ZFS, Actually Storage you can finally reason about.

The Vdev

concept
depth 4/4 − deeper + surface

From above, the pool looks like one undivided pot of storage. Underneath, ZFS never treats your disks as one big group. It binds them into small teams, and each team backs itself up. That team is a vdev ("virtual device"). A team might be two disks keeping identical copies (a mirror), a group of disks sharing recovery math so any one can die (RAIDZ), or— dangerously—one lone disk backing up nothing. A pool is built by joining teams together.

Here is the sentence this entire site most wants you to hold: spare copies exist only inside each team, never between teams—so losing any one whole team loses the whole pool. Not damaged. Gone. The pool spread everyone's data across all its teams, so a dead team is a bite out of everything.

This is why the classic homelab move—"the pool's getting full, I'll just add this spare disk"—is a trap. zpool add tank /dev/sdX with a bare disk welds a redundancy-less vdev into your pool of mirrors, permanently. Now that one consumer drive can take down everything, and the two mirrors you paid for protect nothing. The pool is exactly as strong as its weakest vdev.

Audit your real layout in ten seconds: zpool status indents the truth—vdevs (mirror-0, raidz1-0) at one level, member disks under them. A disk sitting at vdev indentation with no group above it is a single-disk vdev. If you find one under data you care about, that's a rebuild-the-pool conversation, not a someday conversation.

The vdev is where every serious layout question actually lives:

  • Mirrors vs RAIDZ is a per-vdev choice (and don't mix types under one pool without a reason).
  • Performance scales with vdev count, not disk count—the pool stripes across vdevs, so six disks as three mirrors substantially out-IOPS six disks as one RAIDZ2.
  • A resilver is a vdev event: one vdev rebuilding a member while the others carry on.
  • Growth happens vdev-wise: add a vdev, or swap every disk in one vdev for bigger ones. (RAIDZ widening exists now—see RAIDZ Expansion—but the vdev remains the unit you plan around.)

Special vdev classes exist (log, cache, spare, special)—supporting cast, not data vdevs; the log and cache ones have their own nodes.

Why did ZFS put redundancy inside the stripe instead of across the whole pool, when whole-pool parity sounds safer? Because the vdev boundary is what keeps redundancy composable and repairable:

  • Each vdev handles its own failures locally—a resilver touches one vdev's disks, not the whole pool.
  • Vdevs of different shapes, ages, and sizes can join one pool over the years; a global parity scheme would have to be re-solved on every change.
  • The allocator balances new writes across vdevs by free space and load—striping is a placement policy, cheap and adaptive, while redundancy stays a structural property, local and stable.

The cost of that composability is the brutal rule at depth 2: the pool stripes with no parity of its own, so vdev survival is pool survival. ZFS chose "redundancy is local, loss is total" over "redundancy is global, everything is coupled"—and it expects you to know it chose that.