Composition of Database Relations

The “standard” method of implementing composition is to perform join and projection as one combined operation and then to remove the duplicates from the result of this operation as a separate step. When a hash-based or sort-merge algorithm is used for the join operation and the size of the result is large, duplicates cannot be eliminated at the same time the join is being performed. In this paper, we argue for implementing composition as a primitive operation, and present a single-sided composition algorithm that performs join, projection, and duplicate elimination as one unilied operation. We also present experimental results that show an operating region in which this algorithm outperforms composition by the standard method. This operating region is characterized by a join result many times larger than the source relations, and many duplicates after projection over the non-join attributes. This occurs, for example, in deductive databases when computing transitive closures of dense graphs.