Optimizing chain queries in a distributed database system.

Reviewer: Mark Robbin Brown

The paper considers the problem of evaluating a relational query consisting of the join of N distinct relations, projected to the columns of the first relation. The main result of the paper is an efficient algorithm that, given such a query, finds the lowest-cost program for it that consists of a sequence of semijoins. The semijoin of two relations is just the join of the two relations, projected to the columns of the first relation. Semijoins may be interesting in distributed systems because to compute a semijoin, only the joining column of the second relation needs to be communicated to the site of the first relation. Perhaps the most surprising result of the paper is that in some cases the optimal program is quite long (containing roughly N :9I 2/2 semijoins). It is quite difficult to produce mathematical results that are relevant to the problem of computing distributed queries on a real database. The impact of this paper is lessened in several ways, some of which are mentionaled by the authors in their concluding section. The paper focuses on queries that perform joins in order to select data from a single relation, but most programs perform joins in order to combine data from several tables. The paper focuses on semijoin programs, so the optimizer produces programs that do not use joins even when this would reduce the total cost. The cost measure is an estimate of total communication time, and it does not account for processing and disk I/O time or for potential parallel communication. Relations are assumed to be stored at distinct sites. Finally, the optimization algorithm requires knowledge of the database that is expensive to obtain and is invalidated by database updates. This need for detailed information may be inherent in algorithms that optimize performance against an arbitrary database, as opposed to those which optimize in some average sense given incomplete information about the database or those which optimize per instance over a restricted set of databases.