ML Engineer (MLE) interviews at FAANG are different from SWE interviews. The coding round is shorter and easier. The system-design round is harder and ML-specific. Most candidates over-prepare on LeetCode and under-prepare on ML system design — which is exactly backward.
The Five Rounds You Should Expect
Round 1: ML coding (45 min)
Python + NumPy / PyTorch + scikit-learn. Typical problems: implement K-means from scratch, implement logistic regression with gradient descent, implement a simple transformer attention block, implement a sampling algorithm. Easier than SWE coding rounds but tests numerical fluency.
Round 2: ML system design — recommendations / ranking
Most common ML system design topic. Design YouTube recommendations, design Amazon product ranking, design Spotify Discover Weekly, design Meta ad ranking. Expect: candidate generation → ranking model → re-ranking → exploration/exploitation balance. Know two-tower retrieval and DLRM cold.
Round 3: ML system design — content moderation / fraud / search
Second most common topic. Design fraud detection for a payments product, design content moderation for a social platform, design search ranking. Expect heavy emphasis on labelling strategy, online vs offline metrics, and human-in-the-loop.
Round 4: ML fundamentals + statistics (45 min)
Bias-variance trade-off, regularisation (L1 vs L2 vs dropout), bagging vs boosting, A/B testing power analysis, p-hacking awareness, transformer mechanics, attention mechanisms, embedding spaces. Conceptual, not coded.
Round 5: Behavioural + cross-functional (45 min)
ML projects involve product, engineering, data science, and ops. Behavioural rounds test: How did you handle disagreement with PM on metrics? How did you respond when your model was failing in production? How did you communicate model uncertainty to a non-technical stakeholder?
The Offline / Online Metric Bridge (Most Common Trap)
Strong MLE candidates always discuss the relationship between offline metrics (AUC, precision/recall, MAP, NDCG) and online metrics (revenue per session, click-through rate, retention). Weak candidates discuss offline metrics in isolation. Always bridge: 'Our offline AUC improved by 3%, but in the online A/B test we saw +0.5% CTR with no impact on revenue per session, suggesting the model picked up on a feature that correlates with clicks but not with purchase intent.' That phrasing is what gets you to Senior MLE.
Universal ML System Design Cheat Sheet
- Always state the ML problem framing first: regression, binary classification, multi-class, ranking, generation. The problem framing drives every downstream choice.
- Always discuss labelling strategy. Where do labels come from? Implicit (clicks) vs explicit (ratings) vs human-annotated (Mechanical Turk)? Cost vs latency vs quality trade-offs.
- Always discuss training data freshness. Daily retrain? Weekly? Online learning? Cite the trade-off between freshness and stability.
- Always discuss serving architecture. Batch inference, near-real-time, real-time? Latency budget?
- Always discuss feature stores. Online vs offline parity is the #1 source of train-serve skew bugs in production ML.
- Always discuss exploration/exploitation. Pure exploitation collapses long-tail content; pure exploration tanks short-term metrics. Multi-armed bandit framing.
- Always discuss fairness and bias. Disparate impact, demographic parity, counterfactual fairness. Even if the interviewer doesn't ask, mentioning fairness signals MLE maturity.
InterviewLift's mock interview simulator includes MLE mode with the full ML system-design rubric and offline/online bridge scaffolds. The verified question bank has 800+ MLE-specific questions across Google, Meta, Amazon, Microsoft, Apple, Netflix, and Stripe — indexed by team and level.
매주 지원자의 실제 면접 준비를 돕는 InterviewLift 고객 성공 팀이 검토했습니다.