Abstract: When it comes to the cloud data store, due to data segmentation, multiple copies of each division and frequent node failures, the problem of atomic commitment of distributed transactions has changed, and the two-phase commitment protocol (2PC) is no longer fit for it. Therefore, the authors describe the typical scene and redefine the problem. A consensus-based approach, called ODTrans, is proposed to deal with distributed commitment and replicas update separately. Compared with the traditional 2PC and the primary-backup scheme, it could make progress all the while in a cluster without the Byzantine Generals Problem. Experimental results demonstrate that proposed method not only could achieve performance close to the state-of-the-art method but also is able to tolerate the faults of both master and slaves.

Key words: distributed transaction, fault tolerance, cloud data store

摘要： 云数据存储系统中, 由于数据的分片、冗余存储以及频繁的节点故障, 导致分布式事务的原子提交问题发生变化, 使得原先的两阶段提交协议(2PC)不再适用。作者对云存储中事务提交的典型场景进行刻画, 重新定义分布式事务的提交问题, 并提出基于一致性协议的方法ODTrans, 用于处理分布式提交和副本更新。相比于传统的2PC 和主备复制方法, 在一个没有拜占庭将军问题的集群中, 它可以始终取得进展。实验结果表明, 所提方法可以获得与已有的方法相近的性能, 既可以容忍从节点发生错误, 也可以应对主节点的故障。

关键词: 分布式事务, 容错性, 云数据存储