DiskANN is a graph-based algorithm designed for approximate nearest neighbor (ANN) search on datasets too large to fit in memory. It addresses the challenge of efficiently searching high-dimensional data (like vectors from images or text embeddings) by storing most of the index on disk while keeping a small, frequently accessed portion in memory. Unlike in-memory ANN methods (e.g., HNSW or FAISS), which require the entire index to reside in RAM, DiskANN leverages disk storage to handle massive datasets. The core idea is to organize data into a graph structure where nodes represent data points, and edges connect each node to its nearest neighbors. During a search, the algorithm navigates this graph, loading only the necessary nodes from disk into memory as needed, minimizing costly disk reads.
The algorithm works by first constructing a graph where each node is linked to its approximate nearest neighbors. During a query, the search starts at one or more entry points and traverses the graph, comparing the query vector to nearby nodes. Since the graph is stored on disk, DiskANN uses a caching mechanism to keep recently accessed nodes in memory, reducing repeated disk access. For example, when searching a billion-scale image dataset, DiskANN might load 1% of the graph into memory initially, then fetch additional nodes from disk as the traversal progresses. This approach ensures that even with limited RAM, the system can handle large datasets by trading off some speed (due to disk latency) for scalability. Optimizations like prioritizing high-degree nodes in the cache or using SSDs for faster disk access further improve performance.
DiskANN balances accuracy, speed, and memory usage through configurable parameters. For instance, increasing the number of edges per node improves search accuracy but requires more disk space and I/O operations. Developers can tune the cache size to match available memory: a larger cache reduces disk reads but uses more RAM. A practical use case is recommendation systems where embedding datasets exceed terabytes—DiskANN allows querying these without requiring expensive RAM-heavy servers. However, its performance depends on disk speed, making it less suitable for real-time applications unless paired with fast storage. By focusing on disk-friendly graph traversal and smart caching, DiskANN enables scalable ANN search for datasets that would otherwise be impractical to handle with traditional in-memory methods.