So the symptom appears to be a missing teardown of some fixtures in case of an error? I think we might be able to make things more flexible by using finalizers from pytest: https://docs.pytest.org/en/6.2.x/fixture.html#adding-finalizers-directly

They are essentially callbacks that are guaranteed to be called during teardown. That could be used as a last-ditch attempt to close the FDs even though the node_factory fixture was not torn down completely due to a cleanup error.

So the symptom appears to be a missing teardown of some fixtures in case of an error?

Actually, the FDs stick around even if the test passed successfully. lsof reports 4 FDs for each test (in a deleted state), but they hang out while the entire test suite runs synchronously. When running in parallel using PYTEST_PAR, they clean themselves up.

I will review the finalizer documentation and try some experiments.

edit: If it was unclear, the subset of local test failures reported in the issue were against my initial changes which closed the files in the TailableProc::stop() function, not the current master branch.

Updated issue description for clarity by removing references to earlier prototype code cross-referenced in #7108 (comment) and describing test results.