[
https://issues.apache.org/jira/browse/HIVE-15767?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16082948#comment-16082948
]
Peter Cseh commented on HIVE-15767:
-----------------------------------
This happens with HiveCLI, not with HS2.
The exception is coming from the spark driver.
When the HiveCLI is executed from shell, the mapreduce.job.credentials.binary
is empty in the configuration as spark-submit is called from the RemoteClient.
When it's executed from Oozie's LauncherMapper, Hive picks up this property
from the Oozie launcher's configuration which is correct, but passes it to
Spark. Spark runs in yarn-cluster mode so the Spark driver gets it's own
container (which may be on an other machine). It look for the credential files
in the folder where the Oozie Launcher ran. That's on a different machine, so
it can't pick up the conatiner_tokens file which leaves the spark driver with
no tokens so it fails.
I don't know how Hive-on-MR works in this regards, but we had no similar issues
with the HiveAction before, so I assume it works differently.
I don't think it's possible to reproduce it using MiniClusters as the local
folders will be available in the test so the Spark driver will be able to
access it.
> Hive On Spark is not working on secure clusters from Oozie
> ----------------------------------------------------------
>
> Key: HIVE-15767
> URL: https://issues.apache.org/jira/browse/HIVE-15767
> Project: Hive
> Issue Type: Bug
> Components: Spark
> Affects Versions: 1.2.1, 2.1.1
> Reporter: Peter Cseh
> Assignee: Peter Cseh
> Attachments: HIVE-15767-001.patch, HIVE-15767-002.patch
>
>
> When a HiveAction is launched form Oozie with Hive On Spark enabled, we're
> getting errors:
> {noformat}
> Caused by: java.io.IOException: Exception reading
> file:/yarn/nm/usercache/yshi/appcache/application_1485271416004_0022/container_1485271416004_0022_01_000002/container_tokens
> at
> org.apache.hadoop.security.Credentials.readTokenStorageFile(Credentials.java:188)
> at
> org.apache.hadoop.mapreduce.security.TokenCache.mergeBinaryTokens(TokenCache.java:155)
> {noformat}
> This is caused by passing the {{mapreduce.job.credentials.binary}} property
> to the Spark configuration in RemoteHiveSparkClient.
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
