PromQL: histogram_quantiles() and non-literal scalar k/φ parameters unsupported

[zzylol]

Summary

Two PromQL constructs parse cleanly but are rejected by the lowerer. Both are genuine coverage gaps rather than parser issues, so they are fixable entirely within crates/frontend-promql.

Corpus impact

30 rejections on origin/main @ 9fdd3d7 (of 322 total), plus ~21 more in the "arguably correct" bucket described under Notes.

Corrected 2026-07-09: figures re-measured on current main. The two counts below are unchanged; only the overall denominator moved (387 → 322) after the #40 offset/@ work landed.

Construct	Rejections	Error variant
histogram_quantiles(v, "label", φ…)	19	UnsupportedFunction
Non-literal scalar k / φ parameter	11	InvalidParameter

Examples

# histogram_quantiles — upstream experimental, variadic (up to 9 quantiles),
# returns one series per φ, tagged with the given label
histogram_quantiles(testhistogram3, "q", 0, 0.25, 0.5, 0.75, 1)

# non-literal k: the parameter is a scalar-valued *expression*, not a literal
topk(scalar(foo), http_requests)

Notes

• histogram_quantiles is a real upstream function (promql/parser/functions.go:213, Experimental: true, Variadic: 9). It is effectively histogram_quantile fanned out over N quantiles with the results labelled — so it should decompose into the existing AggIntent::Quantile path once the fan-out is expressed. Related but distinct from PromQL: native-histogram accessor functions (histogram_*) unsupported #43 (histogram accessors) and PromQL: histogram_quantile classic-bucket vs sketch-able quantile is a structural heuristic — drive it from sample type/metadata #79 (classic-bucket vs sketch-able quantile), neither of which covers it.
• The non-literal-k rejection comes from crates/frontend-promql/src/promql.rs:1560 (expected a numeric scalar, got {:?}). Today topk/bottomk/quantile require a literal scalar. topk(scalar(foo), v) supplies a scalar-valued subexpression, which is legal PromQL. Note the error message leaks a raw Discriminant(9) rather than naming the offending node — worth fixing regardless.
• Possibly working as intended, flagged for a decision, not filed as part of this issue: ~22 further rejections come from out-of-domain constants that Prometheus accepts and answers with ±Inf/NaN plus a warn annotation, whereas we hard-error:

  histogram_quantile(1.001, testhistogram)   → InvalidParameter (Prometheus: +Inf + warn)
  quantile_over_time(-1, data[2m])           → InvalidParameter (Prometheus: -Inf + warn)
  topk(NaN, non_existent)                    → InvalidParameter
  

  (17 out-of-range quantile + 3 NaN k + 1 non-finite limit_ratio ratio = 21.)
  These originate from promql.rs:1617. For an optimizer IR, rejecting may well be the right call — but the divergence from PromQL semantics should be a conscious choice and ideally a documented one.
• A further 5 rejections are bare string literals (("Foo")), rejected as UnsupportedFeature("bare string literal"). String-typed expressions have no vector semantics; almost certainly WAI.

[zzylol]

Status

• histogram_quantiles — implemented in feat(promql): lower histogram_quantiles as a per-φ Merge fan-out (#109) #121 (open for review). Lowers to a Merge of one Relabel-wrapped quantile branch per φ, reusing the histogram_quantile classic-vs-native decision. Corpus: testdata 1502 → 1512 lowered.
• Non-literal scalar k / φ — split out to PromQL: non-literal scalar parameters (topk(scalar(foo), v), quantile(scalar(f), v), limit_ratio(time()%17/17, v)) #120. It needs a scalar subtree in the intent (TopK.k: usize, Quantile.q: f64) and at the L4 sketch readout (SketchQuery::TopK { k }), plus a rethink of the φ-derived output column name. Nothing in common with the histogram_quantiles work beyond this issue title.

The out-of-range-φ question raised in the Notes here is still open and now has a concrete instance: 7 of the 9 histogram_quantiles lines that still fail do so because quantile_param rejects -0.1 / 1.01 / NaN, where Prometheus returns ±Inf/NaN with a warn annotation. Worth deciding deliberately rather than inheriting.

[zzylol]

Closing — both halves are now accounted for.

• histogram_quantiles — implemented in feat(promql): lower histogram_quantiles as a per-φ Merge fan-out (#109) #121 (cb2f354). Corpus: testdata 1502 → 1512 lowered.
• Non-literal scalar k / φ — tracked in PromQL: non-literal scalar parameters (topk(scalar(foo), v), quantile(scalar(f), v), limit_ratio(time()%17/17, v)) #120, which carries the full analysis. Nothing left in this issue that PromQL: non-literal scalar parameters (topk(scalar(foo), v), quantile(scalar(f), v), limit_ratio(time()%17/17, v)) #120 does not cover.

The out-of-range-φ question from the Notes above is not closed by this — it is a deliberate-behaviour decision (histogram_quantile(1.001, …) errors here, returns +Inf with a warn annotation in Prometheus). It is recorded in #120 under "Notes", since that is where the validation would move if scalar parameters become expressions.