Walk me through a postmortem you wrote after an incident you caused.

The incident was mine. A cooldown value I tuned to stop node flapping ended up applying to a pool class it was never meant to touch, and scale-up on that class stalled long enough that the shared queue backed up through a morning peak. I wrote the write-up the next day and published it in the repository, which is the norm on that project. It opens with the timeline and one plain line saying the change was mine, and then deliberately says nothing more about that, because a document that reads as an apology stops other people volunteering facts. The cause was a label selector dropped during a rebase, so the value matched every pool instead of one. I listed the conditions separately and did not blend them in: nobody who reviewed the patch owned that file, the documented default in our reference page was stale, and we alerted on failed jobs rather than on autoscale delay itself, which is why a downstream project told us before our own dashboards did. We are eleven maintainers across a lot of time zones, so an unowned list would have died. Seven actions, each with a named maintainer and a target release. Five shipped within eighteen days, including the delay alert and a check that fails when a selector matches more pools than the change claims to touch. Two I cut back in the open rather than let them sit. The next selector mistake was caught in review, by that check.

My own incident is the one that changed how the project handles all of them. I had shipped a scheduler change that pushed scale-up past four minutes for most of a working day, and while writing it up I noticed I was inventing the format from scratch, exactly as the three authors before me had. So the document had two halves. The first was ordinary: my change, the mechanism, the conditions, the fixes. The second argued that our review process could not have caught this class of bug for anybody, not just for me, because no file in the scaling path had an owner of record and we had no shared definition of what degraded even meant. Out of that I proposed three things to the maintainer group. A template that forces contributing factors to be listed apart from the cause. A published expectation for queue wait, so an incident gets declared by a number rather than by whoever is most annoyed. And a rotating reviewer for write-ups, so the author is never their own only editor. All three were adopted, and I edited the first four reviews myself so the norm was not just a file in the repository. The number I care about is not from my incident. It is that write-ups now close 23 of every 28 actions inside a release cycle, where most used to stay open, and new contributors reach for the template without being told.