[pull] main from pydata:main

doc/whats-new.rst

@@ -29,6 +29,10 @@ Bug Fixes
 - Fix a major performance regression in :py:meth:`Coordinates.to_index` (and
   consequently :py:meth:`Dataset.to_dataframe`) caused by converting the cached
   code ndarrays into Python lists (:issue:`11305`).
+- Preserve the Zarr array ``fill_value`` in the variable ``encoding`` when reading
+  a ``zarr_format=3`` store with ``use_zarr_fill_value_as_mask=False``, so it is no
+  longer silently lost on round-trip (:issue:`10269`).
+  By `Davis Bennett <https://github.com/d-v-b>`_.
 
 
 Documentation

xarray/backends/zarr.py

@@ -938,10 +938,19 @@ def open_store_variable(self, name):
             # by interpreting Zarr's fill_value to mean the same as netCDF's _FillValue
             if zarr_array.fill_value is not None:
                 attributes["_FillValue"] = zarr_array.fill_value
-        elif "_FillValue" in attributes:
-            attributes["_FillValue"] = FillValueCoder.decode(
-                attributes["_FillValue"], zarr_array.dtype
-            )
+        else:
+            # Preserve the Zarr array fill_value in the encoding so it is not
+            # lost on round-trip. The write path reads it back from here.
+            # Only zarr_format 3 supports `fill_value` as an encoding key
+            # (in zarr_format 2 the fill_value is set via `_FillValue`).
+            # See https://github.com/pydata/xarray/issues/10269
+            zarr_format_3 = _zarr_v3() and self.zarr_group.metadata.zarr_format == 3
+            if zarr_format_3:
+                encoding["fill_value"] = zarr_array.fill_value
+            if "_FillValue" in attributes:
+                attributes["_FillValue"] = FillValueCoder.decode(
+                    attributes["_FillValue"], zarr_array.dtype
+                )
 
         return Variable(dimensions, data, attributes, encoding)
 

xarray/tests/test_backends.py

@@ -3869,6 +3869,27 @@ def test_zarr_fill_value_setting(self, dtype):
             # ``raise_on_invalid=vn in check_encoding_set`` line in zarr.py
             # ds.foo.encoding["fill_value"] = fv
 
+    def test_zarr_fill_value_in_encoding_on_read(self) -> None:
+        # GH #10269: the Zarr array fill_value should be preserved in the
+        # variable encoding on read, so that it is not lost on round-trip.
+        # `fill_value` is an independent encoding key only for zarr_format 3;
+        # for zarr_format 2 the fill_value is set via `_FillValue`.
+        if not has_zarr_v3 or zarr.config.get("default_zarr_format") != 3:
+            pytest.skip("fill_value is only an encoding key for zarr_format 3")
+
+        ds = xr.Dataset({"foo": ("x", [1, 2, 3])})
+        ds.foo.encoding = {"fill_value": -99}
+
+        open_kwargs = {"consolidated": False, "use_zarr_fill_value_as_mask": False}
+        with self.roundtrip(ds, open_kwargs=open_kwargs) as actual:
+            assert actual.foo.encoding["fill_value"] == -99
+
+        # the fill_value must survive an open -> write -> open round-trip even
+        # when the user never touches the encoding explicitly
+        with self.roundtrip(ds, open_kwargs=open_kwargs) as opened:
+            with self.roundtrip(opened, open_kwargs=open_kwargs) as actual:
+                assert actual.foo.encoding["fill_value"] == -99
+
 
 @requires_zarr
 @pytest.mark.skipif(