Deploying a Machine Learning Project with AWS CloudFormation

Turning a manual, click-through AWS deployment into a reviewable, repeatable CloudFormation stack.

Deploying a Machine Learning Project with AWS CloudFormation

Clicking through the AWS console to deploy a model works once. Infrastructure as code makes it reproducible: the same template spins up dev, staging, and prod, and can be reviewed and versioned like any other code.

Core Building Blocks

A typical ML inference stack needs:

• An S3 bucket to store model artifacts
• An IAM role with least-privilege access for the compute layer
• A compute target — a SageMaker endpoint for managed hosting, or Lambda/ECS for lighter-weight inference

A Minimal Stack

Resources:
  ModelArtifactsBucket:
    Type: AWS::S3::Bucket

  SageMakerExecutionRole:
    Type: AWS::IAM::Role
    Properties:
      AssumeRolePolicyDocument:
        Statement:
          - Effect: Allow
            Principal:
              Service: sagemaker.amazonaws.com
            Action: sts:AssumeRole
      ManagedPolicyArns:
        - arn:aws:iam::aws:policy/AmazonSageMakerFullAccess

  Model:
    Type: AWS::SageMaker::Model
    Properties:
      ExecutionRoleArn: !GetAtt SageMakerExecutionRole.Arn
      PrimaryContainer:
        Image: <ecr-image-uri>
        ModelDataUrl: !Sub "s3://${ModelArtifactsBucket}/model.tar.gz"

  EndpointConfig:
    Type: AWS::SageMaker::EndpointConfig
    Properties:
      ProductionVariants:
        - ModelName: !GetAtt Model.ModelName
          VariantName: AllTraffic
          InitialInstanceCount: 1
          InstanceType: ml.m5.large

  Endpoint:
    Type: AWS::SageMaker::Endpoint
    Properties:
      EndpointConfigName: !GetAtt EndpointConfig.EndpointConfigName

Deploying the Stack

aws cloudformation deploy \
  --template-file stack.yaml \
  --stack-name ml-inference-stack \
  --capabilities CAPABILITY_IAM

Tips

• Parameterize the instance type and image URI so the same template works across environments.
• Use one stack per environment (dev/staging/prod) rather than branching logic inside a single template.
• Tear down with aws cloudformation delete-stack when you’re done — an idle SageMaker endpoint bills by the hour.

CloudFormation won’t replace judgment about architecture, but it turns “how did we deploy this?” from a Slack archaeology exercise into a file you can read.