Multi-criteria: Degrees of recoverability in distributed databases

Abstract

This paper concerns distributed databases and transactions with a distinction between global and local correctness criteria. The global requirements per system are weaker than the local requirements per site. The paper investigates an application which suits such a two-level division. The main motivation for our investigation is based on the fact that the commonly used correctness criteria for concurrency control and recovery, serializability and total recoverability, are very strict criteria. The use of more relaxed criteria (allowing more true parallel behaviour and more true partial behaviour) is therefore very appealing — as long as this can be achieved without compromising safety or applicability. The main paradigm in our approach is based on the observation that relatively little knowledge about the databases and transactions can lead to major gains in system throughput. This allows specific systems to have more tailor-made correctness criteria.

[Nygå93a] introduced and analysed a 2-level model for non-serializability with distributed databases and transactions, while [Nygå93b] presented and discussed a system which suited such a 2-level division. Here the primary goal is to analyse and discuss n-level partial recoverability with distributed databases and transactions. The secondary goal is to integrate this with the 2-level model/system introduced in the above mentioned papers. Hence we both present a model and apply it to the given system.

We analyse the span between total recoverability per site and total recoverability per system. This requires local total recoverability but allows global partial recoverability. We discuss both single-level and multi-level recovery criteria. We relate and compare the resulting concepts to other established and proposed criteria. Our notions are brand new. We consider step-wise committing transactions. Our main point is not that partial commitment is being used, but rather to discuss how it should be controlled. The main result is a set of recovery rules which may be combined in an orthogonal way. The concepts and notions seem especially appropriate for systems/models which make use of added priority ruling.

The appropriate choices with respect to recovery for the application from [Nygå93b] corresponds to a 2-level criterion — with one part stemming from the existence of intra-site integrity-constraints and value-dependencies and the other part stemming from the superimposition of priority which originates from the designated criteria for concurrency control. Effectively, knowledge about some missing integrity constraints in a distributed database opens the way for non-serializability and partial recoverability globally — and knowledge about some existing overall goals of distributed transactions leads the way to the add-on to serializability and total recoverability locally.