BUG: make Index.where behavior mirror Index.putmask behavior (#39412)

Author: jbrockmendel

Diff for: doc/source/whatsnew/v1.3.0.rst

@@ -340,6 +340,7 @@ Indexing
 - Bug in :meth:`Index.get_loc` not raising ``KeyError`` when method is specified for ``NaN`` value when ``NaN`` is not in :class:`Index` (:issue:`39382`)
 - Bug in incorrectly raising in :meth:`Index.insert`, when setting a new column that cannot be held in the existing ``frame.columns``, or in :meth:`Series.reset_index` or :meth:`DataFrame.reset_index` instead of casting to a compatible dtype (:issue:`39068`)
 - Bug in :meth:`RangeIndex.append` where a single object of length 1 was concatenated incorrectly (:issue:`39401`)
+- Bug in setting ``numpy.timedelta64`` values into an object-dtype :class:`Series` using a boolean indexer (:issue:`39488`)
 
 Missing
 ^^^^^^^
@@ -434,7 +435,9 @@ Other
 - Bug in :class:`Index` constructor sometimes silently ignorning a specified ``dtype`` (:issue:`38879`)
 - Bug in constructing a :class:`Series` from a list and a :class:`PandasDtype` (:issue:`39357`)
 - Bug in :class:`Styler` which caused CSS to duplicate on multiple renders. (:issue:`39395`)
+- :meth:`Index.where` behavior now mirrors :meth:`Index.putmask` behavior, i.e. ``index.where(mask, other)`` matches ``index.putmask(~mask, other)`` (:issue:`39412`)
 - Bug in :func:`pandas.testing.assert_series_equal`, :func:`pandas.testing.assert_frame_equal`, :func:`pandas.testing.assert_index_equal` and :func:`pandas.testing.assert_extension_array_equal` incorrectly raising when an attribute has an unrecognized NA type (:issue:`39461`)
+-
 
 .. ---------------------------------------------------------------------------
 

Diff for: pandas/core/arrays/interval.py

@@ -1413,6 +1413,16 @@ def to_tuples(self, na_tuple=True):
 
     # ---------------------------------------------------------------------
 
+    def putmask(self, mask: np.ndarray, value) -> None:
+        value_left, value_right = self._validate_setitem_value(value)
+
+        if isinstance(self._left, np.ndarray):
+            np.putmask(self._left, mask, value_left)
+            np.putmask(self._right, mask, value_right)
+        else:
+            self._left.putmask(mask, value_left)
+            self._right.putmask(mask, value_right)
+
     def delete(self: IntervalArrayT, loc) -> IntervalArrayT:
         if isinstance(self._left, np.ndarray):
             new_left = np.delete(self._left, loc)

Diff for: pandas/core/indexes/base.py

@@ -115,7 +115,7 @@
 )
 
 if TYPE_CHECKING:
-    from pandas import MultiIndex, RangeIndex, Series
+    from pandas import IntervalIndex, MultiIndex, RangeIndex, Series
     from pandas.core.indexes.datetimelike import DatetimeIndexOpsMixin
 
 
@@ -4316,27 +4316,8 @@ def where(self, cond, other=None):
         >>> idx.where(idx.isin(['car', 'train']), 'other')
         Index(['car', 'other', 'train', 'other'], dtype='object')
         """
-        if other is None:
-            other = self._na_value
-
-        values = self.values
-
-        try:
-            self._validate_fill_value(other)
-        except (ValueError, TypeError):
-            return self.astype(object).where(cond, other)
-
-        if isinstance(other, np.timedelta64) and self.dtype == object:
-            # https://github.com/numpy/numpy/issues/12550
-            #  timedelta64 will incorrectly cast to int
-            other = [other] * (~cond).sum()
-            values = cast(np.ndarray, values).copy()
-            # error: Unsupported target for indexed assignment ("ArrayLike")
-            values[~cond] = other  # type:ignore[index]
-        else:
-            values = np.where(cond, values, other)
-
-        return Index(values, name=self.name)
+        cond = np.asarray(cond, dtype=bool)
+        return self.putmask(~cond, other)
 
     # construction helpers
     @final
@@ -4551,17 +4532,32 @@ def putmask(self, mask, value):
         numpy.ndarray.putmask : Changes elements of an array
             based on conditional and input values.
         """
-        values = self._values.copy()
+        mask = np.asarray(mask, dtype=bool)
+        if mask.shape != self.shape:
+            raise ValueError("putmask: mask and data must be the same size")
+        if not mask.any():
+            return self.copy()
+
+        if value is None:
+            value = self._na_value
         try:
             converted = self._validate_fill_value(value)
         except (ValueError, TypeError) as err:
             if is_object_dtype(self):
                 raise err
 
-            # coerces to object
-            return self.astype(object).putmask(mask, value)
+            dtype = self._find_common_type_compat(value)
+            return self.astype(dtype).putmask(mask, value)
+
+        values = self._values.copy()
+        if isinstance(converted, np.timedelta64) and self.dtype == object:
+            # https://github.com/numpy/numpy/issues/12550
+            #  timedelta64 will incorrectly cast to int
+            converted = [converted] * mask.sum()
+            values[mask] = converted
+        else:
+            np.putmask(values, mask, converted)
 
-        np.putmask(values, mask, converted)
         return type(self)._simple_new(values, name=self.name)
 
     def equals(self, other: Any) -> bool:
@@ -5198,18 +5194,31 @@ def _maybe_promote(self, other: Index):
 
         return self, other
 
-    def _find_common_type_compat(self, target: Index) -> DtypeObj:
+    @final
+    def _find_common_type_compat(self, target) -> DtypeObj:
         """
         Implementation of find_common_type that adjusts for Index-specific
         special cases.
         """
-        dtype = find_common_type([self.dtype, target.dtype])
+        if is_interval_dtype(self.dtype) and is_valid_nat_for_dtype(target, self.dtype):
+            # e.g. setting NA value into IntervalArray[int64]
+            self = cast("IntervalIndex", self)
+            return IntervalDtype(np.float64, closed=self.closed)
+
+        target_dtype, _ = infer_dtype_from(target, pandas_dtype=True)
+        dtype = find_common_type([self.dtype, target_dtype])
         if dtype.kind in ["i", "u"]:
             # TODO: what about reversed with self being categorical?
-            if is_categorical_dtype(target.dtype) and target.hasnans:
+            if (
+                isinstance(target, Index)
+                and is_categorical_dtype(target.dtype)
+                and target.hasnans
+            ):
                 # FIXME: find_common_type incorrect with Categorical GH#38240
                 # FIXME: some cases where float64 cast can be lossy?
                 dtype = np.dtype(np.float64)
+        if dtype.kind == "c":
+            dtype = np.dtype(object)
         return dtype
 
     @final

Diff for: pandas/core/indexes/interval.py

@@ -799,29 +799,22 @@ def length(self):
         return Index(self._data.length, copy=False)
 
     def putmask(self, mask, value):
-        arr = self._data.copy()
+        mask = np.asarray(mask, dtype=bool)
+        if mask.shape != self.shape:
+            raise ValueError("putmask: mask and data must be the same size")
+        if not mask.any():
+            return self.copy()
+
         try:
-            value_left, value_right = arr._validate_setitem_value(value)
+            self._validate_fill_value(value)
         except (ValueError, TypeError):
-            return self.astype(object).putmask(mask, value)
+            dtype = self._find_common_type_compat(value)
+            return self.astype(dtype).putmask(mask, value)
 
-        if isinstance(self._data._left, np.ndarray):
-            np.putmask(arr._left, mask, value_left)
-            np.putmask(arr._right, mask, value_right)
-        else:
-            # TODO: special case not needed with __array_function__
-            arr._left.putmask(mask, value_left)
-            arr._right.putmask(mask, value_right)
+        arr = self._data.copy()
+        arr.putmask(mask, value)
         return type(self)._simple_new(arr, name=self.name)
 
-    @Appender(Index.where.__doc__)
-    def where(self, cond, other=None):
-        if other is None:
-            other = self._na_value
-        values = np.where(cond, self._values, other)
-        result = IntervalArray(values)
-        return type(self)._simple_new(result, name=self.name)
-
     def insert(self, loc, item):
         """
         Return a new IntervalIndex inserting new item at location. Follows
@@ -998,6 +991,9 @@ def func(self, other, sort=sort):
 
     # --------------------------------------------------------------------
 
+    def _validate_fill_value(self, value):
+        return self._data._validate_setitem_value(value)
+
     @property
     def _is_all_dates(self) -> bool:
         """

Diff for: pandas/tests/series/indexing/test_setitem.py

@@ -3,8 +3,6 @@
 import numpy as np
 import pytest
 
-from pandas.compat import np_version_under1p20
-
 from pandas import (
     DatetimeIndex,
     Index,
@@ -380,16 +378,14 @@ def test_index_where(self, obj, key, expected, val, request):
         mask = np.zeros(obj.shape, dtype=bool)
         mask[key] = True
 
-        if obj.dtype == bool and not mask.all():
-            # When mask is all True, casting behavior does not apply
+        if obj.dtype == bool:
             msg = "Index/Series casting behavior inconsistent GH#38692"
             mark = pytest.mark.xfail(reason=msg)
             request.node.add_marker(mark)
 
         res = Index(obj).where(~mask, val)
         tm.assert_index_equal(res, Index(expected))
 
-    @pytest.mark.xfail(reason="Index/Series casting behavior inconsistent GH#38692")
     def test_index_putmask(self, obj, key, expected, val):
         if Index(obj).dtype != obj.dtype:
             pytest.skip("test not applicable for this dtype")
@@ -629,10 +625,6 @@ def test_index_where(self, obj, key, expected, val, request):
         res = Index(obj).where(~mask, val)
         tm.assert_index_equal(res, Index(expected))
 
-    @pytest.mark.xfail(
-        np_version_under1p20,
-        reason="Index/Series casting behavior inconsistent GH#38692",
-    )
     def test_index_putmask(self, obj, key, expected, val):
         if Index(obj).dtype != obj.dtype:
             pytest.skip("test not applicable for this dtype")