Hadoop MapReduce链式实践--ChainReducer

版本号：CDH5.0.0，HDFS：2.3.0，Mapreduce:2.3.0,Yarn:2.3.0。

场景描写叙述：求一组数据中依照不同类别的最大值，比方，例如以下的数据：

data1:

A,10
A,11
A,12
A,13
B,21
B,31
B,41
B,51

data2:

A,20
A,21
A,22
A,23
B,201
B,301
B,401
B,501

最后输出为：

A,23
B,501

假如这种逻辑的mapreduce数据流例如以下：

假设C组数据比較多，同一时候假设集群有2个节点，那么这个任务分配2个reducer，且C组数据平均分布到两个reducer中，（这样做是为了效率考虑，假设仅仅有一个reducer，那么当一个节点在执行reducer的时候另外一个节点会处于空暇状态）那么假设在reducer之后，还能够再次做一个reducer，那么不就能够整合数据到一个文件了么，同一时候还能够再次比較C组数据中，以得到真正比較大的数据。

首先说下，不用上面假设的方式进行操作，那么一般的操作方法。一般有两种方法：其一，直接读出HDFS数据，然后进行整合；其二，新建另外一个Job来进行整合。这两种方法，假设就效率来说的话，可能第一种效率会高点。

考虑到前面提出的mapreduce数据流，曾经曾对ChainReducer有点印象，好像能够做这个，所以就拿ChainReducer来试，同一时候为了学多点知识，也是用了多个Mapper（即使用ChainMapper）。

主程序代码例如以下：

package chain;

import org.apache.hadoop.conf.Configuration;
import org.apache.hadoop.conf.Configured;
import org.apache.hadoop.fs.Path;
import org.apache.hadoop.io.IntWritable;
import org.apache.hadoop.io.LongWritable;
import org.apache.hadoop.io.Text;
import org.apache.hadoop.mapred.FileInputFormat;
import org.apache.hadoop.mapred.FileOutputFormat;
import org.apache.hadoop.mapred.JobClient;
import org.apache.hadoop.mapred.JobConf;
import org.apache.hadoop.mapred.TextInputFormat;
import org.apache.hadoop.mapred.TextOutputFormat;
import org.apache.hadoop.mapred.lib.ChainMapper;
import org.apache.hadoop.mapred.lib.ChainReducer;
import org.apache.hadoop.util.Tool;
import org.apache.hadoop.util.ToolRunner;

public class ChainDriver2 extends Configured implements Tool{

	/**
	 * ChainReducer 实战
	 * 验证多个reducer的整合
	 * 逻辑：寻找最大值
	 * @param args
	 */
	
	private String input=null;
	private String output=null;
	private String delimiter=null;
	private int reducer=1;
	public static void main(String[] args) throws Exception {
		ToolRunner.run(new Configuration(), new ChainDriver2(),args);
	}
	
	@Override
	public int run(String[] arg0) throws Exception {
		configureArgs(arg0);
		checkArgs();
		Configuration conf = getConf();
		conf.set("delimiter", delimiter);
		JobConf  job= new JobConf(conf,ChainDriver2.class);
		
		ChainMapper.addMapper(job, MaxMapper.class, LongWritable.class,
				Text.class, Text.class, IntWritable.class, true, new JobConf(false)) ;
		
		ChainMapper.addMapper(job, MergeMaxMapper.class, Text.class,
				IntWritable.class, Text.class, IntWritable.class, true, new JobConf(false));
		
		ChainReducer.setReducer(job, MaxReducer.class, Text.class, IntWritable.class,
				Text.class, IntWritable.class, true, new JobConf(false));
		ChainReducer.addMapper(job, MergeMaxMapper.class, Text.class,
				IntWritable.class, Text.class, IntWritable.class, false, new JobConf(false));
		job.setJarByClass(ChainDriver2.class);
		job.setJobName("ChainReducer test job");
		
        job.setMapOutputKeyClass(Text.class);
        job.setMapOutputValueClass(IntWritable.class);
        job.setOutputKeyClass(Text.class);
        job.setOutputValueClass(IntWritable.class);
        
       /* job.setMapperClass(MaxMapper.class);
        job.setReducerClass(MaxReducer.class);*/
        job.setInputFormat(TextInputFormat.class);;
        job.setOutputFormat(TextOutputFormat.class);
        job.setNumReduceTasks(reducer);
        
        FileInputFormat.addInputPath(job, new Path(input));
        FileOutputFormat.setOutputPath(job, new Path(output));
        
        JobClient.runJob(job);
		return 0;
	}
	
	
	/**
	 * check the args 
	 */
	private void checkArgs() {
		if(input==null||"".equals(input)){
			System.out.println("no input...");
			printUsage();
			System.exit(-1);
		}
		if(output==null||"".equals(output)){
			System.out.println("no output...");
			printUsage();
			System.exit(-1);
		}
		if(delimiter==null||"".equals(delimiter)){
			System.out.println("no delimiter...");
			printUsage();
			System.exit(-1);
		}
		if(reducer==0){
			System.out.println("no reducer...");
			printUsage();
			System.exit(-1);
		}
	}

	/**
	 * configuration the args
	 * @param args
	 */
	private void configureArgs(String[] args) {
    	for(int i=0;i<args.length;i++){
    		if("-i".equals(args[i])){
    			input=args[++i];
    		}
    		if("-o".equals(args[i])){
    			output=args[++i];
    		}
    		
    		if("-delimiter".equals(args[i])){
    			delimiter=args[++i];
    		}
    		if("-reducer".equals(args[i])){
    			try {
    				reducer=Integer.parseInt(args[++i]);
				} catch (Exception e) {
					reducer=0;
				}
    		}
    	}
	}
	public static void printUsage(){
    	System.err.println("Usage:");
    	System.err.println("-i input 	 cell data path.");
    	System.err.println("-o output 	 output data path.");
    	System.err.println("-delimiter  data delimiter , default is blanket  .");
    	System.err.println("-reducer  reducer number , default is 1  .");
    }
	
}

MaxMapper：

package chain;

import java.io.IOException;

import org.apache.hadoop.io.IntWritable;
import org.apache.hadoop.io.LongWritable;
import org.apache.hadoop.io.Text;
import org.apache.hadoop.mapred.JobConf;
import org.apache.hadoop.mapred.MapReduceBase;
import org.apache.hadoop.mapred.Mapper;
import org.apache.hadoop.mapred.OutputCollector;
import org.apache.hadoop.mapred.Reporter;
import org.slf4j.Logger;
import org.slf4j.LoggerFactory;

public class MaxMapper extends MapReduceBase implements Mapper<LongWritable ,Text,Text,IntWritable>{
	private Logger log = LoggerFactory.getLogger(MaxMapper.class);
	private String delimiter=null;
	@Override
	public void configure(JobConf conf){
		delimiter=conf.get("delimiter");
		log.info("delimiter:"+delimiter);
		log.info("This is the begin of MaxMapper");
	}
	
	@Override
	public void map(LongWritable key, Text value,
			OutputCollector<Text, IntWritable> out, Reporter reporter)
			throws IOException {
		// TODO Auto-generated method stub
		String[] values= value.toString().split(delimiter);
		log.info(values[0]+"-->"+values[1]);
		out.collect(new Text(values[0]), new IntWritable(Integer.parseInt(values[1])));
		
	}
	public void close(){
		log.info("This is the end of MaxMapper");
	}
}

MaxReducer：

package chain;

import java.io.IOException;
import java.util.Iterator;

import org.apache.hadoop.io.IntWritable;
import org.apache.hadoop.io.Text;
import org.apache.hadoop.mapred.JobConf;
import org.apache.hadoop.mapred.MapReduceBase;
import org.apache.hadoop.mapred.OutputCollector;
import org.apache.hadoop.mapred.Reducer;
import org.apache.hadoop.mapred.Reporter;
import org.slf4j.Logger;
import org.slf4j.LoggerFactory;

public   class MaxReducer extends MapReduceBase implements Reducer<Text,IntWritable,Text,IntWritable>{
	private Logger log = LoggerFactory.getLogger(MaxReducer.class);
	@Override
	public void configure(JobConf conf){
		log.info("This is the begin of the MaxReducer");
	}
	@Override
	public void reduce(Text key, Iterator<IntWritable> values,
			OutputCollector<Text, IntWritable> out, Reporter reporter)
			throws IOException {
		// TODO Auto-generated method stub
		int max=-1;
		while(values.hasNext()){
			int value=values.next().get();
			if(value>max){
				max=value;
			}
		}
		log.info(key+"-->"+max);
		out.collect(key, new IntWritable(max));
		
	}
	
	@Override
	public void close(){
		log.info("This is the end of the MaxReducer");
	}
}

MergeMaxMapper：

package chain;

import java.io.IOException;
//import java.util.ArrayList;
//import java.util.HashMap;
//import java.util.Map;

import org.apache.hadoop.io.IntWritable;
import org.apache.hadoop.io.Text;
import org.apache.hadoop.mapred.JobConf;
import org.apache.hadoop.mapred.MapReduceBase;
import org.apache.hadoop.mapred.Mapper;
import org.apache.hadoop.mapred.OutputCollector;
import org.apache.hadoop.mapred.Reporter;
import org.slf4j.Logger;
import org.slf4j.LoggerFactory;

public class MergeMaxMapper extends MapReduceBase implements Mapper<Text ,IntWritable,Text,IntWritable>{
	private Logger log = LoggerFactory.getLogger(MergeMaxMapper.class);
//	private Map<Text,ArrayList<IntWritable>> outMap= new HashMap<Text,ArrayList<IntWritable>>();
	@Override
	public void configure(JobConf conf){
		log.info("This is the begin of MergeMaxMapper");
	}
	
	@Override
	public void map(Text key, IntWritable value,
			OutputCollector<Text, IntWritable> out, Reporter reporter)
			throws IOException {
		log.info(key.toString()+"_MergeMaxMapper"+"-->"+value.get());
		out.collect(new Text(key.toString()+"_MergeMaxMapper"), value);
		
	}
	
	@Override
	public void close(){
		log.info("this is the end of MergeMaxMapper");
	}
}

编程思路例如以下：原始測试数据data1、data2首先经过MaxMapper（因为两个文件，所以生成了2个map），然后经过MergeMaxMapper，到MaxReducer，最后再次经过MergeMaxMapper。

在程序中加入了输出数据的log，能够通过log来查看各个map和reduce的数据流程。

mapper端的log（当中的一个mapper）：

2014-05-14 17:23:51,307 INFO [main] chain.MaxMapper: delimiter:,
2014-05-14 17:23:51,307 INFO [main] chain.MaxMapper: This is the begin of MaxMapper
2014-05-14 17:23:51,454 INFO [main] chain.MergeMaxMapper: This is the begin of MergeMaxMapper
2014-05-14 17:23:51,471 INFO [main] chain.MaxMapper: A-->20
2014-05-14 17:23:51,476 INFO [main] chain.MergeMaxMapper: A_MergeMaxMapper-->20
2014-05-14 17:23:51,476 INFO [main] chain.MaxMapper: A-->21
2014-05-14 17:23:51,477 INFO [main] chain.MergeMaxMapper: A_MergeMaxMapper-->21
2014-05-14 17:23:51,477 INFO [main] chain.MaxMapper: A-->22
2014-05-14 17:23:51,477 INFO [main] chain.MergeMaxMapper: A_MergeMaxMapper-->22
2014-05-14 17:23:51,477 INFO [main] chain.MaxMapper: A-->23
2014-05-14 17:23:51,477 INFO [main] chain.MergeMaxMapper: A_MergeMaxMapper-->23
2014-05-14 17:23:51,477 INFO [main] chain.MaxMapper: B-->201
2014-05-14 17:23:51,477 INFO [main] chain.MergeMaxMapper: B_MergeMaxMapper-->201
2014-05-14 17:23:51,477 INFO [main] chain.MaxMapper: B-->301
2014-05-14 17:23:51,477 INFO [main] chain.MergeMaxMapper: B_MergeMaxMapper-->301
2014-05-14 17:23:51,478 INFO [main] chain.MaxMapper: B-->401
2014-05-14 17:23:51,478 INFO [main] chain.MergeMaxMapper: B_MergeMaxMapper-->401
2014-05-14 17:23:51,478 INFO [main] chain.MaxMapper: B-->501
2014-05-14 17:23:51,478 INFO [main] chain.MergeMaxMapper: B_MergeMaxMapper-->501
2014-05-14 17:23:51,481 INFO [main] chain.MaxMapper: This is the end of MaxMapper
2014-05-14 17:23:51,481 INFO [main] chain.MergeMaxMapper: this is the end of MergeMaxMapper

通过上面log，能够看出，通过ChainMapper加入mapper的方式的mapper的处理顺序为：首先初始化第一个mapper（即调用configure方法）；接着初始第二个mapper（调用configure方法）；然后開始map函数，map函数针对一条记录，首先採用mapper1进行处理，然后使用mapper2进行处理；最后是关闭阶段，关闭的顺序同样是首先关闭mapper1（调用close方法），然后关闭mapper2。

reducer端的log（当中一个reducer）

2014-05-14 17:24:10,171 INFO [main] chain.MergeMaxMapper: This is the begin of MergeMaxMapper
2014-05-14 17:24:10,311 INFO [main] chain.MaxReducer: This is the begin of the MaxReducer
2014-05-14 17:24:10,671 INFO [main] chain.MaxReducer: B_MergeMaxMapper-->501
2014-05-14 17:24:10,672 INFO [main] chain.MergeMaxMapper: B_MergeMaxMapper_MergeMaxMapper-->501
2014-05-14 17:24:10,673 INFO [main] chain.MergeMaxMapper: this is the end of MergeMaxMapper
2014-05-14 17:24:10,673 INFO [main] chain.MaxReducer: This is the end of the MaxReducer

通过上面的log能够看出，通过ChainReducer加入mapper的方式，其数据处理顺序为：首先初始化Reducer之后的Mapper，接着初始化Reducer（看configure函数就可以知道）；然后处理reducer，reducer的输出接着交给mapper处理；最后先关闭Mapper，接着关闭reducer。

同一时候，注意到，reducer后面的mapper也是两个的，即有多少个reducer，就有多少个mapper。

通过实验得到上面的ChainReducer的数据处理流程，且ChainReducer没有addReducer的方法，也即是不能加入reducer了，那么最開始提出的mapreduce数据流程就不能採用这种方式实现了。

最后，前面提出的mapreduce数据流程应该是错的，在reducer out里面C组数据不会被拆分为两个reducer，同样的key仅仅会向同一个reducer传输。这里同样做了个试验，通过对接近90M的数据（仅仅有一个分组A）执行上面的程序，能够看到有2个mapper，2个reducer（此数值为设置值），可是在当中一个reducer中并没有A分组的不论什么数据，在另外一个reducer中才有数据。事实上，不用试验也是能够的，曾经看的书上一般都会说同样的key进入同一个reducer中。只是，假设是这种话，那么这种数据效率应该不高。

返回最開始提出的场景，最開始提出的问题，假设同样的key仅仅会进入一个reducer中，那么最后的2个数据文件（2个reducer生成2个数据文件）事实上里面不会有key冲突的数据，所以在进行后面的操作的时候能够直接读多个文件就可以，就像是读一个文件一样。

会产生这种认知错误，应该是对mapreduce 原理不清楚导致。

分享，成长，快乐

转载请注明blog地址：http://blog.csdn.net/fansy1990