[ 
https://issues.apache.org/jira/browse/YARN-10229?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17143294#comment-17143294
 ] 

Young Chen commented on YARN-10229:
-----------------------------------

Two thoughts:
 * We're checking both the amrmProxyEnabled flag and the federationEnabled flag 
before forwarding the requests. Is just checking amrmProxyEnabled sufficient 
since that's a federation component? Not sure if there is some edge case usage 
for amrm proxy without federation.
 * Can we move the bulk of this logic into AMRMProxyService? E.g. the state 
store check, etc. Since the federation facade is already loaded inside 
amrmproxy service, it may be better to reduce ContainerManagerImpl dependency 
on Federation implementation and keep this logic inside Federation components. 
This way we also don't have to expose getFederationStateStoreFacade from 
AMRMProxyService.

> [Federation] Client should be able to submit application to RM directly using 
> normal client conf
> ------------------------------------------------------------------------------------------------
>
>                 Key: YARN-10229
>                 URL: https://issues.apache.org/jira/browse/YARN-10229
>             Project: Hadoop YARN
>          Issue Type: Bug
>          Components: amrmproxy, federation
>    Affects Versions: 3.1.1
>            Reporter: JohnsonGuo
>            Assignee: Bilwa S T
>            Priority: Major
>         Attachments: YARN-10229.001.patch, YARN-10229.002.patch, 
> YARN-10229.003.patch, YARN-10229.004.patch
>
>
> Scenario: When enable the yarn federation feature with multi yarn clusters, 
> one can submit their job to yarn-router by *modified* their client 
> configuration with yarn router address.
> But if one still wants to submit their jobs via the original client (before 
> enable federation) to RM directly, it will encounter the AMRMToken exception. 
>  That means once enable federation ,if some one want to submit job, they have 
> to  modify the client conf.
>  
> one possible solution for this Scenario is:
> In NodeManger, when the client ApplicationMaster request comes:
>  * get the client job.xml  from HDFS "".
>  * parse the "yarn.resourcemanager.scheduler.address" parameter in job.xml
>  * if the value of the parameter is "localhost:8049"(AMRM address),then do 
> the AMRMToken valid process
>  * if the value of the parameter is "rm:port"(rm address),then skip the 
> AMRMToken valid process
>  
>  
>  



--
This message was sent by Atlassian Jira
(v8.3.4#803005)

---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]

Reply via email to