[
https://issues.apache.org/jira/browse/SLIDER-494?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14219534#comment-14219534
]
Steve Loughran commented on SLIDER-494:
---------------------------------------
That could work.
Some initial thoughts
# The AM would have to build up some structure of completed instances, maybe
just another counter in the {{RoleStatus}} field.
# It would add more containers until the number of successful completions
matched the number of instances requested.
# status code would come back from the agent.
when a stopped instance is started it would re-run all the commands
where we'd have problems is on AM restart: the restarted AM doesn't remember
anything from the previous instance, so re-execute everything.
Question is: does that matter? We care a lot about AM restart as over a few
weeks it becomes an issue, and on short lived apps the risk is less. We could
even allow app configs to set the option to say "don't keep containers over
restarts" and the AM retry count itself (which can be set in
{{slider-client.xml}} today, in {{slider.yarn.restart.limit}}. That way, run a
short-lived app and if the AM fails, the app itself halts; no attempt to recover
> Support for short-lived services
> --------------------------------
>
> Key: SLIDER-494
> URL: https://issues.apache.org/jira/browse/SLIDER-494
> Project: Slider
> Issue Type: Bug
> Components: app-package, appmaster
> Affects Versions: Slider 0.50
> Reporter: Sumit Mohanty
> Fix For: Slider 2.0.0
>
>
> Slider applications could be short lived in the sense that the component
> instances are not daemons but are processes that stop after they are done.
> These processes may also fail in which case they may be auto-restarted or
> they can complete gracefully in which case the application instance should
> terminate (when all component instances are complete).
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
