Recommend check / test version numbering

[jarrodmillman]

See matplotlib/matplotlib#26736 (comment)

Currently I am not able to figure out how to easily install the recent nightly wheels for matplotlib due to version numbering issues and upper pins.

I wonder if we could automate testing whether it is possible to install the nightly wheels using the recommended pip command:

python -m pip install \
  --upgrade \
  --pre \
  --index-url https://pypi.anaconda.org/scientific-python-nightly-wheels/simple \
  --extra-index-url https://pypi.org/simple \
  matplotlib

[jarrodmillman]

Just so folks don't need to follow the link above:

>>> from packaging import version
>>> version.parse("3.8.0rc1") < version.parse("3.8.0.dev1951+g5f785e3e19")
False
>>> version.parse("3.8.0rc1") < version.parse("3.8.0rc1.dev1951+g5f785e3e19")
False
>>> version.parse("3.8.0rc1") < version.parse("3.8.0rc1.dev0")
False
>>> version.parse("3.8.0rc1") < version.parse("3.8.0rc2.dev0")
True
>>> version.parse("3.8.0rc1") < version.parse("3.8.0rc2.dev1951+g5f785e3e19")
True

So after releasing 3.8.0rc1 the basic version number should be updated to 3.8.0rc2.dev0 or 3.8.0rc2.dev.

[bsipocz]

Thanks for the heads up, the first one is surprising, the rest is what I was expecting.

As for testing, would you suggest to test whether it's possible to install the latest uploaded nightly? E.g. with astropy I ended up in a situation where the pip install worked, but it picked up an older nightly which masked over an issue with the wheels (and a newly added dependency) perfectly.

[jarrodmillman]

As for testing, would you suggest to test whether it's possible to install the latest uploaded nightly?

Yes, exactly. Maybe a cron job once a week or so to test whether the recommended pip command is able to install the most recent nightly wheels.

[martinfleis]

And what would be the action if the job fails? Getting back to a project, suggesting they need to change the version of the dev branch?

[bsipocz]

And what would be the action if the job fails? Getting back to a project, suggesting they need to change the version of the dev branch?

Yes, I suppose an issue being auto-opened here, and then one of us following up with the project. This of course assumes that the problem is scaling well, e.g. we won't host gazillions of wheels as a "conda-forge for wheels", but keep it to the core ones.

[matthewfeickert]

I didn't follow this closeley in 2023 (sorry), but given @ksunden's comment matplotlib/matplotlib#26736 (comment) was this actually a systematic problem that a check would have help avoid? Or was this simply a time delay issue for information to get propaged to matplotlib's main branch so it could end up in the matplotlib nightly wheels?