Definition
Machine learning operations, usually shortened to MLOps, is the engineering and operational discipline for building, versioning, testing, deploying, observing, governing, and maintaining machine-learning systems. It applies DevOps ideas to systems whose behavior depends on code, data, features, training configuration, and learned parameters.
A production MLOps practice connects data lineage, reproducible training, experiment tracking, validation, model registries, deployment, monitoring, rollback, and retraining. Continuous integration tests more than application code; it also checks data, schemas, pipeline components, and model quality. Continuous delivery may promote both a training pipeline and the model service that pipeline produces. Continuous training is optional and should be gated, not assumed.
How the term is used
The label became common as teams discovered that deploying a model endpoint was the small part of operating an ML system. Google Cloud's reference architecture defines MLOps as a culture and practice joining ML development with operation, with automation and monitoring across construction and production. That is an influential implementation model, not a universal maturity standard or proof that every model should retrain automatically.
Distinguish it from nearby terms
DevOps primarily manages software and infrastructure delivery. MLOps adds learned artifacts and the changing data that produces and evaluates them. DataOps focuses on the quality and movement of data products. LLMOps is an application-heavy specialization for large-language-model systems, often with less control over model training.
Check your understanding
A fraud model degrades after customer behavior changes. Which version do you roll back: application code, features, training data, or weights? An MLOps system should reconstruct the lineage among them, show which evaluation and approval promoted the model, and support a safe response rather than treating the endpoint as an isolated binary.