一、背景html
最近在搜索Netty和Zookeeper方面的文章時,看到了這篇文章《輕量級分佈式 RPC 框架》,做者用Zookeeper、Netty和Spring寫了一個輕量級的分佈式RPC框架。花了一些時間看了下他的代碼,寫的乾淨簡單,寫的RPC框架能夠算是一個簡易版的dubbo。這個RPC框架雖小,可是麻雀雖小,五臟俱全,有興趣的能夠學習一下。node
本人在這個簡易版的RPC上添加了以下特性:git
* 服務異步調用的支持,回調函數callback的支持github
* 客戶端使用長鏈接(在屢次調用共享鏈接)數據庫
* 服務端異步多線程處理RPC請求數組
項目地址:https://github.com/luxiaoxun/NettyRpc網絡
二、簡介多線程
RPC,即 Remote Procedure Call(遠程過程調用),調用遠程計算機上的服務,就像調用本地服務同樣。RPC能夠很好的解耦系統,如WebService就是一種基於Http協議的RPC。併發
這個RPC總體框架以下:框架
這個RPC框架使用的一些技術所解決的問題:
服務發佈與訂閱:服務端使用Zookeeper註冊服務地址,客戶端從Zookeeper獲取可用的服務地址。
通訊:使用Netty做爲通訊框架。
Spring:使用Spring配置服務,加載Bean,掃描註解。
動態代理:客戶端使用代理模式透明化服務調用。
消息編解碼:使用Protostuff序列化和反序列化消息。
三、服務端發佈服務
使用註解標註要發佈的服務
服務註解
@Target({ElementType.TYPE}) @Retention(RetentionPolicy.RUNTIME) @Component public @interface RpcService { Class<?> value(); }
一個服務接口:
public interface HelloService { String hello(String name); String hello(Person person); }
一個服務實現:使用註解標註
@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(); } }
服務在啓動的時候掃描獲得全部的服務接口及其實現:
@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集羣上註冊服務地址:
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); } } }
這裏在原文的基礎上加了AddRootNode()判斷服務父節點是否存在,若是不存在則添加一個PERSISTENT的服務父節點,這樣雖然啓動服務時多了點判斷,可是不須要手動命令添加服務父節點了。
關於Zookeeper的使用原理,能夠看這裏《ZooKeeper基本原理》。
四、客戶端調用服務
使用代理模式調用服務:
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!"); } } } ); } }
這裏每次使用代理遠程調用服務,從Zookeeper上獲取可用的服務地址,經過RpcClient send一個Request,等待該Request的Response返回。這裏原文有個比較嚴重的bug,在原文給出的簡單的Test中是很難測出來的,原文使用了obj的wait和notifyAll來等待Response返回,會出現「假死等待」的狀況:一個Request發送出去後,在obj.wait()調用以前可能Response就返回了,這時候在channelRead0裏已經拿到了Response而且obj.notifyAll()已經在obj.wait()以前調用了,這時候send後再obj.wait()就出現了假死等待,客戶端就一直等待在這裏。使用CountDownLatch能夠解決這個問題。
注意:這裏每次調用的send時候纔去和服務端創建鏈接,使用的是短鏈接,這種短鏈接在高併發時會有鏈接數問題,也會影響性能。
從Zookeeper上獲取服務地址:
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); } } }
每次服務地址節點發生變化,都須要再次watchNode,獲取新的服務地址列表。
五、消息編碼
請求消息:
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; } }
響應消息:
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; } }
消息序列化和反序列化工具:(基於 Protostuff 實現)
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(對象 -> 字節數組) */ @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(字節數組 -> 對象) */ 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); } } }
因爲處理的是TCP消息,本人加了TCP的粘包處理Handler
channel.pipeline().addLast(new LengthFieldBasedFrameDecoder(65536,0,4,0,0))
消息編解碼時開始4個字節表示消息的長度,也就是消息編碼的時候,先寫消息的長度,再寫消息。
六、性能改進
1)服務端請求異步處理
Netty自己就是一個高性能的網絡框架,從網絡IO方面來講並無太大的問題。
從這個RPC框架自己來講,在原文的基礎上把Server端處理請求的過程改爲了多線程異步:
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處理在IO線程中,若是Handler處理中有耗時的操做(如數據庫相關),會讓IO線程等待,影響性能。
2)服務端長鏈接的管理
客戶端保持和服務進行長鏈接,不須要每次調用服務的時候進行鏈接,長鏈接的管理(經過Zookeeper獲取有效的地址)。
經過監聽Zookeeper服務節點值的變化,動態更新客戶端和服務端保持的長鏈接。這個事情如今放在客戶端在作,客戶端保持了和全部可用服務的長鏈接,給客戶端和服務端都形成了壓力,須要解耦這個實現。
3)客戶端請求異步處理
客戶端請求異步處理的支持,不須要同步等待:發送一個異步請求,返回Feature,經過Feature的callback機制獲取結果。
IAsyncObjectProxy client = rpcClient.createAsync(HelloService.class); RPCFuture helloFuture = client.call("hello", Integer.toString(i)); String result = (String) helloFuture.get(3000, TimeUnit.MILLISECONDS);
我的以爲該RPC的待改進項:
* 編碼序列化的多協議支持。
項目持續更新中。
項目地址:https://github.com/luxiaoxun/NettyRpc
參考:
輕量級分佈式 RPC 框架:http://my.oschina.net/huangyong/blog/361751
你應該知道的RPC原理:http://www.cnblogs.com/LBSer/p/4853234.html