[jira] Commented: (PIG-274) changing combiner behavior past hadoop 18

["Olga Natkovich (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/PIG-274?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12613673#action_12613673 ] 

Olga Natkovich commented on PIG-274:
------------------------------------

Some additional thoughts.

Hash based approach would work well for group by keys with very low cardinality since the data given to each map is usually pretty small. For other cases hashing would actually introduce a significant overhead. Since we don't currently have metadata, we will not be able to make an intelligent choice. Moreover, if the combiner is called after the aggregation, it is completely wasted since the data is already reduced.

Alan suggested that we could opportunistically reduce data if we see adjacent keys that are the same but I am wondering whether the overhead of compare is justified on this case.

Comments are welocome.

> changing combiner behavior past hadoop 18
> -----------------------------------------
>
>                 Key: PIG-274
>                 URL: https://issues.apache.org/jira/browse/PIG-274
>             Project: Pig
>          Issue Type: Bug
>            Reporter: Olga Natkovich
>
> In hadoop 18, the way commbiners are handled is changing. The hadoop team agreed to keep things backward compatible for now but will depricate the current behavior in the future (likely in hadoop 19) so pig needs to adjust to the new behavior. This should be done in the post 2.0 code base.
> Old behavior: combiner is called once and only once per map task
> New behavior: combiner can be run 0 or more times on both map and reduce sides. 0 times happens if only a single <K, V> fits into sort buffer. Multiple time can happen in case of a hierarchical merge.
> The main issue that causes problem for pig is that we would not know in advance whether the combiner will run 0,1 or more times. This causes several issues:
> (1) Lets assume that we compute count. If we enable combiner, reducer expects to get numbers not values as its input. Hadoop team suggested that we could annotate each tuple with a byte that tells if it want through combiner. This could be expensive computatinally as well as will use extra memory. One things to notice is that some algebraics (like SUM, MIN, MAX) don't care whether the data was precombined as they always to the same thing. Perhaps we can make algebaic functions declare if they care or not. Then we only anotate the ones that need it.
> (2) Since combiner can be called 1 or more times, getInitial and getIntermediate have to do the same thing. So again, we need to change the interface to reflcat that.
> (3) current combiner code assumes that it only works with 1 input. When it runs on the reduce side, it can be dealing with tuples from multiple inputs. 

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