[
https://issues.apache.org/jira/browse/MAPREDUCE-1555?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12853314#action_12853314
]
Amareshwari Sriramadasu commented on MAPREDUCE-1555:
----------------------------------------------------
bq. so that MapTask object in JobTracker doesn't have a reference to the
input-split. Why?
This is to save memory on JobTracker. We have seen scenarios where the input
split size so huge which causes OOM on JT. See HADOOP-3670
If we have to remove the nullification in serialization to avoid NPE, we should
nullify the splitMetaInfo either after the task completion or after the first
update from the map task.
Thoughts?
> Maybe some code of MapTask is wrong.
> ------------------------------------
>
> Key: MAPREDUCE-1555
> URL: https://issues.apache.org/jira/browse/MAPREDUCE-1555
> Project: Hadoop Map/Reduce
> Issue Type: Bug
> Components: task
> Reporter: Ruyue Ma
>
> {code:title=MapTask.java|borderStyle=solid}
> public void write(DataOutput out) throws IOException {
> super.write(out);
> if (isMapOrReduce()) {
> splitMetaInfo.write(out);
> splitMetaInfo = null; // HERE: why set null ??????
> }
> }
> {code}
> In above code, if the splitMetaInfo is set null, Second Serialization (will
> invoke write) will throw NullPointerException.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
