An implementation of the M-tree index structure for PostgreSQL using GiST

Several index structures are competing for efficient operations on various types of data, but there are some forms of information that cannot fit into the existing models, because by forcing them into one of the currently available representations they either lose some of their value or they do not have the necessary properties which would allow them to be used in the first place. In light of this, we implemented the M-tree [1] [2] index structure under the PostgreSQL database management system through a GiST [3] extension to facilitate efficient range-based queries and k-nearest neighbor searches on data that resides only in a metric space. The M-tree structure has parts that require heuristics, of which the implementations are not entirely specified, only suggestions are given. This allows flexibility for adapting to different situations and leaves space for further improvements. We implemented several of the strategies proposed in the original paper alongside some of our own ideas and saw significant speed gains when compared to a sequential scan, our main reference, as the other built-in index structures are not applicable to metric spaces.