Learn MLOps
7 lessons · ~66 minutes of reading
A model that scores well offline is roughly a quarter of the work. The rest is getting it deployed reliably, knowing when it stops working, and being able to explain and reproduce any prediction it made six months ago.
MLOps is that discipline. It borrows from DevOps but has to handle something DevOps does not: the system degrades even when nothing about it changes, because the world it models moves.
This track follows the lifecycle in order — data, training and reproducibility, evaluation, deployment, monitoring, and governance — and each lesson names the failure it exists to prevent.
Assumes: You have trained a model and can read Python. This track is about everything that happens after the notebook.
The lessons
- 1
Why ML systems fail differently
What makes a machine learning system harder to operate than ordinary software, and the maturity levels teams pass through.
8 minPractise: Easy - 2
Data pipelines, versioning and feature stores
Make training data reproducible, keep serving consistent with training, and validate data before it corrupts a model.
10 minPractise: Medium - 3
Experiment tracking and reproducibility
Make every result explainable six months later, and get models into a registry with a real promotion path.
9 minPractise: Medium - 4
Evaluating models honestly
Pick metrics that reflect the decision being made, and validate a model the way it will actually be used.
10 minPractise: Medium - 5
Deployment and serving patterns
Batch, online and streaming inference — and how to release a new model without betting the traffic on it.
10 minPractise: Medium - 6
Monitoring, drift and retraining
Detect degradation before users report it, distinguish the kinds of drift, and decide when to retrain.
10 minPractise: Hard - 7
Governance, cost and responsible deployment
Documentation, fairness, privacy and the cost controls that keep an ML platform sustainable.
9 minPractise: Hard
What the practice bank covers
The lessons above are written against this syllabus, so anything in the question bank has been taught before you meet it.
- ML lifecycle, technical debt, CACE principle
- Data versioning, lineage, reproducibility (DVC, LakeFS)
- Feature stores, online/offline parity, training–serving skew
- Experiment tracking & model registries (MLflow, W&B)
- CI/CD/CT, pipeline orchestration (Airflow, Kubeflow, SageMaker)
- Serving: batch, online, streaming, shadow, canary, A/B
- Drift: covariate, prior, concept; detection tests
- Monitoring, alerting, ground-truth lag, proxy metrics
- Model governance, model cards, bias & fairness audits
- Cost, latency, autoscaling, GPU utilisation
- Case studies: recommender, fraud, forecasting, LLM/RAG in production