[milindsrivastava1997]

Finishes the remaining parts of S1 and S6 from #143. #145 already did metric/grouping_labels (S1) and the scrape-interval multiples (S6).

Interface changes later sessions depend on

Rqe → Raqe (rqe-optimizer/src/lib.rs)

• accuracy_tolerance → accuracy_sla.
• New spatial_filter: String: canonical PromQL matcher text, "" for none. It is part of stream identity, compared as an opaque string. validate_facts rejects a non-empty filter for now, because facts have no per-filter cardinality and a filtered stream would be costed as the whole metric.
• New latency_sla_ms: Option<f64>: a hard ceiling on modeled query latency. Only the MILP enforces it. validate_facts rejects values that are non-finite or ≤ 0.
• Stream identity for grouping, eligibility, sharing and pruning is now (capability, metric, spatial_filter, grouping_labels). Deployment carries spatial_filter too.

Milliseconds everywhere. Seconds → Millis (u64), and *_secs → *_ms for lookback, interval, window, slide and MetricFacts::scrape_interval. Rates (CPU-sec/sec, samples/sec) stay per second. Latency is in ms too: PlanCost::query_latency_secs → query_latency_ms. validate_facts rejects a zero lookback or interval, and any lookback or interval that isn't a multiple of the metric's scrape interval. That catches values left in seconds.

MilpBounds removed. milp::minimize(raqes, deployments, facts, objective) reads each Raqe's latency_sla_ms. In the example, --latency-limit RAQE_ID=SECONDS becomes --latency-sla RAQE_ID=MS.

Allowed families (S6). These are hardcoded, not taken as input: DEPLOYABLE_FAMILIES lists the variants ASAPQuery's engine runs (middle column). The rest of Capability::families() is measured but undeployable:

Capability	Deployable (ASAPQuery AggregationType)	Undeployable (planned only with --allow-undeployable-families)
SumOrCount	exact-sum (MultipleSum)	—
Min	exact-min (MultipleMinMax, sub_type "min")	—
Max	exact-max (MultipleMinMax, sub_type "max")	—
RateOrIncrease	exact-increase (MultipleIncrease)	—
Quantile	kll-percall (DatasketchesKLL)	dd
Cardinality	hll (HLL)	univmon-cardinality
TopK	cms-heap-topk-fastpath-vector2d (CountMinSketchWithHeap)	countsketch-heap-topk-fastpath-vector2d, univmon-topk

ASAPQuery engine types not in this table yet: HydraKLL (Quantile) joins families() and DEPLOYABLE_FAMILIES once S5/S4b measure it. SetAggregator and DeltaSetAggregator (exact Cardinality) have no sketch-bench family; DeltaSet is planned only as HydraKLL's key tracker (S5/S4b).

By default, candidate generation and the AutoSketch baseline only use deployable families, so dd, univmon-* and countsketch-heap are measured but not planned.

--allow-undeployable-families. build_all_candidates, build_all_candidates_unpruned, autosketch::plan and autosketch::search take allow_undeployable_families: bool. All of them choose families through one rule, Capability::candidate_families(allow_undeployable_families). When the flag is true, every measured family in Capability::families() is a candidate. Use it for studies only: the resulting plan can't be deployed. The small_problem example exposes it as --allow-undeployable-families. AutoSketch takes the same flag, so the baseline and the MILP always plan with the same sketches. In ASAPQuery, A3/A5 can pass the flag through to print the plan and its cost without writing YAML. hydra-kll joins families() and this list once S5/S4b measure it. ASAPQuery's sketch_bench_key doesn't map CountMinSketchWithHeap yet, so A3 needs to add that mapping.

Notes

• AutoSketch's cross-variant test runs search(..., allow_undeployable_families: true) over both TopK families. No capability has two deployable variants until hydra-kll.
• The crate, file and script names keep rqe (rqe-optimizer, export_rqe_optimizer_costs.sh). Only the types and identifiers were renamed.

Test plan

• cargo test -p rqe-optimizer (39 pass): new tests for spatial-filter separation, undeployable families being candidates only with the flag, each deployable family being in some families() (iterating Capability::ALL), and validate_facts rejecting bad SLAs, spatial filters, zero times, and times that aren't multiples of the scrape interval.
• Workspace fmt, clippy and test via pre-commit.
• small_problem --milp --latency-sla … builds candidates. A full solve needs a fresh S5 export: the local cost table predates rqe-optimizer: exact sum/min-max/increase accumulators, CountSketch+heap and UnivMon top-k #144 and has no exact-* rows.

🤖 Generated with Claude Code

[milindsrivastava1997]

Code review findings (10). Nine are inline below. One is on a file this PR doesn't touch:

scripts/export_rqe_optimizer_costs.sh: the export script still benchmarks dd, univmon-cardinality, univmon-topk and countsketch-heap-topk, which can no longer become candidates (deployable_families() discards their rows). Its header still says it measures what Capability::families() names. That makes every cost-table regeneration slower than needed, and the header is misleading.