For building DSLs the following features are useful: Infix/postfix functions, Operator overloading, Varargs/spread operator, Scope controling what is allowed. Other features??
Also you can really, really change both syntax and semantics: https://docs.racket-lang.org/2d/
Although not similar technically speaking, Graal/Truffle is a better comparison in spirit, IMO. But Graal/Truffle is hugely complex, with many many man-hours in it.
Classic example and discussion: https://www.reddit.com/r/programming/comments/i1slm/comment/... https://www.hashcollision.org/brainfudge/index.html
Also you can really, really change both syntax and semantics: https://docs.racket-lang.org/2d/ Exactly, racket give you much more freedom because of this, I don't know the use cases but it's indeed more powerful. A comparison with Truffle would be great! I wonder if a higher level/simpler but less powerful abstraction on top of Truffle could bridge the difficulty gap.
This proposition is often criticized by software engineers for resulting in code that is potentially inscrutable to anyone but the writer. IMO, this misses the forest for the tree. Sure, one-off DSLs probably aren't suited to entreprise codebases you'd use Java for. But what the critics miss is that this truly enables domain experts. In that way, it's a bit like the MS Excel value proposition cast into a (very well-designed) programming language. Which makes it ideal for solo devs, scientists, small teams, etc. (as most lisps). That's what makes Racket interesting to me, at least.