The problem
Every model team ends up rebuilding the same features, with different freshness and different bugs, and serving each model on its own stack.
How it works
Kafka streams transaction events to feature processors that compute real-time features and store them in Redis through Feast, with online and offline serving. Multiple models share those features. Ray Serve handles deployment with A/B testing and traffic splitting, Prometheus monitors it, and MLflow tracks experiments. A FastAPI service sits in front.
03 /
Key decisions
- One feature store, many models
Fraud detection, recommendation and forecasting models share the same feature infrastructure instead of each computing their own.
- A/B testing in the serving layer
Ray Serve splits traffic between model versions, with Prometheus metrics and MLflow experiment tracking.
04 /
Results
- Events / sec
- 10K
- Feature lookup
- 5–10 ms
- Inference
- 20–50 ms