Reuse entry point scores and provide mechanisms to provide scores for directly entry points #14256
Conversation
… entrypoint iterators
|
Hi, @benwtrent , this bakes support for this supplying scores feature pretty deeply. I was thinking if we were to use this only for the SeededKnnVectorQuery, it might suffice to create a wrapping |
| * | ||
| * @lucene.internal | ||
| */ | ||
| public sealed class MappedDISI extends DocIdSetIterator { |
There was a problem hiding this comment.
whoa I had to look up what a sealed class is - I guess it is like final but you can allow some classes to inherit?
| eps = ArrayUtil.growExact(eps, candidates.size()); | ||
| epsScores = ArrayUtil.growExact(epsScores, candidates.size()); | ||
| epCount = candidates.size(); | ||
| candidates.collectNodesAndScores(eps, epsScores); |
There was a problem hiding this comment.
you are re-using the scores from upper levels when scoring the lower levels?
|
I also tried something based on the simpler approach I mentioned and also saw very minor gains in the seeded search with reentry when reusing scores. |
Yeah, I don't expect this to be a huge performance gain. I would expect the actual bottlenecks addressed in your changes (thrashing, over exploration, etc.) would dominate any savings. |
Spinning out of: #14226
That particular evolution of kNN querying is attempting to re-entry individual segment graphs with new exit and search criteria. To prevent having to rescore the new entry points, this PR provides the ability to (optionally) keep track of the scores for entry points.
Additionally, this will take advantage of entry point score retention during graph building and searching. The performance improvements are marginal.