FAQ of AI Model Experiments

We want to give our customers the information they need to get the most from our STACKIT AI Model Experiments. This FAQ section answers common questions. This helps you quickly find solutions and improve your experience. We encourage you to check these FAQ before contacting our support team, as you might find your answer here.

• General information

  Expand all

  Can I use the service with my existing ML code?

  Absolutely. The service is compatible with the standard MLflow™ Python SDK and REST API. If you are already using MLflow™ locally, you only need to update your environment variables to point to the STACKIT instance. To ensure a seamless connection, ensure that your local SDK version is supported by the server, typically by using an SDK version that is not higher than the server’s current version. This prevents compatibility issues where newer client features are not yet recognized by the tracking service.

• Setup and configuration

  Expand all

  Why does the activation fail if I don't have Object Storage enabled?

  AI Model Experiments requires STACKIT Object Storage to store your model artifacts. A dedicated bucket is automatically provisioned for every MLflow™ instance you create, so Object Storage must be active in your region first.

  Are the Object Storage buckets and credentials created automatically?

  Yes. For every AI Model Experiments instance you create, STACKIT automatically provisions a dedicated bucket (model-experience-{bucket-id}) and a corresponding credential group (aiexp-{bucket-id}-client). This ensures that your model artifacts are stored securely and in isolation from your other data.

  I see the aiexp credential group, but where are my keys?

  While the Credential Group is auto generated, the actual Access Key and Secret Key are not displayed for security reasons. You must navigate to Object Storage > Credentials & Groups, select the aiexp group, and click Create credentials to generate the keys you need for your AWS_ACCESS_KEY_ID and AWS_SECRET_ACCESS_KEY environment variables.

  Can I use a different Object Storage bucket instead of the auto generated one?

  Yes. You can configure the MLflow™ SDK to point to any S3-compatible bucket by changing your environment variables. However, there is a significant trade-off: The managed MLflow™ UI is only integrated with the auto generated model-experience bucket. If you use a custom bucket:

  • SDK/API: You can still log and retrieve artifacts programmatically using your own tools or the MLflow™ Python library.

  • MLflow™ UI: The “Artifacts” tab in the hosted dashboard will be unable to display or preview your files (plots, models, etc.), as it does not have permissions or connectivity to your external buckets.

  What is the difference between a STACKIT Auth Token and a Tracking Token?

  • STACKIT Auth Token: Used to interact with the STACKIT API/Portal to manage the lifecycle of the service (e.g., creating or deleting instances).

  • Tracking Token: A scoped token generated for a specific MLflow™ instance. Engineers use this in their Python code (MLFLOW_TRACKING_TOKEN) to log experiment data.

  How do I find my instance URI?

  • Log into the STACKIT Portal.

  • Navigate to AI Model Experiments.

  • Your instance URI is linked directly within the specific instance.

  How do I find a bucket linked to an instance?

  • Log into the STACKIT Portal.

  • Navigate to AI Model Experiments.

  • Your bucket is linked directly within the specific instance.

  Why can’t I see my token value after I created it?

  The token value is only displayed once at the moment of creation for security purposes. STACKIT does not store the plain-text token. If you lose the token or it is no longer accessible, you will need to create a new one.

• Running experiments

  Expand all

  Which machine learning libraries are supported?

  Since the service uses standard MLflow™, it supports all major frameworks including XGBoost, PyTorch, TensorFlow, Scikit-learn, and more. You simply use the corresponding “flavor” (e.g., mlflow.xgboost.log_model). For a complete and up-to-date list of all built-in and community-supported flavors, we recommend visiting the official MLflow™ Built-in Model Flavors documentation.

  What happens if I delete an experiment?

  Experiments follow a Soft Delete policy. When deleted, they are marked for deletion but remain in the system for a grace period defined by the DeletedExperimentRetention setting (default is 30 days). After this period, the data is permanently deleted and cannot be recovered.

  How do I log in to the visual dashboard?

  You can access the MLflow™ UI using your standard STACKIT credentials via:

  • The STACKIT Portal: Click the direct link provided in your AI Model Experiments instance view.

  • Via API: When getting an instance via the API, copy the URL field from the response and paste it into your web browser.

• Troubleshooting

  Expand all

  I lost my tracking token. Can I recover it?

  No. For security reasons, tokens are only shown once during creation. If you lose a token, you must delete the old one and generate a new one in the Portal or via the API.

  I'm getting a libxgboost.dylib error. How do I fix it?

  This is a common local environment issue. You need to install the OpenMP runtime. Run the following command based on your operating system:

  On macOS

  If you are using Homebrew, run:

  brew install libomp

  On Linux (Debian/Ubuntu)

  Use the advanced package tool to install the runtime:

  sudo apt-get update && sudo apt-get install libomp-development

  Why can’t I create a new instance or generate a tracking token?

  In STACKIT AI Model Experiments, these actions are restricted to project administrators.

  • Admins manage the infrastructure: they provision instances, set up the storage buckets, and issue the tracking tokens.

  • Engineers focus on the experiments: they use the tokens provided by the admin to log training data and use the UI to analyze results.

  If you need a new environment or a fresh token, please contact your project administrator.

  I cannot access the hosted UI and receive an "Unauthorized" error. What should I do?

  If you receive an unauthorized error when trying to access the hosted UI, it is likely that your STACKIT user account lacks the necessary permissions for that specific project.

  Ensure your user has been assigned the correct role within the STACKIT project. At a minimum, your user requires the permission Model Experiments Reader to view and access instances. If you are an Administrator, you can grant this access via the IAM & Access section of the STACKIT Portal.

  My tracking token has expired. Can I extend or renew it?

  No. For security reasons, existing tokens cannot be prolonged or renewed once they reach their expiration date. If your token has expired, you must generate a new one.

  Why is my code creating a local folder instead of connecting to the managed AI Model Experiments server?

  If a local folder is generated with the same name as your MLFLOW_TRACKING_URI, the MLflow™ client has mistaken your remote server address for a local directory path. This occurs because the https:// prefix is missing; without it, MLflow™ assumes the URI is a local path rather than a network address. To ensure your data is sent to the managed server instead of being stored locally, you must include the full protocol in your tracking URI:

  Replace

  export MLFLOW_TRACKING_URI="<instance-id>.model-experiments.eu01.onstackit.cloud"

  with

  export MLFLOW_TRACKING_URI="https://<instance-id>.model-experiments.eu01.onstackit.cloud"