Advances in the design of the BANG file

Abstract

A recent paper described a new kind of grid file — the BANG file — which, unlike previous grid file designs, has the important property that the directory never expands faster than the data, whatever the dimensionality of the file or the form of the data distribution. Experience with the first implementation of the BANG file has given new insights into its properties and behaviour, which have enabled significant improvements in performance to be obtained without compromising the conceptual simplicity of the original design:•access to an individual tuple now always succeeds in a single pass down the directory tree, as with a B-tree;•overflow or underflow of data or directory pages never leads to more page splits or merges than in a B-tree of the same directory depth;•every direct representation of a sub-space of the data space is guaranteed to be a minimal representation — substantially improving the efficiency of range searches, partial match searches and joins. The paper begins with a review of the original design in the light of its implementation. A detailed description is then given of the enhancements which have been found possible, with the theoretical foundations on which they depend. Finally, a selection of performance figures are given to demonstrate its characteristics.