New multiprocess manager

[abersheeran]

Summary

About #2164.

The multiprocess manager introduced by this PR includes process keep-alive and process hung detection. It also imitates gunicorn and uses hup, ttin, and ttou signals to control child processes.

Checklist

• I understand that this PR may be closed in case there was no previous discussion. (This doesn't apply to typos!)
• I've added a test for each change that was introduced, and I tried as much as possible to make a single atomic change.
• I've updated the documentation accordingly.

[Kludex]

I'm not willing to introduce this feature without testing it.

[abersheeran]

Yes, I just want to see your attitude towards this PR before adding the test. Because I remember you have implemented something similar before.

[Kludex]

Ah, then let me review and get back to you.

[Kludex]

Checklist

• Increase coverage for the new multiprocess manager to 100% (if tests are slow, run on a separate job in the CI, and only if the multiprocess.py is changed).
• Docstrings.
• Documentation (I can help here).
• Answer the question: how this should interact with --reload?
• Move UvicornWorker out from Uvicorn's source code (I'll do it).

[Kludex]

@abersheeran I'm fully available to move this forward. 👍

[Kludex]

Pinging @ahopkins to review the initial steps since he offered helped on #2164. 🙏

[Kludex]

Ok. 👍

We can move forward with the checklist above 🙏

[ahopkins]

I see what the PR is doing, but I somewhat fail to understand the why and the goals. Is the pinging the ultimate feature add here?

[abersheeran]

Also, if I press CTRL + C on uvicorn main:app --workers 2, it doesn't work as expected.

Can you try the latest commit version? This should solve the problem.

[baseplate-admin]

Please use the English version of the documentation instead of chinese version

[Kludex]

I think the PID should be shown in the "uvicorn.access" logger (only when workers > 1).

Is it normal for the child processes to be alive after I kill the manager process?

[abersheeran]

I think the PID should be shown in the "uvicorn.access" logger (only when workers > 1).

Here I just reused the previous logger. If we want to refactor the logs, I think it can be done in another PR.

Is it normal for the child processes to be alive after I kill the manager process?

No, normally only killing the manager process through the SIGKILL signal will result in the survival of child processes (this is due to system signal mechanism reasons, and it also exists in all other multi-process managers), under other circumstances, child processes should be killed.

[Kludex]

Thanks @abersheeran 🙏