Random Sampling from Pseudo-Ranked B+ Trees
Gennady Antoshenkov · Very Large Data Bases · 1992
In the past, two basic approaches for sampling f5-om B+ trees have been suggested: sampling from the ranked trees and acceptance/rejection sampling i?om non-ranked trees. The first approach requires the entire root-to-leaf path to be updated with each insertion and deletion. The second has no update overhead, but incurs a high rejection rate for the compressed-key B+ trees commonly used in practice. In this paper we introduce a new sampling method based on pseudo-ranked B+ trees, which are B+ trees supplemented with information loosely describing the estimated rank limits. This new method exhibits a very small rejection rate while paying only a marginal cost of the tree update overhead. We also present comparative efficiency measurements of different methods that were run on production databases and on several prototype workload simulations.