手写一个RPC框架

一、前言

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

主要改动点如下：

1. 除了Java序列化协议，增加了protobuf和kryo序列化协议，配置即用。
2. 增加多种负载均衡算法（随机、轮询、加权轮询、平滑加权轮询），配置即用。
3. 客户端增加本地服务列表缓存，提高性能。
4. 修复高并发情况下，netty导致的内存泄漏问题
5. 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。
6. 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。

调用示意图

总的来说，就如下几个步骤：

1. 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。
2. 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）
3. 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。
4. 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、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

负载均衡算法实现类

2.自定义注解 @LoadBalanceAno

1. /** 
   
   

   ---

2. * 负载均衡注解 
   
   

   ---

3. */ 
   
   

   ---

4. @Target(ElementType.TYPE) 
   
   

   ---

5. @Retention(RetentionPolicy.RUNTIME) 
   
   

   ---

6. @Documented 
   
   

   ---

7. public @interface LoadBalanceAno { 
   
   

   ---

8. ---

9. String value() default ""; 
   
   

   ---

10. } 
    
    

    ---

11. ---

12. /** 
    
    

    ---

13. * 轮询算法 
    
    

    ---

14. */ 
    
    

    ---

15. @LoadBalanceAno(RpcConstant.BALANCE_ROUND) 
    
    

    ---

16. public class FullRoundBalance implements LoadBalance { 
    
    

    ---

17. ---

18. private static Logger logger = LoggerFactory.getLogger(FullRoundBalance.class); 
    
    

    ---

19. ---

20. private volatile int index; 
    
    

    ---

21. ---

22. @Override 
    
    

    ---

23. public synchronized Service chooseOne(List<Service> services) { 
    
    

    ---

24. // 加锁防止多线程情况下，index超出services.size() 
    
    

    ---

25. if (index == services.size()) { 
    
    

    ---

26. index = 0; 
    
    

    ---

27. } 
    
    

    ---

28. return services.get(index++); 
    
    

    ---

29. } 
    
    

    ---

30. } 
    
    

    ---

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

enter description here

4.RpcConfig增加配置项loadBalance

1. /** 
   
   

   ---

2. * @author 2YSP 
   
   

   ---

3. * @date 2020/7/26 15:13 
   
   

   ---

4. */ 
   
   

   ---

5. @ConfigurationProperties(prefix = "sp.rpc") 
   
   

   ---

6. public class RpcConfig { 
   
   

   ---

7. ---

8. /** 
   
   

   ---

9. * 服务注册中心地址 
   
   

   ---

10. */ 
    
    

    ---

11. private String registerAddress = "127.0.0.1:2181"; 
    
    

    ---

12. ---

13. /** 
    
    

    ---

14. * 服务暴露端口 
    
    

    ---

15. */ 
    
    

    ---

16. private Integer serverPort = 9999; 
    
    

    ---

17. /** 
    
    

    ---

18. * 服务协议 
    
    

    ---

19. */ 
    
    

    ---

20. private String protocol = "java"; 
    
    

    ---

21. /** 
    
    

    ---

22. * 负载均衡算法 
    
    

    ---

23. */ 
    
    

    ---

24. private String loadBalance = "random"; 
    
    

    ---

25. /** 
    
    

    ---

26. * 权重，默认为1 
    
    

    ---

27. */ 
    
    

    ---

28. private Integer weight = 1; 
    
    

    ---

29. ---

30. // 省略getter setter 
    
    

    ---

31. } 
    
    

    ---

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

1. /** 
   
   

   ---

2. * 使用spi匹配符合配置的负载均衡算法 
   
   

   ---

3. * 
   
   

   ---

4. * @param name 
   
   

   ---

5. * @return 
   
   

   ---

6. */ 
   
   

   ---

7. private LoadBalance getLoadBalance(String name) { 
   
   

   ---

8. ServiceLoader<LoadBalance> loader = ServiceLoader.load(LoadBalance.class); 
   
   

   ---

9. Iterator<LoadBalance> iterator = loader.iterator(); 
   
   

   ---

10. while (iterator.hasNext()) { 
    
    

    ---

11. LoadBalance loadBalance = iterator.next(); 
    
    

    ---

12. LoadBalanceAno ano = loadBalance.getClass().getAnnotation(LoadBalanceAno.class); 
    
    

    ---

13. Assert.notNull(ano, "load balance name can not be empty!"); 
    
    

    ---

14. if (name.equals(ano.value())) { 
    
    

    ---

15. return loadBalance; 
    
    

    ---

16. } 
    
    

    ---

17. } 
    
    

    ---

18. throw new RpcException("invalid load balance config"); 
    
    

    ---

19. } 
    
    

    ---

20. ---

21. @Bean 
    
    

    ---

22. public ClientProxyFactory proxyFactory(@Autowired RpcConfig rpcConfig) { 
    
    

    ---

23. ClientProxyFactory clientProxyFactory = new ClientProxyFactory(); 
    
    

    ---

24. // 设置服务发现着 
    
    

    ---

25. clientProxyFactory.setServerDiscovery(new ZookeeperServerDiscovery(rpcConfig.getRegisterAddress())); 
    
    

    ---

26. ---

27. // 设置支持的协议 
    
    

    ---

28. Map<String, MessageProtocol> supportMessageProtocols = buildSupportMessageProtocols(); 
    
    

    ---

29. clientProxyFactory.setSupportMessageProtocols(supportMessageProtocols); 
    
    

    ---

30. // 设置负载均衡算法 
    
    

    ---

31. LoadBalance loadBalance = getLoadBalance(rpcConfig.getLoadBalance()); 
    
    

    ---

32. clientProxyFactory.setLoadBalance(loadBalance); 
    
    

    ---

33. // 设置网络层实现 
    
    

    ---

34. clientProxyFactory.setNetClient(new NettyNetClient()); 
    
    

    ---

35. ---

36. return clientProxyFactory; 
    
    

    ---

37. } 
    
    

    ---

3.2本地服务列表缓存

使用Map来缓存数据

1. /** 
   
   

   ---

2. * 服务发现本地缓存 
   
   

   ---

3. */ 
   
   

   ---

4. public class ServerDiscoveryCache { 
   
   

   ---

5. /** 
   
   

   ---

6. * key: serviceName 
   
   

   ---

7. */ 
   
   

   ---

8. private static final Map<String, List<Service>> SERVER_MAP = new ConcurrentHashMap<>(); 
   
   

   ---

9. /** 
   
   

   ---

10. * 客户端注入的远程服务service class 
    
    

    ---

11. */ 
    
    

    ---

12. public static final List<String> SERVICE_CLASS_NAMES = new ArrayList<>(); 
    
    

    ---

13. ---

14. public static void put(String serviceName, List<Service> serviceList) { 
    
    

    ---

15. SERVER_MAP.put(serviceName, serviceList); 
    
    

    ---

16. } 
    
    

    ---

17. ---

18. /** 
    
    

    ---

19. * 去除指定的值 
    
    

    ---

20. * @param serviceName 
    
    

    ---

21. * @param service 
    
    

    ---

22. */ 
    
    

    ---

23. public static void remove(String serviceName, Service service) { 
    
    

    ---

24. SERVER_MAP.computeIfPresent(serviceName, (key, value) -> 
    
    

    ---

25. value.stream().filter(o -> !o.toString().equals(service.toString())).collect(Collectors.toList()) 
    
    

    ---

26. ); 
    
    

    ---

27. } 
    
    

    ---

28. ---

29. public static void removeAll(String serviceName) { 
    
    

    ---

30. SERVER_MAP.remove(serviceName); 
    
    

    ---

31. } 
    
    

    ---

32. ---

33. ---

34. public static boolean isEmpty(String serviceName) { 
    
    

    ---

35. return SERVER_MAP.get(serviceName) == null || SERVER_MAP.get(serviceName).size() == 0; 
    
    

    ---

36. } 
    
    

    ---

37. ---

38. public static List<Service> get(String serviceName) { 
    
    

    ---

39. return SERVER_MAP.get(serviceName); 
    
    

    ---

40. } 
    
    

    ---

41. } 
    
    

    ---

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

1. /** 
   
   

   ---

2. * 根据服务名获取可用的服务地址列表 
   
   

   ---

3. * @param serviceName 
   
   

   ---

4. * @return 
   
   

   ---

5. */ 
   
   

   ---

6. private List<Service> getServiceList(String serviceName) { 
   
   

   ---

7. List<Service> services; 
   
   

   ---

8. synchronized (serviceName){ 
   
   

   ---

9. if (ServerDiscoveryCache.isEmpty(serviceName)) { 
   
   

   ---

10. services = serverDiscovery.findServiceList(serviceName); 
    
    

    ---

11. if (services == null || services.size() == 0) { 
    
    

    ---

12. throw new RpcException("No provider available!"); 
    
    

    ---

13. } 
    
    

    ---

14. ServerDiscoveryCache.put(serviceName, services); 
    
    

    ---

15. } else { 
    
    

    ---

16. services = ServerDiscoveryCache.get(serviceName); 
    
    

    ---

17. } 
    
    

    ---

18. } 
    
    

    ---

19. return services; 
    
    

    ---

20. } 
    
    

    ---

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。

DefaultRpcProcessor

1. /** 
   
   

   ---

2. * Rpc处理者，支持服务启动暴露，自动注入Service 
   
   

   ---

3. * @author 2YSP 
   
   

   ---

4. * @date 2020/7/26 14:46 
   
   

   ---

5. */ 
   
   

   ---

6. public class DefaultRpcProcessor implements ApplicationListener<ContextRefreshedEvent> { 
   
   

   ---

7. ---

8. ---

9. ---

10. @Override 
    
    

    ---

11. public void onApplicationEvent(ContextRefreshedEvent event) { 
    
    

    ---

12. // Spring启动完毕过后会收到一个事件通知 
    
    

    ---

13. if (Objects.isNull(event.getApplicationContext().getParent())){ 
    
    

    ---

14. ApplicationContext context = event.getApplicationContext(); 
    
    

    ---

15. // 开启服务 
    
    

    ---

16. startServer(context); 
    
    

    ---

17. // 注入Service 
    
    

    ---

18. injectService(context); 
    
    

    ---

19. } 
    
    

    ---

20. } 
    
    

    ---

21. ---

22. private void injectService(ApplicationContext context) { 
    
    

    ---

23. String[] names = context.getBeanDefinitionNames(); 
    
    

    ---

24. for(String name : names){ 
    
    

    ---

25. Class<?> clazz = context.getType(name); 
    
    

    ---

26. if (Objects.isNull(clazz)){ 
    
    

    ---

27. continue; 
    
    

    ---

28. } 
    
    

    ---

29. ---

30. Field[] declaredFields = clazz.getDeclaredFields(); 
    
    

    ---

31. for(Field field : declaredFields){ 
    
    

    ---

32. // 找出标记了InjectService注解的属性 
    
    

    ---

33. InjectService injectService = field.getAnnotation(InjectService.class); 
    
    

    ---

34. if (injectService == null){ 
    
    

    ---

35. continue; 
    
    

    ---

36. } 
    
    

    ---

37. ---

38. Class<?> fieldClass = field.getType(); 
    
    

    ---

39. Object object = context.getBean(name); 
    
    

    ---

40. field.setAccessible(true); 
    
    

    ---

41. try { 
    
    

    ---

42. field.set(object,clientProxyFactory.getProxy(fieldClass)); 
    
    

    ---

43. } catch (IllegalAccessException e) { 
    
    

    ---

44. e.printStackTrace(); 
    
    

    ---

45. } 
    
    

    ---

46. // 添加本地服务缓存 
    
    

    ---

47. ServerDiscoveryCache.SERVICE_CLASS_NAMES.add(fieldClass.getName()); 
    
    

    ---

48. } 
    
    

    ---

49. } 
    
    

    ---

50. // 注册子节点监听 
    
    

    ---

51. if (clientProxyFactory.getServerDiscovery() instanceof ZookeeperServerDiscovery){ 
    
    

    ---

52. ZookeeperServerDiscovery serverDiscovery = (ZookeeperServerDiscovery) clientProxyFactory.getServerDiscovery(); 
    
    

    ---

53. ZkClient zkClient = serverDiscovery.getZkClient(); 
    
    

    ---

54. ServerDiscoveryCache.SERVICE_CLASS_NAMES.forEach(name ->{ 
    
    

    ---

55. String servicePath = RpcConstant.ZK_SERVICE_PATH + RpcConstant.PATH_DELIMITER + name + "/service"; 
    
    

    ---

56. zkClient.subscribeChildChanges(servicePath, new ZkChildListenerImpl()); 
    
    

    ---

57. }); 
    
    

    ---

58. logger.info("subscribe service zk node successfully"); 
    
    

    ---

59. } 
    
    

    ---

60. ---

61. } 
    
    

    ---

62. ---

63. private void startServer(ApplicationContext context) { 
    
    

    ---

64. ... 
    
    

    ---

65. ---

66. } 
    
    

    ---

67. } 
    
    

    ---

68. ---

ZkChildListenerImpl

1. /** 
   
   

   ---

2. * 子节点事件监听处理类 
   
   

   ---

3. */ 
   
   

   ---

4. public class ZkChildListenerImpl implements IZkChildListener { 
   
   

   ---

5. ---

6. private static Logger logger = LoggerFactory.getLogger(ZkChildListenerImpl.class); 
   
   

   ---

7. ---

8. /** 
   
   

   ---

9. * 监听子节点的删除和新增事件 
   
   

   ---

10. * @param parentPath /rpc/serviceName/service 
    
    

    ---

11. * @param childList 
    
    

    ---

12. * @throws Exception 
    
    

    ---

13. */ 
    
    

    ---

14. @Override 
    
    

    ---

15. public void handleChildChange(String parentPath, List<String> childList) throws Exception { 
    
    

    ---

16. logger.debug("Child change parentPath:[{}] -- childList:[{}]", parentPath, childList); 
    
    

    ---

17. // 只要子节点有改动就清空缓存 
    
    

    ---

18. String[] arr = parentPath.split("/"); 
    
    

    ---

19. ServerDiscoveryCache.removeAll(arr[2]); 
    
    

    ---

20. } 
    
    

    ---

21. } 
    
    

    ---

3.3nettyClient支持TCP长连接

这部分的改动最多，先增加新的sendRequest接口。

添加接口

实现类NettyNetClient

1. /** 
   
   

   ---

2. * @author 2YSP 
   
   

   ---

3. * @date 2020/7/25 20:12 
   
   

   ---

4. */ 
   
   

   ---

5. public class NettyNetClient implements NetClient { 
   
   

   ---

6. ---

7. private static Logger logger = LoggerFactory.getLogger(NettyNetClient.class); 
   
   

   ---

8. ---

9. private static ExecutorService threadPool = new ThreadPoolExecutor(4, 10, 200, 
   
   

   ---

10. TimeUnit.SECONDS, new LinkedBlockingQueue<>(1000), new ThreadFactoryBuilder() 
    
    

    ---

11. .setNameFormat("rpcClient-%d") 
    
    

    ---

12. .build()); 
    
    

    ---

13. ---

14. private EventLoopGroup loopGroup = new NioEventLoopGroup(4); 
    
    

    ---

15. ---

16. /** 
    
    

    ---

17. * 已连接的服务缓存 
    
    

    ---

18. * key: 服务地址，格式：ip:port 
    
    

    ---

19. */ 
    
    

    ---

20. public static Map<String, SendHandlerV2> connectedServerNodes = new ConcurrentHashMap<>(); 
    
    

    ---

21. ---

22. @Override 
    
    

    ---

23. public byte[] sendRequest(byte[] data, Service service) throws InterruptedException { 
    
    

    ---

24. .... 
    
    

    ---

25. return respData; 
    
    

    ---

26. } 
    
    

    ---

27. ---

28. @Override 
    
    

    ---

29. public RpcResponse sendRequest(RpcRequest rpcRequest, Service service, MessageProtocol messageProtocol) { 
    
    

    ---

30. ---

31. String address = service.getAddress(); 
    
    

    ---

32. synchronized (address) { 
    
    

    ---

33. if (connectedServerNodes.containsKey(address)) { 
    
    

    ---

34. SendHandlerV2 handler = connectedServerNodes.get(address); 
    
    

    ---

35. logger.info("使用现有的连接"); 
    
    

    ---

36. return handler.sendRequest(rpcRequest); 
    
    

    ---

37. } 
    
    

    ---

38. ---

39. String[] addrInfo = address.split(":"); 
    
    

    ---

40. final String serverAddress = addrInfo[0]; 
    
    

    ---

41. final String serverPort = addrInfo[1]; 
    
    

    ---

42. final SendHandlerV2 handler = new SendHandlerV2(messageProtocol, address); 
    
    

    ---

43. threadPool.submit(() -> { 
    
    

    ---

44. // 配置客户端 
    
    

    ---

45. Bootstrap b = new Bootstrap(); 
    
    

    ---

46. b.group(loopGroup).channel(NioSocketChannel.class) 
    
    

    ---

47. .option(ChannelOption.TCP_NODELAY, true) 
    
    

    ---

48. .handler(new ChannelInitializer<SocketChannel>() { 
    
    

    ---

49. @Override 
    
    

    ---

50. protected void initChannel(SocketChannel socketChannel) throws Exception { 
    
    

    ---

51. ChannelPipeline pipeline = socketChannel.pipeline(); 
    
    

    ---

52. pipeline 
    
    

    ---

53. .addLast(handler); 
    
    

    ---

54. } 
    
    

    ---

55. }); 
    
    

    ---

56. // 启用客户端连接 
    
    

    ---

57. ChannelFuture channelFuture = b.connect(serverAddress, Integer.parseInt(serverPort)); 
    
    

    ---

58. channelFuture.addListener(new ChannelFutureListener() { 
    
    

    ---

59. @Override 
    
    

    ---

60. public void operationComplete(ChannelFuture channelFuture) throws Exception { 
    
    

    ---

61. connectedServerNodes.put(address, handler); 
    
    

    ---

62. } 
    
    

    ---

63. }); 
    
    

    ---

64. } 
    
    

    ---

65. ); 
    
    

    ---

66. logger.info("使用新的连接。。。"); 
    
    

    ---

67. return handler.sendRequest(rpcRequest); 
    
    

    ---

68. } 
    
    

    ---

69. } 
    
    

    ---

70. } 
    
    

    ---

71. ---

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

1. /** 
   
   

   ---

2. * @author 2YSP 
   
   

   ---

3. * @date 2020/8/19 20:06 
   
   

   ---

4. */ 
   
   

   ---

5. public class SendHandlerV2 extends ChannelInboundHandlerAdapter { 
   
   

   ---

6. ---

7. private static Logger logger = LoggerFactory.getLogger(SendHandlerV2.class); 
   
   

   ---

8. ---

9. /** 
   
   

   ---

10. * 等待通道建立最大时间 
    
    

    ---

11. */ 
    
    

    ---

12. static final int CHANNEL_WAIT_TIME = 4; 
    
    

    ---

13. /** 
    
    

    ---

14. * 等待响应最大时间 
    
    

    ---

15. */ 
    
    

    ---

16. static final int RESPONSE_WAIT_TIME = 8; 
    
    

    ---

17. ---

18. private volatile Channel channel; 
    
    

    ---

19. ---

20. private String remoteAddress; 
    
    

    ---

21. ---

22. private static Map<String, RpcFuture<RpcResponse>> requestMap = new ConcurrentHashMap<>(); 
    
    

    ---

23. ---

24. private MessageProtocol messageProtocol; 
    
    

    ---

25. ---

26. private CountDownLatch latch = new CountDownLatch(1); 
    
    

    ---

27. ---

28. public SendHandlerV2(MessageProtocol messageProtocol,String remoteAddress) { 
    
    

    ---

29. this.messageProtocol = messageProtocol; 
    
    

    ---

30. this.remoteAddress = remoteAddress; 
    
    

    ---

31. } 
    
    

    ---

32. ---

33. @Override 
    
    

    ---

34. public void channelRegistered(ChannelHandlerContext ctx) throws Exception { 
    
    

    ---

35. this.channel = ctx.channel(); 
    
    

    ---

36. latch.countDown(); 
    
    

    ---

37. } 
    
    

    ---

38. ---

39. @Override 
    
    

    ---

40. public void channelActive(ChannelHandlerContext ctx) throws Exception { 
    
    

    ---

41. logger.debug("Connect to server successfully:{}", ctx); 
    
    

    ---

42. } 
    
    

    ---

43. ---

44. @Override 
    
    

    ---

45. public void channelRead(ChannelHandlerContext ctx, Object msg) throws Exception { 
    
    

    ---

46. logger.debug("Client reads message:{}", msg); 
    
    

    ---

47. ByteBuf byteBuf = (ByteBuf) msg; 
    
    

    ---

48. byte[] resp = new byte[byteBuf.readableBytes()]; 
    
    

    ---

49. byteBuf.readBytes(resp); 
    
    

    ---

50. // 手动回收 
    
    

    ---

51. ReferenceCountUtil.release(byteBuf); 
    
    

    ---

52. RpcResponse response = messageProtocol.unmarshallingResponse(resp); 
    
    

    ---

53. RpcFuture<RpcResponse> future = requestMap.get(response.getRequestId()); 
    
    

    ---

54. future.setResponse(response); 
    
    

    ---

55. } 
    
    

    ---

56. ---

57. @Override 
    
    

    ---

58. public void exceptionCaught(ChannelHandlerContext ctx, Throwable cause) throws Exception { 
    
    

    ---

59. cause.printStackTrace(); 
    
    

    ---

60. logger.error("Exception occurred:{}", cause.getMessage()); 
    
    

    ---

61. ctx.close(); 
    
    

    ---

62. } 
    
    

    ---

63. ---

64. @Override 
    
    

    ---

65. public void channelReadComplete(ChannelHandlerContext ctx) throws Exception { 
    
    

    ---

66. ctx.flush(); 
    
    

    ---

67. } 
    
    

    ---

68. ---

69. @Override 
    
    

    ---

70. public void channelInactive(ChannelHandlerContext ctx) throws Exception { 
    
    

    ---

71. super.channelInactive(ctx); 
    
    

    ---

72. logger.error("channel inactive with remoteAddress:[{}]",remoteAddress); 
    
    

    ---

73. NettyNetClient.connectedServerNodes.remove(remoteAddress); 
    
    

    ---

74. ---

75. } 
    
    

    ---

76. ---

77. @Override 
    
    

    ---

78. public void userEventTriggered(ChannelHandlerContext ctx, Object evt) throws Exception { 
    
    

    ---

79. super.userEventTriggered(ctx, evt); 
    
    

    ---

80. } 
    
    

    ---

81. ---

82. public RpcResponse sendRequest(RpcRequest request) { 
    
    

    ---

83. RpcResponse response; 
    
    

    ---

84. RpcFuture<RpcResponse> future = new RpcFuture<>(); 
    
    

    ---

85. requestMap.put(request.getRequestId(), future); 
    
    

    ---

86. try { 
    
    

    ---

87. byte[] data = messageProtocol.marshallingRequest(request); 
    
    

    ---

88. ByteBuf reqBuf = Unpooled.buffer(data.length); 
    
    

    ---

89. reqBuf.writeBytes(data); 
    
    

    ---

90. if (latch.await(CHANNEL_WAIT_TIME,TimeUnit.SECONDS)){ 
    
    

    ---

91. channel.writeAndFlush(reqBuf); 
    
    

    ---

92. // 等待响应 
    
    

    ---

93. response = future.get(RESPONSE_WAIT_TIME, TimeUnit.SECONDS); 
    
    

    ---

94. }else { 
    
    

    ---

95. throw new RpcException("establish channel time out"); 
    
    

    ---

96. } 
    
    

    ---

97. } catch (Exception e) { 
    
    

    ---

98. throw new RpcException(e.getMessage()); 
    
    

    ---

99. } finally { 
    
    

    ---

100. requestMap.remove(request.getRequestId()); 
     
     

     ---

101. } 
     
     

     ---

102. return response; 
     
     

     ---

103. } 
     
     

     ---

104. } 
     
     

     ---

105. ---

RpcFuture

1. package cn.sp.rpc.client.net; 
   
   

   ---

2. ---

3. import java.util.concurrent.*; 
   
   

   ---

4. ---

5. /** 
   
   

   ---

6. * @author 2YSP 
   
   

   ---

7. * @date 2020/8/19 22:31 
   
   

   ---

8. */ 
   
   

   ---

9. public class RpcFuture<T> implements Future<T> { 
   
   

   ---

10. ---

11. private T response; 
    
    

    ---

12. /** 
    
    

    ---

13. * 因为请求和响应是一一对应的，所以这里是1 
    
    

    ---

14. */ 
    
    

    ---

15. private CountDownLatch countDownLatch = new CountDownLatch(1); 
    
    

    ---

16. /** 
    
    

    ---

17. * Future的请求时间，用于计算Future是否超时 
    
    

    ---

18. */ 
    
    

    ---

19. private long beginTime = System.currentTimeMillis(); 
    
    

    ---

20. ---

21. @Override 
    
    

    ---

22. public boolean cancel(boolean mayInterruptIfRunning) { 
    
    

    ---

23. return false; 
    
    

    ---

24. } 
    
    

    ---

25. ---

26. @Override 
    
    

    ---

27. public boolean isCancelled() { 
    
    

    ---

28. return false; 
    
    

    ---

29. } 
    
    

    ---

30. ---

31. @Override 
    
    

    ---

32. public boolean isDone() { 
    
    

    ---

33. if (response != null) { 
    
    

    ---

34. return true; 
    
    

    ---

35. } 
    
    

    ---

36. return false; 
    
    

    ---

37. } 
    
    

    ---

38. ---

39. /** 
    
    

    ---

40. * 获取响应，直到有结果才返回 
    
    

    ---

41. * @return 
    
    

    ---

42. * @throws InterruptedException 
    
    

    ---

43. * @throws ExecutionException 
    
    

    ---

44. */ 
    
    

    ---

45. @Override 
    
    

    ---

46. public T get() throws InterruptedException, ExecutionException { 
    
    

    ---

47. countDownLatch.await(); 
    
    

    ---

48. return response; 
    
    

    ---

49. } 
    
    

    ---

50. ---

51. @Override 
    
    

    ---

52. public T get(long timeout, TimeUnit unit) throws InterruptedException, ExecutionException, TimeoutException { 
    
    

    ---

53. if (countDownLatch.await(timeout,unit)){ 
    
    

    ---

54. return response; 
    
    

    ---

55. } 
    
    

    ---

56. return null; 
    
    

    ---

57. } 
    
    

    ---

58. ---

59. public void setResponse(T response) { 
    
    

    ---

60. this.response = response; 
    
    

    ---

61. countDownLatch.countDown(); 
    
    

    ---

62. } 
    
    

    ---

63. ---

64. public long getBeginTime() { 
    
    

    ---

65. return beginTime; 
    
    

    ---

66. } 
    
    

    ---

67. } 
    
    

    ---

68. ---

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：

测试结果

从图片可以看出，10000个请求只用了11s，比之前的130+秒耗时减少了10倍以上。

代码地址：

https://github.com/2YSP/rpc-spring-boot-starter

https://github.com/2YSP/rpc-example

参考：

《看了这篇你就会手写RPC框架了》