Amazon SageMaker

Amazon SageMaker (sagemaker) on fakecloud: the full 403-operation ML control plane — models, endpoints, endpoint configs, training/processing/transform/AutoML/tuning jobs, notebook instances, pipelines, feature groups, domains, model packages — at 100% conformance. awsJson1.1.

fakecloud implements the Amazon SageMaker control plane (sagemaker, SDK id SageMaker, endpoint prefix api.sagemaker) as an awsJson1.1 service. All 403 operations ship with 100% conformance against AWS's own Smithy model, backed by account-partitioned state that persists across restarts in persistent mode. SageMaker signs SigV4 with the sagemaker scope; every request is a POST / whose operation is selected by the X-Amz-Target: SageMaker.<Op> header, with all inputs carried in the JSON body (no HTTP path / label / query bindings).

The operation table, the per-operation model-derived input constraints, the output member shapes, the list element shapes, and each operation's resource family + identifier member are all generated directly from the Smithy model (scripts/generate-sagemaker-tables.py), so the control plane tracks the model exactly rather than by hand.

Uniform resource engine

SageMaker's ~130 resource families all follow the same shape: Create<X> takes an <X>Name, Describe<X> / Delete<X> / Update<X> take the same name, and List<X>s returns an array of <X>Summary structs. A single generic engine serves every family:

  • Create mints a proper arn:aws:sagemaker:<region>:<account>:<kind>/<name> ARN, stamps a numeric CreationTime / LastModifiedTime, persists every accepted input field, and returns the family's <X>Arn. A duplicate name returns the family's declared conflict error (ResourceInUse / ConflictException).
  • Describe echoes the persisted record — every field the create accepted, the minted ARN, and the timestamps.
  • List projects each record onto its <X>Summary (Name / Arn / CreationTime and any other summary members), paginating with NextToken.
  • Update merges the new fields and refreshes LastModifiedTime; Delete is idempotent.

A resource whose Describe identifier is a service-minted Id or ARN distinct from the create-time Name (for example Domain, ImageVersion, ModelCardExportJob) is resolved by scanning the family's minted identifiers, so a describe by the returned Id / ARN still round-trips.

What is real

Every modelled named resource mints proper ARNs, persists its attributes, and round-trips on read / list / update:

  • Models — CreateModel / DescribeModel / ListModels / DeleteModel, round-tripping the PrimaryContainer / Containers / ExecutionRoleArn / VpcConfig.
  • Endpoint configs and endpoints — production variants, data-capture config, async-inference config, all persisted and echoed on describe.
  • Jobs — training, processing, transform, labeling, compilation, AutoML (v1 + v2), and hyper-parameter-tuning jobs are created, persisted, described, and listed; Stop* is accepted.
  • Model packages (+ groups), pipelines, feature groups, domains, user profiles, spaces, apps, images (+ versions), experiments, trials (+ components), actions, artifacts, contexts, clusters, inference components, monitoring schedules, notebook instances (+ lifecycle configs), code repositories, workteams, workforces, and the rest of the ~130 families.
  • Tags — ARN-keyed AddTags / ListTags / DeleteTags.

Input validation is model-derived: required members, string @length, numeric @range, and @enum constraints are enforced, returning SageMaker's ValidationException. A missing resource returns ResourceNotFound; a duplicate create returns ResourceInUse. Timestamps are emitted as awsJson1.1 epoch-second JSON numbers. State is account-partitioned and persisted across restarts.

Honest emulation choices

This is the SageMaker control plane only. There is no ML execution plane: training / processing / transform / AutoML / tuning jobs are created, persisted, and described, but no container is scheduled, no model is trained, and no inference endpoint serves traffic. Jobs and endpoints are not advanced through a live lifecycle by a background scheduler — a described job reflects what was submitted, not a running computation. Presigned-URL operations (CreatePresignedNotebookInstanceUrl, CreatePresignedDomainUrl) and the lineage / search query operations are accepted as control-plane no-ops.