Sanoid and Syncoid
conceptA snapshot routine that actually protects you—taken every hour, old ones cleaned out, fresh copies shipped to another machine—is a chore no human keeps up by hand at three in the morning. So most serious ZFS setups run two small helper tools, usually installed together and easily mistaken for one thing called "the backup system." Actually, they're two separate helpers, each automating one job you can do yourself. Sanoid is the snapshot-taker: "a policy-driven snapshot management tool" (its README)—it takes snapshots on a schedule (hourly, daily, monthly) and deletes old ones according to rules you set. Syncoid is the shipper: it "facilitates the asynchronous incremental replication of ZFS filesystems"— it finds the last moment both machines share, and sends over only what's new since.
Neither adds any storage machinery of its own. They're scheduling and judgment layered on top of snapshots and send/receive—which is why understanding those two ideas is understanding these tools.
The working setup (the practitioner's homelab runs exactly this shape):
/etc/sanoid/sanoid.conf assigns each dataset a template—
[tank/services/postgres]
use_template = production # hourlies pruned in days; frequent history
[tank/media]
use_template = archive # weekly, kept for months; slow history
—and a timer runs sanoid --cron. Separately, a scheduled
syncoid --recursive tank/services backup-host:vault/services replicates.
Two decisions worth making deliberately rather than by default: retention
per dataset (churn-heavy data accrues snapshot rent fast—see Snapshots
and Disappearing Space), and replication cadence per dataset (how much
history are you willing to lose—that's your real RPO, in systems terms:
the age of the newest copy elsewhere when disaster hits).
The relationship that bites when misunderstood: sanoid's pruning and syncoid's incrementals share the snapshot chain. Incremental replication needs a common snapshot on both ends; a retention policy that destroys it forces the next replication to start from zero. Design retention and replication as one system (syncoid's own sync snapshots exist to hold the reference point)—and treat hand-deleting snapshots under a policy engine as off-limits; change the policy instead.
Automation's quiet cost, named once: a year of these tools working
flawlessly is a year of never being forced to learn what they do—the
naive models survive because nothing fails. Run the primitives by hand
once (zfs snapshot, zfs send -i ... | zfs recv), and the config file
stops being an incantation. Then let the robots have it back.