Machine Learning Model Deployment with Git and MLflow

This repository demonstrates a workflow for deploying machine learning models through Git instead of making manual changes directly using the MLflow GUI.

Overview

The goal of this project is to introduce engineering best practices to ML model deployment, including:

• Pull requests (PR) for changes
• Code reviews
• Code quality checks
• Model deployment rollback
• Automatically starting A/B testing

By adopting this workflow, data scientists can deploy ML models in a more robust and automated way.

Deployment Workflow

1. Update Configuration: Data scientists modify the ML model configuration file for the desired environment (e.g., dev, pre-prod, or production).
2. Create a Pull Request: Submit a PR for the changes.
3. Trigger CI/CD: The PR triggers a GitHub Action that runs the following steps:
   • Code quality checks
   • Unit tests
   • Other CI/CD steps
4. Approval and Deployment: Once the PR is approved, the ML model is deployed to the selected environment.

Project Context

This code is part of a larger MLOps project designed to automate the entire machine learning pipeline. The broader project includes:

• Feature engineering
• Model deployment
• Model monitoring
• Running A/B experiments

Goals of the Project

• Automate ML Pipelines: From idea inception to running A/B experiments in a couple of weeks.
• Increase Experimentation Frequency: By automating the ML pipeline, we significantly improved the number of A/B experiments conducted.
• Deliver Business Value: Faster experimentation and deployment cycles have allowed us to deliver substantial business impact.

Code Scope

This repository contains a small portion of the overall MLOps codebase, focusing on the deployment process. It highlights the transition from manual processes to a CI/CD-driven workflow for ML model deployment.

---

Authored by Rostand