17.1 Model Versioning

Model versioning means giving every trained model a unique, traceable identity — capturing not just the model file but the code, data, and hyperparameters that produced it. This lets you reproduce results, roll back to a previous model if a new one misbehaves, and compare experiments fairly.

Imagine a Telebirr fraud-detection model. Last month's version v1.2 worked well, but this week's v1.3 started flagging too many legitimate transfers. Because each version was tracked with a registry, the team instantly rolled back to v1.2 while investigating.