ZFS, Actually Storage you can finally reason about.

What ZFS Is

concept
depth 3/4 − deeper + surface

People call ZFS a filesystem, because that's the word the documentation uses. A filesystem is the software that keeps track of your files—which pieces of the disk belong to which document, what's named what, what lives in which folder. You've used them all your life without naming them: FAT32 on a USB stick, NTFS on Windows, APFS on a Mac.

Actually, ZFS is more than that. It does the filesystem's job—and also the jobs storage normally spreads across other tools: combining many disks into one supply of space, checking everything it reads for damage and repairing it from spare copies, freezing moments in time so you can rewind a mistake, and shipping those moments to another machine. One system, doing the whole job of storage. People reach for it when a computer's job is to hold data that matters—and that merger is where everything interesting about it comes from.

Using ZFS feels different from ordinary storage from the first minute: the usual parade of single-purpose utilities—one tool to partition the disk, one to group disks together, one to format, one config file to say what mounts where—simply isn't there. Two commands cover almost everything:

  • zpool—the physical side: disks, redundancy, health.
  • zfs—the logical side: datasets, properties, snapshots, replication.

You meet ZFS most often without choosing it: Proxmox offers it at install, TrueNAS is built on it, a homelab guide says "make a pool." Which is exactly how you end up operating ZFS for a year without being able to say what it is. The commands work whether or not you hold the model—until the first disk dies or the first pool fills, and then the model is the only thing that helps.

Now the category question, because it's where ZFS sits oddly among its neighbors: ZFS is the filesystem, the volume manager (the software that gathers disks into one usable space), and the RAID layer (the software that keeps redundant copies so a dead disk costs nothing) merged into a single system. Elsewhere those are three separate tools, stacked.

Why the merge matters: the traditional stack is a bucket brigade of strangers—the filesystem hands blocks to a volume manager, which hands them to a RAID layer, which hands them to disks—and no layer can check another's work. RAID will faithfully mirror corruption it can't detect; a filesystem can't ask RAID "which of your two copies is right?" because the interface between them has no words for that question.

People called ZFS's merge a layering violation—a serious accusation in systems design. The retort is that the boundary being violated was the problem: repair requires one system that both knows what the data should be (checksums, the filesystem's knowledge) and has another copy to fix it with (redundancy, the RAID layer's knowledge). Split those across an interface and self-healing is structurally impossible.

Nearest kin: Btrfs (same convictions, rockier history—its RAID5/6 never became trustworthy) and APFS (took copy-on-write and snapshots, skipped data checksums and redundancy—the single-device convictions only).