SageMaker Production Defaults
Deploy SageMaker real-time or async endpoints with autoscaling, CloudWatch alarms, and consistent tagging built in.
The evidence shows autoscaling, CloudWatch alarms, consistent tagging, opt-in data capture, and a teardown command that preserves IAM, S3 objects, and model artifacts. However, it directly creates billable SageMaker, S3, CloudWatch, and autoscaling resources without an explicit confirmation step, accepts arbitrary image URIs, IAM role ARNs, environment variables, and S3/SNS destinations, and lacks automatic rollback after partial failure, so points are deducted.
The scripts provide argument validation, endpoint polling, surfaced failure reasons, BOM handling, and staged error feedback, making the happy path reasonably coherent. However, no committed test suite is supplied; boto3/AWS CLI dependencies and API behavior are not verified by the files; later autoscaling or alarm failures can leave orphaned resources; and same-name model reuse does not verify configuration equivalence, so points are deducted.
The description clearly covers real-time and async SageMaker endpoints, vLLM/TEI deployments, long-running workloads, sparse traffic, and some Windows usage. Trigger boundaries, non-SageMaker exclusions, and region/network constraints are incomplete, while core operation depends on AWS, S3, CloudWatch, SNS, and potentially the HF Hub; mainland-China reachability and Chinese-language fit are not evidenced, so points are deducted.
The skill has well-layered instructions, a reference design document, scripts, examples, defaults tables, troubleshooting, teardown guidance, stable-looking names, and repository-level Apache-2.0 licensing. The selected skill lacks its own version, changelog, maintainer, update path, complete dependency installation notes, FAQ, test entry point, and documented API compatibility limits, so points are deducted.
The scripts cover model and endpoint configuration, real-time and async deployment, autoscaling, alarms, smoke invocation, and teardown, and emit machine-readable JSON; static evidence supports completion of the core deployment task. However, execution was not performed, smoke testing is not enforced by the deploy scripts, defaults can create avoidable cost or throughput mismatch, and partial cleanup plus the async no-alarms combination create usability risks, so the static ceiling applies and points are deducted.
The source exposes auditable code, parameters, defaults, logs, and failure paths, with limited repository-level context from README and LICENSE. It contains no committed tests, CI coverage, or third-party execution records; key AWS/Hugging Face claims remain documentary assertions and cannot be reproduced statically, so points are deducted.
- Confirm the image, IAM role, environment variables, S3/SNS destinations, and expected cost before deployment; these parameters create external resources and ongoing charges.
- Real-time deployment registers autoscaling and alarms only after InService; later failures may leave billable resources, so inspect state and run teardown as needed.
- Avoid combining --no-alarms with async min-capacity=0; the documented wake-from-zero alarm may then be absent.
- Data capture writes requests and responses to S3 and may contain sensitive data; confirm permissions, retention, and compliance requirements first.
- The scripts were not executed, so AWS API behavior, dependency versions, and end-to-end results remain unverified; publisher identity is treated as unknown.
What does this skill do, and when should you use it?
This skill from the Hugging Face Skills repository handles the final deployment step after the image URI and IAM role are known. It creates the SageMaker model, endpoint configuration, and endpoint, then enables autoscaling, CloudWatch alarms, and consistent tags by default. It supports real-time inference and asynchronous inference with genuine scale-to-zero behavior. A deployment is not considered complete until a real invocation succeeds and endpoint logs have been checked for worker crashes.
Runs the bundled scripts/deploy.py to create real-time SageMaker models, endpoint configurations, and endpoints, register invocation-based target tracking, and create latency, error, and platform-overhead alarms. Runs scripts/deploy_async.py to create asynchronous endpoints, backlog alarms, and both target-tracking and step-scaling policies needed to wake an endpoint from zero capacity. The scripts accept environment variables, instance settings, S3 model artifacts, data-capture options, SNS notifications, and capacity limits; wait for InService; and emit JSON containing resource names and teardown information. invoke_endpoint.py validates real-time requests while handling JSON and UTF-8 BOM issues. Async invocation uploads input to S3, calls invoke-endpoint-async, and polls the output location.
- A team has selected a serving image and IAM role and needs a production-oriented real-time LLM endpoint with scaling and alarms.
- An engineer is deploying a TEI embedding model and needs a higher per-instance invocation target to avoid unnecessarily aggressive scaling.
- A team handles long-running, large, or sparse workloads and wants an async SageMaker endpoint that scales to zero between batches.
- A deployment operator needs to verify actual inference and diagnose an endpoint that is InService but has a crashing Python worker.
- An AWS user needs consistent resource tags and an ordered teardown path for later cleanup.
What are this skill's strengths and limitations?
- Real-time and async deployments default to autoscaling, CloudWatch alarms, and consistent tags.
- Async mode genuinely supports `MinCapacity=0` and automatically configures wake-from-zero step scaling.
- Bundled helpers cover real-time invocation, async invocation, failure diagnosis, and ordered teardown.
- Data capture is opt-in, avoiding ongoing S3 capture costs unless requested.
- It assumes the image URI, usable IAM role, region, instance type, and required AWS permissions are already available; it is not the full planning or preflight workflow.
- AWS endpoint, monitoring, storage, and related costs still apply; data capture additionally requires S3 write access.
- The supplied source provides no automated test suite or cross-platform test results.
- The default target of 20 invocations per instance can be too low for high-throughput embedding deployments and may require tuning.
How do you install this skill?
Copy or symlink the hf-cloud-sagemaker-production-defaults folder from the repository's skills/ directory into a standard Codex .agents/skills location, such as $REPO_ROOT/.agents/skills or $HOME/.agents/skills. The README does not document a dedicated install command for this individual skill; the repository supports loading Agent Skills through compatible clients. If the installed copy lacks scripts/, obtain them from the source repository rather than reimplementing them from the description.
How do you use this skill?
After AWS context discovery, IAM preparation, and image selection, prompt an agent: “Use the SageMaker production defaults skill to deploy this model, then run the smoke test and CloudWatch worker-crash log check.” For a real-time endpoint, run python scripts/deploy.py --model-name <name> --image-uri <uri> --role-arn <role-arn> --instance-type <type> --region <region>; vLLM deployments also require the model-specific environment variables and --inference-ami-version for cu130+ images. For async inference, run python scripts/deploy_async.py and provide --output-s3-uri. Report completion only after a real request succeeds and the log scan passes. Tear down with python3 scripts/teardown.py <endpoint-name> <region>.
How does this skill compare with similar options?
The source does not provide a direct comparison with other deployment tools or skills; it positions this skill as the final step in the SageMaker deployment workflow.