Having a smaller number of big servers is generally better than having a large number of small servers, as long as you have enough fault tolerance for your use case.

In druid document, above is the recommended for historical service. Couls someone explain to me what is the main reason?

There is some amount of fixed overhead (you can think of it as roughly equating to one vCPU per node) that is carved out from the standard Druid service configurations ... so on a 4 vCPU node you would try to limit your use to 75% of the computing resources ... whereas on a 16 vCPU node you would target to use ~95% of the computing resource.

I see what you means. Thank you.