What is MLOps/ModelOps?
MLOps, told as a story about an automatic proofing machine that has to be retrained and watched, not just built once and forgotten.
The proofing machine built last year has been left untouched.The machine keeps missing today's expressions — language moved on, but it didn't learn.
Unlike ordinary code, the machine learned from piles of examples — so it's harder to fix.So fixing it isn't like changing one line of ordinary code.
MLOps is continuously caring for the proofing machine.Keep teaching the machine with new manuscripts, catch mistakes fast, and roll back to an older version if needed.
Caring for the machine isn't a one-time pass — it loops continuously.Collect, retrain, test, swap, watch — then the loop starts over again.
A smarter machine still doesn't remove the need for people.The tricky calls still go to a human proofreader. Caring for a borrowed big parrot (an LLM) instead is a different story → caring for the parrot.
MLOps/ModelOps = not building the proofing machine once and walking away, but continuously retraining it on new manuscripts, watching for mistakes, and rolling back when needed.
MLOps (Machine Learning Operations) manages a machine learning model's data collection, training, validation, deployment, and monitoring as one repeating pipeline. ModelOps widens the scope beyond ML to AI and decision models such as rules and optimization, covering governance and lifecycle management in production. Because a model's behavior shifts with data as well as code, retraining and drift detection matter more here than in ordinary CI/CD.
When grown-ups say it
- MLOps/ModelOps
- The process of continuously caring for the machine. Not just building it — running it is part of the job too.
- Data Pipeline
- The path that brings in new manuscripts. A steady stream of training examples.
- Retraining
- Teaching the machine again with new material. When the world changes, the machine has to learn again.
- Model Registry
- The shelf that stores each version of the machine. Same idea as configuration management → the shelf that protects the manuscript.
- Drift
- The machine starting to miss things as the world changes. As the data shifts, accuracy quietly slips.
- A/B Testing
- Running the new machine and the old one side by side. Only the more accurate machine stays.
- Model Rollback
- If the new machine makes more mistakes, go back to the old version. Same idea as the rewind in the SRE field → resetting a ride to its old setup.
- Monitoring
- Keeping an eye on the machine's mistakes. It's the first thing to notice when accuracy drops.