feat(planner): emit streaming and inference configs from the MILP plan

[milindsrivastava1997]

Implements A3 (plan → StreamingConfig / InferenceConfig) of #790 (session 3a). It builds on #801, which has the S8 plan output and the capability split.

What

• New optimizer/milp_output.rs: plan_to_planner_output(config, workload, solution) -> PlannerOutput produces streaming and inference YAML in asap-planner's format. rollup_labels is left empty (the legacy planner fills it; the engine doesn't read it). It reuses the PromQL generator's build_streaming_yaml / build_inference_yaml.
• asap-optimizer-cli --milp:
  • --output-dir <dir> writes streaming_config.yaml and inference_config.yaml.
  • --allow-undeployable-families passes true to build_all_candidates. It prints the plan, writes no configs, and can't be combined with --output-dir.
• solve_milp takes allow_undeployable_families.

Mapping

sketch-bench variant (capability)	Aggregation	sub_type	parameters
exact-sum (Sum / Count)	MultipleSum	sum / count	—
exact-min / exact-max	MultipleMinMax	min / max	—
exact-increase	MultipleIncrease	—	—
kll-percall	DatasketchesKLL	—	K = k
hll	HLL	—	precision = lg_k
cms-heap-topk-fastpath-vector2d (TopKByValue / TopKByCount)	CountMinSketchWithHeap	sum / count	depth = rows, width = cols, heapsize = largest k among served queries

• Grouping and aggregated labels come from set_subpopulation_labels, the same split the legacy planner uses.
• A deployment is tumbling if window == slide, otherwise sliding.
• Aggregation ids start at 1 and follow the plan's deployment order.

Design decisions

1. Unmapped variants are errors, including hydra-kll (not deployed: its DeltaSet key tracker has no ASAPQuery pairing). No DeltaSet pairing is built.
2. avg queries are errors (A3b) when writing configs. --output-dir rejects them before the solve; print-only runs still plan them. The avg rewrite plans sum and count, but the user's avg query would get no query_config, and the engine has no Avg (Support Avg (and other multi-statistic aggregations) in query engine #463).
3. A query string served by two deployments is an error. This can happen when the same leaf appears at two cadences or SLAs. One query_config can name only one aggregation (related: Query engine cannot distinguish occurrences of the same query string with different requirements #778).
4. heapsize = the largest k, without the legacy × 4 multiplier. sketch-bench doesn't model heap size: cost rows use a fixed top-k = 32 wrapper.
5. Cleanup: NoCleanup, the same as the greedy translator. Under NoCleanup the YAML carries no retention, so retained_instance_count and merged_instance_count aren't emitted yet and retention is unbounded. Moving to ReadBased is tracked in planner: MILP plan → InferenceConfig should use ReadBased cleanup instead of NoCleanup #800.
6. DDSketch is not mapped yet. sketch-bench doesn't list dd as deployable; the follow-up, including value-range facts and the accuracy metric, is planner: deploy DDSketch from MILP plans #802.

Known risk

Count is served by MultipleSum with sub-type count. That works because each query_config names its aggregation id. The engine's fallback matching (compatible_agg_types(Count)) doesn't list MultipleSum, so a Count query without a query_config wouldn't find it. No end-to-end engine run yet.

Interface for 4a (A4+A5)

• plan_to_planner_output is the plan → configs entry point. Its PlannerOutput has to_streaming_yaml_string / to_inference_yaml_string.
• promql::generator::{build_streaming_yaml, build_inference_yaml} are now pub(crate).

Test plan

• 8 new tests in milp_output.rs:
  • Full round trip: generated YAML → InferenceConfig / StreamingConfig as the engine parses it, checking type, sub_type, params, labels and no retention for sum, count, quantile, and value- and count-ranked topk.
  • Shared heap sized for the largest k.
  • Sliding window type.
  • Errors: avg query, hydra-kll, query split across deployments.
  • avg detection inside binary queries.
  • Parsing topk's k.
• cargo test -p asap_planner, clippy with -D warnings, fmt; pre-commit ran workspace check, clippy and test.
• CLI smoke run on sketch-bench/out_10_6_26_1342/rqe_atomic_costs.json (scrape 15s; sum, count, rate, quantile, both topk kinds, max[1h]) wrote 7 aggregations as in the table. --allow-undeployable-families prints the plan and is rejected with --output-dir.

🤖 Generated with Claude Code