Hi all, It seems that the changes introduced in 10.5.1 revealed a dormant bug that can silently regress KNN recall 5-7% (#16552 [1]), and I'm wondering if the community agrees this warrants a second bugfix release for 10.5.
Some background on the bug: #15003 [2] added support for HNSW re-use across merges for segments with deletes in 10.4.0, but re-use was silently disabled under per-field codecs the whole time, until #16400 [3] fixed it in 10.5.1. The issue is that this 're-use with deletes' path that is now enabled had a latent bug, which skipped repairing nodes where connections deteriorated slowly over many merges. After many repeated merges on segments with low delete percents, connections continue to erode, and this bug ends up reducing out-degree 40+% and dropping recall by an observed 5-7%. The bug was fixed in #16559 [4], but this seems like a big drop in recall that should be resolved quickly. Assuming others agree it's needed, would anyone more experienced than me be willing to help release such a patch? Thanks, Ethan [1] https://github.com/apache/lucene/issues/16552 [2] https://github.com/apache/lucene/pull/15003 [3] https://github.com/apache/lucene/issues/16400 [4] https://github.com/apache/lucene/pull/16559
