Skip to content

[CI] RandomForestRegressor sklearn compatibility check is flaky #8457

Description

@csadorf

Summary

RandomForestRegressor()-check_regressor_data_not_an_array is flaky in the native scikit-learn compatibility suite.

The check creates two estimator clones, assigns the same random state, fits them on equivalent array-like and NumPy inputs, and compares their predictions. The predictions usually differ beyond the check's tolerance, but intermittently agree closely enough for the expected failure to XPASS. Because the suite uses strict xfails, an intermittent XPASS fails CI.

Failing test/component: python/cuml/tests/test_sklearn_compatibility.py::test_sklearn_compatible_estimator[RandomForestRegressor()-check_regressor_data_not_an_array]

Failure observed in:

Environment

  • OS: Ubuntu 22.04
  • Python: 3.12
  • CUDA: 12.2.2
  • Architecture: amd64
  • GPU: L4
  • Dependencies: latest-driver, latest-deps

Test details

FAILED test_sklearn_compatibility.py::test_sklearn_compatible_estimator[RandomForestRegressor()-check_regressor_data_not_an_array]
[XPASS(strict)] Predictions from repeated fits on equivalent inputs can differ beyond the check's tolerance

The check can be exercised repeatedly with:

pytest -q python/cuml/tests/test_sklearn_compatibility.py \
  -k "RandomForestRegressor and check_regressor_data_not_an_array"

Repeated runs can produce both XFAIL and XPASS outcomes.

Expected behavior

Seeded fits on equivalent inputs should produce predictions that consistently satisfy the compatibility check.

Until the underlying instability is resolved, this individual check should remain skipped rather than converted to a non-strict xfail. Once fixed, remove the skip and restore the check.

Metadata

Metadata

Assignees

No one assigned

    Labels

    bugSomething isn't workingci

    Type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions