Introduction to MLOps

MLOps applies DevOps principles — automation, continuous delivery, and monitoring — to the machine learning lifecycle, enabling teams to deploy, monitor, and reliably maintain ML models in production at scale.

The ML Lifecycle and Where MLOps Fits

The ML lifecycle spans data collection, feature engineering, training, evaluation, deployment, and monitoring. MLOps provides the tooling and processes to automate and govern each of these stages reliably.

Core MLOps Components

  • Experiment Tracking: Log metrics, parameters, and artifacts (MLflow, W&B)
  • Model Registry: Version, stage, and audit trained models
  • CI/CD Pipelines: Automate testing, training, and deployment
  • Model Serving: REST APIs, batch jobs, or streaming inference
  • Monitoring: Track data drift, performance degradation, and system health

MLOps Maturity Levels

Google defines three MLOps maturity levels:

  • Level 0: Manual training and deployment — scripts run by hand
  • Level 1: Automated training pipelines, manual deployment triggers
  • Level 2: Fully automated CI/CD, continuous training and deployment based on triggers

Most teams start at Level 0 and incrementally add automation as their model inventory grows.

Key Differences from Traditional DevOps

ML systems have additional complexity: the code, the data, and the model all change independently and all affect system behaviour.

The Three Axes of Change

  • Code changes: Updated model architecture or preprocessing logic
  • Data changes: New training data, schema drift, label changes
  • Model changes: Retrained weights with the same code and new data

Traditional CI/CD only tracks code; MLOps must version and test all three axes together to ensure reproducibility and auditability.

A Minimal MLOps Stack

You don't need enterprise infrastructure to start — a lightweight MLOps setup can be assembled from a few open-source tools.

Starter MLOps Toolchain

  • Experiment Tracking: MLflow (local or remote)
  • Data Versioning: DVC (Data Version Control)
  • Model Serving: FastAPI + Uvicorn
  • Containerisation: Docker
  • CI/CD: GitHub Actions
  • Monitoring: Evidently AI or Prometheus + Grafana