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, ensure_list, 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        finished = set()
 23        queue = set(plan.leaves)
 24        contexts = {}
 25
 26        while queue:
 27            node = queue.pop()
 28            try:
 29                context = self.context(
 30                    {
 31                        name: table
 32                        for dep in node.dependencies
 33                        for name, table in contexts[dep].tables.items()
 34                    }
 35                )
 36
 37                if isinstance(node, planner.Scan):
 38                    contexts[node] = self.scan(node, context)
 39                elif isinstance(node, planner.Aggregate):
 40                    contexts[node] = self.aggregate(node, context)
 41                elif isinstance(node, planner.Join):
 42                    contexts[node] = self.join(node, context)
 43                elif isinstance(node, planner.Sort):
 44                    contexts[node] = self.sort(node, context)
 45                elif isinstance(node, planner.SetOperation):
 46                    contexts[node] = self.set_operation(node, context)
 47                else:
 48                    raise NotImplementedError
 49
 50                finished.add(node)
 51
 52                for dep in node.dependents:
 53                    if all(d in contexts for d in dep.dependencies):
 54                        queue.add(dep)
 55
 56                for dep in node.dependencies:
 57                    if all(d in finished for d in dep.dependents):
 58                        contexts.pop(dep)
 59            except Exception as e:
 60                raise ExecuteError(f"Step '{node.id}' failed: {e}") from e
 61
 62        root = plan.root
 63        return contexts[root].tables[root.name]
 64
 65    def generate(self, expression):
 66        """Convert a SQL expression into literal Python code and compile it into bytecode."""
 67        if not expression:
 68            return None
 69
 70        sql = self.generator.generate(expression)
 71        return compile(sql, sql, "eval", optimize=2)
 72
 73    def generate_tuple(self, expressions):
 74        """Convert an array of SQL expressions into tuple of Python byte code."""
 75        if not expressions:
 76            return tuple()
 77        return tuple(self.generate(expression) for expression in expressions)
 78
 79    def context(self, tables):
 80        return Context(tables, env=self.env)
 81
 82    def table(self, expressions):
 83        return Table(
 84            expression.alias_or_name if isinstance(expression, exp.Expression) else expression
 85            for expression in expressions
 86        )
 87
 88    def scan(self, step, context):
 89        source = step.source
 90
 91        if source and isinstance(source, exp.Expression):
 92            source = source.name or source.alias
 93
 94        if source is None:
 95            context, table_iter = self.static()
 96        elif source in context:
 97            if not step.projections and not step.condition:
 98                return self.context({step.name: context.tables[source]})
 99            table_iter = context.table_iter(source)
100        elif isinstance(step.source, exp.Table) and isinstance(step.source.this, exp.ReadCSV):
101            table_iter = self.scan_csv(step)
102            context = next(table_iter)
103        else:
104            context, table_iter = self.scan_table(step)
105
106        return self.context({step.name: self._project_and_filter(context, step, table_iter)})
107
108    def _project_and_filter(self, context, step, table_iter):
109        sink = self.table(step.projections if step.projections else context.columns)
110        condition = self.generate(step.condition)
111        projections = self.generate_tuple(step.projections)
112
113        for reader in table_iter:
114            if len(sink) >= step.limit:
115                break
116
117            if condition and not context.eval(condition):
118                continue
119
120            if projections:
121                sink.append(context.eval_tuple(projections))
122            else:
123                sink.append(reader.row)
124
125        return sink
126
127    def static(self):
128        return self.context({}), [RowReader(())]
129
130    def scan_table(self, step):
131        table = self.tables.find(step.source)
132        context = self.context({step.source.alias_or_name: table})
133        return context, iter(table)
134
135    def scan_csv(self, step):
136        alias = step.source.alias
137        source = step.source.this
138
139        with csv_reader(source) as reader:
140            columns = next(reader)
141            table = Table(columns)
142            context = self.context({alias: table})
143            yield context
144            types = []
145            for row in reader:
146                if not types:
147                    for v in row:
148                        try:
149                            types.append(type(ast.literal_eval(v)))
150                        except (ValueError, SyntaxError):
151                            types.append(str)
152
153                # We can't cast empty values ('') to non-string types, so we convert them to None instead
154                context.set_row(
155                    tuple(None if (t is not str and v == "") else t(v) for t, v in zip(types, row))
156                )
157                yield context.table.reader
158
159    def join(self, step, context):
160        source = step.source_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
280        def add_row():
281            table.append(group + context.eval_tuple(aggregations))
282
283        if length:
284            for i in range(length):
285                context.set_index(i)
286                key = context.eval_tuple(group_by)
287                group = key if group is None else group
288                end += 1
289                if key != group:
290                    context.set_range(start, end - 2)
291                    add_row()
292                    group = key
293                    start = end - 2
294                if len(table.rows) >= step.limit:
295                    break
296                if i == length - 1:
297                    context.set_range(start, end - 1)
298                    add_row()
299        elif step.limit > 0 and not group_by:
300            context.set_range(0, 0)
301            table.append(context.eval_tuple(aggregations))
302
303        context = self.context({step.name: table, **{name: table for name in context.tables}})
304
305        if step.projections or step.condition:
306            return self.scan(step, context)
307        return context
308
309    def sort(self, step, context):
310        projections = self.generate_tuple(step.projections)
311        projection_columns = [p.alias_or_name for p in step.projections]
312        all_columns = list(context.columns) + projection_columns
313        sink = self.table(all_columns)
314        for reader, ctx in context:
315            sink.append(reader.row + ctx.eval_tuple(projections))
316
317        sort_ctx = self.context(
318            {
319                None: sink,
320                **{table: sink for table in context.tables},
321            }
322        )
323        sort_ctx.sort(self.generate_tuple(step.key))
324
325        if not math.isinf(step.limit):
326            sort_ctx.table.rows = sort_ctx.table.rows[0 : step.limit]
327
328        output = Table(
329            projection_columns,
330            rows=[r[len(context.columns) : len(all_columns)] for r in sort_ctx.table.rows],
331        )
332        return self.context({step.name: output})
333
334    def set_operation(self, step, context):
335        left = context.tables[step.left]
336        right = context.tables[step.right]
337
338        sink = self.table(left.columns)
339
340        if issubclass(step.op, exp.Intersect):
341            sink.rows = list(set(left.rows).intersection(set(right.rows)))
342        elif issubclass(step.op, exp.Except):
343            sink.rows = list(set(left.rows).difference(set(right.rows)))
344        elif issubclass(step.op, exp.Union) and step.distinct:
345            sink.rows = list(set(left.rows).union(set(right.rows)))
346        else:
347            sink.rows = left.rows + right.rows
348
349        if not math.isinf(step.limit):
350            sink.rows = sink.rows[0 : step.limit]
351
352        return self.context({step.name: sink})
353
354
355def _ordered_py(self, expression):
356    this = self.sql(expression, "this")
357    desc = "True" if expression.args.get("desc") else "False"
358    nulls_first = "True" if expression.args.get("nulls_first") else "False"
359    return f"ORDERED({this}, {desc}, {nulls_first})"
360
361
362def _rename(self, e):
363    try:
364        values = list(e.args.values())
365
366        if len(values) == 1:
367            values = values[0]
368            if not isinstance(values, list):
369                return self.func(e.key, values)
370            return self.func(e.key, *values)
371
372        if isinstance(e, exp.Func) and e.is_var_len_args:
373            *head, tail = values
374            return self.func(e.key, *head, *ensure_list(tail))
375
376        return self.func(e.key, *values)
377    except Exception as ex:
378        raise Exception(f"Could not rename {repr(e)}") from ex
379
380
381def _case_sql(self, expression):
382    this = self.sql(expression, "this")
383    chain = self.sql(expression, "default") or "None"
384
385    for e in reversed(expression.args["ifs"]):
386        true = self.sql(e, "true")
387        condition = self.sql(e, "this")
388        condition = f"{this} = ({condition})" if this else condition
389        chain = f"{true} if {condition} else ({chain})"
390
391    return chain
392
393
394def _lambda_sql(self, e: exp.Lambda) -> str:
395    names = {e.name.lower() for e in e.expressions}
396
397    e = e.transform(
398        lambda n: (
399            exp.var(n.name) if isinstance(n, exp.Identifier) and n.name.lower() in names else n
400        )
401    ).assert_is(exp.Lambda)
402
403    return f"lambda {self.expressions(e, flat=True)}: {self.sql(e, 'this')}"
404
405
406def _div_sql(self: generator.Generator, e: exp.Div) -> str:
407    denominator = self.sql(e, "expression")
408
409    if e.args.get("safe"):
410        denominator += " or None"
411
412    sql = f"DIV({self.sql(e, 'this')}, {denominator})"
413
414    if e.args.get("typed"):
415        sql = f"int({sql})"
416
417    return sql
418
419
420class Python(Dialect):
421    class Tokenizer(tokens.Tokenizer):
422        STRING_ESCAPES = ["\\"]
423
424    class Generator(generator.Generator):
425        TRANSFORMS = {
426            **{klass: _rename for klass in subclasses(exp.__name__, exp.Binary)},
427            **{klass: _rename for klass in exp.ALL_FUNCTIONS},
428            exp.Case: _case_sql,
429            exp.Alias: lambda self, e: self.sql(e.this),
430            exp.Array: inline_array_sql,
431            exp.And: lambda self, e: self.binary(e, "and"),
432            exp.Between: _rename,
433            exp.Boolean: lambda self, e: "True" if e.this else "False",
434            exp.Cast: lambda self, e: f"CAST({self.sql(e.this)}, exp.DataType.Type.{e.args['to']})",
435            exp.Column: lambda self,
436            e: f"scope[{self.sql(e, 'table') or None}][{self.sql(e.this)}]",
437            exp.Concat: lambda self, e: self.func(
438                "SAFECONCAT" if e.args.get("safe") else "CONCAT", *e.expressions
439            ),
440            exp.Distinct: lambda self, e: f"set({self.sql(e, 'this')})",
441            exp.Div: _div_sql,
442            exp.Extract: lambda self,
443            e: f"EXTRACT('{e.name.lower()}', {self.sql(e, 'expression')})",
444            exp.In: lambda self,
445            e: f"{self.sql(e, 'this')} in {{{self.expressions(e, flat=True)}}}",
446            exp.Interval: lambda self, e: f"INTERVAL({self.sql(e.this)}, '{self.sql(e.unit)}')",
447            exp.Is: lambda self, e: (
448                self.binary(e, "==") if isinstance(e.this, exp.Literal) else self.binary(e, "is")
449            ),
450            exp.JSONExtract: lambda self, e: self.func(e.key, e.this, e.expression, *e.expressions),
451            exp.JSONPath: lambda self, e: f"[{','.join(self.sql(p) for p in e.expressions[1:])}]",
452            exp.JSONPathKey: lambda self, e: f"'{self.sql(e.this)}'",
453            exp.JSONPathSubscript: lambda self, e: f"'{e.this}'",
454            exp.Lambda: _lambda_sql,
455            exp.Not: lambda self, e: f"not {self.sql(e.this)}",
456            exp.Null: lambda *_: "None",
457            exp.Or: lambda self, e: self.binary(e, "or"),
458            exp.Ordered: _ordered_py,
459            exp.Star: lambda *_: "1",
460        }