Hadoop DistributedCache对象在作业过程中发生变化

Question

我试图在AWS上运行KMeans，当我试图从DistributedCache读取更新的集群质心时，我遇到了以下异常:

java.io.IOException: The distributed cache object s3://mybucket/centroids_6/part-r-00009 changed during the job from 4/8/13 2:20 PM to 4/8/13 2:20 PM
at org.apache.hadoop.filecache.TrackerDistributedCacheManager.downloadCacheObject(TrackerDistributedCacheManager.java:401)
at org.apache.hadoop.filecache.TrackerDistributedCacheManager.localizePublicCacheObject(TrackerDistributedCacheManager.java:475)
at org.apache.hadoop.filecache.TrackerDistributedCacheManager.getLocalCache(TrackerDistributedCacheManager.java:191)
at org.apache.hadoop.filecache.TaskDistributedCacheManager.setupCache(TaskDistributedCacheManager.java:182)
at org.apache.hadoop.mapred.TaskTracker$4.run(TaskTracker.java:1246)
at java.security.AccessController.doPrivileged(Native Method)
at javax.security.auth.Subject.doAs(Subject.java:396)
at org.apache.hadoop.security.UserGroupInformation.doAs(UserGroupInformation.java:1132)
at org.apache.hadoop.mapred.TaskTracker.initializeJob(TaskTracker.java:1237)
at org.apache.hadoop.mapred.TaskTracker.localizeJob(TaskTracker.java:1152)
at org.apache.hadoop.mapred.TaskTracker$5.run(TaskTracker.java:2541)
at java.lang.Thread.run(Thread.java:662)

这个问题与这个问题的区别在于这个错误是间歇性出现的。我在一个较小的数据集上成功地运行了相同的代码。此外，当我将质心的数量从12(如上面代码所示)更改为8时，它会在第5次迭代而不是6次迭代时失败(您可以在上面的centroids_6名称中看到)。

下面是运行KMeans循环的主驱动程序中的相关DistributedCache代码:

    int iteration = 1;
    long changes = 0; 
    do {
        // First, write the previous iteration's centroids to the dist cache.
        Configuration iterConf = new Configuration();
        Path prevIter = new Path(centroidsPath.getParent(),
                String.format("centroids_%s", iteration - 1));
        FileSystem fs = prevIter.getFileSystem(iterConf);
        Path pathPattern = new Path(prevIter, "part-*");
        FileStatus [] list = fs.globStatus(pathPattern);
        for (FileStatus status : list) {
            DistributedCache.addCacheFile(status.getPath().toUri(), iterConf);
        }
        // Now, set up the job.
        Job iterJob = new Job(iterConf);
        iterJob.setJobName("KMeans " + iteration);
        iterJob.setJarByClass(KMeansDriver.class);
        Path nextIter = new Path(centroidsPath.getParent(), 
                String.format("centroids_%s", iteration));
        KMeansDriver.delete(iterConf, nextIter);
        // Set input/output formats.
        iterJob.setInputFormatClass(SequenceFileInputFormat.class);
        iterJob.setOutputFormatClass(SequenceFileOutputFormat.class);
        // Set Mapper, Reducer, Combiner
        iterJob.setMapperClass(KMeansMapper.class);
        iterJob.setCombinerClass(KMeansCombiner.class);
        iterJob.setReducerClass(KMeansReducer.class);
        // Set MR formats.
        iterJob.setMapOutputKeyClass(IntWritable.class);
        iterJob.setMapOutputValueClass(VectorWritable.class);
        iterJob.setOutputKeyClass(IntWritable.class);
        iterJob.setOutputValueClass(VectorWritable.class);
        // Set input/output paths.
        FileInputFormat.addInputPath(iterJob, data);
        FileOutputFormat.setOutputPath(iterJob, nextIter);
        iterJob.setNumReduceTasks(nReducers);
        if (!iterJob.waitForCompletion(true)) {
            System.err.println("ERROR: Iteration " + iteration + " failed!");
            System.exit(1);
        }
        iteration++;
        changes = iterJob.getCounters().findCounter(KMeansDriver.Counter.CONVERGED).getValue();
        iterJob.getCounters().findCounter(KMeansDriver.Counter.CONVERGED).setValue(0);
    } while (changes > 0);

还可以如何修改文件?我能想到的唯一可能性是，在一次迭代完成时，在前一次作业的质心完成写入之前，循环再次开始。但是在注释中，我使用waitForCompletion(true)调用作业，因此当循环重新开始时，作业的任何残余部分都不应该运行。什么好主意吗?

Answer 1

这不是一个真正的答案，但我确实意识到以我的方式使用DistributedCache是愚蠢的，而不是直接从HDFS读取前一次迭代的结果。我在主驱动程序中编写了这个方法:

public static HashMap<Integer, VectorWritable> readCentroids(Configuration conf, Path path)
        throws IOException {
    HashMap<Integer, VectorWritable> centroids = new HashMap<Integer, VectorWritable>();
    FileSystem fs = FileSystem.get(path.toUri(), conf);
    FileStatus [] list = fs.globStatus(new Path(path, "part-*"));
    for (FileStatus status : list) {
        SequenceFile.Reader reader = new SequenceFile.Reader(fs, status.getPath(), conf);
        IntWritable key = null;
        VectorWritable value = null;
        try {
            key = (IntWritable)reader.getKeyClass().newInstance();
            value = (VectorWritable)reader.getValueClass().newInstance();
        } catch (InstantiationException e) {
            e.printStackTrace();
        } catch (IllegalAccessException e) {
            e.printStackTrace();
        }
        while (reader.next(key, value)) {
            centroids.put(new Integer(key.get()),
                    new VectorWritable(value.get(), value.getClusterId(), value.getNumInstances()));
        }
        reader.close();
    }
    return centroids;
}

在每次迭代时Mapper和Reducer的setup()方法中调用，读取前一次迭代的质心。

protected void setup(Context context) throws IOException {
    Configuration conf = context.getConfiguration();
    Path centroidsPath = new Path(conf.get(KMeansDriver.CENTROIDS));
    centroids = KMeansDriver.readCentroids(conf, centroidsPath);
}

这允许我在我的原始问题中删除循环中的代码块，该问题将质心写入DistributedCache。我对它进行了测试，现在它可以在大型和小型数据集上工作。

我仍然不知道为什么我得到的错误我张贴(如何在只读分布式缓存的东西被改变?特别是当我在每次迭代中更改HDFS路径时?)，但这似乎既有效，又不是一种读取质心的hack方式。