[milindsrivastava1997]

Part of #143 (S5). The export itself hasn't been run yet; it will be run after review.

Changes

• exact-delta-set: mirrors ASAPQuery's DeltaSetAggregatorAccumulator, but without churn. Only the added set is kept (a key per group, as Vec<String>), so removed stays empty and merging is a union. It reuses the exact Multiple<C> wrapper with an empty Present cell. A new key-set comparator (the Presence aggregate: 1 for every key) scores it, and a missing key counts as infinitely wrong. Capability: KeySet.
  • Memory is divided by key count through the existing groups_per_instance normalization in atomic_costs.rs.
  • Our memory counts the label string bytes. ASAPQuery's memory_usage_bytes reports only size_of::<KeyByLabelValues>() (24 bytes) per key.
• HydraKLL: 5 configs on the new hydra_kll_disjoint_labels.yaml, using comparator subpop-rank-error. The new spec is hydra_columns_f64.yaml with a separate alphabet per label column (Overlapping label-column domains alias into one Hydra cell, and the comparator scores it as sketch error #74). On the shared alphabet, even the exact polars baseline shows a max rank error of 0.036; on the disjoint spec it scores 0. These rows are whole-sketch: memory and merge cost aren't divided per group, so they must not become optimizer candidates until rqe-optimizer: support shared multi-subpopulation sketches (HydraKLL) in the cost model #142 lands, or they'd be overpriced by card(G). Accuracy is measured at this spec's group count only.
• Export script: one measure helper runs both passes (point for inline zipf data, point_spec for spec files), replacing three near-copies. Every swept parameter is now in one block at the top of scripts/export_rqe_optimizer_costs.sh, and the expected row count goes from 18 to 24.

Swept parameters

Variant	Swept
kll-percall	k ∈ {200, 500}
dd	alpha ∈ {0.01, 0.02}
hydra-kll	rows=3 cols=128/256/512 cell_k=200, rows=5 cols=256 cell_k=200, rows=3 cols=256 cell_k=500
hll	lg_k ∈ {12, 14}
univmon-cardinality, univmon-topk	(heap 1000, rows 5, cols 2048, layers 8), (500, 3, 1024, 6)
cms-heap-topk, countsketch-heap-topk	rows ∈ {3, 5} × cols ∈ {2048}
exact-sum, exact-min, exact-max, exact-increase, exact-delta-set	none

Smoke run (1 run each, through flatten and atomic-costs)

• exact-delta-set: relative_error 0, 0 missing groups, about 104 B per key over 9,950 groups.
• hydra-kll (rows=3 cols=128 cell_k=200): 3.4 MB for the whole sketch, max rank error about 0.047 on the disjoint spec (0.042 on the shared one, where label collisions were part of the error).

For later sessions

rqe-optimizer doesn't reference either new variant yet. Adding them as candidates and costing the key tracker is S4b.

🤖 Generated with Claude Code