[
https://issues.apache.org/jira/browse/YARN-8044?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16416093#comment-16416093
]
Shane Kumpf commented on YARN-8044:
-----------------------------------
{quote}What if binary doesn't exist on one of the faulty node due to disk
failure, and exit code is -1. We will want the retry to happen on some other
nodes.
{quote}
I agree that we would want to retry in that case and can see the challenge with
using exit codes.
{quote}We might want to use the heuristic approach with failure validity
intervals. We might be able to count number of failures within the time frame
to decide if we should abort the containers.
{quote}
Make sense to me. It seems YARN-5015 / YARN-8032 addresses this approach.
Given the above, would it make more sense to re-purpose this issue to expose
the retry policy used by Native Services to the end user? We could use
RETRY_ON_ALL_ERRORS as the default.
> Determine the appropriate default ContainerRetryPolicy
> ------------------------------------------------------
>
> Key: YARN-8044
> URL: https://issues.apache.org/jira/browse/YARN-8044
> Project: Hadoop YARN
> Issue Type: Sub-task
> Reporter: Shane Kumpf
> Priority: Major
>
> {{AbstractLauncher}} sets the retry policy to {{RETRY_ON_ALL_ERRORS}}, which
> may be too inclusive. Some error codes, such as -1, should likely result in a
> hard fail.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
