[LinuxBender]

If using the public DNS servers to resolve things then certainly have a few of them. 1.1.1.1 and 8.8.4.4 in case CF has problems.

[Edit] Testing 1.1.1.1 it seems I have finally built cache on all their nodes. It took about 25 to 30 requests. Second test took only a few requests for a different record, same domain. Now I am more curious about their back-end.

If you have your own recursive servers in your datacenter then you can entirely control how many things are cached and how that cache behaves. Unbound is a really good option for this as it is fast, has controls around memory, threads, min/max TTL and you could even push your authoritative zones to the edge Unbound nodes if desired so that those records never expire. Some people take this a step further in their datacenter and have Unbound running on every instance to keep response latency low and handle upstream recursive fail-over better than the OS resolver does.

Another benefit to running your own caching servers is that you can purge records that you know are out of date during outages.