Maintaining consistency in distributed systems

• Authors: Kenneth P. Birman
• Venue: EW 5 Proceedings of the 5th workshop on ACM SIGOPS European workshop: Models and paradigms for distributed systems structuring
• Year: 1992

The emerging generation of database systems and general purpose operating systems share many characteristics: object orientation, a stress on distribution, and the utilization of concurrency to increase performance. A consequence is that both types of systems are confronted with the problem of maintaining the consistency of multi-component distributed applications in the face of concurrency and failures. Moreover, large applications can be expected to combine database and general purpose components. This paper reviews four basic approaches to the distributed consistency problem as it arises in such hybrid applications:• Transactional serializability, a widely used database execution model, which has been adapted to distributed and object-oriented settings by several research efforts.• Traditional operating systems synchronization constructs, such as monitors, used within individual system components, and with no system-wide mechanism for inter-object synchronization.• Linearizability, an execution model for object-oriented systems with internal concurrency proposed by Herlihy and Wing [HW90] (similarly restricted to synchronization within individual objects).• Virtual synchrony, a non-transactional execution model used to characterize consistency and correctness in groups of cooperating processes (or groups of objects, in object-oriented systems) [BJ87].We suggest that no single method can cover the spectrum of issues that arise in general purpose distributed systems, and that a composite approach must therefore be adopted. The alternative proposed here uses virtual synchrony and linearizability at a high level, while including transactional mechanisms and monitors for synchronization in embedded subsystems. Such a hybrid solution requires some changes to both the virtual synchrony and transactional model, which we outline. The full-length version of the paper gives details on this, and also explores the problem in the context of a series of examples.The organization of the presentation is as follows. We begin by reviewing the database data and execution models and presenting the transactional approach to concurrency control and failure atomicity. We then turn to distributed systems, focusing on aspects related to synchronization and fault-tolerance and introducing virtually synchronous process groups. The last part of the paper focuses on an object oriented view of distributed systems, and suggests that the linearizability model of Herlihy and Wing might be used to link the virtual synchrony approach with transactions and "internal" synchronization mechanisms such as monitors, arriving at a flexible, general approach to concurrency control in systems built of typed objects. We identify some technical problems raised by this merging of models and propose solutions.