[
https://issues.apache.org/jira/browse/HADOOP-15419?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16456005#comment-16456005
]
Tao Jie commented on HADOOP-15419:
----------------------------------
[~hexiaoqiao] thank you for your comment.
{quote}
In fact, some request path may be found after task running (for one simple
instance: hard code path in MapReduce/Spark job), if we don't obtain delegation
token for that NameNode, the Job will be fail due to all tasks can not pass
authentication.
{quote}
Agree. For the cluster, it is not easy to know which namenodes a certain job
would access. I think the mechanism could be more flexible and obtaining tokens
from all namenodes seems to be too crude.
1, We can have a option maybe {{fs.viewfs.use.specific.filesystem}}, only when
this option is true, the following logic works.
2, When submit a mr/spark job, if the input/output path is a viewfs path,
instead of obtaining token from all namenode, we would visit and fetch token
from only a SET of filesystem.
3, The raw filesystem of the input/output path should be in the SET
4, We may have a global option like {{fs.viewfs.global.filesystem}} which
defines filesystems that all jobs may visit(Eg. the filesystem of tmp dir,
scratch dir), and it should be added into the SET
5, Job-level option like {{fs.viewfs.additional.filesystem}} which defiles
extra filesystem that the certain job need.
Since obtaining delegation tokens happens on the client side, the effect of the
modification would be controllable.
Any thought?
> Should not obtain delegationTokens from all namenodes when using ViewFS
> -----------------------------------------------------------------------
>
> Key: HADOOP-15419
> URL: https://issues.apache.org/jira/browse/HADOOP-15419
> Project: Hadoop Common
> Issue Type: Bug
> Affects Versions: 3.0.2
> Reporter: Tao Jie
> Priority: Major
>
> Today when submit a job to a viewfs cluster, the client will try to obtain
> delegation token from all namenodes under the viewfs while only one namespace
> is actually used in this job. It would create many unnecessary rpc call to
> the whole cluster.
> In viewfs situation, we can just obtain delegation token from specific
> namenode rather than all namenodes.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
