Deploy a model to an Azure Kubernetes Service cluster

Learn how to use Azure Machine Learning to deploy a model as a web service on Azure Kubernetes Service (AKS). Azure Kubernetes Service is good for high-scale production deployments. Use Azure Kubernetes service if you need one or more of the following capabilities:

• Fast response time
• Autoscaling of the deployed service
• Logging
• Model data collection
• Authentication
• TLS termination
• Hardware acceleration options such as GPU and field-programmable gate arrays (FPGA)

When deploying to Azure Kubernetes Service, you deploy to an AKS cluster that is connected to your workspace. For information on connecting an AKS cluster to your workspace, see Create and attach an Azure Kubernetes Service cluster.

Prerequisites

• An Azure Machine Learning workspace. For more information, see Create an Azure Machine Learning workspace.

• A machine learning model registered in your workspace. If you don't have a registered model, see How and where to deploy models.

• The Azure CLI extension (v1) for Machine Learning service, Azure Machine Learning Python SDK, or the Azure Machine Learning Visual Studio Code extension.

• The Python code snippets in this article assume that the following variables are set:

  • ws - Set to your workspace.
  • model - Set to your registered model.
  • inference_config - Set to the inference configuration for the model.

  For more information on setting these variables, see How and where to deploy models.

• The CLI snippets in this article assume that you've created an inferenceconfig.json document. For more information on creating this document, see How and where to deploy models.

• An Azure Kubernetes Service cluster connected to your workspace. For more information, see Create and attach an Azure Kubernetes Service cluster.

  • If you want to deploy models to GPU nodes or FPGA nodes (or any specific SKU), then you must create a cluster with the specific SKU. There is no support for creating a secondary node pool in an existing cluster and deploying models in the secondary node pool.

Understand the deployment processes

The word "deployment" is used in both Kubernetes and Azure Machine Learning. "Deployment" has different meanings in these two contexts. In Kubernetes, a Deployment is a concrete entity, specified with a declarative YAML file. A Kubernetes Deployment has a defined lifecycle and concrete relationships to other Kubernetes entities such as Pods and ReplicaSets. You can learn about Kubernetes from docs and videos at What is Kubernetes?.

In Azure Machine Learning, "deployment" is used in the more general sense of making available and cleaning up your project resources. The steps that Azure Machine Learning considers part of deployment are:

1. Zipping the files in your project folder, ignoring those specified in .amlignore or .gitignore
2. Scaling up your compute cluster (Relates to Kubernetes)
3. Building or downloading the dockerfile to the compute node (Relates to Kubernetes)
   1. The system calculates a hash of:
      • The base image
      • Custom docker steps (see Deploy a model using a custom Docker base image)
      • The conda definition YAML (see Create & use software environments in Azure Machine Learning)
   2. The system uses this hash as the key in a lookup of the workspace Azure Container Registry (ACR)
   3. If it is not found, it looks for a match in the global ACR
   4. If it is not found, the system builds a new image (which will be cached and pushed to the workspace ACR)
4. Downloading your zipped project file to temporary storage on the compute node
5. Unzipping the project file
6. The compute node executing python <entry script> <arguments>
7. Saving logs, model files, and other files written to ./outputs to the storage account associated with the workspace
8. Scaling down compute, including removing temporary storage (Relates to Kubernetes)

Azure ML router

The front-end component (azureml-fe) that routes incoming inference requests to deployed services automatically scales as needed. Scaling of azureml-fe is based on the AKS cluster purpose and size (number of nodes). The cluster purpose and nodes are configured when you create or attach an AKS cluster. There is one azureml-fe service per cluster, which may be running on multiple pods.

Azureml-fe scales both up (vertically) to use more cores, and out (horizontally) to use more pods. When making the decision to scale up, the time that it takes to route incoming inference requests is used. If this time exceeds the threshold, a scale-up occurs. If the time to route incoming requests continues to exceed the threshold, a scale-out occurs.

When scaling down and in, CPU usage is used. If the CPU usage threshold is met, the front end will first be scaled down. If the CPU usage drops to the scale-in threshold, a scale-in operation happens. Scaling up and out will only occur if there are enough cluster resources available.

When scale-up or scale-down, azureml-fe pods will be restarted to apply the cpu/memory changes. Inferencing requests are not affected by the restarts.

Understand connectivity requirements for AKS inferencing cluster

When Azure Machine Learning creates or attaches an AKS cluster, AKS cluster is deployed with one of the following two network models:

• Kubenet networking - The network resources are typically created and configured as the AKS cluster is deployed.
• Azure Container Networking Interface (CNI) networking - The AKS cluster is connected to an existing virtual network resource and configurations.

For Kubenet networking, the network is created and configured properly for Azure Machine Learning service. For the CNI networking, you need to understand the connectivity requirements and ensure DNS resolution and outbound connectivity for AKS inferencing. For example, you may be using a firewall to block network traffic.

The following diagram shows the connectivity requirements for AKS inferencing. Black arrows represent actual communication, and blue arrows represent the domain names. You may need to add entries for these hosts to your firewall or to your custom DNS server.

For general AKS connectivity requirements, see Control egress traffic for cluster nodes in Azure Kubernetes Service.

For accessing Azure ML services behind a firewall, see How to access azureml behind firewall.

Overall DNS resolution requirements

DNS resolution within an existing VNet is under your control. For example, a firewall or custom DNS server. The following hosts must be reachable:

Connectivity requirements in chronological order: from cluster creation to model deployment

In the process of AKS create or attach, Azure ML router (azureml-fe) is deployed into the AKS cluster. In order to deploy Azure ML router, AKS node should be able to:

• Resolve DNS for AKS API server
• Resolve DNS for MCR in order to download docker images for Azure ML router
• Download images from MCR, where outbound connectivity is required

Right after azureml-fe is deployed, it will attempt to start and this requires to:

• Resolve DNS for AKS API server
• Query AKS API server to discover other instances of itself (it is a multi-pod service)
• Connect to other instances of itself

Once azureml-fe is started, it requires the following connectivity to function properly:

• Connect to Azure Storage to download dynamic configuration
• Resolve DNS for AAD authentication server api.ml.azure.cn and communicate with it when the deployed service uses Azure AD authentication.
• Query AKS API server to discover deployed models
• Communicate to deployed model PODs

At model deployment time, for a successful model deployment AKS node should be able to:

• Resolve DNS for customer's ACR
• Download images from customer's ACR
• Resolve DNS for Azure BLOBs where model is stored
• Download models from Azure BLOBs

After the model is deployed and service starts, azureml-fe will automatically discover it using AKS API and will be ready to route request to it. It must be able to communicate to model PODs.

Deploy to AKS

To deploy a model to Azure Kubernetes Service, create a deployment configuration that describes the compute resources needed. For example, number of cores and memory. You also need an inference configuration, which describes the environment needed to host the model and web service. For more information on creating the inference configuration, see How and where to deploy models.

from azureml.core.webservice import AksWebservice, Webservice
from azureml.core.model import Model
from azureml.core.compute import AksCompute

aks_target = AksCompute(ws,"myaks")
# If deploying to a cluster configured for dev/test, ensure that it was created with enough
# cores and memory to handle this deployment configuration. Note that memory is also used by
# things such as dependencies and AML components.
deployment_config = AksWebservice.deploy_configuration(cpu_cores = 1, memory_gb = 1)
service = Model.deploy(ws, "myservice", [model], inference_config, deployment_config, aks_target)
service.wait_for_deployment(show_output = True)
print(service.state)
print(service.get_logs())

az ml model deploy --ct myaks -m mymodel:1 -n myservice --ic inferenceconfig.json --dc deploymentconfig.json

{
    "computeType": "aks",
    "autoScaler":
    {
        "autoscaleEnabled": true,
        "minReplicas": 1,
        "maxReplicas": 3,
        "refreshPeriodInSeconds": 1,
        "targetUtilization": 70
    },
    "dataCollection":
    {
        "storageEnabled": true
    },
    "authEnabled": true,
    "containerResourceRequirements":
    {
        "cpu": 0.5,
        "memoryInGB": 1.0
    }
}

Autoscaling

The component that handles autoscaling for Azure ML model deployments is azureml-fe, which is a smart request router. Since all inference requests go through it, it has the necessary data to automatically scale the deployed model(s).

Autoscaling can be controlled by setting autoscale_target_utilization, autoscale_min_replicas, and autoscale_max_replicas for the AKS web service. The following example demonstrates how to enable autoscaling:

aks_config = AksWebservice.deploy_configuration(autoscale_enabled=True, 
                                                autoscale_target_utilization=30,
                                                autoscale_min_replicas=1,
                                                autoscale_max_replicas=4)

Decisions to scale up/down is based off of utilization of the current container replicas. The number of replicas that are busy (processing a request) divided by the total number of current replicas is the current utilization. If this number exceeds autoscale_target_utilization, then more replicas are created. If it is lower, then replicas are reduced. By default, the target utilization is 70%.

Decisions to add replicas are eager and fast (around 1 second). Decisions to remove replicas are conservative (around 1 minute).

You can calculate the required replicas by using the following code:

from math import ceil
# target requests per second
targetRps = 20
# time to process the request (in seconds)
reqTime = 10
# Maximum requests per container
maxReqPerContainer = 1
# target_utilization. 70% in this example
targetUtilization = .7

concurrentRequests = targetRps * reqTime / targetUtilization

# Number of container replicas
replicas = ceil(concurrentRequests / maxReqPerContainer)

For more information on setting autoscale_target_utilization, autoscale_max_replicas, and autoscale_min_replicas, see the AksWebservice module reference.

Web service authentication

When deploying to Azure Kubernetes Service, key-based authentication is enabled by default. You can also enable token-based authentication. Token-based authentication requires clients to use an Azure Active Directory account to request an authentication token, which is used to make requests to the deployed service.

To disable authentication, set the auth_enabled=False parameter when creating the deployment configuration. The following example disables authentication using the SDK:

deployment_config = AksWebservice.deploy_configuration(cpu_cores=1, memory_gb=1, auth_enabled=False)

For information on authenticating from a client application, see the Consume an Azure Machine Learning model deployed as a web service.

Authentication with keys

If key authentication is enabled, you can use the get_keys method to retrieve a primary and secondary authentication key:

primary, secondary = service.get_keys()
print(primary)

Authentication with tokens

To enable token authentication, set the token_auth_enabled=True parameter when you are creating or updating a deployment. The following example enables token authentication using the SDK:

deployment_config = AksWebservice.deploy_configuration(cpu_cores=1, memory_gb=1, token_auth_enabled=True)

If token authentication is enabled, you can use the get_token method to retrieve a JWT token and that token's expiration time:

token, refresh_by = service.get_token()
print(token)

Vulnerability scanning

Microsoft Defender for Cloud provides unified security management and advanced threat protection across hybrid cloud workloads. You should allow Microsoft Defender for Cloud to scan your resources and follow its recommendations. For more, see Azure Kubernetes Services integration with Defender for Cloud.

Next steps

• Use Azure RBAC for Kubernetes authorization
• Secure inferencing environment with Azure Virtual Network
• How to deploy a model using a custom Docker image
• Deployment troubleshooting
• Update web service
• Use TLS to secure a web service through Azure Machine Learning
• Consume a ML Model deployed as a web service
• Monitor your Azure Machine Learning models with Application Insights
• Collect data for models in production