[
https://issues.apache.org/jira/browse/SOLR-6790?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14223960#comment-14223960
]
Timothy Potter commented on SOLR-6790:
--------------------------------------
Errrr ... no, I'm wrong ;-) I had the graphs for the leader and replica
switched, meaning there was higher load on the replica than on the leader in
this case. That must be because there are more requests coming into the replica
because batches sent to the leader are broken up and then many more batches are
sent to the replica (via CUSS). I'll table this feature for now.
> Implement a leader-serves-reads flag
> ------------------------------------
>
> Key: SOLR-6790
> URL: https://issues.apache.org/jira/browse/SOLR-6790
> Project: Solr
> Issue Type: New Feature
> Components: SolrCloud
> Reporter: Timothy Potter
>
> ZooKeeper has a flag that allows you to disable leaders serving reads, from
> some basic indexing performance analysis, leaders have visibly higher CPU and
> heap usage. Here are some screen shots that show the difference between a
> leader and replica in a 1x2 collection with no query activity (just indexing):
> leader - http://www.dropmocks.com/mHoPUx
> replica - http://www.dropmocks.com/mHoWpX
> There's a clear difference in peak & avg CPU and heap usage. I'm guessing the
> difference would be even more pronounced if there were more replicas involved.
> Setting the leader-serves-reads=false allows the leader to only do indexing
> and serve no queries so long as there are other active replicas. I'm thinking
> this will be true by default but able to set to false per collection. Of
> course if only the leader is up, then it will serve reads.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
