For the last month I have been working on restructuring my Kubernetes cluster to make everything a bit more organised and sensible. During the process I also scripted the setup of the host using Ansible and made some changes to the way I do things.
If you’re just interested in seeing the GitOps repository, you can check that out here: https://github.com/Jonnobrow/coffee-shop.
Motivation for change
Before I dive in to the new structure of my cluster, the things I have added and the things I have changed, I want to explain why this was necessary (or why I felt it was anyway).
Three main things prompted the changes:
- Messy Manifests: I mentioned this towards the end of my second post and although functionally everything was fine, having your manifests organised makes everything a little easier to maintain.
- Backups: Everything important is backed up so I can mess around with my cluster knowing a restore is (a) possible, and (b) quite straightforward thanks to the work I did in part 3.
- Learning and Automation: My initial cluster had been somewhat adapted from a friends cluster1, so I hadn’t really learnt all that much about what I was doing. Also, I had left a lot of manual processes creep into my maintenance and I wanted to learn how to remove as much of these as possible - automate everything!
Some new tools
A couple of new tools were integrated into my workflow during this process. Some replaced tools I was already using while others simply improved my quality of life or were necessary for the automation I was trying to accomplish.
The Github repository describes SOPS (Secrets OPerationS) as being a “simple and flexible tool for managing secrets”. It supports a wide range of file formats and can encrypt using a few different mechanisms depending on your needs.
For me it replaces Bitnami Sealed Secrets3, but also gives me a way to encrypt secrets that don’t like within my cluster, like secrets used in Ansible deployments.
Some examples of usage:
- To Encrypt a file in place:
sops -e -i /path/to/file
- To Decrypt a file in place:
sops -d -i /path/to/file
- To edit an encrypted file with $EDITOR:
FluxCD provides a guide4 on using Mozilla SOPS, as does Ansible5.
Task is a task runner … that aims to be simpler and easier to use than … GNU Make.
Instead on the Makefile format, Task uses yaml, which I am much more familiar with. It also has a more natural feel to usage (for me) and I found the ability to use templates, include other task files and define dynamic variables really useful for the kinds of things I wanted to do with it.
I predominantly use Task for running small admin tasks like:
- Enabling/Disabling Nextcloud Maintenance Mode
- Running an Ansible playbook
- Running a flux reconciliation
The documentation6 explains how to install and use Task and I recommend checking it out for a little bit more automation in all of your projects. For an example of multiple task files, see my GitOps repository.
This tool managed pre-commit hooks for Git. These hooks mean that on every commit my code or manifests are run through a linter and any other tools I want.
So far I am only using this for a few checks for my manifests such as:
- Checking for trailing whitespace and line endings
- Checking for consistent yaml using yamllint8
- Checking for unencrypted secrets9
However, in the future I can see me using this to run linters and tests before committing python code and much more.
A small configuration file is included in my GitOps repository and simply
pre-commit install-hooks will add the hooks to that repository, causing them
to run on every
git commit command, stopping the commit if a check fails.
Still Messy Manifests
My original plan was to just put the manifests in to a new directory structure, perhaps move around some services into more sensible namespaces and then be done. However, I am somewhat prone to diving deep into the rabbit hole10 and found myself looking at other GitOps Kubernetes Clusters on k8s-at-home/awesome-home-kubernetes. This quickly led to me excitedly creating a new branch on my own GitOps repository and the procrastination had officially begun.
A new cluster is born
I run a single node cluster, so by “a new cluster” what I really mean is “a new node”. In order to test the re-done manifests I decided it was best to create a separate environment because I wasn’t sure how long it would take and I didn’t want to be without some of my services for an unknown amount of time.
For ease, I decided to look into using k3s11, rather than doing everything from scratch, and here begins the first descent into the rabbit hole.
Many of the aforementioned repositories contain a directory with deployments using either Ansible, Terraform or a combination of the two. Of course, I liked the idea of having a single command I could run that would install dependencies, configuration files and setup k3s - so I made it happen. It was a chance to learn some more about Ansible12 and also make my life easier in the future should I decide to move to new hardware or rebuild my cluster again.
So after about two days work I had a new cluster up an running, but still hadn’t done anything that I originally set out to do, oh well.
Making use of Kustomize
Another thing that was common across most of the repositories linked in the awesome-home-kubernetes project was the use of Kustomize13 and the Flux Kustomize Controller14.
Kustomize is a powerful tool that can do some really cool stuff like generating secrets and config maps from the actual files, so you aren’t managing two copies of something. It can also apply patches, perform replacements and much more. When coupled with the Kustomize Controller for flux it allows me to have quite fine-grained control over my manifests and I had to have that.
These are the things I mainly use Kustomize for:
- Decrypting secrets
- Substituting cluster wide values in all my manifests
- Applying the resulting manifests
Kustomizations also mean that I can specify exactly which manifests should be applied, rather than the alternative which seemed quite flakey to me. I can reconcile a whole kustomiztion in the knowledge that my volumes, secrets and configs will also get applied whereas before only the helm release itself would be updated.
Once again, this is all very cool but at this point I’m three days in an not a lot has changed in the way of removing messy manifests.
Setting up Tasks and SOPS
Of course, things must come in threes, so here we are with a third change before the real work even begins. As mentioned in the “Some new tools” section, I wanted to use SOPS to manage secrets and Task to run tasks. I set up both of these tools when creating the Ansible deployments so that I could easily keep secret variables and run playbooks. However I needed to add a couple more tasks and generate some extra keys before I could use everything with Kubernetes.
I wrote a couple of tasks that simply ran Flux commands I use all the time, so
instead of running
flux reconcile source git flux-system I could type
task flux:sync which saves me quite a few key presses over a day of
Additionally, I followed the instructions on the flux website4 for using SOPS and generated a separate key that the cluster would use to decrypt secrets. I then added some tasks to my task files for generating those secrets again in the future, although hopefully I will never have to as that would be very bad! I also created a task to generate the secret in the cluster that flux uses to decrypt everything and that was that.
Finally getting somewhere
Now that there is new cluster, I have decided to use Kustomize and I’ve set up some new tools, I can finally start re-organising my services. I started by copying a structure from the repos I was using as inspiration.
/coffee-shop ├── cluster │ ├── apps │ ├── base │ ├── core │ └── crds ├── server # Notes live elsewhere, see above └── Taskfile.yml # Contains tasks
Let’s break that down a bit:
clustercontains all the manifests for my K8S cluster
appscontains the services I am running on the cluster and inside lives a directory for each namespace, which then holds a directory for each service. More on that later.
baseis the directory that flux uses as a source. It contains the cluster settings and secrets, the core, apps and crd kustomizations and the gotk manifests that contain everything required by Flux. I also put other sources for my services, such as helm repositories, here.
corecontains the namespaces themselves and the
cert-managermanifests. It gets applied before anything else as the namespaces and
cert-managerare crucial to the whole cluster.
crdscontains custom resource definitions used by some of the services. Although all of the services have an option to manage crds with the helm chart, by having them here I can ensure they exist and have finer control over their versions.
servercontains the Ansible roles, playbooks and inventory files
Taskfile.ymlcontains tasks for the project
- There is also a
.taskfilesdirectory with more task files that relate to specific things like nextcloud or ansible.
- There is also a
Okay, so now a nice new structure is in place, what does a typical service look like in that structure. Here is an example using Jellyfin:
/coffee-shop └── cluster # Top Level └── apps # Jellyfin is a service, so under apps └── media # Jellyfin is a "Free Software Media System" ├── _pvc # Namespace level Persistent Volume Claims └── jellyfin # Jellyfin get's its own directory ├── helm-release.yaml # The helm release itself ├── kustomization.yaml # A kustomization to say what should be included └── config-pvc.yaml # Service level Persistent Volume Claims
By grouping resources into common namespaces I avoid having separate persistent
volume claims and persistent volumes for each service. I have six services in
media namespace that all use the same volumes, but I only have one
definition for each, not six like in my old setup.
The logical next step would be moving everything over, but of course the rabbit hole re-opened and I couldn’t help but jump in.
Switching to Traefik
What’s New? (and a recap of everything else)
a complexly bizarre or difficult state or situation conceived of as a hole into which one falls or descends, especially one in which the pursuit of something leads to other questions, problems or pursuits. — Merriam-Webster ↩︎