[
https://issues.apache.org/jira/browse/YARN-1927?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13968830#comment-13968830
]
Carlo Curino commented on YARN-1927:
------------------------------------
I agree with Chris's explanation, we are "sustaining" an ask, sort of
continuously publishing our latest belief about preemption (e.g., a massive job
could release lots of resources, and relief the pressure that caused us to ask
containers back)... The choice of how to propagate this information to the AM
is somewhat irrelevant. The current solution tries to be more or less aligned
with the general style of protocols and it has a couple of nice properties: 1)
losing a message is not a problem, and 2) it is simple for the AM to
reconstruct the latest RM opinion about resource preemption (simply re-stated
in each message). Repeating the ask, confirms to the AM that the need for
preemption is still there... this does convey some useful information to the
AM. Given the scale/frequency of operations this doesn't seem to be a perf
concern either. It is a matter of taste whether one likes better sustaining an
ask or start-stopping it.
Since this is a public protocol and I would suggest we consider to change it
only if there is a substantial gain in expressivity or performance... I don't
see one at the moment (but I might be missing your point). What you propose
seems a plausible alternative, but not substantially better than what's already
there, so I would lean towards leaving it be.
> Preemption message shouldn’t be created multiple times for same container-id
> in ProportionalCapacityPreemptionPolicy
> --------------------------------------------------------------------------------------------------------------------
>
> Key: YARN-1927
> URL: https://issues.apache.org/jira/browse/YARN-1927
> Project: Hadoop YARN
> Issue Type: Bug
> Components: capacityscheduler
> Affects Versions: 2.4.0
> Reporter: Wangda Tan
> Assignee: Wangda Tan
> Priority: Minor
> Attachments: YARN-1927.patch
>
>
> Currently, after each editSchedule() called, preemption message will be
> created and sent to scheduler. ProportionalCapacityPreemptionPolicy should
> only send preemption message once for each container.
--
This message was sent by Atlassian JIRA
(v6.2#6252)
