[
https://issues.apache.org/jira/browse/HIVE-20853?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16677589#comment-16677589
]
Sergey Shelukhin commented on HIVE-20853:
-----------------------------------------
To fetch something, you first need to write something... currently that always
happens in the same DAG I think. So in order to fetch smth from a node current
DAG has to have written something there before or after restart. Before should
be solved by this patch as far as I can tell, and after is solved because the
task that wrote stuff already added credentials.
[~gopalv] correct me if I'm wrong about shuffle and fetch :)
> Expose ShuffleHandler.registerDag in the llap daemon API
> --------------------------------------------------------
>
> Key: HIVE-20853
> URL: https://issues.apache.org/jira/browse/HIVE-20853
> Project: Hive
> Issue Type: Improvement
> Components: llap
> Affects Versions: 3.1.0
> Reporter: Jaume M
> Assignee: Jaume M
> Priority: Critical
> Attachments: HIVE-20853.1.patch, HIVE-20853.2.patch,
> HIVE-20853.3.patch, HIVE-20853.4.patch
>
>
> Currently DAGs are only registered when a submitWork is called for that DAG.
> At this point the crendentials are added to the ShuffleHandler and it can
> start serving.
> However Tez might (and will) schedule tasks to fetch from the ShuffleHandler
> before anything of this happens and all this tasks will fail which may
> results in the query failing.
> This happens in the scenario in which a LlapDaemon just comes up and tez
> fetchers try to open a connection before a DAG has been registered.
> Adding this API will allow to register the DAG against the Daemon when the AM
> notices that a new Daemon is up.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
