How does gunicorn handle connection draining / graceful termination?

[tnusser]

When running my FastAPI app via uvicorn and gunicorn on kubernetes, I sometimes have the issue of some requests failing during a rolling update. I suspect there is either an issue on how gunicorn is handling shutdown signals or how we setup our readiness and liveness checks.
More specifically, when k8s sends a SIGTERM how is this distributed from gunicorn to the uvicorn workers and is there a connection draining mechanism? Or do I just have to set the kubernentes gracefulTermination flag to higher values (default 30s). Somehow I have the feeling that even though gunicorn gets a SIGTERM, new requests are still being processed OR old ones are not able to finish before the SIGKILL is issued or the gracefulTermination flag is not propagated to the workers.
Are there any specific flags or mechanisms on how to handle that? E.g. also regarding background tasks or longer running async db calls etc?

[benoitc]

this is documented there: https://docs.gunicorn.org/en/stable/signals.html