Autoscaling a
multi-platform Kubernetes cluster built with kubeadm

A Swedish-speaking second-year Upper Secondary School (High School) Student from Finland

A person that has never attended a computing class :)

A maintainer of Kubernetes since a year back

The "kubernetes-on-arm" guy

$ whoami

I worked on and maintained minikube in the early days of the project, until I...

However, I wasn't satisfied with a side-project, I wanted it in core, so I implemented multiarch support for Kubernetes in the Spring 2016. I also wrote a multi-platform proposal

My first open source project was kubernetes-on-arm. It was the first easy solution to run Kubernetes on Raspberry Pi's

...moved on to kubeadm in August 2016, and started focusing on SIG-Cluster-Lifecycle issues, which I find very interesting and challenging.

What have I been tinkering with?

What am I gonna talk about?




What now?

Multi-platform Kuberentes

Definition and

Motivation and reasoning

How-to info and
a demo

Roadmap and
help-wanted issues


Definition and concepts

Briefly, when I talk about multi-platform Kubernetes I mean both:

 - the ability to run Kubernetes on a non-linux/amd64 computer

 - the ability to create clusters with computers of mixed platforms and use it smoothly

$ kubectl explain multiplatform

Binaries and docker images released by Kubernetes are cross-compiled and cross-built for non-amd64 architectures. A quick recap:

$ # Cross-compile main.go to ARM 32-bit
$ GOOS=linux GOARCH=arm CGO_ENABLED=0 go build main.go
$ # Cross-compile main.go (which contains CGO code) to ARM 32-bit
$ GOOS=linux GOARCH=arm CGO_ENABLED=1 CC=arm-linux-gnueabihf go build main.go

Cross-compilation with Go is relatively easy. Cross-building is a little bit harder, one may have to use QEMU to emulate another arch:

$ # Cross-build an armhf image with a RUN command that is executed on an amd64 host
$ cat Dockerfile
FROM armhf/debian:jessie
COPY qemu-arm-static /usr/bin/
RUN apt-get install iptables nfs-common
COPY hyperkube /

$ # Register the binfmt_misc module in the kernel and download QEMU
$ docker run --rm --privileged multiarch/qemu-user-static:register --reset
$ curl -sSL | tar -xz

$ docker build -t .
$ kubectl explain multiplatform


 - The first release I participated in, I made the release bundle include ARM 32-bit binaries


 - Server docker images are released for ARM, both 32 and 64-bit

 - kubelet chooses the right pause image and registers itself with{os,arch}


 - kubeadm released as an official deployment method that supports ARM 32 and 64-bit

 - Unfortunately, I had to use a patched Golang version for building ARM 32-bit binaries...


 - The patched Golang version for ARM could be removed.

 - I reenabled ppc64le builds and the community contributed s390x builds.

$ kubectl logs multiplatform


Motivation and reasoning

Platform agnostic. The specifications developed will not be platform specific such that they can be implemented on a variety of architectures and operating systems.

 -- CNCF Charter

We don’t want to make it be an exclusive choice. We don’t want to make people have to decide. We want to allow people to find the place that works for them.

-- Brendan Burns

Why is the multi-platform functionality important for Kubernetes long-term?

$ kubectl motivate multiplatform

1. We don't know which platform will be the dominating one in 20 years from now

2. By letting new architectures join the project, and more people with them, we'll see a stronger ecosystem and a sound competition.

3. The risk of vendor lock-in on the default platform is significantly reduced

What could Kubernetes on ARM be used for right now?

KubeCloud: A Small-Scale Tangible Cloud Computing Environment

 - A 163-page(!) master's thesis about educating Kubernetes' concepts by letting the students use Kubernetes on small Raspberry Pi clusters.

Microsoft Pledges to Use ARM Server Chips, Threatening Intel's Dominance

- The world's first 10nm processor is an ARM processor, exciting times!

In classrooms -- learning others how Kubernetes works by using Raspberry Pi's is the ideal way of letting newcomers actually see what it's all about


How-to info and a demo

Since kubeadm was announced, it has been super-easy to set up Kubernetes in an official way on ARM and now also on ppc64le and s390x

Example setup on an ARM machine:

$ curl -s | apt-key add -
$ cat <<EOF > /etc/apt/sources.list.d/kubernetes.list
deb kubernetes-xenial main
$ apt-get update && apt-get install -y kubeadm

$ kubeadm init

$ kubectl apply -f

$ # DONE!

TL;DR; Kubernetes shouldn't have different install paths for different platforms, it should just work out-of-the-box

How can I set up Kubernetes on an other architecture?

Kubernetes releases server binaries for all supported architectures (amd64, arm, arm64, ppc64le, s390x) and node binaries for all supported platforms (+windows/amd64)

All docker images in the core k8s repo are built and pushed for all architectures using a semi-standardized Makefile.

Debian packages are provided for all architectures as well, basically just downloads the binaries and makes debs of them

kubeadm is aware of which architecture it's running on on init and generates manifests for the right architecture.

How does it work under the hood?

I don't want to have the architecture in the image name!!

Me neither. Enter manifest lists.

Imagine this scenario...

$ go build my-cool-app.go
$ docker build -t luxas/my-cool-app-amd64:v1.0.0 .
$ docker push luxas/my-cool-app-amd64:v1.0.0

$ # ARM
$ GOARCH=arm go build my-cool-app.go
$ docker build -t luxas/my-cool-app-arm:v1.0.0 .
$ docker push luxas/my-cool-app-arm:v1.0.0

$ # ARM 64-bit
$ GOARCH=arm64 go build my-cool-app.go
$ docker build -t luxas/my-cool-app-arm64:v1.0.0 .
$ docker push luxas/my-cool-app-arm64:v1.0.0

Then you get excited and create a k8s cluster of amd64, arm and arm64 nodes

and try to run your application on that cluster. But what architecture should you use?

$ kubectl run --image luxas/my-cool-app-???:v1.0.0 my-cool-app --port 80
$ kubectl expose deployment my-cool-app --port 80

This the hardest problem with a multi-platform cluster, if you hardcode the architecture here, it will fail on all other machines. Ideally I would like to do this:

$ kubectl run --image luxas/my-cool-app:v1.0.0 my-cool-app --port 80
$ kubectl expose deployment my-cool-app --port 80

Fortunately, that's totally possible!

"Manifest list" is currently a Docker registry and client feature and I hope the general idea can propagate to other CRI implementations in the future.

The idea is very simple, you have one tag (e.g. luxas/my-cool-app:v1.0.0) that serves as a "redirector" to platform-specific images. The client will then download the right image digest based on what platform it's running on.

Ok, so now that I know what a manifest list is, how do I create it?

$ go build my-app.go
$ docker build -t luxas/my-app-amd64:v1.0.0 .
$ docker push luxas/my-app-amd64:v1.0.0

$ # ARM
$ GOARCH=arm go build my-app.go
$ docker build -t luxas/my-app-arm:v1.0.0 .
$ docker push luxas/my-app-arm:v1.0.0

$ # ARM 64-bit
$ GOARCH=arm64 go build my-app.go
$ docker build -t luxas/my-app-arm64:v1.0.0 .
$ docker push luxas/my-app-arm64:v1.0.0

$ wget
$ mv manifest-tool-linux-amd64 manifest-tool && chmod +x manifest-tool
$ export PLATFORMS=linux/amd64,linux/arm,linux/arm64
$ ./manifest-tool push from-args \
    --platforms $PLATFORMS \ # Which platforms the manifest list include
    --template luxas/my-app-ARCH:v1.0.0 \ # ARCH is a placeholder for the real architecture
    --target luxas/my-app:v1.0.0 # The name of the resulting manifest list

Finally, demo time!

Reproduce yourself by following this guide:

1. Creating the cluster with "kubeadm init"

2. Joining all nodes with "kubeadm join"


3. The Pod network, in this case Weave Net

4. The Kubernetes Dashboard and Heapster

5. Traefik as the Ingress Controller and Ngrok as proxy

6. InfluxDB and Grafana for storing and visualizing CPU/memory metrics

7. The Prometheus Operator and a Prometheus TPR

8. A sample Custom Metrics API Server that queries Prom

9. A sample app that serves a /metrics endpoint and a HPA v2

Autoscaling based on Custom Metrics?

Possible with v1.6 and some hacks (alpha)

Why do you do autoscaling here?

It's quite unrelated.

Well, it's very cool!

I wanted to demo it to ultimately show that it's possible for multiple platforms to coexist, even in a dynamic system.

(but very alpha at the same time)

What now?

Roadmap and help-wanted issues

The current situation is ok and works, but it could obviously be improved. Here are some shout-outs to the community:

- Automated CI testing for the other architectures using kubeadm

    - We might be able to use the CNCF cluster here?

- Formalize a standard specification for how Kubernetes binaries should be compiled and how server images should be built

    - Official Kubernetes projects should publish binaries for at least             amd64, arm, arm64, ppc64le, s390x and windows (node only)

- Manifest lists should be built for the server images

    - This is blocked on not supporting v2 schema 2 :(

- Implement this feature in other CRI-compliant implementations

- Creating an external Admission Controller that applies platform data 

What's yet to be done here?

Thank you for listening!


KubeCon Presentation

By lucask

KubeCon Presentation

  • 4,899