Pointers in Far Memory: A rethink of how data & computations should be organized

invariant pointers in a distributed global address space? how very database-y.

"pointer" implies ephemeral, to me; if you're in a cluster or something then you extend the pointer to include a "arena" component; be that process id or a cluster:procid pair or whatever. They're still invalid once the process is gone but then you're not chasing indexes and indirection tables quite so far for every single access.