By carefully reading the code in `job_controller.go`, I finally understood that
the back-off count is reset when `forget` is true, which happens when `active`
or `successful` changes without any new failures right at the moment. That
happens in this code:
dd649bb7ef/pkg/controller/job/job_controller.go (L588)
That behavior does not match what this document says. My change fixes the doc
to match the code.
It might be better to fix the behavior to match the doc, since the behavior is
kind of weird to describe. But I imagine that the Kubernetes team will need to
consider factors I'm not aware of before deciding to change job back-off
behavior, so I am not going to the effort of proposing a change like that.
7b92c46503