[jira] [Commented] (SOLR-11730) Test NodeLost / NodeAdded dynamics

["Andrzej Bialecki (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/SOLR-11730?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16282027#comment-16282027 ] 

Andrzej Bialecki  commented on SOLR-11730:
------------------------------------------

Simulations indicate that even with significant flakiness the framework may not take any actions if there are other events happening too, because even if a node lost trigger creates an event that event may be discarded due to the cooldown period. And after the cooldown period has passed the flaky node may be back up again, so the event would not be generated again.

> Test NodeLost / NodeAdded dynamics
> ----------------------------------
>
>                 Key: SOLR-11730
>                 URL: https://issues.apache.org/jira/browse/SOLR-11730
>             Project: Solr
>          Issue Type: Sub-task
>      Security Level: Public(Default Security Level. Issues are Public) 
>          Components: AutoScaling
>            Reporter: Andrzej Bialecki 
>
> Let's consider a "flaky node" scenario.
> A node is going up and down at short intervals (eg. due to a flaky network cable). If the frequency of these events coincides with {{waitFor}} interval in {{nodeLost}} trigger configuration, the node may never be reported to the autoscaling framework as lost. Similarly it may never be reported as added back if it's lost again within the {{waitFor}} period of {{nodeAdded}} trigger.
> Other scenarios are possible here too, depending on timing:
> * node being constantly reported as lost
> * node being constantly reported as added
> One possible solution for the autoscaling triggers is that the framework should keep a short-term ({{waitFor * 2}} long?) memory of a node state that the trigger is tracking in order to eliminate flaky nodes (ie. those that transitioned between states more than once within the period).
> Situation like this is detrimental to SolrCloud behavior regardless of autoscaling actions, so it should probably be addressed at a node level by eg. shutting down Solr node after the number of disconnects in a time window reaches a certain threshold.



--
This message was sent by Atlassian JIRA
(v6.4.14#64029)

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