Handle scalar dataset variables

virtualizarr/tests/test_xarray.py

@@ -419,6 +419,32 @@ def test_open_virtual_dataset_passes_expected_args(
         }
         mock_read_kerchunk.assert_called_once_with(**args)
 
+    @patch("virtualizarr.kerchunk.parse_array_refs")
+    def test_open_dataset_with_scalars(self, mock_parse_array_refs, tmpdir):
+        expected_attrs = {"units": "K", "long_name": "scalar"}
+        mock_parse_array_refs.return_value = (
+            None,
+            ZArray(
+                chunks=(),
+                shape=(),
+                compressor={},
+                dtype=np.dtype("float32"),
+                fill_value=0.0,
+                filters=None,
+                order="C",
+                zarr_format=2,
+            ),
+            {**expected_attrs, **{"_ARRAY_DIMENSIONS": []}},
+        )
+
+        ds = xr.Dataset()
+        ds["scalar"] = xr.Variable(dims=(), data=None, attrs=expected_attrs)
+        ds.to_netcdf(f"{tmpdir}/scalar.nc")
+        vds = open_virtual_dataset(f"{tmpdir}/scalar.nc")
+        assert vds.scalar.dims == ()
+        assert vds.scalar.data == 0.0
+        assert vds.scalar.attrs == expected_attrs
+
 
 class TestRenamePaths:
     def test_rename_to_str(self, netcdf4_file):

virtualizarr/xarray.py

@@ -333,13 +333,16 @@ def variable_from_kerchunk_refs(
 
     arr_refs = kerchunk.extract_array_refs(refs, var_name)
     chunk_dict, zarray, zattrs = kerchunk.parse_array_refs(arr_refs)
-
-    manifest = ChunkManifest._from_kerchunk_chunk_dict(chunk_dict)
-
     # we want to remove the _ARRAY_DIMENSIONS from the final variables' .attrs
     dims = zattrs.pop("_ARRAY_DIMENSIONS")
-
-    varr = virtual_array_class(zarray=zarray, chunkmanifest=manifest)
+    if chunk_dict:
+        manifest = ChunkManifest._from_kerchunk_chunk_dict(chunk_dict)
+        varr = virtual_array_class(zarray=zarray, chunkmanifest=manifest)
+    else:
+        # This means we encountered a scalar variable of dimension 0,
+        # very likely that it actually has no numeric value and its only purpose
+        # is to communicate dataset attributes.
+        varr = zarray.fill_value
 
     return xr.Variable(data=varr, dims=dims, attrs=zattrs)