Quake: Adaptive Indexing for Vector Search
Abstract
Vector search, the task of finding the k-nearest neighbors of high-dimensional vectors, underpins many machine learning applications, including recommendation systems and information retrieval. However, existing approximate nearest neighbor (ANN) methods perform poorly under dynamic, skewed workloads where data distributions evolve. We introduce Quake, an adaptive indexing system that maintains low latency and high recall in such environments. Quake employs a hierarchical partitioning scheme that adjusts to updates and changing access patterns, guided by a cost model that predicts query latency based on partition sizes and access frequencies. Quake also dynamically optimizes query execution parameters to meet recall targets using a novel recall estimation model. Furthermore, Quake utilizes optimized query processing, leveraging NUMA-aware parallelism for improved memory bandwidth utilization. To evaluate Quake, we prepare a Wikipedia vector search workload and develop a workload generator to create vector search workloads with configurable access patterns. Our evaluation shows that on dynamic workloads, Quake achieves query latency reductions of 1.5-22x and update latency reductions of 6-83x compared to state-of-the-art indexes SVS, DiskANN, HNSW, and SCANN.