PERF: avoid copies in frame[col].fillna(val, inplace=True) GH#46149 (#46204)

Author: jbrockmendel

doc/source/whatsnew/v1.4.2.rst

@@ -15,6 +15,7 @@ including other versions of pandas.
 Fixed regressions
 ~~~~~~~~~~~~~~~~~
 - Fixed regression in :meth:`DataFrame.drop` and :meth:`Series.drop` when :class:`Index` had extension dtype and duplicates (:issue:`45860`)
+- Fixed memory performance regression in :meth:`Series.fillna` when called on a :class:`DataFrame` column with ``inplace=True`` (:issue:`46149`)
 -
 
 .. ---------------------------------------------------------------------------

pandas/core/frame.py

@@ -3960,6 +3960,12 @@ def _maybe_cache_changed(self, item, value: Series, inplace: bool) -> None:
         """
         loc = self._info_axis.get_loc(item)
         arraylike = value._values
+
+        old = self._ixs(loc, axis=1)
+        if old._values is value._values and inplace:
+            # GH#46149 avoid making unnecessary copies/block-splitting
+            return
+
         self._mgr.iset(loc, arraylike, inplace=inplace)
 
     # ----------------------------------------------------------------------

pandas/core/generic.py

@@ -4368,7 +4368,7 @@ def _update_inplace(self, result, verify_is_copy: bool_t = True) -> None:
         self._reset_cache()
         self._clear_item_cache()
         self._mgr = result._mgr
-        self._maybe_update_cacher(verify_is_copy=verify_is_copy)
+        self._maybe_update_cacher(verify_is_copy=verify_is_copy, inplace=True)
 
     @final
     def add_prefix(self: NDFrameT, prefix: str) -> NDFrameT:

pandas/tests/frame/methods/test_fillna.py

@@ -1,6 +1,8 @@
 import numpy as np
 import pytest
 
+import pandas.util._test_decorators as td
+
 from pandas import (
     Categorical,
     DataFrame,
@@ -17,6 +19,19 @@
 
 
 class TestFillNA:
+    @td.skip_array_manager_not_yet_implemented
+    def test_fillna_on_column_view(self):
+        # GH#46149 avoid unnecessary copies
+        arr = np.full((40, 50), np.nan)
+        df = DataFrame(arr)
+
+        df[0].fillna(-1, inplace=True)
+        assert (arr[:, 0] == -1).all()
+
+        # i.e. we didn't create a new 49-column block
+        assert len(df._mgr.arrays) == 1
+        assert np.shares_memory(df.values, arr)
+
     def test_fillna_datetime(self, datetime_frame):
         tf = datetime_frame
         tf.loc[tf.index[:5], "A"] = np.nan