Running the Feast Feature Store with Dragonfly
In this blog post, we explore the seamless integration of Dragonfly as a drop-in replacement for Redis in Feast—an acclaimed feature storage and server project widely recognized in the machine...
Real-time Data Infrastructure for Mission Critical Feature Stores
Dragonfly delivers sub-millisecond latency and unlimited scalability, making it the ideal choice for large, high throughput feature stores that are critical to revenue.
- Sign up for freeReceive $100 credit
- Book a demo
Dragonfly for Machine Learning
Sub-Millisecond Latency, Crash-Proof Feature Serving for ML Inference
Dragonfly is a leader in providing high-performance in-memory storage for ML feature stores. It serves features in under a millisecond, handles millions of queries per second, and stays stable under extreme throughput. It's a drop-in Redis replacement and plugs into platforms like Feast without code changes or tuning.
When integrated with frameworks like Feast, Dragonfly’s multi-threaded architecture delivers up to 25x higher throughput than Redis, sub-millisecond latency, consistent feature access between training and serving, and vertical scalability to handle terabyte-scale feature sets. ML Platform teams can deploy faster, reduce infrastructure complexity, and serve real-time predictions without making the cost and complexity tradeoffs that are inherent with Redis.
“By migrating our feature store from ElastiCache to Dragonfly Cloud we've been able to cut our cluster size by 70% while also reducing our average and P99 latencies by 40-50%.”
Tristan Fletcher
Tech Lead - Infrastructure Platforms at Instacart
Reduce Cost & Complexity
25% Better Memory Efficiency
Dragonfly's data structures delivery dramatically better memory efficiency, with most workloads experiencing around a 25% drop in memory when migrating from Redis to Dragonfly.
100% Hardware Utilization
Dragonfly's multi-threaded architecture means you can get dramatically more throughput on a fraction of the infrastructure, resulting in smaller cluster sizes and lower cloud bills.
Intelligent Data Tiering
Dragonfly's Intelligent Data Tiering can spill select data to disk that you are already paying for, resulting in dramatically lower bills with a negligible cost to performance.
Unlimited Scale Without Compromises
Don't drop features. Don't split up your cluster.
ML platform teams running feature stores on legacy infrastructure often have you to make tradeoffs in order to scale. That might mean dropping certain features or splitting in to multiple clusters. Dragonfly is built for modern scale, negating the need for these tradeoffs. Scale to hundreds of TBs of data and hundreds of million of QPS without fear of outages or performance degredation.
Free In-memory Data Resources
Get started for free
Create a Dragonfly Cloud account and receive $100 in credit to try it out.
Sign up for FreeReceive $100 credit
Request Demo