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
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();
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:
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;}
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。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();}
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
FSOutputSummer的write1的方法如下:
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
writeChecksumChunk的实现如下:
//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);}
writeChunk由子类DFSOutputStream实现,如下:
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; ...... } }
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:
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; } } } ...... }
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:
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; } }}}
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dataQueue3.删除坏datanode4.通过RPC调用datanode的recoverBlock方法来恢复块,如果有错,返回true5.如果没有可用的datanode,关闭DFSOutputStream和streamer,返回false6.创建块输出流,如果不成功,转到3实现如下:
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 = false; while (!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}
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)
原文:http://www.cnblogs.com/ggjucheng/