The Databricks MLOps Stacks repository provides a customizable template and framework for initializing new machine learning projects on Databricks with production-ready best practices built in from the start. Written in Python, the project is designed to bridge the gap between data scientists who need to iterate quickly on ML code and operations engineers who need to establish CI/CD pipelines and manage ML resources at scale. The repository is currently in public preview and is documented at the Databricks dev tools bundles documentation site.
The core value proposition of MLOps Stacks is enabling rapid project initialization with three modular components already integrated. The first component is an example ML code structure that includes training pipelines and batch inference modules organized as unit-tested Python packages and notebooks, allowing data scientists to start iterating on ML problems without needing to refactor code into production-ready modules later. The second component is ML Resources as Code, which defines ML pipeline resources including training and batch inference jobs through Databricks CLI bundles, enabling teams to govern and audit ML resource changes through pull requests rather than ad-hoc UI modifications. The third component is CI/CD automation, supporting both GitHub Actions and Azure DevOps workflows that test and deploy ML code and resources automatically.
The repository implements a structured development workflow across three execution environments: development, staging, and production. Data scientists work in the dev environment, file pull requests that trigger unit and integration tests in an isolated staging workspace, and upon merge to the main branch, staging jobs automatically update with the latest code. Production deployments occur through a separate release branch process, allowing teams to control when code changes move to production on a scheduled basis. This separation of concerns enables data scientists to move quickly while maintaining production stability and auditability.
To use MLOps Stacks, users run the databricks bundle init command with the mlops-stacks template, which prompts for configuration parameters. The initialization process offers flexibility through three setup modes: CICD_and_Project for complete setup, Project_Only for data scientists getting started, and CICD_Only for machine learning engineers setting up CI/CD on existing projects. Required configuration includes the cloud provider (AWS, Azure, or GCP), CI/CD platform choice, staging and production workspace URLs, and branch naming conventions. The tool requires Python 3.8 or higher and Databricks CLI version 0.236.0 or later.
The project is classified across multiple domains including machine learning operations, pipeline automation, model lifecycle management, continuous integration and delivery, and scalable ML infrastructure, reflecting its comprehensive approach to operationalizing machine learning on Databricks.
