[ 
https://issues.apache.org/jira/browse/PHOENIX-2405?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15003357#comment-15003357
 ] 

Andrew Purtell commented on PHOENIX-2405:
-----------------------------------------

bq. This would basically mean that we'd be writing files when the sort spills 
to disk instead of using memory mapped files.

I think you're at the mercy of the JVM and OS when using mapped files, and this 
has already proven to be fatal route for some users. "Manually" spilling 
instead if in-effect letting the OS do that by paging in and out the mmapped 
file is a fine solution. 

> Use Java heap memory instead of memory mapped files during ORDER BY
> -------------------------------------------------------------------
>
>                 Key: PHOENIX-2405
>                 URL: https://issues.apache.org/jira/browse/PHOENIX-2405
>             Project: Phoenix
>          Issue Type: Bug
>            Reporter: James Taylor
>
> We currently use memory mapped files to buffer data as it's being sorted in 
> an ORDER BY (see MappedByteBufferQueue). The following types of exceptions 
> have been seen to occur:
> {code}
> Caused by: java.lang.OutOfMemoryError: Map failed
>         at sun.nio.ch.FileChannelImpl.map0(Native Method)
>         at sun.nio.ch.FileChannelImpl.map(FileChannelImpl.java:904)
> {code}
> [~apurtell] has read that memory mapped files are not cleaned up after very 
> well in Java:
> {quote}
> "Map failed" means the JVM ran out of virtual address space. If you search 
> around stack overflow for suggestions on what to do when your app (in this 
> case Phoenix) encounters this issue when using mapped buffers, the answers 
> tend toward manually cleaning up the mapped buffers or explicitly triggering 
> a full GC. See 
> http://stackoverflow.com/questions/8553158/prevent-outofmemory-when-using-java-nio-mappedbytebuffer
>  for example. There are apparently long standing JVM/JRE problems with 
> reclamation of mapped buffers. I think we may want to explore in Phoenix a 
> different way to achieve what the current code is doing.
> {quote}
> Instead of using memory mapped files, we could use heap memory, or perhaps 
> there are other mechanisms too.



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)

Reply via email to