run CI on python=3.13

[keewis]

• Closes Add python 3.13 to CI #9664

Missing libraries:

• dask: dask-expr depends on pyarrow, which will support 3.13 starting with pyarrow>=18.0
• pydap: depends on webob which uses the removed cgi stdlib (see Resolve Deprecation Warnings  pydap/pydap#319 (comment) in the "out-of-scope" section)
• numba: as usual will take some time to support it (and thus we can't test numbagg and sparse)

[keewis]

Interestingly it's conda that you can't install along with python=3.13, but cftime, pandas, and possibly h5py also raise errors (I can't reproduce that locally, though). So I guess we'll have to revisit this in a couple of weeks.

[mathause]

Is there a reason we are installing conda? For the output of conda info and conda list? There is a micromamba info and micromamba list output in the Setup micromamba step which we could refer to instead.

[keewis]

we use conda to call conda info -a, which contains information micromamba info does not include (and I don't know whether micromamba info -a exists). In theory we could use mamba info -a as well, but that also doesn't provide packages for python 3.13.

[keewis]

does numba support python=3.13 now? If not we're losing some test coverage by not testing on 3.12 (but not sure how important that is?)

[dcherian]

there is an RC out so I guess we wait a little bit more

[max-sixty]

(fwiw unless it's more work — and tests seem to be passing already — I would vote to merge sooner rather than wait on upstream. better to have something working and fewer things in buffer...)

[keewis]

numba=0.61.0 has been released last week with support for python=3.13 and numpy<=2.1 (numpy=2.2 support is scheduled for numba=0.61.1), so we should be able to finally merge this should CI pass.

[keewis]

looks like there's no python=3.13 build for numba on conda-forge, yet (the PR that adds this is still open). If we want to get this in now I'd propose to add back the 3.12 CI (which we can remove again once py3.13 builds are on conda-forge) and have the 3.13 CI without numba for now.

[dcherian]

how about we add a python3.12 + numba env and skip numba on the rest?

[keewis]

That could work. My only worry is that we'd be giving up some test coverage (but maybe that's small enough to ignore?)

[dcherian]

IMO it's ok to risk that and switch back if issues get reported.

[dcherian]

I took a pass at this. The windows failure looks flaky. I think we can merge.

[keewis]

Looks good, let's do the suggestion above on moving numba into a separate job in a different PR. I've rerun the flaky job and it passes now.