Pradeep Kamath commented on PIG-841:

One issue with implementing this might be that now all the sample records do 
not come in one bag - they come in multiple bags from POPackage - the 
FindQuantiles udf needs all the samples to compute the weighted range partition 
information. It may need to cache its input into a bag and then in finish(), do 
the computation - however then finish would need to write out the information 
to dfs. There would already be a store in the reduce plan with the output 
filename. If the udf writes out the output to dfs in finish(), the store would 
not be writing any output and this can be confusing. So this needs to be 
thought through a little more.

> PERFORMANCE: The sample MR job in order by implementation can use Hadoop 
> sorting instead of doing a POSort
> ----------------------------------------------------------------------------------------------------------
>                 Key: PIG-841
>                 URL: https://issues.apache.org/jira/browse/PIG-841
>             Project: Pig
>          Issue Type: Improvement
>    Affects Versions: 0.2.1
>            Reporter: Pradeep Kamath
>             Fix For: 0.3.0
> Currently the sample map reduce job in order by implementation does the 
> following:
>  - sample 100 records from each map
>  - group all on the above output
>  - sort the output bag from the above grouping on keys of the order by
>  - give the sorted bag to FindQuantiles udf
> The steps 2 and 3 above can be replaced by
> - group the sample output by the order by key and set parallelism of the 
> group to 1 so that output of the group goes to one reducer. Since Hadoop 
> ensures the output of the group is sorted by key we get sorting for free 
> without using POSort 

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