True. We did a fair bit of experimenting with this (for both load balancing and failover) a few years back, and so long as you stay at 300sec or greater it works fine for well over 99% of all the traffic we tested with. Once you dropped below 300 seconds problems started appearing - from memory some older versions of Windows would default to 900secs, 3600secs, or 86400secs.

I haven't revisited that research for 3 or 4 years, but our findings then led to a policy of "if you're prepared to accept ~15mins of mixed availability (after you've identified a problem and hit the panic button) then DNS based failover works well enough. If you need significantly better response times than that, you need proper hardware/network based failover"

Yup.

The day I got a bunch of Netflix traffic sent to servers I was running thanks to ELB and caching resolvers not honoring TTL values was great fun.

(That was also the day we learnt that our 500 page was actually really expensive to render, which added to the excitement...)