[GordonYuanyc]

Summary

• Dependency. The workspace pins asap_sketchlib to git rev a05b35c, which is 0.4.0 plus main through #262. aqpbm-planeval takes the same workspace dependency, so Cargo.lock holds two copies, ours and ASAPPlanner's own 21b6357, instead of three. To move to another sketchlib commit, edit rev in the root Cargo.toml.

• 0.4.0 names.

  • Every write is now update (UnivMon's is update_weighted).
  • Type renames: CountMinSketch, CountSketch, CountMinSketchWithHeap, CountSketchWithHeap, HeapEntry, Kll (with_k, Cdf::value_at), DdSketch, Ertl, HyperLogLogHip, HllRegistersP*.
  • UnivMon: new, layer_heaps, estimate_*.
  • The HydraCounter variants follow the new names.
  • The registry descriptions and existing comments are renamed to match.

• Memory. Every sketchlib row now reports the library's own memory_bytes(). The parallel-insert rows report workers × the library's theoretical_memory_bytes for the shape each worker uses. Three hand-written formulas copied 0.3's internals and are deleted, along with the tests that asserted them:

  • aqpbm-core's heap_footprint (the 0.3 HHHeap layout)
  • the KLL slot count
  • the Hydra grid overhead

  The DDSketch formula and the oxide rows' formulas stay.

• rows/ moves from aqpbm-cli to sketch-bench. aqpbm-cli/src/rows is now sketch-bench/src/rows, and the CLI calls sketch_bench::rows::measurements. aqpbm-cli's source no longer names any asap_sketchlib type; it still links the library through sketch-bench. The re-export of sketchlib types in wrappers/mod.rs is gone.

Numbers that change

• Memory for every sketchlib row now comes from the library. It includes the struct, and for heaps and KLL it is the state at the moment of measurement, not a modelled steady state.
• Accuracy changes because 0.4.0 fixed the UnivMon cardinality and L2 estimators.
• Throughput may change from #245 (heap rescoring), #246 (KLL compaction moved out of the insert path) and #247 (Count Sketch medians taken on the stack).

Earlier results should be re-baselined rather than compared directly.

Docs

• docs/rqe_optimizer_hydra.md points at sketch-bench/src/rows/mod.rs, drops the 0.3.0 mentions, and states the footprint as Hydra::memory_bytes.
• docs/e2e_planning_flow.md drops the 0.3.0 mention.
• README.md shows the hll sample's memory_bytes as the library reports it (16400 at lg_k = 14).

The results READMEs under rqe-optimizer/results/ still say 0.3.0, because those runs were measured on it.

Test plan

• cargo build --release
• cargo test --workspace --no-fail-fast: all pass
• cargo fmt --all --check
• approxbench sketchbench --variant cms-heap-topk-fastpath-vector2d --library lib --config 'rows=5 cols=2048 heap=128' on 1M Zipf(1.1) keys: insert 30.3 M items/s, memory_bytes 58,904 (40,960 bytes of counters plus the struct and a 128-entry heap)

🤖 Generated with Claude Code