聊聊storm nimbus的LeaderElector

序

本文主要研究一下storm nimbus的LeaderElector

Nimbus

org/apache/storm/daemon/nimbus/Nimbus.java

public static void main(String[] args) throws Exception {
        Utils.setupDefaultUncaughtExceptionHandler();
        launch(new StandaloneINimbus());
    }

    public static Nimbus launch(INimbus inimbus) throws Exception {
        Map<String, Object> conf = Utils.merge(ConfigUtils.readStormConfig(),
                                               ConfigUtils.readYamlConfig("storm-cluster-auth.yaml", false));
        boolean fixupAcl = (boolean) conf.get(DaemonConfig.STORM_NIMBUS_ZOOKEEPER_ACLS_FIXUP);
        boolean checkAcl = fixupAcl || (boolean) conf.get(DaemonConfig.STORM_NIMBUS_ZOOKEEPER_ACLS_CHECK);
        if (checkAcl) {
            AclEnforcement.verifyAcls(conf, fixupAcl);
        }
        return launchServer(conf, inimbus);
    }

    private static Nimbus launchServer(Map<String, Object> conf, INimbus inimbus) throws Exception {
        StormCommon.validateDistributedMode(conf);
        validatePortAvailable(conf);
        StormMetricsRegistry metricsRegistry = new StormMetricsRegistry();
        final Nimbus nimbus = new Nimbus(conf, inimbus, metricsRegistry);
        nimbus.launchServer();
        final ThriftServer server = new ThriftServer(conf, new Processor<>(nimbus), ThriftConnectionType.NIMBUS);
        metricsRegistry.startMetricsReporters(conf);
        Utils.addShutdownHookWithDelayedForceKill(() -> {
            metricsRegistry.stopMetricsReporters();
            nimbus.shutdown();
            server.stop();
        }, 10);
        if (ClientAuthUtils.areWorkerTokensEnabledServer(server, conf)) {
            nimbus.initWorkerTokenManager();
        }
        LOG.info("Starting nimbus server for storm version '{}'", STORM_VERSION);
        server.serve();
        return nimbus;
    }

    public Nimbus(Map<String, Object> conf, INimbus inimbus, IStormClusterState stormClusterState, NimbusInfo hostPortInfo,
                  BlobStore blobStore, TopoCache topoCache, ILeaderElector leaderElector, IGroupMappingServiceProvider groupMapper,
                  StormMetricsRegistry metricsRegistry)
        throws Exception {
        //......

        if (blobStore == null) {
            blobStore = ServerUtils.getNimbusBlobStore(conf, this.nimbusHostPortInfo, null);
        }
        this.blobStore = blobStore;

        if (topoCache == null) {
            topoCache = new TopoCache(blobStore, conf);
        }
        if (leaderElector == null) {
            leaderElector = Zookeeper.zkLeaderElector(conf, zkClient, blobStore, topoCache, stormClusterState, getNimbusAcls(conf),
                metricsRegistry);
        }
        this.leaderElector = leaderElector;
        this.blobStore.setLeaderElector(this.leaderElector);

        //......
    }

    public void launchServer() throws Exception {
        try {
            BlobStore store = blobStore;
            IStormClusterState state = stormClusterState;
            NimbusInfo hpi = nimbusHostPortInfo;

            LOG.info("Starting Nimbus with conf {}", ConfigUtils.maskPasswords(conf));
            validator.prepare(conf);

            //add to nimbuses
            state.addNimbusHost(hpi.getHost(),
                                new NimbusSummary(hpi.getHost(), hpi.getPort(), Time.currentTimeSecs(), false, STORM_VERSION));
            leaderElector.addToLeaderLockQueue();
            this.blobStore.startSyncBlobs();
            
            for (ClusterMetricsConsumerExecutor exec: clusterConsumerExceutors) {
                exec.prepare();
            }

            if (isLeader()) {
                for (String topoId : state.activeStorms()) {
                    transition(topoId, TopologyActions.STARTUP, null);
                }
                clusterMetricSet.setActive(true);
            }

            //......
        } catch (Exception e) {
            if (Utils.exceptionCauseIsInstanceOf(InterruptedException.class, e)) {
                throw e;
            }

            if (Utils.exceptionCauseIsInstanceOf(InterruptedIOException.class, e)) {
                throw e;
            }
            LOG.error("Error on initialization of nimbus", e);
            Utils.exitProcess(13, "Error on initialization of nimbus");
        }
    }

• Nimbus在构造器里头调用Zookeeper.zkLeaderElector建立leaderElector
• launchServer方法调用了leaderElector.addToLeaderLockQueue()参与leader选举

Zookeeper.zkLeaderElector

storm-core-1.1.0-sources.jar!/org/apache/storm/zookeeper/Zookeeper.java

public static ILeaderElector zkLeaderElector(Map conf, BlobStore blobStore) throws UnknownHostException {
        return _instance.zkLeaderElectorImpl(conf, blobStore);
    }

    protected ILeaderElector zkLeaderElectorImpl(Map conf, BlobStore blobStore) throws UnknownHostException {
        List<String> servers = (List<String>) conf.get(Config.STORM_ZOOKEEPER_SERVERS);
        Object port = conf.get(Config.STORM_ZOOKEEPER_PORT);
        CuratorFramework zk = mkClientImpl(conf, servers, port, "", conf);
        String leaderLockPath = conf.get(Config.STORM_ZOOKEEPER_ROOT) + "/leader-lock";
        String id = NimbusInfo.fromConf(conf).toHostPortString();
        AtomicReference<LeaderLatch> leaderLatchAtomicReference = new AtomicReference<>(new LeaderLatch(zk, leaderLockPath, id));
        AtomicReference<LeaderLatchListener> leaderLatchListenerAtomicReference =
                new AtomicReference<>(leaderLatchListenerImpl(conf, zk, blobStore, leaderLatchAtomicReference.get()));
        return new LeaderElectorImp(conf, servers, zk, leaderLockPath, id, leaderLatchAtomicReference,
            leaderLatchListenerAtomicReference, blobStore);
    }

• 这里使用/leader-lock路径建立了LeaderLatch，而后使用leaderLatchListenerImpl建立了LeaderLatchListener
• 最后使用LeaderElectorImp建立ILeaderElector

leaderLatchListenerImpl

storm-core-1.1.0-sources.jar!/org/apache/storm/zookeeper/Zookeeper.java

// Leader latch listener that will be invoked when we either gain or lose leadership
    public static LeaderLatchListener leaderLatchListenerImpl(final Map conf, final CuratorFramework zk, final BlobStore blobStore, final LeaderLatch leaderLatch) throws UnknownHostException {
        final String hostName = InetAddress.getLocalHost().getCanonicalHostName();
        return new LeaderLatchListener() {
            final String STORM_JAR_SUFFIX = "-stormjar.jar";
            final String STORM_CODE_SUFFIX = "-stormcode.ser";
            final String STORM_CONF_SUFFIX = "-stormconf.ser";

            @Override
            public void isLeader() {
                Set<String> activeTopologyIds = new TreeSet<>(Zookeeper.getChildren(zk, conf.get(Config.STORM_ZOOKEEPER_ROOT) + ClusterUtils.STORMS_SUBTREE, false));

                Set<String> activeTopologyBlobKeys = populateTopologyBlobKeys(activeTopologyIds);
                Set<String> activeTopologyCodeKeys = filterTopologyCodeKeys(activeTopologyBlobKeys);
                Set<String> allLocalBlobKeys = Sets.newHashSet(blobStore.listKeys());
                Set<String> allLocalTopologyBlobKeys = filterTopologyBlobKeys(allLocalBlobKeys);

                // this finds all active topologies blob keys from all local topology blob keys
                Sets.SetView<String> diffTopology = Sets.difference(activeTopologyBlobKeys, allLocalTopologyBlobKeys);
                LOG.info("active-topology-blobs [{}] local-topology-blobs [{}] diff-topology-blobs [{}]",
                        generateJoinedString(activeTopologyIds), generateJoinedString(allLocalTopologyBlobKeys),
                        generateJoinedString(diffTopology));

                if (diffTopology.isEmpty()) {
                    Set<String> activeTopologyDependencies = getTopologyDependencyKeys(activeTopologyCodeKeys);

                    // this finds all dependency blob keys from active topologies from all local blob keys
                    Sets.SetView<String> diffDependencies = Sets.difference(activeTopologyDependencies, allLocalBlobKeys);
                    LOG.info("active-topology-dependencies [{}] local-blobs [{}] diff-topology-dependencies [{}]",
                            generateJoinedString(activeTopologyDependencies), generateJoinedString(allLocalBlobKeys),
                            generateJoinedString(diffDependencies));

                    if (diffDependencies.isEmpty()) {
                        LOG.info("Accepting leadership, all active topologies and corresponding dependencies found locally.");
                    } else {
                        LOG.info("Code for all active topologies is available locally, but some dependencies are not found locally, giving up leadership.");
                        closeLatch();
                    }
                } else {
                    LOG.info("code for all active topologies not available locally, giving up leadership.");
                    closeLatch();
                }
            }

            @Override
            public void notLeader() {
                LOG.info("{} lost leadership.", hostName);
            }

            //......

            private void closeLatch() {
                try {
                    leaderLatch.close();
                } catch (IOException e) {
                    throw new RuntimeException(e);
                }
            }
        };
    }

• leaderLatchListenerImpl返回一个LeaderLatchListener接口的实现类
• isLeader接口里头作了一些校验，即当被zookeeper选中为leader的时候，若是本地没有全部的active topologies或者本地没有全部dependencies，那么就须要调用leaderLatch.close()放弃leadership
• notLeader接口主要打印一下log

LeaderElectorImp

org/apache/storm/zookeeper/LeaderElectorImp.java

public class LeaderElectorImp implements ILeaderElector {
    private static Logger LOG = LoggerFactory.getLogger(LeaderElectorImp.class);
    private final Map<String, Object> conf;
    private final List<String> servers;
    private final CuratorFramework zk;
    private final String leaderlockPath;
    private final String id;
    private final AtomicReference<LeaderLatch> leaderLatch;
    private final AtomicReference<LeaderLatchListener> leaderLatchListener;
    private final BlobStore blobStore;
    private final TopoCache tc;
    private final IStormClusterState clusterState;
    private final List<ACL> acls;
    private final StormMetricsRegistry metricsRegistry;

    public LeaderElectorImp(Map<String, Object> conf, List<String> servers, CuratorFramework zk, String leaderlockPath, String id,
                            AtomicReference<LeaderLatch> leaderLatch, AtomicReference<LeaderLatchListener> leaderLatchListener,
                            BlobStore blobStore, final TopoCache tc, IStormClusterState clusterState, List<ACL> acls,
                            StormMetricsRegistry metricsRegistry) {
        this.conf = conf;
        this.servers = servers;
        this.zk = zk;
        this.leaderlockPath = leaderlockPath;
        this.id = id;
        this.leaderLatch = leaderLatch;
        this.leaderLatchListener = leaderLatchListener;
        this.blobStore = blobStore;
        this.tc = tc;
        this.clusterState = clusterState;
        this.acls = acls;
        this.metricsRegistry = metricsRegistry;
    }

    @Override
    public void prepare(Map<String, Object> conf) {
        // no-op for zookeeper implementation
    }

    @Override
    public void addToLeaderLockQueue() throws Exception {
        // if this latch is already closed, we need to create new instance.
        if (LeaderLatch.State.CLOSED.equals(leaderLatch.get().getState())) {
            leaderLatch.set(new LeaderLatch(zk, leaderlockPath));
            LeaderListenerCallback callback = new LeaderListenerCallback(conf, zk, leaderLatch.get(), blobStore, tc, clusterState, acls,
                metricsRegistry);
            leaderLatchListener.set(Zookeeper.leaderLatchListenerImpl(callback));
            LOG.info("LeaderLatch was in closed state. Resetted the leaderLatch and listeners.");
        }
        // Only if the latch is not already started we invoke start
        if (LeaderLatch.State.LATENT.equals(leaderLatch.get().getState())) {
            leaderLatch.get().addListener(leaderLatchListener.get());
            leaderLatch.get().start();
            LOG.info("Queued up for leader lock.");
        } else {
            LOG.info("Node already in queue for leader lock.");
        }
    }

    @Override
    // Only started latches can be closed.
    public void removeFromLeaderLockQueue() throws Exception {
        if (LeaderLatch.State.STARTED.equals(leaderLatch.get().getState())) {
            leaderLatch.get().close();
            LOG.info("Removed from leader lock queue.");
        } else {
            LOG.info("leader latch is not started so no removeFromLeaderLockQueue needed.");
        }
    }

    @Override
    public boolean isLeader() throws Exception {
        return leaderLatch.get().hasLeadership();
    }

    @Override
    public NimbusInfo getLeader() {
        try {
            return Zookeeper.toNimbusInfo(leaderLatch.get().getLeader());
        } catch (Exception e) {
            throw Utils.wrapInRuntime(e);
        }
    }

    @Override
    public List<NimbusInfo> getAllNimbuses() throws Exception {
        List<NimbusInfo> nimbusInfos = new ArrayList<>();
        Collection<Participant> participants = leaderLatch.get().getParticipants();
        for (Participant participant : participants) {
            nimbusInfos.add(Zookeeper.toNimbusInfo(participant));
        }
        return nimbusInfos;
    }

    @Override
    public void close() {
        //Do nothing now.
    }
}

• LeaderElectorImp实现了ILeaderElector接口
• addToLeaderLockQueue方法检测若是latch已经closed，则从新建立一个新的，而后检测latch的状态，若是尚未start的话，则调用start参与选举
• 之因此对closed状态的latch建立一个，主要有两个缘由：一是对已经closed的latch进行方法调用会抛异常，二是被zk选举为leader，可是不满意storm的一些leader条件会放弃leadership即close掉

小结

• storm nimbus的LeaderElector主要是基于zookeeper recipies的LeaderLatch来实现
• storm nimbus自定义了LeaderLatchListener，对成为leader以后的nimbus进行校验，须要本地拥有全部的active topologies以及全部dependencies，不然放弃leadership

doc

• Highly Available Nimbus Design