PromQL: rate/increase and counter-derivative functions reject a sub-query argument (generalize #42)

[zzylol]

Summary

Range-vector functions accept a sub-query (expr[range:res]) wherever they accept a matrix selector — it's the other way (besides m[w]) to produce a range vector, and it is legitimate, parser-valid PromQL. Today only the *_over_time family accepts one (added in #42); rate/increase/irate and all the counter-derivative functions still reject a sub-query argument.

Current state (verified on main)

LOWERS    max_over_time(rate(m[5m])[1h:])        # *_over_time — done in #42
REJECTED  rate(sum(m)[5m:])                        # rate/increase/irate — NOT covered
REJECTED  increase(sum(m)[5m:])
REJECTED  irate(sum(m)[5m:])
REJECTED  changes(rate(m[5m])[1h:])                # counter-derivatives (#44) — NOT covered
REJECTED  delta(sum(m)[5m:])
REJECTED  deriv(sum(m)[10m:])
REJECTED  resets(rate(m[5m])[1h:])
REJECTED  idelta(sum(m)[5m:])
REJECTED  predict_linear(sum(m)[1h:], 3600)
REJECTED  double_exponential_smoothing(sum(m)[10m:], 0.5, 0.3)

All reject with unsupported feature: expected a range-vector (matrix) argument, got Discriminant(4) — i.e. our lowering refuses the sub-query; the parser accepts it (Discriminant(4) = Subquery). So these are valid PromQL queries Prometheus executes, e.g. changes(rate(m[5m])[1h:]) = "how many times the 5m rate changed over the past hour, per series".

Root cause

extract_matrix (crates/lower/src/promql.rs) only accepts a (parenthesised) MatrixSelector. #42 special-cased the *_over_time reducers to lower a sub-query argument into a PromQLSubquery and reduce it per series (over_time_reducer + the is_subquery branch in walk_call). rate/increase/irate and the counter-derivatives (changes/delta/idelta/deriv/resets/predict_linear/double_exponential_smoothing) go through lower_inner_call → extract_matrix and were never wired into that path.

Proposed fix

Generalize the #42 sub-query path to the whole range-vector-function family: when the matrix argument slot is a Subquery, lower it via walk into a PromQLSubquery and wrap it in the function's per-series intent (Rate/Increase/Changes/Delta/…), instead of calling extract_matrix. The L3 per-series marker already covers a Subquery child (added in #42), so the schema side needs no change — only the front end.

Watch-outs:

• rate/increase carry their window in the AggFunc variant (no L2 Window node); over a sub-query the "window" is the sub-query's own range, so the shape is Aggregate{Rate} → PromQLSubquery{range,res} → <inner> (no separate TimeRange — the sub-query is the range marker). Confirm is_per_series/schema derivation is happy with Rate directly over a Subquery.
• Thread scalar params (predict_linear horizon, double_exponential_smoothing factors) as usual.

Tests

• Pinned today by counter_derivative_over_a_subquery_is_rejected__GAP (crates/lower/tests/promql_conformance.rs) — flip it to a passing conformance test when implemented.
• Add positive tests for rate/increase over a sub-query alongside the existing over_time_of_subquery_* tests.

Related

• feat(lower): support *_over_time over a sub-query (#27) #42 — *_over_time over a sub-query (the pattern to generalize)
• PromQL: counter-derivative range functions (changes/delta/deriv/resets/predict_linear/idelta) unsupported #44 — counter-derivative range functions (the functions to extend)
• Support nested query functions for PromQL and SQL (testing + implementation) #27 — nested query functions (umbrella)