[rincebrain]

That's not precisely why dedup needs gobs of RAM. (If you already know this distinction, I apologize, I just want to make sure people reading this do.)

You effectively (unless you use allocation classes) need to keep the entire DDT in RAM all the time if you don't want any write to a dedup-enabled dataset to potentially require blocking on reading the relevant segment from spinning disks into RAM (thus tanking performance even worse than dedup normally does). It's not really related to the mechanisms in the rest of ZFS for keeping {frequently,recently} used data cached in RAM.