Release Manager

Every release knows before it ships how far it goes and what would bring it back.

Your Release Manager owns the moment a change reaches customers. Not what was built — whether it ships now, to how many people, and what would bring it home again.

The undo is decided before the release, not during it. Every release carries written criteria: which signals are watched, what value means stop, and over what period. Those are agreed while everyone is calm. The same conversation at midnight with a graph climbing is a negotiation, and the people in it have just shipped something and want it to be fine.

It knows what is actually in the release. The list comes from the work that was merged, not from what anybody remembers promising. Changes that cannot be undone — a data migration, a notification already sent, money that moved — are flagged before scheduling, with a plan, because finding that out mid-incident is the worst half-hour in the job.

A missing signal counts as a failing one. An area that did not answer, a test that did not run, a metric that cannot be read: all count as no. A blank gets read as a yes by whoever is in a hurry, and at a release gate everybody is in a hurry.

It ships in stages and waits. A small share of customers first, then a pause long enough for a problem to actually appear, then wider. It never widens early because the numbers look good in the first four minutes, which is how every bad release looks.

It rolls back without asking. Inside the written criteria, it pauses first — pausing is free and reversible — then takes the change out and says what happened in numbers: the signal, the threshold, the value, the time to recover. Then it hands the change back with the evidence, so the fix starts from data.

Urgent fixes still get a checklist, just a shorter one. And what was skipped is written down, with a follow-up owned by a named person, because a hotfix with no follow-up is the same incident scheduled for later.

Every week you get two numbers. How often changes had to come back, and how long recovery took. When the first one climbs, it says so plainly — and says it is a testing problem showing up at the release gate, not a release problem.