azure-k3s-cluster

A (WIP) dynamically resizable k3s cluster for Azure, based on my azure-docker-swarm-cluster project.

What

This is an Azure Resource Manager template that automatically deploys a k3s cluster atop Ubuntu LTS. This cluster has a single master VM and a VM scaleset for workers/agents, plus required network infrastructure.

The template defaults to deploying B-Series VMs (B1ls) with the smallest possible managed disk size (S4, 32GB). It also deploys (and mounts) an Azure File Share on all machines with (very) permissive access at /srv, which makes it quite easy to run stateful services without messing about with volume claims.

The key aspect of this template is that you can add and remove agents at will simply by resizing the VM scaleset, which is very handy when running the node pool as spot instances - the cluster comes with a few (very simple) helper scripts that allow nodes to join and leave the cluster as they are created/destroyed, and the k3s scheduler will redeploy pods as needed.

Why

This was originally built as a Docker Swarm template, and even though Azure has a perfectly serviceable Kubernetes managed service, I enjoy the challenge of building my own stuff and fine-tuning it.

k3s is a breath of fresh air, and an opportunity to play around with a simpler, slimmer version of Kubernetes--and break it to see what happens.

Also, a lot of the ARM templating involved (for metrics, managed identities, etc.) lacked comprehensive samples when I started the project, so this was also a way for me to provide a fully working example that other people can learn from.

Roadmap

• air-gapped (i.e., standalone) install without curl
• test metrics server
• document cloud-config
• variables for more permissive connectivity levels
• clean kubernetes-dashboard deployment
• WIP: sample deployments/pods/charts
• TODO: Leverage Instance Protection and Scale-In Policies
• WIP: simple Python scale-down helper (blog post on how I'm going to do that with managed service identities and the instance metadata service)
• support an easy way to automatically redeploy charts and manifests when the compute resource group is recreated (/srv/autostart?)
• Tailscale installed for accessing the master securely without a public IP (you still need to log in once and set it up manually with sudo tailscale up after deployment)
• Preliminary NVIDIA support (for agents only - installing the drivers takes a LONG time and does not work on all SKUs yet)
• VM SKUs and agent priorities can now be set from environment variables in Makefile
• Button down NSGs to allow access from fewer IP addresses by default
• linted ARM template using VS Code
• upgrade to k3s v1.25.2+k3s1 and allow setting cluster version in Makefile
• add pre-defined NSGs for Cloudflare ingress and origin SSH access (TODO: Separate handling for IPv6)
• support an (insecure) private registry hosted on the master node (requires using docker instead of containerd, but saves a lot of hassle when doing tests)
• allow setting SSH_PORT via Makefile
• upgrade to k3s v1.22.3+k3s1
• temporarily remove Docker support so that I can explore k3c
• temporarily remove advanced Azure Linux diagnostics extensions (3.0) due to incompatibility with Ubuntu 20.04 (extension looks for python instead of python3)
• update Azure templates, helper scripts and cloud-config for Ubuntu 20.04 and python3
• upgrade to k3s v1.19.4+k3s1
• Handle eviction notifications
• Use Spot Instances for node pool
• upgrade to k3s v1.17.0+k3s.1
• upgrade to k3s 1.0.1
• upgrade to k3s 1.0.0
• upgrade to k3s 0.8.0
• upgrade to k3s 0.7.0
• upgrade to k3s 0.6.0
• re-usable user-assigned service identity instead of system (per-machine)
• Managed Service Identity for master and role allocations to allow it to manage the scaleset (and the rest of the resource group)
• add Linux Monitoring Extension (3.x) to master and agents (visible in the "Guest (classic)" metrics namespace in Azure Portal)
• scratch folder on agents' temporary storage volume (on-hypervisor SSD), available as /mnt/scratch
• set timezone
• bash completion for kubectl in master node
• remove scale set load balancer (everything must go through traefik on the master)
• re-enable first-time reboot after OS package updates
• private registry on master node
• trivial ingress through master node (built-in)
• Set node role labels
• install k3s viacloud-config
• change cloud-config to expose k3s token to agents
• remove unused packages from cloud-config
• remove unnecessary commands from Makefile
• remove unnecessary files from repo and trim history
• fork, new README

Makefile commands

• make keys - generates an SSH key for provisioning
• make deploy-storage - deploys shared storage
• make params - generates ARM template parameters
• make deploy-compute - deploys cluster resources and pre-provisions Docker on all machines
• make view-deployment - view deployment status
• make watch-deployment - watch deployment progress
• make list-agents - lists all agent VMs
• make scale-agents-<number> - scales the agent VM scale set to <number> instances, i.e., make scale-10 will resize it (up or down) to 10 VMs
• make stop-agents - stops all agents
• make start-agents - starts all agents
• make reimage-agents-parallel - nukes and paves all agents
• make reimage-agents-serial - reimages all agents in sequence
• make chaos-monkey - restarts all agents in random order
• make proxy - opens an SSH session to master0 and sets up TCP forwarding to localhost
• make tail-helper - opens an SSH session to master0 and tails the k3s-helper log
• make list-endpoints - list DNS aliases
• make destroy-cluster - destroys the entire cluster (should not be the default)
• make destroy-compute - destroys only the compute cluster (should be the default if you want to save costs)
• make destroy-storage - destroys the storage (should be avoided)

Recommended Sequence

az login
make keys
make deploy-storage
make params
make deploy-compute
make view-deployment
# Go to the Azure portal and check the deployment progress

# Clean up after we're done working for the day, to save costs (preserves storage)
make destroy-compute

# Clean up the whole thing (destroys storage as well)
make destroy-cluster

Requirements

Azure Cloud Shell (which includes all the below in bash mode) or:

• Python 3
• The Azure CLI (pip install -U -r requirements.txt will install it)
• GNU make (you can just read through the Makefile and type the commands yourself)

Internals

master0 runs a very simple HTTP server (only accessible inside the cluster) that provides tokens for new VMs to join the cluster and an endpoint for them to signal that they're leaving. That server also cleans up the node table once agents are gone.

Upon provisioning, all agents try to obtain a token and join the cluster. Upon rebooting, they signal they're leaving the cluster and re-join it again.

This is done in the simplest possible way, by using cloud-init to bootstrap a few helper scripts that are invoked upon shutdown and (re)boot. Check the YAML files for details.

Provisioning Notes

The cluster is actually split across two resource groups (-storage and -compute).

The -storage resource group contains an Azure Storage Account with an Azure Files (SMB) share that is mounted on all the nodes. This makes it trivial to deploy the cluster, work on it for a few hours, store your manifests and data on /srv, destroy the -compute resources to save costs and spin them up again against the same -storage the next day.

NVIDIA Support

Since it is possible to run machines like Standard_NV6ads_A10_v5 as spot instances, you can now try to run k3s on these with nvidia-docker2, which should be considered highly experimental since Ubuntu 22.04 is not yet supported on those machines (although older versions are).

Deployment Notes

To avoid using VM extensions (which are nice, but opaque to most people used to using cloud-init) and to ensure each fresh deployment runs the latest Docker version, VMs are provisioned using customData in their respective ARM templates.

cloud-init files and SSH keys are then packed into the JSON parameters file and submitted as a single provisioning transaction, and upon first boot of a node (master or agent) Ubuntu takes the cloud-init file and provisions the machine accordingly.

See azure-docker-swarm-cluster for more details.

Private Registry

Deploying registry/registry.yml will set up a container registry on the master node that uses the shared Azure Files storage (mounted in /srv) as backing store.

The container registry is "insecure" in that it does not require authentication nor uses HTTPS, but it is accessible only to the cluster nodes and does not require setting up any kind of certificates.

This makes it very easy to use docker on the master node to manage images, and having the registry mapped to the SMB share ensures it is persistent across cluster deployments if you don't delete the -storage resource group.

Pro Tip: You can set STORAGE_ACCOUNT_GROUP and STORAGE_ACCOUNT_NAME inside an .env file if you want to use a pre-existing storage account. As long as you use make to do everything, the value will be automatically overridden.

Disclaimers

Keep in mind that this was written for conciseness and ease of experimentation -- look to AKS for a production service.