[
https://issues.apache.org/jira/browse/SOLR-18444?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=18118133#comment-18118133
]
David Smiley commented on SOLR-18444:
-------------------------------------
To take advantage of this fully, this requires that queries all go to the
leader
{\{shards.preference=replica.leader:true}} – AKA Query-Only-Leader (QOL) +or+
query NRT replicas. Extending support to non-leader TLOG/PULL would require
smarter just-in-time replica synching – e.g. SOLR-18331
RE NRT: granted NRT kind of misses the point/implication here -- you are
probably committing too frequently to benefit. If you aren't committing
frequently with NRT then you should probably be using TLOG instead.
> Time coordinated/synchronized autoCommit
> ----------------------------------------
>
> Key: SOLR-18444
> URL: https://issues.apache.org/jira/browse/SOLR-18444
> Project: Solr
> Issue Type: Bug
> Reporter: David Smiley
> Priority: Major
>
> Distributed search across shards is about as slow as the slowest shard. The
> response time for a shard (a replica), in turn, can be heavily influenced by
> when the searcher was opened, because, over time, a searcher (including
> caches and lower level OS caching etc.) will get warmed up. I've seen a
> factor of 10 difference between shards (e.g. 25ms vs 250ms) for this
> difference, even ruling out queryResultCache and filterCache.
> This issue proposes that autoCommits are aligned; perhaps to an interval from
> the 1970 epoch
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
