Secure an Azure Machine Learning training environment with virtual networks

APPLIES TO: Python SDK azure-ai-ml v2 (current)

Azure Machine Learning compute instance, serverless compute, and compute cluster can be used to securely train models in an Azure Virtual Network. When planning your environment, you can configure the compute instance/cluster or serverless compute with or without a public IP address. The general differences between the two are:

• No public IP: Reduces costs as it doesn't have the same networking resource requirements. Improves security by removing the requirement for inbound traffic from the internet. However, there are additional configuration changes required to enable outbound access to required resources (Microsoft Entra ID, Azure Resource Manager, etc.).
• Public IP: Works by default, but costs more due to additional Azure networking resources. Requires inbound communication from the Azure Machine Learning service over the public internet.

The following table contains the differences between these configurations:

You can also use Azure Databricks or HDInsight to train models in a virtual network.

This article is part of a series on securing an Azure Machine Learning workflow. See the other articles in this series:

• Virtual network overview
• Secure the workspace resources
• Secure the inference environment
• Enable studio functionality
• Use custom DNS
• Use a firewall

For a tutorial on creating a secure workspace, see Tutorial: Create a secure workspace, Bicep template, or Terraform template.

In this article you learn how to secure the following training compute resources in a virtual network:

Prerequisites

• Read the Network security overview article to understand common virtual network scenarios and overall virtual network architecture.

• An existing virtual network and subnet to use with your compute resources. This VNet must be in the same subscription as your Azure Machine Learning workspace.

  • We recommend putting the storage accounts used by your workspace and training jobs in the same Azure region that you plan to use for compute instances, serverless compute, and clusters. If they aren't in the same Azure region, you may incur data transfer costs and increased network latency.
  • Make sure that WebSocket communication is allowed to *.instances.ml.azure.cn and *.instances.ml.azure.cn in your VNet. WebSockets are used by Jupyter on compute instances.

• An existing subnet in the virtual network. This subnet is used when creating compute instances, clusters, and nodes for serverless compute.

  • Make sure that the subnet isn't delegated to other Azure services.
  • Make sure that the subnet contains enough free IP addresses. Each compute instance requires one IP address. Each node within a compute cluster and each serverless compute node requires one IP address.

• If you have your own DNS server, we recommend using DNS forwarding to resolve the fully qualified domain names (FQDN) of compute instances and clusters. For more information, see Use a custom DNS with Azure Machine Learning.

• To deploy resources into a virtual network or subnet, your user account must have permissions to the following actions in Azure role-based access control (Azure RBAC):

  • "Microsoft.Network/*/read" on the virtual network resource. This permission isn't needed for Azure Resource Manager (ARM) template deployments.
  • "Microsoft.Network/virtualNetworks/join/action" on the virtual network resource.
  • "Microsoft.Network/virtualNetworks/subnets/join/action" on the subnet resource.

  For more information on Azure RBAC with networking, see the Networking built-in roles

Limitations

• Compute cluster/instance and serverless compute deployment in virtual network isn't supported with Azure Lighthouse.

• Port 445 must be open for private network communications between your compute instances and the default storage account during training. For example, if your computes are in one VNet and the storage account is in another, don't block port 445 to the storage account VNet.

Compute cluster in a different VNet/region from workspace

To create a compute cluster in an Azure Virtual Network in a different region than your workspace virtual network, you have couple of options to enable communication between the two VNets.

• Use VNet Peering.
• Add a private endpoint for your workspace in the virtual network that will contain the compute cluster.

Scenario: VNet peering

1. Configure your workspace to use an Azure Virtual Network. For more information, see Secure your workspace resources.

2. Create a second Azure Virtual Network that will be used for your compute clusters. It can be in a different Azure region than the one used for your workspace.

3. Configure VNet Peering between the two VNets.

   Tip

   Wait until the VNet Peering status is Connected before continuing.

4. Modify the privatelink.api.azureml.ms DNS zone to add a link to the VNet for the compute cluster. This zone is created by your Azure Machine Learning workspace when it uses a private endpoint to participate in a VNet.

   1. Add a new virtual network link to the DNS zone. You can do this multiple ways:

      • From the Azure portal, navigate to the DNS zone and select Virtual network links. Then select + Add and select the VNet that you created for your compute clusters.
      • From the Azure CLI, use the az network private-dns link vnet create command. For more information, see az network private-dns link vnet create.
      • From Azure PowerShell, use the New-AzPrivateDnsVirtualNetworkLink command. For more information, see New-AzPrivateDnsVirtualNetworkLink.

5. Repeat the previous step and sub-steps for the privatelink.notebooks.chinacloudapi.cn DNS zone.

6. Configure the following Azure resources to allow access from both VNets.

   • The default storage account for the workspace.
   • The Azure Container registry for the workspace.
   • The Azure Key Vault for the workspace.

   Tip

   There are multiple ways that you might configure these services to allow access to the VNets. For example, you might create a private endpoint for each resource in both VNets. Or you might configure the resources to allow access from both VNets.

7. Create a compute cluster as you normally would when using a VNet, but select the VNet that you created for the compute cluster. If the VNet is in a different region, select that region when creating the compute cluster.

   Warning

   When setting the region, if it is a different region than your workspace or datastores you may see increased network latency and data transfer costs. The latency and costs can occur when creating the cluster, and when running jobs on it.

Scenario: Private endpoint

1. Configure your workspace to use an Azure Virtual Network. For more information, see Secure your workspace resources.

2. Create a second Azure Virtual Network that will be used for your compute clusters. It can be in a different Azure region than the one used for your workspace.

3. Create a new private endpoint for your workspace in the VNet that will contain the compute cluster.

   • To add a new private endpoint using the Azure portal, select your workspace and then select Networking. Select Private endpoint connections, + Private endpoint and use the fields to create a new private endpoint.

     • When selecting the Region, select the same region as your virtual network.
     • When selecting Resource type, use Microsoft.MachineLearningServices/workspaces.
     • Set the Resource to your workspace name.
     • Set the Virtual network and Subnet to the VNet and subnet that you created for your compute clusters.

     Finally, select Create to create the private endpoint.

   • To add a new private endpoint using the Azure CLI, use the az network private-endpoint create. For an example of using this command, see Configure a private endpoint for Azure Machine Learning workspace.

4. Create a compute cluster as you normally would when using a VNet, but select the VNet that you created for the compute cluster. If the VNet is in a different region, select that region when creating the compute cluster.

   Warning

   When setting the region, if it is a different region than your workspace or datastores you may see increased network latency and data transfer costs. The latency and costs can occur when creating the cluster, and when running jobs on it.

Compute instance/cluster or serverless compute with no public IP

The following configurations are in addition to those listed in the Prerequisites section, and are specific to creating a compute instances/clusters configured for no public IP. They also apply to serverless compute:

• You must use a workspace private endpoint for the compute resource to communicate with Azure Machine Learning services from the VNet. For more information, see Configure a private endpoint for Azure Machine Learning workspace.

• In your VNet, allow outbound traffic to the following service tags or fully qualified domain names (FQDN):

  Service tag	Protocol	Port	Notes
  AzureMachineLearning	TCP UDP	443/8787/18881 5831	Communication with the Azure Machine Learning service.
  BatchNodeManagement.<region>	ANY	443	Replace <region> with the Azure region that contains your Azure Machine Learning workspace. Communication with Azure Batch. Compute instance and compute cluster are implemented using the Azure Batch service.
  Storage.<region>	TCP	443	Replace <region> with the Azure region that contains your Azure Machine Learning workspace. This service tag is used to communicate with the Azure Storage account used by Azure Batch.

  Important

  The outbound access to Storage.<region> could potentially be used to exfiltrate data from your workspace. By using a Service Endpoint Policy, you can mitigate this vulnerability. For more information, see the Azure Machine Learning data exfiltration prevention article.

  FQDN	Protocol	Port	Notes
  <region>.tundra.ml.azure.cn	UDP	5831	Replace <region> with the Azure region that contains your Azure Machine Learning workspace.
  microsoftgraph.chinacloudapi.cn	TCP	443	Communication with the Microsoft Graph API.
  *.instances.ml.azure.cn	TCP	443/8787/18881	Communication with Azure Machine Learning.
  *.<region>.batch.chinacloudapi.cn	ANY	443	Replace <region> with the Azure region that contains your Azure Machine Learning workspace. Communication with Azure Batch.
  *.<region>.service.batch.chinacloudapi.cn	ANY	443	Replace <region> with the Azure region that contains your Azure Machine Learning workspace. Communication with Azure Batch.
  *.blob.core.chinacloudapi.cn	TCP	443	Communication with Azure Blob storage.
  *.queue.core.chinacloudapi.cn	TCP	443	Communication with Azure Queue storage.
  *.table.core.chinacloudapi.cn	TCP	443	Communication with Azure Table storage.

• By default, a compute instance/cluster configured for no public IP doesn't have outbound access to the internet. If you can access the internet from it, it is because of Azure default outbound access and you have an NSG that allows outbound to the internet. However, we don't recommend using the default outbound access. If you need outbound access to the internet, we recommend using either a firewall and outbound rules or a NAT gateway and network service groups to allow outbound traffic instead.

  For more information on the outbound traffic that is used by Azure Machine Learning, see the following articles:

  • Configure inbound and outbound network traffic.
  • Azure's outbound connectivity methods.

  For more information on service tags that can be used with Azure Firewall, see the Virtual network service tags article.

Use the following information to create a compute instance or cluster with no public IP address:

# create a compute cluster with no public IP
az ml compute create --name cpu-cluster --resource-group rg --workspace-name ws --vnet-name yourvnet --subnet yoursubnet --type AmlCompute --set enable_node_public_ip=False

# create a compute instance with no public IP
az ml compute create --name myci --resource-group rg --workspace-name ws --vnet-name yourvnet --subnet yoursubnet --type ComputeInstance --set enable_node_public_ip=False

from azure.ai.ml.entities import AmlCompute, NetworkSettings

network_settings = NetworkSettings(vnet_name="<vnet-name>", subnet="<subnet-name>")
compute = AmlCompute(
    name=cpu_compute_target,
    size="STANDARD_D2_V2",
    min_instances=0,
    max_instances=4,
    enable_node_public_ip=False,
    network_settings=network_settings
)
ml_client.begin_create_or_update(entity=compute)

Use the following information to configure serverless compute nodes with no public IP address in the VNet for a given workspace:

az ml workspace create -n <workspace-name> -g <resource-group-name> --file serverlesscomputevnetsettings.yml

name: testserverlesswithnpip
location: eastus
public_network_access: Disabled
serverless_compute:
  custom_subnet: /subscriptions/<sub id>/resourceGroups/<resource group>/providers/Microsoft.Network/virtualNetworks/<vnet name>/subnets/<subnet name>
  no_public_ip: true

az ml workspace update -n <workspace-name> -g <resource-group-name> --file serverlesscomputevnetsettings.yml

serverless_compute:
  custom_subnet: /subscriptions/<sub id>/resourceGroups/<resource group>/providers/Microsoft.Network/virtualNetworks/<vnet name>/subnets/<subnet name>
  no_public_ip: true

from azure.ai.ml import MLClient
from azure.ai.ml.entities import ServerlessComputeSettings, Workspace
from azure.identity import DefaultAzureCredential

subscription_id = <sub id>
resource_group = <resource group>
workspace_name = <workspace name>
ml_client = MLClient(
    DefaultAzureCredential(), subscription_id, resource_group
)

workspace = Workspace(
    name=workspace_name,
    serverless_compute=ServerlessComputeSettings(
        custom_subnet=<subnet id>,
        no_public_ip=true,
    )
)

workspace = ml_client.workspaces.begin_update(workspace)

Compute instance/cluster or serverless compute with public IP

The following configurations are in addition to those listed in the Prerequisites section, and are specific to creating compute instances/clusters that have a public IP. They also apply to serverless compute:

• If you put multiple compute instances/clusters in one virtual network, you may need to request a quota increase for one or more of your resources. The Machine Learning compute instance or cluster automatically allocates networking resources in the resource group that contains the virtual network. For each compute instance or cluster, the service allocates the following resources:

  • A network security group (NSG) is automatically created. This NSG allows inbound TCP traffic on port 44224 from the AzureMachineLearning service tag.

    Important

    Compute instance and compute cluster automatically create an NSG with the required rules.

    If you have another NSG at the subnet level, the rules in the subnet level NSG mustn't conflict with the rules in the automatically created NSG.

    To learn how the NSGs filter your network traffic, see How network security groups filter network traffic.

  • One load balancer

  For compute clusters, these resources are deleted every time the cluster scales down to 0 nodes and created when scaling up.

  For a compute instance, these resources are kept until the instance is deleted. Stopping the instance doesn't remove the resources.

  Important

  These resources are limited by the subscription's resource quotas. If the virtual network resource group is locked then deletion of compute cluster/instance will fail. Load balancer cannot be deleted until the compute cluster/instance is deleted. Also please ensure there is no Azure Policy assignment which prohibits creation of network security groups.

• In your VNet, allow inbound TCP traffic on port 44224 from the AzureMachineLearning service tag.

  Important

  The compute instance/cluster is dynamically assigned an IP address when it is created. Since the address is not known before creation, and inbound access is required as part of the creation process, you cannot statically assign it on your firewall. Instead, if you are using a firewall with the VNet you must create a user-defined route to allow this inbound traffic.

• In your VNet, allow outbound traffic to the following service tags:

  Service tag	Protocol	Port	Notes
  AzureMachineLearning	TCP UDP	443/8787/18881 5831	Communication with the Azure Machine Learning service.
  BatchNodeManagement.<region>	ANY	443	Replace <region> with the Azure region that contains your Azure Machine Learning workspace. Communication with Azure Batch. Compute instance and compute cluster are implemented using the Azure Batch service.
  Storage.<region>	TCP	443	Replace <region> with the Azure region that contains your Azure Machine Learning workspace. This service tag is used to communicate with the Azure Storage account used by Azure Batch.

  Important

  The outbound access to Storage.<region> could potentially be used to exfiltrate data from your workspace. By using a Service Endpoint Policy, you can mitigate this vulnerability. For more information, see the Azure Machine Learning data exfiltration prevention article.

  FQDN	Protocol	Port	Notes
  <region>.tundra.ml.azure.cn	UDP	5831	Replace <region> with the Azure region that contains your Azure Machine Learning workspace.
  microsoftgraph.chinacloudapi.cn	TCP	443	Communication with the Microsoft Graph API.
  *.instances.ml.azure.cn	TCP	443/8787/18881	Communication with Azure Machine Learning.
  *.<region>.batch.chinacloudapi.cn	ANY	443	Replace <region> with the Azure region that contains your Azure Machine Learning workspace. Communication with Azure Batch.
  *.<region>.service.batch.chinacloudapi.cn	ANY	443	Replace <region> with the Azure region that contains your Azure Machine Learning workspace. Communication with Azure Batch.
  *.blob.core.chinacloudapi.cn	TCP	443	Communication with Azure Blob storage.
  *.queue.core.chinacloudapi.cn	TCP	443	Communication with Azure Queue storage.
  *.table.core.chinacloudapi.cn	TCP	443	Communication with Azure Table storage.

Use the following information to create a compute instance or cluster with a public IP address in the VNet:

# create a compute cluster with a public IP
az ml compute create --name cpu-cluster --resource-group rg --workspace-name ws --vnet-name yourvnet --subnet yoursubnet --type AmlCompute

# create a compute instance with a public IP
az ml compute create --name myci --resource-group rg --workspace-name ws --vnet-name yourvnet --subnet yoursubnet --type ComputeInstance

from azure.ai.ml.entities import AmlCompute, NetworkSettings

network_settings = NetworkSettings(vnet_name="<vnet-name>", subnet="<subnet-name>")
compute = AmlCompute(
    name=cpu_compute_target,
    size="STANDARD_D2_V2",
    min_instances=0,
    max_instances=4,
    network_settings=network_settings
)
ml_client.begin_create_or_update(entity=compute)

Use the following information to configure serverless compute nodes with a public IP address in the VNet for a given workspace:

az ml workspace create -n <workspace-name> -g <resource-group-name> --file serverlesscomputevnetsettings.yml

name: testserverlesswithvnet
location: eastus
serverless_compute:
  custom_subnet: /subscriptions/<sub id>/resourceGroups/<resource group>/providers/Microsoft.Network/virtualNetworks/<vnet name>/subnets/<subnet name>
  no_public_ip: false

az ml workspace update -n <workspace-name> -g <resource-group-name> --file serverlesscomputevnetsettings.yml

serverless_compute:
  custom_subnet: /subscriptions/<sub id>/resourceGroups/<resource group>/providers/Microsoft.Network/virtualNetworks/<vnet name>/subnets/<subnet name>
  no_public_ip: false

from azure.ai.ml import MLClient
from azure.ai.ml.entities import ServerlessComputeSettings, Workspace
from azure.identity import DefaultAzureCredential

subscription_id = <sub id>
resource_group = <resource group>
workspace_name = <workspace name>
ml_client = MLClient(
    DefaultAzureCredential(), subscription_id, resource_group
)

workspace = Workspace(
    name=workspace_name,
    serverless_compute=ServerlessComputeSettings(
        custom_subnet=<subnet id>,
        no_public_ip=false,
    )
)

workspace = ml_client.workspaces.begin_update(workspace)

Azure Databricks

• The virtual network must be in the same subscription and region as the Azure Machine Learning workspace.
• If the Azure Storage Account(s) for the workspace are also secured in a virtual network, they must be in the same virtual network as the Azure Databricks cluster.
• In addition to the databricks-private and databricks-public subnets used by Azure Databricks, the default subnet created for the virtual network is also required.
• Azure Databricks doesn't use a private endpoint to communicate with the virtual network.

For specific information on using Azure Databricks with a virtual network, see Deploy Azure Databricks in your Azure Virtual Network.

Virtual machine or HDInsight cluster

In this section, you learn how to use a virtual machine or Azure HDInsight cluster in a virtual network with your workspace.

Create the VM or HDInsight cluster

Create a VM or HDInsight cluster by using the Azure portal or the Azure CLI, and put the cluster in an Azure virtual network. For more information, see the following articles:

• Create and manage Azure virtual networks for Linux VMs

• Extend HDInsight using an Azure virtual network

Configure network ports

Allow Azure Machine Learning to communicate with the SSH port on the VM or cluster, configure a source entry for the network security group. The SSH port is usually port 22. To allow traffic from this source, do the following actions:

1. In the Source drop-down list, select Service Tag.

2. In the Source service tag drop-down list, select AzureMachineLearning.

   

3. In the Source port ranges drop-down list, select *.

4. In the Destination drop-down list, select Any.

5. In the Destination port ranges drop-down list, select 22.

6. Under Protocol, select Any.

7. Under Action, select Allow.

Keep the default outbound rules for the network security group. For more information, see the default security rules in Security groups.

If you don't want to use the default outbound rules and you do want to limit the outbound access of your virtual network, see the required public internet access section.

Attach the VM or HDInsight cluster

Attach the VM or HDInsight cluster to your Azure Machine Learning workspace. For more information, see Manage compute resources for model training and deployment in studio.

Required public internet access to train models

Azure Machine Learning requires both inbound and outbound access to the public internet. The following tables provide an overview of the required access and what purpose it serves. For service tags that end in .region, replace region with the Azure region that contains your workspace. For example, Storage.chinanorth:

You may also need to allow outbound traffic to Visual Studio Code and non-Microsoft sites for the installation of packages required by your machine learning project. The following table lists commonly used repositories for machine learning:

When using Azure Kubernetes Service (AKS) with Azure Machine Learning, allow the following traffic to the AKS VNet:

• General inbound/outbound requirements for AKS as described in the Restrict egress traffic in Azure Kubernetes Service article.
• Outbound to mcr.microsoft.com.
• When deploying a model to an AKS cluster, use the guidance in the Deploy ML models to Azure Kubernetes Service article.

For information on using a firewall solution, see Use a firewall with Azure Machine Learning.

Next steps

This article is part of a series on securing an Azure Machine Learning workflow. See the other articles in this series:

• Virtual network overview
• Secure the workspace resources
• Secure the inference environment
• Enable studio functionality
• Use custom DNS
• Use a firewall