PERF-#7230: Don't preserve bad partition for merge

modin/core/storage_formats/pandas/merge.py

@@ -13,15 +13,23 @@
 
 """Contains implementations for Merge/Join."""
 
+from __future__ import annotations
+
+from typing import TYPE_CHECKING
+
 import pandas
 from pandas.core.dtypes.common import is_list_like
 from pandas.errors import MergeError
 
+from modin.config import MinPartitionSize, NPartitions
 from modin.core.dataframe.base.dataframe.utils import join_columns
 from modin.core.dataframe.pandas.metadata import ModinDtypes
 
 from .utils import merge_partitioning
 
+if TYPE_CHECKING:
+    from modin.core.storage_formats.pandas.query_compiler import PandasQueryCompiler
+
 
 # TODO: add methods for 'join' here
 class MergeImpl:
@@ -93,7 +101,9 @@
         ).reset_index(drop=True)
 
     @classmethod
-    def row_axis_merge(cls, left, right, kwargs):
+    def row_axis_merge(
+        cls, left: PandasQueryCompiler, right: PandasQueryCompiler, kwargs: dict
+    ):
         """
         Execute merge using row-axis implementation.
 
@@ -164,6 +174,16 @@
                 left, right, on, left_on, right_on, kwargs.get("suffixes", ("_x", "_y"))
             )
 
+            # We rebalance when the ratio of the number of existing partitions to
+            # the ideal number of partitions is smaller than this threshold. The
+            # threshold is a heuristic that may need to be tuned for performance.
+            if (
+                left._modin_frame._partitions.shape[0] < 0.3 * NPartitions.get()
+                # to avoid empty partitions after repartition; can materialize index
+                and len(left._modin_frame) > NPartitions.get() * MinPartitionSize.get()
+            ):
+                left = left.repartition(axis=0)
+
             new_left = left.__constructor__(
                 left._modin_frame.broadcast_apply_full_axis(
                     axis=1,

modin/tests/core/storage_formats/pandas/test_internals.py

@@ -811,6 +811,41 @@ def test_merge_partitioning(
     )
 
 
+def test_merge_with_bad_partitioning():
+    # https://github.com/modin-project/modin/pull/7229
+
+    left_partitioning = [256]
+    right_partitioning = [32, 32, 32, 32]
+
+    left_df = pandas.DataFrame(
+        [np.arange(sum(left_partitioning)) for _ in range(sum(left_partitioning))]
+    )
+    right_df = pandas.DataFrame(
+        [np.arange(sum(right_partitioning)) for _ in range(sum(right_partitioning))]
+    )
+
+    left = construct_modin_df_by_scheme(
+        left_df, {"row_lengths": left_partitioning, "column_widths": left_partitioning}
+    )
+    right = construct_modin_df_by_scheme(
+        right_df,
+        {"row_lengths": right_partitioning, "column_widths": right_partitioning},
+    )
+
+    left_frame = left._query_compiler._modin_frame
+    right_frame = right._query_compiler._modin_frame
+    assert left_frame.row_lengths == left_frame.column_widths == left_partitioning
+    assert right_frame.row_lengths == right_frame.column_widths == right_partitioning
+
+    # just a dummy value
+    return_value = pd.DataFrame([1, 2, 3, 4])._query_compiler
+    with mock.patch.object(
+        left._query_compiler, "repartition", return_value=return_value
+    ) as repartition:
+        _ = left.merge(right)
+        repartition.assert_called_once_with(axis=0)
+
+
 def test_groupby_with_empty_partition():
     # see #5461 for details
     md_df = construct_modin_df_by_scheme(