Simple Random Sampling from Relational Databases

Sampling is a fundamental operation for the auditing and statistical analysis of large databases. It is not well supported in existing relational database management systems. We discuss how to obtain samples from the results of relational queries without first performing the query. Specifically, we examine simple random sampling from selections, projections, joins, unions, and intersections. We discuss data structures and algorithms for sampling, and their performance. We show that samples of relational queries can often be computed for a small fraction of the effort of computing the entire relational query, i.e., in time proportional to sample size, rather than time proportional to the size of the full result of the relational query.