I also wonder if the test failures with numpy.linalg.LinAlgError: Eigenvalues did not converge are related to atlas vs openblas.

Note that numpy has similar ARM64 failures (numpy.linalg.LinAlgError: Eigenvalues did not converge) on their dev build on travis:

https://travis-ci.com/github/numpy/numpy/jobs/521809468

and I also observed similar failures today on our travis arm64 build when working with [cd build] commit flag on an unrelated change in this PR: #20468 (comment).

We did get the failure on our weekly arm64 build on travis.

https://travis-ci.com/github/scikit-learn/scikit-learn/jobs/521594473

but not 8 days ago:

https://travis-ci.com/github/scikit-learn/scikit-learn/builds/230982232

The last successful arm64 nightly build for numpy was 4 days ago: https://travis-ci.com/github/numpy/numpy/builds/231467572

I opened numpy/numpy#19411 to track the issue upstream in numpy but this is probably something that changed in the runtime environment other than numpy / scipy and openblas since those jobs use the same released wheels for those libs...

It means that we can probably remove the following entry in the travis CI config and update the documentation to remove the [arm64] manual flag in the scikit-learn CI documentation.

https://github.com/scikit-learn/scikit-learn/blob/main/.travis.yml#L33-L40

I have removed the trigger and the related documentation.
I agree that travis builds should be considered in another pull request.

@ogrisel the numpy.linalg.LinAlgError: Eigenvalues did not converge has disappeared... I don't know why and when...

@ogrisel the numpy.linalg.LinAlgError: Eigenvalues did not converge has disappeared... I don't know why and when...

After investigation, it seems it was fixed in OpenBLAS 0.3.16:

• fixed missing restore of a register in the recently rewritten DNRM2 kernel
  for ThunderX2 and Neoverse N1 that could cause spurious failures in e.g.
  DGEEV

Something has changed in openmp management? Two azure builds are failing with

=================================== FAILURES ===================================
_______________________ test_openmp_parallelism_enabled ________________________
[gw0] linux -- Python 3.9.6 /usr/share/miniconda/envs/testvenv/bin/python

    def test_openmp_parallelism_enabled():
        # Check that sklearn is built with OpenMP-based parallelism enabled.
        # This test can be skipped by setting the environment variable
        # ``SKLEARN_SKIP_OPENMP_TEST``.
        if os.getenv("SKLEARN_SKIP_OPENMP_TEST"):
            pytest.skip("test explicitly skipped (SKLEARN_SKIP_OPENMP_TEST)")
    
        base_url = "dev" if __version__.endswith(".dev0") else "stable"
        err_msg = textwrap.dedent(
            """
            This test fails because scikit-learn has been built without OpenMP.
            This is not recommended since some estimators will run in sequential
            mode instead of leveraging thread-based parallelism.
    
            You can find instructions to build scikit-learn with OpenMP at this
            address:
    
                https://scikit-learn.org/{}/developers/advanced_installation.html
    
            You can skip this test by setting the environment variable
            SKLEARN_SKIP_OPENMP_TEST to any value.
            """
        ).format(base_url)
    
>       assert _openmp_parallelism_enabled(), err_msg
E       AssertionError: 
E         This test fails because scikit-learn has been built without OpenMP.
E         This is not recommended since some estimators will run in sequential
E         mode instead of leveraging thread-based parallelism.
E         
E         You can find instructions to build scikit-learn with OpenMP at this
E         address:
E         
E             https://scikit-learn.org/dev/developers/advanced_installation.html
E         
E         You can skip this test by setting the environment variable
E         SKLEARN_SKIP_OPENMP_TEST to any value.
E         
E       assert False
E        +  where False = _openmp_parallelism_enabled()

base_url   = 'dev'
err_msg    = '\nThis test fails because scikit-learn has been built without OpenMP.\nThis is not recommended since some estimators ...tallation.html\n\nYou can skip this test by setting the environment variable\nSKLEARN_SKIP_OPENMP_TEST to any value.\n'

../1/s/sklearn/tests/test_build.py:33: AssertionError

Probably the issue reported as #20640? Hopefully merging main with #20654 will help.

I made a mistake because the github UI was not up to date and applied some of the review suggestion twice ... I will try to fix. Sorry for the mess.

I will merge as soon as the last CI run passes.

It's green! Merging. Thank you very much @cmarmo for the improved CI setup!

@cmarmo are you interested in a follow-up PR to also migrate the [cd build] part of the ARM-based travis config? Because of weird travis errors, the weekly ARM64 wheel build is broken at the moment and it would be great to get it back up and working.

@cmarmo are you interested in a follow-up PR to also migrate the [cd build] part of the ARM-based travis config?

Yes, I'm interested. I will take care of it ASAP. Thanks @ogrisel for the suggestion.

Nice Thanks @cmarmo

This is great work! Thank you @cmarmo :)