[
https://issues.apache.org/jira/browse/HIVE-12568?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15036886#comment-15036886
]
Xuefu Zhang commented on HIVE-12568:
------------------------------------
Yeah. I'm not really sure of that or whether it's a common practice. However,
current implementation is to pick (randomly) any possible interface to bind,
which can lead to problem. On the other hand, making such a change does require
that HS2 can be reached by the NM nodes. Maybe we should give an option to
overwrite this? Any thoughts to share, [~vanzin]?
> Use the same logic finding HS2 host name in Spark client [Spark Branch]
> -----------------------------------------------------------------------
>
> Key: HIVE-12568
> URL: https://issues.apache.org/jira/browse/HIVE-12568
> Project: Hive
> Issue Type: Bug
> Components: Spark
> Affects Versions: 1.1.0
> Reporter: Xuefu Zhang
> Assignee: Xuefu Zhang
> Attachments: HIVE-12568.0-spark.patch, HIVE-12568.1-spark.patch
>
>
> Spark client sends a pair of host name and port number to the remote driver
> so that the driver can connects back to HS2 where the user session is. Spark
> client has its own way determining the host name, and pick one network
> interface if the host happens to have multiple network interfaces. This can
> be problematic. For that, there is parameter,
> hive.spark.client.server.address, which user can pick an interface.
> Unfortunately, this interface isn't exposed.
> Instead of exposing this parameter, we can use the same logic as Hive in
> determining the host name. Therefore, the remote driver connecting to HS2
> using the same network interface as a HS2 client would do.
> There might be a case where user may want the remote driver to use a
> different network. This is rare if at all. Thus, for now it should be
> sufficient to use the same network interface.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
