[
https://issues.apache.org/jira/browse/HIVE-9007?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Szehon Ho updated HIVE-9007:
----------------------------
Attachment: HIVE-9007-spark.patch
IdentityProjectRemover removes unnecessary Select/Fil operators.
I tried some hacks on Spark compiler's mapjoin optimizers to work with this
kind of graph (where multiple RS are connected to each other without
intermediate Sel/Fil). But it's too hard as this affects where they cut the
ReduceWork, and loses the rowSchema from for ReduceSink -> HTSink conversion.
So disabling for spark compilation.
Wrote a new test that stresses this case. Interestingly, it does not seem to
work for MapReduce either.
> Hive may generate wrong plan for map join queries due to
> IdentityProjectRemover [Spark Branch]
> ----------------------------------------------------------------------------------------------
>
> Key: HIVE-9007
> URL: https://issues.apache.org/jira/browse/HIVE-9007
> Project: Hive
> Issue Type: Sub-task
> Components: Spark
> Affects Versions: spark-branch
> Reporter: Chao
> Assignee: Szehon Ho
> Attachments: HIVE-9007-spark.patch
>
>
> HIVE-8435 introduces a new logical optimizer called IdentityProjectRemover,
> which may cause map join in spark branch to generate wrong plan.
> Currently, the map join conversion in spark branch first goes through a
> method {{convertJoinMapJoin}}, which replaces a join op with a mapjoin op,
> removes RS associated with big table, and keep RSs for all small tables.
> Afterwards, in {{SparkReduceSinkMapJoinProc}} it replaces all parent RSs of
> the mapjoin op with HTS (note it doesn't check whether the RS belongs to
> small table or big table.)
> The issue arises, when IdentityProjectRemover comes into play, which may
> result into a situation that a operator tree has two consecutive RSs. Imaging
> the following example:
> {noformat}
> Join MapJoin
> / \ / \
> RS RS ---> RS RS
> / \ / \
> TS RS TS TS (big table)
> \ (small table)
> TS
> {noformat}
> In this case, all parents of the mapjoin op will be RS, even the branch for
> big table! In {{SparkReduceSinkMapJoinProc}}, they will be replaced with HTS,
> which is obviously incorrect.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
