mlflow-auth-coverage

name	mlflow-auth-coverage
description	Adds auth coverage to the Kubernetes auth plugin for a new MLflow version. Use when the user asks to add auth support for a new MLflow release, update mlflow-integration for a new version, or create auth rules for new MLflow endpoints.

MLflow Auth Coverage

Add authorization rule coverage to the kubeflow/mlflow-integration Kubernetes auth plugin for a new MLflow minor release.

Before writing code, read reference.md for file patterns, auth mapping conventions, and review-established conventions.

Prerequisites

1. The target MLflow version is released on PyPI.
2. The local mlflow-integration repo is checked out with a remote pointing to kubeflow/mlflow-integration.
3. The pyproject.toml upper bound currently excludes the target version.

Workflow

1. Create a branch from the upstream main.
2. Discover new endpoints — install the target MLflow version and run the startup validator to find uncovered endpoints. Also check the MLflow changelog and release notes for new API surface.
3. Inspect proto fields — for each new proto request class, check which fields carry identifiers (experiment_id, schema_id, etc.) to determine the correct resource name parser. For path-based endpoints, read the handler source to find required parameters.
4. Present a checkpoint — summarize all discovered endpoints, proposed verb mappings, resource scoping decisions, and which endpoints get experiment-level vs workspace-level access.
5. Implement — modify files in the order listed in reference.md: pyproject.toml, .github/workflows/tests.yml, mlflow_kubernetes_plugins/auth/_compat.py, mlflow_kubernetes_plugins/auth/rules_v3_XX.py (new), mlflow_kubernetes_plugins/auth/rules.py, then tests/test_auth.py.
6. Test across versions — run the full test suite with the target version, the previous version, and the oldest supported version (currently 3.10). Confirm the startup validator passes on the target version.
7. Lint and build — make python-lint and uv build.
8. End with a summary — show the final diff, test results across all versions, and any open items. Do not push or create a PR. The user will request those explicitly when ready.

Endpoint Discovery

Install the target version and use the startup validator to find gaps:

make install-dev
uv pip install "mlflow==<version>"
uv run --no-sync pytest -xvs -k "validate" tests/test_auth_fastapi.py

Any MlflowException listing uncovered endpoints tells you exactly what needs rules.

For proto-defined endpoints, find which proto module they live in and inspect fields:

uv run --no-sync python3 -c "
import importlib
pb = importlib.import_module('mlflow.protos.<module>_pb2')
for name in ['CreateFoo', 'GetFoo', 'ListFoos']:
    cls = getattr(pb, name)
    fields = [f.name for f in cls.DESCRIPTOR.fields]
    print(f'{name}: {fields}')
"

For path-based (non-proto) endpoints, check handler source for required parameters:

uv run --no-sync python3 -c "
import inspect
from mlflow.server.handlers import _some_handler
print(inspect.getsource(_some_handler))
"

Check whether any path-based endpoint carries experiment_id as a required parameter. If so, it needs resource_name_parsers for experiment-level RBAC.

Auth Mapping Conventions

• All MLflow "list" operations map to K8s get verb, not list. The get verb is treated as blanket read access.
• All MLflow "create", "update", "delete", "add", "remove", "set" operations map to update verb.
• Experiment-scoped resources use _experiments_rule(...) with RESOURCE_NAME_PARSER_EXPERIMENT_ID_TO_NAME when experiment_id is in the request.
• ID-only endpoints (schema_id, queue_id, etc.) with no tracking store lookup API use workspace-level access (no resource_name_parsers). Do not add raw database queries for ID resolution.
• New version-specific routes go ONLY in the delta file (rules_v3_XX.py), never in rules_base.py.
• Gateway proxy routes use _gateway_endpoints_use_rule with RESOURCE_NAME_PARSER_GATEWAY_PROXY_ENDPOINT_NAME.
• Gateway guardrails are workspace-scoped (not experiment sub-resources), so they use their own verbs (create, delete, list) rather than always mapping to update.

Stop And Ask

Stop and ask the user when:

• A new endpoint introduces a resource type not in the current ALLOWED_RESOURCES set.
• The endpoint has a novel auth pattern not covered by existing AuthorizationRule fields.
• Proto imports fail or the endpoint registration pattern is unclear.
• The startup validator still fails after adding rules.
• The user asks to push, create a PR, or make any remote change — use the commit/PR title format fix: add MLflow <version> auth coverage and the official repo PR template from .github/PULL_REQUEST_TEMPLATE.md.

Examples

User request examples that should trigger this skill:

• "Add auth coverage for MLflow 3.15."
• "Update mlflow-integration to support the new MLflow release."
• "Create auth rules for the new MLflow endpoints."
• "Add 3.16 support to the Kubernetes auth plugin."