SAFE: Sampling-Assisted Fast Learned Cardinality Estimation for Dynamic Spatial Data

Yuchen Ji, Daichi Amagata, Yuya Sasaki, Takahiro Hara

Abstract: Cardinality estimation for spatial queries plays an important role in query scheduling and optimization. Spatial datasets are fully dynamic, and this setting necessitates an update-friendly, low-latency, and accurate cardinality estimator. However, existing cardinality estimation methods suffer from time-consuming updates and/or inefficient estimation. This work proposes SAFE (Sampling-Assisted Fast learned cardinality Estimator), which is carefully designed for dynamic spatial data. We specifically develop a sampling strategy that uses a quad-tree-based data partitioning and extracts a small subset, to enable fast training of cardinality estimation models. In addition, we employ 2-tier regression models to approximate the spatial data distribution while achieving accurate and fast cardinality estimation. We furthermore provide an incremental model update strategy to avoid re-training all models from scratch when we receive updates. We conduct experiments on real and synthetic datasets. Their results demonstrate that SAFE (i) outperforms state-of-the-art cardinality estimation models and (ii) efficiently handles data updates while ensuring accurate and low-latency estimation.