[原始碼閱讀] 阿里SOFA服務註冊中心MetaServer(3)

羅西的思考發表於2020-10-15

[原始碼閱讀] 阿里SOFA服務註冊中心MetaServer(3)

0x00 摘要

SOFARegistry 是螞蟻金服開源的一個生產級、高時效、高可用的服務註冊中心。本系列將帶領大家一起分析其MetaServer的實現機制。本文為第三篇,介紹MetaServer如何基於raft實現了資料一致性。

因為篇幅限制,本文不會涉及 Raft 和 JRaft 的原理和實現,只是講解MetaServer如何基於 JRaft 的實現。

0x01 概念

1.1 分散式一致性

分散式一致性 (distributed consensus) 是分散式系統中最基本的問題,用來保證一個分散式系統的可靠性以及容災能力。

簡單的來講,就是如何在多個機器間對某一個值達成一致,並且當達成一致之後,無論之後這些機器間發生怎樣的故障,這個值能保持不變。抽象定義上, 一個分散式系統裡的所有程式要確定一個值 v,如果這個系統滿足如下幾個性質,就可以認為它解決了分散式一致性問題,幾個性質分別是 :

  • Termination: 所有正常程式都會決定 v 具體的值,不會存在一直在迴圈的程式。

  • Validity: 任何正常程式如果有一個確定的值 v’,那麼 v’ 肯定是某個程式提交的。比如隨機數生成器就不滿足這個性質。

  • Agreement: 所有正常程式選擇的值都是一樣的。

1.2 SOFAJRaft

SOFAJRaft 是一個基於 Raft 一致性演算法的生產級高效能 Java 實現,支援 MULTI-RAFT-GROUP,適用於高負載低延遲的場景。

因為 SOFARegistry 叢集節點列表資料並不是很多,因此不需要使用資料分片的方式在 MetaServer 中儲存。叢集節點列表儲存在 Repository 中,通過 Raft 強一致性協議對外提供節點註冊、續約、列表查詢等 Bolt 請求,從而保障叢集獲得的資料是強一致性的。

0x02 基礎架構

這裡的基礎架構指的是 :基於JRaft之上的,在SOFARegistry之中構建的基礎架構,包括StateMachine,Handler,RaftServer,RaftClient 等等。

2.1 RaftExchanger

Exchange 作為 Client / Server 連線的抽象,負責節點之間的連線。RaftExchanger就是Raft協議連線的抽象,可以看到其中包含配置,Registry和Raft元件。

public class RaftExchanger {
    @Autowired
    private MetaServerConfig    metaServerConfig;
    @Autowired
    private NodeConfig          nodeConfig;
    @Autowired
    private Registry            metaServerRegistry;
    private RaftServer          raftServer;
    private RaftClient          raftClient;
    private CliService          cliService;
}

在系統啟動時候,會把Raft這幾個元件啟動。

private void initRaft() {
    raftExchanger.startRaftServer(executorManager);
    raftExchanger.startRaftClient();
    raftExchanger.startCliService();
}

2.2 RaftServer

RaftServer是Raft協議的服務端,主要有如下成員或者行為:

  • 啟動一個 raft node節點,提供分散式服務。
  • 內部使用 jraft 提供的 RaftGroupService 服務框架。
  • fsm是業務狀態機,其實現類是ServiceStateMachine,其行為handler針對Leader和follower分別有leaderProcessListener 和 followerProcessListener。
  • boltServer 是Blot服務。因為JRraft基於bolt,所以設定了RaftServerHandlerRaftServerConnectionHandler

具體類實現如下:

public class RaftServer { 
    private RaftGroupService        raftGroupService; // jraft 服務端服務框架
    private Node                    node; // raft 節點
    private ServiceStateMachine     fsm; // 業務狀態機
    private PeerId                  serverId;
    private Configuration           initConf;
    private String                  groupId;
    private String                  dataPath;
    private List<ChannelHandler>    serverHandlers = new ArrayList<>();
    private LeaderProcessListener   leaderProcessListener;
    private FollowerProcessListener followerProcessListener;
    private BoltServer              boltServer;	
    
    public void start(RaftServerConfig raftServerConfig) throws IOException {

        FileUtils.forceMkdir(new File(dataPath));

        // 構建服務端,設定handler
        serverHandlers.add(new RaftServerHandler(this));
        serverHandlers.add(new RaftServerConnectionHandler());
        boltServer = new BoltServer(new URL(NetUtil.getLocalAddress().getHostAddress(),
            serverId.getPort()), serverHandlers);

        // 啟動服務端
        boltServer.initServer();

        RpcServer rpcServer = boltServer.getRpcServer();

        RaftRpcServerFactory.addRaftRequestProcessors(rpcServer);

        // 設定狀態機的handler
        this.fsm = ServiceStateMachine.getInstance();
        this.fsm.setLeaderProcessListener(leaderProcessListener);
        this.fsm.setFollowerProcessListener(followerProcessListener);

        NodeOptions nodeOptions = initNodeOptions(raftServerConfig);

        this.raftGroupService = new RaftGroupService(groupId, serverId, nodeOptions, rpcServer);
        //start
        this.node = this.raftGroupService.start();

        // 啟動客戶端
        RpcClient raftClient = ((AbstractBoltClientService) (((NodeImpl) node).getRpcService()))
            .getRpcClient();

        NotifyLeaderChangeHandler notifyLeaderChangeHandler = new NotifyLeaderChangeHandler(
            groupId, null);
        raftClient.registerUserProcessor(new SyncUserProcessorAdapter(notifyLeaderChangeHandler));
    }
    
}	

2.2.1 RaftServerHandler

RaftServerHandler是服務端相應handler,首先接受Bolt訊息,然後轉化成 processRequest,傳送給node。

received:84, RaftServerHandler (com.alipay.sofa.registry.jraft.handler)
handleRequest:55, AsyncUserProcessorAdapter (com.alipay.sofa.registry.remoting.bolt)
dispatchToUserProcessor:224, RpcRequestProcessor (com.alipay.remoting.rpc.protocol)
doProcess:145, RpcRequestProcessor (com.alipay.remoting.rpc.protocol)
run:366, RpcRequestProcessor$ProcessTask (com.alipay.remoting.rpc.protocol)
runWorker:1149, ThreadPoolExecutor (java.util.concurrent)
run:624, ThreadPoolExecutor$Worker (java.util.concurrent)
run:748, Thread (java.lang)

RaftServerHandler 會根據本身是Leader還是Follower做不同處理。

然後會在createTask之中進行 Hessian協議 處理,進而呼叫 raftServer.getNode().apply(task);

大致邏輯如下:

  • 從訊息中解析出請求;
  • 根據請求解析出對應的處理函式;
  • 如果就是簡單讀取,就直接處理,然後返回;
  • 如果需要task處理,就生成closure;
    • 生成處理closure的task;
    • 執行task;

具體程式碼如下:

public class RaftServerHandler implements ChannelHandler {
    protected RaftServer        raftServer;
    
    @Override
    public void received(Channel channel, Object message) throws RemotingException {

        BoltChannel boltChannel = (BoltChannel) channel;
        AsyncContext asyncContext = boltChannel.getAsyncContext();

        if (!raftServer.getFsm().isLeader()) {
            asyncContext.sendResponse(ProcessResponse.redirect(raftServer.redirect()).build());
            return;
        }
        
        // 從訊息中解析出請求
        ProcessRequest processRequest = (ProcessRequest) message;
        long start = System.currentTimeMillis();
        // 根據請求解析出對應的處理函式
        Method method = Processor.getInstance().getWorkMethod(processRequest);

        if (Processor.getInstance().isLeaderReadMethod(method)) {
            // 如果就是簡單讀取,就直接處理,然後返回
            Object obj = Processor.getInstance().process(method, processRequest);
            long cost = System.currentTimeMillis() - start;
            asyncContext.sendResponse(obj);
        } else {
            // 如果需要task處理,就生成closure
            LeaderTaskClosure closure = new LeaderTaskClosure();
            closure.setRequest(processRequest);
            closure.setDone(status -> {
                long cost = System.currentTimeMillis() - start;
                if (status.isOk()) {
                    asyncContext.sendResponse(closure.getResponse());
                } else {
                    asyncContext.sendResponse(ProcessResponse.fail(status.getErrorMsg()).build());
                }
            });

            // 生成處理closure的task
            Task task = createTask(closure, processRequest);
            // 執行task
            raftServer.getNode().apply(task);
        }
    }  
}

2.2.2 ServiceStateMachine

ServiceStateMachine 是服務端的狀態機,MetaServer這裡主要是實現核心的 onApply(iterator) 方法,應用使用者提交的請求到Processor處理。

關於快照的部分我們會在後續講解。

public class ServiceStateMachine extends StateMachineAdapter {;

    private LeaderProcessListener               leaderProcessListener;
    private FollowerProcessListener             followerProcessListener;
    private static volatile ServiceStateMachine instance;
  
    @Override
    public void onApply(Iterator iter) {
        while (iter.hasNext()) {
            Closure done = iter.done();
            ByteBuffer data = iter.getData();
            ProcessRequest request;
            LeaderTaskClosure closure = null;

            if (done != null) {
                closure = (LeaderTaskClosure) done;
                request = closure.getRequest();
            } else {
                Hessian2Input input = new Hessian2Input(new ByteArrayInputStream(data.array()));
                SerializerFactory serializerFactory = new SerializerFactory();
                input.setSerializerFactory(serializerFactory);
                request = (ProcessRequest) input.readObject();
                input.close();
            }

            ProcessResponse response = Processor.getInstance().process(request);
            if (closure != null) {
                closure.setResponse(response);
                closure.run(Status.OK());
            }
            iter.next();
        }
    }
}                                                              

2.3 RaftClient

客戶端 Client 比較簡單,主要使用 jraft 提供的 RouteTable 來重新整理獲取最新的 leader 節點,然後傳送請求到 leader節點。

public class RaftClient {
    private BoltCliClientService cliClientService;
    private RpcClient            rpcClient;
    private CliOptions           cliOptions;
    private String               groupId;
    private Configuration        conf;
}

0x03 相關配置

JRaft相關配置主要是在 MetaServerRepositoryConfiguration 之中完成的。

因為各種節點列表是儲存在Repository之中,而Repository是由JRaft來保證資料一致性,所以配置中主要是和Repository相關,比如三個RepositoryService。

  • dataRepositoryService
  • metaRepositoryService
  • sessionRepositoryService

其次是Session版本服務和兩個Confirm服務

  • SessionVersionRepositoryService
  • DataConfirmStatusService
  • SessionConfirmStatusService

然後是RaftExchanger,這是一個網路互動的抽象。

最後是RaftAnnotationBeanPostProcessor,這是用來在執行時候處理Bean。

程式碼如下:

@Configuration
public static class MetaServerRepositoryConfiguration {
    @Bean
    public RepositoryService dataRepositoryService() {
        return new DataRepositoryService();
    }

    @Bean
    public RepositoryService metaRepositoryService() {
        return new MetaRepositoryService();
    }

    @Bean
    public RepositoryService sessionRepositoryService() {
        return new SessionRepositoryService();
    }
  
    @Bean
    public VersionRepositoryService sessionVersionRepositoryService() {
        return new SessionVersionRepositoryService();
    }
  
    @Bean
    public NodeConfirmStatusService dataConfirmStatusService() {
        return new DataConfirmStatusService();
    }

    @Bean
    public NodeConfirmStatusService sessionConfirmStatusService() {
        return new SessionConfirmStatusService();
    }

    @Bean
    public RaftExchanger raftExchanger() {
        return new RaftExchanger();
    }

    @Bean
    public RaftAnnotationBeanPostProcessor raftAnnotationBeanPostProcessor() {
        return new RaftAnnotationBeanPostProcessor();
    }
}

另外,MetaDBConfiguration 也實現了一個Bean。

@Configuration
public static class MetaDBConfiguration {
    @Bean
    public DBService persistenceDataDBService() {
        return new PersistenceDataDBService();
    }
}

3.1 RepositoryService介面

因為Raft主要作用於RepositoryService介面,所以首先講解RepositoryService介面。

針對Repository所有的操作都是直接呼叫的 RepositoryService 等介面,DataRepositoryService 等類實現了這個介面。

@RaftService(uniqueId = "dataServer")
public class DataRepositoryService extends AbstractSnapshotProcess
                                   implements
                                   RepositoryService<String, RenewDecorate<DataNode>> {
}

比如 DataStoreService 就會直接呼叫 dataRepositoryService進行各種操作。

public class DataStoreService implements StoreService<DataNode> {
    @RaftReference(uniqueId = "dataServer")
    private RepositoryService<String, RenewDecorate<DataNode>> dataRepositoryService;
  
    ......
    dataRepositoryService.replaceAll(dataCenter, dataCenterNodesMapTemp, version);
    ......
}

3.2 RaftReference & RaftService

這兩個註解可以認為是封裝好Raft的從而呈現給Registry的介面。RaftReference 對應了客戶端代理,RaftService對應著服務端的實現

為什麼要這麼做?因為需要維護資料一致性,所以必須把單純的本地呼叫轉換為非同步網路呼叫,這樣才能用raft協議保證資料一致性

3.2.1 註解定義

RaftService定義如下:

@Retention(RetentionPolicy.RUNTIME)
@Target(ElementType.TYPE)
public @interface RaftService {
    Class<?> interfaceType() default void.class;
    String uniqueId() default "";
}

RaftReference定義如下:

@Retention(RetentionPolicy.RUNTIME)
@Target(ElementType.FIELD)
public @interface RaftReference {
    Class<?> interfaceType() default void.class;
    String uniqueId() default "";
}

3.2.2 註解使用

凡是需要由Raft控制的服務都加上了RaftService這個註解。

  • dataRepositoryService
  • metaRepositoryService
  • sessionRepositoryService
  • SessionVersionRepositoryService
  • DataConfirmStatusService
  • SessionConfirmStatusService
  • PersistenceDataDBService

凡是 RaftService具體相關實現類都加了 @RaftReference 註解,因為根據id進行區分,所以有些服務設定了uniqueId。

    @RaftReference
    private DBService           persistenceDataDBService;

    @RaftReference(uniqueId = "dataServer")
    private RepositoryService<String, RenewDecorate<DataNode>> dataRepositoryService;
    @RaftReference(uniqueId = "dataServer")
    private NodeConfirmStatusService<DataNode>                 dataConfirmStatusService;

    @RaftReference(uniqueId = "metaServer")
    private RepositoryService<String, RenewDecorate<MetaNode>> metaRepositoryService;

	  @RaftReference(uniqueId = "sessionServer")
    private RepositoryService<String, RenewDecorate<SessionNode>> sessionRepositoryService;
    @RaftReference(uniqueId = "sessionServer")
    private VersionRepositoryService<String>                      sessionVersionRepositoryService;
    @RaftReference(uniqueId = "sessionServer")
    private NodeConfirmStatusService<SessionNode>                 sessionConfirmStatusService;

3.2.3 註解實現

RaftAnnotationBeanPostProcessor 是 BeanPostProcessor 的實現,在這裡就對 RaftReference & RaftService 這兩個註解進行了處理。

BeanPostProcessor介面作用如下:如果我們想在Spring容器中完成bean例項化、配置以及其他初始化方法前後要新增一些自己邏輯處理。我們需要定義一個或多個BeanPostProcessor介面實現類,然後註冊到Spring IoC容器中。

public class RaftAnnotationBeanPostProcessor implements BeanPostProcessor, Ordered {
    @Autowired
    private RaftExchanger       raftExchanger;
    @Override
    public Object postProcessBeforeInitialization(Object bean, String beanName) {
        processRaftReference(bean);
        return bean;
    }
    @Override
    public Object postProcessAfterInitialization(Object bean, String beanName) {
        processRaftService(bean, beanName);
        return bean;
    }  
}

對於兩個註解,有不同的處理方式。

3.2.3.1 客戶端processRaftReference

針對processRaftReference的處理就是:把加了 @RaftReference 註解的屬性替換成動態代理,進而替換成客戶端呼叫。即在 processRaftReference 方法中,凡是加了 @RaftReference 註解的屬性,都會被動態代理類替換,其代理實現見 ProxyHandler 類,即將方法呼叫,封裝為 ProcessRequest,通過 RaftClient 傳送給 RaftServer

private void processRaftReference(Object bean) {
    final Class<?> beanClass = bean.getClass();

    ReflectionUtils.doWithFields(beanClass, field -> {
        RaftReference referenceAnnotation = field.getAnnotation(RaftReference.class);

        Class<?> interfaceType = referenceAnnotation.interfaceType();
        String serviceId = getServiceId(interfaceType, referenceAnnotation.uniqueId());
        Object proxy = getProxy(interfaceType, serviceId);
        ReflectionUtils.makeAccessible(field);
        ReflectionUtils.setField(field, bean, proxy); // 設定代理

    }, field -> !Modifier.isStatic(field.getModifiers())
            && field.isAnnotationPresent(RaftReference.class));
}

private Object getProxy(Class<?> interfaceType, String serviceId) {
        RaftClient client = raftExchanger.getRaftClient();
        return Proxy.newProxyInstance(Thread.currentThread().getContextClassLoader(),
            new Class<?>[] { interfaceType }, new ProxyHandler(interfaceType, serviceId,
                raftExchanger.getRaftClient()));
}

field = {Field@3824} "private com.alipay.sofa.registry.store.api.DBService com.alipay.sofa.registry.server.meta.remoting.handler.FetchProvideDataRequestHandler.persistenceDataDBService"
referenceAnnotation = {$Proxy42@3825} "@com.alipay.sofa.registry.store.api.annotation.RaftReference(interfaceType=void, uniqueId=)"
interfaceType = {Class@3150} "interface com.alipay.sofa.registry.store.api.DBService"
serviceId = "com.alipay.sofa.registry.store.api.DBService"

假設 DataStoreService ,在註解處理之前是:

bean = {DataStoreService@4053} 
 dataRepositoryService = null 

+-----------------------------+
|     DataStoreService        |
|                             |
|  +-----------------------+  |
|  | dataRepositoryService +---------> Null
|  +-----------------------+  |
+-----------------------------+

註解處理之後是

bean = {DataStoreService@4053} 
 dataRepositoryService = {$Proxy46@4057} Method threw 'java.lang.RuntimeException' exception. Cannot evaluate com.sun.proxy.$Proxy46.toString()
   
   
proxy = {$Proxy46@4057} Method threw 'java.lang.RuntimeException' exception. Cannot evaluate com.sun.proxy.$Proxy46.toString()
 h = {ProxyHandler@4098} 
  interfaceType = {Class@4018} "interface com.alipay.sofa.registry.server.meta.repository.RepositoryService"
  serviceId = "com.alipay.sofa.registry.server.meta.repository.RepositoryService:dataServer"
  client = {RaftClient@4077} 
   cliClientService = {BoltCliClientService@4144} 
   rpcClient = {RpcClient@4145} 
   cliOptions = {CliOptions@4146} "RpcOptions{rpcConnectTimeoutMs=1000, rpcDefaultTimeout=5000, rpcInstallSnapshotTimeout=300000, rpcProcessorThreadPoolSize=80, enableRpcChecksum=false, metricRegistry=null}"
   groupId = "RegistryGroup_DefaultDataCenter"
   conf = {Configuration@4148} "192.168.1.2:9614"
   started = {AtomicBoolean@4149} "true"   

即如下圖

+-----------------------------+
|     DataStoreService        |                  +-------------------+
|                             |                  |   ProxyHandler    |
|  +-----------------------+  |       +-----+    | +---------------+ |
|  | dataRepositoryService +--------->+Proxy+--->+ | interfaceType | |
|  +-----------------------+  |       +-----+    | |               | |
+-----------------------------+                  | | serviceId     | |
                                                 | |               | |
                                                 | | RpcClient     | |
                                                 | +---------------+ |
                                                 +-------------------+

這樣就被動態轉移到了ProxyHandler,這樣如果呼叫成員函式,就會通過rpc進行呼叫。

3.2.3.2 服務端processRaftService

針對processRaftService的處理就是,把加了 @RaftService 的類對映成 Processor 類,進而實現為SOFAJRaft 的狀態機 ServiceStateMachine即被加了 @RaftService 的類會被新增到 Processor 類 中,通過 serviceId(interfaceName + uniqueId) 進行區分。

RaftServer 收到請求後,會把它生效到 SOFAJRaft 的狀態機,具體實現類為 ServiceStateMachine,即會呼叫 Processor 方法,通過 serviceId 找到這個實現類,執行對應的方法呼叫。

private void processRaftService(Object bean, String beanName) {
    final Class<?> beanClass = AopProxyUtils.ultimateTargetClass(bean);
    RaftService raftServiceAnnotation = beanClass.getAnnotation(RaftService.class);
    Class<?> interfaceType = raftServiceAnnotation.interfaceType();

    String serviceUniqueId = getServiceId(interfaceType, raftServiceAnnotation.uniqueId());
    Processor.getInstance().addWorker(serviceUniqueId, interfaceType, bean);
}

其中部分變數如下:
bean = {DataRepositoryService@3805} 
beanName = "dataRepositoryService"
beanClass = {Class@3796} "class com.alipay.sofa.registry.server.meta.repository.service.DataRepositoryService"
raftServiceAnnotation = {$Proxy41@3807} "@com.alipay.sofa.registry.store.api.annotation.RaftService(interfaceType=void, uniqueId=dataServer)"
interfaceType = {Class@3795} "interface com.alipay.sofa.registry.server.meta.repository.RepositoryService"
serviceUniqueId = "com.alipay.sofa.registry.server.meta.repository.RepositoryService:dataServer"

在處理註解時候,通過 addWorker 來把類和成員變數設定到map中。注意 workerMethods 是一個雙層HashMap,第一層是以服務名為key,value是一個HashMap,第二層是以函式名為key,具體函式為value。

public void addWorker(String serviceId, Class interfaceClazz, Object target) {
    Map<String, Method> publicMethods = new HashMap();
    for (Method m : interfaceClazz.getMethods()) {
        StringBuilder mSigs = new StringBuilder();
        mSigs.append(m.getName());
        for (Class<?> paramType : m.getParameterTypes()) {
            mSigs.append(paramType.getName());
        }
        publicMethods.put(mSigs.toString(), m);
    }

    workerMethods.put(serviceId, publicMethods);
    workers.put(serviceId, target);
}

serviceId = "com.alipay.sofa.registry.store.api.DBService"
interfaceClazz = {Class@3118} "interface com.alipay.sofa.registry.store.api.DBService"
target = {PersistenceDataDBService@3124} 

this = {Processor@3812} 
 workerMethods = {HashMap@3815}  size = 2
  "com.alipay.sofa.registry.server.meta.repository.RepositoryService:dataServer" -> {HashMap@3813}  size = 13
   key = "com.alipay.sofa.registry.server.meta.repository.RepositoryService:dataServer"
   value = {HashMap@3813}  size = 13
    "getNodeRepositories" -> {Method@3856} "public abstract java.util.Map com.alipay.sofa.registry.server.meta.repository.RepositoryService.getNodeRepositories()"
    "replaceAlljava.lang.Stringjava.util.Mapjava.lang.Long" -> {Method@3858} "public abstract java.util.Map com.alipay.sofa.registry.server.meta.repository.RepositoryService.replaceAll(java.lang.String,java.util.Map,java.lang.Long)"
    "checkVersionjava.lang.Objectjava.lang.Long" -> {Method@3860} "public abstract boolean com.alipay.sofa.registry.server.meta.repository.RepositoryService.checkVersion(java.lang.Object,java.lang.Long)"
    "replacejava.lang.Objectjava.lang.Object" -> {Method@3862} "public default java.lang.Object com.alipay.sofa.registry.server.meta.repository.RepositoryService.replace(java.lang.Object,java.lang.Object)"
    "removejava.lang.Object" -> {Method@3864} "public default java.lang.Object com.alipay.sofa.registry.server.meta.repository.RepositoryService.remove(java.lang.Object)"
    "putjava.lang.Objectjava.lang.Objectjava.lang.Long" -> {Method@3866} "public abstract java.lang.Object com.alipay.sofa.registry.server.meta.repository.RepositoryService.put(java.lang.Object,java.lang.Object,java.lang.Long)"
    "getVersionjava.lang.Object" -> {Method@3868} "public abstract java.lang.Long com.alipay.sofa.registry.server.meta.repository.RepositoryService.getVersion(java.lang.Object)"
    "putjava.lang.Objectjava.lang.Object" -> {Method@3870} "public default java.lang.Object com.alipay.sofa.registry.server.meta.repository.RepositoryService.put(java.lang.Object,java.lang.Object)"
    "getAllData" -> {Method@3872} "public abstract java.util.Map com.alipay.sofa.registry.server.meta.repository.RepositoryService.getAllData()"
    "removejava.lang.Objectjava.lang.Long" -> {Method@3874} "public abstract java.lang.Object com.alipay.sofa.registry.server.meta.repository.RepositoryService.remove(java.lang.Object,java.lang.Long)"
    "getjava.lang.Object" -> {Method@3876} "public abstract java.lang.Object com.alipay.sofa.registry.server.meta.repository.RepositoryService.get(java.lang.Object)"
    "getAllDataMap" -> {Method@3878} "public abstract java.util.Map com.alipay.sofa.registry.server.meta.repository.RepositoryService.getAllDataMap()"
    "replacejava.lang.Objectjava.lang.Objectjava.lang.Long" -> {Method@3880} "public abstract java.lang.Object com.alipay.sofa.registry.server.meta.repository.RepositoryService.replace(java.lang.Object,java.lang.Object,java.lang.Long)"
  "com.alipay.sofa.registry.store.api.DBService" -> {HashMap@3828}  size = 5
   
 workers = {HashMap@3814}  size = 2
  "com.alipay.sofa.registry.server.meta.repository.RepositoryService:dataServer" -> {DataRepositoryService@3805} 
   key = "com.alipay.sofa.registry.server.meta.repository.RepositoryService:dataServer"
   value = {DataRepositoryService@3805} 
  "com.alipay.sofa.registry.store.api.DBService" -> {PersistenceDataDBService@3117} 
   key = "com.alipay.sofa.registry.store.api.DBService"
   value = {PersistenceDataDBService@3117} 
 methodHandleMap = {ConcurrentHashMap@3817}  size = 0

即如下圖所示:

                                        +-------------------+
                                        |      Processor    |        +> DataRepositoryService              +-> getNodeRepositories +-> getNodeRepositories()
                                        | +---------------+ |        |                                     |
                                        | |  workers+--------------->+> PersistenceDataDBService           +-> replaceAll+--> replaceAll
                                        | |               | |                                              |
+-----------------------+  addWorker    | |               | |                      {HashMap}               +-> checkVersion +--> ...
| DataRepositoryService +-------------->+ | {HashMap}     | |       +---> +----------------------------+   |
+-----------------------+               | | workerMethods +-------->+     |RepositoryService:dataServer+-->-->-replace +----->  ...
                                        | +---------------+ |       |     +----------------------------+   |
                                        +-------------------+       |                                      +-> remove  +---->  ...
                                                                    |                                      |
                                                                    +--->  ......                          +-> put  +-----> ...
                                                                    |                                      |
                                                                    |                                      +-> getAllDataMap +----> ...
                                                                    |                                      |
                                                                    |       {HashMap}       +--->   ...    +-> ......
                                                                    |     +---------+       |
                                                                    +---> |DBService+----------->   ...
                                                                          +---------+       |
                                                                                            |
                                                                                            +--->   ...

手機上參見如下:

其呼叫棧如下:

addWorker:69, Processor (com.alipay.sofa.registry.jraft.processor)
processRaftService:123, RaftAnnotationBeanPostProcessor (com.alipay.sofa.registry.server.meta.repository.annotation)
postProcessAfterInitialization:60, RaftAnnotationBeanPostProcessor (com.alipay.sofa.registry.server.meta.repository.annotation)
applyBeanPostProcessorsAfterInitialization:421, AbstractAutowireCapableBeanFactory (org.springframework.beans.factory.support)
initializeBean:1635, AbstractAutowireCapableBeanFactory (org.springframework.beans.factory.support)
doCreateBean:553, AbstractAutowireCapableBeanFactory 

以上過程其實和 RPC 呼叫非常類似,在引用方發起的方法呼叫,並不會真正的執行方法,而是封裝成請求傳送到 Raft 服務,由 Raft 狀態機進行真正的方法呼叫,比如把節點資訊儲存到 Map 中。所有節點之間的資料一致由Raft協議進行保證。當然如果本機就是主節點, 對於一些查詢請求不需要走Raft協議而直接呼叫本地實現方法。

0x04 網路互動

當Registry需要進行業務呼叫時候,就會隱形使用Raft。

比如 DataStoreService 會進行如下呼叫:

Map<String/*dataCenter*/, NodeRepository> dataNodeRepositoryMap = dataRepositoryService
        .getNodeRepositories();

getNodeRepositories會使用 Proxy 呼叫到 ProxyHandler # invoke。

public class ProxyHandler implements InvocationHandler {
    private final Class<?>      interfaceType;
    private final String        serviceId;
    private final RaftClient    client;

    @Override
    public Object invoke(Object proxy, Method method, Object[] args) {
        try {
            ProcessRequest request = new ProcessRequest();
            request.setMethodArgSigs(createParamSignature(method.getParameterTypes()));
            request.setMethodName(method.getName());
            request.setMethodArgs(args);
            request.setServiceName(serviceId);

            if (Processor.getInstance().isLeaderReadMethod(method)) {
                return doInvokeMethod(request); // 如果本身就是leader,則直接呼叫JVM函式 
            }
            return client.sendRequest(request); // 否則發起client呼叫
        }
    }
}

其呼叫棧如下:

invoke:69, ProxyHandler (com.alipay.sofa.registry.jraft.processor)
getNodeRepositories:-1, $Proxy46 (com.sun.proxy)
getNodeChangeResult:238, DataStoreService (com.alipay.sofa.registry.server.meta.store)
getAllNodes:96, MetaServerRegistry (com.alipay.sofa.registry.server.meta.registry)
getRegisterNodeByType:81, MetaDigestResource (com.alipay.sofa.registry.server.meta.resource)
lambda$init$1:70, MetaDigestResource (com.alipay.sofa.registry.server.meta.resource)

然後在 RaftClient # sendRequest 中有對 Raft 的進一步呼叫

public Object sendRequest(ProcessRequest request) {
    try {
        PeerId peer = getLeader();
        Object response = this.rpcClient.invokeSync(peer.getEndpoint().toString(), request,
            cliOptions.getRpcDefaultTimeout());
        ProcessResponse cmd = (ProcessResponse) response;
        if (cmd.getSuccess()) {
            return cmd.getEntity();
        } 
    } 
}

當在服務端,呼叫棧如下

process:123, Processor (com.alipay.sofa.registry.jraft.processor)
onApply:133, ServiceStateMachine (com.alipay.sofa.registry.jraft.bootstrap)
doApplyTasks:534, FSMCallerImpl (com.alipay.sofa.jraft.core)
doCommitted:503, FSMCallerImpl (com.alipay.sofa.jraft.core)
runApplyTask:431, FSMCallerImpl (com.alipay.sofa.jraft.core)
access$100:72, FSMCallerImpl (com.alipay.sofa.jraft.core)
onEvent:147, FSMCallerImpl$ApplyTaskHandler (com.alipay.sofa.jraft.core)
onEvent:141, FSMCallerImpl$ApplyTaskHandler (com.alipay.sofa.jraft.core)
run:137, BatchEventProcessor (com.lmax.disruptor)
run:748, Thread (java.lang)

最後圖例如下:

+---------------------------------------+        +---------------------------------------------+
| +------------------------------+      |        |       +----------------------------------+  |
| | +----------------+  registry |Client|        | Server| +----------------------+registry |  |
| | |DataStoreService|           |      |        |       | | DataRepositoryService|         |  |
| | +-----+----------+           |      |        |       | +---------+------------+         |  |
| |       | getNodeRepositories  |      |        |       |           ^  getNodeRepositories |  |
| |       |                      |      |        |       |           |                      |  |
| |       v                      |      |        |       |    +------+----+                 |  |
| | +-----+-----------------+    |      |        |       |    | Processor |                 |  |
| | |DataRepositoryService  |    |      |        |       |    +------+----+                 |  |
| | +-----+-----------------+    |      |        |       |           ^  onApply             |  |
| |       |                      |      |        |       |           |                      |  |
| |       v                      |      |        |       |   +-------+------+               |  |
| |     +-+---+                  |      |        |       |   | StateMachine |               |  |
| |     |Proxy|                  |      |        |       |   +-------+------+               |  |
| |     +-+---+                  |      |        |       |           ^  process             |  |
| |       | invoke               |      |        |       |           |                      |  |
| |       v                      |      |        |       |           |                      |  |
| |  +----+-------+              |      |        |       |    +------+------+               |  |
| |  |ProxyHandler|              |      |        |       |    |FSMCallerImpl|               |  |
| |  +----+-------+              |      |        |       |    +------+------+               |  |
| |       | sendRequest          |      |        |       |           ^                      |  |
| |       v                      |      |        |       |           |  received            |  |
| |   +---+------+               |      |        |       |           |                      |  |
| |   |RaftClient|               |      |        |       |   +-----------------+            |  |
| |   +---+------+               |      |        |       |   |RaftServerHandler|            |  |
| |       | invokeSync           |      |        |       |   +-----------------+            |  |
| +------------------------------+      |        |       +----------------------------------+  |
|         |                             |        |                   |                         |
|         |                             |        |                   |                         |
| +------------------------------+      |        |         +--------------------------+        |
| |       |         remoting.rpc |      |        |         |         |    remoting.rpc|        |
| |  +----v------+               | bolt | Network|         | +-------+-----------+    |        |
| |  | RpcClient |               | +---------------------> | |RpcRequestProcessor|    |        |
| |  +-----------+               |      |        |         | +-------------------+    |        |
| +------------------------------+      |        |         +--------------------------+        |
+---------------------------------------+        +---------------------------------------------+

在手機上如圖

0x05 快照儲存

首先我們需要看看為什麼要有快照機制。

5.1 儲存模組

SOFAJRaft 儲存模組分為:

  • Log 儲存記錄 Raft 配置變更和使用者提交任務日誌;
  • Meta 儲存即元資訊儲存記錄 Raft 實現的內部狀態;
  • Snapshot 儲存用於存放使用者的狀態機 Snapshot 及元資訊,Snapshot 是快照,是對資料當前值的一個記錄;

5.2 問題

當 Raft 節點 Node 重啟時,記憶體中狀態機的狀態資料丟失,觸發啟動過程重新存放日誌儲存 LogStorage 的所有日誌重建整個狀態機例項,此種場景會導致三個問題:

  • 如果任務提交比較頻繁,例如訊息中介軟體場景導致整個重建過程很長啟動緩慢;
  • 如果日誌非常多並且節點需要儲存所有的日誌,對儲存來說是資源佔用不可持續;
  • 如果增加 Node 節點,新節點需要從 Leader 獲取所有的日誌重新存放至狀態機,對於 Leader 和網路頻寬都是不小的負擔。

5.3 Snapshot 機制

因此通過引入 Snapshot 機制來解決此三個問題。

所謂快照 Snapshot 即對資料當前值的記錄,是為當前狀態機的最新狀態構建”映象”單獨儲存,儲存成功刪除此時刻之前的日誌減少日誌儲存佔用;啟動的時候直接載入最新的 Snapshot 映象,然後重放在此之後的日誌即可,如果 Snapshot 間隔合理,整個重放到狀態機過程較快,加速啟動過程。最後新節點的加入先從 Leader 拷貝最新的 Snapshot 安裝到本地狀態機,然後只要拷貝後續的日誌即可,能夠快速跟上整個 Raft Group 的進度。

Leader 生成快照有幾個作用:

  • 當有新的節點 Node 加入叢集不用只靠日誌複製、回放機制和 Leader 保持資料一致,通過安裝 Leader 的快照方式跳過早期大量日誌的回放;
  • Leader 用快照替代 Log 複製減少網路端的資料量;
  • 用快照替代早期的 Log 節省儲存佔用空間。

5.4 ServiceStateMachine

在狀態機層面上來說,使用 snapshot 機制,也就是為狀態機做一個 checkpoint,儲存當時狀態機的狀態,刪除在此之前的所有日誌,核心是實現 StateMachine的兩個方法:

  • onSnapshotLoad,啟動或者安裝 snapshot 後載入 snapshot;
  • onSnapshotSave ,定期儲存 snapshot;

從具體使用 Raft 的相關服務層面來說,每個服務提供了自己不同的業務實現。

  • 在Registry處理註解時候,會通過 addWorker 來把使用 Raft 的相關服務類和成員變數設定到map中;
  • 於是在狀態機呼叫快照相關函式時候,狀態機會遍歷 Processor.getInstance().getWorkers() ,從而呼叫每個類的具體處理函式;

具體work變數如下:

workers = {HashMap@6176}  size = 7
 "com.alipay.sofa.registry.server.meta.repository.RepositoryService:metaServer" -> {MetaRepositoryService@6201} 
 "com.alipay.sofa.registry.server.meta.repository.RepositoryService:dataServer" -> {DataRepositoryService@6162} 
 "com.alipay.sofa.registry.store.api.DBService" -> {PersistenceDataDBService@6203} 
 "com.alipay.sofa.registry.server.meta.repository.RepositoryService:sessionServer" -> {SessionRepositoryService@6205} 
 "com.alipay.sofa.registry.server.meta.repository.VersionRepositoryService:sessionServer" -> {SessionVersionRepositoryService@6207} 
 "com.alipay.sofa.registry.server.meta.repository.NodeConfirmStatusService:sessionServer" -> {SessionConfirmStatusService@6209} 
 "com.alipay.sofa.registry.server.meta.repository.NodeConfirmStatusService:dataServer" -> {DataConfirmStatusService@6211} 

狀態機具體程式碼如下:

public class ServiceStateMachine extends StateMachineAdapter {
    @Override
    public void onSnapshotSave(final SnapshotWriter writer, final Closure done) {

        Map<String, Object> workers = Processor.getInstance().getWorkers();
        Map<String, SnapshotProcess> snapshotProcessors = new HashMap<>();
        if (workers != null) {
            // 遍歷
            workers.forEach((serviceId, worker) -> {
                if (worker instanceof SnapshotProcess) {
                    SnapshotProcess snapshotProcessor = (SnapshotProcess) worker;
                    snapshotProcessors.put(serviceId, snapshotProcessor.copy());
                }
            });
        }
        Utils.runInThread(() -> {
            String errors = null;
            outer:
            // 遍歷
            for (Map.Entry<String, SnapshotProcess> entry : snapshotProcessors.entrySet()) {
                String serviceId = entry.getKey();
                SnapshotProcess snapshotProcessor = entry.getValue();
                Set<String> fileNames = snapshotProcessor.getSnapshotFileNames();
                for (String fileName : fileNames) {
                    String savePath = writer.getPath() + File.separator + fileName;
                    boolean ret = snapshotProcessor.save(savePath);   // 呼叫具體實現
                    if (ret) {
                        if (!writer.addFile(fileName)) {
                            break outer;
                        }
                    } else {
                        break outer;
                    }
                }
            }
            if (errors != null) {
                done.run(new Status(RaftError.EIO, errors));
            } else {
                done.run(Status.OK());
            }
        });
    }

    @Override
    public boolean onSnapshotLoad(SnapshotReader reader) {
        List<String> failServices = new ArrayList<>();
        Map<String, Object> workers = Processor.getInstance().getWorkers();
        if (workers != null) {
            // 遍歷
            outer: for (Map.Entry<String, Object> entry : workers.entrySet()) {
                String serviceId = entry.getKey();
                Object worker = entry.getValue();
                if (worker instanceof SnapshotProcess) {
                    SnapshotProcess snapshotProcess = (SnapshotProcess) worker;
                    Set<String> fileNames = snapshotProcess.getSnapshotFileNames();

                    for (String fileName : fileNames) {
                        if (reader.getFileMeta(fileName) == null) {
                            failServices.add(serviceId);
                            break outer;
                        }

                        String savePath = reader.getPath() + File.separator + fileName;
                        boolean ret = snapshotProcess.load(savePath);  // 呼叫具體實現
                        if (!ret) {
                            failServices.add(serviceId);
                            break outer;
                        }
                    }
                }
            }
        }
        return true;
    }
}

5.5 XXXRepositoryService

關於具體服務,我們可以參見XXXRepositoryService。

在ServiceStateMachine中,會用 snapshotProcess.load(savePath); 呼叫具體服務的特殊實現,這從呼叫棧中可以清晰見到。

load:317, DataRepositoryService (com.alipay.sofa.registry.server.meta.repository.service)
onSnapshotLoad:212, ServiceStateMachine (com.alipay.sofa.registry.jraft.bootstrap)
doSnapshotLoad:641, FSMCallerImpl (com.alipay.sofa.jraft.core)
runApplyTask:389, FSMCallerImpl (com.alipay.sofa.jraft.core)
access$100:72, FSMCallerImpl (com.alipay.sofa.jraft.core)
onEvent:147, FSMCallerImpl$ApplyTaskHandler (com.alipay.sofa.jraft.core)
onEvent:141, FSMCallerImpl$ApplyTaskHandler (com.alipay.sofa.jraft.core)
run:137, BatchEventProcessor (com.lmax.disruptor)
run:748, Thread (java.lang)

以DataRepositoryService為例,其基類AbstractSnapshotProcess做了一些基礎實現。

public abstract class AbstractSnapshotProcess implements SnapshotProcess {
    public boolean save(String path, Object values) {
            FileUtils.writeByteArrayToFile(new File(path), CommandCodec.encodeCommand(values),
                false);
            return true;
    }
    public <T> T load(String path, Class<T> clazz) throws IOException {
        byte[] bs = FileUtils.readFileToByteArray(new File(path));
        if (bs != null && bs.length > 0) {
            return CommandCodec.decodeCommand(bs, clazz);
        }
    }
}

在DataRepositoryService之中,又對load做了一些適配。

@RaftService(uniqueId = "dataServer")
public class DataRepositoryService extends AbstractSnapshotProcess
                                                                  implements
                                                                  RepositoryService<String, RenewDecorate<DataNode>> {
    @Override
    public boolean save(String path) {
        return save(path, registry);
    }

    @Override
    public synchronized boolean load(String path) {
            Map<String, NodeRepository> map = load(path, registry.getClass());
            registry.clear();
            registry.putAll(map);
            return true;
    }  
}

0xFF 參考

服務註冊中心 MetaServer 功能介紹和實現剖析 | SOFARegistry 解析

服務註冊中心如何實現 DataServer 平滑擴縮容 | SOFARegistry 解析

服務註冊中心資料一致性方案分析 | SOFARegistry 解析

服務註冊中心如何實現秒級服務上下線通知 | SOFARegistry 解析

服務註冊中心 Session 儲存策略 | SOFARegistry 解析

服務註冊中心資料分片和同步方案詳解 | SOFARegistry 解析

服務註冊中心 SOFARegistry 解析 | 服務發現優化之路

海量資料下的註冊中心 - SOFARegistry 架構介紹

服務端部署

客戶端使用

全面理解Raft協議

詳解螞蟻金服 SOFAJRaft | 生產級高效能 Java 實現

從JRaft來看Raft協議實現細節

SOFAJRaft—初次使用

JRaft 使用者指南 & API 詳解

怎樣打造一個分散式資料庫——rocksDB, raft, mvcc,本質上是為了解決跨資料中心的複製

sofa-bolt原始碼閱讀(5)-日誌

Raft 為什麼是更易理解的分散式一致性演算法

SOFAJRaft 原始碼分析一(啟動流程和節點變化)

SOFAJRaft 實現原理 - 生產級 Raft 演算法庫儲存模組剖析

客戶端使用

相關文章