SPFresh: Incremental In-Place Update for Billion-Scale Vector Search

TL;DR

SPFresh introduces an in-place update system for billion-scale vector search that significantly reduces latency and resource usage compared to traditional global rebuild methods, enabling efficient, accurate, and scalable vector index updates.

Contribution

The paper presents SPFresh, a novel system with LIRE protocol for low-overhead, in-place vector index updates supporting billion-scale data, improving efficiency and accuracy.

Findings

SPFresh achieves lower latency than global rebuild methods.

Uses only 1% of DRAM and less than 10% of cores at peak.

Supports 1% daily vector update rate at billion scale.

Abstract

Approximate Nearest Neighbor Search (ANNS) is now widely used in various applications, ranging from information retrieval, question answering, and recommendation, to search for similar high-dimensional vectors. As the amount of vector data grows continuously, it becomes important to support updates to vector index, the enabling technique that allows for efficient and accurate ANNS on vectors. Because of the curse of high dimensionality, it is often costly to identify the right neighbors of a single new vector, a necessary process for index update. To amortize update costs, existing systems maintain a secondary index to accumulate updates, which are merged by the main index by global rebuilding the entire index periodically. However, this approach has high fluctuations of search latency and accuracy, not even to mention that it requires substantial resources and is extremely time-consuming for rebuilds. We introduce SPFresh, a system that supports in-place vector updates. At the heart of SPFresh is LIRE, a lightweight incremental rebalancing protocol to split vector partitions and reassign vectors in the nearby partitions to adapt to data distribution shift. LIRE achieves low-overhead vector updates by only reassigning vectors at the boundary between partitions, where in a high-quality vector index the amount of such vectors are deemed small. With LIRE, SPFresh provides superior query latency and accuracy to solutions based on global rebuild, with only 1% of DRAM and less than 10% cores needed at the peak compared to the state-of-the-art, in a billion scale vector index with 1% of daily vector update rate.