ScyllaDB has announced the general availability of its new Vector Search feature, designed to handle massive datasets with impressive speed. This update allows users to search through billions of vectors quickly and efficiently, making it ideal for real-time AI applications. Companies working with large-scale machine learning, predictive analytics, and fraud detection now have a powerful new tool to improve performance and reduce costs.

High-Performance Vector Search at Scale

ScyllaDB’s Vector Search can manage datasets with up to 1 billion vectors, offering a P99 latency as low as 1.7 milliseconds and supporting up to 252,000 queries per second. This makes it one of the fastest options available for similarity search at such a large scale. The system is built on ScyllaDB’s existing shard-per-core architecture, which is optimized for high throughput and low latency.

The feature is integrated directly into ScyllaDB Cloud, removing the need for separate, standalone vector databases that can be complex and costly to operate at scale. By combining structured data and vector embeddings within the same distributed table, users can simplify their infrastructure while maintaining high performance.

Innovative Architecture and Performance Boosts

The Vector Search is powered by a Rust-based extension that leverages the USearch library for approximate nearest neighbor (ANN) search. This separation of storage and indexing responsibilities allows each layer to scale independently based on workload needs, reducing resource contention and improving efficiency.

This architecture enables each component to perform optimally—while ScyllaDB handles high-speed queries using its hardware-optimized shard-per-core design, USearch provides rapid vector similarity searches with performance gains up to ten times higher than some alternatives like FAISS. This combination results in faster response times and better handling of large AI workloads.

Real-World Benchmarks and Industry Impact

Recent benchmarks on a dataset with one billion vectors showed that ScyllaDB’s Vector Search achieved sub-2 millisecond latency at the 99th percentile, even supporting close to 250,000 queries per second. These tests used a realistic setup with three nodes for both data storage and processing, reflecting typical production environments.

According to Dor Laor, CEO of ScyllaDB, this performance demonstrates the platform’s ability to support the most demanding AI inference workloads. He emphasizes that the combination of speed, scalability, and cost-effectiveness helps companies support their largest AI models without having to compromise on performance or expenses.

Overall, the new Vector Search feature positions ScyllaDB as a leader in large-scale, real-time AI data management. It offers a practical, high-performance solution that simplifies infrastructure and improves efficiency for organizations working with massive AI datasets.

Inspired by

• https://ai-techpark.com/scylladb-brings-massive-scale-vector-search-to-real-time-ai/