[ndriscoll]

Open a TCP connection from the instance to the cloud service. I don't know about all consumer routers, but I just checked mine and the default TCP established timeout is 7440 seconds. Idle timeouts are supposed to be at least 2 hours.

If you served the entire US (130 million households) and had a 1 hour keepalive, that's only 36k packets per second, which is nothing.

You could also auto-train the idle timeout by using a pair of TCP connections. One uses a known good value while the other probes upwards until it finds its connections start getting closed (with some optional binary search fanciness), feeding new known good values back to the first.

(Obviously the no-cloud solution is better still)