[
https://issues.apache.org/jira/browse/PIG-802?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12707064#action_12707064
]
Pradeep Kamath commented on PIG-802:
------------------------------------
PIG-744 is a duplicate - will be marking that one as duplicate.
Pasting the summary from PIG-744 which has a little more detail:
Currently order by results in multiple map reduce jobs (2 or 3 depending on the
script) of which the last one does the actual ordering. In this last map reduce
job, we create a bag of values (each value being the entire tuple that is
getting sorted) for each sort key(s) using POPackage in the reduce phase. Then
we turn around and flatten the bag in the foreach following the package. So
there is really no need for the bag. But to be generic and use the existing
operators, we can be more efficient by tagging the POPackage to create bags
which are backed by the Hadoop iterator itself. This way we do not create a bag
by making a copy of each tuple from the hadoop iterator. This should help both
performance and scalability by making better use of memory.
> PERFORMANCE: not creating bags for ORDER BY
> -------------------------------------------
>
> Key: PIG-802
> URL: https://issues.apache.org/jira/browse/PIG-802
> Project: Pig
> Issue Type: Improvement
> Reporter: Olga Natkovich
>
> Order by should be changed to not use POPackage to put all of the tuples in a
> bag on the reduce side, as the bag is just immediately flattened. It can
> instead work like join does for the last input in the join.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
