[ 
https://issues.apache.org/jira/browse/MAPREDUCE-5221?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13670163#comment-13670163
 ] 

Tsuyoshi OZAWA commented on MAPREDUCE-5221:
-------------------------------------------

I've confirmed all tests are pass correctly. It's absolutely odd, because of 
the difference of o.a.h.mapred.Reducer and o.a.h.mapreduce.Reduce as you 
pointed out. Therefore, I've analysed the source code, and I've found some 
problems:

1. There are no codes to test with MiniMRYarnCluster with launching reduce-side 
combiner. TestMRAppWithCombiner seems to be enough, but reduce-side combiners 
are actually not launched because data size is too small to launch combiners.
2. ShuffleConsumerPlugin.Contexts only accepts Class<? extends 
o.a.h.mapred.Reducer> when constructing. It's needed to add new 
ShuffleConsumerPlugin.Contexts which accepts  Class<? extends 
o.a.h.mapreduce.Reducer>. # Or, change huffleConsumerPlugin.Contexts to accepts 
Class<?> and check the type with Java reflection API when constructing.
3. Shuffle#MergeManager only accepts Class<? extends o.a.h.mapred.Reducer>. 
It's needed to add new MergeManager which accepts  Class<? extends 
o.a.h.mapreduce.Reducer>.

I'll create a subtasks to deal with these problems step by step.
                
> Reduce side Combiner is not used when using the new API
> -------------------------------------------------------
>
>                 Key: MAPREDUCE-5221
>                 URL: https://issues.apache.org/jira/browse/MAPREDUCE-5221
>             Project: Hadoop Map/Reduce
>          Issue Type: Bug
>    Affects Versions: 2.0.4-alpha
>            Reporter: Siddharth Seth
>            Assignee: Tsuyoshi OZAWA
>         Attachments: MAPREDUCE-5221.1.patch, MAPREDUCE-5221.2.patch, 
> MAPREDUCE-5221.3.patch
>
>
> If a combiner is specified using o.a.h.mapreduce.Job.setCombinerClass - this 
> will silently ignored on the reduce side since the reduce side usage is only 
> aware of the old api combiner.
> This doesn't fail the job - since the new combiner key does not deprecate the 
> old key.

--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators
For more information on JIRA, see: http://www.atlassian.com/software/jira

Reply via email to