IAM for ML Workloads Advanced

ML workloads require access to compute, storage, container registries, and ML services. Implementing least-privilege IAM policies for these cross-service workflows is challenging but essential. This lesson provides practical IAM patterns for common ML scenarios on AWS.

Key IAM Roles for ML

RoleUsed ByKey Permissions
SageMaker ExecutionTraining jobs, endpointsS3 read/write, ECR pull, CloudWatch write
EC2 Training InstanceSelf-managed trainingS3 read, ECR pull, EFA access
Data ScientistHuman usersSageMaker full, S3 scoped, notebook access
ML PipelineStep Functions/AirflowSageMaker create/describe, S3, Lambda invoke

Least-Privilege SageMaker Execution Role

JSON
{
  "Version": "2012-10-17",
  "Statement": [
    {
      "Effect": "Allow",
      "Action": ["s3:GetObject", "s3:PutObject", "s3:ListBucket"],
      "Resource": [
        "arn:aws:s3:::ml-data-lake-*",
        "arn:aws:s3:::ml-artifacts-*"
      ]
    },
    {
      "Effect": "Allow",
      "Action": ["ecr:GetDownloadUrlForLayer", "ecr:BatchGetImage"],
      "Resource": "arn:aws:ecr:*:*:repository/ml-*"
    },
    {
      "Effect": "Allow",
      "Action": ["logs:CreateLogStream", "logs:PutLogEvents"],
      "Resource": "arn:aws:logs:*:*:log-group:/aws/sagemaker/*"
    }
  ]
}

Permission Boundaries

Use permission boundaries to cap the maximum permissions that ML roles can have, even if a data scientist creates their own roles:

  • Prevent launching instances larger than allowed (cost control)
  • Restrict regions to approved locations only
  • Block access to sensitive S3 buckets outside the ML domain
  • Prevent IAM escalation by limiting role creation permissions
Security Principle: Never attach the AmazonSageMakerFullAccess managed policy to production roles. It grants far more access than needed. Always create custom policies scoped to specific resources.

Ready for Best Practices?

The final lesson covers AWS Well-Architected ML practices and operational excellence.

Next: Best Practices →
← VPC SetupBest Practices →