The contract here is that we think about programs in a declarative, mathematical way, and the DSL (e.g. GraphLab) takes care of the rest. In some sense, we would think in logic and math (e.g. latex), not code.

DSLs invent language so that we may do what we do best as humans.

I am thinking about it as "Abstractions with gears." Imagining compute blended into the edges of these abstractions (and also within them).

Have graph libraries like these ever been used to optimize programs in other domains (like how cuda was used for more than just graphics back in the day)?