[1SaltwaterC]

I can provide you some pointers as we made it go away. Some data from one of our production virtual hosts:

  cat access.log.1 | wc -l
  1054423 # no static objects served here
  cat error.log.1 | grep timed

For PHP-FPM we use static pm and unix domain sockets. This virtual host if fairly busy with some slow (~200 ms) requests, therefore it uses 96 processes per pool. listen.backlog = -1 in php-fpm.ini for letting the kernel decide the size of the actual connection backlog. UDS is getting filled faster than TCP and nginx starts responding with 502's. Throw net.core.somaxconn = 65535 somewhere in /etc/sysctl.d for increasing the actual backlog since even if you specify a high listen.backlog value, the actual value is truncated to SOMAXCONN. Couple of years ago I wrote an article about stuff like this: http://www.saltwaterc.eu/nginx-php-fpm-for-high-loaded-websi... (shameless plug, I know, but you may get some useful info). As a side note, I am curious how the backend persistency for nginx is playing. Our production still uses the same config since 2011 as it isn't broken, but it may be more efficient.

[pavs]

I tried your suggestions listen backlog and somaxconn, are they dependent on unix socket or do I get benefit from it either way?

Thanks.

[1SaltwaterC]

Both TCP and UDS depend on it. UDS uses an API that follows the BSD sockets standard. However, TCP makes sense when you use nginx as load balancer between multiple PHP-FPM backends. A really tricky setup if you ask me. For UDS, besides lower latency, the namespace is cleaner (filesystem paths vs numeric ports - much easier to automate the configuration), and (at least under Linux) they follow the filesystem ACL. For example, under my setup, only nginx's user is allowed to read / write to the PHP-FPM sockets.