DGAI: Decoupled On-Disk Graph-Based ANN Index for Efficient Updates and Queries

2025-10-29

AI Summary

Existing disk-based graph indexes for approximate nearest neighbor (ANN) search suffer from high redundant I/O and low update efficiency due to tightly coupled vector storage and graph topology. To address this, we propose a decoupled disk-resident graph index architecture that separates vector storage from graph topology, and introduce a three-stage query processing mechanism along with an incremental, page-level topological reordering strategy. We further innovate by integrating multi-level product quantization (PQ) for compressed vector filtering, jointly reducing both I/O and computational overhead. Experimental results demonstrate that, compared to conventional coupled architectures, our approach achieves 10.05× and 6.89× speedups in insertion and deletion throughput, respectively, and improves query throughput by 2.66×—all while maintaining strong query accuracy and latency performance.

Technology Category

• Search and Optimization: Distributed Search
• Machine Learning: Graph-based Machine Learning
• Data Mining & Knowledge Management: Intelligent Query Processing

Application Category

• Graph Algorithms and Modeling for the Web: Querying, indexing, and retrieval in Web-related graphs
• Search and Retrieval-Augmented AI: Web crawling and indexing
• Semantics and Knowledge: Scalable techniques for the creation, curation, publication, maintenance, and consumption of large, Web-based, structured, reusable, knowledge graphs and ontologies

Abstract

On-disk graph-based indexes are widely used in approximate nearest neighbor (ANN) search systems for large-scale, high-dimensional vectors. However, traditional coupled storage methods, which store vectors within the index, are inefficient for index updates. Coupled storage incurs excessive redundant vector reads and writes when updating the graph topology, leading to significant invalid I/O. To address this issue, we propose a decoupled storage architecture. While a decoupled architecture reduces query performance. To overcome this limitation, we design two tailored strategies: (i) a three-stage query mechanism that leverages multiple PQ compressed vectors to filter invalid I/O and computations, and (ii) an incremental page-level topological reordering strategy that incrementally inserts new nodes into pages containing their most similar neighbors to mitigate read amplification. Together, these techniques substantially reduce both I/O and computational overhead during ANN search. Experimental results show that the decoupled architecture improves update speed by 10.05x for insertions and 6.89x for deletions, while the three-stage query and incremental reordering enhance query efficiency by 2.66x compared to the traditional coupled architecture.

Problem

• Improving update efficiency in on-disk graph-based ANN indexes
• Reducing query performance degradation in decoupled storage architectures
• Minimizing I/O and computational overhead during nearest neighbor search

Innovation

• Decoupled storage architecture separates vectors from index
• Three-stage query mechanism filters invalid I/O and computations
• Incremental page-level reordering inserts nodes near similar neighbors