I agree that the behavior has been with us for a long time but still I think it is a bug that can be painful for big vector deployments. I would like to consider it for the bugfix release as the change looks pretty safe.
On Wed, Aug 5, 2026 at 4:06 PM Michael Sokolov <[email protected]> wrote: > is this really a candidate for a bugfix release? We've had this > behavior ever since 9.0 -- seems more like an improvement that could > wait for the next minor release. But, you know if the Ignacio wants to > take it on, I wouldn't block it. > > On Wed, Aug 5, 2026 at 10:52 AM Eran Weiss <[email protected]> wrote: > > > > Hi all, > > > > I've opened a backport of #16368 to branch_10_5: > https://github.com/apache/lucene/pull/16490 . > > > > The fix makes HNSW graph construction check the merge abort flag before > each node insertion, so IndexWriter#rollback and abortMerges no longer > block until the entire graph is built. Without it, rollback on a large > vector segment can block for tens of minutes. The fix has already been > merged to main and branch_10x (#16391). > > > > Is it possible to include it in 10.5.1? > > > > Thanks, > > Eran. > > --------------------------------------------------------------------- > To unsubscribe, e-mail: [email protected] > For additional commands, e-mail: [email protected] > >
