MLOps
The set of practices, tools, and automation for taking machine learning models from experimentation to reliable production — covering data pipelines, training, deployment, monitoring, and retraining.
MLOps applies DevOps discipline to machine learning. Where traditional software is deterministic, ML systems depend on data that drifts, models that decay, and behaviour that can only be judged statistically. MLOps exists to make those systems reproducible, observable, and safe to update.
Core concerns:
- Data pipeline management: versioned datasets and features, with lineage from raw source to training run.
- Experiment tracking: recording code, hyperparameters, and metrics so results can be reproduced.
- Deployment: serving models behind APIs or at the edge, with canary rollouts and rollback.
- Monitoring: watching for data drift, quality regressions, latency, and cost in production.
- Retraining and governance: defining when and how models are refreshed, and keeping an audit trail.
The rise of foundation models has shifted parts of MLOps toward prompt and retrieval management, evaluation harnesses, and GPU capacity planning — sometimes called LLMOps. The underlying goal is unchanged: make the system’s behaviour predictable enough to trust in production.