JStorm之Supervisor简介

一、简介
Supervisor是JStorm中的工作节点，类似于MR的TT，subscribe zookeeper的任务调度结果数据，根据任务调度情况启动/停止工作进程Worker。同时Supervisor需要定期向zookeeper写入活跃端口信息以便Nimbus监控。Supervisor不执行具体处理工作，所有的计算任务都交Worker完成。从整个架构上看，Supervisor处在整个JStorm三级管理架构的中间环节，辅助管理任务调度和资源管理工作。
二、架构

1.Supervisor


 

Supervisor单节点架构如上图所示，初始化时启动进程Supervisor，根据Nimbus分配的任务情况触发启动/停用Worker JVM进程，其中每个Worker进程启动一个或多个Task线程，其中Task须同属单个Topology。从整个Supervisor节点来看运行多个JVM进程，包括一个Supervisor进程和一个或多个Worker进程。
不同角色状态通过不同的方式维护。其中Task通过hb直接将包括时间信息和当前Task的统计信息写到zookeeper；Worker定期将包括Topology id，端口，Task id集合及当前时间写入本地；Supervisor定期将包括时间及节点资源（端口集合）写到zookeeper，同时从zookeeper读取任务调度结果，根据结果启动/停用Worker进程。
2.Worker


 

在Worker JVM进程内部，除了相互独立的Task线程外，Task线程会共享数据收发和节点之间连接管理等Worker进程内的公共资源，如图所示。其中：
VirtualPort：数据接收线程；
KeyoTupleSerialize：Tuple数据序列化；
TransferQueue：数据发送管道；
DrainerRunnable：数据发送线程；
RefreshConnections：节点之间连接管理线程。
三、实现与代码剖析

1.Supervisor

在jstorm-0.7.1中，Supervisor daemon实现在jstorm-server/src/main/java目录下com.alipay.dw.jstorm.daemon.supervisor包里。Supervisor.java是Supervisor daemon的入口，Supervisor进程主要做以下几件事情。
初始化

1、清理本地临时目录下数据$jstorm-local-dir/supervisor/tmp；
2、创建zk操作实例；
3、本地新建状态文件，$jstorm-local-dir/supervisor/localstate；
4、生成supervisor-id并写入localstate，其中key=”supervisor-id”；如果supervisor重启，先检查supervisor-id是否已经存在，若存在直接读取即可；
5、初始化并启动Heartbeat线程；
6、初始化并启动SyncProcessEvent线程；
7、初始化并启动SyncProcessEvent线程；
8、注册主进程退出数据清理Hook in SupervisorManger。
@SuppressWarnings("rawtypes")public SupervisorManger mkSupervisor(Map conf, MQContext sharedContext)throwsException{
    LOG.info("Starting Supervisor with conf "+ conf);
    active =new AtomicBoolean(true);/*
     * Step 1: cleanup all files in /storm-local-dir/supervisor/tmp
     */String path = StormConfig.supervisorTmpDir(conf);
    FileUtils.cleanDirectory(newFile(path));/*
     * Step 2: create ZK operation instance
     * StromClusterState
     */
    StormClusterState stormClusterState = Cluster
            .mk_storm_cluster_state(conf);/*
     * Step 3, create LocalStat
     * LocalStat is one KV database
     * 4.1 create LocalState instance
     * 4.2 get supervisorId, if no supervisorId, create one
     */
    LocalState localState = StormConfig.supervisorState(conf);String supervisorId =(String) localState.get(Common.LS_ID);if(supervisorId ==null){
        supervisorId = UUID.randomUUID().toString();
        localState.put(Common.LS_ID, supervisorId);}Vector threads =newVector();// Step 5 create HeartBeat// every supervisor.heartbeat.frequency.secs, write SupervisorInfo to ZKString myHostName = NetWorkUtils.hostname();int startTimeStamp = TimeUtils.current_time_secs();
    Heartbeat hb =new Heartbeat(conf, stormClusterState, supervisorId,
            myHostName, startTimeStamp, active);
    hb.update();
    AsyncLoopThread heartbeat =new AsyncLoopThread(hb, false, null,
            Thread.MIN_PRIORITY, true);
    threads.add(heartbeat);// Step 6 create and start sync Supervisor thread// every supervisor.monitor.frequency.secs second run SyncSupervisor
    EventManager processEventManager =new EventManagerImp(false);
    ConcurrentHashMap workerThreadPids =new ConcurrentHashMap();//读取$jstorm-local-dir/supervior/localstate中key=local-assignments的value值，根据该值执行workers的kill/start
    SyncProcessEvent syncProcessEvent =new SyncProcessEvent(supervisorId,
            conf, localState, workerThreadPids, sharedContext);
    EventManager syncSupEventManager =new EventManagerImp(false);//通过比较$zkroot/assignments/{topologyid}全量数据和本地STORM-LOCAL-DIR/supervisor/stormdist/{topologyid}://1.从nimbus下载有任务分配到本节点的topology的jar和配置数据//2.从本地删除已经失效的topology的jar和配置数据
    SyncSupervisorEvent syncSupervisorEvent =new SyncSupervisorEvent(
            supervisorId, conf, processEventManager, syncSupEventManager,
            stormClusterState, localState, syncProcessEvent);int syncFrequence =(Integer) conf
            .get(Config.SUPERVISOR_MONITOR_FREQUENCY_SECS);
    EventManagerPusher syncSupervisorPusher =new EventManagerPusher(
            syncSupEventManager, syncSupervisorEvent, active, syncFrequence);
    AsyncLoopThread syncSupervisorThread =new AsyncLoopThread(
            syncSupervisorPusher);
    threads.add(syncSupervisorThread);
    LOG.info("Starting supervisor with id "+ supervisorId +" at host "+ myHostName);// SupervisorManger which can shutdown all supervisor and workersreturnnew SupervisorManger(conf, supervisorId, active, threads,
            syncSupEventManager, processEventManager, stormClusterState,
            workerThreadPids);}
Heartbeat线程

1、默认间隔60s向zookeeper汇报supervisor信息，汇报内容打包成SupervisorInfo，包括hostname，workerports，current time和during time等信息；
@SuppressWarnings("unchecked")publicvoid update(){
    SupervisorInfo sInfo =new SupervisorInfo(
            TimeUtils.current_time_secs(), myHostName,
            (List) conf.get(Config.SUPERVISOR_SLOTS_PORTS),
            (int)(TimeUtils.current_time_secs()- startTime));try{
        stormClusterState.supervisor_heartbeat(supervisorId, sInfo);}catch(Exception e){
        LOG.error("Failed to update SupervisorInfo to ZK", e);

    }}
SyncProcessEvent线程

1、定期从本地文件$jstorm-local-dir/supervisor/localstate中读取key=”local-assignments”数据；该数据会由SyncSupervisorEvent线程定期写入；
2、读取本地$jstorm-local-dir /worker/ids/heartbeat中Worker状态数据；
3、对比local-assignments及worker的状态数据，执行操作start/kill worker进程；其中Worker和Supervisor属于不同JVM进程，Supervisor通过Shell命令启动Worker：
nohup java –server
-Djava.library.path="$JAVA.LIBRARY.PATH"
-Dlogfile.name="$topologyid-worker-$port.log"
-Dlog4j.configuration=jstorm.log4j.properties
-Djstorm.home="$JSTORM_HOME"
-cp $JAVA_CLASSSPATH:$JSTORM_CLASSPATH
com.alipay.dw.jstorm.daemon.worker.Worker
topologyid supervisorid port workerid
SyncProcessEvent线程执行流程如下：

@SuppressWarnings("unchecked")
@Override
publicvoid run(){
    LOG.debug("Syncing processes");try{/**
         * Step 1: get assigned tasks from localstat Map
         *///1.从本地文件$jstorm-local-dir/supervisor/localstate里读取key=“local-assignments”数据Map localAssignments =null;try{
            localAssignments =(Map) localState
                    .get(Common.LS_LOCAL_ASSIGNMENTS);}catch(IOException e){
            LOG.error("Failed to get LOCAL_ASSIGNMENTS from LocalState", e);throw e;}if(localAssignments ==null){
            localAssignments =newHashMap();}
        LOG.debug("Assigned tasks: "+ localAssignments);/**
         * Step 2: get local WorkerStats from local_dir/worker/ids/heartbeat
         * Map
         *///2.根据localAssignments与workers的hb比对结果得到workers的状态Map localWorkerStats =null;try{
            localWorkerStats = getLocalWorkerStats(conf, localState,
                    localAssignments);}catch(IOException e){
            LOG.error("Failed to get Local worker stats");throw e;}
        LOG.debug("Allocated: "+ localWorkerStats);/**
         * Step 3: kill Invalid Workers and remove killed worker from
         * localWorkerStats
         *///3.根据workers的状态值启动/停用相关workerSet keepPorts = killUselessWorkers(localWorkerStats);// start new workers
        startNewWorkers(keepPorts, localAssignments);}catch(Exception e){
        LOG.error("Failed Sync Process", e);// throw e}}
SyncSupervisorEvent线程

1、从$zk-root/assignments/{topologyid}下载所有任务调度结果，并筛选出分配到当前supervisor的任务集合，验证单个端口仅分配了单个Topology的任务通过后，将上述任务集合写入本地文件$jstorm-local-dir/supervisor/localstate，以便SyncProcessEvent读取及后续操作；
2、对比任务分配结果与已经存在的Topology，从Nimbus下载新分配过来的Topology，同时删除过期Topology。
SyncSupervisorEvent线程执行流程如下：

@Override
publicvoid run(){
    LOG.debug("Synchronizing supervisor");try{
        RunnableCallback syncCallback =new EventManagerZkPusher(this,
                syncSupEventManager);/**
         * Step 1: get all assignments
         * and register /ZK-dir/assignment and every assignment watch
         *
         *///1.从zk获取分配完成的任务集assignments:(topologyid -> Assignment)//$zkroot/assignments/{topologyid}Map assignments = Cluster.get_all_assignment(
                stormClusterState, syncCallback);
        LOG.debug("Get all assignments "+ assignments);/**
         * Step 2: get topologyIds list from
         * STORM-LOCAL-DIR/supervisor/stormdist/
         *///2.本地已经下载的topology集合$jstorm-local-dir/supervisor/stormdist/{topologyid}List downloadedTopologyIds = StormConfig
                .get_supervisor_toplogy_list(conf);
        LOG.debug("Downloaded storm ids: "+ downloadedTopologyIds);/**
         * Step 3: get  from ZK local node's
         * assignment
         *///3.从assignments里筛选出分配到当前supervisor的任务集合Map localAssignment = getLocalAssign(
                stormClusterState, supervisorId, assignments);/**
         * Step 4: writer local assignment to LocalState
         *///4.将步骤3得到的结果写本地文件$jstorm-local-dir/supervisor/localstatetry{
            LOG.debug("Writing local assignment "+ localAssignment);
            localState.put(Common.LS_LOCAL_ASSIGNMENTS, localAssignment);}catch(IOException e){
            LOG.error("put LS_LOCAL_ASSIGNMENTS "+ localAssignment
                    +" of localState failed");throw e;}// Step 5: download code from ZK//5.下载新分配任务的TopologyMap topologyCodes = getTopologyCodeLocations(assignments);
        downloadTopology(topologyCodes, downloadedTopologyIds);/**
         * Step 6: remove any downloaded useless topology
         */6.删除过期任务的Topology
        removeUselessTopology(topologyCodes, downloadedTopologyIds);/**
         * Step 7: push syncProcesses Event
         */
        processEventManager.add(syncProcesses);}catch(Exception e){
        LOG.error("Failed to Sync Supervisor", e);// throw new RuntimeException(e);}}
2.Worker

在jstorm-0.7.1里，Worker daemon实现在jstorm-server/src/main/java目录下com.alipay.dw.jstorm.daemon.worker包。其中Worker.java是Worker daemon的入口。Worker进程的生命周期：
1、初始化Tuple序列化功能和数据发送管道；
2、创建分配到当前Worker的Tasks；
3、初始化并启动接收Tuple dispatcher；
4、初始化并启动用于维护Worker间连接线程RefreshConnections，包括创建/维护/销毁节点之间的连接等功能；
5、初始化并启动心跳线程WorkerHeartbeatRunable，更新本地目录：$jstorm_local_dir/worker/{workerid}/heartbeats/{workerid}；
6、初始化并启动发送Tuple线程DrainerRunable；
7、注册主线程退出现场数据清理Hook。
Worker Daemon初始化流程如下：

public WorkerShutdown execute()throwsException{//1. Tuple序列化+发送管道LinkedBlockingQueue
    WorkerTransfer workerTransfer = getSendingTransfer();// shutdown task callbacks//2. 初始化task线程List shutdowntasks = createTasks(workerTransfer);
    workerData.setShutdownTasks(shutdowntasks);//3. WorkerVirtualPort：tuple接收dispatcher// create virtual port object// when worker receives tupls, dispatch targetTask according to task_id// conf, supervisorId, topologyId, port, mqContext, taskids
    WorkerVirtualPort virtual_port =new WorkerVirtualPort(workerData);
    Shutdownable virtual_port_shutdown = virtual_port.launch();//3. RefreshConnections：维护节点间的连接：创建新连接|维护已建立连接|销毁无用连接// refresh connection
    RefreshConnections refreshConn = makeRefreshConnections();
    AsyncLoopThread refreshconn =new AsyncLoopThread(refreshConn);// refresh ZK active status
    RefreshActive refreshZkActive =new RefreshActive(workerData);
    AsyncLoopThread refreshzk =new AsyncLoopThread(refreshZkActive);//4. WorkerHeartbeatRunable：心跳线程// 每次心跳更新本地目录数据 $LOCAL_PATH/workers/{worker-id}/Heartbeats/{worker-id}// refresh hearbeat to Local dir
    RunnableCallback heartbeat_fn =new WorkerHeartbeatRunable(workerData);
    AsyncLoopThread hb =new AsyncLoopThread(heartbeat_fn, false, null,
            Thread.NORM_PRIORITY, true);//5. DrainerRunable：发送tuple线程// transferQueue, nodeportSocket, taskNodeport
    DrainerRunable drainer =new DrainerRunable(workerData);
    AsyncLoopThread dr =new AsyncLoopThread(drainer, false, null,
            Thread.MAX_PRIORITY, true);
    AsyncLoopThread[] threads ={ refreshconn, refreshzk, hb, dr };//6. 注册主线程退出数据清理hookreturnnew WorkerShutdown(workerData, shutdowntasks,
            virtual_port_shutdown, threads);}
3.Task

根据任务在Topology中不同节点角色，Task相应也会分成SpoutTask和BoltTask，二者除Task心跳及公共数据初始化等相同以外，各自有独立处理逻辑。核心实现在SpoutExecutors.java/BoltExecutors.java。
SpoutExecutors主要做两件事情：
1、作为DAG起点，负责发送原始Tuple数据；
2、如果Topology定义了Acker，SpoutExecutors会启动接收ack线程，根据接收到的ack决定是否重发Tuple；
BoltExecutor相比SpoutExecutor功能会稍微复杂：
1、接收从上游发送过来的Tuple，并根据Topology中定义的处理逻辑进行处理；
2、如果该Bolt存在下游，需要向下游发送新生成的Tuple；
3、如果Topology中定义了Acker，Bolt需要将经过简单计算的ack返回给根Spout。