⚠️ BentoCTL project has been deprecated

Plese see the latest BentoML documentation on OCI-container based deployment workflow: https://docs.bentoml.com/

AWS Sagemaker Operator

Sagemaker is a fully managed service for building ML models. BentoML provides great support for deploying BentoService to AWS Sagemaker without the additional process and work from users. With BentoML serving framework and bentoctl users can enjoy the performance and scalability of Sagemaker with any popular ML frameworks.

Note: This operator is compatible with BentoML version 1.0.0 and above. For older versions, please switch to the branch pre-v1.0 and follow the instructions in the README.md.

Table of Contents

• Quickstart with bentoctl
• Configuration options
• Troubleshooting

Quickstart with bentoctl

This quickstart will walk you through deploying a bento as an AWS Sagemaker Endpoint. Make sure to go through the prerequisites section and follow the instructions to set everything up.

Prerequisites

1. BentoML version 1.0 or above. Please follow the Installation guide.
2. Terraform - Terraform is a tool for building, configuring, and managing infrastructure. Installation instruction: www.terraform.io/downloads
3. AWS CLI - installed and configured with an AWS account with permission to Sagemaker, Lambda and ECR. Please follow the Installation guide.
4. Docker - Install instruction: docs.docker.com/install
5. A built Bento project. For this guide, we will use the Iris classifier bento from the BentoML quickstart guide. You can also use your own Bentos that are available locally.

Steps

1. Install bentoctl via pip

   $ pip install bentoctl
   

2. Install AWS Sagemaker operator

   Bentoctl will install the official AWS Sagemaker operator and its dependencies.

   $ bentoctl operator install aws-sagemaker
   

3. Initialize deployment with bentoctl

   Follow the interactive guide to initialize the deployment project.

   $ bentoctl init
   
   Bentoctl Interactive Deployment Config Builder
   
   Welcome! You are now in interactive mode.
   
   This mode will help you setup the deployment_config.yaml file required for
   deployment. Fill out the appropriate values for the fields.
   
   (deployment config will be saved to: ./deployment_config.yaml)
   
   api_version: v1
   name: quickstart
   operator: aws-sagemaker
   template: terraform
   spec:
       region: ap-south-1
       instance_type: ml.t2.medium
       initial_instance_count: 1
       timeout: 60
       enable_data_capture: False
       destination_s3_uri:
       initial_sampling_percentage: 1
   filename for deployment_config [deployment_config.yaml]:
   deployment config generated to: deployment_config.yaml
   ✨ generated template files.
     - ./main.tf
     - ./bentoctl.tfvars

   This will also run the bentoctl generate command for you and will generate the main.tf terraform file, which specifies the resources to be created and the bentoctl.tfvars file which contains the values for the variables used in the main.tf file.

4. Build and push AWS sagemaker compatible docker image to the registry

   Bentoctl will build and push the sagemaker compatible docker image to the AWS ECR repository.

   bentoctl build -b iris_classifier:latest -f deployment_config.yaml
   
   Step 1/22 : FROM bentoml/bento-server:1.0.0a6-python3.8-debian-runtime
    ---> 046bc2e28220
   Step 2/22 : ARG UID=1034
    ---> Using cache
    ---> f44cfa910c52
   Step 3/22 : ARG GID=1034
    ---> Using cache
    ---> e4d5aed007af
   Step 4/22 : RUN groupadd -g $GID -o bentoml && useradd -m -u $UID -g $GID -o -r bentoml
    ---> Using cache
    ---> fa8ddcfa15cf
   ...
   Step 22/22 : CMD ["bentoml", "serve", ".", "--production"]
    ---> Running in 28eccee2f650
    ---> 98bc66e49cd9
   Successfully built 98bc66e49cd9
   Successfully tagged quickstart:kiouq7wmi2gmockr
   🔨 Image build!
   Created the repository quickstart
   The push refers to repository
   [213386773652.dkr.ecr.ap-south-1.amazonaws.com/quickstart]
   kiouq7wmi2gmockr: digest:
   sha256:e1a468e6b9ceeed65b52d0ee2eac9e3cd1a57074eb94db9c263be60e4db98881 size: 3250
   63984d77b4da: Pushed
   2bc5eef20c91: Pushed
   ...
   da0af9cdde98: Layer already exists
   e5baccb54724: Layer already exists
   🚀 Image pushed!
   ✨ generated template files.
     - ./bentoctl.tfvars
     - ./startup_script.sh

   The iris-classifier service is now built and pushed into the container registry and the required terraform files have been created. Now we can use terraform to perform the deployment.

5. Apply Deployment with Terraform

   1. Initialize terraform project. This installs the AWS provider and sets up the terraform folders.

      $ terraform init

   2. Apply terraform project to create Sagemaker deployment

      $ terraform apply -var-file=bentoctl.tfvars -auto-approve
      
      aws_iam_role.iam_role_lambda: Creating...
      aws_iam_role.iam_role_sagemaker: Creating...
      aws_apigatewayv2_api.lambda: Creating...
      aws_apigatewayv2_api.lambda: Creation complete after 1s [id=rwfej5qsf6]
      aws_cloudwatch_log_group.api_gw: Creating...
      aws_cloudwatch_log_group.api_gw: Creation complete after 1s [id=/aws/api_gw/quickstart-gw]
      aws_apigatewayv2_stage.lambda: Creating...
      aws_apigatewayv2_stage.lambda: Creation complete after 3s [id=$default]
      aws_iam_role.iam_role_sagemaker: Creation complete after 7s [id=quickstart-sagemaker-iam-role]
      aws_sagemaker_model.sagemaker_model: Creating...
      aws_iam_role.iam_role_lambda: Creation complete after 8s [id=quickstart-lambda-iam-role]
      aws_lambda_function.fn: Creating...
      ...
      
      
      Apply complete! Resources: 1 added, 0 changed, 0 destroyed.
      
      Outputs:
      
      endpoint = "https://rwfej5qsf6.execute-api.ap-south-1.amazonaws.com/"
      ecr_image_tag = "213386773652.dkr.ecr.ap-south-1.amazonaws.com/quickstart:sfx3dagmpogmockr"

6. Test deployed endpoint

   The iris_classifier uses the /classify endpoint for receiving requests so the full URL for the classifier will be in the form {EndpointUrl}/classify.

   URL=$(terraform output -json | jq -r .endpoint.value)classify
   curl -i \
     --header "Content-Type: application/json" \
     --request POST \
     --data '[5.1, 3.5, 1.4, 0.2]' \
     $URL
   
   HTTP/2 200
   date: Thu, 14 Apr 2022 23:02:45 GMT
   content-type: application/json
   content-length: 1
   apigw-requestid: Ql8zbicdSK4EM5g=
   
   0%

Note: You can also invoke the Sagemaker endpoint directly. If there is only one service, SageMaker deployment will choose that one. If there is more than one, you can specify which service to use by passing the X-Amzn-SageMaker-Custom-Attributes header with the name of the service as value.

7. Delete deployment Use the bentoctl destroy command to remove the registry and the deployment

   bentoctl destroy -f deployment_config.yaml
   

Configuration Options

A sample configuration file has been given has been provided here. Feel free to copy it over and change it for you specific deployment values

• region: AWS region where Sagemaker endpoint is deploying to
• instance_type: The ML compute instance type for Sagemaker endpoint. See https://docs.aws.amazon.com/cli/latest/reference/sagemaker/create-endpoint-config.html for available instance types
• initial_instance_count: Number of instances to launch initially.
• timeout: timeout for API request in seconds
• enable_data_capture: Enable Sagemaker capture data from requests and responses and store the captured data to AWS S3
• destination_s3_uri: S3 bucket path for store captured data
• initial_sampling_percentage: Percentage of the data will be captured to S3 bucket.

Troubleshooting

By default sagemaker is configured with cloudwatch for metrics and logs. To see the cloudwatch logs for the deployment

1. Open the Amazon Cloudwatch console at https://console.aws.amazon.com/cloudwatch/.
2. In the navigation pane, choose Logs -> Log groups.
3. Head over to /aws/sagemaker/Endpoints/<deployment_name>-endpoint
4. Choose the latest logs streams