Fix regression on datetime in MultiIndex

pandas/core/indexing.py

@@ -1165,6 +1165,9 @@ def _convert_to_indexer(self, key, axis: int, is_setter: bool = False):
                     if len(key) == labels.nlevels:
                         return {"key": key}
                     raise
+            except InvalidIndexError:
+                if not isinstance(labels, ABCMultiIndex):
+                    raise
             except TypeError:
                 pass
             except ValueError:

pandas/tests/indexing/multiindex/test_datetime.py

@@ -2,7 +2,16 @@
 
 import numpy as np
 
-from pandas import Index, Period, Series, period_range
+from pandas import (
+    DataFrame,
+    DatetimeIndex,
+    Index,
+    MultiIndex,
+    Period,
+    Series,
+    period_range,
+    to_datetime,
+)
 
 
 def test_multiindex_period_datetime():
@@ -20,3 +29,16 @@ def test_multiindex_period_datetime():
     # try datetime as index
     result = s.loc["a", datetime(2012, 1, 1)]
     assert result == expected
+
+
+def test_multiindex_datetime_columns():
+    # GH35015, using datetime as column indices raises exception
+
+    mi = MultiIndex.from_tuples(
+        [(to_datetime("02/29/2020"), to_datetime("03/01/2020"))], names=["a", "b"]
+    )
+
+    df = DataFrame([], columns=mi)
+    assert list(df.columns.names) == ["a", "b"]
+
+    assert all(isinstance(mi.get_level_values(i), DatetimeIndex) for i in range(2))