Data replication is employed in distributed database systems to enhance data availability and performance. However, the benefit of data replication is only realized at the cost of maintaining the consistency of data. In particular, network partition failures in distributed systems make it more difficult to achieve high data availability while ensuring strong correctness criteria such as 1SR. In this thesis, we propose a data replication method to improve the availability of data in the presence of network partition failures by using the relaxed correctness criteria, which is called insular consistency.
Our method adopts and extends the traditional primary copy method for large-scale distributed systems, where partition failures frequently occur. An asynchronous update propagation mechanism is also employed to improve the performance of update operations. We focus on increasing the availability of data for two particular classes of transactions. One is the read-only transactions which are allowed to read old values, and the other is the update transactions whose write operations are all blind-writes. In our method, a mechanism is presented that makes these transactions to be executed at any partition as long as insular consistency is satisfied. We show that our mechanism guarantees the correctness criteria.
