From the manpage: "[rotate] causes round robin selection of nameservers from among those listed". This means in your configuration, the broken nameserver will be hit once every 3 times, resulting in a 1second timeout before trying the other 2. So in every 9 requests your setup results in 3 seconds delay for as long as the nameserver is down.

In case of Amazon, you actually prefer their nameserver as it resolves names to LAN IPs where possible. With your strategy, ~66% of my instance hostnames would be resolved to the less efficient WAN IPs, that's with all 3 nameservers being reachable.

As for data, there's papertrail and Premium Support tickets with Amazon admitting to their downtimes. In most cases (e.g. the one on Jan 14, 2013 11:27 PM PST) "there was a problem with the underlying host".

Remember it doesn't even need to be the service itself going down, there's weaker links in this chain like misconfiguration of the virtual host. Cables can get tripped over. Anycast is not possible here.

I'm not blaming them. This article was about mitigating suckyness in case of outages, not proving that such a thing even exists : )