BUG: get_level_values() on int level upcasts to Float64Index if needed

closes pandas-dev#17924

Author: toobaz

doc/source/whatsnew/v0.21.0.txt

@@ -1061,6 +1061,7 @@ Indexing
 - Fixes ``DataFrame.loc`` for setting with alignment and tz-aware ``DatetimeIndex`` (:issue:`16889`)
 - Avoids ``IndexError`` when passing an Index or Series to ``.iloc`` with older numpy (:issue:`17193`)
 - Allow unicode empty strings as placeholders in multilevel columns in Python 2 (:issue:`17099`)
+- Bug in :meth:`MultiIndex.get_level_values` which would return an invalid index on level of ints with missing values (:issue:`17924`)
 - Bug in ``.iloc`` when used with inplace addition or assignment and an int indexer on a ``MultiIndex`` causing the wrong indexes to be read from and written to (:issue:`17148`)
 - Bug in ``.isin()`` in which checking membership in empty ``Series`` objects raised an error (:issue:`16991`)
 - Bug in ``CategoricalIndex`` reindexing in which specified indices containing duplicates were not being respected (:issue:`17323`)

pandas/core/indexes/numeric.py

@@ -11,6 +11,7 @@
     is_bool_dtype,
     is_scalar)
 from pandas.core.common import _asarray_tuplesafe, _values_from_object
+from pandas.core.dtypes.missing import isna
 
 from pandas import compat
 from pandas.core import algorithms
@@ -62,6 +63,17 @@ def _maybe_cast_slice_bound(self, label, side, kind):
         # we will try to coerce to integers
         return self._maybe_cast_indexer(label)
 
+    @Appender(_index_shared_docs['_shallow_copy'])
+    def _shallow_copy(self, values=None, **kwargs):
+        if values is not None and (not self._can_hold_na
+                                   and not is_integer_dtype(values)
+                                   and isna(values).any()):
+            attributes = self._get_attributes_dict()
+            attributes.update(kwargs)
+            return Float64Index._simple_new(values, **attributes)
+        return (super(NumericIndex, self)._shallow_copy(values=values,
+                                                        **kwargs))
+
     def _convert_for_op(self, value):
         """ Convert value to be insertable to ndarray """
 

pandas/tests/indexes/test_multi.py

@@ -997,8 +997,8 @@ def test_get_level_values(self):
         exp = CategoricalIndex([1, 2, 3, 1, 2, 3])
         tm.assert_index_equal(index.get_level_values(1), exp)
 
-    @pytest.mark.xfail(reason='GH 17924 (returns Int64Index with float data)')
     def test_get_level_values_int_with_na(self):
+        # GH 17924
         arrays = [['a', 'b', 'b'], [1, np.nan, 2]]
         index = pd.MultiIndex.from_arrays(arrays)
         result = index.get_level_values(1)
@@ -1024,14 +1024,26 @@ def test_get_level_values_na(self):
 
         arrays = [['a', 'b', 'b'], pd.DatetimeIndex([0, 1, pd.NaT])]
         index = pd.MultiIndex.from_arrays(arrays)
-        values = index.get_level_values(1)
+        result = index.get_level_values(1)
         expected = pd.DatetimeIndex([0, 1, pd.NaT])
-        tm.assert_index_equal(values, expected)
+        tm.assert_index_equal(result, expected)
 
         arrays = [[], []]
         index = pd.MultiIndex.from_arrays(arrays)
-        values = index.get_level_values(0)
-        assert values.shape == (0, )
+        result = index.get_level_values(0)
+        expected = pd.Index([], dtype=object)
+        tm.assert_index_equal(result, expected)
+
+    def test_get_level_values_all_na(self):
+        arrays = [[np.nan, np.nan, np.nan], ['a', np.nan, 1]]
+        index = pd.MultiIndex.from_arrays(arrays)
+        result = index.get_level_values(0)
+        expected = pd.Index([np.nan, np.nan, np.nan], dtype=np.float64)
+        tm.assert_index_equal(result, expected)
+
+        result = index.get_level_values(1)
+        expected = pd.Index(['a', np.nan, 1], dtype=object)
+        tm.assert_index_equal(result, expected)
 
     def test_reorder_levels(self):
         # this blows up