Yornik Heyl

Rotterdam

Cloud reliability engineer at Siemens since 2021. I handle incidents, design safe rollouts and build monitoring that still makes sense at 3 in the morning.

How I work

I build platforms that are safe to change, not just safe to leave alone. Declarative by default: reproducible rollouts, clear ownership, and security built into the platform rather than bolted on. Systems should be explainable under pressure, not just work when everything is fine.

AI tooling where it pays, automating the boring parts and catching mistakes early, not as a buzzword.

The homelab

At home, four small Intel N150 machines run a Kubernetes cluster with 59 applications on it. It hosts my own GitLab, my password manager, my media server, a Fediverse instance and a site where people learn Go in the browser.

It also runs the house. Every quarter of an hour it reads the electricity price and decides whether to heat the water boiler. That job used to be a cron script on a Raspberry Pi. It keeps a dashboard for my car too. That one can only read: the build fails if the code could ever send the car a command.

All of it is declarative and every change is a pull request. YAML lint, helm template, kubeconform, per‑chart values rendering and a KSOPS hash guard must pass in CI before merge.

Tools the homelab runs on and how each is set up
Tool Setup
Talos Linux A cluster of four nodes: three control-plane and one worker, on Intel N150.
Argo CD App-of-apps with 59 applications and secrets in SOPS with KSOPS.
Prometheus, Grafana and Loki Alloy ships the logs, retention is 60 days, and the dashboards are my own.
Traefik TLS 1.3 only, with ECDSA P-384 certificates from Let's Encrypt over ACME and HSTS preload.
external-dns Works with Cloudflare, plus cron jobs for DANE with TLSA and for CAA, with SOPS-encrypted credentials.
OpenTofu Hetzner edge nodes, HAProxy ingress to Tailscale, and infrastructure as code end to end.

What went wrong

Things break here too. When they do, I write down what happened, why, how I fixed it and what stops it next time. There are 24 of these in git so far. A few favourites:

The cache that did nothing
Loki reserved almost 10 GB of memory for a cache that answered 37 requests out of 12,000. I measured it, then turned it off.
The alert that gave up
After a storage update I wrote an alert for volumes that go read-only. It cleared itself two hours later while the volumes were still broken, because it watched for the change and not the state.
Eight quiet months
The update bot skipped one image for eight months and nothing complained. I found out by wondering where a release had gone.
The update that broke nothing
A routine minor bump stopped all DNS automation for days. Every site stayed up, so nobody noticed until a new hostname never appeared.
One second
I deleted a music library, the media server got busy, and its health check gave it one second to answer. Kubernetes killed it for being slow.

How it stays up

Backups
Four HA Postgres clusters, with WAL and a base backup every 3 hours, 14-day point-in-time recovery and a restore runbook.
Dependencies
Renovate automerges patch and minor image bumps. Chart and major updates wait for a human.
Admin access
Admin UIs are reachable over the tailnet only, apart from one public webhook path.
Email
SPF, DKIM, DMARC and DANE with TLSA.
IPv6
End to end.
Audit
The Dutch NCSC rates every category “good”, and the site scores 100% on Internet.nl.

Get in touch

I'm always happy to network with people in platform and reliability work. I work in English and Dutch.

If you are human, add your favourite song to your email.