BUG: Series[object].fillna ignoring downcast='infer'

doc/source/whatsnew/v1.4.0.rst

@@ -732,6 +732,8 @@ Missing
 - Bug in :meth:`Series.interpolate` and :meth:`DataFrame.interpolate` with ``inplace=True`` not writing to the underlying array(s) in-place (:issue:`44749`)
 - Bug in :meth:`Index.fillna` incorrectly returning an un-filled :class:`Index` when NA values are present and ``downcast`` argument is specified. This now raises ``NotImplementedError`` instead; do not pass ``downcast`` argument (:issue:`44873`)
 - Bug in :meth:`DataFrame.dropna` changing :class:`Index` even if no entries were dropped (:issue:`41965`)
+- Bug in :meth:`Series.fillna` with an object-dtype incorrectly ignoring ``downcast="infer"`` (:issue:`44241`)
+-
 
 MultiIndex
 ^^^^^^^^^^

pandas/core/generic.py

@@ -6399,6 +6399,11 @@ def fillna(
         else:
             if self.ndim == 1:
                 if isinstance(value, (dict, ABCSeries)):
+                    if not len(value):
+                        # test_fillna_nonscalar
+                        if inplace:
+                            return None
+                        return self.copy()
                     value = create_series_with_explicit_dtype(
                         value, dtype_if_empty=object
                     )

pandas/core/internals/blocks.py

@@ -518,11 +518,11 @@ def split_and_operate(self, func, *args, **kwargs) -> list[Block]:
     def _maybe_downcast(self, blocks: list[Block], downcast=None) -> list[Block]:
 
         if self.dtype == _dtype_obj:
-            # TODO: why is behavior different for object dtype?
-            if downcast is not None:
-                return blocks
-
+            # GH#44241 We downcast regardless of the argument;
+            #  respecting 'downcast=None' may be worthwhile at some point,
+            #  but ATM it breaks too much existing code.
             # split and convert the blocks
+
             return extend_blocks(
                 [blk.convert(datetime=True, numeric=False) for blk in blocks]
             )

pandas/tests/series/methods/test_fillna.py

@@ -189,6 +189,42 @@ def test_fillna_downcast(self):
         expected = Series([1, 0])
         tm.assert_series_equal(result, expected)
 
+    def test_fillna_downcast_infer_objects_to_numeric(self):
+        # GH#44241 if we have object-dtype, 'downcast="infer"' should
+        #  _actually_ infer
+
+        arr = np.arange(5).astype(object)
+        arr[3] = np.nan
+
+        ser = Series(arr)
+
+        res = ser.fillna(3, downcast="infer")
+        expected = Series(np.arange(5), dtype=np.int64)
+        tm.assert_series_equal(res, expected)
+
+        res = ser.ffill(downcast="infer")
+        expected = Series([0, 1, 2, 2, 4], dtype=np.int64)
+        tm.assert_series_equal(res, expected)
+
+        res = ser.bfill(downcast="infer")
+        expected = Series([0, 1, 2, 4, 4], dtype=np.int64)
+        tm.assert_series_equal(res, expected)
+
+        # with a non-round float present, we will downcast to float64
+        ser[2] = 2.5
+
+        expected = Series([0, 1, 2.5, 3, 4], dtype=np.float64)
+        res = ser.fillna(3, downcast="infer")
+        tm.assert_series_equal(res, expected)
+
+        res = ser.ffill(downcast="infer")
+        expected = Series([0, 1, 2.5, 2.5, 4], dtype=np.float64)
+        tm.assert_series_equal(res, expected)
+
+        res = ser.bfill(downcast="infer")
+        expected = Series([0, 1, 2.5, 4, 4], dtype=np.float64)
+        tm.assert_series_equal(res, expected)
+
     def test_timedelta_fillna(self, frame_or_series):
         # GH#3371
         ser = Series(