RuntimeWarning "invalid value encountered in reciprocal" when running tests on macOS

[ogrisel]

This test is unexpectedly raised when running test_linear_model_normalize_deprecation_message when numpy is compiled with a recent but not too recent versions of clang on macOS.

This use to make the tests fail only on Apple Silicon M1 machine but now it also happens with macOS Intel machines on the CI for some reason.

The problem was already reported upstream as and fixed in numpy (fix to be released in numpy 1.22.0):

• RuntimeWarning: divide by zero encountered in reciprocal for np.ones(3) ** -1 numpy/numpy#18555
  • fixed by BUG: Resolve Divide by Zero on Apple silicon + test failures numpy/numpy#19926
• numpy 1.20 on MacOSX: spurious RuntimeWarning: invalid value encountered in reciprocal conda-forge/numpy-feedstock#229

Example failure on our CI:

https://dev.azure.com/scikit-learn/scikit-learn/_build/results?buildId=33848&view=logs&j=97641769-79fb-5590-9088-a30ce9b850b9&t=4dc8e08b-9a37-5d92-b6d2-246a6a63d51e

I can reproduce locally. Will work on a workaround to make those tests XFAIL on macos.

[ogrisel]

We might need to backport this to 1.0.X to make it possible to release macos packages on conda-forge without test failures.