zenDiscovery和master选举
上一篇通过 ElectMasterService源码，分析了master选举的原理的大部分内容：master候选节点ID排序保证选举一致性及通过设置最小可见候选节点数目避免brain split。节点排序后选举只能保证局部一致性，如果发生节点接收到了错误的集群状态就会选举出错误的master，因此必须有其它措施来保证选举的一致性。这就是上一篇所提到的第二点：被选举的数量达到一定的数目同时自己也选举自己，这个节点才能成为master。这一点体现在zenDiscovery中，本篇将结合节点的发现过程进一步介绍master选举机制。
节点启动后首先启动join线程，join线程会寻找cluster的master节点，如果集群之前已经启动，并且运行良好，则试图连接集群的master节点，加入集群。否则（集群正在启动）选举master节点，如果自己被选为master，则向集群中其它节点发送一个集群状态更新的task，如果master是其它节点则试图加入该集群。
join的代码如下所示：
private void innerJoinCluster() { DiscoveryNode masterNode = null; final Thread currentThread = Thread.currentThread();
//一直阻塞直到找到master节点，在集群刚刚启动，或者集群master丢失的情况，这种阻塞能够保证集群一致性 while (masterNode == null && joinThreadControl.joinThreadActive(currentThread)) { masterNode = findMaster(); } //有可能自己会被选举为master（集群启动，或者加入时正在选举）
if (clusterService.localNode().equals(masterNode)) {
//如果本身是master，则需要向其它所有节点发送集群状态更新 clusterService.submitStateUpdateTask("zen-disco-join (elected_as_master)", Priority.IMMEDIATE, new ProcessedClusterStateNonMasterUpdateTask() { @Override public ClusterState execute(ClusterState currentState) { //选举时错误的，之前的master状态良好，则不更新状态，仍旧使用之前状态。 if (currentState.nodes().masterNode() != null) { return currentState; } DiscoveryNodes.Builder builder = new DiscoveryNodes.Builder(currentState.nodes()).masterNodeId(currentState.nodes().localNode().id()); // update the fact that we are the master... ClusterBlocks clusterBlocks = ClusterBlocks.builder().blocks(currentState.blocks()).removeGlobalBlock(discoverySettings.getNoMasterBlock()).build(); currentState = ClusterState.builder(currentState).nodes(builder).blocks(clusterBlocks).build(); // eagerly run reroute to remove dead nodes from routing table RoutingAllocation.Result result = allocationService.reroute(currentState); return ClusterState.builder(currentState).routingResult(result).build(); } @Override public void onFailure(String source, Throwable t) { logger.error("unexpected failure during [{}]", t, source); joinThreadControl.markThreadAsDoneAndStartNew(currentThread); } @Override public void clusterStateProcessed(String source, ClusterState oldState, ClusterState newState) { if (newState.nodes().localNodeMaster()) { // we only starts nodesFD if we are master (it may be that we received a cluster state while pinging) joinThreadControl.markThreadAsDone(currentThread); nodesFD.updateNodesAndPing(newState); // start the nodes FD } else { // if we're not a master it means another node published a cluster state while we were pinging // make sure we go through another pinging round and actively join it joinThreadControl.markThreadAsDoneAndStartNew(currentThread); } sendInitialStateEventIfNeeded(); long count = clusterJoinsCounter.incrementAndGet(); logger.trace("cluster joins counter set to [{}] (elected as master)", count); } }); } else { // 找到的节点不是我，试图连接该master final boolean success = joinElectedMaster(masterNode); // finalize join through the cluster state update thread final DiscoveryNode finalMasterNode = masterNode; clusterService.submitStateUpdateTask("finalize_join (" + masterNode + ")", new ClusterStateNonMasterUpdateTask() { @Override public ClusterState execute(ClusterState currentState) throws Exception { if (!success) { // failed to join. Try again... joinThreadControl.markThreadAsDoneAndStartNew(currentThread); return currentState; } if (currentState.getNodes().masterNode() == null) { // Post 1.3.0, the master should publish a new cluster state before acking our join request. we now should have // a valid master. logger.debug("no master node is set, despite of join request completing. retrying pings."); joinThreadControl.markThreadAsDoneAndStartNew(currentThread); return currentState; } if (!currentState.getNodes().masterNode().equals(finalMasterNode)) { return joinThreadControl.stopRunningThreadAndRejoin(currentState, "master_switched_while_finalizing_join"); } // Note: we do not have to start master fault detection here because it's set at {@link #handleNewClusterStateFromMaster } // when the first cluster state arrives. joinThreadControl.markThreadAsDone(currentThread); return currentState; } @Override public void onFailure(String source, @Nullable Throwable t) { logger.error("unexpected error while trying to finalize cluster join", t); joinThreadControl.markThreadAsDoneAndStartNew(currentThread); } }); } }
以上就是join的过程。zenDiscovery在启动时会启动一个join线程，这个线程调用了该方法。同时在节点离开，master丢失等情况下也会重启这一线程仍然运行join方法。接下来看一下findMaster这个方法。这个方法体现了master选举的机制。代码如下：
private DiscoveryNode findMaster() {
//ping集群中的节点 ZenPing.PingResponse[] fullPingResponses = pingService.pingAndWait(pingTimeout); if (fullPingResponses == null) {return null; }// 过滤所得到的ping响应，虑除client节点，单纯的data节点 List<ZenPing.PingResponse> pingResponses = Lists.newArrayList(); for (ZenPing.PingResponse pingResponse : fullPingResponses) { DiscoveryNode node = pingResponse.node(); if (masterElectionFilterClientNodes && (node.clientNode() || (!node.masterNode() && !node.dataNode()))) { // filter out the client node, which is a client node, or also one that is not data and not master (effectively, client) } else if (masterElectionFilterDataNodes && (!node.masterNode() && node.dataNode())) { // filter out data node that is not also master } else { pingResponses.add(pingResponse); } } final DiscoveryNode localNode = clusterService.localNode(); List<DiscoveryNode> pingMasters = newArrayList();
//获取所有ping响应中的master节点，如果master节点是节点本身则过滤掉。pingMasters列表结果要么为空（本节点是master）要么是同一个节点（出现不同节点则集群出现了问题
不过没关系，后面会进行选举） for (ZenPing.PingResponse pingResponse : pingResponses) { if (pingResponse.master() != null) { if (!localNode.equals(pingResponse.master())) { pingMasters.add(pingResponse.master()); } } } // nodes discovered during pinging Set<DiscoveryNode> activeNodes = Sets.newHashSet(); // nodes discovered who has previously been part of the cluster and do not ping for the very first time Set<DiscoveryNode> joinedOnceActiveNodes = Sets.newHashSet();
Version minimumPingVersion = localNode.version();
for (ZenPing.PingResponse pingResponse : pingResponses) {
activeNodes.add(pingResponse.node());
minimumPingVersion = Version.smallest(pingResponse.node().version(), minimumPingVersion);
if (pingResponse.hasJoinedOnce() != null && pingResponse.hasJoinedOnce()) {
joinedOnceActiveNodes.add(pingResponse.node());
}
}
//本节点暂时是master也要加入候选节点进行选举 if (localNode.masterNode()) { activeNodes.add(localNode); long joinsCounter = clusterJoinsCounter.get(); if (joinsCounter > 0) { logger.trace("adding local node to the list of active nodes who has previously joined the cluster (joins counter is [{}})", joinsCounter); joinedOnceActiveNodes.add(localNode); } }
//pingMasters为空，则本节点是master节点， if (pingMasters.isEmpty()) { if (electMaster.hasEnoughMasterNodes(activeNodes)) {//保证选举数量,说明有足够多的节点选举本节点为master，但是这还不够，本节点还需要再选举一次，如果
本次选举节点仍旧是自己，那么本节点才能成为master。这里就体现了master选举的第二条原则。 DiscoveryNode master = electMaster.electMaster(joinedOnceActiveNodes); if (master != null) { return master; } return electMaster.electMaster(activeNodes); } else { // if we don't have enough master nodes, we bail, because there are not enough master to elect from logger.trace("not enough master nodes [{}]", activeNodes); return null; } } else {
//pingMasters不为空（pingMasters列表中应该都是同一个节点），本节点没有被选举为master，那就接受之前的选举。 return electMaster.electMaster(pingMasters); } }
上面的重点部分都做了标注，就不再分析。除了findMaster方法，还有一个方法也体现了master选举，那就是handleMasterGone。下面是它的部分代码，提交master丢失task部分，
clusterService.submitStateUpdateTask("zen-disco-master_failed (" + masterNode + ")", Priority.IMMEDIATE, new ProcessedClusterStateNonMasterUpdateTask() {
@Override public ClusterState execute(ClusterState currentState) { //获取到当前集群状态下的所有节点 DiscoveryNodes discoveryNodes = DiscoveryNodes.builder(currentState.nodes()) // make sure the old master node, which has failed, is not part of the nodes we publish .remove(masterNode.id()) .masterNodeId(null).build(); //rejoin过程仍然是重复findMaster过程
if (rejoin) { return rejoin(ClusterState.builder(currentState).nodes(discoveryNodes).build(), "master left (reason = " + reason + ")"); } //无法达到选举数量，进行findMaster过程 if (!electMaster.hasEnoughMasterNodes(discoveryNodes)) { return rejoin(ClusterState.builder(currentState).nodes(discoveryNodes).build(), "not enough master nodes after master left (reason = " + reason + ")"); }
//在当前集群状态下，如果候选节点数量达到预期数量，那么选举出来的节点一定是同一个节点，因为所有的节点看到的集群states是一致的 final DiscoveryNode electedMaster = electMaster.electMaster(discoveryNodes); // elect master final DiscoveryNode localNode = currentState.nodes().localNode(); .... }
从以上的代码可以看到master选举节点的应用场景，无论是findMaster还是handlemasterGone,他们都保证了选举一致性。那就是所选节点数量必须要达到一定的数量，否则不能认为选举成功，进入等待环境。如果当前节点被其它节点选举为master，仍然要进行选举一次以保证选举的一致性。这样在保证了选举数量同时对候选节点排序从而保证选举的一致性。
发现和加入集群是zenDiscovery的主要功能，当然它还有一些其它功能，如处理节点离开（handleLeaveRequest），处理master发送的最小clustersates（handleNewClusterStateFromMaster）等功能。这里就不一一介绍，有兴趣请参考相关源码。
总结一下：本节结合zenDiscovery，分析了master选举的另外一部分内容。同时zenDiscovery是节点发现集群功能的集合，它主要功能是发现（选举）出集群的master节点，并试图加入集群。同时如果 本机是master还会处理节点的离开和节点丢失，如果不是master则会处理来自master的节点状态更新。
