Some seriously cool stuff here with some interesting insights which we should take into account when working on our own revamp. Having said that, even without worrying about multi-threading (it's still an area I'm not that familiar/comfortable dealing with yet... and it'd certainly be worth having some multithread gurus around if/when we go down this path) we still have a lot of design issues to tackle to even get a basic system that tackles most of the issues that we're currently facing with our old antiquated beast.
One thing that bothers me a bit with their approach is that they resorted to having to restrict everything to C++ nodes only to gain "optimal" performance. Perhaps it's just my unadulterated hatred of C++ and some of the syntactic abominations it's been inflicting on the world since it's release, but surely we must be able to derive some kind of efficient down-compilers or even VM-type approaches which will allow the usage of nicer syntax while approaching "reasonable" performance without problems.
I also wonder a bit about how much leverage we might be able to get out of the "non-editable" assumption they state. It seems that what we do now doesn't actually provide that much "editable graph" functionality currently (i.e. everytime we do make dependency changes, we just do a full rebuild anyway). So, perhaps we will be able to do some optimisations there...
Anyways, at least it's good to know that someone out there is still working on these problems these days (AFAIK, there's fairly scant info published out there about the intricacies of building these sorts of systems). Seeing a modern approach to this problem is certainly quite enlightening, and inspiring too.