Measurement correctness lab

Latency Percentile Aggregation Explorer

Compare the true pooled nearest-rank p95 with worker-level p95 averages and a merged classic-histogram interpolation estimate. The methods are labelled separately because they answer different mathematical questions.

Browser-only calculation. Nothing is uploaded or stored.

Synthetic observations

Choose a fixture or edit the JSON. Limits: 50 workers, 5,000 observations, 100 boundaries, and 120 KB per JSON field.

Worker groups JSON

Required shape: an array of objects containing a unique name and a non-empty values array.

Shared duration unit

Bucket boundaries JSON

Population assumptions

All groups come from the same observation window.

Groups contain disjoint requests; no observation is duplicated across workers.

Comparison

The pooled result uses raw observations only because this lab already has them. It never attempts to reconstruct observations from percentile summaries.

Load a valid fixture or enter your own bounded synthetic data, then calculate.

Pooled nearest-rank p95

—

Unweighted worker p95 mean

—

Request-weighted worker p95 mean

—

Histogram linear estimate

—

Worker-level nearest-rank p95

Worker	Requests	Rank	p95

Merged cumulative buckets

Upper bound	In bucket	Cumulative	Share	Visual

Warnings and interpretation

Method boundary

Classic cumulative buckets can be summed across disjoint populations that share boundaries and an observation window. That makes the counts aggregatable, but it does not make within-bucket linear interpolation exact. Worker percentile summaries alone are not sufficient to recover the pooled percentile, and this explorer does not present classic histograms as the recommended default for new deployments.