MLOps applies DevOps discipline to ML: reproducible training, tested deployments, monitoring, and collaboration between data scientists and platform engineers.
MLOps capabilities
- Experiment tracking (metrics, parameters, artifacts)
- Feature stores for consistent online/offline features
- CI/CD for models with canary releases
- Monitoring: latency, skew, drift, business KPIs
Training vs serving skew
Features computed differently in notebook vs production API cause silent quality drops. Feature stores and shared transformation libraries reduce skew.
Deployment sketch
# Blue/green model deploy (conceptual)
traffic = {"churn_v2": 0.9, "churn_v3_canary": 0.1}
assert abs(sum(traffic.values()) - 1.0) < 1e-9Practice: Sketch product flows on paper or in a doc—optional Python illustrates API response shapes only.
Important interview questions and answers
- Q: Feature store?
A: Central definitions of features for training and realtime serving. - Q: Canary deploy?
A: Route small traffic to new model before full cutover.
Self-check
- List three MLOps capabilities.
- What is training-serving skew?
Pitfall: Training-serving skew—share feature definitions between notebook and API.
Interview prep
- Training-serving skew?
- Different feature code offline vs online silently hurts quality.
- Canary deploy?
- Route small traffic fraction to new model before full promotion.