AlloyDB boosts pgvector HNSW vector search speed by 4x

AlloyDB now offers columnar engine accelerated HNSW, a feature designed to significantly improve vector search performance for pgvector. This enhancement provides up to 4x higher QPS compared to standard PostgreSQL HNSW, while maintaining or improving search recall.

The columnar engine stores frequently accessed index data in an optimized, in-memory columnar format. This bypasses standard PostgreSQL buffer manager overhead and uses a vectorized memory layout for faster traversals, leading to substantial QPS and recall gains.

This feature allows for reduced infrastructure costs by achieving higher performance with fewer compute resources. It also enables AI applications to attain better search accuracy at speeds previously only possible for lower-quality searches, with no application code modifications needed.

Beyond HNSW, AlloyDB also supports ScaNN, a Google Research-backed index, and standard k-nearest neighbor (KNN) search for use cases requiring absolute precision and 100% recall.

To utilize this feature, users need to enable the google_columnar_engine.enabled and google_columnar_engine.enable_index_caching flags. After creating an HNSW index via pgvector, a specific SQL command caches the index within the columnar engine.