Infrastructure

Eight Services, Shipped by One Engineer, Without a Platform Team

Section 5 of 5

Eight independently deployable services is a lot of surface for one engineer to keep shippable. The only way that stays manageable is if every service is deployed the same way, so that knowing how to ship one means knowing how to ship all of them. Every repository in this product carries the same three things: a Dockerfile, a build configuration for development, and a build configuration for production.

Everything runs on Cloud Run in australia-southeast1. That is a deliberate choice rather than a default: the users are Australian, the data is about Australian families and their children, and keeping both the compute and the data in-region is the right answer for latency and for data residency at the same time.

8

independently deployable services

2

build pipelines per service — dev and prod

1

region, chosen for residency and latency

0

credentials baked into any image

Delivery Pipeline

The same shape for every service. Two build configurations separate the environments, but they produce the same kind of artifact and deploy it the same way.

Service RepositoryDockerfile · cloudbuild.yaml · prodbuild.yamlDevelopment Buildfast iteration, dev configurationProduction Buildrelease artifact, production configurationMulti-Stage Docker Buildbuild toolchain discarded — only the runtime artifact shipsArtifact Registryimmutable, versioned imagesCloud Run Revision — australia-southeast1configuration and secrets injected as environment variables at start-up

Operational Decisions

Every service builds in stages. Dependencies are installed and the TypeScript or frontend bundle is compiled in a build stage; the runtime stage then copies only the compiled output and the production dependencies. The compiler, the development dependencies and the original source never enter the final image.

The benefit is threefold and all three matter on Cloud Run. Smaller images pull faster, which directly reduces cold-start time on a service that scales to zero. Less software in the image means less attack surface and fewer vulnerabilities to triage. And the separation makes it structurally impossible for a build-time-only tool to be accidentally depended on at runtime — a class of bug that otherwise only appears in production.