Multi-App Hosting Platform

Overview

Rather than one AWS resource per demo app, both Timeline and Paperwork run as separate Docker containers on a single $12/mo Lightsail instance, behind Caddy doing path-based routing (/timeline, /paperwork) with fully automatic Let's Encrypt HTTPS. Each app has its own independent GitHub Actions deploy pipeline, triggered on push to main.

Architecture

• Docker Compose as the single source of truth for the stack: Caddy, Timeline, Paperwork, and a containerized PostgreSQL 17 for Paperwork (Timeline uses SQLite on its own volume).
• Caddy handling TLS certificate issuance/renewal and reverse-proxying both apps by path, so neither app needs to know it's sharing a domain.
• Per-app GitHub Actions pipelines that package the source, scp it to the instance, and rebuild just that one container — serialized across both repos with a shared flock lock file, so a push to either repo can't build concurrently with the other and starve the instance's 2GB of RAM (this exact thing happened once and needed a reboot before the lock was added).

Real production issues, root-caused

• A DNS change that "propagated everywhere" but never actually worked. The domain had two separate Route53 hosted zones — the one being edited, and a second, older one from the Lightsail account's own DNS feature that was silently the one actually delegated at the registrar. Every public DNS checker showed the new IP had spread globally; it hadn't, because the real authoritative zone was never touched. Found by comparing the registrar's actual delegated nameservers against both zones directly, not by waiting longer.
• A Prisma 6 packaging bug under Docker. Docker's cross-stage COPY flattens node_modules/.bin/prisma from a symlink into a plain file copy, which breaks its relative lookup of a sibling WASM file the newer Prisma schema engine needs — a silent MODULE_NOT_FOUND at container start. Fixed by invoking Prisma's real entry point directly instead of the symlinked CLI shim.
• 37GB of stale Docker build cache quietly consuming 80% of the instance's disk — traced with docker system df's build-cache accounting (separate from image/volume size, easy to miss) and reclaimed with no impact to the running containers.

Cost cleanup (same AWS account)

Beyond the hosting itself: a systematic pass through the account found and removed an orphaned Lightsail instance with a stopped-but-still-billing EC2 counterpart, unattached static IPs, two unused VPCs, and a dozen multi-year-old Lightsail snapshots — plus tuning an Aurora Serverless v2 cluster's min capacity to 0 so it scales to zero when idle instead of running (and billing) continuously.

Stack