Snapshots Are Not Backups
conceptThe feeling is earned: snapshots every hour, every day, every month—you can bring back anything from any point in time. That's what backups do, so surely you have backups. Actually, ask one question of anything that claims to protect you: which disasters does it survive? A snapshot lives on the same disks, in the same machine, in the same building, under the same hands as the data itself. Fire, theft, the whole pool dying, malicious software that gets full control, one catastrophic mistyped command—the snapshot dies in every single scenario the word "backup" exists for.
Snapshots answer "the data went wrong." Backups answer "the place went wrong." Different questions, and you need answers to both.
What snapshots are, said positively: the versioning half of a backup strategy—recovery from deletion, bad upgrades, corruption-by-software, your own past self. Instant, free, high-resolution. Most restores you'll ever perform come from a snapshot, because most disasters are small.
The other half is replication: zfs send the snapshots to a different
pool on a different machine (syncoid on a schedule does exactly this). Now
the failure domains actually separate—and the classic 3-2-1 shape
(three copies, two media, one offsite) maps cleanly onto ZFS: live pool +
snapshots, replicated pool in the house, one leg somewhere else entirely.
The test that keeps everyone honest: a backup you haven't restored from is
a hypothesis. Browse the replica's .zfs/snapshot/, restore one real file,
once a quarter.