A New Generic Indexing Technology

There has been no fundamental change in the dynamic indexing methods supporting database systems since the invention of the B-tree twenty-five years ago. And yet the whole classical approach to dynamic database indexing has long since become inappropriate and increasingly inadequate. We are moving rapidly from the conventional one-dimensional world of fixed-structure text and numbers to a multi-dimensional world of variable structures, objects and images, in space and time. But, even before leaving the confines of conventional database indexing, the situation is highly unsatisfactory. In fact, our research has led us to question the basic assumptions of conventional database indexing. We have spent the past ten years studying the properties of multi-dimensional indexing methods, and in this paper we draw the strands of a number of developments together some quite old, some very new, to show how we now have the basis for a new generic indexing technology for the next generation of database systems. Multi-dimensional indexing There has been no fundamental change in the dynamic indexing methods supporting database systems since the invention of the B-tree twenty-five years ago. And yet the whole classical approach to dynamic database indexing has long since become inappropriate and increasingly inadequate. We are moving rapidly from the conventional one-dimensional world of fixed-structure text and numbers to a multi-dimensional world of variable structures, objects and images, in space and time. But, even before leaving the confines of conventional (i.e. relational) database indexing, the situation is highly unsatisfactory. In fact, our research has lead us to question the basic assumptions of conventional database indexing. The concept of a set of one-dimensional primary and secondary indexes, tuned to a particular query pattern, emerged in the days of menu-driven applications. This concept has not evolved in any way to match the flexibility of 4GL query languages. And today two new factors make such tuning increasingly impractical: internal system inferencing where even the system itself cannot predict the progression of queries and an increasing number of very large scale applications where the overhead of periodic re-indexing is unacceptable .