Project: Demo Garden

The idea started from wanting a single, owned place to host technical experiments — without scattering them across Vercel, Fly.io, Heroku, and GCP trial accounts that eventually expire. The goal was a system that was stateless enough to rebuild quickly if something went wrong, and cheap enough to run indefinitely.

The result runs on an AWS EC2-micro instance (1GB RAM, slow CPU) and hosts dozens of demos. When no traffic arrives, nginx and the lazy loader together use around 10MB of RAM. When a request comes in for a sleeping container, the lazy loader wakes it, proxies the request, then watches for idle time before putting it back to sleep — closely mirroring how platforms like Cloud Run and Heroku manage scale-to-zero internally.

The lazy loader was initially written in Python (~50MB memory footprint) and later rewritten in Go to get it down to near-zero overhead. It was a good excuse to learn Go properly.

CI/CD runs on GitHub Actions: images are pre-built on the runner (not on the tiny VM), cached aggressively, then copied to the server and started via SSH. The full deploy cycle — push to live — takes under a minute.

Each demo gets a page with a description sidebar and an embedded live preview. The project is open source: github.com/pnettto/demo-garden.