Quantile sketch
A quantile sketch is a compact summary of observed values used to estimate quantiles under an algorithm-specific approximation contract.
Retain the summary state, not only an exported p99. Precomputed quantiles generally cannot be averaged into a pooled request percentile; compatible histogram counts retain more information for aggregation. [2]
Rank error limits displacement in the ordered observations; relative-value error limits the returned value's deviation from the selected sample quantile. Neither is automatically a confidence interval for an unobserved population. Record the estimator and error contract. [1]
Engineering implication: mergeability does not choose which stream version or window contribution belongs in the requested population. Preserve source, units, time bounds, version and coverage separately. Repeated delivery or overlapping windows can count observations twice even when the summaries are numerically compatible.
Relations
- related
- Quantile, Latency histogram bin, DDSketch