Follow the ML lifecycle — MLOps Chapter 2

$ run as a normal user# run as root (or with sudo)

See the lifecycle

Each step has its own tools and checks.

Step	Work	Typical tools
Data	Collect, clean, version, validate	DVC, Great Expectations, data lakes
Features	Compute inputs the same way in training and serving	Feature stores (Feast)
Training	Run experiments, track results	Notebooks, MLflow, Weights & Biases
Evaluation	Compare with the current model, check fairness	MLflow, custom tests
Registry	Store approved model versions	MLflow Model Registry
Deployment	Serve the model as an API or batch job	FastAPI, KServe, Seldon, BentoML
Monitoring	Watch quality, drift, latency, cost	Evidently, Prometheus, Grafana

2.2

Know the maturity levels

Teams grow from manual work to full automation.

Level	What it looks like
0 — manual	Notebook → model file → someone copies it to a server
1 — pipeline	Training is an automated pipeline; models are retrained on new data
2 — CI/CD for pipelines	Pipeline code is tested and deployed by CI/CD; models promoted by rules

2.3

Version everything

Reproducible = you can say exactly which code, data, parameters and environment made a model.

$git tag model-fraud-v12
$dvc add data/transactions.parquet
$git add data/transactions.parquet.dvc .gitignore
$dvc push
$pip freeze > requirements.lock

What	How
Code	Git
Data	DVC, lakeFS, or dated snapshots in object storage
Parameters and metrics	Experiment tracking (MLflow)
Environment	Container image with pinned packages
Model	Model registry with version and stage

2.4

Add DevSecOps to ML

Models and data are assets — and new attack surfaces.

Scan training and serving images like any image (container security).
Protect training data (PII): access control, encryption, data minimisation.
Pickled model files can run code when loaded — load only trusted, signed artifacts; prefer safe formats (ONNX, safetensors).
Limit and log access to model APIs (rate limits, auth) — models can leak data through their answers.

All commands in this chapter

git