9.2.2 hadoop采样分组源码解析SplitSampler、RandomSampler、IntervalSampler

采样分组

为了实现输出的全局排序，可以对温度数据进行分组处理，实现多个reduce处理，组间有序，组内有序，从而实现全局有序。而如何分组才能保证每个reduce分到的数据差不多，这样作业中的任务执行时间也差不多。例如将处理温度数据，要求温度按顺序输出。分成4组个分组，<-10℃，-10℃~0℃, 0℃~10℃,>10℃。

<-10℃	-10℃~0℃	0℃~10℃	>10℃
1%	%10	29%	60%

显然这样分区后，每个ruduce获取的数据量相差很大。调整下分区后使每个分区获取到的数据量差不多，reduce任务分析起来处理时间差不多。而这些分区的边界值就需要通过采样来决定。

<-4℃	4℃~13℃	13℃~25℃	>5℃
23%	%24	27%	26%

 

用InputSampler对象对输入数据进行采样，得到数据的采样区间分隔值，将这些值写入到一个文件中。然后TotalOrderPartitioner类读取这些边界值作为分区依据。采样分组就是通过采集输入的部分数据，得到相对均匀的分布区间，每个区间的数据量差不多。InputSampler是采样方式有三种：前n条记录采样SplitSample，随机采样RandomSample，固定间隔采样？IntervalSample。

类名称	采样方式	构造方法	效率
SplitSampler(int numSamples, int maxSplitsSampled)	对输入分片均匀采样，每个分片取前n个。	采样总数，用于采样的分片数	最高
RandomSampler(double freq, int numSamples, int maxSplitsSampled)	遍历所有数据，随机采样	采样频率，采样总数，划分数	最低
IntervalSampler(double freq, int maxSplitsSampled)	固定间隔采样对有序的数据十分适用	采样频率，划分数	中

（4）InputSampler原理

1）InputSampler是个hadoop任务类，继承Configured，实现Tool接口，main函数作为入口函数，run函数用来执行任务，InputSampler还有另外一个writePartitionFile函数，它是将采样的值排序，然后按照分区的数量进行划分，得到边界值写入分区文件，其定义为如下：

public class InputSampler extends Configured implements Tool {

    private static final Log LOG = LogFactory.getLog(InputSampler.class);



    static int printUsage() {

        System.out.println("sampler -r \n      [-inFormat ]\n      [-keyClass 

    //有numPartitions个reduce任务就有numPartitions分区，产生numPartitions个文件

        int numPartitions = job.getNumReduceTasks();

      //获取采样的值

        K[] samples = (Object[])sampler.getSample(inf, job);

        LOG.info("Using " + samples.length + " samples");

      //获取排序函数，对采样值进行排序

        RawComparator comparator = job.getSortComparator();

        Arrays.sort(samples, comparator);

      //获取分区文件的路径

        Path dst = new Path(TotalOrderPartitioner.getPartitionFile(conf));

        FileSystem fs = dst.getFileSystem(conf);

      //如果存在，删除原来的分区文件

        if (fs.exists(dst)) {

            fs.delete(dst, false);

        }

      创建写入对象，创建新的文件

        Writer writer = SequenceFile.createWriter(fs, conf, dst, job.getMapOutputKeyClass(), NullWritable.class);

        NullWritable nullValue = NullWritable.get();

      //获取间隔值的步长，已经排好序之后，每隔stepSize取一个值作为分组的边界值

        float stepSize = (float)samples.length / (float)numPartitions;

        int last = -1;



        for(int i = 1; i < numPartitions; ++i) {

            int k;

            for(k = Math.round(stepSize * (float)i); last >= k && comparator.compare(samples[last], samples[k]) == 0; ++k) {

                ;

            }



            writer.append(samples[k], nullValue);

            last = k;

        }



        writer.close();

    }

//run函数执行采样任务，入参是采样类型

    public int run(String[] args) throws Exception {

        Job job = new Job(this.getConf());

        ArrayList otherArgs = new ArrayList();

        InputSampler.Sampler sampler = null;



        for(int i = 0; i < args.length; ++i) {

            try {

                if ("-r".equals(args[i])) {

                    ++i;

                    job.setNumReduceTasks(Integer.parseInt(args[i]));

                } else if ("-inFormat".equals(args[i])) {

                    ++i;

                    job.setInputFormatClass(Class.forName(args[i]).asSubclass(InputFormat.class));

                } else if ("-keyClass".equals(args[i])) {

                    ++i;

                    job.setMapOutputKeyClass(Class.forName(args[i]).asSubclass(WritableComparable.class));

                } else if ("-splitSample".equals(args[i])) {

                    ++i;

                    int numSamples = Integer.parseInt(args[i]);

                    ++i;

                    int maxSplits = Integer.parseInt(args[i]);

                    if (0 >= maxSplits) {

                        maxSplits = 2147483647;

                    }

分区采样

                    sampler = new InputSampler.SplitSampler(numSamples, maxSplits);

                } else {

                    int maxSplits;

                    double pcnt;

                    if ("-splitRandom".equals(args[i])) {

                        ++i;

                        pcnt = Double.parseDouble(args[i]);

                        ++i;

                        maxSplits = Integer.parseInt(args[i]);

                        ++i;

                        int maxSplits = Integer.parseInt(args[i]);

                        if (0 >= maxSplits) {

                            maxSplits = 2147483647;

                        }

                     //随机采样

                        sampler = new InputSampler.RandomSampler(pcnt, maxSplits, maxSplits);

                    } else if ("-splitInterval".equals(args[i])) {

                        ++i;

                        pcnt = Double.parseDouble(args[i]);

                        ++i;

                        maxSplits = Integer.parseInt(args[i]);

                        if (0 >= maxSplits) {

                            maxSplits = 2147483647;

                        }

                     //间隔采样

                        sampler = new InputSampler.IntervalSampler(pcnt, maxSplits);

                    } else {

                        otherArgs.add(args[i]);

                    }

                }

            } catch (NumberFormatException var10) {

                System.out.println("ERROR: Integer expected instead of " + args[i]);

                return printUsage();

            } catch (ArrayIndexOutOfBoundsException var11) {

                System.out.println("ERROR: Required parameter missing from " + args[i - 1]);

                return printUsage();

            }

        }

      // reduce任务数量不能<=2，否则分组就没有了任何意义



        if (job.getNumReduceTasks() <= 1) {

            System.err.println("Sampler requires more than one reducer");

            return printUsage();

        } else if (otherArgs.size() < 2) {

            System.out.println("ERROR: Wrong number of parameters: ");

            return printUsage();

        } else {

            if (null == sampler) {

               //默认采用随机采样

                sampler = new InputSampler.RandomSampler(0.1D, 10000, 10);

            }



            Path outf = new Path((String)otherArgs.remove(otherArgs.size() - 1));

            TotalOrderPartitioner.setPartitionFile(this.getConf(), outf);

            Iterator i$ = otherArgs.iterator();



            while(i$.hasNext()) {

                String s = (String)i$.next();

                FileInputFormat.addInputPath(job, new Path(s));

            }

           //默任执行写入分区文件

            writePartitionFile(job, (InputSampler.Sampler)sampler);

            return 0;

        }

    }



    public static void main(String[] args) throws Exception {

        InputSampler sampler = new InputSampler(new Configuration());

        int res = ToolRunner.run(sampler, args);

        System.exit(res);

    }

public interface Sampler {

        K[] getSample(InputFormat var1, Job var2) throws IOException, InterruptedException;

    }

}

 

2）InputSampler类定义个一个采样接口Sample接口，定义方法getSample，SplitSample、RandomSample、IntervalSample类都实现了这个接口，采用不同的方法获取采样值。三个类都是InputSampler的内部静态类，实现了getSample方法，下面分别阐述。

SplitSample类定义

总的采样数除以用于采样的分片数量，得到每个分片的取样数n，采取每个分片的前n个数据。

public static class SplitSampler implements InputSampler.Sampler {

    protected final int numSamples;

    protected final int maxSplitsSampled;

    public SplitSampler(int numSamples) {

        this(numSamples, 2147483647);

    }



    public SplitSampler(int numSamples, int maxSplitsSampled) {

        this.numSamples = numSamples;//采样总数

        this.maxSplitsSampled = maxSplitsSampled;// 用于取样的分片数量，不大于实际分片数

    }



    public K[] getSample(InputFormat inf, Job job) throws IOException, InterruptedException {

        //获取分片数

    List splits = inf.getSplits(job);

    //采样总数创建数组

        ArrayList samples = new ArrayList(this.numSamples);       

    //用于取样的分片数量

        int splitsToSample = Math.min(this.maxSplitsSampled, splits.size());

        //每个分片需要采集多少个数据

    int samplesPerSplit = this.numSamples / splitsToSample;

        long records = 0L;



        for(int i = 0; i < splitsToSample; ++i) {

            TaskAttemptContext samplingContext = new TaskAttemptContextImpl(job.getConfiguration(), new TaskAttemptID());

//创建读取记录的reader

            RecordReader reader = inf.createRecordReader((InputSplit)splits.get(i), samplingContext);

            reader.initialize((InputSplit)splits.get(i), samplingContext);



            while(reader.nextKeyValue()) {

//采样数据写入smaple数组

                samples.add(ReflectionUtils.copy(job.getConfiguration(), reader.getCurrentKey(), (Object)null));

                ++records;

             //每个分片只采集前个samplesPerSplit数据，超出则退出

                if ((long)((i + 1) * samplesPerSplit) <= records) {

                    break;

                }

            }



            reader.close();

        }



        return (Object[])samples.toArray();

    }

}

IntervalSample类定义

遍历用于采样的分片数据，根据采样率来等间隔采集数据，例如采样率是0.1，则每隔10个采集一个数据。

public static class IntervalSampler implements InputSampler.Sampler {

    protected final double freq;

    protected final int maxSplitsSampled;



    public IntervalSampler(double freq) {

        this(freq, 2147483647);

    }



    public IntervalSampler(double freq, int maxSplitsSampled) {

        this.freq = freq;//采样率

        this.maxSplitsSampled = maxSplitsSampled;//用于采样的分片数

    }



    public K[] getSample(InputFormat inf, Job job) throws IOException, InterruptedException {

        List splits = inf.getSplits(job);

        ArrayList samples = new ArrayList();

        int splitsToSample = Math.min(this.maxSplitsSampled, splits.size());

        long records = 0L;//遍历的记录数

        long kept = 0L;//采集的记录数



        for(int i = 0; i < splitsToSample; ++i) {

            TaskAttemptContext samplingContext = new TaskAttemptContextImpl(job.getConfiguration(), new TaskAttemptID());

            RecordReader reader = inf.createRecordReader((InputSplit)splits.get(i), samplingContext);

            reader.initialize((InputSplit)splits.get(i), samplingContext);



            while(reader.nextKeyValue()) {

             //假设freq为0.1,第一次循环，record为1，kept为0，0/1小于freq0.1，第一条记录会被采到，kept变为1；第二次循环，record=2，kept=1，1/2大于freq0.1，第二条记录不会取到；kept/records的值从1/2,1/3,1/4……1/10大于等于freq0.1，第11条记录时，1/11小于0.1，第11条记录会被取到，kept变成2，只有到2/21时，才会取第三条数据，所以是每隔10条取一个，是等间隔取数据。

             ++records;

                if ((double)kept / (double)records < this.freq) {

                    samples.add(ReflectionUtils.copy(job.getConfiguration(), reader.getCurrentKey(), (Object)null));

                    ++kept;

                }

            }



            reader.close();

        }



        return (Object[])samples.toArray();

    }

}

RandomSample类定义

随机采样输入参数是采样频率，采样总数，用于采样的的分片数。遍历用于采样的分片中的记录，如果随机数小于采样率则进行采样，添加进入采样数组，或者更换已满数组中的值。同时减小采样率，越往后面，采集到数据的概率越小。

public static class RandomSampler implements InputSampler.Sampler {

    protected double freq;

    protected final int numSamples;

    protected final int maxSplitsSampled;



    public RandomSampler(double freq, int numSamples) {

        this(freq, numSamples, 2147483647);

    }



    public RandomSampler(double freq, int numSamples, int maxSplitsSampled) {

        this.freq = freq;//采样率

        this.numSamples = numSamples;//采样总数

        this.maxSplitsSampled = maxSplitsSampled;//用于采样的分片数

    }



    public K[] getSample(InputFormat inf, Job job) throws IOException, InterruptedException {

        List splits = inf.getSplits(job);

        ArrayList samples = new ArrayList(this.numSamples);//采样保存申请空间

        int splitsToSample = Math.min(this.maxSplitsSampled, splits.size());//计算用于采样的分片数

        Random r = new Random();//创建随机对象

        long seed = r.nextLong();//创建随机种子

        r.setSeed(seed);

        InputSampler.LOG.debug("seed: " + seed);



        int i;//将分片打乱顺序，随机获取第j个分片，和第i个分片进行交换

        for(i = 0; i < splits.size(); ++i) {

            InputSplit tmp = (InputSplit)splits.get(i);

            int j = r.nextInt(splits.size());

            splits.set(i, splits.get(j));

            splits.set(j, tmp);

        }

    //循环从用于采样的分片中随机获取数据，直到采样分片遍历完（可能数量不够numSamples个），或者已经采集到numSamples个数据

        for(i = 0; i < splitsToSample || i < splits.size() && samples.size() < this.numSamples; ++i) {

            TaskAttemptContext samplingContext = new TaskAttemptContextImpl(job.getConfiguration(), new TaskAttemptID());

            RecordReader reader = inf.createRecordReader((InputSplit)splits.get(i), samplingContext);

            reader.initialize((InputSplit)splits.get(i), samplingContext);



            while(reader.nextKeyValue()) {

//随机double值小于采样率，符合条件，进行获取当前值，这样有可能，遍历所有的值，可能没有获取到指定的numSamples记录？

                if (r.nextDouble() <= this.freq) {

//采样数组中数据还不足则add进去，如果已经采集到了numSamples个记录，则随机替换set到sample数组中

                    if (samples.size() < this.numSamples) {

                        samples.add(ReflectionUtils.copy(job.getConfiguration(), reader.getCurrentKey(), (Object)null));

                    } else {

                        int ind = r.nextInt(this.numSamples);

                        if (ind != this.numSamples) {

                            samples.set(ind, ReflectionUtils.copy(job.getConfiguration(), reader.getCurrentKey(), (Object)null));

                        }

//每采样到一个数据，采样率会减小，r.nextDouble() <= this.freq采样到的数据概率会减小

                        this.freq *= (double)(this.numSamples - 1) / (double)this.numSamples;

                    }

                }

            }



            reader.close();

        }



        return (Object[])samples.toArray();

    }

}