How this site runs
Files on a CDN, and one small service that is allowed to sleep
The site argues that a system should keep itself running and say when it cannot. This page is how it tries to live up to that, for about the price of nothing.
Scroll sideways for all of it →
firebase.json, the workflows and the Terraform.
The pieces
-
The pages
Astro builds every page to a plain HTML file. No page is rendered by a server when you ask for it.
-
Hosting
Firebase Hosting serves the files from a CDN with HTTPS, clean URLs, a year of caching for fingerprinted assets and five minutes for pages. Requests under /api/ are passed to the status service.
-
The status service
One Cloud Run service, portfolio-status, in us-east4. It runs no instances when nobody is asking and at most two. Python 3.12, FastAPI and httpx, in a container that runs as an unprivileged user.
-
Infrastructure
Terraform declares the Hosting sites, the container registry, the Cloud Run service and its own service account, the keyless sign-in for GitHub Actions and a budget alert. State is kept in a Cloud Storage bucket.
-
Checks
Every pull request is type-checked, unit-tested (site and service), built, smoke-tested in a browser against the status service running on recorded responses, and held to a Lighthouse floor of 95 in all four categories.
-
Deploys
A merge to main deploys the files and the service from GitHub Actions. It signs in to Google Cloud through Workload Identity Federation, so there is no key to store or leak. Each pull request gets its own preview channel.
The status panel
The panel on the home page is the only part of the site that is not a file. After the page
has painted, the browser asks /api/status, and the service answers with one
row per system: Sightline, Orbital Watch, the AWS platform’s last live drill, and this
site.
The service asks each system’s own public endpoint, all at once. Sightline’s status endpoint names the revision that is serving and its region. Orbital Watch publishes a health file that its own page reads. GitHub’s API lists the runs of the drill workflow. Each call gets four seconds, each source six in all, and the page gives up on the service after ten.
Answers are kept in memory for a minute, and GitHub’s for five, because GitHub allows an anonymous caller 60 requests an hour. The CDN may hold the whole answer for another minute, or fifteen seconds when a check in it failed, so that recovery shows sooner. Every row carries the time it was really checked, measured on the service’s clock.
A check that fails never becomes an error page and never becomes “up”. It becomes “could not check”, with the reason. That state means the check failed, not the system: from one vantage point the two cannot be told apart, and the panel does not pretend otherwise.
| When | The panel |
|---|---|
| One source is slow or down | Its row says “could not check” with the reason and the last known answer. The other rows are unaffected. |
| GitHub’s rate limit is used up | The AWS row says so, with the time the limit resets, and the service stops asking until then. |
| The status service is cold | The panel shows “checking”, then a line explaining the wait. Pages are already on screen. |
| The status service is down | Every row says “could not check”, the panel says nothing on it is live, and offers to try again. Pages still load. |
| A source reports its own trouble | The row says “degraded” and repeats what the source said, for example how many of Orbital Watch’s feeds are stale. |
What it costs
About $0 to $2 a month. A budget alert emails me at $10, long before a mistake could become expensive.
| Piece | A month | Why |
|---|---|---|
| Firebase Hosting | $0 | No charge up to 10 GB stored and 360 MB transferred a day. This site is a few megabytes. |
| Cloud Run | $0 to about $1 | It scales to zero, and the free tier covers 2 million requests, 180,000 vCPU-seconds and 360,000 GiB-seconds a month. That allowance is shared by every project on the billing account, so if others use it up this service pays the list price of $0.000024 per vCPU-second. The CDN holds each answer for a minute, so at most about 43,200 requests a month can reach it. |
| Artifact Registry | $0 to a few cents | The first 0.5 GB is free. One image is about 230 MB uncompressed, and a cleanup policy keeps only the latest few. |
| State bucket and logs | Cents | Terraform state is a few kilobytes. The service logs one line per request. |
Prices and free allowances read from Google’s pricing pages on 18 November 2026.
Left out on purpose
- No load balancer
- A load balancer is billed by the hour whether or not anyone visits. Firebase Hosting already gives a CDN, a certificate and the /api rewrite.
- No warm instances
- Keeping one instance running would remove the cold start and would be billed around the clock. Instead the first request after a quiet spell waits a few seconds, and the panel says so while it waits.
- No server rendering
- A page that is a file cannot be down because a server is. Only the panel depends on the service, and it is built to say when the service is not there.
- No database, cookies or third-party scripts
- Nothing here needs them. The two typefaces, Newsreader and IBM Plex Mono, are open-licensed and served from this site. Page views are counted by one request to the status service that carries the page’s path and nothing else, and goes to its log. It is not sent when the browser asks not to be tracked.
- No secrets
- Every source the service reads is public. A GitHub token can be supplied to raise the rate limit, and the service works without one.
Read it yourself
The site, the status service, the hosting configuration and the workflows are in one public repository. The Terraform is kept in a separate repository that is not published.