[ 
https://issues.apache.org/jira/browse/SOLR-13491?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16848250#comment-16848250
 ] 

Erick Erickson commented on SOLR-13491:
---------------------------------------

We're routinely seeing many hundreds of replicas, so I should think anything we 
can do to reduce the load on ZK would be A Good Thing.

[~gus_heck] Do you have any inkling of whether moving to ZK 3.5.5 would affect 
either the problem or the solution? I'm fairly close to being able to commit 
SOLR-8346.

> SolrZkClient can register the same watch twice.
> -----------------------------------------------
>
>                 Key: SOLR-13491
>                 URL: https://issues.apache.org/jira/browse/SOLR-13491
>             Project: Solr
>          Issue Type: Bug
>      Security Level: Public(Default Security Level. Issues are Public) 
>          Components: SolrCloud
>    Affects Versions: 8.0
>            Reporter: Gus Heck
>            Priority: Major
>
> While working on SOLR-13420 and related tickets I noticed that we presently 
> use the wrapWatch() method to wrap all zookeeper watches. This allows the 
> watch to be processed asynchronously by an executor which is good, but it 
> subtly changes the way zookeeper behaves. Normally zookeeper keeps a 
> Map<String,Set<Watch>> that ensures that if the same watch is registered 
> twice it's only notified once. However, our wrapper breaks this because now, 
> if the same watch is added twice each one gets it's own wrapper with default 
> Object.equals() and thus the two wrapped versions of the Watch aren not 
> .equals() and may co-exist in the Set<Watch>
> I do not know of any current cases for which this would cause a problem, but 
> changes I am contemplating for SOLR-13375 are likely to hit this case and at 
> a minimum create wasted re-notifications of the same event and excess watch 
> load in Zk.
> Moving the anonymous class wrapping to an explicit inner class and writing 
> hashcode/equals carefully to act as a pass through such that any to wrappers 
> with the same watch object are .equals() should fix this.



--
This message was sent by Atlassian JIRA
(v7.6.3#76005)

---------------------------------------------------------------------
To unsubscribe, e-mail: dev-unsubscr...@lucene.apache.org
For additional commands, e-mail: dev-h...@lucene.apache.org

Reply via email to