MLflow SharingHub¶

MLflow is a platform to streamline machine learning development, including tracking experiments, packaging code into reproducible runs, and sharing and deploying models.

For SharingHub, we developed a plugin for MLflow, MLflow SharingHub, that enables model tracking per AI Model project in SharingHub. To be more clear, each AI model project in SharingHub have a corresponding MLflow tracking URI, like https://sharinghub.example.com/mlflow/<gitlab-project-path>/tracking/.

Permissions¶

The permission system is based on the permission system of SharingHub, based upon GitLab. That means only collaborators of the GitLab repository can use the tracking URI, people with read-only access on the project will have read-only access in mlflow, and no access will not display anything.

Permissions mapping with GitLab (for experiments, runs, model registry...):

• Create: Owner, Maintainer, Developer
• Read: Owner, Maintainer, Developer, Reporter, Guest
• Update: Owner, Maintainer, Developer
• Delete: Owner, Maintainer, Developer

Authentication¶

Web¶

MLflow SharingHub uses SharingHub authentication session, meaning that if you login in SharingHub, you should also be authenticated in MLflow SharingHub.

API¶

You can authenticate with the MLflow API by setting an environment variable, MLFLOW_TRACKING_TOKEN, with an access token as value.

The MLflow code generator can help you:

Limitations¶

It is important to note that there are some limitations and constraints on MLflow SharingHub. It is a plugin, on top of MLflow, meaning MLflow particularities impose restrictions. Despite having multiple tracking URIs, we have a single server and database. In MLflow, experiments names, and model names (model registry) are unique, if I name an experiment, no one can create another experiment with the same name.

If you use MLflow SharingHub, and use project A tracking URI, another person on project B could want to use the same name as you for an experiment, creating name collision, throwing an error. MLflow experiments and models (in Model Registry) requires a suffix with the GitLab project ID, of the format <experiment-name> (<project-id>). This avoid collision across all projects.