大厂防止超卖的7种实现,很受用!

B站影视 2025-01-17 08:59 3

摘要:@ApiOperation(value="秒杀实现方式——Lock加锁")@PostMapping("/start/lock")public Result startLock(long skgId){ try { log.info("开始秒杀方式一...");

高并发场景在现场的日常工作中很常见,特别是在互联网公司中,这篇文章就来通过秒杀商品来模拟高并发的场景。

本文环境: SpringBoot 2.5.7 + MySQL 8.0 X + MybatisPlus + Swagger2.9.2模拟工具: Jmeter模拟场景: 减库存->创建订单->模拟支付

在开发中,对于下面的代码,可能很熟悉:在Service里面加上@Transactional事务注解和lock锁。

控制层:Controller

@ApiOperation(value="秒杀实现方式——Lock加锁")@PostMapping("/start/lock")public Result startLock(long skgId){ try { log.info("开始秒杀方式一..."); final long userId = (int) (new Random.nextDouble * (99999 - 10000 + 1)) + 10000; Result result = secondKillService.startSecondKillByLock(skgId, userId); if(result != null){ log.info("用户:{}--{}", userId, result.get("msg")); }else{ log.info("用户:{}--{}", userId, "哎呦喂,人也太多了,请稍后!"); } } catch (Exception e) { e.printStackTrace; } finally { } return Result.ok;}

业务层:Service

@Override@Transactional(rollbackFor = Exception.class)public Result startSecondKillByLock(long skgId, long userId) { lock.lock; try { // 校验库存 SecondKill secondKill = secondKillMapper.selectById(skgId); Integer number = secondKill.getNumber; if (number > 0) { // 扣库存 secondKill.setNumber(number - 1); secondKillMapper.updateById(secondKill); // 创建订单 SuccessKilled killed = new SuccessKilled; killed.setSeckillId(skgId); killed.setUserId(userId); killed.setState((short) 0); killed.setCreateTime(new Timestamp(System.currentTimeMillis)); successKilledMapper.insert(killed); // 模拟支付 Payment payment = new Payment; payment.setSeckillId(skgId); payment.setSeckillId(skgId); payment.setUserId(userId); payment.setMoney(40); payment.setState((short) 1); payment.setCreateTime(new Timestamp(System.currentTimeMillis)); paymentMapper.insert(payment); } else { return Result.error(SecondKillStateEnum.END); } } catch (Exception e) { throw new ScorpiosException("异常了个乖乖"); } finally { lock.unlock; } return Result.ok(SecondKillStateEnum.SUCCESS);}

对于上面的代码应该没啥问题吧,业务方法上加事务,在处理业务的时候加锁。

但上面这样写法是有问题的,会出现超卖的情况,看下测试结果:模拟1000个并发,抢100商品。

这里在业务方法开始加了锁,在业务方法结束后释放了锁。但这里的事务提交却不是这样的,有可能在事务提交之前,就已经把锁释放了,这样会导致商品超卖现象。所以加锁的时机很重要!

对于上面超卖现象,主要问题出现在事务中锁释放的时机,事务未提交之前,锁已经释放。(事务提交是在整个方法执行完)。如何解决这个问题呢,就是把加锁步骤提前

可以在controller层进行加锁可以使用Aop在业务方法执行之前进行加锁@ApiOperation(value="秒杀实现方式——Lock加锁")@PostMapping("/start/lock")public Result startLock(long skgId){ // 在此处加锁 lock.lock; try { log.info("开始秒杀方式一..."); final long userId = (int) (new Random.nextDouble * (99999 - 10000 + 1)) + 10000; Result result = secondKillService.startSecondKillByLock(skgId, userId); if(result != null){ log.info("用户:{}--{}", userId, result.get("msg")); }else{ log.info("用户:{}--{}", userId, "哎呦喂,人也太多了,请稍后!"); } } catch (Exception e) { e.printStackTrace; } finally { // 在此处释放锁 lock.unlock; } return Result.ok;}

上面这样的加锁就可以解决事务未提交之前,锁释放的问题,可以分三种情况进行压力测试:

并发数1000,商品100并发数1000,商品1000并发数2000,商品1000

对于并发量大于商品数的情况,商品秒杀一般不会出现少卖的请况,但对于并发数小于等于商品数的时候可能会出现商品少卖情况,这也很好理解。

对于没有问题的情况就不贴图了,因为有很多种方式,贴图会太多

对于上面在控制层进行加锁的方式,可能显得不优雅,那就还有另一种方式进行在事务之前加锁,那就是AOP

自定义AOP注解

@Target({ElementType.PARAMETER, ElementType.METHOD})@Retention(RetentionPolicy.RUNTIME)@Documentedpublic @interface ServiceLock { String description default "";}

定义切面类

@Slf4j@Component@Scope@Aspect@order(1) //order越小越是最先执行,但更重要的是最先执行的最后结束public class LockAspect { /** * 思考:为什么不用synchronized * service 默认是单例的,并发下lock只有一个实例 */ private static Lock lock = new ReentrantLock(true); // 互斥锁 参数默认false,不公平锁 // Service层切点 用于记录错误日志 @Pointcut("@annotation(com.scorpios.secondkill.aop.ServiceLock)") public void lockAspect { } @Around("lockAspect") public Object around(ProceedingJoinPoint joinPoint) { lock.lock; Object obj = null; try { obj = joinPoint.proceed; } catch (Throwable e) { e.printStackTrace; throw new RuntimeException; } finally{ lock.unlock; } return obj; }}

在业务方法上添加AOP注解

@Override@ServiceLock // 使用Aop进行加锁@Transactional(rollbackFor = Exception.class)public Result startSecondKillByAop(long skgId, long userId) { try { // 校验库存 SecondKill secondKill = secondKillMapper.selectById(skgId); Integer number = secondKill.getNumber; if (number > 0) { //扣库存 secondKill.setNumber(number - 1); secondKillMapper.updateById(secondKill); //创建订单 SuccessKilled killed = new SuccessKilled; killed.setSeckillId(skgId); killed.setUserId(userId); killed.setState((short) 0); killed.setCreateTime(new Timestamp(System.currentTimeMillis)); successKilledMapper.insert(killed); //支付 Payment payment = new Payment; payment.setSeckillId(skgId); payment.setSeckillId(skgId); payment.setUserId(userId); payment.setMoney(40); payment.setState((short) 1); payment.setCreateTime(new Timestamp(System.currentTimeMillis)); paymentMapper.insert(payment); } else { return Result.error(SecondKillStateEnum.END); } } catch (Exception e) { throw new ScorpiosException("异常了个乖乖"); } return Result.ok(SecondKillStateEnum.SUCCESS);}

控制层:

@ApiOperation(value="秒杀实现方式二——Aop加锁")@PostMapping("/start/aop")public Result startAop(long skgId){ try { log.info("开始秒杀方式二..."); final long userId = (int) (new Random.nextDouble * (99999 - 10000 + 1)) + 10000; Result result = secondKillService.startSecondKillByAop(skgId, userId); if(result != null){ log.info("用户:{}--{}", userId, result.get("msg")); }else{ log.info("用户:{}--{}", userId, "哎呦喂,人也太多了,请稍后!"); } } catch (Exception e) { e.printStackTrace; } return Result.ok;}

这种方式在对锁的使用上,更高阶、更美观!

除了上面在业务代码层面加锁外,还可以使用数据库自带的锁进行并发控制。

悲观锁,什么是悲观锁呢?通俗的说,在做任何事情之前,都要进行加锁确认。这种数据库级加锁操作效率较低。

使用for update一定要加上事务,当事务处理完后,for update才会将行级锁解除

如果请求数和秒杀商品数量一致,会出现少卖

@ApiOperation(value="秒杀实现方式三——悲观锁")@PostMapping("/start/pes/lock/one")public Result startPesLockOne(long skgId){ try { log.info("开始秒杀方式三..."); final long userId = (int) (new Random.nextDouble * (99999 - 10000 + 1)) + 10000; Result result = secondKillService.startSecondKillByUpdate(skgId, userId); if(result != null){ log.info("用户:{}--{}", userId, result.get("msg")); }else{ log.info("用户:{}--{}", userId, "哎呦喂,人也太多了,请稍后!"); } } catch (Exception e) { e.printStackTrace; } return Result.ok;}

业务逻辑

@Override@Transactional(rollbackFor = Exception.class)public Result startSecondKillByUpdate(long skgId, long userId) { try { // 校验库存-悲观锁 SecondKill secondKill = secondKillMapper.querySecondKillForUpdate(skgId); Integer number = secondKill.getNumber; if (number > 0) { //扣库存 secondKill.setNumber(number - 1); secondKillMapper.updateById(secondKill); //创建订单 SuccessKilled killed = new SuccessKilled; killed.setSeckillId(skgId); killed.setUserId(userId); killed.setState((short) 0); killed.setCreateTime(new Timestamp(System.currentTimeMillis)); successKilledMapper.insert(killed); //支付 Payment payment = new Payment; payment.setSeckillId(skgId); payment.setSeckillId(skgId); payment.setUserId(userId); payment.setMoney(40); payment.setState((short) 1); payment.setCreateTime(new Timestamp(System.currentTimeMillis)); paymentMapper.insert(payment); } else { return Result.error(SecondKillStateEnum.END); } } catch (Exception e) { throw new ScorpiosException("异常了个乖乖"); } finally { } return Result.ok(SecondKillStateEnum.SUCCESS);}

Dao层

@Repositorypublic interface SecondKillMapper extends BaseMapper { /** * 将此行数据进行加锁,当整个方法将事务提交后,才会解锁 * @param skgId * @return */ @Select(value = "SELECT * FROM seckill WHERE seckill_id=#{skgId} FOR UPDATE") SecondKill querySecondKillForUpdate(@Param("skgId") Long skgId);}

上面是利用for update进行对查询数据加锁,加的是行锁。

悲观锁的第二种方式就是利用update更新命令来加表锁

/** * UPDATE锁表 * @param skgId 商品id * @param userId 用户id * @return */@Override@Transactional(rollbackFor = Exception.class)public Result startSecondKillByUpdateTwo(long skgId, long userId) { try { // 不校验,直接扣库存更新 int result = secondKillMapper.updateSecondKillById(skgId); if (result > 0) { //创建订单 SuccessKilled killed = new SuccessKilled; killed.setSeckillId(skgId); killed.setUserId(userId); killed.setState((short) 0); killed.setCreateTime(new Timestamp(System.currentTimeMillis)); successKilledMapper.insert(killed); //支付 Payment payment = new Payment; payment.setSeckillId(skgId); payment.setSeckillId(skgId); payment.setUserId(userId); payment.setMoney(40); payment.setState((short) 1); payment.setCreateTime(new Timestamp(System.currentTimeMillis)); paymentMapper.insert(payment); } else { return Result.error(SecondKillStateEnum.END); } } catch (Exception e) { throw new ScorpiosException("异常了个乖乖"); } finally { } return Result.ok(SecondKillStateEnum.SUCCESS);}

Dao层

@Repositorypublic interface SecondKillMapper extends BaseMapper { /** * 将此行数据进行加锁,当整个方法将事务提交后,才会解锁 * @param skgId * @return */ @Select(value = "SELECT * FROM seckill WHERE seckill_id=#{skgId} FOR UPDATE") SecondKill querySecondKillForUpdate(@Param("skgId") Long skgId); @Update(value = "UPDATE seckill SET number=number-1 WHERE seckill_id=#{skgId} AND number > 0") int updateSecondKillById(@Param("skgId") long skgId);}

乐观锁,顾名思义,就是对操作结果很乐观,通过利用version字段来判断数据是否被修改。

乐观锁,不进行库存数量的校验,直接做库存扣减。

这里使用的乐观锁会出现大量的数据更新异常(抛异常就会导致购买失败)、如果配置的抢购人数比较少、比如120:100(人数:商品) 会出现少买的情况,不推荐使用乐观锁。

@ApiOperation(value="秒杀实现方式五——乐观锁")@PostMapping("/start/opt/lock")public Result startOptLock(long skgId){ try { log.info("开始秒杀方式五..."); final long userId = (int) (new Random.nextDouble * (99999 - 10000 + 1)) + 10000; // 参数添加了购买数量 Result result = secondKillService.startSecondKillByPesLock(skgId, userId,1); if(result != null){ log.info("用户:{}--{}", userId, result.get("msg")); }else{ log.info("用户:{}--{}", userId, "哎呦喂,人也太多了,请稍后!"); } } catch (Exception e) { e.printStackTrace; } return Result.ok;}@Override@Transactional(rollbackFor = Exception.class)public Result startSecondKillByPesLock(long skgId, long userId, int number) { // 乐观锁,不进行库存数量的校验,直接 try { SecondKill kill = secondKillMapper.selectById(skgId); // 剩余的数量应该要大于等于秒杀的数量 if(kill.getNumber >= number) { int result = secondKillMapper.updateSecondKillByVersion(number,skgId,kill.getVersion); if (result > 0) { //创建订单 SuccessKilled killed = new SuccessKilled; killed.setSeckillId(skgId); killed.setUserId(userId); killed.setState((short) 0); killed.setCreateTime(new Timestamp(System.currentTimeMillis)); successKilledMapper.insert(killed); //支付 Payment payment = new Payment; payment.setSeckillId(skgId); payment.setSeckillId(skgId); payment.setUserId(userId); payment.setMoney(40); payment.setState((short) 1); payment.setCreateTime(new Timestamp(System.currentTimeMillis)); paymentMapper.insert(payment); } else { return Result.error(SecondKillStateEnum.END); } } } catch (Exception e) { throw new ScorpiosException("异常了个乖乖"); } finally { } return Result.ok(SecondKillStateEnum.SUCCESS);}@Repositorypublic interface SecondKillMapper extends BaseMapper { /** * 将此行数据进行加锁,当整个方法将事务提交后,才会解锁 * @param skgId * @return */ @Select(value = "SELECT * FROM seckill WHERE seckill_id=#{skgId} FOR UPDATE") SecondKill querySecondKillForUpdate(@Param("skgId") Long skgId); @Update(value = "UPDATE seckill SET number=number-1 WHERE seckill_id=#{skgId} AND number > 0") int updateSecondKillById(@Param("skgId") long skgId); @Update(value = "UPDATE seckill SET number=number-#{number},version=version+1 WHERE seckill_id=#{skgId} AND version = #{version}") int updateSecondKillByVersion(@Param("number") int number, @Param("skgId") long skgId, @Param("version")int version);}

乐观锁会出现大量的数据更新异常(抛异常就会导致购买失败),会出现少买的情况,不推荐使用乐观锁。

利用阻塞队类,也可以解决高并发问题。其思想就是把接收到的请求按顺序存放到队列中,消费者线程逐一从队列里取数据进行处理,看下具体代码。

阻塞队列:这里使用静态内部类的方式来实现单例模式,在并发条件下不会出现问题。

// 秒杀队列(固定长度为100)public class SecondKillQueue { // 队列大小 static final int QUEUE_MAX_SIZE = 100; // 用于多线程间下单的队列 static BlockingQueue blockingQueue = new LinkedBlockingQueue(QUEUE_MAX_SIZE); // 使用静态内部类,实现单例模式 private SecondKillQueue{}; private static class SingletonHolder{ // 静态初始化器,由JVM来保证线程安全 private static SecondKillQueue queue = new SecondKillQueue; } /** * 单例队列 * @return */ public static SecondKillQueue getSkillQueue{ return SingletonHolder.queue; } /** * 生产入队 * @param kill * @throws InterruptedException * add(e) 队列未满时,返回true;队列满则抛出IllegalStateException(“Queue full”)异常——AbstractQueue * put(e) 队列未满时,直接插入没有返回值;队列满时会阻塞等待,一直等到队列未满时再插入。 * offer(e) 队列未满时,返回true;队列满时返回false。非阻塞立即返回。 * offer(e, time, unit) 设定等待的时间,如果在指定时间内还不能往队列中插入数据则返回false,插入成功返回true。 */ public Boolean produce(SuccessKilled kill) { return blockingQueue.offer(kill); } /** * 消费出队 * poll 获取并移除队首元素,在指定的时间内去轮询队列看有没有首元素有则返回,否者超时后返回null * take 与带超时时间的poll类似不同在于take时候如果当前队列空了它会一直等待其他线程调用notEmpty.signal才会被唤醒 */ public SuccessKilled consume throws InterruptedException { return blockingQueue.take; } /** * 获取队列大小 * @return */ public int size { return blockingQueue.size; }}

消费秒杀队列:实现ApplicationRunner接口

// 消费秒杀队列@Slf4j@Componentpublic class TaskRunner implements ApplicationRunner{ @Autowired private SecondKillService seckillService; @Override public void run(ApplicationArguments var){ new Thread( -> { log.info("队列启动成功"); while(true){ try { // 进程内队列 SuccessKilled kill = SecondKillQueue.getSkillQueue.consume; if(kill != null){ Result result = seckillService.startSecondKillByAop(kill.getSeckillId, kill.getUserId); if(result != null && result.equals(Result.ok(SecondKillStateEnum.SUCCESS))){ log.info("TaskRunner,result:{}",result); log.info("TaskRunner从消息队列取出用户,用户:{}{}",kill.getUserId,"秒杀成功"); } } } catch (InterruptedException e) { e.printStackTrace; } } }).start; }}@ApiOperation(value="秒杀实现方式六——消息队列")@PostMapping("/start/queue")public Result startQueue(long skgId){ try { log.info("开始秒杀方式六..."); final long userId = (int) (new Random.nextDouble * (99999 - 10000 + 1)) + 10000; SuccessKilled kill = new SuccessKilled; kill.setSeckillId(skgId); kill.setUserId(userId); Boolean flag = SecondKillQueue.getSkillQueue.produce(kill); // 虽然进入了队列,但是不一定能秒杀成功 进队出队有时间间隙 if(flag){ log.info("用户:{}{}",kill.getUserId,"秒杀成功"); }else{ log.info("用户:{}{}",userId,"秒杀失败"); } } catch (Exception e) { e.printStackTrace; } return Result.ok;}

注意:在业务层和AOP方法中,不能抛出任何异常, throw new RuntimeException这些抛异常代码要注释掉。因为一旦程序抛出异常就会停止,导致消费秒杀队列进程终止!

使用阻塞队列来实现秒杀,有几点要注意:

消费秒杀队列中调用业务方法加锁与不加锁情况一样,也就是seckillService.startSecondKillByAop、seckillService.startSecondKillByLock方法结果一样,这也很好理解当队列长度与商品数量一致时,会出现少卖的现象,可以调大数值下面是队列长度1000,商品数量1000,并发数2000情况下出现的少卖

Disruptor是个高性能队列,研发的初衷是解决内存队列的延迟问题,在性能测试中发现竟然与I/O操作处于同样的数量级,基于Disruptor开发的系统单线程能支撑每秒600万订单。

// 事件生成工厂(用来初始化预分配事件对象)public class SecondKillEventFactory implements EventFactory { @Override public SecondKillEvent newInstance { return new SecondKillEvent; }}// 事件对象(秒杀事件)public class SecondKillEvent implements Serializable { private static final long serialVersionUID = 1L; private long seckillId; private long userId; // set/get方法略}// 使用translator方式生产者public class SecondKillEventProducer { private final static EventTranslatorVararg translator = (seckillEvent, seq, objs) -> { seckillEvent.setSeckillId((Long) objs[0]); seckillEvent.setUserId((Long) objs[1]); }; private final RingBuffer ringBuffer; public SecondKillEventProducer(RingBuffer ringBuffer){ this.ringBuffer = ringBuffer; } public void secondKill(long seckillId, long userId){ this.ringBuffer.publishEvent(translator, seckillId, userId); }}// 消费者(秒杀处理器)@Slf4jpublic class SecondKillEventConsumer implements EventHandler { private SecondKillService secondKillService = (SecondKillService) SpringUtil.getBean("secondKillService"); @Override public void onEvent(SecondKillEvent seckillEvent, long seq, boolean bool) { Result result = secondKillService.startSecondKillByAop(seckillEvent.getSeckillId, seckillEvent.getUserId); if(result.equals(Result.ok(SecondKillStateEnum.SUCCESS))){ log.info("用户:{}{}",seckillEvent.getUserId,"秒杀成功"); } }}public class DisruptorUtil { static Disruptor disruptor; static{ SecondKillEventFactory factory = new SecondKillEventFactory; int ringBufferSize = 1024; ThreadFactory threadFactory = runnable -> new Thread(runnable); disruptor = new Disruptor(factory, ringBufferSize, threadFactory); disruptor.handleEventsWith(new SecondKillEventConsumer); disruptor.start; } public static void producer(SecondKillEvent kill){ RingBuffer ringBuffer = disruptor.getRingBuffer; SecondKillEventProducer producer = new SecondKillEventProducer(ringBuffer); producer.secondKill(kill.getSeckillId,kill.getUserId); }}@ApiOperation(value="秒杀实现方式七——Disruptor队列")@PostMapping("/start/disruptor")public Result startDisruptor(long skgId){ try { log.info("开始秒杀方式七..."); final long userId = (int) (new Random.nextDouble * (99999 - 10000 + 1)) + 10000; SecondKillEvent kill = new SecondKillEvent; kill.setSeckillId(skgId); kill.setUserId(userId); DisruptorUtil.producer(kill); } catch (Exception e) { e.printStackTrace; } return Result.ok;}

经过测试,发现使用Disruptor队列队列,与自定义队列有着同样的问题,也会出现超卖的情况,但效率有所提高。

一、二方式是在代码中利用锁和事务的方式解决了并发问题,主要解决的是锁要加载事务之前三、四、五方式主要是数据库的锁来解决并发问题,方式三是利用for upate对表加行锁,方式四是利用update来对表加锁,方式五是通过增加version字段来控制数据库的更新操作,方式五的效果最差六、七方式是通过队列来解决并发问题,这里需要特别注意的是,在代码中不能通过throw抛异常,否则消费线程会终止,而且由于进队和出队存在时间间隙,会导致商品少卖

来源:散文随风想

相关推荐