Repository navigation
eval: RQEs of one e2e query map to different deployments #214
Description
Problem
Templates 14, 15 and 17 list some of the same groupings, so they produce identical queries with different accuracy scopes. For example, COUNT(DISTINCT user_id) … GROUP BY service, endpoint comes from t14 (groups ≥1%), t15 (≥5%) and t17 (every group).
#209 merges them into one e2e entry that reads all of their RQEs. The MILP still assigns each RQE separately, so one query can be mapped to two sketch deployments. However, ASAPQuery maps a query string to one aggregation (ProjectASAP/ASAPQuery#778), so --planner saved refuses these plans (ProjectASAP/ASAPQuery#856).
Affected: 8 multigroup queries (t14+t15 on region, service and region, service; t14+t15+t17 on service, endpoint; 5m and 15m each).
In #212's 10/s plans, 13 of the 16 asap SQL conversions of multigroup and all are refused. Example (multigroup, cpu, cheapest, 5m service, endpoint): t14 and t15 are on deployment 98 (HLL by region, service, endpoint), and t17 is on deployment 114 (HLL by service, endpoint, status). The 200/s plans happen not to split.
Options
- The MILP requires the RQEs of one e2e query to share a deployment.
- Give the duplicates distinct queries, or keep only one of them in the e2e workload.
Activity
[-]eval: RQEs of one e2e query land on different deployments[/-][+]eval: RQEs of one e2e query map to different deployments[/+]milindsrivastava1997 commentedon Oct 14, 2026
Decisions
e2eblock would leave plan deployments the run never deploys, so measured cost wouldn't match the plan's prediction (docs(eval): plan the end-to-end evaluation (paper §6.2) ASAPQuery#862).covers_sharenone (every group), then 0.01, then 0.05. That's t17 forservice, endpoint, t14 forregion,serviceandregion, service, and t15 elsewhere. The export errors if the strictest copy is left out at the table's rate; it doesn't fall back to a weaker copy.--keep-duplicate-queriesto turn it off. The §6.3 commands inautosketch-vs-asap-hydra/README.mdpass the flag, so those tables still reproduce byte for byte.main. Once it and eval: instant RQEs read one scrape, not 1 s #215 have merged, the e2e 10/s tables are re-exported and re-planned once.asapplan of the regenerated tables converts with ASAPQuery's--planner saved: SQL formultigroupandall(3 of 16 today), PromQL forclassicandall. ASAPQuery#856 keeps refusing split queries.eval: one RQE per schema query, the strictest template's (#214)
eval: one RQE per schema query, the strictest template's (#216)