I looked at the Thin code again and it uses Open3 to spawn multiple instances of itself as subprocesses. Intuitively I'd expect that method to consume more Ram

I looked at the Thin code again and it uses Open3 to spawn multiple instances of itself as subprocesses. Intuitively I'd expect that method to consume more Ram than forking because there's no opportunity for COW at all. However it'd be interesting to see what you measured, looking forward to your write-up if you get around to it.