← AI Terminology
MLOps
MLOps is the set of practices and tools that take ML models from experiment to reliable production — data/versioning, training pipelines, registries, deployment, and monitoring.
It applies DevOps ideas to the ML lifecycle.
It applies DevOps ideas to the ML lifecycle.
Why It Matters in AI
Models rot without pipelines and monitoring. MLOps is how organisations industrialise classical ML and deep learning. LLMOps specialises it further, but the foundations (CI, registry, drift) remain essential.
Key Points
| Aspect | Description |
|---|---|
| Risks | Training-serving skew, silent drift |
| Tools | MLflow, Kubeflow, SageMaker, Vertex, Airflow |
| Culture | Reproducibility and automated tests |
| Related | LLMOps, model registry, experiment tracking |
| Artifacts | Datasets, models, metrics, feature defs |
| Lifecycle | Data → train → validate → deploy → monitor → retrain |
Simple Analogy
A factory line for models: not a one-off lab prototype, but versioned parts, QA gates, and continuous delivery.
Common Usage Examples
- MLflow tracking + model registry
- Feature store → training job → endpoint
- Data drift monitors in prod
- Rollback bad model versions
Summary
In short: MLOps is DevOps for machine learning — pipelines, registries, and monitoring that make models production-grade.