From what I understood, the problem is that creating a SemLock has an impact on the resource tracker.

The failing unit tests checks something about the resource tracker.

The test runs a function in a child process. The function is defined in _test_multiprocessing.py. _test_multiprocessing.py calls skip_if_broken_multiprocessing_synchronize() which creates a SemLock(). At the end, the child process modifies the resource tracker because it creates a SemLock().

There are different options to fix the issue:

* Modify skip_if_broken_multiprocessing_synchronize() only only attempt to create a SemLock on Linux.
* Modify skip_if_broken_multiprocessing_synchronize() to only check if the current user is allowed to write into /dev/shm/ directory: that would allow touching the resource tracker.
* Modify the test to avoid improting _test_multiprocessing in the child process.

I don't know why the test only fails on macOS.

Note: I ignored macOS failure when I merged my PR because the macOS job of GitHub Action is failing randomly for at least 3 months and nobody attempted to fix it. So I learnt to ignore this job.