New Collections protocol breaks xarray objects

[dcherian]

The recent merge of the new Collections protocol (#8674) broke Xarray's tests (pydata/xarray#6578)

Xarray objects return None from __dask_graph__ if there no dask arrays being wrapped.

It seems like the current version of is_dask_collection only checks for the presence of __dask_graph__. Xarray objects are a case where this check is not sufficient (they look like dask collections in general, you can call dask.visualize and dask.compute on xarray objects).

It isn't clear to me that there is a good solution other than explicitly checking the return value in is_dask_collection but if there is something we could do on the Xarray side, please let us know!

[jrbourbeau]

Thanks for reporting @dcherian (and hooray for upstream CI testing 🎉 ). I've not had a chance to dig into this yet, but plan to do so tomorrow. cc'ing @douglasdavis @ian-r-rose in the meantime

[ian-r-rose]

Thanks for flagging this @dcherian! We had some discussion about exactly this change here, and were not aware of any downstream usages of returning None. At the time it seemed relatively safe since there are many places in dask's codebase that don't check for None and happily assume the result is a graph. But evidently not.

It's simple enough fix to roll back that change, though after we do it it might be nice if we could continue discussing how to tighten the protocol a bit.

[douglasdavis]

I agree with @ian-r-rose. I just created a PR (#9062) to revert the change to is_dask_collection.