Averaging Site p95 Values? Rebuild the Combined Latency Distribution
A multi-site report displays one p95 response time per site, then averages those figures into a headline p95. That operation generally does not produce the percentile of all responses. Site summaries have already discarded information about the underlying observations.

Original conceptual illustration; not a live software screen, specific product or customer installation.
Key takeaways
A percentile describes an ordered observation population.
Site percentile summaries do not retain the full combined distribution.
Keep the percentile method, counts and excluded observations visible.
1. Name the observations behind the label
First define the population: device reads, complete uploads or user-visible refreshes are different timing events. Record the start and finish boundaries, units, reporting window and treatment of failed attempts. A report cannot combine measurements that describe different operations merely because they all use milliseconds.
Discuss those requirements against the ProSight Software platform overview before selecting a deployment scope. The examples here are acceptance requirements for a reporting project, not claims that a named package exposes a particular percentile engine or latency service level.
2. Specify the percentile method
A percentile is obtained from the ordered data using a stated estimation rule. NIST documents several commonly used percentile methods and notes that they can produce different results, especially for small samples. Record the selected method rather than treating every p95 label as identical.
Keep the observation count and eligible population alongside the result. A percentile based on a small test set has different evidential strength from one based on a large representative set. More rows also do not guarantee representative coverage if a collector omitted failures or busy periods.
Decide whether the report describes successful operations only or includes another explicitly defined outcome treatment. A timeout is not an observed completed duration. Preserve timeout counts separately unless the reporting specification establishes a justified censored-data method.
3. Use a simple counterexample
For a deliberately simplified nearest-rank test, define p95 as the observation at position ceiling(0.95 × N) after sorting. Site A contributes one response of 100 ms. Site B contributes ninety-nine responses of 1 ms. Their individual p95 values are 100 ms and 1 ms.
The equal average of those site p95 values is 50.5 ms. The combined one-hundred-observation population contains ninety-nine values of 1 ms followed by 100 ms, so its nearest-rank p95 is the ninety-fifth observation: 1 ms.
Even weighting the two site p95 summaries by their counts gives 1.99 ms, still not the combined p95. This invented fixture demonstrates why summary averaging loses distribution information. It is not a benchmark, a real installation result or a universal choice of percentile algorithm.

Original explanatory infographic. Values and test cases are hypothetical; equipment selection and implementation require project-specific verification.
Diagnostic comparison
Nearest-rank fixture | p95 |
A: one × 100 ms | 100 ms |
B: 99 × 1 ms | 1 ms |
Combined: 100 observations | 1 ms; not 50.5 ms |
4. Retain suitable evidence for aggregation
When permitted and practical, retain the eligible raw durations so that the combined distribution can be calculated reproducibly. Apply access, retention and data-minimisation rules. Avoid including credentials, payload contents or unnecessary personal information in timing fixtures.
If storage uses histograms or another summary representation, confirm whether that representation can be combined and what approximation it introduces. Matching bucket boundaries, units and populations matters. Do not assume every statistical sketch is mergeable or that a displayed percentile is an exact raw-data result.
Keep per-site percentiles for diagnosing site behaviour, while labelling the overall calculation according to its actual population. A low combined p95 can coexist with poor performance at a low-volume site. Include site counts and exceptions so that a large site cannot hide a smaller site’s operational problem.
Define one latency population and timing boundary.
Record percentile method, counts and timeout treatment.
Combine eligible distribution evidence, not displayed p95 values.
Verify unequal-volume and empty-population cases.
5. Test the whole reporting path
Run the counterexample through the intended aggregation and export workflow. Confirm the chosen method, observation count and combined result. Include an empty site, excluded invalid durations and unequal observation volumes as separate test cases.
Record software versions and any approximation settings. Reconcile the headline with an independently calculated fixture after changes to report grouping or retention. A chart matching another chart is weak evidence when both use the same incorrect intermediate summary.
For ProSight Software assistance, provide the timing definition, a sanitised dataset, expected percentile method and required site breakdown. Ask which supported reporting or engineered integration can preserve the necessary evidence. Agree an acceptance result before using the metric for a service or purchasing decision.
Recommended equipment / related solutions
Start with the ProSight Software platform overview to discuss the supported software scope and the reporting evidence your team needs.
Where the requirement needs engineered integration or a tailored reporting workflow, review ProSight Custom project scope. Confirm deliverables and the exact supported deployment before selecting an implementation.
Related practical guide: NaN Is Not a Sensor Reading: Validate Finite Telemetry Before Charting
Frequently asked question
Does weighting site p95 by sample count solve it?
Not generally. Counts plus percentile summaries still omit the distribution needed to determine the combined percentile; the worked example demonstrates the difference.
Technical references
Need help with your application?
Send the equipment models, current documentation and a sanitised test example. Discuss this integration and equipment requirement with ProSense; confirm the actual functions, interfaces and acceptance scope before purchase or deployment.

.png)
Comments