Fix bug with NaNs in by and method='blockwise' (#384)

xref pydata/xarray#9320

flox/core.py

@@ -2663,10 +2663,18 @@ def groupby_reduce(
                 groups = (groups[0][sorted_idx],)
 
     if factorize_early:
+        assert len(groups) == 1
+        (groups_,) = groups
         # nan group labels are factorized to -1, and preserved
         # now we get rid of them by reindexing
-        # This also handles bins with no data
-        result = reindex_(result, from_=groups[0], to=expected_, fill_value=fill_value).reshape(
+        # First, for "blockwise", we can have -1 repeated in different blocks
+        # This breaks the reindexing so remove those first.
+        if method == "blockwise" and (mask := groups_ == -1).sum(axis=-1) > 1:
+            result = result[..., ~mask]
+            groups_ = groups_[..., ~mask]
+
+        # This reindex also handles bins with no data
+        result = reindex_(result, from_=groups_, to=expected_, fill_value=fill_value).reshape(
             result.shape[:-1] + grp_shape
         )
         groups = final_groups

tests/test_core.py

@@ -1929,3 +1929,17 @@ def test_ffill_bfill(chunks, size, add_nan_by, func):
     expected = flox.groupby_scan(array.compute(), by, func=func)
     actual = flox.groupby_scan(array, by, func=func)
     assert_equal(expected, actual)
+
+
+@requires_dask
+def test_blockwise_nans():
+    array = dask.array.ones((1, 10), chunks=2)
+    by = np.array([-1, 0, -1, 1, -1, 2, -1, 3, 4, 4])
+    actual, actual_groups = flox.groupby_reduce(
+        array, by, func="sum", expected_groups=pd.RangeIndex(0, 5)
+    )
+    expected, expected_groups = flox.groupby_reduce(
+        array.compute(), by, func="sum", expected_groups=pd.RangeIndex(0, 5)
+    )
+    assert_equal(expected_groups, actual_groups)
+    assert_equal(expected, actual)