A thin monitoring layer for top-k aggregation queries over a database

Abstract

We consider the problem of maintaining a large set of top-k rankings over the update stream of a database. The rankings stem from top-k aggregation queries that are given a-priori based on the application scenario, for instance created along dimensions of a traditional data warehouse, for efficient automated reporting/detection of changes. The focus on only the top part of a ranking enables efficient buffering techniques to limit expensive interactions with the underlying database, while still guaranteeing correct top-k rankings at all times. This is achieved by employing conservative rank (score) estimates of previously unseen items that are not in the top-k result so far. The proposed family of maintenance algorithms further exploits the relations between the monitored rankings known from multi-query optimisation. We present results of a preliminary experimental evaluation using TPC-H data to study the performance of our algorithms.