Thrift's RPC is (optionally) built on top of HTTP, but provides support for custom typesafe (kind-of) data structures and the conveniences of RPC. If you're sending data that's more complex than a single homogenous type, then delivering a structured message over RPC is a necessary abstraction.

I don't think network overhead is the problem here, it is about ease of development. Thrift generates working server code and you just have to implement RPC functions. On PB, you need a stack to handle connections, parse messages, dispatch, etc.

Note that Thrift supports HTTP as a transport.