【Flink源码】JobManager源码之启动WebMonitorEndpoint

在【Flink源码】JobManager 启动流程 一文中，笔者分析了 JobManager 的启动流程，并讲到 JobManager 启动了 ResourceManager、Dispatcher 和 WebMonitorEndpoint 三大核心组件。
本文将就其中 WebMonitorEndpoint 启动过程的源码进行探究

首先，我们先要讲述一下 WebMonitorEndpoint 的主要功能：
WebMonitorEndpoint 里面维护了很多 Handler，如果客户端通过 flink run 的方式来提交一个 job 到 flink 集群，最终，是由 WebMonitorEndpoint 来接收，并且决定使用哪一个 Handler 来执行处理

---

WebMonitorEndpoint 启动流程解析

按前文所述，在 JobManager 创建完工厂类后，首先创建的实例为 WebMonitorEndpoint 实例

DefaultDispatcherResourceManagerComponentFactory.java

// 创建 WebMonitorEndpoint 实例，在 Standalone 模式下为：DispatcherRestEndpoint
// 该实例内部会启动一个 Netty 服务端，绑定了一堆 Handler
webMonitorEndpoint =
        restEndpointFactory.createRestEndpoint(
                configuration,
                dispatcherGatewayRetriever,
                resourceManagerGatewayRetriever,
                blobServer,
                executor,
                metricFetcher,
                highAvailabilityServices.getClusterRestEndpointLeaderElectionService(),
                fatalErrorHandler);
// 启动 WebMonitorEndpoint
log.debug("Starting Dispatcher REST endpoint.");
webMonitorEndpoint.start();

下面我们从 start 方法讲起
该方法是 WebMonitorEndpoint 类下的方法，继承了 RestServerEndpoint 类，找到该类
由于代码很长，我们分段来看

Handler 相关操作

RestServerEndpoint.java

// 路由器，解析请求寻找对应 Handler
final Router router = new Router();
final CompletableFuture<String> restAddressFuture = new CompletableFuture<>();
// 初始化 Handlers
handlers = initializeHandlers(restAddressFuture);

/* sort the handlers such that they are ordered the following:
    * /jobs
    * /jobs/overview
    * /jobs/:jobid
    * /jobs/:jobid/config
    * /:*
    */
// 排序，为了确认 URL 和 Handler 的一一对应的关系，不应出现一对多的情况
Collections.sort(handlers, RestHandlerUrlComparator.INSTANCE);
// 确认唯一性方法
checkAllEndpointsAndHandlersAreUnique(handlers);
// 注册 Handler
handlers.forEach(handler -> registerHandler(router, handler, log));

这一段主要包含 Handler 相关操作，包括：

1. 首先创建 Router 来解析 Client 的请求并寻找对应的 Handler
2. 通过 initializeHandlers 方法注册了一堆 Handler
3. 将这些 Handler 进行排序，这里的排序是为了确认 URL 和 Handler 一对一的关系
4. 排序好后通过 checkAllEndpointsAndHandlersAreUnique 方法来确认唯一性
5. 确认唯一性后将 Handler 逐一注册

Netty 启动的相关操作

NioEventLoopGroup bossGroup =
        new NioEventLoopGroup(
                1, new ExecutorThreadFactory("flink-rest-server-netty-boss"));
NioEventLoopGroup workerGroup =
        new NioEventLoopGroup(
                0, new ExecutorThreadFactory("flink-rest-server-netty-worker"));

bootstrap = new ServerBootstrap();
bootstrap
        .group(bossGroup, workerGroup)
        .channel(NioServerSocketChannel.class)
        .childHandler(initializer);

Iterator<Integer> portsIterator;
try {
    portsIterator = NetUtils.getPortRangeFromString(restBindPortRange);
} catch (IllegalConfigurationException e) {
    throw e;
} catch (Exception e) {
    throw new IllegalArgumentException(
            "Invalid port range definition: " + restBindPortRange);
}

这一段主要是 Netty 启动的相关操作，为了解决端口冲突问题，Flink 做了如下操作

// 为防止端口冲突，将逐一尝试端口是否可用
int chosenPort = 0;
while (portsIterator.hasNext()) {
    try {
        chosenPort = portsIterator.next();
        final ChannelFuture channel;
        if (restBindAddress == null) {
            channel = bootstrap.bind(chosenPort);
        } else {
            channel = bootstrap.bind(restBindAddress, chosenPort);
        }
        serverChannel = channel.syncUninterruptibly().channel();
        break;
    } catch (final Exception e) {
        // syncUninterruptibly() throws checked exceptions via Unsafe
        // continue if the exception is due to the port being in use, fail early
        // otherwise
        if (!(e instanceof java.net.BindException)) {
            throw e;
        }
    }
}

修改状态并启动其他基础服务

// 修改状态
state = State.RUNNING;
// 启动其他基础服务
startInternal();

在方法末尾修改一下 EndPoint 状态为 RUNNING，至此，WebMonitorEndpoint 的 Netty 服务就启动完毕了，接下来通过 startInternal 方法启动其他基础服务

节点选举以及方法回调

我们进入 startInternal 方法内部

public void startInternal() throws Exception {
    // 开始选举，去找回调方法
    leaderElectionService.start(this);
    startExecutionGraphCacheCleanupTask();

    if (hasWebUI) {
        log.info("Web frontend listening at {}.", getRestBaseUrl());
    }
}

可以看到 WebMonitorEndpoint 准备开始进行 Leader 竞选
首先介绍一下主节点内部组件的 Leader 选举机制

Flink 的选举使用的是 Curator 框架，节点的选举针对每一个参选对象，会创建一个选举驱动 leaderElectionDriver，在完成选举之后，会回调两个方法。
如果选举成功会回调 isLeader 方法，如果失败则回调 notLeader 方法

基于此，我们来看 leaderElectionService.start 方法

DefaultLeaderElectionService.java

public final void start(LeaderContender contender) throws Exception {
    checkNotNull(contender, "Contender must not be null.");
    Preconditions.checkState(leaderContender == null, "Contender was already set.");

    synchronized (lock) {
        // 在 WebMonitorEndpoint 中调用时，在 contender 为 DispatcherRestEndPoint
        // 在 ResourceManager 中调用时，contender 为 ResourceManager
        // 在 DispatcherRunner 中调用时，contender 为 DisoatcherRunner
        running = true;
        leaderContender = contender;
        // 创建选举对象 leaderElectionDriver
        leaderElectionDriver =
                leaderElectionDriverFactory.createLeaderElectionDriver(
                        this,
                        new LeaderElectionFatalErrorHandler(),
                        leaderContender.getDescription());
        LOG.info("Starting DefaultLeaderElectionService with {}.", leaderElectionDriver);
    }
}

在 Standalone 模式下，WebMonitorEndpoint、ResourceManager、DispatcherRunner 都会使用该模式来进行竞选，此刻我们是从 WebMonitorEndpoint 进入的此方法，此时 contender 对象实际为 DispatcherRestEndpoint。
我们继续看 leaderElectionDriver 的构建
进入 createLeaderElectionDriver 方法，选择 ZooKeeperLeaderElectionDriverFactory 实现

ZooKeeperLeaderElectionDriverFactory.java

public ZooKeeperLeaderElectionDriver createLeaderElectionDriver(
        LeaderElectionEventHandler leaderEventHandler,
        FatalErrorHandler fatalErrorHandler,
        String leaderContenderDescription)
        throws Exception {
    return new ZooKeeperLeaderElectionDriver(
            client, path, leaderEventHandler, fatalErrorHandler, leaderContenderDescription);
}

ZooKeeperLeaderElectionDriver.java

public ZooKeeperLeaderElectionDriver(
        CuratorFramework client,
        String path,
        LeaderElectionEventHandler leaderElectionEventHandler,
        FatalErrorHandler fatalErrorHandler,
        String leaderContenderDescription)
        throws Exception {
    checkNotNull(path);
    this.client = checkNotNull(client);
    this.connectionInformationPath = ZooKeeperUtils.generateConnectionInformationPath(path);
    this.leaderElectionEventHandler = checkNotNull(leaderElectionEventHandler);
    this.fatalErrorHandler = checkNotNull(fatalErrorHandler);
    this.leaderContenderDescription = checkNotNull(leaderContenderDescription);

    leaderLatchPath = ZooKeeperUtils.generateLeaderLatchPath(path);
    leaderLatch = new LeaderLatch(client, leaderLatchPath);
    this.cache =
            ZooKeeperUtils.createTreeCache(
                    client,
                    connectionInformationPath,
                    this::retrieveLeaderInformationFromZooKeeper);

    running = true;
    // 开始选举
    leaderLatch.addListener(this);
    leaderLatch.start();
    // 选举开始后，会接收到响应：
    // 1.如果竞选成功，则回调该类的 isLeader 方法
    // 2.如果竞选失败，则回调该类的 notLeader 方法
    // 每一个竞选者对应一个竞选 Driver
    cache.start();

    client.getConnectionStateListenable().addListener(listener);
}

接下来我们继续看该类的 isLeader 方法

public void isLeader() {
    leaderElectionEventHandler.onGrantLeadership(UUID.randomUUID());
}

再进入 onGrantLeadership 方法

@Override
@GuardedBy("lock")
public void onGrantLeadership(UUID newLeaderSessionId) {
    synchronized (lock) {
        if (running) {
            issuedLeaderSessionID = newLeaderSessionId;
            clearConfirmedLeaderInformation();

            if (LOG.isDebugEnabled()) {
                LOG.debug(
                        "Grant leadership to contender {} with session ID {}.",
                        leaderContender.getDescription(),
                        issuedLeaderSessionID);
            }

            /**
                * 有 4 种竞选者类型，LeaderContender 有 4 种情况
                * 1.Dispatcher = DefaultDispatcherRunner
                * 2.JobMaster = JobManagerRunnerImpl
                * 3.ResourceManager = ResourceManager
                * 4.WebMonitorEndpoint = WebMonitorEndpoint
                */
            leaderContender.grantLeadership(issuedLeaderSessionID);
        } else {
            if (LOG.isDebugEnabled()) {
                LOG.debug(
                        "Ignoring the grant leadership notification since the {} has "
                                + "already been closed.",
                        leaderElectionDriver);
            }
        }
    }
}

再进入 leaderContender.grantLeadership 方法，引入当前是 WebMonitorEndpoint 的选举，所以我们进入 WebMonitorEndpoint 的实现

@Override
public void grantLeadership(final UUID leaderSessionID) {
    log.info(
            "{} was granted leadership with leaderSessionID={}",
            getRestBaseUrl(),
            leaderSessionID);
    leaderElectionService.confirmLeadership(leaderSessionID, getRestBaseUrl());
}

@Override
public void confirmLeadership(UUID leaderSessionID, String leaderAddress) {
    if (LOG.isDebugEnabled()) {
        LOG.debug(
                "Confirm leader session ID {} for leader {}.", leaderSessionID, leaderAddress);
    }

    checkNotNull(leaderSessionID);

    synchronized (lock) {
        if (hasLeadership(leaderSessionID)) {
            if (running) {
                // 确认 Leader 信息，并将节点信息写入 Zookeeper
                confirmLeaderInformation(leaderSessionID, leaderAddress);
            } else {
                if (LOG.isDebugEnabled()) {
                    LOG.debug(
                            "Ignoring the leader session Id {} confirmation, since the "
                                    + "LeaderElectionService has already been stopped.",
                            leaderSessionID);
                }
            }
        } else {
            // Received an old confirmation call
            if (!leaderSessionID.equals(this.issuedLeaderSessionID)) {
                if (LOG.isDebugEnabled()) {
                    LOG.debug(
                            "Receive an old confirmation call of leader session ID {}, "
                                    + "current issued session ID is {}",
                            leaderSessionID,
                            issuedLeaderSessionID);
                }
            } else {
                LOG.warn(
                        "The leader session ID {} was confirmed even though the "
                                + "corresponding JobManager was not elected as the leader.",
                        leaderSessionID);
            }
        }
    }
}

@GuardedBy("lock")
private void confirmLeaderInformation(UUID leaderSessionID, String leaderAddress) {
    confirmedLeaderInformation = LeaderInformation.known(leaderSessionID, leaderAddress);
    leaderElectionDriver.writeLeaderInformation(confirmedLeaderInformation);
}

WebMonitorEndpoint 在选举 Leader 成功后，并没有做什么，只是将自己的信息写入 Zookeeper
信息写入完毕后，WebMonitorEndpoint 就算是启动完成了

---

总结
WebMonitorEndpoint 的启动流程并不复杂，主要如下：

1. 初始化一堆 Handler
2. 启动 Netty 服务，注册 Handler
3. 启动内部服务：执行竞选，WebMonitorEndpoint 本身就是一个 LeaderContender 角色
4. 竞选成功，其实只是把 WebMonitorEndpoint 的 address 以及和 zookeeper 的 sessionId 写入 znode 中