When the balancer gets a request from a client, it will try to send the request to a back-end. If the back-end fails, it will re-send the request to another back-end (if possible). There is an upper limit on the number of times that a request will be retried, controllable with the tunable tuning!max_connect_tries. If after all these attempts, no connection has been successful, then the balancer will 'give in' and send back the error page that you're seeing.