[
https://issues.apache.org/jira/browse/HIVE-19439?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16466772#comment-16466772
]
Rui Li commented on HIVE-19439:
-------------------------------
Hi [~vihangk1], the task is retried by Spark, and it calls
SparkMapRecordHandler::init to initialize the map operator. This is where we
retrieve the MapWork [from
cache|https://github.com/apache/hive/blob/rel/release-2.2.0/ql/src/java/org/apache/hadoop/hive/ql/exec/spark/SparkMapRecordHandler.java#L75].
I'm not sure whether we have a way to reset the operators to UNINIT state. If
not, guess we have to clear the cache when initialization fails.
> MapWork shouldn't be reused when Spark task fails during initialization
> -----------------------------------------------------------------------
>
> Key: HIVE-19439
> URL: https://issues.apache.org/jira/browse/HIVE-19439
> Project: Hive
> Issue Type: Bug
> Components: Spark
> Reporter: Rui Li
> Priority: Major
>
> Issue identified in HIVE-19388. When a Spark task fails during initializing
> the map operator, the task is retried with the same MapWork retrieved from
> cache. This can be problematic because the MapWork may be partially
> initialized, e.g. some operators are already in INIT state.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
