Rebuild From Scratch

Observant readers of this blog, refreshing every day desperate for new content, will have noticed that the last blog post - dated 2022-12-31 - actually went live in the middle of January. My k3s cluster, which had always been a bit rickety, finally gave up the ghost in late December, and two of the nodes needed to be fully reimaged before I could start it back up again.

It wasn’t a total loss, though - I learnt some things along the way!

Avoid circular dependencies at cold-start

My previous setup for this cluster had a circular dependency:

• I use Cloudflare tunnels¹ to expose my Kubernetes pods to the outside world - but since adding a new domain name also requires updating Cloudflare’s DNS to point at the tunnel, I used this code as an initContainer to run those updates. The image for this code is hosted on my Gitea server (which, as well as a Source Control system, is also an Image Registry).
• Pulling the image from Gitea requires that it’s available on the external domain name - Gitea only supports interactions on a single URL, even if your internal DNS is set up to have an internal-only name as well as the externally-available one.

So, at cold-start, we have a deadlock - the tunnels won’t start up because they can’t access the initContainer image, which is unavailable because the tunnels aren’t started up.

There are a couple of ways around this:

• Set up an internal-only Gitea server which only hosts the Cloudflare initContainer script
• Rely on the fact that the DNS entry for Gitea is probably already present on a cold-start, and have the initContainers “fail open” (i.e. progress to starting the Cloudflare tunnels even if the initContainer image can’t be found²)

Both of which sounded like more hassle than they were worth. I took a bodgey-but-effective solution of extracting the actual script logic to a ConfigMap within the Kubernetes manifest. This way, there’s no Gitea dependency - the Cloudflare tunnel Pod definition contains the initContainer’s code within itself.

It ain’t pretty, but it works! If I had the dev-hours to make this properly enterprise-grade, I’d either do the internal-only Gitea instance approach, or see if there’s some magic possible with nginx to make Gitea think it’s being called on a given name when called from an external name or an internal one - but ain’t nobody got time for that.

Don’t overtax your SD cards

The fact that Raspberries Pi run on SD cards is both a strength and a weakness. A strength in that their storage is inexpensive and widely available; a weakness in that their storage is among the more fallible formats in common usage. I was running into particular problems because the /var/lib/rancher/k3s directory would rapidly fill up the cards, leading to instability. It’s possible to move that data to a different location with judicious use of symbolic links - or you can even start up the nodes with a --data-dir argument pointing elsewhere - but I wanted to go one step further and hard-limit the size this directory could grow to³, by creating a virtual drive (physically on an external hard drive, but logically mounted on the SD card) with limited space:

$ mkdir -p /mnt/EXTERNAL_DRIVE/k3s-data
$ cd /mnt/EXTERNAL_DRIVE/k3-data
# The below creates a 20Gb file, since dd uses a block size of 512 bytes
# At the time of checking (2023-02-03), k3s' data was ~6Gb, this leaves room for expansion
$ dd if=/dev/zero of=k3s.ext4 count=40960000 status=progress
$ sudo /sbin/mkfs -t ext4 -q k3s.ext4 -F
$ sudo systemctl stop k3s-agent # Or `k3s` if this is a control-plane node
$ sudo mount -o rw k3s.ext4 /mnt/EXTERNAL_DRIVE/temp_mount
$ sudo mv /var/lib/rancher/k3s /mnt/EXTERNAL_DRIVE/temp_mount
$ sudo umount /mnt/EXTERNAL_DRIVE/temp_mount
$ echo "/mnt/EXTERNAL_DRIVE/k3s-data/k3s.ext4 /var/lib/rancher ext4 defaults 0 2" >> /etc/fstab
$ sudo mount -a
$ sudo systemctl restart k3s-agent # or `k3s`

Oncall is overkill for a homelab

I’d spent ages trying to get Grafana Oncall working on my setup - a lot of the helm configuration was poorly-documented or apparently-incorrect. Nevertheless, I persisted, since I knew that it would be pointless to keep implementing features on a possibly-unstable cluster if I wasn’t able to receive alerts when it was having issues.

Until one night I was poking around and found that vanilla Grafana has the ability to send Telegram alerts directly. Given that I’m just a single person (who doesn’t need oncall rotation support), this is perfectly fine for my needs!

That said, recent changes have meant that Oncall is now easier to set up than I had previously found - RabbitMQ clusters can now be run directly from the Helm chart on an arm64 machine rather than having to install them separately, for instance! Still, though - not worth it for a single-operator system.

• CI/CD
• homelab
• k8s
• observability