MAINT reduce scope of test_linear_models_cv_fit_for_all_backends to reduce CI usage

[ogrisel]

Alternative to #21907.
Towards: #21407.

Do not test for the threading backend which was not involved in the original problem.
Do not test for the multitask variants that are much more computationally intensive. LassoCV and ElasticNetCV should be enough to cover as a non-regression test for the original problem.
Use the minimal dataset for the fewest number of features to trigger memmaping in joblib.

[thomasjpfan]

I left comment about future API for joblib.

Overall this PR LGTM.

[thomasjpfan]

Do you think there should be an API for changing max_nbytes in the inner Parallel call? Something like:

with parallel_backend("loky", max_nbytes=1000):
    results = Parallel(n_jobs=4)(delayed(func)(x, y) for x, y in data)

[ogrisel]

Yes that would be nice but unfortunately this conflicts with the current joblib backend API design and dealing with backward compat is...

[thomasjpfan]

If it is hard for joblib API wise, what do you think about adding a parallel_kwargs parameter to estimators that creates a Parallel object?

(I know this is a little counter to how we have been removing pre_dispatch from the estimator's __init__.)

[ogrisel]

We should probably find a way to change the parallel_backend to detect Parallel kwargs and treat time specifically instead of passing them as constructor arguments for the backend.

[ogrisel]

@jeremiedbb you might want to give this PR a second review.

[jeremiedbb]

This test is a non regression test for a fix regarding using the loky backend. By default these estimators use the threading backend which has always been working. We need to keep testing with the "loky" backend

[ogrisel]

This test is a non regression test for a fix regarding using the loky backend. By default these estimators use the threading backend which has always been working. We need to keep testing with the "loky" backend

Indeed, good catch, I had not realized. Let's see if the CI stays green.

[jeremiedbb]

LGTM. Do you know how much time it saves ?

[ogrisel]

When running sklearn/linear_model/tests/test_coordinate_descent.py in with pytest-xdist in oversubscription context (-n 8 on a 8 core machine):

• on main:

================================================================== slowest 10 durations ===================================================================
7.37s call     sklearn/linear_model/tests/test_coordinate_descent.py::test_linear_models_cv_fit_for_all_backends[MultiTaskElasticNetCV-threading]
6.97s call     sklearn/linear_model/tests/test_coordinate_descent.py::test_linear_models_cv_fit_for_all_backends[MultiTaskLassoCV-threading]
5.18s call     sklearn/linear_model/tests/test_coordinate_descent.py::test_linear_models_cv_fit_for_all_backends[MultiTaskElasticNetCV-loky]
4.70s call     sklearn/linear_model/tests/test_coordinate_descent.py::test_linear_models_cv_fit_for_all_backends[MultiTaskLassoCV-loky]
1.10s call     sklearn/linear_model/tests/test_coordinate_descent.py::test_linear_models_cv_fit_for_all_backends[ElasticNetCV-loky]
... other unrelated tests

• on this branch:

================================================================== slowest 10 durations ===================================================================
0.89s call     sklearn/linear_model/tests/test_coordinate_descent.py::test_linear_models_cv_fit_with_loky[ElasticNetCV]
0.88s call     sklearn/linear_model/tests/test_coordinate_descent.py::test_linear_models_cv_fit_with_loky[LassoCV]
... other unrelated tests