The ML lifecycle: experiments, registry, pipelines and deployment
MLOps turns a model from a notebook into a product that can be rebuilt, released and rolled back. This page walks through each stage of the lifecycle and the tools for it.
Overview
A model in production is the result of data, code, configuration and an environment. MLOps keeps all four versioned and connected, so any model can be explained, reproduced and replaced safely.
The stages
| Stage | Purpose | Tools |
|---|---|---|
| Experiment tracking | Record parameters, metrics and artefacts of every run | MLflow, Weights & Biases |
| Features | Compute features once for training and serving | Feature store |
| Pipelines | Repeatable training and evaluation | Kubeflow, Airflow, Dagster |
| Tuning | Search hyperparameters efficiently | Optuna, Ray Tune |
| Registry | Versioned models with stage and approval | MLflow, Vertex AI, SageMaker |
| Packaging | Containers and model formats | Docker, ONNX Runtime, BentoML |
| Deployment | Release to users safely | Kubernetes, KServe, Argo CD |
| Documentation | What the model is for and its limits | Model cards |
Release strategies
- Shadow: the new model runs on live traffic but its answers are not shown; compare offline.
- Canary: a small share of traffic goes to the new model, increasing as metrics hold.
- A/B test: split traffic to measure business impact.
- Rollback: the registry keeps the previous version one command away.
CI/CD for models
Code changes run unit tests; data changes run data contract validation; model changes run the evaluation set and fail the release below a threshold. Pipelines and infrastructure live in code with Terraform or Helm and deploy through GitLab CI/CD or Argo CD.
LLM applications
For LLM apps the "model" also includes prompts, retrieval settings and tools. Version them together, trace every request, and run evaluations on each change; see evals and tracing.
Frequently asked questions
Do small teams need all of this?
No. Start with experiment tracking, a registry and an evaluation gate; add feature stores and orchestration when several models share data.
What is the difference between MLOps and DevOps?
MLOps adds data and model versioning, evaluation gates and drift monitoring to the usual software release practices.
Summary
Track every experiment, version data and models, automate training and evaluation, release gradually and keep rollback ready. The MLOps in production path covers each step.
Back to the topic
This page is part of MLOps and AI infrastructure, which maps every concept, tool, simulation and article on the subject.