ZFS, Actually Storage you can finally reason about.

Resilvering

concept
depth 3/3 − deeper + surface

Resilvering is what happens after you swap in a replacement disk: ZFS refills it from the surviving team members. Older storage systems rebuild a replacement by blindly copying the entire disk, end to end, because they have no idea which parts hold anything. ZFS knows exactly what's real data, so a resilver copies the data—not the disk—and fingerprint-checks every piece as it goes ("resilvering only examines data that ZFS knows to be out of date"—zpool-scrub(8)). A half-full pool recovers in roughly half the time. And the time genuinely matters: until the resilver finishes, you're running with less protection than you paid for.

Living through one: the pool stays up and serving—a resilver is a background process, visible in zpool status with progress and an estimate (early estimates lie; trust them only after they settle). The system deliberately throttles resilver I/O behind live traffic, so a busy pool resilvers slower. Don't reboot casually mid-resilver—it resumes, but the window extends.

Vdev geometry sets the character of the event: a mirror resilver is a straight copy from the healthy twin—fast, gentle. A RAIDZ resilver reconstructs the missing disk from all survivors, hammering every one of them for hours or days at exactly the moment redundancy is reduced—correlated disks (same batch, same age) make that a real stress test. This asymmetry is a load-bearing input to the mirrors-vs-RAIDZ decision.

Placed against its relatives:

  • vs a scrub—same walk, different scope: scrub reads everything to find silent damage; resilver reads what's needed to rebuild one member. The pairing is standard: resilver to restore redundancy, then scrub to verify the whole (see Replacing a Disk).
  • vs sequential rebuild—the block-tree walk has an honest cost: on a very full, fragmented pool it's seek-bound and can be slower than a dumb sector copy (zpool replace -s offers a sequential mode for mirrors that trades immediate verification for speed). "Smarter" wins most of the time, not always.
  • The window logic—everything about resilvers feeds one number: how long you're exposed to a second failure. Emptier pools, mirror vdevs, RAIDZ2-over-RAIDZ1—most healthy-paranoia layout advice is resilver-window management by another name.