Offer maintained-population top-k candidates in Pass 1 #681

Activity

1. zzylol

   added 3 commits that reference this issue on Oct 16, 2026

   refactor(asap-aware-mapping)!: delete the legacy replacement search a…

   ...

   544bc09

   refactor(asap-aware-mapping)!: delete the legacy replacement search a…

   ...

   8a6aa83

   refactor(asap-aware-mapping)!: delete the legacy replacement search a…

   ...

   775e21e
2. zzylol commented on Oct 27, 2026

   

   zzylol

   on Oct 27, 2026

   ContributorAuthor

   More actions

   #961 deletes the unwired MaintainedPopulationStrategy rule and its rule tests. The cases they covered are acceptance tests for reimplementing the rule as a Pass 1 candidate here. The deleted files at ebfad9cd:

   • crates/logical-asap-optimization/src/maintained_population.rs (rule and unit tests)
   • crates/frontend-sql/tests/maintained_population.rs
   • crates/frontend-promql/tests/maintained_population_horizon.rs

   Cases to cover again:

   1. Sharing across readouts. quantile by(job)(0.5,a), quantile by(job)(0.99,a), topk by(job)(1,a) and topk by(job)(5,a) share one population producer, with max_k = 5 and quantiles = true. sum, count and avg, with and without by(job), are current-series candidates that compile.
   2. SQL table rows. Quantiles and scalar readouts over one table share rows without a PromQL lookback. max_k is shared. Populations separate by grouping and filters, projections are preserved, and an invalid value column is rejected.
   3. Recognition limits. These are not instant populations: quantile_over_time(0.5,a[1m]), quantile(0.5,sum by(job)(a)), topk(5,a offset 1m), topk(5,a @ 100), bottomk(5,a). Different sources, groupings or matchers separate populations: topk(5,a) vs topk(10,b), and quantile(0.9,a{job="api"}). without(instance) with a regex matcher keeps without, the grouping and the regex.
   4. Limit/Sort partition check. A global Limit over a per-job Sort, or a per-job Limit over a global Sort, is not the population's top-k.
   5. Selector horizon. The population of sum(a) over a 1-second selector has lookback_ms = 1000 and matches its source; with 300000 it does not.

   #961 keeps the IR coverage (invalid value column, lookback agreement, timing follows materialization, malformed DAGs fail closed) as hand-built tests in crates/types/tests/schema_rebuilding.rs.
3. zzylol

   added a commit that references this issue on Oct 27, 2026

   refactor(optimizer): delete the unwired maintained-population rule (#961

   ...

   d4152bc