They wrote a blog post about it but it's been a few years since I read it. IIRC it centered on the inherently stateless request model and how that interacts with the way they wanted to do concurrency.