A service’s average response time looks fine, but a few times an hour it spikes badly. How do you figure out what is going on?