[jira] [Commented] (REEF-501) Distinguish different types of FailedEvaluator in Vortex

["John Yang (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/REEF-501?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14640436#comment-14640436 ] 

John Yang commented on REEF-501:
--------------------------------

Note: We may need to implement FailedContextHandler as well.

> Distinguish different types of FailedEvaluator in Vortex
> --------------------------------------------------------
>
>                 Key: REEF-501
>                 URL: https://issues.apache.org/jira/browse/REEF-501
>             Project: REEF
>          Issue Type: Sub-task
>          Components: Vortex
>            Reporter: John Yang
>            Assignee: John Yang
>
> Currently Vortex considers all failures the same, via FailedEvaluatorHandler. We should handle different types of failures differently.
> Type 1: Resource preemption
> We react based on a configured policy. (e.g. re-request infinitely) If needed we can even add a new event handler to REEF Driver named PreemptedEvaluatorHandler just for this type(a separate JIRA issue outside of the Vortex umbrella JIRA).
> Type 2: Internal Vortex code failure
> Can happen nondeterministically and even result in an infinite resource release+request. In such case, we should probably shut down the Driver immediately for the ease of debugging and to prevent it from interefereing with other jobs in the cluster.
> Type 3: Other types of failures
> If the failure is caused by issues like OOM then we also treat such case differently.



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)