Measuring proxy speed with median and p95 latency
Benchmark proxy speed using median and p95 latency, defined timing boundaries, sample counts and the customer's actual route.

Start the stopwatch at a defined event
A buyer comparing proxy speed needs to know what the number measures. Timing from the beginning of a request to the first response byte answers a different question from timing until the final byte arrives. Choose the measure that corresponds to the application. A small stock-check request and a downloadable product catalogue can have very different bottlenecks even on the same route.
Keep the slow observations
The median describes the middle of a set of observations. A p95 value describes a point toward the slow end of the distribution. For example, if the median run finishes within a second but customers regularly wait much longer, a median-only chart can hide the problem. Keep the sample size, failures and test window next to both percentiles. A handful of requests has very little to say about the extreme tail.
A meaningful sample can also be slow for an understandable reason. Record planned deployments and maintenance windows with the results, so that an unusual spike can be evaluated on its own evidence before it affects a purchasing decision. Google SRE on service level objectives.
Turn the measurement into a buying decision
Use the same approved endpoint, payload, client region and time window across candidates. Run enough samples across your normal business hours to capture variation without overloading the service. An XtremeProxy speed evaluation should end with a target-specific report that another engineer can reproduce. It is more useful than a single fastest request taken from an especially quiet moment.
A good median hides a poor deadline result
In a fictional batch of twenty completed requests, eighteen take 200 milliseconds and two take 1,500 milliseconds. Under the nearest-rank convention the median is 200 and p95 is the nineteenth ordered observation, 1,500. With a one-second deadline, eighteen of twenty complete in time. Keep the sample size and percentile method with the report. A separate timeout must also appear in the attempt count even if it has no completed response duration.
Handle timeouts outside the completion distribution
A request that never returns has no observed full completion time. Keep it in the attempt and deadline counts, and specify any censoring convention if the analysis models the tail. Show completed latency and the fraction that completed. A route can look faster among successes while failing more often.
Attach comparable timeouts to each route in the comparison. A route that drops its slowest requests at two seconds can appear faster among completed responses than a route allowed to finish within ten seconds. Report completion by the common deadline alongside the percentile of successful responses.
Show how a bucket threshold changes the story
For a pilot with 18 responses below one second and two above, a median-only dashboard looks comfortable. An objective requiring 95 percent below one second fails because only 90 percent met the threshold. Keep the percentile estimate and threshold attainment as distinct figures. The same observations can legitimately support different conclusions under different service objectives.
The nearest-rank p95 can move sharply with small batches. In twenty observations it is the nineteenth ordered duration. In twenty-one it is the twentieth. One added slow completion may therefore change the reported tail even though the first twenty measurements were untouched. Put the number of on-time completions beside the percentile and keep the sorted values for a small pilot. A buyer with a one-second deadline can interpret “19 of 21 on time” more directly than an unexplained change in a single percentile.
A buyer comparing two latency charts also needs to know what each request did. A route returning a small generic page can appear faster than one completing the correct stock lookup. Establish a common content marker, size range and deadline for the comparison. Keep failed validation visible even when its transport time is extremely short. Only form the latency distribution for the declared population and place the validation failure count beside it, so a superficially faster wrong response cannot win the buying decision.
Related fieldwork
A useful performance summary has a defined starting event, the observed completion distribution and the number of missing or late results. Keep the actual deadline with the statistics. Regional download release checks gives a release download with a known output. How many samples make a regional proxy benchmark useful? addresses the sampling plan needed before extending a pilot conclusion to normal operations. Technical references include libcurl timing information. Compare XtremeProxy network types.
Frequently asked questions
Can we average daily p95 values to obtain the monthly p95?
Generally the resulting average represents daily summaries, not the combined request distribution. Retain raw observations or compatible histograms and calculate the desired percentile for the full population. Keep timeout counts visible alongside completed durations.


