[
https://issues.apache.org/jira/browse/HIVE-16329?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15948019#comment-15948019
]
Gopal V commented on HIVE-16329:
--------------------------------
The issue is immediately resolved if a full GC is triggered via jmap -histo:live
> TopN: TopNHash totalFreeMemory computation ignores free memory
> --------------------------------------------------------------
>
> Key: HIVE-16329
> URL: https://issues.apache.org/jira/browse/HIVE-16329
> Project: Hive
> Issue Type: Bug
> Reporter: Gopal V
>
> {code}
> // TODO: For LLAP, assumption is off-heap cache.
> final long memoryUsedPerExecutor =
> (memoryMXBean.getHeapMemoryUsage().getUsed() / numExecutors);
> // this is total free memory available per executor in case of LLAP
> totalFreeMemory = conf.getMaxMemoryAvailable() - memoryUsedPerExecutor;
> {code}
> {code}
> exec.TopNHash: isTez parameters -615768144 = 5312782848 - 71142611912 / 12
> {code}
> This turns off the TopNHash entirely causing something trivial like
> {code}
> select c_custkey, count(1) from customer group by c_custkey limit 10;
> {code}
> To shuffle 30M rows instead of 10.
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
