[jameslamb]

More changes to estimator checks have been made in scikit-learn, and those broke CI here.

CI is failing like this when using the latest scikit-learn nightlies:

=========================== short test summary info ============================
FAILED tests/python_package_test/test_sklearn.py::test_sklearn_integration[LGBMClassifier()-check_sample_weight_equivalence_on_dense_data]
FAILED tests/python_package_test/test_sklearn.py::test_sklearn_integration[LGBMClassifier()-check_sample_weight_equivalence_on_sparse_data]
===== 2 failed, 767 passed, 40 skipped, 2 xfailed, 584 warnings in 36.86s ======

(build link)

Looks like scikit-learn/scikit-learn#30137 removed check_sample_weight_equivalence, in favor of 2 new functions names check_sample_weight_equivalence_on_{dense,sparse}_data. This updates lightgbm's xfail_checks tag to include all 3 names, so those tests are correctly xfail'd for the full range of scikit-learn versions lightgbm supports.

update: this also removes the R 3.6 Linux CI job to fix another blocking CI issue, see #6733 (comment)

[jameslamb]

Noticing that the R 3.6 CI job has been failing for a couple days like this:

Post job cleanup.
/usr/bin/docker exec 0f363447ec94ed26e9ae21b628f3850fe9f1e15a4c8a341b96fcdc540018e374 sh -c "cat /etc/*release | grep ^ID"
/__e/node20/bin/node: /lib/x86_64-linux-gnu/libc.so.6: version `GLIBC_2.28' not found (required by /__e/node20/bin/node)

(build link)

That's happening here:

https://github.com/microsoft/LightGBM/blob/83c0ff3de1925b0e2d4831a9ccb6ffc196aa795b/.github/workflows/r_package.yml#L176-L180

Will have to fix that in this PR too. I'll try that.

Related prior conversations:

• checkout@v4 is not working in containers actions/checkout#1487
• [R-package] [ci] newer versions of actions/setup-pandoc are incompatible with Ubuntu 18.04 #6298
• /__e/node20/bin/node: /lib/x86_64-linux-gnu/libc.so.6: version GLIBC_2.28' not found (required by /__e/node20/bin/node)` actions/checkout#1590 (comment)

[jameslamb]

Trying to fix #6733 (comment), it seems that the workaround from actions/checkout#1590 (comment) is no longer working.

See actions/checkout#1590 (comment).

Between this and all the work that was required with package-pinning in #6718, I think we should just fully stop testing R 3.6 on Linux in CI here. As we did for macOS in #5859.

Note that we'll still have test coverage on Windows:

https://github.com/microsoft/LightGBM/blob/83c0ff3de1925b0e2d4831a9ccb6ffc196aa795b/.github/workflows/r_package.yml#L81-L87

https://github.com/microsoft/LightGBM/blob/83c0ff3de1925b0e2d4831a9ccb6ffc196aa795b/.github/workflows/r_package.yml#L95-L102

Dropping this testing gets us the following:

• reduces CI configuration complexity (see all the code removed in this PR)
• allows upgrading to actions/checkout@v4 (no more skipping PRs like [ci]: Bump actions/checkout from 3 to 4 in the ci-dependencies group #6712)
• reduced ongoing maintenance burden, via avoiding things like these:
  • [python-package][R-package] adapt to scikit-learn 1.6 testing changes, pin more packages in R 3.6 CI jobs #6718
  • [R-package] [ci] newer versions of actions/setup-pandoc are incompatible with Ubuntu 18.04 #6298
  • [ci] use an MRAN snapshot for R 3.6 Linux jobs (fixes #5898) #5903

[StrikerRUS]

LGTM! Thank you very much for quick fix! Nice to see this simplification diff for R.

[jameslamb]

Thanks very much for the quick review! Merging this to unblock CI.

[github-actions]

This pull request has been automatically locked since there has not been any recent activity since it was closed.
To start a new related discussion, open a new issue at https://github.com/microsoft/LightGBM/issues including a reference to this.