Looks like the metadata service uses HTTP long polling (“ This request also includes a timeout (timeout_sec=<TIME>)”), so I think they had a <TIME> second head-start in racing against the real metadata response.