[
https://issues.apache.org/jira/browse/YARN-7630?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16288315#comment-16288315
]
Botong Huang edited comment on YARN-7630 at 12/12/17 9:44 PM:
--------------------------------------------------------------
I've validated the fix in our 1k node test cluster, where we found the issue
initially. The RM key rolling interval is set at once every hour. After the fix
is applied last week, there's no related issue since then. So I am pretty
confident about it.
was (Author: botong):
I've validated the fix in our 1k node test cluster, where we found the issue
initially. After the fix is applied last week, there's no related issue since
then. So I am pretty confident about it.
> Fix AMRMToken handling in AMRMProxy
> -----------------------------------
>
> Key: YARN-7630
> URL: https://issues.apache.org/jira/browse/YARN-7630
> Project: Hadoop YARN
> Issue Type: Sub-task
> Reporter: Botong Huang
> Assignee: Botong Huang
> Priority: Minor
> Attachments: YARN-7630.v1.patch, YARN-7630.v1.patch
>
>
> Symptom: after RM rolls over the master key for AMRMToken, whenever the RPC
> connection from FederationInterceptor to RM breaks due to transient network
> issue and reconnects, heartbeat to RM starts failing because of the “Invalid
> AMRMToken” exception. Whenever it hits, it happens for both home RM and
> secondary RMs.
> Related facts:
> 1. When RM issues a new AMRMToken, it always send with service name field as
> empty string. RPC layer in AM side will set it properly before start using
> it.
> 2. UGI keeps all tokens using a map from serviceName->Token. Initially
> AMRMClientUtils.createRMProxy() is used to load the first token and start the
> RM connection.
> 3. When RM renew the token, YarnServerSecurityUtils.updateAMRMToken() is used
> to load it into UGI and replace the existing token (with the same serviceName
> key).
> Bug:
> The bug is that 2-AMRMClientUtils.createRMProxy() and
> 3-YarnServerSecurityUtils.updateAMRMToken() is not handling the sequence
> consistently. We always need to load the token (with empty service name) into
> UGI first before we set the serviceName, so that the previous AMRMToken will
> be overridden. But 2 is doing it reversely. That’s why after RM rolls the
> amrmToken, the UGI end up with two tokens. Whenever the RPC connection break
> and reconnect, the wrong token could be picked and thus trigger the
> exception.
> Fix:
> Should load the AMRMToken into UGI first and then update the service name
> field for RPC
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
