Fix pickling regression with scipy backend after multiple opens

xarray/backends/scipy_.py

@@ -127,11 +127,18 @@ def __setitem__(self, key, value):
 # TODO: Remove this after upstreaming the fixes to scipy.
 class _PickleWorkaround:
     flush_only_netcdf_file: type[scipy.io.netcdf_file]
+    _initialized: bool = False
 
     @classmethod
     def add_cls(cls, new_class: type[Any]) -> None:
-        setattr(cls, new_class.__name__, new_class)
-        new_class.__qualname__ = cls.__qualname__ + "." + new_class.__name__
+        # Only set the class once to ensure pickling stability.
+        # Each call to _open_scipy_netcdf creates a new class, but we only need
+        # the first one. Subsequent calls would overwrite the class definition,
+        # breaking pickle's class-identity check for instances created earlier.
+        if not cls._initialized:
+            setattr(cls, new_class.__name__, new_class)
+            new_class.__qualname__ = cls.__qualname__ + "." + new_class.__name__
+            cls._initialized = True
 
 
 def _open_scipy_netcdf(

xarray/tests/test_backends.py

@@ -2523,6 +2523,14 @@ def test_pickle_open_dataset_from_bytes(self) -> None:
             with pickle.loads(pickle.dumps(roundtrip)) as unpickled:
                 assert_identical(unpickled, original)
 
+    def test_pickle_open_dataset_after_multiple_opens(self) -> None:
+        original = Dataset({"foo": ("x", [1, 2, 3])})
+        netcdf_bytes = bytes(original.to_netcdf(engine=self.engine))
+        with open_dataset(netcdf_bytes, engine=self.engine) as ds1:
+            with open_dataset(netcdf_bytes, engine=self.engine):
+                with pickle.loads(pickle.dumps(ds1)) as unpickled:
+                    assert_identical(unpickled, original)
+
     def test_compute_false(self) -> None:
         original = create_test_data()
         with pytest.raises(