BUG: Fix fillna on multi indexed Dataframe doesn't work (#47774)

Author: xr-chen

doc/source/whatsnew/v1.4.4.rst

@@ -14,6 +14,7 @@ including other versions of pandas.
 
 Fixed regressions
 ~~~~~~~~~~~~~~~~~
+- Fixed regression in :meth:`DataFrame.fillna` not working :class:`DataFrame` with :class:`MultiIndex` (:issue:`47649`)
 - Fixed regression in taking NULL :class:`objects` from a :class:`DataFrame` causing a segmentation violation. These NULL values are created by :meth:`numpy.empty_like` (:issue:`46848`)
 - Fixed regression in :func:`concat` materializing :class:`Index` during sorting even if :class:`Index` was already sorted (:issue:`47501`)
 - Fixed regression in calling bitwise numpy ufuncs (for example, ``np.bitwise_and``) on Index objects (:issue:`46769`)

pandas/core/generic.py

@@ -6861,16 +6861,22 @@ def fillna(
                 for k, v in value.items():
                     if k not in result:
                         continue
+
                     # error: Item "None" of "Optional[Dict[Any, Any]]" has no
                     # attribute "get"
                     downcast_k = (
                         downcast
                         if not is_dict
                         else downcast.get(k)  # type: ignore[union-attr]
                     )
-                    result.loc[:, k] = result[k].fillna(
-                        v, limit=limit, downcast=downcast_k
+                    # GH47649
+                    result.loc[:, k] = (
+                        result[k].fillna(v, limit=limit, downcast=downcast_k).values
                     )
+                    # TODO: result.loc[:, k] = result.loc[:, k].fillna(
+                    #     v, limit=limit, downcast=downcast_k
+                    # )
+                    # Revert when GH45751 is fixed
                 return result if not inplace else None
 
             elif not is_list_like(value):

pandas/tests/frame/methods/test_fillna.py

@@ -715,6 +715,34 @@ def test_single_block_df_with_horizontal_axis(self):
         )
         tm.assert_frame_equal(result, expected)
 
+    def test_fillna_with_multi_index_frame(self):
+        # GH 47649
+        pdf = DataFrame(
+            {
+                ("x", "a"): [np.nan, 2.0, 3.0],
+                ("x", "b"): [1.0, 2.0, np.nan],
+                ("y", "c"): [1.0, 2.0, np.nan],
+            }
+        )
+        expected = DataFrame(
+            {
+                ("x", "a"): [-1.0, 2.0, 3.0],
+                ("x", "b"): [1.0, 2.0, -1.0],
+                ("y", "c"): [1.0, 2.0, np.nan],
+            }
+        )
+        tm.assert_frame_equal(pdf.fillna({"x": -1}), expected)
+        tm.assert_frame_equal(pdf.fillna({"x": -1, ("x", "b"): -2}), expected)
+
+        expected = DataFrame(
+            {
+                ("x", "a"): [-1.0, 2.0, 3.0],
+                ("x", "b"): [1.0, 2.0, -2.0],
+                ("y", "c"): [1.0, 2.0, np.nan],
+            }
+        )
+        tm.assert_frame_equal(pdf.fillna({("x", "b"): -2, "x": -1}), expected)
+
 
 def test_fillna_nonconsolidated_frame():
     # https://github.com/pandas-dev/pandas/issues/36495