Proxy Latency Test P95 and P99 Interpretation
A proxy latency test can have an acceptable average while a meaningful share of tasks finish much later. P95 and P99 describe positions in a measured latency distribution, but they are only useful when the sample and timing definition are clear. Use Go2Proxy proxy checking tools with application-level measurements that preserve those conditions.
Do not treat a percentile label as an automatic quality guarantee. A small or mixed sample can produce a precise-looking number that answers no stable operational question.

Define the measured population
State whether the report covers request attempts, completed requests or complete business tasks. A multi-step task can have a different latency distribution from its individual requests.
Record the start and end events used for timing. Connection setup, first byte and validated task completion are different measures. Do not combine them under one latency column.
Group observations by conditions that materially affect the path, such as worker location, destination, response type and connection state. Combining unrelated groups can conceal a slow segment behind a large fast segment.
Explain the percentile meaning
Under the chosen calculation method, P95 marks a value near the point at which roughly 95 percent of the measured durations are at or below it. It does not mean that every future request has a guaranteed 95 percent chance of meeting that value.
Document the calculation method or monitoring system used. Different percentile estimators can differ, especially with small samples. Keep raw counts alongside the displayed statistic.
Use enough observations for the question being asked. A high percentile calculated from a handful of requests is unstable and should not be presented as a dependable description of rare slow behavior.
Keep failures visible
If percentiles cover only successful responses, say so and report the excluded failure counts. A route with many timeouts can look fast when every timeout disappears from the latency calculation.
Separate content-valid responses from merely completed transfers. A fast error page should not improve the performance score of a task that needs a particular dataset.
Report first-attempt and eventual task outcomes separately when retries are involved. A quick successful retry can conceal a longer user-visible task duration if the first attempt's time is omitted.
Compare equivalent observation windows
Use comparable time windows and request conditions when evaluating a change. A quiet-hour sample and a peak-hour sample may differ for reasons unrelated to the configuration being tested.
Keep the sample count, concurrency and destination mix visible. A new release that changes which pages are requested can alter the percentile without changing the underlying route quality.
Avoid selecting only the most favorable window. Preserve the planned measurement period and note interruptions or missing observations that limit the comparison.
A misleading improvement example
Suppose a route's reported P99 improves after an application update. The new report excludes timed-out tasks and measures only the final successful request. The business task still waits through failed attempts before that request begins.
Restore a task-level measurement and show the successful-request distribution as a separate view. Keep failure counts and retry time visible. The earlier improvement was a change in measurement scope, not established evidence of a faster experience.
Use the corrected report to decide whether slow tasks need a client, scheduler, destination or route investigation. A percentile alone does not identify the responsible component.
Percentile questions
Can P99 be trusted from a tiny sample?
It can be calculated, but its operational interpretation is weak. Show the count and avoid claiming that a few requests characterize rare slow events reliably.
Should timeouts be removed from the report?
They may sit outside a successful-duration distribution, but their counts and task impact must remain visible. Do not exclude them silently.
Does a better percentile prove the proxy caused the improvement?
No. Check whether workload, timing definition and environment stayed comparable. Attribution requires more than a changed summary statistic.
Publish the metric with its population
Pair each tail-latency value with a definition, count and failure context. This makes proxy performance reports understandable enough to guide an actual engineering decision.



