JingsongLi opened a new pull request, #8569:
URL: https://github.com/apache/paimon/pull/8569

   ## Summary
   
   Adds the bucket-local search kernel for primary-key vector indexes. A bucket 
can now merge ANN results with exact fallback results from active data files 
that are not covered by ANN yet, while returning physical source positions for 
the later read path.
   
   ## Changes
   
   - Introduce a shared `PkVectorSearchResult` containing data file name, 
zero-based physical row position, and normalized distance, and use it directly 
from ANN search instead of the unreleased nested candidate type.
   - Add bounded-memory exact Top-K search for `l2`, `cosine`, and 
`inner_product`, preserving physical positions across null vectors and applying 
deletion-vector exclusions.
   - Add bucket-local search that validates ANN source metadata against active 
files, avoids rescanning ANN-covered files, forwards deletion vectors and 
search options, and deterministically merges ANN and exact candidates.
   
   ## Testing
   
   - [x] Verify multi-source ANN results map to physical data-file positions 
through the shared result type.
   - [x] Verify exact search supports all configured metrics and rejects 
invalid queries.
   - [x] Verify null vectors and deletion vectors preserve physical row 
addressing.
   - [x] Verify bucket search merges ANN and exact fallback Top-K without 
rescanning covered files.
   - [x] Run the focused Core suite: 9 tests passed.
   - [x] Run non-fast Core compilation with Checkstyle, Spotless, and Maven 
enforcers.
   
   ## Notes
   
   - Follows #8563, which publishes and restores bucket-level primary-key 
vector indexes.
   - Table scan routing, split serialization, physical-row materialization, 
cross-bucket Top-K, and Flink source integration remain follow-up work.
   


-- 
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