[jira] Updated: (PIG-1516) finalize in bag implementations causes pig to run out of memory in reduce

["Thejas M Nair (JIRA)" <ji...@apache.org>]

     [ https://issues.apache.org/jira/browse/PIG-1516?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]

Thejas M Nair updated PIG-1516:
-------------------------------

    Attachment: PIG-1516.patch

I haven't removed the use of finalize in this patch,  but with the patch the number of objects with finalize() method that get created should be much smaller, and if all the bags used in a query are small enough, there will not be any such objects created. 
Even in case of large bags that spill to disk, since finalize() is not a method of the bags, the tuples in the bags can be freed by GC without waiting on finalization.
This should stop queries from running out of memory because of of the wait on finalize().

The creation of spill files happens only if the bag is very large, and the processing of the tuples in those bags is likely to give enough time for the finalization thread to catch up. 

The changes - 
1. As I proposed in the solution, the finalize has been removed from DefaultAbstractBag and its subclasses, and a FileList class with a finalize is used as container for the list of spill files.
2. Removed the finalize() method in InternalCachedBag.CachedBagIterator . It was used to call close on DataInputStream. The DataInputStream contains a FileInputStream which would have non-memory resources to be freed. But the FileInputStream already has a finalize() method, so the finalize() method in InternalCachedBag.CachedBagIterator is unnecessary.
3. In the bags that have code to pre-merge files when there are large number of spill files, the files that have been merged into larger files are deleted.


Using WeakReferences as Scott suggested, we can get rid of the finalization completely. I have created a separate jira for that - PIG-1519 .


> finalize in bag implementations causes pig to run out of memory in reduce 
> --------------------------------------------------------------------------
>
>                 Key: PIG-1516
>                 URL: https://issues.apache.org/jira/browse/PIG-1516
>             Project: Pig
>          Issue Type: Bug
>    Affects Versions: 0.7.0
>            Reporter: Thejas M Nair
>            Assignee: Thejas M Nair
>             Fix For: 0.8.0
>
>         Attachments: PIG-1516.patch
>
>
> *Problem:*
> pig bag implementations that are subclasses of DefaultAbstractBag, have finalize methods implemented. As a result, the garbage collector moves them to a finalization queue, and the memory used is freed only after the finalization happens on it.
> If the bags are not finalized fast enough, a lot of memory is consumed by the finalization queue, and pig runs out of memory. This can happen if large number of small bags are being created.
> *Solution:*
> The finalize function exists for the purpose of deleting the spill files that are created when the bag is too large. But if the bags are small enough, no spill files are created, and there is no use of the finalize function.
>  A new class that holds a list of files will be introduced (FileList). This class will have a finalize method that deletes the files. The bags will no longer have finalize methods, and the bags will use FileList instead of ArrayList<File>.
> *Possible workaround for earlier releases:*
> Since the fix is going into 0.8, here is a workaround -
> Disabling the combiner will reduce the number of bags getting created, as there will not be the stage of combining intermediate merge results. But I would recommend disabling it only if you have this problem as it is likely to slow down the query .
> To disable combiner, set the property: -Dpig.exec.nocombiner=true

-- 
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.