It's a very cleanly layered system, it just doesn't bother end user with details (as implementor, you can play with them, thus LustreZFS): there's separate SPA (block), DMU (OSD) and ZPL (FS) & ZVOL (emulated block device) layers.
Compression and encryption are integrated at DMU level because that's a logical place for them.
NFS actually calls OS nfs server.
Interesting; it'd been claimed to me before that ZFS had its own NFS server (or I guess the OpenSolaris NFS server) included but that nobody used it because it was old/buggy. A quick glance at https://github.com/openzfs/openzfs/ (the old archived version based on illumos, if I read correctly) implies that this might have been true at one point, but indeed https://github.com/openzfs/zfs doesn't seem to do its own NFS so it's not true now if it ever was. Thanks for correcting my understanding.
I believe that the modularity will only proof itself when external (as in, from unrelated people) projects becomes established and we see how well the original project maintains compatibility.
(I must say I wonder how big the intersection of people-who-like-ZFS and people-who-like-systemd is; they seemed to originate from very different cliques but there's no reason people who like one would dislike the other…)
As for external clients for the layers, LustreZFS is a separate project though it had started with certain intersection of ZFS devs. However the general division of labour between layer is pretty strict (except for - now extinct - FreeBSD TRIM support), it's just that there isn't any work being done to use it outside of OpenZFS.
The boundaries are pretty clear, it's just that ZPL and ZVOL build up on all of them. Linux has /some/ related features, but nothing that was feature-parity: SPA roughly corresponds with MD/DM subsystem assuming certain plugins in use, DMU is very roughly the equivalent of OSD subsystem, but that one supports only SCSI OSD which has incompatible assumptions etc. - in fact, an OSD implementation on top of DMU should be pretty simple (main differences are due to DMU being a bit explicit on redundancy features, iirc).
