I noticed that if you have too few pm_children set then some requests hang until timeout. This surprised me - I’d expect an immediate error, but it’s more like a tarpit! For ages I was thinking my server was not performant, then I noticed via top that it wasn’t doing or waiting while the browser was.
I have two questions:
-
If you have
pm_max_children=1and you occupy that and submit another request, what actually happens? (I’m proxying through nginx.) HTTP doesn’t have a “40_ Come back later”. -
(if life deals you lemons…) if you can generate a tarpit that doesn’t use server resources, this could be quite useful to know about too!
FPM uses a request queue, so when all workers are busy they wait in line until a worker frees up to process it. With just 1 worker, this can take a long time to catch up, but with a reasonable number of workers you’ll only wait a few seconds or even milliseconds depending on the app and how fast it can process requests.
The setting for that is
listen.backlogwhich defaults to 511 on Linux.Right Thanks. So I think php might queue then forget some requests, because I’ve had this behaviour without too much traffic. I just don’t know how to reproduce it, but maybe I’ll try with that listen backlog setting, thanks for tip.
It doesn’t forget requests, but NGINX will drop them after it times out if it takes too long.
