[
https://issues.apache.org/jira/browse/HDFS-14035?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16670465#comment-16670465
]
Chen Liang commented on HDFS-14035:
-----------------------------------
Had some discussion offline. Turns out one issue with this approach is that it
will associate HDFS DelegationToken with HAServiceProtocol ALL the time. Say,
in the future if we have another use case using HAServiceProtocol but has
nothing to do with HDFS (HAServiceProtocol is meant to be generic after all),
it will be tricky to prevent returning an out-of-nowhere HDFS delegation token
selector in that case. Actually, this rises the high level question of whether
HAServiceProtocol should be using HDFS delegation token at all.
In the long term, a better approach might be to introduce a new protocol such
as HAStatusProtocol which just separates getServiceStatus functionality out of
HAServiceProtocol and start from there. But that is not in the scope of this
feature branch and should probably be discussed separately with more folks from
trunk.
Additionally, in our case, getServiceStatus returns a HAServiceStatus instance,
but HAServiceStatus has more than what is actually needed, what we really need
is just HAServiceStatus#HAServiceState.
So for now, I propose to just add the functionality of server state discover to
ClientNamenodeProtocol and replace using HAServiceProtocol here. Specifically,
introduce a new RPC call such as getServerHAState in ClientNamenodeProtocol,
which just returns something like a HAServiceState instance. If we are to
separate getServiceStatus out of HAServiceProtocol, we can revisit and refactor
in future. Will repurpose this JIRA.
cc [~shv], [~xkrogen], [~csun]
> NN status discovery does not leverage delegation token
> ------------------------------------------------------
>
> Key: HDFS-14035
> URL: https://issues.apache.org/jira/browse/HDFS-14035
> Project: Hadoop HDFS
> Issue Type: Sub-task
> Reporter: Chen Liang
> Assignee: Chen Liang
> Priority: Major
> Attachments: HDFS-14035-HDFS-12943.001.patch,
> HDFS-14035-HDFS-12943.002.patch
>
>
> Currently ObserverReadProxyProvider uses
> {{HAServiceProtocol#getServiceStatus}} to get the status of each NN. However
> {{HAServiceProtocol}} does not leverage delegation token. So when running an
> application on YARN and when YARN node manager makes this call
> getServiceStatus, token authentication will fail, causing the application to
> fail.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
