Investigate cibuildwheel

[rth]

I recently added a setup to build wheels in scikit-learn-contrib/scikit-learn-extra#66 with cibuildwheel and it was pleasantly straightforward. Or maybe I haven't run into issues yet )

In any case I think it might be worth considering it for scikit-learn. We could also add it to the main repo (related to #27) and build wheels on tags. It would very easy to test before the release just by uncommenting some CI config lines in a PR. It's probably a bit less flexible that our current setup, but it might be sufficient.

In the long run it would also allow building wheels for other architecture (aarc64, ppc64le) or Python distributions (PyPy), if we want to do that.

Unless there some blocker that would prevent us from using it?

[jnothman]

Sounds worthwhile to try, to avoid this separate repo and process.

[alfaro96]

If you think it is okay, I may open a (Draft) PR in the scikit-learn repository to investigate how cibuildwheel performs.

[rth]

That would be great @alfaro96 ! We could probably try with Github Actions there.

[thomasjpfan]

I suspect that integrating https://github.com/MacPython/scikit-learn-wheels/blob/master/azure/vendor_vcomp140.py with cibuildwheel may be nontrival.