The environment variable it not meant to be used this way. Please use the global_random_seed pytest fixture as explained in the issue #22827 instead of manually inspecting the environment variables.

Please carefully read the guide lines to update existing test that I wrote in the description of #22827 and also have a look at other merged PRs linked in that issue for examples on how to use the fixture.

You can also have a look at the documentation of the fixture itself to run the fixtured tests locally:

https://scikit-learn.org/dev/computing/parallelism.html#sklearn-tests-global-random-seed

If you want to know more details on how the fixture is coded, please have a look at its source code:

https://github.com/scikit-learn/scikit-learn/blob/main/sklearn/tests/random_seed.py

@ogrisel Ah makes sense. I had a feeling, it did not seem idiomatic. I think the difficulty for these tests are that they depend on top-of-scope variables. They are generated without knowledge of theglobal_random_seed fixture. I guess the fix would be to generate those variables (from rng) using the fixture within each test? Otherwise, perhaps we can define another fixture to pass to the tests to make it more DRY?

I think it's fine to have a bit of repetition in this case.