Using the last rate counters, we now compute the queue, connect, response and total times per server and per backend with a 95% accuracy over the last 1024 samples. The operation is cheap so we don't need to condition it.