Video Recommendation: Problem Framing & Requirements
Explore how to effectively frame the problem of designing a video recommendation system by asking structured clarifying questions. Understand how to map business objectives to machine learning metrics and consider scale and latency constraints. Learn what is expected at different interview levels and why problem framing is crucial for downstream design decisions.
You are sitting across from an interviewer at a top tech company, and they say: “Design a video recommendation system for a platform like YouTube.” The next sixty seconds determine whether you spend the interview floundering through architecture diagrams or methodically building a system that demonstrates real design fluency. Video recommendation is the single most popular ML system design interview question at MAANG companies, and for good reason. It compresses nearly every hard problem in applied machine learning into one scenario: candidate generation over massive catalogs, real-time ranking under strict latency budgets, feedback loops that can silently degrade model quality, and business objectives that pull in competing directions.
The dominant paradigm for these systems is a two-stage architecture where a fast candidate generation layer narrows billions of videos down to hundreds, and a heavier ranking layer scores those candidates for the final recommendation slate. But this lesson is not about architecture yet. It is about everything that must happen before you draw a single box on the whiteboard. Problem framing and requirements gathering is the upstream work that shapes every downstream decision, and skipping it is the single most common failure mode in ML system design interviews.
Most candidates rush to modeling. The strongest candidates slow down here.
Clarifying questions for video recommendation
A well-structured set of clarifying questions signals to the interviewer that you understand the problem space deeply enough to know where ambiguity hides. Rather than asking generic questions, each question should directly influence an architectural or modeling decision downstream.
Organizing questions by category
The following categories cover the critical dimensions of a video recommendation problem.
Product context: Asking whether the recommendation surface is a homepage feed, a watch-next sidebar, or a search results page determines whether candidate generation must cast a wide net across diverse topics or stay contextually anchored to the video currently playing.
User context: Understanding whether users are logged in or anonymous dictates the balance between collaborative filtering (leveraging behavioral patterns across similar users) and content-based signals (relying on video metadata alone). This category also surfaces the cold-start problem for new users with no watch history.
Content context: The size and nature ...