Minimizing Detail Data in Data Warehouses

Data warehouses collect and maintain large amounts of data from several distributed and heterogeneous data sources. Because of security reasons, operational requirements, and technical feasibility it is often impossible for data warehouses to access the data sources directly. Instead data warehouses have to replicate legacy information as detail data in order to be able to maintain their summary data. In this paper we investigate how to minimize the amount of detail data stored in a data warehouse. More speci cally, we identify the minimal amount of data that has to be replicated in order to maintain, either incrementally or by recomputation, summary data de ned in terms of generalized project-select-join (GPSJ) views. We show how to minimize the number of tuples and attributes in the current detail tables and even aggregate them where possible. The amount of data to be stored in current detail tables is minimized by exploiting smart duplicate compression in addition to local and join reductions. We identify situations where it becomes possible to omit the typically huge fact table and prove that these techniques in concert ensure that the current detail data is minimal in the sense that no subset of it permits to accurately maintain the same summary data. Finally, we sketch how existing maintenance methods can be adapted to use the minimal detail tables we propose.