The merge queue runs itself

Every change to the platform is tested before it lands. Until now, two changes that were each fine on their own could still break things when they landed together, because nothing had ever built the combination. A merge queue fixes that: changes wait in line and each one is built on top of the ones ahead of it, so what lands is exactly what was tested.

The queue's first outing, at the end of August, needed constant attention. A test that fails only occasionally would knock a change out of the line, and someone had to notice and put it back. And with several changes being built at once, every arrival or departure restarted all the builds — for over an hour nothing landed and nothing failed either.

Both are addressed. The queue now builds one change at a time, so there is nothing to restart. And a steward watches every change that leaves the queue without landing: if the build timed out, or failed only on a test that is already known to be unreliable, or died on a machine problem before it could reach a verdict, the steward puts the change back in line by itself — at most twice, and it says what it did on the change. If the failure is real, it leaves the change out, names the failing test and the build, and marks it so a person can pick it up.

For anyone contributing: mark a change ready and it lands when it is green. The queue and its steward are documented under Architecture → The Merge Queue.