Release record · 03 of 03 · Honest notes
Built on assumptions, and how I would test them
With no users, I wrote the release record for the people I imagine using it. These are the guesses it rests on. Each one is built, and none of them has been tested against anyone but me.
Built on assumptions, and how I would test them
Three columns: what I assumed, what I built on it, and how I would find out if I was right. Notes are written late, so they are thin or skipped: I built a draft that exists from day one, pre-filled from merged PRs; I would check how many releases have notes before Promote, and how often the pre-fill is kept as is. Approval after the fact is only an acknowledgement: I moved sign-off onto the release PR, Hand-flown mode, a soft gate; I would check whether it is a rubber stamp by the time from approval to merge. People say a name, not an image tag like 1.8.0-4f9c2ab: I built tree names such as DOGWOOD, one per release and never reused; I would listen for them in standups and pull requests. Evidence goes missing because pipeline runs are pruned and tracking state is swept: I freeze the gates, the commit proof and the test evidence into the record when it is sealed; I would run one dry-run audit with a real auditor. Seven days means clean, a bake window I picked: the outcome is written after the bake, as clean, rolled back or stopped; I would compare outcomes with what on-call actually saw. Autoland at an average readiness of 80 and under 15% rolled back, thresholds I made up: Tower shows eligibility, display only, and Autoland itself is not built; I would watch which release mode teams choose, and keep. None of these tests has been run.
WHAT I ASSUMED WHAT I BUILT HOW I WOULD FIND OUT Notes are written late so they are thin or skipped A draft from day one pre-filled from merged PRs Notes before Promote? and how often the pre-fill is kept Approval after the fact is only an acknowledgement Sign-off on the PR Hand-flown · a soft gate Is it a rubber stamp? time from approval to merge People say a name not 1.8.0-4f9c2ab Tree names: DOGWOOD one per release, never reused Do the names stick? listen in standups and PRs Evidence goes missing runs pruned, state swept Frozen at the seal gates · commit proof · tests Would an audit pass? one dry run with a real auditor Seven days means clean a bake window I picked Outcome after the bake clean · rolled back · stopped Does it match incidents? compare with what on-call saw Autoland at 80 and 15% thresholds I made up Eligibility, display only Autoland itself not built Do teams trust it? which mode they choose, and keep Every feature here is a hypothesis with one customer: me. LEGEND An assumption Built on it Not built Not yet tested The test I would run
Each row starts as something I believe about how teams release software, from experience and from reading, not from talking to Hangar's users. It has none.
Everything in the middle column is in Tower today, except Autoland itself. Building first was cheap in a lab, and it gives me something real to put in front of people.
The right column is the discovery work a real team would do first. Here it comes last, and I would expect at least one row to be wrong.