Lexicon
Histogram
A histogram is a metric that captures value distributions in buckets, powering p95 and p99 latency. How bucket boundaries work and how percentiles go wrong.
Definition
A histogram is a metric type that captures how values distribute, by counting observations into buckets, rather than recording what any single value was. It is the type behind latency percentiles: p50, p95, and p99 are all read from histograms.
What it means in observability
Averages lie about experience. A service with 50 millisecond median latency and a two second p99 is fast on average and slow for every hundredth request, and at scale every hundredth request is thousands of people. Histograms exist to keep the whole shape of the distribution visible at metric prices: instead of storing every measurement, they store bucket counts, from which percentiles can be estimated cheaply over any time range.
How it works in practice
Each histogram tracks bucket boundaries, per bucket counts, a total count, and a sum, and percentiles are estimated from how observations spread across the buckets. Accuracy therefore depends on boundary placement, which is why exponential histograms, whose buckets scale automatically to the data, have become the OpenTelemetry default and remove the boundary guessing game. Histograms also aggregate soundly across series by summing bucket counts, which is exactly the property percentiles themselves lack.
Where it gets hard
The universal mistake is averaging percentiles: the average of ten services' p95s is not the p95 of the whole, and dashboards built that way report a fiction. Fixed bucket boundaries mis-sized for the data yield percentiles that are estimates in name and guesses in practice. And histograms multiply cardinality, since every bucket of every attribute combination is a stored series, which is where high resolution meets high invoice on per series pricing.
Where Tsuga fits
Tsuga supports percentile aggregations over histogram data along with a distribution graph type for seeing the whole shape, and per GB pricing means bucket resolution is a fidelity decision rather than a billing one.
Related terms
- CardinalityCardinality is the number of distinct values, or distinct value combinations, that an attribute or set of attributes can take.
- CounterA counter is a metric type for values that only accumulate: requests served, errors thrown, bytes transmitted, jobs completed.
- GaugeA gauge is a metric type that records a point in time value that can move in both directions: memory in use, queue depth, active connections, temperature, items in a cart.
- MetricA metric is a named numeric measurement tracked over time, request counts, memory usage, response latency, carrying a type, a unit, and attributes that slice it into series.