Notes on measuring things that move
A small, slow-moving log about latency, sampling and the parts of a measurement that are easy to get wrong.
Most of what is written here started as a note to self. The recurring theme is that a number is only as good as the question it answers — and that the question is usually less obvious than the number.
Recent
Percentiles lie about tails
Why a p95 that looks flat can hide a queue that is quietly getting longer.
Sampling a signal you cannot pause
Notes on picking an interval when the thing you measure reacts to being measured.
Two points beat one
A short argument for deriving a rate from a slope instead of a single reading.