经验首页 前端设计 程序设计 Java相关 移动开发 数据库/运维 软件/图像 大数据/云计算 其他经验
当前位置:技术经验 » 大数据/云/AI » Flink » 查看文章
Flink源码分析 - 剖析一个简单的Flink程序
来源:cnblogs  作者:chaojianok  时间:2019/6/6 8:34:01  对本文有异议

本篇文章首发于头条号Flink程序是如何执行的?通过源码来剖析一个简单的Flink程序,欢迎关注头条号和微信公众号“大数据技术和人工智能”(微信搜索bigdata_ai_tech)获取更多干货,也欢迎关注我的CSDN博客

在这之前已经介绍了如何在本地搭建Flink环境和如何创建Flink应用如何构建Flink源码,这篇文章用官方提供的SocketWindowWordCount例子来解析一下一个常规Flink程序的每一个基本步骤。

示例程序

  1. public class SocketWindowWordCount {
  2. public static void main(String[] args) throws Exception {
  3. // the host and the port to connect to
  4. final String hostname;
  5. final int port;
  6. try {
  7. final ParameterTool params = ParameterTool.fromArgs(args);
  8. hostname = params.has("hostname") ? params.get("hostname") : "localhost";
  9. port = params.getInt("port");
  10. } catch (Exception e) {
  11. System.err.println("No port specified. Please run 'SocketWindowWordCount " +
  12. "--hostname <hostname> --port <port>', where hostname (localhost by default) " +
  13. "and port is the address of the text server");
  14. System.err.println("To start a simple text server, run 'netcat -l <port>' and " +
  15. "type the input text into the command line");
  16. return;
  17. }
  18. // get the execution environment
  19. final StreamExecutionEnvironment env = StreamExecutionEnvironment.getExecutionEnvironment();
  20. // get input data by connecting to the socket
  21. DataStream<String> text = env.socketTextStream(hostname, port, "\n");
  22. // parse the data, group it, window it, and aggregate the counts
  23. DataStream<WordWithCount> windowCounts = text
  24. .flatMap(new FlatMapFunction<String, WordWithCount>() {
  25. @Override
  26. public void flatMap(String value, Collector<WordWithCount> out) {
  27. for (String word : value.split("\\s")) {
  28. out.collect(new WordWithCount(word, 1L));
  29. }
  30. }
  31. })
  32. .keyBy("word")
  33. .timeWindow(Time.seconds(5))
  34. .reduce(new ReduceFunction<WordWithCount>() {
  35. @Override
  36. public WordWithCount reduce(WordWithCount a, WordWithCount b) {
  37. return new WordWithCount(a.word, a.count + b.count);
  38. }
  39. });
  40. // print the results with a single thread, rather than in parallel
  41. windowCounts.print().setParallelism(1);
  42. env.execute("Socket Window WordCount");
  43. }
  44. // ------------------------------------------------------------------------
  45. /**
  46. * Data type for words with count.
  47. */
  48. public static class WordWithCount {
  49. public String word;
  50. public long count;
  51. public WordWithCount() {}
  52. public WordWithCount(String word, long count) {
  53. this.word = word;
  54. this.count = count;
  55. }
  56. @Override
  57. public String toString() {
  58. return word + " : " + count;
  59. }
  60. }
  61. }

上面这个是官网的SocketWindowWordCount程序示例,它首先从命令行中获取socket连接的host和port,然后获取执行环境、从socket连接中读取数据、解析和转换数据,最后输出结果数据。
每个Flink程序都包含以下几个相同的基本部分:

  1. 获得一个execution environment,
  2. 加载/创建初始数据,
  3. 指定此数据的转换,
  4. 指定放置计算结果的位置,
  5. 触发程序执行

Flink执行环境

  1. final StreamExecutionEnvironment env = StreamExecutionEnvironment.getExecutionEnvironment();

Flink程序都是从这句代码开始,这行代码会返回一个执行环境,表示当前执行程序的上下文。如果程序是独立调用的,则此方法返回一个由createLocalEnvironment()创建的本地执行环境LocalStreamEnvironment。从其源码里可以看出来:

  1. //代码目录:org/apache/flink/streaming/api/environment/StreamExecutionEnvironment.java
  2. public static StreamExecutionEnvironment getExecutionEnvironment() {
  3. if (contextEnvironmentFactory != null) {
  4. return contextEnvironmentFactory.createExecutionEnvironment();
  5. }
  6. ExecutionEnvironment env = ExecutionEnvironment.getExecutionEnvironment();
  7. if (env instanceof ContextEnvironment) {
  8. return new StreamContextEnvironment((ContextEnvironment) env);
  9. } else if (env instanceof OptimizerPlanEnvironment || env instanceof PreviewPlanEnvironment) {
  10. return new StreamPlanEnvironment(env);
  11. } else {
  12. return createLocalEnvironment();
  13. }
  14. }

获取输入数据

  1. DataStream<String> text = env.socketTextStream(hostname, port, "\n");

这个例子里的源数据来自于socket,这里会根据指定的socket配置创建socket连接,然后创建一个新数据流,包含从套接字无限接收的字符串,接收的字符串由系统的默认字符集解码。当socket连接关闭时,数据读取会立即终止。通过查看源码可以发现,这里实际上是通过指定的socket配置来构造一个SocketTextStreamFunction实例,然后源源不断的从socket连接里读取输入的数据创建数据流。

  1. //代码目录:org/apache/flink/streaming/api/environment/StreamExecutionEnvironment.java
  2. @PublicEvolving
  3. public DataStreamSource<String> socketTextStream(String hostname, int port, String delimiter, long maxRetry) {
  4. return addSource(new SocketTextStreamFunction(hostname, port, delimiter, maxRetry),
  5. "Socket Stream");
  6. }

SocketTextStreamFunction的类继承关系如下:
SocketTextStreamFunction类关系图

可以看出SocketTextStreamFunctionSourceFunction的子类,SourceFunction是Flink中所有流数据源的基本接口。SourceFunction的定义如下:

  1. //代码目录:org/apache/flink/streaming/api/functions/source/SourceFunction.java
  2. @Public
  3. public interface SourceFunction<T> extends Function, Serializable {
  4. void run(SourceContext<T> ctx) throws Exception;
  5. void cancel();
  6. @Public
  7. interface SourceContext<T> {
  8. void collect(T element);
  9. @PublicEvolving
  10. void collectWithTimestamp(T element, long timestamp);
  11. @PublicEvolving
  12. void emitWatermark(Watermark mark);
  13. @PublicEvolving
  14. void markAsTemporarilyIdle();
  15. Object getCheckpointLock();
  16. void close();
  17. }
  18. }

SourceFunction定义了runcancel两个方法和SourceContext内部接口。

  • run(SourceContex):实现数据获取逻辑,并可以通过传入的参数ctx进行向下游节点的数据转发。
  • cancel():用来取消数据源,一般在run方法中,会存在一个循环来持续产生数据,cancel方法则可以使该循环终止。
  • SourceContext:source函数用于发出元素和可能的watermark的接口,返回source生成的元素的类型。

了解了SourceFunction这个接口,再来看下SocketTextStreamFunction的具体实现(主要是run方法),逻辑就已经很清晰了,就是从指定的hostname和port持续不断的读取数据,按回车换行分隔符划分成一个个字符串,然后再将数据转发到下游。现在回到StreamExecutionEnvironmentsocketTextStream方法,它通过调用addSource返回一个DataStreamSource实例。思考一下,例子里的text变量是DataStream类型,为什么源码里的返回类型却是DataStreamSource呢?这是因为DataStreamDataStreamSource的父类,下面的类关系图可以看出来,这也体现出了Java的多态的特性。
DataStreamSource类关系图

数据流操作

对上面取到的DataStreamSource,进行flatMapkeyBytimeWindowreduce转换操作。

  1. DataStream<WordWithCount> windowCounts = text
  2. .flatMap(new FlatMapFunction<String, WordWithCount>() {
  3. @Override
  4. public void flatMap(String value, Collector<WordWithCount> out) {
  5. for (String word : value.split("\\s")) {
  6. out.collect(new WordWithCount(word, 1L));
  7. }
  8. }
  9. })
  10. .keyBy("word")
  11. .timeWindow(Time.seconds(5))
  12. .reduce(new ReduceFunction<WordWithCount>() {
  13. @Override
  14. public WordWithCount reduce(WordWithCount a, WordWithCount b) {
  15. return new WordWithCount(a.word, a.count + b.count);
  16. }
  17. });

这段逻辑中,对上面取到的DataStreamSource数据流分别做了flatMapkeyBytimeWindowreduce四个转换操作,下面说一下flatMap转换,其他三个转换操作读者可以试着自己查看源码理解一下。

先看一下flatMap方法的源码吧,如下。

  1. //代码目录:org/apache/flink/streaming/api/datastream/DataStream.java
  2. public <R> SingleOutputStreamOperator<R> flatMap(FlatMapFunction<T, R> flatMapper) {
  3. TypeInformation<R> outType = TypeExtractor.getFlatMapReturnTypes(clean(flatMapper),
  4. getType(), Utils.getCallLocationName(), true);
  5. return transform("Flat Map", outType, new StreamFlatMap<>(clean(flatMapper)));
  6. }

这里面做了两件事,一是用反射拿到了flatMap算子的输出类型,二是生成了一个operator。flink流式计算的核心概念就是将数据从输入流一个个传递给operator进行链式处理,最后交给输出流的过程。对数据的每一次处理在逻辑上成为一个operator。上面代码中的最后一行transform方法的作用是返回一个SingleOutputStreamOperator,它继承了Datastream类并且定义了一些辅助方法,方便对流的操作。在返回之前,transform方法还把它注册到了执行环境中。下面这张图是一个由Flink程序映射为Streaming Dataflow的示意图:
Flink基本编程模型

结果输出

  1. windowCounts.print().setParallelism(1);

每个Flink程序都是以source开始以sink结尾,这里的print方法就是把计算出来的结果sink标准输出流。在实际开发中,一般会通过官网提供的各种Connectors或者自定义的Connectors把计算好的结果数据sink到指定的地方,比如Kafka、HBase、FileSystem、Elasticsearch等等。这里的setParallelism是设置此接收器的并行度的,值必须大于零。

执行程序

  1. env.execute("Socket Window WordCount");

Flink有远程模式和本地模式两种执行模式,这两种模式有一点不同,这里按本地模式来解析。先看下execute方法的源码,如下:

  1. //代码目录:org/apache/flink/streaming/api/environment/LocalStreamEnvironment.java
  2. @Override
  3. public JobExecutionResult execute(String jobName) throws Exception {
  4. // transform the streaming program into a JobGraph
  5. StreamGraph streamGraph = getStreamGraph();
  6. streamGraph.setJobName(jobName);
  7. JobGraph jobGraph = streamGraph.getJobGraph();
  8. jobGraph.setAllowQueuedScheduling(true);
  9. Configuration configuration = new Configuration();
  10. configuration.addAll(jobGraph.getJobConfiguration());
  11. configuration.setString(TaskManagerOptions.MANAGED_MEMORY_SIZE, "0");
  12. // add (and override) the settings with what the user defined
  13. configuration.addAll(this.configuration);
  14. if (!configuration.contains(RestOptions.BIND_PORT)) {
  15. configuration.setString(RestOptions.BIND_PORT, "0");
  16. }
  17. int numSlotsPerTaskManager = configuration.getInteger(TaskManagerOptions.NUM_TASK_SLOTS, jobGraph.getMaximumParallelism());
  18. MiniClusterConfiguration cfg = new MiniClusterConfiguration.Builder()
  19. .setConfiguration(configuration)
  20. .setNumSlotsPerTaskManager(numSlotsPerTaskManager)
  21. .build();
  22. if (LOG.isInfoEnabled()) {
  23. LOG.info("Running job on local embedded Flink mini cluster");
  24. }
  25. MiniCluster miniCluster = new MiniCluster(cfg);
  26. try {
  27. miniCluster.start();
  28. configuration.setInteger(RestOptions.PORT, miniCluster.getRestAddress().get().getPort());
  29. return miniCluster.executeJobBlocking(jobGraph);
  30. }
  31. finally {
  32. transformations.clear();
  33. miniCluster.close();
  34. }
  35. }

这个方法包含三部分:将流程序转换为JobGraph、使用用户定义的内容添加(或覆盖)设置、启动一个miniCluster并执行任务。关于JobGraph暂先不讲,这里就只说一下执行任务,跟进下return miniCluster.executeJobBlocking(jobGraph);这行的源码,如下:

  1. //代码目录:org/apache/flink/runtime/minicluster/MiniCluster.java
  2. @Override
  3. public JobExecutionResult executeJobBlocking(JobGraph job) throws JobExecutionException, InterruptedException {
  4. checkNotNull(job, "job is null");
  5. final CompletableFuture<JobSubmissionResult> submissionFuture = submitJob(job);
  6. final CompletableFuture<JobResult> jobResultFuture = submissionFuture.thenCompose(
  7. (JobSubmissionResult ignored) -> requestJobResult(job.getJobID()));
  8. final JobResult jobResult;
  9. try {
  10. jobResult = jobResultFuture.get();
  11. } catch (ExecutionException e) {
  12. throw new JobExecutionException(job.getJobID(), "Could not retrieve JobResult.", ExceptionUtils.stripExecutionException(e);
  13. }
  14. try {
  15. return jobResult.toJobExecutionResult(Thread.currentThread().getContextClassLoader());
  16. } catch (IOException | ClassNotFoundException e) {
  17. throw new JobExecutionException(job.getJobID(), e);
  18. }
  19. }

这段代码的核心逻辑就是final CompletableFuture<JobSubmissionResult> submissionFuture = submitJob(job);,调用了MiniCluster类的submitJob方法,接着看这个方法:

  1. //代码目录:org/apache/flink/runtime/minicluster/MiniCluster.java
  2. public CompletableFuture<JobSubmissionResult> submitJob(JobGraph jobGraph) {
  3. final CompletableFuture<DispatcherGateway> dispatcherGatewayFuture = getDispatcherGatewayFuture();
  4. // we have to allow queued scheduling in Flip-6 mode because we need to request slots
  5. // from the ResourceManager
  6. jobGraph.setAllowQueuedScheduling(true);
  7. final CompletableFuture<InetSocketAddress> blobServerAddressFuture = createBlobServerAddress(dispatcherGatewayFuture);
  8. final CompletableFuture<Void> jarUploadFuture = uploadAndSetJobFiles(blobServerAddressFuture, jobGraph);
  9. final CompletableFuture<Acknowledge> acknowledgeCompletableFuture = jarUploadFuture
  10. .thenCombine(
  11. dispatcherGatewayFuture,
  12. (Void ack, DispatcherGateway dispatcherGateway) -> dispatcherGateway.submitJob(jobGraph, rpcTimeout))
  13. .thenCompose(Function.identity());
  14. return acknowledgeCompletableFuture.thenApply(
  15. (Acknowledge ignored) -> new JobSubmissionResult(jobGraph.getJobID()));
  16. }

这里的Dispatcher组件负责接收作业提交,持久化它们,生成JobManagers来执行作业并在主机故障时恢复它们。Dispatcher有两个实现,在本地环境下启动的是MiniDispatcher,在集群环境上启动的是StandaloneDispatcher。下面是类结构图:
MiniDispatcher类结构图

这里的Dispatcher启动了一个JobManagerRunner,委托JobManagerRunner去启动该Job的JobMaster。对应的代码如下:

  1. //代码目录:org/apache/flink/runtime/jobmaster/JobManagerRunner.java
  2. private CompletableFuture<Void> verifyJobSchedulingStatusAndStartJobManager(UUID leaderSessionId) {
  3. final CompletableFuture<JobSchedulingStatus> jobSchedulingStatusFuture = getJobSchedulingStatus();
  4. return jobSchedulingStatusFuture.thenCompose(
  5. jobSchedulingStatus -> {
  6. if (jobSchedulingStatus == JobSchedulingStatus.DONE) {
  7. return jobAlreadyDone();
  8. } else {
  9. return startJobMaster(leaderSessionId);
  10. }
  11. });
  12. }

JobMaster经过一系列方法嵌套调用之后,最终执行到下面这段逻辑:

  1. //代码目录:org/apache/flink/runtime/jobmaster/JobMaster.java
  2. private void scheduleExecutionGraph() {
  3. checkState(jobStatusListener == null);
  4. // register self as job status change listener
  5. jobStatusListener = new JobManagerJobStatusListener();
  6. executionGraph.registerJobStatusListener(jobStatusListener);
  7. try {
  8. executionGraph.scheduleForExecution();
  9. }
  10. catch (Throwable t) {
  11. executionGraph.failGlobal(t);
  12. }
  13. }

这里executionGraph.scheduleForExecution();调用了ExecutionGraph的启动方法。在Flink的图结构中,ExecutionGraph是真正被执行的地方,所以到这里为止,一个任务从提交到真正执行的流程就结束了,下面再回顾一下本地环境下的执行流程:

  1. 客户端执行execute方法;
  2. MiniCluster完成了大部分任务后把任务直接委派给MiniDispatcher
  3. Dispatcher接收job之后,会实例化一个JobManagerRunner,然后用这个实例启动job;
  4. JobManagerRunner接下来把job交给JobMaster去处理;
  5. JobMaster使用ExecutionGraph的方法启动整个执行图,整个任务就启动起来了。

原文链接:http://www.cnblogs.com/cjblogs/p/10978975.html

 友情链接:直通硅谷  点职佳  北美留学生论坛

本站QQ群:前端 618073944 | Java 606181507 | Python 626812652 | C/C++ 612253063 | 微信 634508462 | 苹果 692586424 | C#/.net 182808419 | PHP 305140648 | 运维 608723728

W3xue 的所有内容仅供测试,对任何法律问题及风险不承担任何责任。通过使用本站内容随之而来的风险与本站无关。
关于我们  |  意见建议  |  捐助我们  |  报错有奖  |  广告合作、友情链接(目前9元/月)请联系QQ:27243702 沸活量
皖ICP备17017327号-2 皖公网安备34020702000426号