深度探索Hadoop分布式文件系统（HDFS）数据读取流程

{"type":"doc","content":[{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"1. 开篇","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"Hadoop分布式文件系统（HDFS）是Hadoop大数据生态最底层的数据存储设施。因其具备了海量数据分布式存储能力，针对不同批处理业务的大吞吐数据计算承载力，使其综合复杂度要远远高于其他数据存储系统。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"因此对Hadoop分布式文件系统（HDFS）的深入研究，了解其架构特征、读写流程、分区模式、高可用思想、数据存储规划等知识，对学习大数据技术大有裨益，尤其是面临开发生产环境时，能做到胸中有数。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"本文重点从客户端读取HDFS数据的角度切入，通过Hadoop源代码跟踪手段，层层拨开，渐渐深入Hadoop机制内部，使其读取流程逐渐明朗化。","attrs":{}}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"2. HDFS数据读取整体架构流程","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https://static001.geekbang.org/infoq/a2/a29439d647f693a701769d73f0d1e98a.png","alt":null,"title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":"center","origin":null},"content":[{"type":"text","text":" HDFS数据访问整体架构流程","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"如上图所示：描绘了客户端访问HDFS数据的简化后整体架构流程。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"（1） 客户端向hdfs namenode节点发送Path文件路径的数据访问的请求","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"（2） Namenode会根据文件路径收集所有数据块（block）的位置信息，并根据数据块在文件中的先后顺序，按次序组成数据块定位集合（located blocks），回应给客户端","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"（3） 客户端拿到数据块定位集合后，创建HDFS输入流，定位第一个数据块所在的位置，并读取datanode的数据流。之后根据读取偏移量定位下一个datanode并创建新的数据块读取数据流，以此类推，完成对HDFS文件的整个读取。","attrs":{}}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"3. Hadoop源代码分析","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"经过上述简单描述，我们对客户端读取HDFS文件数据有了一个整体上概念，那么这一节，我们开始从源代码跟踪的方向，深度去分析一下HDFS的数据访问内部机制。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong","attrs":{}}],"text":"(一) namenode代理类生成的源代码探索","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"为什么我们要先从namenode代理生成说起呢？原因就是先了解清楚客户端与namenode之间的来龙去脉，再看之后的数据获取过程就有头绪了。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"（1） 首先我们先从一个hdfs-site.xml配置看起","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"codeblock","attrs":{"lang":"html"},"content":[{"type":"text","text":"\n \n dfs.client.failover.proxy.provider.fszx\n org.apache.hadoop.hdfs.server.namenode.ha.\n ConfiguredFailoverProxyProvider\n","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"配置中定义了namenode代理的提供者为ConfiguredFailoverProxyProvider。什么叫namenode代理？其实本质上就是连接namenode服务的客户端网络通讯对象，用于客户端和namenode服务端的交流。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"（2） 接着我们看看ConfiguredFailoverProxyProvider的源代码继承关系结构","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https://static001.geekbang.org/infoq/02/02946eeea933fbff9bc896ba1b7566c1.png","alt":null,"title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":"center","origin":null},"content":[{"type":"text","text":"ConfiguredFailoverProxyProvider继承关系图","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"上图是ConfiguredFailoverProxyProvider的继承关系，顶端接口是FailoverProxyProvider，它包含了一段代码:","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"codeblock","attrs":{"lang":"java"},"content":[{"type":"text","text":" /**\n * Get the proxy object which should be used until the next failover event\n * occurs.\n * @return the proxy object to invoke methods upon\n */\n public ProxyInfo getProxy();","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"这个方法返回的ProxyInfo就是namenode代理对象，当然客户端获取的ProxyInfo整个过程非常复杂，甚至还用到了动态代理，但本质上就是通过此接口拿到了namenode代理。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"（3） 此时类关系演化成如下图所示：","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https://static001.geekbang.org/infoq/1c/1cfd58ea29192a69ff092a46a1207c3f.png","alt":null,"title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":"center","origin":null},"content":[{"type":"text","text":"namonode创建过程类关系图","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"上图ProxyInfo就是namenode的代理类，继承的子类NNProxyInfo就是具体指定是高可用代理类。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"（4） 那么费了这么大劲搞清楚的namenode代理，它的作用在哪里呢？","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"这就需要关注一个极为重要的对象DFSClient了，它是所有客户端向HDFS发起输入输出流的起点，如下图所示：","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https://static001.geekbang.org/infoq/6a/6aa0c2ef2c01321a146e7515707a443a.png","alt":null,"title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":"center","origin":null},"content":[{"type":"text","text":"DFSClient初始化过程类关系图","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"上图实线代表了真实的调用过程，虚线代表了对象之间的间接关系。我们可以看到DFSClient是一个关键角色，它由分布式文件系统对象(DistributeFileSystem)初始化，并在初始化中调用NameNodeProxiesClient等一系列操作，实现了高可用NNproxyInfo对象创建，也就是namenode代理，并最终作为DFSClient对象的一个成员，在创建数据流等过程中使用。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong","attrs":{}}],"text":"(二) 读取文件流的深入源代码探索","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"（1） 首先方法一样，先找一个切入口。建立从HDFS下载文件到本地的一个简单场景，以下是代码片段：","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"codeblock","attrs":{"lang":"java"},"content":[{"type":"text","text":"……\n//打开HDFS文件输入流\ninput = fileSystem.open(new Path(hdfs_file_path));\n//创建本地文件输出流\noutput = new FileOutputStream(local_file_path);\n//通过IOUtils工具实现数据流字节循环复制\nIOUtils.copyBytes(input, output, 4096, true);\n……","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"咱们再看看IOUtils的一段文件流读写的方法代码：","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"codeblock","attrs":{"lang":"java"},"content":[{"type":"text","text":"/**\n * Copies from one stream to another.\n * \n * @param in InputStrem to read from\n * @param out OutputStream to write to\n * @param buffSize the size of the buffer \n */\n public static void copyBytes(InputStream in, OutputStream out, int buffSize) \n throws IOException {\n PrintStream ps = out instanceof PrintStream ? (PrintStream)out : null;\n byte buf[] = new byte[buffSize];\n int bytesRead = in.read(buf);\n while (bytesRead >= 0) {\n out.write(buf, 0, bytesRead);\n if ((ps != null) && ps.checkError()) {\n throw new IOException(\"Unable to write to output stream.\");\n }\n bytesRead = in.read(buf);\n }\n }","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"这段代码是个标准的循环读取HDFS InputStream数据流，然后向本地文件OutputStream输出流写数据的过程。我们的目标是深入到HDFS InputStream数据流的创建和使用过程。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"（2） 接下来我们开始分析InputStream的产生过程，如下图所示：","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https://static001.geekbang.org/infoq/ec/ec8a36797060fa31ba13f7fc985e9d40.png","alt":null,"title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":"center","origin":null},"content":[{"type":"text","text":"InputStream打开流程类关系图","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"上图实线代表了真实的调用过程，虚线代表了对象之间的间接关系。其代码内部结构极为复杂，我用此图用最简化的方式让我们能快速的理解清楚他的原理。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"我来简单讲解一下这个过程：","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"第一步是DistributeFileSystem通过调用DFSClient对象的open方法，实现对DFSInputStream对象的创建，DFSInputStream对象是真正读取数据块（LocationBlock）以及与datanode交互的实现逻辑，是真正的核心类。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"第二步，DFSClient在创建DFSInputStream的过程中，需要为其传入调用namenode代理而返回的数据块集合（LocationBlocks）。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"第三步，DFSClient创建一个装饰器类HDFSDataInputStream，封装了DFSInputStream，由装饰器的父类FSDataInputStream最终返回给DistributeFileSystem，由客户端开发者使用。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"（3） 最后我们再深入到数据块读取机制的源代码上看看，如下图所示：","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https://static001.geekbang.org/infoq/c1/c1e86a454da4a48215a319d3523b4486.png","alt":null,"title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":"center","origin":null},"content":[{"type":"text","text":"DFSInputStream数据读取流程类关系图","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"上图实线代表了真实的调用过程，虚线代表了对象之间的间接关系。实际的代码逻辑比较复杂，此图也是尽量简化展现，方便我们理解。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"一样的，我来简单讲解一下这个过程：","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"第一步FSDataInputStream装饰器接受客户端的读取调用对DFSInputStream对象进行read(...)方法调用。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"第二步 DFSInputStream会调用自身的blockSeekTo(long offset)方法，一方面根据offset数据偏移量，定位当前是否要读取新的数据块(LocationBlock)，另一方面新的数据块从数据块集合(LocationBlocks)中找到后，寻找最佳的数据节点，也就是Hadoop所谓的就近原则，先看看本机数据节点有没有副本，再次根据网络距离着就近获取副本。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"第三步通过FSDataInputStream副本上数据块（LocationBlock）构建BlockReader对象，它就是真正读取数据块的对象。BlockReader对象它有不同的实现，由BlockReaderFactory.build根据条件最优选择具体实现，BlockReaderLocal和BlockReaderLocalLegacy（based on HDFS-2246）是优选方案，也是short-circuit block readers方案，相当于直接从本地文件系统读数据了，若short-circuit因为安全等因素不可用，就会尝试UNIX domain sockets的优化方案，再不行才考虑BlockReaderRemote建立TCP sockets的连接方案了。BlockReader的细节原理也非常值得深入一探究竟，待下次我专门写一篇针对BlockReader原理机制文章。","attrs":{}}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"4. 结束","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"非常感觉您能看完。下一篇我会对“","attrs":{}},{"type":"link","attrs":{"href":"https://xie.infoq.cn/article/4438a7b40a9d832e9eb7c4e67","title":""},"content":[{"type":"text","text":"Hadoop分布式文件系统（HDFS）数据写入流程","attrs":{}}]},{"type":"text","text":"”做一篇深度探索分析。期盼您的关注。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong","attrs":{}}],"text":"作者：方顺 西安守护石信息科技创始人 致力于IT工程师在大数据领域的技术提升","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"link","attrs":{"href":"https://www.zhihu.com/column/c_151487501","title":null},"content":[{"type":"text","text":"前往我的知乎专栏——了解更多关于大数据的知识","attrs":{}}]},{"type":"text","marks":[{"type":"strong","attrs":{}}],"text":" ","attrs":{}}]},{"type":"image","attrs":{"src":"https://static001.geekbang.org/infoq/c1/c18eee3b3b896897a333305bc70cfc88.jpeg","alt":null,"title":"公众号：守护石论数据","style":[{"key":"width","value":"25%"},{"key":"bordertype","value":"none"}],"href":"","fromPaste":false,"pastePass":false}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}}]}