Reading latency without guessing the cause.
Latency matters to users, but a smooth chart can conceal slow requests and measurement gaps.
Editorial team ·
Choose the measured boundary
A browser navigation includes several stages, while a server request timer may measure only application handling. Record where the timer starts and stops before comparing values. Network delays, queueing and retries can dominate what a visitor experiences without appearing in the application metric. A claim about end-to-end latency needs end-to-end evidence. Otherwise, label the number as the narrower component it actually measures.
Use distributions carefully
An average can hide a small but painful group of slow requests. Percentiles offer another view, but they are not interchangeable across services or time windows. The 95th percentile in two groups cannot be averaged to obtain the combined 95th percentile. Preserve the underlying samples or appropriate distribution summaries when aggregating. Consider request volume alongside each percentile, because a tiny window can produce unstable values that look dramatic on a dashboard.
Correlate before assigning blame
When response times increase, look for changes in request rate, saturation and errors in the same interval. A rising queue can explain why even healthy downstream components appear slow. Compare the affected route with a stable control route if one exists, and mark any monitoring gaps. Correlation narrows the investigation; it does not establish causality. Avoid treating the most recent deployment as the culprit until the evidence connects it to the observed behaviour.
Turn a chart into a question
A useful latency panel should answer what became slower, for whom and since when. Annotate changes only when their times are known. If client-side experience differs from server metrics, inspect the network path and resource loading rather than changing an unrelated application threshold. Decide whether a separate alert is actionable before adding it. The goal is not the lowest-looking line; it is a measurement that helps someone diagnose a real delay.
Further reading
Continue with the field notes index, or explore the overview for the wider context. Compare the advice with a clearly defined measurement on a system you are allowed to inspect. Record the observation window and the affected operation, then describe what the evidence can and cannot establish. A hypothetical example is never a substitute for service data.