Machine Learning System Design Interview, issue 40, May 28, 2026
The Look-Ahead Trap
Machine Learning Engineer interview at Netflix, and the interviewer asks:
“You train a predictive model on user activity logs using a standard random 80/20 split and hit a spectacular 98% offline accuracy. But the minute you push it to production, online performance crashes to 55%. What structurally broke, and how do you fix it?”
Why predicting the past with future logs quietly destroys your live performance, and how lagging your training features fixes the hidden pipeline latency killing your metrics.
The full answer, with the mechanism and the arithmetic, is for paid subscribers on Substack.