java

关注公众号 jb51net

关闭
首页 > 软件编程 > java > tcc分布式事务框架

tcc分布式事务框架体系解析

作者:kl

这篇文章主要为大家介绍了tcc分布式事务框架体系结构的解析说明,有需要的朋友可以借鉴参考下,希望能够有所帮助,祝大家多多进步

前言碎语

楼主之前推荐过2pc的分布式事务框架LCN。今天来详细聊聊TCC事务协议。

首先我们了解下什么是tcc,如下图

tcc分布式事务协议控制整体业务事务分为三个阶段。

try:执行业务逻辑

confirm:确定业务逻辑执行无误后,确定业务逻辑执行完成

cancel:假如try阶段有问题,执行cancel阶段逻辑,取消try阶段的数据

这就需要我们在设计业务时,在try阶段多想想业务处理的折中状态,比如,处理中,支付中,进行中等,在confirm阶段变更为处理完成,或者在cancel阶段变更为处理失败。

以电商下单为例

假设我们有一个电商下单的业务,有三个服务组成,订单服务处理下单逻辑,库存服务处理减库存逻辑,支付服务处理减账户余额逻辑。在下单服务里先后调用减库存和减余额的方法。如果使用tcc分布式事务来协调事务,我们服务就要做如下设计:

订单服务:

库存服务:

多加一个锁定库存的字段记录,用于记录业务处理中状态

支付服务:

多加一个冻结金额的字段记录,用于记录业务处理中状态

tcc分布式事务在这里起到了一个事务协调者的角色。真实业务只需要调用try阶段的方法。confirm和cancel阶段的额方法由tcc框架来帮我们调用完成最终业务逻辑。下面我们假设如下三个场景的业务情况,看tcc如何协调业务最终一致的。

hmily事务框架怎么做的?

通过上面对tcc事务协议说明大家应该都了解了tcc的处理协调机制,下面我们来看看hmily是怎么做到的,我们以接入支持dubbo服务为例。

概要:首先最基础两个应用点是aop和dubbo的filter机制,其次针对一组事务,定义了启动事务处理器,参与事务处理器去协调处理不同的事务单元。外加一个disruptor+ScheduledService处理事务日志,补偿处理失败的事务。

hmily框架以@Hmily注解为切入点,定义了一个环绕织入的切面,注解必填两个参数confirmMethod和cancelMethod,也就是tcc协调的两个阶段方法。在需要tcc事务的方法上面加上这个注解,也就托管了tcc三个阶段的处理流程。下面是aspect切面的抽象类,不同的RPC框架支持会有不同的实现 。其中真正处理业务逻辑需要实现HmilyTransactionInterceptor接口。

实现HmilyTransactionInterceptor接口

@Aspect
public abstract class AbstractHmilyTransactionAspect {
    private HmilyTransactionInterceptor hmilyTransactionInterceptor;
    protected void setHmilyTransactionInterceptor(final HmilyTransactionInterceptor hmilyTransactionInterceptor) {
        this.hmilyTransactionInterceptor = hmilyTransactionInterceptor;
    }
    /**
     * this is point cut with {@linkplain Hmily }.
     */
    @Pointcut("@annotation(org.dromara.hmily.annotation.Hmily)")
    public void hmilyInterceptor() {
    }
    /**
     * this is around in {@linkplain Hmily }.
     * @param proceedingJoinPoint proceedingJoinPoint
     * @return Object
     * @throws Throwable  Throwable
     */
    @Around("hmilyInterceptor()")
    public Object interceptTccMethod(final ProceedingJoinPoint proceedingJoinPoint) throws Throwable {
        return hmilyTransactionInterceptor.interceptor(proceedingJoinPoint);
    }
    /**
     * spring Order.
     *
     * @return int
     */
    public abstract int getOrder();
}

dubbo的aspect抽象实现

@Aspect
@Component
public class DubboHmilyTransactionAspect extends AbstractHmilyTransactionAspect implements Ordered {
    @Autowired
    public DubboHmilyTransactionAspect(final DubboHmilyTransactionInterceptor dubboHmilyTransactionInterceptor) {
        super.setHmilyTransactionInterceptor(dubboHmilyTransactionInterceptor);
    }
    @Override
    public int getOrder() {
        return Ordered.HIGHEST_PRECEDENCE;
    }
}

dubbo的HmilyTransactionInterceptor实现

@Component
public class DubboHmilyTransactionInterceptor implements HmilyTransactionInterceptor {
    private final HmilyTransactionAspectService hmilyTransactionAspectService;
    @Autowired
    public DubboHmilyTransactionInterceptor(final HmilyTransactionAspectService hmilyTransactionAspectService) {
        this.hmilyTransactionAspectService = hmilyTransactionAspectService;
    }
  @Override
public Object interceptor(final ProceedingJoinPoint pjp) throws Throwable {
    final String context = RpcContext.getContext().getAttachment(CommonConstant.HMILY_TRANSACTION_CONTEXT);
    HmilyTransactionContext hmilyTransactionContext;
    //判断dubbo上下文中是否携带了tcc事务,如果有就取出反序列化为事务上下文对象
    if (StringUtils.isNoneBlank(context)) {
        hmilyTransactionContext = GsonUtils.getInstance().fromJson(context, HmilyTransactionContext.class);
        RpcContext.getContext().getAttachments().remove(CommonConstant.HMILY_TRANSACTION_CONTEXT);
    } else {
        //如果dubbo上下文中没有,就从当前上下文中获取。如果是事务发起者,这里其实也获取不到事务
        hmilyTransactionContext = HmilyTransactionContextLocal.getInstance().get();
    }
    return hmilyTransactionAspectService.invoke(hmilyTransactionContext, pjp);
}
}

这里主要判断了dubbo上下文中是否携带了tcc事务。如果没有就从当前线程上下文中获取,如果是事务的发起者,这里其实获取不到事务上下文对象的。在invoke里有个获取事务处理器的逻辑,如果事务上下文入参 为null,那么获取到的就是启动事务处理器。

启动事务处理器处理逻辑如下

public Object handler(final ProceedingJoinPoint point, final HmilyTransactionContext context)
        throws Throwable {
    System.err.println("StarterHmilyTransactionHandler");
    Object returnValue;
    try {
        HmilyTransaction hmilyTransaction = hmilyTransactionExecutor.begin(point);
        try {
            //execute try
            returnValue = point.proceed();
            hmilyTransaction.setStatus(HmilyActionEnum.TRYING.getCode());
            hmilyTransactionExecutor.updateStatus(hmilyTransaction);
        } catch (Throwable throwable) {
            //if exception ,execute cancel
            final HmilyTransaction currentTransaction = hmilyTransactionExecutor.getCurrentTransaction();
            executor.execute(() -> hmilyTransactionExecutor
                    .cancel(currentTransaction));
            throw throwable;
        }
        //execute confirm
        final HmilyTransaction currentTransaction = hmilyTransactionExecutor.getCurrentTransaction();
        executor.execute(() -> hmilyTransactionExecutor.confirm(currentTransaction));
    } finally {
        HmilyTransactionContextLocal.getInstance().remove();
        hmilyTransactionExecutor.remove();
    }
    return returnValue;
}

真正业务处理方法,point.proceed();被try,catch包起来了,如果try里面的方法出现异常,就会走hmilyTransactionExecutor.cancel(currentTransaction)的逻辑,如果成功,就走hmilyTransactionExecutor.confirm(currentTransaction)逻辑。其中cancel和confirm里都有协调参与者事务的处理逻辑,以confirm逻辑为例。

public void confirm(final HmilyTransaction currentTransaction) throws HmilyRuntimeException {
    LogUtil.debug(LOGGER, () -> "tcc confirm .......!start");
    if (Objects.isNull(currentTransaction) || CollectionUtils.isEmpty(currentTransaction.getHmilyParticipants())) {
        return;
    }
    currentTransaction.setStatus(HmilyActionEnum.CONFIRMING.getCode());
    updateStatus(currentTransaction);
    final ListhmilyParticipants = currentTransaction.getHmilyParticipants();
    ListfailList = Lists.newArrayListWithCapacity(hmilyParticipants.size());
    boolean success = true;
    if (CollectionUtils.isNotEmpty(hmilyParticipants)) {
        for (HmilyParticipant hmilyParticipant : hmilyParticipants) {
            try {
                HmilyTransactionContext context = new HmilyTransactionContext();
                context.setAction(HmilyActionEnum.CONFIRMING.getCode());
                context.setRole(HmilyRoleEnum.START.getCode());
                context.setTransId(hmilyParticipant.getTransId());
                HmilyTransactionContextLocal.getInstance().set(context);
                executeParticipantMethod(hmilyParticipant.getConfirmHmilyInvocation());
            } catch (Exception e) {
                LogUtil.error(LOGGER, "execute confirm :{}", () -> e);
                success = false;
                failList.add(hmilyParticipant);
            } finally {
                HmilyTransactionContextLocal.getInstance().remove();
            }
        }
        executeHandler(success, currentTransaction, failList);
    }
}

可以看到executeParticipantMethod(hmilyParticipant.getConfirmHmilyInvocation()),这里执行了事务参与者的confirm方法。同理cancel里面也有类似代码,执行事务参与者的cancel方法。那么事务参与者的信息是怎么获取到的呢?我们需要回到一开始提到的dubbo的filter机制。

@Activate(group = {Constants.SERVER_KEY, Constants.CONSUMER})
public class DubboHmilyTransactionFilter implements Filter {
    private HmilyTransactionExecutor hmilyTransactionExecutor;
    /**
     * this is init by dubbo spi
     * set hmilyTransactionExecutor.
     *
     * @param hmilyTransactionExecutor {@linkplain HmilyTransactionExecutor }
     */
    public void setHmilyTransactionExecutor(final HmilyTransactionExecutor hmilyTransactionExecutor) {
        this.hmilyTransactionExecutor = hmilyTransactionExecutor;
    }
    @Override
    @SuppressWarnings("unchecked")
    public Result invoke(final Invoker invoker, final Invocation invocation) throws RpcException {
        String methodName = invocation.getMethodName();
        Class clazz = invoker.getInterface();
        Class[] args = invocation.getParameterTypes();
        final Object[] arguments = invocation.getArguments();
        converterParamsClass(args, arguments);
        Method method = null;
        Hmily hmily = null;
        try {
            method = clazz.getMethod(methodName, args);
            hmily = method.getAnnotation(Hmily.class);
        } catch (NoSuchMethodException e) {
            e.printStackTrace();
        }
        if (Objects.nonNull(hmily)) {
            try {
                final HmilyTransactionContext hmilyTransactionContext = HmilyTransactionContextLocal.getInstance().get();
                if (Objects.nonNull(hmilyTransactionContext)) {
                    if (hmilyTransactionContext.getRole() == HmilyRoleEnum.LOCAL.getCode()) {
                        hmilyTransactionContext.setRole(HmilyRoleEnum.INLINE.getCode());
                    }
                    RpcContext.getContext().setAttachment(CommonConstant.HMILY_TRANSACTION_CONTEXT, GsonUtils.getInstance().toJson(hmilyTransactionContext));
                }
                final Result result = invoker.invoke(invocation);
                //if result has not exception
                if (!result.hasException()) {
                    final HmilyParticipant hmilyParticipant = buildParticipant(hmilyTransactionContext, hmily, method, clazz, arguments, args);
                    if (hmilyTransactionContext.getRole() == HmilyRoleEnum.INLINE.getCode()) {
                        hmilyTransactionExecutor.registerByNested(hmilyTransactionContext.getTransId(),
                                hmilyParticipant);
                    } else {
                        hmilyTransactionExecutor.enlistParticipant(hmilyParticipant);
                    }
                } else {
                    throw new HmilyRuntimeException("rpc invoke exception{}", result.getException());
                }
                return result;
            } catch (RpcException e) {
                e.printStackTrace();
                throw e;
            }
        } else {
            return invoker.invoke(invocation);
        }
    }
    @SuppressWarnings("unchecked")
    private HmilyParticipant buildParticipant(final HmilyTransactionContext hmilyTransactionContext,
                                              final Hmily hmily,
                                              final Method method, final Class clazz,
                                              final Object[] arguments, final Class... args) throws HmilyRuntimeException {
        if (Objects.isNull(hmilyTransactionContext)
                || (HmilyActionEnum.TRYING.getCode() != hmilyTransactionContext.getAction())) {
            return null;
        }
        //获取协调方法
        String confirmMethodName = hmily.confirmMethod();
        if (StringUtils.isBlank(confirmMethodName)) {
            confirmMethodName = method.getName();
        }
        String cancelMethodName = hmily.cancelMethod();
        if (StringUtils.isBlank(cancelMethodName)) {
            cancelMethodName = method.getName();
        }
        HmilyInvocation confirmInvocation = new HmilyInvocation(clazz, confirmMethodName, args, arguments);
        HmilyInvocation cancelInvocation = new HmilyInvocation(clazz, cancelMethodName, args, arguments);
        //封装调用点
        return new HmilyParticipant(hmilyTransactionContext.getTransId(), confirmInvocation, cancelInvocation);
    }
    private void converterParamsClass(final Class[] args, final Object[] arguments) {
        if (arguments == null || arguments.length < 1) {
            return;
        }
        for (int i = 0; i < arguments.length; i++) {
            args[i] = arguments[i].getClass();
        }
    }
}

需要注意三个地方

参数者事务处理器

public Object handler(final ProceedingJoinPoint point, final HmilyTransactionContext context) throws Throwable {
    HmilyTransaction hmilyTransaction = null;
    HmilyTransaction currentTransaction;
    switch (HmilyActionEnum.acquireByCode(context.getAction())) {
        case TRYING:
            try {
                hmilyTransaction = hmilyTransactionExecutor.beginParticipant(context, point);
                final Object proceed = point.proceed();
                hmilyTransaction.setStatus(HmilyActionEnum.TRYING.getCode());
                //update log status to try
                hmilyTransactionExecutor.updateStatus(hmilyTransaction);
                return proceed;
            } catch (Throwable throwable) {
                //if exception ,delete log.
                hmilyTransactionExecutor.deleteTransaction(hmilyTransaction);
                throw throwable;
            } finally {
               HmilyTransactionContextLocal.getInstance().remove();
            }
        case CONFIRMING:
            currentTransaction = HmilyTransactionCacheManager.getInstance().getTccTransaction(context.getTransId());
            hmilyTransactionExecutor.confirm(currentTransaction);
            break;
        case CANCELING:
            currentTransaction = HmilyTransactionCacheManager.getInstance().getTccTransaction(context.getTransId());
            hmilyTransactionExecutor.cancel(currentTransaction);
            break;
        default:
            break;
    }
    Method method = ((MethodSignature) (point.getSignature())).getMethod();
    logger.error(HmilyActionEnum.acquireByCode(context.getAction()).getDesc());
    return DefaultValueUtils.getDefaultValue(method.getReturnType());
}

参与者事务处理器的逻辑比启动事务处理器要简单很多,try阶段记录事务日志用于事务补偿的时候使用。其他的confirm和cancel都是由启动事务管理器来触发调用执行的。这个地方之前纠结了楼主几个小时,怎么一个环绕织入的切面会被触发执行两次,其实是启动事务处理器里的confirm或cancel触发的。

disruptor+ScheduledService处理事务日志,补偿处理失败的事务

这个不细聊了,简述下。disruptor是一个高性能的队列。对事务日志落地的所有操作都是通过disruptor来异步完成的。ScheduledService默认128秒执行一次,来检查是否有处理失败的事务日志,用于补偿事务协调失败的事务

文末结语

相比较2pc的LCN而言,tcc分布式事务对业务侵入性更高。也因2pc的长时间占用事务资源,tcc的性能肯定比2pc要好。两者之间本身不存在谁优谁劣的问题。所以在做分布式事务选型时,选一个对的适合自身业务的分布式事务框架就比较重要了。

以上就是tcc分布式事务框架体系解析的详细内容,更多关于tcc分布式事务框架的资料请关注脚本之家其它相关文章!

您可能感兴趣的文章:
阅读全文