Distinct Values Estimators for Power Law Distributions

The number of distinct values in a relation is an important statistic for database query optimization. As databases have grown in size, scalability of distinct values estimators has become extremely important, since a näıve linear scan through the data is no longer feasible. An approach that scales very well involves taking a sample of the data, and performing the estimate on the sample. Unfortunately, it has been shown that obtaining estimators with guaranteed small error bounds requires an extremely large sample size in the worst case. On the other hand, it is typically the case that the data is not worst-case, but follows some form of a Power Law or Zipfian distribution. We exploit data distribution assumptions to devise distinct-values estimators with analytic error guarantees for Zipfian distributions. Our estimators are the first to have the required number of samples depend only on the number of distinct values present, D, and not the database size, n. This allows the estimators to scale well with the size of the database, particularly if the growth is due to multiple copies of the data. In addition to theoretical analysis, we also provide experimental evidence of the effectiveness of our estimators by benchmarking their performance against previously best known heuristic and analytic estimators on both synthetic and real-world datasets.