MLOps Engineering Practice: Machine Learning Lifecycle and Deployment
MLOps Engineering Practice: Lifecycle and Deployment
This domain tests whether you can operationalize a model from experiment to production and keep it healthy. Expect questions on reproducible pipelines, artifact and version control, release strategies, serving patterns, and monitoring/rollback. The exam favors end-to-end reasoning: a model is only deployed when data, code, config, and environment are traceable together.
Core rules
- Version everything that affects predictions: training data snapshot, feature definitions, hyperparameters, code commit, container image, and model artifact. A model registry should store immutable artifacts with stage or alias metadata.
- Build pipelines as directed acyclic graphs with explicit data validation and schema checks. Point-in-time-correct feature joins prevent training-serving skew and label leakage.
- Separate offline evaluation from online guardrails. Offline metrics such as precision at fixed recall gate promotion; online SLOs such as latency, error rate, and prediction distribution gate rollout.
- Automate retraining triggers but require validation gates. Drift detection is a signal, not proof of degradation.
Deployment patterns
Shadow deployment sends live traffic to a candidate without serving its output; it validates behavior safely. Canary releases route a small percentage of traffic and expand only when guardrails pass. Blue-green swaps environments for fast rollback. Choose based on blast radius and rollback speed, not novelty.
Worked example
A fraud model v2 improves offline recall. Deploy it in shadow mode for a week, comparing score distributions and precision at the current review budget. Then route 5% of traffic as a canary with automated rollback if p95 latency exceeds 150 ms or precision drops more than 2 points. Only after the canary passes, promote to 50% and then 100%, keeping v1 warm for instant rollback. This sequence separates model-quality risk from serving risk.
Traps
Confusing data drift with concept drift; retraining on unvalidated labels; promoting on training metrics alone; skipping point-in-time joins; and deploying without a tested rollback path. Also watch for model registry equals deployment: a registry stores artifacts, not traffic routing.