REGR: groupby.value_counts with all NA values (#59999)

* REGR: groupby.value_counts with all NA values

* Better implementation

Author: rhshadrach

pandas/core/groupby/ops.py

@@ -867,7 +867,7 @@ def _ob_index_and_ids(
             names=names,
             verify_integrity=False,
         )
-        if not consistent_sorting:
+        if not consistent_sorting and len(ob_index) > 0:
             # Sort by the levels where the corresponding sort argument is True
             n_levels = len(sorts)
             drop_levels = [

pandas/tests/groupby/methods/test_value_counts.py

@@ -1219,3 +1219,25 @@ def test_value_counts_sort_categorical(sort, vc_sort, normalize):
     expected = expected.take(taker)
 
     tm.assert_series_equal(result, expected)
+
+
+@pytest.mark.parametrize("groupby_sort", [True, False])
+def test_value_counts_all_na(sort, dropna, groupby_sort):
+    # GH#59989
+    df = DataFrame({"a": [2, 1, 1], "b": np.nan})
+    gb = df.groupby("a", sort=groupby_sort)
+    result = gb.value_counts(sort=sort, dropna=dropna)
+
+    kwargs = {"levels": [[1, 2], [np.nan]], "names": ["a", "b"]}
+    if dropna:
+        data = []
+        index = MultiIndex(codes=[[], []], **kwargs)
+    elif not groupby_sort and not sort:
+        data = [1, 2]
+        index = MultiIndex(codes=[[1, 0], [0, 0]], **kwargs)
+    else:
+        data = [2, 1]
+        index = MultiIndex(codes=[[0, 1], [0, 0]], **kwargs)
+    expected = Series(data, index=index, dtype="int64", name="count")
+
+    tm.assert_series_equal(result, expected)