Machine Learning System Design Interview #7 - The 10-Minute Horizon

Why TikTok-level recommendation systems retrain every few minutes - not nightly.

You’re in a System Design interview at TikTok. The interviewer sets a trap:

“How often should we retrain the core recommendation model?”

95% of candidates walk right into it.

The Instinct: Most engineers default to the standard MLOps playbook. “We should retrain weekly, or maybe nightly if compute allows. This balances cost with managing concept drift.”

It sounds reasonable. It’s efficient. It’s what you learned in bootcamps. It is also wrong.

The Turn: On high-velocity platforms, “nightly” is an eternity. You are treating User Intent as a static variable. It isn’t.

If a viral trend explodes at 2:00 PM, and your model was trained at 4:00 AM, your system is statistically blind. You aren’t suffering from “Concept Drift” - you are suffering from immediate irrelevance.

The Solution: The best engineers understand the 𝟏𝟎-𝐌𝐢𝐧𝐮𝐭𝐞 𝐇𝐨𝐫𝐢𝐳𝐨𝐧.