kaivalnp commented on PR #16771: URL: https://github.com/apache/lucene/pull/16771#issuecomment-6025360897
> can't we do similar alignment to Bytes and F16's? I think the alignment is beneficial for something that is repeatedly read from the index on the "hot" path, but cannot be kept in-memory. Vector re-scoring might be a good use-case, because raw vectors can be "cold". But we don't allow re-scoring by bytes, so the additional padding is useless there? (assuming they fit in memory to be used by HNSW search directly, the padding would help otherwise). Separately: I wonder if we should align the quantized bytes too? (making no assumptions about whether the application has enough RAM for these). This could be complemented with some prefetching for neighbors known to be expanded soon. This change makes sense to me overall, but another set of eyes would be appreciated. We also need to update [this comment](https://github.com/apache/lucene/blob/5a812aa07091fad934665b4530674923c44df312/lucene/core/src/java/org/apache/lucene/codecs/lucene99/Lucene99FlatVectorsFormat.java#L66-L70). -- 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] --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
