshotOnNextOperation();
+  auto snapshot = txn1->GetSnapshot();
+  ASSERT_FALSE(snapshot);
+
+  s = txn2->Put("A", "a2");
+  ASSERT_OK(s);
+  s = txn2->Commit();
+  ASSERT_OK(s);
+  delete txn2;
+
+  s = txn1->GetForUpdate(read_options, "A", &value);
+  // Should not conflict with txn2 since snapshot wasn't set until
+  // GetForUpdate was called.
+  ASSERT_OK(s);
+  ASSERT_EQ("a2", value);
+
+  s = txn1->Put("A", "a1");
+  ASSERT_OK(s);
+
+  s = db->Put(write_options, "B", "b0");
+  ASSERT_OK(s);
+
+  // Cannot lock B since it was written after the snapshot was set
+  s = txn1->Put("B", "b1");
+  ASSERT_TRUE(s.IsBusy());
+
+  s = txn1->Commit();
+  ASSERT_OK(s);
+  delete txn1;
+
+  s = db->Get(read_options, "A", &value);
+  ASSERT_OK(s);
+  ASSERT_EQ("a1", value);
+
+  s = db->Get(read_options, "B", &value);
+  ASSERT_OK(s);
+  ASSERT_EQ("b0", value);
+}
+
+TEST_P(TransactionTest, DeferSnapshotTest2) {
+  WriteOptions write_options;
+  ReadOptions read_options, snapshot_read_options;
+  std::string value;
+  Status s;
+
+  Transaction* txn1 = db->BeginTransaction(write_options);
+
+  txn1->SetSnapshot();
+
+  s = txn1->Put("A", "a1");
+  ASSERT_OK(s);
+
+  s = db->Put(write_options, "C", "c0");
+  ASSERT_OK(s);
+  s = db->Put(write_options, "D", "d0");
+  ASSERT_OK(s);
+
+  snapshot_read_options.snapshot = txn1->GetSnapshot();
+
+  txn1->SetSnapshotOnNextOperation();
+
+  s = txn1->Get(snapshot_read_options, "C", &value);
+  // Snapshot was set before C was written
+  ASSERT_TRUE(s.IsNotFound());
+  s = txn1->Get(snapshot_read_options, "D", &value);
+  // Snapshot was set before D was written
+  ASSERT_TRUE(s.IsNotFound());
+
+  // Snapshot should not have changed yet.
+  snapshot_read_options.snapshot = txn1->GetSnapshot();
+
+  s = txn1->Get(snapshot_read_options, "C", &value);
+  // Snapshot was set before C was written
+  ASSERT_TRUE(s.IsNotFound());
+  s = txn1->Get(snapshot_read_options, "D", &value);
+  // Snapshot was set before D was written
+  ASSERT_TRUE(s.IsNotFound());
+
+  s = txn1->GetForUpdate(read_options, "C", &value);
+  ASSERT_OK(s);
+  ASSERT_EQ("c0", value);
+
+  s = db->Put(write_options, "D", "d00");
+  ASSERT_OK(s);
+
+  // Snapshot is now set
+  snapshot_read_options.snapshot = txn1->GetSnapshot();
+  s = txn1->Get(snapshot_read_options, "D", &value);
+  ASSERT_OK(s);
+  ASSERT_EQ("d0", value);
+
+  s = txn1->Commit();
+  ASSERT_OK(s);
+  delete txn1;
+}
+
+TEST_P(TransactionTest, DeferSnapshotSavePointTest) {
+  WriteOptions write_options;
+  ReadOptions read_options, snapshot_read_options;
+  std::string value;
+  Status s;
+
+  Transaction* txn1 = db->BeginTransaction(write_options);
+
+  txn1->SetSavePoint();  // 1
+
+  s = db->Put(write_options, "T", "1");
+  ASSERT_OK(s);
+
+  txn1->SetSnapshotOnNextOperation();
+
+  s = db->Put(write_options, "T", "2");
+  ASSERT_OK(s);
+
+  txn1->SetSavePoint();  // 2
+
+  s = db->Put(write_options, "T", "3");
+  ASSERT_OK(s);
+
+  s = txn1->Put("A", "a");
+  ASSERT_OK(s);
+
+  txn1->SetSavePoint();  // 3
+
+  s = db->Put(write_options, "T", "4");
+  ASSERT_OK(s);
+
+  txn1->SetSnapshot();
+  txn1->SetSnapshotOnNextOperation();
+
+  txn1->SetSavePoint();  // 4
+
+  s = db->Put(write_options, "T", "5");
+  ASSERT_OK(s);
+
+  snapshot_read_options.snapshot = txn1->GetSnapshot();
+  s = txn1->Get(snapshot_read_options, "T", &value);
+  ASSERT_OK(s);
+  ASSERT_EQ("4", value);
+
+  s = txn1->Put("A", "a1");
+  ASSERT_OK(s);
+
+  snapshot_read_options.snapshot = txn1->GetSnapshot();
+  s = txn1->Get(snapshot_read_options, "T", &value);
+  ASSERT_OK(s);
+  ASSERT_EQ("5", value);
+
+  s = txn1->RollbackToSavePoint();  // Rollback to 4
+  ASSERT_OK(s);
+
+  snapshot_read_options.snapshot = txn1->GetSnapshot();
+  s = txn1->Get(snapshot_read_options, "T", &value);
+  ASSERT_OK(s);
+  ASSERT_EQ("4", value);
+
+  s = txn1->RollbackToSavePoint();  // Rollback to 3
+  ASSERT_OK(s);
+
+  snapshot_read_options.snapshot = txn1->GetSnapshot();
+  s = txn1->Get(snapshot_read_options, "T", &value);
+  ASSERT_OK(s);
+  ASSERT_EQ("3", value);
+
+  s = txn1->Get(read_options, "T", &value);
+  ASSERT_OK(s);
+  ASSERT_EQ("5", value);
+
+  s = txn1->RollbackToSavePoint();  // Rollback to 2
+  ASSERT_OK(s);
+
+  snapshot_read_options.snapshot = txn1->GetSnapshot();
+  ASSERT_FALSE(snapshot_read_options.snapshot);
+  s = txn1->Get(snapshot_read_options, "T", &value);
+  ASSERT_OK(s);
+  ASSERT_EQ("5", value);
+
+  s = txn1->Delete("A");
+  ASSERT_OK(s);
+
+  snapshot_read_options.snapshot = txn1->GetSnapshot();
+  ASSERT_TRUE(snapshot_read_options.snapshot);
+  s = txn1->Get(snapshot_read_options, "T", &value);
+  ASSERT_OK(s);
+  ASSERT_EQ("5", value);
+
+  s = txn1->RollbackToSavePoint();  // Rollback to 1
+  ASSERT_OK(s);
+
+  s = txn1->Delete("A");
+  ASSERT_OK(s);
+
+  snapshot_read_options.snapshot = txn1->GetSnapshot();
+  ASSERT_FALSE(snapshot_read_options.snapshot);
+  s = txn1->Get(snapshot_read_options, "T", &value);
+  ASSERT_OK(s);
+  ASSERT_EQ("5", value);
+
+  s = txn1->Commit();
+  ASSERT_OK(s);
+
+  delete txn1;
+}
+
+TEST_P(TransactionTest, SetSnapshotOnNextOperationWithNotification) {
+  WriteOptions write_options;
+  ReadOptions read_options;
+  std::string value;
+
+  class Notifier : public TransactionNotifier {
+   private:
+    const Snapshot** snapshot_ptr_;
+
+   public:
+    explicit Notifier(const Snapshot** snapshot_ptr)
+        : snapshot_ptr_(snapshot_ptr) {}
+
+    void SnapshotCreated(const Snapshot* newSnapshot) override {
+      *snapshot_ptr_ = newSnapshot;
+    }
+  };
+
+  std::shared_ptr<Notifier> notifier =
+      std::make_shared<Notifier>(&read_options.snapshot);
+  Status s;
+
+  s = db->Put(write_options, "B", "0");
+  ASSERT_OK(s);
+
+  Transaction* txn1 = db->BeginTransaction(write_options);
+
+  txn1->SetSnapshotOnNextOperation(notifier);
+  ASSERT_FALSE(read_options.snapshot);
+
+  s = db->Put(write_options, "B", "1");
+  ASSERT_OK(s);
+
+  // A Get does not generate the snapshot
+  s = txn1->Get(read_options, "B", &value);
+  ASSERT_OK(s);
+  ASSERT_FALSE(read_options.snapshot);
+  ASSERT_EQ(value, "1");
+
+  // Any other operation does
+  s = txn1->Put("A", "0");
+  ASSERT_OK(s);
+
+  // Now change "B".
+  s = db->Put(write_options, "B", "2");
+  ASSERT_OK(s);
+
+  // The original value should still be read
+  s = txn1->Get(read_options, "B", &value);
+  ASSERT_OK(s);
+  ASSERT_TRUE(read_options.snapshot);
+  ASSERT_EQ(value, "1");
+
+  s = txn1->Commit();
+  ASSERT_OK(s);
+
+  delete txn1;
+}
+
+TEST_P(TransactionTest, ClearSnapshotTest) {
+  WriteOptions write_options;
+  ReadOptions read_options, snapshot_read_options;
+  std::string value;
+  Status s;
+
+  s = db->Put(write_options, "foo", "0");
+  ASSERT_OK(s);
+
+  Transaction* txn = db->BeginTransaction(write_options);
+  ASSERT_TRUE(txn);
+
+  s = db->Put(write_options, "foo", "1");
+  ASSERT_OK(s);
+
+  snapshot_read_options.snapshot = txn->GetSnapshot();
+  ASSERT_FALSE(snapshot_read_options.snapshot);
+
+  // No snapshot created yet
+  s = txn->Get(snapshot_read_options, "foo", &value);
+  ASSERT_EQ(value, "1");
+
+  txn->SetSnapshot();
+  snapshot_read_options.snapshot = txn->GetSnapshot();
+  ASSERT_TRUE(snapshot_read_options.snapshot);
+
+  s = db->Put(write_options, "foo", "2");
+  ASSERT_OK(s);
+
+  // Snapshot was created before change to '2'
+  s = txn->Get(snapshot_read_options, "foo", &value);
+  ASSERT_EQ(value, "1");
+
+  txn->ClearSnapshot();
+  snapshot_read_options.snapshot = txn->GetSnapshot();
+  ASSERT_FALSE(snapshot_read_options.snapshot);
+
+  // Snapshot has now been cleared
+  s = txn->Get(snapshot_read_options, "foo", &value);
+  ASSERT_EQ(value, "2");
+
+  s = txn->Commit();
+  ASSERT_OK(s);
+
+  delete txn;
+}
+
+TEST_P(TransactionTest, ToggleAutoCompactionTest) {
+  Status s;
+
+  ColumnFamilyHandle *cfa, *cfb;
+  ColumnFamilyOptions cf_options;
+
+  // Create 2 new column families
+  s = db->CreateColumnFamily(cf_options, "CFA", &cfa);
+  ASSERT_OK(s);
+  s = db->CreateColumnFamily(cf_options, "CFB", &cfb);
+  ASSERT_OK(s);
+
+  delete cfa;
+  delete cfb;
+  delete db;
+
+  // open DB with three column families
+  std::vector<ColumnFamilyDescriptor> column_families;
+  // have to open default column family
+  column_families.push_back(
+      ColumnFamilyDescriptor(kDefaultColumnFamilyName, ColumnFamilyOptions()));
+  // open the new column families
+  column_families.push_back(
+      ColumnFamilyDescriptor("CFA", ColumnFamilyOptions()));
+  column_families.push_back(
+      ColumnFamilyDescriptor("CFB", ColumnFamilyOptions()));
+
+  ColumnFamilyOptions* cf_opt_default = &column_families[0].options;
+  ColumnFamilyOptions* cf_opt_cfa = &column_families[1].options;
+  ColumnFamilyOptions* cf_opt_cfb = &column_families[2].options;
+  cf_opt_default->disable_auto_compactions = false;
+  cf_opt_cfa->disable_auto_compactions = true;
+  cf_opt_cfb->disable_auto_compactions = false;
+
+  std::vector<ColumnFamilyHandle*> handles;
+
+  s = TransactionDB::Open(options, txn_db_options, dbname, column_families,
+                          &handles, &db);
+  ASSERT_OK(s);
+
+  auto cfh_default = static_cast_with_check<ColumnFamilyHandleImpl>(handles[0]);
+  auto opt_default = *cfh_default->cfd()->GetLatestMutableCFOptions();
+
+  auto cfh_a = static_cast_with_check<ColumnFamilyHandleImpl>(handles[1]);
+  auto opt_a = *cfh_a->cfd()->GetLatestMutableCFOptions();
+
+  auto cfh_b = static_cast_with_check<ColumnFamilyHandleImpl>(handles[2]);
+  auto opt_b = *cfh_b->cfd()->GetLatestMutableCFOptions();
+
+  ASSERT_EQ(opt_default.disable_auto_compactions, false);
+  ASSERT_EQ(opt_a.disable_auto_compactions, true);
+  ASSERT_EQ(opt_b.disable_auto_compactions, false);
+
+  for (auto handle : handles) {
+    delete handle;
+  }
+}
+
+TEST_P(TransactionStressTest, ExpiredTransactionDataRace1) {
+  // In this test, txn1 should succeed committing,
+  // as the callback is called after txn1 starts committing.
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->LoadDependency(
+      {{"TransactionTest::ExpirableTransactionDataRace:1"}});
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
+      "TransactionTest::ExpirableTransactionDataRace:1", [&](void* /*arg*/) {
+        WriteOptions write_options;
+        TransactionOptions txn_options;
+
+        // Force txn1 to expire
+        /* sleep override */
+        std::this_thread::sleep_for(std::chrono::milliseconds(1500));
+
+        Transaction* txn2 = db->BeginTransaction(write_options, txn_options);
+        Status s;
+        s = txn2->Put("X", "2");
+        ASSERT_TRUE(s.IsTimedOut());
+        s = txn2->Commit();
+        ASSERT_OK(s);
+        delete txn2;
+      });
+
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->EnableProcessing();
+
+  WriteOptions write_options;
+  TransactionOptions txn_options;
+
+  txn_options.expiration = 1000;  // 1 second
+  Transaction* txn1 = db->BeginTransaction(write_options, txn_options);
+
+  Status s;
+  s = txn1->Put("X", "1");
+  ASSERT_OK(s);
+  s = txn1->Commit();
+  ASSERT_OK(s);
+
+  ReadOptions read_options;
+  std::string value;
+  s = db->Get(read_options, "X", &value);
+  ASSERT_OK(s);
+  ASSERT_EQ("1", value);
+
+  delete txn1;
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->DisableProcessing();
+}
+
+#if !defined(ROCKSDB_VALGRIND_RUN) || defined(ROCKSDB_FULL_VALGRIND_RUN)
+namespace {
+// cmt_delay_ms is the delay between prepare and commit
+// first_id is the id of the first transaction
+Status TransactionStressTestInserter(
+    TransactionDB* db, const size_t num_transactions, const size_t num_sets,
+    const size_t num_keys_per_set, Random64* rand,
+    const uint64_t cmt_delay_ms = 0, const uint64_t first_id = 0) {
+  WriteOptions write_options;
+  ReadOptions read_options;
+  TransactionOptions txn_options;
+  txn_options.use_only_the_last_commit_time_batch_for_recovery = true;
+
+  // Inside the inserter we might also retake the snapshot. We do both since two
+  // separte functions are engaged for each.
+  txn_options.set_snapshot = rand->OneIn(2);
+
+  RandomTransactionInserter inserter(
+      rand, write_options, read_options, num_keys_per_set,
+      static_cast<uint16_t>(num_sets), cmt_delay_ms, first_id);
+
+  for (size_t t = 0; t < num_transactions; t++) {
+    bool success = inserter.TransactionDBInsert(db, txn_options);
+    if (!success) {
+      // unexpected failure
+      return inserter.GetLastStatus();
+    }
+  }
+  inserter.GetLastStatus().PermitUncheckedError();
+
+  // Make sure at least some of the transactions succeeded.  It's ok if
+  // some failed due to write-conflicts.
+  if (num_transactions != 1 &&
+      inserter.GetFailureCount() > num_transactions / 2) {
+    return Status::TryAgain("Too many transactions failed! " +
+                            std::to_string(inserter.GetFailureCount()) + " / " +
+                            std::to_string(num_transactions));
+  }
+
+  return Status::OK();
+}
+}  // namespace
+
+// Worker threads add a number to a key from each set of keys. The checker
+// threads verify that the sum of all keys in each set are equal.
+TEST_P(MySQLStyleTransactionTest, TransactionStressTest) {
+  // Small write buffer to trigger more compactions
+  options.write_buffer_size = 1024;
+  txn_db_options.rollback_deletion_type_callback =
+      [](TransactionDB*, ColumnFamilyHandle*, const Slice& key) {
+        return RandomTransactionInserter::RollbackDeletionTypeCallback(key);
+      };
+  ASSERT_OK(ReOpenNoDelete());
+  constexpr size_t num_workers = 4;        // worker threads count
+  constexpr size_t num_checkers = 2;       // checker threads count
+  constexpr size_t num_slow_checkers = 2;  // checker threads emulating backups
+  constexpr size_t num_slow_workers = 1;   // slow worker threads count
+  constexpr size_t num_transactions_per_thread = 1000;
+  constexpr uint16_t num_sets = 3;
+  constexpr size_t num_keys_per_set = 100;
+  // Setting the key-space to be 100 keys should cause enough write-conflicts
+  // to make this test interesting.
+
+  std::vector<port::Thread> threads;
+  std::atomic<uint32_t> finished = {0};
+  constexpr bool TAKE_SNAPSHOT = true;
+  uint64_t time_seed = env->NowMicros();
+  printf("time_seed is %" PRIu64 "\n", time_seed);  // would help to reproduce
+
+  std::function<void()> call_inserter = [&] {
+    size_t thd_seed = std::hash<std::thread::id>()(std::this_thread::get_id());
+    Random64 rand(time_seed * thd_seed);
+    ASSERT_OK(TransactionStressTestInserter(db, num_transactions_per_thread,
+                                            num_sets, num_keys_per_set, &rand));
+    finished++;
+  };
+  std::function<void()> call_checker = [&] {
+    size_t thd_seed = std::hash<std::thread::id>()(std::this_thread::get_id());
+    Random64 rand(time_seed * thd_seed);
+    // Verify that data is consistent
+    while (finished < num_workers) {
+      ASSERT_OK(RandomTransactionInserter::Verify(
+          db, num_sets, num_keys_per_set, TAKE_SNAPSHOT, &rand));
+    }
+  };
+  std::function<void()> call_slow_checker = [&] {
+    size_t thd_seed = std::hash<std::thread::id>()(std::this_thread::get_id());
+    Random64 rand(time_seed * thd_seed);
+    // Verify that data is consistent
+    while (finished < num_workers) {
+      uint64_t delay_ms = rand.Uniform(100) + 1;
+      Status s = RandomTransactionInserter::Verify(
+          db, num_sets, num_keys_per_set, TAKE_SNAPSHOT, &rand, delay_ms);
+      ASSERT_OK(s);
+    }
+  };
+  std::function<void()> call_slow_inserter = [&] {
+    size_t thd_seed = std::hash<std::thread::id>()(std::this_thread::get_id());
+    Random64 rand(time_seed * thd_seed);
+    uint64_t id = 0;
+    // Verify that data is consistent
+    while (finished < num_workers) {
+      uint64_t delay_ms = rand.Uniform(500) + 1;
+      ASSERT_OK(TransactionStressTestInserter(db, 1, num_sets, num_keys_per_set,
+                                              &rand, delay_ms, id++));
+    }
+  };
+
+  for (uint32_t i = 0; i < num_workers; i++) {
+    threads.emplace_back(call_inserter);
+  }
+  for (uint32_t i = 0; i < num_checkers; i++) {
+    threads.emplace_back(call_checker);
+  }
+  if (with_slow_threads_) {
+    for (uint32_t i = 0; i < num_slow_checkers; i++) {
+      threads.emplace_back(call_slow_checker);
+    }
+    for (uint32_t i = 0; i < num_slow_workers; i++) {
+      threads.emplace_back(call_slow_inserter);
+    }
+  }
+
+  // Wait for all threads to finish
+  for (auto& t : threads) {
+    t.join();
+  }
+
+  // Verify that data is consistent
+  Status s = RandomTransactionInserter::Verify(db, num_sets, num_keys_per_set,
+                                               !TAKE_SNAPSHOT);
+  ASSERT_OK(s);
+}
+#endif  // !defined(ROCKSDB_VALGRIND_RUN) || defined(ROCKSDB_FULL_VALGRIND_RUN)
+
+TEST_P(TransactionTest, MemoryLimitTest) {
+  TransactionOptions txn_options;
+  // Header (12 bytes) + NOOP (1 byte) + 2 * 8 bytes for data.
+  txn_options.max_write_batch_size = 29;
+  // Set threshold to unlimited so that the write batch does not get flushed,
+  // and can hit the memory limit.
+  txn_options.write_batch_flush_threshold = 0;
+  std::string value;
+  Status s;
+
+  Transaction* txn = db->BeginTransaction(WriteOptions(), txn_options);
+  ASSERT_TRUE(txn);
+
+  ASSERT_EQ(0, txn->GetNumPuts());
+  ASSERT_LE(0, txn->GetID());
+
+  s = txn->Put(Slice("a"), Slice("...."));
+  ASSERT_OK(s);
+  ASSERT_EQ(1, txn->GetNumPuts());
+
+  s = txn->Put(Slice("b"), Slice("...."));
+  ASSERT_OK(s);
+  ASSERT_EQ(2, txn->GetNumPuts());
+
+  s = txn->Put(Slice("b"), Slice("...."));
+  ASSERT_TRUE(s.IsMemoryLimit());
+  ASSERT_EQ(2, txn->GetNumPuts());
+
+  ASSERT_OK(txn->Rollback());
+  delete txn;
+}
+
+#if !defined(ROCKSDB_VALGRIND_RUN) || defined(ROCKSDB_FULL_VALGRIND_RUN)
+// This test clarifies the existing expectation from the sequence number
+// algorithm. It could detect mistakes in updating the code but it is not
+// necessarily the one acceptable way. If the algorithm is legitimately changed,
+// this unit test should be updated as well.
+TEST_P(TransactionStressTest, SeqAdvanceTest) {
+  // TODO(myabandeh): must be test with false before new releases
+  const bool short_test = true;
+  WriteOptions wopts;
+  FlushOptions fopt;
+
+  options.disable_auto_compactions = true;
+  ASSERT_OK(ReOpen());
+
+  // Do the test with NUM_BRANCHES branches in it. Each run of a test takes some
+  // of the branches. This is the same as counting a binary number where i-th
+  // bit represents whether we take branch i in the represented by the number.
+  const size_t NUM_BRANCHES = short_test ? 6 : 10;
+  // Helper function that shows if the branch is to be taken in the run
+  // represented by the number n.
+  auto branch_do = [&](size_t n, size_t* branch) {
+    assert(*branch < NUM_BRANCHES);
+    const size_t filter = static_cast<size_t>(1) << *branch;
+    return n & filter;
+  };
+  const size_t max_n = static_cast<size_t>(1) << NUM_BRANCHES;
+  for (size_t n = 0; n < max_n; n++) {
+    DBImpl* db_impl = static_cast_with_check<DBImpl>(db->GetRootDB());
+    size_t branch = 0;
+    auto seq = db_impl->GetLatestSequenceNumber();
+    exp_seq = seq;
+    TestTxn0(0);
+    seq = db_impl->TEST_GetLastVisibleSequence();
+    ASSERT_EQ(exp_seq, seq);
+
+    if (branch_do(n, &branch)) {
+      ASSERT_OK(db_impl->Flush(fopt));
+      seq = db_impl->TEST_GetLastVisibleSequence();
+      ASSERT_EQ(exp_seq, seq);
+    }
+    if (!short_test && branch_do(n, &branch)) {
+      ASSERT_OK(db_impl->FlushWAL(true));
+      ASSERT_OK(ReOpenNoDelete());
+      db_impl = static_cast_with_check<DBImpl>(db->GetRootDB());
+      seq = db_impl->GetLatestSequenceNumber();
+      ASSERT_EQ(exp_seq, seq);
+    }
+
+    // Doing it twice might detect some bugs
+    TestTxn0(1);
+    seq = db_impl->TEST_GetLastVisibleSequence();
+    ASSERT_EQ(exp_seq, seq);
+
+    TestTxn1(0);
+    seq = db_impl->TEST_GetLastVisibleSequence();
+    ASSERT_EQ(exp_seq, seq);
+
+    if (branch_do(n, &branch)) {
+      ASSERT_OK(db_impl->Flush(fopt));
+      seq = db_impl->TEST_GetLastVisibleSequence();
+      ASSERT_EQ(exp_seq, seq);
+    }
+    if (!short_test && branch_do(n, &branch)) {
+      ASSERT_OK(db_impl->FlushWAL(true));
+      ASSERT_OK(ReOpenNoDelete());
+      db_impl = static_cast_with_check<DBImpl>(db->GetRootDB());
+      seq = db_impl->GetLatestSequenceNumber();
+      ASSERT_EQ(exp_seq, seq);
+    }
+
+    TestTxn3(0);
+    seq = db_impl->TEST_GetLastVisibleSequence();
+    ASSERT_EQ(exp_seq, seq);
+
+    if (branch_do(n, &branch)) {
+      ASSERT_OK(db_impl->Flush(fopt));
+      seq = db_impl->TEST_GetLastVisibleSequence();
+      ASSERT_EQ(exp_seq, seq);
+    }
+    if (!short_test && branch_do(n, &branch)) {
+      ASSERT_OK(db_impl->FlushWAL(true));
+      ASSERT_OK(ReOpenNoDelete());
+      db_impl = static_cast_with_check<DBImpl>(db->GetRootDB());
+      seq = db_impl->GetLatestSequenceNumber();
+      ASSERT_EQ(exp_seq, seq);
+    }
+
+    TestTxn4(0);
+    seq = db_impl->TEST_GetLastVisibleSequence();
+
+    ASSERT_EQ(exp_seq, seq);
+
+    if (branch_do(n, &branch)) {
+      ASSERT_OK(db_impl->Flush(fopt));
+      seq = db_impl->TEST_GetLastVisibleSequence();
+      ASSERT_EQ(exp_seq, seq);
+    }
+    if (!short_test && branch_do(n, &branch)) {
+      ASSERT_OK(db_impl->FlushWAL(true));
+      ASSERT_OK(ReOpenNoDelete());
+      db_impl = static_cast_with_check<DBImpl>(db->GetRootDB());
+      seq = db_impl->GetLatestSequenceNumber();
+      ASSERT_EQ(exp_seq, seq);
+    }
+
+    TestTxn2(0);
+    seq = db_impl->TEST_GetLastVisibleSequence();
+    ASSERT_EQ(exp_seq, seq);
+
+    if (branch_do(n, &branch)) {
+      ASSERT_OK(db_impl->Flush(fopt));
+      seq = db_impl->TEST_GetLastVisibleSequence();
+      ASSERT_EQ(exp_seq, seq);
+    }
+    if (!short_test && branch_do(n, &branch)) {
+      ASSERT_OK(db_impl->FlushWAL(true));
+      ASSERT_OK(ReOpenNoDelete());
+      db_impl = static_cast_with_check<DBImpl>(db->GetRootDB());
+      seq = db_impl->GetLatestSequenceNumber();
+      ASSERT_EQ(exp_seq, seq);
+    }
+    ASSERT_OK(ReOpen());
+  }
+}
+#endif  // !defined(ROCKSDB_VALGRIND_RUN) || defined(ROCKSDB_FULL_VALGRIND_RUN)
+
+// Verify that the optimization would not compromize the correctness
+TEST_P(TransactionTest, Optimizations) {
+  size_t comb_cnt = size_t(1) << 2;  // 2 is number of optimization vars
+  for (size_t new_comb = 0; new_comb < comb_cnt; new_comb++) {
+    TransactionDBWriteOptimizations optimizations;
+    optimizations.skip_concurrency_control = IsInCombination(0, new_comb);
+    optimizations.skip_duplicate_key_check = IsInCombination(1, new_comb);
+
+    ASSERT_OK(ReOpen());
+    WriteOptions write_options;
+    WriteBatch batch;
+    ASSERT_OK(batch.Put(Slice("k"), Slice("v1")));
+    ASSERT_OK(db->Write(write_options, &batch));
+
+    ReadOptions ropt;
+    PinnableSlice pinnable_val;
+    ASSERT_OK(db->Get(ropt, db->DefaultColumnFamily(), "k", &pinnable_val));
+    ASSERT_TRUE(pinnable_val == ("v1"));
+  }
+}
+
+// A comparator that uses only the first three bytes
+class ThreeBytewiseComparator : public Comparator {
+ public:
+  ThreeBytewiseComparator() {}
+  const char* Name() const override { return "test.ThreeBytewiseComparator"; }
+  int Compare(const Slice& a, const Slice& b) const override {
+    Slice na = Slice(a.data(), a.size() < 3 ? a.size() : 3);
+    Slice nb = Slice(b.data(), b.size() < 3 ? b.size() : 3);
+    return na.compare(nb);
+  }
+  bool Equal(const Slice& a, const Slice& b) const override {
+    Slice na = Slice(a.data(), a.size() < 3 ? a.size() : 3);
+    Slice nb = Slice(b.data(), b.size() < 3 ? b.size() : 3);
+    return na == nb;
+  }
+  // These methods below don't seem relevant to this test. Implement them if
+  // proven othersize.
+  void FindShortestSeparator(std::string* start,
+                             const Slice& limit) const override {
+    const Comparator* bytewise_comp = BytewiseComparator();
+    bytewise_comp->FindShortestSeparator(start, limit);
+  }
+  void FindShortSuccessor(std::string* key) const override {
+    const Comparator* bytewise_comp = BytewiseComparator();
+    bytewise_comp->FindShortSuccessor(key);
+  }
+};
+
+#if !defined(ROCKSDB_VALGRIND_RUN) || defined(ROCKSDB_FULL_VALGRIND_RUN)
+TEST_P(TransactionTest, GetWithoutSnapshot) {
+  WriteOptions write_options;
+  std::atomic<bool> finish = {false};
+  ASSERT_OK(db->Put(write_options, "key", "value"));
+  ROCKSDB_NAMESPACE::port::Thread commit_thread([&]() {
+    for (int i = 0; i < 100; i++) {
+      TransactionOptions txn_options;
+      Transaction* txn = db->BeginTransaction(write_options, txn_options);
+      ASSERT_OK(txn->SetName("xid"));
+      ASSERT_OK(txn->Put("key", "overridedvalue"));
+      ASSERT_OK(txn->Put("key", "value"));
+      ASSERT_OK(txn->Prepare());
+      ASSERT_OK(txn->Commit());
+      delete txn;
+    }
+    finish = true;
+  });
+  ROCKSDB_NAMESPACE::port::Thread read_thread([&]() {
+    while (!finish) {
+      ReadOptions ropt;
+      PinnableSlice pinnable_val;
+      ASSERT_OK(db->Get(ropt, db->DefaultColumnFamily(), "key", &pinnable_val));
+      ASSERT_TRUE(pinnable_val == ("value"));
+    }
+  });
+  commit_thread.join();
+  read_thread.join();
+}
+#endif  // !defined(ROCKSDB_VALGRIND_RUN) || defined(ROCKSDB_FULL_VALGRIND_RUN)
+
+// Test that the transactional db can handle duplicate keys in the write batch
+TEST_P(TransactionTest, DuplicateKeys) {
+  ColumnFamilyOptions cf_options;
+  std::string cf_name = "two";
+  ColumnFamilyHandle* cf_handle = nullptr;
+  {
+    ASSERT_OK(db->CreateColumnFamily(cf_options, cf_name, &cf_handle));
+    WriteOptions write_options;
+    WriteBatch batch;
+    ASSERT_OK(batch.Put(Slice("key"), Slice("value")));
+    ASSERT_OK(batch.Put(Slice("key2"), Slice("value2")));
+    // duplicate the keys
+    ASSERT_OK(batch.Put(Slice("key"), Slice("value3")));
+    // duplicate the 2nd key. It should not be counted duplicate since a
+    // sub-patch is cut after the last duplicate.
+    ASSERT_OK(batch.Put(Slice("key2"), Slice("value4")));
+    // duplicate the keys but in a different cf. It should not be counted as
+    // duplicate keys
+    ASSERT_OK(batch.Put(cf_handle, Slice("key"), Slice("value5")));
+
+    ASSERT_OK(db->Write(write_options, &batch));
+
+    ReadOptions ropt;
+    PinnableSlice pinnable_val;
+    auto s = db->Get(ropt, db->DefaultColumnFamily(), "key", &pinnable_val);
+    ASSERT_OK(s);
+    ASSERT_TRUE(pinnable_val == ("value3"));
+    s = db->Get(ropt, db->DefaultColumnFamily(), "key2", &pinnable_val);
+    ASSERT_OK(s);
+    ASSERT_TRUE(pinnable_val == ("value4"));
+    s = db->Get(ropt, cf_handle, "key", &pinnable_val);
+    ASSERT_OK(s);
+    ASSERT_TRUE(pinnable_val == ("value5"));
+
+    delete cf_handle;
+  }
+
+  // Test with non-bytewise comparator
+  {
+    ASSERT_OK(ReOpen());
+    std::unique_ptr<const Comparator> comp_gc(new ThreeBytewiseComparator());
+    cf_options.comparator = comp_gc.get();
+    ASSERT_OK(db->CreateColumnFamily(cf_options, cf_name, &cf_handle));
+    WriteOptions write_options;
+    WriteBatch batch;
+    ASSERT_OK(batch.Put(cf_handle, Slice("key"), Slice("value")));
+    // The first three bytes are the same, do it must be counted as duplicate
+    ASSERT_OK(batch.Put(cf_handle, Slice("key2"), Slice("value2")));
+    // check for 2nd duplicate key in cf with non-default comparator
+    ASSERT_OK(batch.Put(cf_handle, Slice("key2b"), Slice("value2b")));
+    ASSERT_OK(db->Write(write_options, &batch));
+
+    // The value must be the most recent value for all the keys equal to "key",
+    // including "key2"
+    ReadOptions ropt;
+    PinnableSlice pinnable_val;
+    ASSERT_OK(db->Get(ropt, cf_handle, "key", &pinnable_val));
+    ASSERT_TRUE(pinnable_val == ("value2b"));
+
+    // Test duplicate keys with rollback
+    TransactionOptions txn_options;
+    Transaction* txn0 = db->BeginTransaction(write_options, txn_options);
+    ASSERT_OK(txn0->SetName("xid"));
+    ASSERT_OK(txn0->Put(cf_handle, Slice("key3"), Slice("value3")));
+    ASSERT_OK(txn0->Merge(cf_handle, Slice("key4"), Slice("value4")));
+    ASSERT_OK(txn0->Rollback());
+    ASSERT_OK(db->Get(ropt, cf_handle, "key5", &pinnable_val));
+    ASSERT_TRUE(pinnable_val == ("value2b"));
+    delete txn0;
+
+    delete cf_handle;
+    cf_options.comparator = BytewiseComparator();
+  }
+
+  for (bool do_prepare : {true, false}) {
+    for (bool do_rollback : {true, false}) {
+      for (bool with_commit_batch : {true, false}) {
+        if (with_commit_batch && !do_prepare) {
+          continue;
+        }
+        if (with_commit_batch && do_rollback) {
+          continue;
+        }
+        ASSERT_OK(ReOpen());
+        ASSERT_OK(db->CreateColumnFamily(cf_options, cf_name, &cf_handle));
+        TransactionOptions txn_options;
+        txn_options.use_only_the_last_commit_time_batch_for_recovery = true;
+        WriteOptions write_options;
+        Transaction* txn0 = db->BeginTransaction(write_options, txn_options);
+        auto s = txn0->SetName("xid");
+        ASSERT_OK(s);
+        s = txn0->Put(Slice("foo0"), Slice("bar0a"));
+        ASSERT_OK(s);
+        s = txn0->Put(Slice("foo0"), Slice("bar0b"));
+        ASSERT_OK(s);
+        s = txn0->Put(Slice("foo1"), Slice("bar1"));
+        ASSERT_OK(s);
+        s = txn0->Merge(Slice("foo2"), Slice("bar2a"));
+        ASSERT_OK(s);
+        // Repeat a key after the start of a sub-patch. This should not cause a
+        // duplicate in the most recent sub-patch and hence not creating a new
+        // sub-patch.
+        s = txn0->Put(Slice("foo0"), Slice("bar0c"));
+        ASSERT_OK(s);
+        s = txn0->Merge(Slice("foo2"), Slice("bar2b"));
+        ASSERT_OK(s);
+        // duplicate the keys but in a different cf. It should not be counted as
+        // duplicate.
+        s = txn0->Put(cf_handle, Slice("foo0"), Slice("bar0-cf1"));
+        ASSERT_OK(s);
+        s = txn0->Put(Slice("foo3"), Slice("bar3"));
+        ASSERT_OK(s);
+        s = txn0->Merge(Slice("foo3"), Slice("bar3"));
+        ASSERT_OK(s);
+        s = txn0->Put(Slice("foo4"), Slice("bar4"));
+        ASSERT_OK(s);
+        s = txn0->Delete(Slice("foo4"));
+        ASSERT_OK(s);
+        s = txn0->SingleDelete(Slice("foo4"));
+        ASSERT_OK(s);
+        if (do_prepare) {
+          s = txn0->Prepare();
+          ASSERT_OK(s);
+        }
+        if (do_rollback) {
+          // Test rolling back the batch with duplicates
+          s = txn0->Rollback();
+          ASSERT_OK(s);
+        } else {
+          if (with_commit_batch) {
+            assert(do_prepare);
+            auto cb = txn0->GetCommitTimeWriteBatch();
+            // duplicate a key in the original batch
+            // TODO(myabandeh): the behavior of GetCommitTimeWriteBatch
+            // conflicting with the prepared batch is currently undefined and
+            // gives different results in different implementations.
+
+            // s = cb->Put(Slice("foo0"), Slice("bar0d"));
+            // ASSERT_OK(s);
+            // add a new duplicate key
+            s = cb->Put(Slice("foo6"), Slice("bar6a"));
+            ASSERT_OK(s);
+            s = cb->Put(Slice("foo6"), Slice("bar6b"));
+            ASSERT_OK(s);
+            // add a duplicate key that is removed in the same batch
+            s = cb->Put(Slice("foo7"), Slice("bar7a"));
+            ASSERT_OK(s);
+            s = cb->Delete(Slice("foo7"));
+            ASSERT_OK(s);
+          }
+          s = txn0->Commit();
+          ASSERT_OK(s);
+        }
+        delete txn0;
+        ReadOptions ropt;
+        PinnableSlice pinnable_val;
+
+        if (do_rollback) {
+          s = db->Get(ropt, db->DefaultColumnFamily(), "foo0", &pinnable_val);
+          ASSERT_TRUE(s.IsNotFound());
+          s = db->Get(ropt, cf_handle, "foo0", &pinnable_val);
+          ASSERT_TRUE(s.IsNotFound());
+          s = db->Get(ropt, db->DefaultColumnFamily(), "foo1", &pinnable_val);
+          ASSERT_TRUE(s.IsNotFound());
+          s = db->Get(ropt, db->DefaultColumnFamily(), "foo2", &pinnable_val);
+          ASSERT_TRUE(s.IsNotFound());
+          s = db->Get(ropt, db->DefaultColumnFamily(), "foo3", &pinnable_val);
+          ASSERT_TRUE(s.IsNotFound());
+          s = db->Get(ropt, db->DefaultColumnFamily(), "foo4", &pinnable_val);
+          ASSERT_TRUE(s.IsNotFound());
+        } else {
+          s = db->Get(ropt, db->DefaultColumnFamily(), "foo0", &pinnable_val);
+          ASSERT_OK(s);
+          ASSERT_TRUE(pinnable_val == ("bar0c"));
+          s = db->Get(ropt, cf_handle, "foo0", &pinnable_val);
+          ASSERT_OK(s);
+          ASSERT_TRUE(pinnable_val == ("bar0-cf1"));
+          s = db->Get(ropt, db->DefaultColumnFamily(), "foo1", &pinnable_val);
+          ASSERT_OK(s);
+          ASSERT_TRUE(pinnable_val == ("bar1"));
+          s = db->Get(ropt, db->DefaultColumnFamily(), "foo2", &pinnable_val);
+          ASSERT_OK(s);
+          ASSERT_TRUE(pinnable_val == ("bar2a,bar2b"));
+          s = db->Get(ropt, db->DefaultColumnFamily(), "foo3", &pinnable_val);
+          ASSERT_OK(s);
+          ASSERT_TRUE(pinnable_val == ("bar3,bar3"));
+          s = db->Get(ropt, db->DefaultColumnFamily(), "foo4", &pinnable_val);
+          ASSERT_TRUE(s.IsNotFound());
+          if (with_commit_batch) {
+            s = db->Get(ropt, db->DefaultColumnFamily(), "foo6", &pinnable_val);
+            if (txn_db_options.write_policy ==
+                TxnDBWritePolicy::WRITE_COMMITTED) {
+              ASSERT_OK(s);
+              ASSERT_TRUE(pinnable_val == ("bar6b"));
+            } else {
+              ASSERT_TRUE(s.IsNotFound());
+            }
+            s = db->Get(ropt, db->DefaultColumnFamily(), "foo7", &pinnable_val);
+            ASSERT_TRUE(s.IsNotFound());
+          }
+        }
+        delete cf_handle;
+      }  // with_commit_batch
+    }    // do_rollback
+  }      // do_prepare
+
+  if (!options.unordered_write) {
+    // Also test with max_successive_merges > 0. max_successive_merges will not
+    // affect our algorithm for duplicate key insertion but we add the test to
+    // verify that.
+    cf_options.max_successive_merges = 2;
+    cf_options.merge_operator = MergeOperators::CreateStringAppendOperator();
+    ASSERT_OK(ReOpen());
+    db->CreateColumnFamily(cf_options, cf_name, &cf_handle);
+    WriteOptions write_options;
+    // Ensure one value for the key
+    ASSERT_OK(db->Put(write_options, cf_handle, Slice("key"), Slice("value")));
+    WriteBatch batch;
+    // Merge more than max_successive_merges times
+    ASSERT_OK(batch.Merge(cf_handle, Slice("key"), Slice("1")));
+    ASSERT_OK(batch.Merge(cf_handle, Slice("key"), Slice("2")));
+    ASSERT_OK(batch.Merge(cf_handle, Slice("key"), Slice("3")));
+    ASSERT_OK(batch.Merge(cf_handle, Slice("key"), Slice("4")));
+    ASSERT_OK(db->Write(write_options, &batch));
+    ReadOptions read_options;
+    std::string value;
+    ASSERT_OK(db->Get(read_options, cf_handle, "key", &value));
+    ASSERT_EQ(value, "value,1,2,3,4");
+    delete cf_handle;
+  }
+
+  {
+    // Test that the duplicate detection is not compromised after rolling back
+    // to a save point
+    TransactionOptions txn_options;
+    WriteOptions write_options;
+    Transaction* txn0 = db->BeginTransaction(write_options, txn_options);
+    ASSERT_OK(txn0->Put(Slice("foo0"), Slice("bar0a")));
+    ASSERT_OK(txn0->Put(Slice("foo0"), Slice("bar0b")));
+    txn0->SetSavePoint();
+    ASSERT_OK(txn0->RollbackToSavePoint());
+    ASSERT_OK(txn0->Commit());
+    delete txn0;
+  }
+
+  // Test sucessfull recovery after a crash
+  {
+    ASSERT_OK(ReOpen());
+    TransactionOptions txn_options;
+    WriteOptions write_options;
+    ReadOptions ropt;
+    Transaction* txn0;
+    PinnableSlice pinnable_val;
+    Status s;
+
+    std::unique_ptr<const Comparator> comp_gc(new ThreeBytewiseComparator());
+    cf_options.comparator = comp_gc.get();
+    cf_options.merge_operator = MergeOperators::CreateStringAppendOperator();
+    ASSERT_OK(db->CreateColumnFamily(cf_options, cf_name, &cf_handle));
+    delete cf_handle;
+    std::vector<ColumnFamilyDescriptor> cfds{
+        ColumnFamilyDescriptor(kDefaultColumnFamilyName,
+                               ColumnFamilyOptions(options)),
+        ColumnFamilyDescriptor(cf_name, cf_options),
+    };
+    std::vector<ColumnFamilyHandle*> handles;
+    ASSERT_OK(ReOpenNoDelete(cfds, &handles));
+
+    assert(db != nullptr);
+    ASSERT_OK(db->Put(write_options, "foo0", "init"));
+    ASSERT_OK(db->Put(write_options, "foo1", "init"));
+    ASSERT_OK(db->Put(write_options, handles[1], "foo0", "init"));
+    ASSERT_OK(db->Put(write_options, handles[1], "foo1", "init"));
+
+    // one entry
+    txn0 = db->BeginTransaction(write_options, txn_options);
+    ASSERT_OK(txn0->SetName("xid"));
+    ASSERT_OK(txn0->Put(Slice("foo0"), Slice("bar0a")));
+    ASSERT_OK(txn0->Prepare());
+    delete txn0;
+    // This will check the asserts inside recovery code
+    ASSERT_OK(db->FlushWAL(true));
+    reinterpret_cast<PessimisticTransactionDB*>(db)->TEST_Crash();
+    ASSERT_OK(ReOpenNoDelete(cfds, &handles));
+    txn0 = db->GetTransactionByName("xid");
+    ASSERT_TRUE(txn0 != nullptr);
+    ASSERT_OK(txn0->Commit());
+    delete txn0;
+    s = db->Get(ropt, db->DefaultColumnFamily(), "foo0", &pinnable_val);
+    ASSERT_OK(s);
+    ASSERT_TRUE(pinnable_val == ("bar0a"));
+
+    // two entries, no duplicate
+    txn0 = db->BeginTransaction(write_options, txn_options);
+    ASSERT_OK(txn0->SetName("xid"));
+    ASSERT_OK(txn0->Put(handles[1], Slice("foo0"), Slice("bar0b")));
+    ASSERT_OK(txn0->Put(handles[1], Slice("fol1"), Slice("bar1b")));
+    ASSERT_OK(txn0->Put(Slice("foo0"), Slice("bar0b")));
+    ASSERT_OK(txn0->Put(Slice("foo1"), Slice("bar1b")));
+    ASSERT_OK(txn0->Prepare());
+    delete txn0;
+    // This will check the asserts inside recovery code
+    ASSERT_OK(db->FlushWAL(true));
+    // Flush only cf 1
+    ASSERT_OK(static_cast_with_check<DBImpl>(db->GetRootDB())
+                  ->TEST_FlushMemTable(true, false, handles[1]));
+    reinterpret_cast<PessimisticTransactionDB*>(db)->TEST_Crash();
+    ASSERT_OK(ReOpenNoDelete(cfds, &handles));
+    txn0 = db->GetTransactionByName("xid");
+    ASSERT_TRUE(txn0 != nullptr);
+    ASSERT_OK(txn0->Commit());
+    delete txn0;
+    pinnable_val.Reset();
+    s = db->Get(ropt, db->DefaultColumnFamily(), "foo0", &pinnable_val);
+    ASSERT_OK(s);
+    ASSERT_TRUE(pinnable_val == ("bar0b"));
+    pinnable_val.Reset();
+    s = db->Get(ropt, db->DefaultColumnFamily(), "foo1", &pinnable_val);
+    ASSERT_OK(s);
+    ASSERT_TRUE(pinnable_val == ("bar1b"));
+    pinnable_val.Reset();
+    s = db->Get(ropt, handles[1], "foo0", &pinnable_val);
+    ASSERT_OK(s);
+    ASSERT_TRUE(pinnable_val == ("bar0b"));
+    pinnable_val.Reset();
+    s = db->Get(ropt, handles[1], "fol1", &pinnable_val);
+    ASSERT_OK(s);
+    ASSERT_TRUE(pinnable_val == ("bar1b"));
+
+    // one duplicate with ::Put
+    txn0 = db->BeginTransaction(write_options, txn_options);
+    ASSERT_OK(txn0->SetName("xid"));
+    ASSERT_OK(txn0->Put(handles[1], Slice("key-nonkey0"), Slice("bar0c")));
+    ASSERT_OK(txn0->Put(handles[1], Slice("key-nonkey1"), Slice("bar1d")));
+    ASSERT_OK(txn0->Put(Slice("foo0"), Slice("bar0c")));
+    ASSERT_OK(txn0->Put(Slice("foo1"), Slice("bar1c")));
+    ASSERT_OK(txn0->Put(Slice("foo0"), Slice("bar0d")));
+    ASSERT_OK(txn0->Prepare());
+    delete txn0;
+    // This will check the asserts inside recovery code
+    ASSERT_OK(db->FlushWAL(true));
+    // Flush only cf 1
+    ASSERT_OK(static_cast_with_check<DBImpl>(db->GetRootDB())
+                  ->TEST_FlushMemTable(true, false, handles[1]));
+    reinterpret_cast<PessimisticTransactionDB*>(db)->TEST_Crash();
+    ASSERT_OK(ReOpenNoDelete(cfds, &handles));
+    txn0 = db->GetTransactionByName("xid");
+    ASSERT_TRUE(txn0 != nullptr);
+    ASSERT_OK(txn0->Commit());
+    delete txn0;
+    pinnable_val.Reset();
+    s = db->Get(ropt, db->DefaultColumnFamily(), "foo0", &pinnable_val);
+    ASSERT_OK(s);
+    ASSERT_TRUE(pinnable_val == ("bar0d"));
+    pinnable_val.Reset();
+    s = db->Get(ropt, db->DefaultColumnFamily(), "foo1", &pinnable_val);
+    ASSERT_OK(s);
+    ASSERT_TRUE(pinnable_val == ("bar1c"));
+    pinnable_val.Reset();
+    s = db->Get(ropt, handles[1], "key-nonkey2", &pinnable_val);
+    ASSERT_OK(s);
+    ASSERT_TRUE(pinnable_val == ("bar1d"));
+
+    // Duplicate with ::Put, ::Delete
+    txn0 = db->BeginTransaction(write_options, txn_options);
+    ASSERT_OK(txn0->SetName("xid"));
+    ASSERT_OK(txn0->Put(handles[1], Slice("key-nonkey0"), Slice("bar0e")));
+    ASSERT_OK(txn0->Delete(handles[1], Slice("key-nonkey1")));
+    ASSERT_OK(txn0->Put(Slice("foo0"), Slice("bar0e")));
+    ASSERT_OK(txn0->Delete(Slice("foo0")));
+    ASSERT_OK(txn0->Prepare());
+    delete txn0;
+    // This will check the asserts inside recovery code
+    ASSERT_OK(db->FlushWAL(true));
+    // Flush only cf 1
+    ASSERT_OK(static_cast_with_check<DBImpl>(db->GetRootDB())
+                  ->TEST_FlushMemTable(true, false, handles[1]));
+    reinterpret_cast<PessimisticTransactionDB*>(db)->TEST_Crash();
+    ASSERT_OK(ReOpenNoDelete(cfds, &handles));
+    txn0 = db->GetTransactionByName("xid");
+    ASSERT_TRUE(txn0 != nullptr);
+    ASSERT_OK(txn0->Commit());
+    delete txn0;
+    pinnable_val.Reset();
+    s = db->Get(ropt, db->DefaultColumnFamily(), "foo0", &pinnable_val);
+    ASSERT_TRUE(s.IsNotFound());
+    pinnable_val.Reset();
+    s = db->Get(ropt, handles[1], "key-nonkey2", &pinnable_val);
+    ASSERT_TRUE(s.IsNotFound());
+
+    // Duplicate with ::Put, ::SingleDelete
+    txn0 = db->BeginTransaction(write_options, txn_options);
+    ASSERT_OK(txn0->SetName("xid"));
+    ASSERT_OK(txn0->Put(handles[1], Slice("key-nonkey0"), Slice("bar0g")));
+    ASSERT_OK(txn0->SingleDelete(handles[1], Slice("key-nonkey1")));
+    ASSERT_OK(txn0->Put(Slice("foo0"), Slice("bar0e")));
+    ASSERT_OK(txn0->SingleDelete(Slice("foo0")));
+    ASSERT_OK(txn0->Prepare());
+    delete txn0;
+    // This will check the asserts inside recovery code
+    ASSERT_OK(db->FlushWAL(true));
+    // Flush only cf 1
+    ASSERT_OK(static_cast_with_check<DBImpl>(db->GetRootDB())
+                  ->TEST_FlushMemTable(true, false, handles[1]));
+    reinterpret_cast<PessimisticTransactionDB*>(db)->TEST_Crash();
+    ASSERT_OK(ReOpenNoDelete(cfds, &handles));
+    txn0 = db->GetTransactionByName("xid");
+    ASSERT_TRUE(txn0 != nullptr);
+    ASSERT_OK(txn0->Commit());
+    delete txn0;
+    pinnable_val.Reset();
+    s = db->Get(ropt, db->DefaultColumnFamily(), "foo0", &pinnable_val);
+    ASSERT_TRUE(s.IsNotFound());
+    pinnable_val.Reset();
+    s = db->Get(ropt, handles[1], "key-nonkey2", &pinnable_val);
+    ASSERT_TRUE(s.IsNotFound());
+
+    // Duplicate with ::Put, ::Merge
+    txn0 = db->BeginTransaction(write_options, txn_options);
+    ASSERT_OK(txn0->SetName("xid"));
+    ASSERT_OK(txn0->Put(handles[1], Slice("key-nonkey0"), Slice("bar1i")));
+    ASSERT_OK(txn0->Merge(handles[1], Slice("key-nonkey1"), Slice("bar1j")));
+    ASSERT_OK(txn0->Put(Slice("foo0"), Slice("bar0f")));
+    ASSERT_OK(txn0->Merge(Slice("foo0"), Slice("bar0g")));
+    ASSERT_OK(txn0->Prepare());
+    delete txn0;
+    // This will check the asserts inside recovery code
+    ASSERT_OK(db->FlushWAL(true));
+    // Flush only cf 1
+    ASSERT_OK(static_cast_with_check<DBImpl>(db->GetRootDB())
+                  ->TEST_FlushMemTable(true, false, handles[1]));
+    reinterpret_cast<PessimisticTransactionDB*>(db)->TEST_Crash();
+    ASSERT_OK(ReOpenNoDelete(cfds, &handles));
+    txn0 = db->GetTransactionByName("xid");
+    ASSERT_TRUE(txn0 != nullptr);
+    ASSERT_OK(txn0->Commit());
+    delete txn0;
+    pinnable_val.Reset();
+    s = db->Get(ropt, db->DefaultColumnFamily(), "foo0", &pinnable_val);
+    ASSERT_OK(s);
+    ASSERT_TRUE(pinnable_val == ("bar0f,bar0g"));
+    pinnable_val.Reset();
+    s = db->Get(ropt, handles[1], "key-nonkey2", &pinnable_val);
+    ASSERT_OK(s);
+    ASSERT_TRUE(pinnable_val == ("bar1i,bar1j"));
+
+    for (auto h : handles) {
+      delete h;
+    }
+    delete db;
+    db = nullptr;
+  }
+}
+
+// Test that the reseek optimization in iterators will not result in an infinite
+// loop if there are too many uncommitted entries before the snapshot.
+TEST_P(TransactionTest, ReseekOptimization) {
+  WriteOptions write_options;
+  write_options.sync = true;
+  write_options.disableWAL = false;
+  ColumnFamilyDescriptor cfd;
+  ASSERT_OK(db->DefaultColumnFamily()->GetDescriptor(&cfd));
+  auto max_skip = cfd.options.max_sequential_skip_in_iterations;
+
+  ASSERT_OK(db->Put(write_options, Slice("foo0"), Slice("initv")));
+
+  TransactionOptions txn_options;
+  Transaction* txn0 = db->BeginTransaction(write_options, txn_options);
+  ASSERT_OK(txn0->SetName("xid"));
+  // Duplicate keys will result into separate sequence numbers in WritePrepared
+  // and WriteUnPrepared
+  for (size_t i = 0; i < 2 * max_skip; i++) {
+    ASSERT_OK(txn0->Put(Slice("foo1"), Slice("bar")));
+  }
+  ASSERT_OK(txn0->Prepare());
+  ASSERT_OK(db->Put(write_options, Slice("foo2"), Slice("initv")));
+
+  ReadOptions read_options;
+  // To avoid loops
+  read_options.max_skippable_internal_keys = 10 * max_skip;
+  Iterator* iter = db->NewIterator(read_options);
+  ASSERT_OK(iter->status());
+  size_t cnt = 0;
+  iter->SeekToFirst();
+  while (iter->Valid()) {
+    iter->Next();
+    ASSERT_OK(iter->status());
+    cnt++;
+  }
+  ASSERT_EQ(cnt, 2);
+  cnt = 0;
+  iter->SeekToLast();
+  while (iter->Valid()) {
+    iter->Prev();
+    ASSERT_OK(iter->status());
+    cnt++;
+  }
+  ASSERT_EQ(cnt, 2);
+  delete iter;
+  ASSERT_OK(txn0->Rollback());
+  delete txn0;
+}
+
+// After recovery in kPointInTimeRecovery mode, the corrupted log file remains
+// there. The new log files should be still read succesfully during recovery of
+// the 2nd crash.
+TEST_P(TransactionTest, DoubleCrashInRecovery) {
+  for (const bool manual_wal_flush : {false, true}) {
+    for (const bool write_after_recovery : {false, true}) {
+      options.wal_recovery_mode = WALRecoveryMode::kPointInTimeRecovery;
+      options.manual_wal_flush = manual_wal_flush;
+      ASSERT_OK(ReOpen());
+      std::string cf_name = "two";
+      ColumnFamilyOptions cf_options;
+      ColumnFamilyHandle* cf_handle = nullptr;
+      ASSERT_OK(db->CreateColumnFamily(cf_options, cf_name, &cf_handle));
+
+      // Add a prepare entry to prevent the older logs from being deleted.
+      WriteOptions write_options;
+      TransactionOptions txn_options;
+      Transaction* txn = db->BeginTransaction(write_options, txn_options);
+      ASSERT_OK(txn->SetName("xid"));
+      ASSERT_OK(txn->Put(Slice("foo-prepare"), Slice("bar-prepare")));
+      ASSERT_OK(txn->Prepare());
+
+      FlushOptions flush_ops;
+      ASSERT_OK(db->Flush(flush_ops));
+      // Now we have a log that cannot be deleted
+
+      ASSERT_OK(db->Put(write_options, cf_handle, "foo1", "bar1"));
+      // Flush only the 2nd cf
+      ASSERT_OK(db->Flush(flush_ops, cf_handle));
+
+      // The value is large enough to be touched by the corruption we ingest
+      // below.
+      std::string large_value(400, ' ');
+      // key/value not touched by corruption
+      ASSERT_OK(db->Put(write_options, "foo2", "bar2"));
+      // key/value touched by corruption
+      ASSERT_OK(db->Put(write_options, "foo3", large_value));
+      // key/value not touched by corruption
+      ASSERT_OK(db->Put(write_options, "foo4", "bar4"));
+
+      ASSERT_OK(db->FlushWAL(true));
+      DBImpl* db_impl = static_cast_with_check<DBImpl>(db->GetRootDB());
+      uint64_t wal_file_id = db_impl->TEST_LogfileNumber();
+      std::string fname = LogFileName(dbname, wal_file_id);
+      reinterpret_cast<PessimisticTransactionDB*>(db)->TEST_Crash();
+      delete txn;
+      delete cf_handle;
+      delete db;
+      db = nullptr;
+
+      // Corrupt the last log file in the middle, so that it is not corrupted
+      // in the tail.
+      std::string file_content;
+      ASSERT_OK(ReadFileToString(env, fname, &file_content));
+      file_content[400] = 'h';
+      file_content[401] = 'a';
+      ASSERT_OK(env->DeleteFile(fname));
+      ASSERT_OK(WriteStringToFile(env, file_content, fname, true));
+
+      // Recover from corruption
+      std::vector<ColumnFamilyHandle*> handles;
+      std::vector<ColumnFamilyDescriptor> column_families;
+      column_families.push_back(ColumnFamilyDescriptor(kDefaultColumnFamilyName,
+                                                       ColumnFamilyOptions()));
+      column_families.push_back(
+          ColumnFamilyDescriptor("two", ColumnFamilyOptions()));
+      ASSERT_OK(ReOpenNoDelete(column_families, &handles));
+      assert(db != nullptr);
+
+      if (write_after_recovery) {
+        // Write data to the log right after the corrupted log
+        ASSERT_OK(db->Put(write_options, "foo5", large_value));
+      }
+
+      // Persist data written to WAL during recovery or by the last Put
+      ASSERT_OK(db->FlushWAL(true));
+      // 2nd crash to recover while having a valid log after the corrupted one.
+      ASSERT_OK(ReOpenNoDelete(column_families, &handles));
+      assert(db != nullptr);
+      txn = db->GetTransactionByName("xid");
+      ASSERT_TRUE(txn != nullptr);
+      ASSERT_OK(txn->Commit());
+      delete txn;
+      for (auto handle : handles) {
+        delete handle;
+      }
+    }
+  }
+}
+
+TEST_P(TransactionTest, CommitWithoutPrepare) {
+  {
+    // skip_prepare = false.
+    WriteOptions write_options;
+    TransactionOptions txn_options;
+    txn_options.skip_prepare = false;
+    Transaction* txn = db->BeginTransaction(write_options, txn_options);
+    ASSERT_TRUE(txn->Commit().IsTxnNotPrepared());
+    delete txn;
+  }
+
+  {
+    // skip_prepare = true.
+    WriteOptions write_options;
+    TransactionOptions txn_options;
+    txn_options.skip_prepare = true;
+    Transaction* txn = db->BeginTransaction(write_options, txn_options);
+    ASSERT_OK(txn->Commit());
+    delete txn;
+  }
+}
+
+TEST_P(TransactionTest, OpenAndEnableU64Timestamp) {
+  ASSERT_OK(ReOpenNoDelete());
+
+  assert(db);
+
+  const std::string test_cf_name = "test_cf";
+  ColumnFamilyOptions cf_opts;
+  cf_opts.comparator = test::BytewiseComparatorWithU64TsWrapper();
+  {
+    ColumnFamilyHandle* cfh = nullptr;
+    const Status s = db->CreateColumnFamily(cf_opts, test_cf_name, &cfh);
+    if (txn_db_options.write_policy == WRITE_COMMITTED) {
+      ASSERT_OK(s);
+      delete cfh;
+    } else {
+      ASSERT_TRUE(s.IsNotSupported());
+      assert(!cfh);
+    }
+  }
+
+  // Bypass transaction db layer.
+  if (txn_db_options.write_policy != WRITE_COMMITTED) {
+    DBImpl* db_impl = static_cast_with_check<DBImpl>(db->GetRootDB());
+    assert(db_impl);
+    ColumnFamilyHandle* cfh = nullptr;
+    ASSERT_OK(db_impl->CreateColumnFamily(cf_opts, test_cf_name, &cfh));
+    delete cfh;
+  }
+
+  {
+    std::vector<ColumnFamilyDescriptor> cf_descs;
+    cf_descs.emplace_back(kDefaultColumnFamilyName, options);
+    cf_descs.emplace_back(test_cf_name, cf_opts);
+    std::vector<ColumnFamilyHandle*> handles;
+    const Status s = ReOpenNoDelete(cf_descs, &handles);
+    if (txn_db_options.write_policy == WRITE_COMMITTED) {
+      ASSERT_OK(s);
+      for (auto* h : handles) {
+        delete h;
+      }
+    } else {
+      ASSERT_TRUE(s.IsNotSupported());
+    }
+  }
+}
+
+TEST_P(TransactionTest, OpenAndEnableU32Timestamp) {
+  class DummyComparatorWithU32Ts : public Comparator {
+   public:
+    DummyComparatorWithU32Ts() : Comparator(sizeof(uint32_t)) {}
+    const char* Name() const override { return "DummyComparatorWithU32Ts"; }
+    void FindShortSuccessor(std::string*) const override {}
+    void FindShortestSeparator(std::string*, const Slice&) const override {}
+    int Compare(const Slice&, const Slice&) const override { return 0; }
+  };
+
+  std::unique_ptr<Comparator> dummy_ucmp(new DummyComparatorWithU32Ts());
+
+  ASSERT_OK(ReOpenNoDelete());
+
+  assert(db);
+
+  const std::string test_cf_name = "test_cf";
+
+  ColumnFamilyOptions cf_opts;
+  cf_opts.comparator = dummy_ucmp.get();
+  {
+    ColumnFamilyHandle* cfh = nullptr;
+    ASSERT_TRUE(db->CreateColumnFamily(cf_opts, test_cf_name, &cfh)
+                    .IsInvalidArgument());
+  }
+
+  // Bypass transaction db layer.
+  {
+    ColumnFamilyHandle* cfh = nullptr;
+    DBImpl* db_impl = static_cast_with_check<DBImpl>(db->GetRootDB());
+    assert(db_impl);
+    ASSERT_OK(db_impl->CreateColumnFamily(cf_opts, test_cf_name, &cfh));
+    delete cfh;
+  }
+
+  {
+    std::vector<ColumnFamilyDescriptor> cf_descs;
+    cf_descs.emplace_back(kDefaultColumnFamilyName, options);
+    cf_descs.emplace_back(test_cf_name, cf_opts);
+    std::vector<ColumnFamilyHandle*> handles;
+    ASSERT_TRUE(ReOpenNoDelete(cf_descs, &handles).IsInvalidArgument());
+  }
+}
+
+TEST_P(TransactionTest, WriteWithBulkCreatedColumnFamilies) {
+  ColumnFamilyOptions cf_options;
+  WriteOptions write_options;
+
+  std::vector<std::string> cf_names;
+  std::vector<ColumnFamilyHandle*> cf_handles;
+
+  cf_names.push_back("test_cf");
+
+  ASSERT_OK(db->CreateColumnFamilies(cf_options, cf_names, &cf_handles));
+  ASSERT_OK(db->Put(write_options, cf_handles[0], "foo", "bar"));
+  ASSERT_OK(db->DropColumnFamilies(cf_handles));
+
+  for (auto* h : cf_handles) {
+    delete h;
+  }
+  cf_handles.clear();
+
+  std::vector<ColumnFamilyDescriptor> cf_descriptors;
+
+  cf_descriptors.emplace_back("test_cf", ColumnFamilyOptions());
+
+  ASSERT_OK(db->CreateColumnFamilies(cf_options, cf_names, &cf_handles));
+  ASSERT_OK(db->Put(write_options, cf_handles[0], "foo", "bar"));
+  ASSERT_OK(db->DropColumnFamilies(cf_handles));
+  for (auto* h : cf_handles) {
+    delete h;
+  }
+  cf_handles.clear();
+}
+
+}  // namespace ROCKSDB_NAMESPACE
+
+int main(int argc, char** argv) {
+  ROCKSDB_NAMESPACE::port::InstallStackTraceHandler();
+  ::testing::InitGoogleTest(&argc, argv);
+  return RUN_ALL_TESTS();
+}
+
+#else
+#include <stdio.h>
+
+int main(int /*argc*/, char** /*argv*/) {
+  fprintf(stderr,
+          "SKIPPED as Transactions are not supported in ROCKSDB_LITE\n");
+  return 0;
+}
+
+#endif  // ROCKSDB_LITE
diff --git a/src/rocksdb/utilities/transactions/transaction_test.h b/src/rocksdb/utilities/transactions/transaction_test.h
new file mode 100644
index 000000000..0b86453a4
--- /dev/null
+++ b/src/rocksdb/utilities/transactions/transaction_test.h
@@ -0,0 +1,578 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#pragma once
+
+#include <algorithm>
+#include <cinttypes>
+#include <functional>
+#include <string>
+#include <thread>
+
+#include "db/db_impl/db_impl.h"
+#include "db/db_test_util.h"
+#include "port/port.h"
+#include "rocksdb/db.h"
+#include "rocksdb/options.h"
+#include "rocksdb/utilities/transaction.h"
+#include "rocksdb/utilities/transaction_db.h"
+#include "table/mock_table.h"
+#include "test_util/sync_point.h"
+#include "test_util/testharness.h"
+#include "test_util/testutil.h"
+#include "test_util/transaction_test_util.h"
+#include "util/random.h"
+#include "util/string_util.h"
+#include "utilities/fault_injection_env.h"
+#include "utilities/merge_operators.h"
+#include "utilities/merge_operators/string_append/stringappend.h"
+#include "utilities/transactions/pessimistic_transaction_db.h"
+#include "utilities/transactions/write_unprepared_txn_db.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+// Return true if the ith bit is set in combination represented by comb
+bool IsInCombination(size_t i, size_t comb) { return comb & (size_t(1) << i); }
+
+enum WriteOrdering : bool { kOrderedWrite, kUnorderedWrite };
+
+class TransactionTestBase : public ::testing::Test {
+ public:
+  TransactionDB* db;
+  SpecialEnv special_env;
+  FaultInjectionTestEnv* env;
+  std::string dbname;
+  Options options;
+
+  TransactionDBOptions txn_db_options;
+  bool use_stackable_db_;
+
+  TransactionTestBase(bool use_stackable_db, bool two_write_queue,
+                      TxnDBWritePolicy write_policy,
+                      WriteOrdering write_ordering)
+      : db(nullptr),
+        special_env(Env::Default()),
+        env(nullptr),
+        use_stackable_db_(use_stackable_db) {
+    options.create_if_missing = true;
+    options.max_write_buffer_number = 2;
+    options.write_buffer_size = 4 * 1024;
+    options.unordered_write = write_ordering == kUnorderedWrite;
+    options.level0_file_num_compaction_trigger = 2;
+    options.merge_operator = MergeOperators::CreateFromStringId("stringappend");
+    special_env.skip_fsync_ = true;
+    env = new FaultInjectionTestEnv(&special_env);
+    options.env = env;
+    options.two_write_queues = two_write_queue;
+    dbname = test::PerThreadDBPath("transaction_testdb");
+
+    EXPECT_OK(DestroyDB(dbname, options));
+    txn_db_options.transaction_lock_timeout = 0;
+    txn_db_options.default_lock_timeout = 0;
+    txn_db_options.write_policy = write_policy;
+    txn_db_options.rollback_merge_operands = true;
+    // This will stress write unprepared, by forcing write batch flush on every
+    // write.
+    txn_db_options.default_write_batch_flush_threshold = 1;
+    // Write unprepared requires all transactions to be named. This setting
+    // autogenerates the name so that existing tests can pass.
+    txn_db_options.autogenerate_name = true;
+    Status s;
+    if (use_stackable_db == false) {
+      s = TransactionDB::Open(options, txn_db_options, dbname, &db);
+    } else {
+      s = OpenWithStackableDB();
+    }
+    EXPECT_OK(s);
+  }
+
+  ~TransactionTestBase() {
+    delete db;
+    db = nullptr;
+    // This is to skip the assert statement in FaultInjectionTestEnv. There
+    // seems to be a bug in btrfs that the makes readdir return recently
+    // unlink-ed files. By using the default fs we simply ignore errors resulted
+    // from attempting to delete such files in DestroyDB.
+    if (getenv("KEEP_DB") == nullptr) {
+      options.env = Env::Default();
+      EXPECT_OK(DestroyDB(dbname, options));
+    } else {
+      fprintf(stdout, "db is still in %s\n", dbname.c_str());
+    }
+    delete env;
+  }
+
+  Status ReOpenNoDelete() {
+    delete db;
+    db = nullptr;
+    env->AssertNoOpenFile();
+    env->DropUnsyncedFileData();
+    env->ResetState();
+    Status s;
+    if (use_stackable_db_ == false) {
+      s = TransactionDB::Open(options, txn_db_options, dbname, &db);
+    } else {
+      s = OpenWithStackableDB();
+    }
+    assert(!s.ok() || db != nullptr);
+    return s;
+  }
+
+  Status ReOpenNoDelete(std::vector<ColumnFamilyDescriptor>& cfs,
+                        std::vector<ColumnFamilyHandle*>* handles) {
+    for (auto h : *handles) {
+      delete h;
+    }
+    handles->clear();
+    delete db;
+    db = nullptr;
+    env->AssertNoOpenFile();
+    env->DropUnsyncedFileData();
+    env->ResetState();
+    Status s;
+    if (use_stackable_db_ == false) {
+      s = TransactionDB::Open(options, txn_db_options, dbname, cfs, handles,
+                              &db);
+    } else {
+      s = OpenWithStackableDB(cfs, handles);
+    }
+    assert(!s.ok() || db != nullptr);
+    return s;
+  }
+
+  Status ReOpen() {
+    delete db;
+    db = nullptr;
+    DestroyDB(dbname, options);
+    Status s;
+    if (use_stackable_db_ == false) {
+      s = TransactionDB::Open(options, txn_db_options, dbname, &db);
+    } else {
+      s = OpenWithStackableDB();
+    }
+    assert(db != nullptr);
+    return s;
+  }
+
+  Status OpenWithStackableDB(std::vector<ColumnFamilyDescriptor>& cfs,
+                             std::vector<ColumnFamilyHandle*>* handles) {
+    std::vector<size_t> compaction_enabled_cf_indices;
+    TransactionDB::PrepareWrap(&options, &cfs, &compaction_enabled_cf_indices);
+    DB* root_db = nullptr;
+    Options options_copy(options);
+    const bool use_seq_per_batch =
+        txn_db_options.write_policy == WRITE_PREPARED ||
+        txn_db_options.write_policy == WRITE_UNPREPARED;
+    const bool use_batch_per_txn =
+        txn_db_options.write_policy == WRITE_COMMITTED ||
+        txn_db_options.write_policy == WRITE_PREPARED;
+    Status s = DBImpl::Open(options_copy, dbname, cfs, handles, &root_db,
+                            use_seq_per_batch, use_batch_per_txn);
+    auto stackable_db = std::make_unique<StackableDB>(root_db);
+    if (s.ok()) {
+      assert(root_db != nullptr);
+      // If WrapStackableDB() returns non-ok, then stackable_db is already
+      // deleted within WrapStackableDB().
+      s = TransactionDB::WrapStackableDB(stackable_db.release(), txn_db_options,
+                                         compaction_enabled_cf_indices,
+                                         *handles, &db);
+    }
+    return s;
+  }
+
+  Status OpenWithStackableDB() {
+    std::vector<size_t> compaction_enabled_cf_indices;
+    std::vector<ColumnFamilyDescriptor> column_families{ColumnFamilyDescriptor(
+        kDefaultColumnFamilyName, ColumnFamilyOptions(options))};
+
+    TransactionDB::PrepareWrap(&options, &column_families,
+                               &compaction_enabled_cf_indices);
+    std::vector<ColumnFamilyHandle*> handles;
+    DB* root_db = nullptr;
+    Options options_copy(options);
+    const bool use_seq_per_batch =
+        txn_db_options.write_policy == WRITE_PREPARED ||
+        txn_db_options.write_policy == WRITE_UNPREPARED;
+    const bool use_batch_per_txn =
+        txn_db_options.write_policy == WRITE_COMMITTED ||
+        txn_db_options.write_policy == WRITE_PREPARED;
+    Status s = DBImpl::Open(options_copy, dbname, column_families, &handles,
+                            &root_db, use_seq_per_batch, use_batch_per_txn);
+    if (!s.ok()) {
+      delete root_db;
+      return s;
+    }
+    StackableDB* stackable_db = new StackableDB(root_db);
+    assert(root_db != nullptr);
+    assert(handles.size() == 1);
+    s = TransactionDB::WrapStackableDB(stackable_db, txn_db_options,
+                                       compaction_enabled_cf_indices, handles,
+                                       &db);
+    delete handles[0];
+    if (!s.ok()) {
+      delete stackable_db;
+    }
+    return s;
+  }
+
+  std::atomic<size_t> linked = {0};
+  std::atomic<size_t> exp_seq = {0};
+  std::atomic<size_t> commit_writes = {0};
+  std::atomic<size_t> expected_commits = {0};
+  // Without Prepare, the commit does not write to WAL
+  std::atomic<size_t> with_empty_commits = {0};
+  void TestTxn0(size_t index) {
+    // Test DB's internal txn. It involves no prepare phase nor a commit marker.
+    auto s = db->Put(WriteOptions(), "key" + std::to_string(index), "value");
+    ASSERT_OK(s);
+    if (txn_db_options.write_policy == TxnDBWritePolicy::WRITE_COMMITTED) {
+      // Consume one seq per key
+      exp_seq++;
+    } else {
+      // Consume one seq per batch
+      exp_seq++;
+      if (options.two_write_queues) {
+        // Consume one seq for commit
+        exp_seq++;
+      }
+    }
+    with_empty_commits++;
+  }
+
+  void TestTxn1(size_t index) {
+    // Testing directly writing a write batch. Functionality-wise it is
+    // equivalent to commit without prepare.
+    WriteBatch wb;
+    auto istr = std::to_string(index);
+    ASSERT_OK(wb.Put("k1" + istr, "v1"));
+    ASSERT_OK(wb.Put("k2" + istr, "v2"));
+    ASSERT_OK(wb.Put("k3" + istr, "v3"));
+    auto s = db->Write(WriteOptions(), &wb);
+    if (txn_db_options.write_policy == TxnDBWritePolicy::WRITE_COMMITTED) {
+      // Consume one seq per key
+      exp_seq += 3;
+    } else {
+      // Consume one seq per batch
+      exp_seq++;
+      if (options.two_write_queues) {
+        // Consume one seq for commit
+        exp_seq++;
+      }
+    }
+    ASSERT_OK(s);
+    with_empty_commits++;
+  }
+
+  void TestTxn2(size_t index) {
+    // Commit without prepare. It should write to DB without a commit marker.
+    Transaction* txn =
+        db->BeginTransaction(WriteOptions(), TransactionOptions());
+    auto istr = std::to_string(index);
+    ASSERT_OK(txn->SetName("xid" + istr));
+    ASSERT_OK(txn->Put(Slice("foo" + istr), Slice("bar")));
+    ASSERT_OK(txn->Put(Slice("foo2" + istr), Slice("bar2")));
+    ASSERT_OK(txn->Put(Slice("foo3" + istr), Slice("bar3")));
+    ASSERT_OK(txn->Put(Slice("foo4" + istr), Slice("bar4")));
+    ASSERT_OK(txn->Commit());
+    if (txn_db_options.write_policy == TxnDBWritePolicy::WRITE_COMMITTED) {
+      // Consume one seq per key
+      exp_seq += 4;
+    } else if (txn_db_options.write_policy ==
+               TxnDBWritePolicy::WRITE_PREPARED) {
+      // Consume one seq per batch
+      exp_seq++;
+      if (options.two_write_queues) {
+        // Consume one seq for commit
+        exp_seq++;
+      }
+    } else {
+      // Flushed after each key, consume one seq per flushed batch
+      exp_seq += 4;
+      // WriteUnprepared implements CommitWithoutPrepareInternal by simply
+      // calling Prepare then Commit. Consume one seq for the prepare.
+      exp_seq++;
+    }
+    delete txn;
+    with_empty_commits++;
+  }
+
+  void TestTxn3(size_t index) {
+    // A full 2pc txn that also involves a commit marker.
+    Transaction* txn =
+        db->BeginTransaction(WriteOptions(), TransactionOptions());
+    auto istr = std::to_string(index);
+    ASSERT_OK(txn->SetName("xid" + istr));
+    ASSERT_OK(txn->Put(Slice("foo" + istr), Slice("bar")));
+    ASSERT_OK(txn->Put(Slice("foo2" + istr), Slice("bar2")));
+    ASSERT_OK(txn->Put(Slice("foo3" + istr), Slice("bar3")));
+    ASSERT_OK(txn->Put(Slice("foo4" + istr), Slice("bar4")));
+    ASSERT_OK(txn->Put(Slice("foo5" + istr), Slice("bar5")));
+    expected_commits++;
+    ASSERT_OK(txn->Prepare());
+    commit_writes++;
+    ASSERT_OK(txn->Commit());
+    if (txn_db_options.write_policy == TxnDBWritePolicy::WRITE_COMMITTED) {
+      // Consume one seq per key
+      exp_seq += 5;
+    } else if (txn_db_options.write_policy ==
+               TxnDBWritePolicy::WRITE_PREPARED) {
+      // Consume one seq per batch
+      exp_seq++;
+      // Consume one seq per commit marker
+      exp_seq++;
+    } else {
+      // Flushed after each key, consume one seq per flushed batch
+      exp_seq += 5;
+      // Consume one seq per commit marker
+      exp_seq++;
+    }
+    delete txn;
+  }
+
+  void TestTxn4(size_t index) {
+    // A full 2pc txn that also involves a commit marker.
+    Transaction* txn =
+        db->BeginTransaction(WriteOptions(), TransactionOptions());
+    auto istr = std::to_string(index);
+    ASSERT_OK(txn->SetName("xid" + istr));
+    ASSERT_OK(txn->Put(Slice("foo" + istr), Slice("bar")));
+    ASSERT_OK(txn->Put(Slice("foo2" + istr), Slice("bar2")));
+    ASSERT_OK(txn->Put(Slice("foo3" + istr), Slice("bar3")));
+    ASSERT_OK(txn->Put(Slice("foo4" + istr), Slice("bar4")));
+    ASSERT_OK(txn->Put(Slice("foo5" + istr), Slice("bar5")));
+    expected_commits++;
+    ASSERT_OK(txn->Prepare());
+    commit_writes++;
+    ASSERT_OK(txn->Rollback());
+    if (txn_db_options.write_policy == TxnDBWritePolicy::WRITE_COMMITTED) {
+      // No seq is consumed for deleting the txn buffer
+      exp_seq += 0;
+    } else if (txn_db_options.write_policy ==
+               TxnDBWritePolicy::WRITE_PREPARED) {
+      // Consume one seq per batch
+      exp_seq++;
+      // Consume one seq per rollback batch
+      exp_seq++;
+      if (options.two_write_queues) {
+        // Consume one seq for rollback commit
+        exp_seq++;
+      }
+    } else {
+      // Flushed after each key, consume one seq per flushed batch
+      exp_seq += 5;
+      // Consume one seq per rollback batch
+      exp_seq++;
+      if (options.two_write_queues) {
+        // Consume one seq for rollback commit
+        exp_seq++;
+      }
+    }
+    delete txn;
+  }
+
+  // Test that we can change write policy after a clean shutdown (which would
+  // empty the WAL)
+  void CrossCompatibilityTest(TxnDBWritePolicy from_policy,
+                              TxnDBWritePolicy to_policy, bool empty_wal) {
+    TransactionOptions txn_options;
+    ReadOptions read_options;
+    WriteOptions write_options;
+    uint32_t index = 0;
+    Random rnd(1103);
+    options.write_buffer_size = 1024;  // To create more sst files
+    std::unordered_map<std::string, std::string> committed_kvs;
+    Transaction* txn;
+
+    txn_db_options.write_policy = from_policy;
+    if (txn_db_options.write_policy == WRITE_COMMITTED) {
+      options.unordered_write = false;
+    }
+    ASSERT_OK(ReOpen());
+
+    for (int i = 0; i < 1024; i++) {
+      auto istr = std::to_string(index);
+      auto k = Slice("foo-" + istr).ToString();
+      auto v = Slice("bar-" + istr).ToString();
+      // For test the duplicate keys
+      auto v2 = Slice("bar2-" + istr).ToString();
+      auto type = rnd.Uniform(4);
+      switch (type) {
+        case 0:
+          committed_kvs[k] = v;
+          ASSERT_OK(db->Put(write_options, k, v));
+          committed_kvs[k] = v2;
+          ASSERT_OK(db->Put(write_options, k, v2));
+          break;
+        case 1: {
+          WriteBatch wb;
+          committed_kvs[k] = v;
+          ASSERT_OK(wb.Put(k, v));
+          committed_kvs[k] = v2;
+          ASSERT_OK(wb.Put(k, v2));
+          ASSERT_OK(db->Write(write_options, &wb));
+
+        } break;
+        case 2:
+        case 3:
+          txn = db->BeginTransaction(write_options, txn_options);
+          ASSERT_OK(txn->SetName("xid" + istr));
+          committed_kvs[k] = v;
+          ASSERT_OK(txn->Put(k, v));
+          committed_kvs[k] = v2;
+          ASSERT_OK(txn->Put(k, v2));
+
+          if (type == 3) {
+            ASSERT_OK(txn->Prepare());
+          }
+          ASSERT_OK(txn->Commit());
+          delete txn;
+          break;
+        default:
+          FAIL();
+      }
+
+      index++;
+    }  // for i
+
+    txn_db_options.write_policy = to_policy;
+    if (txn_db_options.write_policy == WRITE_COMMITTED) {
+      options.unordered_write = false;
+    }
+    auto db_impl = static_cast_with_check<DBImpl>(db->GetRootDB());
+    // Before upgrade/downgrade the WAL must be emptied
+    if (empty_wal) {
+      ASSERT_OK(db_impl->TEST_FlushMemTable());
+    } else {
+      ASSERT_OK(db_impl->FlushWAL(true));
+    }
+    auto s = ReOpenNoDelete();
+    if (empty_wal) {
+      ASSERT_OK(s);
+    } else {
+      // Test that we can detect the WAL that is produced by an incompatible
+      // WritePolicy and fail fast before mis-interpreting the WAL.
+      ASSERT_TRUE(s.IsNotSupported());
+      return;
+    }
+    db_impl = static_cast_with_check<DBImpl>(db->GetRootDB());
+    // Check that WAL is empty
+    VectorLogPtr log_files;
+    ASSERT_OK(db_impl->GetSortedWalFiles(log_files));
+    ASSERT_EQ(0, log_files.size());
+
+    for (auto& kv : committed_kvs) {
+      std::string value;
+      s = db->Get(read_options, kv.first, &value);
+      if (s.IsNotFound()) {
+        printf("key = %s\n", kv.first.c_str());
+      }
+      ASSERT_OK(s);
+      if (kv.second != value) {
+        printf("key = %s\n", kv.first.c_str());
+      }
+      ASSERT_EQ(kv.second, value);
+    }
+  }
+};
+
+class TransactionTest
+    : public TransactionTestBase,
+      virtual public ::testing::WithParamInterface<
+          std::tuple<bool, bool, TxnDBWritePolicy, WriteOrdering>> {
+ public:
+  TransactionTest()
+      : TransactionTestBase(std::get<0>(GetParam()), std::get<1>(GetParam()),
+                            std::get<2>(GetParam()), std::get<3>(GetParam())){};
+};
+
+class TransactionStressTest : public TransactionTest {};
+
+class MySQLStyleTransactionTest
+    : public TransactionTestBase,
+      virtual public ::testing::WithParamInterface<
+          std::tuple<bool, bool, TxnDBWritePolicy, WriteOrdering, bool>> {
+ public:
+  MySQLStyleTransactionTest()
+      : TransactionTestBase(std::get<0>(GetParam()), std::get<1>(GetParam()),
+                            std::get<2>(GetParam()), std::get<3>(GetParam())),
+        with_slow_threads_(std::get<4>(GetParam())) {
+    if (with_slow_threads_ &&
+        (txn_db_options.write_policy == WRITE_PREPARED ||
+         txn_db_options.write_policy == WRITE_UNPREPARED)) {
+      // The corner case with slow threads involves the caches filling
+      // over which would not happen even with artifial delays. To help
+      // such cases to show up we lower the size of the cache-related data
+      // structures.
+      txn_db_options.wp_snapshot_cache_bits = 1;
+      txn_db_options.wp_commit_cache_bits = 10;
+      options.write_buffer_size = 1024;
+      EXPECT_OK(ReOpen());
+    }
+  };
+
+ protected:
+  // Also emulate slow threads by addin artiftial delays
+  const bool with_slow_threads_;
+};
+
+class WriteCommittedTxnWithTsTest
+    : public TransactionTestBase,
+      public ::testing::WithParamInterface<std::tuple<bool, bool, bool>> {
+ public:
+  WriteCommittedTxnWithTsTest()
+      : TransactionTestBase(std::get<0>(GetParam()), std::get<1>(GetParam()),
+                            WRITE_COMMITTED, kOrderedWrite) {}
+  ~WriteCommittedTxnWithTsTest() override {
+    for (auto* h : handles_) {
+      delete h;
+    }
+  }
+
+  Status GetFromDb(ReadOptions read_opts, ColumnFamilyHandle* column_family,
+                   const Slice& key, TxnTimestamp ts, std::string* value) {
+    std::string ts_buf;
+    PutFixed64(&ts_buf, ts);
+    Slice ts_slc = ts_buf;
+    read_opts.timestamp = &ts_slc;
+    assert(db);
+    return db->Get(read_opts, column_family, key, value);
+  }
+
+  Transaction* NewTxn(WriteOptions write_opts, TransactionOptions txn_opts) {
+    assert(db);
+    auto* txn = db->BeginTransaction(write_opts, txn_opts);
+    assert(txn);
+    const bool enable_indexing = std::get<2>(GetParam());
+    if (enable_indexing) {
+      txn->EnableIndexing();
+    } else {
+      txn->DisableIndexing();
+    }
+    return txn;
+  }
+
+ protected:
+  std::vector<ColumnFamilyHandle*> handles_{};
+};
+
+class TimestampedSnapshotWithTsSanityCheck
+    : public TransactionTestBase,
+      public ::testing::WithParamInterface<
+          std::tuple<bool, bool, TxnDBWritePolicy, WriteOrdering>> {
+ public:
+  explicit TimestampedSnapshotWithTsSanityCheck()
+      : TransactionTestBase(std::get<0>(GetParam()), std::get<1>(GetParam()),
+                            std::get<2>(GetParam()), std::get<3>(GetParam())) {}
+  ~TimestampedSnapshotWithTsSanityCheck() override {
+    for (auto* h : handles_) {
+      delete h;
+    }
+  }
+
+ protected:
+  std::vector<ColumnFamilyHandle*> handles_{};
+};
+
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/utilities/transactions/transaction_util.cc b/src/rocksdb/utilities/transactions/transaction_util.cc
new file mode 100644
index 000000000..360edc8ec
--- /dev/null
+++ b/src/rocksdb/utilities/transactions/transaction_util.cc
@@ -0,0 +1,206 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#ifndef ROCKSDB_LITE
+
+#include "utilities/transactions/transaction_util.h"
+
+#include <cinttypes>
+#include <string>
+#include <vector>
+
+#include "db/db_impl/db_impl.h"
+#include "rocksdb/status.h"
+#include "rocksdb/utilities/write_batch_with_index.h"
+#include "util/cast_util.h"
+#include "util/string_util.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+Status TransactionUtil::CheckKeyForConflicts(
+    DBImpl* db_impl, ColumnFamilyHandle* column_family, const std::string& key,
+    SequenceNumber snap_seq, const std::string* const read_ts, bool cache_only,
+    ReadCallback* snap_checker, SequenceNumber min_uncommitted) {
+  Status result;
+
+  auto cfh = static_cast_with_check<ColumnFamilyHandleImpl>(column_family);
+  auto cfd = cfh->cfd();
+  SuperVersion* sv = db_impl->GetAndRefSuperVersion(cfd);
+
+  if (sv == nullptr) {
+    result = Status::InvalidArgument("Could not access column family " +
+                                     cfh->GetName());
+  }
+
+  if (result.ok()) {
+    SequenceNumber earliest_seq =
+        db_impl->GetEarliestMemTableSequenceNumber(sv, true);
+
+    result = CheckKey(db_impl, sv, earliest_seq, snap_seq, key, read_ts,
+                      cache_only, snap_checker, min_uncommitted);
+
+    db_impl->ReturnAndCleanupSuperVersion(cfd, sv);
+  }
+
+  return result;
+}
+
+Status TransactionUtil::CheckKey(DBImpl* db_impl, SuperVersion* sv,
+                                 SequenceNumber earliest_seq,
+                                 SequenceNumber snap_seq,
+                                 const std::string& key,
+                                 const std::string* const read_ts,
+                                 bool cache_only, ReadCallback* snap_checker,
+                                 SequenceNumber min_uncommitted) {
+  // When `min_uncommitted` is provided, keys are not always committed
+  // in sequence number order, and `snap_checker` is used to check whether
+  // specific sequence number is in the database is visible to the transaction.
+  // So `snap_checker` must be provided.
+  assert(min_uncommitted == kMaxSequenceNumber || snap_checker != nullptr);
+
+  Status result;
+  bool need_to_read_sst = false;
+
+  // Since it would be too slow to check the SST files, we will only use
+  // the memtables to check whether there have been any recent writes
+  // to this key after it was accessed in this transaction.  But if the
+  // Memtables do not contain a long enough history, we must fail the
+  // transaction.
+  if (earliest_seq == kMaxSequenceNumber) {
+    // The age of this memtable is unknown.  Cannot rely on it to check
+    // for recent writes.  This error shouldn't happen often in practice as
+    // the Memtable should have a valid earliest sequence number except in some
+    // corner cases (such as error cases during recovery).
+    need_to_read_sst = true;
+
+    if (cache_only) {
+      result = Status::TryAgain(
+          "Transaction could not check for conflicts as the MemTable does not "
+          "contain a long enough history to check write at SequenceNumber: ",
+          std::to_string(snap_seq));
+    }
+  } else if (snap_seq < earliest_seq || min_uncommitted <= earliest_seq) {
+    // Use <= for min_uncommitted since earliest_seq is actually the largest sec
+    // before this memtable was created
+    need_to_read_sst = true;
+
+    if (cache_only) {
+      // The age of this memtable is too new to use to check for recent
+      // writes.
+      char msg[300];
+      snprintf(msg, sizeof(msg),
+               "Transaction could not check for conflicts for operation at "
+               "SequenceNumber %" PRIu64
+               " as the MemTable only contains changes newer than "
+               "SequenceNumber %" PRIu64
+               ".  Increasing the value of the "
+               "max_write_buffer_size_to_maintain option could reduce the "
+               "frequency "
+               "of this error.",
+               snap_seq, earliest_seq);
+      result = Status::TryAgain(msg);
+    }
+  }
+
+  if (result.ok()) {
+    SequenceNumber seq = kMaxSequenceNumber;
+    std::string timestamp;
+    bool found_record_for_key = false;
+
+    // When min_uncommitted == kMaxSequenceNumber, writes are committed in
+    // sequence number order, so only keys larger than `snap_seq` can cause
+    // conflict.
+    // When min_uncommitted != kMaxSequenceNumber, keys lower than
+    // min_uncommitted will not triggered conflicts, while keys larger than
+    // min_uncommitted might create conflicts, so we need  to read them out
+    // from the DB, and call callback to snap_checker to determine. So only
+    // keys lower than min_uncommitted can be skipped.
+    SequenceNumber lower_bound_seq =
+        (min_uncommitted == kMaxSequenceNumber) ? snap_seq : min_uncommitted;
+    Status s = db_impl->GetLatestSequenceForKey(
+        sv, key, !need_to_read_sst, lower_bound_seq, &seq,
+        !read_ts ? nullptr : &timestamp, &found_record_for_key,
+        /*is_blob_index=*/nullptr);
+
+    if (!(s.ok() || s.IsNotFound() || s.IsMergeInProgress())) {
+      result = s;
+    } else if (found_record_for_key) {
+      bool write_conflict = snap_checker == nullptr
+                                ? snap_seq < seq
+                                : !snap_checker->IsVisible(seq);
+      // Perform conflict checking based on timestamp if applicable.
+      if (!write_conflict && read_ts != nullptr) {
+        ColumnFamilyData* cfd = sv->cfd;
+        assert(cfd);
+        const Comparator* const ucmp = cfd->user_comparator();
+        assert(ucmp);
+        assert(read_ts->size() == ucmp->timestamp_size());
+        assert(read_ts->size() == timestamp.size());
+        // Write conflict if *ts < timestamp.
+        write_conflict = ucmp->CompareTimestamp(*read_ts, timestamp) < 0;
+      }
+      if (write_conflict) {
+        result = Status::Busy();
+      }
+    }
+  }
+
+  return result;
+}
+
+Status TransactionUtil::CheckKeysForConflicts(DBImpl* db_impl,
+                                              const LockTracker& tracker,
+                                              bool cache_only) {
+  Status result;
+
+  std::unique_ptr<LockTracker::ColumnFamilyIterator> cf_it(
+      tracker.GetColumnFamilyIterator());
+  assert(cf_it != nullptr);
+  while (cf_it->HasNext()) {
+    ColumnFamilyId cf = cf_it->Next();
+
+    SuperVersion* sv = db_impl->GetAndRefSuperVersion(cf);
+    if (sv == nullptr) {
+      result = Status::InvalidArgument("Could not access column family " +
+                                       std::to_string(cf));
+      break;
+    }
+
+    SequenceNumber earliest_seq =
+        db_impl->GetEarliestMemTableSequenceNumber(sv, true);
+
+    // For each of the keys in this transaction, check to see if someone has
+    // written to this key since the start of the transaction.
+    std::unique_ptr<LockTracker::KeyIterator> key_it(
+        tracker.GetKeyIterator(cf));
+    assert(key_it != nullptr);
+    while (key_it->HasNext()) {
+      const std::string& key = key_it->Next();
+      PointLockStatus status = tracker.GetPointLockStatus(cf, key);
+      const SequenceNumber key_seq = status.seq;
+
+      // TODO: support timestamp-based conflict checking.
+      // CheckKeysForConflicts() is currently used only by optimistic
+      // transactions.
+      result = CheckKey(db_impl, sv, earliest_seq, key_seq, key,
+                        /*read_ts=*/nullptr, cache_only);
+      if (!result.ok()) {
+        break;
+      }
+    }
+
+    db_impl->ReturnAndCleanupSuperVersion(cf, sv);
+
+    if (!result.ok()) {
+      break;
+    }
+  }
+
+  return result;
+}
+
+}  // namespace ROCKSDB_NAMESPACE
+
+#endif  // ROCKSDB_LITE
diff --git a/src/rocksdb/utilities/transactions/transaction_util.h b/src/rocksdb/utilities/transactions/transaction_util.h
new file mode 100644
index 000000000..a349ba87a
--- /dev/null
+++ b/src/rocksdb/utilities/transactions/transaction_util.h
@@ -0,0 +1,85 @@
+// Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#pragma once
+
+#ifndef ROCKSDB_LITE
+
+#include <string>
+#include <unordered_map>
+
+#include "db/dbformat.h"
+#include "db/read_callback.h"
+#include "rocksdb/db.h"
+#include "rocksdb/slice.h"
+#include "rocksdb/status.h"
+#include "rocksdb/types.h"
+#include "utilities/transactions/lock/lock_tracker.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+class DBImpl;
+struct SuperVersion;
+class WriteBatchWithIndex;
+
+class TransactionUtil {
+ public:
+  // Verifies there have been no commits to this key in the db since this
+  // sequence number. If user-defined timestamp is enabled, then also check
+  // no commits to this key in the db since the given ts.
+  //
+  // If cache_only is true, then this function will not attempt to read any
+  // SST files.  This will make it more likely this function will
+  // return an error if it is unable to determine if there are any conflicts.
+  //
+  // See comment of CheckKey() for explanation of `snap_seq`, `ts`,
+  // `snap_checker` and `min_uncommitted`.
+  //
+  // Returns OK on success, BUSY if there is a conflicting write, or other error
+  // status for any unexpected errors.
+  static Status CheckKeyForConflicts(
+      DBImpl* db_impl, ColumnFamilyHandle* column_family,
+      const std::string& key, SequenceNumber snap_seq,
+      const std::string* const ts, bool cache_only,
+      ReadCallback* snap_checker = nullptr,
+      SequenceNumber min_uncommitted = kMaxSequenceNumber);
+
+  // For each key,SequenceNumber pair tracked by the LockTracker, this function
+  // will verify there have been no writes to the key in the db since that
+  // sequence number.
+  //
+  // Returns OK on success, BUSY if there is a conflicting write, or other error
+  // status for any unexpected errors.
+  //
+  // REQUIRED:
+  // This function should only be called on the write thread or if the
+  // mutex is held.
+  // tracker must support point lock.
+  static Status CheckKeysForConflicts(DBImpl* db_impl,
+                                      const LockTracker& tracker,
+                                      bool cache_only);
+
+ private:
+  // If `snap_checker` == nullptr, writes are always commited in sequence number
+  // order. All sequence number <= `snap_seq` will not conflict with any
+  // write, and all keys > `snap_seq` of `key` will trigger conflict.
+  // If `snap_checker` != nullptr, writes may not commit in sequence number
+  // order. In this case `min_uncommitted` is a lower bound.
+  //  seq < `min_uncommitted`: no conflict
+  //  seq > `snap_seq`: applicable to conflict
+  //  `min_uncommitted` <= seq <= `snap_seq`: call `snap_checker` to determine.
+  //
+  // If user-defined timestamp is enabled, a write conflict is detected if an
+  // operation for `key` with timestamp greater than `ts` exists.
+  static Status CheckKey(DBImpl* db_impl, SuperVersion* sv,
+                         SequenceNumber earliest_seq, SequenceNumber snap_seq,
+                         const std::string& key, const std::string* const ts,
+                         bool cache_only, ReadCallback* snap_checker = nullptr,
+                         SequenceNumber min_uncommitted = kMaxSequenceNumber);
+};
+
+}  // namespace ROCKSDB_NAMESPACE
+
+#endif  // ROCKSDB_LITE
diff --git a/src/rocksdb/utilities/transactions/write_committed_transaction_ts_test.cc b/src/rocksdb/utilities/transactions/write_committed_transaction_ts_test.cc
new file mode 100644
index 000000000..94b8201f7
--- /dev/null
+++ b/src/rocksdb/utilities/transactions/write_committed_transaction_ts_test.cc
@@ -0,0 +1,588 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#include "rocksdb/db.h"
+#include "rocksdb/options.h"
+#include "rocksdb/utilities/transaction_db.h"
+#include "utilities/merge_operators.h"
+#ifndef ROCKSDB_LITE
+
+#include "test_util/testutil.h"
+#include "utilities/transactions/transaction_test.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+INSTANTIATE_TEST_CASE_P(
+    DBAsBaseDB, WriteCommittedTxnWithTsTest,
+    ::testing::Values(std::make_tuple(false, /*two_write_queue=*/false,
+                                      /*enable_indexing=*/false),
+                      std::make_tuple(false, /*two_write_queue=*/true,
+                                      /*enable_indexing=*/false),
+                      std::make_tuple(false, /*two_write_queue=*/false,
+                                      /*enable_indexing=*/true),
+                      std::make_tuple(false, /*two_write_queue=*/true,
+                                      /*enable_indexing=*/true)));
+
+INSTANTIATE_TEST_CASE_P(
+    DBAsStackableDB, WriteCommittedTxnWithTsTest,
+    ::testing::Values(std::make_tuple(true, /*two_write_queue=*/false,
+                                      /*enable_indexing=*/false),
+                      std::make_tuple(true, /*two_write_queue=*/true,
+                                      /*enable_indexing=*/false),
+                      std::make_tuple(true, /*two_write_queue=*/false,
+                                      /*enable_indexing=*/true),
+                      std::make_tuple(true, /*two_write_queue=*/true,
+                                      /*enable_indexing=*/true)));
+
+TEST_P(WriteCommittedTxnWithTsTest, SanityChecks) {
+  ASSERT_OK(ReOpenNoDelete());
+
+  ColumnFamilyOptions cf_opts;
+  cf_opts.comparator = test::BytewiseComparatorWithU64TsWrapper();
+  const std::string test_cf_name = "test_cf";
+  ColumnFamilyHandle* cfh = nullptr;
+  assert(db);
+  ASSERT_OK(db->CreateColumnFamily(cf_opts, test_cf_name, &cfh));
+  delete cfh;
+  cfh = nullptr;
+
+  std::vector<ColumnFamilyDescriptor> cf_descs;
+  cf_descs.emplace_back(kDefaultColumnFamilyName, options);
+  cf_descs.emplace_back(test_cf_name, cf_opts);
+  ASSERT_OK(ReOpenNoDelete(cf_descs, &handles_));
+
+  std::unique_ptr<Transaction> txn(
+      NewTxn(WriteOptions(), TransactionOptions()));
+  assert(txn);
+  ASSERT_OK(txn->Put(handles_[1], "foo", "value"));
+  ASSERT_TRUE(txn->Commit().IsInvalidArgument());
+
+  auto* pessimistic_txn =
+      static_cast_with_check<PessimisticTransaction>(txn.get());
+  ASSERT_TRUE(
+      pessimistic_txn->CommitBatch(/*batch=*/nullptr).IsInvalidArgument());
+
+  {
+    WriteBatchWithIndex* wbwi = txn->GetWriteBatch();
+    assert(wbwi);
+    WriteBatch* wb = wbwi->GetWriteBatch();
+    assert(wb);
+    // Write a key to the batch for nonexisting cf.
+    ASSERT_OK(WriteBatchInternal::Put(wb, /*column_family_id=*/10, /*key=*/"",
+                                      /*value=*/""));
+  }
+
+  ASSERT_OK(txn->SetCommitTimestamp(20));
+
+  ASSERT_TRUE(txn->Commit().IsInvalidArgument());
+  txn.reset();
+
+  std::unique_ptr<Transaction> txn1(
+      NewTxn(WriteOptions(), TransactionOptions()));
+  assert(txn1);
+  ASSERT_OK(txn1->SetName("txn1"));
+  ASSERT_OK(txn1->Put(handles_[1], "foo", "value"));
+  {
+    WriteBatchWithIndex* wbwi = txn1->GetWriteBatch();
+    assert(wbwi);
+    WriteBatch* wb = wbwi->GetWriteBatch();
+    assert(wb);
+    // Write a key to the batch for non-existing cf.
+    ASSERT_OK(WriteBatchInternal::Put(wb, /*column_family_id=*/10, /*key=*/"",
+                                      /*value=*/""));
+  }
+  ASSERT_OK(txn1->Prepare());
+  ASSERT_OK(txn1->SetCommitTimestamp(21));
+  ASSERT_TRUE(txn1->Commit().IsInvalidArgument());
+  txn1.reset();
+}
+
+TEST_P(WriteCommittedTxnWithTsTest, ReOpenWithTimestamp) {
+  options.merge_operator = MergeOperators::CreateUInt64AddOperator();
+  ASSERT_OK(ReOpenNoDelete());
+
+  ColumnFamilyOptions cf_opts;
+  cf_opts.comparator = test::BytewiseComparatorWithU64TsWrapper();
+  const std::string test_cf_name = "test_cf";
+  ColumnFamilyHandle* cfh = nullptr;
+  assert(db);
+  ASSERT_OK(db->CreateColumnFamily(cf_opts, test_cf_name, &cfh));
+  delete cfh;
+  cfh = nullptr;
+
+  std::vector<ColumnFamilyDescriptor> cf_descs;
+  cf_descs.emplace_back(kDefaultColumnFamilyName, options);
+  cf_descs.emplace_back(test_cf_name, cf_opts);
+  ASSERT_OK(ReOpenNoDelete(cf_descs, &handles_));
+
+  std::unique_ptr<Transaction> txn0(
+      NewTxn(WriteOptions(), TransactionOptions()));
+  assert(txn0);
+  ASSERT_OK(txn0->Put(handles_[1], "foo", "value"));
+  ASSERT_OK(txn0->SetName("txn0"));
+  ASSERT_OK(txn0->Prepare());
+  ASSERT_TRUE(txn0->Commit().IsInvalidArgument());
+  txn0.reset();
+
+  std::unique_ptr<Transaction> txn1(
+      NewTxn(WriteOptions(), TransactionOptions()));
+  assert(txn1);
+  ASSERT_OK(txn1->Put(handles_[1], "foo", "value1"));
+  {
+    std::string buf;
+    PutFixed64(&buf, 23);
+    ASSERT_OK(txn1->Put("id", buf));
+    ASSERT_OK(txn1->Merge("id", buf));
+  }
+  ASSERT_OK(txn1->SetName("txn1"));
+  ASSERT_OK(txn1->Prepare());
+  ASSERT_OK(txn1->SetCommitTimestamp(/*ts=*/23));
+  ASSERT_OK(txn1->Commit());
+  txn1.reset();
+
+  {
+    std::string value;
+    const Status s =
+        GetFromDb(ReadOptions(), handles_[1], "foo", /*ts=*/23, &value);
+    ASSERT_OK(s);
+    ASSERT_EQ("value1", value);
+  }
+
+  {
+    std::string value;
+    const Status s = db->Get(ReadOptions(), handles_[0], "id", &value);
+    ASSERT_OK(s);
+    uint64_t ival = 0;
+    Slice value_slc = value;
+    bool result = GetFixed64(&value_slc, &ival);
+    assert(result);
+    ASSERT_EQ(46, ival);
+  }
+}
+
+TEST_P(WriteCommittedTxnWithTsTest, RecoverFromWal) {
+  ASSERT_OK(ReOpenNoDelete());
+
+  ColumnFamilyOptions cf_opts;
+  cf_opts.comparator = test::BytewiseComparatorWithU64TsWrapper();
+  const std::string test_cf_name = "test_cf";
+  ColumnFamilyHandle* cfh = nullptr;
+  assert(db);
+  ASSERT_OK(db->CreateColumnFamily(cf_opts, test_cf_name, &cfh));
+  delete cfh;
+  cfh = nullptr;
+
+  std::vector<ColumnFamilyDescriptor> cf_descs;
+  cf_descs.emplace_back(kDefaultColumnFamilyName, options);
+  cf_descs.emplace_back(test_cf_name, cf_opts);
+  options.avoid_flush_during_shutdown = true;
+  ASSERT_OK(ReOpenNoDelete(cf_descs, &handles_));
+
+  std::unique_ptr<Transaction> txn0(
+      NewTxn(WriteOptions(), TransactionOptions()));
+  assert(txn0);
+  ASSERT_OK(txn0->Put(handles_[1], "foo", "foo_value"));
+  ASSERT_OK(txn0->SetName("txn0"));
+  ASSERT_OK(txn0->Prepare());
+
+  WriteOptions write_opts;
+  write_opts.sync = true;
+  std::unique_ptr<Transaction> txn1(NewTxn(write_opts, TransactionOptions()));
+  assert(txn1);
+  ASSERT_OK(txn1->Put("bar", "bar_value_1"));
+  ASSERT_OK(txn1->Put(handles_[1], "bar", "bar_value_1"));
+  ASSERT_OK(txn1->SetName("txn1"));
+  ASSERT_OK(txn1->Prepare());
+  ASSERT_OK(txn1->SetCommitTimestamp(/*ts=*/23));
+  ASSERT_OK(txn1->Commit());
+  txn1.reset();
+
+  std::unique_ptr<Transaction> txn2(NewTxn(write_opts, TransactionOptions()));
+  assert(txn2);
+  ASSERT_OK(txn2->Put("key1", "value_3"));
+  ASSERT_OK(txn2->Put(handles_[1], "key1", "value_3"));
+  ASSERT_OK(txn2->SetCommitTimestamp(/*ts=*/24));
+  ASSERT_OK(txn2->Commit());
+  txn2.reset();
+
+  txn0.reset();
+
+  ASSERT_OK(ReOpenNoDelete(cf_descs, &handles_));
+
+  {
+    std::string value;
+    Status s = GetFromDb(ReadOptions(), handles_[1], "foo", /*ts=*/23, &value);
+    ASSERT_TRUE(s.IsNotFound());
+
+    s = db->Get(ReadOptions(), handles_[0], "bar", &value);
+    ASSERT_OK(s);
+    ASSERT_EQ("bar_value_1", value);
+
+    value.clear();
+    s = GetFromDb(ReadOptions(), handles_[1], "bar", /*ts=*/23, &value);
+    ASSERT_OK(s);
+    ASSERT_EQ("bar_value_1", value);
+
+    s = GetFromDb(ReadOptions(), handles_[1], "key1", /*ts=*/23, &value);
+    ASSERT_TRUE(s.IsNotFound());
+
+    s = db->Get(ReadOptions(), handles_[0], "key1", &value);
+    ASSERT_OK(s);
+    ASSERT_EQ("value_3", value);
+
+    s = GetFromDb(ReadOptions(), handles_[1], "key1", /*ts=*/24, &value);
+    ASSERT_OK(s);
+    ASSERT_EQ("value_3", value);
+  }
+}
+
+TEST_P(WriteCommittedTxnWithTsTest, TransactionDbLevelApi) {
+  ASSERT_OK(ReOpenNoDelete());
+
+  ColumnFamilyOptions cf_options;
+  cf_options.merge_operator = MergeOperators::CreateStringAppendOperator();
+  cf_options.comparator = test::BytewiseComparatorWithU64TsWrapper();
+  const std::string test_cf_name = "test_cf";
+  ColumnFamilyHandle* cfh = nullptr;
+  assert(db);
+  ASSERT_OK(db->CreateColumnFamily(cf_options, test_cf_name, &cfh));
+  delete cfh;
+  cfh = nullptr;
+
+  std::vector<ColumnFamilyDescriptor> cf_descs;
+  cf_descs.emplace_back(kDefaultColumnFamilyName, options);
+  cf_descs.emplace_back(test_cf_name, cf_options);
+
+  ASSERT_OK(ReOpenNoDelete(cf_descs, &handles_));
+
+  std::string key_str = "tes_key";
+  std::string ts_str;
+  std::string value_str = "test_value";
+  PutFixed64(&ts_str, 100);
+  Slice value = value_str;
+
+  assert(db);
+  ASSERT_TRUE(
+      db->Put(WriteOptions(), handles_[1], "foo", "bar").IsNotSupported());
+  ASSERT_TRUE(db->Delete(WriteOptions(), handles_[1], "foo").IsNotSupported());
+  ASSERT_TRUE(
+      db->SingleDelete(WriteOptions(), handles_[1], "foo").IsNotSupported());
+  ASSERT_TRUE(
+      db->Merge(WriteOptions(), handles_[1], "foo", "+1").IsNotSupported());
+  WriteBatch wb1(/*reserved_bytes=*/0, /*max_bytes=*/0,
+                 /*protection_bytes_per_key=*/0, /*default_cf_ts_sz=*/0);
+  ASSERT_OK(wb1.Put(handles_[1], key_str, ts_str, value));
+  ASSERT_TRUE(db->Write(WriteOptions(), &wb1).IsNotSupported());
+  ASSERT_TRUE(db->Write(WriteOptions(), TransactionDBWriteOptimizations(), &wb1)
+                  .IsNotSupported());
+  auto* pessimistic_txn_db =
+      static_cast_with_check<PessimisticTransactionDB>(db);
+  assert(pessimistic_txn_db);
+  ASSERT_TRUE(
+      pessimistic_txn_db->WriteWithConcurrencyControl(WriteOptions(), &wb1)
+          .IsNotSupported());
+
+  ASSERT_OK(db->Put(WriteOptions(), "foo", "value"));
+  ASSERT_OK(db->Put(WriteOptions(), "bar", "value"));
+  ASSERT_OK(db->Delete(WriteOptions(), "bar"));
+  ASSERT_OK(db->SingleDelete(WriteOptions(), "foo"));
+  ASSERT_OK(db->Put(WriteOptions(), "key", "value"));
+  ASSERT_OK(db->Merge(WriteOptions(), "key", "_more"));
+  WriteBatch wb2(/*reserved_bytes=*/0, /*max_bytes=*/0,
+                 /*protection_bytes_per_key=*/0, /*default_cf_ts_sz=*/0);
+  ASSERT_OK(wb2.Put(key_str, value));
+  ASSERT_OK(db->Write(WriteOptions(), &wb2));
+  ASSERT_OK(db->Write(WriteOptions(), TransactionDBWriteOptimizations(), &wb2));
+  ASSERT_OK(
+      pessimistic_txn_db->WriteWithConcurrencyControl(WriteOptions(), &wb2));
+
+  std::unique_ptr<Transaction> txn(
+      NewTxn(WriteOptions(), TransactionOptions()));
+  assert(txn);
+
+  WriteBatch wb3(/*reserved_bytes=*/0, /*max_bytes=*/0,
+                 /*protection_bytes_per_key=*/0, /*default_cf_ts_sz=*/0);
+
+  ASSERT_OK(wb3.Put(handles_[1], "key", "value"));
+  auto* pessimistic_txn =
+      static_cast_with_check<PessimisticTransaction>(txn.get());
+  assert(pessimistic_txn);
+  ASSERT_TRUE(pessimistic_txn->CommitBatch(&wb3).IsNotSupported());
+
+  txn.reset();
+}
+
+TEST_P(WriteCommittedTxnWithTsTest, Merge) {
+  options.merge_operator = MergeOperators::CreateStringAppendOperator();
+  ASSERT_OK(ReOpenNoDelete());
+
+  ColumnFamilyOptions cf_options;
+  cf_options.comparator = test::BytewiseComparatorWithU64TsWrapper();
+  cf_options.merge_operator = MergeOperators::CreateStringAppendOperator();
+  const std::string test_cf_name = "test_cf";
+  ColumnFamilyHandle* cfh = nullptr;
+  assert(db);
+  ASSERT_OK(db->CreateColumnFamily(cf_options, test_cf_name, &cfh));
+  delete cfh;
+  cfh = nullptr;
+
+  std::vector<ColumnFamilyDescriptor> cf_descs;
+  cf_descs.emplace_back(kDefaultColumnFamilyName, options);
+  cf_descs.emplace_back(test_cf_name, Options(DBOptions(), cf_options));
+  options.avoid_flush_during_shutdown = true;
+
+  ASSERT_OK(ReOpenNoDelete(cf_descs, &handles_));
+
+  std::unique_ptr<Transaction> txn(
+      NewTxn(WriteOptions(), TransactionOptions()));
+  assert(txn);
+  ASSERT_OK(txn->Put(handles_[1], "foo", "bar"));
+  ASSERT_OK(txn->Merge(handles_[1], "foo", "1"));
+  ASSERT_OK(txn->SetCommitTimestamp(24));
+  ASSERT_OK(txn->Commit());
+  txn.reset();
+  {
+    std::string value;
+    const Status s =
+        GetFromDb(ReadOptions(), handles_[1], "foo", /*ts=*/24, &value);
+    ASSERT_OK(s);
+    ASSERT_EQ("bar,1", value);
+  }
+}
+
+TEST_P(WriteCommittedTxnWithTsTest, GetForUpdate) {
+  ASSERT_OK(ReOpenNoDelete());
+
+  ColumnFamilyOptions cf_options;
+  cf_options.comparator = test::BytewiseComparatorWithU64TsWrapper();
+  const std::string test_cf_name = "test_cf";
+  ColumnFamilyHandle* cfh = nullptr;
+  assert(db);
+  ASSERT_OK(db->CreateColumnFamily(cf_options, test_cf_name, &cfh));
+  delete cfh;
+  cfh = nullptr;
+
+  std::vector<ColumnFamilyDescriptor> cf_descs;
+  cf_descs.emplace_back(kDefaultColumnFamilyName, options);
+  cf_descs.emplace_back(test_cf_name, Options(DBOptions(), cf_options));
+  options.avoid_flush_during_shutdown = true;
+
+  ASSERT_OK(ReOpenNoDelete(cf_descs, &handles_));
+
+  std::unique_ptr<Transaction> txn0(
+      NewTxn(WriteOptions(), TransactionOptions()));
+
+  std::unique_ptr<Transaction> txn1(
+      NewTxn(WriteOptions(), TransactionOptions()));
+  ASSERT_OK(txn1->Put(handles_[1], "key", "value1"));
+  ASSERT_OK(txn1->SetCommitTimestamp(24));
+  ASSERT_OK(txn1->Commit());
+  txn1.reset();
+
+  std::string value;
+  ASSERT_OK(txn0->SetReadTimestampForValidation(23));
+  ASSERT_TRUE(
+      txn0->GetForUpdate(ReadOptions(), handles_[1], "key", &value).IsBusy());
+  ASSERT_OK(txn0->Rollback());
+  txn0.reset();
+
+  std::unique_ptr<Transaction> txn2(
+      NewTxn(WriteOptions(), TransactionOptions()));
+  ASSERT_OK(txn2->SetReadTimestampForValidation(25));
+  ASSERT_OK(txn2->GetForUpdate(ReadOptions(), handles_[1], "key", &value));
+  ASSERT_OK(txn2->SetCommitTimestamp(26));
+  ASSERT_OK(txn2->Commit());
+  txn2.reset();
+}
+
+TEST_P(WriteCommittedTxnWithTsTest, BlindWrite) {
+  ASSERT_OK(ReOpenNoDelete());
+
+  ColumnFamilyOptions cf_options;
+  cf_options.comparator = test::BytewiseComparatorWithU64TsWrapper();
+  const std::string test_cf_name = "test_cf";
+  ColumnFamilyHandle* cfh = nullptr;
+  assert(db);
+  ASSERT_OK(db->CreateColumnFamily(cf_options, test_cf_name, &cfh));
+  delete cfh;
+  cfh = nullptr;
+
+  std::vector<ColumnFamilyDescriptor> cf_descs;
+  cf_descs.emplace_back(kDefaultColumnFamilyName, options);
+  cf_descs.emplace_back(test_cf_name, Options(DBOptions(), cf_options));
+  options.avoid_flush_during_shutdown = true;
+  ASSERT_OK(ReOpenNoDelete(cf_descs, &handles_));
+
+  std::unique_ptr<Transaction> txn0(
+      NewTxn(WriteOptions(), TransactionOptions()));
+  assert(txn0);
+  std::unique_ptr<Transaction> txn1(
+      NewTxn(WriteOptions(), TransactionOptions()));
+  assert(txn1);
+
+  {
+    std::string value;
+    ASSERT_OK(txn0->SetReadTimestampForValidation(100));
+    // Lock "key".
+    ASSERT_TRUE(txn0->GetForUpdate(ReadOptions(), handles_[1], "key", &value)
+                    .IsNotFound());
+  }
+
+  ASSERT_OK(txn0->Put(handles_[1], "key", "value0"));
+  ASSERT_OK(txn0->SetCommitTimestamp(101));
+  ASSERT_OK(txn0->Commit());
+
+  ASSERT_OK(txn1->Put(handles_[1], "key", "value1"));
+  // In reality, caller needs to ensure commit_ts of txn1 is greater than the
+  // commit_ts of txn0, which is true for lock-based concurrency control.
+  ASSERT_OK(txn1->SetCommitTimestamp(102));
+  ASSERT_OK(txn1->Commit());
+
+  txn0.reset();
+  txn1.reset();
+}
+
+TEST_P(WriteCommittedTxnWithTsTest, RefineReadTimestamp) {
+  ASSERT_OK(ReOpenNoDelete());
+
+  ColumnFamilyOptions cf_options;
+  cf_options.comparator = test::BytewiseComparatorWithU64TsWrapper();
+  const std::string test_cf_name = "test_cf";
+  ColumnFamilyHandle* cfh = nullptr;
+  assert(db);
+  ASSERT_OK(db->CreateColumnFamily(cf_options, test_cf_name, &cfh));
+  delete cfh;
+  cfh = nullptr;
+
+  std::vector<ColumnFamilyDescriptor> cf_descs;
+  cf_descs.emplace_back(kDefaultColumnFamilyName, options);
+  cf_descs.emplace_back(test_cf_name, Options(DBOptions(), cf_options));
+  options.avoid_flush_during_shutdown = true;
+
+  ASSERT_OK(ReOpenNoDelete(cf_descs, &handles_));
+
+  std::unique_ptr<Transaction> txn0(
+      NewTxn(WriteOptions(), TransactionOptions()));
+  assert(txn0);
+
+  std::unique_ptr<Transaction> txn1(
+      NewTxn(WriteOptions(), TransactionOptions()));
+  assert(txn1);
+
+  {
+    ASSERT_OK(txn0->SetReadTimestampForValidation(100));
+    // Lock "key0", "key1", ..., "key4".
+    for (int i = 0; i < 5; ++i) {
+      std::string value;
+      ASSERT_TRUE(txn0->GetForUpdate(ReadOptions(), handles_[1],
+                                     "key" + std::to_string(i), &value)
+                      .IsNotFound());
+    }
+  }
+  ASSERT_OK(txn1->Put(handles_[1], "key5", "value5_0"));
+  ASSERT_OK(txn1->SetName("txn1"));
+  ASSERT_OK(txn1->Prepare());
+  ASSERT_OK(txn1->SetCommitTimestamp(101));
+  ASSERT_OK(txn1->Commit());
+  txn1.reset();
+
+  {
+    std::string value;
+    ASSERT_TRUE(txn0->GetForUpdate(ReadOptions(), handles_[1], "key5", &value)
+                    .IsBusy());
+    ASSERT_OK(txn0->SetReadTimestampForValidation(102));
+    ASSERT_OK(txn0->GetForUpdate(ReadOptions(), handles_[1], "key5", &value));
+    ASSERT_EQ("value5_0", value);
+  }
+
+  for (int i = 0; i < 6; ++i) {
+    ASSERT_OK(txn0->Put(handles_[1], "key" + std::to_string(i),
+                        "value" + std::to_string(i)));
+  }
+  ASSERT_OK(txn0->SetName("txn0"));
+  ASSERT_OK(txn0->Prepare());
+  ASSERT_OK(txn0->SetCommitTimestamp(103));
+  ASSERT_OK(txn0->Commit());
+  txn0.reset();
+}
+
+TEST_P(WriteCommittedTxnWithTsTest, CheckKeysForConflicts) {
+  options.comparator = test::BytewiseComparatorWithU64TsWrapper();
+  ASSERT_OK(ReOpen());
+
+  std::unique_ptr<Transaction> txn1(
+      db->BeginTransaction(WriteOptions(), TransactionOptions()));
+  assert(txn1);
+
+  std::unique_ptr<Transaction> txn2(
+      db->BeginTransaction(WriteOptions(), TransactionOptions()));
+  assert(txn2);
+  ASSERT_OK(txn2->Put("foo", "v0"));
+  ASSERT_OK(txn2->SetCommitTimestamp(10));
+  ASSERT_OK(txn2->Commit());
+  txn2.reset();
+
+  // txn1 takes a snapshot after txn2 commits. The writes of txn2 have
+  // a smaller seqno than txn1's snapshot, thus should not affect conflict
+  // checking.
+  txn1->SetSnapshot();
+
+  std::unique_ptr<Transaction> txn3(
+      db->BeginTransaction(WriteOptions(), TransactionOptions()));
+  assert(txn3);
+  ASSERT_OK(txn3->SetReadTimestampForValidation(20));
+  std::string dontcare;
+  ASSERT_OK(txn3->GetForUpdate(ReadOptions(), "foo", &dontcare));
+  ASSERT_OK(txn3->SingleDelete("foo"));
+  ASSERT_OK(txn3->SetName("txn3"));
+  ASSERT_OK(txn3->Prepare());
+  ASSERT_OK(txn3->SetCommitTimestamp(30));
+  // txn3 reads at ts=20 > txn2's commit timestamp, and commits at ts=30.
+  // txn3 can commit successfully, leaving a tombstone with ts=30.
+  ASSERT_OK(txn3->Commit());
+  txn3.reset();
+
+  bool called = false;
+  SyncPoint::GetInstance()->DisableProcessing();
+  SyncPoint::GetInstance()->ClearAllCallBacks();
+  SyncPoint::GetInstance()->SetCallBack(
+      "DBImpl::GetLatestSequenceForKey:mem", [&](void* arg) {
+        auto* const ts_ptr = reinterpret_cast<std::string*>(arg);
+        assert(ts_ptr);
+        Slice ts_slc = *ts_ptr;
+        uint64_t last_ts = 0;
+        ASSERT_TRUE(GetFixed64(&ts_slc, &last_ts));
+        ASSERT_EQ(30, last_ts);
+        called = true;
+      });
+  SyncPoint::GetInstance()->EnableProcessing();
+
+  // txn1's read timestamp is 25 < 30 (commit timestamp of txn3). Therefore,
+  // the tombstone written by txn3 causes the conflict checking to fail.
+  ASSERT_OK(txn1->SetReadTimestampForValidation(25));
+  ASSERT_TRUE(txn1->GetForUpdate(ReadOptions(), "foo", &dontcare).IsBusy());
+  ASSERT_TRUE(called);
+
+  SyncPoint::GetInstance()->DisableProcessing();
+  SyncPoint::GetInstance()->ClearAllCallBacks();
+}
+
+}  // namespace ROCKSDB_NAMESPACE
+
+int main(int argc, char** argv) {
+  ROCKSDB_NAMESPACE::port::InstallStackTraceHandler();
+  ::testing::InitGoogleTest(&argc, argv);
+  return RUN_ALL_TESTS();
+}
+
+#else
+#include <cstdio>
+
+int main(int /*argc*/, char** /*argv*/) {
+  fprintf(stderr, "SKIPPED as Transactions not supported in ROCKSDB_LITE\n");
+  return 0;
+}
+
+#endif  // ROCKSDB_LITE
diff --git a/src/rocksdb/utilities/transactions/write_prepared_transaction_test.cc b/src/rocksdb/utilities/transactions/write_prepared_transaction_test.cc
new file mode 100644
index 000000000..86a9511a4
--- /dev/null
+++ b/src/rocksdb/utilities/transactions/write_prepared_transaction_test.cc
@@ -0,0 +1,4078 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#ifndef ROCKSDB_LITE
+
+#include <algorithm>
+#include <atomic>
+#include <cinttypes>
+#include <functional>
+#include <string>
+#include <thread>
+
+#include "db/db_impl/db_impl.h"
+#include "db/dbformat.h"
+#include "port/port.h"
+#include "port/stack_trace.h"
+#include "rocksdb/db.h"
+#include "rocksdb/options.h"
+#include "rocksdb/types.h"
+#include "rocksdb/utilities/debug.h"
+#include "rocksdb/utilities/transaction.h"
+#include "rocksdb/utilities/transaction_db.h"
+#include "table/mock_table.h"
+#include "test_util/sync_point.h"
+#include "test_util/testharness.h"
+#include "test_util/testutil.h"
+#include "test_util/transaction_test_util.h"
+#include "util/mutexlock.h"
+#include "util/random.h"
+#include "util/string_util.h"
+#include "utilities/fault_injection_env.h"
+#include "utilities/merge_operators.h"
+#include "utilities/merge_operators/string_append/stringappend.h"
+#include "utilities/transactions/pessimistic_transaction_db.h"
+#include "utilities/transactions/transaction_test.h"
+#include "utilities/transactions/write_prepared_txn_db.h"
+
+using std::string;
+
+namespace ROCKSDB_NAMESPACE {
+
+using CommitEntry = WritePreparedTxnDB::CommitEntry;
+using CommitEntry64b = WritePreparedTxnDB::CommitEntry64b;
+using CommitEntry64bFormat = WritePreparedTxnDB::CommitEntry64bFormat;
+
+TEST(PreparedHeap, BasicsTest) {
+  WritePreparedTxnDB::PreparedHeap heap;
+  {
+    MutexLock ml(heap.push_pop_mutex());
+    heap.push(14l);
+    // Test with one element
+    ASSERT_EQ(14l, heap.top());
+    heap.push(24l);
+    heap.push(34l);
+    // Test that old min is still on top
+    ASSERT_EQ(14l, heap.top());
+    heap.push(44l);
+    heap.push(54l);
+    heap.push(64l);
+    heap.push(74l);
+    heap.push(84l);
+  }
+  // Test that old min is still on top
+  ASSERT_EQ(14l, heap.top());
+  heap.erase(24l);
+  // Test that old min is still on top
+  ASSERT_EQ(14l, heap.top());
+  heap.erase(14l);
+  // Test that the new comes to the top after multiple erase
+  ASSERT_EQ(34l, heap.top());
+  heap.erase(34l);
+  // Test that the new comes to the top after single erase
+  ASSERT_EQ(44l, heap.top());
+  heap.erase(54l);
+  ASSERT_EQ(44l, heap.top());
+  heap.pop();  // pop 44l
+  // Test that the erased items are ignored after pop
+  ASSERT_EQ(64l, heap.top());
+  heap.erase(44l);
+  // Test that erasing an already popped item would work
+  ASSERT_EQ(64l, heap.top());
+  heap.erase(84l);
+  ASSERT_EQ(64l, heap.top());
+  {
+    MutexLock ml(heap.push_pop_mutex());
+    heap.push(85l);
+    heap.push(86l);
+    heap.push(87l);
+    heap.push(88l);
+    heap.push(89l);
+  }
+  heap.erase(87l);
+  heap.erase(85l);
+  heap.erase(89l);
+  heap.erase(86l);
+  heap.erase(88l);
+  // Test top remains the same after a random order of many erases
+  ASSERT_EQ(64l, heap.top());
+  heap.pop();
+  // Test that pop works with a series of random pending erases
+  ASSERT_EQ(74l, heap.top());
+  ASSERT_FALSE(heap.empty());
+  heap.pop();
+  // Test that empty works
+  ASSERT_TRUE(heap.empty());
+}
+
+// This is a scenario reconstructed from a buggy trace. Test that the bug does
+// not resurface again.
+TEST(PreparedHeap, EmptyAtTheEnd) {
+  WritePreparedTxnDB::PreparedHeap heap;
+  {
+    MutexLock ml(heap.push_pop_mutex());
+    heap.push(40l);
+  }
+  ASSERT_EQ(40l, heap.top());
+  // Although not a recommended scenario, we must be resilient against erase
+  // without a prior push.
+  heap.erase(50l);
+  ASSERT_EQ(40l, heap.top());
+  {
+    MutexLock ml(heap.push_pop_mutex());
+    heap.push(60l);
+  }
+  ASSERT_EQ(40l, heap.top());
+
+  heap.erase(60l);
+  ASSERT_EQ(40l, heap.top());
+  heap.erase(40l);
+  ASSERT_TRUE(heap.empty());
+
+  {
+    MutexLock ml(heap.push_pop_mutex());
+    heap.push(40l);
+  }
+  ASSERT_EQ(40l, heap.top());
+  heap.erase(50l);
+  ASSERT_EQ(40l, heap.top());
+  {
+    MutexLock ml(heap.push_pop_mutex());
+    heap.push(60l);
+  }
+  ASSERT_EQ(40l, heap.top());
+
+  heap.erase(40l);
+  // Test that the erase has not emptied the heap (we had a bug doing that)
+  ASSERT_FALSE(heap.empty());
+  ASSERT_EQ(60l, heap.top());
+  heap.erase(60l);
+  ASSERT_TRUE(heap.empty());
+}
+
+// Generate random order of PreparedHeap access and test that the heap will be
+// successfully emptied at the end.
+TEST(PreparedHeap, Concurrent) {
+  const size_t t_cnt = 10;
+  ROCKSDB_NAMESPACE::port::Thread t[t_cnt + 1];
+  WritePreparedTxnDB::PreparedHeap heap;
+  port::RWMutex prepared_mutex;
+  std::atomic<size_t> last;
+
+  for (size_t n = 0; n < 100; n++) {
+    last = 0;
+    t[0] = ROCKSDB_NAMESPACE::port::Thread([&]() {
+      Random rnd(1103);
+      for (size_t seq = 1; seq <= t_cnt; seq++) {
+        // This is not recommended usage but we should be resilient against it.
+        bool skip_push = rnd.OneIn(5);
+        if (!skip_push) {
+          MutexLock ml(heap.push_pop_mutex());
+          std::this_thread::yield();
+          heap.push(seq);
+          last.store(seq);
+        }
+      }
+    });
+    for (size_t i = 1; i <= t_cnt; i++) {
+      t[i] =
+          ROCKSDB_NAMESPACE::port::Thread([&heap, &prepared_mutex, &last, i]() {
+            auto seq = i;
+            do {
+              std::this_thread::yield();
+            } while (last.load() < seq);
+            WriteLock wl(&prepared_mutex);
+            heap.erase(seq);
+          });
+    }
+    for (size_t i = 0; i <= t_cnt; i++) {
+      t[i].join();
+    }
+    ASSERT_TRUE(heap.empty());
+  }
+}
+
+// Test that WriteBatchWithIndex correctly counts the number of sub-batches
+TEST(WriteBatchWithIndex, SubBatchCnt) {
+  ColumnFamilyOptions cf_options;
+  std::string cf_name = "two";
+  DB* db;
+  Options options;
+  options.create_if_missing = true;
+  const std::string dbname = test::PerThreadDBPath("transaction_testdb");
+  EXPECT_OK(DestroyDB(dbname, options));
+  ASSERT_OK(DB::Open(options, dbname, &db));
+  ColumnFamilyHandle* cf_handle = nullptr;
+  ASSERT_OK(db->CreateColumnFamily(cf_options, cf_name, &cf_handle));
+  WriteOptions write_options;
+  size_t batch_cnt = 1;
+  size_t save_points = 0;
+  std::vector<size_t> batch_cnt_at;
+  WriteBatchWithIndex batch(db->DefaultColumnFamily()->GetComparator(), 0, true,
+                            0);
+  ASSERT_EQ(batch_cnt, batch.SubBatchCnt());
+  batch_cnt_at.push_back(batch_cnt);
+  batch.SetSavePoint();
+  save_points++;
+  ASSERT_OK(batch.Put(Slice("key"), Slice("value")));
+  ASSERT_EQ(batch_cnt, batch.SubBatchCnt());
+  batch_cnt_at.push_back(batch_cnt);
+  batch.SetSavePoint();
+  save_points++;
+  ASSERT_OK(batch.Put(Slice("key2"), Slice("value2")));
+  ASSERT_EQ(batch_cnt, batch.SubBatchCnt());
+  // duplicate the keys
+  batch_cnt_at.push_back(batch_cnt);
+  batch.SetSavePoint();
+  save_points++;
+  ASSERT_OK(batch.Put(Slice("key"), Slice("value3")));
+  batch_cnt++;
+  ASSERT_EQ(batch_cnt, batch.SubBatchCnt());
+  // duplicate the 2nd key. It should not be counted duplicate since a
+  // sub-patch is cut after the last duplicate.
+  batch_cnt_at.push_back(batch_cnt);
+  batch.SetSavePoint();
+  save_points++;
+  ASSERT_OK(batch.Put(Slice("key2"), Slice("value4")));
+  ASSERT_EQ(batch_cnt, batch.SubBatchCnt());
+  // duplicate the keys but in a different cf. It should not be counted as
+  // duplicate keys
+  batch_cnt_at.push_back(batch_cnt);
+  batch.SetSavePoint();
+  save_points++;
+  ASSERT_OK(batch.Put(cf_handle, Slice("key"), Slice("value5")));
+  ASSERT_EQ(batch_cnt, batch.SubBatchCnt());
+
+  // Test that the number of sub-batches matches what we count with
+  // SubBatchCounter
+  std::map<uint32_t, const Comparator*> comparators;
+  comparators[0] = db->DefaultColumnFamily()->GetComparator();
+  comparators[cf_handle->GetID()] = cf_handle->GetComparator();
+  SubBatchCounter counter(comparators);
+  ASSERT_OK(batch.GetWriteBatch()->Iterate(&counter));
+  ASSERT_EQ(batch_cnt, counter.BatchCount());
+
+  // Test that RollbackToSavePoint will properly resets the number of
+  // sub-batches
+  for (size_t i = save_points; i > 0; i--) {
+    ASSERT_OK(batch.RollbackToSavePoint());
+    ASSERT_EQ(batch_cnt_at[i - 1], batch.SubBatchCnt());
+  }
+
+  // Test the count is right with random batches
+  {
+    const size_t TOTAL_KEYS = 20;  // 20 ~= 10 to cause a few randoms
+    Random rnd(1131);
+    std::string keys[TOTAL_KEYS];
+    for (size_t k = 0; k < TOTAL_KEYS; k++) {
+      int len = static_cast<int>(rnd.Uniform(50));
+      keys[k] = test::RandomKey(&rnd, len);
+    }
+    for (size_t i = 0; i < 1000; i++) {  // 1000 random batches
+      WriteBatchWithIndex rndbatch(db->DefaultColumnFamily()->GetComparator(),
+                                   0, true, 0);
+      for (size_t k = 0; k < 10; k++) {  // 10 key per batch
+        size_t ki = static_cast<size_t>(rnd.Uniform(TOTAL_KEYS));
+        Slice key = Slice(keys[ki]);
+        std::string tmp = rnd.RandomString(16);
+        Slice value = Slice(tmp);
+        ASSERT_OK(rndbatch.Put(key, value));
+      }
+      SubBatchCounter batch_counter(comparators);
+      ASSERT_OK(rndbatch.GetWriteBatch()->Iterate(&batch_counter));
+      ASSERT_EQ(rndbatch.SubBatchCnt(), batch_counter.BatchCount());
+    }
+  }
+
+  delete cf_handle;
+  delete db;
+}
+
+TEST(CommitEntry64b, BasicTest) {
+  const size_t INDEX_BITS = static_cast<size_t>(21);
+  const size_t INDEX_SIZE = static_cast<size_t>(1ull << INDEX_BITS);
+  const CommitEntry64bFormat FORMAT(static_cast<size_t>(INDEX_BITS));
+
+  // zero-initialized CommitEntry64b should indicate an empty entry
+  CommitEntry64b empty_entry64b;
+  uint64_t empty_index = 11ul;
+  CommitEntry empty_entry;
+  bool ok = empty_entry64b.Parse(empty_index, &empty_entry, FORMAT);
+  ASSERT_FALSE(ok);
+
+  // the zero entry is reserved for un-initialized entries
+  const size_t MAX_COMMIT = (1 << FORMAT.COMMIT_BITS) - 1 - 1;
+  // Samples over the numbers that are covered by that many index bits
+  std::array<uint64_t, 4> is = {{0, 1, INDEX_SIZE / 2 + 1, INDEX_SIZE - 1}};
+  // Samples over the numbers that are covered by that many commit bits
+  std::array<uint64_t, 4> ds = {{0, 1, MAX_COMMIT / 2 + 1, MAX_COMMIT}};
+  // Iterate over prepare numbers that have i) cover all bits of a sequence
+  // number, and ii) include some bits that fall into the range of index or
+  // commit bits
+  for (uint64_t base = 1; base < kMaxSequenceNumber; base *= 2) {
+    for (uint64_t i : is) {
+      for (uint64_t d : ds) {
+        uint64_t p = base + i + d;
+        for (uint64_t c : {p, p + d / 2, p + d}) {
+          uint64_t index = p % INDEX_SIZE;
+          CommitEntry before(p, c), after;
+          CommitEntry64b entry64b(before, FORMAT);
+          ok = entry64b.Parse(index, &after, FORMAT);
+          ASSERT_TRUE(ok);
+          if (!(before == after)) {
+            printf("base %" PRIu64 " i %" PRIu64 " d %" PRIu64 " p %" PRIu64
+                   " c %" PRIu64 " index %" PRIu64 "\n",
+                   base, i, d, p, c, index);
+          }
+          ASSERT_EQ(before, after);
+        }
+      }
+    }
+  }
+}
+
+class WritePreparedTxnDBMock : public WritePreparedTxnDB {
+ public:
+  WritePreparedTxnDBMock(DBImpl* db_impl, TransactionDBOptions& opt)
+      : WritePreparedTxnDB(db_impl, opt) {}
+  void SetDBSnapshots(const std::vector<SequenceNumber>& snapshots) {
+    snapshots_ = snapshots;
+  }
+  void TakeSnapshot(SequenceNumber seq) { snapshots_.push_back(seq); }
+
+ protected:
+  const std::vector<SequenceNumber> GetSnapshotListFromDB(
+      SequenceNumber /* unused */) override {
+    return snapshots_;
+  }
+
+ private:
+  std::vector<SequenceNumber> snapshots_;
+};
+
+class WritePreparedTransactionTestBase : public TransactionTestBase {
+ public:
+  WritePreparedTransactionTestBase(bool use_stackable_db, bool two_write_queue,
+                                   TxnDBWritePolicy write_policy,
+                                   WriteOrdering write_ordering)
+      : TransactionTestBase(use_stackable_db, two_write_queue, write_policy,
+                            write_ordering){};
+
+ protected:
+  void UpdateTransactionDBOptions(size_t snapshot_cache_bits,
+                                  size_t commit_cache_bits) {
+    txn_db_options.wp_snapshot_cache_bits = snapshot_cache_bits;
+    txn_db_options.wp_commit_cache_bits = commit_cache_bits;
+  }
+  void UpdateTransactionDBOptions(size_t snapshot_cache_bits) {
+    txn_db_options.wp_snapshot_cache_bits = snapshot_cache_bits;
+  }
+  // If expect_update is set, check if it actually updated old_commit_map_. If
+  // it did not and yet suggested not to check the next snapshot, do the
+  // opposite to check if it was not a bad suggestion.
+  void MaybeUpdateOldCommitMapTestWithNext(uint64_t prepare, uint64_t commit,
+                                           uint64_t snapshot,
+                                           uint64_t next_snapshot,
+                                           bool expect_update) {
+    WritePreparedTxnDB* wp_db = dynamic_cast<WritePreparedTxnDB*>(db);
+    // reset old_commit_map_empty_ so that its value indicate whether
+    // old_commit_map_ was updated
+    wp_db->old_commit_map_empty_ = true;
+    bool check_next = wp_db->MaybeUpdateOldCommitMap(prepare, commit, snapshot,
+                                                     snapshot < next_snapshot);
+    if (expect_update == wp_db->old_commit_map_empty_) {
+      printf("prepare: %" PRIu64 " commit: %" PRIu64 " snapshot: %" PRIu64
+             " next: %" PRIu64 "\n",
+             prepare, commit, snapshot, next_snapshot);
+    }
+    EXPECT_EQ(!expect_update, wp_db->old_commit_map_empty_);
+    if (!check_next && wp_db->old_commit_map_empty_) {
+      // do the opposite to make sure it was not a bad suggestion
+      const bool dont_care_bool = true;
+      wp_db->MaybeUpdateOldCommitMap(prepare, commit, next_snapshot,
+                                     dont_care_bool);
+      if (!wp_db->old_commit_map_empty_) {
+        printf("prepare: %" PRIu64 " commit: %" PRIu64 " snapshot: %" PRIu64
+               " next: %" PRIu64 "\n",
+               prepare, commit, snapshot, next_snapshot);
+      }
+      EXPECT_TRUE(wp_db->old_commit_map_empty_);
+    }
+  }
+
+  // Test that a CheckAgainstSnapshots thread reading old_snapshots will not
+  // miss a snapshot because of a concurrent update by UpdateSnapshots that is
+  // writing new_snapshots. Both threads are broken at two points. The sync
+  // points to enforce them are specified by a1, a2, b1, and b2. CommitEntry
+  // entry is expected to be vital for one of the snapshots that is common
+  // between the old and new list of snapshots.
+  void SnapshotConcurrentAccessTestInternal(
+      WritePreparedTxnDB* wp_db,
+      const std::vector<SequenceNumber>& old_snapshots,
+      const std::vector<SequenceNumber>& new_snapshots, CommitEntry& entry,
+      SequenceNumber& version, size_t a1, size_t a2, size_t b1, size_t b2) {
+    // First reset the snapshot list
+    const std::vector<SequenceNumber> empty_snapshots;
+    wp_db->old_commit_map_empty_ = true;
+    wp_db->UpdateSnapshots(empty_snapshots, ++version);
+    // Then initialize it with the old_snapshots
+    wp_db->UpdateSnapshots(old_snapshots, ++version);
+
+    // Starting from the first thread, cut each thread at two points
+    ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->LoadDependency({
+        {"WritePreparedTxnDB::CheckAgainstSnapshots:p:" + std::to_string(a1),
+         "WritePreparedTxnDB::UpdateSnapshots:s:start"},
+        {"WritePreparedTxnDB::UpdateSnapshots:p:" + std::to_string(b1),
+         "WritePreparedTxnDB::CheckAgainstSnapshots:s:" + std::to_string(a1)},
+        {"WritePreparedTxnDB::CheckAgainstSnapshots:p:" + std::to_string(a2),
+         "WritePreparedTxnDB::UpdateSnapshots:s:" + std::to_string(b1)},
+        {"WritePreparedTxnDB::UpdateSnapshots:p:" + std::to_string(b2),
+         "WritePreparedTxnDB::CheckAgainstSnapshots:s:" + std::to_string(a2)},
+        {"WritePreparedTxnDB::CheckAgainstSnapshots:p:end",
+         "WritePreparedTxnDB::UpdateSnapshots:s:" + std::to_string(b2)},
+    });
+    ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->EnableProcessing();
+    {
+      ASSERT_TRUE(wp_db->old_commit_map_empty_);
+      ROCKSDB_NAMESPACE::port::Thread t1(
+          [&]() { wp_db->UpdateSnapshots(new_snapshots, version); });
+      wp_db->CheckAgainstSnapshots(entry);
+      t1.join();
+      ASSERT_FALSE(wp_db->old_commit_map_empty_);
+    }
+    ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->DisableProcessing();
+
+    wp_db->old_commit_map_empty_ = true;
+    wp_db->UpdateSnapshots(empty_snapshots, ++version);
+    wp_db->UpdateSnapshots(old_snapshots, ++version);
+    // Starting from the second thread, cut each thread at two points
+    ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->LoadDependency({
+        {"WritePreparedTxnDB::UpdateSnapshots:p:" + std::to_string(a1),
+         "WritePreparedTxnDB::CheckAgainstSnapshots:s:start"},
+        {"WritePreparedTxnDB::CheckAgainstSnapshots:p:" + std::to_string(b1),
+         "WritePreparedTxnDB::UpdateSnapshots:s:" + std::to_string(a1)},
+        {"WritePreparedTxnDB::UpdateSnapshots:p:" + std::to_string(a2),
+         "WritePreparedTxnDB::CheckAgainstSnapshots:s:" + std::to_string(b1)},
+        {"WritePreparedTxnDB::CheckAgainstSnapshots:p:" + std::to_string(b2),
+         "WritePreparedTxnDB::UpdateSnapshots:s:" + std::to_string(a2)},
+        {"WritePreparedTxnDB::UpdateSnapshots:p:end",
+         "WritePreparedTxnDB::CheckAgainstSnapshots:s:" + std::to_string(b2)},
+    });
+    ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->EnableProcessing();
+    {
+      ASSERT_TRUE(wp_db->old_commit_map_empty_);
+      ROCKSDB_NAMESPACE::port::Thread t1(
+          [&]() { wp_db->UpdateSnapshots(new_snapshots, version); });
+      wp_db->CheckAgainstSnapshots(entry);
+      t1.join();
+      ASSERT_FALSE(wp_db->old_commit_map_empty_);
+    }
+    ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->DisableProcessing();
+  }
+
+  // Verify value of keys.
+  void VerifyKeys(const std::unordered_map<std::string, std::string>& data,
+                  const Snapshot* snapshot = nullptr) {
+    std::string value;
+    ReadOptions read_options;
+    read_options.snapshot = snapshot;
+    for (auto& kv : data) {
+      auto s = db->Get(read_options, kv.first, &value);
+      ASSERT_TRUE(s.ok() || s.IsNotFound());
+      if (s.ok()) {
+        if (kv.second != value) {
+          printf("key = %s\n", kv.first.c_str());
+        }
+        ASSERT_EQ(kv.second, value);
+      } else {
+        ASSERT_EQ(kv.second, "NOT_FOUND");
+      }
+
+      // Try with MultiGet API too
+      std::vector<std::string> values;
+      auto s_vec = db->MultiGet(read_options, {db->DefaultColumnFamily()},
+                                {kv.first}, &values);
+      ASSERT_EQ(1, values.size());
+      ASSERT_EQ(1, s_vec.size());
+      s = s_vec[0];
+      ASSERT_TRUE(s.ok() || s.IsNotFound());
+      if (s.ok()) {
+        ASSERT_TRUE(kv.second == values[0]);
+      } else {
+        ASSERT_EQ(kv.second, "NOT_FOUND");
+      }
+    }
+  }
+
+  // Verify all versions of keys.
+  void VerifyInternalKeys(const std::vector<KeyVersion>& expected_versions) {
+    std::vector<KeyVersion> versions;
+    const size_t kMaxKeys = 100000;
+    ASSERT_OK(GetAllKeyVersions(db, expected_versions.front().user_key,
+                                expected_versions.back().user_key, kMaxKeys,
+                                &versions));
+    ASSERT_EQ(expected_versions.size(), versions.size());
+    for (size_t i = 0; i < versions.size(); i++) {
+      ASSERT_EQ(expected_versions[i].user_key, versions[i].user_key);
+      ASSERT_EQ(expected_versions[i].sequence, versions[i].sequence);
+      ASSERT_EQ(expected_versions[i].type, versions[i].type);
+      if (versions[i].type != kTypeDeletion &&
+          versions[i].type != kTypeSingleDeletion) {
+        ASSERT_EQ(expected_versions[i].value, versions[i].value);
+      }
+      // Range delete not supported.
+      ASSERT_NE(expected_versions[i].type, kTypeRangeDeletion);
+    }
+  }
+};
+
+class WritePreparedTransactionTest
+    : public WritePreparedTransactionTestBase,
+      virtual public ::testing::WithParamInterface<
+          std::tuple<bool, bool, TxnDBWritePolicy, WriteOrdering>> {
+ public:
+  WritePreparedTransactionTest()
+      : WritePreparedTransactionTestBase(
+            std::get<0>(GetParam()), std::get<1>(GetParam()),
+            std::get<2>(GetParam()), std::get<3>(GetParam())){};
+};
+
+#if !defined(ROCKSDB_VALGRIND_RUN) || defined(ROCKSDB_FULL_VALGRIND_RUN)
+class SnapshotConcurrentAccessTest
+    : public WritePreparedTransactionTestBase,
+      virtual public ::testing::WithParamInterface<std::tuple<
+          bool, bool, TxnDBWritePolicy, WriteOrdering, size_t, size_t>> {
+ public:
+  SnapshotConcurrentAccessTest()
+      : WritePreparedTransactionTestBase(
+            std::get<0>(GetParam()), std::get<1>(GetParam()),
+            std::get<2>(GetParam()), std::get<3>(GetParam())),
+        split_id_(std::get<4>(GetParam())),
+        split_cnt_(std::get<5>(GetParam())){};
+
+ protected:
+  // A test is split into split_cnt_ tests, each identified with split_id_ where
+  // 0 <= split_id_ < split_cnt_
+  size_t split_id_;
+  size_t split_cnt_;
+};
+#endif  // !defined(ROCKSDB_VALGRIND_RUN) || defined(ROCKSDB_FULL_VALGRIND_RUN)
+
+class SeqAdvanceConcurrentTest
+    : public WritePreparedTransactionTestBase,
+      virtual public ::testing::WithParamInterface<std::tuple<
+          bool, bool, TxnDBWritePolicy, WriteOrdering, size_t, size_t>> {
+ public:
+  SeqAdvanceConcurrentTest()
+      : WritePreparedTransactionTestBase(
+            std::get<0>(GetParam()), std::get<1>(GetParam()),
+            std::get<2>(GetParam()), std::get<3>(GetParam())),
+        split_id_(std::get<4>(GetParam())),
+        split_cnt_(std::get<5>(GetParam())) {
+    special_env.skip_fsync_ = true;
+  };
+
+ protected:
+  // A test is split into split_cnt_ tests, each identified with split_id_ where
+  // 0 <= split_id_ < split_cnt_
+  size_t split_id_;
+  size_t split_cnt_;
+};
+
+INSTANTIATE_TEST_CASE_P(
+    WritePreparedTransaction, WritePreparedTransactionTest,
+    ::testing::Values(
+        std::make_tuple(false, false, WRITE_PREPARED, kOrderedWrite),
+        std::make_tuple(false, true, WRITE_PREPARED, kOrderedWrite),
+        std::make_tuple(false, true, WRITE_PREPARED, kUnorderedWrite)));
+
+#if !defined(ROCKSDB_VALGRIND_RUN) || defined(ROCKSDB_FULL_VALGRIND_RUN)
+INSTANTIATE_TEST_CASE_P(
+    TwoWriteQueues, SnapshotConcurrentAccessTest,
+    ::testing::Values(
+        std::make_tuple(false, true, WRITE_PREPARED, kOrderedWrite, 0, 20),
+        std::make_tuple(false, true, WRITE_PREPARED, kOrderedWrite, 1, 20),
+        std::make_tuple(false, true, WRITE_PREPARED, kOrderedWrite, 2, 20),
+        std::make_tuple(false, true, WRITE_PREPARED, kOrderedWrite, 3, 20),
+        std::make_tuple(false, true, WRITE_PREPARED, kOrderedWrite, 4, 20),
+        std::make_tuple(false, true, WRITE_PREPARED, kOrderedWrite, 5, 20),
+        std::make_tuple(false, true, WRITE_PREPARED, kOrderedWrite, 6, 20),
+        std::make_tuple(false, true, WRITE_PREPARED, kOrderedWrite, 7, 20),
+        std::make_tuple(false, true, WRITE_PREPARED, kOrderedWrite, 8, 20),
+        std::make_tuple(false, true, WRITE_PREPARED, kOrderedWrite, 9, 20),
+        std::make_tuple(false, true, WRITE_PREPARED, kOrderedWrite, 10, 20),
+        std::make_tuple(false, true, WRITE_PREPARED, kOrderedWrite, 11, 20),
+        std::make_tuple(false, true, WRITE_PREPARED, kOrderedWrite, 12, 20),
+        std::make_tuple(false, true, WRITE_PREPARED, kOrderedWrite, 13, 20),
+        std::make_tuple(false, true, WRITE_PREPARED, kOrderedWrite, 14, 20),
+        std::make_tuple(false, true, WRITE_PREPARED, kOrderedWrite, 15, 20),
+        std::make_tuple(false, true, WRITE_PREPARED, kOrderedWrite, 16, 20),
+        std::make_tuple(false, true, WRITE_PREPARED, kOrderedWrite, 17, 20),
+        std::make_tuple(false, true, WRITE_PREPARED, kOrderedWrite, 18, 20),
+        std::make_tuple(false, true, WRITE_PREPARED, kOrderedWrite, 19, 20),
+
+        std::make_tuple(false, true, WRITE_PREPARED, kUnorderedWrite, 0, 20),
+        std::make_tuple(false, true, WRITE_PREPARED, kUnorderedWrite, 1, 20),
+        std::make_tuple(false, true, WRITE_PREPARED, kUnorderedWrite, 2, 20),
+        std::make_tuple(false, true, WRITE_PREPARED, kUnorderedWrite, 3, 20),
+        std::make_tuple(false, true, WRITE_PREPARED, kUnorderedWrite, 4, 20),
+        std::make_tuple(false, true, WRITE_PREPARED, kUnorderedWrite, 5, 20),
+        std::make_tuple(false, true, WRITE_PREPARED, kUnorderedWrite, 6, 20),
+        std::make_tuple(false, true, WRITE_PREPARED, kUnorderedWrite, 7, 20),
+        std::make_tuple(false, true, WRITE_PREPARED, kUnorderedWrite, 8, 20),
+        std::make_tuple(false, true, WRITE_PREPARED, kUnorderedWrite, 9, 20),
+        std::make_tuple(false, true, WRITE_PREPARED, kUnorderedWrite, 10, 20),
+        std::make_tuple(false, true, WRITE_PREPARED, kUnorderedWrite, 11, 20),
+        std::make_tuple(false, true, WRITE_PREPARED, kUnorderedWrite, 12, 20),
+        std::make_tuple(false, true, WRITE_PREPARED, kUnorderedWrite, 13, 20),
+        std::make_tuple(false, true, WRITE_PREPARED, kUnorderedWrite, 14, 20),
+        std::make_tuple(false, true, WRITE_PREPARED, kUnorderedWrite, 15, 20),
+        std::make_tuple(false, true, WRITE_PREPARED, kUnorderedWrite, 16, 20),
+        std::make_tuple(false, true, WRITE_PREPARED, kUnorderedWrite, 17, 20),
+        std::make_tuple(false, true, WRITE_PREPARED, kUnorderedWrite, 18, 20),
+        std::make_tuple(false, true, WRITE_PREPARED, kUnorderedWrite, 19, 20)));
+
+INSTANTIATE_TEST_CASE_P(
+    OneWriteQueue, SnapshotConcurrentAccessTest,
+    ::testing::Values(
+        std::make_tuple(false, false, WRITE_PREPARED, kOrderedWrite, 0, 20),
+        std::make_tuple(false, false, WRITE_PREPARED, kOrderedWrite, 1, 20),
+        std::make_tuple(false, false, WRITE_PREPARED, kOrderedWrite, 2, 20),
+        std::make_tuple(false, false, WRITE_PREPARED, kOrderedWrite, 3, 20),
+        std::make_tuple(false, false, WRITE_PREPARED, kOrderedWrite, 4, 20),
+        std::make_tuple(false, false, WRITE_PREPARED, kOrderedWrite, 5, 20),
+        std::make_tuple(false, false, WRITE_PREPARED, kOrderedWrite, 6, 20),
+        std::make_tuple(false, false, WRITE_PREPARED, kOrderedWrite, 7, 20),
+        std::make_tuple(false, false, WRITE_PREPARED, kOrderedWrite, 8, 20),
+        std::make_tuple(false, false, WRITE_PREPARED, kOrderedWrite, 9, 20),
+        std::make_tuple(false, false, WRITE_PREPARED, kOrderedWrite, 10, 20),
+        std::make_tuple(false, false, WRITE_PREPARED, kOrderedWrite, 11, 20),
+        std::make_tuple(false, false, WRITE_PREPARED, kOrderedWrite, 12, 20),
+        std::make_tuple(false, false, WRITE_PREPARED, kOrderedWrite, 13, 20),
+        std::make_tuple(false, false, WRITE_PREPARED, kOrderedWrite, 14, 20),
+        std::make_tuple(false, false, WRITE_PREPARED, kOrderedWrite, 15, 20),
+        std::make_tuple(false, false, WRITE_PREPARED, kOrderedWrite, 16, 20),
+        std::make_tuple(false, false, WRITE_PREPARED, kOrderedWrite, 17, 20),
+        std::make_tuple(false, false, WRITE_PREPARED, kOrderedWrite, 18, 20),
+        std::make_tuple(false, false, WRITE_PREPARED, kOrderedWrite, 19, 20)));
+
+INSTANTIATE_TEST_CASE_P(
+    TwoWriteQueues, SeqAdvanceConcurrentTest,
+    ::testing::Values(
+        std::make_tuple(false, true, WRITE_PREPARED, kOrderedWrite, 0, 10),
+        std::make_tuple(false, true, WRITE_PREPARED, kOrderedWrite, 1, 10),
+        std::make_tuple(false, true, WRITE_PREPARED, kOrderedWrite, 2, 10),
+        std::make_tuple(false, true, WRITE_PREPARED, kOrderedWrite, 3, 10),
+        std::make_tuple(false, true, WRITE_PREPARED, kOrderedWrite, 4, 10),
+        std::make_tuple(false, true, WRITE_PREPARED, kOrderedWrite, 5, 10),
+        std::make_tuple(false, true, WRITE_PREPARED, kOrderedWrite, 6, 10),
+        std::make_tuple(false, true, WRITE_PREPARED, kOrderedWrite, 7, 10),
+        std::make_tuple(false, true, WRITE_PREPARED, kOrderedWrite, 8, 10),
+        std::make_tuple(false, true, WRITE_PREPARED, kOrderedWrite, 9, 10),
+        std::make_tuple(false, true, WRITE_PREPARED, kUnorderedWrite, 0, 10),
+        std::make_tuple(false, true, WRITE_PREPARED, kUnorderedWrite, 1, 10),
+        std::make_tuple(false, true, WRITE_PREPARED, kUnorderedWrite, 2, 10),
+        std::make_tuple(false, true, WRITE_PREPARED, kUnorderedWrite, 3, 10),
+        std::make_tuple(false, true, WRITE_PREPARED, kUnorderedWrite, 4, 10),
+        std::make_tuple(false, true, WRITE_PREPARED, kUnorderedWrite, 5, 10),
+        std::make_tuple(false, true, WRITE_PREPARED, kUnorderedWrite, 6, 10),
+        std::make_tuple(false, true, WRITE_PREPARED, kUnorderedWrite, 7, 10),
+        std::make_tuple(false, true, WRITE_PREPARED, kUnorderedWrite, 8, 10),
+        std::make_tuple(false, true, WRITE_PREPARED, kUnorderedWrite, 9, 10)));
+
+INSTANTIATE_TEST_CASE_P(
+    OneWriteQueue, SeqAdvanceConcurrentTest,
+    ::testing::Values(
+        std::make_tuple(false, false, WRITE_PREPARED, kOrderedWrite, 0, 10),
+        std::make_tuple(false, false, WRITE_PREPARED, kOrderedWrite, 1, 10),
+        std::make_tuple(false, false, WRITE_PREPARED, kOrderedWrite, 2, 10),
+        std::make_tuple(false, false, WRITE_PREPARED, kOrderedWrite, 3, 10),
+        std::make_tuple(false, false, WRITE_PREPARED, kOrderedWrite, 4, 10),
+        std::make_tuple(false, false, WRITE_PREPARED, kOrderedWrite, 5, 10),
+        std::make_tuple(false, false, WRITE_PREPARED, kOrderedWrite, 6, 10),
+        std::make_tuple(false, false, WRITE_PREPARED, kOrderedWrite, 7, 10),
+        std::make_tuple(false, false, WRITE_PREPARED, kOrderedWrite, 8, 10),
+        std::make_tuple(false, false, WRITE_PREPARED, kOrderedWrite, 9, 10)));
+#endif  // !defined(ROCKSDB_VALGRIND_RUN) || defined(ROCKSDB_FULL_VALGRIND_RUN)
+
+TEST_P(WritePreparedTransactionTest, CommitMap) {
+  WritePreparedTxnDB* wp_db = dynamic_cast<WritePreparedTxnDB*>(db);
+  ASSERT_NE(wp_db, nullptr);
+  ASSERT_NE(wp_db->db_impl_, nullptr);
+  size_t size = wp_db->COMMIT_CACHE_SIZE;
+  CommitEntry c = {5, 12}, e;
+  bool evicted = wp_db->AddCommitEntry(c.prep_seq % size, c, &e);
+  ASSERT_FALSE(evicted);
+
+  // Should be able to read the same value
+  CommitEntry64b dont_care;
+  bool found = wp_db->GetCommitEntry(c.prep_seq % size, &dont_care, &e);
+  ASSERT_TRUE(found);
+  ASSERT_EQ(c, e);
+  // Should be able to distinguish between overlapping entries
+  found = wp_db->GetCommitEntry((c.prep_seq + size) % size, &dont_care, &e);
+  ASSERT_TRUE(found);
+  ASSERT_NE(c.prep_seq + size, e.prep_seq);
+  // Should be able to detect non-existent entry
+  found = wp_db->GetCommitEntry((c.prep_seq + 1) % size, &dont_care, &e);
+  ASSERT_FALSE(found);
+
+  // Reject an invalid exchange
+  CommitEntry e2 = {c.prep_seq + size, c.commit_seq + size};
+  CommitEntry64b e2_64b(e2, wp_db->FORMAT);
+  bool exchanged = wp_db->ExchangeCommitEntry(e2.prep_seq % size, e2_64b, e);
+  ASSERT_FALSE(exchanged);
+  // check whether it did actually reject that
+  found = wp_db->GetCommitEntry(e2.prep_seq % size, &dont_care, &e);
+  ASSERT_TRUE(found);
+  ASSERT_EQ(c, e);
+
+  // Accept a valid exchange
+  CommitEntry64b c_64b(c, wp_db->FORMAT);
+  CommitEntry e3 = {c.prep_seq + size, c.commit_seq + size + 1};
+  exchanged = wp_db->ExchangeCommitEntry(c.prep_seq % size, c_64b, e3);
+  ASSERT_TRUE(exchanged);
+  // check whether it did actually accepted that
+  found = wp_db->GetCommitEntry(c.prep_seq % size, &dont_care, &e);
+  ASSERT_TRUE(found);
+  ASSERT_EQ(e3, e);
+
+  // Rewrite an entry
+  CommitEntry e4 = {e3.prep_seq + size, e3.commit_seq + size + 1};
+  evicted = wp_db->AddCommitEntry(e4.prep_seq % size, e4, &e);
+  ASSERT_TRUE(evicted);
+  ASSERT_EQ(e3, e);
+  found = wp_db->GetCommitEntry(e4.prep_seq % size, &dont_care, &e);
+  ASSERT_TRUE(found);
+  ASSERT_EQ(e4, e);
+}
+
+TEST_P(WritePreparedTransactionTest, MaybeUpdateOldCommitMap) {
+  // If prepare <= snapshot < commit we should keep the entry around since its
+  // nonexistence could be interpreted as committed in the snapshot while it is
+  // not true. We keep such entries around by adding them to the
+  // old_commit_map_.
+  uint64_t p /*prepare*/, c /*commit*/, s /*snapshot*/, ns /*next_snapshot*/;
+  p = 10l, c = 15l, s = 20l, ns = 21l;
+  MaybeUpdateOldCommitMapTestWithNext(p, c, s, ns, false);
+  // If we do not expect the old commit map to be updated, try also with a next
+  // snapshot that is expected to update the old commit map. This would test
+  // that MaybeUpdateOldCommitMap would not prevent us from checking the next
+  // snapshot that must be checked.
+  p = 10l, c = 15l, s = 20l, ns = 11l;
+  MaybeUpdateOldCommitMapTestWithNext(p, c, s, ns, false);
+
+  p = 10l, c = 20l, s = 20l, ns = 19l;
+  MaybeUpdateOldCommitMapTestWithNext(p, c, s, ns, false);
+  p = 10l, c = 20l, s = 20l, ns = 21l;
+  MaybeUpdateOldCommitMapTestWithNext(p, c, s, ns, false);
+
+  p = 20l, c = 20l, s = 20l, ns = 21l;
+  MaybeUpdateOldCommitMapTestWithNext(p, c, s, ns, false);
+  p = 20l, c = 20l, s = 20l, ns = 19l;
+  MaybeUpdateOldCommitMapTestWithNext(p, c, s, ns, false);
+
+  p = 10l, c = 25l, s = 20l, ns = 21l;
+  MaybeUpdateOldCommitMapTestWithNext(p, c, s, ns, true);
+
+  p = 20l, c = 25l, s = 20l, ns = 21l;
+  MaybeUpdateOldCommitMapTestWithNext(p, c, s, ns, true);
+
+  p = 21l, c = 25l, s = 20l, ns = 22l;
+  MaybeUpdateOldCommitMapTestWithNext(p, c, s, ns, false);
+  p = 21l, c = 25l, s = 20l, ns = 19l;
+  MaybeUpdateOldCommitMapTestWithNext(p, c, s, ns, false);
+}
+
+// Trigger the condition where some old memtables are skipped when doing
+// TransactionUtil::CheckKey(), and make sure the result is still correct.
+TEST_P(WritePreparedTransactionTest, CheckKeySkipOldMemtable) {
+  const int kAttemptHistoryMemtable = 0;
+  const int kAttemptImmMemTable = 1;
+  for (int attempt = kAttemptHistoryMemtable; attempt <= kAttemptImmMemTable;
+       attempt++) {
+    options.max_write_buffer_number_to_maintain = 3;
+    ASSERT_OK(ReOpen());
+
+    WriteOptions write_options;
+    ReadOptions read_options;
+    TransactionOptions txn_options;
+    txn_options.set_snapshot = true;
+    string value;
+
+    ASSERT_OK(db->Put(write_options, Slice("foo"), Slice("bar")));
+    ASSERT_OK(db->Put(write_options, Slice("foo2"), Slice("bar")));
+
+    Transaction* txn = db->BeginTransaction(write_options, txn_options);
+    ASSERT_TRUE(txn != nullptr);
+    ASSERT_OK(txn->SetName("txn"));
+
+    Transaction* txn2 = db->BeginTransaction(write_options, txn_options);
+    ASSERT_TRUE(txn2 != nullptr);
+    ASSERT_OK(txn2->SetName("txn2"));
+
+    // This transaction is created to cause potential conflict.
+    Transaction* txn_x = db->BeginTransaction(write_options);
+    ASSERT_OK(txn_x->SetName("txn_x"));
+    ASSERT_OK(txn_x->Put(Slice("foo"), Slice("bar3")));
+    ASSERT_OK(txn_x->Prepare());
+
+    // Create snapshots after the prepare, but there should still
+    // be a conflict when trying to read "foo".
+
+    if (attempt == kAttemptImmMemTable) {
+      // For the second attempt, hold flush from beginning. The memtable
+      // will be switched to immutable after calling TEST_SwitchMemtable()
+      // while CheckKey() is called.
+      ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->LoadDependency(
+          {{"WritePreparedTransactionTest.CheckKeySkipOldMemtable",
+            "FlushJob::Start"}});
+      ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->EnableProcessing();
+    }
+
+    // force a memtable flush. The memtable should still be kept
+    FlushOptions flush_ops;
+    if (attempt == kAttemptHistoryMemtable) {
+      ASSERT_OK(db->Flush(flush_ops));
+    } else {
+      ASSERT_EQ(attempt, kAttemptImmMemTable);
+      DBImpl* db_impl = static_cast<DBImpl*>(db->GetRootDB());
+      ASSERT_OK(db_impl->TEST_SwitchMemtable());
+    }
+    uint64_t num_imm_mems;
+    ASSERT_TRUE(db->GetIntProperty(DB::Properties::kNumImmutableMemTable,
+                                   &num_imm_mems));
+    if (attempt == kAttemptHistoryMemtable) {
+      ASSERT_EQ(0, num_imm_mems);
+    } else {
+      ASSERT_EQ(attempt, kAttemptImmMemTable);
+      ASSERT_EQ(1, num_imm_mems);
+    }
+
+    // Put something in active memtable
+    ASSERT_OK(db->Put(write_options, Slice("foo3"), Slice("bar")));
+
+    // Create txn3 after flushing, but this transaction also needs to
+    // check all memtables because of they contains uncommitted data.
+    Transaction* txn3 = db->BeginTransaction(write_options, txn_options);
+    ASSERT_TRUE(txn3 != nullptr);
+    ASSERT_OK(txn3->SetName("txn3"));
+
+    // Commit the pending write
+    ASSERT_OK(txn_x->Commit());
+
+    // Commit txn, txn2 and tx3. txn and tx3 will conflict but txn2 will
+    // pass. In all cases, both memtables are queried.
+    SetPerfLevel(PerfLevel::kEnableCount);
+    get_perf_context()->Reset();
+    ASSERT_TRUE(txn3->GetForUpdate(read_options, "foo", &value).IsBusy());
+    // We should have checked two memtables, active and either immutable
+    // or history memtable, depending on the test case.
+    ASSERT_EQ(2, get_perf_context()->get_from_memtable_count);
+
+    get_perf_context()->Reset();
+    ASSERT_TRUE(txn->GetForUpdate(read_options, "foo", &value).IsBusy());
+    // We should have checked two memtables, active and either immutable
+    // or history memtable, depending on the test case.
+    ASSERT_EQ(2, get_perf_context()->get_from_memtable_count);
+
+    get_perf_context()->Reset();
+    ASSERT_OK(txn2->GetForUpdate(read_options, "foo2", &value));
+    ASSERT_EQ(value, "bar");
+    // We should have checked two memtables, and since there is no
+    // conflict, another Get() will be made and fetch the data from
+    // DB. If it is in immutable memtable, two extra memtable reads
+    // will be issued. If it is not (in history), only one will
+    // be made, which is to the active memtable.
+    if (attempt == kAttemptHistoryMemtable) {
+      ASSERT_EQ(3, get_perf_context()->get_from_memtable_count);
+    } else {
+      ASSERT_EQ(attempt, kAttemptImmMemTable);
+      ASSERT_EQ(4, get_perf_context()->get_from_memtable_count);
+    }
+
+    Transaction* txn4 = db->BeginTransaction(write_options, txn_options);
+    ASSERT_TRUE(txn4 != nullptr);
+    ASSERT_OK(txn4->SetName("txn4"));
+    get_perf_context()->Reset();
+    ASSERT_OK(txn4->GetForUpdate(read_options, "foo", &value));
+    if (attempt == kAttemptHistoryMemtable) {
+      // Active memtable will be checked in snapshot validation and when
+      // getting the value.
+      ASSERT_EQ(2, get_perf_context()->get_from_memtable_count);
+    } else {
+      // Only active memtable will be checked in snapshot validation but
+      // both of active and immutable snapshot will be queried when
+      // getting the value.
+      ASSERT_EQ(attempt, kAttemptImmMemTable);
+      ASSERT_EQ(3, get_perf_context()->get_from_memtable_count);
+    }
+
+    ASSERT_OK(txn2->Commit());
+    ASSERT_OK(txn4->Commit());
+
+    TEST_SYNC_POINT("WritePreparedTransactionTest.CheckKeySkipOldMemtable");
+    ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->DisableProcessing();
+
+    SetPerfLevel(PerfLevel::kDisable);
+
+    delete txn;
+    delete txn2;
+    delete txn3;
+    delete txn4;
+    delete txn_x;
+  }
+}
+
+// Reproduce the bug with two snapshots with the same seuqence number and test
+// that the release of the first snapshot will not affect the reads by the other
+// snapshot
+TEST_P(WritePreparedTransactionTest, DoubleSnapshot) {
+  TransactionOptions txn_options;
+  Status s;
+
+  // Insert initial value
+  ASSERT_OK(db->Put(WriteOptions(), "key", "value1"));
+
+  WritePreparedTxnDB* wp_db = dynamic_cast<WritePreparedTxnDB*>(db);
+  Transaction* txn =
+      wp_db->BeginTransaction(WriteOptions(), txn_options, nullptr);
+  ASSERT_OK(txn->SetName("txn"));
+  ASSERT_OK(txn->Put("key", "value2"));
+  ASSERT_OK(txn->Prepare());
+  // Three snapshots with the same seq number
+  const Snapshot* snapshot0 = wp_db->GetSnapshot();
+  const Snapshot* snapshot1 = wp_db->GetSnapshot();
+  const Snapshot* snapshot2 = wp_db->GetSnapshot();
+  ASSERT_OK(txn->Commit());
+  SequenceNumber cache_size = wp_db->COMMIT_CACHE_SIZE;
+  SequenceNumber overlap_seq = txn->GetId() + cache_size;
+  delete txn;
+
+  // 4th snapshot with a larger seq
+  const Snapshot* snapshot3 = wp_db->GetSnapshot();
+  // Cause an eviction to advance max evicted seq number
+  // This also fetches the 4 snapshots from db since their seq is lower than the
+  // new max
+  wp_db->AddCommitted(overlap_seq, overlap_seq);
+
+  ReadOptions ropt;
+  // It should see the value before commit
+  ropt.snapshot = snapshot2;
+  PinnableSlice pinnable_val;
+  s = wp_db->Get(ropt, wp_db->DefaultColumnFamily(), "key", &pinnable_val);
+  ASSERT_OK(s);
+  ASSERT_TRUE(pinnable_val == "value1");
+  pinnable_val.Reset();
+
+  wp_db->ReleaseSnapshot(snapshot1);
+
+  // It should still see the value before commit
+  s = wp_db->Get(ropt, wp_db->DefaultColumnFamily(), "key", &pinnable_val);
+  ASSERT_OK(s);
+  ASSERT_TRUE(pinnable_val == "value1");
+  pinnable_val.Reset();
+
+  // Cause an eviction to advance max evicted seq number and trigger updating
+  // the snapshot list
+  overlap_seq += cache_size;
+  wp_db->AddCommitted(overlap_seq, overlap_seq);
+
+  // It should still see the value before commit
+  s = wp_db->Get(ropt, wp_db->DefaultColumnFamily(), "key", &pinnable_val);
+  ASSERT_OK(s);
+  ASSERT_TRUE(pinnable_val == "value1");
+  pinnable_val.Reset();
+
+  wp_db->ReleaseSnapshot(snapshot0);
+  wp_db->ReleaseSnapshot(snapshot2);
+  wp_db->ReleaseSnapshot(snapshot3);
+}
+
+size_t UniqueCnt(std::vector<SequenceNumber> vec) {
+  std::set<SequenceNumber> aset;
+  for (auto i : vec) {
+    aset.insert(i);
+  }
+  return aset.size();
+}
+// Test that the entries in old_commit_map_ get garbage collected properly
+TEST_P(WritePreparedTransactionTest, OldCommitMapGC) {
+  const size_t snapshot_cache_bits = 0;
+  const size_t commit_cache_bits = 0;
+  DBImpl* mock_db = new DBImpl(options, dbname);
+  UpdateTransactionDBOptions(snapshot_cache_bits, commit_cache_bits);
+  std::unique_ptr<WritePreparedTxnDBMock> wp_db(
+      new WritePreparedTxnDBMock(mock_db, txn_db_options));
+
+  SequenceNumber seq = 0;
+  // Take the first snapshot that overlaps with two txn
+  auto prep_seq = ++seq;
+  wp_db->AddPrepared(prep_seq);
+  auto prep_seq2 = ++seq;
+  wp_db->AddPrepared(prep_seq2);
+  auto snap_seq1 = seq;
+  wp_db->TakeSnapshot(snap_seq1);
+  auto commit_seq = ++seq;
+  wp_db->AddCommitted(prep_seq, commit_seq);
+  wp_db->RemovePrepared(prep_seq);
+  auto commit_seq2 = ++seq;
+  wp_db->AddCommitted(prep_seq2, commit_seq2);
+  wp_db->RemovePrepared(prep_seq2);
+  // Take the 2nd and 3rd snapshot that overlap with the same txn
+  prep_seq = ++seq;
+  wp_db->AddPrepared(prep_seq);
+  auto snap_seq2 = seq;
+  wp_db->TakeSnapshot(snap_seq2);
+  seq++;
+  auto snap_seq3 = seq;
+  wp_db->TakeSnapshot(snap_seq3);
+  seq++;
+  commit_seq = ++seq;
+  wp_db->AddCommitted(prep_seq, commit_seq);
+  wp_db->RemovePrepared(prep_seq);
+  // Make sure max_evicted_seq_ will be larger than 2nd snapshot by evicting the
+  // only item in the commit_cache_ via another commit.
+  prep_seq = ++seq;
+  wp_db->AddPrepared(prep_seq);
+  commit_seq = ++seq;
+  wp_db->AddCommitted(prep_seq, commit_seq);
+  wp_db->RemovePrepared(prep_seq);
+
+  // Verify that the evicted commit entries for all snapshots are in the
+  // old_commit_map_
+  {
+    ASSERT_FALSE(wp_db->old_commit_map_empty_.load());
+    ReadLock rl(&wp_db->old_commit_map_mutex_);
+    ASSERT_EQ(3, wp_db->old_commit_map_.size());
+    ASSERT_EQ(2, UniqueCnt(wp_db->old_commit_map_[snap_seq1]));
+    ASSERT_EQ(1, UniqueCnt(wp_db->old_commit_map_[snap_seq2]));
+    ASSERT_EQ(1, UniqueCnt(wp_db->old_commit_map_[snap_seq3]));
+  }
+
+  // Verify that the 2nd snapshot is cleaned up after the release
+  wp_db->ReleaseSnapshotInternal(snap_seq2);
+  {
+    ASSERT_FALSE(wp_db->old_commit_map_empty_.load());
+    ReadLock rl(&wp_db->old_commit_map_mutex_);
+    ASSERT_EQ(2, wp_db->old_commit_map_.size());
+    ASSERT_EQ(2, UniqueCnt(wp_db->old_commit_map_[snap_seq1]));
+    ASSERT_EQ(1, UniqueCnt(wp_db->old_commit_map_[snap_seq3]));
+  }
+
+  // Verify that the 1st snapshot is cleaned up after the release
+  wp_db->ReleaseSnapshotInternal(snap_seq1);
+  {
+    ASSERT_FALSE(wp_db->old_commit_map_empty_.load());
+    ReadLock rl(&wp_db->old_commit_map_mutex_);
+    ASSERT_EQ(1, wp_db->old_commit_map_.size());
+    ASSERT_EQ(1, UniqueCnt(wp_db->old_commit_map_[snap_seq3]));
+  }
+
+  // Verify that the 3rd snapshot is cleaned up after the release
+  wp_db->ReleaseSnapshotInternal(snap_seq3);
+  {
+    ASSERT_TRUE(wp_db->old_commit_map_empty_.load());
+    ReadLock rl(&wp_db->old_commit_map_mutex_);
+    ASSERT_EQ(0, wp_db->old_commit_map_.size());
+  }
+}
+
+TEST_P(WritePreparedTransactionTest, CheckAgainstSnapshots) {
+  std::vector<SequenceNumber> snapshots = {100l, 200l, 300l, 400l, 500l,
+                                           600l, 700l, 800l, 900l};
+  const size_t snapshot_cache_bits = 2;
+  const uint64_t cache_size = 1ul << snapshot_cache_bits;
+  // Safety check to express the intended size in the test. Can be adjusted if
+  // the snapshots lists changed.
+  ASSERT_EQ((1ul << snapshot_cache_bits) * 2 + 1, snapshots.size());
+  DBImpl* mock_db = new DBImpl(options, dbname);
+  UpdateTransactionDBOptions(snapshot_cache_bits);
+  std::unique_ptr<WritePreparedTxnDBMock> wp_db(
+      new WritePreparedTxnDBMock(mock_db, txn_db_options));
+  SequenceNumber version = 1000l;
+  ASSERT_EQ(0, wp_db->snapshots_total_);
+  wp_db->UpdateSnapshots(snapshots, version);
+  ASSERT_EQ(snapshots.size(), wp_db->snapshots_total_);
+  // seq numbers are chosen so that we have two of them between each two
+  // snapshots. If the diff of two consecutive seq is more than 5, there is a
+  // snapshot between them.
+  std::vector<SequenceNumber> seqs = {50l,  55l,  150l, 155l, 250l, 255l, 350l,
+                                      355l, 450l, 455l, 550l, 555l, 650l, 655l,
+                                      750l, 755l, 850l, 855l, 950l, 955l};
+  ASSERT_GT(seqs.size(), 1);
+  for (size_t i = 0; i + 1 < seqs.size(); i++) {
+    wp_db->old_commit_map_empty_ = true;  // reset
+    CommitEntry commit_entry = {seqs[i], seqs[i + 1]};
+    wp_db->CheckAgainstSnapshots(commit_entry);
+    // Expect update if there is snapshot in between the prepare and commit
+    bool expect_update = commit_entry.commit_seq - commit_entry.prep_seq > 5 &&
+                         commit_entry.commit_seq >= snapshots.front() &&
+                         commit_entry.prep_seq <= snapshots.back();
+    ASSERT_EQ(expect_update, !wp_db->old_commit_map_empty_);
+  }
+
+  // Test that search will include multiple snapshot from snapshot cache
+  {
+    // exclude first and last item in the cache
+    CommitEntry commit_entry = {snapshots.front() + 1,
+                                snapshots[cache_size - 1] - 1};
+    wp_db->old_commit_map_empty_ = true;  // reset
+    wp_db->old_commit_map_.clear();
+    wp_db->CheckAgainstSnapshots(commit_entry);
+    ASSERT_EQ(wp_db->old_commit_map_.size(), cache_size - 2);
+  }
+
+  // Test that search will include multiple snapshot from old snapshots
+  {
+    // include two in the middle
+    CommitEntry commit_entry = {snapshots[cache_size] + 1,
+                                snapshots[cache_size + 2] + 1};
+    wp_db->old_commit_map_empty_ = true;  // reset
+    wp_db->old_commit_map_.clear();
+    wp_db->CheckAgainstSnapshots(commit_entry);
+    ASSERT_EQ(wp_db->old_commit_map_.size(), 2);
+  }
+
+  // Test that search will include both snapshot cache and old snapshots
+  // Case 1: includes all in snapshot cache
+  {
+    CommitEntry commit_entry = {snapshots.front() - 1, snapshots.back() + 1};
+    wp_db->old_commit_map_empty_ = true;  // reset
+    wp_db->old_commit_map_.clear();
+    wp_db->CheckAgainstSnapshots(commit_entry);
+    ASSERT_EQ(wp_db->old_commit_map_.size(), snapshots.size());
+  }
+
+  // Case 2: includes all snapshot caches except the smallest
+  {
+    CommitEntry commit_entry = {snapshots.front() + 1, snapshots.back() + 1};
+    wp_db->old_commit_map_empty_ = true;  // reset
+    wp_db->old_commit_map_.clear();
+    wp_db->CheckAgainstSnapshots(commit_entry);
+    ASSERT_EQ(wp_db->old_commit_map_.size(), snapshots.size() - 1);
+  }
+
+  // Case 3: includes only the largest of snapshot cache
+  {
+    CommitEntry commit_entry = {snapshots[cache_size - 1] - 1,
+                                snapshots.back() + 1};
+    wp_db->old_commit_map_empty_ = true;  // reset
+    wp_db->old_commit_map_.clear();
+    wp_db->CheckAgainstSnapshots(commit_entry);
+    ASSERT_EQ(wp_db->old_commit_map_.size(), snapshots.size() - cache_size + 1);
+  }
+}
+
+#if !defined(ROCKSDB_VALGRIND_RUN) || defined(ROCKSDB_FULL_VALGRIND_RUN)
+// Test that CheckAgainstSnapshots will not miss a live snapshot if it is run in
+// parallel with UpdateSnapshots.
+TEST_P(SnapshotConcurrentAccessTest, SnapshotConcurrentAccess) {
+  // We have a sync point in the method under test after checking each snapshot.
+  // If you increase the max number of snapshots in this test, more sync points
+  // in the methods must also be added.
+  const std::vector<SequenceNumber> snapshots = {10l, 20l, 30l, 40l, 50l,
+                                                 60l, 70l, 80l, 90l, 100l};
+  const size_t snapshot_cache_bits = 2;
+  // Safety check to express the intended size in the test. Can be adjusted if
+  // the snapshots lists changed.
+  ASSERT_EQ((1ul << snapshot_cache_bits) * 2 + 2, snapshots.size());
+  SequenceNumber version = 1000l;
+  // Choose the cache size so that the new snapshot list could replace all the
+  // existing items in the cache and also have some overflow.
+  DBImpl* mock_db = new DBImpl(options, dbname);
+  UpdateTransactionDBOptions(snapshot_cache_bits);
+  std::unique_ptr<WritePreparedTxnDBMock> wp_db(
+      new WritePreparedTxnDBMock(mock_db, txn_db_options));
+  const size_t extra = 2;
+  size_t loop_id = 0;
+  // Add up to extra items that do not fit into the cache
+  for (size_t old_size = 1; old_size <= wp_db->SNAPSHOT_CACHE_SIZE + extra;
+       old_size++) {
+    const std::vector<SequenceNumber> old_snapshots(
+        snapshots.begin(), snapshots.begin() + old_size);
+
+    // Each member of old snapshot might or might not appear in the new list. We
+    // create a common_snapshots for each combination.
+    size_t new_comb_cnt = size_t(1) << old_size;
+    for (size_t new_comb = 0; new_comb < new_comb_cnt; new_comb++, loop_id++) {
+      if (loop_id % split_cnt_ != split_id_) continue;
+      printf(".");  // To signal progress
+      fflush(stdout);
+      std::vector<SequenceNumber> common_snapshots;
+      for (size_t i = 0; i < old_snapshots.size(); i++) {
+        if (IsInCombination(i, new_comb)) {
+          common_snapshots.push_back(old_snapshots[i]);
+        }
+      }
+      // And add some new snapshots to the common list
+      for (size_t added_snapshots = 0;
+           added_snapshots <= snapshots.size() - old_snapshots.size();
+           added_snapshots++) {
+        std::vector<SequenceNumber> new_snapshots = common_snapshots;
+        for (size_t i = 0; i < added_snapshots; i++) {
+          new_snapshots.push_back(snapshots[old_snapshots.size() + i]);
+        }
+        for (auto it = common_snapshots.begin(); it != common_snapshots.end();
+             ++it) {
+          auto snapshot = *it;
+          // Create a commit entry that is around the snapshot and thus should
+          // be not be discarded
+          CommitEntry entry = {static_cast<uint64_t>(snapshot - 1),
+                               snapshot + 1};
+          // The critical part is when iterating the snapshot cache. Afterwards,
+          // we are operating under the lock
+          size_t a_range =
+              std::min(old_snapshots.size(), wp_db->SNAPSHOT_CACHE_SIZE) + 1;
+          size_t b_range =
+              std::min(new_snapshots.size(), wp_db->SNAPSHOT_CACHE_SIZE) + 1;
+          // Break each thread at two points
+          for (size_t a1 = 1; a1 <= a_range; a1++) {
+            for (size_t a2 = a1 + 1; a2 <= a_range; a2++) {
+              for (size_t b1 = 1; b1 <= b_range; b1++) {
+                for (size_t b2 = b1 + 1; b2 <= b_range; b2++) {
+                  SnapshotConcurrentAccessTestInternal(
+                      wp_db.get(), old_snapshots, new_snapshots, entry, version,
+                      a1, a2, b1, b2);
+                }
+              }
+            }
+          }
+        }
+      }
+    }
+  }
+  printf("\n");
+}
+#endif  // !defined(ROCKSDB_VALGRIND_RUN) || defined(ROCKSDB_FULL_VALGRIND_RUN)
+
+// This test clarifies the contract of AdvanceMaxEvictedSeq method
+TEST_P(WritePreparedTransactionTest, AdvanceMaxEvictedSeqBasic) {
+  DBImpl* mock_db = new DBImpl(options, dbname);
+  std::unique_ptr<WritePreparedTxnDBMock> wp_db(
+      new WritePreparedTxnDBMock(mock_db, txn_db_options));
+
+  // 1. Set the initial values for max, prepared, and snapshots
+  SequenceNumber zero_max = 0l;
+  // Set the initial list of prepared txns
+  const std::vector<SequenceNumber> initial_prepared = {10,  30,  50, 100,
+                                                        150, 200, 250};
+  for (auto p : initial_prepared) {
+    wp_db->AddPrepared(p);
+  }
+  // This updates the max value and also set old prepared
+  SequenceNumber init_max = 100;
+  wp_db->AdvanceMaxEvictedSeq(zero_max, init_max);
+  const std::vector<SequenceNumber> initial_snapshots = {20, 40};
+  wp_db->SetDBSnapshots(initial_snapshots);
+  // This will update the internal cache of snapshots from the DB
+  wp_db->UpdateSnapshots(initial_snapshots, init_max);
+
+  // 2. Invoke AdvanceMaxEvictedSeq
+  const std::vector<SequenceNumber> latest_snapshots = {20, 110, 220, 300};
+  wp_db->SetDBSnapshots(latest_snapshots);
+  SequenceNumber new_max = 200;
+  wp_db->AdvanceMaxEvictedSeq(init_max, new_max);
+
+  // 3. Verify that the state matches with AdvanceMaxEvictedSeq contract
+  // a. max should be updated to new_max
+  ASSERT_EQ(wp_db->max_evicted_seq_, new_max);
+  // b. delayed prepared should contain every txn <= max and prepared should
+  // only contain txns > max
+  auto it = initial_prepared.begin();
+  for (; it != initial_prepared.end() && *it <= new_max; ++it) {
+    ASSERT_EQ(1, wp_db->delayed_prepared_.erase(*it));
+  }
+  ASSERT_TRUE(wp_db->delayed_prepared_.empty());
+  for (; it != initial_prepared.end() && !wp_db->prepared_txns_.empty();
+       ++it, wp_db->prepared_txns_.pop()) {
+    ASSERT_EQ(*it, wp_db->prepared_txns_.top());
+  }
+  ASSERT_TRUE(it == initial_prepared.end());
+  ASSERT_TRUE(wp_db->prepared_txns_.empty());
+  // c. snapshots should contain everything below new_max
+  auto sit = latest_snapshots.begin();
+  for (size_t i = 0; sit != latest_snapshots.end() && *sit <= new_max &&
+                     i < wp_db->snapshots_total_;
+       sit++, i++) {
+    ASSERT_TRUE(i < wp_db->snapshots_total_);
+    // This test is in small scale and the list of snapshots are assumed to be
+    // within the cache size limit. This is just a safety check to double check
+    // that assumption.
+    ASSERT_TRUE(i < wp_db->SNAPSHOT_CACHE_SIZE);
+    ASSERT_EQ(*sit, wp_db->snapshot_cache_[i]);
+  }
+}
+
+// A new snapshot should always be always larger than max_evicted_seq_
+// Otherwise the snapshot does not go through AdvanceMaxEvictedSeq
+TEST_P(WritePreparedTransactionTest, NewSnapshotLargerThanMax) {
+  WriteOptions woptions;
+  TransactionOptions txn_options;
+  WritePreparedTxnDB* wp_db = dynamic_cast<WritePreparedTxnDB*>(db);
+  Transaction* txn0 = db->BeginTransaction(woptions, txn_options);
+  ASSERT_OK(txn0->Put(Slice("key"), Slice("value")));
+  ASSERT_OK(txn0->Commit());
+  const SequenceNumber seq = txn0->GetId();  // is also prepare seq
+  delete txn0;
+  std::vector<Transaction*> txns;
+  // Inc seq without committing anything
+  for (int i = 0; i < 10; i++) {
+    Transaction* txn = db->BeginTransaction(woptions, txn_options);
+    ASSERT_OK(txn->SetName("xid" + std::to_string(i)));
+    ASSERT_OK(txn->Put(Slice("key" + std::to_string(i)), Slice("value")));
+    ASSERT_OK(txn->Prepare());
+    txns.push_back(txn);
+  }
+
+  // The new commit is seq + 10
+  ASSERT_OK(db->Put(woptions, "key", "value"));
+  auto snap = wp_db->GetSnapshot();
+  const SequenceNumber last_seq = snap->GetSequenceNumber();
+  wp_db->ReleaseSnapshot(snap);
+  ASSERT_LT(seq, last_seq);
+  // Otherwise our test is not effective
+  ASSERT_LT(last_seq - seq, wp_db->INC_STEP_FOR_MAX_EVICTED);
+
+  // Evict seq out of commit cache
+  const SequenceNumber overwrite_seq = seq + wp_db->COMMIT_CACHE_SIZE;
+  // Check that the next write could make max go beyond last
+  auto last_max = wp_db->max_evicted_seq_.load();
+  wp_db->AddCommitted(overwrite_seq, overwrite_seq);
+  // Check that eviction has advanced the max
+  ASSERT_LT(last_max, wp_db->max_evicted_seq_.load());
+  // Check that the new max has not advanced the last seq
+  ASSERT_LT(wp_db->max_evicted_seq_.load(), last_seq);
+  for (auto txn : txns) {
+    txn->Rollback();
+    delete txn;
+  }
+}
+
+// A new snapshot should always be always larger than max_evicted_seq_
+// In very rare cases max could be below last published seq. Test that
+// taking snapshot will wait for max to catch up.
+TEST_P(WritePreparedTransactionTest, MaxCatchupWithNewSnapshot) {
+  const size_t snapshot_cache_bits = 7;  // same as default
+  const size_t commit_cache_bits = 0;    // only 1 entry => frequent eviction
+  UpdateTransactionDBOptions(snapshot_cache_bits, commit_cache_bits);
+  ASSERT_OK(ReOpen());
+  WriteOptions woptions;
+  WritePreparedTxnDB* wp_db = dynamic_cast<WritePreparedTxnDB*>(db);
+
+  const int writes = 50;
+  const int batch_cnt = 4;
+  ROCKSDB_NAMESPACE::port::Thread t1([&]() {
+    for (int i = 0; i < writes; i++) {
+      WriteBatch batch;
+      // For duplicate keys cause 4 commit entries, each evicting an entry that
+      // is not published yet, thus causing max evicted seq go higher than last
+      // published.
+      for (int b = 0; b < batch_cnt; b++) {
+        ASSERT_OK(batch.Put("foo", "foo"));
+      }
+      ASSERT_OK(db->Write(woptions, &batch));
+    }
+  });
+
+  ROCKSDB_NAMESPACE::port::Thread t2([&]() {
+    while (wp_db->max_evicted_seq_ == 0) {  // wait for insert thread
+      std::this_thread::yield();
+    }
+    for (int i = 0; i < 10; i++) {
+      SequenceNumber max_lower_bound = wp_db->max_evicted_seq_;
+      auto snap = db->GetSnapshot();
+      if (snap->GetSequenceNumber() != 0) {
+        // Value of max_evicted_seq_ when snapshot was taken in unknown. We thus
+        // compare with the lower bound instead as an approximation.
+        ASSERT_LT(max_lower_bound, snap->GetSequenceNumber());
+      }  // seq 0 is ok to be less than max since nothing is visible to it
+      db->ReleaseSnapshot(snap);
+    }
+  });
+
+  t1.join();
+  t2.join();
+
+  // Make sure that the test has worked and seq number has advanced as we
+  // thought
+  auto snap = db->GetSnapshot();
+  ASSERT_GT(snap->GetSequenceNumber(), batch_cnt * writes - 1);
+  db->ReleaseSnapshot(snap);
+}
+
+// Test that reads without snapshots would not hit an undefined state
+TEST_P(WritePreparedTransactionTest, MaxCatchupWithUnbackedSnapshot) {
+  const size_t snapshot_cache_bits = 7;  // same as default
+  const size_t commit_cache_bits = 0;    // only 1 entry => frequent eviction
+  UpdateTransactionDBOptions(snapshot_cache_bits, commit_cache_bits);
+  ASSERT_OK(ReOpen());
+  WriteOptions woptions;
+  WritePreparedTxnDB* wp_db = dynamic_cast<WritePreparedTxnDB*>(db);
+
+  const int writes = 50;
+  ROCKSDB_NAMESPACE::port::Thread t1([&]() {
+    for (int i = 0; i < writes; i++) {
+      WriteBatch batch;
+      ASSERT_OK(batch.Put("key", "foo"));
+      ASSERT_OK(db->Write(woptions, &batch));
+    }
+  });
+
+  ROCKSDB_NAMESPACE::port::Thread t2([&]() {
+    while (wp_db->max_evicted_seq_ == 0) {  // wait for insert thread
+      std::this_thread::yield();
+    }
+    ReadOptions ropt;
+    PinnableSlice pinnable_val;
+    TransactionOptions txn_options;
+    for (int i = 0; i < 10; i++) {
+      auto s = db->Get(ropt, db->DefaultColumnFamily(), "key", &pinnable_val);
+      ASSERT_TRUE(s.ok() || s.IsTryAgain());
+      pinnable_val.Reset();
+      Transaction* txn = db->BeginTransaction(woptions, txn_options);
+      s = txn->Get(ropt, db->DefaultColumnFamily(), "key", &pinnable_val);
+      ASSERT_TRUE(s.ok() || s.IsTryAgain());
+      pinnable_val.Reset();
+      std::vector<std::string> values;
+      auto s_vec =
+          txn->MultiGet(ropt, {db->DefaultColumnFamily()}, {"key"}, &values);
+      ASSERT_EQ(1, values.size());
+      ASSERT_EQ(1, s_vec.size());
+      s = s_vec[0];
+      ASSERT_TRUE(s.ok() || s.IsTryAgain());
+      Slice key("key");
+      txn->MultiGet(ropt, db->DefaultColumnFamily(), 1, &key, &pinnable_val, &s,
+                    true);
+      ASSERT_TRUE(s.ok() || s.IsTryAgain());
+      delete txn;
+    }
+  });
+
+  t1.join();
+  t2.join();
+
+  // Make sure that the test has worked and seq number has advanced as we
+  // thought
+  auto snap = db->GetSnapshot();
+  ASSERT_GT(snap->GetSequenceNumber(), writes - 1);
+  db->ReleaseSnapshot(snap);
+}
+
+// Check that old_commit_map_ cleanup works correctly if the snapshot equals
+// max_evicted_seq_.
+TEST_P(WritePreparedTransactionTest, CleanupSnapshotEqualToMax) {
+  const size_t snapshot_cache_bits = 7;  // same as default
+  const size_t commit_cache_bits = 0;    // only 1 entry => frequent eviction
+  UpdateTransactionDBOptions(snapshot_cache_bits, commit_cache_bits);
+  ASSERT_OK(ReOpen());
+  WriteOptions woptions;
+  WritePreparedTxnDB* wp_db = dynamic_cast<WritePreparedTxnDB*>(db);
+  // Insert something to increase seq
+  ASSERT_OK(db->Put(woptions, "key", "value"));
+  auto snap = db->GetSnapshot();
+  auto snap_seq = snap->GetSequenceNumber();
+  // Another insert should trigger eviction + load snapshot from db
+  ASSERT_OK(db->Put(woptions, "key", "value"));
+  // This is the scenario that we check agaisnt
+  ASSERT_EQ(snap_seq, wp_db->max_evicted_seq_);
+  // old_commit_map_ now has some data that needs gc
+  ASSERT_EQ(1, wp_db->snapshots_total_);
+  ASSERT_EQ(1, wp_db->old_commit_map_.size());
+
+  db->ReleaseSnapshot(snap);
+
+  // Another insert should trigger eviction + load snapshot from db
+  ASSERT_OK(db->Put(woptions, "key", "value"));
+
+  // the snapshot and related metadata must be properly garbage collected
+  ASSERT_EQ(0, wp_db->snapshots_total_);
+  ASSERT_TRUE(wp_db->snapshots_all_.empty());
+  ASSERT_EQ(0, wp_db->old_commit_map_.size());
+}
+
+TEST_P(WritePreparedTransactionTest, AdvanceSeqByOne) {
+  auto snap = db->GetSnapshot();
+  auto seq1 = snap->GetSequenceNumber();
+  db->ReleaseSnapshot(snap);
+
+  WritePreparedTxnDB* wp_db = dynamic_cast<WritePreparedTxnDB*>(db);
+  wp_db->AdvanceSeqByOne();
+
+  snap = db->GetSnapshot();
+  auto seq2 = snap->GetSequenceNumber();
+  db->ReleaseSnapshot(snap);
+
+  ASSERT_LT(seq1, seq2);
+}
+
+// Test that the txn Initilize calls the overridden functions
+TEST_P(WritePreparedTransactionTest, TxnInitialize) {
+  TransactionOptions txn_options;
+  WriteOptions write_options;
+  ASSERT_OK(db->Put(write_options, "key", "value"));
+  Transaction* txn0 = db->BeginTransaction(write_options, txn_options);
+  ASSERT_OK(txn0->SetName("xid"));
+  ASSERT_OK(txn0->Put(Slice("key"), Slice("value1")));
+  ASSERT_OK(txn0->Prepare());
+
+  // SetSnapshot is overridden to update min_uncommitted_
+  txn_options.set_snapshot = true;
+  Transaction* txn1 = db->BeginTransaction(write_options, txn_options);
+  auto snap = txn1->GetSnapshot();
+  auto snap_impl = reinterpret_cast<const SnapshotImpl*>(snap);
+  // If ::Initialize calls the overriden SetSnapshot, min_uncommitted_ must be
+  // udpated
+  ASSERT_GT(snap_impl->min_uncommitted_, kMinUnCommittedSeq);
+
+  ASSERT_OK(txn0->Rollback());
+  ASSERT_OK(txn1->Rollback());
+  delete txn0;
+  delete txn1;
+}
+
+// This tests that transactions with duplicate keys perform correctly after max
+// is advancing their prepared sequence numbers. This will not be the case if
+// for example the txn does not add the prepared seq for the second sub-batch to
+// the PreparedHeap structure.
+TEST_P(WritePreparedTransactionTest, AdvanceMaxEvictedSeqWithDuplicates) {
+  const size_t snapshot_cache_bits = 7;  // same as default
+  const size_t commit_cache_bits = 1;    // disable commit cache
+  UpdateTransactionDBOptions(snapshot_cache_bits, commit_cache_bits);
+  ASSERT_OK(ReOpen());
+
+  ReadOptions ropt;
+  PinnableSlice pinnable_val;
+  WriteOptions write_options;
+  TransactionOptions txn_options;
+  Transaction* txn0 = db->BeginTransaction(write_options, txn_options);
+  ASSERT_OK(txn0->SetName("xid"));
+  ASSERT_OK(txn0->Put(Slice("key"), Slice("value1")));
+  ASSERT_OK(txn0->Put(Slice("key"), Slice("value2")));
+  ASSERT_OK(txn0->Prepare());
+
+  ASSERT_OK(db->Put(write_options, "key2", "value"));
+  // Will cause max advance due to disabled commit cache
+  ASSERT_OK(db->Put(write_options, "key3", "value"));
+
+  auto s = db->Get(ropt, db->DefaultColumnFamily(), "key", &pinnable_val);
+  ASSERT_TRUE(s.IsNotFound());
+  delete txn0;
+
+  WritePreparedTxnDB* wp_db = dynamic_cast<WritePreparedTxnDB*>(db);
+  ASSERT_OK(wp_db->db_impl_->FlushWAL(true));
+  wp_db->TEST_Crash();
+  ASSERT_OK(ReOpenNoDelete());
+  ASSERT_NE(db, nullptr);
+  s = db->Get(ropt, db->DefaultColumnFamily(), "key", &pinnable_val);
+  ASSERT_TRUE(s.IsNotFound());
+
+  txn0 = db->GetTransactionByName("xid");
+  ASSERT_OK(txn0->Rollback());
+  delete txn0;
+}
+
+#if !defined(ROCKSDB_VALGRIND_RUN) || defined(ROCKSDB_FULL_VALGRIND_RUN)
+// Stress SmallestUnCommittedSeq, which reads from both prepared_txns_ and
+// delayed_prepared_, when is run concurrently with advancing max_evicted_seq,
+// which moves prepared txns from prepared_txns_ to delayed_prepared_.
+TEST_P(WritePreparedTransactionTest, SmallestUnCommittedSeq) {
+  const size_t snapshot_cache_bits = 7;  // same as default
+  const size_t commit_cache_bits = 1;    // disable commit cache
+  UpdateTransactionDBOptions(snapshot_cache_bits, commit_cache_bits);
+  ASSERT_OK(ReOpen());
+  WritePreparedTxnDB* wp_db = dynamic_cast<WritePreparedTxnDB*>(db);
+  ReadOptions ropt;
+  PinnableSlice pinnable_val;
+  WriteOptions write_options;
+  TransactionOptions txn_options;
+  std::vector<Transaction*> txns, committed_txns;
+
+  const int cnt = 100;
+  for (int i = 0; i < cnt; i++) {
+    Transaction* txn = db->BeginTransaction(write_options, txn_options);
+    ASSERT_OK(txn->SetName("xid" + std::to_string(i)));
+    auto key = "key1" + std::to_string(i);
+    auto value = "value1" + std::to_string(i);
+    ASSERT_OK(txn->Put(Slice(key), Slice(value)));
+    ASSERT_OK(txn->Prepare());
+    txns.push_back(txn);
+  }
+
+  port::Mutex mutex;
+  Random rnd(1103);
+  ROCKSDB_NAMESPACE::port::Thread commit_thread([&]() {
+    for (int i = 0; i < cnt; i++) {
+      uint32_t index = rnd.Uniform(cnt - i);
+      Transaction* txn;
+      {
+        MutexLock l(&mutex);
+        txn = txns[index];
+        txns.erase(txns.begin() + index);
+      }
+      // Since commit cache is practically disabled, commit results in immediate
+      // advance in max_evicted_seq_ and subsequently moving some prepared txns
+      // to delayed_prepared_.
+      ASSERT_OK(txn->Commit());
+      committed_txns.push_back(txn);
+    }
+  });
+  ROCKSDB_NAMESPACE::port::Thread read_thread([&]() {
+    while (1) {
+      MutexLock l(&mutex);
+      if (txns.empty()) {
+        break;
+      }
+      auto min_uncommitted = wp_db->SmallestUnCommittedSeq();
+      ASSERT_LE(min_uncommitted, (*txns.begin())->GetId());
+    }
+  });
+
+  commit_thread.join();
+  read_thread.join();
+  for (auto txn : committed_txns) {
+    delete txn;
+  }
+}
+#endif  // !defined(ROCKSDB_VALGRIND_RUN) || defined(ROCKSDB_FULL_VALGRIND_RUN)
+
+TEST_P(SeqAdvanceConcurrentTest, SeqAdvanceConcurrent) {
+  // Given the sequential run of txns, with this timeout we should never see a
+  // deadlock nor a timeout unless we have a key conflict, which should be
+  // almost infeasible.
+  txn_db_options.transaction_lock_timeout = 1000;
+  txn_db_options.default_lock_timeout = 1000;
+  ASSERT_OK(ReOpen());
+  FlushOptions fopt;
+
+  // Number of different txn types we use in this test
+  const size_t type_cnt = 5;
+  // The size of the first write group
+  // TODO(myabandeh): This should be increase for pre-release tests
+  const size_t first_group_size = 2;
+  // Total number of txns we run in each test
+  // TODO(myabandeh): This should be increase for pre-release tests
+  const size_t txn_cnt = first_group_size + 1;
+
+  size_t base[txn_cnt + 1] = {
+      1,
+  };
+  for (size_t bi = 1; bi <= txn_cnt; bi++) {
+    base[bi] = base[bi - 1] * type_cnt;
+  }
+  const size_t max_n = static_cast<size_t>(std::pow(type_cnt, txn_cnt));
+  printf("Number of cases being tested is %" ROCKSDB_PRIszt "\n", max_n);
+  for (size_t n = 0; n < max_n; n++) {
+    if (n > 0) {
+      ASSERT_OK(ReOpen());
+    }
+
+    if (n % split_cnt_ != split_id_) continue;
+    if (n % 1000 == 0) {
+      printf("Tested %" ROCKSDB_PRIszt " cases so far\n", n);
+    }
+    DBImpl* db_impl = static_cast_with_check<DBImpl>(db->GetRootDB());
+    auto seq = db_impl->TEST_GetLastVisibleSequence();
+    with_empty_commits = 0;
+    exp_seq = seq;
+    // This is increased before writing the batch for commit
+    commit_writes = 0;
+    // This is increased before txn starts linking if it expects to do a commit
+    // eventually
+    expected_commits = 0;
+    std::vector<port::Thread> threads;
+
+    linked.store(0, std::memory_order_release);
+    std::atomic<bool> batch_formed(false);
+    ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
+        "WriteThread::EnterAsBatchGroupLeader:End",
+        [&](void* /*arg*/) { batch_formed = true; });
+    ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
+        "WriteThread::JoinBatchGroup:Wait", [&](void* /*arg*/) {
+          size_t orig_linked = linked.fetch_add(1, std::memory_order_acq_rel);
+          if (orig_linked == 0) {
+            // Wait until the others are linked too.
+            while (linked.load(std::memory_order_acquire) < first_group_size) {
+            }
+          } else if (orig_linked == first_group_size) {
+            // Make the 2nd batch of the rest of writes plus any followup
+            // commits from the first batch
+            while (linked.load(std::memory_order_acquire) <
+                   txn_cnt + commit_writes) {
+            }
+          }
+          // Then we will have one or more batches consisting of follow-up
+          // commits from the 2nd batch. There is a bit of non-determinism here
+          // but it should be tolerable.
+        });
+
+    ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->EnableProcessing();
+    for (size_t bi = 0; bi < txn_cnt; bi++) {
+      // get the bi-th digit in number system based on type_cnt
+      size_t d = (n % base[bi + 1]) / base[bi];
+      switch (d) {
+        case 0:
+          threads.emplace_back(&TransactionTestBase::TestTxn0, this, bi);
+          break;
+        case 1:
+          threads.emplace_back(&TransactionTestBase::TestTxn1, this, bi);
+          break;
+        case 2:
+          threads.emplace_back(&TransactionTestBase::TestTxn2, this, bi);
+          break;
+        case 3:
+          threads.emplace_back(&TransactionTestBase::TestTxn3, this, bi);
+          break;
+        case 4:
+          threads.emplace_back(&TransactionTestBase::TestTxn3, this, bi);
+          break;
+        default:
+          FAIL();
+      }
+      // wait to be linked
+      while (linked.load(std::memory_order_acquire) <= bi) {
+      }
+      // after a queue of size first_group_size
+      if (bi + 1 == first_group_size) {
+        while (!batch_formed) {
+        }
+        // to make it more deterministic, wait until the commits are linked
+        while (linked.load(std::memory_order_acquire) <=
+               bi + expected_commits) {
+        }
+      }
+    }
+    for (auto& t : threads) {
+      t.join();
+    }
+    if (options.two_write_queues) {
+      // In this case none of the above scheduling tricks to deterministically
+      // form merged batches works because the writes go to separate queues.
+      // This would result in different write groups in each run of the test. We
+      // still keep the test since although non-deterministic and hard to debug,
+      // it is still useful to have.
+      // TODO(myabandeh): Add a deterministic unit test for two_write_queues
+    }
+
+    // Check if memtable inserts advanced seq number as expected
+    seq = db_impl->TEST_GetLastVisibleSequence();
+    ASSERT_EQ(exp_seq, seq);
+
+    ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->DisableProcessing();
+    ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->ClearAllCallBacks();
+
+    // Check if recovery preserves the last sequence number
+    ASSERT_OK(db_impl->FlushWAL(true));
+    ASSERT_OK(ReOpenNoDelete());
+    ASSERT_NE(db, nullptr);
+    db_impl = static_cast_with_check<DBImpl>(db->GetRootDB());
+    seq = db_impl->TEST_GetLastVisibleSequence();
+    ASSERT_LE(exp_seq, seq + with_empty_commits);
+
+    // Check if flush preserves the last sequence number
+    ASSERT_OK(db_impl->Flush(fopt));
+    seq = db_impl->GetLatestSequenceNumber();
+    ASSERT_LE(exp_seq, seq + with_empty_commits);
+
+    // Check if recovery after flush preserves the last sequence number
+    ASSERT_OK(db_impl->FlushWAL(true));
+    ASSERT_OK(ReOpenNoDelete());
+    ASSERT_NE(db, nullptr);
+    db_impl = static_cast_with_check<DBImpl>(db->GetRootDB());
+    seq = db_impl->GetLatestSequenceNumber();
+    ASSERT_LE(exp_seq, seq + with_empty_commits);
+  }
+}
+
+// Run a couple of different txns among them some uncommitted. Restart the db at
+// a couple points to check whether the list of uncommitted txns are recovered
+// properly.
+TEST_P(WritePreparedTransactionTest, BasicRecovery) {
+  options.disable_auto_compactions = true;
+  ASSERT_OK(ReOpen());
+  WritePreparedTxnDB* wp_db = dynamic_cast<WritePreparedTxnDB*>(db);
+
+  TestTxn0(0);
+
+  TransactionOptions txn_options;
+  WriteOptions write_options;
+  size_t index = 1000;
+  Transaction* txn0 = db->BeginTransaction(write_options, txn_options);
+  auto istr0 = std::to_string(index);
+  auto s = txn0->SetName("xid" + istr0);
+  ASSERT_OK(s);
+  s = txn0->Put(Slice("foo0" + istr0), Slice("bar0" + istr0));
+  ASSERT_OK(s);
+  s = txn0->Prepare();
+  ASSERT_OK(s);
+  auto prep_seq_0 = txn0->GetId();
+
+  TestTxn1(0);
+
+  index++;
+  Transaction* txn1 = db->BeginTransaction(write_options, txn_options);
+  auto istr1 = std::to_string(index);
+  s = txn1->SetName("xid" + istr1);
+  ASSERT_OK(s);
+  s = txn1->Put(Slice("foo1" + istr1), Slice("bar"));
+  ASSERT_OK(s);
+  s = txn1->Prepare();
+  ASSERT_OK(s);
+  auto prep_seq_1 = txn1->GetId();
+
+  TestTxn2(0);
+
+  ReadOptions ropt;
+  PinnableSlice pinnable_val;
+  // Check the value is not committed before restart
+  s = db->Get(ropt, db->DefaultColumnFamily(), "foo0" + istr0, &pinnable_val);
+  ASSERT_TRUE(s.IsNotFound());
+  pinnable_val.Reset();
+
+  delete txn0;
+  delete txn1;
+  ASSERT_OK(wp_db->db_impl_->FlushWAL(true));
+  wp_db->TEST_Crash();
+  ASSERT_OK(ReOpenNoDelete());
+  ASSERT_NE(db, nullptr);
+  wp_db = dynamic_cast<WritePreparedTxnDB*>(db);
+  // After recovery, all the uncommitted txns (0 and 1) should be inserted into
+  // delayed_prepared_
+  ASSERT_TRUE(wp_db->prepared_txns_.empty());
+  ASSERT_FALSE(wp_db->delayed_prepared_empty_);
+  ASSERT_LE(prep_seq_0, wp_db->max_evicted_seq_);
+  ASSERT_LE(prep_seq_1, wp_db->max_evicted_seq_);
+  {
+    ReadLock rl(&wp_db->prepared_mutex_);
+    ASSERT_EQ(2, wp_db->delayed_prepared_.size());
+    ASSERT_TRUE(wp_db->delayed_prepared_.find(prep_seq_0) !=
+                wp_db->delayed_prepared_.end());
+    ASSERT_TRUE(wp_db->delayed_prepared_.find(prep_seq_1) !=
+                wp_db->delayed_prepared_.end());
+  }
+
+  // Check the value is still not committed after restart
+  s = db->Get(ropt, db->DefaultColumnFamily(), "foo0" + istr0, &pinnable_val);
+  ASSERT_TRUE(s.IsNotFound());
+  pinnable_val.Reset();
+
+  TestTxn3(0);
+
+  // Test that a recovered txns will be properly marked committed for the next
+  // recovery
+  txn1 = db->GetTransactionByName("xid" + istr1);
+  ASSERT_NE(txn1, nullptr);
+  ASSERT_OK(txn1->Commit());
+  delete txn1;
+
+  index++;
+  Transaction* txn2 = db->BeginTransaction(write_options, txn_options);
+  auto istr2 = std::to_string(index);
+  s = txn2->SetName("xid" + istr2);
+  ASSERT_OK(s);
+  s = txn2->Put(Slice("foo2" + istr2), Slice("bar"));
+  ASSERT_OK(s);
+  s = txn2->Prepare();
+  ASSERT_OK(s);
+  auto prep_seq_2 = txn2->GetId();
+
+  delete txn2;
+  ASSERT_OK(wp_db->db_impl_->FlushWAL(true));
+  wp_db->TEST_Crash();
+  ASSERT_OK(ReOpenNoDelete());
+  ASSERT_NE(db, nullptr);
+  wp_db = dynamic_cast<WritePreparedTxnDB*>(db);
+  ASSERT_TRUE(wp_db->prepared_txns_.empty());
+  ASSERT_FALSE(wp_db->delayed_prepared_empty_);
+
+  // 0 and 2 are prepared and 1 is committed
+  {
+    ReadLock rl(&wp_db->prepared_mutex_);
+    ASSERT_EQ(2, wp_db->delayed_prepared_.size());
+    const auto& end = wp_db->delayed_prepared_.end();
+    ASSERT_NE(wp_db->delayed_prepared_.find(prep_seq_0), end);
+    ASSERT_EQ(wp_db->delayed_prepared_.find(prep_seq_1), end);
+    ASSERT_NE(wp_db->delayed_prepared_.find(prep_seq_2), end);
+  }
+  ASSERT_LE(prep_seq_0, wp_db->max_evicted_seq_);
+  ASSERT_LE(prep_seq_2, wp_db->max_evicted_seq_);
+
+  // Commit all the remaining txns
+  txn0 = db->GetTransactionByName("xid" + istr0);
+  ASSERT_NE(txn0, nullptr);
+  ASSERT_OK(txn0->Commit());
+  txn2 = db->GetTransactionByName("xid" + istr2);
+  ASSERT_NE(txn2, nullptr);
+  ASSERT_OK(txn2->Commit());
+
+  // Check the value is committed after commit
+  s = db->Get(ropt, db->DefaultColumnFamily(), "foo0" + istr0, &pinnable_val);
+  ASSERT_TRUE(s.ok());
+  ASSERT_TRUE(pinnable_val == ("bar0" + istr0));
+  pinnable_val.Reset();
+
+  delete txn0;
+  delete txn2;
+  ASSERT_OK(wp_db->db_impl_->FlushWAL(true));
+  ASSERT_OK(ReOpenNoDelete());
+  ASSERT_NE(db, nullptr);
+  wp_db = dynamic_cast<WritePreparedTxnDB*>(db);
+  ASSERT_TRUE(wp_db->prepared_txns_.empty());
+  ASSERT_TRUE(wp_db->delayed_prepared_empty_);
+
+  // Check the value is still committed after recovery
+  s = db->Get(ropt, db->DefaultColumnFamily(), "foo0" + istr0, &pinnable_val);
+  ASSERT_TRUE(s.ok());
+  ASSERT_TRUE(pinnable_val == ("bar0" + istr0));
+  pinnable_val.Reset();
+}
+
+// After recovery the commit map is empty while the max is set. The code would
+// go through a different path which requires a separate test. Test that the
+// committed data before the restart is visible to all snapshots.
+TEST_P(WritePreparedTransactionTest, IsInSnapshotEmptyMap) {
+  for (bool end_with_prepare : {false, true}) {
+    ASSERT_OK(ReOpen());
+    WriteOptions woptions;
+    ASSERT_OK(db->Put(woptions, "key", "value"));
+    ASSERT_OK(db->Put(woptions, "key", "value"));
+    ASSERT_OK(db->Put(woptions, "key", "value"));
+    SequenceNumber prepare_seq = kMaxSequenceNumber;
+    if (end_with_prepare) {
+      TransactionOptions txn_options;
+      Transaction* txn = db->BeginTransaction(woptions, txn_options);
+      ASSERT_OK(txn->SetName("xid0"));
+      ASSERT_OK(txn->Prepare());
+      prepare_seq = txn->GetId();
+      delete txn;
+    }
+    dynamic_cast<WritePreparedTxnDB*>(db)->TEST_Crash();
+    auto db_impl = static_cast_with_check<DBImpl>(db->GetRootDB());
+    ASSERT_OK(db_impl->FlushWAL(true));
+    ASSERT_OK(ReOpenNoDelete());
+    WritePreparedTxnDB* wp_db = dynamic_cast<WritePreparedTxnDB*>(db);
+    ASSERT_NE(wp_db, nullptr);
+    ASSERT_GT(wp_db->max_evicted_seq_, 0);  // max after recovery
+    // Take a snapshot right after recovery
+    const Snapshot* snap = db->GetSnapshot();
+    auto snap_seq = snap->GetSequenceNumber();
+    ASSERT_GT(snap_seq, 0);
+
+    for (SequenceNumber seq = 0;
+         seq <= wp_db->max_evicted_seq_ && seq != prepare_seq; seq++) {
+      ASSERT_TRUE(wp_db->IsInSnapshot(seq, snap_seq));
+    }
+    if (end_with_prepare) {
+      ASSERT_FALSE(wp_db->IsInSnapshot(prepare_seq, snap_seq));
+    }
+    // trivial check
+    ASSERT_FALSE(wp_db->IsInSnapshot(snap_seq + 1, snap_seq));
+
+    db->ReleaseSnapshot(snap);
+
+    ASSERT_OK(db->Put(woptions, "key", "value"));
+    // Take a snapshot after some writes
+    snap = db->GetSnapshot();
+    snap_seq = snap->GetSequenceNumber();
+    for (SequenceNumber seq = 0;
+         seq <= wp_db->max_evicted_seq_ && seq != prepare_seq; seq++) {
+      ASSERT_TRUE(wp_db->IsInSnapshot(seq, snap_seq));
+    }
+    if (end_with_prepare) {
+      ASSERT_FALSE(wp_db->IsInSnapshot(prepare_seq, snap_seq));
+    }
+    // trivial check
+    ASSERT_FALSE(wp_db->IsInSnapshot(snap_seq + 1, snap_seq));
+
+    db->ReleaseSnapshot(snap);
+  }
+}
+
+// Shows the contract of IsInSnapshot when called on invalid/released snapshots
+TEST_P(WritePreparedTransactionTest, IsInSnapshotReleased) {
+  WritePreparedTxnDB* wp_db = dynamic_cast<WritePreparedTxnDB*>(db);
+  WriteOptions woptions;
+  ASSERT_OK(db->Put(woptions, "key", "value"));
+  // snap seq = 1
+  const Snapshot* snap1 = db->GetSnapshot();
+  ASSERT_OK(db->Put(woptions, "key", "value"));
+  ASSERT_OK(db->Put(woptions, "key", "value"));
+  // snap seq = 3
+  const Snapshot* snap2 = db->GetSnapshot();
+  const SequenceNumber seq = 1;
+  // Evict seq out of commit cache
+  size_t overwrite_seq = wp_db->COMMIT_CACHE_SIZE + seq;
+  wp_db->AddCommitted(overwrite_seq, overwrite_seq);
+  SequenceNumber snap_seq;
+  uint64_t min_uncommitted = kMinUnCommittedSeq;
+  bool released;
+
+  released = false;
+  snap_seq = snap1->GetSequenceNumber();
+  ASSERT_LE(seq, snap_seq);
+  // Valid snapshot lower than max
+  ASSERT_LE(snap_seq, wp_db->max_evicted_seq_);
+  ASSERT_TRUE(wp_db->IsInSnapshot(seq, snap_seq, min_uncommitted, &released));
+  ASSERT_FALSE(released);
+
+  released = false;
+  snap_seq = snap1->GetSequenceNumber();
+  // Invaid snapshot lower than max
+  ASSERT_LE(snap_seq + 1, wp_db->max_evicted_seq_);
+  ASSERT_TRUE(
+      wp_db->IsInSnapshot(seq, snap_seq + 1, min_uncommitted, &released));
+  ASSERT_TRUE(released);
+
+  db->ReleaseSnapshot(snap1);
+
+  released = false;
+  // Released snapshot lower than max
+  ASSERT_TRUE(wp_db->IsInSnapshot(seq, snap_seq, min_uncommitted, &released));
+  // The release does not take affect until the next max advance
+  ASSERT_FALSE(released);
+
+  released = false;
+  // Invaid snapshot lower than max
+  ASSERT_TRUE(
+      wp_db->IsInSnapshot(seq, snap_seq + 1, min_uncommitted, &released));
+  ASSERT_TRUE(released);
+
+  // This make the snapshot release to reflect in txn db structures
+  wp_db->AdvanceMaxEvictedSeq(wp_db->max_evicted_seq_,
+                              wp_db->max_evicted_seq_ + 1);
+
+  released = false;
+  // Released snapshot lower than max
+  ASSERT_TRUE(wp_db->IsInSnapshot(seq, snap_seq, min_uncommitted, &released));
+  ASSERT_TRUE(released);
+
+  released = false;
+  // Invaid snapshot lower than max
+  ASSERT_TRUE(
+      wp_db->IsInSnapshot(seq, snap_seq + 1, min_uncommitted, &released));
+  ASSERT_TRUE(released);
+
+  snap_seq = snap2->GetSequenceNumber();
+
+  released = false;
+  // Unreleased snapshot lower than max
+  ASSERT_TRUE(wp_db->IsInSnapshot(seq, snap_seq, min_uncommitted, &released));
+  ASSERT_FALSE(released);
+
+  db->ReleaseSnapshot(snap2);
+}
+
+// Test WritePreparedTxnDB's IsInSnapshot against different ordering of
+// snapshot, max_committed_seq_, prepared, and commit entries.
+TEST_P(WritePreparedTransactionTest, IsInSnapshot) {
+  WriteOptions wo;
+  // Use small commit cache to trigger lots of eviction and fast advance of
+  // max_evicted_seq_
+  const size_t commit_cache_bits = 3;
+  // Same for snapshot cache size
+  const size_t snapshot_cache_bits = 2;
+
+  // Take some preliminary snapshots first. This is to stress the data structure
+  // that holds the old snapshots as it will be designed to be efficient when
+  // only a few snapshots are below the max_evicted_seq_.
+  for (int max_snapshots = 1; max_snapshots < 20; max_snapshots++) {
+    // Leave some gap between the preliminary snapshots and the final snapshot
+    // that we check. This should test for also different overlapping scenarios
+    // between the last snapshot and the commits.
+    for (int max_gap = 1; max_gap < 10; max_gap++) {
+      // Since we do not actually write to db, we mock the seq as it would be
+      // increased by the db. The only exception is that we need db seq to
+      // advance for our snapshots. for which we apply a dummy put each time we
+      // increase our mock of seq.
+      uint64_t seq = 0;
+      // At each step we prepare a txn and then we commit it in the next txn.
+      // This emulates the consecutive transactions that write to the same key
+      uint64_t cur_txn = 0;
+      // Number of snapshots taken so far
+      int num_snapshots = 0;
+      // Number of gaps applied so far
+      int gap_cnt = 0;
+      // The final snapshot that we will inspect
+      uint64_t snapshot = 0;
+      bool found_committed = false;
+      // To stress the data structure that maintain prepared txns, at each cycle
+      // we add a new prepare txn. These do not mean to be committed for
+      // snapshot inspection.
+      std::set<uint64_t> prepared;
+      // We keep the list of txns committed before we take the last snapshot.
+      // These should be the only seq numbers that will be found in the snapshot
+      std::set<uint64_t> committed_before;
+      // The set of commit seq numbers to be excluded from IsInSnapshot queries
+      std::set<uint64_t> commit_seqs;
+      DBImpl* mock_db = new DBImpl(options, dbname);
+      UpdateTransactionDBOptions(snapshot_cache_bits, commit_cache_bits);
+      std::unique_ptr<WritePreparedTxnDBMock> wp_db(
+          new WritePreparedTxnDBMock(mock_db, txn_db_options));
+      // We continue until max advances a bit beyond the snapshot.
+      while (!snapshot || wp_db->max_evicted_seq_ < snapshot + 100) {
+        // do prepare for a transaction
+        seq++;
+        wp_db->AddPrepared(seq);
+        prepared.insert(seq);
+
+        // If cur_txn is not started, do prepare for it.
+        if (!cur_txn) {
+          seq++;
+          cur_txn = seq;
+          wp_db->AddPrepared(cur_txn);
+        } else {  // else commit it
+          seq++;
+          wp_db->AddCommitted(cur_txn, seq);
+          wp_db->RemovePrepared(cur_txn);
+          commit_seqs.insert(seq);
+          if (!snapshot) {
+            committed_before.insert(cur_txn);
+          }
+          cur_txn = 0;
+        }
+
+        if (num_snapshots < max_snapshots - 1) {
+          // Take preliminary snapshots
+          wp_db->TakeSnapshot(seq);
+          num_snapshots++;
+        } else if (gap_cnt < max_gap) {
+          // Wait for some gap before taking the final snapshot
+          gap_cnt++;
+        } else if (!snapshot) {
+          // Take the final snapshot if it is not already taken
+          snapshot = seq;
+          wp_db->TakeSnapshot(snapshot);
+          num_snapshots++;
+        }
+
+        // If the snapshot is taken, verify seq numbers visible to it. We redo
+        // it at each cycle to test that the system is still sound when
+        // max_evicted_seq_ advances.
+        if (snapshot) {
+          for (uint64_t s = 1;
+               s <= seq && commit_seqs.find(s) == commit_seqs.end(); s++) {
+            bool was_committed =
+                (committed_before.find(s) != committed_before.end());
+            bool is_in_snapshot = wp_db->IsInSnapshot(s, snapshot);
+            if (was_committed != is_in_snapshot) {
+              printf("max_snapshots %d max_gap %d seq %" PRIu64 " max %" PRIu64
+                     " snapshot %" PRIu64
+                     " gap_cnt %d num_snapshots %d s %" PRIu64 "\n",
+                     max_snapshots, max_gap, seq,
+                     wp_db->max_evicted_seq_.load(), snapshot, gap_cnt,
+                     num_snapshots, s);
+            }
+            ASSERT_EQ(was_committed, is_in_snapshot);
+            found_committed = found_committed || is_in_snapshot;
+          }
+        }
+      }
+      // Safety check to make sure the test actually ran
+      ASSERT_TRUE(found_committed);
+      // As an extra check, check if prepared set will be properly empty after
+      // they are committed.
+      if (cur_txn) {
+        wp_db->AddCommitted(cur_txn, seq);
+        wp_db->RemovePrepared(cur_txn);
+      }
+      for (auto p : prepared) {
+        wp_db->AddCommitted(p, seq);
+        wp_db->RemovePrepared(p);
+      }
+      ASSERT_TRUE(wp_db->delayed_prepared_.empty());
+      ASSERT_TRUE(wp_db->prepared_txns_.empty());
+    }
+  }
+}
+
+void ASSERT_SAME(ReadOptions roptions, TransactionDB* db, Status exp_s,
+                 PinnableSlice& exp_v, Slice key) {
+  Status s;
+  PinnableSlice v;
+  s = db->Get(roptions, db->DefaultColumnFamily(), key, &v);
+  ASSERT_EQ(exp_s, s);
+  ASSERT_TRUE(s.ok() || s.IsNotFound());
+  if (s.ok()) {
+    ASSERT_TRUE(exp_v == v);
+  }
+
+  // Try with MultiGet API too
+  std::vector<std::string> values;
+  auto s_vec =
+      db->MultiGet(roptions, {db->DefaultColumnFamily()}, {key}, &values);
+  ASSERT_EQ(1, values.size());
+  ASSERT_EQ(1, s_vec.size());
+  s = s_vec[0];
+  ASSERT_EQ(exp_s, s);
+  ASSERT_TRUE(s.ok() || s.IsNotFound());
+  if (s.ok()) {
+    ASSERT_TRUE(exp_v == values[0]);
+  }
+}
+
+void ASSERT_SAME(TransactionDB* db, Status exp_s, PinnableSlice& exp_v,
+                 Slice key) {
+  ASSERT_SAME(ReadOptions(), db, exp_s, exp_v, key);
+}
+
+TEST_P(WritePreparedTransactionTest, Rollback) {
+  ReadOptions roptions;
+  WriteOptions woptions;
+  TransactionOptions txn_options;
+  const size_t num_keys = 4;
+  const size_t num_values = 5;
+  for (size_t ikey = 1; ikey <= num_keys; ikey++) {
+    for (size_t ivalue = 0; ivalue < num_values; ivalue++) {
+      for (bool crash : {false, true}) {
+        ASSERT_OK(ReOpen());
+        WritePreparedTxnDB* wp_db = dynamic_cast<WritePreparedTxnDB*>(db);
+        std::string key_str = "key" + std::to_string(ikey);
+        switch (ivalue) {
+          case 0:
+            break;
+          case 1:
+            ASSERT_OK(db->Put(woptions, key_str, "initvalue1"));
+            break;
+          case 2:
+            ASSERT_OK(db->Merge(woptions, key_str, "initvalue2"));
+            break;
+          case 3:
+            ASSERT_OK(db->Delete(woptions, key_str));
+            break;
+          case 4:
+            ASSERT_OK(db->SingleDelete(woptions, key_str));
+            break;
+          default:
+            FAIL();
+        }
+
+        PinnableSlice v1;
+        auto s1 =
+            db->Get(roptions, db->DefaultColumnFamily(), Slice("key1"), &v1);
+        PinnableSlice v2;
+        auto s2 =
+            db->Get(roptions, db->DefaultColumnFamily(), Slice("key2"), &v2);
+        PinnableSlice v3;
+        auto s3 =
+            db->Get(roptions, db->DefaultColumnFamily(), Slice("key3"), &v3);
+        PinnableSlice v4;
+        auto s4 =
+            db->Get(roptions, db->DefaultColumnFamily(), Slice("key4"), &v4);
+        Transaction* txn = db->BeginTransaction(woptions, txn_options);
+        auto s = txn->SetName("xid0");
+        ASSERT_OK(s);
+        s = txn->Put(Slice("key1"), Slice("value1"));
+        ASSERT_OK(s);
+        s = txn->Merge(Slice("key2"), Slice("value2"));
+        ASSERT_OK(s);
+        s = txn->Delete(Slice("key3"));
+        ASSERT_OK(s);
+        s = txn->SingleDelete(Slice("key4"));
+        ASSERT_OK(s);
+        s = txn->Prepare();
+        ASSERT_OK(s);
+
+        {
+          ReadLock rl(&wp_db->prepared_mutex_);
+          ASSERT_FALSE(wp_db->prepared_txns_.empty());
+          ASSERT_EQ(txn->GetId(), wp_db->prepared_txns_.top());
+        }
+
+        ASSERT_SAME(db, s1, v1, "key1");
+        ASSERT_SAME(db, s2, v2, "key2");
+        ASSERT_SAME(db, s3, v3, "key3");
+        ASSERT_SAME(db, s4, v4, "key4");
+
+        if (crash) {
+          delete txn;
+          auto db_impl = static_cast_with_check<DBImpl>(db->GetRootDB());
+          ASSERT_OK(db_impl->FlushWAL(true));
+          dynamic_cast<WritePreparedTxnDB*>(db)->TEST_Crash();
+          ASSERT_OK(ReOpenNoDelete());
+          ASSERT_NE(db, nullptr);
+          wp_db = dynamic_cast<WritePreparedTxnDB*>(db);
+          txn = db->GetTransactionByName("xid0");
+          ASSERT_FALSE(wp_db->delayed_prepared_empty_);
+          ReadLock rl(&wp_db->prepared_mutex_);
+          ASSERT_TRUE(wp_db->prepared_txns_.empty());
+          ASSERT_FALSE(wp_db->delayed_prepared_.empty());
+          ASSERT_TRUE(wp_db->delayed_prepared_.find(txn->GetId()) !=
+                      wp_db->delayed_prepared_.end());
+        }
+
+        ASSERT_SAME(db, s1, v1, "key1");
+        ASSERT_SAME(db, s2, v2, "key2");
+        ASSERT_SAME(db, s3, v3, "key3");
+        ASSERT_SAME(db, s4, v4, "key4");
+
+        s = txn->Rollback();
+        ASSERT_OK(s);
+
+        {
+          ASSERT_TRUE(wp_db->delayed_prepared_empty_);
+          ReadLock rl(&wp_db->prepared_mutex_);
+          ASSERT_TRUE(wp_db->prepared_txns_.empty());
+          ASSERT_TRUE(wp_db->delayed_prepared_.empty());
+        }
+
+        ASSERT_SAME(db, s1, v1, "key1");
+        ASSERT_SAME(db, s2, v2, "key2");
+        ASSERT_SAME(db, s3, v3, "key3");
+        ASSERT_SAME(db, s4, v4, "key4");
+        delete txn;
+      }
+    }
+  }
+}
+
+TEST_P(WritePreparedTransactionTest, DisableGCDuringRecovery) {
+  // Use large buffer to avoid memtable flush after 1024 insertions
+  options.write_buffer_size = 1024 * 1024;
+  ASSERT_OK(ReOpen());
+  std::vector<KeyVersion> versions;
+  uint64_t seq = 0;
+  for (uint64_t i = 1; i <= 1024; i++) {
+    std::string v = "bar" + std::to_string(i);
+    ASSERT_OK(db->Put(WriteOptions(), "foo", v));
+    VerifyKeys({{"foo", v}});
+    seq++;  // one for the key/value
+    KeyVersion kv = {"foo", v, seq, kTypeValue};
+    if (options.two_write_queues) {
+      seq++;  // one for the commit
+    }
+    versions.emplace_back(kv);
+  }
+  std::reverse(std::begin(versions), std::end(versions));
+  VerifyInternalKeys(versions);
+  DBImpl* db_impl = static_cast_with_check<DBImpl>(db->GetRootDB());
+  ASSERT_OK(db_impl->FlushWAL(true));
+  // Use small buffer to ensure memtable flush during recovery
+  options.write_buffer_size = 1024;
+  ASSERT_OK(ReOpenNoDelete());
+  VerifyInternalKeys(versions);
+}
+
+TEST_P(WritePreparedTransactionTest, SequenceNumberZero) {
+  ASSERT_OK(db->Put(WriteOptions(), "foo", "bar"));
+  VerifyKeys({{"foo", "bar"}});
+  const Snapshot* snapshot = db->GetSnapshot();
+  ASSERT_OK(db->Flush(FlushOptions()));
+  // Dummy keys to avoid compaction trivially move files and get around actual
+  // compaction logic.
+  ASSERT_OK(db->Put(WriteOptions(), "a", "dummy"));
+  ASSERT_OK(db->Put(WriteOptions(), "z", "dummy"));
+  ASSERT_OK(db->CompactRange(CompactRangeOptions(), nullptr, nullptr));
+  // Compaction will output keys with sequence number 0, if it is visible to
+  // earliest snapshot. Make sure IsInSnapshot() report sequence number 0 is
+  // visible to any snapshot.
+  VerifyKeys({{"foo", "bar"}});
+  VerifyKeys({{"foo", "bar"}}, snapshot);
+  VerifyInternalKeys({{"foo", "bar", 0, kTypeValue}});
+  db->ReleaseSnapshot(snapshot);
+}
+
+// Compaction should not remove a key if it is not committed, and should
+// proceed with older versions of the key as-if the new version doesn't exist.
+TEST_P(WritePreparedTransactionTest, CompactionShouldKeepUncommittedKeys) {
+  options.disable_auto_compactions = true;
+  ASSERT_OK(ReOpen());
+  DBImpl* db_impl = static_cast_with_check<DBImpl>(db->GetRootDB());
+  // Snapshots to avoid keys get evicted.
+  std::vector<const Snapshot*> snapshots;
+  // Keep track of expected sequence number.
+  SequenceNumber expected_seq = 0;
+
+  auto add_key = [&](std::function<Status()> func) {
+    ASSERT_OK(func());
+    expected_seq++;
+    if (options.two_write_queues) {
+      expected_seq++;  // 1 for commit
+    }
+    ASSERT_EQ(expected_seq, db_impl->TEST_GetLastVisibleSequence());
+    snapshots.push_back(db->GetSnapshot());
+  };
+
+  // Each key here represent a standalone test case.
+  add_key([&]() { return db->Put(WriteOptions(), "key1", "value1_1"); });
+  add_key([&]() { return db->Put(WriteOptions(), "key2", "value2_1"); });
+  add_key([&]() { return db->Put(WriteOptions(), "key3", "value3_1"); });
+  add_key([&]() { return db->Put(WriteOptions(), "key4", "value4_1"); });
+  add_key([&]() { return db->Merge(WriteOptions(), "key5", "value5_1"); });
+  add_key([&]() { return db->Merge(WriteOptions(), "key5", "value5_2"); });
+  add_key([&]() { return db->Put(WriteOptions(), "key6", "value6_1"); });
+  add_key([&]() { return db->Put(WriteOptions(), "key7", "value7_1"); });
+  ASSERT_OK(db->Flush(FlushOptions()));
+  add_key([&]() { return db->Delete(WriteOptions(), "key6"); });
+  add_key([&]() { return db->SingleDelete(WriteOptions(), "key7"); });
+
+  auto* transaction = db->BeginTransaction(WriteOptions());
+  ASSERT_OK(transaction->SetName("txn"));
+  ASSERT_OK(transaction->Put("key1", "value1_2"));
+  ASSERT_OK(transaction->Delete("key2"));
+  ASSERT_OK(transaction->SingleDelete("key3"));
+  ASSERT_OK(transaction->Merge("key4", "value4_2"));
+  ASSERT_OK(transaction->Merge("key5", "value5_3"));
+  ASSERT_OK(transaction->Put("key6", "value6_2"));
+  ASSERT_OK(transaction->Put("key7", "value7_2"));
+  // Prepare but not commit.
+  ASSERT_OK(transaction->Prepare());
+  ASSERT_EQ(++expected_seq, db->GetLatestSequenceNumber());
+  ASSERT_OK(db->Flush(FlushOptions()));
+  for (auto* s : snapshots) {
+    db->ReleaseSnapshot(s);
+  }
+  // Dummy keys to avoid compaction trivially move files and get around actual
+  // compaction logic.
+  ASSERT_OK(db->Put(WriteOptions(), "a", "dummy"));
+  ASSERT_OK(db->Put(WriteOptions(), "z", "dummy"));
+  ASSERT_OK(db->CompactRange(CompactRangeOptions(), nullptr, nullptr));
+  VerifyKeys({
+      {"key1", "value1_1"},
+      {"key2", "value2_1"},
+      {"key3", "value3_1"},
+      {"key4", "value4_1"},
+      {"key5", "value5_1,value5_2"},
+      {"key6", "NOT_FOUND"},
+      {"key7", "NOT_FOUND"},
+  });
+  VerifyInternalKeys({
+      {"key1", "value1_2", expected_seq, kTypeValue},
+      {"key1", "value1_1", 0, kTypeValue},
+      {"key2", "", expected_seq, kTypeDeletion},
+      {"key2", "value2_1", 0, kTypeValue},
+      {"key3", "", expected_seq, kTypeSingleDeletion},
+      {"key3", "value3_1", 0, kTypeValue},
+      {"key4", "value4_2", expected_seq, kTypeMerge},
+      {"key4", "value4_1", 0, kTypeValue},
+      {"key5", "value5_3", expected_seq, kTypeMerge},
+      {"key5", "value5_1,value5_2", 0, kTypeValue},
+      {"key6", "value6_2", expected_seq, kTypeValue},
+      {"key7", "value7_2", expected_seq, kTypeValue},
+  });
+  ASSERT_OK(transaction->Commit());
+  VerifyKeys({
+      {"key1", "value1_2"},
+      {"key2", "NOT_FOUND"},
+      {"key3", "NOT_FOUND"},
+      {"key4", "value4_1,value4_2"},
+      {"key5", "value5_1,value5_2,value5_3"},
+      {"key6", "value6_2"},
+      {"key7", "value7_2"},
+  });
+  delete transaction;
+}
+
+// Compaction should keep keys visible to a snapshot based on commit sequence,
+// not just prepare sequence.
+TEST_P(WritePreparedTransactionTest, CompactionShouldKeepSnapshotVisibleKeys) {
+  options.disable_auto_compactions = true;
+  ASSERT_OK(ReOpen());
+  // Keep track of expected sequence number.
+  SequenceNumber expected_seq = 0;
+  auto* txn1 = db->BeginTransaction(WriteOptions());
+  ASSERT_OK(txn1->SetName("txn1"));
+  ASSERT_OK(txn1->Put("key1", "value1_1"));
+  ASSERT_OK(txn1->Prepare());
+  ASSERT_EQ(++expected_seq, db->GetLatestSequenceNumber());
+  ASSERT_OK(txn1->Commit());
+  DBImpl* db_impl = static_cast_with_check<DBImpl>(db->GetRootDB());
+  ASSERT_EQ(++expected_seq, db_impl->TEST_GetLastVisibleSequence());
+  delete txn1;
+  // Take a snapshots to avoid keys get evicted before compaction.
+  const Snapshot* snapshot1 = db->GetSnapshot();
+  auto* txn2 = db->BeginTransaction(WriteOptions());
+  ASSERT_OK(txn2->SetName("txn2"));
+  ASSERT_OK(txn2->Put("key2", "value2_1"));
+  ASSERT_OK(txn2->Prepare());
+  ASSERT_EQ(++expected_seq, db->GetLatestSequenceNumber());
+  // txn1 commit before snapshot2 and it is visible to snapshot2.
+  // txn2 commit after snapshot2 and it is not visible.
+  const Snapshot* snapshot2 = db->GetSnapshot();
+  ASSERT_OK(txn2->Commit());
+  ASSERT_EQ(++expected_seq, db_impl->TEST_GetLastVisibleSequence());
+  delete txn2;
+  // Take a snapshots to avoid keys get evicted before compaction.
+  const Snapshot* snapshot3 = db->GetSnapshot();
+  ASSERT_OK(db->Put(WriteOptions(), "key1", "value1_2"));
+  expected_seq++;  // 1 for write
+  SequenceNumber seq1 = expected_seq;
+  if (options.two_write_queues) {
+    expected_seq++;  // 1 for commit
+  }
+  ASSERT_EQ(expected_seq, db_impl->TEST_GetLastVisibleSequence());
+  ASSERT_OK(db->Put(WriteOptions(), "key2", "value2_2"));
+  expected_seq++;  // 1 for write
+  SequenceNumber seq2 = expected_seq;
+  if (options.two_write_queues) {
+    expected_seq++;  // 1 for commit
+  }
+  ASSERT_EQ(expected_seq, db_impl->TEST_GetLastVisibleSequence());
+  ASSERT_OK(db->Flush(FlushOptions()));
+  db->ReleaseSnapshot(snapshot1);
+  db->ReleaseSnapshot(snapshot3);
+  // Dummy keys to avoid compaction trivially move files and get around actual
+  // compaction logic.
+  ASSERT_OK(db->Put(WriteOptions(), "a", "dummy"));
+  ASSERT_OK(db->Put(WriteOptions(), "z", "dummy"));
+  ASSERT_OK(db->CompactRange(CompactRangeOptions(), nullptr, nullptr));
+  VerifyKeys({{"key1", "value1_2"}, {"key2", "value2_2"}});
+  VerifyKeys({{"key1", "value1_1"}, {"key2", "NOT_FOUND"}}, snapshot2);
+  VerifyInternalKeys({
+      {"key1", "value1_2", seq1, kTypeValue},
+      // "value1_1" is visible to snapshot2. Also keys at bottom level visible
+      // to earliest snapshot will output with seq = 0.
+      {"key1", "value1_1", 0, kTypeValue},
+      {"key2", "value2_2", seq2, kTypeValue},
+  });
+  db->ReleaseSnapshot(snapshot2);
+}
+
+TEST_P(WritePreparedTransactionTest, SmallestUncommittedOptimization) {
+  const size_t snapshot_cache_bits = 7;  // same as default
+  const size_t commit_cache_bits = 0;    // disable commit cache
+  for (bool has_recent_prepare : {true, false}) {
+    UpdateTransactionDBOptions(snapshot_cache_bits, commit_cache_bits);
+    ASSERT_OK(ReOpen());
+
+    ASSERT_OK(db->Put(WriteOptions(), "key1", "value1"));
+    auto* transaction =
+        db->BeginTransaction(WriteOptions(), TransactionOptions(), nullptr);
+    ASSERT_OK(transaction->SetName("txn"));
+    ASSERT_OK(transaction->Delete("key1"));
+    ASSERT_OK(transaction->Prepare());
+    // snapshot1 should get min_uncommitted from prepared_txns_ heap.
+    auto snapshot1 = db->GetSnapshot();
+    ASSERT_EQ(transaction->GetId(),
+              ((SnapshotImpl*)snapshot1)->min_uncommitted_);
+    // Add a commit to advance max_evicted_seq and move the prepared transaction
+    // into delayed_prepared_ set.
+    ASSERT_OK(db->Put(WriteOptions(), "key2", "value2"));
+    Transaction* txn2 = nullptr;
+    if (has_recent_prepare) {
+      txn2 =
+          db->BeginTransaction(WriteOptions(), TransactionOptions(), nullptr);
+      ASSERT_OK(txn2->SetName("txn2"));
+      ASSERT_OK(txn2->Put("key3", "value3"));
+      ASSERT_OK(txn2->Prepare());
+    }
+    // snapshot2 should get min_uncommitted from delayed_prepared_ set.
+    auto snapshot2 = db->GetSnapshot();
+    ASSERT_EQ(transaction->GetId(),
+              ((SnapshotImpl*)snapshot1)->min_uncommitted_);
+    ASSERT_OK(transaction->Commit());
+    delete transaction;
+    if (has_recent_prepare) {
+      ASSERT_OK(txn2->Commit());
+      delete txn2;
+    }
+    VerifyKeys({{"key1", "NOT_FOUND"}});
+    VerifyKeys({{"key1", "value1"}}, snapshot1);
+    VerifyKeys({{"key1", "value1"}}, snapshot2);
+    db->ReleaseSnapshot(snapshot1);
+    db->ReleaseSnapshot(snapshot2);
+  }
+}
+
+// Insert two values, v1 and v2, for a key. Between prepare and commit of v2
+// take two snapshots, s1 and s2. Release s1 during compaction.
+// Test to make sure compaction doesn't get confused and think s1 can see both
+// values, and thus compact out the older value by mistake.
+TEST_P(WritePreparedTransactionTest, ReleaseSnapshotDuringCompaction) {
+  const size_t snapshot_cache_bits = 7;  // same as default
+  const size_t commit_cache_bits = 0;    // minimum commit cache
+  UpdateTransactionDBOptions(snapshot_cache_bits, commit_cache_bits);
+  options.disable_auto_compactions = true;
+  ASSERT_OK(ReOpen());
+
+  ASSERT_OK(db->Put(WriteOptions(), "key1", "value1_1"));
+  auto* transaction =
+      db->BeginTransaction(WriteOptions(), TransactionOptions(), nullptr);
+  ASSERT_OK(transaction->SetName("txn"));
+  ASSERT_OK(transaction->Put("key1", "value1_2"));
+  ASSERT_OK(transaction->Prepare());
+  auto snapshot1 = db->GetSnapshot();
+  // Increment sequence number.
+  ASSERT_OK(db->Put(WriteOptions(), "key2", "value2"));
+  auto snapshot2 = db->GetSnapshot();
+  ASSERT_OK(transaction->Commit());
+  delete transaction;
+  VerifyKeys({{"key1", "value1_2"}});
+  VerifyKeys({{"key1", "value1_1"}}, snapshot1);
+  VerifyKeys({{"key1", "value1_1"}}, snapshot2);
+  // Add a flush to avoid compaction to fallback to trivial move.
+
+  // The callback might be called twice, record the calling state to
+  // prevent double calling.
+  bool callback_finished = false;
+  auto callback = [&](void*) {
+    if (callback_finished) {
+      return;
+    }
+    // Release snapshot1 after CompactionIterator init.
+    // CompactionIterator need to figure out the earliest snapshot
+    // that can see key1:value1_2 is kMaxSequenceNumber, not
+    // snapshot1 or snapshot2.
+    db->ReleaseSnapshot(snapshot1);
+    // Add some keys to advance max_evicted_seq.
+    ASSERT_OK(db->Put(WriteOptions(), "key3", "value3"));
+    ASSERT_OK(db->Put(WriteOptions(), "key4", "value4"));
+    callback_finished = true;
+  };
+  SyncPoint::GetInstance()->SetCallBack("CompactionIterator:AfterInit",
+                                        callback);
+  SyncPoint::GetInstance()->EnableProcessing();
+
+  ASSERT_OK(db->Flush(FlushOptions()));
+  VerifyKeys({{"key1", "value1_2"}});
+  VerifyKeys({{"key1", "value1_1"}}, snapshot2);
+  db->ReleaseSnapshot(snapshot2);
+  SyncPoint::GetInstance()->ClearAllCallBacks();
+}
+
+// Insert two values, v1 and v2, for a key. Take two snapshots, s1 and s2,
+// after committing v2. Release s1 during compaction, right after compaction
+// processes v2 and before processes v1. Test to make sure compaction doesn't
+// get confused and believe v1 and v2 are visible to different snapshot
+// (v1 by s2, v2 by s1) and refuse to compact out v1.
+TEST_P(WritePreparedTransactionTest, ReleaseSnapshotDuringCompaction2) {
+  const size_t snapshot_cache_bits = 7;  // same as default
+  const size_t commit_cache_bits = 0;    // minimum commit cache
+  UpdateTransactionDBOptions(snapshot_cache_bits, commit_cache_bits);
+  options.disable_auto_compactions = true;
+  ASSERT_OK(ReOpen());
+
+  ASSERT_OK(db->Put(WriteOptions(), "key1", "value1"));
+  ASSERT_OK(db->Put(WriteOptions(), "key1", "value2"));
+  SequenceNumber v2_seq = db->GetLatestSequenceNumber();
+  auto* s1 = db->GetSnapshot();
+  // Advance sequence number.
+  ASSERT_OK(db->Put(WriteOptions(), "key2", "dummy"));
+  auto* s2 = db->GetSnapshot();
+
+  int count_value = 0;
+  auto callback = [&](void* arg) {
+    auto* ikey = reinterpret_cast<ParsedInternalKey*>(arg);
+    if (ikey->user_key == "key1") {
+      count_value++;
+      if (count_value == 2) {
+        // Processing v1.
+        db->ReleaseSnapshot(s1);
+        // Add some keys to advance max_evicted_seq and update
+        // old_commit_map.
+        ASSERT_OK(db->Put(WriteOptions(), "key3", "dummy"));
+        ASSERT_OK(db->Put(WriteOptions(), "key4", "dummy"));
+      }
+    }
+  };
+  SyncPoint::GetInstance()->SetCallBack("CompactionIterator:ProcessKV",
+                                        callback);
+  SyncPoint::GetInstance()->EnableProcessing();
+
+  ASSERT_OK(db->Flush(FlushOptions()));
+  // value1 should be compact out.
+  VerifyInternalKeys({{"key1", "value2", v2_seq, kTypeValue}});
+
+  // cleanup
+  db->ReleaseSnapshot(s2);
+  SyncPoint::GetInstance()->ClearAllCallBacks();
+}
+
+// Insert two values, v1 and v2, for a key. Insert another dummy key
+// so to evict the commit cache for v2, while v1 is still in commit cache.
+// Take two snapshots, s1 and s2. Release s1 during compaction.
+// Since commit cache for v2 is evicted, and old_commit_map don't have
+// s1 (it is released),
+// TODO(myabandeh): how can we be sure that the v2's commit info is evicted
+// (and not v1's)? Instead of putting a dummy, we can directly call
+// AddCommitted(v2_seq + cache_size, ...) to evict v2's entry from commit cache.
+TEST_P(WritePreparedTransactionTest, ReleaseSnapshotDuringCompaction3) {
+  const size_t snapshot_cache_bits = 7;  // same as default
+  const size_t commit_cache_bits = 1;    // commit cache size = 2
+  UpdateTransactionDBOptions(snapshot_cache_bits, commit_cache_bits);
+  options.disable_auto_compactions = true;
+  ASSERT_OK(ReOpen());
+
+  // Add a dummy key to evict v2 commit cache, but keep v1 commit cache.
+  // It also advance max_evicted_seq and can trigger old_commit_map cleanup.
+  auto add_dummy = [&]() {
+    auto* txn_dummy =
+        db->BeginTransaction(WriteOptions(), TransactionOptions(), nullptr);
+    ASSERT_OK(txn_dummy->SetName("txn_dummy"));
+    ASSERT_OK(txn_dummy->Put("dummy", "dummy"));
+    ASSERT_OK(txn_dummy->Prepare());
+    ASSERT_OK(txn_dummy->Commit());
+    delete txn_dummy;
+  };
+
+  ASSERT_OK(db->Put(WriteOptions(), "key1", "value1"));
+  auto* txn =
+      db->BeginTransaction(WriteOptions(), TransactionOptions(), nullptr);
+  ASSERT_OK(txn->SetName("txn"));
+  ASSERT_OK(txn->Put("key1", "value2"));
+  ASSERT_OK(txn->Prepare());
+  // TODO(myabandeh): replace it with GetId()?
+  auto v2_seq = db->GetLatestSequenceNumber();
+  ASSERT_OK(txn->Commit());
+  delete txn;
+  auto* s1 = db->GetSnapshot();
+  // Dummy key to advance sequence number.
+  add_dummy();
+  auto* s2 = db->GetSnapshot();
+
+  // The callback might be called twice, record the calling state to
+  // prevent double calling.
+  bool callback_finished = false;
+  auto callback = [&](void*) {
+    if (callback_finished) {
+      return;
+    }
+    db->ReleaseSnapshot(s1);
+    // Add some dummy entries to trigger s1 being cleanup from old_commit_map.
+    add_dummy();
+    add_dummy();
+    callback_finished = true;
+  };
+  SyncPoint::GetInstance()->SetCallBack("CompactionIterator:AfterInit",
+                                        callback);
+  SyncPoint::GetInstance()->EnableProcessing();
+
+  ASSERT_OK(db->Flush(FlushOptions()));
+  // value1 should be compact out.
+  VerifyInternalKeys({{"key1", "value2", v2_seq, kTypeValue}});
+
+  db->ReleaseSnapshot(s2);
+  SyncPoint::GetInstance()->ClearAllCallBacks();
+}
+
+TEST_P(WritePreparedTransactionTest, ReleaseEarliestSnapshotDuringCompaction) {
+  const size_t snapshot_cache_bits = 7;  // same as default
+  const size_t commit_cache_bits = 0;    // minimum commit cache
+  UpdateTransactionDBOptions(snapshot_cache_bits, commit_cache_bits);
+  options.disable_auto_compactions = true;
+  ASSERT_OK(ReOpen());
+
+  ASSERT_OK(db->Put(WriteOptions(), "key1", "value1"));
+  SequenceNumber put_seq = db->GetLatestSequenceNumber();
+  auto* transaction =
+      db->BeginTransaction(WriteOptions(), TransactionOptions(), nullptr);
+  ASSERT_OK(transaction->SetName("txn"));
+  ASSERT_OK(transaction->Delete("key1"));
+  ASSERT_OK(transaction->Prepare());
+  SequenceNumber del_seq = db->GetLatestSequenceNumber();
+  auto snapshot1 = db->GetSnapshot();
+  // Increment sequence number.
+  ASSERT_OK(db->Put(WriteOptions(), "key2", "value2"));
+  auto snapshot2 = db->GetSnapshot();
+  ASSERT_OK(transaction->Commit());
+  delete transaction;
+  VerifyKeys({{"key1", "NOT_FOUND"}});
+  VerifyKeys({{"key1", "value1"}}, snapshot1);
+  VerifyKeys({{"key1", "value1"}}, snapshot2);
+  ASSERT_OK(db->Flush(FlushOptions()));
+
+  auto callback = [&](void* compaction) {
+    // Release snapshot1 after CompactionIterator init.
+    // CompactionIterator need to double check and find out snapshot2 is now
+    // the earliest existing snapshot.
+    if (compaction != nullptr) {
+      db->ReleaseSnapshot(snapshot1);
+      // Add some keys to advance max_evicted_seq.
+      ASSERT_OK(db->Put(WriteOptions(), "key3", "value3"));
+      ASSERT_OK(db->Put(WriteOptions(), "key4", "value4"));
+    }
+  };
+  SyncPoint::GetInstance()->SetCallBack("CompactionIterator:AfterInit",
+                                        callback);
+  SyncPoint::GetInstance()->EnableProcessing();
+
+  // Dummy keys to avoid compaction trivially move files and get around actual
+  // compaction logic.
+  ASSERT_OK(db->Put(WriteOptions(), "a", "dummy"));
+  ASSERT_OK(db->Put(WriteOptions(), "z", "dummy"));
+  ASSERT_OK(db->CompactRange(CompactRangeOptions(), nullptr, nullptr));
+  // Only verify for key1. Both the put and delete for the key should be kept.
+  // Since the delete tombstone is not visible to snapshot2, we need to keep
+  // at least one version of the key, for write-conflict check.
+  VerifyInternalKeys({{"key1", "", del_seq, kTypeDeletion},
+                      {"key1", "value1", put_seq, kTypeValue}});
+  db->ReleaseSnapshot(snapshot2);
+  SyncPoint::GetInstance()->ClearAllCallBacks();
+}
+
+TEST_P(WritePreparedTransactionTest,
+       ReleaseEarliestSnapshotDuringCompaction_WithSD) {
+  constexpr size_t kSnapshotCacheBits = 7;  // same as default
+  constexpr size_t kCommitCacheBits = 0;    // minimum commit cache
+  UpdateTransactionDBOptions(kSnapshotCacheBits, kCommitCacheBits);
+  options.disable_auto_compactions = true;
+  ASSERT_OK(ReOpen());
+
+  ASSERT_OK(db->Put(WriteOptions(), "key", "value"));
+  ASSERT_OK(db->Put(WriteOptions(), "foo", "value"));
+  ASSERT_OK(db->Flush(FlushOptions()));
+
+  auto* txn = db->BeginTransaction(WriteOptions(), TransactionOptions(),
+                                   /*old_txn=*/nullptr);
+  ASSERT_OK(txn->SingleDelete("key"));
+  ASSERT_OK(txn->Put("wow", "value"));
+  ASSERT_OK(txn->SetName("txn"));
+  ASSERT_OK(txn->Prepare());
+  ASSERT_OK(db->Flush(FlushOptions()));
+
+  const bool two_write_queues = std::get<1>(GetParam());
+  if (two_write_queues) {
+    // In the case of two queues, commit another txn just to bump
+    // last_published_seq so that a subsequent GetSnapshot() call can return
+    // a snapshot with higher sequence.
+    auto* dummy_txn = db->BeginTransaction(WriteOptions(), TransactionOptions(),
+                                           /*old_txn=*/nullptr);
+    ASSERT_OK(dummy_txn->Put("haha", "value"));
+    ASSERT_OK(dummy_txn->Commit());
+    delete dummy_txn;
+  }
+  auto* snapshot = db->GetSnapshot();
+
+  ASSERT_OK(txn->Commit());
+  delete txn;
+
+  SyncPoint::GetInstance()->SetCallBack(
+      "CompactionIterator::NextFromInput:SingleDelete:1", [&](void* arg) {
+        if (!arg) {
+          return;
+        }
+        db->ReleaseSnapshot(snapshot);
+
+        // Advance max_evicted_seq
+        ASSERT_OK(db->Put(WriteOptions(), "bar", "value"));
+      });
+  SyncPoint::GetInstance()->EnableProcessing();
+  ASSERT_OK(db->CompactRange(CompactRangeOptions(), /*begin=*/nullptr,
+                             /*end=*/nullptr));
+  SyncPoint::GetInstance()->ClearAllCallBacks();
+}
+
+TEST_P(WritePreparedTransactionTest,
+       ReleaseEarliestSnapshotDuringCompaction_WithSD2) {
+  constexpr size_t kSnapshotCacheBits = 7;  // same as default
+  constexpr size_t kCommitCacheBits = 0;    // minimum commit cache
+  UpdateTransactionDBOptions(kSnapshotCacheBits, kCommitCacheBits);
+  options.disable_auto_compactions = true;
+  ASSERT_OK(ReOpen());
+
+  ASSERT_OK(db->Put(WriteOptions(), "foo", "value"));
+  ASSERT_OK(db->Put(WriteOptions(), "key", "value"));
+  ASSERT_OK(db->Flush(FlushOptions()));
+
+  auto* txn = db->BeginTransaction(WriteOptions(), TransactionOptions(),
+                                   /*old_txn=*/nullptr);
+  ASSERT_OK(txn->Put("bar", "value"));
+  ASSERT_OK(txn->SingleDelete("key"));
+  ASSERT_OK(txn->SetName("txn"));
+  ASSERT_OK(txn->Prepare());
+  ASSERT_OK(db->Flush(FlushOptions()));
+
+  ASSERT_OK(txn->Commit());
+  delete txn;
+
+  ASSERT_OK(db->Put(WriteOptions(), "haha", "value"));
+
+  // Create a dummy transaction to take a snapshot for ww-conflict detection.
+  TransactionOptions txn_opts;
+  txn_opts.set_snapshot = true;
+  auto* dummy_txn =
+      db->BeginTransaction(WriteOptions(), txn_opts, /*old_txn=*/nullptr);
+
+  SyncPoint::GetInstance()->SetCallBack(
+      "CompactionIterator::NextFromInput:SingleDelete:2", [&](void* /*arg*/) {
+        ASSERT_OK(dummy_txn->Rollback());
+        delete dummy_txn;
+
+        ASSERT_OK(db->Put(WriteOptions(), "dontcare", "value"));
+      });
+  SyncPoint::GetInstance()->EnableProcessing();
+
+  ASSERT_OK(db->Put(WriteOptions(), "haha2", "value"));
+  auto* snapshot = db->GetSnapshot();
+
+  ASSERT_OK(db->CompactRange(CompactRangeOptions(), nullptr, nullptr));
+  db->ReleaseSnapshot(snapshot);
+  SyncPoint::GetInstance()->ClearAllCallBacks();
+}
+
+TEST_P(WritePreparedTransactionTest,
+       ReleaseEarliestSnapshotDuringCompaction_WithDelete) {
+  constexpr size_t kSnapshotCacheBits = 7;  // same as default
+  constexpr size_t kCommitCacheBits = 0;    // minimum commit cache
+  UpdateTransactionDBOptions(kSnapshotCacheBits, kCommitCacheBits);
+  options.disable_auto_compactions = true;
+  ASSERT_OK(ReOpen());
+
+  ASSERT_OK(db->Put(WriteOptions(), "a", "value"));
+  ASSERT_OK(db->Put(WriteOptions(), "b", "value"));
+  ASSERT_OK(db->Put(WriteOptions(), "c", "value"));
+  ASSERT_OK(db->Flush(FlushOptions()));
+
+  auto* txn = db->BeginTransaction(WriteOptions(), TransactionOptions(),
+                                   /*old_txn=*/nullptr);
+  ASSERT_OK(txn->Delete("b"));
+  ASSERT_OK(txn->SetName("txn"));
+  ASSERT_OK(txn->Prepare());
+
+  const bool two_write_queues = std::get<1>(GetParam());
+  if (two_write_queues) {
+    // In the case of two queues, commit another txn just to bump
+    // last_published_seq so that a subsequent GetSnapshot() call can return
+    // a snapshot with higher sequence.
+    auto* dummy_txn = db->BeginTransaction(WriteOptions(), TransactionOptions(),
+                                           /*old_txn=*/nullptr);
+    ASSERT_OK(dummy_txn->Put("haha", "value"));
+    ASSERT_OK(dummy_txn->Commit());
+    delete dummy_txn;
+  }
+  auto* snapshot1 = db->GetSnapshot();
+  ASSERT_OK(txn->Commit());
+  delete txn;
+  auto* snapshot2 = db->GetSnapshot();
+
+  SyncPoint::GetInstance()->SetCallBack(
+      "CompactionIterator::NextFromInput:BottommostDelete:1", [&](void* arg) {
+        if (!arg) {
+          return;
+        }
+        db->ReleaseSnapshot(snapshot1);
+
+        // Advance max_evicted_seq
+        ASSERT_OK(db->Put(WriteOptions(), "dummy1", "value"));
+      });
+  SyncPoint::GetInstance()->EnableProcessing();
+
+  ASSERT_OK(db->CompactRange(CompactRangeOptions(), /*begin=*/nullptr,
+                             /*end=*/nullptr));
+  db->ReleaseSnapshot(snapshot2);
+  SyncPoint::GetInstance()->ClearAllCallBacks();
+}
+
+TEST_P(WritePreparedTransactionTest,
+       ReleaseSnapshotBetweenSDAndPutDuringCompaction) {
+  constexpr size_t kSnapshotCacheBits = 7;  // same as default
+  constexpr size_t kCommitCacheBits = 0;    // minimum commit cache
+  UpdateTransactionDBOptions(kSnapshotCacheBits, kCommitCacheBits);
+  options.disable_auto_compactions = true;
+  ASSERT_OK(ReOpen());
+
+  // Create a dummy transaction to take a snapshot for ww-conflict detection.
+  TransactionOptions txn_opts;
+  txn_opts.set_snapshot = true;
+  auto* dummy_txn =
+      db->BeginTransaction(WriteOptions(), txn_opts, /*old_txn=*/nullptr);
+  // Increment seq
+  ASSERT_OK(db->Put(WriteOptions(), "bar", "value"));
+
+  ASSERT_OK(db->Put(WriteOptions(), "foo", "value"));
+  ASSERT_OK(db->SingleDelete(WriteOptions(), "foo"));
+  auto* snapshot1 = db->GetSnapshot();
+  // Increment seq
+  ASSERT_OK(db->Put(WriteOptions(), "dontcare", "value"));
+  auto* snapshot2 = db->GetSnapshot();
+
+  SyncPoint::GetInstance()->SetCallBack(
+      "CompactionIterator::NextFromInput:KeepSDForWW", [&](void* /*arg*/) {
+        db->ReleaseSnapshot(snapshot1);
+
+        ASSERT_OK(db->Put(WriteOptions(), "dontcare2", "value2"));
+      });
+  SyncPoint::GetInstance()->EnableProcessing();
+
+  ASSERT_OK(db->Flush(FlushOptions()));
+  db->ReleaseSnapshot(snapshot2);
+  ASSERT_OK(dummy_txn->Commit());
+  delete dummy_txn;
+  SyncPoint::GetInstance()->ClearAllCallBacks();
+}
+
+TEST_P(WritePreparedTransactionTest,
+       ReleaseEarliestWriteConflictSnapshot_SingleDelete) {
+  constexpr size_t kSnapshotCacheBits = 7;  // same as default
+  constexpr size_t kCommitCacheBits = 0;    // minimum commit cache
+  UpdateTransactionDBOptions(kSnapshotCacheBits, kCommitCacheBits);
+  options.disable_auto_compactions = true;
+  ASSERT_OK(ReOpen());
+
+  ASSERT_OK(db->Put(WriteOptions(), "a", "value"));
+  ASSERT_OK(db->Put(WriteOptions(), "b", "value"));
+  ASSERT_OK(db->Put(WriteOptions(), "c", "value"));
+  ASSERT_OK(db->Flush(FlushOptions()));
+
+  {
+    CompactRangeOptions cro;
+    cro.change_level = true;
+    cro.target_level = 2;
+    ASSERT_OK(db->CompactRange(cro, /*begin=*/nullptr, /*end=*/nullptr));
+  }
+
+  std::unique_ptr<Transaction> txn;
+  txn.reset(db->BeginTransaction(WriteOptions(), TransactionOptions(),
+                                 /*old_txn=*/nullptr));
+  ASSERT_OK(txn->SetName("txn1"));
+  ASSERT_OK(txn->SingleDelete("b"));
+  ASSERT_OK(txn->Prepare());
+  ASSERT_OK(txn->Commit());
+
+  auto* snapshot1 = db->GetSnapshot();
+
+  // Bump seq of the db by performing writes so that
+  // earliest_snapshot_ < earliest_write_conflict_snapshot_ in
+  // CompactionIterator.
+  ASSERT_OK(db->Put(WriteOptions(), "z", "dontcare"));
+
+  // Create another snapshot for write conflict checking
+  std::unique_ptr<Transaction> txn2;
+  {
+    TransactionOptions txn_opts;
+    txn_opts.set_snapshot = true;
+    txn2.reset(
+        db->BeginTransaction(WriteOptions(), txn_opts, /*old_txn=*/nullptr));
+  }
+
+  // Bump seq so that the subsequent bg flush won't create a snapshot with the
+  // same seq as the previous snapshot for conflict checking.
+  ASSERT_OK(db->Put(WriteOptions(), "y", "dont"));
+
+  ASSERT_OK(db->Flush(FlushOptions()));
+
+  SyncPoint::GetInstance()->DisableProcessing();
+  SyncPoint::GetInstance()->ClearAllCallBacks();
+  SyncPoint::GetInstance()->SetCallBack(
+      "CompactionIterator::NextFromInput:SingleDelete:1", [&](void* /*arg*/) {
+        // Rolling back txn2 should release its snapshot(for ww checking).
+        ASSERT_OK(txn2->Rollback());
+        txn2.reset();
+        // Advance max_evicted_seq
+        ASSERT_OK(db->Put(WriteOptions(), "x", "value"));
+      });
+  SyncPoint::GetInstance()->EnableProcessing();
+
+  ASSERT_OK(db->CompactRange(CompactRangeOptions(), /*begin=*/nullptr,
+                             /*end=*/nullptr));
+
+  SyncPoint::GetInstance()->DisableProcessing();
+  SyncPoint::GetInstance()->ClearAllCallBacks();
+
+  db->ReleaseSnapshot(snapshot1);
+}
+
+TEST_P(WritePreparedTransactionTest, ReleaseEarliestSnapshotAfterSeqZeroing) {
+  constexpr size_t kSnapshotCacheBits = 7;  // same as default
+  constexpr size_t kCommitCacheBits = 0;    // minimum commit cache
+  UpdateTransactionDBOptions(kSnapshotCacheBits, kCommitCacheBits);
+  options.disable_auto_compactions = true;
+  ASSERT_OK(ReOpen());
+
+  ASSERT_OK(db->Put(WriteOptions(), "a", "value"));
+  ASSERT_OK(db->Put(WriteOptions(), "b", "value"));
+  ASSERT_OK(db->Put(WriteOptions(), "c", "value"));
+  ASSERT_OK(db->Flush(FlushOptions()));
+
+  {
+    CompactRangeOptions cro;
+    cro.change_level = true;
+    cro.target_level = 2;
+    ASSERT_OK(db->CompactRange(cro, /*begin=*/nullptr, /*end=*/nullptr));
+  }
+
+  ASSERT_OK(db->SingleDelete(WriteOptions(), "b"));
+
+  // Take a snapshot so that the SD won't be dropped during flush.
+  auto* tmp_snapshot = db->GetSnapshot();
+
+  ASSERT_OK(db->Put(WriteOptions(), "b", "value2"));
+  auto* snapshot = db->GetSnapshot();
+  ASSERT_OK(db->Flush(FlushOptions()));
+
+  db->ReleaseSnapshot(tmp_snapshot);
+
+  // Bump the sequence so that the below bg compaction job's snapshot will be
+  // different from snapshot's sequence.
+  ASSERT_OK(db->Put(WriteOptions(), "z", "foo"));
+
+  SyncPoint::GetInstance()->DisableProcessing();
+  SyncPoint::GetInstance()->ClearAllCallBacks();
+  SyncPoint::GetInstance()->SetCallBack(
+      "CompactionIterator::PrepareOutput:ZeroingSeq", [&](void* arg) {
+        const auto* const ikey =
+            reinterpret_cast<const ParsedInternalKey*>(arg);
+        assert(ikey);
+        if (ikey->user_key == "b") {
+          assert(ikey->type == kTypeValue);
+          db->ReleaseSnapshot(snapshot);
+
+          // Bump max_evicted_seq.
+          ASSERT_OK(db->Put(WriteOptions(), "z", "dontcare"));
+        }
+      });
+  SyncPoint::GetInstance()->EnableProcessing();
+
+  ASSERT_OK(db->CompactRange(CompactRangeOptions(), /*begin=*/nullptr,
+                             /*end=*/nullptr));
+
+  SyncPoint::GetInstance()->DisableProcessing();
+  SyncPoint::GetInstance()->ClearAllCallBacks();
+}
+
+TEST_P(WritePreparedTransactionTest, ReleaseEarliestSnapshotAfterSeqZeroing2) {
+  constexpr size_t kSnapshotCacheBits = 7;  // same as default
+  constexpr size_t kCommitCacheBits = 0;    // minimum commit cache
+  UpdateTransactionDBOptions(kSnapshotCacheBits, kCommitCacheBits);
+  options.disable_auto_compactions = true;
+  ASSERT_OK(ReOpen());
+
+  // Generate an L0 with only SD for one key "b".
+  ASSERT_OK(db->Put(WriteOptions(), "a", "value"));
+  ASSERT_OK(db->Put(WriteOptions(), "b", "value"));
+  // Take a snapshot so that subsequent flush outputs the SD for "b".
+  auto* tmp_snapshot = db->GetSnapshot();
+  ASSERT_OK(db->SingleDelete(WriteOptions(), "b"));
+  ASSERT_OK(db->Put(WriteOptions(), "c", "value"));
+
+  SyncPoint::GetInstance()->DisableProcessing();
+  SyncPoint::GetInstance()->ClearAllCallBacks();
+  SyncPoint::GetInstance()->SetCallBack(
+      "CompactionIterator::NextFromInput:SingleDelete:3", [&](void* arg) {
+        if (!arg) {
+          db->ReleaseSnapshot(tmp_snapshot);
+          // Bump max_evicted_seq
+          ASSERT_OK(db->Put(WriteOptions(), "x", "dontcare"));
+        }
+      });
+  SyncPoint::GetInstance()->EnableProcessing();
+
+  ASSERT_OK(db->Flush(FlushOptions()));
+  // Finish generating L0 with only SD for "b".
+
+  SyncPoint::GetInstance()->DisableProcessing();
+  SyncPoint::GetInstance()->ClearAllCallBacks();
+
+  // Move the L0 to L2.
+  {
+    CompactRangeOptions cro;
+    cro.change_level = true;
+    cro.target_level = 2;
+    ASSERT_OK(db->CompactRange(cro, /*begin=*/nullptr, /*end=*/nullptr));
+  }
+
+  ASSERT_OK(db->Put(WriteOptions(), "b", "value1"));
+
+  auto* snapshot = db->GetSnapshot();
+
+  // Bump seq so that a subsequent flush/compaction job's snapshot is larger
+  // than the above snapshot's seq.
+  ASSERT_OK(db->Put(WriteOptions(), "x", "dontcare"));
+
+  // Generate a second L0.
+  ASSERT_OK(db->Flush(FlushOptions()));
+
+  SyncPoint::GetInstance()->SetCallBack(
+      "CompactionIterator::PrepareOutput:ZeroingSeq", [&](void* arg) {
+        const auto* const ikey =
+            reinterpret_cast<const ParsedInternalKey*>(arg);
+        assert(ikey);
+        if (ikey->user_key == "b") {
+          assert(ikey->type == kTypeValue);
+          db->ReleaseSnapshot(snapshot);
+
+          // Bump max_evicted_seq.
+          ASSERT_OK(db->Put(WriteOptions(), "z", "dontcare"));
+        }
+      });
+  SyncPoint::GetInstance()->EnableProcessing();
+
+  ASSERT_OK(db->CompactRange(CompactRangeOptions(), /*begin=*/nullptr,
+                             /*end=*/nullptr));
+
+  SyncPoint::GetInstance()->DisableProcessing();
+  SyncPoint::GetInstance()->ClearAllCallBacks();
+}
+
+// Although the user-contract indicates that a SD can only be issued for a key
+// that exists and has not been overwritten, it is still possible for a Delete
+// to be present when write-prepared transaction is rolled back.
+TEST_P(WritePreparedTransactionTest, SingleDeleteAfterRollback) {
+  constexpr size_t kSnapshotCacheBits = 7;  // same as default
+  constexpr size_t kCommitCacheBits = 0;    // minimum commit cache
+  txn_db_options.rollback_deletion_type_callback =
+      [](TransactionDB*, ColumnFamilyHandle*, const Slice&) { return true; };
+  UpdateTransactionDBOptions(kSnapshotCacheBits, kCommitCacheBits);
+  options.disable_auto_compactions = true;
+  ASSERT_OK(ReOpen());
+
+  // Get a write conflict snapshot by creating a transaction with
+  // set_snapshot=true.
+  TransactionOptions txn_opts;
+  txn_opts.set_snapshot = true;
+  std::unique_ptr<Transaction> dummy_txn(
+      db->BeginTransaction(WriteOptions(), txn_opts));
+
+  std::unique_ptr<Transaction> txn0(
+      db->BeginTransaction(WriteOptions(), TransactionOptions()));
+  ASSERT_OK(txn0->Put("foo", "value"));
+  ASSERT_OK(txn0->SetName("xid0"));
+  ASSERT_OK(txn0->Prepare());
+
+  // Create an SST with only {"foo": "value"}.
+  ASSERT_OK(db->Flush(FlushOptions()));
+
+  // Insert a Delete to cancel out the prior Put by txn0.
+  ASSERT_OK(txn0->Rollback());
+  txn0.reset();
+
+  // Create a second SST.
+  ASSERT_OK(db->Flush(FlushOptions()));
+
+  ASSERT_OK(db->Put(WriteOptions(), "foo", "value1"));
+
+  auto* snapshot = db->GetSnapshot();
+
+  ASSERT_OK(db->SingleDelete(WriteOptions(), "foo"));
+
+  int count = 0;
+  SyncPoint::GetInstance()->DisableProcessing();
+  SyncPoint::GetInstance()->ClearAllCallBacks();
+  SyncPoint::GetInstance()->SetCallBack(
+      "CompactionIterator::NextFromInput:SingleDelete:1", [&](void* arg) {
+        const auto* const c = reinterpret_cast<const Compaction*>(arg);
+        assert(!c);
+        // Trigger once only for SingleDelete during flush.
+        if (0 == count) {
+          ++count;
+          db->ReleaseSnapshot(snapshot);
+          // Bump max_evicted_seq
+          ASSERT_OK(db->Put(WriteOptions(), "x", "dontcare"));
+        }
+      });
+  SyncPoint::GetInstance()->EnableProcessing();
+
+  // Create a third SST containing a SD without its matching PUT.
+  ASSERT_OK(db->Flush(FlushOptions()));
+
+  SyncPoint::GetInstance()->DisableProcessing();
+  SyncPoint::GetInstance()->ClearAllCallBacks();
+  SyncPoint::GetInstance()->EnableProcessing();
+
+  DBImpl* dbimpl = static_cast_with_check<DBImpl>(db->GetRootDB());
+  assert(dbimpl);
+  ASSERT_OK(dbimpl->TEST_CompactRange(
+      /*level=*/0, /*begin=*/nullptr, /*end=*/nullptr,
+      /*column_family=*/nullptr, /*disallow_trivial_mode=*/true));
+
+  SyncPoint::GetInstance()->DisableProcessing();
+  SyncPoint::GetInstance()->ClearAllCallBacks();
+
+  // Release the conflict-checking snapshot.
+  ASSERT_OK(dummy_txn->Rollback());
+}
+
+// A more complex test to verify compaction/flush should keep keys visible
+// to snapshots.
+TEST_P(WritePreparedTransactionTest,
+       CompactionKeepSnapshotVisibleKeysRandomized) {
+  constexpr size_t kNumTransactions = 10;
+  constexpr size_t kNumIterations = 1000;
+
+  std::vector<Transaction*> transactions(kNumTransactions, nullptr);
+  std::vector<size_t> versions(kNumTransactions, 0);
+  std::unordered_map<std::string, std::string> current_data;
+  std::vector<const Snapshot*> snapshots;
+  std::vector<std::unordered_map<std::string, std::string>> snapshot_data;
+
+  Random rnd(1103);
+  options.disable_auto_compactions = true;
+  ASSERT_OK(ReOpen());
+
+  for (size_t i = 0; i < kNumTransactions; i++) {
+    std::string key = "key" + std::to_string(i);
+    std::string value = "value0";
+    ASSERT_OK(db->Put(WriteOptions(), key, value));
+    current_data[key] = value;
+  }
+  VerifyKeys(current_data);
+
+  for (size_t iter = 0; iter < kNumIterations; iter++) {
+    auto r = rnd.Next() % (kNumTransactions + 1);
+    if (r < kNumTransactions) {
+      std::string key = "key" + std::to_string(r);
+      if (transactions[r] == nullptr) {
+        std::string value = "value" + std::to_string(versions[r] + 1);
+        auto* txn = db->BeginTransaction(WriteOptions());
+        ASSERT_OK(txn->SetName("txn" + std::to_string(r)));
+        ASSERT_OK(txn->Put(key, value));
+        ASSERT_OK(txn->Prepare());
+        transactions[r] = txn;
+      } else {
+        std::string value = "value" + std::to_string(++versions[r]);
+        ASSERT_OK(transactions[r]->Commit());
+        delete transactions[r];
+        transactions[r] = nullptr;
+        current_data[key] = value;
+      }
+    } else {
+      auto* snapshot = db->GetSnapshot();
+      VerifyKeys(current_data, snapshot);
+      snapshots.push_back(snapshot);
+      snapshot_data.push_back(current_data);
+    }
+    VerifyKeys(current_data);
+  }
+  // Take a last snapshot to test compaction with uncommitted prepared
+  // transaction.
+  snapshots.push_back(db->GetSnapshot());
+  snapshot_data.push_back(current_data);
+
+  ASSERT_EQ(snapshots.size(), snapshot_data.size());
+  for (size_t i = 0; i < snapshots.size(); i++) {
+    VerifyKeys(snapshot_data[i], snapshots[i]);
+  }
+  ASSERT_OK(db->Flush(FlushOptions()));
+  for (size_t i = 0; i < snapshots.size(); i++) {
+    VerifyKeys(snapshot_data[i], snapshots[i]);
+  }
+  // Dummy keys to avoid compaction trivially move files and get around actual
+  // compaction logic.
+  ASSERT_OK(db->Put(WriteOptions(), "a", "dummy"));
+  ASSERT_OK(db->Put(WriteOptions(), "z", "dummy"));
+  ASSERT_OK(db->CompactRange(CompactRangeOptions(), nullptr, nullptr));
+  for (size_t i = 0; i < snapshots.size(); i++) {
+    VerifyKeys(snapshot_data[i], snapshots[i]);
+  }
+  // cleanup
+  for (size_t i = 0; i < kNumTransactions; i++) {
+    if (transactions[i] == nullptr) {
+      continue;
+    }
+    ASSERT_OK(transactions[i]->Commit());
+    delete transactions[i];
+  }
+  for (size_t i = 0; i < snapshots.size(); i++) {
+    db->ReleaseSnapshot(snapshots[i]);
+  }
+}
+
+// Compaction should not apply the optimization to output key with sequence
+// number equal to 0 if the key is not visible to earliest snapshot, based on
+// commit sequence number.
+TEST_P(WritePreparedTransactionTest,
+       CompactionShouldKeepSequenceForUncommittedKeys) {
+  options.disable_auto_compactions = true;
+  ASSERT_OK(ReOpen());
+  // Keep track of expected sequence number.
+  SequenceNumber expected_seq = 0;
+  auto* transaction = db->BeginTransaction(WriteOptions());
+  ASSERT_OK(transaction->SetName("txn"));
+  ASSERT_OK(transaction->Put("key1", "value1"));
+  ASSERT_OK(transaction->Prepare());
+  ASSERT_EQ(++expected_seq, db->GetLatestSequenceNumber());
+  SequenceNumber seq1 = expected_seq;
+  ASSERT_OK(db->Put(WriteOptions(), "key2", "value2"));
+  DBImpl* db_impl = static_cast_with_check<DBImpl>(db->GetRootDB());
+  expected_seq++;  // one for data
+  if (options.two_write_queues) {
+    expected_seq++;  // one for commit
+  }
+  ASSERT_EQ(expected_seq, db_impl->TEST_GetLastVisibleSequence());
+  ASSERT_OK(db->Flush(FlushOptions()));
+  // Dummy keys to avoid compaction trivially move files and get around actual
+  // compaction logic.
+  ASSERT_OK(db->Put(WriteOptions(), "a", "dummy"));
+  ASSERT_OK(db->Put(WriteOptions(), "z", "dummy"));
+  ASSERT_OK(db->CompactRange(CompactRangeOptions(), nullptr, nullptr));
+  VerifyKeys({
+      {"key1", "NOT_FOUND"},
+      {"key2", "value2"},
+  });
+  VerifyInternalKeys({
+      // "key1" has not been committed. It keeps its sequence number.
+      {"key1", "value1", seq1, kTypeValue},
+      // "key2" is committed and output with seq = 0.
+      {"key2", "value2", 0, kTypeValue},
+  });
+  ASSERT_OK(transaction->Commit());
+  VerifyKeys({
+      {"key1", "value1"},
+      {"key2", "value2"},
+  });
+  delete transaction;
+}
+
+TEST_P(WritePreparedTransactionTest, CommitAndSnapshotDuringCompaction) {
+  options.disable_auto_compactions = true;
+  ASSERT_OK(ReOpen());
+
+  const Snapshot* snapshot = nullptr;
+  ASSERT_OK(db->Put(WriteOptions(), "key1", "value1"));
+  auto* txn = db->BeginTransaction(WriteOptions());
+  ASSERT_OK(txn->SetName("txn"));
+  ASSERT_OK(txn->Put("key1", "value2"));
+  ASSERT_OK(txn->Prepare());
+
+  auto callback = [&](void*) {
+    // Snapshot is taken after compaction start. It should be taken into
+    // consideration for whether to compact out value1.
+    snapshot = db->GetSnapshot();
+    ASSERT_OK(txn->Commit());
+    delete txn;
+  };
+  SyncPoint::GetInstance()->SetCallBack("CompactionIterator:AfterInit",
+                                        callback);
+  SyncPoint::GetInstance()->EnableProcessing();
+  ASSERT_OK(db->Flush(FlushOptions()));
+  ASSERT_NE(nullptr, snapshot);
+  VerifyKeys({{"key1", "value2"}});
+  VerifyKeys({{"key1", "value1"}}, snapshot);
+  db->ReleaseSnapshot(snapshot);
+}
+
+TEST_P(WritePreparedTransactionTest, Iterate) {
+  auto verify_state = [](Iterator* iter, const std::string& key,
+                         const std::string& value) {
+    ASSERT_TRUE(iter->Valid());
+    ASSERT_OK(iter->status());
+    ASSERT_EQ(key, iter->key().ToString());
+    ASSERT_EQ(value, iter->value().ToString());
+  };
+
+  auto verify_iter = [&](const std::string& expected_val) {
+    // Get iterator from a concurrent transaction and make sure it has the
+    // same view as an iterator from the DB.
+    auto* txn = db->BeginTransaction(WriteOptions());
+
+    for (int i = 0; i < 2; i++) {
+      Iterator* iter = (i == 0) ? db->NewIterator(ReadOptions())
+                                : txn->GetIterator(ReadOptions());
+      // Seek
+      iter->Seek("foo");
+      verify_state(iter, "foo", expected_val);
+      // Next
+      iter->Seek("a");
+      verify_state(iter, "a", "va");
+      iter->Next();
+      verify_state(iter, "foo", expected_val);
+      // SeekForPrev
+      iter->SeekForPrev("y");
+      verify_state(iter, "foo", expected_val);
+      // Prev
+      iter->SeekForPrev("z");
+      verify_state(iter, "z", "vz");
+      iter->Prev();
+      verify_state(iter, "foo", expected_val);
+      delete iter;
+    }
+    delete txn;
+  };
+
+  ASSERT_OK(db->Put(WriteOptions(), "foo", "v1"));
+  auto* transaction = db->BeginTransaction(WriteOptions());
+  ASSERT_OK(transaction->SetName("txn"));
+  ASSERT_OK(transaction->Put("foo", "v2"));
+  ASSERT_OK(transaction->Prepare());
+  VerifyKeys({{"foo", "v1"}});
+  // dummy keys
+  ASSERT_OK(db->Put(WriteOptions(), "a", "va"));
+  ASSERT_OK(db->Put(WriteOptions(), "z", "vz"));
+  verify_iter("v1");
+  ASSERT_OK(transaction->Commit());
+  VerifyKeys({{"foo", "v2"}});
+  verify_iter("v2");
+  delete transaction;
+}
+
+TEST_P(WritePreparedTransactionTest, IteratorRefreshNotSupported) {
+  Iterator* iter = db->NewIterator(ReadOptions());
+  ASSERT_OK(iter->status());
+  ASSERT_TRUE(iter->Refresh().IsNotSupported());
+  delete iter;
+}
+
+// Committing an delayed prepared has two non-atomic steps: update commit cache,
+// remove seq from delayed_prepared_. The read in IsInSnapshot also involves two
+// non-atomic steps of checking these two data structures. This test breaks each
+// in the middle to ensure correctness in spite of non-atomic execution.
+// Note: This test is limitted to the case where snapshot is larger than the
+// max_evicted_seq_.
+TEST_P(WritePreparedTransactionTest, NonAtomicCommitOfDelayedPrepared) {
+  const size_t snapshot_cache_bits = 7;  // same as default
+  const size_t commit_cache_bits = 3;    // 8 entries
+  for (auto split_read : {true, false}) {
+    std::vector<bool> split_options = {false};
+    if (split_read) {
+      // Also test for break before mutex
+      split_options.push_back(true);
+    }
+    for (auto split_before_mutex : split_options) {
+      UpdateTransactionDBOptions(snapshot_cache_bits, commit_cache_bits);
+      ASSERT_OK(ReOpen());
+      WritePreparedTxnDB* wp_db = dynamic_cast<WritePreparedTxnDB*>(db);
+      DBImpl* db_impl = static_cast_with_check<DBImpl>(db->GetRootDB());
+      // Fill up the commit cache
+      std::string init_value("value1");
+      for (int i = 0; i < 10; i++) {
+        ASSERT_OK(db->Put(WriteOptions(), Slice("key1"), Slice(init_value)));
+      }
+      // Prepare a transaction but do not commit it
+      Transaction* txn =
+          db->BeginTransaction(WriteOptions(), TransactionOptions());
+      ASSERT_OK(txn->SetName("xid"));
+      ASSERT_OK(txn->Put(Slice("key1"), Slice("value2")));
+      ASSERT_OK(txn->Prepare());
+      // Commit a bunch of entries to advance max evicted seq and make the
+      // prepared a delayed prepared
+      for (int i = 0; i < 10; i++) {
+        ASSERT_OK(db->Put(WriteOptions(), Slice("key3"), Slice("value3")));
+      }
+      // The snapshot should not see the delayed prepared entry
+      auto snap = db->GetSnapshot();
+
+      if (split_read) {
+        if (split_before_mutex) {
+          // split before acquiring prepare_mutex_
+          ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->LoadDependency(
+              {{"WritePreparedTxnDB::IsInSnapshot:prepared_mutex_:pause",
+                "AtomicCommitOfDelayedPrepared:Commit:before"},
+               {"AtomicCommitOfDelayedPrepared:Commit:after",
+                "WritePreparedTxnDB::IsInSnapshot:prepared_mutex_:resume"}});
+        } else {
+          // split right after reading from the commit cache
+          ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->LoadDependency(
+              {{"WritePreparedTxnDB::IsInSnapshot:GetCommitEntry:pause",
+                "AtomicCommitOfDelayedPrepared:Commit:before"},
+               {"AtomicCommitOfDelayedPrepared:Commit:after",
+                "WritePreparedTxnDB::IsInSnapshot:GetCommitEntry:resume"}});
+        }
+      } else {  // split commit
+        // split right before removing from delayed_prepared_
+        ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->LoadDependency(
+            {{"WritePreparedTxnDB::RemovePrepared:pause",
+              "AtomicCommitOfDelayedPrepared:Read:before"},
+             {"AtomicCommitOfDelayedPrepared:Read:after",
+              "WritePreparedTxnDB::RemovePrepared:resume"}});
+      }
+      SyncPoint::GetInstance()->EnableProcessing();
+
+      ROCKSDB_NAMESPACE::port::Thread commit_thread([&]() {
+        TEST_SYNC_POINT("AtomicCommitOfDelayedPrepared:Commit:before");
+        ASSERT_OK(txn->Commit());
+        if (split_before_mutex) {
+          // Do bunch of inserts to evict the commit entry from the cache. This
+          // would prevent the 2nd look into commit cache under prepare_mutex_
+          // to see the commit entry.
+          auto seq = db_impl->TEST_GetLastVisibleSequence();
+          size_t tries = 0;
+          while (wp_db->max_evicted_seq_ < seq && tries < 50) {
+            ASSERT_OK(db->Put(WriteOptions(), Slice("key3"), Slice("value3")));
+            tries++;
+          };
+          ASSERT_LT(tries, 50);
+        }
+        TEST_SYNC_POINT("AtomicCommitOfDelayedPrepared:Commit:after");
+        delete txn;
+      });
+
+      ROCKSDB_NAMESPACE::port::Thread read_thread([&]() {
+        TEST_SYNC_POINT("AtomicCommitOfDelayedPrepared:Read:before");
+        ReadOptions roptions;
+        roptions.snapshot = snap;
+        PinnableSlice value;
+        auto s = db->Get(roptions, db->DefaultColumnFamily(), "key1", &value);
+        ASSERT_OK(s);
+        // It should not see the commit of delayed prepared
+        ASSERT_TRUE(value == init_value);
+        TEST_SYNC_POINT("AtomicCommitOfDelayedPrepared:Read:after");
+        db->ReleaseSnapshot(snap);
+      });
+
+      read_thread.join();
+      commit_thread.join();
+      ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->DisableProcessing();
+      ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->ClearAllCallBacks();
+    }  // for split_before_mutex
+  }    // for split_read
+}
+
+// When max evicted seq advances a prepared seq, it involves two updates: i)
+// adding prepared seq to delayed_prepared_, ii) updating max_evicted_seq_.
+// ::IsInSnapshot also reads these two values in a non-atomic way. This test
+// ensures correctness if the update occurs after ::IsInSnapshot reads
+// delayed_prepared_empty_ and before it reads max_evicted_seq_.
+// Note: this test focuses on read snapshot larger than max_evicted_seq_.
+TEST_P(WritePreparedTransactionTest, NonAtomicUpdateOfDelayedPrepared) {
+  const size_t snapshot_cache_bits = 7;  // same as default
+  const size_t commit_cache_bits = 3;    // 8 entries
+  UpdateTransactionDBOptions(snapshot_cache_bits, commit_cache_bits);
+  ASSERT_OK(ReOpen());
+  WritePreparedTxnDB* wp_db = dynamic_cast<WritePreparedTxnDB*>(db);
+  // Fill up the commit cache
+  std::string init_value("value1");
+  for (int i = 0; i < 10; i++) {
+    ASSERT_OK(db->Put(WriteOptions(), Slice("key1"), Slice(init_value)));
+  }
+  // Prepare a transaction but do not commit it
+  Transaction* txn = db->BeginTransaction(WriteOptions(), TransactionOptions());
+  ASSERT_OK(txn->SetName("xid"));
+  ASSERT_OK(txn->Put(Slice("key1"), Slice("value2")));
+  ASSERT_OK(txn->Prepare());
+  // Create a gap between prepare seq and snapshot seq
+  ASSERT_OK(db->Put(WriteOptions(), Slice("key3"), Slice("value3")));
+  ASSERT_OK(db->Put(WriteOptions(), Slice("key3"), Slice("value3")));
+  // The snapshot should not see the delayed prepared entry
+  auto snap = db->GetSnapshot();
+  ASSERT_LT(txn->GetId(), snap->GetSequenceNumber());
+
+  // split right after reading delayed_prepared_empty_
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->LoadDependency(
+      {{"WritePreparedTxnDB::IsInSnapshot:delayed_prepared_empty_:pause",
+        "AtomicUpdateOfDelayedPrepared:before"},
+       {"AtomicUpdateOfDelayedPrepared:after",
+        "WritePreparedTxnDB::IsInSnapshot:delayed_prepared_empty_:resume"}});
+  SyncPoint::GetInstance()->EnableProcessing();
+
+  ROCKSDB_NAMESPACE::port::Thread commit_thread([&]() {
+    TEST_SYNC_POINT("AtomicUpdateOfDelayedPrepared:before");
+    // Commit a bunch of entries to advance max evicted seq and make the
+    // prepared a delayed prepared
+    size_t tries = 0;
+    while (wp_db->max_evicted_seq_ < txn->GetId() && tries < 50) {
+      ASSERT_OK(db->Put(WriteOptions(), Slice("key3"), Slice("value3")));
+      tries++;
+    };
+    ASSERT_LT(tries, 50);
+    // This is the case on which the test focuses
+    ASSERT_LT(wp_db->max_evicted_seq_, snap->GetSequenceNumber());
+    TEST_SYNC_POINT("AtomicUpdateOfDelayedPrepared:after");
+  });
+
+  ROCKSDB_NAMESPACE::port::Thread read_thread([&]() {
+    ReadOptions roptions;
+    roptions.snapshot = snap;
+    PinnableSlice value;
+    auto s = db->Get(roptions, db->DefaultColumnFamily(), "key1", &value);
+    ASSERT_OK(s);
+    // It should not see the uncommitted value of delayed prepared
+    ASSERT_TRUE(value == init_value);
+    db->ReleaseSnapshot(snap);
+  });
+
+  read_thread.join();
+  commit_thread.join();
+  ASSERT_OK(txn->Commit());
+  delete txn;
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->DisableProcessing();
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->ClearAllCallBacks();
+}
+
+// Eviction from commit cache and update of max evicted seq are two non-atomic
+// steps. Similarly the read of max_evicted_seq_ in ::IsInSnapshot and reading
+// from commit cache are two non-atomic steps. This tests if the update occurs
+// after reading max_evicted_seq_ and before reading the commit cache.
+// Note: the test focuses on snapshot larger than max_evicted_seq_
+TEST_P(WritePreparedTransactionTest, NonAtomicUpdateOfMaxEvictedSeq) {
+  const size_t snapshot_cache_bits = 7;  // same as default
+  const size_t commit_cache_bits = 3;    // 8 entries
+  UpdateTransactionDBOptions(snapshot_cache_bits, commit_cache_bits);
+  ASSERT_OK(ReOpen());
+  WritePreparedTxnDB* wp_db = dynamic_cast<WritePreparedTxnDB*>(db);
+  // Fill up the commit cache
+  std::string init_value("value1");
+  std::string last_value("value_final");
+  for (int i = 0; i < 10; i++) {
+    ASSERT_OK(db->Put(WriteOptions(), Slice("key1"), Slice(init_value)));
+  }
+  // Do an uncommitted write to prevent min_uncommitted optimization
+  Transaction* txn1 =
+      db->BeginTransaction(WriteOptions(), TransactionOptions());
+  ASSERT_OK(txn1->SetName("xid1"));
+  ASSERT_OK(txn1->Put(Slice("key0"), last_value));
+  ASSERT_OK(txn1->Prepare());
+  // Do a write with prepare to get the prepare seq
+  Transaction* txn = db->BeginTransaction(WriteOptions(), TransactionOptions());
+  ASSERT_OK(txn->SetName("xid"));
+  ASSERT_OK(txn->Put(Slice("key1"), last_value));
+  ASSERT_OK(txn->Prepare());
+  ASSERT_OK(txn->Commit());
+  // Create a gap between commit entry and snapshot seq
+  ASSERT_OK(db->Put(WriteOptions(), Slice("key3"), Slice("value3")));
+  ASSERT_OK(db->Put(WriteOptions(), Slice("key3"), Slice("value3")));
+  // The snapshot should see the last commit
+  auto snap = db->GetSnapshot();
+  ASSERT_LE(txn->GetId(), snap->GetSequenceNumber());
+
+  // split right after reading max_evicted_seq_
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->LoadDependency(
+      {{"WritePreparedTxnDB::IsInSnapshot:max_evicted_seq_:pause",
+        "NonAtomicUpdateOfMaxEvictedSeq:before"},
+       {"NonAtomicUpdateOfMaxEvictedSeq:after",
+        "WritePreparedTxnDB::IsInSnapshot:max_evicted_seq_:resume"}});
+  SyncPoint::GetInstance()->EnableProcessing();
+
+  ROCKSDB_NAMESPACE::port::Thread commit_thread([&]() {
+    TEST_SYNC_POINT("NonAtomicUpdateOfMaxEvictedSeq:before");
+    // Commit a bunch of entries to advance max evicted seq beyond txn->GetId()
+    size_t tries = 0;
+    while (wp_db->max_evicted_seq_ < txn->GetId() && tries < 50) {
+      ASSERT_OK(db->Put(WriteOptions(), Slice("key3"), Slice("value3")));
+      tries++;
+    };
+    ASSERT_LT(tries, 50);
+    // This is the case on which the test focuses
+    ASSERT_LT(wp_db->max_evicted_seq_, snap->GetSequenceNumber());
+    TEST_SYNC_POINT("NonAtomicUpdateOfMaxEvictedSeq:after");
+  });
+
+  ROCKSDB_NAMESPACE::port::Thread read_thread([&]() {
+    ReadOptions roptions;
+    roptions.snapshot = snap;
+    PinnableSlice value;
+    auto s = db->Get(roptions, db->DefaultColumnFamily(), "key1", &value);
+    ASSERT_OK(s);
+    // It should see the committed value of the evicted entry
+    ASSERT_TRUE(value == last_value);
+    db->ReleaseSnapshot(snap);
+  });
+
+  read_thread.join();
+  commit_thread.join();
+  delete txn;
+  ASSERT_OK(txn1->Commit());
+  delete txn1;
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->DisableProcessing();
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->ClearAllCallBacks();
+}
+
+// Test when we add a prepared seq when the max_evicted_seq_ already goes beyond
+// that. The test focuses on a race condition between AddPrepared and
+// AdvanceMaxEvictedSeq functions.
+TEST_P(WritePreparedTransactionTest, AddPreparedBeforeMax) {
+  if (!options.two_write_queues) {
+    // This test is only for two write queues
+    return;
+  }
+  const size_t snapshot_cache_bits = 7;  // same as default
+  // 1 entry to advance max after the 2nd commit
+  const size_t commit_cache_bits = 0;
+  UpdateTransactionDBOptions(snapshot_cache_bits, commit_cache_bits);
+  ASSERT_OK(ReOpen());
+  WritePreparedTxnDB* wp_db = dynamic_cast<WritePreparedTxnDB*>(db);
+  std::string some_value("value_some");
+  std::string uncommitted_value("value_uncommitted");
+  // Prepare two uncommitted transactions
+  Transaction* txn1 =
+      db->BeginTransaction(WriteOptions(), TransactionOptions());
+  ASSERT_OK(txn1->SetName("xid1"));
+  ASSERT_OK(txn1->Put(Slice("key1"), some_value));
+  ASSERT_OK(txn1->Prepare());
+  Transaction* txn2 =
+      db->BeginTransaction(WriteOptions(), TransactionOptions());
+  ASSERT_OK(txn2->SetName("xid2"));
+  ASSERT_OK(txn2->Put(Slice("key2"), some_value));
+  ASSERT_OK(txn2->Prepare());
+  // Start the txn here so the other thread could get its id
+  Transaction* txn = db->BeginTransaction(WriteOptions(), TransactionOptions());
+  ASSERT_OK(txn->SetName("xid"));
+  ASSERT_OK(txn->Put(Slice("key0"), uncommitted_value));
+  port::Mutex txn_mutex_;
+
+  // t1) Insert prepared entry, t2) commit other entries to advance max
+  // evicted sec and finish checking the existing prepared entries, t1)
+  // AddPrepared, t2) update max_evicted_seq_
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->LoadDependency({
+      {"AddPreparedCallback::AddPrepared::begin:pause",
+       "AddPreparedBeforeMax::read_thread:start"},
+      {"AdvanceMaxEvictedSeq::update_max:pause",
+       "AddPreparedCallback::AddPrepared::begin:resume"},
+      {"AddPreparedCallback::AddPrepared::end",
+       "AdvanceMaxEvictedSeq::update_max:resume"},
+  });
+  SyncPoint::GetInstance()->EnableProcessing();
+
+  ROCKSDB_NAMESPACE::port::Thread write_thread([&]() {
+    txn_mutex_.Lock();
+    ASSERT_OK(txn->Prepare());
+    txn_mutex_.Unlock();
+  });
+
+  ROCKSDB_NAMESPACE::port::Thread read_thread([&]() {
+    TEST_SYNC_POINT("AddPreparedBeforeMax::read_thread:start");
+    // Publish seq number with a commit
+    ASSERT_OK(txn1->Commit());
+    // Since the commit cache size is one the 2nd commit evict the 1st one and
+    // invokes AdcanceMaxEvictedSeq
+    ASSERT_OK(txn2->Commit());
+
+    ReadOptions roptions;
+    PinnableSlice value;
+    // The snapshot should not see the uncommitted value from write_thread
+    auto snap = db->GetSnapshot();
+    ASSERT_LT(wp_db->max_evicted_seq_, snap->GetSequenceNumber());
+    // This is the scenario that we test for
+    txn_mutex_.Lock();
+    ASSERT_GT(wp_db->max_evicted_seq_, txn->GetId());
+    txn_mutex_.Unlock();
+    roptions.snapshot = snap;
+    auto s = db->Get(roptions, db->DefaultColumnFamily(), "key0", &value);
+    ASSERT_TRUE(s.IsNotFound());
+    db->ReleaseSnapshot(snap);
+  });
+
+  read_thread.join();
+  write_thread.join();
+  delete txn1;
+  delete txn2;
+  ASSERT_OK(txn->Commit());
+  delete txn;
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->DisableProcessing();
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->ClearAllCallBacks();
+}
+
+// When an old prepared entry gets committed, there is a gap between the time
+// that it is published and when it is cleaned up from old_prepared_. This test
+// stresses such cases.
+TEST_P(WritePreparedTransactionTest, CommitOfDelayedPrepared) {
+  const size_t snapshot_cache_bits = 7;  // same as default
+  for (const size_t commit_cache_bits : {0, 2, 3}) {
+    for (const size_t sub_batch_cnt : {1, 2, 3}) {
+      UpdateTransactionDBOptions(snapshot_cache_bits, commit_cache_bits);
+      ASSERT_OK(ReOpen());
+      std::atomic<const Snapshot*> snap = {nullptr};
+      std::atomic<SequenceNumber> exp_prepare = {0};
+      ROCKSDB_NAMESPACE::port::Thread callback_thread;
+      // Value is synchronized via snap
+      PinnableSlice value;
+      // Take a snapshot after publish and before RemovePrepared:Start
+      auto snap_callback = [&]() {
+        ASSERT_EQ(nullptr, snap.load());
+        snap.store(db->GetSnapshot());
+        ReadOptions roptions;
+        roptions.snapshot = snap.load();
+        auto s = db->Get(roptions, db->DefaultColumnFamily(), "key2", &value);
+        ASSERT_OK(s);
+      };
+      auto callback = [&](void* param) {
+        SequenceNumber prep_seq = *((SequenceNumber*)param);
+        if (prep_seq == exp_prepare.load()) {  // only for write_thread
+          // We need to spawn a thread to avoid deadlock since getting a
+          // snpashot might end up calling AdvanceSeqByOne which needs joining
+          // the write queue.
+          callback_thread = ROCKSDB_NAMESPACE::port::Thread(snap_callback);
+          TEST_SYNC_POINT("callback:end");
+        }
+      };
+      // Wait for the first snapshot be taken in GetSnapshotInternal. Although
+      // it might be updated before GetSnapshotInternal finishes but this should
+      // cover most of the cases.
+      ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->LoadDependency({
+          {"WritePreparedTxnDB::GetSnapshotInternal:first", "callback:end"},
+      });
+      SyncPoint::GetInstance()->SetCallBack("RemovePrepared:Start", callback);
+      SyncPoint::GetInstance()->EnableProcessing();
+      // Thread to cause frequent evictions
+      ROCKSDB_NAMESPACE::port::Thread eviction_thread([&]() {
+        // Too many txns might cause commit_seq - prepare_seq in another thread
+        // to go beyond DELTA_UPPERBOUND
+        for (int i = 0; i < 25 * (1 << commit_cache_bits); i++) {
+          ASSERT_OK(db->Put(WriteOptions(), Slice("key1"), Slice("value1")));
+        }
+      });
+      ROCKSDB_NAMESPACE::port::Thread write_thread([&]() {
+        for (int i = 0; i < 25 * (1 << commit_cache_bits); i++) {
+          Transaction* txn =
+              db->BeginTransaction(WriteOptions(), TransactionOptions());
+          ASSERT_OK(txn->SetName("xid"));
+          std::string val_str = "value" + std::to_string(i);
+          for (size_t b = 0; b < sub_batch_cnt; b++) {
+            ASSERT_OK(txn->Put(Slice("key2"), val_str));
+          }
+          ASSERT_OK(txn->Prepare());
+          // Let an eviction to kick in
+          std::this_thread::yield();
+
+          exp_prepare.store(txn->GetId());
+          ASSERT_OK(txn->Commit());
+          delete txn;
+          // Wait for the snapshot taking that is triggered by
+          // RemovePrepared:Start callback
+          callback_thread.join();
+
+          // Read with the snapshot taken before delayed_prepared_ cleanup
+          ReadOptions roptions;
+          roptions.snapshot = snap.load();
+          ASSERT_NE(nullptr, roptions.snapshot);
+          PinnableSlice value2;
+          auto s =
+              db->Get(roptions, db->DefaultColumnFamily(), "key2", &value2);
+          ASSERT_OK(s);
+          // It should see its own write
+          ASSERT_TRUE(val_str == value2);
+          // The value read by snapshot should not change
+          ASSERT_STREQ(value2.ToString().c_str(), value.ToString().c_str());
+
+          db->ReleaseSnapshot(roptions.snapshot);
+          snap.store(nullptr);
+        }
+      });
+      write_thread.join();
+      eviction_thread.join();
+      ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->DisableProcessing();
+      ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->ClearAllCallBacks();
+    }
+  }
+}
+
+// Test that updating the commit map will not affect the existing snapshots
+TEST_P(WritePreparedTransactionTest, AtomicCommit) {
+  for (bool skip_prepare : {true, false}) {
+    ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->LoadDependency({
+        {"WritePreparedTxnDB::AddCommitted:start",
+         "AtomicCommit::GetSnapshot:start"},
+        {"AtomicCommit::Get:end",
+         "WritePreparedTxnDB::AddCommitted:start:pause"},
+        {"WritePreparedTxnDB::AddCommitted:end", "AtomicCommit::Get2:start"},
+        {"AtomicCommit::Get2:end",
+         "WritePreparedTxnDB::AddCommitted:end:pause:"},
+    });
+    ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->EnableProcessing();
+    ROCKSDB_NAMESPACE::port::Thread write_thread([&]() {
+      if (skip_prepare) {
+        ASSERT_OK(db->Put(WriteOptions(), Slice("key"), Slice("value")));
+      } else {
+        Transaction* txn =
+            db->BeginTransaction(WriteOptions(), TransactionOptions());
+        ASSERT_OK(txn->SetName("xid"));
+        ASSERT_OK(txn->Put(Slice("key"), Slice("value")));
+        ASSERT_OK(txn->Prepare());
+        ASSERT_OK(txn->Commit());
+        delete txn;
+      }
+    });
+    ROCKSDB_NAMESPACE::port::Thread read_thread([&]() {
+      ReadOptions roptions;
+      TEST_SYNC_POINT("AtomicCommit::GetSnapshot:start");
+      roptions.snapshot = db->GetSnapshot();
+      PinnableSlice val;
+      auto s = db->Get(roptions, db->DefaultColumnFamily(), "key", &val);
+      TEST_SYNC_POINT("AtomicCommit::Get:end");
+      TEST_SYNC_POINT("AtomicCommit::Get2:start");
+      ASSERT_SAME(roptions, db, s, val, "key");
+      TEST_SYNC_POINT("AtomicCommit::Get2:end");
+      db->ReleaseSnapshot(roptions.snapshot);
+    });
+    read_thread.join();
+    write_thread.join();
+    ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->DisableProcessing();
+  }
+}
+
+TEST_P(WritePreparedTransactionTest, BasicRollbackDeletionTypeCb) {
+  options.level0_file_num_compaction_trigger = 2;
+  // Always use SingleDelete to rollback Put.
+  txn_db_options.rollback_deletion_type_callback =
+      [](TransactionDB*, ColumnFamilyHandle*, const Slice&) { return true; };
+
+  const auto write_to_db = [&]() {
+    assert(db);
+    std::unique_ptr<Transaction> txn0(
+        db->BeginTransaction(WriteOptions(), TransactionOptions()));
+    ASSERT_OK(txn0->SetName("txn0"));
+    ASSERT_OK(txn0->Put("a", "v0"));
+    ASSERT_OK(txn0->Prepare());
+
+    // Generate sst1: [PUT('a')]
+    ASSERT_OK(db->Flush(FlushOptions()));
+
+    {
+      CompactRangeOptions cro;
+      cro.change_level = true;
+      cro.target_level = options.num_levels - 1;
+      cro.bottommost_level_compaction = BottommostLevelCompaction::kForce;
+      ASSERT_OK(db->CompactRange(cro, /*begin=*/nullptr, /*end=*/nullptr));
+    }
+
+    ASSERT_OK(txn0->Rollback());
+    txn0.reset();
+
+    ASSERT_OK(db->Put(WriteOptions(), "a", "v1"));
+
+    ASSERT_OK(db->SingleDelete(WriteOptions(), "a"));
+    // Generate another SST with a SD to cover the oldest PUT('a')
+    ASSERT_OK(db->Flush(FlushOptions()));
+
+    auto* dbimpl = static_cast_with_check<DBImpl>(db->GetRootDB());
+    assert(dbimpl);
+    ASSERT_OK(dbimpl->TEST_WaitForCompact());
+
+    {
+      CompactRangeOptions cro;
+      cro.bottommost_level_compaction = BottommostLevelCompaction::kForce;
+      ASSERT_OK(db->CompactRange(cro, /*begin=*/nullptr, /*end=*/nullptr));
+    }
+
+    {
+      std::string value;
+      const Status s = db->Get(ReadOptions(), "a", &value);
+      ASSERT_TRUE(s.IsNotFound());
+    }
+  };
+
+  // Destroy and reopen
+  ASSERT_OK(ReOpen());
+  write_to_db();
+}
+
+// Test that we can change write policy from WriteCommitted to WritePrepared
+// after a clean shutdown (which would empty the WAL)
+TEST_P(WritePreparedTransactionTest, WP_WC_DBBackwardCompatibility) {
+  bool empty_wal = true;
+  CrossCompatibilityTest(WRITE_COMMITTED, WRITE_PREPARED, empty_wal);
+}
+
+// Test that we fail fast if WAL is not emptied between changing the write
+// policy from WriteCommitted to WritePrepared
+TEST_P(WritePreparedTransactionTest, WP_WC_WALBackwardIncompatibility) {
+  bool empty_wal = true;
+  CrossCompatibilityTest(WRITE_COMMITTED, WRITE_PREPARED, !empty_wal);
+}
+
+// Test that we can change write policy from WritePrepare back to WriteCommitted
+// after a clean shutdown (which would empty the WAL)
+TEST_P(WritePreparedTransactionTest, WC_WP_ForwardCompatibility) {
+  bool empty_wal = true;
+  CrossCompatibilityTest(WRITE_PREPARED, WRITE_COMMITTED, empty_wal);
+}
+
+// Test that we fail fast if WAL is not emptied between changing the write
+// policy from WriteCommitted to WritePrepared
+TEST_P(WritePreparedTransactionTest, WC_WP_WALForwardIncompatibility) {
+  bool empty_wal = true;
+  CrossCompatibilityTest(WRITE_PREPARED, WRITE_COMMITTED, !empty_wal);
+}
+
+}  // namespace ROCKSDB_NAMESPACE
+
+int main(int argc, char** argv) {
+  ROCKSDB_NAMESPACE::port::InstallStackTraceHandler();
+  ::testing::InitGoogleTest(&argc, argv);
+  if (getenv("CIRCLECI")) {
+    // Looking for backtrace on "Resource temporarily unavailable" exceptions
+    ::testing::FLAGS_gtest_catch_exceptions = false;
+  }
+  return RUN_ALL_TESTS();
+}
+
+#else
+#include <stdio.h>
+
+int main(int /*argc*/, char** /*argv*/) {
+  fprintf(stderr,
+          "SKIPPED as Transactions are not supported in ROCKSDB_LITE\n");
+  return 0;
+}
+
+#endif  // ROCKSDB_LITE
diff --git a/src/rocksdb/utilities/transactions/write_prepared_txn.cc b/src/rocksdb/utilities/transactions/write_prepared_txn.cc
new file mode 100644
index 000000000..16b5cc1cb
--- /dev/null
+++ b/src/rocksdb/utilities/transactions/write_prepared_txn.cc
@@ -0,0 +1,512 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#ifndef ROCKSDB_LITE
+
+#include "utilities/transactions/write_prepared_txn.h"
+
+#include <cinttypes>
+#include <map>
+#include <set>
+
+#include "db/column_family.h"
+#include "db/db_impl/db_impl.h"
+#include "rocksdb/db.h"
+#include "rocksdb/status.h"
+#include "rocksdb/utilities/transaction_db.h"
+#include "util/cast_util.h"
+#include "utilities/transactions/pessimistic_transaction.h"
+#include "utilities/transactions/write_prepared_txn_db.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+struct WriteOptions;
+
+WritePreparedTxn::WritePreparedTxn(WritePreparedTxnDB* txn_db,
+                                   const WriteOptions& write_options,
+                                   const TransactionOptions& txn_options)
+    : PessimisticTransaction(txn_db, write_options, txn_options, false),
+      wpt_db_(txn_db) {
+  // Call Initialize outside PessimisticTransaction constructor otherwise it
+  // would skip overridden functions in WritePreparedTxn since they are not
+  // defined yet in the constructor of PessimisticTransaction
+  Initialize(txn_options);
+}
+
+void WritePreparedTxn::Initialize(const TransactionOptions& txn_options) {
+  PessimisticTransaction::Initialize(txn_options);
+  prepare_batch_cnt_ = 0;
+}
+
+void WritePreparedTxn::MultiGet(const ReadOptions& options,
+                                ColumnFamilyHandle* column_family,
+                                const size_t num_keys, const Slice* keys,
+                                PinnableSlice* values, Status* statuses,
+                                const bool sorted_input) {
+  SequenceNumber min_uncommitted, snap_seq;
+  const SnapshotBackup backed_by_snapshot =
+      wpt_db_->AssignMinMaxSeqs(options.snapshot, &min_uncommitted, &snap_seq);
+  WritePreparedTxnReadCallback callback(wpt_db_, snap_seq, min_uncommitted,
+                                        backed_by_snapshot);
+  write_batch_.MultiGetFromBatchAndDB(db_, options, column_family, num_keys,
+                                      keys, values, statuses, sorted_input,
+                                      &callback);
+  if (UNLIKELY(!callback.valid() ||
+               !wpt_db_->ValidateSnapshot(snap_seq, backed_by_snapshot))) {
+    wpt_db_->WPRecordTick(TXN_GET_TRY_AGAIN);
+    for (size_t i = 0; i < num_keys; i++) {
+      statuses[i] = Status::TryAgain();
+    }
+  }
+}
+
+Status WritePreparedTxn::Get(const ReadOptions& options,
+                             ColumnFamilyHandle* column_family,
+                             const Slice& key, PinnableSlice* pinnable_val) {
+  SequenceNumber min_uncommitted, snap_seq;
+  const SnapshotBackup backed_by_snapshot =
+      wpt_db_->AssignMinMaxSeqs(options.snapshot, &min_uncommitted, &snap_seq);
+  WritePreparedTxnReadCallback callback(wpt_db_, snap_seq, min_uncommitted,
+                                        backed_by_snapshot);
+  Status res = write_batch_.GetFromBatchAndDB(db_, options, column_family, key,
+                                              pinnable_val, &callback);
+  const bool callback_valid =
+      callback.valid();  // NOTE: validity of callback must always be checked
+                         // before it is destructed
+  if (res.ok()) {
+    if (!LIKELY(callback_valid &&
+                wpt_db_->ValidateSnapshot(callback.max_visible_seq(),
+                                          backed_by_snapshot))) {
+      wpt_db_->WPRecordTick(TXN_GET_TRY_AGAIN);
+      res = Status::TryAgain();
+    }
+  }
+
+  return res;
+}
+
+Iterator* WritePreparedTxn::GetIterator(const ReadOptions& options) {
+  // Make sure to get iterator from WritePrepareTxnDB, not the root db.
+  Iterator* db_iter = wpt_db_->NewIterator(options);
+  assert(db_iter);
+
+  return write_batch_.NewIteratorWithBase(db_iter);
+}
+
+Iterator* WritePreparedTxn::GetIterator(const ReadOptions& options,
+                                        ColumnFamilyHandle* column_family) {
+  // Make sure to get iterator from WritePrepareTxnDB, not the root db.
+  Iterator* db_iter = wpt_db_->NewIterator(options, column_family);
+  assert(db_iter);
+
+  return write_batch_.NewIteratorWithBase(column_family, db_iter);
+}
+
+Status WritePreparedTxn::PrepareInternal() {
+  WriteOptions write_options = write_options_;
+  write_options.disableWAL = false;
+  const bool WRITE_AFTER_COMMIT = true;
+  const bool kFirstPrepareBatch = true;
+  auto s = WriteBatchInternal::MarkEndPrepare(GetWriteBatch()->GetWriteBatch(),
+                                              name_, !WRITE_AFTER_COMMIT);
+  assert(s.ok());
+  // For each duplicate key we account for a new sub-batch
+  prepare_batch_cnt_ = GetWriteBatch()->SubBatchCnt();
+  // Having AddPrepared in the PreReleaseCallback allows in-order addition of
+  // prepared entries to PreparedHeap and hence enables an optimization. Refer
+  // to SmallestUnCommittedSeq for more details.
+  AddPreparedCallback add_prepared_callback(
+      wpt_db_, db_impl_, prepare_batch_cnt_,
+      db_impl_->immutable_db_options().two_write_queues, kFirstPrepareBatch);
+  const bool DISABLE_MEMTABLE = true;
+  uint64_t seq_used = kMaxSequenceNumber;
+  s = db_impl_->WriteImpl(write_options, GetWriteBatch()->GetWriteBatch(),
+                          /*callback*/ nullptr, &log_number_, /*log ref*/ 0,
+                          !DISABLE_MEMTABLE, &seq_used, prepare_batch_cnt_,
+                          &add_prepared_callback);
+  assert(!s.ok() || seq_used != kMaxSequenceNumber);
+  auto prepare_seq = seq_used;
+  SetId(prepare_seq);
+  return s;
+}
+
+Status WritePreparedTxn::CommitWithoutPrepareInternal() {
+  // For each duplicate key we account for a new sub-batch
+  const size_t batch_cnt = GetWriteBatch()->SubBatchCnt();
+  return CommitBatchInternal(GetWriteBatch()->GetWriteBatch(), batch_cnt);
+}
+
+Status WritePreparedTxn::CommitBatchInternal(WriteBatch* batch,
+                                             size_t batch_cnt) {
+  return wpt_db_->WriteInternal(write_options_, batch, batch_cnt, this);
+}
+
+Status WritePreparedTxn::CommitInternal() {
+  ROCKS_LOG_DETAILS(db_impl_->immutable_db_options().info_log,
+                    "CommitInternal prepare_seq: %" PRIu64, GetID());
+  // We take the commit-time batch and append the Commit marker.
+  // The Memtable will ignore the Commit marker in non-recovery mode
+  WriteBatch* working_batch = GetCommitTimeWriteBatch();
+  const bool empty = working_batch->Count() == 0;
+  auto s = WriteBatchInternal::MarkCommit(working_batch, name_);
+  assert(s.ok());
+
+  const bool for_recovery = use_only_the_last_commit_time_batch_for_recovery_;
+  if (!empty) {
+    // When not writing to memtable, we can still cache the latest write batch.
+    // The cached batch will be written to memtable in WriteRecoverableState
+    // during FlushMemTable
+    if (for_recovery) {
+      WriteBatchInternal::SetAsLatestPersistentState(working_batch);
+    } else {
+      return Status::InvalidArgument(
+          "Commit-time-batch can only be used if "
+          "use_only_the_last_commit_time_batch_for_recovery is true");
+    }
+  }
+
+  auto prepare_seq = GetId();
+  const bool includes_data = !empty && !for_recovery;
+  assert(prepare_batch_cnt_);
+  size_t commit_batch_cnt = 0;
+  if (UNLIKELY(includes_data)) {
+    ROCKS_LOG_WARN(db_impl_->immutable_db_options().info_log,
+                   "Duplicate key overhead");
+    SubBatchCounter counter(*wpt_db_->GetCFComparatorMap());
+    s = working_batch->Iterate(&counter);
+    assert(s.ok());
+    commit_batch_cnt = counter.BatchCount();
+  }
+  const bool disable_memtable = !includes_data;
+  const bool do_one_write =
+      !db_impl_->immutable_db_options().two_write_queues || disable_memtable;
+  WritePreparedCommitEntryPreReleaseCallback update_commit_map(
+      wpt_db_, db_impl_, prepare_seq, prepare_batch_cnt_, commit_batch_cnt);
+  // This is to call AddPrepared on CommitTimeWriteBatch
+  const bool kFirstPrepareBatch = true;
+  AddPreparedCallback add_prepared_callback(
+      wpt_db_, db_impl_, commit_batch_cnt,
+      db_impl_->immutable_db_options().two_write_queues, !kFirstPrepareBatch);
+  PreReleaseCallback* pre_release_callback;
+  if (do_one_write) {
+    pre_release_callback = &update_commit_map;
+  } else {
+    pre_release_callback = &add_prepared_callback;
+  }
+  uint64_t seq_used = kMaxSequenceNumber;
+  // Since the prepared batch is directly written to memtable, there is already
+  // a connection between the memtable and its WAL, so there is no need to
+  // redundantly reference the log that contains the prepared data.
+  const uint64_t zero_log_number = 0ull;
+  size_t batch_cnt = UNLIKELY(commit_batch_cnt) ? commit_batch_cnt : 1;
+  // If `two_write_queues && includes_data`, then `do_one_write` is false. The
+  // following `WriteImpl` will insert the data of the commit-time-batch into
+  // the database before updating the commit cache. Therefore, the data of the
+  // commmit-time-batch is considered uncommitted. Furthermore, since data of
+  // the commit-time-batch are not locked, it is possible for two uncommitted
+  // versions of the same key to co-exist for a (short) period of time until
+  // the commit cache is updated by the second write. If the two uncommitted
+  // keys are compacted to the bottommost level in the meantime, it is possible
+  // that compaction iterator will zero out the sequence numbers of both, thus
+  // violating the invariant that an SST does not have two identical internal
+  // keys. To prevent this situation, we should allow the usage of
+  // commit-time-batch only if the user sets
+  // TransactionOptions::use_only_the_last_commit_time_batch_for_recovery to
+  // true. See the comments about GetCommitTimeWriteBatch() in
+  // include/rocksdb/utilities/transaction.h.
+  s = db_impl_->WriteImpl(write_options_, working_batch, nullptr, nullptr,
+                          zero_log_number, disable_memtable, &seq_used,
+                          batch_cnt, pre_release_callback);
+  assert(!s.ok() || seq_used != kMaxSequenceNumber);
+  const SequenceNumber commit_batch_seq = seq_used;
+  if (LIKELY(do_one_write || !s.ok())) {
+    if (UNLIKELY(!db_impl_->immutable_db_options().two_write_queues &&
+                 s.ok())) {
+      // Note: RemovePrepared should be called after WriteImpl that publishsed
+      // the seq. Otherwise SmallestUnCommittedSeq optimization breaks.
+      wpt_db_->RemovePrepared(prepare_seq, prepare_batch_cnt_);
+    }  // else RemovePrepared is called from within PreReleaseCallback
+    if (UNLIKELY(!do_one_write)) {
+      assert(!s.ok());
+      // Cleanup the prepared entry we added with add_prepared_callback
+      wpt_db_->RemovePrepared(commit_batch_seq, commit_batch_cnt);
+    }
+    return s;
+  }  // else do the 2nd write to publish seq
+  // Note: the 2nd write comes with a performance penality. So if we have too
+  // many of commits accompanied with ComitTimeWriteBatch and yet we cannot
+  // enable use_only_the_last_commit_time_batch_for_recovery_ optimization,
+  // two_write_queues should be disabled to avoid many additional writes here.
+  const size_t kZeroData = 0;
+  // Update commit map only from the 2nd queue
+  WritePreparedCommitEntryPreReleaseCallback update_commit_map_with_aux_batch(
+      wpt_db_, db_impl_, prepare_seq, prepare_batch_cnt_, kZeroData,
+      commit_batch_seq, commit_batch_cnt);
+  WriteBatch empty_batch;
+  s = empty_batch.PutLogData(Slice());
+  assert(s.ok());
+  // In the absence of Prepare markers, use Noop as a batch separator
+  s = WriteBatchInternal::InsertNoop(&empty_batch);
+  assert(s.ok());
+  const bool DISABLE_MEMTABLE = true;
+  const size_t ONE_BATCH = 1;
+  const uint64_t NO_REF_LOG = 0;
+  s = db_impl_->WriteImpl(write_options_, &empty_batch, nullptr, nullptr,
+                          NO_REF_LOG, DISABLE_MEMTABLE, &seq_used, ONE_BATCH,
+                          &update_commit_map_with_aux_batch);
+  assert(!s.ok() || seq_used != kMaxSequenceNumber);
+  return s;
+}
+
+Status WritePreparedTxn::RollbackInternal() {
+  ROCKS_LOG_WARN(db_impl_->immutable_db_options().info_log,
+                 "RollbackInternal prepare_seq: %" PRIu64, GetId());
+
+  assert(db_impl_);
+  assert(wpt_db_);
+
+  WriteBatch rollback_batch(0 /* reserved_bytes */, 0 /* max_bytes */,
+                            write_options_.protection_bytes_per_key,
+                            0 /* default_cf_ts_sz */);
+  assert(GetId() != kMaxSequenceNumber);
+  assert(GetId() > 0);
+  auto cf_map_shared_ptr = wpt_db_->GetCFHandleMap();
+  auto cf_comp_map_shared_ptr = wpt_db_->GetCFComparatorMap();
+  auto read_at_seq = kMaxSequenceNumber;
+  ReadOptions roptions;
+  // to prevent callback's seq to be overrriden inside DBImpk::Get
+  roptions.snapshot = wpt_db_->GetMaxSnapshot();
+  struct RollbackWriteBatchBuilder : public WriteBatch::Handler {
+    DBImpl* const db_;
+    WritePreparedTxnDB* const wpt_db_;
+    WritePreparedTxnReadCallback callback_;
+    WriteBatch* rollback_batch_;
+    std::map<uint32_t, const Comparator*>& comparators_;
+    std::map<uint32_t, ColumnFamilyHandle*>& handles_;
+    using CFKeys = std::set<Slice, SetComparator>;
+    std::map<uint32_t, CFKeys> keys_;
+    bool rollback_merge_operands_;
+    ReadOptions roptions_;
+
+    RollbackWriteBatchBuilder(
+        DBImpl* db, WritePreparedTxnDB* wpt_db, SequenceNumber snap_seq,
+        WriteBatch* dst_batch,
+        std::map<uint32_t, const Comparator*>& comparators,
+        std::map<uint32_t, ColumnFamilyHandle*>& handles,
+        bool rollback_merge_operands, const ReadOptions& _roptions)
+        : db_(db),
+          wpt_db_(wpt_db),
+          callback_(wpt_db, snap_seq),  // disable min_uncommitted optimization
+          rollback_batch_(dst_batch),
+          comparators_(comparators),
+          handles_(handles),
+          rollback_merge_operands_(rollback_merge_operands),
+          roptions_(_roptions) {}
+
+    Status Rollback(uint32_t cf, const Slice& key) {
+      Status s;
+      CFKeys& cf_keys = keys_[cf];
+      if (cf_keys.size() == 0) {  // just inserted
+        auto cmp = comparators_[cf];
+        keys_[cf] = CFKeys(SetComparator(cmp));
+      }
+      auto it = cf_keys.insert(key);
+      // second is false if a element already existed.
+      if (it.second == false) {
+        return s;
+      }
+
+      PinnableSlice pinnable_val;
+      bool not_used;
+      auto cf_handle = handles_[cf];
+      DBImpl::GetImplOptions get_impl_options;
+      get_impl_options.column_family = cf_handle;
+      get_impl_options.value = &pinnable_val;
+      get_impl_options.value_found = &not_used;
+      get_impl_options.callback = &callback_;
+      s = db_->GetImpl(roptions_, key, get_impl_options);
+      assert(s.ok() || s.IsNotFound());
+      if (s.ok()) {
+        s = rollback_batch_->Put(cf_handle, key, pinnable_val);
+        assert(s.ok());
+      } else if (s.IsNotFound()) {
+        // There has been no readable value before txn. By adding a delete we
+        // make sure that there will be none afterwards either.
+        if (wpt_db_->ShouldRollbackWithSingleDelete(cf_handle, key)) {
+          s = rollback_batch_->SingleDelete(cf_handle, key);
+        } else {
+          s = rollback_batch_->Delete(cf_handle, key);
+        }
+        assert(s.ok());
+      } else {
+        // Unexpected status. Return it to the user.
+      }
+      return s;
+    }
+
+    Status PutCF(uint32_t cf, const Slice& key, const Slice& /*val*/) override {
+      return Rollback(cf, key);
+    }
+
+    Status DeleteCF(uint32_t cf, const Slice& key) override {
+      return Rollback(cf, key);
+    }
+
+    Status SingleDeleteCF(uint32_t cf, const Slice& key) override {
+      return Rollback(cf, key);
+    }
+
+    Status MergeCF(uint32_t cf, const Slice& key,
+                   const Slice& /*val*/) override {
+      if (rollback_merge_operands_) {
+        return Rollback(cf, key);
+      } else {
+        return Status::OK();
+      }
+    }
+
+    Status MarkNoop(bool) override { return Status::OK(); }
+    Status MarkBeginPrepare(bool) override { return Status::OK(); }
+    Status MarkEndPrepare(const Slice&) override { return Status::OK(); }
+    Status MarkCommit(const Slice&) override { return Status::OK(); }
+    Status MarkRollback(const Slice&) override {
+      return Status::InvalidArgument();
+    }
+
+   protected:
+    Handler::OptionState WriteAfterCommit() const override {
+      return Handler::OptionState::kDisabled;
+    }
+  } rollback_handler(db_impl_, wpt_db_, read_at_seq, &rollback_batch,
+                     *cf_comp_map_shared_ptr.get(), *cf_map_shared_ptr.get(),
+                     wpt_db_->txn_db_options_.rollback_merge_operands,
+                     roptions);
+  auto s = GetWriteBatch()->GetWriteBatch()->Iterate(&rollback_handler);
+  if (!s.ok()) {
+    return s;
+  }
+  // The Rollback marker will be used as a batch separator
+  s = WriteBatchInternal::MarkRollback(&rollback_batch, name_);
+  assert(s.ok());
+  bool do_one_write = !db_impl_->immutable_db_options().two_write_queues;
+  const bool DISABLE_MEMTABLE = true;
+  const uint64_t NO_REF_LOG = 0;
+  uint64_t seq_used = kMaxSequenceNumber;
+  const size_t ONE_BATCH = 1;
+  const bool kFirstPrepareBatch = true;
+  // We commit the rolled back prepared batches. Although this is
+  // counter-intuitive, i) it is safe to do so, since the prepared batches are
+  // already canceled out by the rollback batch, ii) adding the commit entry to
+  // CommitCache will allow us to benefit from the existing mechanism in
+  // CommitCache that keeps an entry evicted due to max advance and yet overlaps
+  // with a live snapshot around so that the live snapshot properly skips the
+  // entry even if its prepare seq is lower than max_evicted_seq_.
+  AddPreparedCallback add_prepared_callback(
+      wpt_db_, db_impl_, ONE_BATCH,
+      db_impl_->immutable_db_options().two_write_queues, !kFirstPrepareBatch);
+  WritePreparedCommitEntryPreReleaseCallback update_commit_map(
+      wpt_db_, db_impl_, GetId(), prepare_batch_cnt_, ONE_BATCH);
+  PreReleaseCallback* pre_release_callback;
+  if (do_one_write) {
+    pre_release_callback = &update_commit_map;
+  } else {
+    pre_release_callback = &add_prepared_callback;
+  }
+  // Note: the rollback batch does not need AddPrepared since it is written to
+  // DB in one shot. min_uncommitted still works since it requires capturing
+  // data that is written to DB but not yet committed, while
+  // the rollback batch commits with PreReleaseCallback.
+  s = db_impl_->WriteImpl(write_options_, &rollback_batch, nullptr, nullptr,
+                          NO_REF_LOG, !DISABLE_MEMTABLE, &seq_used, ONE_BATCH,
+                          pre_release_callback);
+  assert(!s.ok() || seq_used != kMaxSequenceNumber);
+  if (!s.ok()) {
+    return s;
+  }
+  if (do_one_write) {
+    assert(!db_impl_->immutable_db_options().two_write_queues);
+    wpt_db_->RemovePrepared(GetId(), prepare_batch_cnt_);
+    return s;
+  }  // else do the 2nd write for commit
+  uint64_t rollback_seq = seq_used;
+  ROCKS_LOG_DETAILS(db_impl_->immutable_db_options().info_log,
+                    "RollbackInternal 2nd write rollback_seq: %" PRIu64,
+                    rollback_seq);
+  // Commit the batch by writing an empty batch to the queue that will release
+  // the commit sequence number to readers.
+  WritePreparedRollbackPreReleaseCallback update_commit_map_with_prepare(
+      wpt_db_, db_impl_, GetId(), rollback_seq, prepare_batch_cnt_);
+  WriteBatch empty_batch;
+  s = empty_batch.PutLogData(Slice());
+  assert(s.ok());
+  // In the absence of Prepare markers, use Noop as a batch separator
+  s = WriteBatchInternal::InsertNoop(&empty_batch);
+  assert(s.ok());
+  s = db_impl_->WriteImpl(write_options_, &empty_batch, nullptr, nullptr,
+                          NO_REF_LOG, DISABLE_MEMTABLE, &seq_used, ONE_BATCH,
+                          &update_commit_map_with_prepare);
+  assert(!s.ok() || seq_used != kMaxSequenceNumber);
+  ROCKS_LOG_DETAILS(db_impl_->immutable_db_options().info_log,
+                    "RollbackInternal (status=%s) commit: %" PRIu64,
+                    s.ToString().c_str(), GetId());
+  // TODO(lth): For WriteUnPrepared that rollback is called frequently,
+  // RemovePrepared could be moved to the callback to reduce lock contention.
+  if (s.ok()) {
+    wpt_db_->RemovePrepared(GetId(), prepare_batch_cnt_);
+  }
+  // Note: RemovePrepared for prepared batch is called from within
+  // PreReleaseCallback
+  wpt_db_->RemovePrepared(rollback_seq, ONE_BATCH);
+
+  return s;
+}
+
+Status WritePreparedTxn::ValidateSnapshot(ColumnFamilyHandle* column_family,
+                                          const Slice& key,
+                                          SequenceNumber* tracked_at_seq) {
+  assert(snapshot_);
+
+  SequenceNumber min_uncommitted =
+      static_cast_with_check<const SnapshotImpl>(snapshot_.get())
+          ->min_uncommitted_;
+  SequenceNumber snap_seq = snapshot_->GetSequenceNumber();
+  // tracked_at_seq is either max or the last snapshot with which this key was
+  // trackeed so there is no need to apply the IsInSnapshot to this comparison
+  // here as tracked_at_seq is not a prepare seq.
+  if (*tracked_at_seq <= snap_seq) {
+    // If the key has been previous validated at a sequence number earlier
+    // than the curent snapshot's sequence number, we already know it has not
+    // been modified.
+    return Status::OK();
+  }
+
+  *tracked_at_seq = snap_seq;
+
+  ColumnFamilyHandle* cfh =
+      column_family ? column_family : db_impl_->DefaultColumnFamily();
+
+  WritePreparedTxnReadCallback snap_checker(wpt_db_, snap_seq, min_uncommitted,
+                                            kBackedByDBSnapshot);
+  // TODO(yanqin): support user-defined timestamp
+  return TransactionUtil::CheckKeyForConflicts(
+      db_impl_, cfh, key.ToString(), snap_seq, /*ts=*/nullptr,
+      false /* cache_only */, &snap_checker, min_uncommitted);
+}
+
+void WritePreparedTxn::SetSnapshot() {
+  const bool kForWWConflictCheck = true;
+  SnapshotImpl* snapshot = wpt_db_->GetSnapshotInternal(kForWWConflictCheck);
+  SetSnapshotInternal(snapshot);
+}
+
+Status WritePreparedTxn::RebuildFromWriteBatch(WriteBatch* src_batch) {
+  auto ret = PessimisticTransaction::RebuildFromWriteBatch(src_batch);
+  prepare_batch_cnt_ = GetWriteBatch()->SubBatchCnt();
+  return ret;
+}
+
+}  // namespace ROCKSDB_NAMESPACE
+
+#endif  // ROCKSDB_LITE
diff --git a/src/rocksdb/utilities/transactions/write_prepared_txn.h b/src/rocksdb/utilities/transactions/write_prepared_txn.h
new file mode 100644
index 000000000..30d9bdb99
--- /dev/null
+++ b/src/rocksdb/utilities/transactions/write_prepared_txn.h
@@ -0,0 +1,119 @@
+// Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#pragma once
+
+#ifndef ROCKSDB_LITE
+
+#include <algorithm>
+#include <atomic>
+#include <mutex>
+#include <stack>
+#include <string>
+#include <unordered_map>
+#include <vector>
+
+#include "db/write_callback.h"
+#include "rocksdb/db.h"
+#include "rocksdb/slice.h"
+#include "rocksdb/snapshot.h"
+#include "rocksdb/status.h"
+#include "rocksdb/types.h"
+#include "rocksdb/utilities/transaction.h"
+#include "rocksdb/utilities/transaction_db.h"
+#include "rocksdb/utilities/write_batch_with_index.h"
+#include "util/autovector.h"
+#include "utilities/transactions/pessimistic_transaction.h"
+#include "utilities/transactions/pessimistic_transaction_db.h"
+#include "utilities/transactions/transaction_base.h"
+#include "utilities/transactions/transaction_util.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+class WritePreparedTxnDB;
+
+// This impl could write to DB also uncommitted data and then later tell apart
+// committed data from uncommitted data. Uncommitted data could be after the
+// Prepare phase in 2PC (WritePreparedTxn) or before that
+// (WriteUnpreparedTxnImpl).
+class WritePreparedTxn : public PessimisticTransaction {
+ public:
+  WritePreparedTxn(WritePreparedTxnDB* db, const WriteOptions& write_options,
+                   const TransactionOptions& txn_options);
+  // No copying allowed
+  WritePreparedTxn(const WritePreparedTxn&) = delete;
+  void operator=(const WritePreparedTxn&) = delete;
+
+  virtual ~WritePreparedTxn() {}
+
+  // To make WAL commit markers visible, the snapshot will be based on the last
+  // seq in the WAL that is also published, LastPublishedSequence, as opposed to
+  // the last seq in the memtable.
+  using Transaction::Get;
+  virtual Status Get(const ReadOptions& options,
+                     ColumnFamilyHandle* column_family, const Slice& key,
+                     PinnableSlice* value) override;
+
+  using Transaction::MultiGet;
+  virtual void MultiGet(const ReadOptions& options,
+                        ColumnFamilyHandle* column_family,
+                        const size_t num_keys, const Slice* keys,
+                        PinnableSlice* values, Status* statuses,
+                        const bool sorted_input = false) override;
+
+  // Note: The behavior is undefined in presence of interleaved writes to the
+  // same transaction.
+  // To make WAL commit markers visible, the snapshot will be
+  // based on the last seq in the WAL that is also published,
+  // LastPublishedSequence, as opposed to the last seq in the memtable.
+  using Transaction::GetIterator;
+  virtual Iterator* GetIterator(const ReadOptions& options) override;
+  virtual Iterator* GetIterator(const ReadOptions& options,
+                                ColumnFamilyHandle* column_family) override;
+
+  virtual void SetSnapshot() override;
+
+ protected:
+  void Initialize(const TransactionOptions& txn_options) override;
+  // Override the protected SetId to make it visible to the friend class
+  // WritePreparedTxnDB
+  inline void SetId(uint64_t id) override { Transaction::SetId(id); }
+
+ private:
+  friend class WritePreparedTransactionTest_BasicRecoveryTest_Test;
+  friend class WritePreparedTxnDB;
+  friend class WriteUnpreparedTxnDB;
+  friend class WriteUnpreparedTxn;
+
+  Status PrepareInternal() override;
+
+  Status CommitWithoutPrepareInternal() override;
+
+  Status CommitBatchInternal(WriteBatch* batch, size_t batch_cnt) override;
+
+  // Since the data is already written to memtables at the Prepare phase, the
+  // commit entails writing only a commit marker in the WAL. The sequence number
+  // of the commit marker is then the commit timestamp of the transaction. To
+  // make WAL commit markers visible, the snapshot will be based on the last seq
+  // in the WAL that is also published, LastPublishedSequence, as opposed to the
+  // last seq in the memtable.
+  Status CommitInternal() override;
+
+  Status RollbackInternal() override;
+
+  virtual Status ValidateSnapshot(ColumnFamilyHandle* column_family,
+                                  const Slice& key,
+                                  SequenceNumber* tracked_at_seq) override;
+
+  virtual Status RebuildFromWriteBatch(WriteBatch* src_batch) override;
+
+  WritePreparedTxnDB* wpt_db_;
+  // Number of sub-batches in prepare
+  size_t prepare_batch_cnt_ = 0;
+};
+
+}  // namespace ROCKSDB_NAMESPACE
+
+#endif  // ROCKSDB_LITE
diff --git a/src/rocksdb/utilities/transactions/write_prepared_txn_db.cc b/src/rocksdb/utilities/transactions/write_prepared_txn_db.cc
new file mode 100644
index 000000000..595c3df8f
--- /dev/null
+++ b/src/rocksdb/utilities/transactions/write_prepared_txn_db.cc
@@ -0,0 +1,1030 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#ifndef ROCKSDB_LITE
+
+#include "utilities/transactions/write_prepared_txn_db.h"
+
+#include <algorithm>
+#include <cinttypes>
+#include <string>
+#include <unordered_set>
+#include <vector>
+
+#include "db/arena_wrapped_db_iter.h"
+#include "db/db_impl/db_impl.h"
+#include "logging/logging.h"
+#include "rocksdb/db.h"
+#include "rocksdb/options.h"
+#include "rocksdb/utilities/transaction_db.h"
+#include "test_util/sync_point.h"
+#include "util/cast_util.h"
+#include "util/mutexlock.h"
+#include "util/string_util.h"
+#include "utilities/transactions/pessimistic_transaction.h"
+#include "utilities/transactions/transaction_db_mutex_impl.h"
+
+// This function is for testing only. If it returns true, then all entries in
+// the commit cache will be evicted. Unit and/or stress tests (db_stress)
+// can implement this function and customize how frequently commit cache
+// eviction occurs.
+// TODO: remove this function once we can configure commit cache to be very
+// small so that eviction occurs very frequently. This requires the commit
+// cache entry to be able to encode prepare and commit sequence numbers so that
+// the commit sequence number does not have to be within a certain range of
+// prepare sequence number.
+extern "C" bool rocksdb_write_prepared_TEST_ShouldClearCommitCache(void)
+    __attribute__((__weak__));
+
+namespace ROCKSDB_NAMESPACE {
+
+Status WritePreparedTxnDB::Initialize(
+    const std::vector<size_t>& compaction_enabled_cf_indices,
+    const std::vector<ColumnFamilyHandle*>& handles) {
+  auto dbimpl = static_cast_with_check<DBImpl>(GetRootDB());
+  assert(dbimpl != nullptr);
+  auto rtxns = dbimpl->recovered_transactions();
+  std::map<SequenceNumber, SequenceNumber> ordered_seq_cnt;
+  for (auto rtxn : rtxns) {
+    // There should only one batch for WritePrepared policy.
+    assert(rtxn.second->batches_.size() == 1);
+    const auto& seq = rtxn.second->batches_.begin()->first;
+    const auto& batch_info = rtxn.second->batches_.begin()->second;
+    auto cnt = batch_info.batch_cnt_ ? batch_info.batch_cnt_ : 1;
+    ordered_seq_cnt[seq] = cnt;
+  }
+  // AddPrepared must be called in order
+  for (auto seq_cnt : ordered_seq_cnt) {
+    auto seq = seq_cnt.first;
+    auto cnt = seq_cnt.second;
+    for (size_t i = 0; i < cnt; i++) {
+      AddPrepared(seq + i);
+    }
+  }
+  SequenceNumber prev_max = max_evicted_seq_;
+  SequenceNumber last_seq = db_impl_->GetLatestSequenceNumber();
+  AdvanceMaxEvictedSeq(prev_max, last_seq);
+  // Create a gap between max and the next snapshot. This simplifies the logic
+  // in IsInSnapshot by not having to consider the special case of max ==
+  // snapshot after recovery. This is tested in IsInSnapshotEmptyMapTest.
+  if (last_seq) {
+    db_impl_->versions_->SetLastAllocatedSequence(last_seq + 1);
+    db_impl_->versions_->SetLastSequence(last_seq + 1);
+    db_impl_->versions_->SetLastPublishedSequence(last_seq + 1);
+  }
+
+  db_impl_->SetSnapshotChecker(new WritePreparedSnapshotChecker(this));
+  // A callback to commit a single sub-batch
+  class CommitSubBatchPreReleaseCallback : public PreReleaseCallback {
+   public:
+    explicit CommitSubBatchPreReleaseCallback(WritePreparedTxnDB* db)
+        : db_(db) {}
+    Status Callback(SequenceNumber commit_seq,
+                    bool is_mem_disabled __attribute__((__unused__)), uint64_t,
+                    size_t /*index*/, size_t /*total*/) override {
+      assert(!is_mem_disabled);
+      db_->AddCommitted(commit_seq, commit_seq);
+      return Status::OK();
+    }
+
+   private:
+    WritePreparedTxnDB* db_;
+  };
+  db_impl_->SetRecoverableStatePreReleaseCallback(
+      new CommitSubBatchPreReleaseCallback(this));
+
+  auto s = PessimisticTransactionDB::Initialize(compaction_enabled_cf_indices,
+                                                handles);
+  return s;
+}
+
+Status WritePreparedTxnDB::VerifyCFOptions(
+    const ColumnFamilyOptions& cf_options) {
+  Status s = PessimisticTransactionDB::VerifyCFOptions(cf_options);
+  if (!s.ok()) {
+    return s;
+  }
+  if (!cf_options.memtable_factory->CanHandleDuplicatedKey()) {
+    return Status::InvalidArgument(
+        "memtable_factory->CanHandleDuplicatedKey() cannot be false with "
+        "WritePrpeared transactions");
+  }
+  return Status::OK();
+}
+
+Transaction* WritePreparedTxnDB::BeginTransaction(
+    const WriteOptions& write_options, const TransactionOptions& txn_options,
+    Transaction* old_txn) {
+  if (old_txn != nullptr) {
+    ReinitializeTransaction(old_txn, write_options, txn_options);
+    return old_txn;
+  } else {
+    return new WritePreparedTxn(this, write_options, txn_options);
+  }
+}
+
+Status WritePreparedTxnDB::Write(const WriteOptions& opts,
+                                 WriteBatch* updates) {
+  if (txn_db_options_.skip_concurrency_control) {
+    // Skip locking the rows
+    const size_t UNKNOWN_BATCH_CNT = 0;
+    WritePreparedTxn* NO_TXN = nullptr;
+    return WriteInternal(opts, updates, UNKNOWN_BATCH_CNT, NO_TXN);
+  } else {
+    return PessimisticTransactionDB::WriteWithConcurrencyControl(opts, updates);
+  }
+}
+
+Status WritePreparedTxnDB::Write(
+    const WriteOptions& opts,
+    const TransactionDBWriteOptimizations& optimizations, WriteBatch* updates) {
+  if (optimizations.skip_concurrency_control) {
+    // Skip locking the rows
+    const size_t UNKNOWN_BATCH_CNT = 0;
+    const size_t ONE_BATCH_CNT = 1;
+    const size_t batch_cnt = optimizations.skip_duplicate_key_check
+                                 ? ONE_BATCH_CNT
+                                 : UNKNOWN_BATCH_CNT;
+    WritePreparedTxn* NO_TXN = nullptr;
+    return WriteInternal(opts, updates, batch_cnt, NO_TXN);
+  } else {
+    // TODO(myabandeh): Make use of skip_duplicate_key_check hint
+    // Fall back to unoptimized version
+    return PessimisticTransactionDB::WriteWithConcurrencyControl(opts, updates);
+  }
+}
+
+Status WritePreparedTxnDB::WriteInternal(const WriteOptions& write_options_orig,
+                                         WriteBatch* batch, size_t batch_cnt,
+                                         WritePreparedTxn* txn) {
+  ROCKS_LOG_DETAILS(db_impl_->immutable_db_options().info_log,
+                    "CommitBatchInternal");
+  if (batch->Count() == 0) {
+    // Otherwise our 1 seq per batch logic will break since there is no seq
+    // increased for this batch.
+    return Status::OK();
+  }
+
+  if (write_options_orig.protection_bytes_per_key > 0) {
+    auto s = WriteBatchInternal::UpdateProtectionInfo(
+        batch, write_options_orig.protection_bytes_per_key);
+    if (!s.ok()) {
+      return s;
+    }
+  }
+
+  if (batch_cnt == 0) {  // not provided, then compute it
+    // TODO(myabandeh): add an option to allow user skipping this cost
+    SubBatchCounter counter(*GetCFComparatorMap());
+    auto s = batch->Iterate(&counter);
+    if (!s.ok()) {
+      return s;
+    }
+    batch_cnt = counter.BatchCount();
+    WPRecordTick(TXN_DUPLICATE_KEY_OVERHEAD);
+    ROCKS_LOG_DETAILS(info_log_, "Duplicate key overhead: %" PRIu64 " batches",
+                      static_cast<uint64_t>(batch_cnt));
+  }
+  assert(batch_cnt);
+
+  bool do_one_write = !db_impl_->immutable_db_options().two_write_queues;
+  WriteOptions write_options(write_options_orig);
+  // In the absence of Prepare markers, use Noop as a batch separator
+  auto s = WriteBatchInternal::InsertNoop(batch);
+  assert(s.ok());
+  const bool DISABLE_MEMTABLE = true;
+  const uint64_t no_log_ref = 0;
+  uint64_t seq_used = kMaxSequenceNumber;
+  const size_t ZERO_PREPARES = 0;
+  const bool kSeperatePrepareCommitBatches = true;
+  // Since this is not 2pc, there is no need for AddPrepared but having it in
+  // the PreReleaseCallback enables an optimization. Refer to
+  // SmallestUnCommittedSeq for more details.
+  AddPreparedCallback add_prepared_callback(
+      this, db_impl_, batch_cnt,
+      db_impl_->immutable_db_options().two_write_queues,
+      !kSeperatePrepareCommitBatches);
+  WritePreparedCommitEntryPreReleaseCallback update_commit_map(
+      this, db_impl_, kMaxSequenceNumber, ZERO_PREPARES, batch_cnt);
+  PreReleaseCallback* pre_release_callback;
+  if (do_one_write) {
+    pre_release_callback = &update_commit_map;
+  } else {
+    pre_release_callback = &add_prepared_callback;
+  }
+  s = db_impl_->WriteImpl(write_options, batch, nullptr, nullptr, no_log_ref,
+                          !DISABLE_MEMTABLE, &seq_used, batch_cnt,
+                          pre_release_callback);
+  assert(!s.ok() || seq_used != kMaxSequenceNumber);
+  uint64_t prepare_seq = seq_used;
+  if (txn != nullptr) {
+    txn->SetId(prepare_seq);
+  }
+  if (!s.ok()) {
+    return s;
+  }
+  if (do_one_write) {
+    return s;
+  }  // else do the 2nd write for commit
+  ROCKS_LOG_DETAILS(db_impl_->immutable_db_options().info_log,
+                    "CommitBatchInternal 2nd write prepare_seq: %" PRIu64,
+                    prepare_seq);
+  // Commit the batch by writing an empty batch to the 2nd queue that will
+  // release the commit sequence number to readers.
+  const size_t ZERO_COMMITS = 0;
+  WritePreparedCommitEntryPreReleaseCallback update_commit_map_with_prepare(
+      this, db_impl_, prepare_seq, batch_cnt, ZERO_COMMITS);
+  WriteBatch empty_batch;
+  write_options.disableWAL = true;
+  write_options.sync = false;
+  const size_t ONE_BATCH = 1;  // Just to inc the seq
+  s = db_impl_->WriteImpl(write_options, &empty_batch, nullptr, nullptr,
+                          no_log_ref, DISABLE_MEMTABLE, &seq_used, ONE_BATCH,
+                          &update_commit_map_with_prepare);
+  assert(!s.ok() || seq_used != kMaxSequenceNumber);
+  // Note: RemovePrepared is called from within PreReleaseCallback
+  return s;
+}
+
+Status WritePreparedTxnDB::Get(const ReadOptions& options,
+                               ColumnFamilyHandle* column_family,
+                               const Slice& key, PinnableSlice* value) {
+  SequenceNumber min_uncommitted, snap_seq;
+  const SnapshotBackup backed_by_snapshot =
+      AssignMinMaxSeqs(options.snapshot, &min_uncommitted, &snap_seq);
+  WritePreparedTxnReadCallback callback(this, snap_seq, min_uncommitted,
+                                        backed_by_snapshot);
+  bool* dont_care = nullptr;
+  DBImpl::GetImplOptions get_impl_options;
+  get_impl_options.column_family = column_family;
+  get_impl_options.value = value;
+  get_impl_options.value_found = dont_care;
+  get_impl_options.callback = &callback;
+  auto res = db_impl_->GetImpl(options, key, get_impl_options);
+  if (LIKELY(callback.valid() && ValidateSnapshot(callback.max_visible_seq(),
+                                                  backed_by_snapshot))) {
+    return res;
+  } else {
+    res.PermitUncheckedError();
+    WPRecordTick(TXN_GET_TRY_AGAIN);
+    return Status::TryAgain();
+  }
+}
+
+void WritePreparedTxnDB::UpdateCFComparatorMap(
+    const std::vector<ColumnFamilyHandle*>& handles) {
+  auto cf_map = new std::map<uint32_t, const Comparator*>();
+  auto handle_map = new std::map<uint32_t, ColumnFamilyHandle*>();
+  for (auto h : handles) {
+    auto id = h->GetID();
+    const Comparator* comparator = h->GetComparator();
+    (*cf_map)[id] = comparator;
+    if (id != 0) {
+      (*handle_map)[id] = h;
+    } else {
+      // The pointer to the default cf handle in the handles will be deleted.
+      // Use the pointer maintained by the db instead.
+      (*handle_map)[id] = DefaultColumnFamily();
+    }
+  }
+  cf_map_.reset(cf_map);
+  handle_map_.reset(handle_map);
+}
+
+void WritePreparedTxnDB::UpdateCFComparatorMap(ColumnFamilyHandle* h) {
+  auto old_cf_map_ptr = cf_map_.get();
+  assert(old_cf_map_ptr);
+  auto cf_map = new std::map<uint32_t, const Comparator*>(*old_cf_map_ptr);
+  auto old_handle_map_ptr = handle_map_.get();
+  assert(old_handle_map_ptr);
+  auto handle_map =
+      new std::map<uint32_t, ColumnFamilyHandle*>(*old_handle_map_ptr);
+  auto id = h->GetID();
+  const Comparator* comparator = h->GetComparator();
+  (*cf_map)[id] = comparator;
+  (*handle_map)[id] = h;
+  cf_map_.reset(cf_map);
+  handle_map_.reset(handle_map);
+}
+
+std::vector<Status> WritePreparedTxnDB::MultiGet(
+    const ReadOptions& options,
+    const std::vector<ColumnFamilyHandle*>& column_family,
+    const std::vector<Slice>& keys, std::vector<std::string>* values) {
+  assert(values);
+  size_t num_keys = keys.size();
+  values->resize(num_keys);
+
+  std::vector<Status> stat_list(num_keys);
+  for (size_t i = 0; i < num_keys; ++i) {
+    stat_list[i] = this->Get(options, column_family[i], keys[i], &(*values)[i]);
+  }
+  return stat_list;
+}
+
+// Struct to hold ownership of snapshot and read callback for iterator cleanup.
+struct WritePreparedTxnDB::IteratorState {
+  IteratorState(WritePreparedTxnDB* txn_db, SequenceNumber sequence,
+                std::shared_ptr<ManagedSnapshot> s,
+                SequenceNumber min_uncommitted)
+      : callback(txn_db, sequence, min_uncommitted, kBackedByDBSnapshot),
+        snapshot(s) {}
+
+  WritePreparedTxnReadCallback callback;
+  std::shared_ptr<ManagedSnapshot> snapshot;
+};
+
+namespace {
+static void CleanupWritePreparedTxnDBIterator(void* arg1, void* /*arg2*/) {
+  delete reinterpret_cast<WritePreparedTxnDB::IteratorState*>(arg1);
+}
+}  // anonymous namespace
+
+Iterator* WritePreparedTxnDB::NewIterator(const ReadOptions& options,
+                                          ColumnFamilyHandle* column_family) {
+  constexpr bool expose_blob_index = false;
+  constexpr bool allow_refresh = false;
+  std::shared_ptr<ManagedSnapshot> own_snapshot = nullptr;
+  SequenceNumber snapshot_seq = kMaxSequenceNumber;
+  SequenceNumber min_uncommitted = 0;
+  if (options.snapshot != nullptr) {
+    snapshot_seq = options.snapshot->GetSequenceNumber();
+    min_uncommitted =
+        static_cast_with_check<const SnapshotImpl>(options.snapshot)
+            ->min_uncommitted_;
+  } else {
+    auto* snapshot = GetSnapshot();
+    // We take a snapshot to make sure that the related data in the commit map
+    // are not deleted.
+    snapshot_seq = snapshot->GetSequenceNumber();
+    min_uncommitted =
+        static_cast_with_check<const SnapshotImpl>(snapshot)->min_uncommitted_;
+    own_snapshot = std::make_shared<ManagedSnapshot>(db_impl_, snapshot);
+  }
+  assert(snapshot_seq != kMaxSequenceNumber);
+  auto* cfd =
+      static_cast_with_check<ColumnFamilyHandleImpl>(column_family)->cfd();
+  auto* state =
+      new IteratorState(this, snapshot_seq, own_snapshot, min_uncommitted);
+  auto* db_iter =
+      db_impl_->NewIteratorImpl(options, cfd, snapshot_seq, &state->callback,
+                                expose_blob_index, allow_refresh);
+  db_iter->RegisterCleanup(CleanupWritePreparedTxnDBIterator, state, nullptr);
+  return db_iter;
+}
+
+Status WritePreparedTxnDB::NewIterators(
+    const ReadOptions& options,
+    const std::vector<ColumnFamilyHandle*>& column_families,
+    std::vector<Iterator*>* iterators) {
+  constexpr bool expose_blob_index = false;
+  constexpr bool allow_refresh = false;
+  std::shared_ptr<ManagedSnapshot> own_snapshot = nullptr;
+  SequenceNumber snapshot_seq = kMaxSequenceNumber;
+  SequenceNumber min_uncommitted = 0;
+  if (options.snapshot != nullptr) {
+    snapshot_seq = options.snapshot->GetSequenceNumber();
+    min_uncommitted =
+        static_cast_with_check<const SnapshotImpl>(options.snapshot)
+            ->min_uncommitted_;
+  } else {
+    auto* snapshot = GetSnapshot();
+    // We take a snapshot to make sure that the related data in the commit map
+    // are not deleted.
+    snapshot_seq = snapshot->GetSequenceNumber();
+    own_snapshot = std::make_shared<ManagedSnapshot>(db_impl_, snapshot);
+    min_uncommitted =
+        static_cast_with_check<const SnapshotImpl>(snapshot)->min_uncommitted_;
+  }
+  iterators->clear();
+  iterators->reserve(column_families.size());
+  for (auto* column_family : column_families) {
+    auto* cfd =
+        static_cast_with_check<ColumnFamilyHandleImpl>(column_family)->cfd();
+    auto* state =
+        new IteratorState(this, snapshot_seq, own_snapshot, min_uncommitted);
+    auto* db_iter =
+        db_impl_->NewIteratorImpl(options, cfd, snapshot_seq, &state->callback,
+                                  expose_blob_index, allow_refresh);
+    db_iter->RegisterCleanup(CleanupWritePreparedTxnDBIterator, state, nullptr);
+    iterators->push_back(db_iter);
+  }
+  return Status::OK();
+}
+
+void WritePreparedTxnDB::Init(const TransactionDBOptions& txn_db_opts) {
+  // Adcance max_evicted_seq_ no more than 100 times before the cache wraps
+  // around.
+  INC_STEP_FOR_MAX_EVICTED =
+      std::max(COMMIT_CACHE_SIZE / 100, static_cast<size_t>(1));
+  snapshot_cache_ = std::unique_ptr<std::atomic<SequenceNumber>[]>(
+      new std::atomic<SequenceNumber>[SNAPSHOT_CACHE_SIZE] {});
+  commit_cache_ = std::unique_ptr<std::atomic<CommitEntry64b>[]>(
+      new std::atomic<CommitEntry64b>[COMMIT_CACHE_SIZE] {});
+  dummy_max_snapshot_.number_ = kMaxSequenceNumber;
+  rollback_deletion_type_callback_ =
+      txn_db_opts.rollback_deletion_type_callback;
+}
+
+void WritePreparedTxnDB::CheckPreparedAgainstMax(SequenceNumber new_max,
+                                                 bool locked) {
+  // When max_evicted_seq_ advances, move older entries from prepared_txns_
+  // to delayed_prepared_. This guarantees that if a seq is lower than max,
+  // then it is not in prepared_txns_ and save an expensive, synchronized
+  // lookup from a shared set. delayed_prepared_ is expected to be empty in
+  // normal cases.
+  ROCKS_LOG_DETAILS(
+      info_log_,
+      "CheckPreparedAgainstMax prepared_txns_.empty() %d top: %" PRIu64,
+      prepared_txns_.empty(),
+      prepared_txns_.empty() ? 0 : prepared_txns_.top());
+  const SequenceNumber prepared_top = prepared_txns_.top();
+  const bool empty = prepared_top == kMaxSequenceNumber;
+  // Preliminary check to avoid the synchronization cost
+  if (!empty && prepared_top <= new_max) {
+    if (locked) {
+      // Needed to avoid double locking in pop().
+      prepared_txns_.push_pop_mutex()->Unlock();
+    }
+    WriteLock wl(&prepared_mutex_);
+    // Need to fetch fresh values of ::top after mutex is acquired
+    while (!prepared_txns_.empty() && prepared_txns_.top() <= new_max) {
+      auto to_be_popped = prepared_txns_.top();
+      delayed_prepared_.insert(to_be_popped);
+      ROCKS_LOG_WARN(info_log_,
+                     "prepared_mutex_ overhead %" PRIu64 " (prep=%" PRIu64
+                     " new_max=%" PRIu64 ")",
+                     static_cast<uint64_t>(delayed_prepared_.size()),
+                     to_be_popped, new_max);
+      delayed_prepared_empty_.store(false, std::memory_order_release);
+      // Update prepared_txns_ after updating delayed_prepared_empty_ otherwise
+      // there will be a point in time that the entry is neither in
+      // prepared_txns_ nor in delayed_prepared_, which will not be checked if
+      // delayed_prepared_empty_ is false.
+      prepared_txns_.pop();
+    }
+    if (locked) {
+      prepared_txns_.push_pop_mutex()->Lock();
+    }
+  }
+}
+
+void WritePreparedTxnDB::AddPrepared(uint64_t seq, bool locked) {
+  ROCKS_LOG_DETAILS(info_log_, "Txn %" PRIu64 " Preparing with max %" PRIu64,
+                    seq, max_evicted_seq_.load());
+  TEST_SYNC_POINT("AddPrepared::begin:pause");
+  TEST_SYNC_POINT("AddPrepared::begin:resume");
+  if (!locked) {
+    prepared_txns_.push_pop_mutex()->Lock();
+  }
+  prepared_txns_.push_pop_mutex()->AssertHeld();
+  prepared_txns_.push(seq);
+  auto new_max = future_max_evicted_seq_.load();
+  if (UNLIKELY(seq <= new_max)) {
+    // This should not happen in normal case
+    ROCKS_LOG_ERROR(
+        info_log_,
+        "Added prepare_seq is not larger than max_evicted_seq_: %" PRIu64
+        " <= %" PRIu64,
+        seq, new_max);
+    CheckPreparedAgainstMax(new_max, true /*locked*/);
+  }
+  if (!locked) {
+    prepared_txns_.push_pop_mutex()->Unlock();
+  }
+  TEST_SYNC_POINT("AddPrepared::end");
+}
+
+void WritePreparedTxnDB::AddCommitted(uint64_t prepare_seq, uint64_t commit_seq,
+                                      uint8_t loop_cnt) {
+  ROCKS_LOG_DETAILS(info_log_, "Txn %" PRIu64 " Committing with %" PRIu64,
+                    prepare_seq, commit_seq);
+  TEST_SYNC_POINT("WritePreparedTxnDB::AddCommitted:start");
+  TEST_SYNC_POINT("WritePreparedTxnDB::AddCommitted:start:pause");
+  auto indexed_seq = prepare_seq % COMMIT_CACHE_SIZE;
+  CommitEntry64b evicted_64b;
+  CommitEntry evicted;
+  bool to_be_evicted = GetCommitEntry(indexed_seq, &evicted_64b, &evicted);
+  if (LIKELY(to_be_evicted)) {
+    assert(evicted.prep_seq != prepare_seq);
+    auto prev_max = max_evicted_seq_.load(std::memory_order_acquire);
+    ROCKS_LOG_DETAILS(info_log_,
+                      "Evicting %" PRIu64 ",%" PRIu64 " with max %" PRIu64,
+                      evicted.prep_seq, evicted.commit_seq, prev_max);
+    if (prev_max < evicted.commit_seq) {
+      auto last = db_impl_->GetLastPublishedSequence();  // could be 0
+      SequenceNumber max_evicted_seq;
+      if (LIKELY(evicted.commit_seq < last)) {
+        assert(last > 0);
+        // Inc max in larger steps to avoid frequent updates
+        max_evicted_seq =
+            std::min(evicted.commit_seq + INC_STEP_FOR_MAX_EVICTED, last - 1);
+      } else {
+        // legit when a commit entry in a write batch overwrite the previous one
+        max_evicted_seq = evicted.commit_seq;
+      }
+#ifdef OS_LINUX
+      if (rocksdb_write_prepared_TEST_ShouldClearCommitCache &&
+          rocksdb_write_prepared_TEST_ShouldClearCommitCache()) {
+        max_evicted_seq = last;
+      }
+#endif  // OS_LINUX
+      ROCKS_LOG_DETAILS(info_log_,
+                        "%lu Evicting %" PRIu64 ",%" PRIu64 " with max %" PRIu64
+                        " => %lu",
+                        prepare_seq, evicted.prep_seq, evicted.commit_seq,
+                        prev_max, max_evicted_seq);
+      AdvanceMaxEvictedSeq(prev_max, max_evicted_seq);
+    }
+    if (UNLIKELY(!delayed_prepared_empty_.load(std::memory_order_acquire))) {
+      WriteLock wl(&prepared_mutex_);
+      auto dp_iter = delayed_prepared_.find(evicted.prep_seq);
+      if (dp_iter != delayed_prepared_.end()) {
+        // This is a rare case that txn is committed but prepared_txns_ is not
+        // cleaned up yet. Refer to delayed_prepared_commits_ definition for
+        // why it should be kept updated.
+        delayed_prepared_commits_[evicted.prep_seq] = evicted.commit_seq;
+        ROCKS_LOG_DEBUG(info_log_,
+                        "delayed_prepared_commits_[%" PRIu64 "]=%" PRIu64,
+                        evicted.prep_seq, evicted.commit_seq);
+      }
+    }
+    // After each eviction from commit cache, check if the commit entry should
+    // be kept around because it overlaps with a live snapshot.
+    CheckAgainstSnapshots(evicted);
+  }
+  bool succ =
+      ExchangeCommitEntry(indexed_seq, evicted_64b, {prepare_seq, commit_seq});
+  if (UNLIKELY(!succ)) {
+    ROCKS_LOG_ERROR(info_log_,
+                    "ExchangeCommitEntry failed on [%" PRIu64 "] %" PRIu64
+                    ",%" PRIu64 " retrying...",
+                    indexed_seq, prepare_seq, commit_seq);
+    // A very rare event, in which the commit entry is updated before we do.
+    // Here we apply a very simple solution of retrying.
+    if (loop_cnt > 100) {
+      throw std::runtime_error("Infinite loop in AddCommitted!");
+    }
+    AddCommitted(prepare_seq, commit_seq, ++loop_cnt);
+    return;
+  }
+  TEST_SYNC_POINT("WritePreparedTxnDB::AddCommitted:end");
+  TEST_SYNC_POINT("WritePreparedTxnDB::AddCommitted:end:pause");
+}
+
+void WritePreparedTxnDB::RemovePrepared(const uint64_t prepare_seq,
+                                        const size_t batch_cnt) {
+  TEST_SYNC_POINT_CALLBACK(
+      "RemovePrepared:Start",
+      const_cast<void*>(reinterpret_cast<const void*>(&prepare_seq)));
+  TEST_SYNC_POINT("WritePreparedTxnDB::RemovePrepared:pause");
+  TEST_SYNC_POINT("WritePreparedTxnDB::RemovePrepared:resume");
+  ROCKS_LOG_DETAILS(info_log_,
+                    "RemovePrepared %" PRIu64 " cnt: %" ROCKSDB_PRIszt,
+                    prepare_seq, batch_cnt);
+  WriteLock wl(&prepared_mutex_);
+  for (size_t i = 0; i < batch_cnt; i++) {
+    prepared_txns_.erase(prepare_seq + i);
+    bool was_empty = delayed_prepared_.empty();
+    if (!was_empty) {
+      delayed_prepared_.erase(prepare_seq + i);
+      auto it = delayed_prepared_commits_.find(prepare_seq + i);
+      if (it != delayed_prepared_commits_.end()) {
+        ROCKS_LOG_DETAILS(info_log_, "delayed_prepared_commits_.erase %" PRIu64,
+                          prepare_seq + i);
+        delayed_prepared_commits_.erase(it);
+      }
+      bool is_empty = delayed_prepared_.empty();
+      if (was_empty != is_empty) {
+        delayed_prepared_empty_.store(is_empty, std::memory_order_release);
+      }
+    }
+  }
+}
+
+bool WritePreparedTxnDB::GetCommitEntry(const uint64_t indexed_seq,
+                                        CommitEntry64b* entry_64b,
+                                        CommitEntry* entry) const {
+  *entry_64b = commit_cache_[static_cast<size_t>(indexed_seq)].load(
+      std::memory_order_acquire);
+  bool valid = entry_64b->Parse(indexed_seq, entry, FORMAT);
+  return valid;
+}
+
+bool WritePreparedTxnDB::AddCommitEntry(const uint64_t indexed_seq,
+                                        const CommitEntry& new_entry,
+                                        CommitEntry* evicted_entry) {
+  CommitEntry64b new_entry_64b(new_entry, FORMAT);
+  CommitEntry64b evicted_entry_64b =
+      commit_cache_[static_cast<size_t>(indexed_seq)].exchange(
+          new_entry_64b, std::memory_order_acq_rel);
+  bool valid = evicted_entry_64b.Parse(indexed_seq, evicted_entry, FORMAT);
+  return valid;
+}
+
+bool WritePreparedTxnDB::ExchangeCommitEntry(const uint64_t indexed_seq,
+                                             CommitEntry64b& expected_entry_64b,
+                                             const CommitEntry& new_entry) {
+  auto& atomic_entry = commit_cache_[static_cast<size_t>(indexed_seq)];
+  CommitEntry64b new_entry_64b(new_entry, FORMAT);
+  bool succ = atomic_entry.compare_exchange_strong(
+      expected_entry_64b, new_entry_64b, std::memory_order_acq_rel,
+      std::memory_order_acquire);
+  return succ;
+}
+
+void WritePreparedTxnDB::AdvanceMaxEvictedSeq(const SequenceNumber& prev_max,
+                                              const SequenceNumber& new_max) {
+  ROCKS_LOG_DETAILS(info_log_,
+                    "AdvanceMaxEvictedSeq overhead %" PRIu64 " => %" PRIu64,
+                    prev_max, new_max);
+  // Declare the intention before getting snapshot from the DB. This helps a
+  // concurrent GetSnapshot to wait to catch up with future_max_evicted_seq_ if
+  // it has not already. Otherwise the new snapshot is when we ask DB for
+  // snapshots smaller than future max.
+  auto updated_future_max = prev_max;
+  while (updated_future_max < new_max &&
+         !future_max_evicted_seq_.compare_exchange_weak(
+             updated_future_max, new_max, std::memory_order_acq_rel,
+             std::memory_order_relaxed)) {
+  };
+
+  CheckPreparedAgainstMax(new_max, false /*locked*/);
+
+  // With each change to max_evicted_seq_ fetch the live snapshots behind it.
+  // We use max as the version of snapshots to identify how fresh are the
+  // snapshot list. This works because the snapshots are between 0 and
+  // max, so the larger the max, the more complete they are.
+  SequenceNumber new_snapshots_version = new_max;
+  std::vector<SequenceNumber> snapshots;
+  bool update_snapshots = false;
+  if (new_snapshots_version > snapshots_version_) {
+    // This is to avoid updating the snapshots_ if it already updated
+    // with a more recent vesion by a concrrent thread
+    update_snapshots = true;
+    // We only care about snapshots lower then max
+    snapshots = GetSnapshotListFromDB(new_max);
+  }
+  if (update_snapshots) {
+    UpdateSnapshots(snapshots, new_snapshots_version);
+    if (!snapshots.empty()) {
+      WriteLock wl(&old_commit_map_mutex_);
+      for (auto snap : snapshots) {
+        // This allows IsInSnapshot to tell apart the reads from in valid
+        // snapshots from the reads from committed values in valid snapshots.
+        old_commit_map_[snap];
+      }
+      old_commit_map_empty_.store(false, std::memory_order_release);
+    }
+  }
+  auto updated_prev_max = prev_max;
+  TEST_SYNC_POINT("AdvanceMaxEvictedSeq::update_max:pause");
+  TEST_SYNC_POINT("AdvanceMaxEvictedSeq::update_max:resume");
+  while (updated_prev_max < new_max &&
+         !max_evicted_seq_.compare_exchange_weak(updated_prev_max, new_max,
+                                                 std::memory_order_acq_rel,
+                                                 std::memory_order_relaxed)) {
+  };
+}
+
+const Snapshot* WritePreparedTxnDB::GetSnapshot() {
+  const bool kForWWConflictCheck = true;
+  return GetSnapshotInternal(!kForWWConflictCheck);
+}
+
+SnapshotImpl* WritePreparedTxnDB::GetSnapshotInternal(
+    bool for_ww_conflict_check) {
+  // Note: for this optimization setting the last sequence number and obtaining
+  // the smallest uncommitted seq should be done atomically. However to avoid
+  // the mutex overhead, we call SmallestUnCommittedSeq BEFORE taking the
+  // snapshot. Since we always updated the list of unprepared seq (via
+  // AddPrepared) AFTER the last sequence is updated, this guarantees that the
+  // smallest uncommitted seq that we pair with the snapshot is smaller or equal
+  // the value that would be obtained otherwise atomically. That is ok since
+  // this optimization works as long as min_uncommitted is less than or equal
+  // than the smallest uncommitted seq when the snapshot was taken.
+  auto min_uncommitted = WritePreparedTxnDB::SmallestUnCommittedSeq();
+  SnapshotImpl* snap_impl = db_impl_->GetSnapshotImpl(for_ww_conflict_check);
+  TEST_SYNC_POINT("WritePreparedTxnDB::GetSnapshotInternal:first");
+  assert(snap_impl);
+  SequenceNumber snap_seq = snap_impl->GetSequenceNumber();
+  // Note: Check against future_max_evicted_seq_ (in contrast with
+  // max_evicted_seq_) in case there is a concurrent AdvanceMaxEvictedSeq.
+  if (UNLIKELY(snap_seq != 0 && snap_seq <= future_max_evicted_seq_)) {
+    // There is a very rare case in which the commit entry evicts another commit
+    // entry that is not published yet thus advancing max evicted seq beyond the
+    // last published seq. This case is not likely in real-world setup so we
+    // handle it with a few retries.
+    size_t retry = 0;
+    SequenceNumber max;
+    while ((max = future_max_evicted_seq_.load()) != 0 &&
+           snap_impl->GetSequenceNumber() <= max && retry < 100) {
+      ROCKS_LOG_WARN(info_log_,
+                     "GetSnapshot snap: %" PRIu64 " max: %" PRIu64
+                     " retry %" ROCKSDB_PRIszt,
+                     snap_impl->GetSequenceNumber(), max, retry);
+      ReleaseSnapshot(snap_impl);
+      // Wait for last visible seq to catch up with max, and also go beyond it
+      // by one.
+      AdvanceSeqByOne();
+      snap_impl = db_impl_->GetSnapshotImpl(for_ww_conflict_check);
+      assert(snap_impl);
+      retry++;
+    }
+    assert(snap_impl->GetSequenceNumber() > max);
+    if (snap_impl->GetSequenceNumber() <= max) {
+      throw std::runtime_error(
+          "Snapshot seq " + std::to_string(snap_impl->GetSequenceNumber()) +
+          " after " + std::to_string(retry) +
+          " retries is still less than futre_max_evicted_seq_" +
+          std::to_string(max));
+    }
+  }
+  EnhanceSnapshot(snap_impl, min_uncommitted);
+  ROCKS_LOG_DETAILS(
+      db_impl_->immutable_db_options().info_log,
+      "GetSnapshot %" PRIu64 " ww:%" PRIi32 " min_uncommitted: %" PRIu64,
+      snap_impl->GetSequenceNumber(), for_ww_conflict_check, min_uncommitted);
+  TEST_SYNC_POINT("WritePreparedTxnDB::GetSnapshotInternal:end");
+  return snap_impl;
+}
+
+void WritePreparedTxnDB::AdvanceSeqByOne() {
+  // Inserting an empty value will i) let the max evicted entry to be
+  // published, i.e., max == last_published, increase the last published to
+  // be one beyond max, i.e., max < last_published.
+  WriteOptions woptions;
+  TransactionOptions txn_options;
+  Transaction* txn0 = BeginTransaction(woptions, txn_options, nullptr);
+  std::hash<std::thread::id> hasher;
+  char name[64];
+  snprintf(name, 64, "txn%" ROCKSDB_PRIszt, hasher(std::this_thread::get_id()));
+  assert(strlen(name) < 64 - 1);
+  Status s = txn0->SetName(name);
+  assert(s.ok());
+  if (s.ok()) {
+    // Without prepare it would simply skip the commit
+    s = txn0->Prepare();
+  }
+  assert(s.ok());
+  if (s.ok()) {
+    s = txn0->Commit();
+  }
+  assert(s.ok());
+  delete txn0;
+}
+
+const std::vector<SequenceNumber> WritePreparedTxnDB::GetSnapshotListFromDB(
+    SequenceNumber max) {
+  ROCKS_LOG_DETAILS(info_log_, "GetSnapshotListFromDB with max %" PRIu64, max);
+  InstrumentedMutexLock dblock(db_impl_->mutex());
+  db_impl_->mutex()->AssertHeld();
+  return db_impl_->snapshots().GetAll(nullptr, max);
+}
+
+void WritePreparedTxnDB::ReleaseSnapshotInternal(
+    const SequenceNumber snap_seq) {
+  // TODO(myabandeh): relax should enough since the synchronizatin is already
+  // done by snapshots_mutex_ under which this function is called.
+  if (snap_seq <= max_evicted_seq_.load(std::memory_order_acquire)) {
+    // Then this is a rare case that transaction did not finish before max
+    // advances. It is expected for a few read-only backup snapshots. For such
+    // snapshots we might have kept around a couple of entries in the
+    // old_commit_map_. Check and do garbage collection if that is the case.
+    bool need_gc = false;
+    {
+      WPRecordTick(TXN_OLD_COMMIT_MAP_MUTEX_OVERHEAD);
+      ROCKS_LOG_WARN(info_log_, "old_commit_map_mutex_ overhead for %" PRIu64,
+                     snap_seq);
+      ReadLock rl(&old_commit_map_mutex_);
+      auto prep_set_entry = old_commit_map_.find(snap_seq);
+      need_gc = prep_set_entry != old_commit_map_.end();
+    }
+    if (need_gc) {
+      WPRecordTick(TXN_OLD_COMMIT_MAP_MUTEX_OVERHEAD);
+      ROCKS_LOG_WARN(info_log_, "old_commit_map_mutex_ overhead for %" PRIu64,
+                     snap_seq);
+      WriteLock wl(&old_commit_map_mutex_);
+      old_commit_map_.erase(snap_seq);
+      old_commit_map_empty_.store(old_commit_map_.empty(),
+                                  std::memory_order_release);
+    }
+  }
+}
+
+void WritePreparedTxnDB::CleanupReleasedSnapshots(
+    const std::vector<SequenceNumber>& new_snapshots,
+    const std::vector<SequenceNumber>& old_snapshots) {
+  auto newi = new_snapshots.begin();
+  auto oldi = old_snapshots.begin();
+  for (; newi != new_snapshots.end() && oldi != old_snapshots.end();) {
+    assert(*newi >= *oldi);  // cannot have new snapshots with lower seq
+    if (*newi == *oldi) {    // still not released
+      auto value = *newi;
+      while (newi != new_snapshots.end() && *newi == value) {
+        newi++;
+      }
+      while (oldi != old_snapshots.end() && *oldi == value) {
+        oldi++;
+      }
+    } else {
+      assert(*newi > *oldi);  // *oldi is released
+      ReleaseSnapshotInternal(*oldi);
+      oldi++;
+    }
+  }
+  // Everything remained in old_snapshots is released and must be cleaned up
+  for (; oldi != old_snapshots.end(); oldi++) {
+    ReleaseSnapshotInternal(*oldi);
+  }
+}
+
+void WritePreparedTxnDB::UpdateSnapshots(
+    const std::vector<SequenceNumber>& snapshots,
+    const SequenceNumber& version) {
+  ROCKS_LOG_DETAILS(info_log_, "UpdateSnapshots with version %" PRIu64,
+                    version);
+  TEST_SYNC_POINT("WritePreparedTxnDB::UpdateSnapshots:p:start");
+  TEST_SYNC_POINT("WritePreparedTxnDB::UpdateSnapshots:s:start");
+#ifndef NDEBUG
+  size_t sync_i = 0;
+#endif
+  ROCKS_LOG_DETAILS(info_log_, "snapshots_mutex_ overhead");
+  WriteLock wl(&snapshots_mutex_);
+  snapshots_version_ = version;
+  // We update the list concurrently with the readers.
+  // Both new and old lists are sorted and the new list is subset of the
+  // previous list plus some new items. Thus if a snapshot repeats in
+  // both new and old lists, it will appear upper in the new list. So if
+  // we simply insert the new snapshots in order, if an overwritten item
+  // is still valid in the new list is either written to the same place in
+  // the array or it is written in a higher palce before it gets
+  // overwritten by another item. This guarantess a reader that reads the
+  // list bottom-up will eventaully see a snapshot that repeats in the
+  // update, either before it gets overwritten by the writer or
+  // afterwards.
+  size_t i = 0;
+  auto it = snapshots.begin();
+  for (; it != snapshots.end() && i < SNAPSHOT_CACHE_SIZE; ++it, ++i) {
+    snapshot_cache_[i].store(*it, std::memory_order_release);
+    TEST_IDX_SYNC_POINT("WritePreparedTxnDB::UpdateSnapshots:p:", ++sync_i);
+    TEST_IDX_SYNC_POINT("WritePreparedTxnDB::UpdateSnapshots:s:", sync_i);
+  }
+#ifndef NDEBUG
+  // Release the remaining sync points since they are useless given that the
+  // reader would also use lock to access snapshots
+  for (++sync_i; sync_i <= 10; ++sync_i) {
+    TEST_IDX_SYNC_POINT("WritePreparedTxnDB::UpdateSnapshots:p:", sync_i);
+    TEST_IDX_SYNC_POINT("WritePreparedTxnDB::UpdateSnapshots:s:", sync_i);
+  }
+#endif
+  snapshots_.clear();
+  for (; it != snapshots.end(); ++it) {
+    // Insert them to a vector that is less efficient to access
+    // concurrently
+    snapshots_.push_back(*it);
+  }
+  // Update the size at the end. Otherwise a parallel reader might read
+  // items that are not set yet.
+  snapshots_total_.store(snapshots.size(), std::memory_order_release);
+
+  // Note: this must be done after the snapshots data structures are updated
+  // with the new list of snapshots.
+  CleanupReleasedSnapshots(snapshots, snapshots_all_);
+  snapshots_all_ = snapshots;
+
+  TEST_SYNC_POINT("WritePreparedTxnDB::UpdateSnapshots:p:end");
+  TEST_SYNC_POINT("WritePreparedTxnDB::UpdateSnapshots:s:end");
+}
+
+void WritePreparedTxnDB::CheckAgainstSnapshots(const CommitEntry& evicted) {
+  TEST_SYNC_POINT("WritePreparedTxnDB::CheckAgainstSnapshots:p:start");
+  TEST_SYNC_POINT("WritePreparedTxnDB::CheckAgainstSnapshots:s:start");
+#ifndef NDEBUG
+  size_t sync_i = 0;
+#endif
+  // First check the snapshot cache that is efficient for concurrent access
+  auto cnt = snapshots_total_.load(std::memory_order_acquire);
+  // The list might get updated concurrently as we are reading from it. The
+  // reader should be able to read all the snapshots that are still valid
+  // after the update. Since the survived snapshots are written in a higher
+  // place before gets overwritten the reader that reads bottom-up will
+  // eventully see it.
+  const bool next_is_larger = true;
+  // We will set to true if the border line snapshot suggests that.
+  bool search_larger_list = false;
+  size_t ip1 = std::min(cnt, SNAPSHOT_CACHE_SIZE);
+  for (; 0 < ip1; ip1--) {
+    SequenceNumber snapshot_seq =
+        snapshot_cache_[ip1 - 1].load(std::memory_order_acquire);
+    TEST_IDX_SYNC_POINT("WritePreparedTxnDB::CheckAgainstSnapshots:p:",
+                        ++sync_i);
+    TEST_IDX_SYNC_POINT("WritePreparedTxnDB::CheckAgainstSnapshots:s:", sync_i);
+    if (ip1 == SNAPSHOT_CACHE_SIZE) {  // border line snapshot
+      // snapshot_seq < commit_seq => larger_snapshot_seq <= commit_seq
+      // then later also continue the search to larger snapshots
+      search_larger_list = snapshot_seq < evicted.commit_seq;
+    }
+    if (!MaybeUpdateOldCommitMap(evicted.prep_seq, evicted.commit_seq,
+                                 snapshot_seq, !next_is_larger)) {
+      break;
+    }
+  }
+#ifndef NDEBUG
+  // Release the remaining sync points before accquiring the lock
+  for (++sync_i; sync_i <= 10; ++sync_i) {
+    TEST_IDX_SYNC_POINT("WritePreparedTxnDB::CheckAgainstSnapshots:p:", sync_i);
+    TEST_IDX_SYNC_POINT("WritePreparedTxnDB::CheckAgainstSnapshots:s:", sync_i);
+  }
+#endif
+  TEST_SYNC_POINT("WritePreparedTxnDB::CheckAgainstSnapshots:p:end");
+  TEST_SYNC_POINT("WritePreparedTxnDB::CheckAgainstSnapshots:s:end");
+  if (UNLIKELY(SNAPSHOT_CACHE_SIZE < cnt && search_larger_list)) {
+    // Then access the less efficient list of snapshots_
+    WPRecordTick(TXN_SNAPSHOT_MUTEX_OVERHEAD);
+    ROCKS_LOG_WARN(info_log_,
+                   "snapshots_mutex_ overhead for <%" PRIu64 ",%" PRIu64
+                   "> with %" ROCKSDB_PRIszt " snapshots",
+                   evicted.prep_seq, evicted.commit_seq, cnt);
+    ReadLock rl(&snapshots_mutex_);
+    // Items could have moved from the snapshots_ to snapshot_cache_ before
+    // accquiring the lock. To make sure that we do not miss a valid snapshot,
+    // read snapshot_cache_ again while holding the lock.
+    for (size_t i = 0; i < SNAPSHOT_CACHE_SIZE; i++) {
+      SequenceNumber snapshot_seq =
+          snapshot_cache_[i].load(std::memory_order_acquire);
+      if (!MaybeUpdateOldCommitMap(evicted.prep_seq, evicted.commit_seq,
+                                   snapshot_seq, next_is_larger)) {
+        break;
+      }
+    }
+    for (auto snapshot_seq_2 : snapshots_) {
+      if (!MaybeUpdateOldCommitMap(evicted.prep_seq, evicted.commit_seq,
+                                   snapshot_seq_2, next_is_larger)) {
+        break;
+      }
+    }
+  }
+}
+
+bool WritePreparedTxnDB::MaybeUpdateOldCommitMap(
+    const uint64_t& prep_seq, const uint64_t& commit_seq,
+    const uint64_t& snapshot_seq, const bool next_is_larger = true) {
+  // If we do not store an entry in old_commit_map_ we assume it is committed in
+  // all snapshots. If commit_seq <= snapshot_seq, it is considered already in
+  // the snapshot so we need not to keep the entry around for this snapshot.
+  if (commit_seq <= snapshot_seq) {
+    // continue the search if the next snapshot could be smaller than commit_seq
+    return !next_is_larger;
+  }
+  // then snapshot_seq < commit_seq
+  if (prep_seq <= snapshot_seq) {  // overlapping range
+    WPRecordTick(TXN_OLD_COMMIT_MAP_MUTEX_OVERHEAD);
+    ROCKS_LOG_WARN(info_log_,
+                   "old_commit_map_mutex_ overhead for %" PRIu64
+                   " commit entry: <%" PRIu64 ",%" PRIu64 ">",
+                   snapshot_seq, prep_seq, commit_seq);
+    WriteLock wl(&old_commit_map_mutex_);
+    old_commit_map_empty_.store(false, std::memory_order_release);
+    auto& vec = old_commit_map_[snapshot_seq];
+    vec.insert(std::upper_bound(vec.begin(), vec.end(), prep_seq), prep_seq);
+    // We need to store it once for each overlapping snapshot. Returning true to
+    // continue the search if there is more overlapping snapshot.
+    return true;
+  }
+  // continue the search if the next snapshot could be larger than prep_seq
+  return next_is_larger;
+}
+
+WritePreparedTxnDB::~WritePreparedTxnDB() {
+  // At this point there could be running compaction/flush holding a
+  // SnapshotChecker, which holds a pointer back to WritePreparedTxnDB.
+  // Make sure those jobs finished before destructing WritePreparedTxnDB.
+  if (!db_impl_->shutting_down_) {
+    db_impl_->CancelAllBackgroundWork(true /*wait*/);
+  }
+}
+
+void SubBatchCounter::InitWithComp(const uint32_t cf) {
+  auto cmp = comparators_[cf];
+  keys_[cf] = CFKeys(SetComparator(cmp));
+}
+
+void SubBatchCounter::AddKey(const uint32_t cf, const Slice& key) {
+  CFKeys& cf_keys = keys_[cf];
+  if (cf_keys.size() == 0) {  // just inserted
+    InitWithComp(cf);
+  }
+  auto it = cf_keys.insert(key);
+  if (it.second == false) {  // second is false if a element already existed.
+    batches_++;
+    keys_.clear();
+    InitWithComp(cf);
+    keys_[cf].insert(key);
+  }
+}
+
+}  // namespace ROCKSDB_NAMESPACE
+#endif  // ROCKSDB_LITE
diff --git a/src/rocksdb/utilities/transactions/write_prepared_txn_db.h b/src/rocksdb/utilities/transactions/write_prepared_txn_db.h
new file mode 100644
index 000000000..25a382473
--- /dev/null
+++ b/src/rocksdb/utilities/transactions/write_prepared_txn_db.h
@@ -0,0 +1,1125 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#pragma once
+#ifndef ROCKSDB_LITE
+
+#include <cinttypes>
+#include <mutex>
+#include <queue>
+#include <set>
+#include <string>
+#include <unordered_map>
+#include <vector>
+
+#include "db/db_iter.h"
+#include "db/pre_release_callback.h"
+#include "db/read_callback.h"
+#include "db/snapshot_checker.h"
+#include "logging/logging.h"
+#include "rocksdb/db.h"
+#include "rocksdb/options.h"
+#include "rocksdb/utilities/transaction_db.h"
+#include "util/cast_util.h"
+#include "util/set_comparator.h"
+#include "util/string_util.h"
+#include "utilities/transactions/pessimistic_transaction.h"
+#include "utilities/transactions/pessimistic_transaction_db.h"
+#include "utilities/transactions/write_prepared_txn.h"
+
+namespace ROCKSDB_NAMESPACE {
+enum SnapshotBackup : bool { kUnbackedByDBSnapshot, kBackedByDBSnapshot };
+
+// A PessimisticTransactionDB that writes data to DB after prepare phase of 2PC.
+// In this way some data in the DB might not be committed. The DB provides
+// mechanisms to tell such data apart from committed data.
+class WritePreparedTxnDB : public PessimisticTransactionDB {
+ public:
+  explicit WritePreparedTxnDB(DB* db,
+                              const TransactionDBOptions& txn_db_options)
+      : PessimisticTransactionDB(db, txn_db_options),
+        SNAPSHOT_CACHE_BITS(txn_db_options.wp_snapshot_cache_bits),
+        SNAPSHOT_CACHE_SIZE(static_cast<size_t>(1ull << SNAPSHOT_CACHE_BITS)),
+        COMMIT_CACHE_BITS(txn_db_options.wp_commit_cache_bits),
+        COMMIT_CACHE_SIZE(static_cast<size_t>(1ull << COMMIT_CACHE_BITS)),
+        FORMAT(COMMIT_CACHE_BITS) {
+    Init(txn_db_options);
+  }
+
+  explicit WritePreparedTxnDB(StackableDB* db,
+                              const TransactionDBOptions& txn_db_options)
+      : PessimisticTransactionDB(db, txn_db_options),
+        SNAPSHOT_CACHE_BITS(txn_db_options.wp_snapshot_cache_bits),
+        SNAPSHOT_CACHE_SIZE(static_cast<size_t>(1ull << SNAPSHOT_CACHE_BITS)),
+        COMMIT_CACHE_BITS(txn_db_options.wp_commit_cache_bits),
+        COMMIT_CACHE_SIZE(static_cast<size_t>(1ull << COMMIT_CACHE_BITS)),
+        FORMAT(COMMIT_CACHE_BITS) {
+    Init(txn_db_options);
+  }
+
+  virtual ~WritePreparedTxnDB();
+
+  virtual Status Initialize(
+      const std::vector<size_t>& compaction_enabled_cf_indices,
+      const std::vector<ColumnFamilyHandle*>& handles) override;
+
+  Transaction* BeginTransaction(const WriteOptions& write_options,
+                                const TransactionOptions& txn_options,
+                                Transaction* old_txn) override;
+
+  using TransactionDB::Write;
+  Status Write(const WriteOptions& opts, WriteBatch* updates) override;
+
+  // Optimized version of ::Write that receives more optimization request such
+  // as skip_concurrency_control.
+  using PessimisticTransactionDB::Write;
+  Status Write(const WriteOptions& opts, const TransactionDBWriteOptimizations&,
+               WriteBatch* updates) override;
+
+  // Write the batch to the underlying DB and mark it as committed. Could be
+  // used by both directly from TxnDB or through a transaction.
+  Status WriteInternal(const WriteOptions& write_options, WriteBatch* batch,
+                       size_t batch_cnt, WritePreparedTxn* txn);
+
+  using DB::Get;
+  virtual Status Get(const ReadOptions& options,
+                     ColumnFamilyHandle* column_family, const Slice& key,
+                     PinnableSlice* value) override;
+
+  using DB::MultiGet;
+  virtual std::vector<Status> MultiGet(
+      const ReadOptions& options,
+      const std::vector<ColumnFamilyHandle*>& column_family,
+      const std::vector<Slice>& keys,
+      std::vector<std::string>* values) override;
+
+  using DB::NewIterator;
+  virtual Iterator* NewIterator(const ReadOptions& options,
+                                ColumnFamilyHandle* column_family) override;
+
+  using DB::NewIterators;
+  virtual Status NewIterators(
+      const ReadOptions& options,
+      const std::vector<ColumnFamilyHandle*>& column_families,
+      std::vector<Iterator*>* iterators) override;
+
+  // Check whether the transaction that wrote the value with sequence number seq
+  // is visible to the snapshot with sequence number snapshot_seq.
+  // Returns true if commit_seq <= snapshot_seq
+  // If the snapshot_seq is already released and snapshot_seq <= max, sets
+  // *snap_released to true and returns true as well.
+  inline bool IsInSnapshot(uint64_t prep_seq, uint64_t snapshot_seq,
+                           uint64_t min_uncommitted = kMinUnCommittedSeq,
+                           bool* snap_released = nullptr) const {
+    ROCKS_LOG_DETAILS(info_log_,
+                      "IsInSnapshot %" PRIu64 " in %" PRIu64
+                      " min_uncommitted %" PRIu64,
+                      prep_seq, snapshot_seq, min_uncommitted);
+    assert(min_uncommitted >= kMinUnCommittedSeq);
+    // Caller is responsible to initialize snap_released.
+    assert(snap_released == nullptr || *snap_released == false);
+    // Here we try to infer the return value without looking into prepare list.
+    // This would help avoiding synchronization over a shared map.
+    // TODO(myabandeh): optimize this. This sequence of checks must be correct
+    // but not necessary efficient
+    if (prep_seq == 0) {
+      // Compaction will output keys to bottom-level with sequence number 0 if
+      // it is visible to the earliest snapshot.
+      ROCKS_LOG_DETAILS(
+          info_log_, "IsInSnapshot %" PRIu64 " in %" PRIu64 " returns %" PRId32,
+          prep_seq, snapshot_seq, 1);
+      return true;
+    }
+    if (snapshot_seq < prep_seq) {
+      // snapshot_seq < prep_seq <= commit_seq => snapshot_seq < commit_seq
+      ROCKS_LOG_DETAILS(
+          info_log_, "IsInSnapshot %" PRIu64 " in %" PRIu64 " returns %" PRId32,
+          prep_seq, snapshot_seq, 0);
+      return false;
+    }
+    if (prep_seq < min_uncommitted) {
+      ROCKS_LOG_DETAILS(info_log_,
+                        "IsInSnapshot %" PRIu64 " in %" PRIu64
+                        " returns %" PRId32
+                        " because of min_uncommitted %" PRIu64,
+                        prep_seq, snapshot_seq, 1, min_uncommitted);
+      return true;
+    }
+    // Commit of delayed prepared has two non-atomic steps: add to commit cache,
+    // remove from delayed prepared. Our reads from these two is also
+    // non-atomic. By looking into commit cache first thus we might not find the
+    // prep_seq neither in commit cache not in delayed_prepared_. To fix that i)
+    // we check if there was any delayed prepared BEFORE looking into commit
+    // cache, ii) if there was, we complete the search steps to be these: i)
+    // commit cache, ii) delayed prepared, commit cache again. In this way if
+    // the first query to commit cache missed the commit, the 2nd will catch it.
+    bool was_empty;
+    SequenceNumber max_evicted_seq_lb, max_evicted_seq_ub;
+    CommitEntry64b dont_care;
+    auto indexed_seq = prep_seq % COMMIT_CACHE_SIZE;
+    size_t repeats = 0;
+    do {
+      repeats++;
+      assert(repeats < 100);
+      if (UNLIKELY(repeats >= 100)) {
+        throw std::runtime_error(
+            "The read was intrupted 100 times by update to max_evicted_seq_. "
+            "This is unexpected in all setups");
+      }
+      max_evicted_seq_lb = max_evicted_seq_.load(std::memory_order_acquire);
+      TEST_SYNC_POINT(
+          "WritePreparedTxnDB::IsInSnapshot:max_evicted_seq_:pause");
+      TEST_SYNC_POINT(
+          "WritePreparedTxnDB::IsInSnapshot:max_evicted_seq_:resume");
+      was_empty = delayed_prepared_empty_.load(std::memory_order_acquire);
+      TEST_SYNC_POINT(
+          "WritePreparedTxnDB::IsInSnapshot:delayed_prepared_empty_:pause");
+      TEST_SYNC_POINT(
+          "WritePreparedTxnDB::IsInSnapshot:delayed_prepared_empty_:resume");
+      CommitEntry cached;
+      bool exist = GetCommitEntry(indexed_seq, &dont_care, &cached);
+      TEST_SYNC_POINT("WritePreparedTxnDB::IsInSnapshot:GetCommitEntry:pause");
+      TEST_SYNC_POINT("WritePreparedTxnDB::IsInSnapshot:GetCommitEntry:resume");
+      if (exist && prep_seq == cached.prep_seq) {
+        // It is committed and also not evicted from commit cache
+        ROCKS_LOG_DETAILS(
+            info_log_,
+            "IsInSnapshot %" PRIu64 " in %" PRIu64 " returns %" PRId32,
+            prep_seq, snapshot_seq, cached.commit_seq <= snapshot_seq);
+        return cached.commit_seq <= snapshot_seq;
+      }
+      // else it could be committed but not inserted in the map which could
+      // happen after recovery, or it could be committed and evicted by another
+      // commit, or never committed.
+
+      // At this point we don't know if it was committed or it is still prepared
+      max_evicted_seq_ub = max_evicted_seq_.load(std::memory_order_acquire);
+      if (UNLIKELY(max_evicted_seq_lb != max_evicted_seq_ub)) {
+        continue;
+      }
+      // Note: max_evicted_seq_ when we did GetCommitEntry <= max_evicted_seq_ub
+      if (max_evicted_seq_ub < prep_seq) {
+        // Not evicted from cache and also not present, so must be still
+        // prepared
+        ROCKS_LOG_DETAILS(info_log_,
+                          "IsInSnapshot %" PRIu64 " in %" PRIu64
+                          " returns %" PRId32,
+                          prep_seq, snapshot_seq, 0);
+        return false;
+      }
+      TEST_SYNC_POINT("WritePreparedTxnDB::IsInSnapshot:prepared_mutex_:pause");
+      TEST_SYNC_POINT(
+          "WritePreparedTxnDB::IsInSnapshot:prepared_mutex_:resume");
+      if (!was_empty) {
+        // We should not normally reach here
+        WPRecordTick(TXN_PREPARE_MUTEX_OVERHEAD);
+        ReadLock rl(&prepared_mutex_);
+        ROCKS_LOG_WARN(
+            info_log_, "prepared_mutex_ overhead %" PRIu64 " for %" PRIu64,
+            static_cast<uint64_t>(delayed_prepared_.size()), prep_seq);
+        if (delayed_prepared_.find(prep_seq) != delayed_prepared_.end()) {
+          // This is the order: 1) delayed_prepared_commits_ update, 2) publish
+          // 3) delayed_prepared_ clean up. So check if it is the case of a late
+          // clenaup.
+          auto it = delayed_prepared_commits_.find(prep_seq);
+          if (it == delayed_prepared_commits_.end()) {
+            // Then it is not committed yet
+            ROCKS_LOG_DETAILS(info_log_,
+                              "IsInSnapshot %" PRIu64 " in %" PRIu64
+                              " returns %" PRId32,
+                              prep_seq, snapshot_seq, 0);
+            return false;
+          } else {
+            ROCKS_LOG_DETAILS(info_log_,
+                              "IsInSnapshot %" PRIu64 " in %" PRIu64
+                              " commit: %" PRIu64 " returns %" PRId32,
+                              prep_seq, snapshot_seq, it->second,
+                              snapshot_seq <= it->second);
+            return it->second <= snapshot_seq;
+          }
+        } else {
+          // 2nd query to commit cache. Refer to was_empty comment above.
+          exist = GetCommitEntry(indexed_seq, &dont_care, &cached);
+          if (exist && prep_seq == cached.prep_seq) {
+            ROCKS_LOG_DETAILS(
+                info_log_,
+                "IsInSnapshot %" PRIu64 " in %" PRIu64 " returns %" PRId32,
+                prep_seq, snapshot_seq, cached.commit_seq <= snapshot_seq);
+            return cached.commit_seq <= snapshot_seq;
+          }
+          max_evicted_seq_ub = max_evicted_seq_.load(std::memory_order_acquire);
+        }
+      }
+    } while (UNLIKELY(max_evicted_seq_lb != max_evicted_seq_ub));
+    // When advancing max_evicted_seq_, we move older entires from prepared to
+    // delayed_prepared_. Also we move evicted entries from commit cache to
+    // old_commit_map_ if it overlaps with any snapshot. Since prep_seq <=
+    // max_evicted_seq_, we have three cases: i) in delayed_prepared_, ii) in
+    // old_commit_map_, iii) committed with no conflict with any snapshot. Case
+    // (i) delayed_prepared_ is checked above
+    if (max_evicted_seq_ub < snapshot_seq) {  // then (ii) cannot be the case
+      // only (iii) is the case: committed
+      // commit_seq <= max_evicted_seq_ < snapshot_seq => commit_seq <
+      // snapshot_seq
+      ROCKS_LOG_DETAILS(
+          info_log_, "IsInSnapshot %" PRIu64 " in %" PRIu64 " returns %" PRId32,
+          prep_seq, snapshot_seq, 1);
+      return true;
+    }
+    // else (ii) might be the case: check the commit data saved for this
+    // snapshot. If there was no overlapping commit entry, then it is committed
+    // with a commit_seq lower than any live snapshot, including snapshot_seq.
+    if (old_commit_map_empty_.load(std::memory_order_acquire)) {
+      ROCKS_LOG_DETAILS(info_log_,
+                        "IsInSnapshot %" PRIu64 " in %" PRIu64
+                        " returns %" PRId32 " released=1",
+                        prep_seq, snapshot_seq, 0);
+      assert(snap_released);
+      // This snapshot is not valid anymore. We cannot tell if prep_seq is
+      // committed before or after the snapshot. Return true but also set
+      // snap_released to true.
+      *snap_released = true;
+      return true;
+    }
+    {
+      // We should not normally reach here unless sapshot_seq is old. This is a
+      // rare case and it is ok to pay the cost of mutex ReadLock for such old,
+      // reading transactions.
+      WPRecordTick(TXN_OLD_COMMIT_MAP_MUTEX_OVERHEAD);
+      ReadLock rl(&old_commit_map_mutex_);
+      auto prep_set_entry = old_commit_map_.find(snapshot_seq);
+      bool found = prep_set_entry != old_commit_map_.end();
+      if (found) {
+        auto& vec = prep_set_entry->second;
+        found = std::binary_search(vec.begin(), vec.end(), prep_seq);
+      } else {
+        // coming from compaction
+        ROCKS_LOG_DETAILS(info_log_,
+                          "IsInSnapshot %" PRIu64 " in %" PRIu64
+                          " returns %" PRId32 " released=1",
+                          prep_seq, snapshot_seq, 0);
+        // This snapshot is not valid anymore. We cannot tell if prep_seq is
+        // committed before or after the snapshot. Return true but also set
+        // snap_released to true.
+        assert(snap_released);
+        *snap_released = true;
+        return true;
+      }
+
+      if (!found) {
+        ROCKS_LOG_DETAILS(info_log_,
+                          "IsInSnapshot %" PRIu64 " in %" PRIu64
+                          " returns %" PRId32,
+                          prep_seq, snapshot_seq, 1);
+        return true;
+      }
+    }
+    // (ii) it the case: it is committed but after the snapshot_seq
+    ROCKS_LOG_DETAILS(
+        info_log_, "IsInSnapshot %" PRIu64 " in %" PRIu64 " returns %" PRId32,
+        prep_seq, snapshot_seq, 0);
+    return false;
+  }
+
+  // Add the transaction with prepare sequence seq to the prepared list.
+  // Note: must be called serially with increasing seq on each call.
+  // locked is true if prepared_mutex_ is already locked.
+  void AddPrepared(uint64_t seq, bool locked = false);
+  // Check if any of the prepared txns are less than new max_evicted_seq_. Must
+  // be called with prepared_mutex_ write locked.
+  void CheckPreparedAgainstMax(SequenceNumber new_max, bool locked);
+  // Remove the transaction with prepare sequence seq from the prepared list
+  void RemovePrepared(const uint64_t seq, const size_t batch_cnt = 1);
+  // Add the transaction with prepare sequence prepare_seq and commit sequence
+  // commit_seq to the commit map. loop_cnt is to detect infinite loops.
+  // Note: must be called serially.
+  void AddCommitted(uint64_t prepare_seq, uint64_t commit_seq,
+                    uint8_t loop_cnt = 0);
+
+  struct CommitEntry {
+    uint64_t prep_seq;
+    uint64_t commit_seq;
+    CommitEntry() : prep_seq(0), commit_seq(0) {}
+    CommitEntry(uint64_t ps, uint64_t cs) : prep_seq(ps), commit_seq(cs) {}
+    bool operator==(const CommitEntry& rhs) const {
+      return prep_seq == rhs.prep_seq && commit_seq == rhs.commit_seq;
+    }
+  };
+
+  struct CommitEntry64bFormat {
+    explicit CommitEntry64bFormat(size_t index_bits)
+        : INDEX_BITS(index_bits),
+          PREP_BITS(static_cast<size_t>(64 - PAD_BITS - INDEX_BITS)),
+          COMMIT_BITS(static_cast<size_t>(64 - PREP_BITS)),
+          COMMIT_FILTER(static_cast<uint64_t>((1ull << COMMIT_BITS) - 1)),
+          DELTA_UPPERBOUND(static_cast<uint64_t>((1ull << COMMIT_BITS))) {}
+    // Number of higher bits of a sequence number that is not used. They are
+    // used to encode the value type, ...
+    const size_t PAD_BITS = static_cast<size_t>(8);
+    // Number of lower bits from prepare seq that can be skipped as they are
+    // implied by the index of the entry in the array
+    const size_t INDEX_BITS;
+    // Number of bits we use to encode the prepare seq
+    const size_t PREP_BITS;
+    // Number of bits we use to encode the commit seq.
+    const size_t COMMIT_BITS;
+    // Filter to encode/decode commit seq
+    const uint64_t COMMIT_FILTER;
+    // The value of commit_seq - prepare_seq + 1 must be less than this bound
+    const uint64_t DELTA_UPPERBOUND;
+  };
+
+  // Prepare Seq (64 bits) = PAD ... PAD PREP PREP ... PREP INDEX INDEX ...
+  // INDEX Delta Seq (64 bits)   = 0 0 0 0 0 0 0 0 0  0 0 0 DELTA DELTA ...
+  // DELTA DELTA Encoded Value         = PREP PREP .... PREP PREP DELTA DELTA
+  // ... DELTA DELTA PAD: first bits of a seq that is reserved for tagging and
+  // hence ignored PREP/INDEX: the used bits in a prepare seq number INDEX: the
+  // bits that do not have to be encoded (will be provided externally) DELTA:
+  // prep seq - commit seq + 1 Number of DELTA bits should be equal to number of
+  // index bits + PADs
+  struct CommitEntry64b {
+    constexpr CommitEntry64b() noexcept : rep_(0) {}
+
+    CommitEntry64b(const CommitEntry& entry, const CommitEntry64bFormat& format)
+        : CommitEntry64b(entry.prep_seq, entry.commit_seq, format) {}
+
+    CommitEntry64b(const uint64_t ps, const uint64_t cs,
+                   const CommitEntry64bFormat& format) {
+      assert(ps < static_cast<uint64_t>(
+                      (1ull << (format.PREP_BITS + format.INDEX_BITS))));
+      assert(ps <= cs);
+      uint64_t delta = cs - ps + 1;  // make initialized delta always >= 1
+      // zero is reserved for uninitialized entries
+      assert(0 < delta);
+      assert(delta < format.DELTA_UPPERBOUND);
+      if (delta >= format.DELTA_UPPERBOUND) {
+        throw std::runtime_error(
+            "commit_seq >> prepare_seq. The allowed distance is " +
+            std::to_string(format.DELTA_UPPERBOUND) + " commit_seq is " +
+            std::to_string(cs) + " prepare_seq is " + std::to_string(ps));
+      }
+      rep_ = (ps << format.PAD_BITS) & ~format.COMMIT_FILTER;
+      rep_ = rep_ | delta;
+    }
+
+    // Return false if the entry is empty
+    bool Parse(const uint64_t indexed_seq, CommitEntry* entry,
+               const CommitEntry64bFormat& format) {
+      uint64_t delta = rep_ & format.COMMIT_FILTER;
+      // zero is reserved for uninitialized entries
+      assert(delta < static_cast<uint64_t>((1ull << format.COMMIT_BITS)));
+      if (delta == 0) {
+        return false;  // initialized entry would have non-zero delta
+      }
+
+      assert(indexed_seq < static_cast<uint64_t>((1ull << format.INDEX_BITS)));
+      uint64_t prep_up = rep_ & ~format.COMMIT_FILTER;
+      prep_up >>= format.PAD_BITS;
+      const uint64_t& prep_low = indexed_seq;
+      entry->prep_seq = prep_up | prep_low;
+
+      entry->commit_seq = entry->prep_seq + delta - 1;
+      return true;
+    }
+
+   private:
+    uint64_t rep_;
+  };
+
+  // Struct to hold ownership of snapshot and read callback for cleanup.
+  struct IteratorState;
+
+  std::shared_ptr<std::map<uint32_t, const Comparator*>> GetCFComparatorMap() {
+    return cf_map_;
+  }
+  std::shared_ptr<std::map<uint32_t, ColumnFamilyHandle*>> GetCFHandleMap() {
+    return handle_map_;
+  }
+  void UpdateCFComparatorMap(
+      const std::vector<ColumnFamilyHandle*>& handles) override;
+  void UpdateCFComparatorMap(ColumnFamilyHandle* handle) override;
+
+  virtual const Snapshot* GetSnapshot() override;
+  SnapshotImpl* GetSnapshotInternal(bool for_ww_conflict_check);
+
+ protected:
+  virtual Status VerifyCFOptions(
+      const ColumnFamilyOptions& cf_options) override;
+  // Assign the min and max sequence numbers for reading from the db. A seq >
+  // max is not valid, and a seq < min is valid, and a min <= seq < max requires
+  // further checking. Normally max is defined by the snapshot and min is by
+  // minimum uncommitted seq.
+  inline SnapshotBackup AssignMinMaxSeqs(const Snapshot* snapshot,
+                                         SequenceNumber* min,
+                                         SequenceNumber* max);
+  // Validate is a snapshot sequence number is still valid based on the latest
+  // db status. backed_by_snapshot specifies if the number is baked by an actual
+  // snapshot object. order specified the memory order with which we load the
+  // atomic variables: relax is enough for the default since we care about last
+  // value seen by same thread.
+  inline bool ValidateSnapshot(
+      const SequenceNumber snap_seq, const SnapshotBackup backed_by_snapshot,
+      std::memory_order order = std::memory_order_relaxed);
+  // Get a dummy snapshot that refers to kMaxSequenceNumber
+  Snapshot* GetMaxSnapshot() { return &dummy_max_snapshot_; }
+
+  bool ShouldRollbackWithSingleDelete(ColumnFamilyHandle* column_family,
+                                      const Slice& key) {
+    return rollback_deletion_type_callback_
+               ? rollback_deletion_type_callback_(this, column_family, key)
+               : false;
+  }
+
+  std::function<bool(TransactionDB*, ColumnFamilyHandle*, const Slice&)>
+      rollback_deletion_type_callback_;
+
+ private:
+  friend class AddPreparedCallback;
+  friend class PreparedHeap_BasicsTest_Test;
+  friend class PreparedHeap_Concurrent_Test;
+  friend class PreparedHeap_EmptyAtTheEnd_Test;
+  friend class SnapshotConcurrentAccessTest_SnapshotConcurrentAccess_Test;
+  friend class WritePreparedCommitEntryPreReleaseCallback;
+  friend class WritePreparedTransactionTestBase;
+  friend class WritePreparedTxn;
+  friend class WritePreparedTxnDBMock;
+  friend class WritePreparedTransactionTest_AddPreparedBeforeMax_Test;
+  friend class WritePreparedTransactionTest_AdvanceMaxEvictedSeqBasic_Test;
+  friend class
+      WritePreparedTransactionTest_AdvanceMaxEvictedSeqWithDuplicates_Test;
+  friend class WritePreparedTransactionTest_AdvanceSeqByOne_Test;
+  friend class WritePreparedTransactionTest_BasicRecovery_Test;
+  friend class WritePreparedTransactionTest_CheckAgainstSnapshots_Test;
+  friend class WritePreparedTransactionTest_CleanupSnapshotEqualToMax_Test;
+  friend class WritePreparedTransactionTest_ConflictDetectionAfterRecovery_Test;
+  friend class WritePreparedTransactionTest_CommitMap_Test;
+  friend class WritePreparedTransactionTest_DoubleSnapshot_Test;
+  friend class WritePreparedTransactionTest_IsInSnapshotEmptyMap_Test;
+  friend class WritePreparedTransactionTest_IsInSnapshotReleased_Test;
+  friend class WritePreparedTransactionTest_IsInSnapshot_Test;
+  friend class WritePreparedTransactionTest_NewSnapshotLargerThanMax_Test;
+  friend class WritePreparedTransactionTest_MaxCatchupWithNewSnapshot_Test;
+  friend class WritePreparedTransactionTest_MaxCatchupWithUnbackedSnapshot_Test;
+  friend class
+      WritePreparedTransactionTest_NonAtomicCommitOfDelayedPrepared_Test;
+  friend class
+      WritePreparedTransactionTest_NonAtomicUpdateOfDelayedPrepared_Test;
+  friend class WritePreparedTransactionTest_NonAtomicUpdateOfMaxEvictedSeq_Test;
+  friend class WritePreparedTransactionTest_OldCommitMapGC_Test;
+  friend class WritePreparedTransactionTest_Rollback_Test;
+  friend class WritePreparedTransactionTest_SmallestUnCommittedSeq_Test;
+  friend class WriteUnpreparedTxn;
+  friend class WriteUnpreparedTxnDB;
+  friend class WriteUnpreparedTransactionTest_RecoveryTest_Test;
+  friend class MultiOpsTxnsStressTest;
+
+  void Init(const TransactionDBOptions& txn_db_opts);
+
+  void WPRecordTick(uint32_t ticker_type) const {
+    RecordTick(db_impl_->immutable_db_options_.statistics.get(), ticker_type);
+  }
+
+  // A heap with the amortized O(1) complexity for erase. It uses one extra heap
+  // to keep track of erased entries that are not yet on top of the main heap.
+  class PreparedHeap {
+    // The mutex is required for push and pop from PreparedHeap. ::erase will
+    // use external synchronization via prepared_mutex_.
+    port::Mutex push_pop_mutex_;
+    std::deque<uint64_t> heap_;
+    std::priority_queue<uint64_t, std::vector<uint64_t>, std::greater<uint64_t>>
+        erased_heap_;
+    std::atomic<uint64_t> heap_top_ = {kMaxSequenceNumber};
+    // True when testing crash recovery
+    bool TEST_CRASH_ = false;
+    friend class WritePreparedTxnDB;
+
+   public:
+    ~PreparedHeap() {
+      if (!TEST_CRASH_) {
+        assert(heap_.empty());
+        assert(erased_heap_.empty());
+      }
+    }
+    port::Mutex* push_pop_mutex() { return &push_pop_mutex_; }
+
+    inline bool empty() { return top() == kMaxSequenceNumber; }
+    // Returns kMaxSequenceNumber if empty() and the smallest otherwise.
+    inline uint64_t top() { return heap_top_.load(std::memory_order_acquire); }
+    inline void push(uint64_t v) {
+      push_pop_mutex_.AssertHeld();
+      if (heap_.empty()) {
+        heap_top_.store(v, std::memory_order_release);
+      } else {
+        assert(heap_top_.load() < v);
+      }
+      heap_.push_back(v);
+    }
+    void pop(bool locked = false) {
+      if (!locked) {
+        push_pop_mutex()->Lock();
+      }
+      push_pop_mutex_.AssertHeld();
+      heap_.pop_front();
+      while (!heap_.empty() && !erased_heap_.empty() &&
+             // heap_.top() > erased_heap_.top() could happen if we have erased
+             // a non-existent entry. Ideally the user should not do that but we
+             // should be resilient against it.
+             heap_.front() >= erased_heap_.top()) {
+        if (heap_.front() == erased_heap_.top()) {
+          heap_.pop_front();
+        }
+        uint64_t erased __attribute__((__unused__));
+        erased = erased_heap_.top();
+        erased_heap_.pop();
+        // No duplicate prepare sequence numbers
+        assert(erased_heap_.empty() || erased_heap_.top() != erased);
+      }
+      while (heap_.empty() && !erased_heap_.empty()) {
+        erased_heap_.pop();
+      }
+      heap_top_.store(!heap_.empty() ? heap_.front() : kMaxSequenceNumber,
+                      std::memory_order_release);
+      if (!locked) {
+        push_pop_mutex()->Unlock();
+      }
+    }
+    // Concurrrent calls needs external synchronization. It is safe to be called
+    // concurrent to push and pop though.
+    void erase(uint64_t seq) {
+      if (!empty()) {
+        auto top_seq = top();
+        if (seq < top_seq) {
+          // Already popped, ignore it.
+        } else if (top_seq == seq) {
+          pop();
+#ifndef NDEBUG
+          MutexLock ml(push_pop_mutex());
+          assert(heap_.empty() || heap_.front() != seq);
+#endif
+        } else {  // top() > seq
+          // Down the heap, remember to pop it later
+          erased_heap_.push(seq);
+        }
+      }
+    }
+  };
+
+  void TEST_Crash() override { prepared_txns_.TEST_CRASH_ = true; }
+
+  // Get the commit entry with index indexed_seq from the commit table. It
+  // returns true if such entry exists.
+  bool GetCommitEntry(const uint64_t indexed_seq, CommitEntry64b* entry_64b,
+                      CommitEntry* entry) const;
+
+  // Rewrite the entry with the index indexed_seq in the commit table with the
+  // commit entry <prep_seq, commit_seq>. If the rewrite results into eviction,
+  // sets the evicted_entry and returns true.
+  bool AddCommitEntry(const uint64_t indexed_seq, const CommitEntry& new_entry,
+                      CommitEntry* evicted_entry);
+
+  // Rewrite the entry with the index indexed_seq in the commit table with the
+  // commit entry new_entry only if the existing entry matches the
+  // expected_entry. Returns false otherwise.
+  bool ExchangeCommitEntry(const uint64_t indexed_seq,
+                           CommitEntry64b& expected_entry,
+                           const CommitEntry& new_entry);
+
+  // Increase max_evicted_seq_ from the previous value prev_max to the new
+  // value. This also involves taking care of prepared txns that are not
+  // committed before new_max, as well as updating the list of live snapshots at
+  // the time of updating the max. Thread-safety: this function can be called
+  // concurrently. The concurrent invocations of this function is equivalent to
+  // a serial invocation in which the last invocation is the one with the
+  // largest new_max value.
+  void AdvanceMaxEvictedSeq(const SequenceNumber& prev_max,
+                            const SequenceNumber& new_max);
+
+  inline SequenceNumber SmallestUnCommittedSeq() {
+    // Note: We have two lists to look into, but for performance reasons they
+    // are not read atomically. Since CheckPreparedAgainstMax copies the entry
+    // to delayed_prepared_ before removing it from prepared_txns_, to ensure
+    // that a prepared entry will not go unmissed, we look into them in opposite
+    // order: first read prepared_txns_ and then delayed_prepared_.
+
+    // This must be called before calling ::top. This is because the concurrent
+    // thread would call ::RemovePrepared before updating
+    // GetLatestSequenceNumber(). Reading then in opposite order here guarantees
+    // that the ::top that we read would be lower the ::top if we had otherwise
+    // update/read them atomically.
+    auto next_prepare = db_impl_->GetLatestSequenceNumber() + 1;
+    auto min_prepare = prepared_txns_.top();
+    // Since we update the prepare_heap always from the main write queue via
+    // PreReleaseCallback, the prepared_txns_.top() indicates the smallest
+    // prepared data in 2pc transactions. For non-2pc transactions that are
+    // written in two steps, we also update prepared_txns_ at the first step
+    // (via the same mechanism) so that their uncommitted data is reflected in
+    // SmallestUnCommittedSeq.
+    if (!delayed_prepared_empty_.load()) {
+      ReadLock rl(&prepared_mutex_);
+      if (!delayed_prepared_.empty()) {
+        return *delayed_prepared_.begin();
+      }
+    }
+    bool empty = min_prepare == kMaxSequenceNumber;
+    if (empty) {
+      // Since GetLatestSequenceNumber is updated
+      // after prepared_txns_ are, the value of GetLatestSequenceNumber would
+      // reflect any uncommitted data that is not added to prepared_txns_ yet.
+      // Otherwise, if there is no concurrent txn, this value simply reflects
+      // that latest value in the memtable.
+      return next_prepare;
+    } else {
+      return std::min(min_prepare, next_prepare);
+    }
+  }
+
+  // Enhance the snapshot object by recording in it the smallest uncommitted seq
+  inline void EnhanceSnapshot(SnapshotImpl* snapshot,
+                              SequenceNumber min_uncommitted) {
+    assert(snapshot);
+    assert(min_uncommitted <= snapshot->number_ + 1);
+    snapshot->min_uncommitted_ = min_uncommitted;
+  }
+
+  virtual const std::vector<SequenceNumber> GetSnapshotListFromDB(
+      SequenceNumber max);
+
+  // Will be called by the public ReleaseSnapshot method. Does the maintenance
+  // internal to WritePreparedTxnDB
+  void ReleaseSnapshotInternal(const SequenceNumber snap_seq);
+
+  // Update the list of snapshots corresponding to the soon-to-be-updated
+  // max_evicted_seq_. Thread-safety: this function can be called concurrently.
+  // The concurrent invocations of this function is equivalent to a serial
+  // invocation in which the last invocation is the one with the largest
+  // version value.
+  void UpdateSnapshots(const std::vector<SequenceNumber>& snapshots,
+                       const SequenceNumber& version);
+  // Check the new list of new snapshots against the old one to see  if any of
+  // the snapshots are released and to do the cleanup for the released snapshot.
+  void CleanupReleasedSnapshots(
+      const std::vector<SequenceNumber>& new_snapshots,
+      const std::vector<SequenceNumber>& old_snapshots);
+
+  // Check an evicted entry against live snapshots to see if it should be kept
+  // around or it can be safely discarded (and hence assume committed for all
+  // snapshots). Thread-safety: this function can be called concurrently. If it
+  // is called concurrently with multiple UpdateSnapshots, the result is the
+  // same as checking the intersection of the snapshot list before updates with
+  // the snapshot list of all the concurrent updates.
+  void CheckAgainstSnapshots(const CommitEntry& evicted);
+
+  // Add a new entry to old_commit_map_ if prep_seq <= snapshot_seq <
+  // commit_seq. Return false if checking the next snapshot(s) is not needed.
+  // This is the case if none of the next snapshots could satisfy the condition.
+  // next_is_larger: the next snapshot will be a larger value
+  bool MaybeUpdateOldCommitMap(const uint64_t& prep_seq,
+                               const uint64_t& commit_seq,
+                               const uint64_t& snapshot_seq,
+                               const bool next_is_larger);
+
+  // A trick to increase the last visible sequence number by one and also wait
+  // for the in-flight commits to be visible.
+  void AdvanceSeqByOne();
+
+  // The list of live snapshots at the last time that max_evicted_seq_ advanced.
+  // The list stored into two data structures: in snapshot_cache_ that is
+  // efficient for concurrent reads, and in snapshots_ if the data does not fit
+  // into snapshot_cache_. The total number of snapshots in the two lists
+  std::atomic<size_t> snapshots_total_ = {};
+  // The list sorted in ascending order. Thread-safety for writes is provided
+  // with snapshots_mutex_ and concurrent reads are safe due to std::atomic for
+  // each entry. In x86_64 architecture such reads are compiled to simple read
+  // instructions.
+  const size_t SNAPSHOT_CACHE_BITS;
+  const size_t SNAPSHOT_CACHE_SIZE;
+  std::unique_ptr<std::atomic<SequenceNumber>[]> snapshot_cache_;
+  // 2nd list for storing snapshots. The list sorted in ascending order.
+  // Thread-safety is provided with snapshots_mutex_.
+  std::vector<SequenceNumber> snapshots_;
+  // The list of all snapshots: snapshots_ + snapshot_cache_. This list although
+  // redundant but simplifies CleanupOldSnapshots implementation.
+  // Thread-safety is provided with snapshots_mutex_.
+  std::vector<SequenceNumber> snapshots_all_;
+  // The version of the latest list of snapshots. This can be used to avoid
+  // rewriting a list that is concurrently updated with a more recent version.
+  SequenceNumber snapshots_version_ = 0;
+
+  // A heap of prepared transactions. Thread-safety is provided with
+  // prepared_mutex_.
+  PreparedHeap prepared_txns_;
+  const size_t COMMIT_CACHE_BITS;
+  const size_t COMMIT_CACHE_SIZE;
+  const CommitEntry64bFormat FORMAT;
+  // commit_cache_ must be initialized to zero to tell apart an empty index from
+  // a filled one. Thread-safety is provided with commit_cache_mutex_.
+  std::unique_ptr<std::atomic<CommitEntry64b>[]> commit_cache_;
+  // The largest evicted *commit* sequence number from the commit_cache_. If a
+  // seq is smaller than max_evicted_seq_ is might or might not be present in
+  // commit_cache_. So commit_cache_ must first be checked before consulting
+  // with max_evicted_seq_.
+  std::atomic<uint64_t> max_evicted_seq_ = {};
+  // Order: 1) update future_max_evicted_seq_ = new_max, 2)
+  // GetSnapshotListFromDB(new_max), max_evicted_seq_ = new_max. Since
+  // GetSnapshotInternal guarantess that the snapshot seq is larger than
+  // future_max_evicted_seq_, this guarantes that if a snapshot is not larger
+  // than max has already being looked at via a GetSnapshotListFromDB(new_max).
+  std::atomic<uint64_t> future_max_evicted_seq_ = {};
+  // Advance max_evicted_seq_ by this value each time it needs an update. The
+  // larger the value, the less frequent advances we would have. We do not want
+  // it to be too large either as it would cause stalls by doing too much
+  // maintenance work under the lock.
+  size_t INC_STEP_FOR_MAX_EVICTED = 1;
+  // A map from old snapshots (expected to be used by a few read-only txns) to
+  // prepared sequence number of the evicted entries from commit_cache_ that
+  // overlaps with such snapshot. These are the prepared sequence numbers that
+  // the snapshot, to which they are mapped, cannot assume to be committed just
+  // because it is no longer in the commit_cache_. The vector must be sorted
+  // after each update.
+  // Thread-safety is provided with old_commit_map_mutex_.
+  std::map<SequenceNumber, std::vector<SequenceNumber>> old_commit_map_;
+  // A set of long-running prepared transactions that are not finished by the
+  // time max_evicted_seq_ advances their sequence number. This is expected to
+  // be empty normally. Thread-safety is provided with prepared_mutex_.
+  std::set<uint64_t> delayed_prepared_;
+  // Commit of a delayed prepared: 1) update commit cache, 2) update
+  // delayed_prepared_commits_, 3) publish seq, 3) clean up delayed_prepared_.
+  // delayed_prepared_commits_ will help us tell apart the unprepared txns from
+  // the ones that are committed but not cleaned up yet.
+  std::unordered_map<SequenceNumber, SequenceNumber> delayed_prepared_commits_;
+  // Update when delayed_prepared_.empty() changes. Expected to be true
+  // normally.
+  std::atomic<bool> delayed_prepared_empty_ = {true};
+  // Update when old_commit_map_.empty() changes. Expected to be true normally.
+  std::atomic<bool> old_commit_map_empty_ = {true};
+  mutable port::RWMutex prepared_mutex_;
+  mutable port::RWMutex old_commit_map_mutex_;
+  mutable port::RWMutex commit_cache_mutex_;
+  mutable port::RWMutex snapshots_mutex_;
+  // A cache of the cf comparators
+  // Thread safety: since it is a const it is safe to read it concurrently
+  std::shared_ptr<std::map<uint32_t, const Comparator*>> cf_map_;
+  // A cache of the cf handles
+  // Thread safety: since the handle is read-only object it is a const it is
+  // safe to read it concurrently
+  std::shared_ptr<std::map<uint32_t, ColumnFamilyHandle*>> handle_map_;
+  // A dummy snapshot object that refers to kMaxSequenceNumber
+  SnapshotImpl dummy_max_snapshot_;
+};
+
+class WritePreparedTxnReadCallback : public ReadCallback {
+ public:
+  WritePreparedTxnReadCallback(WritePreparedTxnDB* db, SequenceNumber snapshot)
+      : ReadCallback(snapshot),
+        db_(db),
+        backed_by_snapshot_(kBackedByDBSnapshot) {}
+  WritePreparedTxnReadCallback(WritePreparedTxnDB* db, SequenceNumber snapshot,
+                               SequenceNumber min_uncommitted,
+                               SnapshotBackup backed_by_snapshot)
+      : ReadCallback(snapshot, min_uncommitted),
+        db_(db),
+        backed_by_snapshot_(backed_by_snapshot) {
+    (void)backed_by_snapshot_;  // to silence unused private field warning
+  }
+
+  virtual ~WritePreparedTxnReadCallback() {
+    // If it is not backed by snapshot, the caller must check validity
+    assert(valid_checked_ || backed_by_snapshot_ == kBackedByDBSnapshot);
+  }
+
+  // Will be called to see if the seq number visible; if not it moves on to
+  // the next seq number.
+  inline virtual bool IsVisibleFullCheck(SequenceNumber seq) override {
+    auto snapshot = max_visible_seq_;
+    bool snap_released = false;
+    auto ret =
+        db_->IsInSnapshot(seq, snapshot, min_uncommitted_, &snap_released);
+    assert(!snap_released || backed_by_snapshot_ == kUnbackedByDBSnapshot);
+    snap_released_ |= snap_released;
+    return ret;
+  }
+
+  inline bool valid() {
+    valid_checked_ = true;
+    return snap_released_ == false;
+  }
+
+  // TODO(myabandeh): override Refresh when Iterator::Refresh is supported
+ private:
+  WritePreparedTxnDB* db_;
+  // Whether max_visible_seq_ is backed by a snapshot
+  const SnapshotBackup backed_by_snapshot_;
+  bool snap_released_ = false;
+  // Safety check to ensure that the caller has checked invalid statuses
+  bool valid_checked_ = false;
+};
+
+class AddPreparedCallback : public PreReleaseCallback {
+ public:
+  AddPreparedCallback(WritePreparedTxnDB* db, DBImpl* db_impl,
+                      size_t sub_batch_cnt, bool two_write_queues,
+                      bool first_prepare_batch)
+      : db_(db),
+        db_impl_(db_impl),
+        sub_batch_cnt_(sub_batch_cnt),
+        two_write_queues_(two_write_queues),
+        first_prepare_batch_(first_prepare_batch) {
+    (void)two_write_queues_;  // to silence unused private field warning
+  }
+  virtual Status Callback(SequenceNumber prepare_seq,
+                          bool is_mem_disabled __attribute__((__unused__)),
+                          uint64_t log_number, size_t index,
+                          size_t total) override {
+    assert(index < total);
+    // To reduce the cost of lock acquisition competing with the concurrent
+    // prepare requests, lock on the first callback and unlock on the last.
+    const bool do_lock = !two_write_queues_ || index == 0;
+    const bool do_unlock = !two_write_queues_ || index + 1 == total;
+    // Always Prepare from the main queue
+    assert(!two_write_queues_ || !is_mem_disabled);  // implies the 1st queue
+    TEST_SYNC_POINT("AddPreparedCallback::AddPrepared::begin:pause");
+    TEST_SYNC_POINT("AddPreparedCallback::AddPrepared::begin:resume");
+    if (do_lock) {
+      db_->prepared_txns_.push_pop_mutex()->Lock();
+    }
+    const bool kLocked = true;
+    for (size_t i = 0; i < sub_batch_cnt_; i++) {
+      db_->AddPrepared(prepare_seq + i, kLocked);
+    }
+    if (do_unlock) {
+      db_->prepared_txns_.push_pop_mutex()->Unlock();
+    }
+    TEST_SYNC_POINT("AddPreparedCallback::AddPrepared::end");
+    if (first_prepare_batch_) {
+      assert(log_number != 0);
+      db_impl_->logs_with_prep_tracker()->MarkLogAsContainingPrepSection(
+          log_number);
+    }
+    return Status::OK();
+  }
+
+ private:
+  WritePreparedTxnDB* db_;
+  DBImpl* db_impl_;
+  size_t sub_batch_cnt_;
+  bool two_write_queues_;
+  // It is 2PC and this is the first prepare batch. Always the case in 2PC
+  // unless it is WriteUnPrepared.
+  bool first_prepare_batch_;
+};
+
+class WritePreparedCommitEntryPreReleaseCallback : public PreReleaseCallback {
+ public:
+  // includes_data indicates that the commit also writes non-empty
+  // CommitTimeWriteBatch to memtable, which needs to be committed separately.
+  WritePreparedCommitEntryPreReleaseCallback(
+      WritePreparedTxnDB* db, DBImpl* db_impl, SequenceNumber prep_seq,
+      size_t prep_batch_cnt, size_t data_batch_cnt = 0,
+      SequenceNumber aux_seq = kMaxSequenceNumber, size_t aux_batch_cnt = 0)
+      : db_(db),
+        db_impl_(db_impl),
+        prep_seq_(prep_seq),
+        prep_batch_cnt_(prep_batch_cnt),
+        data_batch_cnt_(data_batch_cnt),
+        includes_data_(data_batch_cnt_ > 0),
+        aux_seq_(aux_seq),
+        aux_batch_cnt_(aux_batch_cnt),
+        includes_aux_batch_(aux_batch_cnt > 0) {
+    assert((prep_batch_cnt_ > 0) != (prep_seq == kMaxSequenceNumber));  // xor
+    assert(prep_batch_cnt_ > 0 || data_batch_cnt_ > 0);
+    assert((aux_batch_cnt_ > 0) != (aux_seq == kMaxSequenceNumber));  // xor
+  }
+
+  virtual Status Callback(SequenceNumber commit_seq,
+                          bool is_mem_disabled __attribute__((__unused__)),
+                          uint64_t, size_t /*index*/,
+                          size_t /*total*/) override {
+    // Always commit from the 2nd queue
+    assert(!db_impl_->immutable_db_options().two_write_queues ||
+           is_mem_disabled);
+    assert(includes_data_ || prep_seq_ != kMaxSequenceNumber);
+    // Data batch is what accompanied with the commit marker and affects the
+    // last seq in the commit batch.
+    const uint64_t last_commit_seq = LIKELY(data_batch_cnt_ <= 1)
+                                         ? commit_seq
+                                         : commit_seq + data_batch_cnt_ - 1;
+    if (prep_seq_ != kMaxSequenceNumber) {
+      for (size_t i = 0; i < prep_batch_cnt_; i++) {
+        db_->AddCommitted(prep_seq_ + i, last_commit_seq);
+      }
+    }  // else there was no prepare phase
+    if (includes_aux_batch_) {
+      for (size_t i = 0; i < aux_batch_cnt_; i++) {
+        db_->AddCommitted(aux_seq_ + i, last_commit_seq);
+      }
+    }
+    if (includes_data_) {
+      assert(data_batch_cnt_);
+      // Commit the data that is accompanied with the commit request
+      for (size_t i = 0; i < data_batch_cnt_; i++) {
+        // For commit seq of each batch use the commit seq of the last batch.
+        // This would make debugging easier by having all the batches having
+        // the same sequence number.
+        db_->AddCommitted(commit_seq + i, last_commit_seq);
+      }
+    }
+    if (db_impl_->immutable_db_options().two_write_queues) {
+      assert(is_mem_disabled);  // implies the 2nd queue
+      // Publish the sequence number. We can do that here assuming the callback
+      // is invoked only from one write queue, which would guarantee that the
+      // publish sequence numbers will be in order, i.e., once a seq is
+      // published all the seq prior to that are also publishable.
+      db_impl_->SetLastPublishedSequence(last_commit_seq);
+      // Note RemovePrepared should be called after publishing the seq.
+      // Otherwise SmallestUnCommittedSeq optimization breaks.
+      if (prep_seq_ != kMaxSequenceNumber) {
+        db_->RemovePrepared(prep_seq_, prep_batch_cnt_);
+      }  // else there was no prepare phase
+      if (includes_aux_batch_) {
+        db_->RemovePrepared(aux_seq_, aux_batch_cnt_);
+      }
+    }
+    // else SequenceNumber that is updated as part of the write already does the
+    // publishing
+    return Status::OK();
+  }
+
+ private:
+  WritePreparedTxnDB* db_;
+  DBImpl* db_impl_;
+  // kMaxSequenceNumber if there was no prepare phase
+  SequenceNumber prep_seq_;
+  size_t prep_batch_cnt_;
+  size_t data_batch_cnt_;
+  // Data here is the batch that is written with the commit marker, either
+  // because it is commit without prepare or commit has a CommitTimeWriteBatch.
+  bool includes_data_;
+  // Auxiliary batch (if there is any) is a batch that is written before, but
+  // gets the same commit seq as prepare batch or data batch. This is used in
+  // two write queues where the CommitTimeWriteBatch becomes the aux batch and
+  // we do a separate write to actually commit everything.
+  SequenceNumber aux_seq_;
+  size_t aux_batch_cnt_;
+  bool includes_aux_batch_;
+};
+
+// For two_write_queues commit both the aborted batch and the cleanup batch and
+// then published the seq
+class WritePreparedRollbackPreReleaseCallback : public PreReleaseCallback {
+ public:
+  WritePreparedRollbackPreReleaseCallback(WritePreparedTxnDB* db,
+                                          DBImpl* db_impl,
+                                          SequenceNumber prep_seq,
+                                          SequenceNumber rollback_seq,
+                                          size_t prep_batch_cnt)
+      : db_(db),
+        db_impl_(db_impl),
+        prep_seq_(prep_seq),
+        rollback_seq_(rollback_seq),
+        prep_batch_cnt_(prep_batch_cnt) {
+    assert(prep_seq != kMaxSequenceNumber);
+    assert(rollback_seq != kMaxSequenceNumber);
+    assert(prep_batch_cnt_ > 0);
+  }
+
+  Status Callback(SequenceNumber commit_seq, bool is_mem_disabled, uint64_t,
+                  size_t /*index*/, size_t /*total*/) override {
+    // Always commit from the 2nd queue
+    assert(is_mem_disabled);  // implies the 2nd queue
+    assert(db_impl_->immutable_db_options().two_write_queues);
+#ifdef NDEBUG
+    (void)is_mem_disabled;
+#endif
+    const uint64_t last_commit_seq = commit_seq;
+    db_->AddCommitted(rollback_seq_, last_commit_seq);
+    for (size_t i = 0; i < prep_batch_cnt_; i++) {
+      db_->AddCommitted(prep_seq_ + i, last_commit_seq);
+    }
+    db_impl_->SetLastPublishedSequence(last_commit_seq);
+    return Status::OK();
+  }
+
+ private:
+  WritePreparedTxnDB* db_;
+  DBImpl* db_impl_;
+  SequenceNumber prep_seq_;
+  SequenceNumber rollback_seq_;
+  size_t prep_batch_cnt_;
+};
+
+// Count the number of sub-batches inside a batch. A sub-batch does not have
+// duplicate keys.
+struct SubBatchCounter : public WriteBatch::Handler {
+  explicit SubBatchCounter(std::map<uint32_t, const Comparator*>& comparators)
+      : comparators_(comparators), batches_(1) {}
+  std::map<uint32_t, const Comparator*>& comparators_;
+  using CFKeys = std::set<Slice, SetComparator>;
+  std::map<uint32_t, CFKeys> keys_;
+  size_t batches_;
+  size_t BatchCount() { return batches_; }
+  void AddKey(const uint32_t cf, const Slice& key);
+  void InitWithComp(const uint32_t cf);
+  Status MarkNoop(bool) override { return Status::OK(); }
+  Status MarkEndPrepare(const Slice&) override { return Status::OK(); }
+  Status MarkCommit(const Slice&) override { return Status::OK(); }
+  Status PutCF(uint32_t cf, const Slice& key, const Slice&) override {
+    AddKey(cf, key);
+    return Status::OK();
+  }
+  Status DeleteCF(uint32_t cf, const Slice& key) override {
+    AddKey(cf, key);
+    return Status::OK();
+  }
+  Status SingleDeleteCF(uint32_t cf, const Slice& key) override {
+    AddKey(cf, key);
+    return Status::OK();
+  }
+  Status MergeCF(uint32_t cf, const Slice& key, const Slice&) override {
+    AddKey(cf, key);
+    return Status::OK();
+  }
+  Status MarkBeginPrepare(bool) override { return Status::OK(); }
+  Status MarkRollback(const Slice&) override { return Status::OK(); }
+  Handler::OptionState WriteAfterCommit() const override {
+    return Handler::OptionState::kDisabled;
+  }
+};
+
+SnapshotBackup WritePreparedTxnDB::AssignMinMaxSeqs(const Snapshot* snapshot,
+                                                    SequenceNumber* min,
+                                                    SequenceNumber* max) {
+  if (snapshot != nullptr) {
+    *min =
+        static_cast_with_check<const SnapshotImpl>(snapshot)->min_uncommitted_;
+    *max = static_cast_with_check<const SnapshotImpl>(snapshot)->number_;
+    // A duplicate of the check in EnhanceSnapshot().
+    assert(*min <= *max + 1);
+    return kBackedByDBSnapshot;
+  } else {
+    *min = SmallestUnCommittedSeq();
+    *max = 0;  // to be assigned later after sv is referenced.
+    return kUnbackedByDBSnapshot;
+  }
+}
+
+bool WritePreparedTxnDB::ValidateSnapshot(
+    const SequenceNumber snap_seq, const SnapshotBackup backed_by_snapshot,
+    std::memory_order order) {
+  if (backed_by_snapshot == kBackedByDBSnapshot) {
+    return true;
+  } else {
+    SequenceNumber max = max_evicted_seq_.load(order);
+    // Validate that max has not advanced the snapshot seq that is not backed
+    // by a real snapshot. This is a very rare case that should not happen in
+    // real workloads.
+    if (UNLIKELY(snap_seq <= max && snap_seq != 0)) {
+      return false;
+    }
+  }
+  return true;
+}
+
+}  // namespace ROCKSDB_NAMESPACE
+#endif  // ROCKSDB_LITE
diff --git a/src/rocksdb/utilities/transactions/write_unprepared_transaction_test.cc b/src/rocksdb/utilities/transactions/write_unprepared_transaction_test.cc
new file mode 100644
index 000000000..6c8c62e0e
--- /dev/null
+++ b/src/rocksdb/utilities/transactions/write_unprepared_transaction_test.cc
@@ -0,0 +1,790 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#ifndef ROCKSDB_LITE
+
+#include "utilities/transactions/transaction_test.h"
+#include "utilities/transactions/write_unprepared_txn.h"
+#include "utilities/transactions/write_unprepared_txn_db.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+class WriteUnpreparedTransactionTestBase : public TransactionTestBase {
+ public:
+  WriteUnpreparedTransactionTestBase(bool use_stackable_db,
+                                     bool two_write_queue,
+                                     TxnDBWritePolicy write_policy)
+      : TransactionTestBase(use_stackable_db, two_write_queue, write_policy,
+                            kOrderedWrite) {}
+};
+
+class WriteUnpreparedTransactionTest
+    : public WriteUnpreparedTransactionTestBase,
+      virtual public ::testing::WithParamInterface<
+          std::tuple<bool, bool, TxnDBWritePolicy>> {
+ public:
+  WriteUnpreparedTransactionTest()
+      : WriteUnpreparedTransactionTestBase(std::get<0>(GetParam()),
+                                           std::get<1>(GetParam()),
+                                           std::get<2>(GetParam())) {}
+};
+
+INSTANTIATE_TEST_CASE_P(
+    WriteUnpreparedTransactionTest, WriteUnpreparedTransactionTest,
+    ::testing::Values(std::make_tuple(false, false, WRITE_UNPREPARED),
+                      std::make_tuple(false, true, WRITE_UNPREPARED)));
+
+enum StressAction { NO_SNAPSHOT, RO_SNAPSHOT, REFRESH_SNAPSHOT };
+class WriteUnpreparedStressTest : public WriteUnpreparedTransactionTestBase,
+                                  virtual public ::testing::WithParamInterface<
+                                      std::tuple<bool, StressAction>> {
+ public:
+  WriteUnpreparedStressTest()
+      : WriteUnpreparedTransactionTestBase(false, std::get<0>(GetParam()),
+                                           WRITE_UNPREPARED),
+        action_(std::get<1>(GetParam())) {}
+  StressAction action_;
+};
+
+INSTANTIATE_TEST_CASE_P(
+    WriteUnpreparedStressTest, WriteUnpreparedStressTest,
+    ::testing::Values(std::make_tuple(false, NO_SNAPSHOT),
+                      std::make_tuple(false, RO_SNAPSHOT),
+                      std::make_tuple(false, REFRESH_SNAPSHOT),
+                      std::make_tuple(true, NO_SNAPSHOT),
+                      std::make_tuple(true, RO_SNAPSHOT),
+                      std::make_tuple(true, REFRESH_SNAPSHOT)));
+
+TEST_P(WriteUnpreparedTransactionTest, ReadYourOwnWrite) {
+  // The following tests checks whether reading your own write for
+  // a transaction works for write unprepared, when there are uncommitted
+  // values written into DB.
+  auto verify_state = [](Iterator* iter, const std::string& key,
+                         const std::string& value) {
+    ASSERT_TRUE(iter->Valid());
+    ASSERT_OK(iter->status());
+    ASSERT_EQ(key, iter->key().ToString());
+    ASSERT_EQ(value, iter->value().ToString());
+  };
+
+  // Test always reseeking vs never reseeking.
+  for (uint64_t max_skip : {0, std::numeric_limits<int>::max()}) {
+    options.max_sequential_skip_in_iterations = max_skip;
+    options.disable_auto_compactions = true;
+    ASSERT_OK(ReOpen());
+
+    TransactionOptions txn_options;
+    WriteOptions woptions;
+    ReadOptions roptions;
+
+    ASSERT_OK(db->Put(woptions, "a", ""));
+    ASSERT_OK(db->Put(woptions, "b", ""));
+
+    Transaction* txn = db->BeginTransaction(woptions, txn_options);
+    WriteUnpreparedTxn* wup_txn = dynamic_cast<WriteUnpreparedTxn*>(txn);
+    txn->SetSnapshot();
+
+    for (int i = 0; i < 5; i++) {
+      std::string stored_value = "v" + std::to_string(i);
+      ASSERT_OK(txn->Put("a", stored_value));
+      ASSERT_OK(txn->Put("b", stored_value));
+      ASSERT_OK(wup_txn->FlushWriteBatchToDB(false));
+
+      // Test Get()
+      std::string value;
+      ASSERT_OK(txn->Get(roptions, "a", &value));
+      ASSERT_EQ(value, stored_value);
+      ASSERT_OK(txn->Get(roptions, "b", &value));
+      ASSERT_EQ(value, stored_value);
+
+      // Test Next()
+      auto iter = txn->GetIterator(roptions);
+      iter->Seek("a");
+      verify_state(iter, "a", stored_value);
+
+      iter->Next();
+      verify_state(iter, "b", stored_value);
+
+      iter->SeekToFirst();
+      verify_state(iter, "a", stored_value);
+
+      iter->Next();
+      verify_state(iter, "b", stored_value);
+
+      delete iter;
+
+      // Test Prev()
+      iter = txn->GetIterator(roptions);
+      iter->SeekForPrev("b");
+      verify_state(iter, "b", stored_value);
+
+      iter->Prev();
+      verify_state(iter, "a", stored_value);
+
+      iter->SeekToLast();
+      verify_state(iter, "b", stored_value);
+
+      iter->Prev();
+      verify_state(iter, "a", stored_value);
+
+      delete iter;
+    }
+
+    delete txn;
+  }
+}
+
+#if !defined(ROCKSDB_VALGRIND_RUN) || defined(ROCKSDB_FULL_VALGRIND_RUN)
+TEST_P(WriteUnpreparedStressTest, ReadYourOwnWriteStress) {
+  // This is a stress test where different threads are writing random keys, and
+  // then before committing or aborting the transaction, it validates to see
+  // that it can read the keys it wrote, and the keys it did not write respect
+  // the snapshot. To avoid row lock contention (and simply stressing the
+  // locking system), each thread is mostly only writing to its own set of keys.
+  const uint32_t kNumIter = 1000;
+  const uint32_t kNumThreads = 10;
+  const uint32_t kNumKeys = 5;
+
+  // Test with
+  // 1. no snapshots set
+  // 2. snapshot set on ReadOptions
+  // 3. snapshot set, and refreshing after every write.
+  StressAction a = action_;
+  WriteOptions write_options;
+  txn_db_options.transaction_lock_timeout = -1;
+  options.disable_auto_compactions = true;
+  ASSERT_OK(ReOpen());
+
+  std::vector<std::string> keys;
+  for (uint32_t k = 0; k < kNumKeys * kNumThreads; k++) {
+    keys.push_back("k" + std::to_string(k));
+  }
+  RandomShuffle(keys.begin(), keys.end());
+
+  // This counter will act as a "sequence number" to help us validate
+  // visibility logic with snapshots. If we had direct access to the seqno of
+  // snapshots and key/values, then we should directly compare those instead.
+  std::atomic<int64_t> counter(0);
+
+  std::function<void(uint32_t)> stress_thread = [&](int id) {
+    size_t tid = std::hash<std::thread::id>()(std::this_thread::get_id());
+    Random64 rnd(static_cast<uint32_t>(tid));
+
+    Transaction* txn;
+    TransactionOptions txn_options;
+    // batch_size of 1 causes writes to DB for every marker.
+    txn_options.write_batch_flush_threshold = 1;
+    ReadOptions read_options;
+
+    for (uint32_t i = 0; i < kNumIter; i++) {
+      std::set<std::string> owned_keys(keys.begin() + id * kNumKeys,
+                                       keys.begin() + (id + 1) * kNumKeys);
+      // Add unowned keys to make the workload more interesting, but this
+      // increases row lock contention, so just do it sometimes.
+      if (rnd.OneIn(2)) {
+        owned_keys.insert(keys[rnd.Uniform(kNumKeys * kNumThreads)]);
+      }
+
+      txn = db->BeginTransaction(write_options, txn_options);
+      ASSERT_OK(txn->SetName(std::to_string(id)));
+      txn->SetSnapshot();
+      if (a >= RO_SNAPSHOT) {
+        read_options.snapshot = txn->GetSnapshot();
+        ASSERT_TRUE(read_options.snapshot != nullptr);
+      }
+
+      uint64_t buf[2];
+      buf[0] = id;
+
+      // When scanning through the database, make sure that all unprepared
+      // keys have value >= snapshot and all other keys have value < snapshot.
+      int64_t snapshot_num = counter.fetch_add(1);
+
+      Status s;
+      for (const auto& key : owned_keys) {
+        buf[1] = counter.fetch_add(1);
+        s = txn->Put(key, Slice((const char*)buf, sizeof(buf)));
+        if (!s.ok()) {
+          break;
+        }
+        if (a == REFRESH_SNAPSHOT) {
+          txn->SetSnapshot();
+          read_options.snapshot = txn->GetSnapshot();
+          snapshot_num = counter.fetch_add(1);
+        }
+      }
+
+      // Failure is possible due to snapshot validation. In this case,
+      // rollback and move onto next iteration.
+      if (!s.ok()) {
+        ASSERT_TRUE(s.IsBusy());
+        ASSERT_OK(txn->Rollback());
+        delete txn;
+        continue;
+      }
+
+      auto verify_key = [&owned_keys, &a, &id, &snapshot_num](
+                            const std::string& key, const std::string& value) {
+        if (owned_keys.count(key) > 0) {
+          ASSERT_EQ(value.size(), 16);
+
+          // Since this key is part of owned_keys, then this key must be
+          // unprepared by this transaction identified by 'id'
+          ASSERT_EQ(((int64_t*)value.c_str())[0], id);
+          if (a == REFRESH_SNAPSHOT) {
+            // If refresh snapshot is true, then the snapshot is refreshed
+            // after every Put(), meaning that the current snapshot in
+            // snapshot_num must be greater than the "seqno" of any keys
+            // written by the current transaction.
+            ASSERT_LT(((int64_t*)value.c_str())[1], snapshot_num);
+          } else {
+            // If refresh snapshot is not on, then the snapshot was taken at
+            // the beginning of the transaction, meaning all writes must come
+            // after snapshot_num
+            ASSERT_GT(((int64_t*)value.c_str())[1], snapshot_num);
+          }
+        } else if (a >= RO_SNAPSHOT) {
+          // If this is not an unprepared key, just assert that the key
+          // "seqno" is smaller than the snapshot seqno.
+          ASSERT_EQ(value.size(), 16);
+          ASSERT_LT(((int64_t*)value.c_str())[1], snapshot_num);
+        }
+      };
+
+      // Validate Get()/Next()/Prev(). Do only one of them to save time, and
+      // reduce lock contention.
+      switch (rnd.Uniform(3)) {
+        case 0:  // Validate Get()
+        {
+          for (const auto& key : keys) {
+            std::string value;
+            s = txn->Get(read_options, Slice(key), &value);
+            if (!s.ok()) {
+              ASSERT_TRUE(s.IsNotFound());
+              ASSERT_EQ(owned_keys.count(key), 0);
+            } else {
+              verify_key(key, value);
+            }
+          }
+          break;
+        }
+        case 1:  // Validate Next()
+        {
+          Iterator* iter = txn->GetIterator(read_options);
+          ASSERT_OK(iter->status());
+          for (iter->SeekToFirst(); iter->Valid(); iter->Next()) {
+            verify_key(iter->key().ToString(), iter->value().ToString());
+          }
+          ASSERT_OK(iter->status());
+          delete iter;
+          break;
+        }
+        case 2:  // Validate Prev()
+        {
+          Iterator* iter = txn->GetIterator(read_options);
+          ASSERT_OK(iter->status());
+          for (iter->SeekToLast(); iter->Valid(); iter->Prev()) {
+            verify_key(iter->key().ToString(), iter->value().ToString());
+          }
+          ASSERT_OK(iter->status());
+          delete iter;
+          break;
+        }
+        default:
+          FAIL();
+      }
+
+      if (rnd.OneIn(2)) {
+        ASSERT_OK(txn->Commit());
+      } else {
+        ASSERT_OK(txn->Rollback());
+      }
+      delete txn;
+    }
+  };
+
+  std::vector<port::Thread> threads;
+  for (uint32_t i = 0; i < kNumThreads; i++) {
+    threads.emplace_back(stress_thread, i);
+  }
+
+  for (auto& t : threads) {
+    t.join();
+  }
+}
+#endif  // !defined(ROCKSDB_VALGRIND_RUN) || defined(ROCKSDB_FULL_VALGRIND_RUN)
+
+// This tests how write unprepared behaves during recovery when the DB crashes
+// after a transaction has either been unprepared or prepared, and tests if
+// the changes are correctly applied for prepared transactions if we decide to
+// rollback/commit.
+TEST_P(WriteUnpreparedTransactionTest, RecoveryTest) {
+  WriteOptions write_options;
+  write_options.disableWAL = false;
+  TransactionOptions txn_options;
+  std::vector<Transaction*> prepared_trans;
+  WriteUnpreparedTxnDB* wup_db;
+  options.disable_auto_compactions = true;
+
+  enum Action { UNPREPARED, ROLLBACK, COMMIT };
+
+  // batch_size of 1 causes writes to DB for every marker.
+  for (size_t batch_size : {1, 1000000}) {
+    txn_options.write_batch_flush_threshold = batch_size;
+    for (bool empty : {true, false}) {
+      for (Action a : {UNPREPARED, ROLLBACK, COMMIT}) {
+        for (int num_batches = 1; num_batches < 10; num_batches++) {
+          // Reset database.
+          prepared_trans.clear();
+          ASSERT_OK(ReOpen());
+          wup_db = dynamic_cast<WriteUnpreparedTxnDB*>(db);
+          if (!empty) {
+            for (int i = 0; i < num_batches; i++) {
+              ASSERT_OK(db->Put(WriteOptions(), "k" + std::to_string(i),
+                                "before value" + std::to_string(i)));
+            }
+          }
+
+          // Write num_batches unprepared batches.
+          Transaction* txn = db->BeginTransaction(write_options, txn_options);
+          WriteUnpreparedTxn* wup_txn = dynamic_cast<WriteUnpreparedTxn*>(txn);
+          ASSERT_OK(txn->SetName("xid"));
+          for (int i = 0; i < num_batches; i++) {
+            ASSERT_OK(
+                txn->Put("k" + std::to_string(i), "value" + std::to_string(i)));
+            if (txn_options.write_batch_flush_threshold == 1) {
+              // WriteUnprepared will check write_batch_flush_threshold and
+              // possibly flush before appending to the write batch. No flush
+              // will happen at the first write because the batch is still
+              // empty, so after k puts, there should be k-1 flushed batches.
+              ASSERT_EQ(wup_txn->GetUnpreparedSequenceNumbers().size(), i);
+            } else {
+              ASSERT_EQ(wup_txn->GetUnpreparedSequenceNumbers().size(), 0);
+            }
+          }
+          if (a == UNPREPARED) {
+            // This is done to prevent the destructor from rolling back the
+            // transaction for us, since we want to pretend we crashed and
+            // test that recovery does the rollback.
+            wup_txn->unprep_seqs_.clear();
+          } else {
+            ASSERT_OK(txn->Prepare());
+          }
+          delete txn;
+
+          // Crash and run recovery code paths.
+          ASSERT_OK(wup_db->db_impl_->FlushWAL(true));
+          wup_db->TEST_Crash();
+          ASSERT_OK(ReOpenNoDelete());
+          assert(db != nullptr);
+
+          db->GetAllPreparedTransactions(&prepared_trans);
+          ASSERT_EQ(prepared_trans.size(), a == UNPREPARED ? 0 : 1);
+          if (a == ROLLBACK) {
+            ASSERT_OK(prepared_trans[0]->Rollback());
+            delete prepared_trans[0];
+          } else if (a == COMMIT) {
+            ASSERT_OK(prepared_trans[0]->Commit());
+            delete prepared_trans[0];
+          }
+
+          Iterator* iter = db->NewIterator(ReadOptions());
+          ASSERT_OK(iter->status());
+          iter->SeekToFirst();
+          // Check that DB has before values.
+          if (!empty || a == COMMIT) {
+            for (int i = 0; i < num_batches; i++) {
+              ASSERT_TRUE(iter->Valid());
+              ASSERT_EQ(iter->key().ToString(), "k" + std::to_string(i));
+              if (a == COMMIT) {
+                ASSERT_EQ(iter->value().ToString(),
+                          "value" + std::to_string(i));
+              } else {
+                ASSERT_EQ(iter->value().ToString(),
+                          "before value" + std::to_string(i));
+              }
+              iter->Next();
+            }
+          }
+          ASSERT_FALSE(iter->Valid());
+          ASSERT_OK(iter->status());
+          delete iter;
+        }
+      }
+    }
+  }
+}
+
+// Basic test to see that unprepared batch gets written to DB when batch size
+// is exceeded. It also does some basic checks to see if commit/rollback works
+// as expected for write unprepared.
+TEST_P(WriteUnpreparedTransactionTest, UnpreparedBatch) {
+  WriteOptions write_options;
+  TransactionOptions txn_options;
+  const int kNumKeys = 10;
+
+  // batch_size of 1 causes writes to DB for every marker.
+  for (size_t batch_size : {1, 1000000}) {
+    txn_options.write_batch_flush_threshold = batch_size;
+    for (bool prepare : {false, true}) {
+      for (bool commit : {false, true}) {
+        ASSERT_OK(ReOpen());
+        Transaction* txn = db->BeginTransaction(write_options, txn_options);
+        WriteUnpreparedTxn* wup_txn = dynamic_cast<WriteUnpreparedTxn*>(txn);
+        ASSERT_OK(txn->SetName("xid"));
+
+        for (int i = 0; i < kNumKeys; i++) {
+          ASSERT_OK(txn->Put("k" + std::to_string(i), "v" + std::to_string(i)));
+          if (txn_options.write_batch_flush_threshold == 1) {
+            // WriteUnprepared will check write_batch_flush_threshold and
+            // possibly flush before appending to the write batch. No flush will
+            // happen at the first write because the batch is still empty, so
+            // after k puts, there should be k-1 flushed batches.
+            ASSERT_EQ(wup_txn->GetUnpreparedSequenceNumbers().size(), i);
+          } else {
+            ASSERT_EQ(wup_txn->GetUnpreparedSequenceNumbers().size(), 0);
+          }
+        }
+
+        if (prepare) {
+          ASSERT_OK(txn->Prepare());
+        }
+
+        Iterator* iter = db->NewIterator(ReadOptions());
+        ASSERT_OK(iter->status());
+        iter->SeekToFirst();
+        assert(!iter->Valid());
+        ASSERT_FALSE(iter->Valid());
+        ASSERT_OK(iter->status());
+        delete iter;
+
+        if (commit) {
+          ASSERT_OK(txn->Commit());
+        } else {
+          ASSERT_OK(txn->Rollback());
+        }
+        delete txn;
+
+        iter = db->NewIterator(ReadOptions());
+        ASSERT_OK(iter->status());
+        iter->SeekToFirst();
+
+        for (int i = 0; i < (commit ? kNumKeys : 0); i++) {
+          ASSERT_TRUE(iter->Valid());
+          ASSERT_EQ(iter->key().ToString(), "k" + std::to_string(i));
+          ASSERT_EQ(iter->value().ToString(), "v" + std::to_string(i));
+          iter->Next();
+        }
+        ASSERT_FALSE(iter->Valid());
+        ASSERT_OK(iter->status());
+        delete iter;
+      }
+    }
+  }
+}
+
+// Test whether logs containing unprepared/prepared batches are kept even
+// after memtable finishes flushing, and whether they are removed when
+// transaction commits/aborts.
+//
+// TODO(lth): Merge with TransactionTest/TwoPhaseLogRollingTest tests.
+TEST_P(WriteUnpreparedTransactionTest, MarkLogWithPrepSection) {
+  WriteOptions write_options;
+  TransactionOptions txn_options;
+  // batch_size of 1 causes writes to DB for every marker.
+  txn_options.write_batch_flush_threshold = 1;
+  const int kNumKeys = 10;
+
+  WriteOptions wopts;
+  wopts.sync = true;
+
+  for (bool prepare : {false, true}) {
+    for (bool commit : {false, true}) {
+      ASSERT_OK(ReOpen());
+      auto wup_db = dynamic_cast<WriteUnpreparedTxnDB*>(db);
+      auto db_impl = wup_db->db_impl_;
+
+      Transaction* txn1 = db->BeginTransaction(write_options, txn_options);
+      ASSERT_OK(txn1->SetName("xid1"));
+
+      Transaction* txn2 = db->BeginTransaction(write_options, txn_options);
+      ASSERT_OK(txn2->SetName("xid2"));
+
+      // Spread this transaction across multiple log files.
+      for (int i = 0; i < kNumKeys; i++) {
+        ASSERT_OK(txn1->Put("k1" + std::to_string(i), "v" + std::to_string(i)));
+        if (i >= kNumKeys / 2) {
+          ASSERT_OK(
+              txn2->Put("k2" + std::to_string(i), "v" + std::to_string(i)));
+        }
+
+        if (i > 0) {
+          ASSERT_OK(db_impl->TEST_SwitchWAL());
+        }
+      }
+
+      ASSERT_GT(txn1->GetLogNumber(), 0);
+      ASSERT_GT(txn2->GetLogNumber(), 0);
+
+      ASSERT_EQ(db_impl->TEST_FindMinLogContainingOutstandingPrep(),
+                txn1->GetLogNumber());
+      ASSERT_GT(db_impl->TEST_LogfileNumber(), txn1->GetLogNumber());
+
+      if (prepare) {
+        ASSERT_OK(txn1->Prepare());
+        ASSERT_OK(txn2->Prepare());
+      }
+
+      ASSERT_GE(db_impl->TEST_LogfileNumber(), txn1->GetLogNumber());
+      ASSERT_GE(db_impl->TEST_LogfileNumber(), txn2->GetLogNumber());
+
+      ASSERT_EQ(db_impl->TEST_FindMinLogContainingOutstandingPrep(),
+                txn1->GetLogNumber());
+      if (commit) {
+        ASSERT_OK(txn1->Commit());
+      } else {
+        ASSERT_OK(txn1->Rollback());
+      }
+
+      ASSERT_EQ(db_impl->TEST_FindMinLogContainingOutstandingPrep(),
+                txn2->GetLogNumber());
+
+      if (commit) {
+        ASSERT_OK(txn2->Commit());
+      } else {
+        ASSERT_OK(txn2->Rollback());
+      }
+
+      ASSERT_EQ(db_impl->TEST_FindMinLogContainingOutstandingPrep(), 0);
+
+      delete txn1;
+      delete txn2;
+    }
+  }
+}
+
+TEST_P(WriteUnpreparedTransactionTest, NoSnapshotWrite) {
+  WriteOptions woptions;
+  TransactionOptions txn_options;
+  txn_options.write_batch_flush_threshold = 1;
+
+  Transaction* txn = db->BeginTransaction(woptions, txn_options);
+
+  // Do some writes with no snapshot
+  ASSERT_OK(txn->Put("a", "a"));
+  ASSERT_OK(txn->Put("b", "b"));
+  ASSERT_OK(txn->Put("c", "c"));
+
+  // Test that it is still possible to create iterators after writes with no
+  // snapshot, if iterator snapshot is fresh enough.
+  ReadOptions roptions;
+  auto iter = txn->GetIterator(roptions);
+  ASSERT_OK(iter->status());
+  int keys = 0;
+  for (iter->SeekToLast(); iter->Valid(); iter->Prev(), keys++) {
+    ASSERT_OK(iter->status());
+    ASSERT_EQ(iter->key().ToString(), iter->value().ToString());
+  }
+  ASSERT_EQ(keys, 3);
+  ASSERT_OK(iter->status());
+
+  delete iter;
+  delete txn;
+}
+
+// Test whether write to a transaction while iterating is supported.
+TEST_P(WriteUnpreparedTransactionTest, IterateAndWrite) {
+  WriteOptions woptions;
+  TransactionOptions txn_options;
+  txn_options.write_batch_flush_threshold = 1;
+
+  enum Action { DO_DELETE, DO_UPDATE };
+
+  for (Action a : {DO_DELETE, DO_UPDATE}) {
+    for (int i = 0; i < 100; i++) {
+      ASSERT_OK(db->Put(woptions, std::to_string(i), std::to_string(i)));
+    }
+
+    Transaction* txn = db->BeginTransaction(woptions, txn_options);
+    // write_batch_ now contains 1 key.
+    ASSERT_OK(txn->Put("9", "a"));
+
+    ReadOptions roptions;
+    auto iter = txn->GetIterator(roptions);
+    ASSERT_OK(iter->status());
+    for (iter->SeekToFirst(); iter->Valid(); iter->Next()) {
+      ASSERT_OK(iter->status());
+      if (iter->key() == "9") {
+        ASSERT_EQ(iter->value().ToString(), "a");
+      } else {
+        ASSERT_EQ(iter->key().ToString(), iter->value().ToString());
+      }
+
+      if (a == DO_DELETE) {
+        ASSERT_OK(txn->Delete(iter->key()));
+      } else {
+        ASSERT_OK(txn->Put(iter->key(), "b"));
+      }
+    }
+    ASSERT_OK(iter->status());
+
+    delete iter;
+    ASSERT_OK(txn->Commit());
+
+    iter = db->NewIterator(roptions);
+    ASSERT_OK(iter->status());
+    if (a == DO_DELETE) {
+      // Check that db is empty.
+      iter->SeekToFirst();
+      ASSERT_FALSE(iter->Valid());
+    } else {
+      int keys = 0;
+      // Check that all values are updated to b.
+      for (iter->SeekToFirst(); iter->Valid(); iter->Next(), keys++) {
+        ASSERT_OK(iter->status());
+        ASSERT_EQ(iter->value().ToString(), "b");
+      }
+      ASSERT_EQ(keys, 100);
+    }
+    ASSERT_OK(iter->status());
+
+    delete iter;
+    delete txn;
+  }
+}
+
+// Test that using an iterator after transaction clear is not supported
+TEST_P(WriteUnpreparedTransactionTest, IterateAfterClear) {
+  WriteOptions woptions;
+  TransactionOptions txn_options;
+  txn_options.write_batch_flush_threshold = 1;
+
+  enum Action { kCommit, kRollback };
+
+  for (Action a : {kCommit, kRollback}) {
+    for (int i = 0; i < 100; i++) {
+      ASSERT_OK(db->Put(woptions, std::to_string(i), std::to_string(i)));
+    }
+
+    Transaction* txn = db->BeginTransaction(woptions, txn_options);
+    ASSERT_OK(txn->Put("9", "a"));
+
+    ReadOptions roptions;
+    auto iter1 = txn->GetIterator(roptions);
+    auto iter2 = txn->GetIterator(roptions);
+    iter1->SeekToFirst();
+    iter2->Seek("9");
+
+    // Check that iterators are valid before transaction finishes.
+    ASSERT_TRUE(iter1->Valid());
+    ASSERT_TRUE(iter2->Valid());
+    ASSERT_OK(iter1->status());
+    ASSERT_OK(iter2->status());
+
+    if (a == kCommit) {
+      ASSERT_OK(txn->Commit());
+    } else {
+      ASSERT_OK(txn->Rollback());
+    }
+
+    // Check that iterators are invalidated after transaction finishes.
+    ASSERT_FALSE(iter1->Valid());
+    ASSERT_FALSE(iter2->Valid());
+    ASSERT_TRUE(iter1->status().IsInvalidArgument());
+    ASSERT_TRUE(iter2->status().IsInvalidArgument());
+
+    delete iter1;
+    delete iter2;
+    delete txn;
+  }
+}
+
+TEST_P(WriteUnpreparedTransactionTest, SavePoint) {
+  WriteOptions woptions;
+  TransactionOptions txn_options;
+  txn_options.write_batch_flush_threshold = 1;
+
+  Transaction* txn = db->BeginTransaction(woptions, txn_options);
+  txn->SetSavePoint();
+  ASSERT_OK(txn->Put("a", "a"));
+  ASSERT_OK(txn->Put("b", "b"));
+  ASSERT_OK(txn->Commit());
+
+  ReadOptions roptions;
+  std::string value;
+  ASSERT_OK(txn->Get(roptions, "a", &value));
+  ASSERT_EQ(value, "a");
+  ASSERT_OK(txn->Get(roptions, "b", &value));
+  ASSERT_EQ(value, "b");
+  delete txn;
+}
+
+TEST_P(WriteUnpreparedTransactionTest, UntrackedKeys) {
+  WriteOptions woptions;
+  TransactionOptions txn_options;
+  txn_options.write_batch_flush_threshold = 1;
+
+  Transaction* txn = db->BeginTransaction(woptions, txn_options);
+  auto wb = txn->GetWriteBatch()->GetWriteBatch();
+  ASSERT_OK(txn->Put("a", "a"));
+  ASSERT_OK(wb->Put("a_untrack", "a_untrack"));
+  txn->SetSavePoint();
+  ASSERT_OK(txn->Put("b", "b"));
+  ASSERT_OK(txn->Put("b_untrack", "b_untrack"));
+
+  ReadOptions roptions;
+  std::string value;
+  ASSERT_OK(txn->Get(roptions, "a", &value));
+  ASSERT_EQ(value, "a");
+  ASSERT_OK(txn->Get(roptions, "a_untrack", &value));
+  ASSERT_EQ(value, "a_untrack");
+  ASSERT_OK(txn->Get(roptions, "b", &value));
+  ASSERT_EQ(value, "b");
+  ASSERT_OK(txn->Get(roptions, "b_untrack", &value));
+  ASSERT_EQ(value, "b_untrack");
+
+  // b and b_untrack should be rolled back.
+  ASSERT_OK(txn->RollbackToSavePoint());
+  ASSERT_OK(txn->Get(roptions, "a", &value));
+  ASSERT_EQ(value, "a");
+  ASSERT_OK(txn->Get(roptions, "a_untrack", &value));
+  ASSERT_EQ(value, "a_untrack");
+  auto s = txn->Get(roptions, "b", &value);
+  ASSERT_TRUE(s.IsNotFound());
+  s = txn->Get(roptions, "b_untrack", &value);
+  ASSERT_TRUE(s.IsNotFound());
+
+  // Everything should be rolled back.
+  ASSERT_OK(txn->Rollback());
+  s = txn->Get(roptions, "a", &value);
+  ASSERT_TRUE(s.IsNotFound());
+  s = txn->Get(roptions, "a_untrack", &value);
+  ASSERT_TRUE(s.IsNotFound());
+  s = txn->Get(roptions, "b", &value);
+  ASSERT_TRUE(s.IsNotFound());
+  s = txn->Get(roptions, "b_untrack", &value);
+  ASSERT_TRUE(s.IsNotFound());
+
+  delete txn;
+}
+
+}  // namespace ROCKSDB_NAMESPACE
+
+int main(int argc, char** argv) {
+  ROCKSDB_NAMESPACE::port::InstallStackTraceHandler();
+  ::testing::InitGoogleTest(&argc, argv);
+  return RUN_ALL_TESTS();
+}
+
+#else
+#include <stdio.h>
+
+int main(int /*argc*/, char** /*argv*/) {
+  fprintf(stderr,
+          "SKIPPED as Transactions are not supported in ROCKSDB_LITE\n");
+  return 0;
+}
+
+#endif  // ROCKSDB_LITE
diff --git a/src/rocksdb/utilities/transactions/write_unprepared_txn.cc b/src/rocksdb/utilities/transactions/write_unprepared_txn.cc
new file mode 100644
index 000000000..6e04d3344
--- /dev/null
+++ b/src/rocksdb/utilities/transactions/write_unprepared_txn.cc
@@ -0,0 +1,1053 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#ifndef ROCKSDB_LITE
+
+#include "utilities/transactions/write_unprepared_txn.h"
+
+#include "db/db_impl/db_impl.h"
+#include "util/cast_util.h"
+#include "utilities/transactions/write_unprepared_txn_db.h"
+#include "utilities/write_batch_with_index/write_batch_with_index_internal.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+bool WriteUnpreparedTxnReadCallback::IsVisibleFullCheck(SequenceNumber seq) {
+  // Since unprep_seqs maps prep_seq => prepare_batch_cnt, to check if seq is
+  // in unprep_seqs, we have to check if seq is equal to prep_seq or any of
+  // the prepare_batch_cnt seq nums after it.
+  //
+  // TODO(lth): Can be optimized with std::lower_bound if unprep_seqs is
+  // large.
+  for (const auto& it : unprep_seqs_) {
+    if (it.first <= seq && seq < it.first + it.second) {
+      return true;
+    }
+  }
+
+  bool snap_released = false;
+  auto ret =
+      db_->IsInSnapshot(seq, wup_snapshot_, min_uncommitted_, &snap_released);
+  assert(!snap_released || backed_by_snapshot_ == kUnbackedByDBSnapshot);
+  snap_released_ |= snap_released;
+  return ret;
+}
+
+WriteUnpreparedTxn::WriteUnpreparedTxn(WriteUnpreparedTxnDB* txn_db,
+                                       const WriteOptions& write_options,
+                                       const TransactionOptions& txn_options)
+    : WritePreparedTxn(txn_db, write_options, txn_options),
+      wupt_db_(txn_db),
+      last_log_number_(0),
+      recovered_txn_(false),
+      largest_validated_seq_(0) {
+  if (txn_options.write_batch_flush_threshold < 0) {
+    write_batch_flush_threshold_ =
+        txn_db_impl_->GetTxnDBOptions().default_write_batch_flush_threshold;
+  } else {
+    write_batch_flush_threshold_ = txn_options.write_batch_flush_threshold;
+  }
+}
+
+WriteUnpreparedTxn::~WriteUnpreparedTxn() {
+  if (!unprep_seqs_.empty()) {
+    assert(log_number_ > 0);
+    assert(GetId() > 0);
+    assert(!name_.empty());
+
+    // We should rollback regardless of GetState, but some unit tests that
+    // test crash recovery run the destructor assuming that rollback does not
+    // happen, so that rollback during recovery can be exercised.
+    if (GetState() == STARTED || GetState() == LOCKS_STOLEN) {
+      auto s = RollbackInternal();
+      assert(s.ok());
+      if (!s.ok()) {
+        ROCKS_LOG_FATAL(
+            wupt_db_->info_log_,
+            "Rollback of WriteUnprepared transaction failed in destructor: %s",
+            s.ToString().c_str());
+      }
+      dbimpl_->logs_with_prep_tracker()->MarkLogAsHavingPrepSectionFlushed(
+          log_number_);
+    }
+  }
+
+  // Clear the tracked locks so that ~PessimisticTransaction does not
+  // try to unlock keys for recovered transactions.
+  if (recovered_txn_) {
+    tracked_locks_->Clear();
+  }
+}
+
+void WriteUnpreparedTxn::Initialize(const TransactionOptions& txn_options) {
+  PessimisticTransaction::Initialize(txn_options);
+  if (txn_options.write_batch_flush_threshold < 0) {
+    write_batch_flush_threshold_ =
+        txn_db_impl_->GetTxnDBOptions().default_write_batch_flush_threshold;
+  } else {
+    write_batch_flush_threshold_ = txn_options.write_batch_flush_threshold;
+  }
+
+  unprep_seqs_.clear();
+  flushed_save_points_.reset(nullptr);
+  unflushed_save_points_.reset(nullptr);
+  recovered_txn_ = false;
+  largest_validated_seq_ = 0;
+  assert(active_iterators_.empty());
+  active_iterators_.clear();
+  untracked_keys_.clear();
+}
+
+Status WriteUnpreparedTxn::HandleWrite(std::function<Status()> do_write) {
+  Status s;
+  if (active_iterators_.empty()) {
+    s = MaybeFlushWriteBatchToDB();
+    if (!s.ok()) {
+      return s;
+    }
+  }
+  s = do_write();
+  if (s.ok()) {
+    if (snapshot_) {
+      largest_validated_seq_ =
+          std::max(largest_validated_seq_, snapshot_->GetSequenceNumber());
+    } else {
+      // TODO(lth): We should use the same number as tracked_at_seq in TryLock,
+      // because what is actually being tracked is the sequence number at which
+      // this key was locked at.
+      largest_validated_seq_ = db_impl_->GetLastPublishedSequence();
+    }
+  }
+  return s;
+}
+
+Status WriteUnpreparedTxn::Put(ColumnFamilyHandle* column_family,
+                               const Slice& key, const Slice& value,
+                               const bool assume_tracked) {
+  return HandleWrite([&]() {
+    return TransactionBaseImpl::Put(column_family, key, value, assume_tracked);
+  });
+}
+
+Status WriteUnpreparedTxn::Put(ColumnFamilyHandle* column_family,
+                               const SliceParts& key, const SliceParts& value,
+                               const bool assume_tracked) {
+  return HandleWrite([&]() {
+    return TransactionBaseImpl::Put(column_family, key, value, assume_tracked);
+  });
+}
+
+Status WriteUnpreparedTxn::Merge(ColumnFamilyHandle* column_family,
+                                 const Slice& key, const Slice& value,
+                                 const bool assume_tracked) {
+  return HandleWrite([&]() {
+    return TransactionBaseImpl::Merge(column_family, key, value,
+                                      assume_tracked);
+  });
+}
+
+Status WriteUnpreparedTxn::Delete(ColumnFamilyHandle* column_family,
+                                  const Slice& key, const bool assume_tracked) {
+  return HandleWrite([&]() {
+    return TransactionBaseImpl::Delete(column_family, key, assume_tracked);
+  });
+}
+
+Status WriteUnpreparedTxn::Delete(ColumnFamilyHandle* column_family,
+                                  const SliceParts& key,
+                                  const bool assume_tracked) {
+  return HandleWrite([&]() {
+    return TransactionBaseImpl::Delete(column_family, key, assume_tracked);
+  });
+}
+
+Status WriteUnpreparedTxn::SingleDelete(ColumnFamilyHandle* column_family,
+                                        const Slice& key,
+                                        const bool assume_tracked) {
+  return HandleWrite([&]() {
+    return TransactionBaseImpl::SingleDelete(column_family, key,
+                                             assume_tracked);
+  });
+}
+
+Status WriteUnpreparedTxn::SingleDelete(ColumnFamilyHandle* column_family,
+                                        const SliceParts& key,
+                                        const bool assume_tracked) {
+  return HandleWrite([&]() {
+    return TransactionBaseImpl::SingleDelete(column_family, key,
+                                             assume_tracked);
+  });
+}
+
+// WriteUnpreparedTxn::RebuildFromWriteBatch is only called on recovery. For
+// WriteUnprepared, the write batches have already been written into the
+// database during WAL replay, so all we have to do is just to "retrack" the key
+// so that rollbacks are possible.
+//
+// Calling TryLock instead of TrackKey is also possible, but as an optimization,
+// recovered transactions do not hold locks on their keys. This follows the
+// implementation in PessimisticTransactionDB::Initialize where we set
+// skip_concurrency_control to true.
+Status WriteUnpreparedTxn::RebuildFromWriteBatch(WriteBatch* wb) {
+  struct TrackKeyHandler : public WriteBatch::Handler {
+    WriteUnpreparedTxn* txn_;
+    bool rollback_merge_operands_;
+
+    TrackKeyHandler(WriteUnpreparedTxn* txn, bool rollback_merge_operands)
+        : txn_(txn), rollback_merge_operands_(rollback_merge_operands) {}
+
+    Status PutCF(uint32_t cf, const Slice& key, const Slice&) override {
+      txn_->TrackKey(cf, key.ToString(), kMaxSequenceNumber,
+                     false /* read_only */, true /* exclusive */);
+      return Status::OK();
+    }
+
+    Status DeleteCF(uint32_t cf, const Slice& key) override {
+      txn_->TrackKey(cf, key.ToString(), kMaxSequenceNumber,
+                     false /* read_only */, true /* exclusive */);
+      return Status::OK();
+    }
+
+    Status SingleDeleteCF(uint32_t cf, const Slice& key) override {
+      txn_->TrackKey(cf, key.ToString(), kMaxSequenceNumber,
+                     false /* read_only */, true /* exclusive */);
+      return Status::OK();
+    }
+
+    Status MergeCF(uint32_t cf, const Slice& key, const Slice&) override {
+      if (rollback_merge_operands_) {
+        txn_->TrackKey(cf, key.ToString(), kMaxSequenceNumber,
+                       false /* read_only */, true /* exclusive */);
+      }
+      return Status::OK();
+    }
+
+    // Recovered batches do not contain 2PC markers.
+    Status MarkBeginPrepare(bool) override { return Status::InvalidArgument(); }
+
+    Status MarkEndPrepare(const Slice&) override {
+      return Status::InvalidArgument();
+    }
+
+    Status MarkNoop(bool) override { return Status::InvalidArgument(); }
+
+    Status MarkCommit(const Slice&) override {
+      return Status::InvalidArgument();
+    }
+
+    Status MarkRollback(const Slice&) override {
+      return Status::InvalidArgument();
+    }
+  };
+
+  TrackKeyHandler handler(this,
+                          wupt_db_->txn_db_options_.rollback_merge_operands);
+  return wb->Iterate(&handler);
+}
+
+Status WriteUnpreparedTxn::MaybeFlushWriteBatchToDB() {
+  const bool kPrepared = true;
+  Status s;
+  if (write_batch_flush_threshold_ > 0 &&
+      write_batch_.GetWriteBatch()->Count() > 0 &&
+      write_batch_.GetDataSize() >
+          static_cast<size_t>(write_batch_flush_threshold_)) {
+    assert(GetState() != PREPARED);
+    s = FlushWriteBatchToDB(!kPrepared);
+  }
+  return s;
+}
+
+Status WriteUnpreparedTxn::FlushWriteBatchToDB(bool prepared) {
+  // If the current write batch contains savepoints, then some special handling
+  // is required so that RollbackToSavepoint can work.
+  //
+  // RollbackToSavepoint is not supported after Prepare() is called, so only do
+  // this for unprepared batches.
+  if (!prepared && unflushed_save_points_ != nullptr &&
+      !unflushed_save_points_->empty()) {
+    return FlushWriteBatchWithSavePointToDB();
+  }
+
+  return FlushWriteBatchToDBInternal(prepared);
+}
+
+Status WriteUnpreparedTxn::FlushWriteBatchToDBInternal(bool prepared) {
+  if (name_.empty()) {
+    assert(!prepared);
+#ifndef NDEBUG
+    static std::atomic_ullong autogen_id{0};
+    // To avoid changing all tests to call SetName, just autogenerate one.
+    if (wupt_db_->txn_db_options_.autogenerate_name) {
+      auto s = SetName(std::string("autoxid") +
+                       std::to_string(autogen_id.fetch_add(1)));
+      assert(s.ok());
+    } else
+#endif
+    {
+      return Status::InvalidArgument("Cannot write to DB without SetName.");
+    }
+  }
+
+  struct UntrackedKeyHandler : public WriteBatch::Handler {
+    WriteUnpreparedTxn* txn_;
+    bool rollback_merge_operands_;
+
+    UntrackedKeyHandler(WriteUnpreparedTxn* txn, bool rollback_merge_operands)
+        : txn_(txn), rollback_merge_operands_(rollback_merge_operands) {}
+
+    Status AddUntrackedKey(uint32_t cf, const Slice& key) {
+      auto str = key.ToString();
+      PointLockStatus lock_status =
+          txn_->tracked_locks_->GetPointLockStatus(cf, str);
+      if (!lock_status.locked) {
+        txn_->untracked_keys_[cf].push_back(str);
+      }
+      return Status::OK();
+    }
+
+    Status PutCF(uint32_t cf, const Slice& key, const Slice&) override {
+      return AddUntrackedKey(cf, key);
+    }
+
+    Status DeleteCF(uint32_t cf, const Slice& key) override {
+      return AddUntrackedKey(cf, key);
+    }
+
+    Status SingleDeleteCF(uint32_t cf, const Slice& key) override {
+      return AddUntrackedKey(cf, key);
+    }
+
+    Status MergeCF(uint32_t cf, const Slice& key, const Slice&) override {
+      if (rollback_merge_operands_) {
+        return AddUntrackedKey(cf, key);
+      }
+      return Status::OK();
+    }
+
+    // The only expected 2PC marker is the initial Noop marker.
+    Status MarkNoop(bool empty_batch) override {
+      return empty_batch ? Status::OK() : Status::InvalidArgument();
+    }
+
+    Status MarkBeginPrepare(bool) override { return Status::InvalidArgument(); }
+
+    Status MarkEndPrepare(const Slice&) override {
+      return Status::InvalidArgument();
+    }
+
+    Status MarkCommit(const Slice&) override {
+      return Status::InvalidArgument();
+    }
+
+    Status MarkRollback(const Slice&) override {
+      return Status::InvalidArgument();
+    }
+  };
+
+  UntrackedKeyHandler handler(
+      this, wupt_db_->txn_db_options_.rollback_merge_operands);
+  auto s = GetWriteBatch()->GetWriteBatch()->Iterate(&handler);
+  assert(s.ok());
+
+  // TODO(lth): Reduce duplicate code with WritePrepared prepare logic.
+  WriteOptions write_options = write_options_;
+  write_options.disableWAL = false;
+  const bool WRITE_AFTER_COMMIT = true;
+  const bool first_prepare_batch = log_number_ == 0;
+  // MarkEndPrepare will change Noop marker to the appropriate marker.
+  s = WriteBatchInternal::MarkEndPrepare(GetWriteBatch()->GetWriteBatch(),
+                                         name_, !WRITE_AFTER_COMMIT, !prepared);
+  assert(s.ok());
+  // For each duplicate key we account for a new sub-batch
+  prepare_batch_cnt_ = GetWriteBatch()->SubBatchCnt();
+  // AddPrepared better to be called in the pre-release callback otherwise there
+  // is a non-zero chance of max advancing prepare_seq and readers assume the
+  // data as committed.
+  // Also having it in the PreReleaseCallback allows in-order addition of
+  // prepared entries to PreparedHeap and hence enables an optimization. Refer
+  // to SmallestUnCommittedSeq for more details.
+  AddPreparedCallback add_prepared_callback(
+      wpt_db_, db_impl_, prepare_batch_cnt_,
+      db_impl_->immutable_db_options().two_write_queues, first_prepare_batch);
+  const bool DISABLE_MEMTABLE = true;
+  uint64_t seq_used = kMaxSequenceNumber;
+  // log_number_ should refer to the oldest log containing uncommitted data
+  // from the current transaction. This means that if log_number_ is set,
+  // WriteImpl should not overwrite that value, so set log_used to nullptr if
+  // log_number_ is already set.
+  s = db_impl_->WriteImpl(write_options, GetWriteBatch()->GetWriteBatch(),
+                          /*callback*/ nullptr, &last_log_number_,
+                          /*log ref*/ 0, !DISABLE_MEMTABLE, &seq_used,
+                          prepare_batch_cnt_, &add_prepared_callback);
+  if (log_number_ == 0) {
+    log_number_ = last_log_number_;
+  }
+  assert(!s.ok() || seq_used != kMaxSequenceNumber);
+  auto prepare_seq = seq_used;
+
+  // Only call SetId if it hasn't been set yet.
+  if (GetId() == 0) {
+    SetId(prepare_seq);
+  }
+  // unprep_seqs_ will also contain prepared seqnos since they are treated in
+  // the same way in the prepare/commit callbacks. See the comment on the
+  // definition of unprep_seqs_.
+  unprep_seqs_[prepare_seq] = prepare_batch_cnt_;
+
+  // Reset transaction state.
+  if (!prepared) {
+    prepare_batch_cnt_ = 0;
+    const bool kClear = true;
+    TransactionBaseImpl::InitWriteBatch(kClear);
+  }
+
+  return s;
+}
+
+Status WriteUnpreparedTxn::FlushWriteBatchWithSavePointToDB() {
+  assert(unflushed_save_points_ != nullptr &&
+         unflushed_save_points_->size() > 0);
+  assert(save_points_ != nullptr && save_points_->size() > 0);
+  assert(save_points_->size() >= unflushed_save_points_->size());
+
+  // Handler class for creating an unprepared batch from a savepoint.
+  struct SavePointBatchHandler : public WriteBatch::Handler {
+    WriteBatchWithIndex* wb_;
+    const std::map<uint32_t, ColumnFamilyHandle*>& handles_;
+
+    SavePointBatchHandler(
+        WriteBatchWithIndex* wb,
+        const std::map<uint32_t, ColumnFamilyHandle*>& handles)
+        : wb_(wb), handles_(handles) {}
+
+    Status PutCF(uint32_t cf, const Slice& key, const Slice& value) override {
+      return wb_->Put(handles_.at(cf), key, value);
+    }
+
+    Status DeleteCF(uint32_t cf, const Slice& key) override {
+      return wb_->Delete(handles_.at(cf), key);
+    }
+
+    Status SingleDeleteCF(uint32_t cf, const Slice& key) override {
+      return wb_->SingleDelete(handles_.at(cf), key);
+    }
+
+    Status MergeCF(uint32_t cf, const Slice& key, const Slice& value) override {
+      return wb_->Merge(handles_.at(cf), key, value);
+    }
+
+    // The only expected 2PC marker is the initial Noop marker.
+    Status MarkNoop(bool empty_batch) override {
+      return empty_batch ? Status::OK() : Status::InvalidArgument();
+    }
+
+    Status MarkBeginPrepare(bool) override { return Status::InvalidArgument(); }
+
+    Status MarkEndPrepare(const Slice&) override {
+      return Status::InvalidArgument();
+    }
+
+    Status MarkCommit(const Slice&) override {
+      return Status::InvalidArgument();
+    }
+
+    Status MarkRollback(const Slice&) override {
+      return Status::InvalidArgument();
+    }
+  };
+
+  // The comparator of the default cf is passed in, similar to the
+  // initialization of TransactionBaseImpl::write_batch_. This comparator is
+  // only used if the write batch encounters an invalid cf id, and falls back to
+  // this comparator.
+  WriteBatchWithIndex wb(wpt_db_->DefaultColumnFamily()->GetComparator(), 0,
+                         true, 0, write_options_.protection_bytes_per_key);
+  // Swap with write_batch_ so that wb contains the complete write batch. The
+  // actual write batch that will be flushed to DB will be built in
+  // write_batch_, and will be read by FlushWriteBatchToDBInternal.
+  std::swap(wb, write_batch_);
+  TransactionBaseImpl::InitWriteBatch();
+
+  size_t prev_boundary = WriteBatchInternal::kHeader;
+  const bool kPrepared = true;
+  for (size_t i = 0; i < unflushed_save_points_->size() + 1; i++) {
+    bool trailing_batch = i == unflushed_save_points_->size();
+    SavePointBatchHandler sp_handler(&write_batch_,
+                                     *wupt_db_->GetCFHandleMap().get());
+    size_t curr_boundary = trailing_batch ? wb.GetWriteBatch()->GetDataSize()
+                                          : (*unflushed_save_points_)[i];
+
+    // Construct the partial write batch up to the savepoint.
+    //
+    // Theoretically, a memcpy between the write batches should be sufficient
+    // since the rewriting into the batch should produce the exact same byte
+    // representation. Rebuilding the WriteBatchWithIndex index is still
+    // necessary though, and would imply doing two passes over the batch though.
+    Status s = WriteBatchInternal::Iterate(wb.GetWriteBatch(), &sp_handler,
+                                           prev_boundary, curr_boundary);
+    if (!s.ok()) {
+      return s;
+    }
+
+    if (write_batch_.GetWriteBatch()->Count() > 0) {
+      // Flush the write batch.
+      s = FlushWriteBatchToDBInternal(!kPrepared);
+      if (!s.ok()) {
+        return s;
+      }
+    }
+
+    if (!trailing_batch) {
+      if (flushed_save_points_ == nullptr) {
+        flushed_save_points_.reset(
+            new autovector<WriteUnpreparedTxn::SavePoint>());
+      }
+      flushed_save_points_->emplace_back(
+          unprep_seqs_, new ManagedSnapshot(db_impl_, wupt_db_->GetSnapshot()));
+    }
+
+    prev_boundary = curr_boundary;
+    const bool kClear = true;
+    TransactionBaseImpl::InitWriteBatch(kClear);
+  }
+
+  unflushed_save_points_->clear();
+  return Status::OK();
+}
+
+Status WriteUnpreparedTxn::PrepareInternal() {
+  const bool kPrepared = true;
+  return FlushWriteBatchToDB(kPrepared);
+}
+
+Status WriteUnpreparedTxn::CommitWithoutPrepareInternal() {
+  if (unprep_seqs_.empty()) {
+    assert(log_number_ == 0);
+    assert(GetId() == 0);
+    return WritePreparedTxn::CommitWithoutPrepareInternal();
+  }
+
+  // TODO(lth): We should optimize commit without prepare to not perform
+  // a prepare under the hood.
+  auto s = PrepareInternal();
+  if (!s.ok()) {
+    return s;
+  }
+  return CommitInternal();
+}
+
+Status WriteUnpreparedTxn::CommitInternal() {
+  // TODO(lth): Reduce duplicate code with WritePrepared commit logic.
+
+  // We take the commit-time batch and append the Commit marker.  The Memtable
+  // will ignore the Commit marker in non-recovery mode
+  WriteBatch* working_batch = GetCommitTimeWriteBatch();
+  const bool empty = working_batch->Count() == 0;
+  auto s = WriteBatchInternal::MarkCommit(working_batch, name_);
+  assert(s.ok());
+
+  const bool for_recovery = use_only_the_last_commit_time_batch_for_recovery_;
+  if (!empty) {
+    // When not writing to memtable, we can still cache the latest write batch.
+    // The cached batch will be written to memtable in WriteRecoverableState
+    // during FlushMemTable
+    if (for_recovery) {
+      WriteBatchInternal::SetAsLatestPersistentState(working_batch);
+    } else {
+      return Status::InvalidArgument(
+          "Commit-time-batch can only be used if "
+          "use_only_the_last_commit_time_batch_for_recovery is true");
+    }
+  }
+
+  const bool includes_data = !empty && !for_recovery;
+  size_t commit_batch_cnt = 0;
+  if (UNLIKELY(includes_data)) {
+    ROCKS_LOG_WARN(db_impl_->immutable_db_options().info_log,
+                   "Duplicate key overhead");
+    SubBatchCounter counter(*wpt_db_->GetCFComparatorMap());
+    s = working_batch->Iterate(&counter);
+    assert(s.ok());
+    commit_batch_cnt = counter.BatchCount();
+  }
+  const bool disable_memtable = !includes_data;
+  const bool do_one_write =
+      !db_impl_->immutable_db_options().two_write_queues || disable_memtable;
+
+  WriteUnpreparedCommitEntryPreReleaseCallback update_commit_map(
+      wpt_db_, db_impl_, unprep_seqs_, commit_batch_cnt);
+  const bool kFirstPrepareBatch = true;
+  AddPreparedCallback add_prepared_callback(
+      wpt_db_, db_impl_, commit_batch_cnt,
+      db_impl_->immutable_db_options().two_write_queues, !kFirstPrepareBatch);
+  PreReleaseCallback* pre_release_callback;
+  if (do_one_write) {
+    pre_release_callback = &update_commit_map;
+  } else {
+    pre_release_callback = &add_prepared_callback;
+  }
+  uint64_t seq_used = kMaxSequenceNumber;
+  // Since the prepared batch is directly written to memtable, there is
+  // already a connection between the memtable and its WAL, so there is no
+  // need to redundantly reference the log that contains the prepared data.
+  const uint64_t zero_log_number = 0ull;
+  size_t batch_cnt = UNLIKELY(commit_batch_cnt) ? commit_batch_cnt : 1;
+  s = db_impl_->WriteImpl(write_options_, working_batch, nullptr, nullptr,
+                          zero_log_number, disable_memtable, &seq_used,
+                          batch_cnt, pre_release_callback);
+  assert(!s.ok() || seq_used != kMaxSequenceNumber);
+  const SequenceNumber commit_batch_seq = seq_used;
+  if (LIKELY(do_one_write || !s.ok())) {
+    if (LIKELY(s.ok())) {
+      // Note RemovePrepared should be called after WriteImpl that publishsed
+      // the seq. Otherwise SmallestUnCommittedSeq optimization breaks.
+      for (const auto& seq : unprep_seqs_) {
+        wpt_db_->RemovePrepared(seq.first, seq.second);
+      }
+    }
+    if (UNLIKELY(!do_one_write)) {
+      wpt_db_->RemovePrepared(commit_batch_seq, commit_batch_cnt);
+    }
+    unprep_seqs_.clear();
+    flushed_save_points_.reset(nullptr);
+    unflushed_save_points_.reset(nullptr);
+    return s;
+  }  // else do the 2nd write to publish seq
+
+  // Populate unprep_seqs_ with commit_batch_seq, since we treat data in the
+  // commit write batch as just another "unprepared" batch. This will also
+  // update the unprep_seqs_ in the update_commit_map callback.
+  unprep_seqs_[commit_batch_seq] = commit_batch_cnt;
+  WriteUnpreparedCommitEntryPreReleaseCallback
+      update_commit_map_with_commit_batch(wpt_db_, db_impl_, unprep_seqs_, 0);
+
+  // Note: the 2nd write comes with a performance penality. So if we have too
+  // many of commits accompanied with ComitTimeWriteBatch and yet we cannot
+  // enable use_only_the_last_commit_time_batch_for_recovery_ optimization,
+  // two_write_queues should be disabled to avoid many additional writes here.
+
+  // Update commit map only from the 2nd queue
+  WriteBatch empty_batch;
+  s = empty_batch.PutLogData(Slice());
+  assert(s.ok());
+  // In the absence of Prepare markers, use Noop as a batch separator
+  s = WriteBatchInternal::InsertNoop(&empty_batch);
+  assert(s.ok());
+  const bool DISABLE_MEMTABLE = true;
+  const size_t ONE_BATCH = 1;
+  const uint64_t NO_REF_LOG = 0;
+  s = db_impl_->WriteImpl(write_options_, &empty_batch, nullptr, nullptr,
+                          NO_REF_LOG, DISABLE_MEMTABLE, &seq_used, ONE_BATCH,
+                          &update_commit_map_with_commit_batch);
+  assert(!s.ok() || seq_used != kMaxSequenceNumber);
+  // Note RemovePrepared should be called after WriteImpl that publishsed the
+  // seq. Otherwise SmallestUnCommittedSeq optimization breaks.
+  for (const auto& seq : unprep_seqs_) {
+    wpt_db_->RemovePrepared(seq.first, seq.second);
+  }
+  unprep_seqs_.clear();
+  flushed_save_points_.reset(nullptr);
+  unflushed_save_points_.reset(nullptr);
+  return s;
+}
+
+Status WriteUnpreparedTxn::WriteRollbackKeys(
+    const LockTracker& lock_tracker, WriteBatchWithIndex* rollback_batch,
+    ReadCallback* callback, const ReadOptions& roptions) {
+  // This assertion can be removed when range lock is supported.
+  assert(lock_tracker.IsPointLockSupported());
+  const auto& cf_map = *wupt_db_->GetCFHandleMap();
+  auto WriteRollbackKey = [&](const std::string& key, uint32_t cfid) {
+    const auto& cf_handle = cf_map.at(cfid);
+    PinnableSlice pinnable_val;
+    bool not_used;
+    DBImpl::GetImplOptions get_impl_options;
+    get_impl_options.column_family = cf_handle;
+    get_impl_options.value = &pinnable_val;
+    get_impl_options.value_found = &not_used;
+    get_impl_options.callback = callback;
+    auto s = db_impl_->GetImpl(roptions, key, get_impl_options);
+
+    if (s.ok()) {
+      s = rollback_batch->Put(cf_handle, key, pinnable_val);
+      assert(s.ok());
+    } else if (s.IsNotFound()) {
+      if (wupt_db_->ShouldRollbackWithSingleDelete(cf_handle, key)) {
+        s = rollback_batch->SingleDelete(cf_handle, key);
+      } else {
+        s = rollback_batch->Delete(cf_handle, key);
+      }
+      assert(s.ok());
+    } else {
+      return s;
+    }
+
+    return Status::OK();
+  };
+
+  std::unique_ptr<LockTracker::ColumnFamilyIterator> cf_it(
+      lock_tracker.GetColumnFamilyIterator());
+  assert(cf_it != nullptr);
+  while (cf_it->HasNext()) {
+    ColumnFamilyId cf = cf_it->Next();
+    std::unique_ptr<LockTracker::KeyIterator> key_it(
+        lock_tracker.GetKeyIterator(cf));
+    assert(key_it != nullptr);
+    while (key_it->HasNext()) {
+      const std::string& key = key_it->Next();
+      auto s = WriteRollbackKey(key, cf);
+      if (!s.ok()) {
+        return s;
+      }
+    }
+  }
+
+  for (const auto& cfkey : untracked_keys_) {
+    const auto cfid = cfkey.first;
+    const auto& keys = cfkey.second;
+    for (const auto& key : keys) {
+      auto s = WriteRollbackKey(key, cfid);
+      if (!s.ok()) {
+        return s;
+      }
+    }
+  }
+
+  return Status::OK();
+}
+
+Status WriteUnpreparedTxn::RollbackInternal() {
+  // TODO(lth): Reduce duplicate code with WritePrepared rollback logic.
+  WriteBatchWithIndex rollback_batch(
+      wpt_db_->DefaultColumnFamily()->GetComparator(), 0, true, 0,
+      write_options_.protection_bytes_per_key);
+  assert(GetId() != kMaxSequenceNumber);
+  assert(GetId() > 0);
+  Status s;
+  auto read_at_seq = kMaxSequenceNumber;
+  ReadOptions roptions;
+  // to prevent callback's seq to be overrriden inside DBImpk::Get
+  roptions.snapshot = wpt_db_->GetMaxSnapshot();
+  // Note that we do not use WriteUnpreparedTxnReadCallback because we do not
+  // need to read our own writes when reading prior versions of the key for
+  // rollback.
+  WritePreparedTxnReadCallback callback(wpt_db_, read_at_seq);
+  // TODO(lth): We write rollback batch all in a single batch here, but this
+  // should be subdivded into multiple batches as well. In phase 2, when key
+  // sets are read from WAL, this will happen naturally.
+  s = WriteRollbackKeys(*tracked_locks_, &rollback_batch, &callback, roptions);
+  if (!s.ok()) {
+    return s;
+  }
+
+  // The Rollback marker will be used as a batch separator
+  s = WriteBatchInternal::MarkRollback(rollback_batch.GetWriteBatch(), name_);
+  assert(s.ok());
+  bool do_one_write = !db_impl_->immutable_db_options().two_write_queues;
+  const bool DISABLE_MEMTABLE = true;
+  const uint64_t NO_REF_LOG = 0;
+  uint64_t seq_used = kMaxSequenceNumber;
+  // Rollback batch may contain duplicate keys, because tracked_keys_ is not
+  // comparator aware.
+  auto rollback_batch_cnt = rollback_batch.SubBatchCnt();
+  // We commit the rolled back prepared batches. Although this is
+  // counter-intuitive, i) it is safe to do so, since the prepared batches are
+  // already canceled out by the rollback batch, ii) adding the commit entry to
+  // CommitCache will allow us to benefit from the existing mechanism in
+  // CommitCache that keeps an entry evicted due to max advance and yet overlaps
+  // with a live snapshot around so that the live snapshot properly skips the
+  // entry even if its prepare seq is lower than max_evicted_seq_.
+  //
+  // TODO(lth): RollbackInternal is conceptually very similar to
+  // CommitInternal, with the rollback batch simply taking on the role of
+  // CommitTimeWriteBatch. We should be able to merge the two code paths.
+  WriteUnpreparedCommitEntryPreReleaseCallback update_commit_map(
+      wpt_db_, db_impl_, unprep_seqs_, rollback_batch_cnt);
+  // Note: the rollback batch does not need AddPrepared since it is written to
+  // DB in one shot. min_uncommitted still works since it requires capturing
+  // data that is written to DB but not yet committed, while the rollback
+  // batch commits with PreReleaseCallback.
+  s = db_impl_->WriteImpl(write_options_, rollback_batch.GetWriteBatch(),
+                          nullptr, nullptr, NO_REF_LOG, !DISABLE_MEMTABLE,
+                          &seq_used, rollback_batch_cnt,
+                          do_one_write ? &update_commit_map : nullptr);
+  assert(!s.ok() || seq_used != kMaxSequenceNumber);
+  if (!s.ok()) {
+    return s;
+  }
+  if (do_one_write) {
+    for (const auto& seq : unprep_seqs_) {
+      wpt_db_->RemovePrepared(seq.first, seq.second);
+    }
+    unprep_seqs_.clear();
+    flushed_save_points_.reset(nullptr);
+    unflushed_save_points_.reset(nullptr);
+    return s;
+  }  // else do the 2nd write for commit
+
+  uint64_t& prepare_seq = seq_used;
+  // Populate unprep_seqs_ with rollback_batch_cnt, since we treat data in the
+  // rollback write batch as just another "unprepared" batch. This will also
+  // update the unprep_seqs_ in the update_commit_map callback.
+  unprep_seqs_[prepare_seq] = rollback_batch_cnt;
+  WriteUnpreparedCommitEntryPreReleaseCallback
+      update_commit_map_with_rollback_batch(wpt_db_, db_impl_, unprep_seqs_, 0);
+
+  ROCKS_LOG_DETAILS(db_impl_->immutable_db_options().info_log,
+                    "RollbackInternal 2nd write prepare_seq: %" PRIu64,
+                    prepare_seq);
+  WriteBatch empty_batch;
+  const size_t ONE_BATCH = 1;
+  s = empty_batch.PutLogData(Slice());
+  assert(s.ok());
+  // In the absence of Prepare markers, use Noop as a batch separator
+  s = WriteBatchInternal::InsertNoop(&empty_batch);
+  assert(s.ok());
+  s = db_impl_->WriteImpl(write_options_, &empty_batch, nullptr, nullptr,
+                          NO_REF_LOG, DISABLE_MEMTABLE, &seq_used, ONE_BATCH,
+                          &update_commit_map_with_rollback_batch);
+  assert(!s.ok() || seq_used != kMaxSequenceNumber);
+  // Mark the txn as rolled back
+  if (s.ok()) {
+    for (const auto& seq : unprep_seqs_) {
+      wpt_db_->RemovePrepared(seq.first, seq.second);
+    }
+  }
+
+  unprep_seqs_.clear();
+  flushed_save_points_.reset(nullptr);
+  unflushed_save_points_.reset(nullptr);
+  return s;
+}
+
+void WriteUnpreparedTxn::Clear() {
+  if (!recovered_txn_) {
+    txn_db_impl_->UnLock(this, *tracked_locks_);
+  }
+  unprep_seqs_.clear();
+  flushed_save_points_.reset(nullptr);
+  unflushed_save_points_.reset(nullptr);
+  recovered_txn_ = false;
+  largest_validated_seq_ = 0;
+  for (auto& it : active_iterators_) {
+    auto bdit = static_cast<BaseDeltaIterator*>(it);
+    bdit->Invalidate(Status::InvalidArgument(
+        "Cannot use iterator after transaction has finished"));
+  }
+  active_iterators_.clear();
+  untracked_keys_.clear();
+  TransactionBaseImpl::Clear();
+}
+
+void WriteUnpreparedTxn::SetSavePoint() {
+  assert((unflushed_save_points_ ? unflushed_save_points_->size() : 0) +
+             (flushed_save_points_ ? flushed_save_points_->size() : 0) ==
+         (save_points_ ? save_points_->size() : 0));
+  PessimisticTransaction::SetSavePoint();
+  if (unflushed_save_points_ == nullptr) {
+    unflushed_save_points_.reset(new autovector<size_t>());
+  }
+  unflushed_save_points_->push_back(write_batch_.GetDataSize());
+}
+
+Status WriteUnpreparedTxn::RollbackToSavePoint() {
+  assert((unflushed_save_points_ ? unflushed_save_points_->size() : 0) +
+             (flushed_save_points_ ? flushed_save_points_->size() : 0) ==
+         (save_points_ ? save_points_->size() : 0));
+  if (unflushed_save_points_ != nullptr && unflushed_save_points_->size() > 0) {
+    Status s = PessimisticTransaction::RollbackToSavePoint();
+    assert(!s.IsNotFound());
+    unflushed_save_points_->pop_back();
+    return s;
+  }
+
+  if (flushed_save_points_ != nullptr && !flushed_save_points_->empty()) {
+    return RollbackToSavePointInternal();
+  }
+
+  return Status::NotFound();
+}
+
+Status WriteUnpreparedTxn::RollbackToSavePointInternal() {
+  Status s;
+
+  const bool kClear = true;
+  TransactionBaseImpl::InitWriteBatch(kClear);
+
+  assert(flushed_save_points_->size() > 0);
+  WriteUnpreparedTxn::SavePoint& top = flushed_save_points_->back();
+
+  assert(save_points_ != nullptr && save_points_->size() > 0);
+  const LockTracker& tracked_keys = *save_points_->top().new_locks_;
+
+  ReadOptions roptions;
+  roptions.snapshot = top.snapshot_->snapshot();
+  SequenceNumber min_uncommitted =
+      static_cast_with_check<const SnapshotImpl>(roptions.snapshot)
+          ->min_uncommitted_;
+  SequenceNumber snap_seq = roptions.snapshot->GetSequenceNumber();
+  WriteUnpreparedTxnReadCallback callback(wupt_db_, snap_seq, min_uncommitted,
+                                          top.unprep_seqs_,
+                                          kBackedByDBSnapshot);
+  s = WriteRollbackKeys(tracked_keys, &write_batch_, &callback, roptions);
+  if (!s.ok()) {
+    return s;
+  }
+
+  const bool kPrepared = true;
+  s = FlushWriteBatchToDBInternal(!kPrepared);
+  if (!s.ok()) {
+    return s;
+  }
+
+  // PessimisticTransaction::RollbackToSavePoint will call also call
+  // RollbackToSavepoint on write_batch_. However, write_batch_ is empty and has
+  // no savepoints because this savepoint has already been flushed. Work around
+  // this by setting a fake savepoint.
+  write_batch_.SetSavePoint();
+  s = PessimisticTransaction::RollbackToSavePoint();
+  assert(s.ok());
+  if (!s.ok()) {
+    return s;
+  }
+
+  flushed_save_points_->pop_back();
+  return s;
+}
+
+Status WriteUnpreparedTxn::PopSavePoint() {
+  assert((unflushed_save_points_ ? unflushed_save_points_->size() : 0) +
+             (flushed_save_points_ ? flushed_save_points_->size() : 0) ==
+         (save_points_ ? save_points_->size() : 0));
+  if (unflushed_save_points_ != nullptr && unflushed_save_points_->size() > 0) {
+    Status s = PessimisticTransaction::PopSavePoint();
+    assert(!s.IsNotFound());
+    unflushed_save_points_->pop_back();
+    return s;
+  }
+
+  if (flushed_save_points_ != nullptr && !flushed_save_points_->empty()) {
+    // PessimisticTransaction::PopSavePoint will call also call PopSavePoint on
+    // write_batch_. However, write_batch_ is empty and has no savepoints
+    // because this savepoint has already been flushed. Work around this by
+    // setting a fake savepoint.
+    write_batch_.SetSavePoint();
+    Status s = PessimisticTransaction::PopSavePoint();
+    assert(!s.IsNotFound());
+    flushed_save_points_->pop_back();
+    return s;
+  }
+
+  return Status::NotFound();
+}
+
+void WriteUnpreparedTxn::MultiGet(const ReadOptions& options,
+                                  ColumnFamilyHandle* column_family,
+                                  const size_t num_keys, const Slice* keys,
+                                  PinnableSlice* values, Status* statuses,
+                                  const bool sorted_input) {
+  SequenceNumber min_uncommitted, snap_seq;
+  const SnapshotBackup backed_by_snapshot =
+      wupt_db_->AssignMinMaxSeqs(options.snapshot, &min_uncommitted, &snap_seq);
+  WriteUnpreparedTxnReadCallback callback(wupt_db_, snap_seq, min_uncommitted,
+                                          unprep_seqs_, backed_by_snapshot);
+  write_batch_.MultiGetFromBatchAndDB(db_, options, column_family, num_keys,
+                                      keys, values, statuses, sorted_input,
+                                      &callback);
+  if (UNLIKELY(!callback.valid() ||
+               !wupt_db_->ValidateSnapshot(snap_seq, backed_by_snapshot))) {
+    wupt_db_->WPRecordTick(TXN_GET_TRY_AGAIN);
+    for (size_t i = 0; i < num_keys; i++) {
+      statuses[i] = Status::TryAgain();
+    }
+  }
+}
+
+Status WriteUnpreparedTxn::Get(const ReadOptions& options,
+                               ColumnFamilyHandle* column_family,
+                               const Slice& key, PinnableSlice* value) {
+  SequenceNumber min_uncommitted, snap_seq;
+  const SnapshotBackup backed_by_snapshot =
+      wupt_db_->AssignMinMaxSeqs(options.snapshot, &min_uncommitted, &snap_seq);
+  WriteUnpreparedTxnReadCallback callback(wupt_db_, snap_seq, min_uncommitted,
+                                          unprep_seqs_, backed_by_snapshot);
+  auto res = write_batch_.GetFromBatchAndDB(db_, options, column_family, key,
+                                            value, &callback);
+  if (LIKELY(callback.valid() &&
+             wupt_db_->ValidateSnapshot(snap_seq, backed_by_snapshot))) {
+    return res;
+  } else {
+    res.PermitUncheckedError();
+    wupt_db_->WPRecordTick(TXN_GET_TRY_AGAIN);
+    return Status::TryAgain();
+  }
+}
+
+namespace {
+static void CleanupWriteUnpreparedWBWIIterator(void* arg1, void* arg2) {
+  auto txn = reinterpret_cast<WriteUnpreparedTxn*>(arg1);
+  auto iter = reinterpret_cast<Iterator*>(arg2);
+  txn->RemoveActiveIterator(iter);
+}
+}  // anonymous namespace
+
+Iterator* WriteUnpreparedTxn::GetIterator(const ReadOptions& options) {
+  return GetIterator(options, wupt_db_->DefaultColumnFamily());
+}
+
+Iterator* WriteUnpreparedTxn::GetIterator(const ReadOptions& options,
+                                          ColumnFamilyHandle* column_family) {
+  // Make sure to get iterator from WriteUnprepareTxnDB, not the root db.
+  Iterator* db_iter = wupt_db_->NewIterator(options, column_family, this);
+  assert(db_iter);
+
+  auto iter = write_batch_.NewIteratorWithBase(column_family, db_iter);
+  active_iterators_.push_back(iter);
+  iter->RegisterCleanup(CleanupWriteUnpreparedWBWIIterator, this, iter);
+  return iter;
+}
+
+Status WriteUnpreparedTxn::ValidateSnapshot(ColumnFamilyHandle* column_family,
+                                            const Slice& key,
+                                            SequenceNumber* tracked_at_seq) {
+  // TODO(lth): Reduce duplicate code with WritePrepared ValidateSnapshot logic.
+  assert(snapshot_);
+
+  SequenceNumber min_uncommitted =
+      static_cast_with_check<const SnapshotImpl>(snapshot_.get())
+          ->min_uncommitted_;
+  SequenceNumber snap_seq = snapshot_->GetSequenceNumber();
+  // tracked_at_seq is either max or the last snapshot with which this key was
+  // trackeed so there is no need to apply the IsInSnapshot to this comparison
+  // here as tracked_at_seq is not a prepare seq.
+  if (*tracked_at_seq <= snap_seq) {
+    // If the key has been previous validated at a sequence number earlier
+    // than the curent snapshot's sequence number, we already know it has not
+    // been modified.
+    return Status::OK();
+  }
+
+  *tracked_at_seq = snap_seq;
+
+  ColumnFamilyHandle* cfh =
+      column_family ? column_family : db_impl_->DefaultColumnFamily();
+
+  WriteUnpreparedTxnReadCallback snap_checker(
+      wupt_db_, snap_seq, min_uncommitted, unprep_seqs_, kBackedByDBSnapshot);
+  // TODO(yanqin): Support user-defined timestamp.
+  return TransactionUtil::CheckKeyForConflicts(
+      db_impl_, cfh, key.ToString(), snap_seq, /*ts=*/nullptr,
+      false /* cache_only */, &snap_checker, min_uncommitted);
+}
+
+const std::map<SequenceNumber, size_t>&
+WriteUnpreparedTxn::GetUnpreparedSequenceNumbers() {
+  return unprep_seqs_;
+}
+
+}  // namespace ROCKSDB_NAMESPACE
+
+#endif  // ROCKSDB_LITE
diff --git a/src/rocksdb/utilities/transactions/write_unprepared_txn.h b/src/rocksdb/utilities/transactions/write_unprepared_txn.h
new file mode 100644
index 000000000..5a3227f4e
--- /dev/null
+++ b/src/rocksdb/utilities/transactions/write_unprepared_txn.h
@@ -0,0 +1,341 @@
+// Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#pragma once
+
+#ifndef ROCKSDB_LITE
+
+#include <set>
+
+#include "utilities/transactions/write_prepared_txn.h"
+#include "utilities/transactions/write_unprepared_txn_db.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+class WriteUnpreparedTxnDB;
+class WriteUnpreparedTxn;
+
+// WriteUnprepared transactions needs to be able to read their own uncommitted
+// writes, and supporting this requires some careful consideration. Because
+// writes in the current transaction may be flushed to DB already, we cannot
+// rely on the contents of WriteBatchWithIndex to determine whether a key should
+// be visible or not, so we have to remember to check the DB for any uncommitted
+// keys that should be visible to us. First, we will need to change the seek to
+// snapshot logic, to seek to max_visible_seq = max(snap_seq, max_unprep_seq).
+// Any key greater than max_visible_seq should not be visible because they
+// cannot be unprepared by the current transaction and they are not in its
+// snapshot.
+//
+// When we seek to max_visible_seq, one of these cases will happen:
+// 1. We hit a unprepared key from the current transaction.
+// 2. We hit a unprepared key from the another transaction.
+// 3. We hit a committed key with snap_seq < seq < max_unprep_seq.
+// 4. We hit a committed key with seq <= snap_seq.
+//
+// IsVisibleFullCheck handles all cases correctly.
+//
+// Other notes:
+// Note that max_visible_seq is only calculated once at iterator construction
+// time, meaning if the same transaction is adding more unprep seqs through
+// writes during iteration, these newer writes may not be visible. This is not a
+// problem for MySQL though because it avoids modifying the index as it is
+// scanning through it to avoid the Halloween Problem. Instead, it scans the
+// index once up front, and modifies based on a temporary copy.
+//
+// In DBIter, there is a "reseek" optimization if the iterator skips over too
+// many keys. However, this assumes that the reseek seeks exactly to the
+// required key. In write unprepared, even after seeking directly to
+// max_visible_seq, some iteration may be required before hitting a visible key,
+// and special precautions must be taken to avoid performing another reseek,
+// leading to an infinite loop.
+//
+class WriteUnpreparedTxnReadCallback : public ReadCallback {
+ public:
+  WriteUnpreparedTxnReadCallback(
+      WritePreparedTxnDB* db, SequenceNumber snapshot,
+      SequenceNumber min_uncommitted,
+      const std::map<SequenceNumber, size_t>& unprep_seqs,
+      SnapshotBackup backed_by_snapshot)
+      // Pass our last uncommitted seq as the snapshot to the parent class to
+      // ensure that the parent will not prematurely filter out own writes. We
+      // will do the exact comparison against snapshots in IsVisibleFullCheck
+      // override.
+      : ReadCallback(CalcMaxVisibleSeq(unprep_seqs, snapshot), min_uncommitted),
+        db_(db),
+        unprep_seqs_(unprep_seqs),
+        wup_snapshot_(snapshot),
+        backed_by_snapshot_(backed_by_snapshot) {
+    (void)backed_by_snapshot_;  // to silence unused private field warning
+  }
+
+  virtual ~WriteUnpreparedTxnReadCallback() {
+    // If it is not backed by snapshot, the caller must check validity
+    assert(valid_checked_ || backed_by_snapshot_ == kBackedByDBSnapshot);
+  }
+
+  virtual bool IsVisibleFullCheck(SequenceNumber seq) override;
+
+  inline bool valid() {
+    valid_checked_ = true;
+    return snap_released_ == false;
+  }
+
+  void Refresh(SequenceNumber seq) override {
+    max_visible_seq_ = std::max(max_visible_seq_, seq);
+    wup_snapshot_ = seq;
+  }
+
+  static SequenceNumber CalcMaxVisibleSeq(
+      const std::map<SequenceNumber, size_t>& unprep_seqs,
+      SequenceNumber snapshot_seq) {
+    SequenceNumber max_unprepared = 0;
+    if (unprep_seqs.size()) {
+      max_unprepared =
+          unprep_seqs.rbegin()->first + unprep_seqs.rbegin()->second - 1;
+    }
+    return std::max(max_unprepared, snapshot_seq);
+  }
+
+ private:
+  WritePreparedTxnDB* db_;
+  const std::map<SequenceNumber, size_t>& unprep_seqs_;
+  SequenceNumber wup_snapshot_;
+  // Whether max_visible_seq_ is backed by a snapshot
+  const SnapshotBackup backed_by_snapshot_;
+  bool snap_released_ = false;
+  // Safety check to ensure that the caller has checked invalid statuses
+  bool valid_checked_ = false;
+};
+
+class WriteUnpreparedTxn : public WritePreparedTxn {
+ public:
+  WriteUnpreparedTxn(WriteUnpreparedTxnDB* db,
+                     const WriteOptions& write_options,
+                     const TransactionOptions& txn_options);
+
+  virtual ~WriteUnpreparedTxn();
+
+  using TransactionBaseImpl::Put;
+  virtual Status Put(ColumnFamilyHandle* column_family, const Slice& key,
+                     const Slice& value,
+                     const bool assume_tracked = false) override;
+  virtual Status Put(ColumnFamilyHandle* column_family, const SliceParts& key,
+                     const SliceParts& value,
+                     const bool assume_tracked = false) override;
+
+  using TransactionBaseImpl::Merge;
+  virtual Status Merge(ColumnFamilyHandle* column_family, const Slice& key,
+                       const Slice& value,
+                       const bool assume_tracked = false) override;
+
+  using TransactionBaseImpl::Delete;
+  virtual Status Delete(ColumnFamilyHandle* column_family, const Slice& key,
+                        const bool assume_tracked = false) override;
+  virtual Status Delete(ColumnFamilyHandle* column_family,
+                        const SliceParts& key,
+                        const bool assume_tracked = false) override;
+
+  using TransactionBaseImpl::SingleDelete;
+  virtual Status SingleDelete(ColumnFamilyHandle* column_family,
+                              const Slice& key,
+                              const bool assume_tracked = false) override;
+  virtual Status SingleDelete(ColumnFamilyHandle* column_family,
+                              const SliceParts& key,
+                              const bool assume_tracked = false) override;
+
+  // In WriteUnprepared, untracked writes will break snapshot validation logic.
+  // Snapshot validation will only check the largest sequence number of a key to
+  // see if it was committed or not. However, an untracked unprepared write will
+  // hide smaller committed sequence numbers.
+  //
+  // TODO(lth): Investigate whether it is worth having snapshot validation
+  // validate all values larger than snap_seq. Otherwise, we should return
+  // Status::NotSupported for untracked writes.
+
+  virtual Status RebuildFromWriteBatch(WriteBatch*) override;
+
+  virtual uint64_t GetLastLogNumber() const override {
+    return last_log_number_;
+  }
+
+  void RemoveActiveIterator(Iterator* iter) {
+    active_iterators_.erase(
+        std::remove(active_iterators_.begin(), active_iterators_.end(), iter),
+        active_iterators_.end());
+  }
+
+ protected:
+  void Initialize(const TransactionOptions& txn_options) override;
+
+  Status PrepareInternal() override;
+
+  Status CommitWithoutPrepareInternal() override;
+  Status CommitInternal() override;
+
+  Status RollbackInternal() override;
+
+  void Clear() override;
+
+  void SetSavePoint() override;
+  Status RollbackToSavePoint() override;
+  Status PopSavePoint() override;
+
+  // Get and GetIterator needs to be overridden so that a ReadCallback to
+  // handle read-your-own-write is used.
+  using Transaction::Get;
+  virtual Status Get(const ReadOptions& options,
+                     ColumnFamilyHandle* column_family, const Slice& key,
+                     PinnableSlice* value) override;
+
+  using Transaction::MultiGet;
+  virtual void MultiGet(const ReadOptions& options,
+                        ColumnFamilyHandle* column_family,
+                        const size_t num_keys, const Slice* keys,
+                        PinnableSlice* values, Status* statuses,
+                        const bool sorted_input = false) override;
+
+  using Transaction::GetIterator;
+  virtual Iterator* GetIterator(const ReadOptions& options) override;
+  virtual Iterator* GetIterator(const ReadOptions& options,
+                                ColumnFamilyHandle* column_family) override;
+
+  virtual Status ValidateSnapshot(ColumnFamilyHandle* column_family,
+                                  const Slice& key,
+                                  SequenceNumber* tracked_at_seq) override;
+
+ private:
+  friend class WriteUnpreparedTransactionTest_ReadYourOwnWrite_Test;
+  friend class WriteUnpreparedTransactionTest_RecoveryTest_Test;
+  friend class WriteUnpreparedTransactionTest_UnpreparedBatch_Test;
+  friend class WriteUnpreparedTxnDB;
+
+  const std::map<SequenceNumber, size_t>& GetUnpreparedSequenceNumbers();
+  Status WriteRollbackKeys(const LockTracker& tracked_keys,
+                           WriteBatchWithIndex* rollback_batch,
+                           ReadCallback* callback, const ReadOptions& roptions);
+
+  Status MaybeFlushWriteBatchToDB();
+  Status FlushWriteBatchToDB(bool prepared);
+  Status FlushWriteBatchToDBInternal(bool prepared);
+  Status FlushWriteBatchWithSavePointToDB();
+  Status RollbackToSavePointInternal();
+  Status HandleWrite(std::function<Status()> do_write);
+
+  // For write unprepared, we check on every writebatch append to see if
+  // write_batch_flush_threshold_ has been exceeded, and then call
+  // FlushWriteBatchToDB if so. This logic is encapsulated in
+  // MaybeFlushWriteBatchToDB.
+  int64_t write_batch_flush_threshold_;
+  WriteUnpreparedTxnDB* wupt_db_;
+
+  // Ordered list of unprep_seq sequence numbers that we have already written
+  // to DB.
+  //
+  // This maps unprep_seq => prepare_batch_cnt for each unprepared batch
+  // written by this transaction.
+  //
+  // Note that this contains both prepared and unprepared batches, since they
+  // are treated similarily in prepare heap/commit map, so it simplifies the
+  // commit callbacks.
+  std::map<SequenceNumber, size_t> unprep_seqs_;
+
+  uint64_t last_log_number_;
+
+  // Recovered transactions have tracked_keys_ populated, but are not actually
+  // locked for efficiency reasons. For recovered transactions, skip unlocking
+  // keys when transaction ends.
+  bool recovered_txn_;
+
+  // Track the largest sequence number at which we performed snapshot
+  // validation. If snapshot validation was skipped because no snapshot was set,
+  // then this is set to GetLastPublishedSequence. This value is useful because
+  // it means that for keys that have unprepared seqnos, we can guarantee that
+  // no committed keys by other transactions can exist between
+  // largest_validated_seq_ and max_unprep_seq. See
+  // WriteUnpreparedTxnDB::NewIterator for an explanation for why this is
+  // necessary for iterator Prev().
+  //
+  // Currently this value only increases during the lifetime of a transaction,
+  // but in some cases, we should be able to restore the previously largest
+  // value when calling RollbackToSavepoint.
+  SequenceNumber largest_validated_seq_;
+
+  struct SavePoint {
+    // Record of unprep_seqs_ at this savepoint. The set of unprep_seq is
+    // used during RollbackToSavepoint to determine visibility when restoring
+    // old values.
+    //
+    // TODO(lth): Since all unprep_seqs_ sets further down the stack must be
+    // subsets, this can potentially be deduplicated by just storing set
+    // difference. Investigate if this is worth it.
+    std::map<SequenceNumber, size_t> unprep_seqs_;
+
+    // This snapshot will be used to read keys at this savepoint if we call
+    // RollbackToSavePoint.
+    std::unique_ptr<ManagedSnapshot> snapshot_;
+
+    SavePoint(const std::map<SequenceNumber, size_t>& seqs,
+              ManagedSnapshot* snapshot)
+        : unprep_seqs_(seqs), snapshot_(snapshot){};
+  };
+
+  // We have 3 data structures holding savepoint information:
+  // 1. TransactionBaseImpl::save_points_
+  // 2. WriteUnpreparedTxn::flushed_save_points_
+  // 3. WriteUnpreparecTxn::unflushed_save_points_
+  //
+  // TransactionBaseImpl::save_points_ holds information about all write
+  // batches, including the current in-memory write_batch_, or unprepared
+  // batches that have been written out. Its responsibility is just to track
+  // which keys have been modified in every savepoint.
+  //
+  // WriteUnpreparedTxn::flushed_save_points_ holds information about savepoints
+  // set on unprepared batches that have already flushed. It holds the snapshot
+  // and unprep_seqs at that savepoint, so that the rollback process can
+  // determine which keys were visible at that point in time.
+  //
+  // WriteUnpreparecTxn::unflushed_save_points_ holds information about
+  // savepoints on the current in-memory write_batch_. It simply records the
+  // size of the write batch at every savepoint.
+  //
+  // TODO(lth): Remove the redundancy between save_point_boundaries_ and
+  // write_batch_.save_points_.
+  //
+  // Based on this information, here are some invariants:
+  // size(unflushed_save_points_) = size(write_batch_.save_points_)
+  // size(flushed_save_points_) + size(unflushed_save_points_)
+  //   = size(save_points_)
+  //
+  std::unique_ptr<autovector<WriteUnpreparedTxn::SavePoint>>
+      flushed_save_points_;
+  std::unique_ptr<autovector<size_t>> unflushed_save_points_;
+
+  // It is currently unsafe to flush a write batch if there are active iterators
+  // created from this transaction. This is because we use WriteBatchWithIndex
+  // to do merging reads from the DB and the write batch. If we flush the write
+  // batch, it is possible that the delta iterator on the iterator will point to
+  // invalid memory.
+  std::vector<Iterator*> active_iterators_;
+
+  // Untracked keys that we have to rollback.
+  //
+  // TODO(lth): Currently we we do not record untracked keys per-savepoint.
+  // This means that when rolling back to savepoints, we have to check all
+  // keys in the current transaction for rollback. Note that this is only
+  // inefficient, but still correct because we take a snapshot at every
+  // savepoint, and we will use that snapshot to construct the rollback batch.
+  // The rollback batch will then contain a reissue of the same marker.
+  //
+  // A more optimal solution would be to only check keys changed since the
+  // last savepoint. Also, it may make sense to merge this into tracked_keys_
+  // and differentiate between tracked but not locked keys to avoid having two
+  // very similar data structures.
+  using KeySet = std::unordered_map<uint32_t, std::vector<std::string>>;
+  KeySet untracked_keys_;
+};
+
+}  // namespace ROCKSDB_NAMESPACE
+
+#endif  // ROCKSDB_LITE
diff --git a/src/rocksdb/utilities/transactions/write_unprepared_txn_db.cc b/src/rocksdb/utilities/transactions/write_unprepared_txn_db.cc
new file mode 100644
index 000000000..2ed2d5c59
--- /dev/null
+++ b/src/rocksdb/utilities/transactions/write_unprepared_txn_db.cc
@@ -0,0 +1,473 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#ifndef ROCKSDB_LITE
+
+#include "utilities/transactions/write_unprepared_txn_db.h"
+
+#include "db/arena_wrapped_db_iter.h"
+#include "rocksdb/utilities/transaction_db.h"
+#include "util/cast_util.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+// Instead of reconstructing a Transaction object, and calling rollback on it,
+// we can be more efficient with RollbackRecoveredTransaction by skipping
+// unnecessary steps (eg. updating CommitMap, reconstructing keyset)
+Status WriteUnpreparedTxnDB::RollbackRecoveredTransaction(
+    const DBImpl::RecoveredTransaction* rtxn) {
+  // TODO(lth): Reduce duplicate code with WritePrepared rollback logic.
+  assert(rtxn->unprepared_);
+  auto cf_map_shared_ptr = WritePreparedTxnDB::GetCFHandleMap();
+  auto cf_comp_map_shared_ptr = WritePreparedTxnDB::GetCFComparatorMap();
+  // In theory we could write with disableWAL = true during recovery, and
+  // assume that if we crash again during recovery, we can just replay from
+  // the very beginning. Unfortunately, the XIDs from the application may not
+  // necessarily be unique across restarts, potentially leading to situations
+  // like this:
+  //
+  // BEGIN_PREPARE(unprepared) Put(a) END_PREPARE(xid = 1)
+  // -- crash and recover with Put(a) rolled back as it was not prepared
+  // BEGIN_PREPARE(prepared) Put(b) END_PREPARE(xid = 1)
+  // COMMIT(xid = 1)
+  // -- crash and recover with both a, b
+  //
+  // We could just write the rollback marker, but then we would have to extend
+  // MemTableInserter during recovery to actually do writes into the DB
+  // instead of just dropping the in-memory write batch.
+  //
+  WriteOptions w_options;
+
+  class InvalidSnapshotReadCallback : public ReadCallback {
+   public:
+    InvalidSnapshotReadCallback(SequenceNumber snapshot)
+        : ReadCallback(snapshot) {}
+
+    inline bool IsVisibleFullCheck(SequenceNumber) override {
+      // The seq provided as snapshot is the seq right before we have locked and
+      // wrote to it, so whatever is there, it is committed.
+      return true;
+    }
+
+    // Ignore the refresh request since we are confident that our snapshot seq
+    // is not going to be affected by concurrent compactions (not enabled yet.)
+    void Refresh(SequenceNumber) override {}
+  };
+
+  // Iterate starting with largest sequence number.
+  for (auto it = rtxn->batches_.rbegin(); it != rtxn->batches_.rend(); ++it) {
+    auto last_visible_txn = it->first - 1;
+    const auto& batch = it->second.batch_;
+    WriteBatch rollback_batch(0 /* reserved_bytes */, 0 /* max_bytes */,
+                              w_options.protection_bytes_per_key,
+                              0 /* default_cf_ts_sz */);
+
+    struct RollbackWriteBatchBuilder : public WriteBatch::Handler {
+      DBImpl* db_;
+      ReadOptions roptions;
+      InvalidSnapshotReadCallback callback;
+      WriteBatch* rollback_batch_;
+      std::map<uint32_t, const Comparator*>& comparators_;
+      std::map<uint32_t, ColumnFamilyHandle*>& handles_;
+      using CFKeys = std::set<Slice, SetComparator>;
+      std::map<uint32_t, CFKeys> keys_;
+      bool rollback_merge_operands_;
+      RollbackWriteBatchBuilder(
+          DBImpl* db, SequenceNumber snap_seq, WriteBatch* dst_batch,
+          std::map<uint32_t, const Comparator*>& comparators,
+          std::map<uint32_t, ColumnFamilyHandle*>& handles,
+          bool rollback_merge_operands)
+          : db_(db),
+            callback(snap_seq),
+            // disable min_uncommitted optimization
+            rollback_batch_(dst_batch),
+            comparators_(comparators),
+            handles_(handles),
+            rollback_merge_operands_(rollback_merge_operands) {}
+
+      Status Rollback(uint32_t cf, const Slice& key) {
+        Status s;
+        CFKeys& cf_keys = keys_[cf];
+        if (cf_keys.size() == 0) {  // just inserted
+          auto cmp = comparators_[cf];
+          keys_[cf] = CFKeys(SetComparator(cmp));
+        }
+        auto res = cf_keys.insert(key);
+        if (res.second ==
+            false) {  // second is false if a element already existed.
+          return s;
+        }
+
+        PinnableSlice pinnable_val;
+        bool not_used;
+        auto cf_handle = handles_[cf];
+        DBImpl::GetImplOptions get_impl_options;
+        get_impl_options.column_family = cf_handle;
+        get_impl_options.value = &pinnable_val;
+        get_impl_options.value_found = &not_used;
+        get_impl_options.callback = &callback;
+        s = db_->GetImpl(roptions, key, get_impl_options);
+        assert(s.ok() || s.IsNotFound());
+        if (s.ok()) {
+          s = rollback_batch_->Put(cf_handle, key, pinnable_val);
+          assert(s.ok());
+        } else if (s.IsNotFound()) {
+          // There has been no readable value before txn. By adding a delete we
+          // make sure that there will be none afterwards either.
+          s = rollback_batch_->Delete(cf_handle, key);
+          assert(s.ok());
+        } else {
+          // Unexpected status. Return it to the user.
+        }
+        return s;
+      }
+
+      Status PutCF(uint32_t cf, const Slice& key,
+                   const Slice& /*val*/) override {
+        return Rollback(cf, key);
+      }
+
+      Status DeleteCF(uint32_t cf, const Slice& key) override {
+        return Rollback(cf, key);
+      }
+
+      Status SingleDeleteCF(uint32_t cf, const Slice& key) override {
+        return Rollback(cf, key);
+      }
+
+      Status MergeCF(uint32_t cf, const Slice& key,
+                     const Slice& /*val*/) override {
+        if (rollback_merge_operands_) {
+          return Rollback(cf, key);
+        } else {
+          return Status::OK();
+        }
+      }
+
+      // Recovered batches do not contain 2PC markers.
+      Status MarkNoop(bool) override { return Status::InvalidArgument(); }
+      Status MarkBeginPrepare(bool) override {
+        return Status::InvalidArgument();
+      }
+      Status MarkEndPrepare(const Slice&) override {
+        return Status::InvalidArgument();
+      }
+      Status MarkCommit(const Slice&) override {
+        return Status::InvalidArgument();
+      }
+      Status MarkRollback(const Slice&) override {
+        return Status::InvalidArgument();
+      }
+    } rollback_handler(db_impl_, last_visible_txn, &rollback_batch,
+                       *cf_comp_map_shared_ptr.get(), *cf_map_shared_ptr.get(),
+                       txn_db_options_.rollback_merge_operands);
+
+    auto s = batch->Iterate(&rollback_handler);
+    if (!s.ok()) {
+      return s;
+    }
+
+    // The Rollback marker will be used as a batch separator
+    s = WriteBatchInternal::MarkRollback(&rollback_batch, rtxn->name_);
+    if (!s.ok()) {
+      return s;
+    }
+
+    const uint64_t kNoLogRef = 0;
+    const bool kDisableMemtable = true;
+    const size_t kOneBatch = 1;
+    uint64_t seq_used = kMaxSequenceNumber;
+    s = db_impl_->WriteImpl(w_options, &rollback_batch, nullptr, nullptr,
+                            kNoLogRef, !kDisableMemtable, &seq_used, kOneBatch);
+    if (!s.ok()) {
+      return s;
+    }
+
+    // If two_write_queues, we must manually release the sequence number to
+    // readers.
+    if (db_impl_->immutable_db_options().two_write_queues) {
+      db_impl_->SetLastPublishedSequence(seq_used);
+    }
+  }
+
+  return Status::OK();
+}
+
+Status WriteUnpreparedTxnDB::Initialize(
+    const std::vector<size_t>& compaction_enabled_cf_indices,
+    const std::vector<ColumnFamilyHandle*>& handles) {
+  // TODO(lth): Reduce code duplication in this function.
+  auto dbimpl = static_cast_with_check<DBImpl>(GetRootDB());
+  assert(dbimpl != nullptr);
+
+  db_impl_->SetSnapshotChecker(new WritePreparedSnapshotChecker(this));
+  // A callback to commit a single sub-batch
+  class CommitSubBatchPreReleaseCallback : public PreReleaseCallback {
+   public:
+    explicit CommitSubBatchPreReleaseCallback(WritePreparedTxnDB* db)
+        : db_(db) {}
+    Status Callback(SequenceNumber commit_seq,
+                    bool is_mem_disabled __attribute__((__unused__)), uint64_t,
+                    size_t /*index*/, size_t /*total*/) override {
+      assert(!is_mem_disabled);
+      db_->AddCommitted(commit_seq, commit_seq);
+      return Status::OK();
+    }
+
+   private:
+    WritePreparedTxnDB* db_;
+  };
+  db_impl_->SetRecoverableStatePreReleaseCallback(
+      new CommitSubBatchPreReleaseCallback(this));
+
+  // PessimisticTransactionDB::Initialize
+  for (auto cf_ptr : handles) {
+    AddColumnFamily(cf_ptr);
+  }
+  // Verify cf options
+  for (auto handle : handles) {
+    ColumnFamilyDescriptor cfd;
+    Status s = handle->GetDescriptor(&cfd);
+    if (!s.ok()) {
+      return s;
+    }
+    s = VerifyCFOptions(cfd.options);
+    if (!s.ok()) {
+      return s;
+    }
+  }
+
+  // Re-enable compaction for the column families that initially had
+  // compaction enabled.
+  std::vector<ColumnFamilyHandle*> compaction_enabled_cf_handles;
+  compaction_enabled_cf_handles.reserve(compaction_enabled_cf_indices.size());
+  for (auto index : compaction_enabled_cf_indices) {
+    compaction_enabled_cf_handles.push_back(handles[index]);
+  }
+
+  // create 'real' transactions from recovered shell transactions
+  auto rtxns = dbimpl->recovered_transactions();
+  std::map<SequenceNumber, SequenceNumber> ordered_seq_cnt;
+  for (auto rtxn : rtxns) {
+    auto recovered_trx = rtxn.second;
+    assert(recovered_trx);
+    assert(recovered_trx->batches_.size() >= 1);
+    assert(recovered_trx->name_.length());
+
+    // We can only rollback transactions after AdvanceMaxEvictedSeq is called,
+    // but AddPrepared must occur before AdvanceMaxEvictedSeq, which is why
+    // two iterations is required.
+    if (recovered_trx->unprepared_) {
+      continue;
+    }
+
+    WriteOptions w_options;
+    w_options.sync = true;
+    TransactionOptions t_options;
+
+    auto first_log_number = recovered_trx->batches_.begin()->second.log_number_;
+    auto first_seq = recovered_trx->batches_.begin()->first;
+    auto last_prepare_batch_cnt =
+        recovered_trx->batches_.begin()->second.batch_cnt_;
+
+    Transaction* real_trx = BeginTransaction(w_options, t_options, nullptr);
+    assert(real_trx);
+    auto wupt = static_cast_with_check<WriteUnpreparedTxn>(real_trx);
+    wupt->recovered_txn_ = true;
+
+    real_trx->SetLogNumber(first_log_number);
+    real_trx->SetId(first_seq);
+    Status s = real_trx->SetName(recovered_trx->name_);
+    if (!s.ok()) {
+      return s;
+    }
+    wupt->prepare_batch_cnt_ = last_prepare_batch_cnt;
+
+    for (auto batch : recovered_trx->batches_) {
+      const auto& seq = batch.first;
+      const auto& batch_info = batch.second;
+      auto cnt = batch_info.batch_cnt_ ? batch_info.batch_cnt_ : 1;
+      assert(batch_info.log_number_);
+
+      ordered_seq_cnt[seq] = cnt;
+      assert(wupt->unprep_seqs_.count(seq) == 0);
+      wupt->unprep_seqs_[seq] = cnt;
+
+      s = wupt->RebuildFromWriteBatch(batch_info.batch_);
+      assert(s.ok());
+      if (!s.ok()) {
+        return s;
+      }
+    }
+
+    const bool kClear = true;
+    wupt->InitWriteBatch(kClear);
+
+    real_trx->SetState(Transaction::PREPARED);
+    if (!s.ok()) {
+      return s;
+    }
+  }
+  // AddPrepared must be called in order
+  for (auto seq_cnt : ordered_seq_cnt) {
+    auto seq = seq_cnt.first;
+    auto cnt = seq_cnt.second;
+    for (size_t i = 0; i < cnt; i++) {
+      AddPrepared(seq + i);
+    }
+  }
+
+  SequenceNumber prev_max = max_evicted_seq_;
+  SequenceNumber last_seq = db_impl_->GetLatestSequenceNumber();
+  AdvanceMaxEvictedSeq(prev_max, last_seq);
+  // Create a gap between max and the next snapshot. This simplifies the logic
+  // in IsInSnapshot by not having to consider the special case of max ==
+  // snapshot after recovery. This is tested in IsInSnapshotEmptyMapTest.
+  if (last_seq) {
+    db_impl_->versions_->SetLastAllocatedSequence(last_seq + 1);
+    db_impl_->versions_->SetLastSequence(last_seq + 1);
+    db_impl_->versions_->SetLastPublishedSequence(last_seq + 1);
+  }
+
+  Status s;
+  // Rollback unprepared transactions.
+  for (auto rtxn : rtxns) {
+    auto recovered_trx = rtxn.second;
+    if (recovered_trx->unprepared_) {
+      s = RollbackRecoveredTransaction(recovered_trx);
+      if (!s.ok()) {
+        return s;
+      }
+      continue;
+    }
+  }
+
+  if (s.ok()) {
+    dbimpl->DeleteAllRecoveredTransactions();
+
+    // Compaction should start only after max_evicted_seq_ is set AND recovered
+    // transactions are either added to PrepareHeap or rolled back.
+    s = EnableAutoCompaction(compaction_enabled_cf_handles);
+  }
+
+  return s;
+}
+
+Transaction* WriteUnpreparedTxnDB::BeginTransaction(
+    const WriteOptions& write_options, const TransactionOptions& txn_options,
+    Transaction* old_txn) {
+  if (old_txn != nullptr) {
+    ReinitializeTransaction(old_txn, write_options, txn_options);
+    return old_txn;
+  } else {
+    return new WriteUnpreparedTxn(this, write_options, txn_options);
+  }
+}
+
+// Struct to hold ownership of snapshot and read callback for iterator cleanup.
+struct WriteUnpreparedTxnDB::IteratorState {
+  IteratorState(WritePreparedTxnDB* txn_db, SequenceNumber sequence,
+                std::shared_ptr<ManagedSnapshot> s,
+                SequenceNumber min_uncommitted, WriteUnpreparedTxn* txn)
+      : callback(txn_db, sequence, min_uncommitted, txn->unprep_seqs_,
+                 kBackedByDBSnapshot),
+        snapshot(s) {}
+  SequenceNumber MaxVisibleSeq() { return callback.max_visible_seq(); }
+
+  WriteUnpreparedTxnReadCallback callback;
+  std::shared_ptr<ManagedSnapshot> snapshot;
+};
+
+namespace {
+static void CleanupWriteUnpreparedTxnDBIterator(void* arg1, void* /*arg2*/) {
+  delete reinterpret_cast<WriteUnpreparedTxnDB::IteratorState*>(arg1);
+}
+}  // anonymous namespace
+
+Iterator* WriteUnpreparedTxnDB::NewIterator(const ReadOptions& options,
+                                            ColumnFamilyHandle* column_family,
+                                            WriteUnpreparedTxn* txn) {
+  // TODO(lth): Refactor so that this logic is shared with WritePrepared.
+  constexpr bool expose_blob_index = false;
+  constexpr bool allow_refresh = false;
+  std::shared_ptr<ManagedSnapshot> own_snapshot = nullptr;
+  SequenceNumber snapshot_seq = kMaxSequenceNumber;
+  SequenceNumber min_uncommitted = 0;
+
+  // Currently, the Prev() iterator logic does not work well without snapshot
+  // validation. The logic simply iterates through values of a key in
+  // ascending seqno order, stopping at the first non-visible value and
+  // returning the last visible value.
+  //
+  // For example, if snapshot sequence is 3, and we have the following keys:
+  // foo: v1 1
+  // foo: v2 2
+  // foo: v3 3
+  // foo: v4 4
+  // foo: v5 5
+  //
+  // Then 1, 2, 3 will be visible, but 4 will be non-visible, so we return v3,
+  // which is the last visible value.
+  //
+  // For unprepared transactions, if we have snap_seq = 3, but the current
+  // transaction has unprep_seq 5, then returning the first non-visible value
+  // would be incorrect, as we should return v5, and not v3. The problem is that
+  // there are committed values at snapshot_seq < commit_seq < unprep_seq.
+  //
+  // Snapshot validation can prevent this problem by ensuring that no committed
+  // values exist at snapshot_seq < commit_seq, and thus any value with a
+  // sequence number greater than snapshot_seq must be unprepared values. For
+  // example, if the transaction had a snapshot at 3, then snapshot validation
+  // would be performed during the Put(v5) call. It would find v4, and the Put
+  // would fail with snapshot validation failure.
+  //
+  // TODO(lth): Improve Prev() logic to continue iterating until
+  // max_visible_seq, and then return the last visible value, so that this
+  // restriction can be lifted.
+  const Snapshot* snapshot = nullptr;
+  if (options.snapshot == nullptr) {
+    snapshot = GetSnapshot();
+    own_snapshot = std::make_shared<ManagedSnapshot>(db_impl_, snapshot);
+  } else {
+    snapshot = options.snapshot;
+  }
+
+  snapshot_seq = snapshot->GetSequenceNumber();
+  assert(snapshot_seq != kMaxSequenceNumber);
+  // Iteration is safe as long as largest_validated_seq <= snapshot_seq. We are
+  // guaranteed that for keys that were modified by this transaction (and thus
+  // might have unprepared values), no committed values exist at
+  // largest_validated_seq < commit_seq (or the contrapositive: any committed
+  // value must exist at commit_seq <= largest_validated_seq). This implies
+  // that commit_seq <= largest_validated_seq <= snapshot_seq or commit_seq <=
+  // snapshot_seq. As explained above, the problem with Prev() only happens when
+  // snapshot_seq < commit_seq.
+  //
+  // For keys that were not modified by this transaction, largest_validated_seq_
+  // is meaningless, and Prev() should just work with the existing visibility
+  // logic.
+  if (txn->largest_validated_seq_ > snapshot->GetSequenceNumber() &&
+      !txn->unprep_seqs_.empty()) {
+    ROCKS_LOG_ERROR(info_log_,
+                    "WriteUnprepared iterator creation failed since the "
+                    "transaction has performed unvalidated writes");
+    return nullptr;
+  }
+  min_uncommitted =
+      static_cast_with_check<const SnapshotImpl>(snapshot)->min_uncommitted_;
+
+  auto* cfd =
+      static_cast_with_check<ColumnFamilyHandleImpl>(column_family)->cfd();
+  auto* state =
+      new IteratorState(this, snapshot_seq, own_snapshot, min_uncommitted, txn);
+  auto* db_iter = db_impl_->NewIteratorImpl(
+      options, cfd, state->MaxVisibleSeq(), &state->callback, expose_blob_index,
+      allow_refresh);
+  db_iter->RegisterCleanup(CleanupWriteUnpreparedTxnDBIterator, state, nullptr);
+  return db_iter;
+}
+
+}  // namespace ROCKSDB_NAMESPACE
+#endif  // ROCKSDB_LITE
diff --git a/src/rocksdb/utilities/transactions/write_unprepared_txn_db.h b/src/rocksdb/utilities/transactions/write_unprepared_txn_db.h
new file mode 100644
index 000000000..c40e96d49
--- /dev/null
+++ b/src/rocksdb/utilities/transactions/write_unprepared_txn_db.h
@@ -0,0 +1,108 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#pragma once
+#ifndef ROCKSDB_LITE
+
+#include "utilities/transactions/write_prepared_txn_db.h"
+#include "utilities/transactions/write_unprepared_txn.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+class WriteUnpreparedTxn;
+
+class WriteUnpreparedTxnDB : public WritePreparedTxnDB {
+ public:
+  using WritePreparedTxnDB::WritePreparedTxnDB;
+
+  Status Initialize(const std::vector<size_t>& compaction_enabled_cf_indices,
+                    const std::vector<ColumnFamilyHandle*>& handles) override;
+
+  Transaction* BeginTransaction(const WriteOptions& write_options,
+                                const TransactionOptions& txn_options,
+                                Transaction* old_txn) override;
+
+  // Struct to hold ownership of snapshot and read callback for cleanup.
+  struct IteratorState;
+
+  using WritePreparedTxnDB::NewIterator;
+  Iterator* NewIterator(const ReadOptions& options,
+                        ColumnFamilyHandle* column_family,
+                        WriteUnpreparedTxn* txn);
+
+ private:
+  Status RollbackRecoveredTransaction(const DBImpl::RecoveredTransaction* rtxn);
+};
+
+class WriteUnpreparedCommitEntryPreReleaseCallback : public PreReleaseCallback {
+  // TODO(lth): Reduce code duplication with
+  // WritePreparedCommitEntryPreReleaseCallback
+ public:
+  // includes_data indicates that the commit also writes non-empty
+  // CommitTimeWriteBatch to memtable, which needs to be committed separately.
+  WriteUnpreparedCommitEntryPreReleaseCallback(
+      WritePreparedTxnDB* db, DBImpl* db_impl,
+      const std::map<SequenceNumber, size_t>& unprep_seqs,
+      size_t data_batch_cnt = 0, bool publish_seq = true)
+      : db_(db),
+        db_impl_(db_impl),
+        unprep_seqs_(unprep_seqs),
+        data_batch_cnt_(data_batch_cnt),
+        includes_data_(data_batch_cnt_ > 0),
+        publish_seq_(publish_seq) {
+    assert(unprep_seqs.size() > 0);
+  }
+
+  virtual Status Callback(SequenceNumber commit_seq,
+                          bool is_mem_disabled __attribute__((__unused__)),
+                          uint64_t, size_t /*index*/,
+                          size_t /*total*/) override {
+    const uint64_t last_commit_seq = LIKELY(data_batch_cnt_ <= 1)
+                                         ? commit_seq
+                                         : commit_seq + data_batch_cnt_ - 1;
+    // Recall that unprep_seqs maps (un)prepared_seq => prepare_batch_cnt.
+    for (const auto& s : unprep_seqs_) {
+      for (size_t i = 0; i < s.second; i++) {
+        db_->AddCommitted(s.first + i, last_commit_seq);
+      }
+    }
+
+    if (includes_data_) {
+      assert(data_batch_cnt_);
+      // Commit the data that is accompanied with the commit request
+      for (size_t i = 0; i < data_batch_cnt_; i++) {
+        // For commit seq of each batch use the commit seq of the last batch.
+        // This would make debugging easier by having all the batches having
+        // the same sequence number.
+        db_->AddCommitted(commit_seq + i, last_commit_seq);
+      }
+    }
+    if (db_impl_->immutable_db_options().two_write_queues && publish_seq_) {
+      assert(is_mem_disabled);  // implies the 2nd queue
+      // Publish the sequence number. We can do that here assuming the callback
+      // is invoked only from one write queue, which would guarantee that the
+      // publish sequence numbers will be in order, i.e., once a seq is
+      // published all the seq prior to that are also publishable.
+      db_impl_->SetLastPublishedSequence(last_commit_seq);
+    }
+    // else SequenceNumber that is updated as part of the write already does the
+    // publishing
+    return Status::OK();
+  }
+
+ private:
+  WritePreparedTxnDB* db_;
+  DBImpl* db_impl_;
+  const std::map<SequenceNumber, size_t>& unprep_seqs_;
+  size_t data_batch_cnt_;
+  // Either because it is commit without prepare or it has a
+  // CommitTimeWriteBatch
+  bool includes_data_;
+  // Should the callback also publishes the commit seq number
+  bool publish_seq_;
+};
+
+}  // namespace ROCKSDB_NAMESPACE
+#endif  // ROCKSDB_LITE
diff --git a/src/rocksdb/utilities/ttl/db_ttl_impl.cc b/src/rocksdb/utilities/ttl/db_ttl_impl.cc
new file mode 100644
index 000000000..6ec9d87b0
--- /dev/null
+++ b/src/rocksdb/utilities/ttl/db_ttl_impl.cc
@@ -0,0 +1,609 @@
+// Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.
+// Copyright (c) 2011 The LevelDB Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file. See the AUTHORS file for names of contributors.
+#ifndef ROCKSDB_LITE
+
+#include "utilities/ttl/db_ttl_impl.h"
+
+#include "db/write_batch_internal.h"
+#include "file/filename.h"
+#include "logging/logging.h"
+#include "rocksdb/convenience.h"
+#include "rocksdb/env.h"
+#include "rocksdb/iterator.h"
+#include "rocksdb/system_clock.h"
+#include "rocksdb/utilities/db_ttl.h"
+#include "rocksdb/utilities/object_registry.h"
+#include "rocksdb/utilities/options_type.h"
+#include "util/coding.h"
+
+namespace ROCKSDB_NAMESPACE {
+static std::unordered_map<std::string, OptionTypeInfo> ttl_merge_op_type_info =
+    {{"user_operator",
+      OptionTypeInfo::AsCustomSharedPtr<MergeOperator>(
+          0, OptionVerificationType::kByName, OptionTypeFlags::kNone)}};
+
+TtlMergeOperator::TtlMergeOperator(
+    const std::shared_ptr<MergeOperator>& merge_op, SystemClock* clock)
+    : user_merge_op_(merge_op), clock_(clock) {
+  RegisterOptions("TtlMergeOptions", &user_merge_op_, &ttl_merge_op_type_info);
+}
+
+bool TtlMergeOperator::FullMergeV2(const MergeOperationInput& merge_in,
+                                   MergeOperationOutput* merge_out) const {
+  const uint32_t ts_len = DBWithTTLImpl::kTSLength;
+  if (merge_in.existing_value && merge_in.existing_value->size() < ts_len) {
+    ROCKS_LOG_ERROR(merge_in.logger,
+                    "Error: Could not remove timestamp from existing value.");
+    return false;
+  }
+
+  // Extract time-stamp from each operand to be passed to user_merge_op_
+  std::vector<Slice> operands_without_ts;
+  for (const auto& operand : merge_in.operand_list) {
+    if (operand.size() < ts_len) {
+      ROCKS_LOG_ERROR(merge_in.logger,
+                      "Error: Could not remove timestamp from operand value.");
+      return false;
+    }
+    operands_without_ts.push_back(operand);
+    operands_without_ts.back().remove_suffix(ts_len);
+  }
+
+  // Apply the user merge operator (store result in *new_value)
+  bool good = true;
+  MergeOperationOutput user_merge_out(merge_out->new_value,
+                                      merge_out->existing_operand);
+  if (merge_in.existing_value) {
+    Slice existing_value_without_ts(merge_in.existing_value->data(),
+                                    merge_in.existing_value->size() - ts_len);
+    good = user_merge_op_->FullMergeV2(
+        MergeOperationInput(merge_in.key, &existing_value_without_ts,
+                            operands_without_ts, merge_in.logger),
+        &user_merge_out);
+  } else {
+    good = user_merge_op_->FullMergeV2(
+        MergeOperationInput(merge_in.key, nullptr, operands_without_ts,
+                            merge_in.logger),
+        &user_merge_out);
+  }
+
+  // Return false if the user merge operator returned false
+  if (!good) {
+    return false;
+  }
+
+  if (merge_out->existing_operand.data()) {
+    merge_out->new_value.assign(merge_out->existing_operand.data(),
+                                merge_out->existing_operand.size());
+    merge_out->existing_operand = Slice(nullptr, 0);
+  }
+
+  // Augment the *new_value with the ttl time-stamp
+  int64_t curtime;
+  if (!clock_->GetCurrentTime(&curtime).ok()) {
+    ROCKS_LOG_ERROR(
+        merge_in.logger,
+        "Error: Could not get current time to be attached internally "
+        "to the new value.");
+    return false;
+  } else {
+    char ts_string[ts_len];
+    EncodeFixed32(ts_string, (int32_t)curtime);
+    merge_out->new_value.append(ts_string, ts_len);
+    return true;
+  }
+}
+
+bool TtlMergeOperator::PartialMergeMulti(const Slice& key,
+                                         const std::deque<Slice>& operand_list,
+                                         std::string* new_value,
+                                         Logger* logger) const {
+  const uint32_t ts_len = DBWithTTLImpl::kTSLength;
+  std::deque<Slice> operands_without_ts;
+
+  for (const auto& operand : operand_list) {
+    if (operand.size() < ts_len) {
+      ROCKS_LOG_ERROR(logger, "Error: Could not remove timestamp from value.");
+      return false;
+    }
+
+    operands_without_ts.push_back(
+        Slice(operand.data(), operand.size() - ts_len));
+  }
+
+  // Apply the user partial-merge operator (store result in *new_value)
+  assert(new_value);
+  if (!user_merge_op_->PartialMergeMulti(key, operands_without_ts, new_value,
+                                         logger)) {
+    return false;
+  }
+
+  // Augment the *new_value with the ttl time-stamp
+  int64_t curtime;
+  if (!clock_->GetCurrentTime(&curtime).ok()) {
+    ROCKS_LOG_ERROR(
+        logger,
+        "Error: Could not get current time to be attached internally "
+        "to the new value.");
+    return false;
+  } else {
+    char ts_string[ts_len];
+    EncodeFixed32(ts_string, (int32_t)curtime);
+    new_value->append(ts_string, ts_len);
+    return true;
+  }
+}
+
+Status TtlMergeOperator::PrepareOptions(const ConfigOptions& config_options) {
+  if (clock_ == nullptr) {
+    clock_ = config_options.env->GetSystemClock().get();
+  }
+  return MergeOperator::PrepareOptions(config_options);
+}
+
+Status TtlMergeOperator::ValidateOptions(
+    const DBOptions& db_opts, const ColumnFamilyOptions& cf_opts) const {
+  if (user_merge_op_ == nullptr) {
+    return Status::InvalidArgument(
+        "UserMergeOperator required by TtlMergeOperator");
+  } else if (clock_ == nullptr) {
+    return Status::InvalidArgument("SystemClock required by TtlMergeOperator");
+  } else {
+    return MergeOperator::ValidateOptions(db_opts, cf_opts);
+  }
+}
+
+void DBWithTTLImpl::SanitizeOptions(int32_t ttl, ColumnFamilyOptions* options,
+                                    SystemClock* clock) {
+  if (options->compaction_filter) {
+    options->compaction_filter =
+        new TtlCompactionFilter(ttl, clock, options->compaction_filter);
+  } else {
+    options->compaction_filter_factory =
+        std::shared_ptr<CompactionFilterFactory>(new TtlCompactionFilterFactory(
+            ttl, clock, options->compaction_filter_factory));
+  }
+
+  if (options->merge_operator) {
+    options->merge_operator.reset(
+        new TtlMergeOperator(options->merge_operator, clock));
+  }
+}
+
+static std::unordered_map<std::string, OptionTypeInfo> ttl_type_info = {
+    {"ttl", {0, OptionType::kInt32T}},
+};
+
+static std::unordered_map<std::string, OptionTypeInfo> ttl_cff_type_info = {
+    {"user_filter_factory",
+     OptionTypeInfo::AsCustomSharedPtr<CompactionFilterFactory>(
+         0, OptionVerificationType::kByNameAllowFromNull,
+         OptionTypeFlags::kNone)}};
+static std::unordered_map<std::string, OptionTypeInfo> user_cf_type_info = {
+    {"user_filter",
+     OptionTypeInfo::AsCustomRawPtr<const CompactionFilter>(
+         0, OptionVerificationType::kByName, OptionTypeFlags::kAllowNull)}};
+
+TtlCompactionFilter::TtlCompactionFilter(
+    int32_t ttl, SystemClock* clock, const CompactionFilter* _user_comp_filter,
+    std::unique_ptr<const CompactionFilter> _user_comp_filter_from_factory)
+    : LayeredCompactionFilterBase(_user_comp_filter,
+                                  std::move(_user_comp_filter_from_factory)),
+      ttl_(ttl),
+      clock_(clock) {
+  RegisterOptions("TTL", &ttl_, &ttl_type_info);
+  RegisterOptions("UserFilter", &user_comp_filter_, &user_cf_type_info);
+}
+
+bool TtlCompactionFilter::Filter(int level, const Slice& key,
+                                 const Slice& old_val, std::string* new_val,
+                                 bool* value_changed) const {
+  if (DBWithTTLImpl::IsStale(old_val, ttl_, clock_)) {
+    return true;
+  }
+  if (user_comp_filter() == nullptr) {
+    return false;
+  }
+  assert(old_val.size() >= DBWithTTLImpl::kTSLength);
+  Slice old_val_without_ts(old_val.data(),
+                           old_val.size() - DBWithTTLImpl::kTSLength);
+  if (user_comp_filter()->Filter(level, key, old_val_without_ts, new_val,
+                                 value_changed)) {
+    return true;
+  }
+  if (*value_changed) {
+    new_val->append(old_val.data() + old_val.size() - DBWithTTLImpl::kTSLength,
+                    DBWithTTLImpl::kTSLength);
+  }
+  return false;
+}
+
+Status TtlCompactionFilter::PrepareOptions(
+    const ConfigOptions& config_options) {
+  if (clock_ == nullptr) {
+    clock_ = config_options.env->GetSystemClock().get();
+  }
+  return LayeredCompactionFilterBase::PrepareOptions(config_options);
+}
+
+Status TtlCompactionFilter::ValidateOptions(
+    const DBOptions& db_opts, const ColumnFamilyOptions& cf_opts) const {
+  if (clock_ == nullptr) {
+    return Status::InvalidArgument(
+        "SystemClock required by TtlCompactionFilter");
+  } else {
+    return LayeredCompactionFilterBase::ValidateOptions(db_opts, cf_opts);
+  }
+}
+
+TtlCompactionFilterFactory::TtlCompactionFilterFactory(
+    int32_t ttl, SystemClock* clock,
+    std::shared_ptr<CompactionFilterFactory> comp_filter_factory)
+    : ttl_(ttl), clock_(clock), user_comp_filter_factory_(comp_filter_factory) {
+  RegisterOptions("UserOptions", &user_comp_filter_factory_,
+                  &ttl_cff_type_info);
+  RegisterOptions("TTL", &ttl_, &ttl_type_info);
+}
+
+std::unique_ptr<CompactionFilter>
+TtlCompactionFilterFactory::CreateCompactionFilter(
+    const CompactionFilter::Context& context) {
+  std::unique_ptr<const CompactionFilter> user_comp_filter_from_factory =
+      nullptr;
+  if (user_comp_filter_factory_) {
+    user_comp_filter_from_factory =
+        user_comp_filter_factory_->CreateCompactionFilter(context);
+  }
+
+  return std::unique_ptr<TtlCompactionFilter>(new TtlCompactionFilter(
+      ttl_, clock_, nullptr, std::move(user_comp_filter_from_factory)));
+}
+
+Status TtlCompactionFilterFactory::PrepareOptions(
+    const ConfigOptions& config_options) {
+  if (clock_ == nullptr) {
+    clock_ = config_options.env->GetSystemClock().get();
+  }
+  return CompactionFilterFactory::PrepareOptions(config_options);
+}
+
+Status TtlCompactionFilterFactory::ValidateOptions(
+    const DBOptions& db_opts, const ColumnFamilyOptions& cf_opts) const {
+  if (clock_ == nullptr) {
+    return Status::InvalidArgument(
+        "SystemClock required by TtlCompactionFilterFactory");
+  } else {
+    return CompactionFilterFactory::ValidateOptions(db_opts, cf_opts);
+  }
+}
+
+int RegisterTtlObjects(ObjectLibrary& library, const std::string& /*arg*/) {
+  library.AddFactory<MergeOperator>(
+      TtlMergeOperator::kClassName(),
+      [](const std::string& /*uri*/, std::unique_ptr<MergeOperator>* guard,
+         std::string* /* errmsg */) {
+        guard->reset(new TtlMergeOperator(nullptr, nullptr));
+        return guard->get();
+      });
+  library.AddFactory<CompactionFilterFactory>(
+      TtlCompactionFilterFactory::kClassName(),
+      [](const std::string& /*uri*/,
+         std::unique_ptr<CompactionFilterFactory>* guard,
+         std::string* /* errmsg */) {
+        guard->reset(new TtlCompactionFilterFactory(0, nullptr, nullptr));
+        return guard->get();
+      });
+  library.AddFactory<CompactionFilter>(
+      TtlCompactionFilter::kClassName(),
+      [](const std::string& /*uri*/,
+         std::unique_ptr<CompactionFilter>* /*guard*/,
+         std::string* /* errmsg */) {
+        return new TtlCompactionFilter(0, nullptr, nullptr);
+      });
+  size_t num_types;
+  return static_cast<int>(library.GetFactoryCount(&num_types));
+}
+// Open the db inside DBWithTTLImpl because options needs pointer to its ttl
+DBWithTTLImpl::DBWithTTLImpl(DB* db) : DBWithTTL(db), closed_(false) {}
+
+DBWithTTLImpl::~DBWithTTLImpl() {
+  if (!closed_) {
+    Close().PermitUncheckedError();
+  }
+}
+
+Status DBWithTTLImpl::Close() {
+  Status ret = Status::OK();
+  if (!closed_) {
+    Options default_options = GetOptions();
+    // Need to stop background compaction before getting rid of the filter
+    CancelAllBackgroundWork(db_, /* wait = */ true);
+    ret = db_->Close();
+    delete default_options.compaction_filter;
+    closed_ = true;
+  }
+  return ret;
+}
+
+void DBWithTTLImpl::RegisterTtlClasses() {
+  static std::once_flag once;
+  std::call_once(once, [&]() {
+    ObjectRegistry::Default()->AddLibrary("TTL", RegisterTtlObjects, "");
+  });
+}
+
+Status DBWithTTL::Open(const Options& options, const std::string& dbname,
+                       DBWithTTL** dbptr, int32_t ttl, bool read_only) {
+  DBOptions db_options(options);
+  ColumnFamilyOptions cf_options(options);
+  std::vector<ColumnFamilyDescriptor> column_families;
+  column_families.push_back(
+      ColumnFamilyDescriptor(kDefaultColumnFamilyName, cf_options));
+  std::vector<ColumnFamilyHandle*> handles;
+  Status s = DBWithTTL::Open(db_options, dbname, column_families, &handles,
+                             dbptr, {ttl}, read_only);
+  if (s.ok()) {
+    assert(handles.size() == 1);
+    // i can delete the handle since DBImpl is always holding a reference to
+    // default column family
+    delete handles[0];
+  }
+  return s;
+}
+
+Status DBWithTTL::Open(
+    const DBOptions& db_options, const std::string& dbname,
+    const std::vector<ColumnFamilyDescriptor>& column_families,
+    std::vector<ColumnFamilyHandle*>* handles, DBWithTTL** dbptr,
+    const std::vector<int32_t>& ttls, bool read_only) {
+  DBWithTTLImpl::RegisterTtlClasses();
+  if (ttls.size() != column_families.size()) {
+    return Status::InvalidArgument(
+        "ttls size has to be the same as number of column families");
+  }
+
+  SystemClock* clock = (db_options.env == nullptr)
+                           ? SystemClock::Default().get()
+                           : db_options.env->GetSystemClock().get();
+
+  std::vector<ColumnFamilyDescriptor> column_families_sanitized =
+      column_families;
+  for (size_t i = 0; i < column_families_sanitized.size(); ++i) {
+    DBWithTTLImpl::SanitizeOptions(
+        ttls[i], &column_families_sanitized[i].options, clock);
+  }
+  DB* db;
+
+  Status st;
+  if (read_only) {
+    st = DB::OpenForReadOnly(db_options, dbname, column_families_sanitized,
+                             handles, &db);
+  } else {
+    st = DB::Open(db_options, dbname, column_families_sanitized, handles, &db);
+  }
+  if (st.ok()) {
+    *dbptr = new DBWithTTLImpl(db);
+  } else {
+    *dbptr = nullptr;
+  }
+  return st;
+}
+
+Status DBWithTTLImpl::CreateColumnFamilyWithTtl(
+    const ColumnFamilyOptions& options, const std::string& column_family_name,
+    ColumnFamilyHandle** handle, int ttl) {
+  RegisterTtlClasses();
+  ColumnFamilyOptions sanitized_options = options;
+  DBWithTTLImpl::SanitizeOptions(ttl, &sanitized_options,
+                                 GetEnv()->GetSystemClock().get());
+
+  return DBWithTTL::CreateColumnFamily(sanitized_options, column_family_name,
+                                       handle);
+}
+
+Status DBWithTTLImpl::CreateColumnFamily(const ColumnFamilyOptions& options,
+                                         const std::string& column_family_name,
+                                         ColumnFamilyHandle** handle) {
+  return CreateColumnFamilyWithTtl(options, column_family_name, handle, 0);
+}
+
+// Appends the current timestamp to the string.
+// Returns false if could not get the current_time, true if append succeeds
+Status DBWithTTLImpl::AppendTS(const Slice& val, std::string* val_with_ts,
+                               SystemClock* clock) {
+  val_with_ts->reserve(kTSLength + val.size());
+  char ts_string[kTSLength];
+  int64_t curtime;
+  Status st = clock->GetCurrentTime(&curtime);
+  if (!st.ok()) {
+    return st;
+  }
+  EncodeFixed32(ts_string, (int32_t)curtime);
+  val_with_ts->append(val.data(), val.size());
+  val_with_ts->append(ts_string, kTSLength);
+  return st;
+}
+
+// Returns corruption if the length of the string is lesser than timestamp, or
+// timestamp refers to a time lesser than ttl-feature release time
+Status DBWithTTLImpl::SanityCheckTimestamp(const Slice& str) {
+  if (str.size() < kTSLength) {
+    return Status::Corruption("Error: value's length less than timestamp's\n");
+  }
+  // Checks that TS is not lesser than kMinTimestamp
+  // Gaurds against corruption & normal database opened incorrectly in ttl mode
+  int32_t timestamp_value = DecodeFixed32(str.data() + str.size() - kTSLength);
+  if (timestamp_value < kMinTimestamp) {
+    return Status::Corruption("Error: Timestamp < ttl feature release time!\n");
+  }
+  return Status::OK();
+}
+
+// Checks if the string is stale or not according to TTl provided
+bool DBWithTTLImpl::IsStale(const Slice& value, int32_t ttl,
+                            SystemClock* clock) {
+  if (ttl <= 0) {  // Data is fresh if TTL is non-positive
+    return false;
+  }
+  int64_t curtime;
+  if (!clock->GetCurrentTime(&curtime).ok()) {
+    return false;  // Treat the data as fresh if could not get current time
+  }
+  int32_t timestamp_value =
+      DecodeFixed32(value.data() + value.size() - kTSLength);
+  return (timestamp_value + ttl) < curtime;
+}
+
+// Strips the TS from the end of the slice
+Status DBWithTTLImpl::StripTS(PinnableSlice* pinnable_val) {
+  if (pinnable_val->size() < kTSLength) {
+    return Status::Corruption("Bad timestamp in key-value");
+  }
+  // Erasing characters which hold the TS
+  pinnable_val->remove_suffix(kTSLength);
+  return Status::OK();
+}
+
+// Strips the TS from the end of the string
+Status DBWithTTLImpl::StripTS(std::string* str) {
+  if (str->length() < kTSLength) {
+    return Status::Corruption("Bad timestamp in key-value");
+  }
+  // Erasing characters which hold the TS
+  str->erase(str->length() - kTSLength, kTSLength);
+  return Status::OK();
+}
+
+Status DBWithTTLImpl::Put(const WriteOptions& options,
+                          ColumnFamilyHandle* column_family, const Slice& key,
+                          const Slice& val) {
+  WriteBatch batch;
+  Status st = batch.Put(column_family, key, val);
+  if (st.ok()) {
+    st = Write(options, &batch);
+  }
+  return st;
+}
+
+Status DBWithTTLImpl::Get(const ReadOptions& options,
+                          ColumnFamilyHandle* column_family, const Slice& key,
+                          PinnableSlice* value) {
+  Status st = db_->Get(options, column_family, key, value);
+  if (!st.ok()) {
+    return st;
+  }
+  st = SanityCheckTimestamp(*value);
+  if (!st.ok()) {
+    return st;
+  }
+  return StripTS(value);
+}
+
+std::vector<Status> DBWithTTLImpl::MultiGet(
+    const ReadOptions& options,
+    const std::vector<ColumnFamilyHandle*>& column_family,
+    const std::vector<Slice>& keys, std::vector<std::string>* values) {
+  auto statuses = db_->MultiGet(options, column_family, keys, values);
+  for (size_t i = 0; i < keys.size(); ++i) {
+    if (!statuses[i].ok()) {
+      continue;
+    }
+    statuses[i] = SanityCheckTimestamp((*values)[i]);
+    if (!statuses[i].ok()) {
+      continue;
+    }
+    statuses[i] = StripTS(&(*values)[i]);
+  }
+  return statuses;
+}
+
+bool DBWithTTLImpl::KeyMayExist(const ReadOptions& options,
+                                ColumnFamilyHandle* column_family,
+                                const Slice& key, std::string* value,
+                                bool* value_found) {
+  bool ret = db_->KeyMayExist(options, column_family, key, value, value_found);
+  if (ret && value != nullptr && value_found != nullptr && *value_found) {
+    if (!SanityCheckTimestamp(*value).ok() || !StripTS(value).ok()) {
+      return false;
+    }
+  }
+  return ret;
+}
+
+Status DBWithTTLImpl::Merge(const WriteOptions& options,
+                            ColumnFamilyHandle* column_family, const Slice& key,
+                            const Slice& value) {
+  WriteBatch batch;
+  Status st = batch.Merge(column_family, key, value);
+  if (st.ok()) {
+    st = Write(options, &batch);
+  }
+  return st;
+}
+
+Status DBWithTTLImpl::Write(const WriteOptions& opts, WriteBatch* updates) {
+  class Handler : public WriteBatch::Handler {
+   public:
+    explicit Handler(SystemClock* clock) : clock_(clock) {}
+    WriteBatch updates_ttl;
+    Status PutCF(uint32_t column_family_id, const Slice& key,
+                 const Slice& value) override {
+      std::string value_with_ts;
+      Status st = AppendTS(value, &value_with_ts, clock_);
+      if (!st.ok()) {
+        return st;
+      }
+      return WriteBatchInternal::Put(&updates_ttl, column_family_id, key,
+                                     value_with_ts);
+    }
+    Status MergeCF(uint32_t column_family_id, const Slice& key,
+                   const Slice& value) override {
+      std::string value_with_ts;
+      Status st = AppendTS(value, &value_with_ts, clock_);
+      if (!st.ok()) {
+        return st;
+      }
+      return WriteBatchInternal::Merge(&updates_ttl, column_family_id, key,
+                                       value_with_ts);
+    }
+    Status DeleteCF(uint32_t column_family_id, const Slice& key) override {
+      return WriteBatchInternal::Delete(&updates_ttl, column_family_id, key);
+    }
+    Status DeleteRangeCF(uint32_t column_family_id, const Slice& begin_key,
+                         const Slice& end_key) override {
+      return WriteBatchInternal::DeleteRange(&updates_ttl, column_family_id,
+                                             begin_key, end_key);
+    }
+    void LogData(const Slice& blob) override { updates_ttl.PutLogData(blob); }
+
+   private:
+    SystemClock* clock_;
+  };
+  Handler handler(GetEnv()->GetSystemClock().get());
+  Status st = updates->Iterate(&handler);
+  if (!st.ok()) {
+    return st;
+  } else {
+    return db_->Write(opts, &(handler.updates_ttl));
+  }
+}
+
+Iterator* DBWithTTLImpl::NewIterator(const ReadOptions& opts,
+                                     ColumnFamilyHandle* column_family) {
+  return new TtlIterator(db_->NewIterator(opts, column_family));
+}
+
+void DBWithTTLImpl::SetTtl(ColumnFamilyHandle* h, int32_t ttl) {
+  std::shared_ptr<TtlCompactionFilterFactory> filter;
+  Options opts;
+  opts = GetOptions(h);
+  filter = std::static_pointer_cast<TtlCompactionFilterFactory>(
+      opts.compaction_filter_factory);
+  if (!filter) return;
+  filter->SetTtl(ttl);
+}
+
+}  // namespace ROCKSDB_NAMESPACE
+#endif  // ROCKSDB_LITE
diff --git a/src/rocksdb/utilities/ttl/db_ttl_impl.h b/src/rocksdb/utilities/ttl/db_ttl_impl.h
new file mode 100644
index 000000000..dd67a6ddc
--- /dev/null
+++ b/src/rocksdb/utilities/ttl/db_ttl_impl.h
@@ -0,0 +1,245 @@
+// Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.
+// Copyright (c) 2011 The LevelDB Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file. See the AUTHORS file for names of contributors.
+
+#pragma once
+
+#ifndef ROCKSDB_LITE
+#include <deque>
+#include <string>
+#include <vector>
+
+#include "db/db_impl/db_impl.h"
+#include "rocksdb/compaction_filter.h"
+#include "rocksdb/db.h"
+#include "rocksdb/merge_operator.h"
+#include "rocksdb/system_clock.h"
+#include "rocksdb/utilities/db_ttl.h"
+#include "utilities/compaction_filters/layered_compaction_filter_base.h"
+
+#ifdef _WIN32
+// Windows API macro interference
+#undef GetCurrentTime
+#endif
+
+namespace ROCKSDB_NAMESPACE {
+struct ConfigOptions;
+class ObjectLibrary;
+class ObjectRegistry;
+class DBWithTTLImpl : public DBWithTTL {
+ public:
+  static void SanitizeOptions(int32_t ttl, ColumnFamilyOptions* options,
+                              SystemClock* clock);
+
+  static void RegisterTtlClasses();
+  explicit DBWithTTLImpl(DB* db);
+
+  virtual ~DBWithTTLImpl();
+
+  virtual Status Close() override;
+
+  Status CreateColumnFamilyWithTtl(const ColumnFamilyOptions& options,
+                                   const std::string& column_family_name,
+                                   ColumnFamilyHandle** handle,
+                                   int ttl) override;
+
+  Status CreateColumnFamily(const ColumnFamilyOptions& options,
+                            const std::string& column_family_name,
+                            ColumnFamilyHandle** handle) override;
+
+  using StackableDB::Put;
+  virtual Status Put(const WriteOptions& options,
+                     ColumnFamilyHandle* column_family, const Slice& key,
+                     const Slice& val) override;
+
+  using StackableDB::Get;
+  virtual Status Get(const ReadOptions& options,
+                     ColumnFamilyHandle* column_family, const Slice& key,
+                     PinnableSlice* value) override;
+
+  using StackableDB::MultiGet;
+  virtual std::vector<Status> MultiGet(
+      const ReadOptions& options,
+      const std::vector<ColumnFamilyHandle*>& column_family,
+      const std::vector<Slice>& keys,
+      std::vector<std::string>* values) override;
+
+  using StackableDB::KeyMayExist;
+  virtual bool KeyMayExist(const ReadOptions& options,
+                           ColumnFamilyHandle* column_family, const Slice& key,
+                           std::string* value,
+                           bool* value_found = nullptr) override;
+
+  using StackableDB::Merge;
+  virtual Status Merge(const WriteOptions& options,
+                       ColumnFamilyHandle* column_family, const Slice& key,
+                       const Slice& value) override;
+
+  virtual Status Write(const WriteOptions& opts, WriteBatch* updates) override;
+
+  using StackableDB::NewIterator;
+  virtual Iterator* NewIterator(const ReadOptions& opts,
+                                ColumnFamilyHandle* column_family) override;
+
+  virtual DB* GetBaseDB() override { return db_; }
+
+  static bool IsStale(const Slice& value, int32_t ttl, SystemClock* clock);
+
+  static Status AppendTS(const Slice& val, std::string* val_with_ts,
+                         SystemClock* clock);
+
+  static Status SanityCheckTimestamp(const Slice& str);
+
+  static Status StripTS(std::string* str);
+
+  static Status StripTS(PinnableSlice* str);
+
+  static const uint32_t kTSLength = sizeof(int32_t);  // size of timestamp
+
+  static const int32_t kMinTimestamp = 1368146402;  // 05/09/2013:5:40PM GMT-8
+
+  static const int32_t kMaxTimestamp = 2147483647;  // 01/18/2038:7:14PM GMT-8
+
+  void SetTtl(int32_t ttl) override { SetTtl(DefaultColumnFamily(), ttl); }
+
+  void SetTtl(ColumnFamilyHandle* h, int32_t ttl) override;
+
+ private:
+  // remember whether the Close completes or not
+  bool closed_;
+};
+
+class TtlIterator : public Iterator {
+ public:
+  explicit TtlIterator(Iterator* iter) : iter_(iter) { assert(iter_); }
+
+  ~TtlIterator() { delete iter_; }
+
+  bool Valid() const override { return iter_->Valid(); }
+
+  void SeekToFirst() override { iter_->SeekToFirst(); }
+
+  void SeekToLast() override { iter_->SeekToLast(); }
+
+  void Seek(const Slice& target) override { iter_->Seek(target); }
+
+  void SeekForPrev(const Slice& target) override { iter_->SeekForPrev(target); }
+
+  void Next() override { iter_->Next(); }
+
+  void Prev() override { iter_->Prev(); }
+
+  Slice key() const override { return iter_->key(); }
+
+  int32_t ttl_timestamp() const {
+    return DecodeFixed32(iter_->value().data() + iter_->value().size() -
+                         DBWithTTLImpl::kTSLength);
+  }
+
+  Slice value() const override {
+    // TODO: handle timestamp corruption like in general iterator semantics
+    assert(DBWithTTLImpl::SanityCheckTimestamp(iter_->value()).ok());
+    Slice trimmed_value = iter_->value();
+    trimmed_value.size_ -= DBWithTTLImpl::kTSLength;
+    return trimmed_value;
+  }
+
+  Status status() const override { return iter_->status(); }
+
+ private:
+  Iterator* iter_;
+};
+
+class TtlCompactionFilter : public LayeredCompactionFilterBase {
+ public:
+  TtlCompactionFilter(int32_t ttl, SystemClock* clock,
+                      const CompactionFilter* _user_comp_filter,
+                      std::unique_ptr<const CompactionFilter>
+                          _user_comp_filter_from_factory = nullptr);
+
+  virtual bool Filter(int level, const Slice& key, const Slice& old_val,
+                      std::string* new_val, bool* value_changed) const override;
+
+  const char* Name() const override { return kClassName(); }
+  static const char* kClassName() { return "TtlCompactionFilter"; }
+  bool IsInstanceOf(const std::string& name) const override {
+    if (name == "Delete By TTL") {
+      return true;
+    } else {
+      return LayeredCompactionFilterBase::IsInstanceOf(name);
+    }
+  }
+
+  Status PrepareOptions(const ConfigOptions& config_options) override;
+  Status ValidateOptions(const DBOptions& db_opts,
+                         const ColumnFamilyOptions& cf_opts) const override;
+
+ private:
+  int32_t ttl_;
+  SystemClock* clock_;
+};
+
+class TtlCompactionFilterFactory : public CompactionFilterFactory {
+ public:
+  TtlCompactionFilterFactory(
+      int32_t ttl, SystemClock* clock,
+      std::shared_ptr<CompactionFilterFactory> comp_filter_factory);
+
+  std::unique_ptr<CompactionFilter> CreateCompactionFilter(
+      const CompactionFilter::Context& context) override;
+  void SetTtl(int32_t ttl) { ttl_ = ttl; }
+
+  const char* Name() const override { return kClassName(); }
+  static const char* kClassName() { return "TtlCompactionFilterFactory"; }
+  Status PrepareOptions(const ConfigOptions& config_options) override;
+  Status ValidateOptions(const DBOptions& db_opts,
+                         const ColumnFamilyOptions& cf_opts) const override;
+  const Customizable* Inner() const override {
+    return user_comp_filter_factory_.get();
+  }
+
+ private:
+  int32_t ttl_;
+  SystemClock* clock_;
+  std::shared_ptr<CompactionFilterFactory> user_comp_filter_factory_;
+};
+
+class TtlMergeOperator : public MergeOperator {
+ public:
+  explicit TtlMergeOperator(const std::shared_ptr<MergeOperator>& merge_op,
+                            SystemClock* clock);
+
+  bool FullMergeV2(const MergeOperationInput& merge_in,
+                   MergeOperationOutput* merge_out) const override;
+
+  bool PartialMergeMulti(const Slice& key,
+                         const std::deque<Slice>& operand_list,
+                         std::string* new_value, Logger* logger) const override;
+
+  static const char* kClassName() { return "TtlMergeOperator"; }
+
+  const char* Name() const override { return kClassName(); }
+  bool IsInstanceOf(const std::string& name) const override {
+    if (name == "Merge By TTL") {
+      return true;
+    } else {
+      return MergeOperator::IsInstanceOf(name);
+    }
+  }
+
+  Status PrepareOptions(const ConfigOptions& config_options) override;
+  Status ValidateOptions(const DBOptions& db_opts,
+                         const ColumnFamilyOptions& cf_opts) const override;
+  const Customizable* Inner() const override { return user_merge_op_.get(); }
+
+ private:
+  std::shared_ptr<MergeOperator> user_merge_op_;
+  SystemClock* clock_;
+};
+extern "C" {
+int RegisterTtlObjects(ObjectLibrary& library, const std::string& /*arg*/);
+}  // extern "C"
+
+}  // namespace ROCKSDB_NAMESPACE
+#endif  // ROCKSDB_LITE
diff --git a/src/rocksdb/utilities/ttl/ttl_test.cc b/src/rocksdb/utilities/ttl/ttl_test.cc
new file mode 100644
index 000000000..a42e0acb4
--- /dev/null
+++ b/src/rocksdb/utilities/ttl/ttl_test.cc
@@ -0,0 +1,912 @@
+// Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.
+// Copyright (c) 2011 The LevelDB Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file. See the AUTHORS file for names of contributors.
+
+#ifndef ROCKSDB_LITE
+
+#include <map>
+#include <memory>
+
+#include "rocksdb/compaction_filter.h"
+#include "rocksdb/convenience.h"
+#include "rocksdb/merge_operator.h"
+#include "rocksdb/utilities/db_ttl.h"
+#include "rocksdb/utilities/object_registry.h"
+#include "test_util/testharness.h"
+#include "util/string_util.h"
+#include "utilities/merge_operators/bytesxor.h"
+#include "utilities/ttl/db_ttl_impl.h"
+#ifndef OS_WIN
+#include <unistd.h>
+#endif
+
+namespace ROCKSDB_NAMESPACE {
+
+namespace {
+
+using KVMap = std::map<std::string, std::string>;
+
+enum BatchOperation { OP_PUT = 0, OP_DELETE = 1 };
+}  // namespace
+
+class SpecialTimeEnv : public EnvWrapper {
+ public:
+  explicit SpecialTimeEnv(Env* base) : EnvWrapper(base) {
+    EXPECT_OK(base->GetCurrentTime(&current_time_));
+  }
+  const char* Name() const override { return "SpecialTimeEnv"; }
+  void Sleep(int64_t sleep_time) { current_time_ += sleep_time; }
+  Status GetCurrentTime(int64_t* current_time) override {
+    *current_time = current_time_;
+    return Status::OK();
+  }
+
+ private:
+  int64_t current_time_ = 0;
+};
+
+class TtlTest : public testing::Test {
+ public:
+  TtlTest() {
+    env_.reset(new SpecialTimeEnv(Env::Default()));
+    dbname_ = test::PerThreadDBPath("db_ttl");
+    options_.create_if_missing = true;
+    options_.env = env_.get();
+    // ensure that compaction is kicked in to always strip timestamp from kvs
+    options_.max_compaction_bytes = 1;
+    // compaction should take place always from level0 for determinism
+    db_ttl_ = nullptr;
+    EXPECT_OK(DestroyDB(dbname_, Options()));
+  }
+
+  ~TtlTest() override {
+    CloseTtl();
+    EXPECT_OK(DestroyDB(dbname_, Options()));
+  }
+
+  // Open database with TTL support when TTL not provided with db_ttl_ pointer
+  void OpenTtl() {
+    ASSERT_TRUE(db_ttl_ ==
+                nullptr);  //  db should be closed before opening again
+    ASSERT_OK(DBWithTTL::Open(options_, dbname_, &db_ttl_));
+  }
+
+  // Open database with TTL support when TTL provided with db_ttl_ pointer
+  void OpenTtl(int32_t ttl) {
+    ASSERT_TRUE(db_ttl_ == nullptr);
+    ASSERT_OK(DBWithTTL::Open(options_, dbname_, &db_ttl_, ttl));
+  }
+
+  // Open with TestFilter compaction filter
+  void OpenTtlWithTestCompaction(int32_t ttl) {
+    options_.compaction_filter_factory =
+        std::shared_ptr<CompactionFilterFactory>(
+            new TestFilterFactory(kSampleSize_, kNewValue_));
+    OpenTtl(ttl);
+  }
+
+  // Open database with TTL support in read_only mode
+  void OpenReadOnlyTtl(int32_t ttl) {
+    ASSERT_TRUE(db_ttl_ == nullptr);
+    ASSERT_OK(DBWithTTL::Open(options_, dbname_, &db_ttl_, ttl, true));
+  }
+
+  // Call db_ttl_->Close() before delete db_ttl_
+  void CloseTtl() { CloseTtlHelper(true); }
+
+  // No db_ttl_->Close() before delete db_ttl_
+  void CloseTtlNoDBClose() { CloseTtlHelper(false); }
+
+  void CloseTtlHelper(bool close_db) {
+    if (db_ttl_ != nullptr) {
+      if (close_db) {
+        EXPECT_OK(db_ttl_->Close());
+      }
+      delete db_ttl_;
+      db_ttl_ = nullptr;
+    }
+  }
+
+  // Populates and returns a kv-map
+  void MakeKVMap(int64_t num_entries) {
+    kvmap_.clear();
+    int digits = 1;
+    for (int64_t dummy = num_entries; dummy /= 10; ++digits) {
+    }
+    int digits_in_i = 1;
+    for (int64_t i = 0; i < num_entries; i++) {
+      std::string key = "key";
+      std::string value = "value";
+      if (i % 10 == 0) {
+        digits_in_i++;
+      }
+      for (int j = digits_in_i; j < digits; j++) {
+        key.append("0");
+        value.append("0");
+      }
+      AppendNumberTo(&key, i);
+      AppendNumberTo(&value, i);
+      kvmap_[key] = value;
+    }
+    ASSERT_EQ(static_cast<int64_t>(kvmap_.size()),
+              num_entries);  // check all insertions done
+  }
+
+  // Makes a write-batch with key-vals from kvmap_ and 'Write''s it
+  void MakePutWriteBatch(const BatchOperation* batch_ops, int64_t num_ops) {
+    ASSERT_LE(num_ops, static_cast<int64_t>(kvmap_.size()));
+    static WriteOptions wopts;
+    static FlushOptions flush_opts;
+    WriteBatch batch;
+    kv_it_ = kvmap_.begin();
+    for (int64_t i = 0; i < num_ops && kv_it_ != kvmap_.end(); i++, ++kv_it_) {
+      switch (batch_ops[i]) {
+        case OP_PUT:
+          ASSERT_OK(batch.Put(kv_it_->first, kv_it_->second));
+          break;
+        case OP_DELETE:
+          ASSERT_OK(batch.Delete(kv_it_->first));
+          break;
+        default:
+          FAIL();
+      }
+    }
+    ASSERT_OK(db_ttl_->Write(wopts, &batch));
+    ASSERT_OK(db_ttl_->Flush(flush_opts));
+  }
+
+  // Puts num_entries starting from start_pos_map from kvmap_ into the database
+  void PutValues(int64_t start_pos_map, int64_t num_entries, bool flush = true,
+                 ColumnFamilyHandle* cf = nullptr) {
+    ASSERT_TRUE(db_ttl_);
+    ASSERT_LE(start_pos_map + num_entries, static_cast<int64_t>(kvmap_.size()));
+    static WriteOptions wopts;
+    static FlushOptions flush_opts;
+    kv_it_ = kvmap_.begin();
+    advance(kv_it_, start_pos_map);
+    for (int64_t i = 0; kv_it_ != kvmap_.end() && i < num_entries;
+         i++, ++kv_it_) {
+      ASSERT_OK(cf == nullptr
+                    ? db_ttl_->Put(wopts, kv_it_->first, kv_it_->second)
+                    : db_ttl_->Put(wopts, cf, kv_it_->first, kv_it_->second));
+    }
+    // Put a mock kv at the end because CompactionFilter doesn't delete last key
+    ASSERT_OK(cf == nullptr ? db_ttl_->Put(wopts, "keymock", "valuemock")
+                            : db_ttl_->Put(wopts, cf, "keymock", "valuemock"));
+    if (flush) {
+      if (cf == nullptr) {
+        ASSERT_OK(db_ttl_->Flush(flush_opts));
+      } else {
+        ASSERT_OK(db_ttl_->Flush(flush_opts, cf));
+      }
+    }
+  }
+
+  // Runs a manual compaction
+  Status ManualCompact(ColumnFamilyHandle* cf = nullptr) {
+    assert(db_ttl_);
+    if (cf == nullptr) {
+      return db_ttl_->CompactRange(CompactRangeOptions(), nullptr, nullptr);
+    } else {
+      return db_ttl_->CompactRange(CompactRangeOptions(), cf, nullptr, nullptr);
+    }
+  }
+
+  // Runs a DeleteRange
+  void MakeDeleteRange(std::string start, std::string end,
+                       ColumnFamilyHandle* cf = nullptr) {
+    ASSERT_TRUE(db_ttl_);
+    static WriteOptions wops;
+    WriteBatch wb;
+    ASSERT_OK(cf == nullptr
+                  ? wb.DeleteRange(db_ttl_->DefaultColumnFamily(), start, end)
+                  : wb.DeleteRange(cf, start, end));
+    ASSERT_OK(db_ttl_->Write(wops, &wb));
+  }
+
+  // checks the whole kvmap_ to return correct values using KeyMayExist
+  void SimpleKeyMayExistCheck() {
+    static ReadOptions ropts;
+    bool value_found;
+    std::string val;
+    for (auto& kv : kvmap_) {
+      bool ret = db_ttl_->KeyMayExist(ropts, kv.first, &val, &value_found);
+      if (ret == false || value_found == false) {
+        fprintf(stderr,
+                "KeyMayExist could not find key=%s in the database but"
+                " should have\n",
+                kv.first.c_str());
+        FAIL();
+      } else if (val.compare(kv.second) != 0) {
+        fprintf(stderr,
+                " value for key=%s present in database is %s but"
+                " should be %s\n",
+                kv.first.c_str(), val.c_str(), kv.second.c_str());
+        FAIL();
+      }
+    }
+  }
+
+  // checks the whole kvmap_ to return correct values using MultiGet
+  void SimpleMultiGetTest() {
+    static ReadOptions ropts;
+    std::vector<Slice> keys;
+    std::vector<std::string> values;
+
+    for (auto& kv : kvmap_) {
+      keys.emplace_back(kv.first);
+    }
+
+    auto statuses = db_ttl_->MultiGet(ropts, keys, &values);
+    size_t i = 0;
+    for (auto& kv : kvmap_) {
+      ASSERT_OK(statuses[i]);
+      ASSERT_EQ(values[i], kv.second);
+      ++i;
+    }
+  }
+
+  void CompactCheck(int64_t st_pos, int64_t span, bool check = true,
+                    bool test_compaction_change = false,
+                    ColumnFamilyHandle* cf = nullptr) {
+    static ReadOptions ropts;
+    kv_it_ = kvmap_.begin();
+    advance(kv_it_, st_pos);
+    std::string v;
+    for (int64_t i = 0; kv_it_ != kvmap_.end() && i < span; i++, ++kv_it_) {
+      Status s = (cf == nullptr) ? db_ttl_->Get(ropts, kv_it_->first, &v)
+                                 : db_ttl_->Get(ropts, cf, kv_it_->first, &v);
+      if (s.ok() != check) {
+        fprintf(stderr, "key=%s ", kv_it_->first.c_str());
+        if (!s.ok()) {
+          fprintf(stderr, "is absent from db but was expected to be present\n");
+        } else {
+          fprintf(stderr, "is present in db but was expected to be absent\n");
+        }
+        FAIL();
+      } else if (s.ok()) {
+        if (test_compaction_change && v.compare(kNewValue_) != 0) {
+          fprintf(stderr,
+                  " value for key=%s present in database is %s but "
+                  " should be %s\n",
+                  kv_it_->first.c_str(), v.c_str(), kNewValue_.c_str());
+          FAIL();
+        } else if (!test_compaction_change && v.compare(kv_it_->second) != 0) {
+          fprintf(stderr,
+                  " value for key=%s present in database is %s but "
+                  " should be %s\n",
+                  kv_it_->first.c_str(), v.c_str(), kv_it_->second.c_str());
+          FAIL();
+        }
+      }
+    }
+  }
+  // Sleeps for slp_tim then runs a manual compaction
+  // Checks span starting from st_pos from kvmap_ in the db and
+  // Gets should return true if check is true and false otherwise
+  // Also checks that value that we got is the same as inserted; and =kNewValue
+  //   if test_compaction_change is true
+  void SleepCompactCheck(int slp_tim, int64_t st_pos, int64_t span,
+                         bool check = true, bool test_compaction_change = false,
+                         ColumnFamilyHandle* cf = nullptr) {
+    ASSERT_TRUE(db_ttl_);
+
+    env_->Sleep(slp_tim);
+    ASSERT_OK(ManualCompact(cf));
+    CompactCheck(st_pos, span, check, test_compaction_change, cf);
+  }
+
+  // Similar as SleepCompactCheck but uses TtlIterator to read from db
+  void SleepCompactCheckIter(int slp, int st_pos, int64_t span,
+                             bool check = true) {
+    ASSERT_TRUE(db_ttl_);
+    env_->Sleep(slp);
+    ASSERT_OK(ManualCompact());
+    static ReadOptions ropts;
+    Iterator* dbiter = db_ttl_->NewIterator(ropts);
+    kv_it_ = kvmap_.begin();
+    advance(kv_it_, st_pos);
+
+    dbiter->Seek(kv_it_->first);
+    if (!check) {
+      if (dbiter->Valid()) {
+        ASSERT_NE(dbiter->value().compare(kv_it_->second), 0);
+      }
+    } else {  // dbiter should have found out kvmap_[st_pos]
+      for (int64_t i = st_pos; kv_it_ != kvmap_.end() && i < st_pos + span;
+           i++, ++kv_it_) {
+        ASSERT_TRUE(dbiter->Valid());
+        ASSERT_EQ(dbiter->value().compare(kv_it_->second), 0);
+        dbiter->Next();
+      }
+    }
+    ASSERT_OK(dbiter->status());
+    delete dbiter;
+  }
+
+  // Set ttl on open db
+  void SetTtl(int32_t ttl, ColumnFamilyHandle* cf = nullptr) {
+    ASSERT_TRUE(db_ttl_);
+    cf == nullptr ? db_ttl_->SetTtl(ttl) : db_ttl_->SetTtl(cf, ttl);
+  }
+
+  class TestFilter : public CompactionFilter {
+   public:
+    TestFilter(const int64_t kSampleSize, const std::string& kNewValue)
+        : kSampleSize_(kSampleSize), kNewValue_(kNewValue) {}
+
+    // Works on keys of the form "key<number>"
+    // Drops key if number at the end of key is in [0, kSampleSize_/3),
+    // Keeps key if it is in [kSampleSize_/3, 2*kSampleSize_/3),
+    // Change value if it is in [2*kSampleSize_/3, kSampleSize_)
+    // Eg. kSampleSize_=6. Drop:key0-1...Keep:key2-3...Change:key4-5...
+    bool Filter(int /*level*/, const Slice& key, const Slice& /*value*/,
+                std::string* new_value, bool* value_changed) const override {
+      assert(new_value != nullptr);
+
+      std::string search_str = "0123456789";
+      std::string key_string = key.ToString();
+      size_t pos = key_string.find_first_of(search_str);
+      int num_key_end;
+      if (pos != std::string::npos) {
+        auto key_substr = key_string.substr(pos, key.size() - pos);
+#ifndef CYGWIN
+        num_key_end = std::stoi(key_substr);
+#else
+        num_key_end = std::strtol(key_substr.c_str(), 0, 10);
+#endif
+
+      } else {
+        return false;  // Keep keys not matching the format "key<NUMBER>"
+      }
+
+      int64_t partition = kSampleSize_ / 3;
+      if (num_key_end < partition) {
+        return true;
+      } else if (num_key_end < partition * 2) {
+        return false;
+      } else {
+        *new_value = kNewValue_;
+        *value_changed = true;
+        return false;
+      }
+    }
+
+    const char* Name() const override { return "TestFilter"; }
+
+   private:
+    const int64_t kSampleSize_;
+    const std::string kNewValue_;
+  };
+
+  class TestFilterFactory : public CompactionFilterFactory {
+   public:
+    TestFilterFactory(const int64_t kSampleSize, const std::string& kNewValue)
+        : kSampleSize_(kSampleSize), kNewValue_(kNewValue) {}
+
+    std::unique_ptr<CompactionFilter> CreateCompactionFilter(
+        const CompactionFilter::Context& /*context*/) override {
+      return std::unique_ptr<CompactionFilter>(
+          new TestFilter(kSampleSize_, kNewValue_));
+    }
+
+    const char* Name() const override { return "TestFilterFactory"; }
+
+   private:
+    const int64_t kSampleSize_;
+    const std::string kNewValue_;
+  };
+
+  // Choose carefully so that Put, Gets & Compaction complete in 1 second buffer
+  static const int64_t kSampleSize_ = 100;
+  std::string dbname_;
+  DBWithTTL* db_ttl_;
+  std::unique_ptr<SpecialTimeEnv> env_;
+
+ private:
+  Options options_;
+  KVMap kvmap_;
+  KVMap::iterator kv_it_;
+  const std::string kNewValue_ = "new_value";
+  std::unique_ptr<CompactionFilter> test_comp_filter_;
+};  // class TtlTest
+
+// If TTL is non positive or not provided, the behaviour is TTL = infinity
+// This test opens the db 3 times with such default behavior and inserts a
+// bunch of kvs each time. All kvs should accumulate in the db till the end
+// Partitions the sample-size provided into 3 sets over boundary1 and boundary2
+TEST_F(TtlTest, NoEffect) {
+  MakeKVMap(kSampleSize_);
+  int64_t boundary1 = kSampleSize_ / 3;
+  int64_t boundary2 = 2 * boundary1;
+
+  OpenTtl();
+  PutValues(0, boundary1);             // T=0: Set1 never deleted
+  SleepCompactCheck(1, 0, boundary1);  // T=1: Set1 still there
+  CloseTtl();
+
+  OpenTtl(0);
+  PutValues(boundary1, boundary2 - boundary1);  // T=1: Set2 never deleted
+  SleepCompactCheck(1, 0, boundary2);           // T=2: Sets1 & 2 still there
+  CloseTtl();
+
+  OpenTtl(-1);
+  PutValues(boundary2, kSampleSize_ - boundary2);  // T=3: Set3 never deleted
+  SleepCompactCheck(1, 0, kSampleSize_, true);  // T=4: Sets 1,2,3 still there
+  CloseTtl();
+}
+
+// Rerun the NoEffect test with a different version of CloseTtl
+// function, where db is directly deleted without close.
+TEST_F(TtlTest, DestructWithoutClose) {
+  MakeKVMap(kSampleSize_);
+  int64_t boundary1 = kSampleSize_ / 3;
+  int64_t boundary2 = 2 * boundary1;
+
+  OpenTtl();
+  PutValues(0, boundary1);             // T=0: Set1 never deleted
+  SleepCompactCheck(1, 0, boundary1);  // T=1: Set1 still there
+  CloseTtlNoDBClose();
+
+  OpenTtl(0);
+  PutValues(boundary1, boundary2 - boundary1);  // T=1: Set2 never deleted
+  SleepCompactCheck(1, 0, boundary2);           // T=2: Sets1 & 2 still there
+  CloseTtlNoDBClose();
+
+  OpenTtl(-1);
+  PutValues(boundary2, kSampleSize_ - boundary2);  // T=3: Set3 never deleted
+  SleepCompactCheck(1, 0, kSampleSize_, true);  // T=4: Sets 1,2,3 still there
+  CloseTtlNoDBClose();
+}
+
+// Puts a set of values and checks its presence using Get during ttl
+TEST_F(TtlTest, PresentDuringTTL) {
+  MakeKVMap(kSampleSize_);
+
+  OpenTtl(2);                  // T=0:Open the db with ttl = 2
+  PutValues(0, kSampleSize_);  // T=0:Insert Set1. Delete at t=2
+  SleepCompactCheck(1, 0, kSampleSize_,
+                    true);  // T=1:Set1 should still be there
+  CloseTtl();
+}
+
+// Puts a set of values and checks its absence using Get after ttl
+TEST_F(TtlTest, AbsentAfterTTL) {
+  MakeKVMap(kSampleSize_);
+
+  OpenTtl(1);                  // T=0:Open the db with ttl = 2
+  PutValues(0, kSampleSize_);  // T=0:Insert Set1. Delete at t=2
+  SleepCompactCheck(2, 0, kSampleSize_, false);  // T=2:Set1 should not be there
+  CloseTtl();
+}
+
+// Resets the timestamp of a set of kvs by updating them and checks that they
+// are not deleted according to the old timestamp
+TEST_F(TtlTest, ResetTimestamp) {
+  MakeKVMap(kSampleSize_);
+
+  OpenTtl(3);
+  PutValues(0, kSampleSize_);             // T=0: Insert Set1. Delete at t=3
+  env_->Sleep(2);                         // T=2
+  PutValues(0, kSampleSize_);             // T=2: Insert Set1. Delete at t=5
+  SleepCompactCheck(2, 0, kSampleSize_);  // T=4: Set1 should still be there
+  CloseTtl();
+}
+
+// Similar to PresentDuringTTL but uses Iterator
+TEST_F(TtlTest, IterPresentDuringTTL) {
+  MakeKVMap(kSampleSize_);
+
+  OpenTtl(2);
+  PutValues(0, kSampleSize_);                 // T=0: Insert. Delete at t=2
+  SleepCompactCheckIter(1, 0, kSampleSize_);  // T=1: Set should be there
+  CloseTtl();
+}
+
+// Similar to AbsentAfterTTL but uses Iterator
+TEST_F(TtlTest, IterAbsentAfterTTL) {
+  MakeKVMap(kSampleSize_);
+
+  OpenTtl(1);
+  PutValues(0, kSampleSize_);  // T=0: Insert. Delete at t=1
+  SleepCompactCheckIter(2, 0, kSampleSize_, false);  // T=2: Should not be there
+  CloseTtl();
+}
+
+// Checks presence while opening the same db more than once with the same ttl
+// Note: The second open will open the same db
+TEST_F(TtlTest, MultiOpenSamePresent) {
+  MakeKVMap(kSampleSize_);
+
+  OpenTtl(2);
+  PutValues(0, kSampleSize_);  // T=0: Insert. Delete at t=2
+  CloseTtl();
+
+  OpenTtl(2);                             // T=0. Delete at t=2
+  SleepCompactCheck(1, 0, kSampleSize_);  // T=1: Set should be there
+  CloseTtl();
+}
+
+// Checks absence while opening the same db more than once with the same ttl
+// Note: The second open will open the same db
+TEST_F(TtlTest, MultiOpenSameAbsent) {
+  MakeKVMap(kSampleSize_);
+
+  OpenTtl(1);
+  PutValues(0, kSampleSize_);  // T=0: Insert. Delete at t=1
+  CloseTtl();
+
+  OpenTtl(1);                                    // T=0.Delete at t=1
+  SleepCompactCheck(2, 0, kSampleSize_, false);  // T=2: Set should not be there
+  CloseTtl();
+}
+
+// Checks presence while opening the same db more than once with bigger ttl
+TEST_F(TtlTest, MultiOpenDifferent) {
+  MakeKVMap(kSampleSize_);
+
+  OpenTtl(1);
+  PutValues(0, kSampleSize_);  // T=0: Insert. Delete at t=1
+  CloseTtl();
+
+  OpenTtl(3);                             // T=0: Set deleted at t=3
+  SleepCompactCheck(2, 0, kSampleSize_);  // T=2: Set should be there
+  CloseTtl();
+}
+
+// Checks presence during ttl in read_only mode
+TEST_F(TtlTest, ReadOnlyPresentForever) {
+  MakeKVMap(kSampleSize_);
+
+  OpenTtl(1);                  // T=0:Open the db normally
+  PutValues(0, kSampleSize_);  // T=0:Insert Set1. Delete at t=1
+  CloseTtl();
+
+  OpenReadOnlyTtl(1);
+  ASSERT_TRUE(db_ttl_);
+
+  env_->Sleep(2);
+  Status s = ManualCompact();  // T=2:Set1 should still be there
+  ASSERT_TRUE(s.IsNotSupported());
+  CompactCheck(0, kSampleSize_);
+  CloseTtl();
+}
+
+// Checks whether WriteBatch works well with TTL
+// Puts all kvs in kvmap_ in a batch and writes first, then deletes first half
+TEST_F(TtlTest, WriteBatchTest) {
+  MakeKVMap(kSampleSize_);
+  BatchOperation batch_ops[kSampleSize_];
+  for (int i = 0; i < kSampleSize_; i++) {
+    batch_ops[i] = OP_PUT;
+  }
+
+  OpenTtl(2);
+  MakePutWriteBatch(batch_ops, kSampleSize_);
+  for (int i = 0; i < kSampleSize_ / 2; i++) {
+    batch_ops[i] = OP_DELETE;
+  }
+  MakePutWriteBatch(batch_ops, kSampleSize_ / 2);
+  SleepCompactCheck(0, 0, kSampleSize_ / 2, false);
+  SleepCompactCheck(0, kSampleSize_ / 2, kSampleSize_ - kSampleSize_ / 2);
+  CloseTtl();
+}
+
+// Checks user's compaction filter for correctness with TTL logic
+TEST_F(TtlTest, CompactionFilter) {
+  MakeKVMap(kSampleSize_);
+
+  OpenTtlWithTestCompaction(1);
+  PutValues(0, kSampleSize_);  // T=0:Insert Set1. Delete at t=1
+  // T=2: TTL logic takes precedence over TestFilter:-Set1 should not be there
+  SleepCompactCheck(2, 0, kSampleSize_, false);
+  CloseTtl();
+
+  OpenTtlWithTestCompaction(3);
+  PutValues(0, kSampleSize_);  // T=0:Insert Set1.
+  int64_t partition = kSampleSize_ / 3;
+  SleepCompactCheck(1, 0, partition, false);                   // Part dropped
+  SleepCompactCheck(0, partition, partition);                  // Part kept
+  SleepCompactCheck(0, 2 * partition, partition, true, true);  // Part changed
+  CloseTtl();
+}
+
+// Insert some key-values which KeyMayExist should be able to get and check that
+// values returned are fine
+TEST_F(TtlTest, KeyMayExist) {
+  MakeKVMap(kSampleSize_);
+
+  OpenTtl();
+  PutValues(0, kSampleSize_, false);
+
+  SimpleKeyMayExistCheck();
+
+  CloseTtl();
+}
+
+TEST_F(TtlTest, MultiGetTest) {
+  MakeKVMap(kSampleSize_);
+
+  OpenTtl();
+  PutValues(0, kSampleSize_, false);
+
+  SimpleMultiGetTest();
+
+  CloseTtl();
+}
+
+TEST_F(TtlTest, ColumnFamiliesTest) {
+  DB* db;
+  Options options;
+  options.create_if_missing = true;
+  options.env = env_.get();
+
+  DB::Open(options, dbname_, &db);
+  ColumnFamilyHandle* handle;
+  ASSERT_OK(db->CreateColumnFamily(ColumnFamilyOptions(options),
+                                   "ttl_column_family", &handle));
+
+  delete handle;
+  delete db;
+
+  std::vector<ColumnFamilyDescriptor> column_families;
+  column_families.push_back(ColumnFamilyDescriptor(
+      kDefaultColumnFamilyName, ColumnFamilyOptions(options)));
+  column_families.push_back(ColumnFamilyDescriptor(
+      "ttl_column_family", ColumnFamilyOptions(options)));
+
+  std::vector<ColumnFamilyHandle*> handles;
+
+  ASSERT_OK(DBWithTTL::Open(DBOptions(options), dbname_, column_families,
+                            &handles, &db_ttl_, {3, 5}, false));
+  ASSERT_EQ(handles.size(), 2U);
+  ColumnFamilyHandle* new_handle;
+  ASSERT_OK(db_ttl_->CreateColumnFamilyWithTtl(options, "ttl_column_family_2",
+                                               &new_handle, 2));
+  handles.push_back(new_handle);
+
+  MakeKVMap(kSampleSize_);
+  PutValues(0, kSampleSize_, false, handles[0]);
+  PutValues(0, kSampleSize_, false, handles[1]);
+  PutValues(0, kSampleSize_, false, handles[2]);
+
+  // everything should be there after 1 second
+  SleepCompactCheck(1, 0, kSampleSize_, true, false, handles[0]);
+  SleepCompactCheck(0, 0, kSampleSize_, true, false, handles[1]);
+  SleepCompactCheck(0, 0, kSampleSize_, true, false, handles[2]);
+
+  // only column family 1 should be alive after 4 seconds
+  SleepCompactCheck(3, 0, kSampleSize_, false, false, handles[0]);
+  SleepCompactCheck(0, 0, kSampleSize_, true, false, handles[1]);
+  SleepCompactCheck(0, 0, kSampleSize_, false, false, handles[2]);
+
+  // nothing should be there after 6 seconds
+  SleepCompactCheck(2, 0, kSampleSize_, false, false, handles[0]);
+  SleepCompactCheck(0, 0, kSampleSize_, false, false, handles[1]);
+  SleepCompactCheck(0, 0, kSampleSize_, false, false, handles[2]);
+
+  for (auto h : handles) {
+    delete h;
+  }
+  delete db_ttl_;
+  db_ttl_ = nullptr;
+}
+
+// Puts a set of values and checks its absence using Get after ttl
+TEST_F(TtlTest, ChangeTtlOnOpenDb) {
+  MakeKVMap(kSampleSize_);
+
+  OpenTtl(1);  // T=0:Open the db with ttl = 2
+  SetTtl(3);
+  PutValues(0, kSampleSize_);  // T=0:Insert Set1. Delete at t=2
+  SleepCompactCheck(2, 0, kSampleSize_, true);  // T=2:Set1 should be there
+  CloseTtl();
+}
+
+// Test DeleteRange for DBWithTtl
+TEST_F(TtlTest, DeleteRangeTest) {
+  OpenTtl();
+  ASSERT_OK(db_ttl_->Put(WriteOptions(), "a", "val"));
+  MakeDeleteRange("a", "b");
+  ASSERT_OK(db_ttl_->Put(WriteOptions(), "c", "val"));
+  MakeDeleteRange("b", "d");
+  ASSERT_OK(db_ttl_->Put(WriteOptions(), "e", "val"));
+  MakeDeleteRange("d", "e");
+  // first iteration verifies query correctness in memtable, second verifies
+  // query correctness for a single SST file
+  for (int i = 0; i < 2; i++) {
+    if (i > 0) {
+      ASSERT_OK(db_ttl_->Flush(FlushOptions()));
+    }
+    std::string value;
+    ASSERT_TRUE(db_ttl_->Get(ReadOptions(), "a", &value).IsNotFound());
+    ASSERT_TRUE(db_ttl_->Get(ReadOptions(), "c", &value).IsNotFound());
+    ASSERT_OK(db_ttl_->Get(ReadOptions(), "e", &value));
+  }
+  CloseTtl();
+}
+
+class DummyFilter : public CompactionFilter {
+ public:
+  bool Filter(int /*level*/, const Slice& /*key*/, const Slice& /*value*/,
+              std::string* /*new_value*/,
+              bool* /*value_changed*/) const override {
+    return false;
+  }
+
+  const char* Name() const override { return kClassName(); }
+  static const char* kClassName() { return "DummyFilter"; }
+};
+
+class DummyFilterFactory : public CompactionFilterFactory {
+ public:
+  const char* Name() const override { return kClassName(); }
+  static const char* kClassName() { return "DummyFilterFactory"; }
+
+  std::unique_ptr<CompactionFilter> CreateCompactionFilter(
+      const CompactionFilter::Context&) override {
+    std::unique_ptr<CompactionFilter> f(new DummyFilter());
+    return f;
+  }
+};
+
+static int RegisterTestObjects(ObjectLibrary& library,
+                               const std::string& /*arg*/) {
+  library.AddFactory<CompactionFilter>(
+      "DummyFilter", [](const std::string& /*uri*/,
+                        std::unique_ptr<CompactionFilter>* /*guard*/,
+                        std::string* /* errmsg */) {
+        static DummyFilter dummy;
+        return &dummy;
+      });
+  library.AddFactory<CompactionFilterFactory>(
+      "DummyFilterFactory", [](const std::string& /*uri*/,
+                               std::unique_ptr<CompactionFilterFactory>* guard,
+                               std::string* /* errmsg */) {
+        guard->reset(new DummyFilterFactory());
+        return guard->get();
+      });
+  return 2;
+}
+
+class TtlOptionsTest : public testing::Test {
+ public:
+  TtlOptionsTest() {
+    config_options_.registry->AddLibrary("RegisterTtlObjects",
+                                         RegisterTtlObjects, "");
+    config_options_.registry->AddLibrary("RegisterTtlTestObjects",
+                                         RegisterTestObjects, "");
+  }
+  ConfigOptions config_options_;
+};
+
+TEST_F(TtlOptionsTest, LoadTtlCompactionFilter) {
+  const CompactionFilter* filter = nullptr;
+
+  ASSERT_OK(CompactionFilter::CreateFromString(
+      config_options_, TtlCompactionFilter::kClassName(), &filter));
+  ASSERT_NE(filter, nullptr);
+  ASSERT_STREQ(filter->Name(), TtlCompactionFilter::kClassName());
+  auto ttl = filter->GetOptions<int32_t>("TTL");
+  ASSERT_NE(ttl, nullptr);
+  ASSERT_EQ(*ttl, 0);
+  ASSERT_OK(filter->ValidateOptions(DBOptions(), ColumnFamilyOptions()));
+  delete filter;
+  filter = nullptr;
+
+  ASSERT_OK(CompactionFilter::CreateFromString(
+      config_options_, "id=TtlCompactionFilter; ttl=123", &filter));
+  ASSERT_NE(filter, nullptr);
+  ttl = filter->GetOptions<int32_t>("TTL");
+  ASSERT_NE(ttl, nullptr);
+  ASSERT_EQ(*ttl, 123);
+  ASSERT_OK(filter->ValidateOptions(DBOptions(), ColumnFamilyOptions()));
+  delete filter;
+  filter = nullptr;
+
+  ASSERT_OK(CompactionFilter::CreateFromString(
+      config_options_,
+      "id=TtlCompactionFilter; ttl=456; user_filter=DummyFilter;", &filter));
+  ASSERT_NE(filter, nullptr);
+  auto inner = filter->CheckedCast<DummyFilter>();
+  ASSERT_NE(inner, nullptr);
+  ASSERT_OK(filter->ValidateOptions(DBOptions(), ColumnFamilyOptions()));
+  std::string mismatch;
+  std::string opts_str = filter->ToString(config_options_);
+  const CompactionFilter* copy = nullptr;
+  ASSERT_OK(
+      CompactionFilter::CreateFromString(config_options_, opts_str, &copy));
+  ASSERT_TRUE(filter->AreEquivalent(config_options_, copy, &mismatch));
+  delete filter;
+  delete copy;
+}
+
+TEST_F(TtlOptionsTest, LoadTtlCompactionFilterFactory) {
+  std::shared_ptr<CompactionFilterFactory> cff;
+
+  ASSERT_OK(CompactionFilterFactory::CreateFromString(
+      config_options_, TtlCompactionFilterFactory::kClassName(), &cff));
+  ASSERT_NE(cff.get(), nullptr);
+  ASSERT_STREQ(cff->Name(), TtlCompactionFilterFactory::kClassName());
+  auto ttl = cff->GetOptions<int32_t>("TTL");
+  ASSERT_NE(ttl, nullptr);
+  ASSERT_EQ(*ttl, 0);
+  ASSERT_OK(cff->ValidateOptions(DBOptions(), ColumnFamilyOptions()));
+
+  ASSERT_OK(CompactionFilterFactory::CreateFromString(
+      config_options_, "id=TtlCompactionFilterFactory; ttl=123", &cff));
+  ASSERT_NE(cff.get(), nullptr);
+  ASSERT_STREQ(cff->Name(), TtlCompactionFilterFactory::kClassName());
+  ttl = cff->GetOptions<int32_t>("TTL");
+  ASSERT_NE(ttl, nullptr);
+  ASSERT_EQ(*ttl, 123);
+  ASSERT_OK(cff->ValidateOptions(DBOptions(), ColumnFamilyOptions()));
+
+  ASSERT_OK(CompactionFilterFactory::CreateFromString(
+      config_options_,
+      "id=TtlCompactionFilterFactory; ttl=456; "
+      "user_filter_factory=DummyFilterFactory;",
+      &cff));
+  ASSERT_NE(cff.get(), nullptr);
+  auto filter = cff->CreateCompactionFilter(CompactionFilter::Context());
+  ASSERT_NE(filter.get(), nullptr);
+  auto ttlf = filter->CheckedCast<TtlCompactionFilter>();
+  ASSERT_EQ(filter.get(), ttlf);
+  auto user = filter->CheckedCast<DummyFilter>();
+  ASSERT_NE(user, nullptr);
+  ASSERT_OK(cff->ValidateOptions(DBOptions(), ColumnFamilyOptions()));
+
+  std::string opts_str = cff->ToString(config_options_);
+  std::string mismatch;
+  std::shared_ptr<CompactionFilterFactory> copy;
+  ASSERT_OK(CompactionFilterFactory::CreateFromString(config_options_, opts_str,
+                                                      &copy));
+  ASSERT_TRUE(cff->AreEquivalent(config_options_, copy.get(), &mismatch));
+}
+
+TEST_F(TtlOptionsTest, LoadTtlMergeOperator) {
+  std::shared_ptr<MergeOperator> mo;
+
+  config_options_.invoke_prepare_options = false;
+  ASSERT_OK(MergeOperator::CreateFromString(
+      config_options_, TtlMergeOperator::kClassName(), &mo));
+  ASSERT_NE(mo.get(), nullptr);
+  ASSERT_STREQ(mo->Name(), TtlMergeOperator::kClassName());
+  ASSERT_NOK(mo->ValidateOptions(DBOptions(), ColumnFamilyOptions()));
+
+  config_options_.invoke_prepare_options = true;
+  ASSERT_OK(MergeOperator::CreateFromString(
+      config_options_, "id=TtlMergeOperator; user_operator=bytesxor", &mo));
+  ASSERT_NE(mo.get(), nullptr);
+  ASSERT_STREQ(mo->Name(), TtlMergeOperator::kClassName());
+  ASSERT_OK(mo->ValidateOptions(DBOptions(), ColumnFamilyOptions()));
+  auto ttl_mo = mo->CheckedCast<TtlMergeOperator>();
+  ASSERT_EQ(mo.get(), ttl_mo);
+  auto user = ttl_mo->CheckedCast<BytesXOROperator>();
+  ASSERT_NE(user, nullptr);
+
+  std::string mismatch;
+  std::string opts_str = mo->ToString(config_options_);
+  std::shared_ptr<MergeOperator> copy;
+  ASSERT_OK(MergeOperator::CreateFromString(config_options_, opts_str, &copy));
+  ASSERT_TRUE(mo->AreEquivalent(config_options_, copy.get(), &mismatch));
+}
+}  // namespace ROCKSDB_NAMESPACE
+
+// A black-box test for the ttl wrapper around rocksdb
+int main(int argc, char** argv) {
+  ROCKSDB_NAMESPACE::port::InstallStackTraceHandler();
+  ::testing::InitGoogleTest(&argc, argv);
+  return RUN_ALL_TESTS();
+}
+
+#else
+#include <stdio.h>
+
+int main(int /*argc*/, char** /*argv*/) {
+  fprintf(stderr, "SKIPPED as DBWithTTL is not supported in ROCKSDB_LITE\n");
+  return 0;
+}
+
+#endif  // !ROCKSDB_LITE
diff --git a/src/rocksdb/utilities/util_merge_operators_test.cc b/src/rocksdb/utilities/util_merge_operators_test.cc
new file mode 100644
index 000000000..fed6f1a75
--- /dev/null
+++ b/src/rocksdb/utilities/util_merge_operators_test.cc
@@ -0,0 +1,100 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#include "test_util/testharness.h"
+#include "test_util/testutil.h"
+#include "utilities/merge_operators.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+class UtilMergeOperatorTest : public testing::Test {
+ public:
+  UtilMergeOperatorTest() {}
+
+  std::string FullMergeV2(std::string existing_value,
+                          std::vector<std::string> operands,
+                          std::string key = "") {
+    std::string result;
+    Slice result_operand(nullptr, 0);
+
+    Slice existing_value_slice(existing_value);
+    std::vector<Slice> operands_slice(operands.begin(), operands.end());
+
+    const MergeOperator::MergeOperationInput merge_in(
+        key, &existing_value_slice, operands_slice, nullptr);
+    MergeOperator::MergeOperationOutput merge_out(result, result_operand);
+    merge_operator_->FullMergeV2(merge_in, &merge_out);
+
+    if (result_operand.data()) {
+      result.assign(result_operand.data(), result_operand.size());
+    }
+    return result;
+  }
+
+  std::string FullMergeV2(std::vector<std::string> operands,
+                          std::string key = "") {
+    std::string result;
+    Slice result_operand(nullptr, 0);
+
+    std::vector<Slice> operands_slice(operands.begin(), operands.end());
+
+    const MergeOperator::MergeOperationInput merge_in(key, nullptr,
+                                                      operands_slice, nullptr);
+    MergeOperator::MergeOperationOutput merge_out(result, result_operand);
+    merge_operator_->FullMergeV2(merge_in, &merge_out);
+
+    if (result_operand.data()) {
+      result.assign(result_operand.data(), result_operand.size());
+    }
+    return result;
+  }
+
+  std::string PartialMerge(std::string left, std::string right,
+                           std::string key = "") {
+    std::string result;
+
+    merge_operator_->PartialMerge(key, left, right, &result, nullptr);
+    return result;
+  }
+
+  std::string PartialMergeMulti(std::deque<std::string> operands,
+                                std::string key = "") {
+    std::string result;
+    std::deque<Slice> operands_slice(operands.begin(), operands.end());
+
+    merge_operator_->PartialMergeMulti(key, operands_slice, &result, nullptr);
+    return result;
+  }
+
+ protected:
+  std::shared_ptr<MergeOperator> merge_operator_;
+};
+
+TEST_F(UtilMergeOperatorTest, MaxMergeOperator) {
+  merge_operator_ = MergeOperators::CreateMaxOperator();
+
+  EXPECT_EQ("B", FullMergeV2("B", {"A"}));
+  EXPECT_EQ("B", FullMergeV2("A", {"B"}));
+  EXPECT_EQ("", FullMergeV2({"", "", ""}));
+  EXPECT_EQ("A", FullMergeV2({"A"}));
+  EXPECT_EQ("ABC", FullMergeV2({"ABC"}));
+  EXPECT_EQ("Z", FullMergeV2({"ABC", "Z", "C", "AXX"}));
+  EXPECT_EQ("ZZZ", FullMergeV2({"ABC", "CC", "Z", "ZZZ"}));
+  EXPECT_EQ("a", FullMergeV2("a", {"ABC", "CC", "Z", "ZZZ"}));
+
+  EXPECT_EQ("z", PartialMergeMulti({"a", "z", "efqfqwgwew", "aaz", "hhhhh"}));
+
+  EXPECT_EQ("b", PartialMerge("a", "b"));
+  EXPECT_EQ("z", PartialMerge("z", "azzz"));
+  EXPECT_EQ("a", PartialMerge("a", ""));
+}
+
+}  // namespace ROCKSDB_NAMESPACE
+
+int main(int argc, char** argv) {
+  ROCKSDB_NAMESPACE::port::InstallStackTraceHandler();
+  ::testing::InitGoogleTest(&argc, argv);
+  return RUN_ALL_TESTS();
+}
diff --git a/src/rocksdb/utilities/wal_filter.cc b/src/rocksdb/utilities/wal_filter.cc
new file mode 100644
index 000000000..98bba3610
--- /dev/null
+++ b/src/rocksdb/utilities/wal_filter.cc
@@ -0,0 +1,23 @@
+// Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.
+// Copyright (c) 2011 The LevelDB Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file. See the AUTHORS file for names of contributors.
+
+#include "rocksdb/wal_filter.h"
+
+#include <memory>
+
+#include "rocksdb/convenience.h"
+#include "rocksdb/options.h"
+#include "rocksdb/utilities/customizable_util.h"
+
+namespace ROCKSDB_NAMESPACE {
+Status WalFilter::CreateFromString(const ConfigOptions& config_options,
+                                   const std::string& value,
+                                   WalFilter** filter) {
+  Status s =
+      LoadStaticObject<WalFilter>(config_options, value, nullptr, filter);
+  return s;
+}
+
+}  // namespace ROCKSDB_NAMESPACE
diff --git a/src/rocksdb/utilities/write_batch_with_index/write_batch_with_index.cc b/src/rocksdb/utilities/write_batch_with_index/write_batch_with_index.cc
new file mode 100644
index 000000000..408243b3f
--- /dev/null
+++ b/src/rocksdb/utilities/write_batch_with_index/write_batch_with_index.cc
@@ -0,0 +1,695 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#ifndef ROCKSDB_LITE
+
+#include "rocksdb/utilities/write_batch_with_index.h"
+
+#include <memory>
+
+#include "db/column_family.h"
+#include "db/db_impl/db_impl.h"
+#include "db/merge_context.h"
+#include "db/merge_helper.h"
+#include "memory/arena.h"
+#include "memtable/skiplist.h"
+#include "options/db_options.h"
+#include "rocksdb/comparator.h"
+#include "rocksdb/iterator.h"
+#include "util/cast_util.h"
+#include "util/string_util.h"
+#include "utilities/write_batch_with_index/write_batch_with_index_internal.h"
+
+namespace ROCKSDB_NAMESPACE {
+struct WriteBatchWithIndex::Rep {
+  explicit Rep(const Comparator* index_comparator, size_t reserved_bytes = 0,
+               size_t max_bytes = 0, bool _overwrite_key = false,
+               size_t protection_bytes_per_key = 0)
+      : write_batch(reserved_bytes, max_bytes, protection_bytes_per_key,
+                    index_comparator ? index_comparator->timestamp_size() : 0),
+        comparator(index_comparator, &write_batch),
+        skip_list(comparator, &arena),
+        overwrite_key(_overwrite_key),
+        last_entry_offset(0),
+        last_sub_batch_offset(0),
+        sub_batch_cnt(1) {}
+  ReadableWriteBatch write_batch;
+  WriteBatchEntryComparator comparator;
+  Arena arena;
+  WriteBatchEntrySkipList skip_list;
+  bool overwrite_key;
+  size_t last_entry_offset;
+  // The starting offset of the last sub-batch. A sub-batch starts right before
+  // inserting a key that is a duplicate of a key in the last sub-batch. Zero,
+  // the default, means that no duplicate key is detected so far.
+  size_t last_sub_batch_offset;
+  // Total number of sub-batches in the write batch. Default is 1.
+  size_t sub_batch_cnt;
+
+  // Remember current offset of internal write batch, which is used as
+  // the starting offset of the next record.
+  void SetLastEntryOffset() { last_entry_offset = write_batch.GetDataSize(); }
+
+  // In overwrite mode, find the existing entry for the same key and update it
+  // to point to the current entry.
+  // Return true if the key is found and updated.
+  bool UpdateExistingEntry(ColumnFamilyHandle* column_family, const Slice& key,
+                           WriteType type);
+  bool UpdateExistingEntryWithCfId(uint32_t column_family_id, const Slice& key,
+                                   WriteType type);
+
+  // Add the recent entry to the update.
+  // In overwrite mode, if key already exists in the index, update it.
+  void AddOrUpdateIndex(ColumnFamilyHandle* column_family, const Slice& key,
+                        WriteType type);
+  void AddOrUpdateIndex(const Slice& key, WriteType type);
+
+  // Allocate an index entry pointing to the last entry in the write batch and
+  // put it to skip list.
+  void AddNewEntry(uint32_t column_family_id);
+
+  // Clear all updates buffered in this batch.
+  void Clear();
+  void ClearIndex();
+
+  // Rebuild index by reading all records from the batch.
+  // Returns non-ok status on corruption.
+  Status ReBuildIndex();
+};
+
+bool WriteBatchWithIndex::Rep::UpdateExistingEntry(
+    ColumnFamilyHandle* column_family, const Slice& key, WriteType type) {
+  uint32_t cf_id = GetColumnFamilyID(column_family);
+  return UpdateExistingEntryWithCfId(cf_id, key, type);
+}
+
+bool WriteBatchWithIndex::Rep::UpdateExistingEntryWithCfId(
+    uint32_t column_family_id, const Slice& key, WriteType type) {
+  if (!overwrite_key) {
+    return false;
+  }
+
+  WBWIIteratorImpl iter(column_family_id, &skip_list, &write_batch,
+                        &comparator);
+  iter.Seek(key);
+  if (!iter.Valid()) {
+    return false;
+  } else if (!iter.MatchesKey(column_family_id, key)) {
+    return false;
+  } else {
+    // Move to the end of this key (NextKey-Prev)
+    iter.NextKey();  // Move to the next key
+    if (iter.Valid()) {
+      iter.Prev();  // Move back one entry
+    } else {
+      iter.SeekToLast();
+    }
+  }
+  WriteBatchIndexEntry* non_const_entry =
+      const_cast<WriteBatchIndexEntry*>(iter.GetRawEntry());
+  if (LIKELY(last_sub_batch_offset <= non_const_entry->offset)) {
+    last_sub_batch_offset = last_entry_offset;
+    sub_batch_cnt++;
+  }
+  if (type == kMergeRecord) {
+    return false;
+  } else {
+    non_const_entry->offset = last_entry_offset;
+    return true;
+  }
+}
+
+void WriteBatchWithIndex::Rep::AddOrUpdateIndex(
+    ColumnFamilyHandle* column_family, const Slice& key, WriteType type) {
+  if (!UpdateExistingEntry(column_family, key, type)) {
+    uint32_t cf_id = GetColumnFamilyID(column_family);
+    const auto* cf_cmp = GetColumnFamilyUserComparator(column_family);
+    if (cf_cmp != nullptr) {
+      comparator.SetComparatorForCF(cf_id, cf_cmp);
+    }
+    AddNewEntry(cf_id);
+  }
+}
+
+void WriteBatchWithIndex::Rep::AddOrUpdateIndex(const Slice& key,
+                                                WriteType type) {
+  if (!UpdateExistingEntryWithCfId(0, key, type)) {
+    AddNewEntry(0);
+  }
+}
+
+void WriteBatchWithIndex::Rep::AddNewEntry(uint32_t column_family_id) {
+  const std::string& wb_data = write_batch.Data();
+  Slice entry_ptr = Slice(wb_data.data() + last_entry_offset,
+                          wb_data.size() - last_entry_offset);
+  // Extract key
+  Slice key;
+  bool success =
+      ReadKeyFromWriteBatchEntry(&entry_ptr, &key, column_family_id != 0);
+#ifdef NDEBUG
+  (void)success;
+#endif
+  assert(success);
+
+  const Comparator* const ucmp = comparator.GetComparator(column_family_id);
+  size_t ts_sz = ucmp ? ucmp->timestamp_size() : 0;
+
+  if (ts_sz > 0) {
+    key.remove_suffix(ts_sz);
+  }
+
+  auto* mem = arena.Allocate(sizeof(WriteBatchIndexEntry));
+  auto* index_entry =
+      new (mem) WriteBatchIndexEntry(last_entry_offset, column_family_id,
+                                     key.data() - wb_data.data(), key.size());
+  skip_list.Insert(index_entry);
+}
+
+void WriteBatchWithIndex::Rep::Clear() {
+  write_batch.Clear();
+  ClearIndex();
+}
+
+void WriteBatchWithIndex::Rep::ClearIndex() {
+  skip_list.~WriteBatchEntrySkipList();
+  arena.~Arena();
+  new (&arena) Arena();
+  new (&skip_list) WriteBatchEntrySkipList(comparator, &arena);
+  last_entry_offset = 0;
+  last_sub_batch_offset = 0;
+  sub_batch_cnt = 1;
+}
+
+Status WriteBatchWithIndex::Rep::ReBuildIndex() {
+  Status s;
+
+  ClearIndex();
+
+  if (write_batch.Count() == 0) {
+    // Nothing to re-index
+    return s;
+  }
+
+  size_t offset = WriteBatchInternal::GetFirstOffset(&write_batch);
+
+  Slice input(write_batch.Data());
+  input.remove_prefix(offset);
+
+  // Loop through all entries in Rep and add each one to the index
+  uint32_t found = 0;
+  while (s.ok() && !input.empty()) {
+    Slice key, value, blob, xid;
+    uint32_t column_family_id = 0;  // default
+    char tag = 0;
+
+    // set offset of current entry for call to AddNewEntry()
+    last_entry_offset = input.data() - write_batch.Data().data();
+
+    s = ReadRecordFromWriteBatch(&input, &tag, &column_family_id, &key, &value,
+                                 &blob, &xid);
+    if (!s.ok()) {
+      break;
+    }
+
+    switch (tag) {
+      case kTypeColumnFamilyValue:
+      case kTypeValue:
+        found++;
+        if (!UpdateExistingEntryWithCfId(column_family_id, key, kPutRecord)) {
+          AddNewEntry(column_family_id);
+        }
+        break;
+      case kTypeColumnFamilyDeletion:
+      case kTypeDeletion:
+        found++;
+        if (!UpdateExistingEntryWithCfId(column_family_id, key,
+                                         kDeleteRecord)) {
+          AddNewEntry(column_family_id);
+        }
+        break;
+      case kTypeColumnFamilySingleDeletion:
+      case kTypeSingleDeletion:
+        found++;
+        if (!UpdateExistingEntryWithCfId(column_family_id, key,
+                                         kSingleDeleteRecord)) {
+          AddNewEntry(column_family_id);
+        }
+        break;
+      case kTypeColumnFamilyMerge:
+      case kTypeMerge:
+        found++;
+        if (!UpdateExistingEntryWithCfId(column_family_id, key, kMergeRecord)) {
+          AddNewEntry(column_family_id);
+        }
+        break;
+      case kTypeLogData:
+      case kTypeBeginPrepareXID:
+      case kTypeBeginPersistedPrepareXID:
+      case kTypeBeginUnprepareXID:
+      case kTypeEndPrepareXID:
+      case kTypeCommitXID:
+      case kTypeCommitXIDAndTimestamp:
+      case kTypeRollbackXID:
+      case kTypeNoop:
+        break;
+      default:
+        return Status::Corruption(
+            "unknown WriteBatch tag in ReBuildIndex",
+            std::to_string(static_cast<unsigned int>(tag)));
+    }
+  }
+
+  if (s.ok() && found != write_batch.Count()) {
+    s = Status::Corruption("WriteBatch has wrong count");
+  }
+
+  return s;
+}
+
+WriteBatchWithIndex::WriteBatchWithIndex(
+    const Comparator* default_index_comparator, size_t reserved_bytes,
+    bool overwrite_key, size_t max_bytes, size_t protection_bytes_per_key)
+    : rep(new Rep(default_index_comparator, reserved_bytes, max_bytes,
+                  overwrite_key, protection_bytes_per_key)) {}
+
+WriteBatchWithIndex::~WriteBatchWithIndex() {}
+
+WriteBatchWithIndex::WriteBatchWithIndex(WriteBatchWithIndex&&) = default;
+
+WriteBatchWithIndex& WriteBatchWithIndex::operator=(WriteBatchWithIndex&&) =
+    default;
+
+WriteBatch* WriteBatchWithIndex::GetWriteBatch() { return &rep->write_batch; }
+
+size_t WriteBatchWithIndex::SubBatchCnt() { return rep->sub_batch_cnt; }
+
+WBWIIterator* WriteBatchWithIndex::NewIterator() {
+  return new WBWIIteratorImpl(0, &(rep->skip_list), &rep->write_batch,
+                              &(rep->comparator));
+}
+
+WBWIIterator* WriteBatchWithIndex::NewIterator(
+    ColumnFamilyHandle* column_family) {
+  return new WBWIIteratorImpl(GetColumnFamilyID(column_family),
+                              &(rep->skip_list), &rep->write_batch,
+                              &(rep->comparator));
+}
+
+Iterator* WriteBatchWithIndex::NewIteratorWithBase(
+    ColumnFamilyHandle* column_family, Iterator* base_iterator,
+    const ReadOptions* read_options) {
+  auto wbwiii =
+      new WBWIIteratorImpl(GetColumnFamilyID(column_family), &(rep->skip_list),
+                           &rep->write_batch, &rep->comparator);
+  return new BaseDeltaIterator(column_family, base_iterator, wbwiii,
+                               GetColumnFamilyUserComparator(column_family),
+                               read_options);
+}
+
+Iterator* WriteBatchWithIndex::NewIteratorWithBase(Iterator* base_iterator) {
+  // default column family's comparator
+  auto wbwiii = new WBWIIteratorImpl(0, &(rep->skip_list), &rep->write_batch,
+                                     &rep->comparator);
+  return new BaseDeltaIterator(nullptr, base_iterator, wbwiii,
+                               rep->comparator.default_comparator());
+}
+
+Status WriteBatchWithIndex::Put(ColumnFamilyHandle* column_family,
+                                const Slice& key, const Slice& value) {
+  rep->SetLastEntryOffset();
+  auto s = rep->write_batch.Put(column_family, key, value);
+  if (s.ok()) {
+    rep->AddOrUpdateIndex(column_family, key, kPutRecord);
+  }
+  return s;
+}
+
+Status WriteBatchWithIndex::Put(const Slice& key, const Slice& value) {
+  rep->SetLastEntryOffset();
+  auto s = rep->write_batch.Put(key, value);
+  if (s.ok()) {
+    rep->AddOrUpdateIndex(key, kPutRecord);
+  }
+  return s;
+}
+
+Status WriteBatchWithIndex::Put(ColumnFamilyHandle* column_family,
+                                const Slice& /*key*/, const Slice& /*ts*/,
+                                const Slice& /*value*/) {
+  if (!column_family) {
+    return Status::InvalidArgument("column family handle cannot be nullptr");
+  }
+  // TODO: support WBWI::Put() with timestamp.
+  return Status::NotSupported();
+}
+
+Status WriteBatchWithIndex::Delete(ColumnFamilyHandle* column_family,
+                                   const Slice& key) {
+  rep->SetLastEntryOffset();
+  auto s = rep->write_batch.Delete(column_family, key);
+  if (s.ok()) {
+    rep->AddOrUpdateIndex(column_family, key, kDeleteRecord);
+  }
+  return s;
+}
+
+Status WriteBatchWithIndex::Delete(const Slice& key) {
+  rep->SetLastEntryOffset();
+  auto s = rep->write_batch.Delete(key);
+  if (s.ok()) {
+    rep->AddOrUpdateIndex(key, kDeleteRecord);
+  }
+  return s;
+}
+
+Status WriteBatchWithIndex::Delete(ColumnFamilyHandle* column_family,
+                                   const Slice& /*key*/, const Slice& /*ts*/) {
+  if (!column_family) {
+    return Status::InvalidArgument("column family handle cannot be nullptr");
+  }
+  // TODO: support WBWI::Delete() with timestamp.
+  return Status::NotSupported();
+}
+
+Status WriteBatchWithIndex::SingleDelete(ColumnFamilyHandle* column_family,
+                                         const Slice& key) {
+  rep->SetLastEntryOffset();
+  auto s = rep->write_batch.SingleDelete(column_family, key);
+  if (s.ok()) {
+    rep->AddOrUpdateIndex(column_family, key, kSingleDeleteRecord);
+  }
+  return s;
+}
+
+Status WriteBatchWithIndex::SingleDelete(const Slice& key) {
+  rep->SetLastEntryOffset();
+  auto s = rep->write_batch.SingleDelete(key);
+  if (s.ok()) {
+    rep->AddOrUpdateIndex(key, kSingleDeleteRecord);
+  }
+  return s;
+}
+
+Status WriteBatchWithIndex::SingleDelete(ColumnFamilyHandle* column_family,
+                                         const Slice& /*key*/,
+                                         const Slice& /*ts*/) {
+  if (!column_family) {
+    return Status::InvalidArgument("column family handle cannot be nullptr");
+  }
+  // TODO: support WBWI::SingleDelete() with timestamp.
+  return Status::NotSupported();
+}
+
+Status WriteBatchWithIndex::Merge(ColumnFamilyHandle* column_family,
+                                  const Slice& key, const Slice& value) {
+  rep->SetLastEntryOffset();
+  auto s = rep->write_batch.Merge(column_family, key, value);
+  if (s.ok()) {
+    rep->AddOrUpdateIndex(column_family, key, kMergeRecord);
+  }
+  return s;
+}
+
+Status WriteBatchWithIndex::Merge(const Slice& key, const Slice& value) {
+  rep->SetLastEntryOffset();
+  auto s = rep->write_batch.Merge(key, value);
+  if (s.ok()) {
+    rep->AddOrUpdateIndex(key, kMergeRecord);
+  }
+  return s;
+}
+
+Status WriteBatchWithIndex::PutLogData(const Slice& blob) {
+  return rep->write_batch.PutLogData(blob);
+}
+
+void WriteBatchWithIndex::Clear() { rep->Clear(); }
+
+Status WriteBatchWithIndex::GetFromBatch(ColumnFamilyHandle* column_family,
+                                         const DBOptions& options,
+                                         const Slice& key, std::string* value) {
+  Status s;
+  WriteBatchWithIndexInternal wbwii(&options, column_family);
+  auto result = wbwii.GetFromBatch(this, key, value, &s);
+
+  switch (result) {
+    case WBWIIteratorImpl::kFound:
+    case WBWIIteratorImpl::kError:
+      // use returned status
+      break;
+    case WBWIIteratorImpl::kDeleted:
+    case WBWIIteratorImpl::kNotFound:
+      s = Status::NotFound();
+      break;
+    case WBWIIteratorImpl::kMergeInProgress:
+      s = Status::MergeInProgress();
+      break;
+    default:
+      assert(false);
+  }
+
+  return s;
+}
+
+Status WriteBatchWithIndex::GetFromBatchAndDB(DB* db,
+                                              const ReadOptions& read_options,
+                                              const Slice& key,
+                                              std::string* value) {
+  assert(value != nullptr);
+  PinnableSlice pinnable_val(value);
+  assert(!pinnable_val.IsPinned());
+  auto s = GetFromBatchAndDB(db, read_options, db->DefaultColumnFamily(), key,
+                             &pinnable_val);
+  if (s.ok() && pinnable_val.IsPinned()) {
+    value->assign(pinnable_val.data(), pinnable_val.size());
+  }  // else value is already assigned
+  return s;
+}
+
+Status WriteBatchWithIndex::GetFromBatchAndDB(DB* db,
+                                              const ReadOptions& read_options,
+                                              const Slice& key,
+                                              PinnableSlice* pinnable_val) {
+  return GetFromBatchAndDB(db, read_options, db->DefaultColumnFamily(), key,
+                           pinnable_val);
+}
+
+Status WriteBatchWithIndex::GetFromBatchAndDB(DB* db,
+                                              const ReadOptions& read_options,
+                                              ColumnFamilyHandle* column_family,
+                                              const Slice& key,
+                                              std::string* value) {
+  assert(value != nullptr);
+  PinnableSlice pinnable_val(value);
+  assert(!pinnable_val.IsPinned());
+  auto s =
+      GetFromBatchAndDB(db, read_options, column_family, key, &pinnable_val);
+  if (s.ok() && pinnable_val.IsPinned()) {
+    value->assign(pinnable_val.data(), pinnable_val.size());
+  }  // else value is already assigned
+  return s;
+}
+
+Status WriteBatchWithIndex::GetFromBatchAndDB(DB* db,
+                                              const ReadOptions& read_options,
+                                              ColumnFamilyHandle* column_family,
+                                              const Slice& key,
+                                              PinnableSlice* pinnable_val) {
+  return GetFromBatchAndDB(db, read_options, column_family, key, pinnable_val,
+                           nullptr);
+}
+
+Status WriteBatchWithIndex::GetFromBatchAndDB(
+    DB* db, const ReadOptions& read_options, ColumnFamilyHandle* column_family,
+    const Slice& key, PinnableSlice* pinnable_val, ReadCallback* callback) {
+  const Comparator* const ucmp = rep->comparator.GetComparator(column_family);
+  size_t ts_sz = ucmp ? ucmp->timestamp_size() : 0;
+  if (ts_sz > 0 && !read_options.timestamp) {
+    return Status::InvalidArgument("Must specify timestamp");
+  }
+
+  Status s;
+  WriteBatchWithIndexInternal wbwii(db, column_family);
+
+  // Since the lifetime of the WriteBatch is the same as that of the transaction
+  // we cannot pin it as otherwise the returned value will not be available
+  // after the transaction finishes.
+  std::string& batch_value = *pinnable_val->GetSelf();
+  auto result = wbwii.GetFromBatch(this, key, &batch_value, &s);
+
+  if (result == WBWIIteratorImpl::kFound) {
+    pinnable_val->PinSelf();
+    return s;
+  } else if (!s.ok() || result == WBWIIteratorImpl::kError) {
+    return s;
+  } else if (result == WBWIIteratorImpl::kDeleted) {
+    return Status::NotFound();
+  }
+  assert(result == WBWIIteratorImpl::kMergeInProgress ||
+         result == WBWIIteratorImpl::kNotFound);
+
+  // Did not find key in batch OR could not resolve Merges.  Try DB.
+  if (!callback) {
+    s = db->Get(read_options, column_family, key, pinnable_val);
+  } else {
+    DBImpl::GetImplOptions get_impl_options;
+    get_impl_options.column_family = column_family;
+    get_impl_options.value = pinnable_val;
+    get_impl_options.callback = callback;
+    s = static_cast_with_check<DBImpl>(db->GetRootDB())
+            ->GetImpl(read_options, key, get_impl_options);
+  }
+
+  if (s.ok() || s.IsNotFound()) {  // DB Get Succeeded
+    if (result == WBWIIteratorImpl::kMergeInProgress) {
+      // Merge result from DB with merges in Batch
+      std::string merge_result;
+      if (s.ok()) {
+        s = wbwii.MergeKey(key, pinnable_val, &merge_result);
+      } else {  // Key not present in db (s.IsNotFound())
+        s = wbwii.MergeKey(key, nullptr, &merge_result);
+      }
+      if (s.ok()) {
+        pinnable_val->Reset();
+        *pinnable_val->GetSelf() = std::move(merge_result);
+        pinnable_val->PinSelf();
+      }
+    }
+  }
+
+  return s;
+}
+
+void WriteBatchWithIndex::MultiGetFromBatchAndDB(
+    DB* db, const ReadOptions& read_options, ColumnFamilyHandle* column_family,
+    const size_t num_keys, const Slice* keys, PinnableSlice* values,
+    Status* statuses, bool sorted_input) {
+  MultiGetFromBatchAndDB(db, read_options, column_family, num_keys, keys,
+                         values, statuses, sorted_input, nullptr);
+}
+
+void WriteBatchWithIndex::MultiGetFromBatchAndDB(
+    DB* db, const ReadOptions& read_options, ColumnFamilyHandle* column_family,
+    const size_t num_keys, const Slice* keys, PinnableSlice* values,
+    Status* statuses, bool sorted_input, ReadCallback* callback) {
+  const Comparator* const ucmp = rep->comparator.GetComparator(column_family);
+  size_t ts_sz = ucmp ? ucmp->timestamp_size() : 0;
+  if (ts_sz > 0 && !read_options.timestamp) {
+    for (size_t i = 0; i < num_keys; ++i) {
+      statuses[i] = Status::InvalidArgument("Must specify timestamp");
+    }
+    return;
+  }
+
+  WriteBatchWithIndexInternal wbwii(db, column_family);
+
+  autovector<KeyContext, MultiGetContext::MAX_BATCH_SIZE> key_context;
+  autovector<KeyContext*, MultiGetContext::MAX_BATCH_SIZE> sorted_keys;
+  // To hold merges from the write batch
+  autovector<std::pair<WBWIIteratorImpl::Result, MergeContext>,
+             MultiGetContext::MAX_BATCH_SIZE>
+      merges;
+  // Since the lifetime of the WriteBatch is the same as that of the transaction
+  // we cannot pin it as otherwise the returned value will not be available
+  // after the transaction finishes.
+  for (size_t i = 0; i < num_keys; ++i) {
+    MergeContext merge_context;
+    std::string batch_value;
+    Status* s = &statuses[i];
+    PinnableSlice* pinnable_val = &values[i];
+    pinnable_val->Reset();
+    auto result =
+        wbwii.GetFromBatch(this, keys[i], &merge_context, &batch_value, s);
+
+    if (result == WBWIIteratorImpl::kFound) {
+      *pinnable_val->GetSelf() = std::move(batch_value);
+      pinnable_val->PinSelf();
+      continue;
+    }
+    if (result == WBWIIteratorImpl::kDeleted) {
+      *s = Status::NotFound();
+      continue;
+    }
+    if (result == WBWIIteratorImpl::kError) {
+      continue;
+    }
+    assert(result == WBWIIteratorImpl::kMergeInProgress ||
+           result == WBWIIteratorImpl::kNotFound);
+    key_context.emplace_back(column_family, keys[i], &values[i],
+                             /*timestamp*/ nullptr, &statuses[i]);
+    merges.emplace_back(result, std::move(merge_context));
+  }
+
+  for (KeyContext& key : key_context) {
+    sorted_keys.emplace_back(&key);
+  }
+
+  // Did not find key in batch OR could not resolve Merges.  Try DB.
+  static_cast_with_check<DBImpl>(db->GetRootDB())
+      ->PrepareMultiGetKeys(key_context.size(), sorted_input, &sorted_keys);
+  static_cast_with_check<DBImpl>(db->GetRootDB())
+      ->MultiGetWithCallback(read_options, column_family, callback,
+                             &sorted_keys);
+
+  for (auto iter = key_context.begin(); iter != key_context.end(); ++iter) {
+    KeyContext& key = *iter;
+    if (key.s->ok() || key.s->IsNotFound()) {  // DB Get Succeeded
+      size_t index = iter - key_context.begin();
+      std::pair<WBWIIteratorImpl::Result, MergeContext>& merge_result =
+          merges[index];
+      if (merge_result.first == WBWIIteratorImpl::kMergeInProgress) {
+        std::string merged_value;
+        // Merge result from DB with merges in Batch
+        if (key.s->ok()) {
+          *key.s = wbwii.MergeKey(*key.key, iter->value, merge_result.second,
+                                  &merged_value);
+        } else {  // Key not present in db (s.IsNotFound())
+          *key.s = wbwii.MergeKey(*key.key, nullptr, merge_result.second,
+                                  &merged_value);
+        }
+        if (key.s->ok()) {
+          key.value->Reset();
+          *key.value->GetSelf() = std::move(merged_value);
+          key.value->PinSelf();
+        }
+      }
+    }
+  }
+}
+
+void WriteBatchWithIndex::SetSavePoint() { rep->write_batch.SetSavePoint(); }
+
+Status WriteBatchWithIndex::RollbackToSavePoint() {
+  Status s = rep->write_batch.RollbackToSavePoint();
+
+  if (s.ok()) {
+    rep->sub_batch_cnt = 1;
+    rep->last_sub_batch_offset = 0;
+    s = rep->ReBuildIndex();
+  }
+
+  return s;
+}
+
+Status WriteBatchWithIndex::PopSavePoint() {
+  return rep->write_batch.PopSavePoint();
+}
+
+void WriteBatchWithIndex::SetMaxBytes(size_t max_bytes) {
+  rep->write_batch.SetMaxBytes(max_bytes);
+}
+
+size_t WriteBatchWithIndex::GetDataSize() const {
+  return rep->write_batch.GetDataSize();
+}
+
+const Comparator* WriteBatchWithIndexInternal::GetUserComparator(
+    const WriteBatchWithIndex& wbwi, uint32_t cf_id) {
+  const WriteBatchEntryComparator& ucmps = wbwi.rep->comparator;
+  return ucmps.GetComparator(cf_id);
+}
+
+}  // namespace ROCKSDB_NAMESPACE
+#endif  // !ROCKSDB_LITE
diff --git a/src/rocksdb/utilities/write_batch_with_index/write_batch_with_index_internal.cc b/src/rocksdb/utilities/write_batch_with_index/write_batch_with_index_internal.cc
new file mode 100644
index 000000000..3c9205bf7
--- /dev/null
+++ b/src/rocksdb/utilities/write_batch_with_index/write_batch_with_index_internal.cc
@@ -0,0 +1,735 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+
+#ifndef ROCKSDB_LITE
+
+#include "utilities/write_batch_with_index/write_batch_with_index_internal.h"
+
+#include "db/column_family.h"
+#include "db/db_impl/db_impl.h"
+#include "db/merge_context.h"
+#include "db/merge_helper.h"
+#include "rocksdb/comparator.h"
+#include "rocksdb/db.h"
+#include "rocksdb/utilities/write_batch_with_index.h"
+#include "util/cast_util.h"
+#include "util/coding.h"
+#include "util/string_util.h"
+
+namespace ROCKSDB_NAMESPACE {
+BaseDeltaIterator::BaseDeltaIterator(ColumnFamilyHandle* column_family,
+                                     Iterator* base_iterator,
+                                     WBWIIteratorImpl* delta_iterator,
+                                     const Comparator* comparator,
+                                     const ReadOptions* read_options)
+    : forward_(true),
+      current_at_base_(true),
+      equal_keys_(false),
+      status_(Status::OK()),
+      base_iterator_(base_iterator),
+      delta_iterator_(delta_iterator),
+      comparator_(comparator),
+      iterate_upper_bound_(read_options ? read_options->iterate_upper_bound
+                                        : nullptr) {
+  assert(comparator_);
+  wbwii_.reset(new WriteBatchWithIndexInternal(column_family));
+}
+
+bool BaseDeltaIterator::Valid() const {
+  return status_.ok() ? (current_at_base_ ? BaseValid() : DeltaValid()) : false;
+}
+
+void BaseDeltaIterator::SeekToFirst() {
+  forward_ = true;
+  base_iterator_->SeekToFirst();
+  delta_iterator_->SeekToFirst();
+  UpdateCurrent();
+}
+
+void BaseDeltaIterator::SeekToLast() {
+  forward_ = false;
+  base_iterator_->SeekToLast();
+  delta_iterator_->SeekToLast();
+  UpdateCurrent();
+}
+
+void BaseDeltaIterator::Seek(const Slice& k) {
+  forward_ = true;
+  base_iterator_->Seek(k);
+  delta_iterator_->Seek(k);
+  UpdateCurrent();
+}
+
+void BaseDeltaIterator::SeekForPrev(const Slice& k) {
+  forward_ = false;
+  base_iterator_->SeekForPrev(k);
+  delta_iterator_->SeekForPrev(k);
+  UpdateCurrent();
+}
+
+void BaseDeltaIterator::Next() {
+  if (!Valid()) {
+    status_ = Status::NotSupported("Next() on invalid iterator");
+    return;
+  }
+
+  if (!forward_) {
+    // Need to change direction
+    // if our direction was backward and we're not equal, we have two states:
+    // * both iterators are valid: we're already in a good state (current
+    // shows to smaller)
+    // * only one iterator is valid: we need to advance that iterator
+    forward_ = true;
+    equal_keys_ = false;
+    if (!BaseValid()) {
+      assert(DeltaValid());
+      base_iterator_->SeekToFirst();
+    } else if (!DeltaValid()) {
+      delta_iterator_->SeekToFirst();
+    } else if (current_at_base_) {
+      // Change delta from larger than base to smaller
+      AdvanceDelta();
+    } else {
+      // Change base from larger than delta to smaller
+      AdvanceBase();
+    }
+    if (DeltaValid() && BaseValid()) {
+      if (0 == comparator_->CompareWithoutTimestamp(
+                   delta_iterator_->Entry().key, /*a_has_ts=*/false,
+                   base_iterator_->key(), /*b_has_ts=*/false)) {
+        equal_keys_ = true;
+      }
+    }
+  }
+  Advance();
+}
+
+void BaseDeltaIterator::Prev() {
+  if (!Valid()) {
+    status_ = Status::NotSupported("Prev() on invalid iterator");
+    return;
+  }
+
+  if (forward_) {
+    // Need to change direction
+    // if our direction was backward and we're not equal, we have two states:
+    // * both iterators are valid: we're already in a good state (current
+    // shows to smaller)
+    // * only one iterator is valid: we need to advance that iterator
+    forward_ = false;
+    equal_keys_ = false;
+    if (!BaseValid()) {
+      assert(DeltaValid());
+      base_iterator_->SeekToLast();
+    } else if (!DeltaValid()) {
+      delta_iterator_->SeekToLast();
+    } else if (current_at_base_) {
+      // Change delta from less advanced than base to more advanced
+      AdvanceDelta();
+    } else {
+      // Change base from less advanced than delta to more advanced
+      AdvanceBase();
+    }
+    if (DeltaValid() && BaseValid()) {
+      if (0 == comparator_->CompareWithoutTimestamp(
+                   delta_iterator_->Entry().key, /*a_has_ts=*/false,
+                   base_iterator_->key(), /*b_has_ts=*/false)) {
+        equal_keys_ = true;
+      }
+    }
+  }
+
+  Advance();
+}
+
+Slice BaseDeltaIterator::key() const {
+  return current_at_base_ ? base_iterator_->key()
+                          : delta_iterator_->Entry().key;
+}
+
+Slice BaseDeltaIterator::value() const {
+  if (current_at_base_) {
+    return base_iterator_->value();
+  } else {
+    WriteEntry delta_entry = delta_iterator_->Entry();
+    if (wbwii_->GetNumOperands() == 0) {
+      return delta_entry.value;
+    } else if (delta_entry.type == kDeleteRecord ||
+               delta_entry.type == kSingleDeleteRecord) {
+      status_ =
+          wbwii_->MergeKey(delta_entry.key, nullptr, merge_result_.GetSelf());
+    } else if (delta_entry.type == kPutRecord) {
+      status_ = wbwii_->MergeKey(delta_entry.key, &delta_entry.value,
+                                 merge_result_.GetSelf());
+    } else if (delta_entry.type == kMergeRecord) {
+      if (equal_keys_) {
+        Slice base_value = base_iterator_->value();
+        status_ = wbwii_->MergeKey(delta_entry.key, &base_value,
+                                   merge_result_.GetSelf());
+      } else {
+        status_ =
+            wbwii_->MergeKey(delta_entry.key, nullptr, merge_result_.GetSelf());
+      }
+    }
+    merge_result_.PinSelf();
+    return merge_result_;
+  }
+}
+
+Status BaseDeltaIterator::status() const {
+  if (!status_.ok()) {
+    return status_;
+  }
+  if (!base_iterator_->status().ok()) {
+    return base_iterator_->status();
+  }
+  return delta_iterator_->status();
+}
+
+void BaseDeltaIterator::Invalidate(Status s) { status_ = s; }
+
+void BaseDeltaIterator::AssertInvariants() {
+#ifndef NDEBUG
+  bool not_ok = false;
+  if (!base_iterator_->status().ok()) {
+    assert(!base_iterator_->Valid());
+    not_ok = true;
+  }
+  if (!delta_iterator_->status().ok()) {
+    assert(!delta_iterator_->Valid());
+    not_ok = true;
+  }
+  if (not_ok) {
+    assert(!Valid());
+    assert(!status().ok());
+    return;
+  }
+
+  if (!Valid()) {
+    return;
+  }
+  if (!BaseValid()) {
+    assert(!current_at_base_ && delta_iterator_->Valid());
+    return;
+  }
+  if (!DeltaValid()) {
+    assert(current_at_base_ && base_iterator_->Valid());
+    return;
+  }
+  // we don't support those yet
+  assert(delta_iterator_->Entry().type != kMergeRecord &&
+         delta_iterator_->Entry().type != kLogDataRecord);
+  int compare = comparator_->CompareWithoutTimestamp(
+      delta_iterator_->Entry().key, /*a_has_ts=*/false, base_iterator_->key(),
+      /*b_has_ts=*/false);
+  if (forward_) {
+    // current_at_base -> compare < 0
+    assert(!current_at_base_ || compare < 0);
+    // !current_at_base -> compare <= 0
+    assert(current_at_base_ && compare >= 0);
+  } else {
+    // current_at_base -> compare > 0
+    assert(!current_at_base_ || compare > 0);
+    // !current_at_base -> compare <= 0
+    assert(current_at_base_ && compare <= 0);
+  }
+  // equal_keys_ <=> compare == 0
+  assert((equal_keys_ || compare != 0) && (!equal_keys_ || compare == 0));
+#endif
+}
+
+void BaseDeltaIterator::Advance() {
+  if (equal_keys_) {
+    assert(BaseValid() && DeltaValid());
+    AdvanceBase();
+    AdvanceDelta();
+  } else {
+    if (current_at_base_) {
+      assert(BaseValid());
+      AdvanceBase();
+    } else {
+      assert(DeltaValid());
+      AdvanceDelta();
+    }
+  }
+  UpdateCurrent();
+}
+
+void BaseDeltaIterator::AdvanceDelta() {
+  if (forward_) {
+    delta_iterator_->NextKey();
+  } else {
+    delta_iterator_->PrevKey();
+  }
+}
+void BaseDeltaIterator::AdvanceBase() {
+  if (forward_) {
+    base_iterator_->Next();
+  } else {
+    base_iterator_->Prev();
+  }
+}
+
+bool BaseDeltaIterator::BaseValid() const { return base_iterator_->Valid(); }
+bool BaseDeltaIterator::DeltaValid() const { return delta_iterator_->Valid(); }
+void BaseDeltaIterator::UpdateCurrent() {
+// Suppress false positive clang analyzer warnings.
+#ifndef __clang_analyzer__
+  status_ = Status::OK();
+  while (true) {
+    auto delta_result = WBWIIteratorImpl::kNotFound;
+    WriteEntry delta_entry;
+    if (DeltaValid()) {
+      assert(delta_iterator_->status().ok());
+      delta_result =
+          delta_iterator_->FindLatestUpdate(wbwii_->GetMergeContext());
+      delta_entry = delta_iterator_->Entry();
+    } else if (!delta_iterator_->status().ok()) {
+      // Expose the error status and stop.
+      current_at_base_ = false;
+      return;
+    }
+    equal_keys_ = false;
+    if (!BaseValid()) {
+      if (!base_iterator_->status().ok()) {
+        // Expose the error status and stop.
+        current_at_base_ = true;
+        return;
+      }
+
+      // Base has finished.
+      if (!DeltaValid()) {
+        // Finished
+        return;
+      }
+      if (iterate_upper_bound_) {
+        if (comparator_->CompareWithoutTimestamp(
+                delta_entry.key, /*a_has_ts=*/false, *iterate_upper_bound_,
+                /*b_has_ts=*/false) >= 0) {
+          // out of upper bound -> finished.
+          return;
+        }
+      }
+      if (delta_result == WBWIIteratorImpl::kDeleted &&
+          wbwii_->GetNumOperands() == 0) {
+        AdvanceDelta();
+      } else {
+        current_at_base_ = false;
+        return;
+      }
+    } else if (!DeltaValid()) {
+      // Delta has finished.
+      current_at_base_ = true;
+      return;
+    } else {
+      int compare =
+          (forward_ ? 1 : -1) * comparator_->CompareWithoutTimestamp(
+                                    delta_entry.key, /*a_has_ts=*/false,
+                                    base_iterator_->key(), /*b_has_ts=*/false);
+      if (compare <= 0) {  // delta bigger or equal
+        if (compare == 0) {
+          equal_keys_ = true;
+        }
+        if (delta_result != WBWIIteratorImpl::kDeleted ||
+            wbwii_->GetNumOperands() > 0) {
+          current_at_base_ = false;
+          return;
+        }
+        // Delta is less advanced and is delete.
+        AdvanceDelta();
+        if (equal_keys_) {
+          AdvanceBase();
+        }
+      } else {
+        current_at_base_ = true;
+        return;
+      }
+    }
+  }
+
+  AssertInvariants();
+#endif  // __clang_analyzer__
+}
+
+void WBWIIteratorImpl::AdvanceKey(bool forward) {
+  if (Valid()) {
+    Slice key = Entry().key;
+    do {
+      if (forward) {
+        Next();
+      } else {
+        Prev();
+      }
+    } while (MatchesKey(column_family_id_, key));
+  }
+}
+
+void WBWIIteratorImpl::NextKey() { AdvanceKey(true); }
+
+void WBWIIteratorImpl::PrevKey() {
+  AdvanceKey(false);  // Move to the tail of the previous key
+  if (Valid()) {
+    AdvanceKey(false);  // Move back another key.  Now we are at the start of
+                        // the previous key
+    if (Valid()) {      // Still a valid
+      Next();           // Move forward one onto this key
+    } else {
+      SeekToFirst();  // Not valid, move to the start
+    }
+  }
+}
+
+WBWIIteratorImpl::Result WBWIIteratorImpl::FindLatestUpdate(
+    MergeContext* merge_context) {
+  if (Valid()) {
+    Slice key = Entry().key;
+    return FindLatestUpdate(key, merge_context);
+  } else {
+    merge_context->Clear();  // Clear any entries in the MergeContext
+    return WBWIIteratorImpl::kNotFound;
+  }
+}
+
+WBWIIteratorImpl::Result WBWIIteratorImpl::FindLatestUpdate(
+    const Slice& key, MergeContext* merge_context) {
+  Result result = WBWIIteratorImpl::kNotFound;
+  merge_context->Clear();  // Clear any entries in the MergeContext
+  // TODO(agiardullo): consider adding support for reverse iteration
+  if (!Valid()) {
+    return result;
+  } else if (comparator_->CompareKey(column_family_id_, Entry().key, key) !=
+             0) {
+    return result;
+  } else {
+    // We want to iterate in the reverse order that the writes were added to the
+    // batch.  Since we don't have a reverse iterator, we must seek past the
+    // end. We do this by seeking to the next key, and then back one step
+    NextKey();
+    if (Valid()) {
+      Prev();
+    } else {
+      SeekToLast();
+    }
+
+    // We are at the end of the iterator for this key.  Search backwards for the
+    // last Put or Delete, accumulating merges along the way.
+    while (Valid()) {
+      const WriteEntry entry = Entry();
+      if (comparator_->CompareKey(column_family_id_, entry.key, key) != 0) {
+        break;  // Unexpected error or we've reached a different next key
+      }
+
+      switch (entry.type) {
+        case kPutRecord:
+          return WBWIIteratorImpl::kFound;
+        case kDeleteRecord:
+          return WBWIIteratorImpl::kDeleted;
+        case kSingleDeleteRecord:
+          return WBWIIteratorImpl::kDeleted;
+        case kMergeRecord:
+          result = WBWIIteratorImpl::kMergeInProgress;
+          merge_context->PushOperand(entry.value);
+          break;
+        case kLogDataRecord:
+          break;  // ignore
+        case kXIDRecord:
+          break;  // ignore
+        default:
+          return WBWIIteratorImpl::kError;
+      }  // end switch statement
+      Prev();
+    }  // End while Valid()
+    // At this point, we have been through the whole list and found no Puts or
+    // Deletes. The iterator points to the previous key.  Move the iterator back
+    // onto this one.
+    if (Valid()) {
+      Next();
+    } else {
+      SeekToFirst();
+    }
+  }
+  return result;
+}
+
+Status ReadableWriteBatch::GetEntryFromDataOffset(size_t data_offset,
+                                                  WriteType* type, Slice* Key,
+                                                  Slice* value, Slice* blob,
+                                                  Slice* xid) const {
+  if (type == nullptr || Key == nullptr || value == nullptr ||
+      blob == nullptr || xid == nullptr) {
+    return Status::InvalidArgument("Output parameters cannot be null");
+  }
+
+  if (data_offset == GetDataSize()) {
+    // reached end of batch.
+    return Status::NotFound();
+  }
+
+  if (data_offset > GetDataSize()) {
+    return Status::InvalidArgument("data offset exceed write batch size");
+  }
+  Slice input = Slice(rep_.data() + data_offset, rep_.size() - data_offset);
+  char tag;
+  uint32_t column_family;
+  Status s = ReadRecordFromWriteBatch(&input, &tag, &column_family, Key, value,
+                                      blob, xid);
+  if (!s.ok()) {
+    return s;
+  }
+
+  switch (tag) {
+    case kTypeColumnFamilyValue:
+    case kTypeValue:
+      *type = kPutRecord;
+      break;
+    case kTypeColumnFamilyDeletion:
+    case kTypeDeletion:
+      *type = kDeleteRecord;
+      break;
+    case kTypeColumnFamilySingleDeletion:
+    case kTypeSingleDeletion:
+      *type = kSingleDeleteRecord;
+      break;
+    case kTypeColumnFamilyRangeDeletion:
+    case kTypeRangeDeletion:
+      *type = kDeleteRangeRecord;
+      break;
+    case kTypeColumnFamilyMerge:
+    case kTypeMerge:
+      *type = kMergeRecord;
+      break;
+    case kTypeLogData:
+      *type = kLogDataRecord;
+      break;
+    case kTypeNoop:
+    case kTypeBeginPrepareXID:
+    case kTypeBeginPersistedPrepareXID:
+    case kTypeBeginUnprepareXID:
+    case kTypeEndPrepareXID:
+    case kTypeCommitXID:
+    case kTypeRollbackXID:
+      *type = kXIDRecord;
+      break;
+    default:
+      return Status::Corruption("unknown WriteBatch tag ",
+                                std::to_string(static_cast<unsigned int>(tag)));
+  }
+  return Status::OK();
+}
+
+// If both of `entry1` and `entry2` point to real entry in write batch, we
+// compare the entries as following:
+// 1. first compare the column family, the one with larger CF will be larger;
+// 2. Inside the same CF, we first decode the entry to find the key of the entry
+//    and the entry with larger key will be larger;
+// 3. If two entries are of the same CF and key, the one with larger offset
+//    will be larger.
+// Some times either `entry1` or `entry2` is dummy entry, which is actually
+// a search key. In this case, in step 2, we don't go ahead and decode the
+// entry but use the value in WriteBatchIndexEntry::search_key.
+// One special case is WriteBatchIndexEntry::key_size is kFlagMinInCf.
+// This indicate that we are going to seek to the first of the column family.
+// Once we see this, this entry will be smaller than all the real entries of
+// the column family.
+int WriteBatchEntryComparator::operator()(
+    const WriteBatchIndexEntry* entry1,
+    const WriteBatchIndexEntry* entry2) const {
+  if (entry1->column_family > entry2->column_family) {
+    return 1;
+  } else if (entry1->column_family < entry2->column_family) {
+    return -1;
+  }
+
+  // Deal with special case of seeking to the beginning of a column family
+  if (entry1->is_min_in_cf()) {
+    return -1;
+  } else if (entry2->is_min_in_cf()) {
+    return 1;
+  }
+
+  Slice key1, key2;
+  if (entry1->search_key == nullptr) {
+    key1 = Slice(write_batch_->Data().data() + entry1->key_offset,
+                 entry1->key_size);
+  } else {
+    key1 = *(entry1->search_key);
+  }
+  if (entry2->search_key == nullptr) {
+    key2 = Slice(write_batch_->Data().data() + entry2->key_offset,
+                 entry2->key_size);
+  } else {
+    key2 = *(entry2->search_key);
+  }
+
+  int cmp = CompareKey(entry1->column_family, key1, key2);
+  if (cmp != 0) {
+    return cmp;
+  } else if (entry1->offset > entry2->offset) {
+    return 1;
+  } else if (entry1->offset < entry2->offset) {
+    return -1;
+  }
+  return 0;
+}
+
+int WriteBatchEntryComparator::CompareKey(uint32_t column_family,
+                                          const Slice& key1,
+                                          const Slice& key2) const {
+  if (column_family < cf_comparators_.size() &&
+      cf_comparators_[column_family] != nullptr) {
+    return cf_comparators_[column_family]->CompareWithoutTimestamp(
+        key1, /*a_has_ts=*/false, key2, /*b_has_ts=*/false);
+  } else {
+    return default_comparator_->CompareWithoutTimestamp(
+        key1, /*a_has_ts=*/false, key2, /*b_has_ts=*/false);
+  }
+}
+
+const Comparator* WriteBatchEntryComparator::GetComparator(
+    const ColumnFamilyHandle* column_family) const {
+  return column_family ? column_family->GetComparator() : default_comparator_;
+}
+
+const Comparator* WriteBatchEntryComparator::GetComparator(
+    uint32_t column_family) const {
+  if (column_family < cf_comparators_.size() &&
+      cf_comparators_[column_family]) {
+    return cf_comparators_[column_family];
+  }
+  return default_comparator_;
+}
+
+WriteEntry WBWIIteratorImpl::Entry() const {
+  WriteEntry ret;
+  Slice blob, xid;
+  const WriteBatchIndexEntry* iter_entry = skip_list_iter_.key();
+  // this is guaranteed with Valid()
+  assert(iter_entry != nullptr &&
+         iter_entry->column_family == column_family_id_);
+  auto s = write_batch_->GetEntryFromDataOffset(
+      iter_entry->offset, &ret.type, &ret.key, &ret.value, &blob, &xid);
+  assert(s.ok());
+  assert(ret.type == kPutRecord || ret.type == kDeleteRecord ||
+         ret.type == kSingleDeleteRecord || ret.type == kDeleteRangeRecord ||
+         ret.type == kMergeRecord);
+  // Make sure entry.key does not include user-defined timestamp.
+  const Comparator* const ucmp = comparator_->GetComparator(column_family_id_);
+  size_t ts_sz = ucmp->timestamp_size();
+  if (ts_sz > 0) {
+    ret.key = StripTimestampFromUserKey(ret.key, ts_sz);
+  }
+  return ret;
+}
+
+bool WBWIIteratorImpl::MatchesKey(uint32_t cf_id, const Slice& key) {
+  if (Valid()) {
+    return comparator_->CompareKey(cf_id, key, Entry().key) == 0;
+  } else {
+    return false;
+  }
+}
+
+WriteBatchWithIndexInternal::WriteBatchWithIndexInternal(
+    ColumnFamilyHandle* column_family)
+    : db_(nullptr), db_options_(nullptr), column_family_(column_family) {}
+
+WriteBatchWithIndexInternal::WriteBatchWithIndexInternal(
+    DB* db, ColumnFamilyHandle* column_family)
+    : db_(db), db_options_(nullptr), column_family_(column_family) {
+  if (db_ != nullptr && column_family_ == nullptr) {
+    column_family_ = db_->DefaultColumnFamily();
+  }
+}
+
+WriteBatchWithIndexInternal::WriteBatchWithIndexInternal(
+    const DBOptions* db_options, ColumnFamilyHandle* column_family)
+    : db_(nullptr), db_options_(db_options), column_family_(column_family) {}
+
+Status WriteBatchWithIndexInternal::MergeKey(const Slice& key,
+                                             const Slice* value,
+                                             const MergeContext& context,
+                                             std::string* result) const {
+  if (column_family_ != nullptr) {
+    auto cfh = static_cast_with_check<ColumnFamilyHandleImpl>(column_family_);
+    const auto merge_operator = cfh->cfd()->ioptions()->merge_operator.get();
+    if (merge_operator == nullptr) {
+      return Status::InvalidArgument(
+          "Merge_operator must be set for column_family");
+    } else if (db_ != nullptr) {
+      const ImmutableDBOptions& immutable_db_options =
+          static_cast_with_check<DBImpl>(db_->GetRootDB())
+              ->immutable_db_options();
+      Statistics* statistics = immutable_db_options.statistics.get();
+      Logger* logger = immutable_db_options.info_log.get();
+      SystemClock* clock = immutable_db_options.clock;
+      return MergeHelper::TimedFullMerge(
+          merge_operator, key, value, context.GetOperands(), result, logger,
+          statistics, clock, /* result_operand */ nullptr,
+          /* update_num_ops_stats */ false);
+    } else if (db_options_ != nullptr) {
+      Statistics* statistics = db_options_->statistics.get();
+      Env* env = db_options_->env;
+      Logger* logger = db_options_->info_log.get();
+      SystemClock* clock = env->GetSystemClock().get();
+      return MergeHelper::TimedFullMerge(
+          merge_operator, key, value, context.GetOperands(), result, logger,
+          statistics, clock, /* result_operand */ nullptr,
+          /* update_num_ops_stats */ false);
+    } else {
+      const auto cf_opts = cfh->cfd()->ioptions();
+      return MergeHelper::TimedFullMerge(
+          merge_operator, key, value, context.GetOperands(), result,
+          cf_opts->logger, cf_opts->stats, cf_opts->clock,
+          /* result_operand */ nullptr, /* update_num_ops_stats */ false);
+    }
+  } else {
+    return Status::InvalidArgument("Must provide a column_family");
+  }
+}
+
+WBWIIteratorImpl::Result WriteBatchWithIndexInternal::GetFromBatch(
+    WriteBatchWithIndex* batch, const Slice& key, MergeContext* context,
+    std::string* value, Status* s) {
+  *s = Status::OK();
+
+  std::unique_ptr<WBWIIteratorImpl> iter(
+      static_cast_with_check<WBWIIteratorImpl>(
+          batch->NewIterator(column_family_)));
+
+  // Search the iterator for this key, and updates/merges to it.
+  iter->Seek(key);
+  auto result = iter->FindLatestUpdate(key, context);
+  if (result == WBWIIteratorImpl::kError) {
+    (*s) = Status::Corruption("Unexpected entry in WriteBatchWithIndex:",
+                              std::to_string(iter->Entry().type));
+    return result;
+  } else if (result == WBWIIteratorImpl::kNotFound) {
+    return result;
+  } else if (result == WBWIIteratorImpl::Result::kFound) {  // PUT
+    Slice entry_value = iter->Entry().value;
+    if (context->GetNumOperands() > 0) {
+      *s = MergeKey(key, &entry_value, *context, value);
+      if (!s->ok()) {
+        result = WBWIIteratorImpl::Result::kError;
+      }
+    } else {
+      value->assign(entry_value.data(), entry_value.size());
+    }
+  } else if (result == WBWIIteratorImpl::kDeleted) {
+    if (context->GetNumOperands() > 0) {
+      *s = MergeKey(key, nullptr, *context, value);
+      if (s->ok()) {
+        result = WBWIIteratorImpl::Result::kFound;
+      } else {
+        result = WBWIIteratorImpl::Result::kError;
+      }
+    }
+  }
+  return result;
+}
+
+}  // namespace ROCKSDB_NAMESPACE
+
+#endif  // !ROCKSDB_LITE
diff --git a/src/rocksdb/utilities/write_batch_with_index/write_batch_with_index_internal.h b/src/rocksdb/utilities/write_batch_with_index/write_batch_with_index_internal.h
new file mode 100644
index 000000000..edabc95bc
--- /dev/null
+++ b/src/rocksdb/utilities/write_batch_with_index/write_batch_with_index_internal.h
@@ -0,0 +1,344 @@
+// Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+#pragma once
+
+#ifndef ROCKSDB_LITE
+
+#include <limits>
+#include <string>
+#include <vector>
+
+#include "db/merge_context.h"
+#include "memtable/skiplist.h"
+#include "options/db_options.h"
+#include "port/port.h"
+#include "rocksdb/comparator.h"
+#include "rocksdb/iterator.h"
+#include "rocksdb/slice.h"
+#include "rocksdb/status.h"
+#include "rocksdb/utilities/write_batch_with_index.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+class MergeContext;
+class WBWIIteratorImpl;
+class WriteBatchWithIndexInternal;
+struct Options;
+
+// when direction == forward
+// * current_at_base_ <=> base_iterator > delta_iterator
+// when direction == backwards
+// * current_at_base_ <=> base_iterator < delta_iterator
+// always:
+// * equal_keys_ <=> base_iterator == delta_iterator
+class BaseDeltaIterator : public Iterator {
+ public:
+  BaseDeltaIterator(ColumnFamilyHandle* column_family, Iterator* base_iterator,
+                    WBWIIteratorImpl* delta_iterator,
+                    const Comparator* comparator,
+                    const ReadOptions* read_options = nullptr);
+
+  ~BaseDeltaIterator() override {}
+
+  bool Valid() const override;
+  void SeekToFirst() override;
+  void SeekToLast() override;
+  void Seek(const Slice& k) override;
+  void SeekForPrev(const Slice& k) override;
+  void Next() override;
+  void Prev() override;
+  Slice key() const override;
+  Slice value() const override;
+  Status status() const override;
+  void Invalidate(Status s);
+
+ private:
+  void AssertInvariants();
+  void Advance();
+  void AdvanceDelta();
+  void AdvanceBase();
+  bool BaseValid() const;
+  bool DeltaValid() const;
+  void UpdateCurrent();
+
+  std::unique_ptr<WriteBatchWithIndexInternal> wbwii_;
+  bool forward_;
+  bool current_at_base_;
+  bool equal_keys_;
+  mutable Status status_;
+  std::unique_ptr<Iterator> base_iterator_;
+  std::unique_ptr<WBWIIteratorImpl> delta_iterator_;
+  const Comparator* comparator_;  // not owned
+  const Slice* iterate_upper_bound_;
+  mutable PinnableSlice merge_result_;
+};
+
+// Key used by skip list, as the binary searchable index of WriteBatchWithIndex.
+struct WriteBatchIndexEntry {
+  WriteBatchIndexEntry(size_t o, uint32_t c, size_t ko, size_t ksz)
+      : offset(o),
+        column_family(c),
+        key_offset(ko),
+        key_size(ksz),
+        search_key(nullptr) {}
+  // Create a dummy entry as the search key. This index entry won't be backed
+  // by an entry from the write batch, but a pointer to the search key. Or a
+  // special flag of offset can indicate we are seek to first.
+  // @_search_key: the search key
+  // @_column_family: column family
+  // @is_forward_direction: true for Seek(). False for SeekForPrev()
+  // @is_seek_to_first: true if we seek to the beginning of the column family
+  //                    _search_key should be null in this case.
+  WriteBatchIndexEntry(const Slice* _search_key, uint32_t _column_family,
+                       bool is_forward_direction, bool is_seek_to_first)
+      // For SeekForPrev(), we need to make the dummy entry larger than any
+      // entry who has the same search key. Otherwise, we'll miss those entries.
+      : offset(is_forward_direction ? 0 : std::numeric_limits<size_t>::max()),
+        column_family(_column_family),
+        key_offset(0),
+        key_size(is_seek_to_first ? kFlagMinInCf : 0),
+        search_key(_search_key) {
+    assert(_search_key != nullptr || is_seek_to_first);
+  }
+
+  // If this flag appears in the key_size, it indicates a
+  // key that is smaller than any other entry for the same column family.
+  static const size_t kFlagMinInCf = std::numeric_limits<size_t>::max();
+
+  bool is_min_in_cf() const {
+    assert(key_size != kFlagMinInCf ||
+           (key_offset == 0 && search_key == nullptr));
+    return key_size == kFlagMinInCf;
+  }
+
+  // offset of an entry in write batch's string buffer. If this is a dummy
+  // lookup key, in which case search_key != nullptr, offset is set to either
+  // 0 or max, only for comparison purpose. Because when entries have the same
+  // key, the entry with larger offset is larger, offset = 0 will make a seek
+  // key small or equal than all the entries with the seek key, so that Seek()
+  // will find all the entries of the same key. Similarly, offset = MAX will
+  // make the entry just larger than all entries with the search key so
+  // SeekForPrev() will see all the keys with the same key.
+  size_t offset;
+  uint32_t column_family;  // column family of the entry.
+  size_t key_offset;       // offset of the key in write batch's string buffer.
+  size_t key_size;         // size of the key. kFlagMinInCf indicates
+                           // that this is a dummy look up entry for
+                           // SeekToFirst() to the beginning of the column
+                           // family. We use the flag here to save a boolean
+                           // in the struct.
+
+  const Slice* search_key;  // if not null, instead of reading keys from
+                            // write batch, use it to compare. This is used
+                            // for lookup key.
+};
+
+class ReadableWriteBatch : public WriteBatch {
+ public:
+  explicit ReadableWriteBatch(size_t reserved_bytes = 0, size_t max_bytes = 0,
+                              size_t protection_bytes_per_key = 0,
+                              size_t default_cf_ts_sz = 0)
+      : WriteBatch(reserved_bytes, max_bytes, protection_bytes_per_key,
+                   default_cf_ts_sz) {}
+  // Retrieve some information from a write entry in the write batch, given
+  // the start offset of the write entry.
+  Status GetEntryFromDataOffset(size_t data_offset, WriteType* type, Slice* Key,
+                                Slice* value, Slice* blob, Slice* xid) const;
+};
+
+class WriteBatchEntryComparator {
+ public:
+  WriteBatchEntryComparator(const Comparator* _default_comparator,
+                            const ReadableWriteBatch* write_batch)
+      : default_comparator_(_default_comparator), write_batch_(write_batch) {}
+  // Compare a and b. Return a negative value if a is less than b, 0 if they
+  // are equal, and a positive value if a is greater than b
+  int operator()(const WriteBatchIndexEntry* entry1,
+                 const WriteBatchIndexEntry* entry2) const;
+
+  int CompareKey(uint32_t column_family, const Slice& key1,
+                 const Slice& key2) const;
+
+  void SetComparatorForCF(uint32_t column_family_id,
+                          const Comparator* comparator) {
+    if (column_family_id >= cf_comparators_.size()) {
+      cf_comparators_.resize(column_family_id + 1, nullptr);
+    }
+    cf_comparators_[column_family_id] = comparator;
+  }
+
+  const Comparator* default_comparator() { return default_comparator_; }
+
+  const Comparator* GetComparator(
+      const ColumnFamilyHandle* column_family) const;
+
+  const Comparator* GetComparator(uint32_t column_family) const;
+
+ private:
+  const Comparator* const default_comparator_;
+  std::vector<const Comparator*> cf_comparators_;
+  const ReadableWriteBatch* const write_batch_;
+};
+
+using WriteBatchEntrySkipList =
+    SkipList<WriteBatchIndexEntry*, const WriteBatchEntryComparator&>;
+
+class WBWIIteratorImpl : public WBWIIterator {
+ public:
+  enum Result : uint8_t {
+    kFound,
+    kDeleted,
+    kNotFound,
+    kMergeInProgress,
+    kError
+  };
+  WBWIIteratorImpl(uint32_t column_family_id,
+                   WriteBatchEntrySkipList* skip_list,
+                   const ReadableWriteBatch* write_batch,
+                   WriteBatchEntryComparator* comparator)
+      : column_family_id_(column_family_id),
+        skip_list_iter_(skip_list),
+        write_batch_(write_batch),
+        comparator_(comparator) {}
+
+  ~WBWIIteratorImpl() override {}
+
+  bool Valid() const override {
+    if (!skip_list_iter_.Valid()) {
+      return false;
+    }
+    const WriteBatchIndexEntry* iter_entry = skip_list_iter_.key();
+    return (iter_entry != nullptr &&
+            iter_entry->column_family == column_family_id_);
+  }
+
+  void SeekToFirst() override {
+    WriteBatchIndexEntry search_entry(
+        nullptr /* search_key */, column_family_id_,
+        true /* is_forward_direction */, true /* is_seek_to_first */);
+    skip_list_iter_.Seek(&search_entry);
+  }
+
+  void SeekToLast() override {
+    WriteBatchIndexEntry search_entry(
+        nullptr /* search_key */, column_family_id_ + 1,
+        true /* is_forward_direction */, true /* is_seek_to_first */);
+    skip_list_iter_.Seek(&search_entry);
+    if (!skip_list_iter_.Valid()) {
+      skip_list_iter_.SeekToLast();
+    } else {
+      skip_list_iter_.Prev();
+    }
+  }
+
+  void Seek(const Slice& key) override {
+    WriteBatchIndexEntry search_entry(&key, column_family_id_,
+                                      true /* is_forward_direction */,
+                                      false /* is_seek_to_first */);
+    skip_list_iter_.Seek(&search_entry);
+  }
+
+  void SeekForPrev(const Slice& key) override {
+    WriteBatchIndexEntry search_entry(&key, column_family_id_,
+                                      false /* is_forward_direction */,
+                                      false /* is_seek_to_first */);
+    skip_list_iter_.SeekForPrev(&search_entry);
+  }
+
+  void Next() override { skip_list_iter_.Next(); }
+
+  void Prev() override { skip_list_iter_.Prev(); }
+
+  WriteEntry Entry() const override;
+
+  Status status() const override {
+    // this is in-memory data structure, so the only way status can be non-ok is
+    // through memory corruption
+    return Status::OK();
+  }
+
+  const WriteBatchIndexEntry* GetRawEntry() const {
+    return skip_list_iter_.key();
+  }
+
+  bool MatchesKey(uint32_t cf_id, const Slice& key);
+
+  // Moves the iterator to first entry of the previous key.
+  void PrevKey();
+  // Moves the iterator to first entry of the next key.
+  void NextKey();
+
+  // Moves the iterator to the Update (Put or Delete) for the current key
+  // If there are no Put/Delete, the Iterator will point to the first entry for
+  // this key
+  // @return kFound if a Put was found for the key
+  // @return kDeleted if a delete was found for the key
+  // @return kMergeInProgress if only merges were fouund for the key
+  // @return kError if an unsupported operation was found for the key
+  // @return kNotFound if no operations were found for this key
+  //
+  Result FindLatestUpdate(const Slice& key, MergeContext* merge_context);
+  Result FindLatestUpdate(MergeContext* merge_context);
+
+ protected:
+  void AdvanceKey(bool forward);
+
+ private:
+  uint32_t column_family_id_;
+  WriteBatchEntrySkipList::Iterator skip_list_iter_;
+  const ReadableWriteBatch* write_batch_;
+  WriteBatchEntryComparator* comparator_;
+};
+
+class WriteBatchWithIndexInternal {
+ public:
+  static const Comparator* GetUserComparator(const WriteBatchWithIndex& wbwi,
+                                             uint32_t cf_id);
+
+  // For GetFromBatchAndDB or similar
+  explicit WriteBatchWithIndexInternal(DB* db,
+                                       ColumnFamilyHandle* column_family);
+  // For GetFromBatchAndDB or similar
+  explicit WriteBatchWithIndexInternal(ColumnFamilyHandle* column_family);
+  // For GetFromBatch or similar
+  explicit WriteBatchWithIndexInternal(const DBOptions* db_options,
+                                       ColumnFamilyHandle* column_family);
+
+  // If batch contains a value for key, store it in *value and return kFound.
+  // If batch contains a deletion for key, return Deleted.
+  // If batch contains Merge operations as the most recent entry for a key,
+  //   and the merge process does not stop (not reaching a value or delete),
+  //   prepend the current merge operands to *operands,
+  //   and return kMergeInProgress
+  // If batch does not contain this key, return kNotFound
+  // Else, return kError on error with error Status stored in *s.
+  WBWIIteratorImpl::Result GetFromBatch(WriteBatchWithIndex* batch,
+                                        const Slice& key, std::string* value,
+                                        Status* s) {
+    return GetFromBatch(batch, key, &merge_context_, value, s);
+  }
+  WBWIIteratorImpl::Result GetFromBatch(WriteBatchWithIndex* batch,
+                                        const Slice& key,
+                                        MergeContext* merge_context,
+                                        std::string* value, Status* s);
+  Status MergeKey(const Slice& key, const Slice* value,
+                  std::string* result) const {
+    return MergeKey(key, value, merge_context_, result);
+  }
+  Status MergeKey(const Slice& key, const Slice* value,
+                  const MergeContext& context, std::string* result) const;
+  size_t GetNumOperands() const { return merge_context_.GetNumOperands(); }
+  MergeContext* GetMergeContext() { return &merge_context_; }
+  Slice GetOperand(int index) const { return merge_context_.GetOperand(index); }
+
+ private:
+  DB* db_;
+  const DBOptions* db_options_;
+  ColumnFamilyHandle* column_family_;
+  MergeContext merge_context_;
+};
+
+}  // namespace ROCKSDB_NAMESPACE
+#endif  // !ROCKSDB_LITE
diff --git a/src/rocksdb/utilities/write_batch_with_index/write_batch_with_index_test.cc b/src/rocksdb/utilities/write_batch_with_index/write_batch_with_index_test.cc
new file mode 100644
index 000000000..350dcc881
--- /dev/null
+++ b/src/rocksdb/utilities/write_batch_with_index/write_batch_with_index_test.cc
@@ -0,0 +1,2419 @@
+//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under both the GPLv2 (found in the
+//  COPYING file in the root directory) and Apache 2.0 License
+//  (found in the LICENSE.Apache file in the root directory).
+//
+// Copyright (c) 2011 The LevelDB Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file. See the AUTHORS file for names of contributors.
+
+#ifndef ROCKSDB_LITE
+
+#include "rocksdb/utilities/write_batch_with_index.h"
+
+#include <map>
+#include <memory>
+
+#include "db/column_family.h"
+#include "port/stack_trace.h"
+#include "test_util/testharness.h"
+#include "test_util/testutil.h"
+#include "util/random.h"
+#include "util/string_util.h"
+#include "utilities/merge_operators.h"
+#include "utilities/merge_operators/string_append/stringappend.h"
+#include "utilities/write_batch_with_index/write_batch_with_index_internal.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+namespace {
+class ColumnFamilyHandleImplDummy : public ColumnFamilyHandleImpl {
+ public:
+  explicit ColumnFamilyHandleImplDummy(int id, const Comparator* comparator)
+      : ColumnFamilyHandleImpl(nullptr, nullptr, nullptr),
+        id_(id),
+        comparator_(comparator) {}
+  uint32_t GetID() const override { return id_; }
+  const Comparator* GetComparator() const override { return comparator_; }
+
+ private:
+  uint32_t id_;
+  const Comparator* comparator_;
+};
+
+struct Entry {
+  std::string key;
+  std::string value;
+  WriteType type;
+};
+
+struct TestHandler : public WriteBatch::Handler {
+  std::map<uint32_t, std::vector<Entry>> seen;
+  Status PutCF(uint32_t column_family_id, const Slice& key,
+               const Slice& value) override {
+    Entry e;
+    e.key = key.ToString();
+    e.value = value.ToString();
+    e.type = kPutRecord;
+    seen[column_family_id].push_back(e);
+    return Status::OK();
+  }
+  Status MergeCF(uint32_t column_family_id, const Slice& key,
+                 const Slice& value) override {
+    Entry e;
+    e.key = key.ToString();
+    e.value = value.ToString();
+    e.type = kMergeRecord;
+    seen[column_family_id].push_back(e);
+    return Status::OK();
+  }
+  void LogData(const Slice& /*blob*/) override {}
+  Status DeleteCF(uint32_t column_family_id, const Slice& key) override {
+    Entry e;
+    e.key = key.ToString();
+    e.value = "";
+    e.type = kDeleteRecord;
+    seen[column_family_id].push_back(e);
+    return Status::OK();
+  }
+};
+
+using KVMap = std::map<std::string, std::string>;
+
+class KVIter : public Iterator {
+ public:
+  explicit KVIter(const KVMap* map) : map_(map), iter_(map_->end()) {}
+  bool Valid() const override { return iter_ != map_->end(); }
+  void SeekToFirst() override { iter_ = map_->begin(); }
+  void SeekToLast() override {
+    if (map_->empty()) {
+      iter_ = map_->end();
+    } else {
+      iter_ = map_->find(map_->rbegin()->first);
+    }
+  }
+  void Seek(const Slice& k) override {
+    iter_ = map_->lower_bound(k.ToString());
+  }
+  void SeekForPrev(const Slice& k) override {
+    iter_ = map_->upper_bound(k.ToString());
+    Prev();
+  }
+  void Next() override { ++iter_; }
+  void Prev() override {
+    if (iter_ == map_->begin()) {
+      iter_ = map_->end();
+      return;
+    }
+    --iter_;
+  }
+  Slice key() const override { return iter_->first; }
+  Slice value() const override { return iter_->second; }
+  Status status() const override { return Status::OK(); }
+
+ private:
+  const KVMap* const map_;
+  KVMap::const_iterator iter_;
+};
+
+static std::string PrintContents(WriteBatchWithIndex* batch,
+                                 ColumnFamilyHandle* column_family,
+                                 bool hex = false) {
+  std::string result;
+
+  WBWIIterator* iter;
+  if (column_family == nullptr) {
+    iter = batch->NewIterator();
+  } else {
+    iter = batch->NewIterator(column_family);
+  }
+
+  iter->SeekToFirst();
+  while (iter->Valid()) {
+    WriteEntry e = iter->Entry();
+
+    if (e.type == kPutRecord) {
+      result.append("PUT(");
+      result.append(e.key.ToString(hex));
+      result.append("):");
+      result.append(e.value.ToString(hex));
+    } else if (e.type == kMergeRecord) {
+      result.append("MERGE(");
+      result.append(e.key.ToString(hex));
+      result.append("):");
+      result.append(e.value.ToString(hex));
+    } else if (e.type == kSingleDeleteRecord) {
+      result.append("SINGLE-DEL(");
+      result.append(e.key.ToString(hex));
+      result.append(")");
+    } else {
+      assert(e.type == kDeleteRecord);
+      result.append("DEL(");
+      result.append(e.key.ToString(hex));
+      result.append(")");
+    }
+
+    result.append(",");
+    iter->Next();
+  }
+
+  delete iter;
+  return result;
+}
+
+static std::string PrintContents(WriteBatchWithIndex* batch, KVMap* base_map,
+                                 ColumnFamilyHandle* column_family) {
+  std::string result;
+
+  Iterator* iter;
+  if (column_family == nullptr) {
+    iter = batch->NewIteratorWithBase(new KVIter(base_map));
+  } else {
+    iter = batch->NewIteratorWithBase(column_family, new KVIter(base_map));
+  }
+
+  iter->SeekToFirst();
+  while (iter->Valid()) {
+    assert(iter->status().ok());
+
+    Slice key = iter->key();
+    Slice value = iter->value();
+
+    result.append(key.ToString());
+    result.append(":");
+    result.append(value.ToString());
+    result.append(",");
+
+    iter->Next();
+  }
+
+  delete iter;
+  return result;
+}
+
+void AssertIter(Iterator* iter, const std::string& key,
+                const std::string& value) {
+  ASSERT_OK(iter->status());
+  ASSERT_TRUE(iter->Valid());
+  ASSERT_EQ(key, iter->key().ToString());
+  ASSERT_EQ(value, iter->value().ToString());
+}
+
+void AssertItersMatch(Iterator* iter1, Iterator* iter2) {
+  ASSERT_EQ(iter1->Valid(), iter2->Valid());
+  if (iter1->Valid()) {
+    ASSERT_EQ(iter1->key().ToString(), iter2->key().ToString());
+    ASSERT_EQ(iter1->value().ToString(), iter2->value().ToString());
+  }
+}
+
+void AssertItersEqual(Iterator* iter1, Iterator* iter2) {
+  iter1->SeekToFirst();
+  iter2->SeekToFirst();
+  while (iter1->Valid()) {
+    ASSERT_EQ(iter1->Valid(), iter2->Valid());
+    ASSERT_EQ(iter1->key().ToString(), iter2->key().ToString());
+    ASSERT_EQ(iter1->value().ToString(), iter2->value().ToString());
+    iter1->Next();
+    iter2->Next();
+  }
+  ASSERT_EQ(iter1->Valid(), iter2->Valid());
+}
+
+void AssertIterEqual(WBWIIteratorImpl* wbwii,
+                     const std::vector<std::string>& keys) {
+  wbwii->SeekToFirst();
+  for (auto k : keys) {
+    ASSERT_TRUE(wbwii->Valid());
+    ASSERT_EQ(wbwii->Entry().key, k);
+    wbwii->NextKey();
+  }
+  ASSERT_FALSE(wbwii->Valid());
+  wbwii->SeekToLast();
+  for (auto kit = keys.rbegin(); kit != keys.rend(); ++kit) {
+    ASSERT_TRUE(wbwii->Valid());
+    ASSERT_EQ(wbwii->Entry().key, *kit);
+    wbwii->PrevKey();
+  }
+  ASSERT_FALSE(wbwii->Valid());
+}
+}  // namespace
+
+class WBWIBaseTest : public testing::Test {
+ public:
+  explicit WBWIBaseTest(bool overwrite) : db_(nullptr) {
+    options_.merge_operator =
+        MergeOperators::CreateFromStringId("stringappend");
+    options_.create_if_missing = true;
+    dbname_ = test::PerThreadDBPath("write_batch_with_index_test");
+    EXPECT_OK(DestroyDB(dbname_, options_));
+    batch_.reset(new WriteBatchWithIndex(BytewiseComparator(), 20, overwrite));
+  }
+
+  virtual ~WBWIBaseTest() {
+    if (db_ != nullptr) {
+      ReleaseSnapshot();
+      delete db_;
+      EXPECT_OK(DestroyDB(dbname_, options_));
+    }
+  }
+
+  std::string AddToBatch(ColumnFamilyHandle* cf, const std::string& key) {
+    std::string result;
+    for (size_t i = 0; i < key.size(); i++) {
+      if (key[i] == 'd') {
+        batch_->Delete(cf, key);
+        result = "";
+      } else if (key[i] == 'p') {
+        result = key + std::to_string(i);
+        batch_->Put(cf, key, result);
+      } else if (key[i] == 'm') {
+        std::string value = key + std::to_string(i);
+        batch_->Merge(cf, key, value);
+        if (result.empty()) {
+          result = value;
+        } else {
+          result = result + "," + value;
+        }
+      }
+    }
+    return result;
+  }
+
+  virtual Status OpenDB() { return DB::Open(options_, dbname_, &db_); }
+
+  void ReleaseSnapshot() {
+    if (read_opts_.snapshot != nullptr) {
+      EXPECT_NE(db_, nullptr);
+      db_->ReleaseSnapshot(read_opts_.snapshot);
+      read_opts_.snapshot = nullptr;
+    }
+  }
+
+ public:
+  DB* db_;
+  std::string dbname_;
+  Options options_;
+  WriteOptions write_opts_;
+  ReadOptions read_opts_;
+  std::unique_ptr<WriteBatchWithIndex> batch_;
+};
+
+class WBWIKeepTest : public WBWIBaseTest {
+ public:
+  WBWIKeepTest() : WBWIBaseTest(false) {}
+};
+
+class WBWIOverwriteTest : public WBWIBaseTest {
+ public:
+  WBWIOverwriteTest() : WBWIBaseTest(true) {}
+};
+class WriteBatchWithIndexTest : public WBWIBaseTest,
+                                public testing::WithParamInterface<bool> {
+ public:
+  WriteBatchWithIndexTest() : WBWIBaseTest(GetParam()) {}
+};
+
+void TestValueAsSecondaryIndexHelper(std::vector<Entry> entries,
+                                     WriteBatchWithIndex* batch) {
+  // In this test, we insert <key, value> to column family `data`, and
+  // <value, key> to column family `index`. Then iterator them in order
+  // and seek them by key.
+
+  // Sort entries by key
+  std::map<std::string, std::vector<Entry*>> data_map;
+  // Sort entries by value
+  std::map<std::string, std::vector<Entry*>> index_map;
+  for (auto& e : entries) {
+    data_map[e.key].push_back(&e);
+    index_map[e.value].push_back(&e);
+  }
+
+  ColumnFamilyHandleImplDummy data(6, BytewiseComparator());
+  ColumnFamilyHandleImplDummy index(8, BytewiseComparator());
+  for (auto& e : entries) {
+    if (e.type == kPutRecord) {
+      ASSERT_OK(batch->Put(&data, e.key, e.value));
+      ASSERT_OK(batch->Put(&index, e.value, e.key));
+    } else if (e.type == kMergeRecord) {
+      ASSERT_OK(batch->Merge(&data, e.key, e.value));
+      ASSERT_OK(batch->Put(&index, e.value, e.key));
+    } else {
+      assert(e.type == kDeleteRecord);
+      std::unique_ptr<WBWIIterator> iter(batch->NewIterator(&data));
+      iter->Seek(e.key);
+      ASSERT_OK(iter->status());
+      auto write_entry = iter->Entry();
+      ASSERT_EQ(e.key, write_entry.key.ToString());
+      ASSERT_EQ(e.value, write_entry.value.ToString());
+      ASSERT_OK(batch->Delete(&data, e.key));
+      ASSERT_OK(batch->Put(&index, e.value, ""));
+    }
+  }
+
+  // Iterator all keys
+  {
+    std::unique_ptr<WBWIIterator> iter(batch->NewIterator(&data));
+    for (int seek_to_first : {0, 1}) {
+      if (seek_to_first) {
+        iter->SeekToFirst();
+      } else {
+        iter->Seek("");
+      }
+      for (auto pair : data_map) {
+        for (auto v : pair.second) {
+          ASSERT_OK(iter->status());
+          ASSERT_TRUE(iter->Valid());
+          auto write_entry = iter->Entry();
+          ASSERT_EQ(pair.first, write_entry.key.ToString());
+          ASSERT_EQ(v->type, write_entry.type);
+          if (write_entry.type != kDeleteRecord) {
+            ASSERT_EQ(v->value, write_entry.value.ToString());
+          }
+          iter->Next();
+        }
+      }
+      ASSERT_TRUE(!iter->Valid());
+    }
+    iter->SeekToLast();
+    for (auto pair = data_map.rbegin(); pair != data_map.rend(); ++pair) {
+      for (auto v = pair->second.rbegin(); v != pair->second.rend(); v++) {
+        ASSERT_OK(iter->status());
+        ASSERT_TRUE(iter->Valid());
+        auto write_entry = iter->Entry();
+        ASSERT_EQ(pair->first, write_entry.key.ToString());
+        ASSERT_EQ((*v)->type, write_entry.type);
+        if (write_entry.type != kDeleteRecord) {
+          ASSERT_EQ((*v)->value, write_entry.value.ToString());
+        }
+        iter->Prev();
+      }
+    }
+    ASSERT_TRUE(!iter->Valid());
+  }
+
+  // Iterator all indexes
+  {
+    std::unique_ptr<WBWIIterator> iter(batch->NewIterator(&index));
+    for (int seek_to_first : {0, 1}) {
+      if (seek_to_first) {
+        iter->SeekToFirst();
+      } else {
+        iter->Seek("");
+      }
+      for (auto pair : index_map) {
+        for (auto v : pair.second) {
+          ASSERT_OK(iter->status());
+          ASSERT_TRUE(iter->Valid());
+          auto write_entry = iter->Entry();
+          ASSERT_EQ(pair.first, write_entry.key.ToString());
+          if (v->type != kDeleteRecord) {
+            ASSERT_EQ(v->key, write_entry.value.ToString());
+            ASSERT_EQ(v->value, write_entry.key.ToString());
+          }
+          iter->Next();
+        }
+      }
+      ASSERT_TRUE(!iter->Valid());
+    }
+
+    iter->SeekToLast();
+    for (auto pair = index_map.rbegin(); pair != index_map.rend(); ++pair) {
+      for (auto v = pair->second.rbegin(); v != pair->second.rend(); v++) {
+        ASSERT_OK(iter->status());
+        ASSERT_TRUE(iter->Valid());
+        auto write_entry = iter->Entry();
+        ASSERT_EQ(pair->first, write_entry.key.ToString());
+        if ((*v)->type != kDeleteRecord) {
+          ASSERT_EQ((*v)->key, write_entry.value.ToString());
+          ASSERT_EQ((*v)->value, write_entry.key.ToString());
+        }
+        iter->Prev();
+      }
+    }
+    ASSERT_TRUE(!iter->Valid());
+  }
+
+  // Seek to every key
+  {
+    std::unique_ptr<WBWIIterator> iter(batch->NewIterator(&data));
+
+    // Seek the keys one by one in reverse order
+    for (auto pair = data_map.rbegin(); pair != data_map.rend(); ++pair) {
+      iter->Seek(pair->first);
+      ASSERT_OK(iter->status());
+      for (auto v : pair->second) {
+        ASSERT_TRUE(iter->Valid());
+        auto write_entry = iter->Entry();
+        ASSERT_EQ(pair->first, write_entry.key.ToString());
+        ASSERT_EQ(v->type, write_entry.type);
+        if (write_entry.type != kDeleteRecord) {
+          ASSERT_EQ(v->value, write_entry.value.ToString());
+        }
+        iter->Next();
+        ASSERT_OK(iter->status());
+      }
+    }
+  }
+
+  // Seek to every index
+  {
+    std::unique_ptr<WBWIIterator> iter(batch->NewIterator(&index));
+
+    // Seek the keys one by one in reverse order
+    for (auto pair = index_map.rbegin(); pair != index_map.rend(); ++pair) {
+      iter->Seek(pair->first);
+      ASSERT_OK(iter->status());
+      for (auto v : pair->second) {
+        ASSERT_TRUE(iter->Valid());
+        auto write_entry = iter->Entry();
+        ASSERT_EQ(pair->first, write_entry.key.ToString());
+        ASSERT_EQ(v->value, write_entry.key.ToString());
+        if (v->type != kDeleteRecord) {
+          ASSERT_EQ(v->key, write_entry.value.ToString());
+        }
+        iter->Next();
+        ASSERT_OK(iter->status());
+      }
+    }
+  }
+
+  // Verify WriteBatch can be iterated
+  TestHandler handler;
+  ASSERT_OK(batch->GetWriteBatch()->Iterate(&handler));
+
+  // Verify data column family
+  {
+    ASSERT_EQ(entries.size(), handler.seen[data.GetID()].size());
+    size_t i = 0;
+    for (auto e : handler.seen[data.GetID()]) {
+      auto write_entry = entries[i++];
+      ASSERT_EQ(e.type, write_entry.type);
+      ASSERT_EQ(e.key, write_entry.key);
+      if (e.type != kDeleteRecord) {
+        ASSERT_EQ(e.value, write_entry.value);
+      }
+    }
+  }
+
+  // Verify index column family
+  {
+    ASSERT_EQ(entries.size(), handler.seen[index.GetID()].size());
+    size_t i = 0;
+    for (auto e : handler.seen[index.GetID()]) {
+      auto write_entry = entries[i++];
+      ASSERT_EQ(e.key, write_entry.value);
+      if (write_entry.type != kDeleteRecord) {
+        ASSERT_EQ(e.value, write_entry.key);
+      }
+    }
+  }
+}
+
+TEST_F(WBWIKeepTest, TestValueAsSecondaryIndex) {
+  Entry entries[] = {
+      {"aaa", "0005", kPutRecord},   {"b", "0002", kPutRecord},
+      {"cdd", "0002", kMergeRecord}, {"aab", "00001", kPutRecord},
+      {"cc", "00005", kPutRecord},   {"cdd", "0002", kPutRecord},
+      {"aab", "0003", kPutRecord},   {"cc", "00005", kDeleteRecord},
+  };
+  std::vector<Entry> entries_list(entries, entries + 8);
+
+  batch_.reset(new WriteBatchWithIndex(nullptr, 20, false));
+
+  TestValueAsSecondaryIndexHelper(entries_list, batch_.get());
+
+  // Clear batch and re-run test with new values
+  batch_->Clear();
+
+  Entry new_entries[] = {
+      {"aaa", "0005", kPutRecord},   {"e", "0002", kPutRecord},
+      {"add", "0002", kMergeRecord}, {"aab", "00001", kPutRecord},
+      {"zz", "00005", kPutRecord},   {"add", "0002", kPutRecord},
+      {"aab", "0003", kPutRecord},   {"zz", "00005", kDeleteRecord},
+  };
+
+  entries_list = std::vector<Entry>(new_entries, new_entries + 8);
+
+  TestValueAsSecondaryIndexHelper(entries_list, batch_.get());
+}
+
+TEST_P(WriteBatchWithIndexTest, TestComparatorForCF) {
+  ColumnFamilyHandleImplDummy cf1(6, nullptr);
+  ColumnFamilyHandleImplDummy reverse_cf(66, ReverseBytewiseComparator());
+  ColumnFamilyHandleImplDummy cf2(88, BytewiseComparator());
+
+  ASSERT_OK(batch_->Put(&cf1, "ddd", ""));
+  ASSERT_OK(batch_->Put(&cf2, "aaa", ""));
+  ASSERT_OK(batch_->Put(&cf2, "eee", ""));
+  ASSERT_OK(batch_->Put(&cf1, "ccc", ""));
+  ASSERT_OK(batch_->Put(&reverse_cf, "a11", ""));
+  ASSERT_OK(batch_->Put(&cf1, "bbb", ""));
+
+  Slice key_slices[] = {"a", "3", "3"};
+  Slice value_slice = "";
+  ASSERT_OK(batch_->Put(&reverse_cf, SliceParts(key_slices, 3),
+                        SliceParts(&value_slice, 1)));
+  ASSERT_OK(batch_->Put(&reverse_cf, "a22", ""));
+
+  {
+    std::unique_ptr<WBWIIterator> iter(batch_->NewIterator(&cf1));
+    iter->Seek("");
+    ASSERT_OK(iter->status());
+    ASSERT_TRUE(iter->Valid());
+    ASSERT_EQ("bbb", iter->Entry().key.ToString());
+    iter->Next();
+    ASSERT_OK(iter->status());
+    ASSERT_TRUE(iter->Valid());
+    ASSERT_EQ("ccc", iter->Entry().key.ToString());
+    iter->Next();
+    ASSERT_OK(iter->status());
+    ASSERT_TRUE(iter->Valid());
+    ASSERT_EQ("ddd", iter->Entry().key.ToString());
+    iter->Next();
+    ASSERT_OK(iter->status());
+    ASSERT_TRUE(!iter->Valid());
+  }
+
+  {
+    std::unique_ptr<WBWIIterator> iter(batch_->NewIterator(&cf2));
+    iter->Seek("");
+    ASSERT_OK(iter->status());
+    ASSERT_TRUE(iter->Valid());
+    ASSERT_EQ("aaa", iter->Entry().key.ToString());
+    iter->Next();
+    ASSERT_OK(iter->status());
+    ASSERT_TRUE(iter->Valid());
+    ASSERT_EQ("eee", iter->Entry().key.ToString());
+    iter->Next();
+    ASSERT_OK(iter->status());
+    ASSERT_TRUE(!iter->Valid());
+  }
+
+  {
+    std::unique_ptr<WBWIIterator> iter(batch_->NewIterator(&reverse_cf));
+    iter->Seek("");
+    ASSERT_OK(iter->status());
+    ASSERT_TRUE(!iter->Valid());
+
+    iter->Seek("z");
+    ASSERT_OK(iter->status());
+    ASSERT_TRUE(iter->Valid());
+    ASSERT_EQ("a33", iter->Entry().key.ToString());
+    iter->Next();
+    ASSERT_OK(iter->status());
+    ASSERT_TRUE(iter->Valid());
+    ASSERT_EQ("a22", iter->Entry().key.ToString());
+    iter->Next();
+    ASSERT_OK(iter->status());
+    ASSERT_TRUE(iter->Valid());
+    ASSERT_EQ("a11", iter->Entry().key.ToString());
+    iter->Next();
+    ASSERT_OK(iter->status());
+    ASSERT_TRUE(!iter->Valid());
+
+    iter->Seek("a22");
+    ASSERT_OK(iter->status());
+    ASSERT_TRUE(iter->Valid());
+    ASSERT_EQ("a22", iter->Entry().key.ToString());
+
+    iter->Seek("a13");
+    ASSERT_OK(iter->status());
+    ASSERT_TRUE(iter->Valid());
+    ASSERT_EQ("a11", iter->Entry().key.ToString());
+  }
+}
+
+TEST_F(WBWIOverwriteTest, TestOverwriteKey) {
+  ColumnFamilyHandleImplDummy cf1(6, nullptr);
+  ColumnFamilyHandleImplDummy reverse_cf(66, ReverseBytewiseComparator());
+  ColumnFamilyHandleImplDummy cf2(88, BytewiseComparator());
+
+  ASSERT_OK(batch_->Merge(&cf1, "ddd", ""));
+  ASSERT_OK(batch_->Put(&cf1, "ddd", ""));
+  ASSERT_OK(batch_->Delete(&cf1, "ddd"));
+  ASSERT_OK(batch_->Put(&cf2, "aaa", ""));
+  ASSERT_OK(batch_->Delete(&cf2, "aaa"));
+  ASSERT_OK(batch_->Put(&cf2, "aaa", "aaa"));
+  ASSERT_OK(batch_->Put(&cf2, "eee", "eee"));
+  ASSERT_OK(batch_->Put(&cf1, "ccc", ""));
+  ASSERT_OK(batch_->Put(&reverse_cf, "a11", ""));
+  ASSERT_OK(batch_->Delete(&cf1, "ccc"));
+  ASSERT_OK(batch_->Put(&reverse_cf, "a33", "a33"));
+  ASSERT_OK(batch_->Put(&reverse_cf, "a11", "a11"));
+  Slice slices[] = {"a", "3", "3"};
+  ASSERT_OK(batch_->Delete(&reverse_cf, SliceParts(slices, 3)));
+
+  {
+    std::unique_ptr<WBWIIterator> iter(batch_->NewIterator(&cf1));
+    iter->Seek("");
+    ASSERT_OK(iter->status());
+    ASSERT_TRUE(iter->Valid());
+    ASSERT_EQ("ccc", iter->Entry().key.ToString());
+    ASSERT_TRUE(iter->Entry().type == WriteType::kDeleteRecord);
+    iter->Next();
+    ASSERT_OK(iter->status());
+    ASSERT_TRUE(iter->Valid());
+    ASSERT_EQ("ddd", iter->Entry().key.ToString());
+    ASSERT_TRUE(iter->Entry().type == WriteType::kDeleteRecord);
+    iter->Next();
+    ASSERT_OK(iter->status());
+    ASSERT_TRUE(!iter->Valid());
+  }
+
+  {
+    std::unique_ptr<WBWIIterator> iter(batch_->NewIterator(&cf2));
+    iter->SeekToLast();
+    ASSERT_OK(iter->status());
+    ASSERT_TRUE(iter->Valid());
+    ASSERT_EQ("eee", iter->Entry().key.ToString());
+    ASSERT_EQ("eee", iter->Entry().value.ToString());
+    iter->Prev();
+    ASSERT_OK(iter->status());
+    ASSERT_TRUE(iter->Valid());
+    ASSERT_EQ("aaa", iter->Entry().key.ToString());
+    ASSERT_EQ("aaa", iter->Entry().value.ToString());
+    iter->Prev();
+    ASSERT_OK(iter->status());
+    ASSERT_TRUE(!iter->Valid());
+
+    iter->SeekToFirst();
+    ASSERT_OK(iter->status());
+    ASSERT_TRUE(iter->Valid());
+    ASSERT_EQ("aaa", iter->Entry().key.ToString());
+    ASSERT_EQ("aaa", iter->Entry().value.ToString());
+    iter->Next();
+    ASSERT_OK(iter->status());
+    ASSERT_TRUE(iter->Valid());
+    ASSERT_EQ("eee", iter->Entry().key.ToString());
+    ASSERT_EQ("eee", iter->Entry().value.ToString());
+    iter->Next();
+    ASSERT_OK(iter->status());
+    ASSERT_TRUE(!iter->Valid());
+  }
+
+  {
+    std::unique_ptr<WBWIIterator> iter(batch_->NewIterator(&reverse_cf));
+    iter->Seek("");
+    ASSERT_OK(iter->status());
+    ASSERT_TRUE(!iter->Valid());
+
+    iter->Seek("z");
+    ASSERT_OK(iter->status());
+    ASSERT_TRUE(iter->Valid());
+    ASSERT_EQ("a33", iter->Entry().key.ToString());
+    ASSERT_TRUE(iter->Entry().type == WriteType::kDeleteRecord);
+    iter->Next();
+    ASSERT_OK(iter->status());
+    ASSERT_TRUE(iter->Valid());
+    ASSERT_EQ("a11", iter->Entry().key.ToString());
+    ASSERT_EQ("a11", iter->Entry().value.ToString());
+    iter->Next();
+    ASSERT_OK(iter->status());
+    ASSERT_TRUE(!iter->Valid());
+
+    iter->SeekToLast();
+    ASSERT_TRUE(iter->Valid());
+    ASSERT_EQ("a11", iter->Entry().key.ToString());
+    ASSERT_EQ("a11", iter->Entry().value.ToString());
+    iter->Prev();
+
+    ASSERT_OK(iter->status());
+    ASSERT_TRUE(iter->Valid());
+    ASSERT_EQ("a33", iter->Entry().key.ToString());
+    ASSERT_TRUE(iter->Entry().type == WriteType::kDeleteRecord);
+    iter->Prev();
+    ASSERT_TRUE(!iter->Valid());
+  }
+}
+
+TEST_P(WriteBatchWithIndexTest, TestWBWIIterator) {
+  ColumnFamilyHandleImplDummy cf1(1, BytewiseComparator());
+  ColumnFamilyHandleImplDummy cf2(2, BytewiseComparator());
+  ASSERT_OK(batch_->Put(&cf1, "a", "a1"));
+  ASSERT_OK(batch_->Put(&cf1, "c", "c1"));
+  ASSERT_OK(batch_->Put(&cf1, "c", "c2"));
+  ASSERT_OK(batch_->Put(&cf1, "e", "e1"));
+  ASSERT_OK(batch_->Put(&cf1, "e", "e2"));
+  ASSERT_OK(batch_->Put(&cf1, "e", "e3"));
+  std::unique_ptr<WBWIIteratorImpl> iter1(
+      static_cast<WBWIIteratorImpl*>(batch_->NewIterator(&cf1)));
+  std::unique_ptr<WBWIIteratorImpl> iter2(
+      static_cast<WBWIIteratorImpl*>(batch_->NewIterator(&cf2)));
+  AssertIterEqual(iter1.get(), {"a", "c", "e"});
+  AssertIterEqual(iter2.get(), {});
+  ASSERT_OK(batch_->Put(&cf2, "a", "a2"));
+  ASSERT_OK(batch_->Merge(&cf2, "b", "b1"));
+  ASSERT_OK(batch_->Merge(&cf2, "b", "b2"));
+  ASSERT_OK(batch_->Delete(&cf2, "d"));
+  ASSERT_OK(batch_->Merge(&cf2, "d", "d2"));
+  ASSERT_OK(batch_->Merge(&cf2, "d", "d3"));
+  ASSERT_OK(batch_->Delete(&cf2, "f"));
+  AssertIterEqual(iter1.get(), {"a", "c", "e"});
+  AssertIterEqual(iter2.get(), {"a", "b", "d", "f"});
+}
+
+TEST_P(WriteBatchWithIndexTest, TestRandomIteraratorWithBase) {
+  std::vector<std::string> source_strings = {"a", "b", "c", "d", "e",
+                                             "f", "g", "h", "i", "j"};
+  for (int rand_seed = 301; rand_seed < 366; rand_seed++) {
+    Random rnd(rand_seed);
+
+    ColumnFamilyHandleImplDummy cf1(6, BytewiseComparator());
+    ColumnFamilyHandleImplDummy cf2(2, BytewiseComparator());
+    ColumnFamilyHandleImplDummy cf3(8, BytewiseComparator());
+    batch_->Clear();
+
+    if (rand_seed % 2 == 0) {
+      ASSERT_OK(batch_->Put(&cf2, "zoo", "bar"));
+    }
+    if (rand_seed % 4 == 1) {
+      ASSERT_OK(batch_->Put(&cf3, "zoo", "bar"));
+    }
+
+    KVMap map;
+    KVMap merged_map;
+    for (auto key : source_strings) {
+      std::string value = key + key;
+      int type = rnd.Uniform(6);
+      switch (type) {
+        case 0:
+          // only base has it
+          map[key] = value;
+          merged_map[key] = value;
+          break;
+        case 1:
+          // only delta has it
+          ASSERT_OK(batch_->Put(&cf1, key, value));
+          map[key] = value;
+          merged_map[key] = value;
+          break;
+        case 2:
+          // both has it. Delta should win
+          ASSERT_OK(batch_->Put(&cf1, key, value));
+          map[key] = "wrong_value";
+          merged_map[key] = value;
+          break;
+        case 3:
+          // both has it. Delta is delete
+          ASSERT_OK(batch_->Delete(&cf1, key));
+          map[key] = "wrong_value";
+          break;
+        case 4:
+          // only delta has it. Delta is delete
+          ASSERT_OK(batch_->Delete(&cf1, key));
+          map[key] = "wrong_value";
+          break;
+        default:
+          // Neither iterator has it.
+          break;
+      }
+    }
+
+    std::unique_ptr<Iterator> iter(
+        batch_->NewIteratorWithBase(&cf1, new KVIter(&map)));
+    std::unique_ptr<Iterator> result_iter(new KVIter(&merged_map));
+
+    bool is_valid = false;
+    for (int i = 0; i < 128; i++) {
+      // Random walk and make sure iter and result_iter returns the
+      // same key and value
+      int type = rnd.Uniform(6);
+      ASSERT_OK(iter->status());
+      switch (type) {
+        case 0:
+          // Seek to First
+          iter->SeekToFirst();
+          result_iter->SeekToFirst();
+          break;
+        case 1:
+          // Seek to last
+          iter->SeekToLast();
+          result_iter->SeekToLast();
+          break;
+        case 2: {
+          // Seek to random key
+          auto key_idx = rnd.Uniform(static_cast<int>(source_strings.size()));
+          auto key = source_strings[key_idx];
+          iter->Seek(key);
+          result_iter->Seek(key);
+          break;
+        }
+        case 3: {
+          // SeekForPrev to random key
+          auto key_idx = rnd.Uniform(static_cast<int>(source_strings.size()));
+          auto key = source_strings[key_idx];
+          iter->SeekForPrev(key);
+          result_iter->SeekForPrev(key);
+          break;
+        }
+        case 4:
+          // Next
+          if (is_valid) {
+            iter->Next();
+            result_iter->Next();
+          } else {
+            continue;
+          }
+          break;
+        default:
+          assert(type == 5);
+          // Prev
+          if (is_valid) {
+            iter->Prev();
+            result_iter->Prev();
+          } else {
+            continue;
+          }
+          break;
+      }
+      AssertItersMatch(iter.get(), result_iter.get());
+      is_valid = iter->Valid();
+    }
+
+    ASSERT_OK(iter->status());
+  }
+}
+
+TEST_P(WriteBatchWithIndexTest, TestIteraratorWithBase) {
+  ColumnFamilyHandleImplDummy cf1(6, BytewiseComparator());
+  ColumnFamilyHandleImplDummy cf2(2, BytewiseComparator());
+  {
+    KVMap map;
+    map["a"] = "aa";
+    map["c"] = "cc";
+    map["e"] = "ee";
+    std::unique_ptr<Iterator> iter(
+        batch_->NewIteratorWithBase(&cf1, new KVIter(&map)));
+
+    iter->SeekToFirst();
+    AssertIter(iter.get(), "a", "aa");
+    iter->Next();
+    AssertIter(iter.get(), "c", "cc");
+    iter->Next();
+    AssertIter(iter.get(), "e", "ee");
+    iter->Next();
+    ASSERT_OK(iter->status());
+    ASSERT_TRUE(!iter->Valid());
+
+    iter->SeekToLast();
+    AssertIter(iter.get(), "e", "ee");
+    iter->Prev();
+    AssertIter(iter.get(), "c", "cc");
+    iter->Prev();
+    AssertIter(iter.get(), "a", "aa");
+    iter->Prev();
+    ASSERT_OK(iter->status());
+    ASSERT_TRUE(!iter->Valid());
+
+    iter->Seek("b");
+    AssertIter(iter.get(), "c", "cc");
+
+    iter->Prev();
+    AssertIter(iter.get(), "a", "aa");
+
+    iter->Seek("a");
+    AssertIter(iter.get(), "a", "aa");
+  }
+
+  // Test the case that there is one element in the write batch
+  ASSERT_OK(batch_->Put(&cf2, "zoo", "bar"));
+  ASSERT_OK(batch_->Put(&cf1, "a", "aa"));
+  {
+    KVMap empty_map;
+    std::unique_ptr<Iterator> iter(
+        batch_->NewIteratorWithBase(&cf1, new KVIter(&empty_map)));
+
+    iter->SeekToFirst();
+    AssertIter(iter.get(), "a", "aa");
+    iter->Next();
+    ASSERT_OK(iter->status());
+    ASSERT_TRUE(!iter->Valid());
+  }
+
+  ASSERT_OK(batch_->Delete(&cf1, "b"));
+  ASSERT_OK(batch_->Put(&cf1, "c", "cc"));
+  ASSERT_OK(batch_->Put(&cf1, "d", "dd"));
+  ASSERT_OK(batch_->Delete(&cf1, "e"));
+
+  {
+    KVMap map;
+    map["b"] = "";
+    map["cc"] = "cccc";
+    map["f"] = "ff";
+    std::unique_ptr<Iterator> iter(
+        batch_->NewIteratorWithBase(&cf1, new KVIter(&map)));
+
+    iter->SeekToFirst();
+    AssertIter(iter.get(), "a", "aa");
+    iter->Next();
+    AssertIter(iter.get(), "c", "cc");
+    iter->Next();
+    AssertIter(iter.get(), "cc", "cccc");
+    iter->Next();
+    AssertIter(iter.get(), "d", "dd");
+    iter->Next();
+    AssertIter(iter.get(), "f", "ff");
+    iter->Next();
+    ASSERT_OK(iter->status());
+    ASSERT_TRUE(!iter->Valid());
+
+    iter->SeekToLast();
+    AssertIter(iter.get(), "f", "ff");
+    iter->Prev();
+    AssertIter(iter.get(), "d", "dd");
+    iter->Prev();
+    AssertIter(iter.get(), "cc", "cccc");
+    iter->Prev();
+    AssertIter(iter.get(), "c", "cc");
+    iter->Next();
+    AssertIter(iter.get(), "cc", "cccc");
+    iter->Prev();
+    AssertIter(iter.get(), "c", "cc");
+    iter->Prev();
+    AssertIter(iter.get(), "a", "aa");
+    iter->Prev();
+    ASSERT_OK(iter->status());
+    ASSERT_TRUE(!iter->Valid());
+
+    iter->Seek("c");
+    AssertIter(iter.get(), "c", "cc");
+
+    iter->Seek("cb");
+    AssertIter(iter.get(), "cc", "cccc");
+
+    iter->Seek("cc");
+    AssertIter(iter.get(), "cc", "cccc");
+    iter->Next();
+    AssertIter(iter.get(), "d", "dd");
+
+    iter->Seek("e");
+    AssertIter(iter.get(), "f", "ff");
+
+    iter->Prev();
+    AssertIter(iter.get(), "d", "dd");
+
+    iter->Next();
+    AssertIter(iter.get(), "f", "ff");
+  }
+
+  {
+    KVMap empty_map;
+    std::unique_ptr<Iterator> iter(
+        batch_->NewIteratorWithBase(&cf1, new KVIter(&empty_map)));
+
+    iter->SeekToFirst();
+    AssertIter(iter.get(), "a", "aa");
+    iter->Next();
+    AssertIter(iter.get(), "c", "cc");
+    iter->Next();
+    AssertIter(iter.get(), "d", "dd");
+    iter->Next();
+    ASSERT_OK(iter->status());
+    ASSERT_TRUE(!iter->Valid());
+
+    iter->SeekToLast();
+    AssertIter(iter.get(), "d", "dd");
+    iter->Prev();
+    AssertIter(iter.get(), "c", "cc");
+    iter->Prev();
+    AssertIter(iter.get(), "a", "aa");
+
+    iter->Prev();
+    ASSERT_OK(iter->status());
+    ASSERT_TRUE(!iter->Valid());
+
+    iter->Seek("aa");
+    AssertIter(iter.get(), "c", "cc");
+    iter->Next();
+    AssertIter(iter.get(), "d", "dd");
+
+    iter->Seek("ca");
+    AssertIter(iter.get(), "d", "dd");
+
+    iter->Prev();
+    AssertIter(iter.get(), "c", "cc");
+  }
+}
+
+TEST_P(WriteBatchWithIndexTest, TestIteraratorWithBaseReverseCmp) {
+  ColumnFamilyHandleImplDummy cf1(6, ReverseBytewiseComparator());
+  ColumnFamilyHandleImplDummy cf2(2, ReverseBytewiseComparator());
+
+  // Test the case that there is one element in the write batch
+  ASSERT_OK(batch_->Put(&cf2, "zoo", "bar"));
+  ASSERT_OK(batch_->Put(&cf1, "a", "aa"));
+  {
+    KVMap empty_map;
+    std::unique_ptr<Iterator> iter(
+        batch_->NewIteratorWithBase(&cf1, new KVIter(&empty_map)));
+
+    iter->SeekToFirst();
+    AssertIter(iter.get(), "a", "aa");
+    iter->Next();
+    ASSERT_OK(iter->status());
+    ASSERT_TRUE(!iter->Valid());
+  }
+
+  ASSERT_OK(batch_->Put(&cf1, "c", "cc"));
+  {
+    KVMap map;
+    std::unique_ptr<Iterator> iter(
+        batch_->NewIteratorWithBase(&cf1, new KVIter(&map)));
+
+    iter->SeekToFirst();
+    AssertIter(iter.get(), "c", "cc");
+    iter->Next();
+    AssertIter(iter.get(), "a", "aa");
+    iter->Next();
+    ASSERT_OK(iter->status());
+    ASSERT_TRUE(!iter->Valid());
+
+    iter->SeekToLast();
+    AssertIter(iter.get(), "a", "aa");
+    iter->Prev();
+    AssertIter(iter.get(), "c", "cc");
+    iter->Prev();
+    ASSERT_OK(iter->status());
+    ASSERT_TRUE(!iter->Valid());
+
+    iter->Seek("b");
+    AssertIter(iter.get(), "a", "aa");
+
+    iter->Prev();
+    AssertIter(iter.get(), "c", "cc");
+
+    iter->Seek("a");
+    AssertIter(iter.get(), "a", "aa");
+  }
+
+  // default column family
+  ASSERT_OK(batch_->Put("a", "b"));
+  {
+    KVMap map;
+    map["b"] = "";
+    std::unique_ptr<Iterator> iter(
+        batch_->NewIteratorWithBase(new KVIter(&map)));
+
+    iter->SeekToFirst();
+    AssertIter(iter.get(), "a", "b");
+    iter->Next();
+    AssertIter(iter.get(), "b", "");
+    iter->Next();
+    ASSERT_OK(iter->status());
+    ASSERT_TRUE(!iter->Valid());
+
+    iter->SeekToLast();
+    AssertIter(iter.get(), "b", "");
+    iter->Prev();
+    AssertIter(iter.get(), "a", "b");
+    iter->Prev();
+    ASSERT_OK(iter->status());
+    ASSERT_TRUE(!iter->Valid());
+
+    iter->Seek("b");
+    AssertIter(iter.get(), "b", "");
+
+    iter->Prev();
+    AssertIter(iter.get(), "a", "b");
+
+    iter->Seek("0");
+    AssertIter(iter.get(), "a", "b");
+  }
+}
+
+TEST_P(WriteBatchWithIndexTest, TestGetFromBatch) {
+  Options options;
+  Status s;
+  std::string value;
+
+  s = batch_->GetFromBatch(options_, "b", &value);
+  ASSERT_TRUE(s.IsNotFound());
+
+  ASSERT_OK(batch_->Put("a", "a"));
+  ASSERT_OK(batch_->Put("b", "b"));
+  ASSERT_OK(batch_->Put("c", "c"));
+  ASSERT_OK(batch_->Put("a", "z"));
+  ASSERT_OK(batch_->Delete("c"));
+  ASSERT_OK(batch_->Delete("d"));
+  ASSERT_OK(batch_->Delete("e"));
+  ASSERT_OK(batch_->Put("e", "e"));
+
+  s = batch_->GetFromBatch(options_, "b", &value);
+  ASSERT_OK(s);
+  ASSERT_EQ("b", value);
+
+  s = batch_->GetFromBatch(options_, "a", &value);
+  ASSERT_OK(s);
+  ASSERT_EQ("z", value);
+
+  s = batch_->GetFromBatch(options_, "c", &value);
+  ASSERT_TRUE(s.IsNotFound());
+
+  s = batch_->GetFromBatch(options_, "d", &value);
+  ASSERT_TRUE(s.IsNotFound());
+
+  s = batch_->GetFromBatch(options_, "x", &value);
+  ASSERT_TRUE(s.IsNotFound());
+
+  s = batch_->GetFromBatch(options_, "e", &value);
+  ASSERT_OK(s);
+  ASSERT_EQ("e", value);
+
+  ASSERT_OK(batch_->Merge("z", "z"));
+
+  s = batch_->GetFromBatch(options_, "z", &value);
+  ASSERT_NOK(s);  // No merge operator specified.
+
+  s = batch_->GetFromBatch(options_, "b", &value);
+  ASSERT_OK(s);
+  ASSERT_EQ("b", value);
+}
+
+TEST_P(WriteBatchWithIndexTest, TestGetFromBatchMerge) {
+  Status s = OpenDB();
+  ASSERT_OK(s);
+
+  ColumnFamilyHandle* column_family = db_->DefaultColumnFamily();
+  std::string value;
+
+  s = batch_->GetFromBatch(options_, "x", &value);
+  ASSERT_TRUE(s.IsNotFound());
+
+  ASSERT_OK(batch_->Put("x", "X"));
+  std::string expected = "X";
+
+  for (int i = 0; i < 5; i++) {
+    ASSERT_OK(batch_->Merge("x", std::to_string(i)));
+    expected = expected + "," + std::to_string(i);
+
+    if (i % 2 == 0) {
+      ASSERT_OK(batch_->Put("y", std::to_string(i / 2)));
+    }
+
+    ASSERT_OK(batch_->Merge("z", "z"));
+
+    s = batch_->GetFromBatch(column_family, options_, "x", &value);
+    ASSERT_OK(s);
+    ASSERT_EQ(expected, value);
+
+    s = batch_->GetFromBatch(column_family, options_, "y", &value);
+    ASSERT_OK(s);
+    ASSERT_EQ(std::to_string(i / 2), value);
+
+    s = batch_->GetFromBatch(column_family, options_, "z", &value);
+    ASSERT_TRUE(s.IsMergeInProgress());
+  }
+}
+
+TEST_F(WBWIOverwriteTest, TestGetFromBatchMerge2) {
+  Status s = OpenDB();
+  ASSERT_OK(s);
+
+  ColumnFamilyHandle* column_family = db_->DefaultColumnFamily();
+  std::string value;
+
+  s = batch_->GetFromBatch(column_family, options_, "X", &value);
+  ASSERT_TRUE(s.IsNotFound());
+
+  ASSERT_OK(batch_->Put(column_family, "X", "x"));
+  ASSERT_OK(batch_->GetFromBatch(column_family, options_, "X", &value));
+  ASSERT_EQ("x", value);
+
+  ASSERT_OK(batch_->Put(column_family, "X", "x2"));
+  ASSERT_OK(batch_->GetFromBatch(column_family, options_, "X", &value));
+  ASSERT_EQ("x2", value);
+
+  ASSERT_OK(batch_->Merge(column_family, "X", "aaa"));
+  ASSERT_OK(batch_->GetFromBatch(column_family, options_, "X", &value));
+  ASSERT_EQ("x2,aaa", value);
+
+  ASSERT_OK(batch_->Merge(column_family, "X", "bbb"));
+  ASSERT_OK(batch_->GetFromBatch(column_family, options_, "X", &value));
+  ASSERT_EQ("x2,aaa,bbb", value);
+
+  ASSERT_OK(batch_->Put(column_family, "X", "x3"));
+  ASSERT_OK(batch_->GetFromBatch(column_family, options_, "X", &value));
+  ASSERT_EQ("x3", value);
+
+  ASSERT_OK(batch_->Merge(column_family, "X", "ccc"));
+  ASSERT_OK(batch_->GetFromBatch(column_family, options_, "X", &value));
+  ASSERT_EQ("x3,ccc", value);
+
+  ASSERT_OK(batch_->Delete(column_family, "X"));
+  s = batch_->GetFromBatch(column_family, options_, "X", &value);
+  ASSERT_TRUE(s.IsNotFound());
+
+  batch_->Merge(column_family, "X", "ddd");
+  ASSERT_OK(batch_->GetFromBatch(column_family, options_, "X", &value));
+  ASSERT_EQ("ddd", value);
+}
+
+TEST_P(WriteBatchWithIndexTest, TestGetFromBatchAndDB) {
+  ASSERT_OK(OpenDB());
+
+  std::string value;
+
+  ASSERT_OK(db_->Put(write_opts_, "a", "a"));
+  ASSERT_OK(db_->Put(write_opts_, "b", "b"));
+  ASSERT_OK(db_->Put(write_opts_, "c", "c"));
+
+  ASSERT_OK(batch_->Put("a", "batch_->a"));
+  ASSERT_OK(batch_->Delete("b"));
+
+  ASSERT_OK(batch_->GetFromBatchAndDB(db_, read_opts_, "a", &value));
+  ASSERT_EQ("batch_->a", value);
+
+  Status s = batch_->GetFromBatchAndDB(db_, read_opts_, "b", &value);
+  ASSERT_TRUE(s.IsNotFound());
+
+  ASSERT_OK(batch_->GetFromBatchAndDB(db_, read_opts_, "c", &value));
+  ASSERT_EQ("c", value);
+
+  s = batch_->GetFromBatchAndDB(db_, read_opts_, "x", &value);
+  ASSERT_TRUE(s.IsNotFound());
+
+  ASSERT_OK(db_->Delete(write_opts_, "x"));
+
+  s = batch_->GetFromBatchAndDB(db_, read_opts_, "x", &value);
+  ASSERT_TRUE(s.IsNotFound());
+}
+
+TEST_P(WriteBatchWithIndexTest, TestGetFromBatchAndDBMerge) {
+  Status s = OpenDB();
+  ASSERT_OK(s);
+
+  std::string value;
+
+  ASSERT_OK(db_->Put(write_opts_, "a", "a0"));
+  ASSERT_OK(db_->Put(write_opts_, "b", "b0"));
+  ASSERT_OK(db_->Merge(write_opts_, "b", "b1"));
+  ASSERT_OK(db_->Merge(write_opts_, "c", "c0"));
+  ASSERT_OK(db_->Merge(write_opts_, "d", "d0"));
+
+  ASSERT_OK(batch_->Merge("a", "a1"));
+  ASSERT_OK(batch_->Merge("a", "a2"));
+  ASSERT_OK(batch_->Merge("b", "b2"));
+  ASSERT_OK(batch_->Merge("d", "d1"));
+  ASSERT_OK(batch_->Merge("e", "e0"));
+
+  ASSERT_OK(batch_->GetFromBatchAndDB(db_, read_opts_, "a", &value));
+  ASSERT_EQ("a0,a1,a2", value);
+
+  ASSERT_OK(batch_->GetFromBatchAndDB(db_, read_opts_, "b", &value));
+  ASSERT_EQ("b0,b1,b2", value);
+
+  ASSERT_OK(batch_->GetFromBatchAndDB(db_, read_opts_, "c", &value));
+  ASSERT_EQ("c0", value);
+
+  ASSERT_OK(batch_->GetFromBatchAndDB(db_, read_opts_, "d", &value));
+  ASSERT_EQ("d0,d1", value);
+
+  ASSERT_OK(batch_->GetFromBatchAndDB(db_, read_opts_, "e", &value));
+  ASSERT_EQ("e0", value);
+
+  ASSERT_OK(db_->Delete(write_opts_, "x"));
+
+  s = batch_->GetFromBatchAndDB(db_, read_opts_, "x", &value);
+  ASSERT_TRUE(s.IsNotFound());
+
+  const Snapshot* snapshot = db_->GetSnapshot();
+  ReadOptions snapshot_read_options;
+  snapshot_read_options.snapshot = snapshot;
+
+  ASSERT_OK(db_->Delete(write_opts_, "a"));
+
+  ASSERT_OK(batch_->GetFromBatchAndDB(db_, read_opts_, "a", &value));
+  ASSERT_EQ("a1,a2", value);
+
+  ASSERT_OK(
+      s = batch_->GetFromBatchAndDB(db_, snapshot_read_options, "a", &value));
+  ASSERT_EQ("a0,a1,a2", value);
+
+  ASSERT_OK(batch_->Delete("a"));
+
+  s = batch_->GetFromBatchAndDB(db_, read_opts_, "a", &value);
+  ASSERT_TRUE(s.IsNotFound());
+
+  s = batch_->GetFromBatchAndDB(db_, snapshot_read_options, "a", &value);
+  ASSERT_TRUE(s.IsNotFound());
+
+  ASSERT_OK(s = db_->Merge(write_opts_, "c", "c1"));
+
+  ASSERT_OK(s = batch_->GetFromBatchAndDB(db_, read_opts_, "c", &value));
+  ASSERT_EQ("c0,c1", value);
+
+  ASSERT_OK(
+      s = batch_->GetFromBatchAndDB(db_, snapshot_read_options, "c", &value));
+  ASSERT_EQ("c0", value);
+
+  ASSERT_OK(db_->Put(write_opts_, "e", "e1"));
+  ASSERT_OK(batch_->GetFromBatchAndDB(db_, read_opts_, "e", &value));
+  ASSERT_EQ("e1,e0", value);
+
+  ASSERT_OK(batch_->GetFromBatchAndDB(db_, snapshot_read_options, "e", &value));
+  ASSERT_EQ("e0", value);
+
+  ASSERT_OK(s = db_->Delete(write_opts_, "e"));
+  ASSERT_OK(batch_->GetFromBatchAndDB(db_, read_opts_, "e", &value));
+  ASSERT_EQ("e0", value);
+
+  ASSERT_OK(batch_->GetFromBatchAndDB(db_, snapshot_read_options, "e", &value));
+  ASSERT_EQ("e0", value);
+
+  db_->ReleaseSnapshot(snapshot);
+}
+
+TEST_F(WBWIOverwriteTest, TestGetFromBatchAndDBMerge2) {
+  Status s = OpenDB();
+  ASSERT_OK(s);
+
+  std::string value;
+
+  s = batch_->GetFromBatchAndDB(db_, read_opts_, "A", &value);
+  ASSERT_TRUE(s.IsNotFound());
+
+  ASSERT_OK(batch_->Merge("A", "xxx"));
+  ASSERT_OK(batch_->GetFromBatchAndDB(db_, read_opts_, "A", &value));
+  ASSERT_EQ(value, "xxx");
+
+  ASSERT_OK(batch_->Merge("A", "yyy"));
+  ASSERT_OK(batch_->GetFromBatchAndDB(db_, read_opts_, "A", &value));
+  ASSERT_EQ(value, "xxx,yyy");
+
+  ASSERT_OK(db_->Put(write_opts_, "A", "a0"));
+  ASSERT_OK(batch_->GetFromBatchAndDB(db_, read_opts_, "A", &value));
+  ASSERT_EQ(value, "a0,xxx,yyy");
+
+  ASSERT_OK(batch_->Delete("A"));
+
+  s = batch_->GetFromBatchAndDB(db_, read_opts_, "A", &value);
+  ASSERT_TRUE(s.IsNotFound());
+}
+
+TEST_P(WriteBatchWithIndexTest, TestGetFromBatchAndDBMerge3) {
+  Status s = OpenDB();
+  ASSERT_OK(s);
+
+  FlushOptions flush_options;
+  std::string value;
+
+  ASSERT_OK(db_->Put(write_opts_, "A", "1"));
+  ASSERT_OK(db_->Flush(flush_options, db_->DefaultColumnFamily()));
+  ASSERT_OK(batch_->Merge("A", "2"));
+
+  ASSERT_OK(batch_->GetFromBatchAndDB(db_, read_opts_, "A", &value));
+  ASSERT_EQ(value, "1,2");
+}
+
+TEST_P(WriteBatchWithIndexTest, TestPinnedGetFromBatchAndDB) {
+  Status s = OpenDB();
+  ASSERT_OK(s);
+
+  PinnableSlice value;
+
+  ASSERT_OK(db_->Put(write_opts_, "a", "a0"));
+  ASSERT_OK(db_->Put(write_opts_, "b", "b0"));
+  ASSERT_OK(db_->Merge(write_opts_, "b", "b1"));
+  ASSERT_OK(db_->Merge(write_opts_, "c", "c0"));
+  ASSERT_OK(db_->Merge(write_opts_, "d", "d0"));
+  ASSERT_OK(batch_->Merge("a", "a1"));
+  ASSERT_OK(batch_->Merge("a", "a2"));
+  ASSERT_OK(batch_->Merge("b", "b2"));
+  ASSERT_OK(batch_->Merge("d", "d1"));
+  ASSERT_OK(batch_->Merge("e", "e0"));
+
+  for (int i = 0; i < 2; i++) {
+    if (i == 1) {
+      // Do it again with a flushed DB...
+      ASSERT_OK(db_->Flush(FlushOptions(), db_->DefaultColumnFamily()));
+    }
+    ASSERT_OK(batch_->GetFromBatchAndDB(db_, read_opts_, "a", &value));
+    ASSERT_EQ("a0,a1,a2", value.ToString());
+
+    ASSERT_OK(batch_->GetFromBatchAndDB(db_, read_opts_, "b", &value));
+    ASSERT_EQ("b0,b1,b2", value.ToString());
+
+    ASSERT_OK(batch_->GetFromBatchAndDB(db_, read_opts_, "c", &value));
+    ASSERT_EQ("c0", value.ToString());
+
+    ASSERT_OK(batch_->GetFromBatchAndDB(db_, read_opts_, "d", &value));
+    ASSERT_EQ("d0,d1", value.ToString());
+
+    ASSERT_OK(batch_->GetFromBatchAndDB(db_, read_opts_, "e", &value));
+    ASSERT_EQ("e0", value.ToString());
+    ASSERT_OK(db_->Delete(write_opts_, "x"));
+
+    s = batch_->GetFromBatchAndDB(db_, read_opts_, "x", &value);
+    ASSERT_TRUE(s.IsNotFound());
+  }
+}
+
+void AssertKey(std::string key, WBWIIterator* iter) {
+  ASSERT_TRUE(iter->Valid());
+  ASSERT_EQ(key, iter->Entry().key.ToString());
+}
+
+void AssertValue(std::string value, WBWIIterator* iter) {
+  ASSERT_TRUE(iter->Valid());
+  ASSERT_EQ(value, iter->Entry().value.ToString());
+}
+
+// Tests that we can write to the WBWI while we iterate (from a single thread).
+// iteration should see the newest writes
+TEST_F(WBWIOverwriteTest, MutateWhileIteratingCorrectnessTest) {
+  for (char c = 'a'; c <= 'z'; ++c) {
+    ASSERT_OK(batch_->Put(std::string(1, c), std::string(1, c)));
+  }
+
+  std::unique_ptr<WBWIIterator> iter(batch_->NewIterator());
+  iter->Seek("k");
+  AssertKey("k", iter.get());
+  iter->Next();
+  AssertKey("l", iter.get());
+  ASSERT_OK(batch_->Put("ab", "cc"));
+  iter->Next();
+  AssertKey("m", iter.get());
+  ASSERT_OK(batch_->Put("mm", "kk"));
+  iter->Next();
+  AssertKey("mm", iter.get());
+  AssertValue("kk", iter.get());
+  ASSERT_OK(batch_->Delete("mm"));
+
+  iter->Next();
+  AssertKey("n", iter.get());
+  iter->Prev();
+  AssertKey("mm", iter.get());
+  ASSERT_EQ(kDeleteRecord, iter->Entry().type);
+
+  iter->Seek("ab");
+  AssertKey("ab", iter.get());
+  ASSERT_OK(batch_->Delete("x"));
+  iter->Seek("x");
+  AssertKey("x", iter.get());
+  ASSERT_EQ(kDeleteRecord, iter->Entry().type);
+  iter->Prev();
+  AssertKey("w", iter.get());
+}
+
+void AssertIterKey(std::string key, Iterator* iter) {
+  ASSERT_TRUE(iter->Valid());
+  ASSERT_EQ(key, iter->key().ToString());
+}
+
+void AssertIterValue(std::string value, Iterator* iter) {
+  ASSERT_TRUE(iter->Valid());
+  ASSERT_EQ(value, iter->value().ToString());
+}
+
+// same thing as above, but testing IteratorWithBase
+TEST_F(WBWIOverwriteTest, MutateWhileIteratingBaseCorrectnessTest) {
+  WriteBatchWithIndex batch(BytewiseComparator(), 0, true);
+  for (char c = 'a'; c <= 'z'; ++c) {
+    ASSERT_OK(batch_->Put(std::string(1, c), std::string(1, c)));
+  }
+
+  KVMap map;
+  map["aa"] = "aa";
+  map["cc"] = "cc";
+  map["ee"] = "ee";
+  map["em"] = "me";
+
+  std::unique_ptr<Iterator> iter(batch_->NewIteratorWithBase(new KVIter(&map)));
+  iter->Seek("k");
+  AssertIterKey("k", iter.get());
+  iter->Next();
+  AssertIterKey("l", iter.get());
+  ASSERT_OK(batch_->Put("ab", "cc"));
+  iter->Next();
+  AssertIterKey("m", iter.get());
+  ASSERT_OK(batch_->Put("mm", "kk"));
+  iter->Next();
+  AssertIterKey("mm", iter.get());
+  AssertIterValue("kk", iter.get());
+  ASSERT_OK(batch_->Delete("mm"));
+  iter->Next();
+  AssertIterKey("n", iter.get());
+  iter->Prev();
+  // "mm" is deleted, so we're back at "m"
+  AssertIterKey("m", iter.get());
+
+  iter->Seek("ab");
+  AssertIterKey("ab", iter.get());
+  iter->Prev();
+  AssertIterKey("aa", iter.get());
+  iter->Prev();
+  AssertIterKey("a", iter.get());
+  ASSERT_OK(batch_->Delete("aa"));
+  iter->Next();
+  AssertIterKey("ab", iter.get());
+  iter->Prev();
+  AssertIterKey("a", iter.get());
+
+  ASSERT_OK(batch_->Delete("x"));
+  iter->Seek("x");
+  AssertIterKey("y", iter.get());
+  iter->Next();
+  AssertIterKey("z", iter.get());
+  iter->Prev();
+  iter->Prev();
+  AssertIterKey("w", iter.get());
+
+  ASSERT_OK(batch_->Delete("e"));
+  iter->Seek("e");
+  AssertIterKey("ee", iter.get());
+  AssertIterValue("ee", iter.get());
+  ASSERT_OK(batch_->Put("ee", "xx"));
+  // still the same value
+  AssertIterValue("ee", iter.get());
+  iter->Next();
+  AssertIterKey("em", iter.get());
+  iter->Prev();
+  // new value
+  AssertIterValue("xx", iter.get());
+
+  ASSERT_OK(iter->status());
+}
+
+// stress testing mutations with IteratorWithBase
+TEST_F(WBWIOverwriteTest, MutateWhileIteratingBaseStressTest) {
+  for (char c = 'a'; c <= 'z'; ++c) {
+    ASSERT_OK(batch_->Put(std::string(1, c), std::string(1, c)));
+  }
+
+  KVMap map;
+  for (char c = 'a'; c <= 'z'; ++c) {
+    map[std::string(2, c)] = std::string(2, c);
+  }
+
+  std::unique_ptr<Iterator> iter(batch_->NewIteratorWithBase(new KVIter(&map)));
+
+  Random rnd(301);
+  for (int i = 0; i < 1000000; ++i) {
+    int random = rnd.Uniform(8);
+    char c = static_cast<char>(rnd.Uniform(26) + 'a');
+    switch (random) {
+      case 0:
+        ASSERT_OK(batch_->Put(std::string(1, c), "xxx"));
+        break;
+      case 1:
+        ASSERT_OK(batch_->Put(std::string(2, c), "xxx"));
+        break;
+      case 2:
+        ASSERT_OK(batch_->Delete(std::string(1, c)));
+        break;
+      case 3:
+        ASSERT_OK(batch_->Delete(std::string(2, c)));
+        break;
+      case 4:
+        iter->Seek(std::string(1, c));
+        break;
+      case 5:
+        iter->Seek(std::string(2, c));
+        break;
+      case 6:
+        if (iter->Valid()) {
+          iter->Next();
+        }
+        break;
+      case 7:
+        if (iter->Valid()) {
+          iter->Prev();
+        }
+        break;
+      default:
+        assert(false);
+    }
+  }
+  ASSERT_OK(iter->status());
+}
+
+TEST_P(WriteBatchWithIndexTest, TestNewIteratorWithBaseFromWbwi) {
+  ColumnFamilyHandleImplDummy cf1(6, BytewiseComparator());
+  KVMap map;
+  map["a"] = "aa";
+  map["c"] = "cc";
+  map["e"] = "ee";
+  std::unique_ptr<Iterator> iter(
+      batch_->NewIteratorWithBase(&cf1, new KVIter(&map)));
+  ASSERT_NE(nullptr, iter);
+  iter->SeekToFirst();
+  ASSERT_TRUE(iter->Valid());
+  ASSERT_OK(iter->status());
+}
+
+TEST_P(WriteBatchWithIndexTest, SavePointTest) {
+  ColumnFamilyHandleImplDummy cf1(1, BytewiseComparator());
+  KVMap empty_map;
+  std::unique_ptr<Iterator> cf0_iter(
+      batch_->NewIteratorWithBase(new KVIter(&empty_map)));
+  std::unique_ptr<Iterator> cf1_iter(
+      batch_->NewIteratorWithBase(&cf1, new KVIter(&empty_map)));
+  Status s;
+  KVMap kvm_cf0_0 = {{"A", "aa"}, {"B", "b"}};
+  KVMap kvm_cf1_0 = {{"A", "a1"}, {"C", "c1"}, {"E", "e1"}};
+  KVIter kvi_cf0_0(&kvm_cf0_0);
+  KVIter kvi_cf1_0(&kvm_cf1_0);
+
+  ASSERT_OK(batch_->Put("A", "a"));
+  ASSERT_OK(batch_->Put("B", "b"));
+  ASSERT_OK(batch_->Put("A", "aa"));
+  ASSERT_OK(batch_->Put(&cf1, "A", "a1"));
+  ASSERT_OK(batch_->Delete(&cf1, "B"));
+  ASSERT_OK(batch_->Put(&cf1, "C", "c1"));
+  ASSERT_OK(batch_->Put(&cf1, "E", "e1"));
+
+  AssertItersEqual(cf0_iter.get(), &kvi_cf0_0);
+  AssertItersEqual(cf1_iter.get(), &kvi_cf1_0);
+  batch_->SetSavePoint();  // 1
+
+  KVMap kvm_cf0_1 = {{"B", "bb"}, {"C", "cc"}};
+  KVMap kvm_cf1_1 = {{"B", "b1"}, {"C", "c1"}};
+  KVIter kvi_cf0_1(&kvm_cf0_1);
+  KVIter kvi_cf1_1(&kvm_cf1_1);
+
+  ASSERT_OK(batch_->Put("C", "cc"));
+  ASSERT_OK(batch_->Put("B", "bb"));
+  ASSERT_OK(batch_->Delete("A"));
+  ASSERT_OK(batch_->Put(&cf1, "B", "b1"));
+  ASSERT_OK(batch_->Delete(&cf1, "A"));
+  ASSERT_OK(batch_->SingleDelete(&cf1, "E"));
+  batch_->SetSavePoint();  // 2
+  AssertItersEqual(cf0_iter.get(), &kvi_cf0_1);
+  AssertItersEqual(cf1_iter.get(), &kvi_cf1_1);
+
+  KVMap kvm_cf0_2 = {{"A", "xxx"}, {"C", "cc"}};
+  KVMap kvm_cf1_2 = {{"B", "b2"}};
+  KVIter kvi_cf0_2(&kvm_cf0_2);
+  KVIter kvi_cf1_2(&kvm_cf1_2);
+
+  ASSERT_OK(batch_->Put("A", "aaa"));
+  ASSERT_OK(batch_->Put("A", "xxx"));
+  ASSERT_OK(batch_->Delete("B"));
+  ASSERT_OK(batch_->Put(&cf1, "B", "b2"));
+  ASSERT_OK(batch_->Delete(&cf1, "C"));
+  batch_->SetSavePoint();  // 3
+  batch_->SetSavePoint();  // 4
+  AssertItersEqual(cf0_iter.get(), &kvi_cf0_2);
+  AssertItersEqual(cf1_iter.get(), &kvi_cf1_2);
+
+  KVMap kvm_cf0_4 = {{"A", "xxx"}, {"C", "cc"}};
+  KVMap kvm_cf1_4 = {{"B", "b2"}};
+  KVIter kvi_cf0_4(&kvm_cf0_4);
+  KVIter kvi_cf1_4(&kvm_cf1_4);
+  ASSERT_OK(batch_->SingleDelete("D"));
+  ASSERT_OK(batch_->Delete(&cf1, "D"));
+  ASSERT_OK(batch_->Delete(&cf1, "E"));
+  AssertItersEqual(cf0_iter.get(), &kvi_cf0_4);
+  AssertItersEqual(cf1_iter.get(), &kvi_cf1_4);
+
+  ASSERT_OK(batch_->RollbackToSavePoint());  // rollback to 4
+  AssertItersEqual(cf0_iter.get(), &kvi_cf0_2);
+  AssertItersEqual(cf1_iter.get(), &kvi_cf1_2);
+
+  ASSERT_OK(batch_->RollbackToSavePoint());  // rollback to 3
+  AssertItersEqual(cf0_iter.get(), &kvi_cf0_2);
+  AssertItersEqual(cf1_iter.get(), &kvi_cf1_2);
+
+  ASSERT_OK(batch_->RollbackToSavePoint());  // rollback to 2
+  AssertItersEqual(cf0_iter.get(), &kvi_cf0_1);
+  AssertItersEqual(cf1_iter.get(), &kvi_cf1_1);
+
+  batch_->SetSavePoint();  // 5
+  ASSERT_OK(batch_->Put("X", "x"));
+
+  KVMap kvm_cf0_5 = {{"B", "bb"}, {"C", "cc"}, {"X", "x"}};
+  KVIter kvi_cf0_5(&kvm_cf0_5);
+  KVIter kvi_cf1_5(&kvm_cf1_1);
+  AssertItersEqual(cf0_iter.get(), &kvi_cf0_5);
+  AssertItersEqual(cf1_iter.get(), &kvi_cf1_5);
+
+  ASSERT_OK(batch_->RollbackToSavePoint());  // rollback to 5
+  AssertItersEqual(cf0_iter.get(), &kvi_cf0_1);
+  AssertItersEqual(cf1_iter.get(), &kvi_cf1_1);
+
+  ASSERT_OK(batch_->RollbackToSavePoint());  // rollback to 1
+  AssertItersEqual(cf0_iter.get(), &kvi_cf0_0);
+  AssertItersEqual(cf1_iter.get(), &kvi_cf1_0);
+
+  s = batch_->RollbackToSavePoint();  // no savepoint found
+  ASSERT_TRUE(s.IsNotFound());
+  AssertItersEqual(cf0_iter.get(), &kvi_cf0_0);
+  AssertItersEqual(cf1_iter.get(), &kvi_cf1_0);
+
+  batch_->SetSavePoint();  // 6
+
+  batch_->Clear();
+  ASSERT_EQ("", PrintContents(batch_.get(), nullptr));
+  ASSERT_EQ("", PrintContents(batch_.get(), &cf1));
+
+  s = batch_->RollbackToSavePoint();  // rollback to 6
+  ASSERT_TRUE(s.IsNotFound());
+}
+
+TEST_P(WriteBatchWithIndexTest, SingleDeleteTest) {
+  Status s;
+  std::string value;
+
+  ASSERT_OK(batch_->SingleDelete("A"));
+
+  s = batch_->GetFromBatch(options_, "A", &value);
+  ASSERT_TRUE(s.IsNotFound());
+  s = batch_->GetFromBatch(options_, "B", &value);
+  ASSERT_TRUE(s.IsNotFound());
+
+  batch_->Clear();
+  ASSERT_OK(batch_->Put("A", "a"));
+  ASSERT_OK(batch_->Put("A", "a2"));
+  ASSERT_OK(batch_->Put("B", "b"));
+  ASSERT_OK(batch_->SingleDelete("A"));
+
+  s = batch_->GetFromBatch(options_, "A", &value);
+  ASSERT_TRUE(s.IsNotFound());
+  s = batch_->GetFromBatch(options_, "B", &value);
+  ASSERT_OK(s);
+  ASSERT_EQ("b", value);
+
+  ASSERT_OK(batch_->Put("C", "c"));
+  ASSERT_OK(batch_->Put("A", "a3"));
+  ASSERT_OK(batch_->Delete("B"));
+  ASSERT_OK(batch_->SingleDelete("B"));
+  ASSERT_OK(batch_->SingleDelete("C"));
+
+  s = batch_->GetFromBatch(options_, "A", &value);
+  ASSERT_OK(s);
+  ASSERT_EQ("a3", value);
+  s = batch_->GetFromBatch(options_, "B", &value);
+  ASSERT_TRUE(s.IsNotFound());
+  s = batch_->GetFromBatch(options_, "C", &value);
+  ASSERT_TRUE(s.IsNotFound());
+  s = batch_->GetFromBatch(options_, "D", &value);
+  ASSERT_TRUE(s.IsNotFound());
+
+  ASSERT_OK(batch_->Put("B", "b4"));
+  ASSERT_OK(batch_->Put("C", "c4"));
+  ASSERT_OK(batch_->Put("D", "d4"));
+  ASSERT_OK(batch_->SingleDelete("D"));
+  ASSERT_OK(batch_->SingleDelete("D"));
+  ASSERT_OK(batch_->Delete("A"));
+
+  s = batch_->GetFromBatch(options_, "A", &value);
+  ASSERT_TRUE(s.IsNotFound());
+  s = batch_->GetFromBatch(options_, "B", &value);
+  ASSERT_OK(s);
+  ASSERT_EQ("b4", value);
+  s = batch_->GetFromBatch(options_, "C", &value);
+  ASSERT_OK(s);
+  ASSERT_EQ("c4", value);
+  s = batch_->GetFromBatch(options_, "D", &value);
+  ASSERT_TRUE(s.IsNotFound());
+}
+
+TEST_P(WriteBatchWithIndexTest, SingleDeleteDeltaIterTest) {
+  std::string value;
+  ASSERT_OK(batch_->Put("A", "a"));
+  ASSERT_OK(batch_->Put("A", "a2"));
+  ASSERT_OK(batch_->Put("B", "b"));
+  ASSERT_OK(batch_->SingleDelete("A"));
+  ASSERT_OK(batch_->Delete("B"));
+
+  KVMap map;
+  value = PrintContents(batch_.get(), &map, nullptr);
+  ASSERT_EQ("", value);
+
+  map["A"] = "aa";
+  map["C"] = "cc";
+  map["D"] = "dd";
+
+  ASSERT_OK(batch_->SingleDelete("B"));
+  ASSERT_OK(batch_->SingleDelete("C"));
+  ASSERT_OK(batch_->SingleDelete("Z"));
+
+  value = PrintContents(batch_.get(), &map, nullptr);
+  ASSERT_EQ("D:dd,", value);
+
+  ASSERT_OK(batch_->Put("A", "a3"));
+  ASSERT_OK(batch_->Put("B", "b3"));
+  ASSERT_OK(batch_->SingleDelete("A"));
+  ASSERT_OK(batch_->SingleDelete("A"));
+  ASSERT_OK(batch_->SingleDelete("D"));
+  ASSERT_OK(batch_->SingleDelete("D"));
+  ASSERT_OK(batch_->Delete("D"));
+
+  map["E"] = "ee";
+
+  value = PrintContents(batch_.get(), &map, nullptr);
+  ASSERT_EQ("B:b3,E:ee,", value);
+}
+
+TEST_P(WriteBatchWithIndexTest, MultiGetTest) {
+  // MultiGet a lot of keys in order to force std::vector reallocations
+  std::vector<std::string> keys;
+  for (int i = 0; i < 100; ++i) {
+    keys.emplace_back(std::to_string(i));
+  }
+
+  ASSERT_OK(OpenDB());
+  ColumnFamilyHandle* cf0 = db_->DefaultColumnFamily();
+
+  // Write some data to the db for the even numbered keys
+  {
+    WriteBatch wb;
+    for (size_t i = 0; i < keys.size(); i += 2) {
+      std::string val = "val" + std::to_string(i);
+      ASSERT_OK(wb.Put(cf0, keys[i], val));
+    }
+    ASSERT_OK(db_->Write(write_opts_, &wb));
+    for (size_t i = 0; i < keys.size(); i += 2) {
+      std::string value;
+      ASSERT_OK(db_->Get(read_opts_, cf0, keys[i], &value));
+    }
+  }
+
+  // Write some data to the batch
+  for (size_t i = 0; i < keys.size(); ++i) {
+    if ((i % 5) == 0) {
+      ASSERT_OK(batch_->Delete(cf0, keys[i]));
+    } else if ((i % 7) == 0) {
+      std::string val = "new" + std::to_string(i);
+      ASSERT_OK(batch_->Put(cf0, keys[i], val));
+    }
+    if (i > 0 && (i % 3) == 0) {
+      ASSERT_OK(batch_->Merge(cf0, keys[i], "merge"));
+    }
+  }
+
+  std::vector<Slice> key_slices;
+  for (size_t i = 0; i < keys.size(); ++i) {
+    key_slices.emplace_back(keys[i]);
+  }
+  std::vector<PinnableSlice> values(keys.size());
+  std::vector<Status> statuses(keys.size());
+
+  batch_->MultiGetFromBatchAndDB(db_, read_opts_, cf0, key_slices.size(),
+                                 key_slices.data(), values.data(),
+                                 statuses.data(), false);
+  for (size_t i = 0; i < keys.size(); ++i) {
+    if (i == 0) {
+      ASSERT_TRUE(statuses[i].IsNotFound());
+    } else if ((i % 3) == 0) {
+      ASSERT_OK(statuses[i]);
+      if ((i % 5) == 0) {  // Merge after Delete
+        ASSERT_EQ(values[i], "merge");
+      } else if ((i % 7) == 0) {  // Merge after Put
+        std::string val = "new" + std::to_string(i);
+        ASSERT_EQ(values[i], val + ",merge");
+      } else if ((i % 2) == 0) {
+        std::string val = "val" + std::to_string(i);
+        ASSERT_EQ(values[i], val + ",merge");
+      } else {
+        ASSERT_EQ(values[i], "merge");
+      }
+    } else if ((i % 5) == 0) {
+      ASSERT_TRUE(statuses[i].IsNotFound());
+    } else if ((i % 7) == 0) {
+      ASSERT_OK(statuses[i]);
+      ASSERT_EQ(values[i], "new" + std::to_string(i));
+    } else if ((i % 2) == 0) {
+      ASSERT_OK(statuses[i]);
+      ASSERT_EQ(values[i], "val" + std::to_string(i));
+    } else {
+      ASSERT_TRUE(statuses[i].IsNotFound());
+    }
+  }
+}
+TEST_P(WriteBatchWithIndexTest, MultiGetTest2) {
+  // MultiGet a lot of keys in order to force std::vector reallocations
+  const int num_keys = 700;
+  const int keys_per_pass = 100;
+  std::vector<std::string> keys;
+  for (size_t i = 0; i < num_keys; ++i) {
+    keys.emplace_back(std::to_string(i));
+  }
+  ASSERT_OK(OpenDB());
+  ColumnFamilyHandle* cf0 = db_->DefaultColumnFamily();
+
+  // Keys   0- 99 have a PUT in the batch but not DB
+  // Keys 100-199 have a PUT in the DB
+  // Keys 200-299 Have a PUT/DELETE
+  // Keys 300-399 Have a PUT/DELETE/MERGE
+  // Keys 400-499 have a PUT/MERGE
+  // Keys 500-599 have a MERGE only
+  // Keys 600-699 were never written
+  {
+    WriteBatch wb;
+    for (size_t i = 100; i < 500; i++) {
+      std::string val = std::to_string(i);
+      ASSERT_OK(wb.Put(cf0, keys[i], val));
+    }
+    ASSERT_OK(db_->Write(write_opts_, &wb));
+  }
+  ASSERT_OK(db_->Flush(FlushOptions(), cf0));
+  for (size_t i = 0; i < 100; i++) {
+    ASSERT_OK(batch_->Put(cf0, keys[i], keys[i]));
+  }
+  for (size_t i = 200; i < 400; i++) {
+    ASSERT_OK(batch_->Delete(cf0, keys[i]));
+  }
+  for (size_t i = 300; i < 600; i++) {
+    std::string val = std::to_string(i) + "m";
+    ASSERT_OK(batch_->Merge(cf0, keys[i], val));
+  }
+
+  Random rnd(301);
+  std::vector<PinnableSlice> values(keys_per_pass);
+  std::vector<Status> statuses(keys_per_pass);
+  for (int pass = 0; pass < 40; pass++) {
+    std::vector<Slice> key_slices;
+    for (size_t i = 0; i < keys_per_pass; i++) {
+      int random = rnd.Uniform(num_keys);
+      key_slices.emplace_back(keys[random]);
+    }
+    batch_->MultiGetFromBatchAndDB(db_, read_opts_, cf0, keys_per_pass,
+                                   key_slices.data(), values.data(),
+                                   statuses.data(), false);
+    for (size_t i = 0; i < keys_per_pass; i++) {
+      int key = ParseInt(key_slices[i].ToString());
+      switch (key / 100) {
+        case 0:  // 0-99 PUT only
+          ASSERT_OK(statuses[i]);
+          ASSERT_EQ(values[i], key_slices[i].ToString());
+          break;
+        case 1:  // 100-199 PUT only
+          ASSERT_OK(statuses[i]);
+          ASSERT_EQ(values[i], key_slices[i].ToString());
+          break;
+        case 2:  // 200-299 Deleted
+          ASSERT_TRUE(statuses[i].IsNotFound());
+          break;
+        case 3:  // 300-399 Delete+Merge
+          ASSERT_OK(statuses[i]);
+          ASSERT_EQ(values[i], key_slices[i].ToString() + "m");
+          break;
+        case 4:  // 400-400 Put+ Merge
+          ASSERT_OK(statuses[i]);
+          ASSERT_EQ(values[i], key_slices[i].ToString() + "," +
+                                   key_slices[i].ToString() + "m");
+          break;
+        case 5:  // Merge only
+          ASSERT_OK(statuses[i]);
+          ASSERT_EQ(values[i], key_slices[i].ToString() + "m");
+          break;
+        case 6:  // Never written
+          ASSERT_TRUE(statuses[i].IsNotFound());
+          break;
+        default:
+          assert(false);
+      }  // end switch
+    }    // End for each key
+  }      // end for passes
+}
+
+// This test has merges, but the merge does not play into the final result
+TEST_P(WriteBatchWithIndexTest, FakeMergeWithIteratorTest) {
+  ASSERT_OK(OpenDB());
+  ColumnFamilyHandle* cf0 = db_->DefaultColumnFamily();
+
+  // The map we are starting with
+  KVMap input = {
+      {"odm", "odm0"},
+      {"omd", "omd0"},
+      {"omp", "omp0"},
+  };
+  KVMap result = {
+      {"odm", "odm2"},  // Orig, Delete, Merge
+      {"mp", "mp1"},    // Merge, Put
+      {"omp", "omp2"},  // Origi, Merge, Put
+      {"mmp", "mmp2"}   // Merge, Merge, Put
+  };
+
+  for (auto& iter : result) {
+    EXPECT_EQ(AddToBatch(cf0, iter.first), iter.second);
+  }
+  AddToBatch(cf0, "md");   // Merge, Delete
+  AddToBatch(cf0, "mmd");  // Merge, Merge, Delete
+  AddToBatch(cf0, "omd");  // Orig, Merge, Delete
+
+  KVIter kvi(&result);
+  // First try just the batch
+  std::unique_ptr<Iterator> iter(
+      batch_->NewIteratorWithBase(cf0, new KVIter(&input)));
+  AssertItersEqual(iter.get(), &kvi);
+}
+
+TEST_P(WriteBatchWithIndexTest, IteratorMergeTest) {
+  ASSERT_OK(OpenDB());
+  ColumnFamilyHandle* cf0 = db_->DefaultColumnFamily();
+
+  KVMap result = {
+      {"m", "m0"},                // Merge
+      {"mm", "mm0,mm1"},          // Merge, Merge
+      {"dm", "dm1"},              // Delete, Merge
+      {"dmm", "dmm1,dmm2"},       // Delete, Merge, Merge
+      {"mdm", "mdm2"},            // Merge, Delete, Merge
+      {"mpm", "mpm1,mpm2"},       // Merge, Put, Merge
+      {"pm", "pm0,pm1"},          // Put, Merge
+      {"pmm", "pmm0,pmm1,pmm2"},  // Put, Merge, Merge
+  };
+
+  for (auto& iter : result) {
+    EXPECT_EQ(AddToBatch(cf0, iter.first), iter.second);
+  }
+
+  KVIter kvi(&result);
+  // First try just the batch
+  KVMap empty_map;
+  std::unique_ptr<Iterator> iter(
+      batch_->NewIteratorWithBase(cf0, new KVIter(&empty_map)));
+  AssertItersEqual(iter.get(), &kvi);
+}
+
+TEST_P(WriteBatchWithIndexTest, IteratorMergeTestWithOrig) {
+  ASSERT_OK(OpenDB());
+  ColumnFamilyHandle* cf0 = db_->DefaultColumnFamily();
+  KVMap original;
+  KVMap results = {
+      {"m", "om,m0"},             // Merge
+      {"mm", "omm,mm0,mm1"},      // Merge, Merge
+      {"dm", "dm1"},              // Delete, Merge
+      {"dmm", "dmm1,dmm2"},       // Delete, Merge, Merge
+      {"mdm", "mdm2"},            // Merge, Delete, Merge
+      {"mpm", "mpm1,mpm2"},       // Merge, Put, Merge
+      {"pm", "pm0,pm1"},          // Put, Merge
+      {"pmm", "pmm0,pmm1,pmm2"},  // Put, Merge, Merge
+  };
+
+  for (auto& iter : results) {
+    AddToBatch(cf0, iter.first);
+    original[iter.first] = "o" + iter.first;
+  }
+
+  KVIter kvi(&results);
+  // First try just the batch
+  std::unique_ptr<Iterator> iter(
+      batch_->NewIteratorWithBase(cf0, new KVIter(&original)));
+  AssertItersEqual(iter.get(), &kvi);
+}
+
+TEST_P(WriteBatchWithIndexTest, GetFromBatchAfterMerge) {
+  std::string value;
+  Status s;
+
+  ASSERT_OK(OpenDB());
+  ASSERT_OK(db_->Put(write_opts_, "o", "aa"));
+  batch_->Merge("o", "bb");  // Merging bb under key "o"
+  batch_->Merge("m", "cc");  // Merging bc under key "m"
+  s = batch_->GetFromBatch(options_, "m", &value);
+  ASSERT_EQ(s.code(), Status::Code::kMergeInProgress);
+  s = batch_->GetFromBatch(options_, "o", &value);
+  ASSERT_EQ(s.code(), Status::Code::kMergeInProgress);
+
+  ASSERT_OK(db_->Write(write_opts_, batch_->GetWriteBatch()));
+  ASSERT_OK(db_->Get(read_opts_, "o", &value));
+  ASSERT_EQ(value, "aa,bb");
+  ASSERT_OK(db_->Get(read_opts_, "m", &value));
+  ASSERT_EQ(value, "cc");
+}
+
+TEST_P(WriteBatchWithIndexTest, GetFromBatchAndDBAfterMerge) {
+  std::string value;
+
+  ASSERT_OK(OpenDB());
+  ASSERT_OK(db_->Put(write_opts_, "o", "aa"));
+  ASSERT_OK(batch_->Merge("o", "bb"));  // Merging bb under key "o"
+  ASSERT_OK(batch_->Merge("m", "cc"));  // Merging bc under key "m"
+  ASSERT_OK(batch_->GetFromBatchAndDB(db_, read_opts_, "o", &value));
+  ASSERT_EQ(value, "aa,bb");
+  ASSERT_OK(batch_->GetFromBatchAndDB(db_, read_opts_, "m", &value));
+  ASSERT_EQ(value, "cc");
+}
+
+TEST_F(WBWIKeepTest, GetAfterPut) {
+  std::string value;
+  ASSERT_OK(OpenDB());
+  ColumnFamilyHandle* cf0 = db_->DefaultColumnFamily();
+
+  ASSERT_OK(db_->Put(write_opts_, "key", "orig"));
+
+  ASSERT_OK(batch_->Put("key", "aa"));  // Writing aa under key
+  ASSERT_OK(batch_->GetFromBatch(cf0, options_, "key", &value));
+  ASSERT_EQ(value, "aa");
+  ASSERT_OK(batch_->GetFromBatchAndDB(db_, read_opts_, "key", &value));
+  ASSERT_EQ(value, "aa");
+
+  ASSERT_OK(batch_->Merge("key", "bb"));  // Merging bb under key
+  ASSERT_OK(batch_->GetFromBatch(cf0, options_, "key", &value));
+  ASSERT_EQ(value, "aa,bb");
+  ASSERT_OK(batch_->GetFromBatchAndDB(db_, read_opts_, "key", &value));
+  ASSERT_EQ(value, "aa,bb");
+
+  ASSERT_OK(batch_->Merge("key", "cc"));  // Merging cc under key
+  ASSERT_OK(batch_->GetFromBatch(cf0, options_, "key", &value));
+  ASSERT_EQ(value, "aa,bb,cc");
+  ASSERT_OK(batch_->GetFromBatchAndDB(db_, read_opts_, "key", &value));
+  ASSERT_EQ(value, "aa,bb,cc");
+}
+
+TEST_P(WriteBatchWithIndexTest, GetAfterMergePut) {
+  std::string value;
+  ASSERT_OK(OpenDB());
+  ColumnFamilyHandle* cf0 = db_->DefaultColumnFamily();
+  ASSERT_OK(db_->Put(write_opts_, "key", "orig"));
+
+  ASSERT_OK(batch_->Merge("key", "aa"));  // Merging aa under key
+  Status s = batch_->GetFromBatch(cf0, options_, "key", &value);
+  ASSERT_EQ(s.code(), Status::Code::kMergeInProgress);
+  ASSERT_OK(batch_->GetFromBatchAndDB(db_, read_opts_, "key", &value));
+  ASSERT_EQ(value, "orig,aa");
+
+  ASSERT_OK(batch_->Merge("key", "bb"));  // Merging bb under key
+  s = batch_->GetFromBatch(cf0, options_, "key", &value);
+  ASSERT_EQ(s.code(), Status::Code::kMergeInProgress);
+  ASSERT_OK(batch_->GetFromBatchAndDB(db_, read_opts_, "key", &value));
+  ASSERT_EQ(value, "orig,aa,bb");
+
+  ASSERT_OK(batch_->Put("key", "cc"));  // Writing cc under key
+  ASSERT_OK(batch_->GetFromBatch(cf0, options_, "key", &value));
+  ASSERT_EQ(value, "cc");
+  ASSERT_OK(batch_->GetFromBatchAndDB(db_, read_opts_, "key", &value));
+  ASSERT_EQ(value, "cc");
+
+  ASSERT_OK(batch_->Merge("key", "dd"));  // Merging dd under key
+  ASSERT_OK(batch_->GetFromBatch(cf0, options_, "key", &value));
+  ASSERT_EQ(value, "cc,dd");
+  ASSERT_OK(batch_->GetFromBatchAndDB(db_, read_opts_, "key", &value));
+  ASSERT_EQ(value, "cc,dd");
+}
+
+TEST_P(WriteBatchWithIndexTest, GetAfterMergeDelete) {
+  std::string value;
+  ASSERT_OK(OpenDB());
+  ColumnFamilyHandle* cf0 = db_->DefaultColumnFamily();
+
+  ASSERT_OK(batch_->Merge("key", "aa"));  // Merging aa under key
+  Status s = batch_->GetFromBatch(cf0, options_, "key", &value);
+  ASSERT_EQ(s.code(), Status::Code::kMergeInProgress);
+  ASSERT_OK(batch_->GetFromBatchAndDB(db_, read_opts_, "key", &value));
+  ASSERT_EQ(value, "aa");
+
+  ASSERT_OK(batch_->Merge("key", "bb"));  // Merging bb under key
+  s = batch_->GetFromBatch(cf0, options_, "key", &value);
+  ASSERT_EQ(s.code(), Status::Code::kMergeInProgress);
+  ASSERT_OK(batch_->GetFromBatchAndDB(db_, read_opts_, "key", &value));
+  ASSERT_EQ(value, "aa,bb");
+
+  ASSERT_OK(batch_->Delete("key"));  // Delete key from batch
+  s = batch_->GetFromBatch(cf0, options_, "key", &value);
+  ASSERT_TRUE(s.IsNotFound());
+  s = batch_->GetFromBatchAndDB(db_, read_opts_, "key", &value);
+  ASSERT_TRUE(s.IsNotFound());
+
+  ASSERT_OK(batch_->Merge("key", "cc"));  // Merging cc under key
+  ASSERT_OK(batch_->GetFromBatch(cf0, options_, "key", &value));
+  ASSERT_EQ(value, "cc");
+  ASSERT_OK(batch_->GetFromBatchAndDB(db_, read_opts_, "key", &value));
+  ASSERT_EQ(value, "cc");
+  ASSERT_OK(batch_->Merge("key", "dd"));  // Merging dd under key
+  ASSERT_OK(batch_->GetFromBatch(cf0, options_, "key", &value));
+  ASSERT_EQ(value, "cc,dd");
+  ASSERT_OK(batch_->GetFromBatchAndDB(db_, read_opts_, "key", &value));
+  ASSERT_EQ(value, "cc,dd");
+}
+
+TEST_F(WBWIOverwriteTest, TestBadMergeOperator) {
+  class FailingMergeOperator : public MergeOperator {
+   public:
+    FailingMergeOperator() {}
+
+    bool FullMergeV2(const MergeOperationInput& /*merge_in*/,
+                     MergeOperationOutput* /*merge_out*/) const override {
+      return false;
+    }
+
+    const char* Name() const override { return "Failing"; }
+  };
+  options_.merge_operator.reset(new FailingMergeOperator());
+  ASSERT_OK(OpenDB());
+
+  ColumnFamilyHandle* column_family = db_->DefaultColumnFamily();
+  std::string value;
+
+  ASSERT_OK(db_->Put(write_opts_, "a", "a0"));
+  ASSERT_OK(batch_->Put("b", "b0"));
+
+  ASSERT_OK(batch_->Merge("a", "a1"));
+  ASSERT_NOK(batch_->GetFromBatchAndDB(db_, read_opts_, "a", &value));
+  ASSERT_NOK(batch_->GetFromBatch(column_family, options_, "a", &value));
+  ASSERT_OK(batch_->GetFromBatchAndDB(db_, read_opts_, "b", &value));
+  ASSERT_OK(batch_->GetFromBatch(column_family, options_, "b", &value));
+}
+
+TEST_P(WriteBatchWithIndexTest, ColumnFamilyWithTimestamp) {
+  ColumnFamilyHandleImplDummy cf2(2,
+                                  test::BytewiseComparatorWithU64TsWrapper());
+
+  // Sanity checks
+  ASSERT_TRUE(batch_->Put(&cf2, "key", "ts", "value").IsNotSupported());
+  ASSERT_TRUE(batch_->Put(/*column_family=*/nullptr, "key", "ts", "value")
+                  .IsInvalidArgument());
+  ASSERT_TRUE(batch_->Delete(&cf2, "key", "ts").IsNotSupported());
+  ASSERT_TRUE(batch_->Delete(/*column_family=*/nullptr, "key", "ts")
+                  .IsInvalidArgument());
+  ASSERT_TRUE(batch_->SingleDelete(&cf2, "key", "ts").IsNotSupported());
+  ASSERT_TRUE(batch_->SingleDelete(/*column_family=*/nullptr, "key", "ts")
+                  .IsInvalidArgument());
+  {
+    std::string value;
+    ASSERT_TRUE(batch_
+                    ->GetFromBatchAndDB(
+                        /*db=*/nullptr, ReadOptions(), &cf2, "key", &value)
+                    .IsInvalidArgument());
+  }
+  {
+    constexpr size_t num_keys = 2;
+    std::array<Slice, num_keys> keys{{Slice(), Slice()}};
+    std::array<PinnableSlice, num_keys> pinnable_vals{
+        {PinnableSlice(), PinnableSlice()}};
+    std::array<Status, num_keys> statuses{{Status(), Status()}};
+    constexpr bool sorted_input = false;
+    batch_->MultiGetFromBatchAndDB(/*db=*/nullptr, ReadOptions(), &cf2,
+                                   num_keys, keys.data(), pinnable_vals.data(),
+                                   statuses.data(), sorted_input);
+    for (const auto& s : statuses) {
+      ASSERT_TRUE(s.IsInvalidArgument());
+    }
+  }
+
+  constexpr uint32_t kMaxKey = 10;
+
+  const auto ts_sz_lookup = [&cf2](uint32_t id) {
+    if (cf2.GetID() == id) {
+      return sizeof(uint64_t);
+    } else {
+      return std::numeric_limits<size_t>::max();
+    }
+  };
+
+  // Put keys
+  for (uint32_t i = 0; i < kMaxKey; ++i) {
+    std::string key;
+    PutFixed32(&key, i);
+    Status s = batch_->Put(&cf2, key, "value" + std::to_string(i));
+    ASSERT_OK(s);
+  }
+
+  WriteBatch* wb = batch_->GetWriteBatch();
+  assert(wb);
+  ASSERT_OK(
+      wb->UpdateTimestamps(std::string(sizeof(uint64_t), '\0'), ts_sz_lookup));
+
+  // Point lookup
+  for (uint32_t i = 0; i < kMaxKey; ++i) {
+    std::string value;
+    std::string key;
+    PutFixed32(&key, i);
+    Status s = batch_->GetFromBatch(&cf2, Options(), key, &value);
+    ASSERT_OK(s);
+    ASSERT_EQ("value" + std::to_string(i), value);
+  }
+
+  // Iterator
+  {
+    std::unique_ptr<WBWIIterator> it(batch_->NewIterator(&cf2));
+    uint32_t start = 0;
+    for (it->SeekToFirst(); it->Valid(); it->Next(), ++start) {
+      std::string key;
+      PutFixed32(&key, start);
+      ASSERT_OK(it->status());
+      ASSERT_EQ(key, it->Entry().key);
+      ASSERT_EQ("value" + std::to_string(start), it->Entry().value);
+      ASSERT_EQ(WriteType::kPutRecord, it->Entry().type);
+    }
+    ASSERT_EQ(kMaxKey, start);
+  }
+
+  // Delete the keys with Delete() or SingleDelete()
+  for (uint32_t i = 0; i < kMaxKey; ++i) {
+    std::string key;
+    PutFixed32(&key, i);
+    Status s;
+    if (0 == (i % 2)) {
+      s = batch_->Delete(&cf2, key);
+    } else {
+      s = batch_->SingleDelete(&cf2, key);
+    }
+    ASSERT_OK(s);
+  }
+
+  ASSERT_OK(wb->UpdateTimestamps(std::string(sizeof(uint64_t), '\xfe'),
+                                 ts_sz_lookup));
+
+  for (uint32_t i = 0; i < kMaxKey; ++i) {
+    std::string value;
+    std::string key;
+    PutFixed32(&key, i);
+    Status s = batch_->GetFromBatch(&cf2, Options(), key, &value);
+    ASSERT_TRUE(s.IsNotFound());
+  }
+
+  // Iterator
+  {
+    const bool overwrite = GetParam();
+    std::unique_ptr<WBWIIterator> it(batch_->NewIterator(&cf2));
+    uint32_t start = 0;
+    for (it->SeekToFirst(); it->Valid(); it->Next(), ++start) {
+      std::string key;
+      PutFixed32(&key, start);
+      ASSERT_EQ(key, it->Entry().key);
+      if (!overwrite) {
+        ASSERT_EQ(WriteType::kPutRecord, it->Entry().type);
+        it->Next();
+        ASSERT_TRUE(it->Valid());
+      }
+      if (0 == (start % 2)) {
+        ASSERT_EQ(WriteType::kDeleteRecord, it->Entry().type);
+      } else {
+        ASSERT_EQ(WriteType::kSingleDeleteRecord, it->Entry().type);
+      }
+    }
+  }
+}
+
+TEST_P(WriteBatchWithIndexTest, IndexNoTs) {
+  const Comparator* const ucmp = test::BytewiseComparatorWithU64TsWrapper();
+  ColumnFamilyHandleImplDummy cf(1, ucmp);
+  WriteBatchWithIndex wbwi;
+  ASSERT_OK(wbwi.Put(&cf, "a", "a0"));
+  ASSERT_OK(wbwi.Put(&cf, "a", "a1"));
+  {
+    std::string ts;
+    PutFixed64(&ts, 10000);
+    ASSERT_OK(wbwi.GetWriteBatch()->UpdateTimestamps(
+        ts, [](uint32_t cf_id) { return cf_id == 1 ? 8 : 0; }));
+  }
+  {
+    std::string value;
+    Status s = wbwi.GetFromBatch(&cf, options_, "a", &value);
+    ASSERT_OK(s);
+    ASSERT_EQ("a1", value);
+  }
+}
+
+INSTANTIATE_TEST_CASE_P(WBWI, WriteBatchWithIndexTest, testing::Bool());
+}  // namespace ROCKSDB_NAMESPACE
+
+int main(int argc, char** argv) {
+  ROCKSDB_NAMESPACE::port::InstallStackTraceHandler();
+  ::testing::InitGoogleTest(&argc, argv);
+  return RUN_ALL_TESTS();
+}
+
+#else
+#include <stdio.h>
+
+int main() {
+  fprintf(stderr, "SKIPPED\n");
+  return 0;
+}
+
+#endif  // !ROCKSDB_LITE