Add support in the "zarr" backend for reading NCZarr data

xarray/backends/zarr.py

@@ -178,19 +178,40 @@ def _determine_zarr_chunks(enc_chunks, var_chunks, ndim, name, safe_chunks):
     raise AssertionError("We should never get here. Function logic must be wrong.")
 
 
+def _get_nczarr_dims(zarr_obj):
+    # NCZarr defines dimensions through metadata in .zarray
+    zarray_path = os.path.join(zarr_obj.path, ".zarray")
+    zarray = zarr.util.json_loads(zarr_obj.store[zarray_path])
+    # NCZarr uses Fully Qualified Names
+    dimensions = [os.path.basename(dim) for dim in zarray["_NCZARR_ARRAY"]["dimrefs"]]
+    return dimensions
+
+
+def _hide_nczarr_attrs(attrs):
+    return HiddenKeyDict(attrs, [attr for attr in attrs if attr.startswith("_NC")])
+
+
 def _get_zarr_dims_and_attrs(zarr_obj, dimension_key):
     # Zarr arrays do not have dimensions. To get around this problem, we add
     # an attribute that specifies the dimension. We have to hide this attribute
     # when we send the attributes to the user.
     # zarr_obj can be either a zarr group or zarr array
+    attributes = _hide_nczarr_attrs(zarr_obj.attrs)
     try:
+        # Xarray-Zarr
         dimensions = zarr_obj.attrs[dimension_key]
     except KeyError:
-        raise KeyError(
-            f"Zarr object is missing the attribute `{dimension_key}`, which is "
-            "required for xarray to determine variable dimensions."
-        )
-    attributes = HiddenKeyDict(zarr_obj.attrs, [dimension_key])
+        try:
+            # NCZarr
+            dimensions = _get_nczarr_dims(zarr_obj)
+            attributes = dict(attributes)
+            attributes[dimension_key] = dimensions
+        except KeyError:
+            raise KeyError(
+                f"Zarr object is missing the attribute `{dimension_key}`, which is "
+                "required for xarray to determine variable dimensions."
+            )
+    attributes = HiddenKeyDict(attributes, [dimension_key])
     return dimensions, attributes
 
 
@@ -430,7 +451,7 @@ def get_variables(self):
         )
 
     def get_attrs(self):
-        return dict(self.zarr_group.attrs.asdict())
+        return dict(_hide_nczarr_attrs(self.zarr_group.attrs.asdict()))
 
     def get_dimensions(self):
         dimensions = {}

xarray/tests/test_backends.py

@@ -4,6 +4,7 @@
 import math
 import os.path
 import pickle
+import platform
 import re
 import shutil
 import sys
@@ -5427,3 +5428,26 @@ def test_write_file_from_np_str(str_type, tmpdir) -> None:
     txr = tdf.to_xarray()
 
     txr.to_netcdf(tmpdir.join("test.nc"))
+
+
+@requires_zarr
+@requires_netCDF4
+def test_nczarr():
+    netcdfc_version = Version(nc4.getlibversion().split()[0])
+    if netcdfc_version < Version("4.8.1"):
+        pytest.skip("requires netcdf-c>=4.8.1")
+    if (platform.system() == "Windows") and (netcdfc_version == Version("4.8.1")):
+        # Bug in netcdf-c==4.8.1 (typo: Nan instead of NaN)
+        # https://github.com/Unidata/netcdf-c/issues/2265
+        pytest.skip("netcdf-c==4.8.1 has issues on Windows")
+
+    expected = create_test_data()
+    # Drop dim3: netcdf-c does not support dtype='<U1'
+    # https://github.com/Unidata/netcdf-c/issues/2259
+    expected = expected.drop_vars("dim3")
+    with create_tmp_file(suffix=".zarr") as tmp:
+        # netcdf-c>4.8.1 will add _ARRAY_DIMENSIONS by default
+        mode = "nczarr" if netcdfc_version == Version("4.8.1") else "nczarr,noxarray"
+        expected.to_netcdf(f"file://{tmp}#mode={mode}")
+        actual = xr.open_zarr(tmp, consolidated=False)
+        assert_identical(expected, actual)