ACI Connector for k8s on a Raspberry Pi Cluster

Running the Azure Container Instances Connector for Kubernetes on a Raspberry Pi Cluster.

One of the most interesting features of Azure Container Instances is the Azure Container Instances Connector for Kubernetes. This adds an ACI “node” to an existing Kubernetes cluster and allows you to deploy pods to it. This “node” will run pods in ACI without having to create or manage and additional Azure VMs, just point-and-shoot a pod at it and it will run with no additional setup required.

By using the ACI Connector for Kubernetes on a Raspberry PI, a cluster can run homogenous ARM containers on-prem, but still have the ability to deploy and manage x86 containers to a cloud provider.

Read more about Azure Container Instances,

Creating an aci-connector-k8s ARM Image for Raspberry Pi

The upstream aci-connector-k8s image is x86 only, but since it’s written in typescript it can easily be run on different architectures. To run on a Raspberry Pi k8s cluster, all that is required is building an armhf Docker image.

Building a Nodejs ARM Docker Image

Note: As of 9/12/2017 Docker Hub Official Images support multi-platform and re-building an image for armhf (or arm64, ppc64le, and s390x) is no longer required if using a supported image (currently Debian based only, Alpine based like below still need to be re-built).

The aci-connector-k8s Dockerfile uses node:8.4.0-alpine as it’s base image. While there are some “unofficial” node ARM images, lets create one from a somewhat official repository. This involves finding an armhf Alpine 3.6 image, copying/pasting the node:8.4-alpine Dockerfile, replacing the FROM with the armhf version of alpine, and building the image.

There are two ways to build an armhf image,

  • Build on Raspberry Pi
    • 100% native build, less likely to run into bugs
    • Painfully slow due to Raspberry Pi hardware
  • Cross-build using Docker for Mac or Multiarch
    • Dramatically speed up builds
    • Could have unforseen issues due to running in an emulated environment

After first attempting option 1, two hours later and losing the ability to ssh into the Raspberry Pi, option 2 is a much faster approach. Building on a MacBook Pro using the built-in multiarch features of Docker for Mac works well, but is still slow even on a 4 core system. Fortunately Using up a 24-core Packet.net Type 2 bare-metal instance to cross-compile using Multiarch is easy to do too.

Note: The official Docker images for armhf are arm32v6 and arm32v7. These will work natively on a Raspberry Pi, Docker for Mac, Multiarch, and a Linux system with qemu-*-static support. For full details on these see my post on cross-building Docker images.

A armhf multiarch nodejs Dockerfile was built on the Packet.net Type 2 instance and pushed to Docker hub as ecliptik/node:8.4.0-alpine-armhf. This only took a few minutes using the Type 2 instance, much faster than Raspberry Pi or Macbook Pro.

Example Using Multiarch to re-build a node armhf alpine image,

curl -sSL https://raw.githubusercontent.com/nodejs/docker-node/c044d61e6d02756bb8ed1557b2f0c7a0d7fead6f/8.4/alpine/Dockerfile | sed "s/alpine:3.6/multiarch\/alpine:armhf-v3.6/" > Dockerfile.node.armhf

docker build -f Dockerfile.node.armhf -t ecliptik/node:8.4.0-alpine-armhf .
docker push ecliptik/node:8.4.0-alpine-armhf

Building an aci-connector-k8s ARM Docker Image

Once a nodejs arm-alpine image is created, clone the aci-connector-k8s repositoriy, and update the Dockerfile to use the ecliptik/node:8.4.0-alpine-armhf image. Additionaly, use the Dockefile below to use multi-stage builds for improved image size.

With an updated Dockerfile in the cloned repo, build the aci-connector-k8s image on a Raspberry Pi,

docker build -t ecliptik/aci-connector-k8s:alpine-armhf .
docker push ecliptik/aci-connector-k8s:alpine-armhf

Note: Trying to build the image on non-native armhf platform like Docker for Mac or Multiarch may result in errors like "SyntaxError: Unexpected end of JSON input". The image only seems to build on native Raspberry Pi or ARM hardware.

Running the ACI Connector

Now that we have a armhf image capable of running on a Raspberry Pi, we can deploy the pod to a Raspberry Pi Kubernetes cluster.

First clone the aci-connector-k8s repository onto the Raspberry Pi cluster master,

git clone https://github.com/Azure/aci-connector-k8s.git

Edit the examples/aci-connector.yaml and update the image to use the ecliptik/aci-connector-k8s:alpine-armhf image.

Next, if you used kubeadm to create your cluster and RBAC is enabled, you’ll need to create a role and set it up for the connector. This is discussed in this Github issue that includes creating a RBAC role and updating the service to use it.

Create the RBAC role for the connector,

curl -sSL https://raw.githubusercontent.com/alexjmoore/aci-connector-k8s-arm/master/aci-connector-rbac.yaml | kubectl create -f -

Under spec in the examples/aci-connector.yaml add the RBAC role,

serviceAccountName: aci-connector-sa

Finally after the connector is setup to use the armhf image and RBAC, follow the rest of the Quickstart guide in the aci-connector-k8s README to set up everything else required to run the connector (Azure keys, deployment of service, etc).

Working Example

Updated examples/aci-connector.yaml with RBAC role and ecliptik/aci-connector-k8s:alpine-armhf image

Deploy the aci-connector pod,

kubectl create -f examples/aci-connector.yaml

Wait a few minutes while the pod comes into service (mostly waiting for the image to pull) on a worker node.

Verify aci-connector pod has started,

kubectl get pods
NAME                             READY     STATUS    RESTARTS   AGE
aci-connector-1252680567-b88w6   1/1       Running   0          3m

Verify aci-connector node is added,

kubectl get nodes -o wide
NAME            STATUS    AGE       VERSION   EXTERNAL-IP   OS-IMAGE                        KERNEL-VERSION
aci-connector   Ready     2m        v1.6.6    <none>        <unknown>                       <unknown>
navi            Ready     1h        v1.7.5    <none>        Raspbian GNU/Linux 8 (jessie)   4.4.50-hypriotos-v7+
tael            Ready     1h        v1.7.5    <none>        Raspbian GNU/Linux 8 (jessie)   4.4.50-hypriotos-v7+
tatl            Ready     1h        v1.7.5    <none>        Raspbian GNU/Linux 8 (jessie)   4.4.50-hypriotos-v7+

Deploy example/nginx-pod.yaml pod from aci-connector-k8s repo,

kubectl create -f examples/nginx-pod.yaml
pod "nginx" created

Verify pod deployed and is running in ACI,

 kubectl get pods -o wide
 NAME                             READY     STATUS    RESTARTS   AGE       IP               NODE
 aci-connector-1696751608-tcjcq   1/1       Running   0          24m       10.244.2.4       tael
 nginx                            1/1       Running   0          10s       104.42.235.280   aci-connector

While researching and setting this up I came across many good resources on running Docker on ARM,

ARM Docker Image Repositories

Written on September 12, 2017