TST Skip test when unstable openblas configuration

[jjerphan]

Reference Issues/PRs

Temporarily fixes #21361.

What does this implement/fix? Explain your changes.

This is a temporary fix for some tests on main.

This fixes some CI jobs on main.

This is a first fix toward preventing contributors spending their days exploring logs
of issues they don't understand the problems being reported because they aren't
related to their PRs.

As informally proposed by @ogrisel, a long term fix would be to pin the maximum
version dependencies and have a process to update each dependency version
when possible.

Any other comments?

This fix can be made more specific on some values of the parametrisation.

[jjerphan]

Force-pushing to include @adrinjalali in 544d564 as a co-author.

[ogrisel]

not sure if xfail is a good idea for code that can segfault. Maybe skip is a better idea.

[ogrisel]

This was marked resolved without being addressed or commented on: I am pretty sure this is a bad idea to execute code that can segfault, especially when pytest-xdist is not used. XFAIL will run the code, SKIP will not.

[jjerphan]

My mistake multitasking and forgetting to take your suggestion in the patch of this commit.

[adrinjalali]

A green CI is a good one.

[jjerphan]

I force-pushed to:

• remove duplicated commits (hence a correct diff now)
• integrate the forgotten suggestion from @ogrisel

[ogrisel]

The prescott kernel will cause a segfault on any OpenBLAS version when called on non-memory aligned data, whatever the version of OpenBLAS. What changes in old version of OpenBLAS is that the prescott kernel is detected instead of Haswell on SkylakeX machines indirectly causing the bug.

But the bug could still happen if the test is actually run on a prescott machine with a recent version of OpenBLAS.

Instead of introducing a utility function in sklearn.utils with a generic name such as _in_unstable_openblas_configuration, I would rather change the function sklearn.utils.estimators_checks.check_classifiers_train

to call create_memmap_backed_data with aligned=has_prescott_openblas once #21654 is merged, has_prescott_openblas is defined as:

    # OpenBLAS is known to segfault with unaligned data on the prescott architecture
    has_prescott_openblas = any(
        True
        for info in threadpool_info()
        if info["internal_api"] == "openblas"
        and info.get("architecture").lower() == "prescott"
    )

Also for the sake of keeping the git history easier to understand, would it be possible to separate the fix for the calibration plots labels separate from the fix for openblas segfaults? Those are unrelated problems, right? Since we squash merge PRs, I would rather have those 2 fixes in 2 separate commits in the git history, hence 2 separate PRs.

[ogrisel]

remove duplicated commits (hence a correct diff now)

I wrote the previous review against before this fix, please ignore that part of the review.

[ogrisel]

Assuming CI is green, +1 for merge with the last comment below:

[ogrisel]

Green! @adrinjalali still +1 for merge with the new way to fix this?

[adrinjalali]

A much cleaner solution indeed

[ogrisel]

Actually I just found out that it does not work :)

$ OPENBLAS_CORETYPE=Prescott pytest -v sklearn/tests/test_common.py -k "check_classifiers_train and svc"
[...]
    def create_memmap_backed_data(data, mmap_mode="r", return_folder=False, aligned=False):
        """
        Parameters
        ----------
        data
        mmap_mode : str, default='r'
        return_folder :  bool, default=False
        aligned : bool, default=False
            If True, if input is a single numpy array and if the input array is aligned,
            the memory mapped array will also be aligned. This is a workaround for
            https://github.com/joblib/joblib/issues/563.
        """
        temp_folder = tempfile.mkdtemp(prefix="sklearn_testing_")
        atexit.register(functools.partial(_delete_folder, temp_folder, warn=True))
        if aligned:
            if isinstance(data, np.ndarray) and data.flags.aligned:
                # https://numpy.org/doc/stable/reference/generated/numpy.memmap.html
                filename = op.join(temp_folder, "data.dat")
                fp = np.memmap(filename, dtype=data.dtype, mode="w+", shape=data.shape)
                fp[:] = data[:]  # write data to memmap array
                fp.flush()
                memmap_backed_data = np.memmap(
                    filename, dtype=data.dtype, mode=mmap_mode, shape=data.shape
                )
            else:
>               raise ValueError("If aligned=True, input must be a single numpy array.")
E               ValueError: If aligned=True, input must be a single numpy array.

[ogrisel]

I will open a PR with a quick fix for the above.