It came up in a discussion around server selection logging that at least some drivers (Java, .NET) only perform server selection once for a bulk write, and that at least the Rust driver re-performs server selection for each batch. This doesn't seem to be covered in any spec right now; it would be good to formalize what the behavior should be in the new bulk spec.

Reusing the server does save the driver some work; but on the flipside it makes it less likely the driver will send a command to a server that is no longer available and end up having to retry.