gistfile1.txt

upstream backends {
     server 127.0.0.1:8081;
     server 127.0.0.1:8082;
}

map $request_method $nonidempotent {
     'POST' 1;
     'PUT' 1;
     'DELETE' 1;
     default 0;
}

server {
     listen 80;
     location / {
         if ($nonidempotent) {
             rewrite ^ @nonidem last;
         }
         proxy_pass http://backends;
     }
     location = @nonidem {
         internal;
         proxy_next_upstream off;
         proxy_pass http://backends$request_uri;
     }
}

Hm, where is @nonidem defined?

Ah, I see. The rewrite defines it.

What if one of the upstream servers is down? Since proxy_next_upstream is off, does that mean 50% of PUT/POST/DELETE calls will fail? Is there a way to still allow proxy_next_upstream during upstream connection failure (eg: connection timeout or connection refused)? Since it is not even connected and request is not sent to failed upstream, it should be safe to try next upstream.