[
https://issues.apache.org/jira/browse/MAPREDUCE-841?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12740853#action_12740853
]
Hong Tang commented on MAPREDUCE-841:
-------------------------------------
For input split objects, we could write the serialized buffer of user input
split objects to a separate file instead of within the RawSplit object. Namely,
change RawSplit from:
{code}
static class RawSplit implements Writable {
private String splitClass;
private BytesWritable bytes = new BytesWritable();
private String[] locations;
long dataLength;
...
{code}
To:
{code}
static class RawSplit implements Writable {
private String splitClass;
private BytesWritable bytes = null;
private long offset; // pointing to the offset to the serialized bytes for
the user input split
private long length; // the length of the serialized bytes for the user
input split
private String[] locations;
long dataLength;
...
{code}
Where the field "bytes" will be loaded from the external file before we send
the object to the TT, and we shall reset the reference to null after that.
> Protect Job Tracker against memory exhaustion due to very large InputSplit or
> JobConf objects
> ---------------------------------------------------------------------------------------------
>
> Key: MAPREDUCE-841
> URL: https://issues.apache.org/jira/browse/MAPREDUCE-841
> Project: Hadoop Map/Reduce
> Issue Type: Improvement
> Components: jobtracker
> Affects Versions: 0.20.1
> Reporter: Hong Tang
> Fix For: 0.21.0
>
>
> JobTracker only needs to examine a subset of information contained by
> InputSplit or JobConf objects. But currently JobTracker loads the complete
> user-defined InputSplit and JobConf objects in memory. This design would
> leave JobTracker susceptible to memory exhaustion particularly in cases when
> some bugs in user code which could result in very large input splits or job
> conf objects (e.g. PIG-901).
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.