Connect to SageMaker Within your VPC

You can connect directly to the SageMaker API or to Amazon SageMaker Runtime through an interface endpoint in your virtual private cloud (VPC) instead of connecting over the internet. When you use a VPC interface endpoint, communication between your VPC and the SageMaker API or Runtime is conducted entirely and securely within an AWS network.

Connect to SageMaker through a VPC interface endpoint

The SageMaker API and SageMaker Runtime support Amazon Virtual Private Cloud (Amazon VPC) interface endpoints that are powered by AWS PrivateLink. Each VPC endpoint is represented by one or more Elastic Network Interfaces with private IP addresses in your VPC subnets. For example, an application inside your VPC uses AWS PrivateLink to communicate with SageMaker Runtime. SageMaker Runtime in turn communicates with the SageMaker endpoint. Using AWS PrivateLink allows you to invoke your SageMaker endpoint from within your VPC, as shown in the following diagram.

The VPC interface endpoint connects your VPC directly to the SageMaker API or SageMaker Runtime using AWS PrivateLink without using an internet gateway, NAT device, VPN connection, or AWS Direct Connect connection. The instances in your VPC do not need to connect to the public internet in order to communicate with the SageMaker API or SageMaker Runtime.

You can create an AWS PrivateLink interface endpoint to connect to SageMaker or to SageMaker Runtime using either the AWS Management Console or AWS Command Line Interface (AWS CLI). For instructions, see Access an AWS service using an interface VPC endpoint.

If you haven't enabled a private Domain Name System (DNS) hostname for your VPC endpoint, after you have created a VPC endpoint, specify the internet endpoint URL to the SageMaker API or SageMaker Runtime. Example code using AWS CLI commands to specify the endpoint-url parameter follows.

aws sagemaker list-notebook-instances --endpoint-url VPC_Endpoint_ID.api.sagemaker.Region.vpce.amazonaws.com

aws sagemaker list-training-jobs --endpoint-url VPC_Endpoint_ID.api.sagemaker.Region.vpce.amazonaws.com

aws sagemaker-runtime invoke-endpoint --endpoint-url https://VPC_Endpoint_ID.runtime.sagemaker.Region.vpce.amazonaws.com  \
    --endpoint-name Endpoint_Name \
    --body "Endpoint_Body" \
    --content-type "Content_Type" \
            Output_File

If you enable private DNS hostnames for your VPC endpoint, you don't need to specify the endpoint URL because the default hostname (https://api.sagemaker.Region.amazon.com) resolves to your VPC endpoint. Similarly, the default SageMaker Runtime DNS hostname (https://runtime.sagemaker.Region.amazonaws.com) also resolves to your VPC endpoint.

The SageMaker API and SageMaker Runtime support VPC endpoints in all AWS Regions where both Amazon VPC and SageMaker ares available. SageMaker supports making calls to all of its Operations inside your VPC. If you use the AuthorizedUrl from the 
 CreatePresignedNotebookInstanceUrl command, your traffic will go over the public internet. You can't only use a VPC endpoint to access the presigned URL, the request must go through the internet gateway.

By default, your users can share the presigned URL to people outside of your corporate network. For additional security, you must add IAM permissions to restrict the URL only be usable within your network. For information about IAM permissions, see How AWS PrivateLink works with IAM.

Note

When setting up a VPC interface endpoint for the SageMaker Runtime service (https://runtime.sagemaker.Region.amazonaws.com), you must ensure that the VPC interface endpoint is activated in the Availability Zone of your client in order for private DNS resolution to work. Otherwise, you may see DNS failures when attempting to resolve the URL.

To learn more about AWS PrivateLink, see the AWS PrivateLink documentation. Refer to AWS PrivateLink Pricing for the price of VPC endpoints. To learn more about VPC and endpoints, see Amazon VPC. For information about how to use identity-based AWS Identity and Access Management policies to restrict access to the SageMaker API and SageMaker Runtime, see Control access to the SageMaker API by using identity-based policies.

Using SageMaker training and hosting with resources inside your VPC

SageMaker uses your execution role to download and upload information from an Amazon S3 bucket and Amazon Elastic Container Registry (Amazon ECR), in isolation from your training or inference container. If you have resources that are located inside your VPC, you can still grant SageMaker access to those resources. The following sections explain how to make your resources available to SageMaker with or without network isolation.

Without network isolation enabled

If you haven't set network isolation on your training job or model, SageMaker can access resources using either of the following methods.

• SageMaker training and deployed inference containers can access the internet by default. SageMaker containers are able to access external services and resources on the public internet as part of your training and inference workloads. SageMaker containers are not able to access resources inside your VPC without a VPC configuration, as shown in the following illustration.

  

• Use a VPC configuration to communicate with resources inside your VPC through an elastic network interface (ENI). The communication between the container and the resources in your VPC takes place securely within your VPC network, as shown in the following illustration. In this case, you manage networking access to your VPC resources and internet.

  

With network isolation

If you employ network isolation, the SageMaker container can't communicate with resources inside your VPC or make any network calls, as shown in the following illustration. If you provide a VPC configuration, the download and upload operations will be run through your VPC. For more information about hosting and training with network isolation while using a VPC, see Network Isolation.

Create a VPC Endpoint Policy for SageMaker

You can create a policy for Amazon VPC endpoints for SageMaker to specify the following:

• The principal that can perform actions.

• The actions that can be performed.

• The resources on which actions can be performed.

For more information, see Controlling Access to Services with VPC Endpoints in the Amazon VPC User Guide.

Note

VPC endpoint policies aren't supported for Federal Information Processing Standard (FIPS) SageMaker runtime endpoints for runtime_InvokeEndpoint.

The following example VPC endpoint policy specifies that all users who have access to the VPC interface endpoint are allowed to invoke the SageMaker hosted endpoint named myEndpoint.

{
  "Statement": [
      {
          "Action": "sagemaker:InvokeEndpoint",
          "Effect": "Allow",
          "Resource": "arn:aws:sagemaker:us-west-2:123456789012:endpoint/myEndpoint",
          "Principal": "*"
      }
  ]
}

In this example, the following are denied:

• Other SageMaker API actions, such as sagemaker:CreateEndpoint and sagemaker:CreateTrainingJob.

• Invoking SageMaker hosted endpoints other than myEndpoint.

Note

In this example, users can still take other SageMaker API actions from outside the VPC. For information about how to restrict API calls to those from within the VPC, see Control access to the SageMaker API by using identity-based policies.

Create a VPC Endpoint Policy for Amazon SageMaker Feature Store

To create a VPC Endpoint for Amazon SageMaker Feature Store, use the following endpoint template, substituting your VPC_Endpoint_ID.api and Region:

VPC_Endpoint_ID.api.featurestore-runtime.sagemaker.Region.vpce.amazonaws.com

Connect Your Private Network to Your VPC

To call the SageMaker API and SageMaker Runtime through your VPC, you have to connect from an instance that is inside the VPC or connect your private network to your VPC by using an AWS Virtual Private Network (AWS VPN) or AWS Direct Connect. For information about AWS VPN, see VPN Connections in the Amazon Virtual Private Cloud User Guide. For information about AWS Direct Connect, see Creating a Connection in the AWS Direct Connect User Guide.