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,
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.
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,
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
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.
### Base Image
# Setup up a base image to use in Build and Runtime images
FROM ecliptik/node:8.4.0-alpine-armhf AS base
WORKDIR /app
COPY package.json .
### Build Image
# Installs build dependencies and npm packages
# Creates artifacts to copy into Runtime image
FROM base AS build
# Install build OS packages
RUN set -ex && \
buildDeps=' \
make \
gcc \
g++ \
python \
py-pip \
curl \
openssl \
' && \
apk add --no-cache \
--virtual .build-deps $buildDeps
#Copy application into build image
COPY . .
# Install npm packages
RUN npm install -g
RUN npm install --silent --save-dev -g \
gulp-cli \
typescript
# Compile typescript sources to javascript artifacts
RUN tsc --target es5 connector.ts
### Runtime Image
# Copy artifacts from Build image and setups up entrypoint/cmd to run app
FROM base AS runtime
# Copy artifacts from Build Image
COPY --from=build /app/node_modules ./node_modules
COPY --from=build /app/*.js ./
COPY --from=build /app/LICENSE ./
# Runtime command
ENTRYPOINT ["node"]
CMD ["connector.js"]
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.
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).
Updated examples/aci-connector.yaml with RBAC role and ecliptik/aci-connector-k8s:alpine-armhf image
apiVersion: extensions/v1beta1
kind: Deployment
metadata:
name: aci-connector
namespace: default
spec:
replicas: 1
template:
metadata:
labels:
app: aci-connector
spec:
serviceAccountName: aci-connector-sa
containers:
- name: aci-connector
image: ecliptik/aci-connector-k8s:alpine-armhf
imagePullPolicy: Always
env:
- name: AZURE_CLIENT_ID
value: 00000-000-00000-0000-0000
- name: AZURE_CLIENT_KEY
value: 00000-000-00000-0000-0000
- name: AZURE_TENANT_ID
value: 00000-000-00000-0000-0000
- name: AZURE_SUBSCRIPTION_ID
value: 100000-000-00000-0000-0000
- name: ACI_RESOURCE_GROUP
value: aci-test
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