分布式事务Seata使用不当引发的全局锁问题

原创已于 2025-08-12 21:05:24 修改 · 497 阅读

2 ·

本内容遵循CC 4.0 BY-SA版权协议

GEO检测

标签

#seata #分布式事务

于 2025-08-12 21:03:20 首次发布

报错信息

2025-08-12T19:36:27.803+08:00 ERROR 1 --- [xxx-inference] [io-32123-exec-7] o.a.s.r.d.exec.AbstractDMLBaseExecutor : execute executeAutoCommitTrue error:Global lock wait timeout org.apache.seata.rm.datasource.exec.LockWaitTimeoutException: Global lock wait timeout at org.apache.seata.rm.datasource.exec.LockRetryController.sleep(LockRetryController.java:66) ~[seata-all-2.1.0.jar!/:2.1.0] at org.apache.seata.rm.datasource.ConnectionProxy$LockRetryPolicy.doRetryOnLockConflict(ConnectionProxy.java:364) ~[seata-all-2.1.0.jar!/:2.1.0] at org.apache.seata.rm.datasource.exec.AbstractDMLBaseExecutor$LockRetryPolicy.execute(AbstractDMLBaseExecutor.java:188) ~[seata-all-2.1.0.jar!/:2.1.0]

错误代码示例

@GlobalTransactional
public void deletePod(PodBo podBo) {
    KubernetesClient kubernetesClient = k8SClientMgr.getConnect("", true);
    
    ModelServiceEntity modelServiceEntity = modelServiceMapper.selectById(podBo.getServiceId());
    if (modelServiceEntity == null) {
         throw new RuntimeException("模型服务不存在");
    }

    if (podBo.getServiceId() != null) {
        UpdateWrapper<ModelServiceEntity> updateWrapper = new UpdateWrapper<>();
        updateWrapper.eq("id", podBo.getServiceId());
        updateWrapper.set("status", 1);
        modelServiceMapper.update(null, updateWrapper);

        // 远程调用删除任务
        inferenceTaskApi.deleteInferenceTask(Arrays.asList(modelServiceEntity.getInferenceTaskId().split(",")));
    }

    ResourceDefinitionContext context = new ResourceDefinitionContext.Builder()
            .withGroup("kubeai.org")
            .withVersion("v1")
            .withPlural("models")
            .withKind("Model")
            .withNamespaced(true)
            .build();

    // 删除 CRD 资源
    kubernetesClient.genericKubernetesResources(context)
            .inNamespace(namespace)
            .withName(podBo.getName().toLowerCase())
            .delete();
}

问题分析

// 这里查询了model_service表，seata会对model_service表进行加锁
ModelServiceEntity modelServiceEntity = modelServiceMapper.selectById(podBo.getServiceId());

// 这里又进行了修改操作，由于前面已经加锁了，导致这里无法执行成功，进而超时
modelServiceMapper.update(null, updateWrapper);

问题解决

查询逻辑和更新逻辑不在一个事务里执行即可

public void deletePod(PodBo podBo) {
    KubernetesClient kubernetesClient = k8SClientMgr.getConnect("", true);
    
    ModelServiceEntity modelServiceEntity = modelServiceMapper.selectById(podBo.getServiceId());
    if (modelServiceEntity == null) {
        throw new RuntimeException("模型服务不存在");
    }

    deleteResource(kubernetesClient, podBo, modelServiceEntity);
}

@GlobalTransactional
public void deleteResource(KubernetesClient kubernetesClient, PodBo podBo, ModelServiceEntity modelServiceEntity) {
    if (podBo.getServiceId() != null) {
        UpdateWrapper<ModelServiceEntity> updateWrapper = new UpdateWrapper<>();
        updateWrapper.eq("id", podBo.getServiceId());
        updateWrapper.set("status", 1);
        modelServiceMapper.update(null, updateWrapper);

        // 远程调用删除任务
        inferenceTaskApi.deleteInferenceTask(Arrays.asList(modelServiceEntity.getInferenceTaskId().split(",")));
    }

    ResourceDefinitionContext context = new ResourceDefinitionContext.Builder()
            .withGroup("kubeai.org")
            .withVersion("v1")
            .withPlural("models")
            .withKind("Model")
            .withNamespaced(true)
            .build();

    // 删除 CRD 资源
    kubernetesClient.genericKubernetesResources(context)
            .inNamespace(namespace)
            .withName(podBo.getName().toLowerCase())
            .delete();
}