JingsongLi opened a new pull request, #122:
URL: https://github.com/apache/paimon-vector-index/pull/122

   IVF-FLAT repeatedly decoded partition payloads across searches, evaluated 
shared batch filters for each query, and rescanned all Top-K slots after 
accepted candidates. Reuse aligned, unfiltered partitions within the reader 
memory budget, compact filters shared by enough queries, batch short-vector L2 
scans, and use an indirect max heap for K >= 32 while preserving tie ordering.
   
   The unified reader now passes its existing memory-budget options to 
IVF-FLAT. Direct `IVFFlatIndexReader::open` remains uncached; the new 
`open_with_options` enables caching. Oversized lists bypass the cache. 
Regression tests cover cache bounds, filter isolation, I/O retries, metadata 
invalidation, heap ties, and range-search semantics. The v1 storage format is 
unchanged.
   
   Local project `ann_bench` measurements (M4 Pro, 8 threads, Top-10, 
nprobe=64, two runs per version) showed warm batch QPS gains of 43% on SIFT1M, 
4% on GIST1M, and 53% on GloVe100, with unchanged recall. Warm single-query 
throughput improved 5.15–5.82x. GIST's first batch regressed approximately 16%, 
so high-dimensional batch gains are limited and cache-fill overhead remains a 
tradeoff.
   
   Validation on an isolated checkout of this commit:
   - `cargo fmt --all -- --check`
   - `cargo clippy --all-targets --workspace -- -D warnings`
   - `cargo test --workspace --quiet`: 723 passed, 2 ignored.
   
   Benchmark harnesses, scripts, and documentation changes are excluded from 
this PR.
   


-- 
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

To unsubscribe, e-mail: [email protected]

For queries about this service, please contact Infrastructure at:
[email protected]

Reply via email to