Skip to main content
Last updated on

Configure a serverless budget policy for an MLflow experiment

Beta

This feature is in Beta. Workspace admins can control access to this feature from the Previews page. See Manage Databricks previews.

Set a serverless budget policy on an MLflow experiment to control which policy MLflow uses for serverless workloads it runs against the experiment. Affected workloads include scheduled scorers, synthetic evaluation set generation, and agent evaluation.

By default, these workloads use the workspace's default serverless budget policy. If your workspace disables the default policy (for example, when each user and service principal must select a dedicated policy), MLflow cannot pick a fallback and registering a scorer or running an evaluation fails with the following error:

Text
403 Client Error: Forbidden
PERMISSION_DENIED: Unable to use fallback policies, please manually select a policy.

Set the policy on the experiment to unblock these workflows. MLflow then uses that policy for every serverless workload it creates for the experiment.

Requirements​

  • Permission to use the budget policy you want to set. Users and service principals can only assign a policy that they are entitled to use.

Set the budget policy in the UI​

  1. Open the MLflow experiment.
  2. In the experiment Details panel, set the Budget policy to a policy you have access to use.

MLflow uses this policy for serverless workloads it creates on behalf of the experiment.

Set the budget policy with the API​

Use mlflow.set_experiment_tag() to set the mlflow.workload_creation_policy_id tag on the experiment:

Python
import mlflow

mlflow.set_experiment_tag(
    experiment_id="<your-experiment-id>",
    key="mlflow.workload_creation_policy_id",
    value="<your-policy-id>",
)

After the tag is set, subsequent calls that create serverless workloads from the experiment (such as Scorer.register()) use the specified policy.

To find the ID of a budget policy, see Attribute usage with serverless usage policies.

Next steps​

On this page