[
https://issues.apache.org/jira/browse/LUCENE-9695?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17273019#comment-17273019
]
ASF subversion and git services commented on LUCENE-9695:
---------------------------------------------------------
Commit 38ec2602cee395029497af992be7a217f97229d3 in lucene-solr's branch
refs/heads/master from Michael Sokolov
[ https://gitbox.apache.org/repos/asf?p=lucene-solr.git;h=38ec260 ]
LUCENE-9695: don't merge deleted vectors (#2239)
> Don't include deleted documents when merging vectors
> ----------------------------------------------------
>
> Key: LUCENE-9695
> URL: https://issues.apache.org/jira/browse/LUCENE-9695
> Project: Lucene - Core
> Issue Type: Improvement
> Reporter: Michael Sokolov
> Priority: Major
> Time Spent: 50m
> Remaining Estimate: 0h
>
> While testing HNSW searches with multi-segment indexes, all kinds of strange
> things were happening; recall performance was radically different for a
> force-merged multi-segment index than for the same index built as a single
> segment. Most testing I've done to date has been with single-segment indexes,
> shame on me.
> One issue is that when merging we iterate over all the vectors from 0 ..
> size-1. But this size was being calculated without taking deletions into
> account, and this caused deleted vectors to be included in the graph leading
> to exceptions and weird inconsistencies.
> The other issue has to do with aliasing in the diverse neighbor selection
> graph construction heuristic introduced recently. Sometimes vectors to be
> compared would be drawn from the same VectorValues, but this is a no-no since
> they are then the same vector (the first one will be overwritten when the
> second one is fetched). This leads to poor results, but not errors per se,
> but the results also became unpredictable in a way that causes the test
> written to reproduce the first issue to fail. Thus I'll include both fixes
> together.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]