[
https://issues.apache.org/jira/browse/SOLR-18331?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=18114705#comment-18114705
]
David Smiley commented on SOLR-18331:
-------------------------------------
Or we write _something_ to ZK (not state.json!) for each shard leader's commit
flush. Followers can "watch". No need for an HTTP message. That _something_
could be interesting info like number of docs, number of segments, number of
bytes, Lucene index generation, commit metadata. This info would in turn be
useful for diagnostic purposes without having to consult (query) replicas.
> TLOG/PULL should be notified to pull just-in-time; no polling
> -------------------------------------------------------------
>
> Key: SOLR-18331
> URL: https://issues.apache.org/jira/browse/SOLR-18331
> Project: Solr
> Issue Type: Improvement
> Components: SolrCloud
> Reporter: David Smiley
> Priority: Major
>
> Proposal: a leader should notify it's followers to pull just-in-time (a fresh
> open searcher). Compared to polling, this can reduce the worst case indexing
> latency by just over 50%, since the default poll time is half the
> auto-commit. TBD if the leader or the follower delays if it's notified too
> quickly. The notification might even provide the list of files & timestamps
> – whenever a follower first needs to get from the leader when it synchronizes.
> Today: It's sad that a TLOG/PULL replica that isn't the leader will poll the
> leader. For some shards, there could be great stretches of time of nothing
> happening. Each poll results in observability noise. Transient cores is now
> gone but nonetheless the polling keept an unused leader from getting evicted.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
