The dashboard is green and the customer is angry
A latency histogram, and five things the team believes about it. The counts are cumulative: 9 850 requests came back within 200 ms, 9 902 within 300 ms, and there are 10 000 in total.
Three of the beliefs are arithmetic that does not hold. Two are the right way to work with a histogram. Nothing here is a tooling problem — every one of these is a statement about what a bucket count can and cannot tell you.
Boundaries, outermost first: Latency histogram (cumulative counts): le = 300 ms (9 902), le = 1 s (9 970), le = 9.5 s (9 999), le = +Inf (10 000), le = 200 ms (9 850) What the team believes: p99 is 9.5 s (steady for weeks), Sum buckets first (then the quantile), The mean improved (so latency did), Alert on burn rate (not on the mean), Average the p99s (one per instance) This diagram draws no connections between its components.