`test_worker_thread_context_not_leaked` sometimes fails

[A5rocks]

cc @EmmanuelNiyonshuti

I haven't dug into this yet but the test is flakey. I'm opening this instead of simply commenting in the flakey-tests thread so we can discuss!

See https://github.com/A5rocks/trio/actions/runs/30208152029 and https://github.com/A5rocks/trio/actions/runs/29643389046

[A5rocks]

First impression: this is failing on PyPy particularly, so it's probably just a GC thing. Actually, the fix is probably pretty simple. I guess this issue was unnecessary :^)

Strangely, I can't reproduce it locally. Maybe it requires a slow machine or something? I'll toss up the PR anyways.

[EmmanuelNiyonshuti]

@A5rocks in anyio, we actually skip testing on pypy but I thought it was because of gc.get_referrers and simplified the test and removed it?

[A5rocks]

Yeah. I mean it's better IMO not to skip; something like this is annoying but very simple to fix + then it's codified in the test case!

[A5rocks]

Wait, it actually fails on CPython!!! https://github.com/A5rocks/trio/actions/runs/30276990757/job/90013769036#step:4:1300

[EmmanuelNiyonshuti]

Wait, it actually fails on CPython!!! https://github.com/A5rocks/trio/actions/runs/30276990757/job/90013769036#step:4:1300

thats weird! I was testing it earlier today.

[EmmanuelNiyonshuti]

Wait, it actually fails on CPython!!! https://github.com/A5rocks/trio/actions/runs/30276990757/job/90013769036#step:4:1300

thats weird! I was testing it earlier today.

I'm not able to see the test in question failing; how are you able to see that? All I get from the linked run is: Error: Process completed with exit code 1

[A5rocks]

I see this in the logs:

  ================================== FAILURES ===================================
  ____________________ test_worker_thread_context_not_leaked ____________________
  
      async def test_worker_thread_context_not_leaked() -> None:
          # Regression test for: https://github.com/python-trio/trio/issues/3472
      
          class Foo:
              pass
      
          def sync_fn() -> None:
              pass
      
          cvar: contextvars.ContextVar[Foo] = contextvars.ContextVar("cvar")
          contextval = Foo()
          ref = weakref.ref(contextval)
          cvar.set(contextval)
          await to_thread_run_sync(sync_fn)
          cvar.set(Foo())
      
          del contextval
          if sys.implementation.name == "pypy":
              import gc
              gc.collect()
      
  >       assert ref() is None
  E       AssertionError: assert <trio._tests.test_threads.test_worker_thread_context_not_leaked.<locals>.Foo object at 0x000001932E4CB740> is None
  E        +  where <trio._tests.test_threads.test_worker_thread_context_not_leaked.<locals>.Foo object at 0x000001932E4CB740> = <weakref at 0x000001932E4FADE0; to 'Foo' at 0x000001932E4CB740>()
  
  C:\hostedtoolcache\windows\Python\3.12.10\x64\Lib\site-packages\trio\_tests\test_threads.py:711: AssertionError
  ------------- generated xml file: D:\a\trio\trio\test-results.xml -------------
  ============================ slowest 10 durations =============================
  0.01s call     _tests/test_threads.py::test_worker_thread_context_not_leaked
  

(notably I changed gc.collect to only happen on pypy, so this is possible if there's a refcycle or something. but.... there shouldn't be a refcycle that only happens like 1% of the time?)

[EmmanuelNiyonshuti]

I see! Are you running the full test suite on free-threaded Python as well? Either way, will need a way to run a full collection gc.collect() on CPython. Otherwise, we will have that problem on free-threaded Python.

[EmmanuelNiyonshuti]

But if you are seeing this problem only on CPython with GIL enabled, well, I think we will have to investigate the worker thread then.