Hi,friend:

    When a MRJob print too much stdout or stderr log, the disk will be filled. Now it has influence our platform management.

    I have improved org.apache.hadoop.mapred.MapReduceChildJVM.java(come from@org.apache.hadoop.mapred.TaskLog) to generate the execute cmd
as follows: 

exec /bin/bash -c "( $JAVA_HOME/bin/java -Djava.net.preferIPv4Stack=true -Dhadoop.metrics.log.level=WARN -Xmx1024m -Djava.io.tmpdir=$PWD/tmp -     Dlog4j.configuration=container-log4j.properties -Dyarn.app.container.log.dir=/logs/userlogs/application_1403930653208_0003/container_1403930653208_0003_01_000002 -Dyarn.app.container.log.filesize=10240 -Dhadoop.root.logger=DEBUG,CLA org.apache.hadoop.mapred.YarnChild $test_IP 53911 attempt_1403930653208_0003_m_000000_0 2 | tail -c 102 >/logs/userlogs/application_1403930653208_0003/container_1403930653208_0003_01_000002/stdout ; exit $PIPESTATUS ) 2>&1 | tail -c 10240 >/logs/userlogs/application_1403930653208_0003/container_1403930653208_0003_01_000002/stderr ; exit $PIPESTATUS "


    But it doesn't take effect.

    And then, when I use "export YARN_NODEMANAGER_OPTS=-Xdebug -Xrunjdwp:transport=dt_socket,address=8788,server=y,suspend=y" for debuging NodeManager, I find when I set the BreakPoints at org.apache.hadoop.util.Shell(line 450:process = builder.start()) and org.apache.hadoop.yarn.server.nodemanager.containermanager.launcher.ContainerLaunch(line 161:List<String> newCmds = new ArrayList<String>(command.size())) the cmd will work.

    I doubt there's concurrency problem caused pipe shell will not perform properly. It matters, and I need help. 

   @https://issues.apache.org/jira/browse/YARN-2231

thanks


Zhanfeng Huo