feat(asap-tools): run the e2e planner through the rqe-optimizer MILP (planner.use_optimizer)

[milindsrivastava1997]

Closes #830.

experiment_run_e2e now starts asap-planner --planner milp (sketch-bench's rqe-optimizer) when the new Hydra option planner.use_optimizer is on (default true).

Config (config/config.yaml):

planner:
  use_optimizer: true
  saturation_dir: null  # remote; atomic costs read from optimizer_cost/rqe_atomic_costs.json inside it
  workload_facts: null  # remote path to the workload-facts YAML
  w_cpu: null           # null = rqe-optimizer default
  w_mem: null

Changes

• MilpPlanner (experiment_utils/services/misc.py): ControllerService.start(milp=...) adds --planner milp --workload-facts --atomic-costs --saturation-dir [--w-cpu] [--w-mem] on bare metal. In the containerized path it passes the inputs to generate_controller_compose.py.
• asap-planner-rs/docker-compose.yml.j2: mounts the saturation dir and the facts file read-only (/app/saturation, /app/input/workload_facts.yaml) and points the MILP flags at them.
• config.milp_planner(cfg) fails fast on settings the MILP can't honor:
  • saturation_dir or workload_facts unset
  • controller.punting=true
  • controller.auto_discover_labels=true
  • windowing set
• experiment_run_e2e leaves aggregate_cleanup and sketch_parameters out of the planner input when the optimizer is on: the MILP rejects both and chooses them itself.
• experiment_run_clickhouse fails fast when use_optimizer=true, because MILP doesn't support SQL. The grafana demo and ingest-path scripts are unchanged.

Notes

• Default true means e2e runs must now set planner.saturation_dir and planner.workload_facts, or pass planner.use_optimizer=false.
• The sketchdb-controller image must be rebuilt so it includes --planner milp.
• MILP output still uses NoCleanup until planner: MILP plan → InferenceConfig should use ReadBased cleanup instead of NoCleanup #800 lands.

Testing

• pytest asap-tools/experiments/tests: 61 passed. New tests are in tests/test_planner_config.py and tests/test_controller_service.py.
• Not yet run end to end on a cluster.

🤖 Generated with Claude Code

[milindsrivastava1997]

Core wiring looks right: the CLI flags match asap-planner-rs/src/main.rs, the compose mounts are correct, and generate_controller_client_configs already handles None for cleanup and sketch params. Inline comments cover the rest. Two points that touch lines outside the diff:

• The Grafana demo and ingest-path scripts silently run the legacy planner when use_optimizer=true. experiment_run_grafana_demo.py and experiment_only_ingest_path.py read the same config and call controller_service.start(...) without milp. ClickHouse got a guard, so this is inconsistent, and it breaks fail-loud: a user who sets use_optimizer=true for a Grafana demo gets a legacy plan with no warning. Either add require_legacy_planner there too (this only works if the default becomes false), or wire milp through.
• Tests: the negative paths are covered well. Nothing covers the default-config interaction, though: a test that loads the real config.yaml defaults and runs the ClickHouse guard would have caught the regression flagged on experiment_run_clickhouse.py.
• Nit (predates this PR): Args.cleanup_policy (config.py:612) is set but never read anywhere, and it's now misleading under MILP. Drop it if you're nearby.

[milindsrivastava1997]

experiment_run_clickhouse now fails by default. config.yaml defaults planner.use_optimizer: true, and this guard runs first, so every ClickHouse run now needs planner.use_optimizer=false. The PR notes cover the e2e break but not this one. Options: default use_optimizer: false (this also avoids breaking e2e runs until the saturation and facts inputs are routinely available), have ClickHouse ignore the key, or keep the guard and document the required override.

[milindsrivastava1997]

The cost-table path is written in two places: here and as COST_TABLE in experiment_utils/services/misc.py. Pass it into the template (e.g. /app/saturation/{{ cost_table }}) so there is one source of truth.

[milindsrivastava1997]

Nit: --w-cpu/--w-mem sit outside {% if saturation_dir %}, and generate_controller_compose.py only checks that saturation-dir and facts are set together. Weights without a saturation dir would render flags the planner rejects (main.rs:125). That fails loudly, so it's harmless, but nesting the weights inside the block or adding them to the generator's parser.error check would be tidier.

[milindsrivastava1997]

The use_optimizer default is set in two places: config.yaml says true, but this .get(..., False) falls back to False (same at L778 in require_legacy_planner). Since config.yaml always defines planner, read cfg.planner.use_optimizer directly.

[milindsrivastava1997]

Nit: with the optimizer on, aggregate_cleanup and sketch_parameters are dropped without a message. That's unavoidable since config.yaml sets defaults for both, but an explicit CLI override (e.g. sketch_parameters.CountMinSketch.width=4096) vanishes quietly. Consider a one-line log when the optimizer is on.