Google AlloyDB’s preview four-level ScaNN index supports vector-search workloads exceeding 10 billion vectors.

Google introduced a preview four-level tree architecture for the ScaNN index in AlloyDB, its fully managed PostgreSQL-compatible database service. Google said the architecture enables the index to operate efficiently at a scale of 10 billion vectors and supports datasets above that level.

Earlier AlloyDB ScaNN tree indexes used two- or three-level configurations. At 10-billion-vector scale, Google identified index-construction compute requirements, query-traversal work and sampling memory usage as major constraints.

The four-level design uses hierarchical partitioning to reduce the number of vectors examined during a query. Google described the basic search complexity as O(N^1/2) for two levels, O(N^1/3) for three levels and O(N^1/4) for four levels.

The design also incorporates Top-K branch selection, Spilling with Orthogonality-Amplified Residuals, or SOAR, centroid adjustment and a balanced tree shape. Google said these additions are intended to preserve performance and limit recall loss while improving build efficiency and memory use.