Rework quantile

Author: rhshadrach

pandas/core/frame.py

@@ -83,6 +83,7 @@
     can_hold_element,
     construct_1d_arraylike_from_scalar,
     construct_2d_arraylike_from_scalar,
+    ensure_dtype_can_hold_na,
     find_common_type,
     infer_dtype_from_scalar,
     invalidate_string_dtypes,
@@ -13046,6 +13047,7 @@ def quantile(
         Name: 0.5, dtype: object
         """
         from pandas.core.dtypes.common import is_object_dtype
+
         validate_percentile(q)
         axis = self._get_axis_number(axis)
 
@@ -13075,6 +13077,10 @@ def quantile(
 
         if axis == 1:
             data = data.T
+            if data.shape[0] == 0:
+                # The transpose has no rows, so the original has no columns, meaning we
+                # have no dtype information. Since this is quantile, default to float64
+                data = data.astype("float64")
 
         if len(data.columns) == 0:
             # GH#23925 _get_numeric_data may have dropped all columns
@@ -13098,14 +13104,17 @@ def quantile(
 
         # handle degenerate case
         if len(data) == 0:
-            dtype = np.float64
-            if data.ndim == 2:
-                cdtype = find_common_type(list(self.dtypes))
-            else:
-                cdtype = self.dtype
-            if needs_i8_conversion(cdtype) or is_object_dtype(cdtype):
-                dtype = cdtype
-            return self._constructor([], index=q, columns=data.columns, dtype=dtype)
+            from pandas import array
+
+            result = self._constructor(
+                {
+                    idx: array(len(q) * [np.nan], dtype=ensure_dtype_can_hold_na(dtype))
+                    for idx, dtype in enumerate(data.dtypes)
+                },
+                index=q,
+            )
+            result.columns = data.columns
+            return result
 
         if method == "single":
             res = data._mgr.quantile(qs=q, interpolation=interpolation)

pandas/tests/frame/methods/test_quantile.py

@@ -319,12 +319,10 @@ def test_quantile_multi_empty(self, interp_method):
         result = DataFrame({"x": [], "y": []}).quantile(
             [0.1, 0.9], axis=0, interpolation=interpolation, method=method
         )
-        # dtype = "float64" if method == "single" else "object"
-        dtype = "object"
         expected = DataFrame(
             {"x": [np.nan, np.nan], "y": [np.nan, np.nan]},
             index=[0.1, 0.9],
-            dtype=dtype,
+            dtype="object",
         )
         tm.assert_frame_equal(result, expected)
 
@@ -692,10 +690,8 @@ def test_quantile_empty_no_rows_dt64(self, interp_method):
         res = df.quantile(
             0.5, numeric_only=False, interpolation=interpolation, method=method
         )
-        exp = exp.astype(object)
-        if True or interpolation == "nearest":
-            # GH#18463 TODO: would we prefer NaTs here?
-            exp = exp.fillna(np.nan)
+        # GH#18463 TODO: would we prefer NaTs here?
+        exp = exp.astype(object).fillna(pd.NaT)
         tm.assert_series_equal(res, exp)
 
         # both dt64tz
@@ -912,7 +908,7 @@ def test_empty_datelike(
     @pytest.mark.parametrize(
         "expected_data, expected_index, axis",
         [
-            [[pd.NaT, pd.NaT], range(2), 1],
+            [[np.nan, np.nan], range(2), 1],
             [[], [], 0],
         ],
     )
@@ -927,6 +923,6 @@ def test_datelike_numeric_only(self, expected_data, expected_index, axis):
         )
         result = df[["a", "c"]].quantile(0.5, axis=axis, numeric_only=True)
         expected = Series(
-            expected_data, name=0.5, index=Index(expected_index)
+            expected_data, name=0.5, index=Index(expected_index), dtype=np.float64
         )
         tm.assert_series_equal(result, expected)