ML System Design Interviews

ML system design interviews test whether a candidate can connect a product problem to a reliable data and serving workflow. Begin by asking what decision the system makes, who consumes it, what happens when it is wrong, and which latency, throughput, freshness, privacy, and reliability constraints apply. Clarify the objective and success metric before drawing components. Avoid assuming that every request requires a deep neural model. Next define inputs, labels, and data collection. Identify data sources, permissions, label quality, feedback loops, leakage risks, and how training examples represent future requests. Choose a baseline and model family appropriate to the task. For recommendation or search, distinguish candidate generation from ranking; for classification, define thresholds and costs of false positives and negatives. Explain offline evaluation and how to validate impact online without exposing users to unreviewed risk. Describe the serving path from request to response. It may include feature retrieval, preprocessing, model inference, business rules, caching, and fallback behavior. Compare batch versus online inference based on freshness and latency. Include versioned models and data, deployment strategy, rollback, capacity planning, and dependency failures. Features used in training must be available and computed consistently at inference time. Monitoring should cover system health and model behavior. Track latency percentiles, errors, throughput, input distribution, output patterns, and delayed outcome metrics. Define alerts and response actions. Privacy and abuse controls affect data retention, access, and model outputs. Failure modes can include missing features, stale models, upstream outages, drift, or feedback loops. A clear design explains tradeoffs and how to test assumptions. State what you would prototype first, what measurements would change your choice, and what the system does when components fail. A diagram without requirements, metrics, data contracts, or operational ownership is incomplete.