How do D templates compare to staging in MetaOCaml? I had always thought D templates only expanded during compile time, like C++'s.

MetaOCaml is about run-time meta-programming. There's your key difference.

Is Scala not practical?

Julia's generated functions are also capable of similar designs.

Both Scala and OCaml are very practical languages, I was referring to the actual multi-stage streams implemented in the proposed library. What I think is interesting is the fact that D arrived to a similar conceptual model from a more _practical_ perspective by extending the applications of templates.

I guess the question is rather whether MetaOCaml or the Scala-Virtualized compiler fork + the LMS framework code are practical.

I am not so sure, but I have also not looked into it too deeply. But in the end it doesn't matter as long as it is just a maturity issue. After all I guess the stream library in the submission is to be seen as a motivation for making MPS more mainstream.