Repository navigation
bench: measure hydra-kll without subset fan-out (one key per grouping) #165
Description
Found in review of #160 (S4b of #143).
Problem
The optimizer models hydra-kll as one sketch keyed by the grouping G: each sample hashes its joined G value into the grid and updates the KLL cell there, so the sketch holds exactly card(G) subpopulations. S4b's eligibility gate (card(G) ≤ subpopulations) assumes this.
The measured rows fan out instead. aqpbm-cli/src/rows/materialise.rs:30 joins the label columns with ;, and asap_sketchlib 0.2.2's Hydra::update splits on ; and inserts into all 2^D − 1 non-empty subsets. On hydra_kll_disjoint_labels.yaml (D = 2) that means 3 inserts per sample: key1, key2, key1;key2. Compared with the intended sketch, the rows:
- overstate insert CPU by about 3×;
- understate accuracy: the cells carry N·(2^D−1) units of mass instead of N, so there is more collision noise in the same r×c;
- misreport load:
subpopulationscounts only the probed key1 groups, not what the sketch held.
Fix
Measure hydra-kll with one subkey per sample. Any of these works:
- 0.2.2: join the grouping values with a separator other than
;, soupdatesees one part (D = 1); asap_sketchlib0.3 frameworkHydra::with_schema(r, c, ["series"], kll)with the joined label set as the single column;- the portable
HydraKllSketch(message_pack_format/portable/hydra_kll.rs), which never fans out.
Then re-measure the hydra-kll rows (ideally at several group counts too, per the S5 note on #143), with subpopulations equal to the number of distinct joined keys.
Also check that ASAPQuery's HydraKLL engine path keys by the joined grouping without fan-out. Otherwise the deployed sketch won't match the costed one.
🤖 Generated with Claude Code
Activity
fix(rqe-optimizer): decide shared-sketch validity in is_eligible, one…
feat(rqe-optimizer): family properties, shared fixed-size sketches an…
milindsrivastava1997 commentedon Oct 7, 2026
Scope decisions
asap_sketchlib's Hydra as it ships, and nothing consumes them.Hydra. The hydra-kll rows join label values with\u{1f}(SERIES_SEPARATOR), whichHydra::updatedoesn't split on, instead of;. Queries join with the same separator. The join happens at materialisation, off the timed insert path.subpop-rank-errorfor hydra-kll groups by every label column, not just key1. The polars baseline groups by the full key (nofan_out), sosubpopulationsis the number of distinct label sets, which is what the optimizer'scard(G) ≤ subpopulationsgate assumes.measured_at_group_countcomment inrqe-optimizer/src/candidates.rs, which cites this issue.HYDRA_KLL_CONFIGSentry on each spec, merge-accuracy runs included. No pipeline change is needed:same_cellkeys on the full table description, so the specs are separate cells, and the optimizer's gate already takes the smallest measured group count ≥card(G). The stale comments in the existing spec (key1 as "the column the comparator groups by", key2 skew for "subset fan-out") get fixed.Out of scope
HydraKllSketch(xxh32, its own KLL,k=20by default), not the frameworkHydrameasured here. Neither fans out, but costs can differ. Tracked in HydraKLL: the costed sketch (sketch-bench Hydra framework) is not the deployed one (portable HydraKllSketch) ASAPQuery#799. We decided to keep measuring the frameworkHydrahere.🤖 Generated with Claude Code
hydra-kll: insert one key per label set, not every label subset (#165)
hydra-kll: measure at ~100, ~1k and ~10k groups (#165)
zzylol commentedon Oct 10, 2026
Closing: the scope here (measure hydra-kll with one key per grouping, no subset fan-out) contradicts the Hydra admission design in #193.
(decided by @zzylol)
🤖 Generated with Claude Code