Notes & Tools

There are roughly two camps for ZFS snapshot retention. One is “snap everything, keep forever, the dedup will save you.” It will not. The other is “snap on a schedule, prune by age.” That one works.

My current zfs-auto-snapshot config:

frequent   keep=4   every 15 minutes
hourly     keep=24
daily      keep=14
weekly     keep=8
monthly    keep=6

Total worst case is 4 + 24 + 14 + 8 + 6 = 56 snapshots per dataset. With recordsize=128K and reasonable churn, that’s a few hundred MB of overhead per snapshot, not a few GB. The trap is the frequent tier: if you have a noisy temp directory inside a snapped dataset, each 15-minute snap pins a whole chunk of writes.

The fix is zfs set com.sun:auto-snapshot=false on tank/var/cache, tank/var/log (if you don’t need it snapped), tank/tmp, and any build-output dataset. It’s the property pattern, not a global exclude — you can flip it per dataset and the agent respects it.

The rule of thumb: a snapshot you cannot identify in 10 seconds (“what state of the world is this?”) is not a backup, it’s clutter. Keep fewer, keep them on a schedule, and always name your manual ones (zfs snapshot -r tank/home@before-upgrade-2024-11-23).