Starling: An I/O-Efficient Disk-Resident Graph Index Framework for High-Dimensional Vector Similarity Search on Data Segment

03/26/24

Optimizing Vector Databases: A Segment-Level Architecture Guide

As high-dimensional vector data becomes central to AI and machine learning applications, vector databases face increasing pressure to efficiently manage unstructured data like images, text, and video. Traditional architectures struggle when a single machine must handle multiple data segments, each constrained by strict memory and disk space limitations.

This paper introduces Starling, a framework that revolutionizes vector database segment handling through a novel dual-component architecture: combining a streamlined in-memory navigation graph with a locality-optimized disk-based graph. Unlike current disk-based solutions that either require excessive storage or suffer from high latency, Starling achieves an optimal balance between search performance, accuracy, and resource utilization.

The framework demonstrates that remarkable performance improvements are possible through intelligent segment-level optimization, achieving 43.9× higher throughput and 98% lower query latency compared to existing methods. This paper provides essential insights for organizations seeking to scale their vector database operations while maintaining high performance and accuracy.

Get the Whitepaper