Bringing Vector Search to OLTP

Google Cloud SQL for MySQL now features a fully-transactional and high-performance Approximate Nearest Neighbor (ANN) search capability integrated with the ScaNN library. This enables semantic search on existing data without the need for separate vector databases or complex external pipelines.

Key technical innovations include:

Deconstructed On-Disk Persistence: Instead of monolithic in-memory objects, ScaNN indices are persisted in standard InnoDB tables. This supports incremental updates, crash safety, and uses a hierarchical algorithm to reduce partitions scanned by 90%.

Model Endpoint Management: SQL functions allow users to generate vector embeddings directly within the database through integrations with Vertex AI and 3rd-party models like OpenAI and Anthropic. SQL-native ANN Pushdown: The APPROX_DISTANCE function is integrated into standard SQL syntax. Iterative scanning logic is pushed to the storage engine to optimize performance for filtered queries through result caching.

This solution provides a seamless vector search experience while maintaining the transactional guarantees and operational simplicity of MySQL.