They were mistaken or maybe referring to an older version, because I definitively use rolling zero downtime updates from commits
It uses docker stop once the new container is healthy with a 30 second timeout, which I believe lets existing connections drain out.