What Is DiskANN? Billion-Scale Vector Search Explained

DiskANN is a graph-based approximate nearest neighbor (ANN) algorithm from Microsoft Research (NeurIPS 2019) that enables billion-scale vector search by storing indexes on SSD rather than RAM. Built on the Vamana directed graph algorithm combined with product quantization (PQ), it keeps only compressed vectors in RAM for fast routing while reading full-precision vectors from SSD for final reranking. This achieves 95%+ recall with sub-5ms latency on 1B vectors using just 64GB RAM — 5-10x more vectors per machine than DRAM-only solutions like HNSW. The post covers how DiskANN works, compares it to HNSW and IVF, details tuning parameters (MaxDegree, SearchListSize, BeamWidthRatio), hardware sizing guidelines, and lists databases that support it including Couchbase 8.0, Azure Cosmos DB, Milvus, and pgvectorscale. FreshDiskANN extends the algorithm to support real-time updates without full index rebuilds.

Source: https://www.couchbase.com/blog/diskann. 8 Sync News only summarizes and links out; content copyright belongs to the authors and original sources.