Any benchmarks compared with Thrift and Protobuf? BTW, Hadoop's new Avro project is yet another serialization project that's designed to be dynamic language friendly.

I personally find thrift IDL trivial to manage and is probably the only sane way to support a dozen languages (statically and dynamically typed) in a type safe and efficient way.

BERT-RPC is simple, but what if you find your server too slow and needs to be rewritten in another language? I think the Thrift approach makes that simpler.