BUG: MemorizedFunc.call output changed between 1.3.2 and 1.4.0

[larsoner]

The short version is that the metadata is no longer returned after #894. You can see it in the docstring of the Returns section of MemorizedFunc.call in the diff here:

https://github.com/joblib/joblib/pull/894/files#diff-892315c4227d75d32ad4675e8aaee76fb62a998e9e609cc204df1dbc4272f84bL849-L850

Was this intentional? It broke some downstream code, e.g.:

https://github.com/mne-tools/mne-bids-pipeline/blob/cbeeb98d0625ac948f8a6951305de11520eac5d5/mne_bids_pipeline/_run.py#L280

https://app.circleci.com/pipelines/github/mne-tools/mne-bids-pipeline/4478/workflows/c9b0bd11-4a7b-4193-aeeb-683979144ef0/jobs/64079

If so maybe the note that says Allow caching co-routines with Memory.cache. in the release log could be updated with this backward incompatible change info?

[tomMoral]

Hello,
Thanks for the report and sorry for the incompatible change. This was not intentional, it came in a large refactor and apparently we had not regression test for this API.

I am not sure what is the best course of action now. I think the API would be cleaner by not returning the metadata as this allows to call either the function of call the same way.
But this means that potentially a lot of code will need to support both or force updates, which is quite troublesome.
I think for the sake of maintainability of downstream packages, I would go for a BF release but happy to take an opinion here (@ogrisel @GaelVaroquaux )

[larsoner]

FWIW it wasn't too difficult in our case to put a simple workaround in place so for us it's no longer a problem. So I guess it's a question of how extensively the API is used by others.

[tomMoral]

I think for the purpose of consistent API, we should make a BF release.
If we change something, we should do it explicitly, not by mistake.. 😓

Thanks for the report, I will revert this change and add some non-regression tests.