What is MLOps, and why isn't a model that works in a notebook enough?
A model trained in a notebook is an experiment. MLOps is what turns it into a reliable system that keeps working well months later, in production.
Training a model with good accuracy in a notebook is only the beginning. Getting it to serve real predictions reliably, and keeping it accurate six months later, is a completely different engineering problem — that's where MLOps comes in.
From notebook to service
A trained model needs to be exposed somehow so other systems can consume it — typically behind an API — with version control over exactly which model is serving at any given moment, not just over the code that trained it.
Monitoring: models degrade over time
Unlike traditional software, a machine learning model can start failing without the code ever changing: the real world changes, and the patterns the model learned stop being valid. This is called data drift or model drift, and without active monitoring it goes unnoticed until the damage is already done.
A concrete drift example
A recommendation model trained on 2024 user behavior can lose accuracy by 2026 simply because consumption habits changed — the model isn't "broken" in the traditional sense, but it no longer reflects current reality.
Retraining and versioning
MLOps establishes a repeatable process to retrain the model with new data, evaluate it before replacing the production version, and roll back to a previous version if something goes wrong — the same CI/CD discipline applied to models instead of just code.
An engineering field, not just data science
MLOps combines data science with software engineering and infrastructure practices — it's the bridge between "the model works in my notebook" and "the model reliably works for real users". Explore the MLOps path to understand that bridge in detail.