BUG: fix reindexing MultiIndex with categorical datetime-like level

pandas/core/indexes/multi.py

@@ -819,11 +819,17 @@ def values(self):
  return self._tuples
 
  values = []
- for lev, lab in zip(self.levels, self.labels):
+ for i, (lev, lab) in enumerate(zip(self.levels, self.labels)):
  # Need to box timestamps, etc.
  box = hasattr(lev, '_box_values')
+ if is_categorical_dtype(lev):
+ # TODO GH-21390
+ taken = self.get_level_values(i)._values.get_values()
+ if not isinstance(taken, np.ndarray):
+ # Datetime/Period index
+ taken = np.array(taken.astype(object))
  # Try to minimize boxing.
- if box and len(lev) > len(lab):
+ elif box and len(lev) > len(lab):
  taken = lev._box_values(algos.take_1d(lev._ndarray_values,
  lab))
  elif box:

pandas/tests/groupby/test_categorical.py

@@ -854,3 +854,23 @@ def test_empty_prod():
  result = df.groupby("A", observed=False).B.prod(min_count=1)
  expected = pd.Series([2, 1, np.nan], expected_idx, name='B')
  tm.assert_series_equal(result, expected)
+
+
+def test_groupby_multiindex_categorical_datetime():
+ # https://github.com/pandas-dev/pandas/issues/21390
+
+ df = pd.DataFrame({
+ 'key1': pd.Categorical(list('abcbabcba')),
+ 'key2': pd.Categorical(
+ list(pd.date_range('2018-06-01 00', freq='1T', periods=3)) * 3),
+ 'values': np.arange(9),
+ })
+ result = df.groupby(['key1', 'key2']).mean()
+
+ idx = pd.MultiIndex.from_product(
+ [pd.Categorical(['a', 'b', 'c']),
+ pd.Categorical(pd.date_range('2018-06-01 00', freq='1T', periods=3))],
+ names=['key1', 'key2'])
+ expected = pd.DataFrame(
+ {'values': [0, 4, 8, 3, 4, 5, 6, np.nan, 2]}, index=idx)
+ assert_frame_equal(result, expected)