ottyl/tickets/098-histogram-quantile-support.md

098 — No histogram_quantile()-style percentile query

Summary

The PromQL layer's histogram() op buckets raw sample values by a fixed width and returns per-bucket counts — it's unrelated to Prometheus's classic histogram bucket/quantile model, and there's no way to ask for p50/p95/p99 out of a metric today.

Where

• src/engine/promql.rs:9-16 — AggOp::Histogram is one of the six supported ops.
• src/engine/snapshot/mod.rs:1002 onward — Histogram evaluation: buckets values by bucket_width into counts, grouped by by_labels. This answers "how many samples fell in each bucket", not "what value is the Nth percentile".
• No histogram_quantile() function or equivalent exists anywhere in promql.rs's parser (parse_query, :261-324 recognizes only sum, avg, count, rate, histogram, gauge).
• Depends on ticket 097 (gauge/counter semantics) for the broader question of whether a metric needs a declared "this holds raw observations for quantile purposes" flag, but quantile computation itself doesn't strictly require that — it can operate on whatever raw values are already stored in a metric's buckets.

Scope

• Decide the quantile computation model: interpolate over the raw stored sample values in a bucket/window (simplest, exact for the data actually retained) vs. building a genuine bucketed-histogram representation with configurable bucket boundaries (closer to Prometheus's classic histograms, more work, needed only if retained raw values are later dropped in favor of pre-aggregated buckets).
• Add a histogram_quantile(p, metric[range])-shaped PromQL construct (or extend histogram() with a quantile argument) to promql.rs's parser and AggOp, plus an evaluator in snapshot/mod.rs alongside the existing Histogram handling.
• Add dashboard/panel UI support for selecting a quantile once the query-language side exists (webapp/src/components/Dashboards.tsx or wherever panel query construction lives).