手写一个RPC框架，理解更透彻（附源码）

发表于 2025-11-04 17:54:06 来源：全栈开发

一、手写前言

前段时间看到一篇不错的个R更透文章《看了这篇你就会手写RPC框架了》，于是框架便来了兴趣对着实现了一遍，后面觉得还有很多优化的理解地方便对其进行了改进。

主要改动点如下：

除了Java序列化协议，彻附增加了protobuf和kryo序列化协议，源码配置即用。手写增加多种负载均衡算法（随机、个R更透轮询、框架加权轮询、理解平滑加权轮询），彻附配置即用。源码客户端增加本地服务列表缓存，手写提高性能。个R更透修复高并发情况下，框架netty导致的内存泄漏问题由原来的每个请求建立一次连接，改为建立TCP长连接，并多次复用。服务端增加线程池提高消息处理能力

二、介绍

RPC，即 Remote Procedure Call（远程过程调用），调用远程计算机上的服务，就像调用本地服务一样。RPC可以很好的解耦系统，如WebService就是一种基于Http协议的RPC。亿华云计算

总的来说，就如下几个步骤：

客户端（ServerA）执行远程方法时就调用client stub传递类名、方法名和参数等信息。 client stub会将参数等信息序列化为二进制流的形式，然后通过Sockect发送给服务端（ServerB）服务端收到数据包后，server stub 需要进行解析反序列化为类名、方法名和参数等信息。 server stub调用对应的本地方法，并把执行结果返回给客户端

所以一个RPC框架有如下角色：

服务消费者远程方法的调用方，即客户端。一个服务既可以是消费者也可以是提供者。服务提供者远程服务的提供方，即服务端。一个服务既可以是消费者也可以是提供者。注册中心保存服务提供者的服务地址等信息，一般由zookeeper、redis等实现。监控运维（可选）监控接口的响应时间、统计请求数量等，及时发现系统问题并发出告警通知。

三、实现

本RPC框架rpc-spring-boot-starter涉及技术栈如下：

使用zookeeper作为注册中心使用netty作为通信框架消息编解码：protostuff、网站模板kryo、java spring 使用SPI来根据配置动态选择负载均衡算法等

由于代码过多，这里只讲几处改动点。

3.1动态负载均衡算法

1.编写LoadBalance的实现类

2.自定义注解 @LoadBalanceAno

/** * 负载均衡注解 */ @Target(ElementType.TYPE) @Retention(RetentionPolicy.RUNTIME) @Documented public @interface LoadBalanceAno { String value() default ""; } /** * 轮询算法 */ @LoadBalanceAno(RpcConstant.BALANCE_ROUND) public class FullRoundBalance implements LoadBalance { private static Logger logger = LoggerFactory.getLogger(FullRoundBalance.class); private volatile int index; @Override public synchronized Service chooseOne(List<Service> services) { // 加锁防止多线程情况下，index超出services.size() if (index == services.size()) { index = 0; } return services.get(index++); } }

3.新建在resource目录下META-INF/servers文件夹并创建文件

4.RpcConfig增加配置项loadBalance

/** * @author 2YSP * @date 2020/7/26 15:13 */ @ConfigurationProperties(prefix = "sp.rpc") public class RpcConfig { /** * 服务注册中心地址 */ private String registerAddress = "127.0.0.1:2181"; /** * 服务暴露端口 */ private Integer serverPort = 9999; /** * 服务协议 */ private String protocol = "java"; /** * 负载均衡算法 */ private String loadBalance = "random"; /** * 权重，默认为1 */ private Integer weight = 1; // 省略getter setter }

5.在自动配置类RpcAutoConfiguration根据配置选择对应的算法实现类

/** * 使用spi匹配符合配置的负载均衡算法 * * @param name * @return */ private LoadBalance getLoadBalance(String name) { ServiceLoader<LoadBalance> loader = ServiceLoader.load(LoadBalance.class); Iterator<LoadBalance> iterator = loader.iterator(); while (iterator.hasNext()) { LoadBalance loadBalance = iterator.next(); LoadBalanceAno ano = loadBalance.getClass().getAnnotation(LoadBalanceAno.class); Assert.notNull(ano, "load balance name can not be empty!"); if (name.equals(ano.value())) { return loadBalance; } } throw new RpcException("invalid load balance config"); } @Bean public ClientProxyFactory proxyFactory(@Autowired RpcConfig rpcConfig) { ClientProxyFactory clientProxyFactory = new ClientProxyFactory(); // 设置服务发现着 clientProxyFactory.setServerDiscovery(new ZookeeperServerDiscovery(rpcConfig.getRegisterAddress())); // 设置支持的协议 Map<String, MessageProtocol> supportMessageProtocols = buildSupportMessageProtocols(); clientProxyFactory.setSupportMessageProtocols(supportMessageProtocols); // 设置负载均衡算法 LoadBalance loadBalance = getLoadBalance(rpcConfig.getLoadBalance()); clientProxyFactory.setLoadBalance(loadBalance); // 设置网络层实现 clientProxyFactory.setNetClient(new NettyNetClient()); return clientProxyFactory; }

3.2本地服务列表缓存

使用Map来缓存数据

/** * 服务发现本地缓存 */ public class ServerDiscoveryCache { /** * key: serviceName */ private static final Map<String, List<Service>> SERVER_MAP = new ConcurrentHashMap<>(); /** * 客户端注入的远程服务service class */ public static final List<String> SERVICE_CLASS_NAMES = new ArrayList<>(); public static void put(String serviceName, List<Service> serviceList) { SERVER_MAP.put(serviceName, serviceList); } /** * 去除指定的值 * @param serviceName * @param service */ public static void remove(String serviceName, Service service) { SERVER_MAP.computeIfPresent(serviceName, (key, value) -> value.stream().filter(o -> !o.toString().equals(service.toString())).collect(Collectors.toList()) ); } public static void removeAll(String serviceName) { SERVER_MAP.remove(serviceName); } public static boolean isEmpty(String serviceName) { return SERVER_MAP.get(serviceName) == null || SERVER_MAP.get(serviceName).size() == 0; } public static List<Service> get(String serviceName) { return SERVER_MAP.get(serviceName); } }

ClientProxyFactory，先查本地缓存，缓存没有再查询zookeeper。

/** * 根据服务名获取可用的服务地址列表 * @param serviceName * @return */ private List<Service> getServiceList(String serviceName) { List<Service> services; synchronized (serviceName){ if (ServerDiscoveryCache.isEmpty(serviceName)) { services = serverDiscovery.findServiceList(serviceName); if (services == null || services.size() == 0) { throw new RpcException("No provider available!"); } ServerDiscoveryCache.put(serviceName, services); } else { services = ServerDiscoveryCache.get(serviceName); } } return services; }

问题：如果服务端因为宕机或网络问题下线了，缓存却还在就会导致客户端请求已经不可用的服务端，增加请求失败率。解决方案：由于服务端注册的是临时节点，所以如果服务端下线节点会被移除。只要监听zookeeper的子节点，如果新增或删除子节点就直接清空本地缓存即可。

推荐：100道Java中高级面试题汇总+详细拆解

DefaultRpcProcessor /** * Rpc处理者，支持服务启动暴露，自动注入Service * @author 2YSP * @date 2020/7/26 14:46 */ public class DefaultRpcProcessor implements ApplicationListener<ContextRefreshedEvent> { @Override public void onApplicationEvent(ContextRefreshedEvent event) { // Spring启动完毕过后会收到一个事件通知 if (Objects.isNull(event.getApplicationContext().getParent())){ ApplicationContext context = event.getApplicationContext(); // 开启服务 startServer(context); // 注入Service injectService(context); } } private void injectService(ApplicationContext context) { String[] names = context.getBeanDefinitionNames(); for(String name : names){ Class<?> clazz = context.getType(name); if (Objects.isNull(clazz)){ continue; } Field[] declaredFields = clazz.getDeclaredFields(); for(Field field : declaredFields){ // 找出标记了InjectService注解的属性 InjectService injectService = field.getAnnotation(InjectService.class); if (injectService == null){ continue; } Class<?> fieldfieldClass = field.getType(); Object object = context.getBean(name); field.setAccessible(true); try { field.set(object,clientProxyFactory.getProxy(fieldClass)); } catch (IllegalAccessException e) { e.printStackTrace(); } // 添加本地服务缓存 ServerDiscoveryCache.SERVICE_CLASS_NAMES.add(fieldClass.getName()); } } // 注册子节点监听 if (clientProxyFactory.getServerDiscovery() instanceof ZookeeperServerDiscovery){ ZookeeperServerDiscovery serverDiscovery = (ZookeeperServerDiscovery) clientProxyFactory.getServerDiscovery(); ZkClient zkClient = serverDiscovery.getZkClient(); ServerDiscoveryCache.SERVICE_CLASS_NAMES.forEach(name ->{ String servicePath = RpcConstant.ZK_SERVICE_PATH + RpcConstant.PATH_DELIMITER + name + "/service"; zkClient.subscribeChildChanges(servicePath, new ZkChildListenerImpl()); }); logger.info("subscribe service zk node successfully"); } } private void startServer(ApplicationContext context) { ... } }

ZkChildListenerImpl

/** * 子节点事件监听处理类 */ public class ZkChildListenerImpl implements IZkChildListener { private static Logger logger = LoggerFactory.getLogger(ZkChildListenerImpl.class); /** * 监听子节点的删除和新增事件 * @param parentPath /rpc/serviceName/service * @param childList * @throws Exception */ @Override public void handleChildChange(String parentPath, List<String> childList) throws Exception { logger.debug("Child change parentPath:[{}] -- childList:[{}]", parentPath, childList); // 只要子节点有改动就清空缓存 String[] arr = parentPath.split("/"); ServerDiscoveryCache.removeAll(arr[2]); } }

3.3nettyClient支持TCP长连接

这部分的改动最多，先增加新的云服务器sendRequest接口。

添加接口

实现类NettyNetClient

/** * @author 2YSP * @date 2020/7/25 20:12 */ public class NettyNetClient implements NetClient { private static Logger logger = LoggerFactory.getLogger(NettyNetClient.class); private static ExecutorService threadPool = new ThreadPoolExecutor(4, 10, 200, TimeUnit.SECONDS, new LinkedBlockingQueue<>(1000), new ThreadFactoryBuilder() .setNameFormat("rpcClient-%d") .build()); private EventLoopGroup loopGroup = new NioEventLoopGroup(4); /** * 已连接的服务缓存 * key: 服务地址，格式：ip:port */ public static Map<String, SendHandlerV2> connectedServerNodes = new ConcurrentHashMap<>(); @Override public byte[] sendRequest(byte[] data, Service service) throws InterruptedException { .... return respData; } @Override public RpcResponse sendRequest(RpcRequest rpcRequest, Service service, MessageProtocol messageProtocol) { String address = service.getAddress(); synchronized (address) { if (connectedServerNodes.containsKey(address)) { SendHandlerV2 handler = connectedServerNodes.get(address); logger.info("使用现有的连接"); return handler.sendRequest(rpcRequest); } String[] addrInfo = address.split(":"); final String serverAddress = addrInfo[0]; final String serverPort = addrInfo[1]; final SendHandlerV2 handler = new SendHandlerV2(messageProtocol, address); threadPool.submit(() -> { // 配置客户端 Bootstrap b = new Bootstrap(); b.group(loopGroup).channel(NioSocketChannel.class) .option(ChannelOption.TCP_NODELAY, true) .handler(new ChannelInitializer<SocketChannel>() { @Override protected void initChannel(SocketChannel socketChannel) throws Exception { ChannelPipeline pipeline = socketChannel.pipeline(); pipeline .addLast(handler); } }); // 启用客户端连接 ChannelFuture channelFuture = b.connect(serverAddress, Integer.parseInt(serverPort)); channelFuture.addListener(new ChannelFutureListener() { @Override public void operationComplete(ChannelFuture channelFuture) throws Exception { connectedServerNodes.put(address, handler); } }); } ); logger.info("使用新的连接。。。"); return handler.sendRequest(rpcRequest); } } }

每次请求都会调用sendRequest()方法，用线程池异步和服务端创建TCP长连接，连接成功后将SendHandlerV2缓存到ConcurrentHashMap中方便复用，后续请求的请求地址（ip+port）如果在connectedServerNodes中存在则使用connectedServerNodes中的handler处理不再重新建立连接。

SendHandlerV2

/** * @author 2YSP * @date 2020/8/19 20:06 */ public class SendHandlerV2 extends ChannelInboundHandlerAdapter { private static Logger logger = LoggerFactory.getLogger(SendHandlerV2.class); /** * 等待通道建立最大时间 */ static final int CHANNEL_WAIT_TIME = 4; /** * 等待响应最大时间 */ static final int RESPONSE_WAIT_TIME = 8; private volatile Channel channel; private String remoteAddress; private static Map<String, RpcFuture<RpcResponse>> requestMap = new ConcurrentHashMap<>(); private MessageProtocol messageProtocol; private CountDownLatch latch = new CountDownLatch(1); public SendHandlerV2(MessageProtocol messageProtocol,String remoteAddress) { this.messageProtocol = messageProtocol; this.remoteAddress = remoteAddress; } @Override public void channelRegistered(ChannelHandlerContext ctx) throws Exception { this.channel = ctx.channel(); latch.countDown(); } @Override public void channelActive(ChannelHandlerContext ctx) throws Exception { logger.debug("Connect to server successfully:{}", ctx); } @Override public void channelRead(ChannelHandlerContext ctx, Object msg) throws Exception { logger.debug("Client reads message:{}", msg); ByteBuf byteBuf = (ByteBuf) msg; byte[] resp = new byte[byteBuf.readableBytes()]; byteBuf.readBytes(resp); // 手动回收 ReferenceCountUtil.release(byteBuf); RpcResponse response = messageProtocol.unmarshallingResponse(resp); RpcFuture<RpcResponse> future = requestMap.get(response.getRequestId()); future.setResponse(response); } @Override public void exceptionCaught(ChannelHandlerContext ctx, Throwable cause) throws Exception { cause.printStackTrace(); logger.error("Exception occurred:{}", cause.getMessage()); ctx.close(); } @Override public void channelReadComplete(ChannelHandlerContext ctx) throws Exception { ctx.flush(); } @Override public void channelInactive(ChannelHandlerContext ctx) throws Exception { super.channelInactive(ctx); logger.error("channel inactive with remoteAddress:[{}]",remoteAddress); NettyNetClient.connectedServerNodes.remove(remoteAddress); } @Override public void userEventTriggered(ChannelHandlerContext ctx, Object evt) throws Exception { super.userEventTriggered(ctx, evt); } public RpcResponse sendRequest(RpcRequest request) { RpcResponse response; RpcFuture<RpcResponse> future = new RpcFuture<>(); requestMap.put(request.getRequestId(), future); try { byte[] data = messageProtocol.marshallingRequest(request); ByteBuf reqBuf = Unpooled.buffer(data.length); reqBuf.writeBytes(data); if (latch.await(CHANNEL_WAIT_TIME,TimeUnit.SECONDS)){ channel.writeAndFlush(reqBuf); // 等待响应 response = future.get(RESPONSE_WAIT_TIME, TimeUnit.SECONDS); }else { throw new RpcException("establish channel time out"); } } catch (Exception e) { throw new RpcException(e.getMessage()); } finally { requestMap.remove(request.getRequestId()); } return response; } }

RpcFuture

package cn.sp.rpc.client.net; import java.util.concurrent.*; /** * @author 2YSP * @date 2020/8/19 22:31 */ public class RpcFuture<T> implements Future<T> { private T response; /** * 因为请求和响应是一一对应的，所以这里是1 */ private CountDownLatch countDownLatch = new CountDownLatch(1); /** * Future的请求时间，用于计算Future是否超时 */ private long beginTime = System.currentTimeMillis(); @Override public boolean cancel(boolean mayInterruptIfRunning) { return false; } @Override public boolean isCancelled() { return false; } @Override public boolean isDone() { if (response != null) { return true; } return false; } /** * 获取响应，直到有结果才返回 * @return * @throws InterruptedException * @throws ExecutionException */ @Override public T get() throws InterruptedException, ExecutionException { countDownLatch.await(); return response; } @Override public T get(long timeout, TimeUnit unit) throws InterruptedException, ExecutionException, TimeoutException { if (countDownLatch.await(timeout,unit)){ return response; } return null; } public void setResponse(T response) { this.response = response; countDownLatch.countDown(); } public long getBeginTime() { return beginTime; } }

此处逻辑，第一次执行 SendHandlerV2#sendRequest() 时channel需要等待通道建立好之后才能发送请求，所以用CountDownLatch来控制，等待通道建立。

自定义Future+requestMap缓存来实现netty的请求和阻塞等待响应，RpcRequest对象在创建时会生成一个请求的唯一标识requestId，发送请求前先将RpcFuture缓存到requestMap中，key为requestId，读取到服务端的响应信息后(channelRead方法)，将响应结果放入对应的RpcFuture中。

SendHandlerV2#channelInactive() 方法中，如果连接的服务端异常断开连接了，则及时清理缓存中对应的serverNode。

四、压力测试

测试环境：

(英特尔)Intel(R) Core(TM) i5-6300HQ CPU @ 2.30GHz 4核 windows10家庭版（64位） 16G内存

1.本地启动zookeeper

2.本地启动一个消费者，两个服务端，轮询算法

3.使用ab进行压力测试，4个线程发送10000个请求

ab -c 4 -n 10000 http://localhost:8080/test/user?id=1

测试结果：