無料でビルドを開始
CircleCI.comアカデミーブログコミュニティサポート

Deploy to Amazon SageMaker

3 months ago6 min read
Cloud
Server v4.x
Server v3.x
このページの内容

This how-to guide walks you through the steps to set up and use the Amazon SageMaker orb to deploy and release models to SageMaker. It uses the sample repository example project repository as a reference, but you can adapt every step to your requirements. Through this guide, you will learn how to orchestrate model deployment to endpoints across different environments, as well as track and manage your releases in the CircleCI release UI.

For further information about the orb and the releases feature see the following:

Prerequisites

To follow along with this how-to guide you will need the following:

  • A CircleCI account connected to your code. You can sign up for free

  • Basic knowledge of Amazon SageMaker

  • An AWS account with access to SageMaker

  • A SageMaker studio domain to follow along the training step, or an existing Model Package Registry if you have already trained one

  • A copy of the example repository building on CircleCI. For steps to get the project set up, see the Create a project in CircleCI page.

1. Set up a CircleCI release environment

First, set up a release environment to monitor your releases from the CircleCI web app, and access options to manage promotions and rollbacks.

  1. In the CircleCI web app, select Releases from the sidebar

  2. If this is your first release environment, you will see an option to Add Release Environment. If not, Click Configure Environments in the top right corner, and then click Configure New Environment

    Screenshot showing option to add a new release environment
  3. In the popup, select Amazon SageMaker from the "Environment" dropdown. Give your new environment a name, and click Create Environment

    Screenshot showing new release environment popup
  4. Click on your new environment name to enter the environment view

  5. Click Create New Token

    Screenshot showing create new token button
  6. Click Create Token. Be sure to save this token, you will create an environment variable later to store it.

2. Set up Amazon SageMaker access

a. OIDC Identity Provider

In this section you will create an IAM (Identity and Access Management) Identity Provider in your AWS IAM for the CircleCI OIDC Provider. The Amazon SageMaker orb uses OIDC. You may already have this set up, in which case you can skip this section. For more information on CircleCI OIDC functionality, see the Using OpenID Connect tokens in jobs page.

  1. Navigate to your AWS Management Console

  2. Go to IAM  Access management  Identity providers

  3. Select Add Provider

    Screenshot showing AWS management console Add Provider option
  4. Under "Provider Type", select OpenID Connect

  5. Enter your "Provider URL". This will be https://oidc.circleci.com/org/<your-organization-id>, where <your-organization-id> is the ID of your CircleCI organization.

  6. Click Get Thumbprint

  7. For "Audience" enter your organization ID

  8. Click Add Provider

b. Policy

  1. From your AWS management console, navigate to IAM  Policies

  2. Click Create Policy

  3. Click JSON to select the JSON policy editor and then copy in the following permissions. We have organized the permissions into two groups. OrbPermissions and S3Access statements are used for the deployment of the model to the endpoints. The S3AccessTrainModel and SageMakerTrainModel statements are needed if you want to train the demo model we provide.

    {
        "Version": "2012-10-17",
        "Statement": [
            {
                "Sid": "OrbPermissions",
                "Effect": "Allow",
                "Action": [
                    "sagemaker:AddTags",
                    "sagemaker:CreateEndpointConfig",
                    "sagemaker:CreateModel",
                    "sagemaker:DescribeEndpoint",
    	 	        "sagemaker:DescribeEndpointConfig",
                    "sagemaker:ListEndpoints",
                    "sagemaker:ListModelPackages",
                    "sagemaker:ListTags",
                    "sagemaker:UpdateEndpoint",
                    "iam:PassRole"
                ],
                "Resource": "*"
            },
            {
                "Sid": "S3Access",
                "Effect": "Allow",
                "Action": [
                    "s3:GetObject",
                    "s3:ListBucket"
                ],
                "Resource": [
                    "arn:aws:s3:::<BUCKET_FOR_MODEL_ASSETS>/*"
                ]
            },
            {
                "Sid": "S3AccessTrainModel",
                "Effect": "Allow",
                "Action": [
                    "s3:GetObject",
                    "s3:ListBucket",
                    "s3:PutObject"
                ],
                "Resource": [
                    "arn:aws:s3:::sagemaker-sample-files/*",
                    "arn:aws:s3:::circleci-sagemaker-pipeline",
                    "arn:aws:s3:::circleci-sagemaker-pipeline/*"
                ]
            },
            {
                "Sid": "SageMakerTrainModel",
                "Effect": "Allow",
                "Action": [
                    "sagemaker:CreateTrainingJob",
                    "sagemaker:DescribeTrainingJob",
                    "logs:DescribeLogStreams",
                    "sagemaker:ListModelPackageGroups",
                    "sagemaker:CreateModelPackage",
                    "sagemaker:UpdateModelPackage"
                ],
                "Resource": "*"
            }
        ]
    }
  4. Scroll down and click Next

  5. Give your policy a name and then click Create Policy

c. Role

  1. From your AWS management console, navigate to IAM  Roles

  2. Click Create Role

  3. Select Web Identity and then select the CircleCI provider you created above, and under Audience, select your org ID

  4. Use the search function to find the policy you created above by name, select it and click Next

  5. Give your Role a name, and then scroll to the Trust policy section. Set up the Trust relationship between the Role and the CircleCI OIDC Provider. Here is an example Policy. Remember to replace the placeholders <CIRCLECI-ORG-ID> and <CIRCLECI-PROJECT-ID> with your values.

    {
    	"Version": "2012-10-17",
    	"Statement": [
            {
                "Effect": "Allow",
                "Principal": {
                    "Federated": "arn:aws:iam::<AWS-ACCOUNT-ID>:oidc-provider/oidc.circleci.com/org/<CIRCLECI-ORG-ID>"
                },
                "Action": "sts:AssumeRoleWithWebIdentity",
                "Condition": {
                    "StringLike": {
                        "oidc.circleci.com/org/<CIRCLECI-ORG-ID>:sub": "org/<CIRCLECI-ORG-ID>/project/<CIRCLECI-PROJECT-ID>/user/*"
                    }
                }
            },
            {
    			"Effect": "Allow",
    			"Principal": {
    				"Service": "sagemaker.amazonaws.com"
    			},
    			"Action": "sts:AssumeRole"
    		}
    
        ]
    }
  6. Click Create Role

  7. Select your role from the list and copy the Role ARN, you will need this in the next section.

    Screenshot showing location of Role ARN

3. Set environment variables

The CircleCI Amazon SageMaker orb requires some environment variables to function. You can store these environment variables at the project level, or you can store them using a context. The following steps show how to add the environment variables at the project level. You need to add two environment variables, as follows:

  • CCI_RELEASE_INTEGRATION_TOKEN: The orb connects your deployment to SageMaker with CircleCI releases. This gives you visibility into the Endpoint Configuration Updates, and what is currently active.

  • SAGEMAKER_EXECUTION_ROLE_ARN: This is the AWS IAM Role you configured with the necessary SageMaker permissions, and the OIDC Trust relationship.

    1. From the CircleCI web app, click Dashboard in the sidebar

    2. Use the project filter dropp-down to choose your project

    3. Click Project Settings and then choose Environment Variables

    4. Click Add Environment Variable button to enter a name and value of the new environment variable.

      Screenshot showing add environment variable popup
    5. Repeat step four for the second environment variable

4. Set up your CircleCI configuration

Take a look at the .circleci/config.yml for the sample project. There is a block of pipeline parameters at the top. You need to update the defaults with your values.

parameters:
  bucket:
    type: string
    default: circleci-sagemaker-pipeline
  model_desc:
    type: string
    default: "Kitten Classifier allowing us to distinguish between giraffes and kittens."
  model_name:
    type: string
    default: kitten-classifier
  project_id:
    type: string
    default: "e47ee9b0-446f-44cf-bec8-5407ceb06930"
  region_name:
    type: string
    default: us-east-1

Make the following updates:

ParameterDefault

bucket

The bucket you set up in the Role Policy

model_desc

A description of your model

model_name

The name of the model

project_id

The CircleCI project ID, you can find this on the Project Settings page in the CircleCI web app

region_name

The region, for example, us-east-1

5. Build a model package version

The sample app used in this how-to guide uses a model commonly found in AWS documentation, Abalone. It has just been renamed. If you already have your own model, feel free to adapt the configuration file to use that one instead. All you will need to do is update the model_name parameter.

Assuming you are using our example repository, follow these steps to build a new model package version:

  1. There is a workflow (model-train) configured to run on the branch model-train. Checkout the model-train branch, and push it up to GitHub

  2. The model-train workflow will make a new model package version in the model registry. If the model package doesn’t already exist, it will create it.

Every time you run this workflow by pushing to the model-train branch, anew model version is created.

6. Use the orb

Let’s break down the deploy-model-through-to-prod workflow. Our first job is aws-sagemaker/create_model. This job creates a model from your latest Model Package in the registry. This will be what we then deploy to the inference endpoints. In general, create-model just needs to be called one time at the beginning of your workflow:

- aws-sagemaker/create_model:
    # job name that will show in Workflow DAG
    name: create-model
    # s3 bucket where asset will be stored
    bucket: << pipeline.parameters.bucket >>
    # Name of the model in SageMaker that we will be deploying.
    model_name: << pipeline.parameters.model_name >>
    # We use the pipeline.id as the unique identifier for some of the configs we create
    circle_pipeline_id: << pipeline.id >>
    # Region where we are deploying to
    region_name: << pipeline.parameters.region_name >>
    filters: *main-branch-only
  • Aside from name and filters, all other parameters are passed in from our pipeline parameters.

  • name controls the name of this job visible in the Workflow graph in the CircleCI web app.

  • filters allows you to control what branch the job runs on.

Next, we need to create the endpoint configuraiton, this happens in the job aws-sagemaker/create_endpoint_configuration:

- aws-sagemaker/create_endpoint_configuration:
    name: dev:create-model-endpoint-config
    bucket: << pipeline.parameters.bucket >>
    deploy_environment: dev
    model_name: << pipeline.parameters.model_name >>
    circle_pipeline_id: << pipeline.id >>
    circle_project_id: << pipeline.parameters.project_id >>
    region_name: << pipeline.parameters.region_name >>
    requires:
        - create-model
    filters: *main-branch-only
  • deploy_environment is an arbitrary string you can use to bucket your model releases, for example, dev, staging, testing, prod. In our example config, you can see we are only using two, dev and prod.

Next, we need to push out the updated configuration, this happens in the job aws-sagemaker/deploy_endpoint:

    - aws-sagemaker/deploy_endpoint:
        name: dev:deploy-model-to-endpoint
        bucket: << pipeline.parameters.bucket >>
        deploy_environment: dev
        model_name: << pipeline.parameters.model_name >>
        # Description for the model. q: can we make it optional?
        model_desc: << pipeline.parameters.model_desc >>
        pipeline_id: << pipeline.id >>
        # You can find this value in the Project Settings in CircleCI
        project_id: << pipeline.parameters.project_id >>
        region_name: << pipeline.parameters.region_name >>
        requires:
            - dev:create-model-endpoint-config
        filters: *main-branch-only
  • The only thing to set here is the deploy_environment so we know what endpoint_configuration to use. This will deploy your endpoint configuration.

Next, is an approval job. This stops the workflow from deploying to production until a human approves it. When this does get approval, we then repeat the above steps of aws-sagemaker/create_endpoint_configuration and aws-sagemaker/deploy_endpoint but with deploy_environment set to prod.:

- promote-model-to-prod-endpoint:
        type: approval
    requires:
        - dev:deploy-model-to-endpoint
    filters: *main-branch-only
Approval job

7. Check in on your release

The CircleCI release UI offers you a single pane of glass to monitor all your deployments across environments. You can view deployment progress in real time, see what versions are currently deployed, and navigate easily to the SageMaker console. If you navigate to the releases dashboard by clicking Releases in the web app sidebar, you should now see something along the lines of the image below:

Releases dashboard

Click on the component name to view details about the component, including the number of instances deployed. From this interface, you can navigate to Amazon SageMaker for live information on your endpoint.

Releases component view

Select the version to see the version details page for version being released. Clicking on specific versions lets you monitor release progress in real-time.

Releases version view

Conclusion


Suggest an edit to this page

Make a contribution
Learn how to contribute