[
https://issues.apache.org/jira/browse/HIVE-17532?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16166272#comment-16166272
]
Peter Vary commented on HIVE-17532:
-----------------------------------
[~pcsaszar]: We recently discussed the same topic in HIVE-17291
There the decision was, that on a production cluster one should use dynamic
allocation, and for the other's it is better to use the actual cores, than the
configured one.
Thanks,
Peter
> Hive on Spark query compilation starts Spark session
> ----------------------------------------------------
>
> Key: HIVE-17532
> URL: https://issues.apache.org/jira/browse/HIVE-17532
> Project: Hive
> Issue Type: Bug
> Components: HiveServer2
> Affects Versions: 2.2.0
> Reporter: Peter Csaszar
> Priority: Minor
>
> Hive on Spark query compilation starts a new Spark session when some kind of
> aggregation is present:
> 0: jdbc:hive2://localhost:10000/default> set hive.execution.engine=spark;
> No rows affected (0.013 seconds)
> 0: jdbc:hive2://localhost:10000/default> explain select distinct label0 from
> iris;
> INFO : Compiling
> command(queryId=hive_20170912151212_914ee322-28dd-442a-9dd9-7ed00a6a8caf):
> explain select distinct label0 from iris
> INFO : Semantic Analysis Completed
> INFO : Returning Hive schema: Schema(fieldSchemas:[FieldSchema(name:Explain,
> type:string, comment:null)], properties:null)
> INFO : Completed compiling
> command(queryId=hive_20170912151212_914ee322-28dd-442a-9dd9-7ed00a6a8caf);
> Time taken: *40.594* seconds
> Spark job started, all consecutive explain statements are fast:
> 0: jdbc:hive2://localhost:10000/default> explain select distinct a1 from iris;
> INFO : Compiling
> command(queryId=hive_20170912151414_faacda24-290e-48bb-9daf-3f301fc170c1):
> explain select distinct label0 from iris
> INFO : Semantic Analysis Completed
> INFO : Returning Hive schema: Schema(fieldSchemas:[FieldSchema(name:Explain,
> type:string, comment:null)], properties:null)
> INFO : Completed compiling
> command(queryId=hive_20170912151414_faacda24-290e-48bb-9daf-3f301fc170c1);
> Time taken: *0.275* seconds
> Killing the Spark job, the same query is still fast, and no new Spark job has
> been started:
> 0: jdbc:hive2://localhost:10000/default> explain select distinct a2 from iris;
> INFO : Compiling
> command(queryId=hive_20170912151616_a7ea83b6-03ce-4636-b3d4-be6feadcde35):
> explain select distinct label0 from iris
> INFO : Semantic Analysis Completed
> INFO : Returning Hive schema: Schema(fieldSchemas:[FieldSchema(name:Explain,
> type:string, comment:null)], properties:null)
> INFO : Completed compiling
> command(queryId=hive_20170912151616_a7ea83b6-03ce-4636-b3d4-be6feadcde35);
> Time taken: *0.213* seconds
> The code in question:
> SetSparkReducerParallelism.java:
> sparkSessionManager = SparkSessionManagerImpl.getInstance();
> sparkSession = SparkUtilities.getSparkSession(context.getConf(),
> sparkSessionManager);
> sparkMemoryAndCores = sparkSession.getMemoryAndCores();
> The created Spark session is used for getting the number of cores and memory
> only. This could be determined from the configurations, without actually
> starting a session.
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
