MLOps
Also known as: machine learning operations
What is MLOps?
The practices and tooling for running machine-learning models reliably in production: training, deploying, monitoring and retraining them as data and behaviour drift. It brings DevOps discipline to machine learning, where a model that tested well can quietly degrade in the real world. In practice, it means automation and monitoring across a model's whole life.
Why it matters
What makes running models harder than running ordinary software is that the thing in production is not just code. It is code plus the data it learned from plus a set of learned parameters, and it degrades in a way software does not. A model that passed every test at launch slowly gets worse as the world it was trained on drifts away from the world it now sees. You cannot fix that by rolling back to a previous version of the code, because the code was never the problem. So the discipline adds practices ordinary deployment never needed: versioning the data and the model together, watching live predictions for drift, and being able to trace exactly which data and which model version produced any given output when a regulator or a customer asks.
In practice
A fraud model that was sharp at launch grows leakier over months as spending habits shift, without a single line of code changing. The pipeline that catches this watches the model’s live accuracy, flags the drift, and triggers a retrain on fresher data. When an auditor later asks why a specific transaction was blocked, the team can point to the exact data and model version that made the call.
