Either expose average (2 counters: total latency and total number of requests) or preferably expose it as a histogram. We will need to add support for histogram metric in streams code.