Limits & FAQ for Git integration with Databricks Git folders

  • Article

Databricks Git folders and Git integration have limits specified in the following sections. For general information, see Databricks limits.

File and repo size limits

Azure Databricks doesn't enforce a limit on the size of a repo. However:

  • Working branches are limited to 500 megabytes (MB).
  • Files larger than 10 MB can't be viewed in the Azure Databricks UI.
  • Individual workspace files are subject to a separate size limit. For more details, read Limitations.

Databricks recommends that in a repo:

  • The total number of all files not exceed 10,000.
  • The total number of notebooks not exceed 5,000.

For any Git operation, memory usage is limited to 2 GB, and disk writes are limited to 4 GB. Since the limit is per-operation, you get a failure if you attempt to clone a Git repo that is 5 GB in current size. However, if you clone a Git repo that is 3 GB in size in one operation and then add 2 GB to it later, the next pull operation will succeed.

You might receive an error message if your repo exceeds these limits. You might also receive a timeout error when you clone the repo, but the operation might complete in the background.

To work with repo larger than the size limits, try sparse checkout.

If you must write temporary files that you do not want to keep after the cluster is shut down, writing the temporary files to $TEMPDIR avoids exceeding branch size limits and yields better performance than writing to the current working directory (CWD) if the CWD is in the workspace filesystem. For more information, see Where should I write temporary files on Azure Databricks?.

Maximum number of Git folders per workspace

You can have a maximum of 2,000 Git folders per workspace. If you require more, contact Databricks support.

Monorepo support

Databricks recommends that you do not create Git folders backed by monorepos, where a monorepo is a large, single-organization Git repository with many thousands of files across many projects.

Git folder configuration

Where is Azure Databricks repo content stored?

The contents of a repo are temporarily cloned onto disk in the control plane. Azure Databricks notebook files are stored in the control plane database just like notebooks in the main workspace. Non-notebook files are stored on disk for up to 30 days.

Do Git folders support on-premises or self-hosted Git servers?

Databricks Git folders supports GitHub Enterprise, Bitbucket Server, Azure DevOps Server, and GitLab Self-managed integration, if the server is internet accessible. For details on integrating Git folders with an on-prem Git server, read Git Proxy Server for Git folders.

To integrate with a Bitbucket Server, GitHub Enterprise Server, or a GitLab self-managed subscription instance that is not internet-accessible, get in touch with your Azure Databricks account team.

What Databricks asset types are supported by Git folders?

For details on supported asset types, read Manage file assets in Databricks Git folders.

Do Git folders support .gitignore files?

Yes. If you add a file to your repo and do not want it to be tracked by Git, create a .gitignore file or use one cloned from your remote repository and add the filename, including the extension.

.gitignore works only for files that are not already tracked by Git. If you add a file that is already tracked by Git to a .gitignore file, the file is still tracked by Git.

Can I create top-level folders that are not user folders?

Yes, admins can create top-level folders to a single depth. Git folders do not support additional folder levels.

Do Git folders support Git submodules?

No. You can clone a repo that contains Git submodules, but the submodule is not cloned.

Does Azure Data Factory (ADF) support Git folders?

Yes.

Source management

Why do notebook dashboards disappear when I pull or checkout a different branch?

This is currently a limitation because Azure Databricks notebook source files don't store notebook dashboard information.

If you want to preserve dashboards in the Git repository, change the notebook format to .ipynb (the Jupyter notebook format). By default, .ipynb supports dashboard and visualization definitions. If you want to preserve graph data (data points), you must commit the notebook with outputs.

To learn about committing .ipynb notebook outputs, see Allow committing .ipynb notebook output.

Do Git folders support branch merging?

Yes. You can also create a pull request and merge through your Git provider.

Can I delete a branch from an Azure Databricks repo?

No. To delete a branch, you must work in your Git provider.

If a library is installed on a cluster, and a library with the same name is included in a folder within a repo, which library is imported?

The library in the repo is imported. For more information about library precedence in Python, see Python library precedence.

Can I pull the latest version of a repository from Git before running a job without relying on an external orchestration tool?

No. Typically you can integrate this as a pre-commit on the Git server so that every push to a branch (main/prod) updates the Production repo.

Can I export a repo?

You can export notebooks, folders, or an entire repo. You cannot export non-notebook files. If you export an entire repo, non-notebook files are not included. To export, use the workspace export command in the Databricks CLI or use the Workspace API.

Security, authentication, and tokens

Issue with a conditional access policy (CAP) for Microsoft Entra ID

When you try to clone a repo, you might get a "denied access" error message when:

  • Azure Databricks is configured to use Azure DevOps with Microsoft Entra ID authentication.
  • You have enabled a conditional access policy in Azure DevOps and an Microsoft Entra ID conditional access policy.

To resolve this, add an exclusion to the conditional access policy (CAP) for the IP address or users of Azure Databricks.

For more information, see Conditional access policies.

Allow list with Azure AD tokens

If you use Azure Active Directory (AAD) for authenticating with Azure DevOps, the default allow list restricts Git URLs to:

  • dev.azure.com
  • visualstudio.com

For more information, see Allow lists restrict remote repo usage.

Are the contents of Azure Databricks Git folders encrypted?

The contents of Azure Databricks Git folders are encrypted by Azure Databricks using a default key. Encryption using customer-managed keys is not supported except when encrypting your Git credentials.

How and where are the GitHub tokens stored in Azure Databricks? Who would have access from Azure Databricks?

  • The authentication tokens are stored in the Azure Databricks control plane, and an Azure Databricks employee can only gain access through a temporary credential that is audited.
  • Azure Databricks logs the creation and deletion of these tokens, but not their usage. Azure Databricks has logging that tracks Git operations that can be used to audit the usage of the tokens by the Azure Databricks application.
  • GitHub enterprise audits token usage. Other Git services might also have Git server auditing.

Do Git folders support GPG signing of commits?

No.

Do Git folders support SSH?

No, only HTTPS.

Error connecting Azure Databricks to an Azure DevOps repo in a different tenancy

When trying to connect to DevOps in a separate tenancy, you might receive the message Unable to parse credentials from Azure Active Directory account. If the Azure DevOps project is in a different Microsoft Entra ID tenancy from Azure Databricks, you need to use an access token from Azure DevOps. See Connect to Azure DevOps using a DevOps token.

CI/CD and MLOps

Incoming changes clear the notebook state

Git operations that alter the notebook source code result in the loss of the notebook state, including cell outputs, comments, version history, and widgets. For example, git pull can change the source code of a notebook. In this case, Databricks Git folders must overwrite the existing notebook to import the changes. git commit and push or creating a new branch do not affect the notebook source code, so the notebook state is preserved in these operations.

Important

MLflow experiments don't work in Git folders with DBR 14.x or lower versions.

Can I create an MLflow experiment in a repo?

There are two types of MLflow experiments: workspace and notebook. For details on the two types of MLflow experiments, see Organize training runs with MLflow experiments.

In Git folders, you can call mlflow.set_experiment("/path/to/experiment") for an MLflow experiment of either type and log runs to it, but that experiment and the associated runs will not be checked into source control.

Workspace MLflow experiments

You cannot create workspace MLflow experiments in a Databricks Git folder (Git folder). If multiple users use separate Git folders to collaborate on the same ML code, log MLflow runs to an MLflow experiment created in a regular workspace folder.

Notebook MLflow experiments

You can create notebook experiments in a Databricks Git folder. If you check your notebook into source control as an .ipynb file, you can log MLflow runs to an automatically created and associated MLflow experiment. For more details, read about creating notebook experiments.

Prevent data loss in MLflow experiments

Notebook MLflow experiments created using Databricks Jobs with source code in a remote repository are stored in a temporary storage location. These experiments persist initially after workflow execution but are at risk of deletion later during scheduled removal of files in temporary storage. Databricks recommends using workspace MLflow experiments with Jobs and remote Git sources.

Warning

Any time you switch to a branch that does not contain the notebook, you risk losing the associated MLflow experiment data. This loss becomes permnanent if the prior branch is not accessed within 30 days.

To recover missing experiment data before the 30 day expiry, rename the notebook back to the original name, open the notebook, click the "experiment" icon on the right side pane (this also effectively calls the mlflow.get_experiment_by_name() API), and you will be able to see the recovered experiment and runs. After 30 days, any orphaned MLflow experiments will be purged to meet GDPR compliance policies.

To prevent this situation, Databricks recommends you either avoid renaming notebooks in repos altogether, or if you do rename a notebook, click the "experiment" icon on the right side pane immediately after renaming a notebook.

What happens if a notebook job is running in a workspace while a Git operation is in progress?

At any point while a Git operation is in progress, some notebooks in the repo might have been updated while others have not. This can cause unpredictable behavior.

For example, suppose notebook A calls notebook Z using a %run command. If a job running during a Git operation starts the most recent version of notebook A, but notebook Z has not yet been updated, the %run command in notebook A might start the older version of notebook Z. During the Git operation, the notebook states are not predictable and the job might fail or run notebook A and notebook Z from different commits.

To avoid this situation, use Git-based jobs (where the source is a Git provider and not a workspace path) instead. For more details, read Use version-controlled source code in an Azure Databricks job.

Resources

For details on Databricks workspace files, see What are workspace files?.