Dubbo（五）：集群容错的实现

前两篇中，我们看到了dubbo在负载均衡和服务路由方面的实现，它为集群功能提供了必要的功能。

今天我们再来看另一个集群组件的实现：集群容错。

1. dubbo 集群容错简介

为了避免单点故障，现在的应用通常至少会部署在两台服务器上。对于一些负载比较高的服务，会部署更多的服务器。对于服务消费者来说，同一环境下出现了多个服务提供者。这时会出现一个问题，服务消费者需要决定选择哪个服务提供者进行调用。另外服务调用失败时的处理措施也是需要考虑的，是重试呢，还是抛出异常，亦或是只打印异常等。为了处理这些问题，Dubbo 定义了集群接口 Cluster 以及 Cluster Invoker。集群 Cluster 用途是将多个服务提供者合并为一个 Cluster Invoker，并将这个 Invoker 暴露给服务消费者。这样一来，服务消费者只需通过这个 Invoker 进行远程调用即可，至于具体调用哪个服务提供者，以及调用失败后如何处理等问题，现在都交给集群模块去处理。集群模块是服务提供者和服务消费者的中间层，为服务消费者屏蔽了服务提供者的情况，这样服务消费者就可以专心处理远程调用相关事宜。

dubbo的集群容错功能由多个组件共同完成：包括 Cluster、Cluster Invoker、Directory、Router 和 LoadBalance 等。它们之间的依赖关系如下：

负载均衡、路由服务是在一次调用中进行的，而容错则是当调用发生异常之后，进行处理策略。

dubbo中主要提供了以下几种容错策略实现：

Failover Cluster - 失败自动切换

Failfast Cluster - 快速失败

Failsafe Cluster - 失败安全

Failback Cluster - 失败自动恢复

Forking Cluster - 并行调用多个服务提供者

2. 集群容错的框架实现

集群接口 Cluster 和 Cluster Invoker，这两者是不同的。Cluster 是接口，而 Cluster Invoker 是一种 Invoker。服务提供者的选择逻辑，以及远程调用失败后的的处理逻辑均是封装在 Cluster Invoker 中。

Cluster 的实现类图如下：

各个Cluster的实现都很简单，也都统一继承了 AbstractCluster, 而该 AbstractCluster 则做了一层统一的拦截器的功能接入，实现如下：

    
public abstract class AbstractCluster implements Cluster {

    private <T> Invoker<T> buildClusterInterceptors(AbstractClusterInvoker<T> clusterInvoker, String key) {
        AbstractClusterInvoker<T> last = clusterInvoker;
        List<ClusterInterceptor> interceptors = ExtensionLoader.getExtensionLoader(ClusterInterceptor.class).getActivateExtension(clusterInvoker.getUrl(), key);
        // 根据需要包装ClusterInvoker, 使用切面的方式进行拦截器接入 
        // 按先后依次强入拦截器
        if (!interceptors.isEmpty()) {
            for (int i = interceptors.size() - 1; i >= 0; i--) {
                final ClusterInterceptor interceptor = interceptors.get(i);
                final AbstractClusterInvoker<T> next = last;
                // 使用内部类进行包装拦截器
                // 先后顺序如: beforeC -> beforeB -> beforeA (spring中还有Around) -> afterA -> afterB -> afterC (spring中还有afterReturn)
                last = new InterceptorInvokerNode<>(clusterInvoker, interceptor, next);
            }
        }
        return last;
    }

    @Override
    public <T> Invoker<T> join(Directory<T> directory) throws RpcException {
        // ClusterInvoker 调用入口, 让具体策略实现 doJoin(), 并在其基础上进行包装拦截器, 依据来源 reference.interceptor=xxx
        return buildClusterInterceptors(doJoin(directory), directory.getUrl().getParameter(REFERENCE_INTERCEPTOR_KEY));
    }
    // 
    protected abstract <T> AbstractClusterInvoker<T> doJoin(Directory<T> directory) throws RpcException;

    protected class InterceptorInvokerNode<T> extends AbstractClusterInvoker<T> {

        private AbstractClusterInvoker<T> clusterInvoker;
        private ClusterInterceptor interceptor;
        private AbstractClusterInvoker<T> next;

        public InterceptorInvokerNode(AbstractClusterInvoker<T> clusterInvoker,
                                      ClusterInterceptor interceptor,
                                      AbstractClusterInvoker<T> next) {
            this.clusterInvoker = clusterInvoker;
            this.interceptor = interceptor;
            this.next = next;
        }

        @Override
        public Class<T> getInterface() {
            return clusterInvoker.getInterface();
        }

        @Override
        public URL getUrl() {
            return clusterInvoker.getUrl();
        }

        @Override
        public boolean isAvailable() {
            return clusterInvoker.isAvailable();
        }

        @Override
        public Result invoke(Invocation invocation) throws RpcException {
            Result asyncResult;
            try {
                // 拦截器的具体处理逻辑
                // 有个 intercept() 的默认方法，其为调用 clusterInvoker.invoke(invocation);  从而实现链式调用
                interceptor.before(next, invocation);
                asyncResult = interceptor.intercept(next, invocation);
            } catch (Exception e) {
                // onError callback
                if (interceptor instanceof ClusterInterceptor.Listener) {
                    ClusterInterceptor.Listener listener = (ClusterInterceptor.Listener) interceptor;
                    listener.onError(e, clusterInvoker, invocation);
                }
                throw e;
            } finally {
                interceptor.after(next, invocation);
            }
            return asyncResult.whenCompleteWithContext((r, t) -> {
                // onResponse callback
                if (interceptor instanceof ClusterInterceptor.Listener) {
                    ClusterInterceptor.Listener listener = (ClusterInterceptor.Listener) interceptor;
                    if (t == null) {
                        listener.onMessage(r, clusterInvoker, invocation);
                    } else {
                        listener.onError(t, clusterInvoker, invocation);
                    }
                }
            });
        }

        @Override
        public void destroy() {
            clusterInvoker.destroy();
        }

        @Override
        public String toString() {
            return clusterInvoker.toString();
        }

        @Override
        protected Result doInvoke(Invocation invocation, List<Invoker<T>> invokers, LoadBalance loadbalance) throws RpcException {
            // The only purpose is to build a interceptor chain, so the cluster related logic doesn't matter.
            return null;
        }
    }
}

接下来，我们详细看看，每个集群容错策略都是如何创建的。

// failover 失败自动切换
public class FailoverCluster extends AbstractCluster {

    public final static String NAME = "failover";

    @Override
    public <T> AbstractClusterInvoker<T> doJoin(Directory<T> directory) throws RpcException {
        return new FailoverClusterInvoker<>(directory);
    }

}
// failfast 快速失败
public class FailfastCluster extends AbstractCluster {

    public final static String NAME = "failfast";

    @Override
    public <T> AbstractClusterInvoker<T> doJoin(Directory<T> directory) throws RpcException {
        return new FailfastClusterInvoker<>(directory);
    }

}
// failsafe 失败安全
public class FailsafeCluster extends AbstractCluster {

    public final static String NAME = "failsafe";

    @Override
    public <T> AbstractClusterInvoker<T> doJoin(Directory<T> directory) throws RpcException {
        return new FailsafeClusterInvoker<>(directory);
    }

}
// failback 失败自动恢复
public class FailbackCluster extends AbstractCluster {

    public final static String NAME = "failback";

    @Override
    public <T> AbstractClusterInvoker<T> doJoin(Directory<T> directory) throws RpcException {
        return new FailbackClusterInvoker<>(directory);
    }

}
// forking 并行调用多个服务提供者
public class ForkingCluster extends AbstractCluster {

    public final static String NAME = "forking";

    @Override
    public <T> AbstractClusterInvoker<T> doJoin(Directory<T> directory) throws RpcException {
        return new ForkingClusterInvoker<>(directory);
    }

}
// mergeable 合并结果容错
public class MergeableCluster extends AbstractCluster {

    public static final String NAME = "mergeable";

    @Override
    public <T> AbstractClusterInvoker<T> doJoin(Directory<T> directory) throws RpcException {
        return new MergeableClusterInvoker<T>(directory);
    }

}

3. 具体集群容错的实现

failover, 失败自动切换。这是dubbo的默认集群容错策略，因为它是一个比较通用的策略，即只需做重试即可，保证高可用。

整个集群容错策略的调用入口在 AbstractClusterInvoker.invoke() 中，经过一些通用过程调用后，再由具体策略实现 doInvoke();

    // org.apache.dubbo.rpc.cluster.support.AbstractClusterInvoker#invoke
    @Override
    public Result invoke(final Invocation invocation) throws RpcException {
        // 有效性检查
        checkWhetherDestroyed();

        // binding attachments into invocation.
        Map<String, Object> contextAttachments = RpcContext.getContext().getObjectAttachments();
        if (contextAttachments != null && contextAttachments.size() != 0) {
            ((RpcInvocation) invocation).addObjectAttachments(contextAttachments);
        }
        // 路由服务提供所有的 invokers
        List<Invoker<T>> invokers = list(invocation);
        // 获取负载均衡器
        LoadBalance loadbalance = initLoadBalance(invokers, invocation);
        RpcUtils.attachInvocationIdIfAsync(getUrl(), invocation);
        // 各子类实现 具体的容错逻辑
        return doInvoke(invocation, invokers, loadbalance);
    }

3.1. failover 失败自动切换实现

    // org.apache.dubbo.rpc.cluster.support.FailoverClusterInvoker#doInvoke
    @Override
    @SuppressWarnings({"unchecked", "rawtypes"})
    public Result doInvoke(Invocation invocation, final List<Invoker<T>> invokers, LoadBalance loadbalance) throws RpcException {
        List<Invoker<T>> copyInvokers = invokers;
        checkInvokers(copyInvokers, invocation);
        String methodName = RpcUtils.getMethodName(invocation);
        int len = getUrl().getMethodParameter(methodName, RETRIES_KEY, DEFAULT_RETRIES) + 1;
        if (len <= 0) {
            len = 1;
        }
        // retry loop.
        RpcException le = null; // last exception.
        List<Invoker<T>> invoked = new ArrayList<Invoker<T>>(copyInvokers.size()); // invoked invokers.
        Set<String> providers = new HashSet<String>(len);
        // 失败自动切换，就是一个重试的过程
        for (int i = 0; i < len; i++) {
            //Reselect before retry to avoid a change of candidate `invokers`.
            //NOTE: if `invokers` changed, then `invoked` also lose accuracy.
            if (i > 0) {
                // 进行重试时，需要刷新invokers
                checkWhetherDestroyed();
                copyInvokers = list(invocation);
                // check again
                checkInvokers(copyInvokers, invocation);
            }
            // 使用负载均衡选取一个 invoker
            Invoker<T> invoker = select(loadbalance, invocation, copyInvokers, invoked);
            // 将选中的invoker添加到 invoked 中，避免反复选择一个失效的invoker
            invoked.add(invoker);
            RpcContext.getContext().setInvokers((List) invoked);
            try {
                // 调用选中的invoker 远程服务，成功直接返回了，失败则容错能力上
                Result result = invoker.invoke(invocation);
                if (le != null && logger.isWarnEnabled()) {
                    logger.warn("Although retry the method " + methodName
                            + " in the service " + getInterface().getName()
                            + " was successful by the provider " + invoker.getUrl().getAddress()
                            + ", but there have been failed providers " + providers
                            + " (" + providers.size() + "/" + copyInvokers.size()
                            + ") from the registry " + directory.getUrl().getAddress()
                            + " on the consumer " + NetUtils.getLocalHost()
                            + " using the dubbo version " + Version.getVersion() + ". Last error is: "
                            + le.getMessage(), le);
                }
                // 调用成功直接返回
                return result;
            } catch (RpcException e) {
                // 业务异常则直接抛出，不再重试
                if (e.isBiz()) { // biz exception.
                    throw e;
                }
                le = e;
            } catch (Throwable e) {
                le = new RpcException(e.getMessage(), e);
            } finally {
                providers.add(invoker.getUrl().getAddress());
            }
        }
        throw new RpcException(le.getCode(), "Failed to invoke the method "
                + methodName + " in the service " + getInterface().getName()
                + ". Tried " + len + " times of the providers " + providers
                + " (" + providers.size() + "/" + copyInvokers.size()
                + ") from the registry " + directory.getUrl().getAddress()
                + " on the consumer " + NetUtils.getLocalHost() + " using the dubbo version "
                + Version.getVersion() + ". Last error is: "
                + le.getMessage(), le.getCause() != null ? le.getCause() : le);
    }

总结：failover 容错，即是自动重试各可用提供者的过程。

3.2. failback 失败自动恢复的实现

    public FailbackClusterInvoker(Directory<T> directory) {
        super(directory);
        // retries=3
        int retriesConfig = getUrl().getParameter(RETRIES_KEY, DEFAULT_FAILBACK_TIMES);
        if (retriesConfig <= 0) {
            retriesConfig = DEFAULT_FAILBACK_TIMES;
        }
        // failbacktasks=100
        int failbackTasksConfig = getUrl().getParameter(FAIL_BACK_TASKS_KEY, DEFAULT_FAILBACK_TASKS);
        if (failbackTasksConfig <= 0) {
            failbackTasksConfig = DEFAULT_FAILBACK_TASKS;
        }
        retries = retriesConfig;
        failbackTasks = failbackTasksConfig;
    }
    // 当调用失败后，将其添加到定时队列中，稍后进行重新请求
    private void addFailed(LoadBalance loadbalance, Invocation invocation, List<Invoker<T>> invokers, Invoker<T> lastInvoker) {
        if (failTimer == null) {
            synchronized (this) {
                if (failTimer == null) {
                    // 以1秒为间隔使用 hash环，扫描任务
                    failTimer = new HashedWheelTimer(
                            new NamedThreadFactory("failback-cluster-timer", true),
                            1,
                            TimeUnit.SECONDS, 32, failbackTasks);
                }
            }
        }
        // 使用 RetryTimerTask 来构建调度的任务
        RetryTimerTask retryTimerTask = new RetryTimerTask(loadbalance, invocation, invokers, lastInvoker, retries, RETRY_FAILED_PERIOD);
        try {
            failTimer.newTimeout(retryTimerTask, RETRY_FAILED_PERIOD, TimeUnit.SECONDS);
        } catch (Throwable e) {
            logger.error("Failback background works error,invocation->" + invocation + ", exception: " + e.getMessage());
        }
    }

    @Override
    protected Result doInvoke(Invocation invocation, List<Invoker<T>> invokers, LoadBalance loadbalance) throws RpcException {
        Invoker<T> invoker = null;
        try {
            checkInvokers(invokers, invocation);
            invoker = select(loadbalance, invocation, invokers, null);
            // 只调用一次，失败即失败
            return invoker.invoke(invocation);
        } catch (Throwable e) {
            logger.error("Failback to invoke method " + invocation.getMethodName() + ", wait for retry in background. Ignored exception: "
                    + e.getMessage() + ", ", e);
            // 添加到失败队列中，稍后进行调度
            addFailed(loadbalance, invocation, invokers, invoker);
            return AsyncRpcResult.newDefaultAsyncResult(null, null, invocation); // ignore
        }
    }

总结：failback 容错，即是只做一次调用，失败后会开启后续定时任务进行重新调用的过程。

3.3. failfast 快速失败的实现

    // org.apache.dubbo.rpc.cluster.support.FailfastClusterInvoker#doInvoke
    @Override
    public Result doInvoke(Invocation invocation, List<Invoker<T>> invokers, LoadBalance loadbalance) throws RpcException {
        checkInvokers(invokers, invocation);
        // 使用负载均衡选取一个 可用的 invoker, 然后进行调用即可
        // selected = null, 即只一次选择即可完成select
        Invoker<T> invoker = select(loadbalance, invocation, invokers, null);
        try {
            return invoker.invoke(invocation);
        } catch (Throwable e) {
            if (e instanceof RpcException && ((RpcException) e).isBiz()) { // biz exception.
                throw (RpcException) e;
            }
            throw new RpcException(e instanceof RpcException ? ((RpcException) e).getCode() : 0,
                    "Failfast invoke providers " + invoker.getUrl() + " " + loadbalance.getClass().getSimpleName()
                            + " select from all providers " + invokers + " for service " + getInterface().getName()
                            + " method " + invocation.getMethodName() + " on consumer " + NetUtils.getLocalHost()
                            + " use dubbo version " + Version.getVersion()
                            + ", but no luck to perform the invocation. Last error is: " + e.getMessage(),
                    e.getCause() != null ? e.getCause() : e);
        }
    }
    
    // org.apache.dubbo.rpc.cluster.support.AbstractClusterInvoker#select
    /**
     * Select a invoker using loadbalance policy.</br>
     * a) Firstly, select an invoker using loadbalance. If this invoker is in previously selected list, or,
     * if this invoker is unavailable, then continue step b (reselect), otherwise return the first selected invoker</br>
     * <p>
     * b) Reselection, the validation rule for reselection: selected > available. This rule guarantees that
     * the selected invoker has the minimum chance to be one in the previously selected list, and also
     * guarantees this invoker is available.
     *
     * @param loadbalance load balance policy
     * @param invocation  invocation
     * @param invokers    invoker candidates
     * @param selected    exclude selected invokers or not
     * @return the invoker which will final to do invoke.
     * @throws RpcException exception
     */
    protected Invoker<T> select(LoadBalance loadbalance, Invocation invocation,
                                List<Invoker<T>> invokers, List<Invoker<T>> selected) throws RpcException {

        if (CollectionUtils.isEmpty(invokers)) {
            return null;
        }
        String methodName = invocation == null ? StringUtils.EMPTY_STRING : invocation.getMethodName();

        boolean sticky = invokers.get(0).getUrl()
                .getMethodParameter(methodName, CLUSTER_STICKY_KEY, DEFAULT_CLUSTER_STICKY);

        //ignore overloaded method
        if (stickyInvoker != null && !invokers.contains(stickyInvoker)) {
            stickyInvoker = null;
        }
        //ignore concurrency problem
        if (sticky && stickyInvoker != null && (selected == null || !selected.contains(stickyInvoker))) {
            if (availablecheck && stickyInvoker.isAvailable()) {
                return stickyInvoker;
            }
        }

        Invoker<T> invoker = doSelect(loadbalance, invocation, invokers, selected);

        if (sticky) {
            stickyInvoker = invoker;
        }
        return invoker;
    }

    private Invoker<T> doSelect(LoadBalance loadbalance, Invocation invocation,
                                List<Invoker<T>> invokers, List<Invoker<T>> selected) throws RpcException {

        if (CollectionUtils.isEmpty(invokers)) {
            return null;
        }
        if (invokers.size() == 1) {
            return invokers.get(0);
        }
        Invoker<T> invoker = loadbalance.select(invokers, getUrl(), invocation);

        //If the `invoker` is in the  `selected` or invoker is unavailable && availablecheck is true, reselect.
        if ((selected != null && selected.contains(invoker))
                || (!invoker.isAvailable() && getUrl() != null && availablecheck)) {
            try {
                Invoker<T> rInvoker = reselect(loadbalance, invocation, invokers, selected, availablecheck);
                if (rInvoker != null) {
                    invoker = rInvoker;
                } else {
                    //Check the index of current selected invoker, if it's not the last one, choose the one at index+1.
                    int index = invokers.indexOf(invoker);
                    try {
                        //Avoid collision
                        invoker = invokers.get((index + 1) % invokers.size());
                    } catch (Exception e) {
                        logger.warn(e.getMessage() + " may because invokers list dynamic change, ignore.", e);
                    }
                }
            } catch (Throwable t) {
                logger.error("cluster reselect fail reason is :" + t.getMessage() + " if can not solve, you can set cluster.availablecheck=false in url", t);
            }
        }
        return invoker;
    }

总结: failfast 容错，使用负载均衡策略选择一次可用的invoker, 进行调用, 异常则抛出，正常则返回结果。

3.4. failsafe 安全失败容错的实现

    @Override
    public Result doInvoke(Invocation invocation, List<Invoker<T>> invokers, LoadBalance loadbalance) throws RpcException {
        try {
            checkInvokers(invokers, invocation);
            // 与failfast 一样，只使用一次负载均衡策略，选择一个invoker调用即可
            // 差别在于返回值，failsafe 不抛出异常，当发生异常时返回一个默认值
            Invoker<T> invoker = select(loadbalance, invocation, invokers, null);
            return invoker.invoke(invocation);
        } catch (Throwable e) {
            logger.error("Failsafe ignore exception: " + e.getMessage(), e);
            // 将异常信息忽略，返回默认值
            return AsyncRpcResult.newDefaultAsyncResult(null, null, invocation); // ignore
        }
    }

总结: failsafe 容错，即忽略掉所有异常，只返回正式结果。当发生异常时，返回 AsyncRpcResult.newDefaultAsyncResult 作为结果，好像没有发生异常一样。

3.5. forking 并发请求容错实现

    // org.apache.dubbo.rpc.cluster.support.ForkingClusterInvoker#doInvoke
    @Override
    @SuppressWarnings({"unchecked", "rawtypes"})
    public Result doInvoke(final Invocation invocation, List<Invoker<T>> invokers, LoadBalance loadbalance) throws RpcException {
        try {
            checkInvokers(invokers, invocation);
            final List<Invoker<T>> selected;
            // forks=2
            final int forks = getUrl().getParameter(FORKS_KEY, DEFAULT_FORKS);
            // timeout=1000
            final int timeout = getUrl().getParameter(TIMEOUT_KEY, DEFAULT_TIMEOUT);
            if (forks <= 0 || forks >= invokers.size()) {
                selected = invokers;
            } else {
                selected = new ArrayList<>(forks);
                while (selected.size() < forks) {
                    Invoker<T> invoker = select(loadbalance, invocation, invokers, selected);
                    if (!selected.contains(invoker)) {
                        //Avoid add the same invoker several times.
                        selected.add(invoker);
                    }
                }
            }
            RpcContext.getContext().setInvokers((List) selected);
            final AtomicInteger count = new AtomicInteger();
            final BlockingQueue<Object> ref = new LinkedBlockingQueue<>();
            for (final Invoker<T> invoker : selected) {
                // 使用线程池进行并发调用 invoker
                // 线程池为无界队列式: executor = Executors.newCachedThreadPool(new NamedInternalThreadFactory("forking-cluster-timer", true));
                executor.execute(() -> {
                    try {
                        Result result = invoker.invoke(invocation);
                        // 只要结果响应，则入队到 ref  中
                        ref.offer(result);
                    } catch (Throwable e) {
                        int value = count.incrementAndGet();
                        if (value >= selected.size()) {
                            // 当超过forks 数量的异常发生后，将异常信息写入ref中，即外部可以获取结果了
                            ref.offer(e);
                        }
                    }
                });
            }
            try {
                // 阻塞获取结果，最长等待 timeout
                // 获取第一个结果作为响应依据
                Object ret = ref.poll(timeout, TimeUnit.MILLISECONDS);
                // 因可以全部异常，获取到的结果可能是个 Throwable 信息，须先判定
                if (ret instanceof Throwable) {
                    Throwable e = (Throwable) ret;
                    throw new RpcException(e instanceof RpcException ? ((RpcException) e).getCode() : 0, "Failed to forking invoke provider " + selected + ", but no luck to perform the invocation. Last error is: " + e.getMessage(), e.getCause() != null ? e.getCause() : e);
                }
                return (Result) ret;
            } catch (InterruptedException e) {
                throw new RpcException("Failed to forking invoke provider " + selected + ", but no luck to perform the invocation. Last error is: " + e.getMessage(), e);
            }
        } finally {
            // clear attachments which is binding to current thread.
            RpcContext.getContext().clearAttachments();
        }
    }

总结: forking 容错，即是同时发起n个并发请求调用提供者，谁最先响应则返回谁的结果。其他结果则全部忽略。可以说是非常耗资源的一种方式了，不过总是有相应的应用场景，所以存在。

3.6. broadcast 广播容错的实现

    // org.apache.dubbo.rpc.cluster.support.BroadcastClusterInvoker#doInvoke
    @Override
    @SuppressWarnings({"unchecked", "rawtypes"})
    public Result doInvoke(final Invocation invocation, List<Invoker<T>> invokers, LoadBalance loadbalance) throws RpcException {
        checkInvokers(invokers, invocation);
        RpcContext.getContext().setInvokers((List) invokers);
        RpcException exception = null;
        Result result = null;
        // 向所有invoker发起调用，只要有一个异常，则抛出异常
        for (Invoker<T> invoker : invokers) {
            try {
                result = invoker.invoke(invocation);
            } catch (RpcException e) {
                exception = e;
                logger.warn(e.getMessage(), e);
            } catch (Throwable e) {
                exception = new RpcException(e.getMessage(), e);
                logger.warn(e.getMessage(), e);
            }
        }
        if (exception != null) {
            throw exception;
        }
        return result;
    }

总结: broadcast 容错，即向所有invoker发起调用（即广播），全部成功才算成功。

3.7. mergeable 归并容错的实现

    // org.apache.dubbo.rpc.cluster.support.MergeableClusterInvoker#doInvoke
    @Override
    protected Result doInvoke(Invocation invocation, List<Invoker<T>> invokers, LoadBalance loadbalance) throws RpcException {
        checkInvokers(invokers, invocation);
        // merger=xxx
        String merger = getUrl().getMethodParameter(invocation.getMethodName(), MERGER_KEY);
        // 没有指定merger, 直接调用一个可用 invoker 即可 
        if (ConfigUtils.isEmpty(merger)) { // If a method doesn't have a merger, only invoke one Group
            for (final Invoker<T> invoker : invokers) {
                if (invoker.isAvailable()) {
                    try {
                        return invoker.invoke(invocation);
                    } catch (RpcException e) {
                        if (e.isNoInvokerAvailableAfterFilter()) {
                            log.debug("No available provider for service" + getUrl().getServiceKey() + " on group " + invoker.getUrl().getParameter(GROUP_KEY) + ", will continue to try another group.");
                        } else {
                            throw e;
                        }
                    }
                }
            }
            // 最后尝试使用第一个 invoker.invoke()
            return invokers.iterator().next().invoke(invocation);
        }

        Class<?> returnType;
        try {
            returnType = getInterface().getMethod(
                    invocation.getMethodName(), invocation.getParameterTypes()).getReturnType();
        } catch (NoSuchMethodException e) {
            returnType = null;
        }

        Map<String, Result> results = new HashMap<>();
        for (final Invoker<T> invoker : invokers) {
            RpcInvocation subInvocation = new RpcInvocation(invocation, invoker);
            subInvocation.setAttachment(ASYNC_KEY, "true");
            // 异步调用所有 invoker
            results.put(invoker.getUrl().getServiceKey(), invoker.invoke(subInvocation));
        }

        Object result = null;

        List<Result> resultList = new ArrayList<Result>(results.size());

        for (Map.Entry<String, Result> entry : results.entrySet()) {
            Result asyncResult = entry.getValue();
            try {
                // 等待所有 invoker 的结果响应
                Result r = asyncResult.get();
                if (r.hasException()) {
                    log.error("Invoke " + getGroupDescFromServiceKey(entry.getKey()) +
                                    " failed: " + r.getException().getMessage(),
                            r.getException());
                } else {
                    // 将所有结果放到 resultList 中
                    resultList.add(r);
                }
            } catch (Exception e) {
                throw new RpcException("Failed to invoke service " + entry.getKey() + ": " + e.getMessage(), e);
            }
        }

        if (resultList.isEmpty()) {
            return AsyncRpcResult.newDefaultAsyncResult(invocation);
        } else if (resultList.size() == 1) {
            // 只有一个结果，则返回一个 Result
            return resultList.iterator().next();
        }

        if (returnType == void.class) {
            return AsyncRpcResult.newDefaultAsyncResult(invocation);
        }

        if (merger.startsWith(".")) {
            merger = merger.substring(1);
            Method method;
            try {
                method = returnType.getMethod(merger, returnType);
            } catch (NoSuchMethodException e) {
                throw new RpcException("Can not merge result because missing method [ " + merger + " ] in class [ " +
                        returnType.getName() + " ]");
            }
            if (!Modifier.isPublic(method.getModifiers())) {
                method.setAccessible(true);
            }
            result = resultList.remove(0).getValue();
            try {
                if (method.getReturnType() != void.class
                        && method.getReturnType().isAssignableFrom(result.getClass())) {
                    for (Result r : resultList) {
                        result = method.invoke(result, r.getValue());
                    }
                } else {
                    for (Result r : resultList) {
                        method.invoke(result, r.getValue());
                    }
                }
            } catch (Exception e) {
                throw new RpcException("Can not merge result: " + e.getMessage(), e);
            }
        } else {
            Merger resultMerger;
            // 解析出 merger, 调用 其 merge 方法，返回结果
            if (ConfigUtils.isDefault(merger)) {
                resultMerger = MergerFactory.getMerger(returnType);
            } else {
                resultMerger = ExtensionLoader.getExtensionLoader(Merger.class).getExtension(merger);
            }
            if (resultMerger != null) {
                List<Object> rets = new ArrayList<Object>(resultList.size());
                for (Result r : resultList) {
                    rets.add(r.getValue());
                }
                // 有很多merger, 都在 org.apache.dubbo.rpc.cluster.merger中，
                // 如: MapMerger/Array/Boolean/Int/List/Set/ByteArray...
                result = resultMerger.merge(
                        rets.toArray((Object[]) Array.newInstance(returnType, 0)));
            } else {
                throw new RpcException("There is no merger to merge result.");
            }
        }
        return AsyncRpcResult.newDefaultAsyncResult(result, invocation);
    }

总结: mergeable 容错，依次调用所有invokers, 并通过使用一个merger进行结果合并处理以返回结果。虽然不知道有啥用，但是感觉很厉害的样子。

dubbo的集群容错实现中，使用了 模板方式模式，责任链模式，工厂模式，使得各个容错的实现显得相当简洁明了和简单容易。这就是优秀框架的特性吧。