hadoop-common-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Michael Fuchs (JIRA)" <j...@apache.org>
Subject [jira] Commented: (HADOOP-5084) Reduce output data is not written to disk
Date Wed, 21 Jan 2009 14:13:59 GMT

    [ https://issues.apache.org/jira/browse/HADOOP-5084?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12665831#action_12665831

Michael Fuchs commented on HADOOP-5084:


I managed to track down the problem. It seems to be a combination of two things which cause
this data loss:

    * I use a string like "hdfs://" as "hadoop.hdfs.defaultfs" and I do
a fs.setWorkingDirectory("hdfs://");
    * The Path object I'm using for the output (TextOutputFormat.setOutputPath) is not created
by myself, it is created by using a Path instance retured by the fs.listStatus[0].getPath().
      In detail, I'm doing a
      Path somePath = fs.listStatus(someOtherPath)[n].getPath()
      Path output = new Path(somePath, "subfolder");
      TextOutputFormat.setOutputPath(jobConf, output);
I created a small application which shows issue.
The application takes three arguments:
1) pass "1" tu use an implementation which fails or "0" to use an implementation which works
2) the "hadoop.hdfs.defaultfs" value
3) the setWorkingDirectory value.

I execute the class like " bin/hadoop jar /tmp/example.jar org.test.TestMR 1 hdfs://BOCK:9000/test/
hdfs://BOCK:9000/test/" and it does not write the output but doees not fail!

package org.test;
import java.io.IOException;
import java.util.Iterator;

import org.apache.hadoop.fs.FSDataOutputStream;
import org.apache.hadoop.fs.FileStatus;
import org.apache.hadoop.fs.FileSystem;
import org.apache.hadoop.fs.Path;
import org.apache.hadoop.io.LongWritable;
import org.apache.hadoop.io.Text;
import org.apache.hadoop.mapred.FileInputFormat;
import org.apache.hadoop.mapred.JobClient;
import org.apache.hadoop.mapred.JobConf;
import org.apache.hadoop.mapred.Mapper;
import org.apache.hadoop.mapred.OutputCollector;
import org.apache.hadoop.mapred.Reducer;
import org.apache.hadoop.mapred.Reporter;
import org.apache.hadoop.mapred.RunningJob;
import org.apache.hadoop.mapred.TextOutputFormat;
import org.apache.log4j.Logger;

public class TestMR implements Mapper<LongWritable, Text, Text, Text>, Reducer<Text,
Text, Text, Text> {
	static Logger logger = Logger.getLogger(TestMR.class);
	public void map(LongWritable key, Text value, OutputCollector<Text, Text> out, Reporter
reporter) throws IOException {
		logger.info("map: key=" + key + ", value=" + value);
		out.collect(new Text(key.toString()), value);

	public void configure(JobConf jobConf) {

	public void close() throws IOException {
	public void reduce(Text key, Iterator<Text> values, OutputCollector<Text, Text>
out, Reporter reporter) throws IOException {
		while(values.hasNext()) {
			Text value = values.next();
			logger.info("reduce: key=" + key + ", value=" + value);
			out.collect(new Text(key), new Text(value));

	public static void main(String[] args) throws IOException {
		JobConf jobConf = new JobConf();
		if(args.length < 3) {
			logger.error("usage: 0|1 hadoop.hdfs.defaultfs hadoop.hdfs.workdir");
		logger.info("setting defaultfs and workdir");
		jobConf.set("hadoop.hdfs.defaultfs", args[1]);

		FileSystem fs = FileSystem.get(jobConf);
		fs.setWorkingDirectory(new Path(args[2]));


		logger.info("in: we use the 'in' directory. We create it and put some dummy data in it.");
		fs.delete(new Path(fs.getWorkingDirectory(), "in"), true);
		fs.mkdirs(new Path(fs.getWorkingDirectory(), "in"));
		FSDataOutputStream dataOut = fs.create(new Path(fs.getWorkingDirectory(), "in/data"));
		dataOut.write("some data".getBytes());
		FileInputFormat.setInputPaths(jobConf, new Path("in"));

		logger.info("out: we use a subdirectory of the 'out' directory as the output of your job.
Therefor we create the out folder first.");
		fs.delete(new Path(fs.getWorkingDirectory(), "out"), true);
		fs.mkdirs(new Path(fs.getWorkingDirectory(), "out"));
		Path out = null;
		if(args.length > 0 && "1".equals(args[0])) {
			logger.info("using broken folder");
			out = getBrokenOutFolder();
		} else {
			logger.info("using working folder");
			out = getWorkingOutFolder(fs);
		logger.info("setting output using path: " + out.toString());
		TextOutputFormat.setOutputPath(jobConf, new Path(out, "subfolder"));

		JobClient jobClient = new JobClient(jobConf);
		RunningJob runningJob = jobClient.submitJob(jobConf);
		while(!runningJob.isComplete()) {
			logger.info("still running...");
			try {
			} catch (InterruptedException e) {

	private static Path getWorkingOutFolder(FileSystem fs) throws IOException {
		Path root = new Path("/test");
		for(FileStatus fileStatus : fs.listStatus(root)) {
			if("out".equals(fileStatus.getPath().getName())) { return fileStatus.getPath(); }
		return null;

	private static Path getBrokenOutFolder() throws IOException {
		return new Path("out");
If I start the application like "bin/hadoop jar /tmp/example.jar org.test.TestMR 1 hdfs://BOCK:9000/test/
/test/" it works.

> Reduce output data is not written to disk
> -----------------------------------------
>                 Key: HADOOP-5084
>                 URL: https://issues.apache.org/jira/browse/HADOOP-5084
>             Project: Hadoop Core
>          Issue Type: Bug
>    Affects Versions: 0.18.2
>         Environment: Linux version 2.6.22-12-generic (buildd@vernadsky) (gcc version
4.1.3 20070831 (prerelease) (Ubuntu 4.1.2-16ubuntu1)) #1 SMP Sun Sep 23 18:11:30 GMT 2007
running Hadoop 18.2 on two nodes
>            Reporter: Michael Fuchs
>            Priority: Critical
> I run into a critical issue with Hadoop 18.2 on my Linux boxes:
> The jobs executes without any complains and they are listed in the
> succeeded list but there is no output data beside the "_logs" directory.
> The same code works with .17.2.1
> Here are some sections of the logs:
> [logfile]
> hadoop@bock:~/logs$ tail hadoop-hadoop-jobtracker-bock.log
> 2008-12-23 13:30:56,707 INFO org.apache.hadoop.mapred.JobInProgress:
> Choosing a data-local task task_200812231229_0031_m_000001 for
> speculation
> 2008-12-23 13:30:56,707 INFO org.apache.hadoop.mapred.JobTracker: Adding
> task 'attempt_200812231229_0031_m_000001_1' to tip
> task_200812231229_0031_m_000001, for tracker
> 'tracker_bock:localhost/'
> 2008-12-23 13:31:01,065 INFO org.apache.hadoop.mapred.JobInProgress:
> Task 'attempt_200812231229_0031_m_000001_1' has completed
> task_200812231229_0031_m_000001 successfully.
> 2008-12-23 13:31:03,177 INFO org.apache.hadoop.mapred.TaskRunner: Saved
> output of task 'attempt_200812231229_0031_r_000000_0' to
> hdfs://BOCK:9000/ana/oiprocessed/2008/12/23/Sen1/92a74190-2038-4c79-82c4-2de6fdc615db
> [/logfile]
> But the folder contains only a "_logs" folder which has a history file
> which contains:
> [logfile]
> Job JOBID="job_200812231415_0001" FINISH_TIME="1230038377844"
> FAILED_MAPS="0" FAILED_REDUCES="0" COUNTERS="Job Counters .Data-local
> map tasks:2,Job Counters .Launched reduce tasks:1,Job Counters .Launched
> map tasks:3,Map-Reduce Framework.Reduce input records:61,Map-Reduce
> Framework.Map output records:61,Map-Reduce Framework.Map output
> bytes:7194,Map-Reduce Framework.Combine output records:0,Map-Reduce
> Framework.Map input records:61,Map-Reduce Framework.Reduce input
> groups:12,Map-Reduce Framework.Combine input records:0,Map-Reduce
> Framework.Map input bytes:36396,Map-Reduce Framework.Reduce output
> records:12,File Systems.HDFS bytes written:1533,File Systems.Local bytes
> written:14858,File Systems.HDFS bytes read:38679,File Systems.Local
> bytes
> read:7388,com..ana.scheduling.HadoopTask$Counter.MAPPEED:61
> "
> [/logfile]
> So what I see is that the system runs successful and it even says it
> writes data! ("Map-Reduce Framework.Reduce output records:12,File Systems.HDFS bytes
> If I run the same code with .17.2.1 or in local mode with .18.2 it works
> and I get a part-0000 file with the expected data.
> Please tell me if you need additional information.

This message is automatically generated by JIRA.
You can reply to this email to add a comment to the issue online.

View raw message