Machine Learning System Design Interview, issue 38, May 26, 2026

The Retraining Window Fallacy

Staff Machine Learning Engineer interview at Netflix, and the interviewer asks:

Your recommendation system works perfectly in staging. But once it goes live, users request products and media that fall completely outside your model’s 30-day training window. How do you handle these out-of-vocabulary (OOV) categories safely at scale without breaking downstream inference SLAs?

Don’t say: Just write a hardcoded default fallback rule to recommend top-trending items, or simply expand the model’s training window to 6 months to capture everything.

The dangerous trap of solving data volatility with frequent model updates, and the low-latency external lookup trick that safely routes out-of-vocabulary features.

The full answer, with the mechanism and the arithmetic, is for paid subscribers on Substack.

Read it on Substack

Get the next one

Free on Substack. Unsubscribe in one click.