[ 
https://issues.apache.org/jira/browse/YARN-4083?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15070131#comment-15070131
 ] 

Arun Suresh commented on YARN-4083:
-----------------------------------

Hoping to get some consensus on this.. since it is required for YARN-2877 as 
well.

I feel having the ContainerExecutor expose a *YARN_SCHEDULER_ADDRESS* 
environment variable (and as [~jianhe] mentioned, maybe let it be a list, with 
the first entry being the local NM and the remaining a list of RM addresses to 
allow for failover) should work across across Java and non-java applications.

It would also be somewhat dynamic, as [~steve_l] mentioned, since the value is 
decided by the NM right before it launches a container, but unlike a global/ZK 
based registry, it can be different for different containers / applications 
(although it would not change during the lifetime of the container).

bq. how do AM IP filters know when to bounce an HTTP Request over to the proxy
My understanding (atleast our requirement for YARN-2877) is that this would be 
used by the AM specifically for resolving the address for the server end of the 
ApplicationMasterProtocol, so HTTP addresses can be specified probably via 
another env variable maybe ?

bq. How does this work when the container is actually a Linux container and not 
a fake yarn-level container ?
[~aw], apologize if I did not fully understand, but I feel an environment 
variable should be accessible by linux, windows and other containers.

Thoughts ?


> Add a discovery mechanism for the scheduler address
> ---------------------------------------------------
>
>                 Key: YARN-4083
>                 URL: https://issues.apache.org/jira/browse/YARN-4083
>             Project: Hadoop YARN
>          Issue Type: Improvement
>          Components: nodemanager, resourcemanager
>            Reporter: Subru Krishnan
>            Assignee: Subru Krishnan
>
> Today many apps like Distributed Shell, REEF, etc rely on the fact that the 
> HADOOP_CONF_DIR of the NM is on the classpath to discover the scheduler 
> address. This JIRA proposes the addition of an explicit discovery mechanism 
> for the scheduler address



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)

Reply via email to