一个轻量级分布式RPC框架--NettyRpc

1、背景

最近在搜索Netty和Zookeeper方面的文章时,看到了这篇文章《轻量级分布式 RPC 框架》,作者用Zookeeper、Netty和Spring写了一个轻量级的分布式RPC框架。花了一些时间看了下他的代码,写的干净简单,写的RPC框架可以算是一个简易版的dubbo。这个RPC框架虽小,但是麻雀虽小,五脏俱全,有兴趣的可以学习一下。

项目地址:https://github.com/luxiaoxun/NettyRpc

自己花了点时间整理了下代码,并修改一些问题,以下是自己学习的一点小结。

2、简介

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。

这个RPC整体框架如下:

这个RPC框架使用的一些技术所解决的问题:

服务发布与订阅:服务端使用Zookeeper注册服务地址,客户端从Zookeeper获取可用的服务地址。

通信:使用Netty作为通信框架。

Spring:使用Spring配置服务,加载Bean,扫描注解。

动态代理:客户端使用代理模式透明化服务调用。

消息编解码:使用Protostuff序列化和反序列化消息。

3、服务端发布服务

使用注解标注要发布的服务

服务注解

@Target({ElementType.TYPE})
@Retention(RetentionPolicy.RUNTIME)
@Component
public @interface RpcService {
    Class<?> value();
}

一个服务接口:

public interface HelloService {

    String hello(String name);

    String hello(Person person);
}

一个服务实现:使用注解标注

@RpcService(HelloService.class)
public class HelloServiceImpl implements HelloService {

    @Override
    public String hello(String name) {
        return "Hello! " + name;
    }

    @Override
    public String hello(Person person) {
        return "Hello! " + person.getFirstName() + " " + person.getLastName();
    }
}

服务在启动的时候扫描得到所有的服务接口及其实现:

@Override
    public void setApplicationContext(ApplicationContext ctx) throws BeansException {
        Map<String, Object> serviceBeanMap = ctx.getBeansWithAnnotation(RpcService.class);
        if (MapUtils.isNotEmpty(serviceBeanMap)) {
            for (Object serviceBean : serviceBeanMap.values()) {
                String interfaceName = serviceBean.getClass().getAnnotation(RpcService.class).value().getName();
                handlerMap.put(interfaceName, serviceBean);
            }
        }
    }

在Zookeeper集群上注册服务地址:

public class ServiceRegistry {

    private static final Logger LOGGER = LoggerFactory.getLogger(ServiceRegistry.class);

    private CountDownLatch latch = new CountDownLatch(1);

    private String registryAddress;

    public ServiceRegistry(String registryAddress) {
        this.registryAddress = registryAddress;
    }

    public void register(String data) {
        if (data != null) {
            ZooKeeper zk = connectServer();
            if (zk != null) {
                AddRootNode(zk); // Add root node if not exist
                createNode(zk, data);
            }
        }
    }

    private ZooKeeper connectServer() {
        ZooKeeper zk = null;
        try {
            zk = new ZooKeeper(registryAddress, Constant.ZK_SESSION_TIMEOUT, new Watcher() {
                @Override
                public void process(WatchedEvent event) {
                    if (event.getState() == Event.KeeperState.SyncConnected) {
                        latch.countDown();
                    }
                }
            });
            latch.await();
        } catch (IOException e) {
            LOGGER.error("", e);
        }
        catch (InterruptedException ex){
            LOGGER.error("", ex);
        }
        return zk;
    }

    private void AddRootNode(ZooKeeper zk){
        try {
            Stat s = zk.exists(Constant.ZK_REGISTRY_PATH, false);
            if (s == null) {
                zk.create(Constant.ZK_REGISTRY_PATH, new byte[0], ZooDefs.Ids.OPEN_ACL_UNSAFE, CreateMode.PERSISTENT);
            }
        } catch (KeeperException e) {
            LOGGER.error(e.toString());
        } catch (InterruptedException e) {
            LOGGER.error(e.toString());
        }
    }

    private void createNode(ZooKeeper zk, String data) {
        try {
            byte[] bytes = data.getBytes();
            String path = zk.create(Constant.ZK_DATA_PATH, bytes, ZooDefs.Ids.OPEN_ACL_UNSAFE, CreateMode.EPHEMERAL_SEQUENTIAL);
            LOGGER.debug("create zookeeper node ({} => {})", path, data);
        } catch (KeeperException e) {
            LOGGER.error("", e);
        }
        catch (InterruptedException ex){
            LOGGER.error("", ex);
        }
    }
}
ServiceRegistry

这里在原文的基础上加了AddRootNode()判断服务父节点是否存在,如果不存在则添加一个PERSISTENT的服务父节点,这样虽然启动服务时多了点判断,但是不需要手动命令添加服务父节点了。

关于Zookeeper的使用原理,可以看这里《ZooKeeper基本原理》。

4、客户端调用服务

使用代理模式调用服务:

public class RpcProxy {

    private String serverAddress;
    private ServiceDiscovery serviceDiscovery;

    public RpcProxy(String serverAddress) {
        this.serverAddress = serverAddress;
    }

    public RpcProxy(ServiceDiscovery serviceDiscovery) {
        this.serviceDiscovery = serviceDiscovery;
    }

    @SuppressWarnings("unchecked")
    public <T> T create(Class<?> interfaceClass) {
        return (T) Proxy.newProxyInstance(
                interfaceClass.getClassLoader(),
                new Class<?>[]{interfaceClass},
                new InvocationHandler() {
                    @Override
                    public Object invoke(Object proxy, Method method, Object[] args) throws Throwable {
                        RpcRequest request = new RpcRequest();
                        request.setRequestId(UUID.randomUUID().toString());
                        request.setClassName(method.getDeclaringClass().getName());
                        request.setMethodName(method.getName());
                        request.setParameterTypes(method.getParameterTypes());
                        request.setParameters(args);

                        if (serviceDiscovery != null) {
                            serverAddress = serviceDiscovery.discover();
                        }
                        if(serverAddress != null){
                            String[] array = serverAddress.split(":");
                            String host = array[0];
                            int port = Integer.parseInt(array[1]);

                            RpcClient client = new RpcClient(host, port);
                            RpcResponse response = client.send(request);

                            if (response.isError()) {
                                throw new RuntimeException("Response error.",new Throwable(response.getError()));
                            } else {
                                return response.getResult();
                            }
                        }
                        else{
                            throw new RuntimeException("No server address found!");
                        }
                    }
                }
        );
    }
}

这里每次使用代理远程调用服务,从Zookeeper上获取可用的服务地址,通过RpcClient send一个Request,等待该Request的Response返回。这里原文有个比较严重的bug,在原文给出的简单的Test中是很难测出来的,原文使用了obj的wait和notifyAll来等待Response返回,会出现“假死等待”的情况:一个Request发送出去后,在obj.wait()调用之前可能Response就返回了,这时候在channelRead0里已经拿到了Response并且obj.notifyAll()已经在obj.wait()之前调用了,这时候send后再obj.wait()就出现了假死等待,客户端就一直等待在这里。使用CountDownLatch可以解决这个问题。

注意:这里每次调用的send时候才去和服务端建立连接,使用的是短连接,这种短连接在高并发时会有连接数问题,也会影响性能。

从Zookeeper上获取服务地址:

public class ServiceDiscovery {

    private static final Logger LOGGER = LoggerFactory.getLogger(ServiceDiscovery.class);

    private CountDownLatch latch = new CountDownLatch(1);

    private volatile List<String> dataList = new ArrayList<>();

    private String registryAddress;

    public ServiceDiscovery(String registryAddress) {
        this.registryAddress = registryAddress;
        ZooKeeper zk = connectServer();
        if (zk != null) {
            watchNode(zk);
        }
    }

    public String discover() {
        String data = null;
        int size = dataList.size();
        if (size > 0) {
            if (size == 1) {
                data = dataList.get(0);
                LOGGER.debug("using only data: {}", data);
            } else {
                data = dataList.get(ThreadLocalRandom.current().nextInt(size));
                LOGGER.debug("using random data: {}", data);
            }
        }
        return data;
    }

    private ZooKeeper connectServer() {
        ZooKeeper zk = null;
        try {
            zk = new ZooKeeper(registryAddress, Constant.ZK_SESSION_TIMEOUT, new Watcher() {
                @Override
                public void process(WatchedEvent event) {
                    if (event.getState() == Event.KeeperState.SyncConnected) {
                        latch.countDown();
                    }
                }
            });
            latch.await();
        } catch (IOException | InterruptedException e) {
            LOGGER.error("", e);
        }
        return zk;
    }

    private void watchNode(final ZooKeeper zk) {
        try {
            List<String> nodeList = zk.getChildren(Constant.ZK_REGISTRY_PATH, new Watcher() {
                @Override
                public void process(WatchedEvent event) {
                    if (event.getType() == Event.EventType.NodeChildrenChanged) {
                        watchNode(zk);
                    }
                }
            });
            List<String> dataList = new ArrayList<>();
            for (String node : nodeList) {
                byte[] bytes = zk.getData(Constant.ZK_REGISTRY_PATH + "/" + node, false, null);
                dataList.add(new String(bytes));
            }
            LOGGER.debug("node data: {}", dataList);
            this.dataList = dataList;
        } catch (KeeperException | InterruptedException e) {
            LOGGER.error("", e);
        }
    }
}
ServiceDiscovery

每次服务地址节点发生变化,都需要再次watchNode,获取新的服务地址列表。

5、消息编码

请求消息:

public class RpcRequest {

    private String requestId;
    private String className;
    private String methodName;
    private Class<?>[] parameterTypes;
    private Object[] parameters;

    public String getRequestId() {
        return requestId;
    }

    public void setRequestId(String requestId) {
        this.requestId = requestId;
    }

    public String getClassName() {
        return className;
    }

    public void setClassName(String className) {
        this.className = className;
    }

    public String getMethodName() {
        return methodName;
    }

    public void setMethodName(String methodName) {
        this.methodName = methodName;
    }

    public Class<?>[] getParameterTypes() {
        return parameterTypes;
    }

    public void setParameterTypes(Class<?>[] parameterTypes) {
        this.parameterTypes = parameterTypes;
    }

    public Object[] getParameters() {
        return parameters;
    }

    public void setParameters(Object[] parameters) {
        this.parameters = parameters;
    }
}
RpcRequest

响应消息:

public class RpcResponse {

    private String requestId;
    private String error;
    private Object result;

    public boolean isError() {
        return error != null;
    }

    public String getRequestId() {
        return requestId;
    }

    public void setRequestId(String requestId) {
        this.requestId = requestId;
    }

    public String getError() {
        return error;
    }

    public void setError(String error) {
        this.error = error;
    }

    public Object getResult() {
        return result;
    }

    public void setResult(Object result) {
        this.result = result;
    }
}
RpcResponse

消息序列化和反序列化工具:(基于 Protostuff 实现)

public class SerializationUtil {

    private static Map<Class<?>, Schema<?>> cachedSchema = new ConcurrentHashMap<>();

    private static Objenesis objenesis = new ObjenesisStd(true);

    private SerializationUtil() {
    }

    @SuppressWarnings("unchecked")
    private static <T> Schema<T> getSchema(Class<T> cls) {
        Schema<T> schema = (Schema<T>) cachedSchema.get(cls);
        if (schema == null) {
            schema = RuntimeSchema.createFrom(cls);
            if (schema != null) {
                cachedSchema.put(cls, schema);
            }
        }
        return schema;
    }

    /**
     * 序列化(对象 -> 字节数组)
     */
    @SuppressWarnings("unchecked")
    public static <T> byte[] serialize(T obj) {
        Class<T> cls = (Class<T>) obj.getClass();
        LinkedBuffer buffer = LinkedBuffer.allocate(LinkedBuffer.DEFAULT_BUFFER_SIZE);
        try {
            Schema<T> schema = getSchema(cls);
            return ProtostuffIOUtil.toByteArray(obj, schema, buffer);
        } catch (Exception e) {
            throw new IllegalStateException(e.getMessage(), e);
        } finally {
            buffer.clear();
        }
    }

    /**
     * 反序列化(字节数组 -> 对象)
     */
    public static <T> T deserialize(byte[] data, Class<T> cls) {
        try {
            T message = (T) objenesis.newInstance(cls);
            Schema<T> schema = getSchema(cls);
            ProtostuffIOUtil.mergeFrom(data, message, schema);
            return message;
        } catch (Exception e) {
            throw new IllegalStateException(e.getMessage(), e);
        }
    }
}
SerializationUtil

由于处理的是TCP消息,本人加了TCP的粘包处理Handler

channel.pipeline().addLast(new LengthFieldBasedFrameDecoder(65536,0,4,0,0))

消息编解码时开始4个字节表示消息的长度,也就是消息编码的时候,先写消息的长度,再写消息。

6、性能改进

Netty本身就是一个高性能的网络框架,从网络IO方面来说并没有太大的问题。

从这个RPC框架本身来说,在原文的基础上把Server端处理请求的过程改成了多线程异步:

 public void channelRead0(final ChannelHandlerContext ctx,final RpcRequest request) throws Exception {
        RpcServer.submit(new Runnable() {
            @Override
            public void run() {
                LOGGER.debug("Receive request " + request.getRequestId());
                RpcResponse response = new RpcResponse();
                response.setRequestId(request.getRequestId());
                try {
                    Object result = handle(request);
                    response.setResult(result);
                } catch (Throwable t) {
                    response.setError(t.toString());
                    LOGGER.error("RPC Server handle request error",t);
                }
                ctx.writeAndFlush(response).addListener(ChannelFutureListener.CLOSE).addListener(new ChannelFutureListener() {
                    @Override
                    public void operationComplete(ChannelFuture channelFuture) throws Exception {
                        LOGGER.debug("Send response for request " + request.getRequestId());
                    }
                });
            }
        });
    }

Netty 4中的Handler处理在IO线程中,如果Handler处理中有耗时的操作(如数据库相关),会让IO线程等待,影响性能。

个人觉得该RPC的待改进项:

1)客户端保持和服务进行长连接,不需要每次调用服务的时候进行连接,长连接的管理(通过Zookeeper获取有效的地址)。

2)客户端请求异步处理的支持,不需要同步等待:发送一个异步请求,返回Feature,通过Feature的callback机制获取结果。

3)编码序列化的多协议支持。

有时间再改改吧。。

 

项目地址:https://github.com/luxiaoxun/NettyRpc

 

参考:

轻量级分布式 RPC 框架:http://my.oschina.net/huangyong/blog/361751

你应该知道的RPC原理:http://www.cnblogs.com/LBSer/p/4853234.html

 

你可能感兴趣的:(一个轻量级分布式RPC框架--NettyRpc)