Add Python 3.12

[cbrnr]

I've replaced Python 3.11 with Python 3.12 in our pip-pre test on ubuntu-latest.

Which versions do we want to test explicitly? Currently (with this change), we're testing 3.9 (ubuntu-latest/minimal, macos-latest/mamba, ubuntu-20.04/old), 3.10 (ubuntu-latest/conda, windows-latest/mamba), and 3.12.

[cbrnr]

Package dependency hell:

The user requested numpy>=2.0.0.dev0
scikit-learn 1.4.dev0 depends on numpy>=1.19.5
scipy 1.12.0rc1 depends on numpy<1.29.0 and >=1.22.4

[larsoner]

I wouldn't do it on the pip-pre job but instead bump a different job. One of the conda ones might actually be easiest

Which versions do we want to test explicitly? Currently (with this change), we're testing 3.9 (ubuntu-latest/minimal, macos-latest/mamba, ubuntu-20.04/old), 3.10 (ubuntu-latest/conda, windows-latest/mamba), and 3.12.

That sounds reasonable. We want to test oldest and newest at a minimum. Beyond that having some coverage is desirable but I don't think full Python version coverage is required.

[cbrnr]

I wouldn't do it on the pip-pre job but instead bump a different job. One of the conda ones might actually be easiest

OK, I bumped the three conda-based jobs. However, conceptually it would make a lot of sense to also use the latest Python for pip-pre IMO.

[larsoner]

However, conceptually it would make a lot of sense to also use the latest Python for pip-pre IMO.

I agree, ideally we can make that switch soon. But the complications added by NumPy 2.0 make it too difficult, and testing against NumPy 2.0 is important -- it has been a pain, but it has also found all sorts of stuff for us and upstream packages to fix and change over the last 6 months or so.

[larsoner]

Looks like numba is a problem on those conda runs. Maybe conda isn't the way to go, then. These CI configs can be a pain and require lots of experimentation and adjustment!

[larsoner]

(Or maybe it's just numba-on-windows-3.12 -- you could try putting that one back to whatever version it was before and see if the Linux and macOS jobs work!)

[cbrnr]

Numba 0.59, which hasn't been released yet, will support Python 3.12 (numba/numba#9246).

[cbrnr]

0.59RC1 is available on PyPI, so maybe use that with pip-pre?

[larsoner]

One solution would be to change just one job (conda linux?) to use 3.12 and before the setup-miniconda or whatever we use nowadays, modify environment.yml inplace to remove the numba line (only in that job) until a compatible numba is released.

[larsoner]

0.59RC1 is available on PyPI, so maybe use that with pip-pre?

I thought pip-pre was the job with the incompat due to NumPy 2.0.0.dev0+Py3.12, no? It's unlikely that that RC is NumPy2.0 compatible in any case (and I do think we want to keep testing NumPy 2.0 in the pre job if possible).

Another option since they're already on RC is just to wait a week or two or whatever, then bump a conda job (or a few as done here).

[cbrnr]

Yes, but the incompatibility was because of SciPy. You're probably right that Numba won't support NumPy 2, but who knows, maybe they have already made their code compatible. So it's probably best to wait a couple of days.

[larsoner]

@cbrnr can I push a commit to try my idea of editing environment.yml? Then when numba lands we can just remove the 4-5 lines in tests.yml that I'd be adding (and would have 3.12 in the meantime)

[cbrnr]

Yes, sure!

[larsoner]

Okay enabling auto-merge, but will need to manually adjust CI reqs in branch protection rules once it comes back green. I'll try to remember to check back in a bit but if someone sees everything green feel free to ping me

[larsoner]

Thanks @cbrnr !

[cbrnr]

Thank you @larsoner, you did all the work!