3 hints available

EasyML EngineeringSystem DesignModel Training Model Serving Feature Stores

Feature Store Architecture for Real-Time Fraud Detection

You are building a real-time fraud detection model. You need features like:

"Number of transactions by this card in the last hour"
"Average transaction amount in the last 30 days"
"Time since last login"

(1) Why can you not compute these from scratch at serving time given a sub-100ms latency budget?

(2) Describe a feature store architecture including both offline and online components. What gets stored where, and how does data flow from raw events to served features?

(3) How does the feature store prevent training-serving skew for these specific features? Walk through what would go wrong without one.

(4) A junior engineer suggests recomputing the 30-day average at serving time using a fresh database query. Why is this a bad idea both for performance AND for correctness?

Sign in to attempt this problem

Free account gives you full access to community problems with the complete solution reveal: golden answer, senior walkthrough, and score breakdown, after submission.

Start free →Already have an account? Sign in