What is MLOps?
MLOps (machine learning operations) describes the principles and tools used to reliably develop, deploy, monitor and update machine learning models. It applies the ideas of DevOps to the lifecycle of ML models, turning a prototype into a system that can be operated permanently.
Also known as: machine learning operations · ML operations · ML ops
Where MLOps is used
MLOps closes the gap between a model that works in a notebook and a system that reliably delivers predictions in daily operation. It governs how data and models are versioned, how training and deployment are automated, and how models are monitored and retrained when their quality declines.
Core building blocks are versioning of data, code and models, automated training and deployment pipelines, a model registry, monitoring of model and data quality, and mechanisms for reproducible retraining. In Azure these building blocks are implemented with Azure Machine Learning and Azure DevOps or GitHub, for example.
A practical example
A model predicts expected demand. Without MLOps it is trained once and gradually becomes outdated as market conditions change (data drift). With MLOps the input data and prediction quality are monitored, and when deviations occur a pipeline automatically triggers retraining and a controlled deployment.
Benefits & typical use cases
MLOps pays off as soon as an ML model is meant to operate permanently rather than serve as a one-off analysis.
- Reproducibility: training can be traced with the same data and parameters
- Automated deployment of new model versions without manual effort
- Monitoring of model quality and input data to detect data drift early
- Clear governance and traceability of which model version was in use and when
How it differs from related terms
DevOps applies to software in general; MLOps extends it with the specifics of data and models, such as data versioning and model monitoring. Machine learning in Azure provides the platform and models, while MLOps ensures their reliable operation. CI/CD is a technique used within MLOps for automated pipelines.
How smiit works with it
smiit helps mid-sized companies not only develop machine learning solutions but make them permanently operable. In the Azure ecosystem, reproducible pipelines, monitoring and controlled deployment are set up so that models reliably create value instead of remaining stuck in the prototype stage.
Common mistakes & misconceptions
- MLOps is not just DevOps for models; it must also cover data and model versioning, drift monitoring and reproducibility of training runs.
- Many assume a trained model stays good forever, but models degrade as data changes and must be monitored and retrained.
- A common error is to think MLOps starts only after deployment. In fact it spans the whole cycle from data preparation through training to operations.
Frequently asked questions
What is the difference between DevOps and MLOps?
DevOps automates the development and deployment of software in general. MLOps applies these principles to machine learning models and adds data versioning, a model registry and monitoring of model quality.
Do we need MLOps for just a single model?
Usually not for a one-off analysis. But as soon as a model is meant to deliver predictions permanently and stay current with new data, MLOps ensures reliable and traceable operation.
What is data drift in the MLOps context?
Data drift describes the change of input data compared to the training data, which gradually makes a model less accurate. MLOps detects this through monitoring and triggers retraining when needed.
Which tools are used for MLOps in Azure?
Typically Azure Machine Learning for training, model registry and deployment, combined with Azure DevOps or GitHub for versioning and pipelines. Azure Databricks is often added for data preparation. Which building blocks are needed depends on the complexity of the models.
How are MLOps and model governance related?
MLOps provides the technical basis for governance: versioning, a model registry and monitoring make it traceable which model version was trained on which data and when it was in use. This is the precondition for audits and for clear responsibilities in model operation.
Related terms
Sources & further reading
Want to put this topic to work in your company?
Updated · Back to the glossary