Tail Latency Is Where Your Cloud Bill Lives
Lorenz and Little: How Much Does Your Tail Cost?
Marc Brooker uses the empirical Lorenz curve to answer a practical question: how much does each latency percentile contribute to mean latency? The answer is startling — requests at or above p99 often account for over half the mean. Since Little's law ties that same share to system concurrency, the slow tail drives capacity demand, lock contention, and cost. Trimming it away would be a mistake, because the tail is often what you're paying for.
We shouldn't make the mistake of trimming them off, because they're often the thing that's driving costs!