[
https://issues.apache.org/jira/browse/HAMA-940?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15266012#comment-15266012
]
Edward J. Yoon commented on HAMA-940:
-------------------------------------
As I mentioned in Description, we can simply check whether there's an newly
appended records to the input file, keeping last read offset.
To implement this, first of all, you should see the InputFormat interface
class. The tricky issue is how we implement the getSplits() method and multiple
tasks.
At the moment, my simple idea is that one bsp task acts as a "Stream input
queue" without implement StreamInputFormat and change the framework core. For
example, we set the file path in job configuration. The master task acts like
below:
{code}
if(isMaster(peer.me)) {
while(true) {
peer.reopen(); // reopen
peer.skip(offset); // jump to last offset
if(peer.readNext()) {
// at here we do load-balance.
sendTo("send a newly appended record to free slave tasks");
} else {
Thread.sleep();
}
}
}
{code}
> Add StreamInputFormat
> ---------------------
>
> Key: HAMA-940
> URL: https://issues.apache.org/jira/browse/HAMA-940
> Project: Hama
> Issue Type: New Feature
> Components: bsp core
> Reporter: Edward J. Yoon
>
> Add StreamInputFormat that reads newly appended records from previous
> superstep.
> I roughly guess it will be possible using reopen() method and file offset.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
