HDFS写入文件的重要概念

               

HDFS一个文件由多个block构成。HDFS在进行block读写的时候是以packet(默认每个packet为64K)为单位进行的。每一个packet由若干个chunk(默认512Byte)组成。Chunk是进行数据校验的基本单位,对每一个chunk生成一个校验和(默认4Byte)并将校验和进行存储。

在写入一个block的时候,数据传输的基本单位是packet,每个packet由若干个chunk组成。

HDFS客户端写文件示例代码

复制代码
FileSystem hdfs = FileSystem.get(new Configuration());Path path = new Path("/testfile");// writingFSDataOutputStream dos = hdfs.create(path);byte[] readBuf = "Hello World".getBytes("UTF-8");dos.write(readBuf, 0, readBuf.length);dos.close();hdfs.close();
复制代码

文件的打开

上传一个文件到hdfs,一般会调用DistributedFileSystem.create,其实现如下:

public FSDataOutputStream create(Path f, FsPermission permission,boolean overwrite,int bufferSize, short replication, long blockSize,Progressable progress) throws IOException {    return new FSDataOutputStream       (dfs.create(getPathName(f), permission,overwrite, replication, blockSize, progress, bufferSize),        statistics);}

其最终生成一个FSDataOutputStream用于向新生成的文件中写入数据。其成员变量dfs的类型为DFSClient,DFSClient的create函数如下:

复制代码
public OutputStream create(String src,FsPermission permission,boolean overwrite,short replication,long blockSize,Progressable progress,int buffersize) throws IOException {    checkOpen();    if (permission == null) {      permission = FsPermission.getDefault();    }    FsPermission masked = permission.applyUMask(FsPermission.getUMask(conf));    OutputStream result = new DFSOutputStream(src, masked,overwrite, replication, blockSize, progress, buffersize,        conf.getInt("io.bytes.per.checksum", 512));    leasechecker.put(src, result);    return result;}
复制代码

其中构造了一个DFSOutputStream,在其构造函数中,同过RPC调用NameNode的create来创建一个文件。 
当然,构造函数中还做了一件重要的事情,就是streamer.start(),也即启动了一个pipeline,用于写数据,在写入数据的过程中,我们会仔细分析。

复制代码
DFSOutputStream(String src, FsPermission masked, boolean overwrite,short replication, long blockSize, Progressable progress,                int buffersize, int bytesPerChecksum) throws IOException {    this(src, blockSize, progress, bytesPerChecksum);    computePacketChunkSize(writePacketSize, bytesPerChecksum);    try {      namenode.create(src, masked, clientName, overwrite, replication, blockSize);    } catch(RemoteException re) {      throw re.unwrapRemoteException(AccessControlException.class,QuotaExceededException.class);    }    streamer.start();}
复制代码

 通过rpc调用NameNode的create函数,调用namesystem.startFile函数,其又调用startFileInternal函数,它创建一个新的文件,状态为under construction,没有任何data block与之对应。

dfsclient文件的写入

下面轮到客户端向新创建的文件中写入数据了,一般会使用FSDataOutputStream的write方法:

按照hdfs的设计,对block的数据写入使用的是pipeline的方式,也即将数据分成一个个的package,如果需要复制三分,分别写入DataNode 1, 2, 3,则会进行如下的过程:

  • 首先将package 1写入DataNode 1
  • 然后由DataNode 1负责将package 1写入DataNode 2,同时客户端可以将pacage 2写入DataNode 1
  • 然后DataNode 2负责将package 1写入DataNode 3, 同时客户端可以讲package 3写入DataNode 1,DataNode 1将package 2写入DataNode 2
  • 就这样将一个个package排着队的传递下去,直到所有的数据全部写入并复制完毕

FSDataOutputStream的write方法会调用DFSOutputStream的write方法,而DFSOutputStream继承自FSOutputSummer,所以实际上是调用FSOutputSummer的write方法,如下:

public synchronized void write(byte b[], int off, int len)  throws IOException {    //参数检查    for (int n=0;n<len;n+=write1(b, off+n, len-n)) {    }  }

FSOutputSummer的write1的方法如下:

复制代码
private int write1(byte b[], int off, int len) throws IOException {    if(count==0 && len>=buf.length) {      // buf初始化的大小是chunk的大小,默认是512,这里的代码会在写入的数据的剩余内容大于或等于一个chunk的大小时调用      // 这里避免多余一次复制      final int length = buf.length;      sum.update(b, off, length);//length是一个完整chunk的大小,默认是512,这里根据一个chunk内容计算校验和      writeChecksumChunk(b, off, length, false);      return length;    }        // buf初始化的大小是chunk的大小,默认是512,这里的代码会在写入的数据的剩余内容小于一个chunk的大小时调用    // 规避了数组越界问题    int bytesToCopy = buf.length-count;    bytesToCopy = (len<bytesToCopy) ? len : bytesToCopy;    sum.update(b, off, bytesToCopy);//bytesToCopy不足一个chunk,是写入的内容的最后一个chunk的剩余字节数目    System.arraycopy(b, off, buf, count, bytesToCopy);    count += bytesToCopy;    if (count == buf.length) {//如果不足一个chunk,就缓存到本地buffer,如果还有下一次写入,就填充这个chunk,满一个chunk再flush,count清0      // local buffer is full      flushBuffer();//最终调用writeChecksumChunk方法实现    }     return bytesToCopy;  }
复制代码

writeChecksumChunk的实现如下:

复制代码
//写入一个chunk的数据长度(默认512),忽略len的长度private void writeChecksumChunk(byte b[], int off, int len, boolean keep)  throws IOException {    int tempChecksum = (int)sum.getValue();    if (!keep) {      sum.reset();    }    int2byte(tempChecksum, checksum);//把当前chunk的校验和从int转换为字节    writeChunk(b, off, len, checksum);}
复制代码

writeChunk由子类DFSOutputStream实现,如下:

复制代码
 protected synchronized void writeChunk(byte[] b, int offset, int len, byte[] checksum)throws IOException {      //创建一个package,并写入数据      currentPacket = new Packet(packetSize, chunksPerPacket,bytesCurBlock);      currentPacket.writeChecksum(checksum, 0, cklen);      currentPacket.writeData(b, offset, len);      currentPacket.numChunks++;      bytesCurBlock += len;      //如果此package已满,则放入队列中准备发送      if (currentPacket.numChunks == currentPacket.maxChunks ||bytesCurBlock == blockSize) {          ......          dataQueue.addLast(currentPacket);          //唤醒等待dataqueue的传输线程,也即DataStreamer          dataQueue.notifyAll();          currentPacket = null;          ......      } }
复制代码

 writeChunk比较简单,就是把数据填充packet,填充完毕,就放到dataQueue,再唤醒DataStreamer。

DataStreamer完成了数据的传输,DataStreamer的run函数如下:

复制代码
  public void run() {    while (!closed && clientRunning) {      Packet one = null;      synchronized (dataQueue) {      boolean doSleep = processDatanodeError(hasError, false);//如果ack出错,则处理IO错误        //如果队列中没有package,则等待        while ((!closed && !hasError && clientRunning && dataQueue.size() == 0) || doSleep) {          try {            dataQueue.wait(1000);          } catch (InterruptedException  e) {          }          doSleep = false;        }        try {          //得到队列中的第一个package          one = dataQueue.getFirst();          long offsetInBlock = one.offsetInBlock;          //由NameNode分配block,并生成一个写入流指向此block          if (blockStream == null) {            nodes = nextBlockOutputStream(src);            response = new ResponseProcessor(nodes);            response.start();          }          ByteBuffer buf = one.getBuffer();          //将packet从dataQueue移至ackQueue,等待确认          dataQueue.removeFirst();          dataQueue.notifyAll();          synchronized (ackQueue) {            ackQueue.addLast(one);            ackQueue.notifyAll();          }          //利用生成的写入流将数据写入DataNode中的block          blockStream.write(buf.array(), buf.position(), buf.remaining());          if (one.lastPacketInBlock) {            blockStream.writeInt(0); //表示此block写入完毕          }          blockStream.flush();        } catch (Throwable e) {        }                if (one.lastPacketInBlock) {            //数据块写满,做一些清理工作,下次再申请块            response.close();        // ignore all errors in Response                        synchronized (dataQueue) {              IOUtils.cleanup(LOG, blockStream, blockReplyStream);              nodes = null;              response = null;              blockStream = null;//设置为null,下次就会判断blockStream为null,申请新的块              blockReplyStream = null;            }        }    }      ......  }
复制代码

DataStreamer线程负责把准备好的数据packet,顺序写入到DataNode,未确认写入成功的packet则移动到ackQueue,等待确认。

DataStreamer线程传输数据到DataNode时,要向namenode申请数据块,方法是nextBlockOutputStream,再调用locateFollowingBlock,通过RPC调用namenode.addBlock(src, clientName),在NameNode分配了DataNode和block以后,createBlockOutputStream开始写入数据。

客户端在DataStreamer的run函数中创建了写入流后,调用blockStream.write将packet写入DataNode

DataStreamer还会启动ResponseProcessor线程,它负责接收datanode的ack,当接收到所有datanode对一个packet确认成功的ack,ResponseProcessor从ackQueue中删除相应的packet。在出错时,从ackQueue中移除packet到dataQueue,移除失败的datanode,恢复数据块,建立新的pipeline。实现如下:

复制代码
public void run() {...PipelineAck ack = new PipelineAck();while (!closed && clientRunning && !lastPacketInBlock) {  try {    // read an ack from the pipeline    ack.readFields(blockReplyStream);    ...    //处理所有DataNode响应的状态    for (int i = ack.getNumOfReplies()-1; i >=0 && clientRunning; i--) {        short reply = ack.getReply(i);        if (reply != DataTransferProtocol.OP_STATUS_SUCCESS) {//ack验证,如果DataNode写入packet失败,则出错            errorIndex = i; //记录损坏的DataNode,会在processDatanodeError方法移除该失败的DataNode        throw new IOException("Bad response " + reply + " for block " + block +  " from datanode " + targets[i].getName());          }       }    long seqno = ack.getSeqno();    if (seqno == Packet.HEART_BEAT_SEQNO) {  // 心跳ack,忽略      continue;    }    Packet one = null;    synchronized (ackQueue) {      one = ackQueue.getFirst();    }    ...    synchronized (ackQueue) {      assert ack.getSeqno() == lastAckedSeqno + 1;//验证ack      lastAckedSeqno = ack.getSeqno();      ackQueue.removeFirst();//移除确认写入成功的packet      ackQueue.notifyAll();    }  } catch (Exception e) {    if (!closed) {      hasError = true;//设置ack错误,让      ...      closed = true;    }  }}}
复制代码

当ResponseProcessor在确认packet失败时,processDatanodeError方法用于处理datanode的错误,当调用返回后需要休眠一段时间时,返回true。下面是其简单的处理流程:

1.关闭blockStream和blockReplyStream
2.将packet从ackQueue移到dataQueue
3.删除坏datanode
4.通过RPC调用datanode的recoverBlock方法来恢复块,如果有错,返回true
5.如果没有可用的datanode,关闭DFSOutputStream和streamer,返回false
6.创建块输出流,如果不成功,转到3

实现如下:

复制代码
private boolean processDatanodeError(boolean hasError, boolean isAppend) {  if (!hasError) {//DataNode没有发生错误,直接返回    return false;  }    //将未确认写入成功的packets从ack queue移动到data queue的前面  synchronized (ackQueue) {    dataQueue.addAll(0, ackQueue);    ackQueue.clear();  }  boolean success = falsewhile (!success && clientRunning) {    DatanodeInfo[] newnodes = null;        //根据errorIndex确定失败的DataNode,从所有的DataNode nodes移除失败的DataNode,复制到newnodes    // 通知primary datanode做数据块恢复,更新合适的时间戳    LocatedBlock newBlock = null;    ClientDatanodeProtocol primary null;    DatanodeInfo primaryNode = null;    try {      // Pick the "least" datanode as the primary datanode to avoid deadlock.      primaryNode = Collections.min(Arrays.asList(newnodes));      primary = createClientDatanodeProtocolProxy(primaryNode, conf, block, accessToken, socketTimeout);      newBlock = primary.recoverBlock(block, isAppend, newnodes);//恢复数据块    } catch (IOException e) {        //循环创建块输出流,如果不成功,移除失败的DataNode          return true;          // 需要休眠    } finally {      RPC.stopProxy(primary);    }    recoveryErrorCount = 0; // 数据块恢复成功    block = newBlock.getBlock();    accessToken = newBlock.getBlockToken();    nodes = newBlock.getLocations();    this.hasError = false;    lastException = null;    errorIndex = 0;    success = createBlockOutputStream(nodes, clientName, true);  }  response = new ResponseProcessor(nodes);  response.start();//启动ResponseProcessor做ack确认处理  return false; // 不休眠,继续处理}
复制代码

总结 

hdfs文件的写入是比较复杂的,所以本文重点介绍了dfsclient端的处理逻辑,对namenode和datanode的响应,就不做详细分析了。

更多参考

      http://www.cnblogs.com/forfuture1978/archive/2010/11/10/1874222.html (HDFS读写过程解析)

      http://blog.jeoygin.org/2012/07/hdfs-source-analysis-hdfs-input-output-stream.html (讲解dfsclient的重要类的职责)

      http://caibinbupt.iteye.com/blog/286259 (datanode对于块写入的处理)

<script type="text/javascript" src="http://pagead2.googlesyndication.com/pagead/show_ads.js"></script>

           

再分享一下我老师大神的人工智能教程吧。零基础!通俗易懂!风趣幽默!还带黄段子!希望你也加入到我们人工智能的队伍中来!https://blog.csdn.net/jiangjunshow

猜你喜欢

转载自blog.csdn.net/ugfdfgg/article/details/86503688