Rigorous Design of Distributed Transactions.
Univsersity of Southampton, Southampton, UK, Dependable Systems and Software Engineering, School of Electronics and Computer Science,
Database replication is traditionally envisaged as a way of increasing fault-tolerance and availability. It is advantageous to replicate the data when transaction workload is predominantly read-only. However, updating replicated data within a transactional framework is a complex affair due to failures and race conditions among conflicting transactions. This thesis investigates various mechanisms for the management of replicas in a large distributed system, formalizing and reasoning about the behavior of such systems using Event-B. We begin by studying current approaches for the management of replicated data and explore the use of broadcast primitives for processing transactions. Subsequently, we outline how a refinement based approach can be used for the development of a reliable replicated database system that ensures atomic commitment of distributed transactions using ordered broadcasts. Event-B is a formal technique that consists of describing rigorously the problem in an abstract model, introducing solutions or design details in refinement steps to obtain more concrete specifications, and verifying that the proposed solutions are correct. This technique requires the discharge of proof obligations for consistency checking and refinement checking. The B tools provide significant automated proof support for generation of the proof obligations and discharging them. The majority of the proof obligations are proved by the automatic prover of the tools. However, some complex proof obligations require interaction with the interactive prover. These proof obligations also help discover new system invariants. The proof obligations and the invariants help us to understand the complexity of the problem and the correctness of the solutions. They also provide a clear insight into the system and enhance our understanding of why a design decision should work. The objective of the research is to demonstrate a technique for the incremental construction of formal models of distributed systems and reasoning about them, to develop the technique for the discovery of gluing invariants due to prover failure to automatically discharge a proof obligation and to develop guidelines for verification of distributed algorithms using the technique of abstraction and refinement.
Actions (login required)