LASE: A Learned Spatial Index for Dynamic Workloads

Recently, several learned spatial indexes have been proposed to improve fast spatial queries, however, they are ineffective when facing highly dynamic scenarios with different read/write ratios. In this paper, we propose LASE, a Learned spatial index for dynamic workloads. We first empirically demonstrate that the poor performance for dynamic workloads comes from high-cost model training. Therefore, LASE takes advantage of both the traditional indexing techniques and the learned indexes to reduce model retraining costs. Specifically, LASE designs effective structural modification operations to support update operations and develops efficient query algorithms. Two aspects are considered in reducing update costs: 1) reduce the scale of model retraining; 2) achieve better locality to accelerate the query performance within leaf nodes. We conduct extensive experiments on various dynamic workloads and different datasets, and the experimental results demonstrate that LASE achieves high-performance updates and queries compared with applicable competitors.