[zzylol]

Denser config grid, measured for the new configs only, and the §6.3 rerun on it, on #210's workload (42 / 62 / 104 RQEs).

Code (study_saturation.py)

• KLL k ∈ {25, 50, 100, 200, 400, 800} (was 50/200/800). DDSketch α adds 0.1.
• The Hydra eval sweep adds W = 65536 for hydra-hll and hydra-kll. hydra-univmon-cardinality is capped at 16384 (HYDRA_MAX_W): its grid at 65536 is about 20 GB, and its error has already saturated.
• --sketches / --configs measure only new configs, so they can be appended to an existing study (merge_study_parts.py refuses duplicate points).
• Hydra cost rows wider than 16384 merge 4 shards instead of 16, because 16 grids of 3.2 GB OOM. Merge cost is still per merge.
• HLL stays at lg_k 12/14/16, the precisions asap_sketchlib builds.

Inputs (new configs only, appended)

• Curves on clnode130, merged into autosketch-vs-asap-inputs/saturation/. Raw records are in raw/dense_*.
• 7 cost rows measured serially on idle clnode138: 63 rows in all.
• 999 Hydra runs at W = 65536 on clnode109/155/167/178: 7,778 rows in all.

Results: the 12 workloads at 95% (results, figures, plans) and at 90% and 99%, on clnode109/155/167/178. 0 RQEs dropped. One sanity check fails by solver tolerance: all-shared8 at 99%, Fargate frontier, 0.01% (see the targets README).

95%, cheapest plan	ASAP, CPU (vCPU), old → new	ASAP, Fargate ($/h), old → new
mixed	2.38 → 2.35	0.170 → 0.149 (−13%)
multi-grouping	0.221 → 0.220	0.0137 → 0.0134
mixed + multi-grouping	2.60 → 2.57	0.184 → 0.162 (−12%)

• CPU cost barely moves. Insert CPU per item is flat across configs (KLL ≈ 30 ns, DDSketch ≈ 19 ns).
• Memory-weighted cost falls up to 13%. Plans pick a KLL or DDSketch closer to the target size.
• 90% still equals 95%. HLL still bottoms out at lg_k=12, and DDSketch α = 0.1 (0.048 error) meets both targets.
• Hydra is never chosen. W = 65536 brings its error on groups holding ≥ 1% from 0.283 to 0.039 (hydra-hll, http), but it is the most expensive grid on every axis.

🤖 Generated with Claude Code