Efficient maintenance of materialized top-k views

摘要

We tackle the problem of maintaining materialized top-k views. Top-k queries, including MIN and MAX as important special cases, occur frequently in common database workloads. A top-k view can be materialized to improve query performance, but in general it is not self-maintainable unless it contains all tuples in the base table. Deletions and updates on the base table may cause tuples to leave the top-k view, resulting in expensive queries over the base table to "refill" the view. We propose an algorithm that reduces the frequency of refills by maintaining a top-k' view instead of a top-k view, where k' changes at runtime between k and some k/sub max//spl ges/k. We show that in most practical cases, our algorithm can reduce the expected amortized cost of refill queries to O(1) while still keeping the view small. The optimal value of k/sub max/ depends on the update pattern and the costs of querying the base table and updating the view. Compared with the simple approach of maintaining either the top-k view itself or a copy of the base table, our algorithm can provide orders-of-magnitude improvements in performance with appropriate k/sub max/ values. We show how to choose k/sub max/ dynamically to adapt to the actual system workload and performance at runtime, without requiring accurate prior knowledge.