Optimize eval-setitem expressions as single eval expressions (#2695)

Author: wjsi

mars/dataframe/indexing/setitem.py

@@ -166,7 +166,7 @@ def tile(cls, op: "DataFrameSetitem"):
             ]
             value_chunk_index_values = [v.index_value for v in value.chunks]
             is_identical = len(target_chunk_index_values) == len(
-                target_chunk_index_values
+                value_chunk_index_values
             ) and all(
                 c.key == v.key
                 for c, v in zip(target_chunk_index_values, value_chunk_index_values)

mars/optimization/logical/tileable/arithmetic_query.py

@@ -13,22 +13,30 @@
 # limitations under the License.
 
 import weakref
-from typing import Optional, Tuple
+from typing import NamedTuple, Optional
 
+import numpy as np
 from pandas.api.types import is_scalar
 
 from .... import dataframe as md
 from ....core import Tileable, get_output_types, ENTITY_TYPE
 from ....dataframe.arithmetic.core import DataFrameUnaryUfunc, DataFrameBinopUfunc
 from ....dataframe.base.eval import DataFrameEval
 from ....dataframe.indexing.getitem import DataFrameIndex
+from ....dataframe.indexing.setitem import DataFrameSetitem
 from ....typing import OperandType
 from ....utils import implements
 from ..core import OptimizationRecord, OptimizationRecordType
 from ..tileable.core import register_tileable_optimization_rule
 from .core import OptimizationRule
 
 
+class EvalExtractRecord(NamedTuple):
+    tileable: Optional[Tileable] = None
+    expr: Optional[str] = None
+    variables: Optional[dict] = None
+
+
 def _get_binop_builder(op_str: str):
     def builder(lhs: str, rhs: str):
         return f"({lhs}) {op_str} ({rhs})"
@@ -60,9 +68,16 @@ def builder(lhs: str, rhs: str):
 
 @register_tileable_optimization_rule([DataFrameUnaryUfunc, DataFrameBinopUfunc])
 class SeriesArithmeticToEval(OptimizationRule):
+    _var_counter = 0
+
+    @classmethod
+    def _next_var_id(cls):
+        cls._var_counter += 1
+        return cls._var_counter
+
     @implements(OptimizationRule.match)
     def match(self, op: OperandType) -> bool:
-        _, expr = self._extract_eval_expression(op.outputs[0])
+        _, expr, _ = self._extract_eval_expression(op.outputs[0])
         return expr is not None
 
     @staticmethod
@@ -91,14 +106,17 @@ def _is_select_dataframe_column(tileable) -> bool:
         )
 
     @classmethod
-    def _extract_eval_expression(
-        cls, tileable
-    ) -> Tuple[Optional[Tileable], Optional[str]]:
+    def _extract_eval_expression(cls, tileable) -> EvalExtractRecord:
         if is_scalar(tileable):
-            return None, repr(tileable)
+            if isinstance(tileable, (int, bool, str, bytes, np.integer, np.bool_)):
+                return EvalExtractRecord(expr=repr(tileable))
+            else:
+                var_name = f"__eval_scalar_var{cls._next_var_id()}"
+                var_dict = {var_name: tileable}
+                return EvalExtractRecord(expr=f"@{var_name}", variables=var_dict)
 
         if not isinstance(tileable, ENTITY_TYPE):  # pragma: no cover
-            return None, None
+            return EvalExtractRecord()
 
         if tileable in _extract_result_cache:
             return _extract_result_cache[tileable]
@@ -109,69 +127,75 @@ def _extract_eval_expression(
             result = cls._extract_unary(tileable)
         elif isinstance(tileable.op, DataFrameBinopUfunc):
             if tileable.op.fill_value is not None or tileable.op.level is not None:
-                result = None, None
+                result = EvalExtractRecord()
             else:
                 result = cls._extract_binary(tileable)
         else:
-            result = None, None
+            result = EvalExtractRecord()
 
         _extract_result_cache[tileable] = result
         return result
 
     @classmethod
-    def _extract_column_select(
-        cls, tileable
-    ) -> Tuple[Optional[Tileable], Optional[str]]:
-        return tileable.inputs[0], f"`{tileable.op.col_names}`"
+    def _extract_column_select(cls, tileable) -> EvalExtractRecord:
+        return EvalExtractRecord(tileable.inputs[0], f"`{tileable.op.col_names}`")
 
     @classmethod
-    def _extract_unary(cls, tileable) -> Tuple[Optional[Tileable], Optional[str]]:
+    def _extract_unary(cls, tileable) -> EvalExtractRecord:
         op = tileable.op
         func_name = getattr(op, "_func_name") or getattr(op, "_bin_func_name")
         if func_name not in _func_name_to_builder:  # pragma: no cover
-            return None, None
+            return EvalExtractRecord()
 
-        in_tileable, expr = cls._extract_eval_expression(op.inputs[0])
+        in_tileable, expr, variables = cls._extract_eval_expression(op.inputs[0])
         if in_tileable is None:
-            return None, None
+            return EvalExtractRecord()
 
         cls._add_collapsable_predecessor(tileable, op.inputs[0])
-        return in_tileable, _func_name_to_builder[func_name](expr)
+        return EvalExtractRecord(
+            in_tileable, _func_name_to_builder[func_name](expr), variables
+        )
 
     @classmethod
-    def _extract_binary(cls, tileable) -> Tuple[Optional[Tileable], Optional[str]]:
+    def _extract_binary(cls, tileable) -> EvalExtractRecord:
         op = tileable.op
         func_name = getattr(op, "_func_name", None) or getattr(op, "_bit_func_name")
         if func_name not in _func_name_to_builder:  # pragma: no cover
-            return None, None
+            return EvalExtractRecord()
 
-        lhs_tileable, lhs_expr = cls._extract_eval_expression(op.lhs)
+        lhs_tileable, lhs_expr, lhs_vars = cls._extract_eval_expression(op.lhs)
         if lhs_tileable is not None:
             cls._add_collapsable_predecessor(tileable, op.lhs)
-        rhs_tileable, rhs_expr = cls._extract_eval_expression(op.rhs)
+        rhs_tileable, rhs_expr, rhs_vars = cls._extract_eval_expression(op.rhs)
         if rhs_tileable is not None:
             cls._add_collapsable_predecessor(tileable, op.rhs)
 
         if lhs_expr is None or rhs_expr is None:
-            return None, None
+            return EvalExtractRecord()
         if (
             lhs_tileable is not None
             and rhs_tileable is not None
             and lhs_tileable.key != rhs_tileable.key
         ):
-            return None, None
+            return EvalExtractRecord()
+
+        variables = (lhs_vars or dict()).copy()
+        variables.update(rhs_vars or dict())
         in_tileable = next(t for t in [lhs_tileable, rhs_tileable] if t is not None)
-        return in_tileable, _func_name_to_builder[func_name](lhs_expr, rhs_expr)
+        return EvalExtractRecord(
+            in_tileable, _func_name_to_builder[func_name](lhs_expr, rhs_expr), variables
+        )
 
     @implements(OptimizationRule.apply)
     def apply(self, op: OperandType):
         node = op.outputs[0]
-        in_tileable, expr = self._extract_eval_expression(node)
+        in_tileable, expr, variables = self._extract_eval_expression(node)
 
         new_op = DataFrameEval(
             _key=node.op.key,
             _output_types=get_output_types(node),
             expr=expr,
+            variables=variables or dict(),
             parser="pandas",
             is_query=False,
         )
@@ -195,39 +219,41 @@ def apply(self, op: OperandType):
             pass
 
 
-@register_tileable_optimization_rule([DataFrameIndex])
-class DataFrameBoolEvalToQuery(OptimizationRule):
-    def match(self, op: "DataFrameIndex") -> bool:
+class _DataFrameEvalRewriteRule(OptimizationRule):
+    def match(self, op: OperandType) -> bool:
+        optimized_eval_op = self._get_optimized_eval_op(op)
         if (
-            op.col_names is not None
-            or not isinstance(op.mask, md.Series)
-            or op.mask.dtype != bool
+            not isinstance(optimized_eval_op, DataFrameEval)
+            or optimized_eval_op.is_query
+            or optimized_eval_op.inputs[0].key != op.inputs[0].key
         ):
             return False
-        optimized = self._records.get_optimization_result(op.mask)
-        mask_op = optimized.op if optimized is not None else op.mask.op
-        if not isinstance(mask_op, DataFrameEval) or mask_op.is_query:
-            return False
         return True
 
-    def apply(self, op: "DataFrameIndex"):
+    def _build_new_eval_op(self, op: OperandType):
+        raise NotImplementedError
+
+    def _get_optimized_eval_op(self, op: OperandType) -> OperandType:
+        in_columnar_node = self._get_input_columnar_node(op)
+        optimized = self._records.get_optimization_result(in_columnar_node)
+        return optimized.op if optimized is not None else in_columnar_node.op
+
+    def _get_input_columnar_node(self, op: OperandType) -> ENTITY_TYPE:
+        raise NotImplementedError
+
+    def apply(self, op: DataFrameIndex):
         node = op.outputs[0]
         in_tileable = op.inputs[0]
+        in_columnar_node = self._get_input_columnar_node(op)
+
+        new_op = self._build_new_eval_op(op)
+        new_op._key = node.op.key
 
-        optimized = self._records.get_optimization_result(op.mask)
-        mask_op = optimized.op if optimized is not None else op.mask.op
-        new_op = DataFrameEval(
-            _key=node.op.key,
-            _output_types=get_output_types(node),
-            expr=mask_op.expr,
-            parser="pandas",
-            is_query=True,
-        )
         new_node = new_op.new_tileable([in_tileable], **node.params).data
         new_node._key = node.key
         new_node._id = node.id
 
-        self._add_collapsable_predecessor(node, op.mask)
+        self._add_collapsable_predecessor(node, in_columnar_node)
         self._remove_collapsable_predecessors(node)
 
         self._replace_node(node, new_node)
@@ -242,3 +268,50 @@ def apply(self, op: "DataFrameIndex"):
             self._graph.results[i] = new_node
         except ValueError:
             pass
+
+
+@register_tileable_optimization_rule([DataFrameIndex])
+class DataFrameBoolEvalToQuery(_DataFrameEvalRewriteRule):
+    def match(self, op: DataFrameIndex) -> bool:
+        if (
+            op.col_names is not None
+            or not isinstance(op.mask, md.Series)
+            or op.mask.dtype != bool
+        ):
+            return False
+        return super().match(op)
+
+    def _get_input_columnar_node(self, op: OperandType) -> ENTITY_TYPE:
+        return op.mask
+
+    def _build_new_eval_op(self, op: OperandType):
+        in_eval_op = self._get_optimized_eval_op(op)
+        return DataFrameEval(
+            _output_types=get_output_types(op.outputs[0]),
+            expr=in_eval_op.expr,
+            variables=in_eval_op.variables,
+            parser="pandas",
+            is_query=True,
+        )
+
+
+@register_tileable_optimization_rule([DataFrameSetitem])
+class DataFrameEvalSetItemToEval(_DataFrameEvalRewriteRule):
+    def match(self, op: DataFrameSetitem):
+        if not isinstance(op.indexes, str) or not isinstance(op.value, md.Series):
+            return False
+        return super().match(op)
+
+    def _get_input_columnar_node(self, op: DataFrameSetitem) -> ENTITY_TYPE:
+        return op.value
+
+    def _build_new_eval_op(self, op: DataFrameSetitem):
+        in_eval_op = self._get_optimized_eval_op(op)
+        return DataFrameEval(
+            _output_types=get_output_types(op.outputs[0]),
+            expr=f"`{op.indexes}` = {in_eval_op.expr}",
+            variables=in_eval_op.variables,
+            parser="pandas",
+            is_query=False,
+            self_target=True,
+        )

mars/optimization/logical/tileable/tests/test_arithmetic_query.py

@@ -12,6 +12,8 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+import re
+
 import numpy as np
 import pandas as pd
 
@@ -22,6 +24,13 @@
 from .. import optimize
 
 
+_var_pattern = re.compile(r"@__eval_scalar_var\d+")
+
+
+def _norm_vars(var_str):
+    return _var_pattern.sub("@scalar", var_str)
+
+
 @enter_mode(build=True)
 def test_arithmetic_query(setup):
     raw = pd.DataFrame(np.random.rand(100, 10), columns=list("ABCDEFGHIJ"))
@@ -62,7 +71,6 @@ def test_arithmetic_query(setup):
 
     pd.testing.assert_series_equal(df2.execute().fetch(), -raw["A"] + raw["B"] * 5)
 
-    raw = pd.DataFrame(np.random.rand(100, 10), columns=list("ABCDEFGHIJ"))
     df1 = md.DataFrame(raw, chunk_size=10)
     df2 = -df1["A"] + df1["B"] * 5 + 3 * df1["C"]
     graph = TileableGraph([df1["A"].data, df2.data])
@@ -74,22 +82,6 @@ def test_arithmetic_query(setup):
     r_df2, _r_col_a = fetch(execute(df2, df1["A"]))
     pd.testing.assert_series_equal(r_df2, -raw["A"] + raw["B"] * 5 + 3 * raw["C"])
 
-    df1 = md.DataFrame(raw, chunk_size=10)
-    df2 = md.DataFrame(raw2, chunk_size=10)
-    df3 = df1.merge(df2, on="A", suffixes=("", "_"))
-    df3["K"] = df4 = df3["A"] * (1 - df3["B"])
-    graph = TileableGraph([df3.data])
-    next(TileableGraphBuilder(graph).build())
-    records = optimize(graph)
-    opt_df4 = records.get_optimization_result(df4.data)
-    assert opt_df4.op.expr == "(`A`) * ((1) - (`B`))"
-    assert len(graph) == 5
-    assert len([n for n in graph if isinstance(n.op, DataFrameEval)]) == 1
-
-    r_df3 = raw.merge(raw2, on="A", suffixes=("", "_"))
-    r_df3["K"] = r_df3["A"] * (1 - r_df3["B"])
-    pd.testing.assert_frame_equal(df3.execute().fetch(), r_df3)
-
 
 @enter_mode(build=True)
 def test_bool_eval_to_query(setup):
@@ -111,7 +103,7 @@ def test_bool_eval_to_query(setup):
     opt_df2 = records.get_optimization_result(df2.data)
     assert isinstance(opt_df2.op, DataFrameEval)
     assert opt_df2.op.is_query
-    assert opt_df2.op.expr == "((`A`) > (0.5)) & ((`C`) < (0.5))"
+    assert _norm_vars(opt_df2.op.expr) == "((`A`) > (@scalar)) & ((`C`) < (@scalar))"
 
     pd.testing.assert_frame_equal(
         df2.execute().fetch(), raw[(raw["A"] > 0.5) & (raw["C"] < 0.5)]
@@ -138,7 +130,37 @@ def test_bool_eval_to_query(setup):
     next(TileableGraphBuilder(graph).build())
     records = optimize(graph)
     opt_df2 = records.get_optimization_result(df2.data)
-    assert opt_df2.op.expr == "(`b`) < (Timestamp('2022-03-20 00:00:00'))"
+    assert _norm_vars(opt_df2.op.expr) == "(`b`) < (@scalar)"
 
     r_df2 = fetch(execute(df2))
     pd.testing.assert_frame_equal(r_df2, raw[raw.b < pd.Timestamp("2022-3-20")])
+
+
+@enter_mode(build=True)
+def test_eval_setitem_to_eval(setup):
+    raw = pd.DataFrame(np.random.rand(100, 10), columns=list("ABCDEFGHIJ"))
+    raw2 = pd.DataFrame(np.random.rand(100, 5), columns=list("ABCDE"))
+
+    # does not support non-eval value setting
+    df1 = md.DataFrame(raw, chunk_size=10)
+    df1["K"] = 345
+    graph = TileableGraph([df1.data])
+    next(TileableGraphBuilder(graph).build())
+    records = optimize(graph)
+    assert records.get_optimization_result(df1.data) is None
+
+    df1 = md.DataFrame(raw, chunk_size=10)
+    df2 = md.DataFrame(raw2, chunk_size=10)
+    df3 = df1.merge(df2, on="A", suffixes=("", "_"))
+    df3["K"] = df3["A"] * (1 - df3["B"])
+    graph = TileableGraph([df3.data])
+    next(TileableGraphBuilder(graph).build())
+    records = optimize(graph)
+    opt_df3 = records.get_optimization_result(df3.data)
+    assert opt_df3.op.expr == "`K` = (`A`) * ((1) - (`B`))"
+    assert len(graph) == 4
+    assert len([n for n in graph if isinstance(n.op, DataFrameEval)]) == 1
+
+    r_df3 = raw.merge(raw2, on="A", suffixes=("", "_"))
+    r_df3["K"] = r_df3["A"] * (1 - r_df3["B"])
+    pd.testing.assert_frame_equal(df3.execute().fetch(), r_df3)