(17)Hive ——MR任务的map与reduce个数由什么决定？

一、MapTask的数量由什么决定？

MapTask的数量由以下参数决定

文件个数
文件大小
blocksize

一般而言，对于每一个输入的文件会有一个map split，每一个分片会开启一个map任务，很容易导致小文件问题（如果不进行小文件合并，极可能导致Hadoop集群资源雪崩）

hive中小文件产生的原因及解决方案见文章：

(14)Hive调优——合并小文件-CSDN博客文章浏览阅读779次，点赞10次，收藏17次。Hive的小文件问题https://blog.csdn.net/SHWAITME/article/details/136108785

maxSize的默认值为256M，minSize的默认值是1byte，切片大小splitSize的计算公式：

splitSize=Min(maxSize,Max(minSize,blockSize)) = Min(256M,Max(1 byte ,128M)) = 128M =blockSize

所以默认splitSize就等于blockSize块大小

# minSize的默认值是1byte
set mapred.min.split.size=1

#maxSize的默认值为256M
set mapred.max.split.size=256000000

#hive.input.format是用来指定输入格式的参数。决定了Hive读取数据时使用的输入格式，
set hive.input.format=org.apache.hadoop.hive.ql.io.CombineHiveInputFormat

二、如何调整MapTask的数量

假设blockSize一直是128M，且splitSize = blockSize = 128M。在不改变blockSize块大小的情况下，如何增加/减少mapTask数量

2.1 增加map的数量

增加map：需要调小maxSize，且要小于blockSize才有效，例如maxSize调成100byte

splitSize=Min(maxSize,Max(minSize,blockSize)) = Min(100,Max(1,128*1000*1000)) =100 byte = maxSize

调整前的map数 = 输入文件的大小/ splitSize = 输入文件的大小/ 128M

调整后的map数 = 输入文件的大小/ splitSize = 输入文件的大小/ 100byte

2.2 减少map的数量

减少map：需要调大minSize ，且要大于blockSize才有效，例如minSize 调成200M

splitSize=Min(maxSize,Max(minSize,blockSize)) = Min(256m, Max(200M,128M)) = 200M=minSize

调整前的map数 = 输入文件的大小/ splitSize = 输入文件的大小/ 128M

调整后的map数 = 输入文件的大小/ splitSize = 输入文件的大小/ 200M

三、ReduceTask的数量决定

reduce的个数决定hdfs上落地文件的个数（即： reduce个数决定文件的输出个数）。 ReduceTask的数量，由参数mapreduce.job.reduces 控制，默认值为 -1 时，代表ReduceTask的数量是根据hive的数据量动态计算的。

总体而言，ReduceTask的数量决定方式有以下两种：

3.1 方式一：hive动态计算

3.1.1 动态计算公式

ReduceTask数量 = min (参数2，输入的总数据量/ 参数1)

参数1：hive.exec.reducers.bytes.per.reducer

含义：每个reduce任务处理的数据量（默认值：256M）

参数2：hive.exec.reducers.max

含义：每个MR任务能开启的reduce任务数的上限值（默认值：1009个）

ps：一般参数2的值不会轻易变动，因此在普通集群规模下，hive根据数据量动态计算reduce的个数，计算公式为：输入总数据量/hive.exec.reducers.bytes.per.reducer

3.1.2 源码分析

（1）通过源码分析 hive是如何动态计算reduceTask的个数的？

	在org.apache.hadoop.hive.ql.exec.mr包下的 MapRedTask类中
	//方法类调用逻辑
	MapRedTask 
	     | ----setNumberOfReducers 
	                  | ---- estimateNumberOfReducers
	                    		 |---- estimateReducers

(2)核心方法setNumberOfReducers（读取用户手动设置的reduce个数）

  /**
   * Set the number of reducers for the mapred work.
   */
  private void setNumberOfReducers() throws IOException {
    ReduceWork rWork = work.getReduceWork();
    // this is a temporary hack to fix things that are not fixed in the compiler
    // 获取通过外部传参设置reduce数量的值 rWork.getNumReduceTasks() 
    Integer numReducersFromWork = rWork == null ? 0 : rWork.getNumReduceTasks();

    if (rWork == null) {
      console
          .printInfo("Number of reduce tasks is set to 0 since there's no reduce operator");
    } else {
       
      if (numReducersFromWork >= 0) {
      //如果手动设置了reduce的数量 大于等于0 ，则进来，控制台打印日志
        console.printInfo("Number of reduce tasks determined at compile time: "
            + rWork.getNumReduceTasks());
      } else if (job.getNumReduceTasks() > 0) {
      //如果手动设置了reduce的数量，获取配置中的值，并传入到work中
        int reducers = job.getNumReduceTasks();
        rWork.setNumReduceTasks(reducers);
        console
            .printInfo("Number of reduce tasks not specified. Defaulting to jobconf value of: "
            + reducers);
      } else {
       //如果没有手动设置reduce的数量，进入方法
        if (inputSummary == null) {
          inputSummary =  Utilities.getInputSummary(driverContext.getCtx(), work.getMapWork(), null);
        }
   // #==========【重中之中】estimateNumberOfReducers 
        int reducers = Utilities.estimateNumberOfReducers(conf, inputSummary, work.getMapWork(),
                                                          work.isFinalMapRed());
        rWork.setNumReduceTasks(reducers);
        console
            .printInfo("Number of reduce tasks not specified. Estimated from input data size: "
            + reducers);

      }
      
      //hive shell中所看到的控制台打印日志就在这里
      console
          .printInfo("In order to change the average load for a reducer (in bytes):");
      console.printInfo("  set " + HiveConf.ConfVars.BYTESPERREDUCER.varname
          + "=<number>");
      console.printInfo("In order to limit the maximum number of reducers:");
      console.printInfo("  set " + HiveConf.ConfVars.MAXREDUCERS.varname
          + "=<number>");
      console.printInfo("In order to set a constant number of reducers:");
      console.printInfo("  set " + HiveConf.ConfVars.HADOOPNUMREDUCERS
          + "=<number>");
    }
  }

(3)如果没有手动设置reduce的个数，hive是如何动态计算reduce个数的？

    int reducers = Utilities.estimateNumberOfReducers(conf, inputSummary, work.getMapWork(),
                                                          work.isFinalMapRed());
                                                          
  /**
   * Estimate the number of reducers needed for this job, based on job input,
   * and configuration parameters.
   *
   * The output of this method should only be used if the output of this
   * MapRedTask is not being used to populate a bucketed table and the user
   * has not specified the number of reducers to use.
   *
   * @return the number of reducers.
   */
  public static int estimateNumberOfReducers(HiveConf conf, ContentSummary inputSummary,
                                             MapWork work, boolean finalMapRed) throws IOException {
    // bytesPerReducer 每个reduce处理的数据量，默认值为256M  BYTESPERREDUCER("hive.exec.reducers.bytes.per.reducer", 256000000L)
    long bytesPerReducer = conf.getLongVar(HiveConf.ConfVars.BYTESPERREDUCER);
    
    //整个mr任务，可以开启的reduce个数的上限值：maxReducers的默认值1009个MAXREDUCERS("hive.exec.reducers.max", 1009)
    int maxReducers = conf.getIntVar(HiveConf.ConfVars.MAXREDUCERS);
   
   //#===========对totalInputFileSize的计算
    double samplePercentage = getHighestSamplePercentage(work);
    long totalInputFileSize = getTotalInputFileSize(inputSummary, work, samplePercentage);

    // if all inputs are sampled, we should shrink the size of reducers accordingly.
    if (totalInputFileSize != inputSummary.getLength()) {
      LOG.info("BytesPerReducer=" + bytesPerReducer + " maxReducers="
          + maxReducers + " estimated totalInputFileSize=" + totalInputFileSize);
    } else {
      LOG.info("BytesPerReducer=" + bytesPerReducer + " maxReducers="
        + maxReducers + " totalInputFileSize=" + totalInputFileSize);
    }

    // If this map reduce job writes final data to a table and bucketing is being inferred,
    // and the user has configured Hive to do this, make sure the number of reducers is a
    // power of two
    boolean powersOfTwo = conf.getBoolVar(HiveConf.ConfVars.HIVE_INFER_BUCKET_SORT_NUM_BUCKETS_POWER_TWO) &&
        finalMapRed && !work.getBucketedColsByDirectory().isEmpty();
    
    //#==============【真正计算reduce个数的方法】看源码的技巧return的方法是重要核心方法
    return estimateReducers(totalInputFileSize, bytesPerReducer, maxReducers, powersOfTwo);
  }

(4) 动态计算reduce个数的方法 estimateReducers

	public static int estimateReducers(long totalInputFileSize, long bytesPerReducer,
      int maxReducers, boolean powersOfTwo) {
    
  
    double bytes = Math.max(totalInputFileSize, bytesPerReducer);
  // 假设totalInputFileSize 1000M
    // bytes=Math.max(1000M,256M)=1000M


    int reducers = (int) Math.ceil(bytes / bytesPerReducer);
    
    //reducers=(int)Math.ceil(1000M/256M)=4  此公式说明如果totalInputFileSize 小于256M ，则reducers=1 ;也就是当输入reduce端的数据量特别小，即使手动设置reduce Task数量为5，最终也只会开启1个reduceTask
    
  
    reducers = Math.max(1, reducers);
  //Math.max(1, 4)=4 ，reducers的结果还是4

    reducers = Math.min(maxReducers, reducers);
    //Math.min(1009,4)=4; reducers的结果还是4

      
    int reducersLog = (int)(Math.log(reducers) / Math.log(2)) + 1;
    int reducersPowerTwo = (int)Math.pow(2, reducersLog);

    if (powersOfTwo) {
      // If the original number of reducers was a power of two, use that
      if (reducersPowerTwo / 2 == reducers) {
        // nothing to do
      } else if (reducersPowerTwo > maxReducers) {
        // If the next power of two greater than the original number of reducers is greater
        // than the max number of reducers, use the preceding power of two, which is strictly
        // less than the original number of reducers and hence the max
        reducers = reducersPowerTwo / 2;
      } else {
        // Otherwise use the smallest power of two greater than the original number of reducers
        reducers = reducersPowerTwo;
      }
    }
    return reducers;
  }

3.2 方式二：用户手动指定

手动调整reduce个数： set mapreduce.job.reduces = 10

需要注意：出现以下几种情况时，手动调整reduce个数不生效。

3.2.1 order by 全局排序

sql中使用了order by全局排序，那只能在一个reduce中完成，无论怎么调整reduce的数量都是无效的。

  
  hive (default)>set mapreduce.job.reduces=5;
  hive (default)> select * from empt order  by length(ename);
  Hadoop job information for Stage-1: number of mappers: 1; number of reducers: 1

3.2.2 map端输出的数据量很小

在【3.1.2 源码分析——(4) 】动态计算reduce个数的核心方法 estimateReducers中，有下面这三行代码：

int reducers = (int) Math.ceil(bytes / bytesPerReducer);
reducers = Math.max(1, reducers);
reducers = Math.min(maxReducers, reducers);

如果map端输出的数据量bytes (假如只有1M) 远小于hive.exec.reducers.bytes.per.reducer (每个reduce处理的数据量默认值为256M) 参数值，maxReducers默认为1009个，计算下列值：

  int reducers  = (int) Math.ceil(1 / 256M)=1;
  reducers = Math.max(1, 1)=1;
  reducers = Math.min(1009, 1)=1;

此时即使用户手动 set mapreduce.job.reduces=10，也不生效，reduce个数最后还是只有1个。

参考文章：

Hive mapreduce的map与reduce个数由什么决定？_hive中map任务和reduce任务数量计算原理-CSDN博客