feat(planner): asap-planner --planner saved writes a sketch-bench eval plan without re-solving

[milindsrivastava1997]

Before this PR

The only way to get streaming and inference configs from an rqe-optimizer plan was --planner milp, which solves the plan itself. The #777 eval's plans were solved in sketch-bench (autosketch_vs_asap, saved in sketch-bench#203), and re-solving gives a different plan: a different solver objective, accuracy source and arrival rates. experiment_run_e2e.py also always runs the planner.

After this PR

B1 + B2: asap-planner --planner saved

asap-planner --planner saved --streaming_engine precompute --data-ingestion-interval-ms 1000 \
  --plan <sketch-bench>/rqe-optimizer/results/autosketch-vs-asap-hydra/plans/synthetic-templatesclassic-shared1.plans.json.gz \
  --workload-table <sketch-bench>/rqe-optimizer/results/autosketch-vs-asap-hydra/inputs/tables/synthetic-templatesclassic-shared1.json \
  --method asap --weights cpu [--sla-ms 1000] --output_dir DIR

• The table's e2e block (sketch-bench#208) becomes a ControllerConfig: one query group per interval, with metrics hints from its labels. workload_from_config (the facts-free half of build_milp_workload) turns it into Raqes.
• Each leaf query maps to its plan RQE through e2e.queries[].rqes, and the RQE to its chosen deployment, sketch, params, window and slide. The result goes through deployments_to_planner_output, the existing converter with MilpSolution replaced by (deployments, Raqe → deployment). plan_to_planner_output is now a thin wrapper around it, so --planner milp output is unchanged.
• Without --sla-ms it picks the cheapest plan, and with it the sla plan at that SLA. A .gz plan file is read directly (flate2).
• It refuses:
  • roll-ups (deployment_grouping)
  • an RQE the plan doesn't serve
  • a plan deployment whose RQEs would need different sketches
  • one query served by two deployments (the existing check, which also covers the same PromQL at two intervals on different deployments)
  • any shape the engine can't serve (the existing pattern check)

B3: controller.prebuilt_output_dir in experiment_run_e2e.py skips controller_service.start and syncs that directory's streaming_config.yaml and inference_config.yaml to the controller output dir. It checks that both files exist at startup.

Departures from the #824 plan:

• B1 passes deployments instead of building a MilpSolution, because PlanCost can't be built honestly from a saved plan.
• B2 takes --sla-ms instead of --kind, because the file has one SLA plan per SLA.

Blocked: 12 classic queries (#827)

On the real classic base table (ASAP cheapest at CPU weights), 12 of the 54 queries are shapes ASAPQuery can't serve: sum by (job) (rate(...)) and topk(32, sum by (label_0) (...)) over rate and sum_over_time. So --planner saved refuses the table:

Error: cannot optimize leaf 'topk(32, sum by (label_0) (sum_over_time(data[15m])))' ...: no supported query pattern matched

With those 12 removed from a copy of the table, it writes 5 aggregations. Their windows, slides and params match the plan: dd 0.02 1s; dd 0.05 300s/60s sliding; exact-increase 60s; exact-sum 1s and 60s. The other 3 plan deployments serve the removed queries. #827 tracks engine support, and the e2e run waits on it.

Validation

• cargo test -p asap_planner: all pass, including 5 new tests in saved_plan.rs. They cover:
  • a ratio whose legs share one sliding DDSketch
  • --sla-ms selection, plus no matching plan
  • roll-up refusal
  • one deployment needing two sketches (this test caught leaves that merge into one item, such as p90 and p50, not being checked)
  • an RQE missing from the plan
• cargo clippy --all-targets -D warnings, cargo fmt, black, flake8: clean.
• The CLI ran on sketch-bench's committed plan file: the refusal above, the 42-query conversion, and --sla-ms 1000.
• B3: the copy-and-sync path was run on its own with a local provider. A full e2e run waits on feat(planner): serve the #777 eval's sum-of-rate and nested topk queries #827 and B4.

Part of #824 (B1, B2, B3).

🤖 Generated with Claude Code