Fix for #16836

* Removed unnecessary conversion to i8
* Fixed failed test (`test_frame_column_inplace_sort_exception`)
* Added check to ensure that the test is performing its intended goal(`test_sort_nan`)

Author: Jean-Mathieu Deschenes

doc/source/whatsnew/v0.21.0.txt

@@ -144,6 +144,7 @@ Bug Fixes
 ~~~~~~~~~
 
 - Fixes regression in 0.20, :func:`Series.aggregate` and :func:`DataFrame.aggregate` allow dictionaries as return values again (:issue:`16741`)
+- Fixes regression when sorting by multiple columns on a datetime array with NaT values (:issue:`16836`)
 
 Conversion
 ^^^^^^^^^^

pandas/core/frame.py

@@ -3303,18 +3303,13 @@ def sort_values(self, by, axis=0, ascending=True, inplace=False,
         if len(by) > 1:
             from pandas.core.sorting import lexsort_indexer
 
-            def trans(v):
-                if needs_i8_conversion(v):
-                    return v.view('i8')
-                return v
-
             keys = []
             for x in by:
                 k = self.xs(x, axis=other_axis).values
                 if k.ndim == 2:
                     raise ValueError('Cannot sort by duplicate column %s' %
                                      str(x))
-                keys.append(trans(k))
+                keys.append(k)
             indexer = lexsort_indexer(keys, orders=ascending,
                                       na_position=na_position)
             indexer = _ensure_platform_int(indexer)

pandas/tests/frame/test_sorting.py

@@ -89,6 +89,22 @@ def test_sort_values(self):
         with tm.assert_raises_regex(ValueError, msg):
             frame.sort_values(by=['A', 'B'], axis=0, ascending=[True] * 5)
 
+        # GH 16836
+
+        d1 = [Timestamp(x) for x in ['2016-01-01', '2015-01-01',
+                                     np.nan, '2016-01-01']]
+        d2 = [Timestamp(x) for x in ['2017-01-01', '2014-01-01',
+                                     '2016-01-01', '2015-01-01']]
+        df = pd.DataFrame({'a': d1, 'b': d2}, index=[0, 1, 2, 3])
+
+        d3 = [Timestamp(x) for x in ['2015-01-01', '2016-01-01',
+                                     '2016-01-01', np.nan]]
+        d4 = [Timestamp(x) for x in ['2014-01-01', '2015-01-01',
+                                     '2017-01-01', '2016-01-01']]
+        expected = pd.DataFrame({'a': d3, 'b': d4}, index=[1, 3, 0, 2])
+        sorted_df = df.sort_values(by=['a', 'b'], )
+        tm.assert_frame_equal(sorted_df, expected)
+
     def test_sort_values_inplace(self):
         frame = DataFrame(np.random.randn(4, 4), index=[1, 2, 3, 4],
                           columns=['A', 'B', 'C', 'D'])
@@ -269,6 +285,11 @@ def test_sort_datetimes(self):
         df2 = df.sort_values(by=['B'])
         assert_frame_equal(df1, df2)
 
+        df1 = df.sort_values(by='B')
+
+        df2 = df.sort_values(by=['C', 'B'])
+        assert_frame_equal(df1, df2)
+
     def test_frame_column_inplace_sort_exception(self):
         s = self.frame['A']
         with tm.assert_raises_regex(ValueError, "This Series is a view"):
@@ -321,7 +342,11 @@ def test_sort_nat_values_in_int_column(self):
         assert_frame_equal(df_sorted, df_reversed)
 
         df_sorted = df.sort_values(["datetime", "float"], na_position="last")
-        assert_frame_equal(df_sorted, df_reversed)
+        assert_frame_equal(df_sorted, df)
+
+        # Ascending should not affect the results.
+        df_sorted = df.sort_values(["datetime", "float"], ascending=False)
+        assert_frame_equal(df_sorted, df)
 
 
 class TestDataFrameSortIndexKinds(TestData):