Lesson 2 / 25
The MLOps Lifecycle
A loop, not a hand-off.
Data, train, deploy, monitor, repeat
The lifecycle is a loop: data collection and validation; experimentation with tracked runs; training pipelines that are automated and reproducible; evaluation against the current model with gates; registration of approved model versions; deployment via batch jobs or services with safe rollout; monitoring of systems, data and model quality; and retraining when monitoring or a schedule says so. Each arrow is a hand-off that automation and versioning make reliable.
A restaurant kitchen, not a cooking competition
A competition rewards one perfect dish; a restaurant must serve the same quality every night with changing ingredients, staff and orders. MLOps is the kitchen system.
Automate the slowest hand-off first
If deploying takes weeks of manual steps, automate deployment before tuning models further.
Quick check: What closes the MLOps loop after deployment?
- Monitoring that triggers retraining or rollback
- Deleting the training data
- Writing a slide deck
- Turning off logging
Answer
Monitoring that triggers retraining or rollback — Monitoring feeds the next iteration.