What MLOps covers
- Reproducible training, so a result from last year can be produced again this year.
- Versioning of data, code and models together, because a model without its data is not reproducible.
- Deployment behind a serving interface, with rollback.
- Monitoring of accuracy and of the inputs, not only of uptime.
- Retraining, triggered by evidence rather than by the calendar.
Why models degrade without it
The world moves and the training data does not. Customer behaviour shifts, an upstream system changes a field, a category is added that the model has never seen. This is drift, and its defining property is silence: accuracy falls without anything breaking, so nothing alerts. The first sign is usually somebody downstream noticing that the predictions have stopped being useful.
What happens without MLOps
The common ending is a model that works, is never measured again, and quietly stops being right. Nobody can retrain it because the person who knew how has moved on, the training data has changed underneath, and the result cannot be reproduced. It is switched off, and the project is recorded as a failure when the model was never the problem.
Is MLOps only for large teams?
No, and the small version is worth having. A reproducible training pipeline, a held out evaluation set, a scheduled accuracy check and a named owner cover most of the value. The elaborate tooling matters at scale; the discipline matters immediately.
Related
- Machine learning development: models deployed with monitoring so accuracy is measured rather than assumed.
- Machine learning consulting: including an MLOps review before anything is built.