SOTA experiment: density aware approximate cache for disk based ANNS

Finding

Aker, arXiv:2609.03712, integrates a density aware approximate cache into pgvector. The originating team reports up to 64 percentage points higher recall than prior approximate caching, up to 3.2x QPS, and 0.6x the memory of pgvector shared buffers.

Evidence class: peer reviewed PVLDB extended paper with originating team measurements. Not independently reproduced by RuVector.

Opportunity Score: 4.529.

RuVector mapping

This is relevant to RuVector disk and hybrid indexes, Cognitum hosted retrieval, Core Memory recurring retrieval, RuView historical spatial indexes, and RuVector WASM only where a bounded local cache is useful.

The reusable primitive is not the pgvector implementation. It is a per query cache admission and hit threshold derived from local neighborhood density plus an update consistency contract that eagerly invalidates deletions and lazily incorporates insertions.

Experiment

Compare under identical vectors, queries, update trace, hardware, storage, seed, and memory budget:

1. current RuVector disk or hybrid search with no approximate result cache
2. simple fixed similarity cache
3. density aware threshold cache
4. density aware cache with deletion eager and insertion lazy refresh

Workloads must include low and high locality query traces, clustered and uniform vector distributions, read heavy and update heavy mixes, deletion bursts, insert bursts, cold start, cache thrash, and adversarial near boundary queries.

Required metrics

Recall at 10, NDCG where applicable, QPS, p50 and p95 and p99 latency, disk reads per query, cache hit rate, false hit rate, stale result rate, memory, build or warmup time, update overhead, CPU, energy if measurable, and modeled infrastructure cost.

Falsification

Reject the mechanism if a fixed threshold cache performs within variance, if gains disappear on low locality workloads, or if update consistency cost erases the QPS benefit.

Promotion gate

At least 30 percent higher QPS or 25 percent fewer disk reads at matched recall, recall at 10 loss no greater than 0.5 absolute points, p99 latency regression below 5 percent, stale deletion results equal to zero after acknowledged deletion, and memory no greater than the stronger baseline.

Governance

Additive experiment only. No default index change, storage migration, autonomous merge, or production rollout. Existing search path remains the rollback.

Start by locating the current RuVector disk or hybrid search path and its existing caching and benchmark entry points. Compare the four cache configurations across the listed workloads and metrics; the experiment is done only if it meets the promotion gate without changing defaults or production behavior.