different file hashes from identical data arrays with non-dimensional coordinates

[djmannion]

I am computing the MD5 hashes of files created using to_netcdf to verify that their contents is as expected. However, I am running into an issue where identical arrays can give different hashes each time they are generated.

This only seems to occur when using non-dimensional coordinates, and I think it is related to the order in which the non-dimensional coordinates are listed in the saved file.

For example, the following array

ds = xr.DataArray(
    data=np.arange(5),
    dims=("x",),
    coords={
        "x": np.arange(5),
        "a": ("x", np.arange(5)),
        "b": ("x", np.arange(5)),
    },
)

will sometimes be saved with the non-dimensional coordinates listed as "a b" and sometimes as "b a" - which causes the file hash to be unpredictable (well, not a single value at least).

I can enforce the order by setting ds.encoding["coordinates"] = "a b", which gives consistent hashes.

Is this the only way to approach it (I find it easy to forget to include and to keep updated) or is the inconsistent ordering a bug to report? It also seems related to some of the behaviour described in #4738. I have tried looking through the code, but I can't find where this inconsistent ordering would be occurring.

This is with xarray==2022.3.0.

Thanks for any assistance!