BYTEHERDERFourteen months behind, on purpose
A cluster sat two minor versions back because nobody had measured the upgrade risk. Unmeasured risk gets treated as infinite, and infinite risk never gets scheduled.
Written from the work.
Notes from the work: what broke, what the fix actually was, and which belief did not survive contact with production.
Provenance is the product
Two dashboards, one metric, two numbers. Working out which one was right took longer than building either of them, and the answer was that neither could be checked.
The plus in the secret
Keycloak returned unauthorized_client for a client secret that was byte-identical on both sides. The character causing it was a plus sign.
A gate that cannot fail is a report
Every pipeline I inherit has a security scanning stage. Almost none of them can stop a release. That is not a tooling problem.
Publish the interface
A platform team without a written-down interface is a ticket queue with a roadmap attached. Here is what writing it down actually looks like.