[
https://issues.apache.org/jira/browse/HADOOP-153?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12610897#action_12610897
]
sharadag edited comment on HADOOP-153 at 7/6/08 11:55 PM:
----------------------------------------------------------------
Had an offline discussion with Eric and Devaraj, and we came up with following:
- Let this issue handle the case of crashes and hangups. For the case of
catching the exception for Java tasks, filed another Jira -> HADOOP-3700
- Design gets impacted based on the assumption of how frequent the failures
would be. At this point of time, design for INFREQUENT failures. This would
simplify the design. Also, bad records can be maintained by the Jobtracker (as
pointed out by Enis), as no of bad records are expected to be quite low.
- Failing Fast the bad jobs is very crucial to avoid wasting Grid resources.
Thresholds should be define in such a way that we identify bad jobs early
enough, say maximum of 10% of the maps can fail. Also, we need to make sure
that we execute failed task VERY FAST.
- Apart from bad data, Task crashes could be due to bad user code (like Out of
memory) or bad nodes. To isolate these cases, on failure, reexecute on another
node as now. If it fails AGAIN, then reexecute a third time, this time in the
special mode where we report every record completion to the Task tracker.
- For the case of Streaming, streaming would have to write the processed record
count to the stderr as a framework counter, to take advantage of this feature.
was (Author: sharadag):
Had an offline discussion with Eric and Devaraj, and we came up with
following:
- Let this issue handle the case of crashes and hangups. For the case of
catching the exception for Java tasks, filed another Jira -> Hadoop-3700
- Design gets impacted based on the assumption of how frequent the failures
would be. At this point of time, design for INFREQUENT failures. This would
simplify the design. Also, bad records can be maintained by the Jobtracker (as
pointed out by Enis), as no of bad records are expected to be quite low.
- Failing Fast the bad jobs is very crucial to avoid wasting Grid resources.
Thresholds should be define in such a way that we identify bad jobs early
enough, say maximum of 10% of the maps can fail. Also, we need to make sure
that we execute failed task VERY FAST.
- Apart from bad data, Task crashes could be due to bad user code (like Out of
memory) or bad nodes. To isolate these cases, on failure, reexecute on another
node as now. If it fails AGAIN, then reexecute a third time, this time in the
special mode where we report every record completion to the Task tracker.
- For the case of Streaming, streaming would have to write the processed record
count to the stderr as a framework counter, to take advantage of this feature.
> skip records that throw exceptions
> ----------------------------------
>
> Key: HADOOP-153
> URL: https://issues.apache.org/jira/browse/HADOOP-153
> Project: Hadoop Core
> Issue Type: New Feature
> Components: mapred
> Affects Versions: 0.2.0
> Reporter: Doug Cutting
> Assignee: Sharad Agarwal
> Attachments: skipRecords_wip1.patch
>
>
> MapReduce should skip records that throw exceptions.
> If the exception is thrown under RecordReader.next() then RecordReader
> implementations should automatically skip to the start of a subsequent record.
> Exceptions in map and reduce implementations can simply be logged, unless
> they happen under RecordWriter.write(). Cancelling partial output could be
> hard. So such output errors will still result in task failure.
> This behaviour should be optional, but enabled by default. A count of errors
> per task and job should be maintained and displayed in the web ui. Perhaps
> if some percentage of records (>50%?) result in exceptions then the task
> should fail. This would stop jobs early that are misconfigured or have buggy
> code.
> Thoughts?
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
