Pradeep Kamath updated PIG-636:

    Status: Patch Available  (was: Open)

Attached patch with implementations for SingleTupleBag and NonSpillableDataBag. 
SingleTupleBag is a simplistic implementation of the DataBag interface which 
has just a single Tuple member representing the contents of the bag. The 
iterator() simply returns this member Tuple object. NonSpillableDataBag is more 
generic and can hold any number of tuples. It essentially has all functionality 
of a DefaultDataBag without the ability to spill to disk. Both these classes 
are to be used by directly calling their constructors without going through the 
BagFactory and are meant only for use in internal operators to transfer data in 

POPreCombinerLocalRearrange has been changed to use SingleTupleBag and 
POCombinerPackage has been changed to use NonSpillableDataBag. The 
CombinerOptimizer has been changed to include a new visitor which visits the 
POProjects in the map plan with result type BAG. Such Projects are annotated 
with a flag to use SingleTupleBags. This is necessary since only changing 
POPrecombinerLocalRearrange to use SingleTupleBag will still result in the 
problem reported in the description for algebraic aggregates which work on 
projects of bags like the following:
a = load...
b = group a by all;
c = foreach b generate SUM(a.$0), GRUNT(a.$1);

POProject has been changed to check the above flag and use a SingleTupleBag in 

> PERFORMANCE: Use lightweight bag implementations which do not register with 
> SpillableMemoryManager with Combiner
> ----------------------------------------------------------------------------------------------------------------
>                 Key: PIG-636
>                 URL: https://issues.apache.org/jira/browse/PIG-636
>             Project: Pig
>          Issue Type: Improvement
>    Affects Versions: types_branch
>            Reporter: Pradeep Kamath
>            Assignee: Pradeep Kamath
>             Fix For: types_branch
> Currently whenever Combiner is used in pig, in the map, the 
> POPrecombinerLocalRearrange operator puts the single "value" tuple 
> corresponding to a key into a DataBag and passes this to the foreach which is 
> being combined. This will generate as many bags as there are input records. 
> These bags all will have a single tuple and hence are small and should not 
> need to be spilt to disk. However since the bags are created through the 
> BagFactory mechanism, each bag creation is registered with the 
> SpillableMemoryManager and a weak reference to the bag is stored in a linked 
> list. This linked list grows really big over time causing unnecessary Garbage 
> collection runs. This can be avoided by having a simple lightweight 
> implementation of the DataBag interface to store the single tuple in a bag. 
> Also these SingleTupleBags should be created without registering with the 
> spillableMemoryManager. Likewise the bags created in POCombinePackage are 
> supposed to fit in Memory and not spill. Again a NonSpillableDataBag 
> implementation of DataBag interface which does not register with the 
> SpillableMemoryManager would help.

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

Reply via email to