add a delay between SIGABRT and SIGKILL for worker timeouts

[tardyp]

candidate fix for #2454

[tilgovi]

Thank you for the patch, but I don't think this is the way I'd like to handle this issue. This function is meant for killing workers that are stuck. A proper implementation of graceful shutdown in the async workers should continue to notify the arbiter of their liveness during shutdown. Not all workers are doing this correctly right now. There is some discussion in #2317 where I'm working on fixing this.

I'll close your PR, but if you think I've misunderstood I'll be happy to discuss further.

[tardyp]

Hi @tilgovi, I indeed think there is a misunderstanding.

This fix is in reality for the worker stuck use case.
When this is the case, there is no delay between the SIKABRT and SIGKILL, because of the logic I am fixing.

In #2454, you say there is 1 second delay, please point me to the actual code that is introducing this delay, because I can't find it, and in my tests and logs I never see a one second delay between those two signals.

[tilgovi]

@tardyp the arbiter process is supposed to sleep for one second if it does not receive any signals: https://github.com/benoitc/gunicorn/blob/master/gunicorn/arbiter.py#L357

We could re-open this PR and modify it if you find that this is not reliable enough, but we should not wait the whole graceful shutdown period since abort is not graceful.