问题

es集群只配置一个节点，client是否能够自动发现集群中的所有节点？是如何发现的？如下配置了一个节点：

单个node配置
es client如何做到负载均衡？
一个es node挂掉之后，es client如何摘掉该节点？
es client node检测分为两种模式，有什么不同?

核心类

TransportClient : es client对外API类
TransportClientNodesService : 维护node节点的类
ScheduledNodeSampler : 定期维护正常节点类
NettyTransport : 进行数据传输
NodeSampler : 节点嗅探类

Client初始化过程

初始化代码

Settings.Builder builder = Settings.settingsBuilder()                                   
                                   .put("cluster.name", clusterName)
                                   .put("client.transport.sniff", true);
Settings settings = builder.build(); 
TransportClient client = TransportClient.builder().settings(settings).build(); 
for (TransportAddress transportAddress : transportAddresses) {
      client.addTransportAddress(transportAddress);
}

ES 通过builder模式构造了基础的配置参数；
通过build构造了client，这个时候包括构造client、初始化ThreadPool、构造TransportClientNodesService、启动定时任务、定制化嗅探类型；
添加集群可用地址，比如我只配了集群中的一个节点；

构建client

调用build API

build code

其中，关于依赖注入的简单说明：Guice 是 Google 用于 Java™ 开发的开放源码依赖项注入框架（感兴趣的可以了解下，这里不做重点讲解），具体可参考如下：

Google Guice Started Guice 博客1 Guice 博客2

初始化TransportClientNodesService

在上一幅图的modules.createInjector对TransportClientNodesService进行实例化，在TransportClient进行注入，可以看到TransportClient里边的绝大部分API都是通过TransportClientNodesService进行代理的：

TransportClient code

Guice通过注解进行注入

Guice 注解注入code

在上图中：注入了集群名称、线程池等，重点是如下代码：该段代码选择了节点嗅探器的类型嗅探同一集群中的所有节点SniffNodesSampler或者是只关注配置文件配置的节点SimpleNodeSampler

if (this.settings.getAsBoolean("client.transport.sniff", false)) {
    this.nodesSampler = new SniffNodesSampler();
} else {
    this.nodesSampler = new SimpleNodeSampler();
}

特点 SniffNodesSampler：client会主动发现集群里的其他节点，会创建fully connect(什么叫fully connect?后边说) SimpleNodeSampler：ping listedNodes中的所有node，区别在于这里创建的都是light connect；其中TransportClientNodesService维护了三个节点存储数据结构：

// nodes that are added to be discovered
1 private volatileListlistedNodes= Collections.emptyList();
2 private volatileListnodes= Collections.emptyList();
3 private volatileListfilteredNodes= Collections.emptyList();

代表配置文件中主动加入的节点；
代表参与请求的节点；
过滤掉的不能进行请求处理的节点；

Client如何做到负载均衡

负载均衡code

如上图，我们发现每次 execute 的时候，是从 nodes 这个数据结构中获取节点，然后通过简单的 rouund-robbin 获取节点服务器，核心代码如下：

private final AtomicInteger randomNodeGenerator = new AtomicInteger();
......
private int getNodeNumber() {
    int index = randomNodeGenerator.incrementAndGet();  
    if (index < ) {
        index = ;
        randomNodeGenerator.set();
    }
    return index;
}

然后通过netty的channel将数据写入，核心代码如下：

public void sendRequest(final DiscoveryNode node, final long requestId, final String action, final TransportRequest request, TransportRequestOptions options) 
throws IOException, TransportException {
1 Channel targetChannel = nodeChannel(node, options);     
  if (compress) {        
      options = TransportRequestOptions.builder(options).withCompress(true).build();    
  }    
byte status = ;    
status = TransportStatus.setRequest(status);    
ReleasableBytesStreamOutput bStream = new ReleasableBytesStreamOutput(bigArrays);    
boolean addedReleaseListener = false;    
try {        
    bStream.skip(NettyHeader.HEADER_SIZE);        
    StreamOutput stream = bStream;        
    // only compress if asked, and, the request is not bytes, since then only
    // the header part is compressed, and the "body" can't be extracted as compressed      
    if (options.compress() && (!(request instanceof BytesTransportRequest))) {            
        status = TransportStatus.setCompress(status);            
        stream = CompressorFactory.defaultCompressor().streamOutput(stream);       
    }
    // we pick the smallest of the 2, to support both backward and forward compatibility
    // note, this is the only place we need to do this, since from here on, we use the serialized version
    // as the version to use also when the node receiving this request will send the     response with
    Version version = Version.smallest(this.version, node.version());
    stream.setVersion(version);
    stream.writeString(action);
    ReleasablePagedBytesReference bytes;
    ChannelBuffer buffer;
    // it might be nice to somehow generalize this optimization, maybe a smart "paged" bytes output        
    // that create paged channel buffers, but its tricky to know when to do it (where this option is
    // more explicit).      
    if (request instanceof BytesTransportRequest) {
        BytesTransportRequest bRequest = (BytesTransportRequest) request;  
        assert node.version().equals(bRequest.version());
        bRequest.writeThin(stream);
        stream.close();
        bytes = bStream.bytes();
        ChannelBuffer headerBuffer = bytes.toChannelBuffer();
        ChannelBuffer contentBuffer = bRequest.bytes().toChannelBuffer();
        buffer = ChannelBuffers.wrappedBuffer(NettyUtils.DEFAULT_GATHERING,         headerBuffer, contentBuffer);      
    } else {
        request.writeTo(stream);
        stream.close();
        bytes = bStream.bytes();            
        buffer = bytes.toChannelBuffer();
    }
    NettyHeader.writeHeader(buffer, requestId, status, version);
2 ChannelFuture future = targetChannel.write(buffer);        
  ReleaseChannelFutureListener listener= new ReleaseChannelFutureListener(bytes);
  future.addListener(listener);
  addedReleaseListener = true;
  transportServiceAdapter.onRequestSent(node, requestId, action, request, options);  
  } finally {      
    if (!addedReleaseListener) {
    Releasables.close(bStream.bytes());       
  }
  }
}

其中最重要的就是1和2

1代表拿到一个连接；
2代表通过拿到的连接写数据；

这时候就会有新的问题

nodes的数据是何时写入的？
连接是什么时候创建的？

Nodes数据何时写入

核心是调用doSampler，代码如下：

protected void doSample() {    
// the nodes we are going to ping include the core listed nodes that were added    
// and the last round of discovered nodes    
SetnodesToPing = Sets.newHashSet();   
for (DiscoveryNode node : listedNodes) {       
    nodesToPing.add(node);  
}
for (DiscoveryNode node : nodes) {        
    nodesToPing.add(node);    
}   
final CountDownLatch latch = new CountDownLatch(nodesToPing.size());    
final ConcurrentMapclusterStateResponses = ConcurrentCollections.newConcurrentMap();   
for (final DiscoveryNode listedNode : nodesToPing) {        
    threadPool.executor(ThreadPool.Names.MANAGEMENT).execute(new Runnable() {           
       @Override   
       public void run() {                
          try {
               if (!transportService.nodeConnected(listedNode)) {                        
                    try {                           
                     // if its one of the actual nodes we will talk to, not to listed nodes, fully connect           
                     if (nodes.contains(listedNode)) {                                        
                        logger.trace("connecting to cluster node [{}]", listedNode);                                    
                        transportService.connectToNode(listedNode);                            
                     } else {
                         // its a listed node, light connect to it...                                    
                        logger.trace("connecting to listed node (light) [{}]", listedNode);                                
                        transportService.connectToNodeLight(listedNode);                                      
                     }
                   } catch (Exception e) {                       
                     logger.debug("failed to connect to node [{}], ignoring...", e, listedNode);
                      latch.countDown();                        
                      return;                       
                   }                
                }
//核心是在这里，刚刚开始初始化的时候，可能只有配置的一个节点，这个时候会通过这个地址发送一个state状态监测                    
//"cluster:monitor/state"                    
transportService.sendRequest(listedNode, ClusterStateAction.NAME,                          
headers.applyTo(Requests.clusterStateRequest().clear().nodes(true).local(true)),                            
TransportRequestOptions.builder().withType(TransportRequestOptions.Type.STAE).withTimeout(pingTimeout).build(),                            
new BaseTransportResponseHandler() {                                
@Override                                
public ClusterStateResponse newInstance() {                                
    return new ClusterStateResponse();                                
}                                
@Override                                
public String executor() {                                    
    return ThreadPool.Names.SAME;                               
}                                
@Override                              
public void handleResponse(ClusterStateResponse response) {
/*通过回调，会在这个地方返回集群中类似下边所有节点的信息
{  "version" : 27,  "state_uuid" : "YSI9d_HiQJ-FFAtGFCVOlw",  "master_node" : "TXHHx-XRQaiXAxtP1EzXMw",  "blocks" : { },  "nodes" : {    "poxubF0LTVue84GMrZ7rwA" : {      "name" : "node1",      "transport_address" : "1.1.1.1:8888",      "attributes" : {        "data" : "false",        "master" : "true"      }    },    "9Cz8m3GkTza7vgmpf3L65Q" : {      "name" : "node2",      "transport_address" : "1.1.1.2:8889",      "attributes" : {        "master" : "false"      }    }  },  "metadata" : {    "cluster_uuid" : "_na_",    "templates" : { },    "indices" : { }  },  "routing_table" : {    "indices" : { }  },  "routing_nodes" : {    "unassigned" : [ ],    "nodes" : {      "lZqD-WExRu-gaSUiCXaJcg" : [ ],      "hR6PbFrgQVSY0MHajNDmgA" : [ ],    }  }}*/                                    
clusterStateResponses.put(listedNode, response);                                  
latch.countDown();                                
}                                
@Override                                
public void handleException(TransportException e) {                                    logger.info("failed to get local cluster state for {}, disconnecting...", e, listedNode);                                    transportService.disconnectFromNode(listedNode);                                    latch.countDown();                                
}                            
});} catch (Throwable e) {                    
logger.info("failed to get local cluster state info for {}, disconnecting...", e, listedNode);                    
transportService.disconnectFromNode(listedNode);                    latch.countDown();                
}}});}    
try {     
   latch.await();    
} catch (InterruptedException e) {       
 return;    
}    
HashSetnewNodes = new HashSet<>();    HashSetnewFilteredNodes = new HashSet<>();   
for (Map.Entryentry : clusterStateResponses.entrySet()) {      
    if (!ignoreClusterName &&!clusterName.equals(entry.getValue().getClusterName())) {            
        logger.warn("node {} not part of the cluster {}, ignoring...",     
        entry.getValue().getState().nodes().localNode(), clusterName);            
        newFilteredNodes.add(entry.getKey());            
        continue;        
}
//接下来在这个地方拿到所有的data nodes 写入到nodes节点里边       
 for (ObjectCursorcursor : entry.getValue().getState().nodes().dataNodes().values()){
    newNodes.add(cursor.value);}}
    nodes = validateNewNodes(newNodes);
    filteredNodes = Collections.unmodifiableList(new ArrayList<(newFilteredNodes));
  }

其中调用时机分为两部分：

client.addTransportAddress(transportAddress);
ScheduledNodeSampler，默认每隔5s会进行一次对各个节点的请求操作；

连接是何时创建的呢

也是在doSampler调用，最终由NettryTransport创建

创建连接code

这个时候发现，如果是light则创建轻连接，也就是，否则创建fully connect，其中包括： recovery：做数据恢复recovery，默认个数2个；

bulk：用于bulk请求，默认个数3个；
med/reg：典型的搜索和单doc索引，默认个数6个；
high:如集群state的发送等，默认个数1个；
ping：就是node之间的ping咯。默认个数1个；

对应的代码为：

public void start() {    
    List<Channel> newAllChannels = new ArrayList<>();    
    newAllChannels.addAll(Arrays.asList(recovery));    
    newAllChannels.addAll(Arrays.asList(bulk));    
    newAllChannels.addAll(Arrays.asList(reg));    
    newAllChannels.addAll(Arrays.asList(state));    
    newAllChannels.addAll(Arrays.asList(ping));    
    this.allChannels = Collections.unmodifiableList(newAllChannels);
}

END

作者：凌渡冰链接：https://www.jianshu.com/p/4a54122aefe7来源：简书著作权归作者所有。商业转载请联系作者获得授权，非商业转载请注明出处。

ITPub博客

Elasticsearch2.3.3 java client search源码分析

问题

核心类