What is MLOps?
Building a model is the start. Keeping it accurate, explainable, and affordable for years is most of the work. MLOps borrows from DevOps, the practices that let software teams ship often without breaking things, and adds what's different about models: they depend on data, and they degrade when data changes.
For language models and agents, the same ideas are sometimes called LLMOps. The pieces shift, with prompts, tools, and evaluation sets alongside or instead of training data, but the discipline is the same.
What are the core MLOps practices?
| Practice | What it does | Tools and terms |
|---|---|---|
| Versioning | Records exactly which data, code, and settings produced each model | Model registry, data versioning |
| Reproducibility | Lets the team rebuild any past model and get the same result | Pinned dependencies, fixed seeds |
| Testing | Checks a model against an evaluation set before release | Evals, regression tests |
| Deployment | Releases changes safely and reversibly | Canary releases, rollback |
| Monitoring | Watches accuracy, cost, and inputs once live | Drift detection, alerts |
| Retraining | Refreshes the model on new data, on a schedule or a trigger | Continuous training, feature stores |
What is an example of MLOps?
An anomaly model that stayed useful
A team runs a model that flags unusual swings in campaign spend. It worked well at launch, then started missing problems during a peak season it had never seen.
Because every version was registered and monitored, the team could see when its accuracy slipped, retrain on the new season's data, test the new version on the old cases, and release it to a few accounts before all of them.
MLOps vs. DevOps
| DevOps | MLOps | |
|---|---|---|
| What ships | Code | Code, data, and models |
| Why it breaks | A code change | A code change, or the world changing |
| Testing | Pass or fail | Scored against an evaluation set |
| After release | Watch errors and speed | Watch errors, speed, accuracy, and drift |
Why does MLOps matter?
Most models that fail in production don't fail at launch. They fail months later, when the data has moved on and nobody noticed. MLOps is what notices.
- Trust. Teams can explain which version produced a number and why.
- Safe change. New models ship without risking the old ones' results.
- Lasting value. Models are kept current instead of being rebuilt from scratch.
How does Alli do this?
Alli by PMG's Generative Workflows keeps each workflow's schedule, notifications, and versions, with a history of every run, and the Testing Suite lets builders run one task at a time and see errors as they happen. Once reviewed, a workflow runs the same steps in the same order every time.
Alli Alerts runs anomaly detection models built by PMG's data science team, which judge each metric against its own historical pattern. Definitions they read live in Alli Semantic Layer, so the models and the reports work from the same governed data.
Explore Alli Generative WorkflowsMLOps FAQ
What does MLOps stand for?
Machine learning operations: the practices and tools for deploying, monitoring, and maintaining machine learning models in production.
What is the difference between MLOps and LLMOps?
LLMOps applies MLOps practices to language models and agents. It puts more weight on prompts, tools, evaluation sets, and token costs, and less on training models from scratch.
What is a model registry?
A catalog of trained models with their versions, the data and code that produced them, their test results, and which one is live.
What is a feature store?
A shared system that stores the inputs models use, called features, so training and live predictions use the same definitions and values.
Last updated Back to the glossary