[
https://issues.apache.org/jira/browse/TEZ-2726?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14701505#comment-14701505
]
Saikat commented on TEZ-2726:
-----------------------------
[~rajesh.balamohan] [~bikassaha]
There are no empty partitions in the example I mentioned. The source vertex has
1 task (used a UnorderedKVOutput, so produced only 1 partition)and sink vertex
has 3 tasks. The edge is of type SCATTER-GATHER.
When http fetchers sent a request for fetching the map outputs, the code in
shufflehandler catches IOException in
IndexCache.java getIndexInformation() function for the condition
[info.mapSpillRecord.size() <= reduce].
2015-08-10 12:36:42,314 [New I/O worker #32] ERROR mapred.ShuffleHandler:
Shuffle error in populating headers :
java.io.IOException: Invalid request Map Id =
attempt_1437478617943_17839_1_05_000000_0_10003 Reducer = 1 Index Info Length =
1
at org.apache.hadoop.mapred.IndexCache.getIndexInformation(IndexCache.java:84)
at
org.apache.hadoop.mapred.ShuffleHandler$Shuffle.getMapOutputInfo(ShuffleHandler.java:855)
at
org.apache.hadoop.mapred.ShuffleHandler$Shuffle.populateHeaders(ShuffleHandler.java:875)
at
org.apache.hadoop.mapred.ShuffleHandler$Shuffle.messageReceived(ShuffleHandler.java:793)
I ll try to get an excerpt of the Fetcher logs for DMEs and post here.
> Handle invalid number of partitions for SCATTER-GATHER edge
> -----------------------------------------------------------
>
> Key: TEZ-2726
> URL: https://issues.apache.org/jira/browse/TEZ-2726
> Project: Apache Tez
> Issue Type: Improvement
> Reporter: Saikat
> Assignee: Saikat
>
> Encountered an issue where the source vertex has M task and sink vertex has N
> tasks (N > M), [e.g. M = 1, N = 3]and the edge is of type SCATTER -GATHER.
> This resulted in sink vertex receiving DMEs with non existent targetIds.
> The fetchers for the sink vertex tasks then try to retrieve the map outputs
> and retrieve invalid headers due to exception in the ShuffleHandler.
> Possible fixes:
> 1. raise proper Tez Exception to indicate this invalid scenario.
> 2. or write appropriate empty partition bits, for the missing partitions
> before sending out the DMEs to sink vertex.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
