MLOps

What MLOps means

MLOps covers the practices that keep a model running in production. The term comes from joining model development with operations, because models behave differently from ordinary software. They do not finish, they degrade, and they need updating because the data has moved.

The difference from development is organisational. A model built once is a project, a model operated is a task. Anyone planning only the build ends up with several models in use whose state nobody knows.

The four basics

• Versioning: Every model needs an identifier, plus the data used, the settings and the date. Without those details it is impossible to reconstruct later why results changed.
• Monitoring: One metric per model, a cadence and a threshold that triggers a review. Without a threshold there is no alarm.
• Retraining: A planned cycle in which the model is retrained on current data. Planned means scheduled, not on demand.
• Rollback: The ability to switch a model off and return to the previous state. That is the most important and most frequently missing basic.

Why data preparation takes the largest share

In practice most of the work falls not on the model but on the data in front of it. Features must be built from orders, sessions and product data, and that construction is itself a program that needs maintenance. When an ERP changes a field, the features change, and the model suddenly receives different values than during training.

The gap between training and production is especially delicate here. If features are computed differently in production than in training, accuracy falls without any visible error, because the model keeps receiving values, just not the expected ones. The rule against it is simple: feature construction must be the same code for training and production, not a reimplementation.

What is realistic in a shop

Full operation with automatic delivery of several model variants is unnecessary for most retailers. Four things are realistic. First, a register of every model and rule in use. Second, one metric per model with a threshold. Third, a scheduled retraining cycle, often quarterly. Fourth, a documented switch-off for every model, so a bad result does not need to be discussed in an emergency meeting.

The most common cause of disappointing results is not model performance but the state after six months. A model with mediocre accuracy that is maintained is worth more than a good model nobody watches.