A status page that cannot report a problem is a decoration. Ours was four cards with hardcoded values — a 14ms API node, three idle build workers — that were never measured against anything.
Checking the actual services
Each row now performs a real operation against the service it names: describe the table, head the bucket, fetch the identity provider's public keys, ask the build service for the project, ask the mail service whether sending is enabled. The latency shown is the round trip that check took.
A name it cannot guess
Resource names come from configuration with no fallback. An earlier version derived a plausible-looking name from the product name when configuration was missing, then checked a resource that had never existed and reported the build pipeline as down for days.
Telling someone
The checks now raise an operator alert when a dependency is not healthy, de-duplicated per hour so an extended outage does not become a stream of identical mail. The page had been correctly reporting a failure for days, to nobody.