Hadoop DistributedCache is deprecated - what is the preferred API?

Question

My map tasks need some configuration data, which I would like to distribute via the Distributed Cache.

The Hadoop MapReduce Tutorial shows the usage of the Distribut

Answer 1

I had the same problem. And not only is DistributedCach deprecated but getLocalCacheFiles and "new Job" too. So what worked for me is the following:

Driver:

Configuration conf = getConf();
Job job = Job.getInstance(conf);
...
job.addCacheFile(new Path(filename).toUri());

In Mapper/Reducer setup:

@Override
protected void setup(Context context) throws IOException, InterruptedException
{
    super.setup(context);

    URI[] files = context.getCacheFiles(); // getCacheFiles returns null

    Path file1path = new Path(files[0])
    ...
}

Answer 2

The APIs for the Distributed Cache can be found in the Job class itself. Check the documentation here: http://hadoop.apache.org/docs/stable2/api/org/apache/hadoop/mapreduce/Job.html The code should be something like

Job job = new Job();
...
job.addCacheFile(new Path(filename).toUri());

In your mapper code:

Path[] localPaths = context.getLocalCacheFiles();
...

Answer 3

None of the solution mentioned worked for me in completeness . It could because Hadoop version keeps changing I am using hadoop 2.6.4. Essentially, DistributedCache is deprecated so I didnt want to use that. As some of the post suggest us to use addCacheFile() however, it has changed a bit. Here is how it worked for me

job.addCacheFile(new URI("hdfs://X.X.X.X:9000/EnglishStop.txt#EnglishStop.txt"));

Here X.X.X.X can be Master IP address or localhost. The EnglishStop.txt was stored in HDFS at / location.

hadoop fs -ls /

The output is

-rw-r--r--   3 centos supergroup       1833 2016-03-12 20:24 /EnglishStop.txt
drwxr-xr-x   - centos supergroup          0 2016-03-12 19:46 /test

Funny but convenient, #EnglishStop.txt means now we can access it as "EnglishStop.txt" in mapper. Here is the code for the same

public void setup(Context context) throws IOException, InterruptedException     
{
    File stopwordFile = new File("EnglishStop.txt");
    FileInputStream fis = new FileInputStream(stopwordFile);
    BufferedReader reader = new BufferedReader(new InputStreamReader(fis));

    while ((stopWord = reader.readLine()) != null) {
        // stopWord is a word read from Cache
    }
}

This just worked for me. You can read line from the file stored in HDFS

Answer 4

I did not use job.addCacheFile(). Instead I used -files option like "-files /path/to/myfile.txt#myfile" as before. Then in the mapper or reducer code I use the method below:

/**
 * This method can be used with local execution or HDFS execution. 
 * 
 * @param context
 * @param symLink
 * @param throwExceptionIfNotFound
 * @return
 * @throws IOException
 */
public static File findDistributedFileBySymlink(JobContext context, String symLink, boolean throwExceptionIfNotFound) throws IOException
{
    URI[] uris = context.getCacheFiles();
    if(uris==null||uris.length==0)
    {
        if(throwExceptionIfNotFound)
            throw new RuntimeException("Unable to find file with symlink '"+symLink+"' in distributed cache");
        return null;
    }
    URI symlinkUri = null;
    for(URI uri: uris)
    {
        if(symLink.equals(uri.getFragment()))
        {
            symlinkUri = uri;
            break;
        }
    }   
    if(symlinkUri==null)
    {
        if(throwExceptionIfNotFound)
            throw new RuntimeException("Unable to find file with symlink '"+symLink+"' in distributed cache");
        return null;
    }
    //if we run this locally the file system URI scheme will be "file" otherwise it should be a symlink
    return "file".equalsIgnoreCase(FileSystem.get(context.getConfiguration()).getScheme())?(new File(symlinkUri.getPath())):new File(symLink);

}

Then in mapper/reducer:

@Override
protected void setup(Context context) throws IOException, InterruptedException
{
    super.setup(context);

    File file = HadoopUtils.findDistributedFileBySymlink(context,"myfile",true);
    ... do work ...
}

Note that if I used "-files /path/to/myfile.txt" directly then I need to use "myfile.txt" to access the file since that is the default symlink name.

Answer 5

To expand on @jtravaglini, the preferred way of using DistributedCache for YARN/MapReduce 2 is as follows:

In your driver, use the Job.addCacheFile()

public int run(String[] args) throws Exception {
    Configuration conf = getConf();

    Job job = Job.getInstance(conf, "MyJob");

    job.setMapperClass(MyMapper.class);

    // ...

    // Mind the # sign after the absolute file location.
    // You will be using the name after the # sign as your
    // file name in your Mapper/Reducer
    job.addCacheFile(new URI("/user/yourname/cache/some_file.json#some"));
    job.addCacheFile(new URI("/user/yourname/cache/other_file.json#other"));

    return job.waitForCompletion(true) ? 0 : 1;
}

And in your Mapper/Reducer, override the setup(Context context) method:

@Override
protected void setup(
        Mapper<LongWritable, Text, Text, Text>.Context context)
        throws IOException, InterruptedException {
    if (context.getCacheFiles() != null
            && context.getCacheFiles().length > 0) {

        File some_file = new File("./some");
        File other_file = new File("./other");

        // Do things to these two files, like read them
        // or parse as JSON or whatever.
    }
    super.setup(context);
}

Answer 6

The new DistributedCache API for YARN/MR2 is found in the org.apache.hadoop.mapreduce.Job class.

   Job.addCacheFile()

Unfortunately, there aren't as of yet many comprehensive tutorial-style examples of this.

http://hadoop.apache.org/docs/stable/api/org/apache/hadoop/mapreduce/Job.html#addCacheFile%28java.net.URI%29