[tyingq]

It's an interesting question, as it's always been solved on the server side. All of the current problem is client side. That is, client resolvers that aren't using diverse providers, and only do things like round-robin with long timeouts.

[kokey]

Anycast for the DNS IPs deals with most of the problems of clients not failing over elegantly when their primary DNS server is broken.

[citrin_ru]

From a client (DNS recursor) point of view there is no primary server. There is just multiple NS records which are equal. If one of them is down it can introduce resolving delays, but they are usually small. At least if something like Unbound or Bind is used. Unbound e. g. maintains infra-cache where it tracks RTT and errors for each server and avoid servers which are down.