BUG: DataFrame reductions with object dtype and axis=1

doc/source/whatsnew/v2.0.0.rst

@@ -793,7 +793,7 @@ Timezones
 Numeric
 ^^^^^^^
 - Bug in :meth:`DataFrame.add` cannot apply ufunc when inputs contain mixed DataFrame type and Series type (:issue:`39853`)
-- Bug in DataFrame reduction methods (e.g. :meth:`DataFrame.sum`) with object dtype, ``axis=1`` and ``numeric_only=False`` would not be coerced to float (:issue:`49551`)
+- Bug in DataFrame reduction methods (e.g. :meth:`DataFrame.sum`) with object dtype, ``axis=1`` and ``numeric_only=False`` would have results unnecessarily coerced to float; coercion still occurs for reductions that necessarily result in floats (``mean``, ``var``, ``std``, ``skew``) (:issue:`49603`)
 - Bug in :meth:`DataFrame.sem` and :meth:`Series.sem` where an erroneous ``TypeError`` would always raise when using data backed by an :class:`ArrowDtype` (:issue:`49759`)
 
 Conversion

pandas/core/frame.py

@@ -10493,10 +10493,6 @@ def _reduce(
         axis = self._get_axis_number(axis)
         assert axis in [0, 1]
 
-        def func(values: np.ndarray):
-            # We only use this in the case that operates on self.values
-            return op(values, axis=axis, skipna=skipna, **kwds)
-
         def blk_func(values, axis: Axis = 1):
             if isinstance(values, ExtensionArray):
                 if not is_1d_only_ea_dtype(values.dtype) and not isinstance(
@@ -10516,51 +10512,40 @@ def _get_data() -> DataFrame:
                 data = self._get_bool_data()
             return data
 
-        if numeric_only or axis == 0:
-            # For numeric_only non-None and axis non-None, we know
-            #  which blocks to use and no try/except is needed.
-            #  For numeric_only=None only the case with axis==0 and no object
-            #  dtypes are unambiguous can be handled with BlockManager.reduce
-            # Case with EAs see GH#35881
-            df = self
-            if numeric_only:
-                df = _get_data()
-            if axis == 1:
-                df = df.T
-                axis = 0
-
-            # After possibly _get_data and transposing, we are now in the
-            #  simple case where we can use BlockManager.reduce
-            res = df._mgr.reduce(blk_func)
-            out = df._constructor(res).iloc[0]
-            if out_dtype is not None:
-                out = out.astype(out_dtype)
-            if axis == 0 and len(self) == 0 and name in ["sum", "prod"]:
-                # Even if we are object dtype, follow numpy and return
-                #  float64, see test_apply_funcs_over_empty
-                out = out.astype(np.float64)
-
-            return out
-
-        assert not numeric_only and axis == 1
-
-        data = self
-        values = data.values
-        result = func(values)
-
-        if hasattr(result, "dtype"):
-            if filter_type == "bool" and notna(result).all():
-                result = result.astype(np.bool_)
-            elif filter_type is None and is_object_dtype(result.dtype):
-                try:
-                    result = result.astype(np.float64)
-                except (ValueError, TypeError):
-                    # try to coerce to the original dtypes item by item if we can
-                    pass
+        # Case with EAs see GH#35881
+        df = self
+        if numeric_only:
+            df = _get_data()
+        if axis == 1:
+            if len(df.index) == 0:
+                # Taking a transpose would result in no columns, losing the dtype.
+                # In the empty case, reducing along axis 0 or 1 gives the same
+                # result dtype, so reduce with axis=0 and ignore values
+                result = df._reduce(
+                    op,
+                    name,
+                    axis=0,
+                    skipna=skipna,
+                    numeric_only=False,
+                    filter_type=filter_type,
+                    **kwds,
+                ).iloc[:0]
+                result.index = df.index
+                return result
+            df = df.T
+
+        # After possibly _get_data and transposing, we are now in the
+        #  simple case where we can use BlockManager.reduce
+        res = df._mgr.reduce(blk_func)
+        out = df._constructor(res).iloc[0]
+        if out_dtype is not None:
+            out = out.astype(out_dtype)
+        if len(self) == 0 and name in ["sum", "prod"]:
+            # Even if we are object dtype, follow numpy and return
+            #  float64, see test_apply_funcs_over_empty
+            out = out.astype(np.float64)
 
-        labels = self._get_agg_axis(axis)
-        result = self._constructor_sliced(result, index=labels)
-        return result
+        return out
 
     def _reduce_axis1(self, name: str, func, skipna: bool) -> Series:
         """

pandas/core/internals/array_manager.py

@@ -986,8 +986,9 @@ def reduce(self: T, func: Callable) -> T:
             else:
                 # error: Argument 1 to "append" of "list" has incompatible type
                 # "ExtensionArray"; expected "ndarray"
+                dtype = arr.dtype if res is NaT else None
                 result_arrays.append(
-                    sanitize_array([res], None)  # type: ignore[arg-type]
+                    sanitize_array([res], None, dtype=dtype)  # type: ignore[arg-type]
                 )
 
         index = Index._simple_new(np.array([None], dtype=object))  # placeholder

pandas/core/nanops.py

@@ -622,7 +622,7 @@ def nansum(
     3.0
     """
     values, mask, dtype, dtype_max, _ = _get_values(
-        values, skipna, fill_value=0, mask=mask
+        values, skipna, fill_value=0.0, mask=mask
     )
     dtype_sum = dtype_max
     if is_float_dtype(dtype):
@@ -1389,7 +1389,7 @@ def nanprod(
 
     if skipna and mask is not None:
         values = values.copy()
-        values[mask] = 1
+        values[mask] = 1.0
     result = values.prod(axis)
     # error: Incompatible return value type (got "Union[ndarray, float]", expected
     # "float")
@@ -1500,7 +1500,15 @@ def _maybe_null_out(
                 result[null_mask] = None
     elif result is not NaT:
         if check_below_min_count(shape, mask, min_count):
-            result = np.nan
+            result_dtype = getattr(result, "dtype", None)
+            if is_float_dtype(result_dtype):
+                # Preserve dtype when possible
+                # mypy doesn't infer result_dtype is not None
+                result = getattr(
+                    np, f"float{8 * result_dtype.itemsize}"  # type: ignore[union-attr]
+                )("nan")
+            else:
+                result = np.nan
 
     return result
 

pandas/tests/apply/test_frame_apply.py

@@ -1432,7 +1432,7 @@ def test_apply_datetime_tz_issue():
 def test_mixed_column_raises(df, method):
     # GH 16832
     if method == "sum":
-        msg = r'can only concatenate str \(not "int"\) to str'
+        msg = r'can only concatenate str \(not "float"\) to str'
     else:
         msg = "not supported between instances of 'str' and 'float'"
     with pytest.raises(TypeError, match=msg):

pandas/tests/frame/test_reductions.py

@@ -318,11 +318,26 @@ def wrapper(x):
             DataFrame({0: [np.nan, 2], 1: [np.nan, 3], 2: [np.nan, 4]}, dtype=object),
         ],
     )
-    def test_stat_operators_attempt_obj_array(self, method, df):
+    def test_stat_operators_attempt_obj_array(
+        self, method, df, using_array_manager, axis
+    ):
         # GH#676
         assert df.values.dtype == np.object_
-        result = getattr(df, method)(1)
-        expected = getattr(df.astype("f8"), method)(1)
+        result = getattr(df, method)(axis=axis)
+        expected = getattr(df.astype("f8"), method)(axis=axis)
+        # With values an np.array with dtype object:
+        #   - When using blocks, `values.sum(axis=1, ...)` returns a np.array of dim 1
+        #     and this remains object dtype
+        #   - When using arrays, `values.sum(axis=0, ...)` returns a Python float
+        if not using_array_manager and method in ("sum", "prod", "min", "max"):
+            expected = expected.astype(object)
+        elif (
+            using_array_manager
+            and axis in (0, "index")
+            and method in ("min", "max")
+            and 0 in df.columns
+        ):
+            expected = expected.astype("int64")
         tm.assert_series_equal(result, expected)
 
     @pytest.mark.parametrize("op", ["mean", "std", "var", "skew", "kurt", "sem"])
@@ -695,6 +710,42 @@ def test_sum_corner(self):
         assert len(axis0) == 0
         assert len(axis1) == 0
 
+    @pytest.mark.parametrize(
+        "index",
+        [
+            tm.makeRangeIndex(0),
+            tm.makeDateIndex(0),
+            tm.makeNumericIndex(0, dtype=int),
+            tm.makeNumericIndex(0, dtype=float),
+            tm.makeDateIndex(0, freq="M"),
+            tm.makePeriodIndex(0),
+        ],
+    )
+    def test_axis_1_empty(self, all_reductions, index, using_array_manager):
+        df = DataFrame(columns=["a"], index=index)
+        result = getattr(df, all_reductions)(axis=1)
+        expected_dtype = {
+            "any": "bool",
+            "all": "bool",
+            "count": "int64",
+            "sum": "float",
+            "prod": "float",
+            "skew": "float",
+            "kurt": "float",
+            "sem": "float",
+        }.get(all_reductions, "object")
+        if using_array_manager and all_reductions in (
+            "max",
+            "min",
+            "mean",
+            "std",
+            "var",
+            "median",
+        ):
+            expected_dtype = "float"
+        expected = Series([], index=index, dtype=expected_dtype)
+        tm.assert_series_equal(result, expected)
+
     @pytest.mark.parametrize("method, unit", [("sum", 0), ("prod", 1)])
     @pytest.mark.parametrize("numeric_only", [None, True, False])
     def test_sum_prod_nanops(self, method, unit, numeric_only):
@@ -1355,13 +1406,11 @@ def test_min_max_dt64_with_NaT_skipna_false(self, request, tz_naive_fixture):
         res = df.min(axis=1, skipna=False)
         expected = Series([df.loc[0, "a"], pd.NaT])
         assert expected.dtype == df["a"].dtype
-
         tm.assert_series_equal(res, expected)
 
         res = df.max(axis=1, skipna=False)
         expected = Series([df.loc[0, "b"], pd.NaT])
         assert expected.dtype == df["a"].dtype
-
         tm.assert_series_equal(res, expected)
 
     def test_min_max_dt64_api_consistency_with_NaT(self):