基于Netty+Zookeeper+Quartz调度分析

前言
前几篇文章分别从使用和源码层面对Quartz做了简单的分析,在分析的过程中也发现了Quartz不足的地方;比如底层调度依赖数据库的悲观锁,谁先抢到谁调度,这样会导致节点负载不均衡;还有调度和执行耦合在一起,导致调度器会受到业务的影响;下面看看如何来解决这几个问题;

思路
调度器和执行器拆成不同的进程,调度器还是依赖Quartz本身的调度方式,但是调度的并不是具体业务的QuartzJobBean,而是统一的一个RemoteQuartzJobBean,在此Bean中通过Netty远程调用执行器去执行具体业务Bean;具体的执行器在启动时注册到Zookeeper中,调度器可以在Zookeeper获取执行器信息,并通过相关的负载算法指定具体的执行器去执行,以下看简单的实现;

执行器
1.执行器配置文件

executor_name=firstExecutor
service_address=127.0.0.1:8000
registry_address=127.0.0.1:2181

配置了执行器的名称,执行器启动的ip和端口以及Zookeeper的地址信息;

2.执行器服务

<bean id="executorServer" class="com.zh.job.executor.ExecutorServer">
		<constructor-arg name="executorName" value="${executor_name}"/>
		<constructor-arg name="serviceAddress" value="${service_address}" />
		<constructor-arg name="serviceRegistry" ref="serviceRegistry" />
	</bean>

ExecutorServer通过Netty启动服务,并向Zookeeper注册服务,部分代码如下:

        EventLoopGroup bossGroup = new NioEventLoopGroup();
		EventLoopGroup workerGroup = new NioEventLoopGroup();
		try {
			// 创建并初始化 Netty 服务端 Bootstrap 对象
			ServerBootstrap bootstrap = new ServerBootstrap();
			bootstrap.group(bossGroup, workerGroup);
			bootstrap.channel(NioServerSocketChannel.class);
			bootstrap.childHandler(new ChannelInitializer<SocketChannel>() {
				@Override
				public void initChannel(SocketChannel channel) throws Exception {
					ChannelPipeline pipeline = channel.pipeline();
					pipeline.addLast(new RpcDecoder(Request.class));
					pipeline.addLast(new RpcEncoder(Response.class));
					pipeline.addLast(new ExecutorServerHandler(handlerMap));
				}
			});
			bootstrap.option(ChannelOption.SO_BACKLOG, 1024);
			bootstrap.childOption(ChannelOption.SO_KEEPALIVE, true);
			// 获取 RPC 服务器的 IP 地址与端口号
			String[] addressArray = StringUtils.splitByWholeSeparator(serviceAddress, ":");
			String ip = addressArray[0];
			int port = Integer.parseInt(addressArray[1]);
			// 启动 RPC 服务器
			ChannelFuture future = bootstrap.bind(ip, port).sync();
			// 注册 RPC 服务地址
			if (serviceRegistry != null) {
				serviceRegistry.register(executorName, serviceAddress);
				LOGGER.info("register service: {} => {}", executorName, serviceAddress);
			}
			LOGGER.info("server started on port {}", port);
			// 关闭 RPC 服务器
			future.channel().closeFuture().sync();
		} finally {
			workerGroup.shutdownGracefully();
			bossGroup.shutdownGracefully();
		}

在Netty中指定了编码器解码器,同时指定了ExecutorServerHandler用来处理调度器发送来的消息(更多代码查看项目源码);最后向Zookeeper注册服务,路径格式如下:

/job_registry/firstExecutor/address-0000000008

job_registry是固定值,firstExecutor是配置的具体执行器名称;

3.配置加载任务
添加注解类,用来指定具体的业务Job:

@Target({ ElementType.TYPE })
@Retention(RetentionPolicy.RUNTIME)
@Component
public @interface ExecutorTask {

	String name();

}

例如具体的业务Task如下所示:

@ExecutorTask(name = "firstTask")
public class FirstTask implements IJobHandler {

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

	@Override
	public Result execute(String param) throws Exception {
		LOGGER.info("execute firstTask");
		return SUCCESS;
	}

}

在启动执行器服务时,加载有ExecutorTask注解的任务类,此处定义的name要和调度端的名称相互匹配;

4.执行具体业务
Netty中指定了ExecutorServerHandler用来处理接受的调度器信息,通过反射的方式来调用具体的业务Job,部分代码如下:

   private Object handle(Request request) throws Exception {
		// 获取服务对象
		String serviceName = request.getInterfaceName();
		Object serviceBean = handlerMap.get(serviceName);
		if (serviceBean == null) {
			throw new RuntimeException(String.format("can not find service bean by key: %s", serviceName));
		}
		// 获取反射调用所需的参数
		Class<?> serviceClass = serviceBean.getClass();
		String methodName = request.getMethodName();
		Class<?>[] parameterTypes = request.getParameterTypes();
		Object[] parameters = request.getParameters();
		// 使用 CGLib 执行反射调用
		FastClass serviceFastClass = FastClass.create(serviceClass);
		FastMethod serviceFastMethod = serviceFastClass.getMethod(methodName, parameterTypes);
		return serviceFastMethod.invoke(serviceBean, parameters);
	}

serviceName对应的就是定义的”firstTask”,然后通过serviceName找到对应的Bean,然后反射调用,最终返回结果;

调度器
调度器还是依赖Quartz的原生调度方式,只不过调度器不在执行相关业务Task,所以相关配置也是类似,同样依赖数据库;
1.定义调度任务

   <bean id="firstTask"
		class="org.springframework.scheduling.quartz.JobDetailFactoryBean">
		<property name="jobClass" value="com.zh.job.scheduler.RemoteQuartzJobBean" />
		<property name="jobDataMap">
			<map>
				<entry key="executorBean" value-ref="firstExecutor" />
			</map>
		</property>
	</bean>

	<bean id="firstExecutor" class="com.zh.job.scheduler.ExecutorBean">
		<constructor-arg name="executorName" value="firstExecutor"></constructor-arg>
		<constructor-arg name="discoveryAddress" value="${discovery_address}"></constructor-arg>
	</bean>

同样在调度端定义了名称问firstTask的任务,可以发现此类是RemoteQuartzJobBean,并不是具体的业务Task;同时也指定了jobDataMap,用来指定执行器名称和发现的Zookeeper地址;

2.RemoteQuartzJobBean

public class RemoteQuartzJobBean extends QuartzJobBean {

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

	private ExecutorBean executorBean;

	@Override
	protected void executeInternal(JobExecutionContext context) throws JobExecutionException {
		JobKey jobKey = context.getTrigger().getJobKey();
		LOGGER.info("jobName:" + jobKey.getName() + ",group:" + jobKey.getGroup());
		IJobHandler executor = JobProxy.create(IJobHandler.class, jobKey, this.executorBean);
		Result result;
		try {
			result = executor.execute("");
			LOGGER.info("result:" + result);
		} catch (Exception e) {
			LOGGER.error("", e);
		}
	}

	public ExecutorBean getExecutorBean() {
		return executorBean;
	}

	public void setExecutorBean(ExecutorBean executorBean) {
		this.executorBean = executorBean;
	}

}

此类同样继承于QuartzJobBean,这样Quartz才能调度Bean,在此Bean中通过jobKey和executorBean创建了IJobHandler的代理类,具体代码如下:

public static <T> T create(final Class<?> interfaceClass, final JobKey jobKey, final ExecutorBean executor) {
		// 创建动态代理对象
		return (T) Proxy.newProxyInstance(interfaceClass.getClassLoader(), new Class<?>[] { interfaceClass },
				new InvocationHandler() {
					@Override
					public Object invoke(Object proxy, Method method, Object[] args) throws Throwable {
						// 创建 RPC 请求对象并设置请求属性
						Request request = new Request();
						request.setRequestId(UUID.randomUUID().toString());
						request.setInterfaceName(jobKey.getName());
						request.setMethodName(method.getName());
						request.setParameterTypes(method.getParameterTypes());
						request.setParameters(args);

						String serviceAddress = null;
						ServiceDiscovery serviceDiscovery = ServiceDiscoveryFactory
								.getServiceDiscovery(executor.getDiscoveryAddress());
						// 获取 RPC 服务地址
						if (serviceDiscovery != null) {
							serviceAddress = serviceDiscovery.discover(executor.getExecutorName());
							LOGGER.debug("discover service: {} => {}", executor.getExecutorName(), serviceAddress);
						}
						if (StringUtil.isEmpty(serviceAddress)) {
							throw new RuntimeException("server address is empty");
						}
						// 从 RPC 服务地址中解析主机名与端口号
						String[] array = StringUtil.split(serviceAddress, ":");
						String host = array[0];
						int port = Integer.parseInt(array[1]);
						// 创建 RPC 客户端对象并发送 RPC 请求
						ExecutorClient client = new ExecutorClient(host, port);
						long time = System.currentTimeMillis();
						Response response = client.send(request);
						LOGGER.debug("time: {}ms", System.currentTimeMillis() - time);
						if (response == null) {
							throw new RuntimeException("response is null");
						}
						// 返回 RPC 响应结果
						if (response.hasException()) {
							throw response.getException();
						} else {
							return response.getResult();
						}
					}
				});
	}

在Request中指定了InterfaceName为jobKey.getName(),也就是这里的firstTask;通过Zookeeper发现服务时指定了executor.getExecutorName(),这样可以在Zookeeper中找到具体的执行器地址,当然这里的地址可能是一个列表,可以通过负载均衡算法(随机,轮询,一致性hash等等)进行分配,获取到地址后通过Netty远程连接执行器,发送执行job等待返回结果;

简单测试
分别执行调度器和执行器,相关日志如下:
1.执行器日志

2018-09-03 11:17:02 [main] 13::: DEBUG com.zh.job.sample.executor.ExecutorBootstrap - start server
2018-09-03 11:17:03 [main] 31::: DEBUG com.zh.job.registry.impl.ZookeeperServiceRegistry - connect zookeeper
2018-09-03 11:17:03 [main] 49::: DEBUG com.zh.job.registry.impl.ZookeeperServiceRegistry - create address node: /job_registry/firstExecutor/address-0000000009
2018-09-03 11:17:03 [main] 107::: INFO  com.zh.job.executor.ExecutorServer - register service: firstExecutor => 127.0.0.1:8000
2018-09-03 11:17:03 [main] 109::: INFO  com.zh.job.executor.ExecutorServer - server started on port 8000
2018-09-03 11:17:15 [nioEventLoopGroup-3-1] 17::: INFO  com.zh.job.sample.executor.task.FirstTask - execute firstTask

2.调度器日志

2018-09-03 11:17:14 [myScheduler_Worker-1] 28::: INFO  com.zh.job.scheduler.RemoteQuartzJobBean - jobName:firstTask,group:DEFAULT
2018-09-03 11:17:15 [myScheduler_Worker-2] 28::: INFO  com.zh.job.scheduler.RemoteQuartzJobBean - jobName:firstTask,group:DEFAULT
2018-09-03 11:17:15 [myScheduler_Worker-1] 33::: DEBUG com.zh.job.registry.impl.ZookeeperServiceDiscovery - connect zookeeper
2018-09-03 11:17:15 [myScheduler_Worker-2] 54::: DEBUG com.zh.job.registry.impl.ZookeeperServiceDiscovery - get only address node: address-0000000009
2018-09-03 11:17:15 [myScheduler_Worker-1] 54::: DEBUG com.zh.job.registry.impl.ZookeeperServiceDiscovery - get only address node: address-0000000009
2018-09-03 11:17:15 [myScheduler_Worker-2] 42::: DEBUG com.zh.job.scheduler.JobProxy$1 - discover service: firstExecutor => 127.0.0.1:8000
2018-09-03 11:17:15 [myScheduler_Worker-1] 42::: DEBUG com.zh.job.scheduler.JobProxy$1 - discover service: firstExecutor => 127.0.0.1:8000
2018-09-03 11:17:15 [myScheduler_Worker-1] 55::: DEBUG com.zh.job.scheduler.JobProxy$1 - time: 369ms
2018-09-03 11:17:15 [myScheduler_Worker-1] 33::: INFO  com.zh.job.scheduler.RemoteQuartzJobBean - result:com.zh.job.common.bean.Result@33b61489

总结
本文通过一个实例来分析如何解决原生Quartz调度存在不足的问题,主要体现在调度器与执行器的隔离上,各司其责发挥各自的优势;

示例代码地址
https://github.com/ksfzhaohui/job