mikemccand opened a new issue, #15079: URL: https://github.com/apache/lucene/issues/15079
### Description [Spinoff from [this comment](https://github.com/apache/lucene/pull/14932#issuecomment-3167938331)] We recently merged a [nice opto to use GroupVarInt instead of VarInt for encoding each node's neighbors list in the HNSW (KNN) graph](https://github.com/apache/lucene/pull/14932). Benchmarks showed improvements to search latency, which we expect since GroupVarInt decodes four ints at once and puts less branch prediction stress on the CPU. But oddly nightly vector indexing benchmarks seemed to slow down, on 2025/08/06, and it's not yet obvious why -- it isn't one of the random "glitches" that that recover after a day or two (those are separately annoying! [`beast3` benchy box is old now](https://blog.mikemccandless.com/2021/01/apache-lucene-performance-on-128-core.html), ~4.5 years, which equates to 135 years in humans years! (1 computer year = 30 human years) Maybe it is getting a bit flakey in its later years...). I'll try to dig some more ... but if I get nowhere (likely!), I think we should temporarily revert GroupVarInt and see if benchy recovers? Especially because this might become a blocker for 10.3 release... ### Version and environment details _No response_ -- 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: issues-unsubscr...@lucene.apache.org.apache.org For queries about this service, please contact Infrastructure at: us...@infra.apache.org --------------------------------------------------------------------- To unsubscribe, e-mail: issues-unsubscr...@lucene.apache.org For additional commands, e-mail: issues-h...@lucene.apache.org
