[
https://issues.apache.org/jira/browse/YARN-5635?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15487832#comment-15487832
]
Ray Chiang commented on YARN-5635:
----------------------------------
I had captured something in a separate JIRA, but putting it here. This was my
attempt to sum up the discussion at the end of YARN-5567:
Done as a alternate design to YARN-5567. Define a specific exit code for the
health checker script (property yarn.nodemanager.health-checker.script.path)
that allows the node to be blacklisted.
As discussed in the latter part of YARN-5567, the current design requirements
are:
# Ignore all exit codes from the script
## _except_ the newly defined error code which will mark the NodeManager as
UNHEALTHY
## This allows any syntax or functional errors in the script to be ignored
# Upon failure (or multiple recorded failures):
## Store the status in the metrics2 state on the NodeManager
## Allow the RM to blacklist the NM or allow the jobs to drain, depending on
how we want UNHEALTHY to be treated
> Better handling when bad script is configured as Node's HealthScript
> --------------------------------------------------------------------
>
> Key: YARN-5635
> URL: https://issues.apache.org/jira/browse/YARN-5635
> Project: Hadoop YARN
> Issue Type: Bug
> Reporter: Allen Wittenauer
>
> Earlier fix to YARN-5567 is reverted because its not ideal to get the whole
> cluster down because of a bad script. At the same time its important to
> report that script is erroneous which is configured as node health script as
> it might miss to detect bad health of a node.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
