gh-102402: Fix logging test_relativeCreated_has_higher_precision() leak

[vstinner]

Fix a reference leak in test_relativeCreated_has_higher_precision() of test_logging: don't reimport the logging the logging module.

• Issue: Logging's msecs doesn't handle "100ms" well. #102402

[vstinner]

@vsajip @serhiy-storchaka: Would you mind to review my fix?

[serhiy-storchaka]

It looks like an intermediate version of this test. I did not like it. It relies on implementation details -- that _startTime exists, that it is initialize with the time as nanoseconds, and that there are no other variables that depend on the starting time. If we rely so much on these details, we could simply not add this test at first place.

I tried to re-import a new logging module with its submodules. It seems not working too. Perhaps the only solution is to run a script which patches time_ns() before the first import of logging.

It is still unsolved mistery why re-importing the logging module leaks.

[vstinner]

It relies on implementation details -- that _startTime exists, that it is initialize with the time as nanoseconds, and that there are no other variables that depend on the starting time.

The test can be marked as @cpython_only if needed. The test already makes the assumption that the logging module uses time.time_ns(). What if it uses time.time() on PyPy, or what if it uses time.monotonic_ns(), etc.?

Creating a new module instance is causing too much trouble. For example, it calls os.register_at_fork() which cannot be unregistered, unless the function is mocked as well. That's just an example: commenting this code is not enough to fix the leak.

[vstinner]

The test can be marked as @cpython_only if needed.

Done.