EHN: Groupby on multiindex with missing data in group keys raises IndexError (#20519)

* if all the values in a level of a MultiIndex were missing, fill with numpy nan

Author: proost

doc/source/whatsnew/v1.0.0.rst

@@ -174,8 +174,10 @@ Groupby/resample/rolling
 ^^^^^^^^^^^^^^^^^^^^^^^^
 
 -
--
+- Bug in :meth:`DataFrame.groupby` with multiple groups where an ``IndexError`` would be raised if any group contained all NA values (:issue:`20519`)
+- Bug in :meth:`DataFrame.rolling` not allowing for rolling over datetimes when ``axis=1`` (:issue: `28192`)
 - Bug in :meth:`DataFrame.groupby` not offering selection by column name when ``axis=1`` (:issue:`27614`)
+- Bug in :meth:`DataFrameGroupby.agg` not able to use lambda function with named aggregation (:issue:`27519`)
 
 Reshaping
 ^^^^^^^^^

pandas/core/indexes/multi.py

@@ -1279,7 +1279,10 @@ def _get_grouper_for_level(self, mapper, level):
             # Remove unobserved levels from level_index
             level_index = level_index.take(uniques)
 
-        grouper = level_index.take(codes)
+        if len(level_index):
+            grouper = level_index.take(codes)
+        else:
+            grouper = level_index.take(codes, fill_value=True)
 
         return grouper, codes, level_index
 

pandas/tests/groupby/test_grouping.py

@@ -628,6 +628,24 @@ def test_groupby_empty(self):
         # check name
         assert s.groupby(s).grouper.names == ["name"]
 
+    def test_groupby_level_index_value_all_na(self):
+        # issue 20519
+        df = DataFrame(
+            [["x", np.nan, 10], [None, np.nan, 20]], columns=["A", "B", "C"]
+        ).set_index(["A", "B"])
+        result = df.groupby(level=["A", "B"]).sum()
+        expected = DataFrame(
+            data=[],
+            index=MultiIndex(
+                levels=[Index(["x"], dtype="object"), Index([], dtype="float64")],
+                codes=[[], []],
+                names=["A", "B"],
+            ),
+            columns=["C"],
+            dtype="int64",
+        )
+        tm.assert_frame_equal(result, expected)
+
 
 # get_group
 # --------------------------------