ScyllaDB Releases Integrated Vector Search: 1B Vectors with 2ms P99s and 250K QPS Throughput

ScyllaDB today announced the general availability of its new Vector Search capability, which is integrated into ScyllaDB X Cloud. This high-performance vector search supports the industry’s largest models with low TCO.

ScyllaDB is commonly used for real-time AI workloads such as latency-sensitive machine learning, predictive analytics, and fraud detection. It is trusted by high-growth companies such as Tripadvisor, ShareChat, and Freshworks to power large-scale latency-sensitive feature stores. As ScyllaDB’s customers began adopting vector search, many found standalone vector databases to be overly complex and costly at scale. In response, ScyllaDB added Vector Search to its ScyllaDB Cloud offering.

ScyllaDB Vector Search is built on ScyllaDB’s shard-per-core architecture with a Rust-based extension that leverages USearch, the industry-standard ANN search library. The architecture separates storage and indexing responsibilities while keeping the system unified from the user’s perspective.

• The ScyllaDB nodes store both the structured attributes and the vector embeddings in the same distributed table.
• The dedicated Vector Store service consumes updates from ScyllaDB via CDC and builds approximate-nearest-neighbor (ANN) indexes in memory.
• Queries are issued to the database, then internally routed to the Vector Store.

This design allows each layer to scale independently, optimizing for its own workload characteristics and eliminating resource interference.

The combination of ScyllaDB’s hardware-optimized shard-per-core architecture for high op/s queries and USearch’s C++ implementation with 10x performance gains over FAISS is a perfect fit for massive real-time AI workloads. Together, this enables industry-leading time-to-first-token.

In recent 1-billion-vector benchmarks, ScyllaDB Vector Search achieved sub-2 ms P99 latency with up to ~250,000 queries per second on large-scale similarity search workloads. These tests were performed with the publicly available yandex-deep_1b dataset, which contains 1 billion vectors, and the 3 + 3 node setup reflects realistic production deployments

“ScyllaDB supports the most scalable vector search deployments at monstrous speed,” said Dor Laor, co-founder and CEO at ScyllaDB. “Based on publicly available benchmarks, ScyllaDB currently demonstrates the fastest vector search performance at billion-vector scale. It also offers excellent TCO across all model sizes. That means teams can support and scale their largest AI inference workloads without the traditional performance-cost tradeoffs.”

The post ScyllaDB Releases Integrated Vector Search: 1B Vectors with 2ms P99s and 250K QPS Throughput appeared first on SD Times.

from SD Times https://ift.tt/Sq6RLU3