PromQL: histogram_quantile classic-bucket vs sketch-able quantile is a structural heuristic — drive it from sample type/metadata

[zzylol]

Context

histogram_quantile(φ, <arg>) has two lowerings (added alongside #43):

• classic cumulative-bucket form → AggIntent::HistogramQuantile { q } — exact interpolation over le buckets; not sketch-able.
• native-histogram / raw-samples form → generic AggIntent::Quantile { q, accuracy } — sketch-able (L4 picks KLL/DDSketch to hit the accuracy target).

The distinction matters operationally: pre-aggregated bucket counts can't be re-sketched (you can't reconstruct the distribution from cumulative counts), while raw samples / native histograms can be, with an accuracy target.

The gap: the discriminator is a structural heuristic

The true signal — the argument's sample type (classic histogram vs native histogram vs raw float samples) — is not visible at lowering time. is_classic_bucket_arg (crates/frontend-promql/src/promql.rs) currently proxies it by recognising, in the argument tree, any of:

• a by (le) grouping (sum by (le) (…)),
• a selector on a _bucket metric (bare name or __name__),
• an le label matcher ({le="…"}).

Everything else → generic (sketch-able) Quantile.

This is right for the common cases but has failure modes:

• False positive — a metric merely named …_bucket that isn't a classic histogram routes to HistogramQuantile (interpolation) instead of the sketch path. (Low impact: histogram_quantile over a non-histogram is meaningless in stock PromQL anyway.)
• False negative — a classic histogram exposed without the _bucket suffix and queried without le grouping/matcher routes to the sketch path — which would be wrong (can't sketch buckets).
• Native histogram named …_bucket (unusual) would misroute to interpolation.
• Relies on the raw-samples extension (histogram_quantile(φ, <raw metric>) → sketch) being distinguishable only by the absence of bucket signals.

Proposed direction

Drive the choice from type/metadata, not query structure, once available:

• a schema/catalog that tags a series (or metric) as classic-histogram / native-histogram / raw, consulted during lowering (or deferred to an L4 rule that has the sample-type), or
• an explicit annotation on the workload / query (e.g. the client declares it holds raw samples it can sketch).

Until then the structural heuristic stands; this issue tracks replacing it with a type-driven decision and pinning the edge cases with tests.

References

• Implementation: is_classic_bucket_arg / walk_histogram in crates/frontend-promql/src/promql.rs
• Tests: histogram_quantile_classic_bucket_vs_native (promql_conformance.rs)
• Related: PromQL: native-histogram accessor functions (histogram_*) unsupported #43 (native-histogram accessors), L3 IR standardization: cross-language correctness and canonicalization #34 (L3 IR standardization)