Fix and test casting tz-aware datetimeindex to object-dtype ndarray or Index

Author: jbrockmendel

doc/source/whatsnew/v0.24.0.txt

@@ -1357,3 +1357,5 @@ Other
 - :meth:`~pandas.io.formats.style.Styler.bar` now also supports tablewise application (in addition to rowwise and columnwise) with ``axis=None`` and setting clipping range with ``vmin`` and ``vmax`` (:issue:`21548` and :issue:`21526`). ``NaN`` values are also handled properly.
 - Logical operations ``&, |, ^`` between :class:`Series` and :class:`Index` will no longer raise ``ValueError`` (:issue:`22092`)
 - Bug in :meth:`DataFrame.combine_first` in which column types were unexpectedly converted to float (:issue:`20699`)
+- Bug in :class:`Index` where passing a timezone-aware :class:`DatetimeIndex` and `dtype=object` would incorrectly raise a ``ValueError`` (:issue:`????`)
+- Bug in :class:`DatetimeIndex` where calling `np.array(dtindex, dtype=object)` would incorrectly return an array of ``long`` objects (:issue:`????`)

pandas/core/arrays/datetimes.py

@@ -374,6 +374,16 @@ def _resolution(self):
     # ----------------------------------------------------------------
     # Array-like Methods
 
+    def __array__(self, dtype=None):
+        if is_object_dtype(dtype):
+            return np.array(list(self), dtype=object)
+        elif dtype == 'i8':
+            return self.asi8
+
+        # TODO: warn that dtype is not used?
+        #  warn that conversion may be lossy?
+        return self._data.view(np.ndarray)  # follow Index.__array__
+
     def __iter__(self):
         """
         Return an iterator over the boxed values

pandas/core/indexes/base.py

@@ -301,11 +301,15 @@ def __new__(cls, data=None, dtype=None, copy=False, name=None,
               (dtype is not None and is_datetime64_any_dtype(dtype)) or
                 'tz' in kwargs):
             from pandas import DatetimeIndex
-            result = DatetimeIndex(data, copy=copy, name=name,
-                                   dtype=dtype, **kwargs)
+
             if dtype is not None and is_dtype_equal(_o_dtype, dtype):
-                return Index(result.to_pydatetime(), dtype=_o_dtype)
+                # if `data` is already a TZaware DatetimeIndex, then passing
+                #  dtype=object to the constructor will raise spuriously
+                result = DatetimeIndex(data, copy=copy, name=name, **kwargs)
+                return Index(list(result), dtype=_o_dtype)
             else:
+                result = DatetimeIndex(data, copy=copy, name=name,
+                                       dtype=dtype, **kwargs)
                 return result
 
         elif (is_timedelta64_dtype(data) or

pandas/tests/arrays/test_datetimelike.py

@@ -57,6 +57,29 @@ def timedelta_index(request):
 
 class TestDatetimeArray(object):
 
+    def test_array_object_dtype(self, tz_naive_fixture):
+        tz = tz_naive_fixture
+        dti = pd.date_range('2016-01-01', periods=3, tz=tz)
+        arr = DatetimeArrayMixin(dti)
+
+        expected = np.array(list(dti))
+
+        result = np.array(arr, dtype=object)
+        tm.assert_numpy_array_equal(result, expected)
+
+        # also test the DatetimeIndex method while we're at it
+        result = np.array(dti, dtype=object)
+        tm.assert_numpy_array_equal(result, expected)
+
+    def test_array(self, tz_naive_fixture):
+        tz = tz_naive_fixture
+        dti = pd.date_range('2016-01-01', periods=3, tz=tz)
+        arr = DatetimeArrayMixin(dti)
+
+        expected = dti.asi8.view('M8[ns]')
+        result = np.array(arr)
+        tm.assert_numpy_array_equal(result, expected)
+
     def test_from_dti(self, tz_naive_fixture):
         tz = tz_naive_fixture
         dti = pd.date_range('2016-01-01', periods=3, tz=tz)

pandas/tests/indexes/test_base.py

@@ -132,7 +132,7 @@ def test_construction_list_tuples_nan(self, na_value, vtype):
     @pytest.mark.parametrize("cast_as_obj", [True, False])
     @pytest.mark.parametrize("index", [
         pd.date_range('2015-01-01 10:00', freq='D', periods=3,
-                      tz='US/Eastern'),  # DTI with tz
+                      tz='US/Eastern', name='Green Eggs & Ham'),  # DTI with tz
         pd.date_range('2015-01-01 10:00', freq='D', periods=3),  # DTI no tz
         pd.timedelta_range('1 days', freq='D', periods=3),  # td
         pd.period_range('2015-01-01', freq='D', periods=3)  # period
@@ -145,8 +145,16 @@ def test_constructor_from_index_dtlike(self, cast_as_obj, index):
 
         tm.assert_index_equal(result, index)
 
-        if isinstance(index, pd.DatetimeIndex) and hasattr(index, 'tz'):
+        if isinstance(index, pd.DatetimeIndex):
             assert result.tz == index.tz
+            if cast_as_obj:
+                # GH#???? check that Index(dti, dtype=object) does not
+                #  incorrectly raise ValueError, and that nanoseconds are not
+                #  dropped
+                index += pd.Timedelta(nanoseconds=50)
+                result = pd.Index(index, dtype=object)
+                assert result.dtype == np.object_
+                assert list(result) == list(index)
 
     @pytest.mark.parametrize("index,has_tz", [
         (pd.date_range('2015-01-01 10:00', freq='D', periods=3,