[
https://issues.apache.org/jira/browse/LUCENE-10374?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17477489#comment-17477489
]
Greg Miller commented on LUCENE-10374:
--------------------------------------
As an update here, merging just part of this original work (counting directly
in the dense vs. sparse values instead of doing a null check on every iteration
of the loops to see which one to increment) resulted in a performance
improvement in the nightly benchmarks. I experimented with different versions
of the second part (avoiding the liveDocs null check on each iteration of the
loop) without success ([https://github.com/apache/lucene/pull/606).] I wasn't
able to observe any qps improvement locally.
At this point I'm not sure what could have caused the difference between the
nightly bench runs and our local runs, or what could have resulted in the
regression seen in the nightly runs. It also looks possible that noise is at
least playing some role in this (have a look at the most recent runs here:
[https://home.apache.org/~mikemccand/lucenebench/BrowseMonthTaxoFacets.html).]
So I'm not sure it's worth chasing this further right now. Or if we do want to
chase it, I think we'll need to work closely with [~mikemccand] and the nightly
bench hardware. I'll leave this open for now but I'm going to table it for the
time being.
> Track down the "browse" taxonomy faceting qps regression
> --------------------------------------------------------
>
> Key: LUCENE-10374
> URL: https://issues.apache.org/jira/browse/LUCENE-10374
> Project: Lucene - Core
> Issue Type: Improvement
> Components: modules/facet
> Reporter: Greg Miller
> Assignee: Greg Miller
> Priority: Major
>
> We need to track down the source of the regression observed here:
> [https://home.apache.org/~mikemccand/lucenebench/2022.01.10.18.03.12.html.]
> Some details on the regression hunting are in
> [https://github.com/apache/lucene/pull/597.]
--
This message was sent by Atlassian Jira
(v8.20.1#820001)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
