[zzylol]

Closes #174. Builds on #171 and #173.

Why

The optimizer read its inputs from two runs of the same binary: export_rqe_optimizer_costs.sh (cost, one accuracy point, a 4–1024-shard merge-accuracy sweep) and study_saturation.py (curves). Since #171 the export's accuracy and merge sweep feed nothing, and the two runs had to be kept in sync by hand.

What

• One study. study_saturation.py --phase optimizer-cost runs one serial cost pass (5 runs, 3 warm-ups) and one accuracy pass per (sketch, config), plus the five exact accumulators. It reduces them with atomic-costs to rqe_atomic_costs.json and fails if any row is skipped. There is no merge sweep. export_rqe_optimizer_costs.sh is deleted.
• One data shape, the AutoSketch vs. ASAP evaluation's dataset (ASAPQuery#777): Zipf θ = 1.1 over 1e4 keys (Pareto a = 2 for quantiles), 1e6 items. Cost is taken as independent of shape.
• Each row names its accuracy. AtomicCostEntry.accuracy_metric is required. It is the query_accuracy key that holds the row's accuracy. atomic-costs --accuracy-metric VARIANT=METRIC replaces --merge-accuracy. measured_at is now required. merge_accuracy is removed. Old tables are not migrated; we rerun.
• rqe-optimizer.
  • Exact accumulators read the row's accuracy without merge brackets (accuracy_at removed; row_accuracy). They merge without loss.
  • check_cost_table also refuses a row whose accuracy_metric isn't its family's or its curve's.
  • small_problem reads the table from --saturation-dir (optimizer_cost/rqe_atomic_costs.json). --cost-dir is removed.
• ASAPQuery's grid. export_atomic_costs.sh passes --accuracy-metric.

Kept on purpose

• query_accuracy: ASAPQuery's copy of AtomicCostEntry requires it.
• The study's per-point cost: crossover.csv and the plots use it.
• Families with no measurements now (univmon-*, countsketch-heap, hydra-kll) and the HydraKLL-only code: left for a separate decision.

Needs attention

• KLL/top-k merge penalty. Since rqe-optimizer: read sketch accuracy from saturation curves #171, nothing reads merge_accuracy for sketches. rqe-optimizer: use merged accuracy curves for KLL and top-k #158 was closed assuming it did. The penalty should come from the study's merge curves (follow-up).
• ASAPQuery. Its AtomicCostEntry rejects unknown fields, so a table with accuracy_metric needs a companion change there.
• Cross-check coverage. θ = 1.1 and K = 1e4 are not on the default curve grid, so check_cost_table reports the Zipf rows as unchecked. Only the Pareto rows are cross-checked.

Verification

• cargo test for aqpbm-core (94), aqpbm-cli and rqe-optimizer (58), on main after rqe-optimizer: check each family by its own accuracy metric #173. cargo clippy --all-targets -D warnings and cargo fmt --check are clean.
• python3 -m unittest test_study_saturation passes (28). It adds a fake-binary test of the phase's commands and its skipped-row failure.
• In progress on clnode109: --phase optimizer-cost --one-config with the real binary, then small_problem against it. Results will be posted here.

🤖 Generated with Claude Code

[zzylol]

End-to-end check on clnode109 (commits through 716d08c):

• study_saturation.py --phase optimizer-cost --one-config with the real binary: 11 rows (6 sketches + 5 exact accumulators), atomic-costs: 11 row(s), 0 skipped. Each row names its metric: are_top100, precision_at_k, relative_error, mean_rank_err, mean_relative_value_error, and the exact rows relative_error = 0.
• small_problem --milp --saturation-dir against that table:
  • The existing curves (from-node0/saturation) predate Report DDSketch relative value error #164/study_saturation: record DDSketch mean relative value error #176. rqe-optimizer: check each family by its own accuracy metric #173's load check refuses their DD curve (dd curve measured mean_rank_err), as intended. With the DD rows removed from a copy of the curves, it plans: 8 RAQEs, 115 candidates, MILP solved.
  • check_cost_table checked the KLL row (Pareto a = 2, N = 1e6) against its curve, and it agrees.
  • The Zipf rows show as unchecked, because the old grid has no (θ = 1.1, K = 1e4) point. 716d08c adds that shape to the grid, so a rerun of the curves covers them.

Also in 716d08c, from review:

• export_atomic_costs.sh now runs the cost pass first. flatten keeps the first pass's timings, so the accuracy-first order skipped every row. The accuracy pass now runs once with no warm-ups, and the script fails on skipped rows.
• --phase optimizer-cost shows atomic-costs' error and matches its row count exactly.

🤖 Generated with Claude Code