h5netcdf: Don't assume there are dimensions when there are chunks

[rho-novatron]

• Closes With backend h5netcdf and driver h5pyd there are cases when len(var.chunks) > 0 but len(var.dimensions) == 0 #10091
• Tests added
• User visible changes (including notable bug fixes) are documented in whats-new.rst
• New functions/methods are listed in api.rst

[welcome]

Thank you for opening this pull request! It may take us a few days to respond here, so thank you for being patient.
If you have questions, some answers may be found in our contributing guidelines.

[rho-novatron]

I added a somewhat artificial test to not have to include an hsds service in the CI chain.

[kmuehlbauer]

Changes look reasonable. @rho-novatron Is this a speciality of h5pyd/hsds to return strings with chunks? I just want to make sure there is no other issue hidden which could be resolved upstream.

[rho-novatron]

That is a good question, and I don't know the answer. I'll check and get back.

[rho-novatron]

It really does look as if chunking isn't really supported for scalar datasets, but that they are stored with layout class H5D_CHUNKED and thus having chunk dimensions == (1,). There are several places, both in h5netcdf and h5pyd that verifies that scalar datasets cannot be created with chunks. However, I can't really find anything in the HSDS documentation that says that a scalar dataset couldn't have a chunk.

Perhaps it would be better to check for scalar datasets in h5netcdf and return None for scalar datasets? I think that might be the best balance between keeping xarray generic, h5pyd slightly raw and h5netcdf the adaptor logic? Do you have a better idea?

[rho-novatron]

I added a draft PR that makes concrete what I meant: h5netcdf/h5netcdf#259

[kmuehlbauer]

Thanks for digging into this @rho-novatron.

I was about suggesting something along your PR h5netcdf/h5netcdf#259. Responded over there.

[kmuehlbauer]

Thanks @rho-novatron for this fast interaction, great experience!