[raza-khan0108]

Description

Fixes #6019: prevents false-positive RoleValidationError exceptions during role resolution and validation (resolve_and_validate_role) when AWS accounts use condition-based Service Control Policies (SCPs) or condition-based IAM permissions boundaries.

Problem

sagemaker-core v2.15.0 introduced a pre-flight permission validation check (_evaluate_permissions via iam:SimulatePrincipalPolicy) that runs during high-level object construction (e.g., ModelTrainer(...), TrainDefaults.get_role()) before job submission.

Per AWS IAM Documentation, the IAM policy simulator does not evaluate conditional SCPs or permissions boundaries. When an AWS account uses condition-based SCPs (e.g., restricting regions, IP ranges, or required tags), iam:SimulatePrincipalPolicy returns:

• EvalDecision: implicitDeny
• OrganizationsDecisionDetail: {"AllowedByOrganizations": false} (or PermissionsBoundaryDecisionDetail: {"AllowedByPermissionsBoundary": "false"})

Because the previous implementation treated any non-allowed SimulatePrincipalPolicy verdict as a hard failure (RoleValidationError), valid roles in enterprise AWS Organizations environments were blocked from using high-level SDK v3 constructs (ModelTrainer, ModelBuilder, Pipeline).

---

Solution & Architectural Changes

This PR refactors the client-side role validation flow to gracefully handle simulator limitations and provides explicit opt-out controls:

1. Unverifiable Denial Detection

• Added helper _is_unverifiable_denial(result: dict) -> bool in sagemaker.core.helper.iam_role_resolver to inspect SimulatePrincipalPolicy evaluation items.
• If EvalDecision is implicitDeny due to AllowedByOrganizations: false or AllowedByPermissionsBoundary: false, the action is classified as unverifiable rather than definitively denied.

2. Partitioned Simulation Results & Warn-and-Proceed Behavior

• _simulate_denied_actions(...) now returns a tuple Tuple[List[str], List[str]] containing (denied_actions, unverifiable_actions).
• When _evaluate_permissions(...) encounters unverifiable actions (and zero definitive denied actions), it:
  1. Logs an informative WARNING message detailing that Organizations / SCP conditions prevented faithful simulation.
  2. Returns verdict = None, [] (unverifiable status) rather than False (denied).
• resolve_and_validate_role(...) treats verdict = None as a soft check: it warns the caller and proceeds with the role without raising RoleValidationError.
• verify_hyperpod_connect_permissions(...) also handles unverifiable decisions gracefully by logging an INFO message and returning None.

3. Client-Side Validation Opt-Out

• Added validate_role: bool = True optional parameter to resolve_and_validate_role(...). When set to False, client-side simulation and trust checks are skipped immediately.
• Added environment variable control SAGEMAKER_VALIDATE_ROLE: setting SAGEMAKER_VALIDATE_ROLE=false (or 0 / no) skips client-side validation globally without requiring code modifications.

4. Object-Oriented Interface Integration

• Exposed resolve_and_validate_role(...) as a method on the IamRoleResolver class so that call sites passing around IamRoleResolver instances can invoke validation directly while supporting the role_arn keyword alias.

---

Changes by File

File	Changes
sagemaker-core/src/sagemaker/core/helper/iam_role_resolver.py	• Added _is_unverifiable_denial helper. • Refactored _simulate_denied_actions to return (denied, unverifiable) tuple. • Updated _evaluate_permissions and verify_hyperpod_connect_permissions to log warnings and proceed on unverifiable verdicts. • Added validate_role parameter and SAGEMAKER_VALIDATE_ROLE env var check to resolve_and_validate_role. • Added resolve_and_validate_role method to IamRoleResolver class.
sagemaker-core/tests/unit/helper/test_iam_role_resolver.py	• Updated existing TestSimulateDeniedActions unit tests for tuple return signature. • Added test_partitions_unverifiable_denials covering conditional SCP and boundary evaluation results. • Added test_conditional_scp_unverifiable_denial_warns_and_proceeds verifying RoleValidationError is bypassed with correct warning log. • Added unit tests for validate_role=False, SAGEMAKER_VALIDATE_ROLE=false, IamRoleResolver().resolve_and_validate_role(), and HyperPod connect unverifiable checks.

---

Verification & Testing

Automated Unit Tests

Ran the full suite of unit tests for the helper module via pytest:

pytest sagemaker-core/tests/unit/helper

Results:

• test_iam_role_resolver.py: 79 passed (100%)
• Total helper suite (test_iam_role_creator.py, test_iam_role_resolver.py, test_session_helper.py): 204 passed in 2.14s

Backward Compatibility

• Fully backward compatible: existing resolve_and_validate_role(...) callers and unit tests continue to function without modification.
• Explicit and definitive IAM policy denials (explicitDeny or implicitDeny due to missing IAM role policy permissions) continue to raise RoleValidationError as expected.

---

Related Issues

• Closes sagemaker-core 2.15.0: role validation raises false-positive RoleValidationError under condition-based SCPs (IAM simulator can't evaluate conditional SCPs) #6019

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[mohamedzeidan2021]

The CI failures here are unrelated to this PR. The diff touches only two files — sagemaker-core/src/sagemaker/core/helper/iam_role_resolver.py and its unit test — and the checks that cover them all pass: unit-tests (sagemaker-core), integ-tests (sagemaker-core), and codecov/patch at 100% of diff.

Failing check	Cause	Related to this PR?
codestyle-doc-tests (core, mlops, train)	black --check wants to reformat 9 files already on master (workflow/job_config_document.py, workflow/pipeline_capture.py, mlops/workflow/steps.py, their tests, test_multi_turn_rl_trainer.py), plus a flake8 F401 unused PropertyMock import in test_multi_turn_rl_trainer.py:5	No
unit-tests (sagemaker-serve)	test_get_source_code_env_vars_with_s3_dir fails with ClientError (Throttling) ... SimulatePrincipalPolicy: Rate exceeded	No
integ-tests (sagemaker-mlops)	test_deployment_steps_execute_end_to_end: expected exactly one EndpointConfigName for this run, got []	No
integ-tests-us-east-1	Two Nova data-mixing tests hit ResourceLimitExceeded — the account is at its quota of 5 concurrent serverless model-customization jobs	No

Two notes:

• The codestyle-doc-tests failures reproduce on a clean master checkout (black --check reports the same files), so they will fail on any branch cut from current master and a rerun won't clear them. They need a separate formatting fix on master.
• The sagemaker-serve failure looks like a test defect worth fixing separately: tests/unit/test_model_builder_methods.py constructs ModelBuilder(model=Mock()) without mocking role resolution, so a unit test makes a live iam:SimulatePrincipalPolicy call and throttles under parallel CI.

The remaining three should clear on a rerun once the throttling and the job quota free up.