elasticsearch源码分析-03选举集群状态

选举集群状态

es中存储的数据有一下几种，state元数据、lucene索引文件、translog事务日志
元数据信息可以分为：

• 集群层面的元信息-对应着metaData数据结构，主要是clusterUUid、settings、templates等
• 索引层面的元信息-对应着indexMetaData数据结构，主要存储分片数量、mappings索引字段映射等
• 分片层面的元信息-对应着shardStateMetaData，主要是version、indexUUid、主分片等
  每个节点可能会有不同的集群状态，需要选择正确的元数据作为权威源数据。状态信息的管理在gatewayService中，它实现了ClusterStateListener接口，当选择完主节点后会发布一个集群状态task，触发回调方法clusterChanged

//恢复分片分配状态
performStateRecovery(enforceRecoverAfterTime, reason);

集群层和索引层元数据恢复在gateway模块完成

public void clusterChanged(final ClusterChangedEvent event) {
   
    if (lifecycle.stoppedOrClosed()) {
   
        return;
    }

    final ClusterState state = event.state();
    //只有主节点才能执行
    if (state.nodes().isLocalNodeElectedMaster() == false) {
   
        // not our job to recover
        return;
    }
    //已经执行过了集群状态和索引状态恢复了
    if (state.blocks().hasGlobalBlock(STATE_NOT_RECOVERED_BLOCK) == false) {
   
        // already recovered
        return;
    }

    //这段省略主要是检查是否达到恢复状态条件
    ......
        //恢复状态
     performStateRecovery(enforceRecoverAfterTime, reason);
}

首先判断只有主节点可以执行状态选举，然后判断是否已经在执行了状态恢复任务了，如果是则直接返回；如果没有则执行恢复状态任务
最终会调用recoveryRunnable.run()

final Gateway gateway = new Gateway(settings, clusterService, listGatewayMetaState);
            recoveryRunnable = () ->
                    gateway.performStateRecovery(new GatewayRecoveryListener());

执行gateway的performStateRecovery方法
首先回去所有master资格的节点信息

//具有master资格的node节点
        final String[] nodesIds = clusterService.state().nodes().getMasterNodes().keys().toArray(String.class);

获取其他master节点的元数据

//获取集群及信息
        final TransportNodesListGatewayMetaState.NodesGatewayMetaState nodesState = listGatewayMetaState.list(nodesIds, null).actionGet();

这里我们看下TransportNodesListGatewayMetaState的构造函数

public TransportNodesListGatewayMetaState(ThreadPool threadPool, ClusterService clusterService, TransportService transportService,
                                          ActionFilters actionFilters, GatewayMetaState metaState) {
   
    super(ACTION_NAME, threadPool, clusterService, transportService, actionFilters,
          Request::new, NodeRequest::new, ThreadPool.Names.GENERIC, NodeGatewayMetaState.class);
    this.metaState = metaState;
}

 //注册action处理类
transportService.registerRequestHandler(actionName, executor, false, canTripCircuitBreaker, requestReader,new TransportHandler());

回到list方法，会调用doExecute方法

public ActionFuture<NodesGatewayMetaState> list(String[] nodesIds, @Nullable TimeValue timeout) {
   
    PlainActionFuture<NodesGatewayMetaState> future = PlainActionFuture.newFuture();
    execute(new Request(nodesIds).timeout(timeout), future);
    return future;
}

protected void doExecute(Task task, NodesRequest request, ActionListener<NodesResponse> listener) {
   
    //执行
    new AsyncAction(task, request, listener).start();
}

发送所有节点获取元数据

void start() {
   
    final DiscoveryNode[] nodes = request.concreteNodes();
    if (nodes.length == 0) {
   //没有需要获取数据的node
        // nothing to notify
        threadPool.generic().execute(() -> listener.onResponse(newResponse(request, responses)));
        return;
    }
    TransportRequestOptions.Builder builder = TransportRequestOptions.builder()