PERF: Merge empty frame (#45838)

* faster merge with empty frame

* whatsnew

* docs, tests, asvs

* fix whatsnew

Co-authored-by: Jeff Reback <[email protected]>

Author: lukemanley

asv_bench/benchmarks/join_merge.py

@@ -216,6 +216,12 @@ def time_merge_dataframe_integer_2key(self, sort):
     def time_merge_dataframe_integer_key(self, sort):
         merge(self.df, self.df2, on="key1", sort=sort)
 
+    def time_merge_dataframe_empty_right(self, sort):
+        merge(self.left, self.right.iloc[:0], sort=sort)
+
+    def time_merge_dataframe_empty_left(self, sort):
+        merge(self.left.iloc[:0], self.right, sort=sort)
+
     def time_merge_dataframes_cross(self, sort):
         merge(self.left.loc[:2000], self.right.loc[:2000], how="cross", sort=sort)
 

doc/source/whatsnew/v1.5.0.rst

@@ -244,6 +244,7 @@ Performance improvements
 - Performance improvement in :meth:`DataFrame.duplicated` when subset consists of only one column (:issue:`45236`)
 - Performance improvement in :meth:`.GroupBy.transform` when broadcasting values for user-defined functions (:issue:`45708`)
 - Performance improvement in :meth:`.GroupBy.transform` for user-defined functions when only a single group exists (:issue:`44977`)
+- Performance improvement in :func:`merge` when left and/or right are empty (:issue:`45838`)
 - Performance improvement in :class:`DataFrame` and :class:`Series` constructors for extension dtype scalars (:issue:`45854`)
 -
 

pandas/core/reshape/merge.py

@@ -977,7 +977,6 @@ def _get_join_info(
                     )
                 else:
                     join_index = self.right.index.take(right_indexer)
-                    left_indexer = np.array([-1] * len(join_index), dtype=np.intp)
             elif self.left_index:
                 if self.how == "asof":
                     # GH#33463 asof should always behave like a left merge
@@ -997,7 +996,6 @@ def _get_join_info(
                     )
                 else:
                     join_index = self.left.index.take(left_indexer)
-                    right_indexer = np.array([-1] * len(join_index), dtype=np.intp)
             else:
                 join_index = Index(np.arange(len(left_indexer)))
 
@@ -1477,6 +1475,20 @@ def get_join_indexers(
         right_keys
     ), "left_key and right_keys must be the same length"
 
+    # fast-path for empty left/right
+    left_n = len(left_keys[0])
+    right_n = len(right_keys[0])
+    if left_n == 0:
+        if how in ["left", "inner", "cross"]:
+            return _get_empty_indexer()
+        elif not sort and how in ["right", "outer"]:
+            return _get_no_sort_one_missing_indexer(right_n, True)
+    elif right_n == 0:
+        if how in ["right", "inner", "cross"]:
+            return _get_empty_indexer()
+        elif not sort and how in ["left", "outer"]:
+            return _get_no_sort_one_missing_indexer(left_n, False)
+
     # get left & right join labels and num. of levels at each location
     mapped = (
         _factorize_keys(left_keys[n], right_keys[n], sort=sort, how=how)
@@ -2055,6 +2067,43 @@ def _get_single_indexer(
     return libjoin.left_outer_join(left_key, right_key, count, sort=sort)
 
 
+def _get_empty_indexer() -> tuple[npt.NDArray[np.intp], npt.NDArray[np.intp]]:
+    """Return empty join indexers."""
+    return (
+        np.array([], dtype=np.intp),
+        np.array([], dtype=np.intp),
+    )
+
+
+def _get_no_sort_one_missing_indexer(
+    n: int, left_missing: bool
+) -> tuple[npt.NDArray[np.intp], npt.NDArray[np.intp]]:
+    """
+    Return join indexers where all of one side is selected without sorting
+    and none of the other side is selected.
+
+    Parameters
+    ----------
+    n : int
+        Length of indexers to create.
+    left_missing : bool
+        If True, the left indexer will contain only -1's.
+        If False, the right indexer will contain only -1's.
+
+    Returns
+    -------
+    np.ndarray[np.intp]
+        Left indexer
+    np.ndarray[np.intp]
+        Right indexer
+    """
+    idx = np.arange(n, dtype=np.intp)
+    idx_missing = np.full(shape=n, fill_value=-1, dtype=np.intp)
+    if left_missing:
+        return idx_missing, idx
+    return idx, idx_missing
+
+
 def _left_join_on_index(
     left_ax: Index, right_ax: Index, join_keys, sort: bool = False
 ) -> tuple[Index, npt.NDArray[np.intp] | None, npt.NDArray[np.intp]]:

pandas/tests/reshape/merge/test_merge.py

@@ -1717,6 +1717,46 @@ def test_merge_ea_with_string(self, join_type, string_dtype):
         )
         tm.assert_series_equal(merged.dtypes, expected)
 
+    @pytest.mark.parametrize(
+        "left_empty, how, exp",
+        [
+            (False, "left", "left"),
+            (False, "right", "empty"),
+            (False, "inner", "empty"),
+            (False, "outer", "left"),
+            (False, "cross", "empty_cross"),
+            (True, "left", "empty"),
+            (True, "right", "right"),
+            (True, "inner", "empty"),
+            (True, "outer", "right"),
+            (True, "cross", "empty_cross"),
+        ],
+    )
+    def test_merge_empty(self, left_empty, how, exp):
+
+        left = DataFrame({"A": [2, 1], "B": [3, 4]})
+        right = DataFrame({"A": [1], "C": [5]}, dtype="int64")
+
+        if left_empty:
+            left = left.head(0)
+        else:
+            right = right.head(0)
+
+        result = left.merge(right, how=how)
+
+        if exp == "left":
+            expected = DataFrame({"A": [2, 1], "B": [3, 4], "C": [np.nan, np.nan]})
+        elif exp == "right":
+            expected = DataFrame({"B": [np.nan], "A": [1], "C": [5]})
+        elif exp == "empty":
+            expected = DataFrame(columns=["A", "B", "C"], dtype="int64")
+            if left_empty:
+                expected = expected[["B", "A", "C"]]
+        elif exp == "empty_cross":
+            expected = DataFrame(columns=["A_x", "B", "A_y", "C"], dtype="int64")
+
+        tm.assert_frame_equal(result, expected)
+
 
 @pytest.fixture
 def left():