AlloyDB supercharges PostgreSQL vector search with accuracy, speed, and 1B+ scale

In our 20 years of experience integrating AI into real-world applications an important theme emerges: the key to building enterprise gen AI applications is having a trustworthy, scalable data foundation that supports the scale and performance needs of the largest workloads. When you’re building a gen AI or search application, you need high-quality results in real time, grounded in your application’s latest data. And you need solutions that are easy for developers to build and manage, and low-cost to run at scale.

Today, we’re announcing the general availability of the ScaNN for AlloyDB index to help enterprises build scalable,performant, and accurate gen AI and search applications. AlloyDB for PostgreSQL is our fully managed PostgreSQL-compatible service for demanding enterprise workloads, including 4x faster transactional workloads and up to 100x faster analytical queries than standard PostgreSQL. Now, the ScaNN index brings 12 years of Google research — the same technology used at Google to power some of our billion+ user services — to AlloyDB for PostgreSQL to supercharge vector database workloads as well.

With the ScaNN index, AlloyDB combines the performance of a vector-optimized database, delivering accurate results in record time and scaling to support more than a billion vectors — with the query flexibility, enterprise features, and multi-workload capabilities of a relational database. 

The ScaNN for AlloyDB index

The ScaNN for AlloyDB index is the first PostgreSQL-compatible index that can scale to support more than a billion vectors while maintaining state-of-the-art query performance across the whole spectrum, delivering high performance at <25ms p95 latency at 1B vectors with 95% recall. It provides up to 4x faster vector queries than the HNSW index in standard PostgreSQL. Moreover, the ScaNN index in AlloyDB offers important benefits for real-world workloads, notably fast index build times — to improve developer productivity and reduce expensive bottlenecks for large-scale workloads. It also has efficient memory usage — typically 3-4x less memory than the HNSW index in standard PostgreSQL — to help larger workloads run on smaller machines and improve performance for hybrid workloads. Finally, AlloyDB supports indices that are much larger than what can fit in main memory, allowing customers to right-size for their specific workloads. Together, these improvements vastly expand the set of vector database workloads that can efficiently run on AlloyDB while reducing the need for specialized tools.

At the same time, AlloyDB offers all the benefits of a commercial-grade relational database. Full PostgreSQL compatibility gives developers a simple and flexible query language that can be used to query vectors and scalar data together, supporting everything from simpler filters to complex joins to hybrid searches on a single database. AlloyDB supports full transactional consistency and high write throughput. In fact, the ScaNN index in AlloyDB has up to 10x higher write throughput than the HNSW index in standard PostgreSQL, so customers with very high update rates can rest assured that writes will be performant and readily available for querying. Finally, AlloyDB has a customer-friendly 99.99% availability SLA that includes maintenance, and critical enterprise capabilities around high availability, disaster recovery, data protection, security, and more.

The new ScaNN index for AlloyDB is fully PostgreSQL- and pgvector-compatible, making this technology easily accessible to the professional developers that now use the most popular open source database, PostgreSQL. 

Next steps

To learn more about the ScaNN for AlloyDB index, check out our overview of the ScaNN for AlloyDB index, or read our ScaNN for AlloyDB whitepaper for a deep dive into Google’s ScaNN algorithm and how we’ve implemented it in PostgreSQL and AlloyDB.

ScaNN for AlloyDB is now generally available in AlloyDB in Google Cloud. Follow our quickstart guide to creating an AlloyDB instance, then follow our documentation to get started with easy and fast vector queries. You can also now try AlloyDB for free with our 30-day free trials.