I can confirm that allowing a longer backlog queue (in unicorn config + net.core.somaxconn and net.core.netdev_max_backlog) did solve the issue (for me).
I posted a bit more details at http://stackoverflow.com/a/15544229/1061997. Posted at Nginx Forum: http://forum.nginx.org/read.php?2,237573,237622#msg-237622 _______________________________________________ nginx mailing list [email protected] http://mailman.nginx.org/mailman/listinfo/nginx
