[
https://issues.apache.org/jira/browse/SOLR-8696?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15172458#comment-15172458
]
Scott Blum edited comment on SOLR-8696 at 2/29/16 7:42 PM:
-----------------------------------------------------------
[~markrmil...@gmail.com] [~shalinmangar]
I think there's a slight problem with the fix as landed (merge artifact?)
{code}
zkStateReader.createClusterStateWatchersAndUpdate();
Stat stat = zkClient.exists(ZkStateReader.LIVE_NODES_ZKNODE, null, true);
if (stat != null && stat.getNumChildren() > 0) {
zkStateReader.createClusterStateWatchersAndUpdate();
publishAndWaitForDownStates();
}
{code}
createClusterStateWatchersAndUpdate() shouldn't be called twice, as it sets up
duplicate watchers. I actually think we should just have a single call at the
top, right after createClusterZkNodes() and right before joining the overseer
election, so that if we get elected we have a valid ClusterState to start with.
(For the record, I'm not super happy with the fact that external code needs to
worry so much about initializing ZkStateReader exactly once)
Attached a patch with a suggested fix.
was (Author: dragonsinth):
[~markrmil...@gmail.com] [~shalinmangar]
I think there's a slight problem with the fix as landed (merge artifact?)
{code}
zkStateReader.createClusterStateWatchersAndUpdate();
Stat stat = zkClient.exists(ZkStateReader.LIVE_NODES_ZKNODE, null, true);
if (stat != null && stat.getNumChildren() > 0) {
zkStateReader.createClusterStateWatchersAndUpdate();
publishAndWaitForDownStates();
}
{code}
createClusterStateWatchersAndUpdate() shouldn't be called twice, as it sets up
duplicate watchers. I actually think we should just have a single call at the
top, right after createClusterZkNodes() and right before joining the overseer
election, so that if we get elected we have a valid ClusterState to start with.
(For the record, I'm not super happy with the fact that external code needs to
worry so much about initializing ZkStateReader exactly once)
> Start the Overseer before actions that need the overseer on init and when
> reconnecting after zk expiration and improve init logic.
> ----------------------------------------------------------------------------------------------------------------------------------
>
> Key: SOLR-8696
> URL: https://issues.apache.org/jira/browse/SOLR-8696
> Project: Solr
> Issue Type: Improvement
> Components: SolrCloud
> Affects Versions: 5.4.1
> Reporter: Scott Blum
> Assignee: Mark Miller
> Labels: patch, performance, solrcloud, startup
> Fix For: master
>
> Attachments: SOLR-8696-followup.patch, SOLR-8696.patch,
> SOLR-8696.patch
>
>
> ZkController.publishAndWaitForDownStates() occurs before overseer election.
> That means if there is currently no overseer, there is ironically no one to
> actually service the down state changes it's waiting on. This particularly
> affects a single-node cluster such as you might run locally for development.
> Additionally, we're doing an unnecessary ZkStateReader forced refresh on all
> Overseer operations. This isn't necessary because ZkStateReader keeps itself
> up to date.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
---------------------------------------------------------------------
To unsubscribe, e-mail: dev-unsubscr...@lucene.apache.org
For additional commands, e-mail: dev-h...@lucene.apache.org
