An Information-Based Model for Failure-Handling in Distributed Database Systems

We consider the failure atomicity problem of distributed transactions in conjunction with the maximization of database availability. We propose a new information-based model for the distributed transaction-execution, which explicitly expresses the information at each stage during a protocol. In addition to rederiving certain existing results, we prove a fundamental relation among the site failures and the network partitioning. We propose a realistic model for site failures under which we show that the costs of commit and termination protocols can be greatly reduced. Finally, we explore the possible recovery strategies for a failed site and show how they are improved under our site failure model.