NOTE: This feature is in public preview.
This repo provides a customizable stack for starting new ML projects on Databricks that follow production best-practices out of the box.
Using Databricks MLOps Stacks, data scientists can quickly get started iterating on ML code for new projects while ops engineers set up CI/CD and ML resources management, with an easy transition to production. You can also use MLOps Stacks as a building block in automation for creating new data science projects with production-grade CI/CD pre-configured. More information can be found at https://docs.databricks.com/en/dev-tools/bundles/mlops-stacks.html.
The default stack in this repo includes three modular components:
| Component | Description | Why it's useful |
|---|---|---|
| ML Code | Example ML project structure (training and batch inference, etc), with unit tested Python modules and notebooks | Quickly iterate on ML problems, without worrying about refactoring your code into tested modules for productionization later on. |
| ML Resources as Code | ML pipeline resources (training and batch inference jobs, etc) defined through databricks CLI bundles | Govern, audit, and deploy changes to your ML resources (e.g. "use a larger instance type for automated model retraining") through pull requests, rather than adhoc changes made via UI. |
| CI/CD(GitHub Actions or Azure DevOps) | GitHub Actions or Azure DevOps workflows to test and deploy ML code and resources | Ship ML code faster and with confidence: ensure all production changes are performed through automation and that only tested code is deployed to prod |
See the FAQ for questions on common use cases.
An ML solution comprises data, code, and models. These resources need to be developed, validated (staging), and deployed (production). In this repository, we use the notion of dev, staging, and prod to represent the execution environments of each stage.
An instantiated project from MLOps Stacks contains an ML pipeline with CI/CD workflows to test and deploy automated model training and batch inference jobs across your dev, staging, and prod Databricks workspaces.
Data scientists can iterate on ML code and file pull requests (PRs). This will trigger unit tests and integration tests in an isolated staging Databricks workspace. Model training and batch inference jobs in staging will immediately update to run the latest code when a PR is merged into main. After merging a PR into main, you can cut a new release branch as part of your regularly scheduled release process to promote ML code changes to production.
mlops_stacks_01_ml_dev.mov
mlops_stacks_02_create_pr.mov
mlops_stacks_03_merge_PR.mov
mlops_stacks_04_deploy_to_staging.mov
mlops_stacks_05_release.mov
See this page for detailed description and diagrams of the ML pipeline structure defined in the default stack.
- Python 3.8+
- Databricks CLI >= v0.221.0
Databricks CLI contains Databricks asset bundle templates for the purpose of project creation.
Please follow the instruction to install and set up databricks CLI. Releases of databricks CLI can be found in the releases section of databricks/cli repository.
Databricks asset bundles and Databricks asset bundle templates are in public preview.
To create a new project, run:
databricks bundle init mlops-stacks
This will prompt for parameters for initialization. Some of these parameters are required to get started:
input_setup_cicd_and_project: If both CI/CD and the project should be set up, or only one of them.CICD_and_Project- Setup both CI/CD and project, the default option.Project_Only- Setup project only, easiest for Data Scientists to get started with.CICD_Only- Setup CI/CD only, likely for monorepo setups or setting up CI/CD on an already initialized project. We expect Data Scientists to specifyProject_Onlyto get started in a development capacity, and when ready to move the project to Staging/Production, CI/CD can be set up. We expect that step to be done by Machine Learning Engineers (MLEs) who can specifyCICD_Onlyduring initialization and use the provided workflow to setup CI/CD for one or more projects.
input_root_dir: name of the root directory. When initializing withCICD_and_Project, this field will automatically be set toinput_project_name.input_cloud: Cloud provider you use with Databricks (AWS, Azure, or GCP).
Others must be correctly specified for CI/CD to work:
input_cicd_platform: CI/CD platform of choice (GitHub Actions or GitHub Actions for GitHub Enterprise Servers or Azure DevOps)input_databricks_staging_workspace_host: URL of staging Databricks workspace, used to run CI tests on PRs and preview config changes before they're deployed to production. We encourage granting data scientists working on the current ML project non-admin (read) access to this workspace, to enable them to view and debug CI test resultsinput_databricks_prod_workspace_host: URL of production Databricks workspace. We encourage granting data scientists working on the current ML project non-admin (read) access to this workspace, to enable them to view production job status and see job logs to debug failures.input_default_branch: Name of the default branch, where the prod and staging ML resources are deployed from and the latest ML code is staged.input_release_branch: Name of the release branch. The production jobs (model training, batch inference) defined in this repo pull ML code from this branch.
Or used for project initialization:
input_project_name: name of the current projectinput_read_user_group: User group name to give READ permissions to for project resources (ML jobs, integration test job runs, and machine learning resources). A group with this name must exist in both the staging and prod workspaces. Defaults to "users", which grants read permission to all users in the staging/prod workspaces. You can specify a custom group name e.g. to restrict read permissions to members of the team working on the current ML project.input_include_models_in_unity_catalog: If selected, models will be registered to Unity Catalog. Models will be registered under a three-level namespace of<catalog>.<schema_name>.<model_name>, according the the target environment in which the model registration code is executed. Thus, if model registration code runs in theprodenvironment, the model will be registered to theprodcatalog under the namespace<prod>.<schema>.<model_name>. This assumes that the respective catalogs exist in Unity Catalog (e.g.dev,stagingandprodcatalogs). Target environment names, and catalogs to be used are defined in the Databricks bundles files, and can be updated as needed.input_schema_name: If using Models in Unity Catalog, specify the name of the schema under which the models should be registered, but we recommend keeping the name the same as the project name. We default to using the sameschema_nameacross catalogs, thus this schema must exist in each catalog used. For example, the training pipeline when executed in the staging environment will register the model tostaging.<schema_name>.<model_name>, whereas the same pipeline executed in the prod environment will register the mode toprod.<schema_name>.<model_name>. Also, be sure that the service principals in each respective environment have the right permissions to access this schema, which would beUSE_CATALOG,USE_SCHEMA,MODIFY,CREATE_MODEL, andCREATE_TABLE.input_unity_catalog_read_user_group: If using Models in Unity Catalog, define the name of the user group to grantEXECUTE(read & use model) privileges for the registered model. Defaults to "account users".input_include_feature_store: If selected, will provide Databricks Feature Store stack components including: project structure and sample feature Python modules, feature engineering notebooks, ML resource configs to provision and manage Feature Store jobs, and automated integration tests covering feature engineering and training.input_include_mlflow_recipes: If selected, will provide MLflow Recipes stack components, dividing the training pipeline into configurable steps and profiles.
See the generated README.md for next steps!
Your organization can use the default stack as is or customize it as needed, e.g. to add/remove components or adapt individual components to fit your organization's best practices. See the stack customization guide for more details.
We recommend using separate dev/staging/prod Databricks workspaces for stronger isolation between environments. For example, Databricks REST API rate limits are applied per-workspace, so if using Databricks Model Serving, using separate workspaces can help prevent high load in staging from DOSing your production model serving endpoints.
However, you can create a single workspace stack, by supplying the same workspace URL for
input_databricks_staging_workspace_host and input_databricks_prod_workspace_host. If you go this route, we
recommend using different service principals to manage staging vs prod resources,
to ensure that CI workloads run in staging cannot interfere with production resources.
Yes. Currently, you can instantiate a new project and copy relevant components
into your existing project to productionize it. MLOps Stacks is modularized, so
you can e.g. copy just the GitHub Actions workflows under .github or ML resource configs
under {{.input_root_dir}}/{{template `project_name_alphanumeric_underscore` .}}/resources
and {{.input_root_dir}}/{{template `project_name_alphanumeric_underscore` .}}/databricks.yml into your existing project.
For this use case, we recommend instantiating via Databricks asset bundle templates
and copying the relevant subdirectories. For example, all ML resource configs
are defined under {{.input_root_dir}}/{{template `project_name_alphanumeric_underscore` .}}/resources
and {{.input_root_dir}}/{{template `project_name_alphanumeric_underscore` .}}/databricks.yml, while CI/CD is defined e.g. under .github
if using GitHub Actions, or under .azure if using Azure DevOps.
Yes. We provide the default stack in this repo as a production-friendly starting point for MLOps. However, in many cases you may need to customize the stack to match your organization's best practices. See the stack customization guide for details on how to do this.
Since MLOps Stacks is based on databricks CLI bundles, it's not limited only to ML workflows and resources - it works for resources across the Databricks Lakehouse. For instance, while the existing ML code samples contain feature engineering, training, model validation, deployment and batch inference workflows, you can use it for Delta Live Tables pipelines as well.
Please provide feedback (bug reports, feature requests, etc) via GitHub issues.
We welcome community contributions. For substantial changes, we ask that you first file a GitHub issue to facilitate discussion, before opening a pull request.
MLOps Stacks is implemented as a Databricks asset bundle template
that generates new projects given user-supplied parameters. Parametrized project code can be found under
the {{.input_root_dir}} directory.
To run tests, install actionlint,
databricks CLI, npm, and
act, then install the Python
dependencies listed in dev-requirements.txt:
pip install -r dev-requirements.txt
NOTE: This section is for open-source developers contributing to the default stack
in this repo. If you are working on an ML project using the stack (e.g. if you ran databricks bundle init
to start a new project), see the README.md within your generated
project directory for detailed instructions on how to make and test changes.
Run unit tests:
pytest tests
Run all tests (unit and slower integration tests):
pytest tests --large
Run integration tests only:
pytest tests --large-only
When making changes to MLOps Stacks, it can be convenient to see how those changes affect a generated new ML project. To do this, you can create an example project from your local checkout of the repo, and inspect its contents/run tests within the project.
We provide example project configs for Azure (using both GitHub and Azure DevOps), AWS (using GitHub), and GCP (using GitHub) under tests/example-project-configs.
To create an example Azure project, using Azure DevOps as the CI/CD platform, run the following from the desired parent directory
of the example project:
# Note: update MLOPS_STACKS_PATH to the path to your local checkout of the MLOps Stacks repo
MLOPS_STACKS_PATH=~/mlops-stacks
databricks bundle init "$MLOPS_STACKS_PATH" --config-file "$MLOPS_STACKS_PATH/tests/example-project-configs/azure/azure-devops.json"
To create an example AWS project, using GitHub Actions for CI/CD, run:
# Note: update MLOPS_STACKS_PATH to the path to your local checkout of the MLOps Stacks repo
MLOPS_STACKS_PATH=~/mlops-stacks
databricks bundle init "$MLOPS_STACKS_PATH" --config-file "$MLOPS_STACKS_PATH/tests/example-project-configs/aws/aws-github.json"
To create an example GCP project, using GitHub Actions for CI/CD, run:
# Note: update MLOPS_STACKS_PATH to the path to your local checkout of the MLOps Stacks repo
MLOPS_STACKS_PATH=~/mlops-stacks
databricks bundle init "$MLOPS_STACKS_PATH" --config-file "$MLOPS_STACKS_PATH/tests/example-project-configs/gcp/gcp-github.json"
