sqlglot.executor.python

  1import ast
  2import collections
  3import itertools
  4import math
  5
  6from sqlglot import exp, generator, planner, tokens
  7from sqlglot.dialects.dialect import Dialect, inline_array_sql
  8from sqlglot.errors import ExecuteError
  9from sqlglot.executor.context import Context
 10from sqlglot.executor.env import ENV
 11from sqlglot.executor.table import RowReader, Table
 12from sqlglot.helper import csv_reader, subclasses
 13
 14
 15class PythonExecutor:
 16    def __init__(self, env=None, tables=None):
 17        self.generator = Python().generator(identify=True, comments=False)
 18        self.env = {**ENV, **(env or {})}
 19        self.tables = tables or {}
 20
 21    def execute(self, plan):
 22        running = set()
 23        finished = set()
 24        queue = set(plan.leaves)
 25        contexts = {}
 26
 27        while queue:
 28            node = queue.pop()
 29            try:
 30                context = self.context(
 31                    {
 32                        name: table
 33                        for dep in node.dependencies
 34                        for name, table in contexts[dep].tables.items()
 35                    }
 36                )
 37                running.add(node)
 38
 39                if isinstance(node, planner.Scan):
 40                    contexts[node] = self.scan(node, context)
 41                elif isinstance(node, planner.Aggregate):
 42                    contexts[node] = self.aggregate(node, context)
 43                elif isinstance(node, planner.Join):
 44                    contexts[node] = self.join(node, context)
 45                elif isinstance(node, planner.Sort):
 46                    contexts[node] = self.sort(node, context)
 47                elif isinstance(node, planner.SetOperation):
 48                    contexts[node] = self.set_operation(node, context)
 49                else:
 50                    raise NotImplementedError
 51
 52                running.remove(node)
 53                finished.add(node)
 54
 55                for dep in node.dependents:
 56                    if dep not in running and all(d in contexts for d in dep.dependencies):
 57                        queue.add(dep)
 58
 59                for dep in node.dependencies:
 60                    if all(d in finished for d in dep.dependents):
 61                        contexts.pop(dep)
 62            except Exception as e:
 63                raise ExecuteError(f"Step '{node.id}' failed: {e}") from e
 64
 65        root = plan.root
 66        return contexts[root].tables[root.name]
 67
 68    def generate(self, expression):
 69        """Convert a SQL expression into literal Python code and compile it into bytecode."""
 70        if not expression:
 71            return None
 72
 73        sql = self.generator.generate(expression)
 74        return compile(sql, sql, "eval", optimize=2)
 75
 76    def generate_tuple(self, expressions):
 77        """Convert an array of SQL expressions into tuple of Python byte code."""
 78        if not expressions:
 79            return tuple()
 80        return tuple(self.generate(expression) for expression in expressions)
 81
 82    def context(self, tables):
 83        return Context(tables, env=self.env)
 84
 85    def table(self, expressions):
 86        return Table(
 87            expression.alias_or_name if isinstance(expression, exp.Expression) else expression
 88            for expression in expressions
 89        )
 90
 91    def scan(self, step, context):
 92        source = step.source
 93
 94        if source and isinstance(source, exp.Expression):
 95            source = source.name or source.alias
 96
 97        if source is None:
 98            context, table_iter = self.static()
 99        elif source in context:
100            if not step.projections and not step.condition:
101                return self.context({step.name: context.tables[source]})
102            table_iter = context.table_iter(source)
103        elif isinstance(step.source, exp.Table) and isinstance(step.source.this, exp.ReadCSV):
104            table_iter = self.scan_csv(step)
105            context = next(table_iter)
106        else:
107            context, table_iter = self.scan_table(step)
108
109        return self.context({step.name: self._project_and_filter(context, step, table_iter)})
110
111    def _project_and_filter(self, context, step, table_iter):
112        sink = self.table(step.projections if step.projections else context.columns)
113        condition = self.generate(step.condition)
114        projections = self.generate_tuple(step.projections)
115
116        for reader in table_iter:
117            if len(sink) >= step.limit:
118                break
119
120            if condition and not context.eval(condition):
121                continue
122
123            if projections:
124                sink.append(context.eval_tuple(projections))
125            else:
126                sink.append(reader.row)
127
128        return sink
129
130    def static(self):
131        return self.context({}), [RowReader(())]
132
133    def scan_table(self, step):
134        table = self.tables.find(step.source)
135        context = self.context({step.source.alias_or_name: table})
136        return context, iter(table)
137
138    def scan_csv(self, step):
139        alias = step.source.alias
140        source = step.source.this
141
142        with csv_reader(source) as reader:
143            columns = next(reader)
144            table = Table(columns)
145            context = self.context({alias: table})
146            yield context
147            types = []
148
149            for row in reader:
150                if not types:
151                    for v in row:
152                        try:
153                            types.append(type(ast.literal_eval(v)))
154                        except (ValueError, SyntaxError):
155                            types.append(str)
156                context.set_row(tuple(t(v) for t, v in zip(types, row)))
157                yield context.table.reader
158
159    def join(self, step, context):
160        source = step.name
161
162        source_table = context.tables[source]
163        source_context = self.context({source: source_table})
164        column_ranges = {source: range(0, len(source_table.columns))}
165
166        for name, join in step.joins.items():
167            table = context.tables[name]
168            start = max(r.stop for r in column_ranges.values())
169            column_ranges[name] = range(start, len(table.columns) + start)
170            join_context = self.context({name: table})
171
172            if join.get("source_key"):
173                table = self.hash_join(join, source_context, join_context)
174            else:
175                table = self.nested_loop_join(join, source_context, join_context)
176
177            source_context = self.context(
178                {
179                    name: Table(table.columns, table.rows, column_range)
180                    for name, column_range in column_ranges.items()
181                }
182            )
183            condition = self.generate(join["condition"])
184            if condition:
185                source_context.filter(condition)
186
187        if not step.condition and not step.projections:
188            return source_context
189
190        sink = self._project_and_filter(
191            source_context,
192            step,
193            (reader for reader, _ in iter(source_context)),
194        )
195
196        if step.projections:
197            return self.context({step.name: sink})
198        else:
199            return self.context(
200                {
201                    name: Table(table.columns, sink.rows, table.column_range)
202                    for name, table in source_context.tables.items()
203                }
204            )
205
206    def nested_loop_join(self, _join, source_context, join_context):
207        table = Table(source_context.columns + join_context.columns)
208
209        for reader_a, _ in source_context:
210            for reader_b, _ in join_context:
211                table.append(reader_a.row + reader_b.row)
212
213        return table
214
215    def hash_join(self, join, source_context, join_context):
216        source_key = self.generate_tuple(join["source_key"])
217        join_key = self.generate_tuple(join["join_key"])
218        left = join.get("side") == "LEFT"
219        right = join.get("side") == "RIGHT"
220
221        results = collections.defaultdict(lambda: ([], []))
222
223        for reader, ctx in source_context:
224            results[ctx.eval_tuple(source_key)][0].append(reader.row)
225        for reader, ctx in join_context:
226            results[ctx.eval_tuple(join_key)][1].append(reader.row)
227
228        table = Table(source_context.columns + join_context.columns)
229        nulls = [(None,) * len(join_context.columns if left else source_context.columns)]
230
231        for a_group, b_group in results.values():
232            if left:
233                b_group = b_group or nulls
234            elif right:
235                a_group = a_group or nulls
236
237            for a_row, b_row in itertools.product(a_group, b_group):
238                table.append(a_row + b_row)
239
240        return table
241
242    def aggregate(self, step, context):
243        group_by = self.generate_tuple(step.group.values())
244        aggregations = self.generate_tuple(step.aggregations)
245        operands = self.generate_tuple(step.operands)
246
247        if operands:
248            operand_table = Table(self.table(step.operands).columns)
249
250            for reader, ctx in context:
251                operand_table.append(ctx.eval_tuple(operands))
252
253            for i, (a, b) in enumerate(zip(context.table.rows, operand_table.rows)):
254                context.table.rows[i] = a + b
255
256            width = len(context.columns)
257            context.add_columns(*operand_table.columns)
258
259            operand_table = Table(
260                context.columns,
261                context.table.rows,
262                range(width, width + len(operand_table.columns)),
263            )
264
265            context = self.context(
266                {
267                    None: operand_table,
268                    **context.tables,
269                }
270            )
271
272        context.sort(group_by)
273
274        group = None
275        start = 0
276        end = 1
277        length = len(context.table)
278        table = self.table(list(step.group) + step.aggregations)
279        condition = self.generate(step.condition)
280
281        def add_row():
282            if not condition or context.eval(condition):
283                table.append(group + context.eval_tuple(aggregations))
284
285        if length:
286            for i in range(length):
287                context.set_index(i)
288                key = context.eval_tuple(group_by)
289                group = key if group is None else group
290                end += 1
291                if key != group:
292                    context.set_range(start, end - 2)
293                    add_row()
294                    group = key
295                    start = end - 2
296                if len(table.rows) >= step.limit:
297                    break
298                if i == length - 1:
299                    context.set_range(start, end - 1)
300                    add_row()
301        elif step.limit > 0 and not group_by:
302            context.set_range(0, 0)
303            table.append(context.eval_tuple(aggregations))
304
305        context = self.context({step.name: table, **{name: table for name in context.tables}})
306
307        if step.projections:
308            return self.scan(step, context)
309        return context
310
311    def sort(self, step, context):
312        projections = self.generate_tuple(step.projections)
313        projection_columns = [p.alias_or_name for p in step.projections]
314        all_columns = list(context.columns) + projection_columns
315        sink = self.table(all_columns)
316        for reader, ctx in context:
317            sink.append(reader.row + ctx.eval_tuple(projections))
318
319        sort_ctx = self.context(
320            {
321                None: sink,
322                **{table: sink for table in context.tables},
323            }
324        )
325        sort_ctx.sort(self.generate_tuple(step.key))
326
327        if not math.isinf(step.limit):
328            sort_ctx.table.rows = sort_ctx.table.rows[0 : step.limit]
329
330        output = Table(
331            projection_columns,
332            rows=[r[len(context.columns) : len(all_columns)] for r in sort_ctx.table.rows],
333        )
334        return self.context({step.name: output})
335
336    def set_operation(self, step, context):
337        left = context.tables[step.left]
338        right = context.tables[step.right]
339
340        sink = self.table(left.columns)
341
342        if issubclass(step.op, exp.Intersect):
343            sink.rows = list(set(left.rows).intersection(set(right.rows)))
344        elif issubclass(step.op, exp.Except):
345            sink.rows = list(set(left.rows).difference(set(right.rows)))
346        elif issubclass(step.op, exp.Union) and step.distinct:
347            sink.rows = list(set(left.rows).union(set(right.rows)))
348        else:
349            sink.rows = left.rows + right.rows
350
351        return self.context({step.name: sink})
352
353
354def _ordered_py(self, expression):
355    this = self.sql(expression, "this")
356    desc = "True" if expression.args.get("desc") else "False"
357    nulls_first = "True" if expression.args.get("nulls_first") else "False"
358    return f"ORDERED({this}, {desc}, {nulls_first})"
359
360
361def _rename(self, e):
362    try:
363        if "expressions" in e.args:
364            this = self.sql(e, "this")
365            this = f"{this}, " if this else ""
366            return f"{e.key.upper()}({this}{self.expressions(e)})"
367        return self.func(e.key, *e.args.values())
368    except Exception as ex:
369        raise Exception(f"Could not rename {repr(e)}") from ex
370
371
372def _case_sql(self, expression):
373    this = self.sql(expression, "this")
374    chain = self.sql(expression, "default") or "None"
375
376    for e in reversed(expression.args["ifs"]):
377        true = self.sql(e, "true")
378        condition = self.sql(e, "this")
379        condition = f"{this} = ({condition})" if this else condition
380        chain = f"{true} if {condition} else ({chain})"
381
382    return chain
383
384
385def _lambda_sql(self, e: exp.Lambda) -> str:
386    names = {e.name.lower() for e in e.expressions}
387
388    e = e.transform(
389        lambda n: exp.Var(this=n.name)
390        if isinstance(n, exp.Identifier) and n.name.lower() in names
391        else n
392    )
393
394    return f"lambda {self.expressions(e, flat=True)}: {self.sql(e, 'this')}"
395
396
397class Python(Dialect):
398    class Tokenizer(tokens.Tokenizer):
399        STRING_ESCAPES = ["\\"]
400
401    class Generator(generator.Generator):
402        TRANSFORMS = {
403            **{klass: _rename for klass in subclasses(exp.__name__, exp.Binary)},
404            **{klass: _rename for klass in exp.ALL_FUNCTIONS},
405            exp.Case: _case_sql,
406            exp.Alias: lambda self, e: self.sql(e.this),
407            exp.Array: inline_array_sql,
408            exp.And: lambda self, e: self.binary(e, "and"),
409            exp.Between: _rename,
410            exp.Boolean: lambda self, e: "True" if e.this else "False",
411            exp.Cast: lambda self, e: f"CAST({self.sql(e.this)}, exp.DataType.Type.{e.args['to']})",
412            exp.Column: lambda self, e: f"scope[{self.sql(e, 'table') or None}][{self.sql(e.this)}]",
413            exp.Distinct: lambda self, e: f"set({self.sql(e, 'this')})",
414            exp.Extract: lambda self, e: f"EXTRACT('{e.name.lower()}', {self.sql(e, 'expression')})",
415            exp.In: lambda self, e: f"{self.sql(e, 'this')} in ({self.expressions(e, flat=True)})",
416            exp.Is: lambda self, e: self.binary(e, "is"),
417            exp.Lambda: _lambda_sql,
418            exp.Not: lambda self, e: f"not {self.sql(e.this)}",
419            exp.Null: lambda *_: "None",
420            exp.Or: lambda self, e: self.binary(e, "or"),
421            exp.Ordered: _ordered_py,
422            exp.Star: lambda *_: "1",
423        }