---

name: core-resources description: "Use SageMaker Python SDK v3 core resources for sessions, URI retrieval, low-level jobs and endpoints, serverless endpoint config, and lineage." metadata: disco-role: operating disable-model-invocation: true license: Apache 2.0

SageMaker Python SDK v3 Core Resources Sub-skill

Use this sub-skill when the task is about foundational sagemaker-core APIs: Session, get_execution_role, image/model/script URI retrieval, direct resource objects, processing jobs, batch transform jobs, endpoint lifecycle, direct training jobs, hyperparameter tuning jobs, serverless or async endpoint config, resource chaining, or lineage primitives.

Trigger phrases

• "core resources", "sagemaker-core", "low-level resource", or "resource object".
• "Session", "default bucket", "get_execution_role", boto_region_name, or AWS profile.
• "image URI", image_uris.retrieve, model_uris.retrieve, or script_uris.retrieve.
• "ProcessingJob", "batch transform", "TransformJob", or "offline inference".
• "Model", "EndpointConfig", "Endpoint", "invoke", or "update_weights_and_capacities".
• "TrainingJob" or "HyperParameterTuningJob" when the user wants the low-level API.
• "serverless endpoint config", ProductionVariantServerlessConfig, or AsyncInferenceConfig.
• "resource chaining", "pass a resource object through", or "get_all / refresh / delete".
• "lineage", sagemaker.core.lineage, Context, Action, Artifact, or Association.
• "migrate from v2 imports" for sagemaker.session, sagemaker.processing, sagemaker.transformer, sagemaker.tuner, sagemaker.model, sagemaker.predictor, sagemaker.image_uris, sagemaker.model_uris, sagemaker.script_uris, or sagemaker.lineage.

Read this when

• The request is direct SageMaker resource management and not a high-level ModelTrainer, ModelBuilder, or MLOps pipeline workflow.
• The user wants a v3 replacement for removed v2 modules or legacy low-level calls.
• The task is to create, list, refresh, delete, wait on, or invoke core resources.
• The task is to retrieve managed framework images or JumpStart model/script artifacts.
• The task is to construct lineage graphs, query associations, or clean up lineage entities.

Do not read this when

• The user wants general training orchestration, local training, distributed training, JumpStart training, or HPO with ModelTrainer; use ../training/SKILL.md.
• The user wants high-level deployment with ModelBuilder, inference schema inference, in-process serving, or optimization; use ../serving/SKILL.md.
• The user wants SageMaker Pipelines, registry flows, feature store, Clarify, or workflow orchestration; use ../mlops/SKILL.md.
• The user wants foundation-model customization or evaluation trainers; use ../model-customization/SKILL.md.

Non-negotiable v3 guardrails

• Use sagemaker.core, sagemaker.train, sagemaker.serve, and sagemaker.mlops imports by default; do not introduce v2 imports in new guidance except in explicitly labeled deprecated migration notes.
• Prefer the SDK over raw boto3.client("sagemaker").create_* calls for normal SageMaker tasks. Raw boto3 is an escape hatch only for unsupported APIs or when the user explicitly asks for the service call.
• Do not hardcode real account IDs, role ARNs, regions, bucket names, subnet IDs, security group IDs, or credentials. Use Session(), get_execution_role(), Session().default_bucket(), placeholders, or AWS configuration.
• Treat create, update, delete, invoke, and wait as cloud operations.
• Old sagemaker.lineage.* imports are deprecated compatibility shims. Prefer sagemaker.core.lineage.* for new lineage code.
• For low-level serverless endpoints, use ProductionVariantServerlessConfig inside ProductionVariant; reserve ServerlessInferenceConfig for helper flows and higher-level deployment code.

Short workflow

1. Classify the request. If it is clearly ModelTrainer, ModelBuilder, or MLOps orchestration, route to the sibling sub-skill instead of staying here.
2. Establish AWS prerequisites: region, credentials/profile, execution role, default bucket/prefix, and whether live service calls are authorized.
3. Create a Session and verify session.boto_region_name. Use get_execution_role(session) only when role discovery is expected to work; otherwise pass an explicit role ARN.
4. Resolve artifacts with image_uris.retrieve, model_uris.retrieve, or script_uris.retrieve. For JumpStart URIs, require model_id and model_version. For serverless image selection, pass a serverless config.
5. Build typed request shapes from sagemaker.core.shapes and sagemaker.core.inference_config.
6. Pick the resource API:
   • ProcessingJob.create/get/get_all/wait
   • TransformJob.create/get/get_all/wait
   • TrainingJob.create/get/get_all/wait
   • Model.create/get/get_all/delete
   • EndpointConfig.create/get/get_all
   • Endpoint.create/get/get_all/wait_for_status/invoke/update_weights_and_capacities
   • HyperParameterTuningJob.create/get/get_all/wait
7. Prefer resource chaining when the signature accepts object or PipelineVariable. Pass upstream resource objects directly, then fall back to get_name() or .name only when the target field requires a plain string.
8. For endpoints, create an EndpointConfig first, then Endpoint, then wait for InService before invoking.
9. For lineage, use sagemaker.core.lineage, delete associations before deleting entities, and prefer the new import path over the legacy shim.
10. Validate before returning: v3 imports only, no hardcoded real identifiers, cleanup path included, and any billable resources have a delete plan.

Reference map

• Detailed signatures, imports, and resource semantics: references/api-reference.md.
• End-to-end operational recipes: references/workflows.md.
• Error boundaries, recovery advice, and migration notes: references/troubleshooting.md.

Handoff expectations

• Produce self-contained v3 guidance; do not require the original repository checkout at runtime.
• Make all AWS calls explicit and distinguish them from local object construction.
• Keep cleanup steps in any example that creates jobs, endpoints, or lineage entities.
• If a user gives v2 code, convert it to v3 and label the v2 original as deprecated before showing the replacement.