Question: Is it possible to create new subprocess for every test?

[kittipatv]

Hi,
I use -d --tx 2*popen//python=python3.7 to try to run my tests in subprocesses. However, it looks like the subprocess is reused between tests. I want each test to run in a fresh subprocess because I suspect that there are some memory leaking between tests. I tried using --forked but it had problem finding dependencies. Is there some flag I could pass to restart subprocess after each test?

Thank you,
Kittipat

[RonnyPfannschmidt]

its not implemented as of now

[kittipatv]

@RonnyPfannschmidt Thank you for quick response. Is that something that could be implemented in current framework?

[RonnyPfannschmidt]

its going to carry a certain kind of complexity as of now

[craigds]

I may need something like this for my work on GDAL

Something like a --max-tests-per-worker=1 flag would suffice (default 0=unlimited). I'd want to use it in conjunction with --dist=loadmodule to run each module in a subprocess rather than each individual test.

@RonnyPfannschmidt do you think it would be hard to do? I would envisage the parent process could just restart worker processes when they exit successfully, same as it restarts them for failures at present.

Happy to submit a PR if you think this is doable/worthwhile; will probably require some help to get it off the ground since I haven't worked on pytest internals before.

[RonnyPfannschmidt]

@craigds i have no idea how hard it would be, i would strongly suggest to start by identifying smaller structural enhancements as shoehorning it in might create certain kinds of trouble

[JuroOravec]

Hi, I've got the same need as @craigds - I'd like to give tests to workers per files, and then restart the worker after the tests from the file/batch are done. I assume that this way each test file would be an isolated subprocess.

My reason for this is that I'm having headaches from managing different Django settings across different test files, because Django settings are global.

For now I've somewhat resolved the issue. But in the meantime I tried to see if I could modify pytest-xdist to add support for this.

Please see #1254.

It's a proof of concept make with Gemini + Cursor. I wonder if it's on the right track? If so I can develop it fullly.

Outstanding questions / thoughts:

1. Is this a valid strategy to force workers to reset after each test / group / scope / file?
2. In the draft PR the CLI option was --dist-reset-workers, but I think @craigds's --max-tests-per-worker=1 is better design choice.
3. Do I need to send separate command, or could we simply reuse and send the shutdown command?
4. Which distmodes should this --max-tests-per-worker=1 work with?

[RonnyPfannschmidt]

for one test per worker we should have a new distmode alltogether