ogcanviz / virtual-kubelet-aci-burst

Run image recognition from Azure Container Service (AKS) and burst into Azure Container Instances (ACI) with the Virtual-Kubelet.

services	platforms	author
containerservice, containerinstances, virtual-kubelet	azure	ribhatia

Bursting from AKS to ACI with the Virtual Kubelet

Description: Run a facial recognition demo across your AKS cluster and use ACI and the ACI Connector to burst into for on-demand compute.

Contact [email protected] if you need help!

Steps to deploy

Part 1 Prerequisites

In order to run this demo, you will need the following:

• An active Microsoft Azure Subscription
• Azure CLI installed
• Kubernetes CLI (kubectl) installed
• Helm client installed

Part 2 Setup

Replace <myResourceGroup> with your expected and run following command to create resource group. Then remember the created resource group name.

$ az group create --name <myResourceGroup> --location eastus

Note: The AKS extending to ACI only support in one location. So far AKS and ACI are both deployed to East US and West Europe. So please create your resource group in East US or West Europe only.

Replace <myResourceGroup>,<myK8sCluster> with your expected and run following command to create the AKS. Then remember the AKS name.

$ az aks create --resource-group <myResourceGroup> --name <myK8sCluster> --node-count 1 --generate-ssh-keys

Replace <myResourceGroup>, <myK8sCluster> with your created in previous steps and run following command to set the AKS as your current connected cluster.

$ az aks get-credentials --resource-group <myResourceGroup> --name <myK8sCluster>

Make sure you're connected.

$ kubectl get nodes

Make sure you have helm installed and initialize this.

$ helm init

Replace <your email> with yours and run to install kube-lego w/ helm for certs.

$ helm install stable/kube-lego --name kube-lego --namespace kube-system --set config.LEGO_EMAIL=<your email>,config.LEGO_URL=https://acme-v01.api.letsencrypt.org/directory

Install ingress controller w/ helm.

$ helm install stable/nginx-ingress --name ingress --namespace kube-system

Get the Public IP of the ingress controller. It may take some times for the IP assigned to the services

$ kubectl get services --namespace kube-system --watch

When the IP assigned, remember the IP address and press Ctrl+Q to exit

Clone this repo

$ git clone https://github.com/rbitia/aci-demos.git

Change the folder to the root of the source code

$ cd aci-demo

Replace <myResourceGroup>, <IP Address> with your created in previous steps. Replace <appName> with your expected and run following shell script to bind FQDN to your IP. Remember the return FQDN name. You will use it to update your configuration file.

$ chmod u+x assignFQDNtoIP.sh
$ ./assignFQDNtoIP.sh -g <myResourceGroup> -d <appName> -i <IP Address>

Edit the values.yaml file and replace all <host name> with the FQDN name in pervious step.

$ vim ./charts/fr-demo/values.yaml 

Start at the top of the aci-demos directory and deploy the Facial Recognition application that consists of a frontend, a backend, and a set of image recognizers.

$ helm install charts/fr-demo --name demo

Checkout the UI that's generated in the output and see the pictures start to get processed The rate will be super slow because we have a 1 node AKS cluster running 1 worker pod.

Deploy the ACI connector : Replace <myResourceGroupmy>, <myK8sCluster> with yours in previous steps and run following command

az aks install-connector --resource-group <myResourceGroup> --name <myK8sCluster> --connector-name myaciconnector

The connector has been deployed and with a kubectl get nodes you can see that the ACI Connector is a new node in your cluster. Now scale up the image recognizer to 10 using the following command

$ kubectl scale deploy demo-fr-ir-aci --replicas 10

Though we are using kubectl, the ACI Connector is dispatching pods to Azure Container Instances transparently, via the ACI connector node. This virtual node has unlimited capacity and a per-second billing model, making it perfect for burst compute scenarios like this one. If we wait a minute or so for the ACI containers to warm up, we should see image recognizer throughput increase dramatically.

Check out the dashboard to see throughput it dramatically increase...

Here we can see throughput really beginning to pick up, thanks to burst capacity provided by ACI.   This is powerful stuff.  Here we can see AKS and ACI combine to provide the best of “serverless” computing – invisible infrastructure and micro-billing – all managed through the open Kubernetes APIs.  This kind of innovation – the marriage of containers and serverless computing -- is important for the industry, and Microsoft is working hard to make it a reality.

Once you've done all the set up you just need these commands during the live demo:

$ helm install charts/fr-demo --name demo
$ az aks install-connector --resource-group <myResourceGroup> --name <myK8sCluster> --connector-name myaciconnector
$ kubectl scale deploy demo-fr-ir-aci --replicas 10

Part 3 Clean up:

Replace <myResourceGroup>, <myK8sCluster> with your created in previous steps. Then run following command to clean up resources

$ helm del --purge demo
$ az aks remove-connector --resource-group <myResourceGroup> --name <myK8sCluster> --connector-name myaciconnector
$ az group delete --resource-group <myResourceGroup> --no-wait -y