Browser-only measurement lab

Latency Percentile Aggregation Explorer

See why averaging worker p95 values is not the same as calculating p95 over the pooled request population. Merge cumulative buckets and inspect the classic-histogram interpolation estimate without reconstructing observations from percentile summaries.

Your data stays in this tab. No uploads, credentials, telemetry, or API calls.

Inputs

Fixtures are explicitly synthetic and editable. Supply raw request durations grouped by worker.

Common observation window

Shared duration unit

I confirm that each worker array represents a disjoint request population from the common observation window. No request appears in more than one group.

Worker durations JSON 0 / 250,000 characters

Schema: an object of worker names mapped to nonempty arrays of finite, nonnegative numbers. Maximum 50 workers and 10,000 total observations.

Positive bucket upper boundaries Maximum 200

Comma or whitespace separated. Boundaries must be finite, positive, unique, and strictly increasing. Values equal to a boundary are included in that bucket.

Results

Nearest-rank calculations use rank ceil(0.95 × N). Histogram interpolation is shown separately.

Choose a synthetic fixture or edit the inputs, then analyze.

Validated results will appear here.

Pooled nearest-rank p95 —

Unweighted worker p95 average —

Request-weighted worker p95 average —

Warnings and boundary notes

Worker p95 values

Worker	Requests	Nearest-rank p95	Weight

Merged cumulative buckets

Upper bound	Interval count	Cumulative	Share	Distribution

Classic-histogram linear-interpolation estimate

Estimated p95, not exact —

Interpretation limits

The pooled nearest-rank p95 is calculated from the supplied raw observations. Worker percentile summaries alone cannot be merged into that value.
Request weighting changes the average of worker p95 values; it does not turn that average into the pooled p95.
Cumulative bucket counts can be added for disjoint populations sharing boundaries and an observation window, but interpolation within a bucket remains an estimate.
If the p95 rank falls into the overflow bucket above the highest finite boundary, this tool does not invent an upper bound or reconstruct raw values.
This bounded teaching tool does not ingest telemetry, handle counter resets, model native exponential histograms, or claim production-system compatibility.