Export the AutoSketch vs ASAP evaluation table from saturation curves

[zzylol]

Before this PR

The AutoSketch vs ASAPQuery evaluation (ProjectASAP/ASAPQuery#777, section 6) had no accuracy/cost table. The only one, the 18-row rqe_atomic_costs.json, measures each config at a single size, with no data-shape matching, saturation or merge evidence.

After this PR

scripts/export_autosketch_eval_table.py writes rqe-optimizer/data/autosketch-eval/table.json.

Workloads. There is one workload per dataset in ASAPQuery's skew_summary.csv:

Dataset	RQEs	Config rows	Saturated rows
alibaba_v2022	51	1020	867
boom	20	140	47
google_2011	27	546	403

Each RQE's data parameters are the worst case over the whole dataset, rounded to the grid exactly as recommend_config.py does. Its grid rounding moved into a shared grid_point(). There is no burst variant: the whole-dataset worst cases replace AutoSketch's synthetic burst injection.

Per config. Every candidate config of every family lists:

• autosketch_error: the single-sketch curve at max_N. AutoSketch never merges, so it has no saturation requirement.
• asap_error_by_m for m = 1, 4, 16, 64. For CMS, CountSketch and DDSketch, every m reads the single-sketch value because they merge exactly. KLL and top-k read the merged curves from Measure accuracy after merging m shards #131.
• saturated (max_N ≥ n_sat): ASAP may only use a value when this is true.
• Costs from recommend_config.cost_of.

Assumptions (README, and the per-RQE assumptions field):

• groups = 1: a key query is one sketch over its group-by keys.
• λ = max_N / lookback.
• An instant query's lookback is one step.
• BOOM's N and window are derived from its 20 equal chunks per series, using the sampling step in the series name.

Inputs. KLL and top-k merged values come from #131 (merged into main as bd644fe, identical to the merged-accuracy head the curves came from). The raw curves (2.8 MB) are not committed; the README gives regeneration commands and input SHA-256s.

Validation

• python3 scripts/test_export_autosketch_eval_table.py: 6 passed. They cover:
  • exact merge giving the same value for every m;
  • the merged KLL curve at m = 4, and the flagged fallback at m = 16;
  • the saturation flag;
  • AutoSketch reading an unsaturated value;
  • the window and label set;
  • BOOM chunk derivation.
• python3 scripts/test_recommend_config.py: 11 passed, after the grid_point extraction.
• Regenerated the table from the full Add saturation study: error vs N, N_saturation and Pareto value data #130 grid with cost, the 1e9 subset, and Measure accuracy after merging m shards #131's top-k and quantile merge grids. Every KLL/top-k point has a merge curve.

🤖 Generated with Claude Code