Spark的二次排序解决方案

一、MapReduce/Hadoop的二次排序解决方案（点击打开）

二、Spark的二次排序解决方案

方案1：同（一）的方案1，将一个给定键的所有值读取并缓存到一个List数组（Array）数据结构中，然后对这些值进行排序。如果内存不够放，则无法实现

方案2：同（一）的方案2，“会为自然键增加部分或整个值来创建一个组合键以实现排序目标”

三、代码实现（一）：使用Spark Java API 中的groupByKey实现方案1（Java原始语法、lambda语法）

package ercipaixu_spark1;

// STEP-0: import required Java/Spark classes.

import java.util.List;
import java.util.SortedMap;
import java.util.TreeMap;
//
import scala.Tuple2;
//
import org.apache.spark.api.java.JavaRDD;
import org.apache.spark.api.java.JavaPairRDD;
import org.apache.spark.api.java.JavaSparkContext;
import org.apache.spark.api.java.function.Function;
import org.apache.spark.api.java.function.PairFunction;
import org.apache.spark.api.java.function.Function2;
//


/** 
 * SecondarySortUsingCombineByKey class implements the secondary sort design pattern 
 * by using combineByKey().
 *
 *
 * Input:
 *
 *    name, time, value  
 *    x,2,9
 *    y,2,5
 *    x,1,3
 *    y,1,7
 *    y,3,1
 *    x,3,6
 *    z,1,4
 *    z,2,8
 *    z,3,7
 *    z,4,0
 *    p,1,10
 *    p,3,60
 *    p,4,40
 *    p,6,20
 *
 * Output: generate a time-series looking like this:
 *
 *       t1   t2   t3   t4  t5     t6
 *  x => [3,  9,   6]
 *  y => [7,  5,   1]
 *  z => [4,  8,   7,   0]
 *  p => [10, null, 60, 40, null , 20]
 * 
 *  x => [(1,3), (2,9), (3,6)]            where 1 < 2 < 3
 *  y => [(1,7), (2,5), (3,1)]            where 1 < 2 < 3 
 *  z => [(1,4), (2,8), (3,7), (4,0)]     where 1 < 2 < 3 < 4
 *  p => [(1,10), (3,60), (4,40), (6,20)] where 1 < 3 < 4 < 6
 *
 * @author Mahmoud Parsian
 *
 */
public class SecondarySortUsingCombineByKey {

    public static void main(String[] args) throws Exception {
    	args = new String[2];
    	args[0] = "/media/chenjie/0009418200012FF3/ubuntu/sample_input.txt";
    	args[1] = "/media/chenjie/0009418200012FF3/ubuntu/sample_output";
        // STEP-1: read input parameters and validate them
        if (args.length < 2) {
            System.err.println("Usage: SecondarySortUsingCombineByKey  
");
            System.exit(1);
        }
        String inputPath = args[0];
        System.out.println("inputPath=" + inputPath);
        String outputPath = args[1];
        System.out.println("outputPath=" + outputPath);

        // STEP-2: Connect to the Sark master by creating JavaSparkContext object
        final JavaSparkContext ctx = SparkUtil.createJavaSparkContext("local","spark1");

        // STEP-3: Use ctx to create JavaRDD
        //  input record format: <,><,>
        JavaRDD lines = ctx.textFile(inputPath, 1);

        // STEP-4: create (key, value) pairs from JavaRDD where
        // key is the {name} and value is a pair of (time, value).
        // The resulting RDD will be JavaPairRDD>.    
        // convert each record into Tuple2(name, time, value)
        // PairFunction	T => Tuple2(K, V) where K=String and V=Tuple2
        //                                                                                     input   K       V
        System.out.println("===  DEBUG STEP-4 ===");
        JavaPairRDD> pairs = lines.mapToPair(new PairFunction>() {
            @Override
            public Tuple2> call(String s) {
                String[] tokens = s.split(","); // x,2,5
                System.out.println(tokens[0] + "," + tokens[1] + "," + tokens[2]);
                Tuple2 timevalue = new Tuple2(Integer.parseInt(tokens[1]), Integer.parseInt(tokens[2]));
                return new Tuple2>(tokens[0], timevalue);
            }
        });

        // STEP-5: validate STEP-4, we collect all values from JavaPairRDD<> and print it.    
        List>> output = pairs.collect();
        for (Tuple2 t : output) {
            Tuple2 timevalue = (Tuple2) t._2;
            System.out.println(t._1 + "," + timevalue._1 + "," + timevalue._1);
        }

        // How to use combineByKey(): to use combineByKey(), you 
        // need to define 3 basic functions f1, f2, f3:
        // and then you invoke it as: combineByKey(f1, f2, f3)
        //    function 1: create a combiner data structure 
        //    function 2: merge a value into a combined data structure
        //    function 3: merge two combiner data structures
        
        
        // function 1: create a combiner data structure         
        // Here, the combiner data structure is a SortedMap,
        // which keeps track of (time, value) for a given key
        // Tuple2 = Tuple2
        // SortedMap = SortedMap
        Function, SortedMap> createCombiner
                = new Function, SortedMap>() {
            @Override
            public SortedMap call(Tuple2 x) {
                Integer time = x._1;
                Integer value = x._2;
                SortedMap map = new TreeMap<>();
                map.put(time, value);
                return map;
            }
        };

        // function 2: merge a value into a combined data structure
        Function2, Tuple2, SortedMap> mergeValue
                = new Function2, Tuple2, SortedMap>() {
            @Override
            public SortedMap call(SortedMap map, Tuple2 x) {
                Integer time = x._1;
                Integer value = x._2;
                map.put(time, value);
                return map;
            }
        };

        // function 3: merge two combiner data structures
        Function2, SortedMap, SortedMap> mergeCombiners
                = new Function2, SortedMap, SortedMap>() {
            @Override
            public SortedMap call(SortedMap map1, SortedMap map2) {
                if (map1.size() < map2.size()) {
                    return DataStructures.merge(map1, map2);
                } else {
                    return DataStructures.merge(map1, map2);
                }
            }
        };

        // STEP-5: create sorted (time, value)
        JavaPairRDD> combined = pairs.combineByKey(
                createCombiner,
                mergeValue,
                mergeCombiners);

        // STEP-7: validate STEP-6, we collect all values from JavaPairRDD<> and print it.    
        System.out.println("===  DEBUG STEP-6 ===");
        List>> output2 = combined.collect();
        for (Tuple2> t : output2) {
            String name = t._1;
            SortedMap map = t._2;
            System.out.println(name);
            System.out.println(map);
        }

        // persist output
        combined.saveAsTextFile(outputPath);

        // done!
        ctx.close();
        
        // exit
        System.exit(0);
    }

}

其中用到的比较器类，因为我们是按时间进行比较，所以比较（int,int）的第一个分量

package ercipaixu_spark1;


import scala.Tuple2;
import java.util.Comparator;
import java.io.Serializable;

/** 
 * The SparkTupleComparator class enable us to compare two 
 * Tuple2 objects based on the first Tuple2
 * argument.
 *
 * @author Mahmoud Parsian
 *
 */
public class SparkTupleComparator 
   implements Comparator>, Serializable {
    
   public static final SparkTupleComparator INSTANCE = new SparkTupleComparator();
   
   private SparkTupleComparator() {
   }
   
   @Override
   public int compare(Tuple2 t1, Tuple2 t2){
      return t1._1.compareTo(t2._1);
   }
}

四、代码实现（二）：使用Spark Java API 中的combineByKey实现方案1（Java原始语法、lambda语法）

五、代码实现（三）：使用Spark Java API 中的repartitionAndSortWithinPartitions实现方案1

六、代码实现（四）：使用Scala语言实现方案2

import org.apache.spark.Partitioner


/***
  * 自定义Partitioner
  * @param partitions
  */
class CustomPartitioner(partitions: Int) extends Partitioner {
  
  require(partitions > 0, s"Number of partitions ($partitions) cannot be negative.")

  def numPartitions: Int = partitions

  def getPartition(key: Any): Int = key match {
    case (k: String, v: Int) => math.abs(k.hashCode % numPartitions)
    case null                => 0
    case _                   => math.abs(key.hashCode % numPartitions)
  }

  override def equals(other: Any): Boolean = other match {
    case h: CustomPartitioner => h.numPartitions == numPartitions
    case _                    => false
  }

  override def hashCode: Int = numPartitions
}

import org.apache.spark.SparkConf
import org.apache.spark.SparkContext
object SecondarySort {
  def main(args: Array[String]): Unit = {
    val partitions = 1//分区数为1
    val inputPath = "file:///media/chenjie/0009418200012FF3/ubuntu/sample_input.txt"//输入文件路径
    val outputPath = "file:///media/chenjie/0009418200012FF3/ubuntu/sample_output"//输出文件路径
    val conf = new SparkConf().setAppName("CJResult").setMaster("local")//设置Spark在本地运行
    val sc = new SparkContext(conf)//新建sc
    val input = sc.textFile(inputPath)//读取输入文件
    val valueToKey = input.map(x => {
      val line = x.split(",")
      ((line(0) + "-" + line(1), line(2).toInt), line(2).toInt)
    })//将输入文件映射为（（名字-时间，值），值）的复合键值对集合，其中的键也为键值对
    implicit def tupleOrderingDesc = new Ordering[Tuple2[String, Int]] {
      override def compare(x: Tuple2[String, Int], y: Tuple2[String, Int]): Int = {
        if (y._1.compare(x._1) == 0) y._2.compare(x._2)
        else y._1.compare(x._1)
      }
    }//定义一个隐式转换，对于（String，Int）类型的键值对（二元组），将其按照：如果键不相同按照键排序，如果键相同再按值排序
    //其实就是将上面的（名字-时间，值）进行排序
    val sorted = valueToKey.repartitionAndSortWithinPartitions(new CustomPartitioner(partitions))
    val result = sorted.map {
      case (k, v) => (k._1, v)
    }//将（（名字-时间，值），值）转为（名字-时间，值）
    result.saveAsTextFile(outputPath)//将结果写入文件
    sc.stop()
  }
}