[
https://issues.apache.org/jira/browse/HIVE-8920?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Xuefu Zhang updated HIVE-8920:
------------------------------
Attachment: HIVE-8920.3-spark.patch
We have to disable caching for MapInput because of IOContext initialization
problem. This would have performance impact, but only for multi-insert cases.
Regardless, correctness goes above performance.
Patch #3 includes a fix as such.
> IOContext problem with multiple MapWorks cloned for multi-insert [Spark
> Branch]
> -------------------------------------------------------------------------------
>
> Key: HIVE-8920
> URL: https://issues.apache.org/jira/browse/HIVE-8920
> Project: Hive
> Issue Type: Sub-task
> Components: Spark
> Affects Versions: spark-branch
> Reporter: Chao
> Assignee: Xuefu Zhang
> Attachments: HIVE-8920.1-spark.patch, HIVE-8920.2-spark.patch,
> HIVE-8920.3-spark.patch
>
>
> The following query will not work:
> {code}
> from (select * from table0 union all select * from table1) s
> insert overwrite table table3 select s.x, count(1) group by s.x
> insert overwrite table table4 select s.y, count(1) group by s.y;
> {code}
> Currently, the plan for this query, before SplitSparkWorkResolver, looks like
> below:
> {noformat}
> M1 M2
> \ / \
> U3 R5
> |
> R4
> {noformat}
> In {{SplitSparkWorkResolver#splitBaseWork}}, it assumes that the
> {{childWork}} is a ReduceWork, but for this case, you can see that for M2 the
> childWork could be UnionWork U3. Thus, the code will fail.
> HIVE-9041 addressed partially addressed the problem by removing union task.
> However, it's still necessary to cloning M1 and M2 to support multi-insert.
> Because M1 and M2 can run in a single JVM, the original solution of storing a
> global IOContext will not work because M1 and M2 have different io contexts,
> both needing to be stored.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
