MLOps (Machine Learning Operations) is the practice of applying DevOps engineering principles to machine learning systems — specifically, to the complete ML lifecycle: data ingestion, feature engineering, model training, evaluation, deployment, monitoring, and retraining. It emerged as a discipline because the operational challenges of ML systems are substantially different from — and more complex than — traditional software systems.
Why ML systems are operationally different
A traditional software system fails in predictable ways: it crashes, it returns an error, it times out. A machine learning model fails silently — it continues to return predictions, but the predictions degrade over time as the real-world distribution of inputs shifts away from the training distribution (a phenomenon called model drift or data drift). Detecting and responding to this requires continuous monitoring of model outputs, not just system metrics.
Additionally, ML experiments are not deterministic in the way that software builds are. The same training code run twice may produce different models depending on random initialisation, data ordering, and hardware. Reproducibility requires logging hyperparameters, dataset versions, environment specifications, and random seeds — a set of concerns that traditional software CI/CD does not address.
Core MLOps capabilities
Experiment tracking: Logging all parameters, metrics, and artefacts from every training run — so that experiments can be compared, reproduced, and audited.
Model registry: A versioned catalogue of trained models — their lineage (what data and code produced them), evaluation metrics, and deployment status.
Automated pipelines: Reproducible, scheduled pipelines for data processing, model training, and evaluation — triggered by new data, schedule, or performance degradation.
Model serving: Scalable, low-latency serving of models via standardised APIs — with traffic management for canary deployments and A/B testing.
Monitoring: Continuous tracking of prediction distributions, data drift, feature drift, and model performance metrics — with alerting when drift exceeds thresholds.
MLOps in regulated environments
For Indian financial entities using ML for credit scoring, fraud detection, or KYC, RBI's guidelines on model risk management effectively require MLOps capabilities: model documentation, validation, monitoring, and governance are all mandated. Sending training data to a cloud-based ML platform (SageMaker, Vertex AI) creates data sovereignty and regulatory issues. A self-hosted MLOps stack is the only compliant path for regulated financial ML.