[
https://issues.apache.org/jira/browse/HIVE-18916?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16511191#comment-16511191
]
Sahil Takiar commented on HIVE-18916:
-------------------------------------
[~aihuaxu] can you take a look. Here is a brief summary of the changes:
* {{SparkClientImpl}} has been modified so that if the thread that is
monitoring the {{bin/spark-submit}} process detects that {{bin/spark-submit}}
fails, it parses the stdout / stderr of {{bin/spark-submit}} and checks for any
log lines that contain "Error" and then includes those lines in the exception
that gets thrown
** {{SparkClientImpl}} was actually already doing this, but the information
wasn't getting propagated all the way to the end user
* A few changes to {{RpcServer}} were necessary to make sure the exception
thrown by the "Driver" thread gets propagated to the user
* A few other minor changes to classes like {{RemoteSparkJobMonitor}},
{{SparkTask}} and the constructor of {{SparkClientImpl}} to prevent double
logging of exceptions
* Added a few unit tests for this, which required masking a few additional
patterns in .q files
The motivation for this is that {{bin/spark-submit}} errors out when certain
parameters are misconfigured. This allows HoS to propagate these error messages
to the end-user, which should improve debuggability. The added .q files are a
good example of this.
> SparkClientImpl doesn't error out if spark-submit fails
> -------------------------------------------------------
>
> Key: HIVE-18916
> URL: https://issues.apache.org/jira/browse/HIVE-18916
> Project: Hive
> Issue Type: Sub-task
> Components: Spark
> Reporter: Sahil Takiar
> Assignee: Sahil Takiar
> Priority: Major
> Attachments: HIVE-18916.1.WIP.patch, HIVE-18916.2.patch,
> HIVE-18916.3.patch
>
>
> If {{spark-submit}} returns a non-zero exit code, {{SparkClientImpl}} will
> simply log the exit code, but won't throw an error. Eventually, the
> connection timeout will get triggered and an exception like {{Timed out
> waiting for client connection}} will be logged, which is pretty misleading.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
