[tomgallard]

But a DNS based failover is still going to take an hour or so to propagate right (given that a lot of browsers/proxies/DNS servers don't respect TTL very well at all)? And then you end up with a system with stale data, and the mess of trying to reconcile it when your other system comes back up.

I'd take an hour long Appengine outage once a year over that anytime!

[peterwwillis]

Your name server or stub resolver is what respects DNS TTL, not your browser or proxy. Everyone - including people hosting on AWS - needs to be able to fail over DNS, if the AWS IP you're using is in a zone that just went down, for example.

Any time you have an outage you need to contact your service provider to get an estimate of downtime. If they can't give you one, assume it'll take forever and cut the DNS over. The worst case is some of your users will start to come back online slowly. If you don't cut over, the worst case is all your users are down until whenever the service provider fixes it, and you get to tell your users "we're waiting for someone else to deal with it", which won't make them very happy.

12 hour stale data sounds kind of long to me. 4 hours sounds more reasonable.

[codeka]

I've seen plenty of crappy ISP DNS servers ignore TTL values and cache DNS entries for many hours longer than they're supposed to. Unfortunately, it's all too common.