hadoop-mapreduce-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Harsh J <ha...@cloudera.com>
Subject Re: determining what files made up a failing task
Date Sun, 04 Dec 2011 15:37:06 GMT

Perhaps you can simply set a percentage of failure toleration for your job.

Doable via http://hadoop.apache.org/common/docs/r0.20.2/api/org/apache/hadoop/mapred/JobConf.html#setMaxMapTaskFailuresPercent(int)
and http://hadoop.apache.org/common/docs/r0.20.2/api/org/apache/hadoop/mapred/JobConf.html#setMaxReduceTaskFailuresPercent(int)

If you set it to 10%, your job still passes if 10% of total Map or Reduce tasks failed. I
think this fits your use-case.

On 04-Dec-2011, at 4:05 AM, Mat Kelcey wrote:

> Hi folks,
> I have a Hadoop 0.20.2 map only job with thousands of inputs tasks;
> I'm using the org.apache.nutch.tools.arc.ArcInputFormat input format
> so each task corresponds to a single file in HDFS
> Most of the way into the job it hits a task that causes the input
> format to OOM. After 4 attempts it fails the job.
> Now this is obviously not great but for the purpose of my job I'd be
> happy to just throw this input file away, it's only one of thousands
> and I don't need exact results.
> The trouble is I can't work out where what file this task corresponds to?
> The closest I can find is that the job history file lists a STATE_STRING
> ( eg STATE_STRING="hdfs://ip-10-115-29-44\.ec2\.internal:9000/user/hadoop/arc_files\.aa/2009/09/17/0/1253240925734_0\.arc\.gz:0+100425468"
> )
> but this is _only_ for the successfully completed ones, for the failed
> one I'm actually interested in there is nothing
> MapAttempt TASK_TYPE="MAP" TASKID="task_201112030459_0011_m_004130"
> TASK_ATTEMPT_ID="attempt_201112030459_0011_m_004130_0"
> HOSTNAME="ip-10-218-57-227\.ec2\.internal" ERROR="Error: null" .
> I grepped through all the hadoop logs and couldn't find anything that
> relates this task to the files in it's split
> Any ideas where this info might be recorded?
> Cheers,
> Mat

View raw message