Permissions for Databricks AI models
Full reference of the privileges required to crawl AI models and extract lineage from Databricks Unity Catalog, including what each privilege enables and how to grant it.
This reference covers the privileges the Atlan service account needs to crawl AI models and extract lineage from Databricks Unity Catalog, what each privilege enables, and how to grant it.
For standard Databricks connector permissions, see Set up Databricks.
Model metadata privileges
Atlan reads models and model versions through the Unity Catalog API. For AI models, the service account needs one privilege: READ METADATA on each catalog that contains models.
Grant READ METADATA on its own. The Data Reader preset also works, but it adds data access (SELECT, READ VOLUME) and EXECUTE, which Atlan doesn't need for AI models.
| Privilege | Scope | What it enables |
|---|---|---|
READ METADATA | Catalog containing models | Lets the service account see the catalog's schemas, models, and model versions, including the MLflow run ID of each version that Atlan uses to build model lineage. Without this, Unity Catalog doesn't return model versions, so Atlan can't crawl versions or build model lineage. |
READ METADATA doesn't let the service account read table data, load a model, or download model files from Unity Catalog. It does let the service account see the metadata of other objects in the catalog, such as table and column names, comments, and grants.
Grant it on each catalog that contains models. It applies to every schema, model, and model version in the catalog. A catalog-level grant doesn't need USE CATALOG or USE SCHEMA:
GRANT READ METADATA ON CATALOG <catalog_name> TO <atlan_user_or_role>;
Earlier versions of this page also listed EXECUTE, USE CATALOG, USE SCHEMA, READ VOLUME, and SELECT. Atlan doesn't need them for AI models. If you granted them only for AI models, you can revoke them after you grant READ METADATA. Keep any that the standard Databricks connector permissions need for crawling tables in the same catalog.
Notebook and experiment privileges
For each model version, Atlan reads the linked MLflow run: its parameters, metrics, timing, and input datasets. Atlan builds model lineage from the input datasets. MLflow controls access to a run through the experiment that holds it, not through Unity Catalog.
Grant CAN READ on each MLflow experiment linked to a model version. If an experiment belongs to a notebook, it takes its permissions from that notebook, so grant CAN READ on the notebook instead.
To avoid granting access one experiment at a time, grant CAN READ on a workspace folder that holds the experiments and notebooks. Objects in a folder inherit its permissions, including objects added later.
Grant experiment access at the same time as READ METADATA. If Atlan can read a model version but not its MLflow run, Atlan may skip that model.
CAN READ also lets the service account list and download a run's artifacts. If a run logged the model, its artifacts include a copy of the model files.
Feature store privileges (optional)
Required only if your workspace uses Databricks Feature Store models and the feature_spec.yaml artifact is stored in an external location.
Atlan reads feature_spec.yaml to discover the upstream feature views, tables, and functions a model was trained on. If this file is inaccessible, Atlan skips Feature Store lineage for that model version and falls back to run-based lineage where available.
| Privilege | Scope | What it enables |
|---|---|---|
READ FILES | External location storing feature_spec.yaml | Lets the service account read the Feature Store artifact. Without this, Feature Store lineage is skipped. |
Grant using:
GRANT READ FILES ON EXTERNAL LOCATION <external_location_name> TO <atlan_user_or_role>;
Repeat for each external location that stores Feature Store artifacts for models you want to crawl.
See also
- Crawl Databricks AI models: Configure the crawler to discover and catalog AI models.
- Extract Databricks AI model lineage: Build lineage from AI model versions to upstream data assets.