Multi-level transaction management, theoretical art or practical need?

A useful approach to the design and description of complex data management systems is the decomposition of a system into a hierarchically organized collection of levels. In such a system, transaction management is distributed among the levels. This paper presents the fundamental theory of multi-level concurrency control and recovery. A model for the computation of multi-level transactions is introduced by generalizing from the well known single-level theory. Three basic principles, called commutation, reduction, and abstraction are explained. Using them enables one to explain and prove seemingly ”tricky” implementation techniques as correct, by regarding them as multi-level algorithms. We show how the theory helps to understand and explain in a systematic framework techniques that are in use in today's DBMSs. We also discuss how and why multi-level algorithms may achieve better performance than single-level ones.