Consistent Return Structure for Rolling Apply (#20984)

Author: WillAyd

doc/source/whatsnew/v0.23.0.txt

@@ -1328,6 +1328,7 @@ Groupby/Resample/Rolling
 - Bug in :func:`DataFrame.groupby` where transformations using ``np.all`` and ``np.any`` were raising a ``ValueError`` (:issue:`20653`)
 - Bug in :func:`DataFrame.resample` where ``ffill``, ``bfill``, ``pad``, ``backfill``, ``fillna``, ``interpolate``, and ``asfreq`` were ignoring ``loffset``. (:issue:`20744`)
 - Bug in :func:`DataFrame.groupby` when applying a function that has mixed data types and the user supplied function can fail on the grouping column (:issue:`20949`)
+- Bug in :func:`DataFrameGroupBy.rolling().apply() <pandas.core.window.Rolling.apply>` where operations performed against the associated :class:`DataFrameGroupBy` object could impact the inclusion of the grouped item(s) in the result (:issue:`14013`)
 
 Sparse
 ^^^^^^

pandas/core/window.py

@@ -837,11 +837,7 @@ def _apply(self, func, name=None, window=None, center=None,
         index, indexi = self._get_index(index=index)
         results = []
         for b in blocks:
-            try:
-                values = self._prep_values(b.values)
-            except TypeError:
-                results.append(b.values.copy())
-                continue
+            values = self._prep_values(b.values)
 
             if values.size == 0:
                 results.append(values.copy())

pandas/tests/test_window.py

@@ -105,17 +105,19 @@ def test_attribute_access(self):
     def tests_skip_nuisance(self):
 
         df = DataFrame({'A': range(5), 'B': range(5, 10), 'C': 'foo'})
-
         r = df.rolling(window=3)
         result = r[['A', 'B']].sum()
         expected = DataFrame({'A': [np.nan, np.nan, 3, 6, 9],
                               'B': [np.nan, np.nan, 18, 21, 24]},
                              columns=list('AB'))
         tm.assert_frame_equal(result, expected)
 
-        expected = concat([r[['A', 'B']].sum(), df[['C']]], axis=1)
-        result = r.sum()
-        tm.assert_frame_equal(result, expected, check_like=True)
+    def test_skip_sum_object_raises(self):
+        df = DataFrame({'A': range(5), 'B': range(5, 10), 'C': 'foo'})
+        r = df.rolling(window=3)
+
+        with tm.assert_raises_regex(TypeError, 'cannot handle this type'):
+            r.sum()
 
     def test_agg(self):
         df = DataFrame({'A': range(5), 'B': range(0, 10, 2)})
@@ -3174,6 +3176,28 @@ def test_rolling_apply(self, raw):
             lambda x: x.rolling(4).apply(lambda y: y.sum(), raw=raw))
         tm.assert_frame_equal(result, expected)
 
+    def test_rolling_apply_mutability(self):
+        # GH 14013
+        df = pd.DataFrame({'A': ['foo'] * 3 + ['bar'] * 3, 'B': [1] * 6})
+        g = df.groupby('A')
+
+        mi = pd.MultiIndex.from_tuples([('bar', 3), ('bar', 4), ('bar', 5),
+                                        ('foo', 0), ('foo', 1), ('foo', 2)])
+
+        mi.names = ['A', None]
+        # Grouped column should not be a part of the output
+        expected = pd.DataFrame([np.nan, 2., 2.] * 2, columns=['B'], index=mi)
+
+        result = g.rolling(window=2).sum()
+        tm.assert_frame_equal(result, expected)
+
+        # Call an arbitrary function on the groupby
+        g.sum()
+
+        # Make sure nothing has been mutated
+        result = g.rolling(window=2).sum()
+        tm.assert_frame_equal(result, expected)
+
     def test_expanding(self):
         g = self.frame.groupby('A')
         r = g.expanding()