You can get a disaggregated version of what ZFS does - you can have filesystem -> LVM -> MD -> devices, where the MD layer does the checksumming at a sequence-of-bits level. It tends to work less well than ZFS though; snapshots are a lot more efficient if the low level knows which bits are which files at the high level, scrubbing can be done more efficiently if you know which files are in use and which aren't, if the parity checking wants to use idle bandwidth then it needs to know about user-requested access. Perhaps most importantly, data writes need to be async for performance, but filesystem metadata needs to be written in a way that will maintain the filesystem's invariants, so the question of when bits hit the physical disk ends up being intimately entangled with the filesystem's internals. It's really difficult to express an interface between the filesystem layer and the parity layer that lets it do the right thing.