How to upgrade embedding models in AlloyDB AI with zero downtime

AlloyDB AI introduces a zero-downtime process for upgrading embedding models in production environments, avoiding the need to pause user queries. The method addresses challenges such as re-embedding entire datasets, rebuilding vector indexes, and managing schema migrations without disrupting service. Traditional upgrades often require dual-write strategies and careful coordination to prevent search degradation due to mathematical incompatibility between old and new embeddings.

The upgrade leverages AlloyDB AI’s features, including model endpoint management for in-database access to any embedding model, and auto embeddings that generate new embeddings in the background at up to 3000 per second. The process uses AlloyDB’s ScaNN index, which builds indexes up to 10 times faster than open-source HNSW, and PostgreSQL-native tools like concurrent index creation and atomic column renames to enable transparent blue/green or canary rollouts.

A practical example demonstrates migrating a production table from text-embedding-005 to gemini-embedding-001 while serving live traffic. The process involves adding a nullable column without default values to update catalog metadata instantly, setting a lock_timeout to 2 seconds to prevent blocking, and using manual mode for auto embeddings to maintain write performance while controlling embedding refresh frequency.

After generating embeddings, the process validates completeness and correctness before traffic cutover, accounting for concurrent modifications. Benchmarks show the new gemini-embedding-001 model improves semantic fidelity, with only 31% overlap in top-10 nearest neighbor results compared to the previous model. A ScaNN index is then built concurrently to enable efficient filtered vector search, with construction times up to 10 times faster than HNSW on open-source PostgreSQL.