BUG: DataFrame[dt64].where downcasting (pandas-dev#45837)

* BUG: DataFrame[dt64].where downcasting

* whatsnew

* GH ref

* punt on dt64

* mypy fixup

Author: jbrockmendel

doc/source/whatsnew/v1.5.0.rst

@@ -318,6 +318,7 @@ Indexing
 - Bug in setting a NA value (``None`` or ``np.nan``) into a :class:`Series` with int-based :class:`IntervalDtype` incorrectly casting to object dtype instead of a float-based :class:`IntervalDtype` (:issue:`45568`)
 - Bug in :meth:`Series.__setitem__` with a non-integer :class:`Index` when using an integer key to set a value that cannot be set inplace where a ``ValueError`` was raised instead of casting to a common dtype (:issue:`45070`)
 - Bug in :meth:`Series.__setitem__` when setting incompatible values into a ``PeriodDtype`` or ``IntervalDtype`` :class:`Series` raising when indexing with a boolean mask but coercing when indexing with otherwise-equivalent indexers; these now consistently coerce, along with :meth:`Series.mask` and :meth:`Series.where` (:issue:`45768`)
+- Bug in :meth:`DataFrame.where` with multiple columns with datetime-like dtypes failing to downcast results consistent with other dtypes (:issue:`45837`)
 - Bug in :meth:`Series.loc.__setitem__` and :meth:`Series.loc.__getitem__` not raising when using multiple keys without using a :class:`MultiIndex` (:issue:`13831`)
 - Bug when setting a value too large for a :class:`Series` dtype failing to coerce to a common type (:issue:`26049`, :issue:`32878`)
 - Bug in :meth:`loc.__setitem__` treating ``range`` keys as positional instead of label-based (:issue:`45479`)

pandas/core/arrays/boolean.py

@@ -214,7 +214,7 @@ def coerce_to_array(
             raise TypeError("Need to pass bool-like values")
 
     if mask is None and mask_values is None:
-        mask = np.zeros(len(values), dtype=bool)
+        mask = np.zeros(values.shape, dtype=bool)
     elif mask is None:
         mask = mask_values
     else:

pandas/core/dtypes/cast.py

@@ -262,7 +262,7 @@ def maybe_downcast_to_dtype(result: ArrayLike, dtype: str | np.dtype) -> ArrayLi
                 dtype = "int64"
             elif inferred_type == "datetime64":
                 dtype = "datetime64[ns]"
-            elif inferred_type == "timedelta64":
+            elif inferred_type in ["timedelta", "timedelta64"]:
                 dtype = "timedelta64[ns]"
 
             # try to upcast here
@@ -290,6 +290,14 @@ def maybe_downcast_to_dtype(result: ArrayLike, dtype: str | np.dtype) -> ArrayLi
     if dtype.kind in ["M", "m"] and result.dtype.kind in ["i", "f"]:
         result = result.astype(dtype)
 
+    elif dtype.kind == "m" and result.dtype == _dtype_obj:
+        # test_where_downcast_to_td64
+        result = cast(np.ndarray, result)
+        result = array_to_timedelta64(result)
+
+    elif dtype == "M8[ns]" and result.dtype == _dtype_obj:
+        return np.asarray(maybe_cast_to_datetime(result, dtype=dtype))
+
     return result
 
 

pandas/core/internals/blocks.py

@@ -1426,21 +1426,40 @@ def where(self, other, cond, _downcast="infer") -> list[Block]:
         except (ValueError, TypeError) as err:
             _catch_deprecated_value_error(err)
 
-            if is_interval_dtype(self.dtype):
-                # TestSetitemFloatIntervalWithIntIntervalValues
-                blk = self.coerce_to_target_dtype(orig_other)
-                nbs = blk.where(orig_other, orig_cond)
-                return self._maybe_downcast(nbs, downcast=_downcast)
+            if self.ndim == 1 or self.shape[0] == 1:
 
-            elif isinstance(self, NDArrayBackedExtensionBlock):
-                # NB: not (yet) the same as
-                #  isinstance(values, NDArrayBackedExtensionArray)
-                blk = self.coerce_to_target_dtype(orig_other)
-                nbs = blk.where(orig_other, orig_cond)
-                return self._maybe_downcast(nbs, "infer")
+                if is_interval_dtype(self.dtype):
+                    # TestSetitemFloatIntervalWithIntIntervalValues
+                    blk = self.coerce_to_target_dtype(orig_other)
+                    nbs = blk.where(orig_other, orig_cond)
+                    return self._maybe_downcast(nbs, downcast=_downcast)
+
+                elif isinstance(self, NDArrayBackedExtensionBlock):
+                    # NB: not (yet) the same as
+                    #  isinstance(values, NDArrayBackedExtensionArray)
+                    blk = self.coerce_to_target_dtype(orig_other)
+                    nbs = blk.where(orig_other, orig_cond)
+                    return self._maybe_downcast(nbs, downcast=_downcast)
+
+                else:
+                    raise
 
             else:
-                raise
+                # Same pattern we use in Block.putmask
+                is_array = isinstance(orig_other, (np.ndarray, ExtensionArray))
+
+                res_blocks = []
+                nbs = self._split()
+                for i, nb in enumerate(nbs):
+                    n = orig_other
+                    if is_array:
+                        # we have a different value per-column
+                        n = orig_other[:, i : i + 1]
+
+                    submask = orig_cond[:, i : i + 1]
+                    rbs = nb.where(n, submask)
+                    res_blocks.extend(rbs)
+                return res_blocks
 
         nb = self.make_block_same_class(res_values)
         return [nb]

pandas/tests/arrays/boolean/test_construction.py

@@ -183,8 +183,11 @@ def test_coerce_to_array():
     values = np.array([True, False, True, False], dtype="bool")
     mask = np.array([False, False, False, True], dtype="bool")
 
+    # passing 2D values is OK as long as no mask
+    coerce_to_array(values.reshape(1, -1))
+
     with pytest.raises(ValueError, match="values.shape and mask.shape must match"):
-        coerce_to_array(values.reshape(1, -1))
+        coerce_to_array(values.reshape(1, -1), mask=mask)
 
     with pytest.raises(ValueError, match="values.shape and mask.shape must match"):
         coerce_to_array(values, mask=mask.reshape(1, -1))

pandas/tests/dtypes/cast/test_downcast.py

@@ -5,7 +5,10 @@
 
 from pandas.core.dtypes.cast import maybe_downcast_to_dtype
 
-from pandas import Series
+from pandas import (
+    Series,
+    Timedelta,
+)
 import pandas._testing as tm
 
 
@@ -34,6 +37,13 @@
             "int64",
             np.array([decimal.Decimal(0.0)]),
         ),
+        (
+            # GH#45837
+            np.array([Timedelta(days=1), Timedelta(days=2)], dtype=object),
+            "infer",
+            np.array([1, 2], dtype="m8[D]").astype("m8[ns]"),
+        ),
+        # TODO: similar for dt64, dt64tz, Period, Interval?
     ],
 )
 def test_downcast(arr, expected, dtype):

pandas/tests/frame/indexing/test_where.py

@@ -965,3 +965,58 @@ def test_where_inplace_casting(data):
     df_copy = df.where(pd.notnull(df), None).copy()
     df.where(pd.notnull(df), None, inplace=True)
     tm.assert_equal(df, df_copy)
+
+
+def test_where_downcast_to_td64():
+    ser = Series([1, 2, 3])
+
+    mask = np.array([False, False, False])
+
+    td = pd.Timedelta(days=1)
+
+    res = ser.where(mask, td)
+    expected = Series([td, td, td], dtype="m8[ns]")
+    tm.assert_series_equal(res, expected)
+
+
+def _check_where_equivalences(df, mask, other, expected):
+    # similar to tests.series.indexing.test_setitem.SetitemCastingEquivalences
+    #  but with DataFrame in mind and less fleshed-out
+    res = df.where(mask, other)
+    tm.assert_frame_equal(res, expected)
+
+    res = df.mask(~mask, other)
+    tm.assert_frame_equal(res, expected)
+
+    # Note: we cannot do the same with frame.mask(~mask, other, inplace=True)
+    #  bc that goes through Block.putmask which does *not* downcast.
+
+
+def test_where_dt64_2d():
+    dti = date_range("2016-01-01", periods=6)
+    dta = dti._data.reshape(3, 2)
+    other = dta - dta[0, 0]
+
+    df = DataFrame(dta, columns=["A", "B"])
+
+    mask = np.asarray(df.isna())
+    mask[:, 1] = True
+
+    # setting all of one column, none of the other
+    expected = DataFrame({"A": other[:, 0], "B": dta[:, 1]})
+    _check_where_equivalences(df, mask, other, expected)
+
+    # setting part of one column, none of the other
+    mask[1, 0] = True
+    expected = DataFrame(
+        {
+            "A": np.array([other[0, 0], dta[1, 0], other[2, 0]], dtype=object),
+            "B": dta[:, 1],
+        }
+    )
+    _check_where_equivalences(df, mask, other, expected)
+
+    # setting nothing in either column
+    mask[:] = True
+    expected = df
+    _check_where_equivalences(df, mask, other, expected)