Overview
MLOps (Machine Learning Operations) applies DevOps principles to machine learning systems — automating the lifecycle of model development, training, deployment, monitoring, and retraining. Unlike traditional software where the artifact is code, ML systems produce models that degrade over time as data distributions shift. MLOps addresses this by treating data, model code, and training infrastructure as first-class versioned artifacts, with pipelines that continuously validate model performance and trigger retraining when quality drops. The discipline bridges the gap between experimental notebooks and production-grade ML systems.
Key Concepts
- Experiment tracking — Recording hyperparameters, metrics, and artifacts for every training run. Tools: MLflow, Weights & Biases, Neptune.
- Model registry — Central repository for versioned models with metadata, lineage, and deployment state (staging, production, archived).
- Feature stores — Centralized systems for computing, storing, and serving features consistently across training and inference.
- Pipeline orchestration — Automating the sequence of data ingestion, preprocessing, training, evaluation, and deployment. Tools: Kubeflow, Airflow, Prefect.
- Model serving — Deploying models as services (REST APIs, gRPC) with scaling, versioning, and A/B testing capabilities.
- Data drift detection — Monitoring whether production data distributions diverge from training data, triggering retraining when necessary.
- Reproducibility — Ensuring any experiment can be recreated with identical results. Requires versioning data, code, environment, and random seeds.
Trade-offs
- Full MLOps infrastructure has significant setup and maintenance cost
- Over-engineering for small teams or few models adds unnecessary complexity
- Automated retraining requires robust evaluation to prevent deploying worse models
- Feature stores add latency for online serving
Related
- Model Evaluation is the quality gate within MLOps pipelines
- CD principles extend into ML pipelines as CI/CT (continuous training)
- Overfitting must be caught by automated evaluation before model promotion