ruby on rails - Nginx proxying websockets, connections not closed

I'm confused about where the problem is located, but basically I have nginx proxying websocket connections to a backend ruby thin server, which services the connections with the websocket-rails module in a Ruby on Rails application. Which all works fine except for that a lot of sockets, maybe all of them, don't get closed, so the thin server relatively quickly runs out of file descriptors.

I'm using nginx 1.4.2 and this is my config:

map $http_upgrade $connection_upgrade {
    default upgrade;
    ''      close;
}
server {
    listen       my.ip.num.ber:80;  
    server_name admin3.mydomain.com;
    root /home/apps/mydomain/current/public;
    try_files $uri/index.html $uri @admin3.mydomain.com;  

    access_log  /var/log/nginx/admin3.access.log  combined;
    error_log  /var/log/nginx/admin3.error.log error;
    location /websocket {  
        proxy_redirect off;
        proxy_pass http://localhost:3008;
        proxy_http_version 1.1;
        proxy_set_header Upgrade $http_upgrade;
        proxy_set_header Connection $connection_upgrade;
        proxy_set_header Host $host;
        proxy_set_header X-Real-IP $remote_addr;

        proxy_set_header X-Forwarded-For $proxy_add_x_forwarded_for;
        keepalive_timeout 90;
        proxy_connect_timeout 10;
        proxy_read_timeout 60;
        proxy_send_timeout 60;
    }
}

I'm using thin 1.5.1 and this is the configuration:

port: 3008
user: ploy
group: ploy
pid: /home/apps/mydomain/shared/pids/thin.pid
timeout: 90
wait: 30
log: /home/apps/mydomain/shared/log/thin.log
max_conns: 1024
require: []

environment: production
max_persistent_conns: 512
servers: 1
threaded: false
#no-epoll: false
daemonize: true
chdir: /home/apps/mydomain/current
tag: admin3

There's only a couple of dozen active websocket connections at a time, and they seem to be established and terminated fine from the perspective of a client browser or the websocket-rails backend. But the thin server ends up with 1025 open file descriptors, mostly sockets.

ls -l /proc/`ps aux | grep "thin server" | grep -v grep | head -n 1 | awk '{print $2}'`/fd

gives this kind of thing:

lrwx------. 1 root root 64 Aug 31 15:15 993 -> socket:[1319549665]
lrwx------. 1 root root 64 Aug 31 15:15 994 -> socket:[1319549762]
lrwx------. 1 root root 64 Aug 31 15:15 995 -> socket:[1319549850]

lrwx------. 1 root root 64 Aug 31 15:15 996 -> socket:[1319549974]
lrwx------. 1 root root 64 Aug 31 15:15 997 -> socket:[1319846052]
lrwx------. 1 root root 64 Aug 31 15:15 998 -> socket:[1319549998]
lrwx------. 1 root root 64 Aug 31 15:15 999 -> socket:[1319550000]

A similar thing seems to subsequently happen for nginx:

ls -l /proc/`ps aux | grep "nginx: worker" | grep -v grep | head -n 1 | awk '{print $2}'`/fd

although the number of socket file descriptors creeps up more slowly and it takes it much longer to get to the 1025. As a matter of fact, I only saw that once.

So, I'm a little at a loss at identifying if there's a problem with my nginx config, or with thin, or it is something in the websocket-rails backend. I hope some of your trained eyes might see something obviously wrong, even if you're not familiar with the backend pieces.