2048 worker_connections are not enough while connecting to upstream

[Maxim Dounin]

Hello!

On Tue, May 10, 2011 at 10:31:06AM +0100, John Moore wrote:

> I'm using nginx as a remote proxy server in front of some Tomcat
> instances, with pretty high traffic. Suddenly yesterday afternoon we
> started getting the error in the message subject, one I'd never seen
> before, and it's occurred tens of thousands of times since. Now,
> I've upped the worker_connections count to a much higher figure and
> the problem has gone away for now. But I'm concerned that it might
> be indicative of something else - some kind of DOS attack, maybe?
> What does this signify and why might we suddenly start getting it?
> Thanks.

Without proper history graphs (use stub_status, Luke!) it's hard 
to say if you are under attack or just your normal load reached 
the bar (or your backend response time degraded and this resulted 
in more connections used, or your network uplink has problems and 
this caused more connections...).

On the other hand, 2048 is pretty low, and easily reached even 
without keepalive.  Usual production values for "high traffic" 
sites is over 9000.

Please also note that using high worker_connections may also 
require tuning of your OS to allow appropriate number of file 
descriptors to be used.

Maxim Dounin