`identical` returns False after read-write round trip

[martibosch]

Following #26.

The identical method returns False after a read-write round trip following the user guide instructions.

Reproducible example:

import geopandas as gpd
import xarray as xr
import xvec  # noqa: F401

cube = xr.Dataset(
    data_vars={"temperature": [20, 20, 20], "relative_humidity": [60, 60, 60]},
    coords={
        "geometry": gpd.points_from_xy(
            [6.7928, 6.1278, 7.3303], [47.0838, 46.2475, 46.2186]
        )
    },
).xvec.set_geom_indexes("geometry", crs="epsg:4326")

encoded = cube.xvec.encode_cf()
encoded.to_zarr("cube.zarr", mode="w")
roundtripped = xr.open_zarr("cube.zarr").xvec.decode_cf()
print("identical: ", roundtripped.identical(encoded))
print(
    "identical (no geometry): ",
    cube.drop_vars("geometry").identical(roundtripped.drop_vars("geometry")),
)

outputs:

identical:  False
identical (no geometry):  True

also note that I am getting this warning:

/path/to/pixi/env/lib/python3.13/site-packages/zarr/api/asynchronous.py:247: ZarrUserWarning: Consolidated metadata is currently not part in the Zarr format 3 specification. It may not be supported by other zarr implementations and may change in the future.
  warnings.warn(

versions:
geopandas 1.1.2
numpy 2.3.5
xarray 2026.1.0
xvec 0.5.2
zarr 3.1.5

[martinfleis]

This is because we don't preserve geopandas.GeometryDtype on the array. Xvec does not strictly depend on GeoPandas so we don't cast GeometryIndex to GeometryDtype but when it is created from one, it is carried through. Until serialisation. I think that the only way around this would be to depend on geopandas as use GeometryDtype every time.

xr.testing.assert_identical(roundtripped.geometry, cube.geometry)

---------------------------------------------------------------------------
AssertionError                            Traceback (most recent call last)
Cell In[22], [line 1](vscode-notebook-cell:?execution_count=22&line=1)
----> [1](vscode-notebook-cell:?execution_count=22&line=1) xr.testing.assert_identical(roundtripped.geometry, cube.geometry)

    [... skipping hidden 1 frame]

File ~/dev/xvec/.pixi/envs/default/lib/python3.13/site-packages/xarray/testing/assertions.py:188, in assert_identical(a, b)
    184 elif isinstance(a, DataArray):
    185     assert a.name == b.name, (
    186         f"DataArray names are different. L: {a.name}, R: {b.name}"
    187     )
--> [188](https://file+.vscode-resource.vscode-cdn.net/Users/martin/dev/xvec/~/dev/xvec/.pixi/envs/default/lib/python3.13/site-packages/xarray/testing/assertions.py:188)     assert a.identical(b), formatting.diff_array_repr(a, b, "identical")
    189 elif isinstance(a, Dataset | Variable):
    190     assert a.identical(b), formatting.diff_dataset_repr(a, b, "identical")

AssertionError: Left and right DataArray objects are not identical
Differing indexes:
L   geometry             Index([POINT (6.7928 47.0838), POINT (6.1278 46.2475), POINT (7.3303 46.2186)], dtype='object', name='geometry')
R   geometry             Index([POINT (6.7928 47.0838), POINT (6.1278 46.2475), POINT (7.3303 46.2186)], dtype='geometry', name='geometry')

I am not sure what is the best pathway forward. Imho this is an acceptable "data loss" but I might be wrong.

[maawoo]

@martibosch regarding the ZarrUserWarning: Just pass consolidated=False to both to_zarr and open_zarr. Here is some information in the Xarray docs but as the warning says, consolidated metadata is not part of the format 3 specs anyway.