[
https://issues.apache.org/jira/browse/YARN-5357?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16755751#comment-16755751
]
Vrushali C commented on YARN-5357:
----------------------------------
Notes from our discussion around this. Abhishek, Sushil, Prabha, Rohith and I
attended the discussion call.
Current situation:
Timeline Service collector's address & port are sent from the RM to NMs which
run containers for this app to publish the system metrics from the NM. In case
of federated yarn clusters, since an Uberized AM is used for coordinating with
containers running in foreign subclusters, the NMs in foreign sub-clusters do
not have the timeline service collector address.
This is not specific to federation but is a design item to be solved for any
unmanaged AM.
We discussed a few approaches. One of which was allowing the AM itself to send
the timeline service address & port to other foreign subclusters' RM and then
propagating it from there. But this has some issues. A malicious AM might try
to "game" the system by sending those a dummy collector address to some NMs.
This way, some NMs' system metrics mysteriously disappear before being reported
and therefore chargeback & accountability will become incorrect.
Another idea Abhishek proposed was around picking a random NM for starting up a
timeline-service collector when the AM is uberized. He will be working on the
design idea further and will discuss with his team mates and we can all review
it when he has consolidated his thoughts.
> Timeline service v2 integration with Federation
> ------------------------------------------------
>
> Key: YARN-5357
> URL: https://issues.apache.org/jira/browse/YARN-5357
> Project: Hadoop YARN
> Issue Type: Sub-task
> Components: timelineserver
> Reporter: Vrushali C
> Assignee: Prabha Manepalli
> Priority: Major
>
> Jira to note the discussion points from an initial chat about integrating
> Timeline Service v2 with Federation (YARN-2915).
> cc [~subru] [~curino]
> For Federation:
> - all entities that belong to the same flow run should have the same cluster
> name
> - app id in the same flow run strongly ordered in time
> - need a logical cluster name and physical cluster name
> - a possibility to implement the Application TimelineCollector as an
> interceptor in the AMRMProxyService.
> For Timeline Service:
> - need to store physical cluster id and logical cluster id so that we don't
> lose information at any level (flow/app/entity etc)
> - add a new table app id to cluster mapping table
> - need a different entity table/some table to store node level metrics for
> physical cluster stats. Once we get to node-level rollup, we probably have to
> store something in a dc, cluster, rack, node hierarchy. In that case a
> physical cluster makes sense, but we'd still need some way to tie physical
> and logical together in order to make automatic error detection etc that
> we're envisioning feasible within a federated setup.
> For the Cluster Naming convention:
> - three situations for cluster name:
> ----> app submitted to router should take federated (aka logical) cluster name
> ----> app submitted directly to RM should take physical cluster name
> ----> Info about the physical cluster in entities?
> - suggestion to set the cluster name as yarn tag at the router level (in the
> app submission context)
> Other points to note:
> - for federation to work smoothly in environments that use HDFS some
> additional considerations are needed, and possibly some solution like what is
> being used at Twitter with the nFly approach.
> Email thread context:
> {code}
> ---------- Forwarded message ----------
> From: Joep Rottinghuis
> Date: Fri, Jul 8, 2016 at 1:22 PM
> Subject: Re: Federation -Timeline Service meeting notes
> To: Subramaniam Venkatraman Krishnan
> Cc: Sangjin Lee, Vrushali Channapattan , Carlo Curino
> Thanks for the notes.
> I think that for federation to work smoothly in environments that use HDFS
> some additional considerations are needed, and possibly some solution like
> what we're using at Twitter with our nFly approach.
> bq. - need a different entity table/some table to store node level metrics
> for physical cluster stats
> Once we get to node-level rollup, we probably have to store something in a
> dc, cluster, rack, node hierarchy. In that case a physical cluster makes
> sense, but we'd still need some way to tie physical and logical together in
> order to make automatic error detection etc that we're envisioning feasible
> within a federated setup.
> Cheers,
> Joep
> On Fri, Jul 8, 2016 at 1:00 PM, Subramaniam Venkatraman Krishnan wrote:
> Thanks Vrushali for crisply capturing the essential from our rambling
> discussion J.
>
> Sangjin, I just want to add one comment to yours – we want to retain the
> physical cluster name (possibly as a new entity type) so that we don’t lose
> information & we can cluster level rollups even if they are not efficient.
>
> Additionally, based on the walkthrough of Federation design:
> · There was general agreement with the proposed approach.
> · There is a possibility to implement the Application
> TimelineCollector as an interceptor in the AMRMProxyService.
> · Joep raised the concern that it would be better if the RMs
> obtain the epoch from FederationStateStore. This is not currently in the
> roadmap of our MVP but we definitely plan to address this in future.
>
> Regards,
> Subru
>
> From: Sangjin Lee
> Sent: Thursday, July 07, 2016 6:22 PM
> To: Vrushali Channapattan
> Cc: Joep Rottinghuis; Carlo Curino; Subramaniam Venkatraman Krishnan
> Subject: Re: Federation -Timeline Service meeting notes
>
> Thanks for the summary Vrushali!
>
> Just so that we're on the same page regarding the terminology, I
> understand we're using the terms "logical cluster" and "federated cluster"
> interchangeably.
>
> Also, between using the federated cluster name and the home cluster name
> as a solution, I think we were leaning towards the federated cluster name
> (although not concluded).
>
> On Thu, Jul 7, 2016 at 4:33 PM, Vrushali Channapattan wrote:
>
> For Federation:
> - all entities that belong to the same flow run should have the same
> cluster name
> - app id in the same flow run strongly ordered in time
> - need a logical cluster name and physical cluster name
> For Timeline Service:
> - need to store physical cluster id and logical cluster id so that we
> don't lose information at any level (flow/app/entity etc)
> - add a new table app id to cluster mapping table
> - need a different entity table/some table to store node level
> metrics for physical cluster stats
> For the Cluster Naming convention:
> - three situations for cluster name:
> ----> app submitted to router should take federated cluster name
> ----> app submitted directly to RM should take physical cluster name
> ----> Info about the physical cluster in entities?
> - suggestion to set the cluster name as yarn tag at the router level
> (in the app submission context)
> {code}
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
---------------------------------------------------------------------
To unsubscribe, e-mail: yarn-issues-unsubscr...@hadoop.apache.org
For additional commands, e-mail: yarn-issues-h...@hadoop.apache.org
