ML & Data Science Interviews · ML System Design · Lesson 1 of 3
The ML design round: a framework
The ML system-design interview (recommendation, ranking/search, fraud, churn) tests whether you can turn a fuzzy business goal into a working, measurable ML system. Follow a repeatable structure instead of jumping to models.
- Clarify & frame: business goal → ML problem (classification / ranking / regression). Nail the objective and constraints (latency, scale, cost).
- Define success: the offline metric (AUC, precision@k, NDCG) AND the online/business metric (CTR, revenue, retention) it should move.
- Data & labels: what's available, how labels are generated (explicit vs implicit feedback), sampling, and imbalance.
- Features: user / item / context features; a feature store; guard against leakage.
- Model: a simple baseline first, then a stronger model; justify the choice.
- Serving & monitoring: training vs serving path, latency budget, A/B rollout, and drift detection.
Structure beats triviaInterviewers score the framework and tradeoffs, not whether you name the fanciest model. Start with the objective and metrics, propose a baseline, then iterate — and think aloud about what could go wrong.
◆ Lock it in
- Go framing → metrics → data/labels → features → model → serving/monitoring.
- Tie every offline metric to an online business metric.
- Start with a baseline; earn complexity; narrate tradeoffs.
Feynman drill — say it out loudWalk through the six steps you'd use to structure any ML system-design question.
Step 1 rate your confidence · Step 2 pick your answer
What should you do FIRST in an ML system-design interview?
Step 1 — how sure are you?