Non-empirical gaps in the Phase 2 greedy solver (asap-planner-rs/src/optimizer/), separate from plugging in real numbers:
Costs are param-invariant: ingest_cost/query_cost ignore sketch params (CMS depth/width, KLL K, HLL precision, heap size) — the whole candidate_gen parameter grid is cost-invisible today. (Optimizer: param-aware atomic costs from sketch-bench #524
No accuracy/error-bound modeling: SketchProperties tracks mergeable/subtractable/subpopulation_aware but nothing about error guarantees, so greedy can pick the smallest param combo with no accuracy constraint. (Optimizer: CMS accuracy constraint (ACC) using accuracy_sla #526 Optimizer: extend sketch-bench cardinality sweep for empirical Error beyond CMS #566
subpopulation_count hardcoded to 1.0 in both cost functions — subpop-aware vs non-aware currently behave identically. (Optimizer: auto-infer N_g from Prometheus #525 No cross-AQE sharing (facility-location) — greedy assigns every AQE independently, deferred to the Phase 3 MIP. (No cross-AQE sharing — greedy deploys one sketch per AQE even when they could share #650 labels_compatible() is exact-match only, not superset matching #649
Multi-statistic AQEs (e.g. avg = Sum+Count) unsupported — always falls back to EXACT. (planner/optimizer: decide support for multi-statistic AQEs (e.g. avg) #419 Support Avg (and other multi-statistic aggregations) in query engine #463
candidate_gen.rs hardcodes aggregation_id = 0, to be replaced by the greedy/MIP solver's assignment. (Optimizer: candidate_gen.rs hardcodes aggregation_id = 0 #564 translator.rs's output structs are empty/stub pending Phase 2/3 sketch configs. (Optimizer: translator.rs output structs are empty stubs #565
For empirical calibration (tracked separately/later):
Non-empirical gaps in the Phase 2 greedy solver (
asap-planner-rs/src/optimizer/), separate from plugging in real numbers:ingest_cost/query_costignore sketch params (CMS depth/width, KLL K, HLL precision, heap size) — the wholecandidate_genparameter grid is cost-invisible today. (Optimizer: param-aware atomic costs from sketch-bench #524)SketchPropertiestracks mergeable/subtractable/subpopulation_aware but nothing about error guarantees, so greedy can pick the smallest param combo with no accuracy constraint. (Optimizer: CMS accuracy constraint (ACC) using accuracy_sla #526, Optimizer: extend sketch-bench cardinality sweep for empirical Error beyond CMS #566)subpopulation_counthardcoded to1.0in both cost functions — subpop-aware vs non-aware currently behave identically. (Optimizer: auto-infer N_g from Prometheus #525)avg= Sum+Count) unsupported — always falls back to EXACT. (planner/optimizer: decide support for multi-statistic AQEs (e.g. avg) #419, Support Avg (and other multi-statistic aggregations) in query engine #463)candidate_gen.rshardcodesaggregation_id = 0, to be replaced by the greedy/MIP solver's assignment. (Optimizer: candidate_gen.rs hardcodes aggregation_id = 0 #564)translator.rs's output structs are empty/stub pending Phase 2/3 sketch configs. (Optimizer: translator.rs output structs are empty stubs #565)For empirical calibration (tracked separately/later):
CostWeightsfrom real cloudarrival_rate_hzper (metric, spatial filter) from Prometheus scrape interval × series count. (Optimizer: arrival_rate_hz is a flat placeholder, not derived from Prometheus #563)subpopulation_countper config from Prometheus series cardinality. (Optimizer: auto-infer N_g from Prometheus #525)