Yeah, I was running it under gunicorn. But that still leaves you with a small number of concurrent connections. If you want to have long connections, for something like long polling or websockets, then being limited to one request per CPU core seems a little sketchy.
Mind you, it could just be that I take high concurrency for granted. I build most web stuff on Node or Clojure, but now that I think about it, apps that require long quiet connections are actually not the norm.