docs/src/reference-dsl-complexity.md
One of Miller's strengths is its brevity: it's much quicker -- and less error-prone -- to type mlr stats1 -a sum -f x,y -g a,b than having to track summation variables as in awk, or using Miller's out-of-stream variables. And the more language features Miller's put-DSL has (for-loops, if-statements, nested control structures, user-defined functions, etc.), then the less powerful it begins to seem: because of the other programming-language features it doesn't have (classes, exceptions, and so on).
When I was initially prototyping Miller in 2015, the primary decision I had was whether to hand-code in a low-level language like C or Rust or Go, with my hand-rolled DSL, or whether to use a higher-level language (like Python or Lua or Nim) and let the put statements be handled by the implementation language's own eval: the implementation language would take the place of a DSL. Multiple performance experiments showed me I could get better throughput using the former, by a wide margin. So Miller is Go under the hood with a hand-rolled DSL.
I want to continue focusing on what Miller excels at — concise notation, low latency, and high throughput — and not add too many high-level language features to the DSL. That said, some customizability is a basic thing to want. As of 4.1.0, we have recursive for/while/if structures on about the same complexity level as awk; as of 5.0.0, we have user-defined functions and map-valued variables, again on about the same complexity level as awk along with optional type-declaration syntax; as of Miller 6, we have full support for arrays. While I'm excited by these powerful language features, I hope to keep new features focused on Miller's sweet spot, which is speed plus simplicity.