[lucasjia-aws]

Summary

The scheduled CI Health V2 workflow (slow-tests-v2 and canaries-v2-master, run against master-v2) has been red every day for weeks. Example run: https://github.com/aws/sagemaker-python-sdk/actions/runs/37412457307. This PR fixes the test-side causes; no SDK source changes.

Root causes and fixes

test_sagemaker_config_cross_context_injection (fails every run): DescribeEndpoint/DescribeEndpointConfig now return MetricsConfig: {EnableDetailedObservability: True} by default, and the default VolumeSizeInGB for ml.c5.xlarge changed from 4 to 8. Neither is set by the SDK or the sagemaker_config, so the test now drops them before comparing.

lineage test_context.py::test_action (ERROR every run since 2026-09-28): The autouse cleanup_endpoints fixture in tests/integ/sagemaker/serve/test_base_model_builder_deploy.py deleted every endpoint on the first ListEndpoints page of the shared account before and after each test. On 2026-09-28 it deleted the lineage static endpoint SdkIntegTestStaticEndpoint20 (confirmed in CloudTrail), along with endpoints other running tests were using. ListEndpoints does not return EndpointConfigName, so the endpoint config was left behind; the lineage fixture's redeploy then failed on CreateEndpointConfig, the ValidationException was swallowed as "already exists", and the test ended in list_contexts(SourceUri=None). The sweep is replaced by per-test cleanup of only the resources each test creates, and the lineage fixture now recreates the endpoint from the surviving config and fails with a clear message if it still cannot.

test_deployment_steps_execute_end_to_end (flaky): SageMaker applies NameContains per result page, so a single unpaginated list_endpoint_configs(NameContains=stamp) returns an empty page once more than 10 newer configs exist, which is routine while the canary runs 120 workers in the same account. The same unpaginated calls in cleanup also leaked the test's endpoint. Lookups and cleanup now page through results, with CreationTimeAfter and page size 100 to keep paging cheap.

test_advanced_right_size_and_deploy_unregistered_model_sklearn: CreateInferenceRecommendationsJob was throttled when several module fixtures started IR jobs at once. retry_and_back_off received the already-evaluated right_size() result, so it never retried, and the fixture swallowed the exception and returned None, so the test reported AttributeError/UnboundLocalError instead. The helper now takes a callable, retries throttling with jittered back-off (deleting the model right_size() pre-created before retrying), fixtures re-raise after cleanup, and tests guard predictor in finally.

Insufficient capacity: test_jumpstart_gated_model (ml.g6e.xlarge), the private hub gated model/estimator tests (ml.g5.2xlarge) and test_huggingface_torch_distributed_g5_glue (ml.g5.4xlarge) fail on InsufficientInstanceCapacity or wait for capacity until timeout. They now use the existing x_fail_if_ice (and fit_estimator_with_capacity_xfail for the HF training job), matching how sibling tests already handle capacity.

Testing

black (24.3.0) and flake8 (7.1.2) pass on all changed files, and pytest --collect-only collects all affected integ modules. Offline mock checks covered the retry helper, the paginated lookups (parameters validated with botocore Stubber), the scoped endpoint cleanup, and all lineage endpoint recovery paths. The integ tests themselves were not run locally because they need the CI account; they will be exercised by the next CI Health V2 run.