Scrubs
conceptThe word sounds like cleaning—you scrub a thing to fix it up. Actually, a scrub is a read-through: ZFS reads every piece of data the pool holds and checks every fingerprint ("a normal scrub examines all data in the specified pools and verifies each block's checksum"—zpool-scrub(8)). Why force that? Because ZFS checks data whenever it's read anyway—but the files nobody has opened in three years never get read. Data can quietly go bad while sitting untouched, and a scrub visits it while a spare copy still exists to fix it with.
The fixing itself only happens where a spare copy lives: "for replicated (mirror, raidz, or draid) devices, ZFS automatically repairs any damage discovered." Scrub a setup with no spare copies and ZFS hands you a list of what went bad—named files, even—and can heal none of it.
The practice: zpool scrub tank, online, background, throttled behind real
traffic—expect hours to a day-plus on big HDD pools. Monthly is the
homelab convention, usually via a systemd timer or cron a distro already
installed (check before adding a second one). Then read the results in
zpool status: repaired 0B ... with 0 errors is the sentence you want.
A scrub that repairs a few blocks did its job; a device whose CKSUM count
rises scrub after scrub is announcing its retirement—that, plus a rising
trend, is your early-warning system working.
Why it must be scheduled at all: your hot data self-verifies on every read. The archive nobody has opened in three years verifies never—unless something walks it. The scrub is that something.