Ejemplo: SageMaker Pipeline

Pipeline completo en AWS SageMaker con model registry y aprobación manual.

Prerrequisitos

pip install godml[aws]

export AWS_DEFAULT_REGION=us-east-1
export SAGEMAKER_ROLE_ARN=arn:aws:iam::123456789012:role/SageMakerRole

# Subir dataset a S3
aws s3 cp data/churn.csv s3://mi-bucket/data/churn.csv

godml.yml

name: churn-sagemaker
version: 1.0.0
provider: sagemaker

dataset:
  uri: s3://mi-bucket/data/churn.csv
  target: churned

governance:
  compliance: pci-dss
  policy: mask_sensitive

aws:
  role_arn: ${SAGEMAKER_ROLE_ARN}
  region: us-east-1
  s3_bucket: mi-bucket
  s3_prefix: godml

compute:
  preprocessing: ml.m5.large
  training: ml.m5.2xlarge
  evaluation: ml.m5.large

model:
  type: xgboost
  hyperparameters:
    max_depth: 6
    eta: 0.1
    n_estimators: 300

metrics:
  - name: auc
    threshold: 0.82

registry:
  model_package_group: godml-churn-models
  approval: manual

Ejecutar

godml run -f godml.yml

Building SageMaker Pipeline: churn-sagemaker
Upserting pipeline definition to SageMaker...
Starting pipeline execution...

Pipeline running. ARN: arn:aws:sagemaker:us-east-1:123456789012:pipeline/churn-sagemaker/execution/abc123

Waiting for pipeline to finish...
  Preprocessing:            Succeeded  (3m 21s)
  Training:                 Succeeded  (8m 47s)
  Evaluation:               Succeeded  (2m 12s)
  CheckMetricAndRegister:   Succeeded

Pipeline completed. AUC: 0.8431 ✅
Model registered: godml-churn-models v1 (PendingManualApproval)

Aprobar el modelo en AWS Console

Una vez que el pipeline termina:

1. Ve a SageMaker → Model Registry → godml-churn-models
2. Busca la versión con status PendingManualApproval
3. Click Update Status → Approved
4. El modelo está listo para desplegarse

O con el SDK:

import boto3
sm = boto3.client("sagemaker", region_name="us-east-1")
sm.update_model_package(
    ModelPackageArn="arn:aws:sagemaker:...",
    ModelApprovalStatus="Approved",
)

Costos estimados

Para un dataset de ~100K filas con XGBoost:

Step	Instancia	Duración	Costo
Preprocessing	ml.m5.large	~3 min	~$0.006
Training	ml.m5.2xlarge	~8 min	~$0.06
Evaluation	ml.m5.large	~2 min	~$0.004
Total		~13 min	~$0.07

Los steps se levantan y apagan solos — solo pagas el tiempo de ejecución.

---

→ Volver a Providers: SageMaker