Sizing Access Structures for Data Warehouses

Data warehousing is an emergent important data management technology having a substantial impact on enterprise decision support and being used in many industries {recent estimates place the market size at $9 billion for 1997. Data warehouses are orders of magnitude larger than conventional databases and are typically subjected to a mix of complex, ad-hoc queries. These queries often access signiicant fractions of the underlying data yet must be answered interactively. Several indexing schemes have been proposed to improve warehouse query performance, namely, value-list, projection and bit-sliced indexes. These access structures are currently being ooered by many of the major warehouse system vendors including Sybase, Oracle and Red Brick. Because of the large size of the data warehouses, however, these indexes come with very high costs (sizes). Thus, choosing to create the wrong index could potentially have very substantial penalties. To date, no work has been done on index selection, even though it is an often crucial step of the physical design process. In this paper, we develop exact cost models of the above warehouse indexes to help guide the practitioner through this process. Based on these cost models, we also perform an in-depth computational study of these indexes, which yields signiicant theoretical insights into the nature and behavior of these indexing schemes.