Availability and performance limitations in multidatabases

摘要

There are several available paradigms to use when constructing the distributed systems of the future. Multidatabase systems (MDBSs) are becoming an increasingly popular paradigm: they offer a new promise for easing the construction of new distributed systems. Instead of developing a whole new distributed system, only a small portion needs to be developed, which integrates existing systems. We believe it is important to investigate the appropriateness of this approach to the construction of future distributed systems. In this paper we begin this investigation by focusing on one of the most important benefits of distributed systems: high availability. We study the problem of replication control in multidatabase systems in which members are autonomous. The requirement for local autonomy creates the need for a new model for replicating data in a MDBS. In this paper we first develop a model for replicated data in a MDBS. Subsequently, we formally characterize the classes of transactions which are allowed to commit, which allows us to formally derive availability upper bounds and define the performance limitations of the replication benefits for one-copy serializable (1-sr) MDBS transactions. We then discuss the requirements of protocols that attempt to provide optimal availability and ensure 1-sr. Finally, we examine some of the available commercial products and relate our formal models and results to them in an effort to show the applicability of the derived results. Because the availability and performance limitations are significant and the protocols that achieve high availability are costly, our results can be viewed in two ways. First, they show a way to replicate data in a 1-sr MDBS, define the optimal availability and discuss the requirements of protocols which achieve it. Second, they constitute a formal justification for researching alternative notions of correctness in a MDBS.