contextlib.redirect_stdout/stderr does not work with `pytest-run-parallel`

[15r10nk]

Bug report

Bug description:

redirect_stdout/stderr does not work if freethreading is enabled.

The following example shows the problem:

# /// script
# dependencies = [
# "pytest",
# "pytest-run-parallel"
# ]
# ///


from contextlib import redirect_stdout
from io import StringIO
import time


def test_redirect():
 text=StringIO()
 with redirect_stdout(text):
 print("hello")
 time.sleep(1)
 print("hello")

 assert text.getvalue()=="hello\nhello\n"


if __name__ == "__main__":
 import pytest
 pytest.main(["--parallel-threads=5",__file__])

output:

❯ uv run -p 3.13t bug.py
================================== test session starts ==================================
platform linux -- Python 3.13.1, pytest-8.3.4, pluggy-1.5.0
rootdir: /home/frank/projects/bugs/pytest_redirect
plugins: run-parallel-0.3.1
collected 1 item 

bug.py F [100%]

======================================= FAILURES ========================================
_____________________________________ test_redirect _____________________________________

 def test_redirect():
 text=StringIO()
 with redirect_stdout(text):
 print("hello")
 time.sleep(1)
 print("hello")
 
> assert text.getvalue()=="hello\nhello\n"
E AssertionError: assert 'hello\n' == 'hello\nhello\n'
E 
E hello
E - hello

bug.py:22: AssertionError
================================ short test summary info ================================
FAILED bug.py::test_redirect - AssertionError: assert 'hello\n' == 'hello\nhello\n'
=================================== 1 failed in 1.04s ===================================

The underlying problem is probably the global sys.stdout/stderr state which is shared between threads.
Is it possible to make these thread local?

CPython versions tested on:

3.13

Operating systems tested on:

No response

[picnixz]

The underlying problem is probably the global sys.stdout/stderr state which is shared between threads. Is it possible to make these thread local?

I'm not sure we can do it. AFAIK, those standard streams are at the interpreter state level. I also don't know why we only have a single write and not more. It's as if one thread is not doing anything or something was not caught at all (I would not be surprised if for instance we had more "hello" but I'm surprised we only have one)

cc @ZeroIntensity

[15r10nk]

The reason for the one hello is that the second thread changes sys.stdout and this thread has more than two hellos. But I get only the first assertion with one hello.

[picnixz]

Ah yes I see! ok nvm it was dumb of me.

[ZeroIntensity]

Is it possible to make these thread local?

I don't think so, at least without making it a huge breaking change. If we were to make sys.stdout/sys.stderr thread-local, we'd break all programs that currently rely on that for catching data from all threads. In theory, I guess we could have some sort of opt-in system to make stdout and stderr thread-local using some dirty hacks, but it would be a lot of work and I'm not convinced it's worth it. I'd just suggest synchronizing your tests that use redirect_stdout with a lock.

[15r10nk]

I don't think so, at least without making it a huge breaking change.

That is what I expected, but I wanted to ask.

I'd just suggest synchronizing your tests that use redirect_stdout with a lock.

I would have to synchronize all Tests which are using sys.stdout. Every test which is using print could use the wrong output.
I see no way how redirect_stdout can be used in a multi threaded environment.