九、MapReduce--input源碼分析

當job提交至yarn之后，就會開始調度運行map任務，這里開始講解map輸入的源碼分析。
一個map任務的入口就是 MapTask.class 中的run() 方法

1、首先看看MapTask.run() 方法

MapTask.class

//---------------------------------MapTask.java
public void run(JobConf job, TaskUmbilicalProtocol umbilical) throws IOException, ClassNotFoundException, InterruptedException { 
    this.umbilical = umbilical;
    if (this.isMapTask()) {
        if (this.conf.getNumReduceTasks() == 0) {
            this.mapPhase = this.getProgress().addPhase("map", 1.0F);
        } else {
            this.mapPhase = this.getProgress().addPhase("map", 0.667F);
            this.sortPhase = this.getProgress().addPhase("sort", 0.333F);
        }
    }

    TaskReporter reporter = this.startReporter(umbilical);
    boolean useNewApi = job.getUseNewMapper();

    //進行map任務的初始化
    this.initialize(job, this.getJobID(), reporter, useNewApi);
    if (this.jobCleanup) {
        this.runJobCleanupTask(umbilical, reporter);
    } else if (this.jobSetup) {
        this.runJobSetupTask(umbilical, reporter);
    } else if (this.taskCleanup) {
        this.runTaskCleanupTask(umbilical, reporter);
    } else {
        //啟動map任務，判斷是使用新的還是舊的api
        if (useNewApi) {
            this.runNewMapper(job, this.splitMetaInfo, umbilical, reporter);
        } else {
            this.runOldMapper(job, this.splitMetaInfo, umbilical, reporter);
        }

        this.done(umbilical, reporter);
    }
}

上面重點有兩個方法，一個是 this.initialize()以及 this.runNewMapper()。

2、下面看看this.initialize()

//---------------------------------Task.java
public void initialize(JobConf job, JobID id, Reporter reporter, boolean useNewApi) throws IOException, ClassNotFoundException, InterruptedException {
    //創建task以及job上下文對象
    this.jobContext = new JobContextImpl(job, id, reporter);
    this.taskContext = new TaskAttemptContextImpl(job, this.taskId, reporter);
    //將task任務的狀態改為正在運行
    if (this.getState() == org.apache.hadoop.mapred.TaskStatus.State.UNASSIGNED) {
        this.setState(org.apache.hadoop.mapred.TaskStatus.State.RUNNING);
    }

    if (useNewApi) {
        if (LOG.isDebugEnabled()) {
            LOG.debug("using new api for output committer");
        }

        //獲取job中配置的輸出格式類，并通過反射獲取該類的Class對象
        this.outputFormat = (OutputFormat)ReflectionUtils.newInstance(this.taskContext.getOutputFormatClass(), job);
        //通過outputformat類獲取commiter
        this.committer = this.outputFormat.getOutputCommitter(this.taskContext);
    } else {
        this.committer = this.conf.getOutputCommitter();
    }

    //從FileOutputFormat獲取任務結果輸出路徑。
    /*
    可能有的人會奇怪，為啥mapper這里要獲取outputformat 的輸出路徑。
    首先我們要知道，一個MapReduce任務可以只有mapper，而沒有reducer的，
    那么這時候程序的輸出是有mapper直接輸出的，這時候自然就需要知道輸出的路徑，這里就派上用場了
    */
    Path outputPath = FileOutputFormat.getOutputPath(this.conf);
    if (outputPath != null) {
        if (this.committer instanceof FileOutputCommitter) {
            FileOutputFormat.setWorkOutputPath(this.conf, ((FileOutputCommitter)this.committer).getTaskAttemptPath(this.taskContext));
        } else {
            FileOutputFormat.setWorkOutputPath(this.conf, outputPath);
        }
    }

    this.committer.setupTask(this.taskContext);
    Class<? extends ResourceCalculatorProcessTree> clazz = this.conf.getClass("mapreduce.job.process-tree.class", (Class)null, ResourceCalculatorProcessTree.class);
    this.pTree = ResourceCalculatorProcessTree.getResourceCalculatorProcessTree((String)System.getenv().get("JVM_PID"), clazz, this.conf);
    LOG.info(" Using ResourceCalculatorProcessTree : " + this.pTree);
    if (this.pTree != null) {
        this.pTree.updateProcessTree();
        this.initCpuCumulativeTime = this.pTree.getCumulativeCpuTime();
    }

}

這個方法主要做了一些初始化工作，比如創建上下文對象，獲取輸出outputFormat類，以及路徑等。

3、下面接著看看this.runNewMapper()

//---------------------------------MapTask.java
private <INKEY, INVALUE, OUTKEY, OUTVALUE> void runNewMapper(JobConf job, TaskSplitIndex splitIndex, TaskUmbilicalProtocol umbilical, TaskReporter reporter) throws IOException, ClassNotFoundException, InterruptedException {
    TaskAttemptContext taskContext = new TaskAttemptContextImpl(job, this.getTaskID(), reporter);
    //通過反射獲取job中配置的mapper實現類
    Mapper<INKEY, INVALUE, OUTKEY, OUTVALUE> mapper = (Mapper)ReflectionUtils.newInstance(taskContext.getMapperClass(), job);
    //通過反射獲取job中配置的輸入格式類,默認是TextInputFormat
    InputFormat<INKEY, INVALUE> inputFormat = (InputFormat)ReflectionUtils.newInstance(taskContext.getInputFormatClass(), job);
    org.apache.hadoop.mapreduce.InputSplit split = null;
    //獲取切片詳細信息,傳入輸出路徑以及偏移量作為參數.也就是當前mapper處理的某個切片
    split = (org.apache.hadoop.mapreduce.InputSplit)this.getSplitDetails(new Path(splitIndex.getSplitLocation()), splitIndex.getStartOffset());
    LOG.info("Processing split: " + split);
    //獲取輸入的讀取數據文件的 RecordReader 的對象，默認inputformat為TextInputFormat，對應默認的RecordReader為LineRecordReader
    org.apache.hadoop.mapreduce.RecordReader<INKEY, INVALUE> input = new MapTask.NewTrackingRecordReader(split, inputFormat, reporter, taskContext);
    job.setBoolean("mapreduce.job.skiprecords", this.isSkipping());
    RecordWriter output = null;
    //獲取RecordWriter輸出對象
    if (job.getNumReduceTasks() == 0) {
        output = new MapTask.NewDirectOutputCollector(taskContext, job, umbilical, reporter);
    } else {
        output = new MapTask.NewOutputCollector(taskContext, job, umbilical, reporter);
    }

    MapContext<INKEY, INVALUE, OUTKEY, OUTVALUE> mapContext = new MapContextImpl(job, this.getTaskID(), input, (RecordWriter)output, this.committer, reporter, split);
    org.apache.hadoop.mapreduce.Mapper.Context mapperContext = (new WrappedMapper()).getMapContext(mapContext);

    try {
        //初始化RecordReader中的數據
        input.initialize(split, mapperContext);
        //運行mapper中的run方法，也就是Mapper類中的run方法，開始運行map任務
        mapper.run(mapperContext);
        this.mapPhase.complete();
        this.setPhase(Phase.SORT);
        this.statusUpdate(umbilical);
        //map運行完，關閉輸入、輸出流
        input.close();
        input = null;
        ((RecordWriter)output).close(mapperContext);
        output = null;
    } finally {
        this.closeQuietly((org.apache.hadoop.mapreduce.RecordReader)input);
        this.closeQuietly((RecordWriter)output, mapperContext);
    }

}

可以看到，這里就是整個map任務的核心流程，做了以下工作：
（1）獲取mapper類對象，下面要執行里面的map方法
（2）獲取InputFormat對象，默認是默認inputformat為TextInputFormat
（3）通過InputFormat對象獲取RecordReader對象，后面用于讀取數據文件
（4）獲取用于輸出map的結果的RecordWriter對象
（5）獲取切片信息，比如切片所在文件的路徑，起始偏移量等
（6）初始化切片數據
（7）開始運行mapper中的run()方法
（8）運行完畢，關閉輸入流，將結果通過RecordWriter刷寫。
（9）刷寫完畢后，關閉輸入流以及輸出流
下面看看其中的核心方法

4、this.getSplitDetails() 獲取切片信息

//---------------------------------MapTask.java
private <T> T getSplitDetails(Path file, long offset) throws IOException {
    //獲取文件系統對象，并打開文件輸出流
    FileSystem fs = file.getFileSystem(this.conf);
    FSDataInputStream inFile = fs.open(file);
    //跳過指定的偏移量，也就是從指定偏移量的位置開始讀取數據，其實就是切片開始的偏移量
    inFile.seek(offset);
    String className = StringInterner.weakIntern(Text.readString(inFile));

    Class cls;
    try {
        cls = this.conf.getClassByName(className);
    } catch (ClassNotFoundException var13) {
        IOException wrap = new IOException("Split class " + className + " not found");
        wrap.initCause(var13);
        throw wrap;
    }

    SerializationFactory factory = new SerializationFactory(this.conf);
    //反序列化方式打開輸入流
    Deserializer<T> deserializer = factory.getDeserializer(cls);
    deserializer.open(inFile);
    T split = deserializer.deserialize((Object)null);
    long pos = inFile.getPos();
    ((Counter)this.getCounters().findCounter(TaskCounter.SPLIT_RAW_BYTES)).increment(pos - offset);
    inFile.close();
    //返回切片經過反序列化之后的可讀取對象
    return split;
}

可以看到這里主要是返回切片的反序列化之后可以讀取的信息對象

5、接著看看 input.initialize()

在看這個方法之前，首先我們看看input這個對象是由哪個類創建的。它是由NewTrackingRecordReader 這個類創建的。這是個靜態內部類

//---------------------------------MapTask.java
static class NewTrackingRecordReader<K, V> extends org.apache.hadoop.mapreduce.RecordReader<K, V> {
    private final org.apache.hadoop.mapreduce.RecordReader<K, V> real;
    private final org.apache.hadoop.mapreduce.Counter inputRecordCounter;
    private final org.apache.hadoop.mapreduce.Counter fileInputByteCounter;
    private final TaskReporter reporter;
    private final List<Statistics> fsStats;

    NewTrackingRecordReader(org.apache.hadoop.mapreduce.InputSplit split, InputFormat<K, V> inputFormat, TaskReporter reporter, TaskAttemptContext taskContext) throws InterruptedException, IOException {
        this.reporter = reporter;
        this.inputRecordCounter = reporter.getCounter(TaskCounter.MAP_INPUT_RECORDS);
        this.fileInputByteCounter = reporter.getCounter(FileInputFormatCounter.BYTES_READ);
        List<Statistics> matchedStats = null;
        if (split instanceof org.apache.hadoop.mapreduce.lib.input.FileSplit) {
            matchedStats = Task.getFsStatistics(((org.apache.hadoop.mapreduce.lib.input.FileSplit)split).getPath(), taskContext.getConfiguration());
        }

        this.fsStats = matchedStats;
        long bytesInPrev = this.getInputBytes(this.fsStats);
        //調用job任務中定義的inputformat類中的createRecordReader方法，獲取RecordReader對象。返回的是 LineRecordReader
        this.real = inputFormat.createRecordReader(split, taskContext);
        long bytesInCurr = this.getInputBytes(this.fsStats);
        this.fileInputByteCounter.increment(bytesInCurr - bytesInPrev);
    }
    ...........
}

我們可以看到構造方法中，是調用 inputFormat對象的createRecordReader() 方法來創建RecordReader對象的，上面也說了默認inputFormat為 TextInputFormat。

//---------------------------TextInputFormat.java
public class TextInputFormat extends FileInputFormat<LongWritable, Text> {
    public TextInputFormat() {
    }

    public RecordReader<LongWritable, Text> createRecordReader(InputSplit split, TaskAttemptContext context) {
        String delimiter = context.getConfiguration().get("textinputformat.record.delimiter");
        byte[] recordDelimiterBytes = null;
        if (null != delimiter) {
            recordDelimiterBytes = delimiter.getBytes(Charsets.UTF_8);
        }

        return new LineRecordReader(recordDelimiterBytes);
    }

可以清楚看到，返回的就是 LineRecordReader 這個reader類。

接著我們繼續看 input.initialize()

static class NewTrackingRecordReader<K, V> extends org.apache.hadoop.mapreduce.RecordReader<K, V> {
    public void initialize(org.apache.hadoop.mapreduce.InputSplit split, TaskAttemptContext context) throws IOException, InterruptedException {
        long bytesInPrev = this.getInputBytes(this.fsStats);
        //調用 RecordReader對象的 initialize方法，初始化輸入。上面說到默認的是LineRecordReader
        //this.real已經在上面初始化了，就是LineRecordReader
        this.real.initialize(split, context);
        long bytesInCurr = this.getInputBytes(this.fsStats);
        this.fileInputByteCounter.increment(bytesInCurr - bytesInPrev);
    }
}

可以看到，調用 RecordReader中的 initialize 方法，也就是調用LineRecordReader 中的 initialize() 方法，下面看看

//---------------------------------------LineRecordReader.java
public void initialize(InputSplit genericSplit, TaskAttemptContext context) throws IOException {
    FileSplit split = (FileSplit)genericSplit;
    Configuration job = context.getConfiguration();
    this.maxLineLength = job.getInt("mapreduce.input.linerecordreader.line.maxlength", 2147483647);
    //獲取切片的數據開始位置以及終止位置
    this.start = split.getStart();
    this.end = this.start + split.getLength();
    //獲取切片對應的文件的輸入流
    Path file = split.getPath();
    FileSystem fs = file.getFileSystem(job);
    this.fileIn = fs.open(file);
    //如果文件有壓縮，則用壓縮類解壓
    CompressionCodec codec = (new CompressionCodecFactory(job)).getCodec(file);
    //以壓縮方式讀取切片
    if (null != codec) {
        this.isCompressedInput = true;
        this.decompressor = CodecPool.getDecompressor(codec);
        if (codec instanceof SplittableCompressionCodec) {
            SplitCompressionInputStream cIn = ((SplittableCompressionCodec)codec).createInputStream(this.fileIn, this.decompressor, this.start, this.end, READ_MODE.BYBLOCK);
            this.in = new CompressedSplitLineReader(cIn, job, this.recordDelimiterBytes);
            this.start = cIn.getAdjustedStart();
            this.end = cIn.getAdjustedEnd();
            this.filePosition = cIn;
        } else {
            if (this.start != 0L) {
                throw new IOException("Cannot seek in " + codec.getClass().getSimpleName() + " compressed stream");
            }

            this.in = new SplitLineReader(codec.createInputStream(this.fileIn, this.decompressor), job, this.recordDelimiterBytes);
            this.filePosition = this.fileIn;
        }
    } else {
        //無壓縮方式讀取切片
        this.fileIn.seek(this.start);
        //這里很重要，是真正用于讀取數據的類
        this.in = new UncompressedSplitLineReader(this.fileIn, job, this.recordDelimiterBytes, split.getLength());
        this.filePosition = this.fileIn;
    }

    //對起始偏移量進行修正，并賦值給pos這個偏移量
    if (this.start != 0L) {
        this.start += (long)this.in.readLine(new Text(), 0, this.maxBytesToConsume(this.start));
    }

    this.pos = this.start;
}

這里的工作主要是給 RecordReader對象讀取文件做初始化工作。主要就是獲取切片的輸入流對象。
this.in 這里就用于后面讀取數據的對象，這里就是完成了這個輸入流對象的初始化。

6、接著我們回到3中，看mapper.run() 方法

這個其實就是寫的mapper 的run方法：

//------------------------Mapper.java   mapper.run(mapperContext);
public void run(Mapper<KEYIN, VALUEIN, KEYOUT, VALUEOUT>.Context context) throws IOException, InterruptedException {
    this.setup(context);

    try {
        //這里循環讀取key和value，給map方法處理
        //關鍵在于 context這個對象,從上面runNewApi中可以看到，是MapContextImpl類型的
        while(context.nextKeyValue()) {
            this.map(context.getCurrentKey(), context.getCurrentValue(), context);
        }
    } finally {
        this.cleanup(context);
    }
}

可以看到，這里是個while循環，通過context上下文對象獲取KV，然后傳入map方法中處理。

7、下面看看 context.nextKeyValue()

從3中可以看到，這個context是 MapContextImpl類型的，看看這個類

//-----------------------MapContextImpl.java..    
public class MapContextImpl<KEYIN, VALUEIN, KEYOUT, VALUEOUT> extends TaskInputOutputContextImpl<KEYIN, VALUEIN, KEYOUT, VALUEOUT> implements MapContext<KEYIN, VALUEIN, KEYOUT, VALUEOUT> {
    private RecordReader<KEYIN, VALUEIN> reader;
    private InputSplit split;

    //構造方法中包括獲取 RecordReader對象，以及split
    public MapContextImpl(Configuration conf, TaskAttemptID taskid, RecordReader<KEYIN, VALUEIN> reader, RecordWriter<KEYOUT, VALUEOUT> writer, OutputCommitter committer, StatusReporter reporter, InputSplit split) {
        super(conf, taskid, writer, committer, reporter);
        this.reader = reader;
        this.split = split;
    }

    public InputSplit getInputSplit() {
        return this.split;
    }

    //下面都是調用 RecordReader 中的get方法獲取key value
    public KEYIN getCurrentKey() throws IOException, InterruptedException {
        return this.reader.getCurrentKey();
    }

    public VALUEIN getCurrentValue() throws IOException, InterruptedException {
        return this.reader.getCurrentValue();
    }

    public boolean nextKeyValue() throws IOException, InterruptedException {
       //這里就是調用reader 的方法
        return this.reader.nextKeyValue();
    }
}

在它的構造方法中，主要從3中傳入了 split切片，以及 RecordReader對象。下面就是三個獲取KV的方法，也就是在 mapper.run() 中調用的方法。

下面看看 this.reader.nextKeyValue()

//----------------------------------LineRecordReader.java
public boolean nextKeyValue() throws IOException {
    if (this.key == null) {
        this.key = new LongWritable();
    }

    //設置key為偏移量
    this.key.set(this.pos);
    if (this.value == null) {
        this.value = new Text();
    }

    int newSize = 0;

    while(this.getFilePosition() <= this.end || this.in.needAdditionalRecordAfterSplit()) {
        if (this.pos == 0L) {
            newSize = this.skipUtfByteOrderMark();
        } else {
            /*讀取數據到value中。this.in是UncompressedSplitLineReader類型的，在LineRecordReader的initialize方法中初始化了。該類父類為LineReader。*/
            //調用 LineRreader 的readline 方法。讀一行數據
            newSize = this.in.readLine(this.value, this.maxLineLength, this.maxBytesToConsume(this.pos));
            this.pos += (long)newSize;
        }

        if (newSize == 0 || newSize < this.maxLineLength) {
            break;
        }

        LOG.info("Skipped line of size " + newSize + " at pos " + (this.pos - (long)newSize));
    }

    if (newSize == 0) {
        this.key = null;
        this.value = null;
        return false;
    } else {
        return true;
    }
}

可以看到，這里已經看到key和value的蹤影了。key就是數據偏移量，value就是通過readLine讀取的數據。如果有數據返回true，mapper.run() 通過getKey和getValue對應的KV。下面看看 this.in.readLine，也就是 LineReader.readLine()。

8、LineReader.readLine() 按行讀取的reader

//---------------------------LineReader.java
public int readLine(Text str, int maxLineLength, int maxBytesToConsume) throws IOException {
    return this.recordDelimiterBytes != null ? this.readCustomLine(str, maxLineLength, maxBytesToConsume) : this.readDefaultLine(str, maxLineLength, maxBytesToConsume);
}

private int readCustomLine(Text str, int maxLineLength, int maxBytesToConsume) throws IOException {
    str.clear();
    int txtLength = 0;
    long bytesConsumed = 0L;
    int delPosn = 0;
    int ambiguousByteCount = 0;

    do {
        int startPosn = this.bufferPosn;
        if (this.bufferPosn >= this.bufferLength) {
            startPosn = this.bufferPosn = 0;
            this.bufferLength = this.fillBuffer(this.in, this.buffer, ambiguousByteCount > 0);
            if (this.bufferLength <= 0) {
                if (ambiguousByteCount > 0) {
                    str.append(this.recordDelimiterBytes, 0, ambiguousByteCount);
                    bytesConsumed += (long)ambiguousByteCount;
                }
                break;
            }
        }

        for(; this.bufferPosn < this.bufferLength; ++this.bufferPosn) {
            if (this.buffer[this.bufferPosn] == this.recordDelimiterBytes[delPosn]) {
                ++delPosn;
                if (delPosn >= this.recordDelimiterBytes.length) {
                    ++this.bufferPosn;
                    break;
                }
            } else if (delPosn != 0) {
                this.bufferPosn -= delPosn;
                if (this.bufferPosn < -1) {
                    this.bufferPosn = -1;
                }

                delPosn = 0;
            }
        }

        int readLength = this.bufferPosn - startPosn;
        bytesConsumed += (long)readLength;
        int appendLength = readLength - delPosn;
        if (appendLength > maxLineLength - txtLength) {
            appendLength = maxLineLength - txtLength;
        }

        bytesConsumed += (long)ambiguousByteCount;
        if (appendLength >= 0 && ambiguousByteCount > 0) {
            //看到這里就很明顯了，將數據追加到 value中
            str.append(this.recordDelimiterBytes, 0, ambiguousByteCount);
            ambiguousByteCount = 0;
            this.unsetNeedAdditionalRecordAfterSplit();
        }

        if (appendLength > 0) {
            str.append(this.buffer, startPosn, appendLength);
            txtLength += appendLength;
        }

        if (this.bufferPosn >= this.bufferLength && delPosn > 0 && delPosn < this.recordDelimiterBytes.length) {
            ambiguousByteCount = delPosn;
            bytesConsumed -= (long)delPosn;
        }
    } while(delPosn < this.recordDelimiterBytes.length && bytesConsumed < (long)maxBytesToConsume);

    if (bytesConsumed > 2147483647L) {
        throw new IOException("Too many bytes before delimiter: " + bytesConsumed);
    } else {
        return (int)bytesConsumed;
    }
}

上面重要就是讀取數據的過程了，過程過于長，抓住關鍵的看，其實就是將讀取的一行數據追加到 this.value中。

9、總結

至此，map的整個輸入流程涉及到兩個重要的類
InputFormat -- 處理原始數據并切片；創建RecordReader 對象
RecordReader -- 讀取切片中的數據，處理成KV，傳遞KV給map方法處理

這兩個都是抽象類：

public abstract class RecordReader<KEYIN, VALUEIN> implements Closeable {
    public RecordReader() {
    }

    public abstract void initialize(InputSplit var1, TaskAttemptContext var2) throws IOException, InterruptedException;

    public abstract boolean nextKeyValue() throws IOException, InterruptedException;

    public abstract KEYIN getCurrentKey() throws IOException, InterruptedException;

    public abstract VALUEIN getCurrentValue() throws IOException, InterruptedException;

    public abstract float getProgress() throws IOException, InterruptedException;

    public abstract void close() throws IOException;
}

public abstract class InputFormat<K, V> {
    public InputFormat() {
    }

    public abstract List<InputSplit> getSplits(JobContext var1) throws IOException, InterruptedException;

    public abstract RecordReader<K, V> createRecordReader(InputSplit var1, TaskAttemptContext var2) throws IOException, InterruptedException;
}

當我們想自定義inputformat類和recordreader類時，就需要繼承這兩個類，并實現其中的方法。