Configure inbound and outbound network traffic

Azure Machine Learning requires access to servers and services on the public internet. When implementing network isolation, you need to understand what access is required and how to enable it.

Common terms and information

The following terms and information are used throughout this article:

• Azure service tags: A service tag is an easy way to specify the IP ranges used by an Azure service. For example, the AzureMachineLearning tag represents the IP addresses used by the Azure Machine Learning service.

  Important

  Azure service tags are only supported by some Azure services. For a list of service tags supported with network security groups and Azure Firewall, see the Virtual network service tags article.

  If you are using a non-Azure solution such as a 3rd party firewall, download a list of Azure IP Ranges and Service Tags. Extract the file and search for the service tag within the file. The IP addresses may change periodically.

• Region: Some service tags allow you to specify an Azure region. This limits access to the service IP addresses in a specific region, usually the one that your service is in. In this article, when you see <region>, substitute your Azure region instead. For example, BatchNodeManagement.<region> would be BatchNodeManagement.chinaeast2 if your Azure Machine Learning workspace is in the China East 2 region.

• Azure Batch: Azure Machine Learning compute clusters and compute instances rely on a back-end Azure Batch instance. This back-end service is hosted in a Azure subscription.

• Ports: The following ports are used in this article. If a port range isn't listed in this table, it's specific to the service and may not have any published information on what it's used for:

  Port	Description
  80	Unsecured web traffic (HTTP)
  443	Secured web traffic (HTTPS)
  445	SMB traffic used to access file shares in Azure File storage
  8787	Used when connecting to RStudio on a compute instance
  18881	Used to connect to the language server to enable IntelliSense for notebooks on a compute instance.

• Protocol: Unless noted otherwise, all network traffic mentioned in this article uses TCP.

Basic configuration

This configuration makes the following assumptions:

• You're using docker images provided by a container registry that you provide, and won't be using images provided by Microsoft.
• You're using a private Python package repository, and won't be accessing public package repositories such as pypi.org, *.anaconda.com, or *.anaconda.org.
• The private endpoints can communicate directly with each other within the VNet. For example, all services have a private endpoint in the same VNet:
  • Azure Machine Learning workspace
  • Azure Storage Account (blob, file, table, queue)

Inbound traffic

Outbound traffic

Recommended configuration for training and deploying models

Outbound traffic

To allow installation of Python packages for training and deployment, allow outbound traffic to the following host names:

Scenario: Install RStudio on compute instance

To allow installation of RStudio on a compute instance, the firewall needs to allow outbound access to the sites to pull the Docker image from. Add the following Application rule to your Azure Firewall policy:

• Name: AllowRStudioInstall
• Source Type: IP Address
• Source IP Addresses: The IP address range of the subnet where you will create the compute instance. For example, 172.16.0.0/24.
• Destination Type: FQDN
• Target FQDN: ghcr.io, pkg-containers.githubusercontent.com
• Protocol: Https:443

To allow the installation of R packages, allow outbound traffic to cloud.r-project.org. This host is used for installing CRAN packages.

Scenario: Using compute cluster or compute instance with a public IP

When using Azure Machine Learning compute instance or compute cluster (with a public IP address), allow inbound traffic from the Azure Machine Learning service. A compute instance or compute cluster with no public IP (preview) doesn't require this inbound communication. A Network Security Group allowing this traffic is dynamically created for you, however you may need to also create user-defined routes (UDR) if you have a firewall. When creating a UDR for this traffic, you can use either IP Addresses or service tags to route the traffic.

az network route-table route create -g MyResourceGroup --route-table-name MyRouteTable -n AzureMLRoute --address-prefix AzureMachineLearning --next-hop-type Internet

For information on configuring UDR, see Route network traffic with a routing table.

Scenario: Firewall between Azure Machine Learning and Azure Storage endpoints

You must also allow outbound access to Storage.<region> on port 445.

Scenario: Workspace created with the hbi_workspace flag enabled

You must also allow outbound access to Keyvault.<region>. This outbound traffic is used to access the key vault instance for the back-end Azure Batch service.

For more information on the hbi_workspace flag, see the data encryption article.

Scenario: Use Kubernetes compute

Kubernetes Cluster running behind an outbound proxy server or firewall needs extra egress network configuration.

• For AKS cluster, configure the AKS extension network requirements.

Besides above requirements, the following outbound URLs are also required for Azure Machine Learning,

In-cluster communication requirements

To install the Azure Machine Learning extension on Kubernetes compute, all Azure Machine Learning related components are deployed in a azureml namespace. The following in-cluster communication is needed to ensure the ML workloads work well in the AKS cluster.

• The components in azureml namespace should be able to communicate with Kubernetes API server.
• The components in azureml namespace should be able to communicate with each other.
• The components in azureml namespace should be able to communicate with kube-dns and konnectivity-agent in kube-system namespace.
• If the cluster is used for real-time inferencing, azureml-fe-xxx PODs should be able to communicate with the deployed model PODs on 5001 port in other namespace. azureml-fe-xxx PODs should open 11001, 12001, 12101, 12201, 20000, 8000, 8001, 9001 ports for internal communication.
• If the cluster is used for real-time inferencing, the deployed model PODs should be able to communicate with amlarc-identity-proxy-xxx PODs on 9999 port.

Scenario: Visual Studio Code

Visual Studio Code relies on specific hosts and ports to establish a remote connection.

Hosts

The hosts in this section are used to install Visual Studio Code packages to establish a remote connection between Visual Studio Code and compute instances in your Azure Machine Learning workspace.

Ports

You must allow network traffic to ports 8704 to 8710. The VS Code server dynamically selects the first available port within this range.

Scenario: Third party firewall or Azure Firewall without service tags

The guidance in this section is generic, as each firewall has its own terminology and specific configurations. If you have questions, check the documentation for the firewall you're using.

If not configured correctly, the firewall can cause problems using your workspace. There are various host names that are used both by the Azure Machine Learning workspace. The following sections list hosts that are required for Azure Machine Learning.

Dependencies API

You can also use the Azure Machine Learning REST API to get a list of hosts and ports that you must allow outbound traffic to. To use this API, use the following steps:

1. Get an authentication token. The following command demonstrates using the Azure CLI to get an authentication token and subscription ID:

   TOKEN=$(az account get-access-token --query accessToken -o tsv)
   SUBSCRIPTION=$(az account show --query id -o tsv)
   

2. Call the API. In the following command, replace the following values:

   • Replace <region> with the Azure region your workspace is in. For example, chinaeast2.
   • Replace <resource-group> with the resource group that contains your workspace.
   • Replace <workspace-name> with the name of your workspace.

   az rest --method GET \
       --url "https://<region>.api.ml.azure.cn/rp/workspaces/subscriptions/$SUBSCRIPTION/resourceGroups/<resource-group>/providers/Microsoft.MachineLearningServices/workspaces/<workspace-name>/outboundNetworkDependenciesEndpoints?api-version=2018-03-01-preview" \
       --header Authorization="Bearer $TOKEN"
   

The result of the API call is a JSON document. The following snippet is an excerpt of this document:

{
  "value": [
    {
      "properties": {
        "category": "Azure Active Directory",
        "endpoints": [
          {
            "domainName": "login.chinacloudapi.cn",
            "endpointDetails": [
              {
                "port": 80
              },
              {
                "port": 443
              }
            ]
          }
        ]
      }
    },
    {
      "properties": {
        "category": "Azure portal",
        "endpoints": [
          {
            "domainName": "management.chinacloudapi.cn",
            "endpointDetails": [
              {
                "port": 443
              }
            ]
          }
        ]
      }
    },
...

Azure hosts

The hosts in the following tables are owned by Microsoft, and provide services required for the proper functioning of your workspace. The tables list hosts for Azure operated by 21Vianet regions.

General Azure hosts

Azure Machine Learning hosts

Azure Machine Learning compute instance and compute cluster hosts

Docker images maintained by Azure Machine Learning

Also, use the information in the compute with public IP section to add IP addresses for BatchNodeManagement and AzureMachineLearning.

For information on restricting access to models deployed to AKS, see Restrict egress traffic in Azure Kubernetes Service.

Monitoring, metrics, and diagnostics

• dc.applicationinsights.azure.com
• dc.applicationinsights.microsoft.com
• dc.services.visualstudio.com
• *.in.applicationinsights.azure.com

For a list of IP addresses for these hosts, see IP addresses used by Azure Monitor.

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

• Enable studio functionality
• Use custom DNS

For more information on configuring Azure Firewall, see Tutorial: Deploy and configure Azure Firewall using the Azure portal.