MLOps without the mythology: what it really takes to keep a model in production

The MLOps literature describes complete platforms that almost no organisation uses in full. The minimum architecture, by contrast, is four things, and it can be built from what you already have.

The first is versioning data, code and model together: without being able to reconstruct exactly how the model was trained six months ago, any investigation into odd behaviour is an exercise in memory.

The second is watching drift on the input distributions and on performance, with thresholds agreed before release. The third is a retraining procedure somebody can run without asking permission, because a model that is only refreshed when the team has time is a model that ages.

The fourth, and the most neglected, is accountability: who authorises the release, who can stop it, who answers for a wrong decision. Without that, the other three are just documentation.