Use shuffle when nunique is calculated

Author: wjsi

mars/dataframe/groupby/aggregation.py

@@ -25,7 +25,7 @@
 from ... import opcodes as OperandDef
 from ...config import options
 from ...core.custom_log import redirect_custom_log
-from ...core import ENTITY_TYPE, OutputType
+from ...core import ENTITY_TYPE, OutputType, recursive_tile
 from ...core.context import get_context
 from ...core.operand import OperandStage
 from ...serialization.serializables import (
@@ -64,6 +64,8 @@
 
 _support_get_group_without_as_index = pd_release_version[:2] > (1, 0)
 
+_FUNCS_PREFER_SHUFFLE = {"nunique"}
+
 
 class SizeRecorder:
     def __init__(self):
@@ -163,6 +165,8 @@ class DataFrameGroupByAgg(DataFrameOperand, DataFrameOperandMixin):
     method = StringField("method")
     use_inf_as_na = BoolField("use_inf_as_na")
 
+    map_on_shuffle = AnyField("map_on_shuffle")
+
     # for chunk
     combine_size = Int32Field("combine_size")
     chunk_store_limit = Int64Field("chunk_store_limit")
@@ -421,10 +425,29 @@ def _tile_with_shuffle(
         in_df: TileableType,
         out_df: TileableType,
         func_infos: ReductionSteps,
+        agg_chunks: List[ChunkType] = None,
     ):
-        # First, perform groupby and aggregation on each chunk.
-        agg_chunks = cls._gen_map_chunks(op, in_df.chunks, out_df, func_infos)
-        return cls._perform_shuffle(op, agg_chunks, in_df, out_df, func_infos)
+        if op.map_on_shuffle is None:
+            op.map_on_shuffle = all(
+                agg_fun.custom_reduction is None for agg_fun in func_infos.agg_funcs
+            )
+
+        if not op.map_on_shuffle:
+            groupby_params = op.groupby_params.copy()
+            selection = groupby_params.pop("selection", None)
+            groupby = in_df.groupby(**groupby_params)
+            if selection:
+                groupby = groupby[selection]
+            result = groupby.transform(
+                op.raw_func, _call_agg=True, index=out_df.index_value
+            )
+            return (yield from recursive_tile(result))
+        else:
+            # First, perform groupby and aggregation on each chunk.
+            agg_chunks = agg_chunks or cls._gen_map_chunks(
+                op, in_df.chunks, out_df, func_infos
+            )
+            return cls._perform_shuffle(op, agg_chunks, in_df, out_df, func_infos)
 
     @classmethod
     def _perform_shuffle(
@@ -624,8 +647,10 @@ def _tile_auto(
         else:
             # otherwise, use shuffle
             logger.debug("Choose shuffle method for groupby operand %s", op)
-            return cls._perform_shuffle(
-                op, chunks + left_chunks, in_df, out_df, func_infos
+            return (
+                yield from cls._tile_with_shuffle(
+                    op, in_df, out_df, func_infos, chunks + left_chunks
+                )
             )
 
     @classmethod
@@ -638,12 +663,16 @@ def tile(cls, op: "DataFrameGroupByAgg"):
         func_infos = cls._compile_funcs(op, in_df)
 
         if op.method == "auto":
-            if len(in_df.chunks) <= op.combine_size:
+            if set(op.func) & _FUNCS_PREFER_SHUFFLE:
+                return (
+                    yield from cls._tile_with_shuffle(op, in_df, out_df, func_infos)
+                )
+            elif len(in_df.chunks) <= op.combine_size:
                 return cls._tile_with_tree(op, in_df, out_df, func_infos)
             else:
                 return (yield from cls._tile_auto(op, in_df, out_df, func_infos))
         if op.method == "shuffle":
-            return cls._tile_with_shuffle(op, in_df, out_df, func_infos)
+            return (yield from cls._tile_with_shuffle(op, in_df, out_df, func_infos))
         elif op.method == "tree":
             return cls._tile_with_tree(op, in_df, out_df, func_infos)
         else:  # pragma: no cover
@@ -1075,7 +1104,15 @@ def execute(cls, ctx, op: "DataFrameGroupByAgg"):
             pd.reset_option("mode.use_inf_as_na")
 
 
-def agg(groupby, func=None, method="auto", combine_size=None, *args, **kwargs):
+def agg(
+    groupby,
+    func=None,
+    method="auto",
+    combine_size=None,
+    map_on_shuffle=None,
+    *args,
+    **kwargs,
+):
     """
     Aggregate using one or more operations on grouped data.
 
@@ -1091,7 +1128,11 @@ def agg(groupby, func=None, method="auto", combine_size=None, *args, **kwargs):
         in distributed mode and use 'tree' in local mode.
     combine_size : int
         The number of chunks to combine when method is 'tree'
-
+    map_on_shuffle : bool
+        When not specified, will decide whether to perform aggregation on the
+        map stage of shuffle (currently no aggregation when there is custom
+        reduction in functions). Otherwise, whether to call map on map stage
+        of shuffle is determined by the value.
 
     Returns
     -------
@@ -1138,5 +1179,6 @@ def agg(groupby, func=None, method="auto", combine_size=None, *args, **kwargs):
         combine_size=combine_size or options.combine_size,
         chunk_store_limit=options.chunk_store_limit,
         use_inf_as_na=use_inf_as_na,
+        map_on_shuffle=map_on_shuffle,
     )
     return agg_op(groupby)

mars/dataframe/reduction/nunique.py

@@ -58,6 +58,8 @@ def _drop_duplicates(self, value, explode=False, agg=False):
             value = value.values
 
         if explode:
+            if len(value) == 0:
+                return [xp.array([], dtype=object)]
             value = xp.concatenate(value)
 
         value = xdf.unique(value)
@@ -79,7 +81,8 @@ def _drop_duplicates(self, value, explode=False, agg=False):
     def pre(self, in_data):  # noqa: W0221  # pylint: disable=arguments-differ
         xp, xdf = self._get_modules()
         if isinstance(in_data, xdf.Series):
-            unique_values = self._drop_duplicates(in_data)
+            # unique_values = self._drop_duplicates(in_data)
+            unique_values = [in_data.values]
             return xdf.Series(unique_values, name=in_data.name, dtype=object)
         else:
             if self._axis == 0: