A public status page has two ways to go wrong. It can leak: names, addresses or links into the private network it describes. Or it can flatter: green gauges and confident targets that nothing measured. The first is a security problem and the second is an honesty problem, and a portfolio page is tempted by both.
Part 6 covered the leak side: the page on the DMZ observer is fed one sanitised, numbers-only file through the one grant that machine has, and nothing flows back. This post is about the other side. Everything on the page that makes a claim about the platform now comes from evidence in the repo or a live measurement, and is built so that a better-looking number would need a better result.
Where it started #
The first version had gauges labelled as live telemetry that were hard-coded, and links to services on the tailnet that no visitor could open. The links also told anyone reading the page source which services existed and what they were called.
All of it came out first (ADR-020). The rules now open the page’s config: no tailnet names or links, no LAN or cluster addresses, and admin tools are named but never linked. Then the page was rebuilt from things it can back up.
Live numbers, labelled with their age #
Node, app, pod and alert counts, capacity gauges and four six-hour charts come from the feed covered in part 6. The feed reports its own age, and the page says either “Live · updated N ago” or “Data delayed since…”. A JSON-query monitor on the observer’s Uptime Kuma alerts me when the feed stops being fresh. A stale number is labelled stale, not shown as current.
The banner at the top works the same way. It says “all systems normal” only when every check behind it passes, and otherwise names each problem: a delayed feed, a node, an app, pods, alerts, scrape targets, or a backup or alerting check.
Recovery targets next to their proof #
Each component’s recovery target sits in a table with two more columns: what a drill has actually proven, and the age of the newest backup, read live. Both come from one file in the repo, objectives.yml, introduced in part 18:
- name: Edge observer (this page)
rto: 30 min
rpo: 24 h
proven: "Full rebuild timed 19 Nov 2026: 48 min from deletion to every monitor green; 16 min of it is the first deploy on the small VM, the rest three breaks it found, all fixed since (INC-2026-016)"
backup: "systemd:kuma-backup.service"So the page says, about itself, that its last measured rebuild took 48 minutes against a 30-minute target. That row is part 19 in one line. The infrastructure-state row has an empty “proven” field, because nothing has been drilled there yet. An empty cell is the honest answer, and the file’s own comment says to leave it empty.
The “newest backup” column is read on the observer from Uptime Kuma’s heartbeats or a systemd unit’s last successful run, so it can’t show a backup that didn’t report.
Incident response, computed from the reviews #
Every incident review in the repo starts with a table that now has four required fields: Started, Detected, Resolved and Cause. At every deploy of the observer, a script reads all of them and publishes the results:
st, de, rs = when(field("Started")), when(field("Detected")), when(field("Resolved"))
if not de or not rs:
sys.exit(f"{r['id']}: Detected and Resolved fields are required")
how = found.get(r["id"], "").strip()
per.append({"found": "alert" if how.lower().startswith("alert") else "other",
"detect_h": span_hours(st, de), "restore_h": span_hours(de, rs),
"change": field("Cause").strip().lower().startswith("change")})Here’s what the page shows today, across 17 reviews:
| Measure | Value |
|---|---|
| Median time to detect, when an alert caught it | 3 min (5 of 17) |
| Found some other way | latent for a median 3 days before an audit or check found it |
| Median time to restore, from detection to a verified fix | 9 h |
| Caused by a change, including setup | 12 of 17 |
None of those numbers is flattering, and that’s the point. Five of 17 found by an alert means most problems were found by me looking, days later. Twelve of 17 caused by my own changes says where the risk is.
It also works as an incentive. A review with a vague or wrong timestamp becomes a wrong public number, so the fields have to be right. Two older reviews had their start time as “unknown”, and their real start times were recovered from the git history. A review missing Detected or Resolved fails the deploy.
Reference: DORA metrics in my knowledge base covers the stability metrics, time to restore and change failure rate, that these numbers echo.
Visits, without a tracking script #
The page shows “N in the last 30 days · N since November · counted without cookies”. It has no analytics script. The count comes from Cloudflare Web Analytics at the edge, which the feed reads once an hour with a token that can only read analytics. It keeps one number per day in a small log on the cloud host, so the total outlives Cloudflare’s retention. Only the status host is counted, and headless browsers (my own test runs) are left out. Re-reading the last three days and storing by date means nothing is counted twice.
It added no new request path into the observer and no new CSP hash. It’s a number in the same sanitised file as everything else.
Three themes under a strict CSP #
The page sends a Content Security Policy that allows scripts only from the site itself, plus one hash for each inline script (runbook). Glance renders a short inline script on every page, which differs only by the page and the theme. Adding a theme picker (Auto, Dark and Light) meant 15 hashes: five pages times three themes.
Glance can’t set response headers, so a Cloudflare rule sets them, and the values come from a file in the repo. The rule is pasted by hand, which is how dashboards drift. So every deploy fetches every page under each theme’s cookie, hashes the inline scripts, and compares the live headers with the file. A Glance upgrade that changes a script fails the deploy and prints the new hashes:
csp_mode: enforce # 19 Nov 2026: 5 pages x 3 themes, zero CSP reports in headless ChromiumIt ran in report-only mode first and moved to enforce only when all 15 combinations were clean in a headless browser.
The light theme itself is built from light-dark() colour tokens, so one stylesheet serves both, and every text element was checked against WCAG AA contrast in each theme. One deploy detail caught me out: Cloudflare caches the stylesheet for four hours, keyed on a version parameter that Glance changes only when it restarts. A CSS-only change was invisible until the cache expired. The deploy now restarts Glance whenever any of its files change.
What I learned #
- A status page that shows numbers it can’t vouch for is worse than none. Label every live number with its age, and say “delayed” when it is. (ADR-020)
- Publish targets next to measurements, and leave the measurement empty when there isn’t one. (
objectives.yml) - Compute the track record from the source documents. If the public number comes from the review, the review has to be accurate.
- Check hand-set config on every deploy. The CSP lives in a dashboard, so the deploy compares the dashboard with the repo. (runbook)
- A cached asset keyed on a version that doesn’t change is a stale asset. Restart, or change the key, when the file changes.
Architecture takeaways #
Expose a purpose-built view, never the system #
Kubernetes never asks you to put the API server on the internet. You publish one narrow endpoint, a Gateway route to one service, and the control plane stays private. The status page follows the same rule: one container, fed one file, built for the public. It can show the cluster’s health without any path to the cluster.
Make the honest number the easy number #
A dashboard someone edits by hand drifts toward what they’d like it to say. Every claim on this page is derived: health from Prometheus, recovery proof from a file next to the runbooks, response times from the reviews, visits from the edge. Improving a number means improving the thing it measures, which is the only way a public page stays worth reading.
Previous: A Fix That Moved the Problem · All posts in the series