MLOps Lecture

Hands-on mini-labs covering the main building blocks of an MLOps workflow: experiment tracking, data versioning, pipelines, monitoring and CI/CD.

Table of contents

• Repository structure
• Set-up
• Quickstarts
  • MLflow: experiment tracking
  • DVC: data versioning
  • Kedro: reproducible pipelines
  • Evidently: data & model monitoring
  • CI/CD with GitHub Actions
• Useful commands

Repository structure

.
├── .dvc/                      # DVC configuration
├── .github/workflow/          # GitHub Actions workflow (CI/CD), to be renamed in your fork
├── cicd-demo/                 # Demo project used for the CI/CD lab
├── dvc_quickstart.ipynb       # Data versioning with DVC
├── evidently_quickstart.ipynb # Monitoring with Evidently
├── kedro_quickstart.ipynb     # Pipelines with Kedro
├── mlflow_quickstart.ipynb    # Experiment tracking with MLflow
├── pyproject.toml             # Project dependencies (managed by uv)
└── uv.lock                    # Locked dependency versions

Set-up

Introducing uv as new standard package-manager

uv is a fast Python package and project manager that replaces pip, venv and pip-tools in a single tool.

# Install uv package manager
pip install uv

# Install Python 3.11
uv python install 3.11

# Pin Python version to 3.11 for this project
uv python pin 3.11

# Sync dependencies from pyproject.toml
uv sync

Quickstarts

Each quickstart is a self-contained notebook. Run uv sync first, then open the notebook of your choice.

MLflow: experiment tracking

Notebook: mlflow_quickstart.ipynb

Track your machine learning experiments: log parameters, metrics and artifacts for each run, compare runs side by side and keep a registry of your models. Launch the tracking UI with:

uv run mlflow ui

DVC: data versioning

Notebook: dvc_quickstart.ipynb

Version large datasets and models alongside your code without storing them in Git. DVC keeps lightweight pointer files in the repository and the actual data in a local or remote storage. Configuration lives in the .dvc/ folder.

# Track a file or folder with DVC
uv run dvc add data/

# Push the data to the configured remote
uv run dvc push

Kedro: reproducible pipelines

Notebook: kedro_quickstart.ipynb

Structure your data science code into modular, testable pipelines made of nodes, with a data catalog to manage inputs and outputs. It is a good starting point for moving from exploratory notebooks to production-ready code.

Evidently: data & model monitoring

Notebook: evidently_quickstart.ipynb

Monitor your models once they are deployed: detect data drift, check data quality and generate visual reports on model performance over time.

CI/CD with GitHub Actions

Demo project: cicd-demo/ · Workflow: .github/workflow/ci.yml

Automate testing and deployment on every push and pull request. The workflow sets up Python 3.11, installs the dependencies, runs the tests with pytest and runs a (simulated) deployment on main.

You can have a look at the file in cicd-demo/ as an example of unit tests: it contains a simple function and its test, which pytest detects and runs automatically in the workflow.

Your turn: the workflow folder is deliberately named .github/workflow (without the s), so it is not picked up by GitHub as is. Fork this repository, then rename the folder yourself so that the workflow appears and runs in the Actions tab of your fork:

git mv .github/workflow .github/workflows
git commit -m "fix: rename workflow folder to workflows"
git push

Remember to enable Actions from the Actions tab of your fork if GitHub asks you to.

Useful commands

# Run the main script
uv run main.py

# Convert requirements.txt to uv like pyproject.toml
uv add -r requirements.txt

# Add a new dependency
uv add <package>

# Run a command inside the project environment
uv run <command>