ENH: consistently cast strings for DTA/TDA/PA.__setitem__ (#36261)

* ENH: consistently cast strings for DTA/TDA/PA.__setitem__

* whatsnew

Author: jbrockmendel

doc/source/whatsnew/v1.2.0.rst

@@ -230,6 +230,8 @@ Datetimelike
 - Bug in :meth:`DatetimeIndex.get_slice_bound` where ``datetime.date`` objects were not accepted or naive :class:`Timestamp` with a tz-aware :class:`DatetimeIndex` (:issue:`35690`)
 - Bug in :meth:`DatetimeIndex.slice_locs` where ``datetime.date`` objects were not accepted (:issue:`34077`)
 - Bug in :meth:`DatetimeIndex.searchsorted`, :meth:`TimedeltaIndex.searchsorted`, :meth:`PeriodIndex.searchsorted`, and :meth:`Series.searchsorted` with ``datetime64``, ``timedelta64`` or ``Period`` dtype placement of ``NaT`` values being inconsistent with ``NumPy`` (:issue:`36176`,:issue:`36254`)
+- Inconsistency in :class:`DatetimeArray`, :class:`TimedeltaArray`, and :class:`PeriodArray`  setitem casting arrays of strings to datetimelike scalars but not scalar strings (:issue:`36261`)
+-
 
 Timedelta
 ^^^^^^^^^

pandas/core/arrays/datetimelike.py

@@ -875,8 +875,7 @@ def _validate_setitem_value(self, value):
         if is_list_like(value):
             value = self._validate_listlike(value, "setitem", cast_str=True)
         else:
-            # TODO: cast_str for consistency?
-            value = self._validate_scalar(value, msg, cast_str=False)
+            value = self._validate_scalar(value, msg, cast_str=True)
 
         return self._unbox(value, setitem=True)
 

pandas/tests/arrays/test_datetimelike.py

@@ -2,6 +2,7 @@
 
 import numpy as np
 import pytest
+import pytz
 
 from pandas._libs import OutOfBoundsDatetime
 from pandas.compat.numpy import np_version_under1p18
@@ -282,15 +283,35 @@ def test_setitem(self):
         expected[:2] = expected[-2:]
         tm.assert_numpy_array_equal(arr.asi8, expected)
 
-    def test_setitem_str_array(self, arr1d):
-        if isinstance(arr1d, DatetimeArray) and arr1d.tz is not None:
-            pytest.xfail(reason="timezone comparisons inconsistent")
+    def test_setitem_strs(self, arr1d):
+        # Check that we parse strs in both scalar and listlike
+        if isinstance(arr1d, DatetimeArray):
+            tz = arr1d.tz
+            if (
+                tz is not None
+                and tz is not pytz.UTC
+                and not isinstance(tz, pytz._FixedOffset)
+            ):
+                # If we have e.g. tzutc(), when we cast to string and parse
+                #  back we get pytz.UTC, and then consider them different timezones
+                #  so incorrectly raise.
+                pytest.xfail(reason="timezone comparisons inconsistent")
+
+        # Setting list-like of strs
         expected = arr1d.copy()
         expected[[0, 1]] = arr1d[-2:]
 
-        arr1d[:2] = [str(x) for x in arr1d[-2:]]
+        result = arr1d.copy()
+        result[:2] = [str(x) for x in arr1d[-2:]]
+        tm.assert_equal(result, expected)
 
-        tm.assert_equal(arr1d, expected)
+        # Same thing but now for just a scalar str
+        expected = arr1d.copy()
+        expected[0] = arr1d[-1]
+
+        result = arr1d.copy()
+        result[0] = str(arr1d[-1])
+        tm.assert_equal(result, expected)
 
     @pytest.mark.parametrize("as_index", [True, False])
     def test_setitem_categorical(self, arr1d, as_index):

pandas/tests/arrays/test_datetimes.py

@@ -197,6 +197,29 @@ def test_tz_setter_raises(self):
         with pytest.raises(AttributeError, match="tz_localize"):
             arr.tz = "UTC"
 
+    def test_setitem_str_impute_tz(self, tz_naive_fixture):
+        # Like for getitem, if we are passed a naive-like string, we impute
+        #  our own timezone.
+        tz = tz_naive_fixture
+
+        data = np.array([1, 2, 3], dtype="M8[ns]")
+        dtype = data.dtype if tz is None else DatetimeTZDtype(tz=tz)
+        arr = DatetimeArray(data, dtype=dtype)
+        expected = arr.copy()
+
+        ts = pd.Timestamp("2020-09-08 16:50").tz_localize(tz)
+        setter = str(ts.tz_localize(None))
+
+        # Setting a scalar tznaive string
+        expected[0] = ts
+        arr[0] = setter
+        tm.assert_equal(arr, expected)
+
+        # Setting a listlike of tznaive strings
+        expected[1] = ts
+        arr[:2] = [setter, setter]
+        tm.assert_equal(arr, expected)
+
     def test_setitem_different_tz_raises(self):
         data = np.array([1, 2, 3], dtype="M8[ns]")
         arr = DatetimeArray(data, copy=False, dtype=DatetimeTZDtype(tz="US/Central"))