Trying to isolate/debug I noticed that the default backlog for unix sockets is nil

Hm, that does seem wrong.

@eprothro I talked to Evan about this, he said this behavior is not intentional and is definitely a bug. We should change the default Unix backlog to 1024.

Hi all, I have updated to latest version of Puma (3.12.0) and I'm still having the same issue: Nginx returns 502 (Bad Gateway) error codes, and I see error messages like this in the logs:

connect() to unix:/path/to/socket/puma.sock failed (11: Resource temporarily unavailable) while connecting to upstream, client...

It only happens when connecting Nginx and Puma using Unix Sockets and under heavy load (300 new connections per second). I have tried setting the backlog to 1024 in the configuration file:

bind 'unix:///tmp/run/puma.sock?backlog=1024'

And I have also tried to leave it empty, which should default the value to 1024 as well, but nothing works.

Any idea?

Any news regarding the 5xx error on load?

I believe the behaviour is influenced by the linux kernel parameters too. I had the same problem but I was able to resolve it as follows:

I set bind 'unix:///tmp/run/puma.sock?backlog=1024' as instructed

I took a look at this example: https://easyengine.io/tutorials/linux/sysctl-conf and changed the default somaxconn from 128 to 4096.

You can see the current somaxconn with cat /proc/sys/net/core/somaxconn

I also used https://github.com/wg/wrk to load test my demo app

Used the default example wrk -t12 -c400 -d30s http://127.0.0.1:3000/

For starters, I ran only one puma worker behind nginx; the response was very slow but I got no errors in /var/logs/nginx/error.log

References:
http://www.tweaked.io/guide/kernel/
https://easyengine.io/tutorials/linux/sysctl-conf
https://gist.github.com/perusio/2154235
https://easyengine.io/tutorials/linux/sysctl-conf

Hi @ernestojpg I'm kinda facing the same issue, and the solution of nkokkos doesn't work for me. I guess there's something else involved (linux kernel settings, linux networking settings...) that I didn't find out yet. Did you resolve your issue? If you did, could you share how? 🙏 Thanks ☺️

any more infos on this? when i set worker to 1 no error with worker 2 i get the same errors any idea how to debug this more?

Same issue here, adding ?backlog=1024 didn't help, any ideas?

There seems to be similar problem here: https://gitlab.com/gitlab-org/omnibus-gitlab/-/issues/5752

From what I read here: https://gitlab.com/gitlab-org/omnibus-gitlab/-/issues/5752, it is resolved by setting somaxconn to puma default backlog. So if you set the same number to both somaxconn and backlog, you should be ok?

Try this and see if it works:

echo 1024 > /proc/sys/net/core/somaxconn

stop your puma server

set bind 'unix:///tmp/run/puma.sock?backlog=1024' in puma.rb

start your puma server

See this too: https://gitlab.com/gitlab-org/omnibus-gitlab/-/merge_requests/4688 :
set net.core.somaxconn kernel parameter to a value equal or higher than the default Puma backlog_socket.