A temporal database management system augments traditional database management systems with the ability to record the time-varying aspects of information. This paper continues the authors’ study of optimization techniques for temporal queries. Specifically, the paper examines the execution of event-join operations, that is, join operations used to group temporal attributes that are stored in separate relations. This paper (and others by the same authors) is significant since it is among the first to address issues in temporal database implementation.
The authors begin by defining the event-join of two temporal relations, r 1 and r 2, as the union of two simpler joins, the temporal equijoin and the temporal outerjoin. Four algorithms are exhibited for computing an event-join. Each algorithm exploits different assumptions about the characteristics of r 1 and r 2. The authors investigate the relative cost of each algorithm under these assumptions. The authors show that for append-only databases (temporal databases ordered on starting timestamp only) event-joins can be computed as efficiently as when both relations are completely sorted on both join key and starting timestamp.
The paper introduces a concept that appears to be fundamental to temporal join processing, the notion of tuple covering. During the join process, the tuple covering of a tuple t ∈ r determines the intervals of t that have already been joined with previously scanned tuples. The algorithms described in the paper attempt to efficiently compute the tuple coverings for each tuple in the joining relations.
Another contribution of the paper is the development of a new index structure optimized for append-only relations. This data structure, the append-only tree, combines characteristics of both B+-trees and ISAM indexes and exploits the time-ordered insertion property of append-only databases.
I recommend this paper for readers interested in temporal databases or query processing. Before starting this paper, it may be useful to read some of the authors’ previous papers (listed in the references) to become familiar with their particular data model and terminology.
