Update open_dataset backend to ensure compatibility with new explicit index model (#7150)

* Update open_dataset backend to ensure compatibility with new explicit index model

* Avoid generation of Indexes object

* Add test ensuring backend compatibility with multiindices

* Update xarray/tests/test_backends_api.py

* Use stack to construct multi index in test

* Mention open_dataset backend multi-index compatibility in whats-new

* remove _create_multiindex utility function

* remove pandas import

* Update doc/whats-new.rst

Co-authored-by: Deepak Cherian <[email protected]>
Co-authored-by: Benoit Bovy <[email protected]>
Co-authored-by: Benoit Bovy <[email protected]>

Author: 3 people

doc/whats-new.rst

@@ -70,6 +70,8 @@ Bug fixes
   :py:meth:`DataArray.to_index` for multi-index levels (convert to single index).
   (:issue:`6836`, :pull:`7105`)
   By `Benoît Bovy <https://github.com/benbovy>`_.
+- Support for open_dataset backends that return datasets containing multi-indexes (:issue:`7139`, :pull:`7150`)
+  By `Lukas Bindreiter <https://github.com/lukasbindreiter>`_.
 
 Documentation
 ~~~~~~~~~~~~~

xarray/backends/api.py

@@ -234,7 +234,7 @@ def _get_mtime(filename_or_obj):
 
 def _protect_dataset_variables_inplace(dataset, cache):
     for name, variable in dataset.variables.items():
-        if name not in variable.dims:
+        if name not in dataset._indexes:
             # no need to protect IndexVariable objects
             data = indexing.CopyOnWriteArray(variable._data)
             if cache:

xarray/tests/test_backends_api.py

@@ -48,6 +48,25 @@ def open_dataset(
     assert_identical(expected, actual)
 
 
+def test_multiindex() -> None:
+    # GH7139
+    # Check that we properly handle backends that change index variables
+    dataset = xr.Dataset(coords={"coord1": ["A", "B"], "coord2": [1, 2]})
+    dataset = dataset.stack(z=["coord1", "coord2"])
+
+    class MultiindexBackend(xr.backends.BackendEntrypoint):
+        def open_dataset(
+            self,
+            filename_or_obj,
+            drop_variables=None,
+            **kwargs,
+        ) -> xr.Dataset:
+            return dataset.copy(deep=True)
+
+    loaded = xr.open_dataset("fake_filename", engine=MultiindexBackend)
+    assert_identical(dataset, loaded)
+
+
 class PassThroughBackendEntrypoint(xr.backends.BackendEntrypoint):
     """Access an object passed to the `open_dataset` method."""