macOS can only have 1 VM

Since we only use it for nightly build and [ci build], it might not be a problem.

Since we only use it for nightly build and [ci build], it might not be a problem.

Currently this PR adds macOS arm64 to our test suite, which means all PRs will run the test.

If we only want to use Cirrus CI for building macOS arm64 wheels, then we can make it nightly and trigger with [cd build]. I'm okay with this alternative.

If we only want to use Cirrus CI for building macOS arm64 wheels, then we can make it nightly and trigger with [cd build]. I'm okay with this alternative.

My reasoning here is that we did not get into trouble by not checking every PR on ARM64. I am thinking that limiting to the nightly build could allow us to save some computation.

I agree with @glemaitre's suggestion to focus its use wheel building and testing (for both nightly and release builds).

Thanks for the PR @thomasjpfan!

+1 to also try to build the linux arm wheels on this platform so as to not rely on travis anymore.

I updated this PR to build macOS wheels arm64 using cibuildwheel and enabled Cirrus CI in the repo. The Cirrus CI results can be seen in the list of checks and is available here.

Follow up PRs include:

• Uploading artifacts to the nightly or staging index
• Opening an tracking issue if the CirrusCI job fails.
• Add configuration for Linux ARM64

@ogrisel: can we merge this?

Should we go back to building the doc on CircleCI i.e. revert #21137? I am happy to do the PR.

It seems like #21137 was part of making room on CircleCI #20711 to move towards building the arm64 wheels on CircleCI but this does not make sense anymore since we are using Cirrus.

#21137 introduced additional complexity and a ~5-10 minutes latency (store artifact on github, get the github artifact on CircleCI) when we are waiting to see the result on the doc build.

I agree the latency is annoying. I propose:

1. doc-min-dependencies on GitHub Actions, but do not upload to CircleCI. I do not think we look at this rendering often. The job is usually for catching incompatible changes with our minimum dependencies when updating the examples.
2. The regular doc goes back to CircleCI. CircleCI focuses on building the version of the docs that is actually shown to users.

I admit this is more complex compared to having everything on CircleCI, but our doc building scripts already work with both, so we will not be adding more complexity.

@lesteve What do you think?

I guess that could get rid of the 5-10 minutes latency, but what is the downside of moving both doc builds back to CircleCI?

Your proposal does not add complexity to what we currently have but still is quite more complex compared to having both builds on CircleCI.

I guess that could get rid of the 5-10 minutes latency, but what is the downside of moving both doc builds back to CircleCI?

Ultimately, I want the setup with the fastest turn around time. Even with the latency, using GitHub Actions + CircleCI was faster to render back when #21137 was developed. Since then, we have gotten access to 60 workers on GitHub Actions.

The last time I checked, CircleCI can only run 2 workers concurrently, but when I look at the docs now we can run 30 workers concurrently. If we can run 30 workers concurrently and have enough compute on CircleCI, then I do prefer moving both doc builds back to CircleCI.

I also see 30 workers if I look at https://app.circleci.com/settings/plan/github/scikit-learn/overview.

About compute time on CircleCI, this is not super clear, the page above mentions 30,000 free credits per month, but also 400,000 credits:

You can also check our historical consumption in https://app.circleci.com/settings/plan/github/scikit-learn/usage and in March 2022, we consumed 600,000 credits so I am not sure CircleCI is actually enforcing these limits (or maybe the rules were different at the time?) ...

@lesteve Yea, I can see that sometimes we sometimes went over the 400k, but CircleCI did not stop working. We always go over their storage limit and that also takes some credits. For example this is for November 2022:

I think it comes down to how much we want to rely on CircleCI being okay with us going over sometimes. Historically, they have been lenient.