Fix non-dim coord grouping by multiple variables.

xref pydata/xarray#9372

Author: dcherian

flox/xarray.py

@@ -250,6 +250,14 @@ def xarray_reduce(
     else:
         ds = obj._to_temp_dataset()
 
+    # These will need to be broadcast/reduced as data_vars
+    reset_non_dim_coords = [
+        name
+        for name in ds._coord_names
+        if any(dim in ds._variables[name].dims for dim in grouper_dims)
+    ]
+    ds = ds.reset_coords(reset_non_dim_coords)
+
     try:
         from xarray.indexes import PandasMultiIndex
     except ImportError:
@@ -475,6 +483,7 @@ def wrapper(array, *by, func, skipna, core_dims, **kwargs):
         if all(d not in ds_broad[var].dims for d in dim_tuple):
             actual[var] = ds_broad[var]
 
+    actual = actual.set_coords(reset_non_dim_coords)
     for newdim in newdims:
         actual.coords[newdim.name] = newdim.values if newdim.is_scalar else np.array(newdim.values)
 

tests/test_xarray.py

@@ -749,3 +749,20 @@ def test_direct_reduction(func):
     with xr.set_options(use_flox=False):
         expected = getattr(data.groupby("x", squeeze=False), func)(**kwargs)
     xr.testing.assert_identical(expected, actual)
+
+
+def test_non_dim_coords_with_core_dim():
+    coords = {"a": ("x", [0, 0, 1, 1]), "b": ("y", [0, 0, 1, 1])}
+    square = xr.DataArray(np.arange(16).reshape(4, 4), coords=coords, dims=["x", "y"])
+    actual = xarray_reduce(square, "a", "b", func="mean")
+    expected = xr.DataArray(
+        np.array([[2.5, 4.5], [10.5, 12.5]]),
+        dims=("a", "b"),
+        coords={"a": [0, 1], "b": [0, 1]},
+    )
+    xr.testing.assert_identical(actual, expected)
+
+    actual = xarray_reduce(square, "x", "y", func="mean")
+    expected = square.astype(np.float64).copy()
+    expected["a"], expected["b"] = xr.broadcast(square.a, square.b)
+    xr.testing.assert_identical(actual, expected)