在学习Spark过程中,资料中介绍的提交Spark Job的方式主要有两种(我所知道的):
第一种:
通过命令行的方式提交Job,使用spark 自带的spark-submit工具提交,官网和大多数参考资料都是已这种方式提交的,提交命令示例如下:
./spark-submit –class com.learn.spark.SimpleApp –master yarn –deploy-mode client –driver-memory 2g –executor-memory 2g –executor-cores 3 ../spark-demo.jar
参数含义就不解释了,请参考官网资料。
第二种:
提交方式是已JAVA API编程的方式提交,这种方式不需要使用命令行,直接可以在IDEA中点击Run 运行包含Job的Main类就行,Spark 提供了以SparkLanuncher 作为唯一入口的API来实现。这种方式很方便(试想如果某个任务需要重复执行,但是又不会写linux 脚本怎么搞?我想到的是以JAV API的方式提交Job, 还可以和Spring整合,让应用在tomcat中运行),官网的示例:http://spark.apache.org/docs/latest/api/java/index.html?org/apache/spark/launcher/package-summary.html
根据官网的示例,通过JAVA API编程的方式提交有两种方式:
第一种是调用SparkLanuncher实例的startApplication方法,但是这种方式在所有配置都正确的情况下使用运行都会失败的,原因是startApplication方法会调用LauncherServer启动一个进程与集群交互,这个操作貌似是异步的,所以可能结果是main主线程结束了这个进程都没有起起来,导致运行失败。解决办法是调用new SparkLanuncher().startApplication后需要让主线程休眠一定的时间后者是使用下面的例子:
1 package com.learn.spark; 2 3 import org.apache.spark.launcher.SparkAppHandle; 4 import org.apache.spark.launcher.SparkLauncher; 5 6 import java.io.IOException; 7 import java.util.HashMap; 8 import java.util.concurrent.CountDownLatch; 9 10 public class LanuncherAppV { 11 public static void main(String[] args) throws IOException, InterruptedException { 12 13 14 HashMap env = new HashMap(); 15 //这两个属性必须设置 16 env.put("HADOOP_CONF_DIR", "/usr/local/hadoop/etc/overriterHaoopConf"); 17 env.put("JAVA_HOME", "/usr/local/java/jdk1.8.0_151"); 18 //可以不设置 19 //env.put("YARN_CONF_DIR",""); 20 CountDownLatch countDownLatch = new CountDownLatch(1); 21 //这里调用setJavaHome()方法后,JAVA_HOME is not set 错误依然存在 22 SparkAppHandle handle = new SparkLauncher(env) 23 .setSparkHome("/usr/local/spark") 24 .setAppResource("/usr/local/spark/spark-demo.jar") 25 .setMainClass("com.learn.spark.SimpleApp") 26 .setMaster("yarn") 27 .setDeployMode("cluster") 28 .setConf("spark.app.id", "11222") 29 .setConf("spark.driver.memory", "2g") 30 .setConf("spark.akka.frameSize", "200") 31 .setConf("spark.executor.memory", "1g") 32 .setConf("spark.executor.instances", "32") 33 .setConf("spark.executor.cores", "3") 34 .setConf("spark.default.parallelism", "10") 35 .setConf("spark.driver.allowMultipleContexts", "true") 36 .setVerbose(true).startApplication(new SparkAppHandle.Listener() { 37 //这里监听任务状态,当任务结束时(不管是什么原因结束),isFinal()方法会返回true,否则返回false 38 @Override 39 public void stateChanged(SparkAppHandle sparkAppHandle) { 40 if (sparkAppHandle.getState().isFinal()) { 41 countDownLatch.countDown(); 42 } 43 System.out.println("state:" + sparkAppHandle.getState().toString()); 44 } 45 46 47 @Override 48 public void infoChanged(SparkAppHandle sparkAppHandle) { 49 System.out.println("Info:" + sparkAppHandle.getState().toString()); 50 } 51 }); 52 System.out.println("The task is executing, please wait ...."); 53 //线程等待任务结束 54 countDownLatch.await(); 55 System.out.println("The task is finished!"); 56 57 58 } 59 }
注意:如果部署模式是cluster,但是代码中有标准输出的话将看不到,需要把结果写到HDFS中,如果是client模式则可以看到输出。
第二种方式是:通过SparkLanuncher.lanunch()方法获取一个进程,然后调用进程的process.waitFor()方法等待线程返回结果,但是使用这种方式需要自己管理运行过程中的输出信息,比较麻烦,好处是一切都在掌握之中,即获取的输出信息和通过命令提交的方式一样,很详细,实现如下:
1 package com.learn.spark; 2 3 import org.apache.spark.launcher.SparkAppHandle; 4 import org.apache.spark.launcher.SparkLauncher; 5 6 import java.io.IOException; 7 import java.util.HashMap; 8 9 public class LauncherApp { 10 11 public static void main(String[] args) throws IOException, InterruptedException { 12 13 HashMap env = new HashMap(); 14 //这两个属性必须设置 15 env.put("HADOOP_CONF_DIR","/usr/local/hadoop/etc/overriterHaoopConf"); 16 env.put("JAVA_HOME","/usr/local/java/jdk1.8.0_151"); 17 //env.put("YARN_CONF_DIR",""); 18 19 SparkLauncher handle = new SparkLauncher(env) 20 .setSparkHome("/usr/local/spark") 21 .setAppResource("/usr/local/spark/spark-demo.jar") 22 .setMainClass("com.learn.spark.SimpleApp") 23 .setMaster("yarn") 24 .setDeployMode("cluster") 25 .setConf("spark.app.id", "11222") 26 .setConf("spark.driver.memory", "2g") 27 .setConf("spark.akka.frameSize", "200") 28 .setConf("spark.executor.memory", "1g") 29 .setConf("spark.executor.instances", "32") 30 .setConf("spark.executor.cores", "3") 31 .setConf("spark.default.parallelism", "10") 32 .setConf("spark.driver.allowMultipleContexts","true") 33 .setVerbose(true); 34 35 36 Process process =handle.launch(); 37 InputStreamReaderRunnable inputStreamReaderRunnable = new InputStreamReaderRunnable(process.getInputStream(), "input"); 38 Thread inputThread = new Thread(inputStreamReaderRunnable, "LogStreamReader input"); 39 inputThread.start(); 40 41 InputStreamReaderRunnable errorStreamReaderRunnable = new InputStreamReaderRunnable(process.getErrorStream(), "error"); 42 Thread errorThread = new Thread(errorStreamReaderRunnable, "LogStreamReader error"); 43 errorThread.start(); 44 45 System.out.println("Waiting for finish..."); 46 int exitCode = process.waitFor(); 47 System.out.println("Finished! Exit code:" + exitCode); 48 49 } 50 }
使用的自定义InputStreamReaderRunnable类实现如下:
1 package com.learn.spark; 2 3 import java.io.BufferedReader; 4 import java.io.IOException; 5 import java.io.InputStream; 6 import java.io.InputStreamReader; 7 8 public class InputStreamReaderRunnable implements Runnable { 9 10 private BufferedReader reader; 11 12 private String name; 13 14 public InputStreamReaderRunnable(InputStream is, String name) { 15 this.reader = new BufferedReader(new InputStreamReader(is)); 16 this.name = name; 17 } 18 19 public void run() { 20 21 System.out.println("InputStream " + name + ":"); 22 try { 23 String line = reader.readLine(); 24 while (line != null) { 25 System.out.println(line); 26 line = reader.readLine(); 27 } 28 reader.close(); 29 } catch (IOException e) { 30 e.printStackTrace(); 31 } 32 } 33 }