[FEATURE] Native SDK support for downloading MLflow artifacts of Unity Catalog model versions

[asanglard]

Problem Statement

There is no method in the Databricks Python SDK to download the artifact files of a Unity Catalog registered model version. The ModelVersionsAPI (w.model_versions) provides no download capability, and the ExperimentsAPI (w.experiments) only exposes list_artifacts() for enumeration — not download.

The only working approach today requires calling an undocumented internal endpoint directly to obtain presigned cloud URLs (Azure SAS / AWS presigned / GCP signed), and then downloading from those URLs manually:

# Currently required workaround — undocumented endpoint:
raw = client.api_client.do(
    "POST",
    "/api/2.0/mlflow/artifacts/credentials-for-read",
    body={"run_id": run_id, "path": ["model/model.pkl"]},
)
signed_uri = raw["credential_infos"][0]["signed_uri"]
# then manually download from the presigned URL

This endpoint (POST /api/2.0/mlflow/artifacts/credentials-for-read) is the same one the mlflow Python package uses internally via DatabricksMlflowArtifactsService.GetCredentialsForRead (defined in databricks_artifacts.proto), but it is not surfaced anywhere in the Databricks SDK.

Proposed Solution

Add a download_artifacts() method (or equivalent) to ExperimentsAPI or ModelVersionsAPI that wraps the presigned URL credential fetch and file download:

# Desired SDK interface — option A (on ExperimentsAPI):
w.experiments.download_artifacts(run_id=run_id, path="model/", dst_path="/local/dir")

# Desired SDK interface — option B (on ModelVersionsAPI):
w.model_versions.download_artifacts(
    full_name="catalog.schema.model", version=1, dst_path="/local/dir"
)

Additional Context

• SDK version: 0.97.0
• The mlflow Python package implements this via MlflowClient.download_artifacts(), but using it in a multi-tenant application requires mutating process-global environment variables (DATABRICKS_HOST, DATABRICKS_TOKEN) protected by a threading lock — which blocks the entire process under async concurrency. A native SDK implementation would be safe for concurrent multi-tenant use since each WorkspaceClient instance is independent.
• The underlying mechanism fetches a temporary presigned cloud URL per artifact path, then downloads directly from cloud storage (no Databricks credentials needed for the download itself). This is already how mlflow works under the hood.
• Related undocumented endpoint: POST /api/2.0/mlflow/artifacts/credentials-for-read (protobuf: DatabricksMlflowArtifactsService.GetCredentialsForRead in databricks_artifacts.proto).
• Related issue: [FEATURE] Native SDK support for reading and writing MLflow tags on Unity Catalog model versions #1313 (UC model version tags)