Dynamic partition lock method to reduce transaction abort rates of cloud database

Abstract

Generally, existing cloud data management systems (CDMSs) support partition tolerance in mandatory and sacrifice consistency according to the CAP theory. Partition tolerance means that the database can be split over multiple servers to guarantee consistency or availability, even though the network is partitioned. This property is mandatory in CDMSs which consist of multiple servers. On the contrary, consistency of database is not strongly required in applications of CDMSs such as online shopping, search engine and so on. However, recently emerging applications such as online gaming, collaborative editing, and social networking require strong consistency, which requires that there be only one copy of each data item, and that only serializable access be permitted. Several approaches have been proposed to support consistency in CDMSs. These can be classified into three groups. In the first group of approaches, transactions are processed over multiple records in only one node. In the second group, limited wider transaction support is given to allow transactions to run in multiple nodes with some constraints. In the third group, transaction services are unbundled from the kernel of database management systems. In this paper, we propose a transaction processing method for CDMSs in which snapshot isolation (SI) techniques are applied to the unbundling transaction approach to increase its transaction throughput. The partition locking used by the unbundling transaction approach is static, and therefore, the number of partitions does not change, even when the number of records or the access frequency increases. This feature of the approach may increase the number of failed transactions when a SI technique is applied to the approach. We propose a dynamic partitioning lock method to solve these problems. Our proposed method monitors the number of transactions that access each partition and the number of records in each partition. Then, it splits any partition that is accessed more times than a particular threshold and maintains the split information using a KD-tree. The results of simulations conducted using our proposed dynamic partitioning lock methods show that less transaction aborts occur with our proposed method than with conventional methods.