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。
扫描二维码关注公众号，回复：
9182252 查看本文章
| <-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<K, V> extends Configured implements Tool {
private static final Log LOG = LogFactory.getLog(InputSampler.class);
static int printUsage() {
System.out.println("sampler -r <reduces>\n [-inFormat <input format class>]\n [-keyClass <map input & output key class>]\n [-splitRandom <double pcnt> <numSamples> <maxsplits> | // Sample from random splits at random (general)\n -splitSample <numSamples> <maxsplits> | // Sample from first records in splits (random data)\n -splitInterval <double pcnt> <maxsplits>] // Sample from splits at intervals (sorted data)");
System.out.println("Default sampler: -splitRandom 0.1 10000 10");
ToolRunner.printGenericCommandUsage(System.out);
return -1;
}
public InputSampler(Configuration conf) {
this.setConf(conf);
}
public static <K, V> void writePartitionFile(Job job, InputSampler.Sampler<K, V> sampler) throws IOException, ClassNotFoundException, InterruptedException {
Configuration conf = job.getConfiguration();
InputFormat inf = (InputFormat)ReflectionUtils.newInstance(job.getInputFormatClass(), conf);
//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<K> 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<String> otherArgs = new ArrayList();
InputSampler.Sampler<K, V> 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, V> {
K[] getSample(InputFormat<K, V> 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<K, V> implements InputSampler.Sampler<K, V> {
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<K, V> inf, Job job) throws IOException, InterruptedException {
//获取分片数
List<InputSplit> splits = inf.getSplits(job);
//采样总数创建数组
ArrayList<K> 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<K, V> 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<K, V> implements InputSampler.Sampler<K, V> {
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<K, V> inf, Job job) throws IOException, InterruptedException {
List<InputSplit> splits = inf.getSplits(job);
ArrayList<K> 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<K, V> 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<K, V> implements InputSampler.Sampler<K, V> {
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<K, V> inf, Job job) throws IOException, InterruptedException {
List<InputSplit> splits = inf.getSplits(job);
ArrayList<K> 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<K, V> 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();
}
}