[
https://issues.apache.org/jira/browse/HIVE-13308?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15280475#comment-15280475
]
Marcelo Vanzin commented on HIVE-13308:
---------------------------------------
There needs to be a synchronization block to access the {{server}} field;
otherwise, it's probably safe to call the {{SparkClientImpl}} constructor
outside of a synchronized block, as long as {{SparkClientFactory.stop()}} is
not called before it finishes. Something like:
{code}
public static SparkClient createClient(Map<String, String> sparkConf,
HiveConf hiveConf)
throws IOException, SparkException {
RpcServer _server;
synchronized (SparkClientFactory.class) {
Preconditions.checkState(server != null, "initialize() not called.");
_server = server;
}
return new SparkClientImpl(_server, sparkConf, hiveConf);
}
{code}
Or maybe just making the {{server}} variable volatile would suffice, too (and
then no synchronization is needed in {{createClient}}).
> HiveOnSpark sumbit query very slow when hundred of beeline exectue at same
> time
> -------------------------------------------------------------------------------
>
> Key: HIVE-13308
> URL: https://issues.apache.org/jira/browse/HIVE-13308
> Project: Hive
> Issue Type: Bug
> Components: Spark
> Affects Versions: 1.2.1
> Reporter: wangwenli
> Original Estimate: 1h
> Remaining Estimate: 1h
>
> backgroud: hive on spark , yarn cluster mode
> details:
> when using hundred of beeline submit query at the same time, we found that
> yarn get application very slow, and hiveserver is blocked at
> SparkClientFactory.createClient method
> after analysis, we think the synchronize on SparkClientFactory.createClient ,
> can be removed
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
