summaryrefslogtreecommitdiffstats
path: root/libnetdata/gorilla/README.md
diff options
context:
space:
mode:
Diffstat (limited to 'libnetdata/gorilla/README.md')
-rw-r--r--libnetdata/gorilla/README.md39
1 files changed, 39 insertions, 0 deletions
diff --git a/libnetdata/gorilla/README.md b/libnetdata/gorilla/README.md
new file mode 100644
index 00000000..dc3718d1
--- /dev/null
+++ b/libnetdata/gorilla/README.md
@@ -0,0 +1,39 @@
+# Gorilla compression and decompression
+
+This provides an alternative way of representing values stored in database
+pages. Instead of allocating and using a page of fixed size, ie. 4096 bytes,
+the Gorilla implementation adds support for dynamically sized pages that
+contain a variable number of Gorilla buffers.
+
+Each buffer takes 512 bytes and compresses incoming data using the Gorilla
+compression:
+
+- The very first value is stored as it is.
+- For each new value, Gorilla compression doesn't store the value itself. Instead,
+it computes the difference (XOR) between the new value and the previous value.
+- If the XOR result is zero (meaning the new value is identical to the previous
+value), we store just a single bit set to `1`.
+- If the XOR result is not zero (meaning the new value differs from the previous):
+ - We store a `0` bit to indicate the change.
+ - We compute the leading-zero count (LZC) of the XOR result, and compare it
+ with the previous LZC. If the two LZCs are equal we store a `1` bit.
+ - If the LZCs are different we use 5 bits to store the new LZC, and we store
+ the rest of the value (ie. without its LZC) in the buffer.
+
+A Gorilla page can have multiple Gorilla buffers. If the values of a metric
+are highly compressible, just one Gorilla buffer is able to store all the values
+that otherwise would require a regular 4096 byte page, ie. we can use just 512
+bytes instead. In the worst case scenario (for metrics whose values are not
+compressible at all), a Gorilla page might end up having `9` Gorilla buffers,
+consuming 4608 bytes. In practice, this is pretty rare and does not negate
+the effect of compression for the metrics.
+
+When a gorilla page is full, ie. it contains 1024 slots/values, we serialize
+the linked-list of gorilla buffers directly to disk. During deserialization,
+eg. when performing a DBEngine query, the Gorilla page is loaded from the disk and
+its linked-list entries are patched to point to the new memory allocated for
+serving the query results.
+
+Overall, on a real-agent the Gorilla compression scheme reduces memory
+consumption approximately by ~30%, which can be several GiB of RAM for parents
+having hundreds, or even thousands of children streaming to them.