ScyllaDB has released Vector Search as part of ScyllaDB X Cloud, achieving sub-2ms P99 latency and ~250K QPS on billion-vector workloads. The solution integrates vector search directly into ScyllaDB's shard-per-core architecture using a Rust-based extension with USearch library. The architecture separates storage (ScyllaDB nodes) from indexing (dedicated Vector Store service consuming CDC updates) while maintaining a unified user experience. Benchmarks using the yandex-deep_1b dataset (1 billion vectors) demonstrate performance advantages over standalone vector databases, targeting real-time AI workloads like feature stores, predictive analytics, and fraud detection.
350 Impressions