[
https://issues.apache.org/jira/browse/PHOENIX-5120?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16759230#comment-16759230
]
Lars Hofhansl edited comment on PHOENIX-5120 at 2/2/19 10:57 PM:
-----------------------------------------------------------------
GROUP BY spilling also uses MappedByteBuffers. So there are three ways (at
least) of spilling in Phoenix: (1) Client side spooling iterators, (2)
OrderedResultIterator, and (3) the GroupByCache.
The GROUP BY spilling is very different and writes/reads pages, so access
patterns there are not clear to me. I won't touch that part here.
was (Author: lhofhansl):
GROUP BY spilling also uses MappedByteBuffers. So there are three ways (at
least) of spilling in Phoenix: (1) Client side spooling iterators, (2)
OrderedResultIterator, and (3) the GroupByCache.
> Avoid using MappedByteBuffers for server side sorting.
> ------------------------------------------------------
>
> Key: PHOENIX-5120
> URL: https://issues.apache.org/jira/browse/PHOENIX-5120
> Project: Phoenix
> Issue Type: Task
> Reporter: Lars Hofhansl
> Assignee: Lars Hofhansl
> Priority: Major
> Attachments: 5120-1.4-wip.txt, 5120-1.4.txt
>
>
> We should had a production outage due to this.
> MappedByteBuffer may leave files around, on top they use direct memory, which
> is not cleared until the JVM executes a full GC.
> See last comment on PHOENIX-2405.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
