3227: shutdown socket when worker is not alive

[michdan]

More info: #3227

Simply close the waiting reads (pending due to keepalive) and allow the worker to restart without having to wait for the graceful period.

[pajod]

Would prefer a more obviously consistent behaviour for both GreenletExit origins. This looks more like a spot-fix to me. Suggest awaiting / working towards something we can run from pytest that gracefully restarts the worker in each of the possible states (idle, keepalive, busy) and asserts the workers behave consistently and without emitting the warning (so far, I only have a test for auto-reload of idle workers, and not even that one is close to ready for merge).

[michdan]

Would prefer a more obviously consistent behaviour for both GreenletExit origins. This looks more like a spot-fix to me. Suggest awaiting / working towards something we can run from pytest that gracefully restarts the worker in each of the possible states (idle, keepalive, busy) and asserts the workers behave consistently and without emitting the warning (so far, I only have a test for auto-reload of idle workers, and not even that one is close to ready for merge).

@pajod Thanks for the quick response!

This is indeed a quick fix. It's up to you whether you want to release it before making major changes, but this fix would have a positive impact on Gunicorn's performance. It's not about handling GreenletExit, but rather preventing it and making worker restarts faster. To demonstrate this, I've created a repository with some showcases and recorded a screen from those showcases. You can watch it on YouTube here: https://youtu.be/xqv4vGKjGBM
The repo is here: https://github.com/michdan/gunicorn_performance_fix_showcase

[michdan]

I have moved it to a draft because it would require comprehensive testing, there is a risk that it can introduce another bug.
However, the problem described in #3227 is still unresolved

[michdan]

New version soon