Google Cloud adds four-level AlloyDB ScaNN preview

Google Cloud has added a four-level tree index to AlloyDB ScaNN in preview, allowing the service to handle more than 10 billion vectors.

The update addresses a scaling problem for database systems used in vector search as AI applications grow. Earlier ScaNN tree structures in AlloyDB were limited to two-level and three-level configurations, which became harder to scale because of rising compute demands and memory constraints.

AlloyDB is Google Cloud's managed PostgreSQL-compatible database service. In this release, the main technical change is a deeper tree structure for the ScaNN index used to organise and search vector data.

How it works

The four-level design uses hierarchical partitioning to reduce the number of vectors scanned during a query. Instead of searching across a broad space, the system narrows the search path through additional layers.

Google Cloud described the progression in terms of search complexity. A two-level tree uses coarse partitioning with a search complexity of O(N1/2), while a three-level tree narrows exploration to O(N1/3). The new four-level tree reduces that to O(N1/4), which Google Cloud said is enough to support workloads above 10 billion vectors.

To limit any drop in recall, the system also uses several supporting methods within the new structure, including Top-K branch, SOAR, centroid adjustment and a balanced tree shape.

Memory limits

Memory use was another constraint at this scale. The sampling processes needed to build indexes for 10 billion vectors can exceed available system memory, so the release includes changes intended to make training and construction more efficient.

One change is a balanced tree configuration designed to work with smaller sampling sets while still producing useful partitions. The system also creates a condensed sampling set when memory limits are reached, aiming to preserve performance and accuracy.

These changes address two bottlenecks that emerge at large scale: the number of operations needed to build the index and traverse it during queries, and the memory required to prepare training samples for the tree.

Performance figures

In internal tests, Google Cloud said AlloyDB with the revised ScaNN index scaled to more than 10 billion vectors. It also reported p95 latency of 51 milliseconds or less with 95% recall at that size.

Those figures matter because vector search is increasingly used in retrieval systems behind AI tools, recommendation engines and search applications. As datasets grow, database providers are trying to reduce latency without pushing hardware demands beyond practical limits.

The move also reflects broader competition among database and cloud groups to position PostgreSQL-based systems for AI workloads. Rather than sending vector data to a separate specialist database, many vendors are adding vector indexing and search features directly into existing relational systems.

For Google Cloud, AlloyDB has become one of its main database products for customers that want PostgreSQL compatibility alongside managed infrastructure. Extending ScaNN to larger vector datasets gives the service a clearer role in AI deployments that need both transactional database features and similarity search on one platform.

The four-level tree is being introduced in preview, meaning the feature is available for testing before wider production use. Users can deploy ScaNN in AlloyDB through the existing setup process and use the new index structure for vector search workloads.

The latest ScaNN release combines the four-level architecture with more efficient memory use to support operation at the 10 billion-vector level.