BUG: Fix bug where sort_index if unsorted levels.

doc/source/whatsnew/v0.18.0.txt

@@ -992,3 +992,4 @@ Bug Fixes
 - Bug in ``crosstab`` where arguments with non-overlapping indexes would return a ``KeyError`` (:issue:`10291`)
 
 - Bug in ``DataFrame.apply`` in which reduction was not being prevented for cases in which ``dtype`` was not a numpy dtype (:issue:`12244`)
+- Bug in ``MultiIndex.sort_index`` where if the levels were out of order, but the labels were monotonic, then the index would return unsorted. (:issue:`12261`)

pandas/core/frame.py

@@ -3236,7 +3236,7 @@ def sort_index(self, axis=0, level=None, ascending=True, inplace=False,
 
             # make sure that the axis is lexsorted to start
             # if not we need to reconstruct to get the correct indexer
-            if not labels.is_lexsorted():
+            if not labels.is_lexsorted() or not all([labels.get_level_values(i).is_monotonic for i in range(labels.nlevels)]):
                 labels = MultiIndex.from_tuples(labels.values)
 
             indexer = _lexsort_indexer(labels.labels, orders=ascending,

pandas/tests/frame/test_sorting.py

@@ -471,3 +471,18 @@ def test_frame_column_inplace_sort_exception(self):
 
         cp = s.copy()
         cp.sort_values()  # it works!
+
+    def test_sort_added_index(self):
+        # GH 12261
+        df = DataFrame(0, columns=[], index=MultiIndex.from_product([[], []]))
+
+        df.loc['b', '2'] = 1
+        df.loc['a', '3'] = 1
+
+        expected = DataFrame([[1.0, np.nan], [np.nan, 1.0]],
+                             index=MultiIndex.from_product([['b', 'a'], ['']]),
+                             columns=['2', '3']).sort_index()
+
+        actual = df.sort_index()
+
+        assert_frame_equal(expected, actual)