工作原理图
源代码分析
包名：org.apache.spark.deploy.worker
启动driver入口点：registerWithMaster方法中的case LaunchDriver
| 1 | ) => DriverState.FINISHED |
| 37 | case _ => DriverState.FAILED |
| 38 | } |
| 39 | } |
| 40 | |
| 41 | finalState = Some(state) |
| 42 | // 向Driver所属worker发送DriverStateChanged消息 |
| 43 | worker ! DriverStateChanged(driverId, state, finalException) |
| 44 | } |
| 45 | }.start() |
| 46 | } |
LaunchExecutor
管理LaunchExecutor的启动
| 1 | case LaunchExecutor(masterUrl, appId, execId, appDesc, cores_, memory_) => |
| 2 | if (masterUrl != activeMasterUrl) { |
| 3 | logWarning("Invalid Master (" + masterUrl + ") attempted to launch executor.") |
| 4 | } else { |
| 5 | try { |
| 6 | logInfo("Asked to launch executor %s/%d for %s".format(appId, execId, appDesc.name)) |
| 7 | |
| 8 | // Create the executor's working directory |
| 9 | // 创建executor本地工作目录 |
| 10 | val executorDir = new File(workDir, appId + "/" + execId) |
| 11 | if (!executorDir.mkdirs()) { |
| 12 | throw new IOException("Failed to create directory " + executorDir) |
| 13 | } |
| 14 | |
| 15 | // Create local dirs for the executor. These are passed to the executor via the |
| 16 | // SPARK_LOCAL_DIRS environment variable, and deleted by the Worker when the |
| 17 | // application finishes. |
| 18 | val appLocalDirs = appDirectories.get(appId).getOrElse { |
| 19 | Utils.getOrCreateLocalRootDirs(conf).map { dir => |
| 20 | Utils.createDirectory(dir).getAbsolutePath() |
| 21 | }.toSeq |
| 22 | } |
| 23 | appDirectories(appId) = appLocalDirs |
| 24 | // 创建ExecutorRunner对象 |
| 25 | val manager = new ExecutorRunner( |
| 26 | appId, |
| 27 | execId, |
| 28 | appDesc.copy(command = Worker.maybeUpdateSSLSettings(appDesc.command, conf)), |
| 29 | cores_, |
| 30 | memory_, |
| 31 | self, |
| 32 | workerId, |
| 33 | host, |
| 34 | webUi.boundPort, |
| 35 | publicAddress, |
| 36 | sparkHome, |
| 37 | executorDir, |
| 38 | akkaUrl, |
| 39 | conf, |
| 40 | appLocalDirs, ExecutorState.LOADING) |
| 41 | // executor加入本地缓存 |
| 42 | executors(appId + "/" + execId) = manager |
| 43 | manager.start() |
| 44 | // 增加worker已使用core |
| 45 | coresUsed += cores_ |
| 46 | // 增加worker已使用memory |
| 47 | memoryUsed += memory_ |
| 48 | // 通知master发送ExecutorStateChanged消息 |
| 49 | master ! ExecutorStateChanged(appId, execId, manager.state, None, None) |
| 50 | } |
| 51 | // 异常情况处理，通知master发送ExecutorStateChanged FAILED消息 |
| 52 | catch { |
| 53 | case e: Exception => { |
| 54 | logError(s"Failed to launch executor $appId/$execId for ${appDesc.name}.", e) |
| 55 | if (executors.contains(appId + "/" + execId)) { |
| 56 | executors(appId + "/" + execId).kill() |
| 57 | executors -= appId + "/" + execId |
| 58 | } |
| 59 | master ! ExecutorStateChanged(appId, execId, ExecutorState.FAILED, |
| 60 | Some(e.toString), None) |
| 61 | } |
| 62 | } |
| 63 | } |
总结
1、Worker、Driver、Application启动后都会向Master进行注册,并缓存到Master内存数据模型中
2、完成注册后发送LaunchExecutor、LaunchDriver到Worker
3、Worker收到消息后启动executor和driver进程，并调用Worker的ExecutorStateChanged和DriverStateChanged方法
4、发送ExecutorStateChanged和DriverStateChanged消息到Master的，根据各自的状态信息进行处理，最重要的是会调用schedule方法进行资源的重新调度