[
https://issues.apache.org/jira/browse/SLIDER-930?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15003818#comment-15003818
]
Steve Loughran commented on SLIDER-930:
---------------------------------------
I thought we'd got some of that failure windowing automatically —I'd missed the
ability/need to set it locally. Make its a major issue. I'm going to change the
title to make it clearer what is happening
We don't need a functional test for this; a minicluster one will suffice
—simpler and something Jenkins will run automatically.
In {{TestStandaloneAMRestart}} the AM is restarted in a minicluster.
For a windowed test we could try to trigger more failures than a non-windowed
run would allow, and, if the AM restarts, we can conclude that the window was
reset.
This would be very similar to the current one, simply starting the YARN cluster
with the window/limit settings we need there, launch the AM with a very small
reset window, try to reset it twice, with a sleep longer than the window each
time, and verify that it always comes back up
> Incorporate Yarn feature of resetting AM failure count into Slider AM
> ---------------------------------------------------------------------
>
> Key: SLIDER-930
> URL: https://issues.apache.org/jira/browse/SLIDER-930
> Project: Slider
> Issue Type: Bug
> Components: appmaster
> Affects Versions: Slider 0.80
> Reporter: Gour Saha
> Assignee: Sherry Guo
> Fix For: Slider 0.90
>
> Attachments: SLIDER-930-001.patch
>
>
> YARN-611 provides this feature. Currently Slider apps are bound by the number
> set for yarn.resourcemanager.am.max-retries in the cluster. By default this
> value is set to 2, which is very low for long running services.
> Slider AM should use the feature provided in YARN-611 and set an interval
> after which the failure count will be reset to 0.
> I believe the API to call on ApplicationSubmissionContext is
> attemptFailuresValidityInterval. To start with Slider can set it to 5 mins
> which should be a reasonable default.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
