MapReduce TotalOrderPartitioner 全局排序<转>

我们知道Mapreduce框架在feed数据给reducer之前会对map output key排序，这种排序机制保证了每一个reducer局部有序，Hadoop 默认的partitioner是HashPartitioner，它依赖于output key的hashcode，使得相同key会去相同reducer，但是不保证全局有序，如果想要获得全局排序结果（比如获取top N, bottom N），就需要用到TotalOrderPartitioner了，它保证了相同key去相同reducer的同时也保证了全局有序。

public class HashPartitioner<K, V> extends Partitioner<K, V> {
  /** Use {@link Object#hashCode()} to partition. */
  public int getPartition(K key, V value,
                          int numReduceTasks) {
    return (key.hashCode() & Integer.MAX_VALUE) % numReduceTasks;
  }
}

/**
 * Partitioner effecting a total order by reading split points from
 * an externally generated source.
 */
@InterfaceAudience.Public
@InterfaceStability.Stable
public class TotalOrderPartitioner<K extends WritableComparable<?>,V>
    extends Partitioner<K,V> implements Configurable {
  // by construction, we know if our keytype
  @SuppressWarnings("unchecked") // is memcmp-able and uses the trie
  public int getPartition(K key, V value, int numPartitions) {
    return partitions.findPartition(key);
  }
}

TotalOrderPartitioner依赖于一个partition file来distribute keys，partition file是一个实现计算好的sequence file，如果我们设置的reducer number是N，那么这个文件包含（N-1）个key分割点，并且是基于key comparator排好序的。TotalOrderPartitioner会检查每一个key属于哪一个reducer的范围内，然后决定分发给哪一个reducer。

InputSampler类的writePartitionFile方法会对input files取样并创建partition file。有三种取样方法：

1. RandomSampler  随机取样
2. IntervalSampler  从s个split里面按照一定间隔取样，通常适用于有序数据
3. SplitSampler  从s个split中选取前n条记录取样

paritition file可以通过TotalOrderPartitioner.setPartitionFile(conf, partitionFile)来设置，在TotalOrderPartitioner instance创建的时候会调用setConf函数，这时会读入partition file中key值，如果key是BinaryComparable(可以认为是字符串类型)的话会构建trie，时间复杂度是O(n), n是树的深度。如果是非BinaryComparable类型就构建BinarySearchNode，用二分查找，时间复杂度O(log(n))，n是reduce数

      boolean natOrder =
        conf.getBoolean(NATURAL_ORDER, true);
      if (natOrder && BinaryComparable.class.isAssignableFrom(keyClass)) {
        partitions = buildTrie((BinaryComparable[])splitPoints, 0,
            splitPoints.length, new byte[0],
            // Now that blocks of identical splitless trie nodes are 
            // represented reentrantly, and we develop a leaf for any trie
            // node with only one split point, the only reason for a depth
            // limit is to refute stack overflow or bloat in the pathological
            // case where the split points are long and mostly look like bytes 
            // iii...iixii...iii  .  Therefore, we make the default depth
            // limit large but not huge.
            conf.getInt(MAX_TRIE_DEPTH, 200));
      } else {
        partitions = new BinarySearchNode(splitPoints, comparator);
      }

示例程序

import org.apache.hadoop.conf.Configuration;
import org.apache.hadoop.fs.Path;
import org.apache.hadoop.io.Text;
import org.apache.hadoop.mapreduce.Job;
import org.apache.hadoop.mapreduce.lib.input.FileInputFormat;
import org.apache.hadoop.mapreduce.lib.input.KeyValueTextInputFormat;
import org.apache.hadoop.mapreduce.lib.output.FileOutputFormat;
import org.apache.hadoop.mapreduce.lib.partition.InputSampler;
import org.apache.hadoop.mapreduce.lib.partition.InputSampler.RandomSampler;
import org.apache.hadoop.mapreduce.lib.partition.TotalOrderPartitioner;

public class TotalSortMR {
 
 public static int runTotalSortJob(String[] args) throws Exception {
  Path inputPath = new Path(args[0]);
  Path outputPath = new Path(args[1]);
  Path partitionFile = new Path(args[2]);
  int reduceNumber = Integer.parseInt(args[3]);
  
  // RandomSampler第一个参数表示key会被选中的概率，第二个参数是一个选取samples数，第三个参数是最大读取input splits数
  RandomSampler<Text, Text> sampler = new InputSampler.RandomSampler<Text, Text>(0.1, 10000, 10);
  
  Configuration conf = new Configuration();
  // 设置partition file全路径到conf
  TotalOrderPartitioner.setPartitionFile(conf, partitionFile);
  
  Job job = new Job(conf);
  job.setJobName("Total-Sort");
  job.setJarByClass(TotalSortMR.class);
  job.setInputFormatClass(KeyValueTextInputFormat.class);
  job.setMapOutputKeyClass(Text.class);
  job.setMapOutputValueClass(Text.class);
  job.setNumReduceTasks(reduceNumber);
  
  // partitioner class设置成TotalOrderPartitioner
  job.setPartitionerClass(TotalOrderPartitioner.class);
  
  FileInputFormat.setInputPaths(job, inputPath);
  FileOutputFormat.setOutputPath(job, outputPath);
  outputPath.getFileSystem(conf).delete(outputPath, true);
  
  // 写partition file到mapreduce.totalorderpartitioner.path
  InputSampler.writePartitionFile(job, sampler);
  
  return job.waitForCompletion(true)? 0 : 1;
  
 }
 
 public static void main(String[] args) throws Exception{
  System.exit(runTotalSortJob(args));
 }
}

上面的例子是采用InputSampler来创建partition file，其实还可以使用mapreduce来创建，可以自定义一个inputformat来取样，将output key输出到一个reducer

ps:hive 0.12实现了parallel ORDER BY(https://issues.apache.org/jira/browse/HIVE-1402)，也是基于TotalOrderPartitioner，非常靠谱的new feature啊