在还没有认真分析源码之前，我根据以前阅读微服务框架的思路，去分析这样几个问题
- 客户端如何获取服务节点列表
- 客户端如何维持与服务之间的链接，这种链接是长连接还是短连接，使用的方式/框架是什么？
- 客户端如何实现服务端调用的负载均衡，还是说负载均衡由硬件来处理（网络层来处理）。
- 客户端与服务端数据是怎样通信的？序列化协议？
- 客户端与服务端连接出现问题如何处理？
这是我暂时想到的，当然微服务框架之间的处理，远不止这些，这些算是入门读懂一些框架，接下来我就来具体分析。
// 从本地启动的配置
CommunicatorConfig cfg = new CommunicatorConfig();
// 从本地启动的Communcator
Communicator communicator = CommunicatorFactory.getInstance().getCommunicator(cfg);
//warn 若是部署在tars平台启动的， 只能使用下面的构造器获取communcator
//Communicator communicator = CommunicatorFactory.getInstance().getCommunicator();
HelloPrx proxy = communicator.stringToProxy(HelloPrx.class, "TestApp.HelloServer.HelloObj@tcp -h 127.0.0.1 -p 18601 -t 60000");
//同步调用
String ret = proxy.hello(1000, "Hello World");
System.out.println(ret);
- CommunicatorConfig:客户端协调器的配置文件（超时配置呀，等等一些列配置信息）
- Communicator：正在的协调调用者。
- stringToProxy：将我们的Servant封装成代理类，ObjectProxy也就是我们后面获取的Servant客户端都是ObjectProxy，直接看invoke方法即可。
几个关键的类
1. CommunicatorConfig
2. Communicator
3. ServantProxyFactory
public <T> Object getServantProxy(Class<T> clazz, String objName, String setDivision, ServantProxyConfig servantProxyConfig,
LoadBalance loadBalance, ProtocolInvoker<T> protocolInvoker) {
String key = setDivision != null ? clazz.getSimpleName() + objName + setDivision : clazz.getSimpleName() + objName;
Object proxy = cache.get(key);
if (proxy == null) {
lock.lock();
try {
proxy = cache.computeIfAbsent(key, param -> {
ObjectProxy<T> objectProxy = communicator.getObjectProxyFactory().getObjectProxy(clazz, objName, setDivision, servantProxyConfig, loadBalance, protocolInvoker);
return createProxy(clazz, objectProxy);
});
} finally {
lock.unlock();
}
}
return proxy;
}
管理ObjectProxyFactory，如果缓存有就从缓存中取，没有就生产
4.ObjectProxyFactory
public <T> ObjectProxy<T> getObjectProxy(Class<T> api, String objName, String setDivision, ServantProxyConfig servantProxyConfig,
LoadBalance<T> loadBalance, ProtocolInvoker<T> protocolInvoker) throws ClientException {
if (servantProxyConfig == null) {
servantProxyConfig = createServantProxyConfig(objName, setDivision);
} else {
servantProxyConfig.setCommunicatorId(communicator.getId());
servantProxyConfig.setModuleName(communicator.getCommunicatorConfig().getModuleName(), communicator.getCommunicatorConfig().isEnableSet(), communicator.getCommunicatorConfig().getSetDivision());
servantProxyConfig.setLocator(communicator.getCommunicatorConfig().getLocator());
if (StringUtils.isNotEmpty(setDivision)) {
servantProxyConfig.setSetDivision(setDivision);
}
}
updateServantEndpoints(servantProxyConfig);
if (loadBalance == null) {
loadBalance = createLoadBalance(servantProxyConfig);
}
if (protocolInvoker == null) {
protocolInvoker = createProtocolInvoker(api, servantProxyConfig);
}
return new ObjectProxy<T>(api, servantProxyConfig, loadBalance, protocolInvoker, communicator);
}
生产ObjectProxy
- 创建ServantProxyConfig：就是Servant服务的配置信息，每个服务配置可能不同
- updateServantEndpoints
private void updateServantEndpoints(ServantProxyConfig cfg) {
CommunicatorConfig communicatorConfig = communicator.getCommunicatorConfig();
String endpoints = null;
if (!ParseTools.hasServerNode(cfg.getObjectName()) && !cfg.isDirectConnection() && !communicatorConfig.getLocator().startsWith(cfg.getSimpleObjectName())) {
try {
/** query server nodes from registerServer */
if (RegisterManager.getInstance().getHandler() != null) {
endpoints = ParseTools.parse(RegisterManager.getInstance().getHandler().query(cfg.getSimpleObjectName()),
cfg.getSimpleObjectName());
} else {
endpoints = communicator.getQueryHelper().getServerNodes(cfg);
}
if (StringUtils.isEmpty(endpoints)) {
throw new CommunicatorConfigException(cfg.getSimpleObjectName(), "servant node is empty on get by registry! communicator id=" + communicator.getId());
}
ServantCacheManager.getInstance().save(communicator.getId(), cfg.getSimpleObjectName(), endpoints, communicatorConfig.getDataPath());
} catch (CommunicatorConfigException e) {
/** If it fails, pull it from the local cache file */
endpoints = ServantCacheManager.getInstance().get(communicator.getId(), cfg.getSimpleObjectName(), communicatorConfig.getDataPath());
logger.error(cfg.getSimpleObjectName() + " error occurred on get by registry, use by local cache=" + endpoints + "|" + e.getLocalizedMessage(), e);
}
if (StringUtils.isEmpty(endpoints)) {
throw new CommunicatorConfigException(cfg.getSimpleObjectName(), "error occurred on create proxy, servant endpoint is empty! locator =" + communicatorConfig.getLocator() + "|communicator id=" + communicator.getId());
}
cfg.setObjectName(endpoints);
}
if (StringUtils.isEmpty(cfg.getObjectName())) {
throw new CommunicatorConfigException(cfg.getSimpleObjectName(), "error occurred on create proxy, servant endpoint is empty!");
}
}
通过ObjectName判断是有设置了服务器节点，如果有（本地只连接），如果没有那就从tars管理中获取服务器节点。放在ServantCacheManager管理起来。
5.LoadBalance
客户端负载均衡选择器
- ConsistentHashLoadBalance：一致hash选择器
- HashLoadBalance：hash选择器
- RoundRobinLoadBalance: 轮询选择器
- DefaultLoadBalance：默认的选择器，先ConsistentHashLoadBalance，HashLoadBalance，RoundRobinLoadBalance
看源码可知提供了3中负载均衡的选择器。默认的是DefaultLoadBalance（这里面设置的是RoundRobinLoadBalance）
我们来分析RoundRobinLoadBalance
public Invoker<T> select(InvokeContext invocation) throws NoInvokerException {
List<Invoker<T>> staticWeightInvokers = staticWeightInvokersCache;
//polling use weights
if (staticWeightInvokers != null && !staticWeightInvokers.isEmpty()) {
Invoker<T> invoker = staticWeightInvokers.get((staticWeightSequence.getAndIncrement() & Integer.MAX_VALUE) % staticWeightInvokers.size());
if (invoker.isAvailable()) return invoker;
ServantInvokerAliveStat stat = ServantInvokerAliveChecker.get(invoker.getUrl());
if (stat.isAlive() || (stat.getLastRetryTime() + (config.getTryTimeInterval() * 1000)) < System.currentTimeMillis()) {
//Try to recall after blocking
logger.info("try to use inactive invoker|" + invoker.getUrl().toIdentityString());
stat.setLastRetryTime(System.currentTimeMillis());
return invoker;
}
}
//polling without weights
List<Invoker<T>> sortedInvokers = sortedInvokersCache;
if (CollectionUtils.isEmpty(sortedInvokers)) {
throw new NoInvokerException("no such active connection invoker");
}
List<Invoker<T>> list = new ArrayList<Invoker<T>>();
for (Invoker<T> invoker : sortedInvokers) {
if (!invoker.isAvailable()) {
/**
* Try to recall after blocking
*/
ServantInvokerAliveStat stat = ServantInvokerAliveChecker.get(invoker.getUrl());
if (stat.isAlive() || (stat.getLastRetryTime() + (config.getTryTimeInterval() * 1000)) < System.currentTimeMillis()) {
list.add(invoker);
}
} else {
list.add(invoker);
}
}
//TODO When all is not available. Whether to randomly extract one
if (list.isEmpty()) {
throw new NoInvokerException(config.getSimpleObjectName() + " try to select active invoker, size=" + sortedInvokers.size() + ", no such active connection invoker");
}
Invoker<T> invoker = list.get((sequence.getAndIncrement() & Integer.MAX_VALUE) % list.size());
if (!invoker.isAvailable()) {
//Try to recall after blocking
logger.info("try to use inactive invoker|" + invoker.getUrl().toIdentityString());
ServantInvokerAliveChecker.get(invoker.getUrl()).setLastRetryTime(System.currentTimeMillis());
}
return invoker;
}
主要就是从Invoke列表中轮询挑选一个可用的Invoke
6.ProtocolInvoker
public interface ProtocolInvoker<T> {
InvokeContext createContext(Object proxy, Method method, Object[] args) throws Exception;
Invoker<T> create(Class<T> api, Url url) throws Exception;
Collection<Invoker<T>> getInvokers();
void refresh();
void destroy();
}
获取invoke上下文信息,创建invoke,获取invokes
@Override
public Invoker<T> create(Class<T> api, Url url) throws Exception {
return new TarsInvoker<T>(servantProxyConfig, api, url, getClients(url));
}
protected ServantClient[] getClients(Url url) throws IOException {
int connections = url.getParameter(Constants.TARS_CLIENT_CONNECTIONS, Constants.default_connections);
ServantClient[] clients = new ServantClient[connections];
for (int i = 0; i < clients.length; i++) {
clients[i] = initClient(url);
}
return clients;
}
protected ServantClient initClient(Url url) {
ServantClient client = null;
try {
boolean tcpNoDelay = url.getParameter(Constants.TARS_CLIENT_TCPNODELAY, false);
long connectTimeout = url.getParameter(Constants.TARS_CLIENT_CONNECTTIMEOUT, Constants.default_connect_timeout);
long syncTimeout = url.getParameter(Constants.TARS_CLIENT_SYNCTIMEOUT, Constants.default_sync_timeout);
long asyncTimeout = url.getParameter(Constants.TARS_CLIENT_ASYNCTIMEOUT, Constants.default_async_timeout);
boolean udpMode = url.getParameter(Constants.TARS_CLIENT_UDPMODE, false);
if (this.selectorManager == null) {
this.selectorManager = ClientPoolManager.getSelectorManager(this.protocolFactory, this.threadPoolExecutor, true, udpMode, this.servantProxyConfig);
}
client = new ServantClient(url.getHost(), url.getPort(), this.selectorManager, udpMode);
client.setConnectTimeout(connectTimeout);
client.setSyncTimeout(syncTimeout);
client.setAsyncTimeout(asyncTimeout);
client.setTcpNoDelay(tcpNoDelay);
} catch (Throwable e) {
throw new ClientException(servantProxyConfig.getSimpleObjectName(), "Fail to create client|" + url.toIdentityString() + "|" + e.getLocalizedMessage(), e);
}
return client;
}
创建一个invoke（连接服务端的处理器），这个url就是我们本地设置的或者是从注册中心获取到的。
refresh刷新现在的invokes
到此我们的核心ObjectProxy就创建成功了。
ObjectProxy
public Object invoke(Object proxy, Method method, Object[] args) throws Throwable {
String methodName = method.getName();
Class<?>[] parameterTypes = method.getParameterTypes();
InvokeContext context = protocolInvoker.createContext(proxy, method, args);
try {
if ("toString".equals(methodName) && parameterTypes.length == 0) {
return this.toString();
} else if ("hashCode".equals(methodName) && parameterTypes.length == 0) {
return this.hashCode();
} else if ("equals".equals(methodName) && parameterTypes.length == 1) {
return this.equals(args[0]);
} else if ("getObjectName".equals(methodName) && parameterTypes.length == 0) {
return this.getObjectName();
} else if ("getApi".equals(methodName) && parameterTypes.length == 0) {
return this.getApi();
} else if ("getConfig".equals(methodName) && parameterTypes.length == 0) {
return this.getConfig();
} else if ("destroy".equals(methodName) && parameterTypes.length == 0) {
this.destroy();
return null;
} else if ("refresh".equals(methodName) && parameterTypes.length == 0) {
this.refresh();
return null;
}
Invoker invoker = loadBalancer.select(context);
return invoker.invoke(context);
} catch (Throwable e) {
e.printStackTrace();
if (logger.isDebugEnabled()) {
logger.debug(servantProxyConfig.getSimpleObjectName() + " error occurred on invoke|" + e.getLocalizedMessage(), e);
}
if (e instanceof NoInvokerException) {
throw new NoConnectionException(servantProxyConfig.getSimpleObjectName(), e.getLocalizedMessage(), e);
}
throw new ClientException(servantProxyConfig.getSimpleObjectName(), e.getLocalizedMessage(), e);
}
}
在执行客户端**Prx 时就是调用了ObjectProxy.invoke方法
- 获取接口的方法名称，方法的函数签名
- 获取InvokeContext
- loadBalancer选择一个可用的invoke(服务端的链接)
- TarsInvoke.invoke
protected Object doInvokeServant(final ServantInvokeContext inv) throws Throwable {
final long begin = System.currentTimeMillis();
int ret = Constants.INVOKE_STATUS_SUCC;
try {
Method method = getApi().getMethod(inv.getMethodName(), inv.getParameterTypes());
if (inv.isAsync()) {
invokeWithAsync(method, inv.getArguments(), inv.getAttachments());
return null;
} else if (inv.isPromiseFuture()) {
return invokeWithPromiseFuture(method, inv.getArguments(), inv.getAttachments());// return Future Result
} else {
TarsServantResponse response = invokeWithSync(method, inv.getArguments(), inv.getAttachments());
ret = response.getRet() == TarsHelper.SERVERSUCCESS ? Constants.INVOKE_STATUS_SUCC : Constants.INVOKE_STATUS_EXEC;
if (response.getRet() != TarsHelper.SERVERSUCCESS) {
throw ServerException.makeException(response.getRet());
}
return response.getResult();
}
} catch (Throwable e) {
if (e instanceof TimeoutException) {
ret = Constants.INVOKE_STATUS_TIMEOUT;
} else if (e instanceof NotConnectedException) {
ret = Constants.INVOKE_STATUS_NETCONNECTTIMEOUT;
} else {
ret = Constants.INVOKE_STATUS_EXEC;
}
throw e;
} finally {
if (inv.isNormal()) {
setAvailable(ServantInvokerAliveChecker.isAlive(getUrl(), config, ret));
InvokeStatHelper.getInstance().addProxyStat(objName)
.addInvokeTimeByClient(config.getMasterName(), config.getSlaveName(), config.getSlaveSetName(), config.getSlaveSetArea(),
config.getSlaveSetID(), inv.getMethodName(), getUrl().getHost(), getUrl().getPort(), ret, System.currentTimeMillis() - begin);
}
}
}
- 通过方法判断是哪种请求方式，同步异步？
- 以同步为例子
private TarsServantResponse invokeWithSync(Method method, Object args[], Map<String, String> context) throws Throwable {
ServantClient client = getClient();
TarsServantRequest request = new TarsServantRequest(client.getIoSession());
request.setVersion(TarsHelper.VERSION);
request.setMessageType(isHashInvoke(context) ? TarsHelper.MESSAGETYPEHASH : TarsHelper.MESSAGETYPENULL);
request.setPacketType(TarsHelper.NORMAL);
request.setServantName(objName);
request.setFunctionName(method.getName());
request.setApi(super.getApi());
request.setMethodInfo(AnalystManager.getInstance().getMethodMap(super.getApi()).get(method));
request.setMethodParameters(args);
request.setContext(context);
request.setInvokeStatus(InvokeStatus.SYNC_CALL);
TarsServantResponse response = new TarsServantResponse(request.getIoSession());
response.setRequest(request);
response.setRequestId(request.getTicketNumber());
response.setVersion(request.getVersion());
response.setPacketType(request.getPacketType());
response.setMessageType(request.getMessageType());
response.setStatus(request.getStatus());
response.setRequest(request);
response.setCharsetName(request.getCharsetName());
response.setTimeout(request.getTimeout());
response.setContext(request.getContext());
DistributedContext distributedContext = DistributedContextManager.getDistributedContext();
Boolean bDyeing = distributedContext.get(DyeingSwitch.BDYEING);
if (bDyeing != null && bDyeing == true) {
request.setMessageType(request.getMessageType() | TarsHelper.MESSAGETYPEDYED);
HashMap<String, String> status = new HashMap<String, String>();
String routeKey = distributedContext.get(DyeingSwitch.DYEINGKEY);
String fileName = distributedContext.get(DyeingSwitch.FILENAME);
status.put(DyeingSwitch.STATUS_DYED_KEY, routeKey == null ? "" : routeKey);
status.put(DyeingSwitch.STATUS_DYED_FILENAME, fileName == null ? "" : fileName);
request.setStatus(status);
}
FilterChain filterChain = new TarsClientFilterChain(filters, objName, FilterKind.CLIENT, client, InvokeStatus.SYNC_CALL, null);
filterChain.doFilter(request, response);
return response;
}
- 创建tars的请求和响应体
- 先处理过滤器，这里就和http的filter是一个道理了
- TarsClientFilterChain.doRealInvoke
@Override
protected void doRealInvoke(Request request, Response response) throws Throwable {
if (request instanceof TarsServantRequest && target != null) {
TarsServantResponse tarsServantResponse = (TarsServantResponse) response;
switch (type) {
case SYNC_CALL:
try {
TarsServantResponse result = target.invokeWithSync((ServantRequest) request);
BeanAccessor.setBeanValue(tarsServantResponse, "cause", result.getCause());
BeanAccessor.setBeanValue(tarsServantResponse, "result", result.getResult());
BeanAccessor.setBeanValue(tarsServantResponse, "ret", result.getRet());
} catch (Exception e) {
BeanAccessor.setBeanValue(tarsServantResponse, "cause", e);
throw e;
}
return;
case ASYNC_CALL:
target.invokeWithAsync((ServantRequest) request, callback);
return;
case FUTURE_CALL:
target.invokeWithFuture((ServantRequest) request, callback);
return;
}
}
}
public <T extends ServantResponse> T invokeWithSync(ServantRequest request) throws IOException {
Ticket<T> ticket = null;
T response = null;
try {
ensureConnected();
request.setInvokeStatus(InvokeStatus.SYNC_CALL);
ticket = TicketManager.createTicket(request, session, this.syncTimeout);
Session current = session;
current.write(request);
if (!ticket.await(this.syncTimeout, TimeUnit.MILLISECONDS)) {
if (current != null && current.getStatus() != SessionStatus.CLIENT_CONNECTED) {
throw new IOException("Connection reset by peer|" + this.getAddress());
} else {
throw new TimeoutException("the operation has timeout, " + this.syncTimeout + "ms|" + this.getAddress());
}
}
response = ticket.response();
if (response == null) {
throw new IOException("the operation is failed.");
}
return response;
} catch (InterruptedException e) {
logger.error(e.getLocalizedMessage());
} finally {
if (ticket != null) {
TicketManager.removeTicket(ticket.getTicketNumber());
}
}
return response;
}
通过同步的方式去调用，调用之前先判断是否有链接上。
public void ensureConnected() throws IOException {
if (isNotConnected()) {
reConnect();
}
}
protected synchronized void reConnect() throws IOException {
if (isNotConnected()) {
SocketAddress server = new InetSocketAddress(this.host, this.port);
SelectableChannel channel = null;
Session temp = null;
int event;
if (this.udpMode) {
channel = DatagramChannel.open();
channel.configureBlocking(false);
temp = new UDPSession(this.selectorManager);
((UDPSession) temp).setBufferSize(bufferSize);
((UDPSession) temp).setTarget(server);
event = SelectionKey.OP_READ;
temp.setStatus(SessionStatus.CLIENT_CONNECTED);
} else {
channel = SocketChannel.open();
channel.configureBlocking(false);
try {
if (this.tc != INVALID_TRAFFIC_CLASS_VALUE) {
((SocketChannel) channel).socket().setTrafficClass(this.tc);
}
} catch (Exception ex) {
logger.error(ex.getLocalizedMessage());
}
((SocketChannel) channel).connect(server);
temp = new TCPSession(this.selectorManager);
((TCPSession) temp).setTcpNoDelay(this.tcpNoDelay);
event = SelectionKey.OP_CONNECT;
}
temp.setChannel(channel);
temp.setKeepAlive(selectorManager.isKeepAlive());
this.selectorManager.nextReactor().registerChannel(channel, event, temp);
if (!this.udpMode) {
if (!temp.waitToConnect(this.connectTimeout)) {
temp.asyncClose();
throw new TimeoutException("connect " + this.connectTimeout + "ms timed out to " + this.getAddress());
}
if (temp.getStatus() == SessionStatus.NOT_CONNECTED) {
temp.asyncClose();
throw new NotConnectedException("connect failed to " + this.getAddress());
} else if (temp.getStatus() == SessionStatus.CLOSED) {
throw new NotConnectedException("connect failed to " + this.getAddress());
}
}
this.session = temp;
}
}
没有连接上就通过tcp(Udp)的方式连接上服务。NIo的方式，进行read，write
然后返回Response进行处理。
到此客户请求已经完毕。这里面的详细，以后再分析。
既然我们知道tars的服务启动和客户端调用，那么springboot使用注解的方式呢，只要熟读了spring源码的，那就是非常简单的了，下篇我就分析springboot怎么使用注解的方式了。
894
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
