Docs: Deployment

Mon, 01 Jan 0001 00:00:00 +0000

Docs: End-to-End Pipeline Example on Azure

Mon, 01 Jan 0001 00:00:00 +0000

Introductions

Overview of Azure and AKS

Microsoft Azure is an open, flexible, enterprise-grade cloud computing platform running on Microsoft infrastructure. The platform has various services, many of which are extremely useful in a pipeline that works with ML models.

The Azure CLI is a set of tools that you can use to interact with Azure from the command line.

Azure Kubernetes Service (AKS) on Azure allows you to deploy containerized applications, within which you describe the resources your application needs, and AKS will manage the underlying resources automatically. This workflow is especially efficient at scale.

The overall workflow

This guide takes you through using your Kubeflow deployment to build a machine learning (ML) pipeline on Azure. This guide uses a sample pipeline to detail the process of creating an ML workflow from scratch. You will learn how to create and run a pipeline that processes data, trains a model, and then registers and deploys that model as a webservice.

To build your pipeline, you must create and build containers using Docker images. Containers are used to abstract the dependencies for each step of the pipeline. You can manage your containers using Azure’s portal, specifically using the Container Registry to store the containers in the cloud. Kubelfow pulls the containers from this registry as they are needed in each step of the pipeline.

By following this guide, you will learn how to:

Set up Kubeflow in an AKS Cluster
Create and compile a pipeline that can:
- Preprocess data
- Train a model
- Register the model to ACR (Azure Container Registry)
- Profile the model to optimize compute resources in AML (Azure Machine Learning)
- Deploy the model to AML
Interact with and customize your deployment
Test and use your deployed model

When your pipeline has finished running, you will be able to see a registered image, model, and deployment in your Azure ML workspace. You will then be able to visit the scoring URI and upload images for scoring in real time.

Set up your environment

Download the project files

This tutorial uses the Azure Pipelines example in the Kubeflow examples repo. You can optionally use a pipeline of your own, but several key steps may differ.

Clone the project files and go to the directory containing the Azure Pipelines (Tacos and Burritos) example:

git clone https://github.com/kubeflow/examples.git
cd examples/pipelines/azurepipeline

As an alternative to cloning, you can download the Kubeflow examples repository zip file.

Deploy Kubeflow

If you don’t already have one, create an Azure account. If you have not used Azure services before, you can recieve up to 1 year of free services and free credits.

Note: that some of the services used in this guide may not be included in the free services, but can be covered by free credits.

First, install the Azure CLI, then follow the instructions in the guide to deploying Kubeflow on Azure.

Ensure that the agent size you use has the proper memory and storage requirements. For the Azure Pipelines example, most machine sizes will work, but premium storage is required. Use this guide to choose the right agent size for your deployment. (We chose an agent size of Standard_D4s_v3.)

Configuring Azure resources

Create an ML workspace in Azure

Throughout your pipeline’s run, all of your models, images, and deployments will be pushed to your ML workspace in Azure. Your ML workspace also has support for managing your active deployments, which will be displayed later in this tutorial.

To create an ML workspace:

Go to the Azure portal and click on your resource group.
Select the add a new resource option.
Search for Machine Learning Studio Workspace and use the default options, taking note of the name you decide for it.

Create an Azure container registry

Kubeflow uses Docker images to describe each pipeline step’s dependencies. You need to create a container registry to store those images in the cloud so that Kubeflow can pull the images as they are needed.

To create a container registry:

Go to the Azure portal and click on your resource group.
From there, select the add a new resource option.
Search for Container Registry and add it to your resource group.
Configure your registry by selecting and noting the name you use for it. Enable an admin user, and change the SKU option to Premium.

Allow your AKS Cluster access to your Azure Container Registry

In order for the AKS cluster to have access to pulling images created for execution of the pipeline, you will need to update your cluster so that it is able to pull the images from the container registry we just created. More references can be found here, on the Microsoft web site.

Using a bash shell, use the following commands to attach the container registry created above to your AKS cluster, using the proper AKS cluster name and resource group:

az aks update -n <myAKSCluster> -g <MyResourceGroup> --attach-acr <REGISTRY_NAME>

The execution of the command will take a few minutes.

Create a persistent volume claim (PVC)

A persistent volume claim is a dynamically provisioned storage resource attached to a Kubernetes cluster. It is used in the pipeline to store data and files across pipeline steps.

Using a bash shell, navigate to the azurepipeline directory. Use the following commands to create a persistent volume claim for your cluster.

cd kubernetes
kubectl apply -f pvc.yaml

Authenticate your service principal

A service principal is used to allow your pipeline to securely interface with your Azure services without having to directly login in the pipeline and use admin privileges. To create a service principal with Contributor access to your Azure account, use the following steps.

Create an App Registration

To create an app registration:

In the Azure Portal, navigate to Azure Active Directory.
Select App registrations and click New registration. Name it, noting the name and use the default options.
Click Register.
You should be redirected to your app registration’s dashboard. Select Overview from the sidebar.
Make note of the Application (client) ID and the Directory (tenant) ID. The client ID is your service principal username. Save these in a secure location.
Select Certificates and Secrets from the sidebar.
Select New client secret. Give the client secret a description and select how long you would like it to remain active for. Once you click the Add button, make sure you take note of the client secret value and save it in a secure place. This is your service principal password.

Add a role assignment

To add a role assignment for your service principal:

Go to your resource group page on the Azure Portal.
Select Access control (IAM) from the sidebar. Select Add a role assignment.
Set the role to Contributor and search for the name you gave your app registration in the Select dropdown.
Click Save.

Creating containers from Docker images

Install Docker

You need to install Docker to be able to push and pull images to/from your Container registry.

For Windows and WSL: Guide

For other OS: Docker Desktop

Build images

To deploy your code to Kubernetes, you must build your local project’s Docker images and push the containers to your Container Registry so that they are available in the cloud.

Set the path in Container Registry that you want to push the containers to:

export REGISTRY_PATH=<REGISTRY_NAME>.azurecr.io

Run the following command to authenticate your Container Registry:

az acr login --name <REGISTRY_NAME>

Create a version, to be associated with your model each time it runs (change this accordingly):

export VERSION_TAG=1

Each docker image will be built and uploaded to the cloud using the Container Registry.

Note: If you would like to test a container locally, you can use the docker run -it ${REGISTRY_PATH}<CONTAINER NAME>:$(VERSION_TAG} before pushing to Container Registry.

//Starting in the 'code' directory of the azurepipeline folder

cd preprocess
docker build . -t ${REGISTRY_PATH}/preprocess:${VERSION_TAG}
docker push ${REGISTRY_PATH}/preprocess:${VERSION_TAG}

cd ../training
docker build . -t ${REGISTRY_PATH}/training:${VERSION_TAG}
docker push ${REGISTRY_PATH}/training:${VERSION_TAG}

cd ../register
docker build . -t ${REGISTRY_PATH}/register:${VERSION_TAG}
docker push ${REGISTRY_PATH}/register:${VERSION_TAG}

cd ../deploy
docker build . -t ${REGISTRY_PATH}/deploy:${VERSION_TAG}
docker push ${REGISTRY_PATH}/deploy:${VERSION_TAG}

When all of the images are pushed successfully, modify the pipeline.py file to use the appropriate image for each pipeline step.

Running and deploying your pipeline

Compile

To compile the pipeline, simply open a terminal and navigate to the azurepipeline/code folder. Run the following command to generate a pipeline in the tar.gz format: python pipeline.py

Run and deploy

Upload the pipeline.tar.gz file to the pipelines dashboard on your Kubeflow deployment.

Create an experiment and then create a run using the pipeline you just uploaded.

The finished pipeline should have four completed steps.

Pushing images for scoring

Once your pipeline has finished successfully, you can visit the Azure portal to find your deployment url. Go to your ML workspace dashboard and select “Deployments” from the sidebar. Click on the most recent deployment. You should see a link under “Scoring URI”. You can whatever method you know best to send a GET/POST request with an image of a taco or a burrito to this url and it should return whether or not the image is of a taco or a burrito.

The easiest method is to find a url of an image of a taco or a burrito and append it to your scoring url as follows: <scoring_url>?image=<image_url>

Clean up your Azure environment

When you are done, make sure you delete your resource group to avoid extra charges.

az group delete -n MyResourceGroup

You can optionally choose to delete individual resources on your clusters using the Azure cluster docs.

Next steps

Build your own pipeline using the Kubeflow Pipelines SDK.

Docs: Access Control for Azure Deployment

Mon, 01 Jan 0001 00:00:00 +0000

This section shows how to restrict access to only certain IP addresses for your LoadBalancer Service on Azure. At a later date, it will also include formal authentication through Azure. This method is not the most ideal way to secure your Kubernetes cluster, as it requires that you access the service from the same IP address every time. This process was adapted from the Kubernetes guide to configuring a firewall.

When using a service with spec.type: LoadBalancer, you can specify the IP ranges that are allowed to access the load balancer by using spec.loadBalancerSourceRanges. This is currently supported on all major cloud providers.

Editing the LoadBalancer Service

Use the kubectl edit svc <loadbalancer-name> -n kubeflow to add your source ranges. This command will open the editor defined by you KUBE_EDITOR or EDITOR environment variables or fall back to ‘vi’ for Linux or ‘notepad’ for Windows. More information about using alternative editors and options for this command can be found in the kubectl edit documentation.

Internal Subnet Access

Assuming 10.0.0.0/8 is the address for the internal subnet, a load balancer will be created such that the deployment is only accessible from internal Kubernetes cluster IPs. This will not allow clients from outside your Kubernetes cluster to access the load balancer.

apiVersion: v1
kind: Service
metadata:
  name: myapp
spec:
  ports:
  - port: 8765
    targetPort: 9376
  selector:
    app: example
  type: LoadBalancer
  loadBalancerSourceRanges:
  - 10.0.0.0/8

External IP Addresses

In the following example, a load balancer will be created that is only accessible to clients with IP addresses from 130.211.204.1 and 130.211.204.2.

apiVersion: v1
kind: Service
metadata:
  name: myapp
spec:
  ports:
  - port: 8765
    targetPort: 9376
  selector:
    app: example
  type: LoadBalancer
  loadBalancerSourceRanges:
  - 130.211.204.1/32
  - 130.211.204.2/32

Docs: Troubleshooting Deployments on Azure AKS

Mon, 01 Jan 0001 00:00:00 +0000

Jupyter Notebook ‘is not a valid page’ when accessing notebook

Restarting the ambassador pods will often fix this issue: kubectl delete pods -l service=ambassador

The client does not have authorization to perform action…

This is likely an issue with ensuring your subscriptions are correctly setup. To fix the issue, list your subscriptions and then change the active subscription.

#list subscription 
az account list --output table 

#change the active subscription 
az account set --subscription "My Demos"

Kubeflow – Kubeflow on Azure

Docs: Deployment

Docs: End-to-End Pipeline Example on Azure

Introductions

Overview of Azure and AKS

The overall workflow

Set up your environment

Download the project files

Deploy Kubeflow

Configuring Azure resources

Create an ML workspace in Azure

Create an Azure container registry

Allow your AKS Cluster access to your Azure Container Registry

Create a persistent volume claim (PVC)

Authenticate your service principal

Create an App Registration

Add a role assignment

Creating containers from Docker images

Install Docker

Build images

Running and deploying your pipeline

Compile

Run and deploy

Pushing images for scoring

Clean up your Azure environment

Next steps

Docs: Access Control for Azure Deployment

Editing the LoadBalancer Service

Internal Subnet Access

External IP Addresses

Docs: Troubleshooting Deployments on Azure AKS

Jupyter Notebook ‘is not a valid page’ when accessing notebook

The client does not have authorization to perform action…