I think this is similar how the D implementation of Thrift works.
Yes, exactly! Except instead of writing code to send/receive messages, I'm thinking that indices are built that are specific to the data. That I think is the harder part as you have to know what is optimal for the data, what operations are expected on the data, and putting it all together. However, I have to wonder if by making the indices very specific query performance is improved by a significant amount.
