Definition
MLOps covers experiment tracking, model registries, automated training pipelines, deployment, and monitoring for drift and performance decay.
It recognizes that models are software artifacts whose data dependencies change over time.
In simple terms
Shipping a model is not a graduation ceremony — it is starting a long maintenance shift with dashboards, alerts, and versioned rollbacks.
Where you see it
- CI that retrains and shadow-tests a recommender weekly.
- Alerting when input feature distributions drift.
How it works
1.Track experiments
Log metrics, data versions, and code.
2.Register and deploy
Promote models through stages.
3.Monitor
Watch quality, latency, and drift; retrain as needed.
Why it matters
- Most ML value is lost between a good notebook and a reliable production system — MLOps closes that gap.
Often confused
MLOps is just Kubernetes.
Orchestration helps, but data versioning, evals, and monitoring are equally central.