[ 
https://issues.apache.org/jira/browse/HADOOP-8632?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13425186#comment-13425186
 ] 

Todd Lipcon commented on HADOOP-8632:
-------------------------------------

Hi Costin. We added the {{CACHE_CLASSES}} member in HADOOP-6133 to solve a 
performance regression which we saw in practice. HADOOP-6502 also talks about 
another case where the performance of this code was critical. So I don't think 
removing the cache is a good idea. 

That said, I do think changing the cache to have weak values as well as weak 
keys makes sense. Guava's MapMaker has a nice utility to do this, or we could 
simply change the class values to be Map<String, WeakReference<Class<?>>>.
                
> Configuration leaking class-loaders
> -----------------------------------
>
>                 Key: HADOOP-8632
>                 URL: https://issues.apache.org/jira/browse/HADOOP-8632
>             Project: Hadoop Common
>          Issue Type: Bug
>          Components: conf
>    Affects Versions: 2.0.0-alpha
>            Reporter: Costin Leau
>
> The newly introduced CACHE_CLASSES leaks class loaders causing associated 
> classes to not be reclaimed.
> One solution is to remove the cache itself since each class loader 
> implementation caches the classes it loads automatically and preventing an 
> exception from being raised is just a micro-optimization that, as one can 
> tell, causes bugs instead of improving anything.
> In fact, I would argue in a highly-concurrent environment, the weakhashmap 
> synchronization/lookup probably costs more then creating the exception itself.
> Another is to prevent the leak from occurring, by inserting the loadedclass 
> into the WeakHashMap wrapped in a WeakReference. Otherwise the class has a 
> strong reference to its classloader (the key) meaning neither gets GC'ed.
> And since the cache_class is static, even if the originating Configuration 
> instance gets GC'ed, its classloader won't.

--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators: 
https://issues.apache.org/jira/secure/ContactAdministrators!default.jspa
For more information on JIRA, see: http://www.atlassian.com/software/jira

        

Reply via email to