Production ML Rollout
An ML system design round. Forty-five minutes, no code. The prototype exists and works for one analyst; you are designing what replaces it. Expect the requirements to move.
The questions in this round
The prototype works for one analyst. Redesign it for real-time production traffic while preserving correctness and rollback safety.
Good. Now the fraud team tells me a decision has to return in 50 milliseconds, and the features you just described take 400 to compute.
One more thing. The labels arrive thirty days late, because that is when chargebacks settle.
Which part of this design do you expect to fail first, and how would you find out?