# SPANN

> SPANN is a highly efficient billion-scale ANN search system using clustered HNSW indexes with dynamic partitioning for balanced load. Key features: disk-based, high recall, low latency on commodity hardware. Use cases: web-scale recommendation, image retrieval. Improves on DiskANN with better build time; competitive FAISS GPU in CPU perf.

- **URL:** https://www.microsoft.com/en-us/research/publication/spann-highly-efficient-billion-scale-approximate-neighborhood-search/
- **Category:** research-papers-surveys
- **Tags:** disk-ann, clustered-hnsw, Billion Scale
- **Updated:** 2026-04-22 22:18
- **Canonical page:** https://vectordb.ever.works/items/spann

## Details

## Overview

SPANN (Highly-efficient Billion-scale Approximate Nearest Neighbor Search) is a memory-disk hybrid vector indexing and search system developed by Microsoft Research that follows the inverted index methodology.

## Architecture

SPANN stores only the centroid points of posting lists in memory while putting the large posting lists on disk. This hybrid approach enables efficient billion-scale vector search with reduced memory requirements.

## Performance

- Significantly outperforms DiskANN in both recall@1 and recall@10
- More than 2x faster than DiskANN, especially in low query latency budgets (less than 4ms)
- Requires more than 30% of the dataset size in memory for graph indexing

## Technical Details

SPANN is an on-disk cluster-based index that stores clusters on disk and maintains a small graph in memory to index the nearest clusters. The system is designed for applications requiring billion-scale vector search with memory efficiency.

## Research

Published by Microsoft Research, SPANN represents a significant advancement in disk-based vector indexing technologies and serves as a benchmark for hybrid vector search systems.

---

_This Markdown mirror is generated by Ever Works for AI agents. The canonical HTML page is at https://vectordb.ever.works/items/spann._