Campus: Making Cluster-based Vector Index Scalable with Transactions

Similarity search over vector databases is important today. Acceleration of read-only parallel search has been studied widely, while online vector insert has not been explored from the viewpoint of scalability. This paper proposes a parallel incremental update algorithm Campus for a vector search index based on an optimistic concurrency control protocol. By performing validation using multiple versions, we have made it possible to execute parallel updates while maintaining index quality. We were able to construct the index that demonstrated high search accuracy and scalability, and construction performance was up to approximately 23.8 times faster than a serial protocol.