advanced

ML Engineering & MLOps

The engineering around the model: turning raw data into leak-free features, shipping and monitoring the trained artifact, and designing the whole ML system end to end.

ml-engineering

Practice: 12 questions · 2 tasks →

Sections

Contents

Detailed explanation

ML Engineering & MLOps

A model that scores well in a notebook is not a product. The distance between the two is engineering — turning raw columns into stable features, shipping the trained artifact so it answers inside a latency budget, and watching it in production so you notice when the world drifts out from under it. This topic covers the disciplines that surround the model rather than the ones inside it.

The single most quoted number in this area is a model that scored 0.91 offline and 0.68 live. The cause is almost never the algorithm; it is a feature computed differently at training and serving (train/serve skew), a column that leaked information from the future, or a distribution that drifted while nobody watched. Everything below exists to prevent, detect, or reason about exactly that.

Topic map

Feature engineering — encodings (one-hot / target / embedding), scaling, missing values, feature stores, and the two production-killers: leakage and train/serve skew.
MLOps & production ML — experiment tracking, model registries, safe deployment (shadow / canary / A/B), drift monitoring and retraining triggers.
ML system design — framing, metric choice, candidate generation and ranking, latency budgets, cold start, feedback loops and A/B validation.

What interviewers probe

Past junior, this is where interviews live: can you frame a problem and name its metric, tell data drift (the input distribution shifts) from concept drift (the target rule shifts) and respond to each correctly, and resist concluding from an offline win before an A/B test confirms it online.

Why it matters

A model that scores 0.91 offline routinely scores 0.68 in production, and the cause is almost never the algorithm — it is train/serve skew, leakage, or drift nobody was watching. Interviews for anything past junior lean hard on this: they want an engineer who can frame a problem, pick the right metric, and keep a model honest after launch, not someone who only tunes hyperparameters.

ML Engineering & MLOps

Topic map

Feature engineering — encodings (one-hot / target / embedding), scaling, missing values, feature stores, and the two production-killers: leakage and train/serve skew.

MLOps & production ML — experiment tracking, model registries, safe deployment (shadow / canary / A/B), drift monitoring and retraining triggers.

ML system design — framing, metric choice, candidate generation and ranking, latency budgets, cold start, feedback loops and A/B validation.