ZFS, Actually Storage you can finally reason about.

Datasets

concept
depth 3/3 − deeper + surface

For your whole computing life, organizing storage has meant one move: make a folder. Folders group things and give them names—and that is all they do. A folder carries no rules of its own.

ZFS adds a second kind of container: the dataset, a named slice of the pool. It looks and acts exactly like a folder—holds files, appears in your folder tree—but a dataset is the thing ZFS applies rules to. Each one can have its own settings (compress or not), its own space limits, its own frozen moments in time, its own copying schedule to another machine.

Two facts make them easy to love: a dataset has no fixed size (they all share the pool's one pot), and creating one costs nothing. So the habit to build: where you'd normally make a folder, consider making a dataset. Make many.

The rule of thumb: anything you might want to snapshot, replicate, quota, or tune separately gets its own dataset. A directory can't be snapshotted, sent to another machine, or given its own recordsize; a dataset can.

This is why a service-deployment setup makes a separate dataset per service rather than a folder per service: each service gets its own snapshot policy (the database hourly, the media library weekly), its own replication schedule, its own properties—and a restore of one service doesn't drag the others backward in time with it.

Datasets nest into a tree (tank/services/postgres), and children inherit properties from parents—so policy gets set once at the sensible level and flows down. Creating one is instant: zfs create tank/services/newthing.

The boundaries that actually matter:

  • Dataset vs directory—invisible day-to-day, decisive the day you need history: snapshots, sends, and rollbacks stop at dataset edges. The granularity you chose at creation is the granularity of your disaster recovery.
  • Dataset vs partition—no fixed size is the liberation, but it cuts both ways: nothing stops one dataset from eating the pool. quota (ceiling) and reservation (guaranteed floor) reintroduce limits as policy where you actually want them, instead of as geometry everywhere.
  • Dataset vs zvol—same carving, different face: a dataset presents a filesystem; a zvol presents a raw block device (what Proxmox gives VMs).

One warm-blooded warning: zfs destroy on a dataset takes its snapshots and its children with it, and there is no undo. The cheapness of creation does not extend to deletion.