ZFS, Actually Storage you can finally reason about.

Snapshots and Disappearing Space

troubleshooting
depth 3/3 − deeper + surface

The lived version: you delete a hundred gigabytes of junk to free up room, check the free space—and almost nothing came back. A lifetime of computers said deleting frees space. Actually, deleting a file only frees its space if nothing else is still holding it—and every snapshot taken while those files existed is still holding them. That's what a snapshot is: a promise that a past moment won't be thrown away. The space didn't leak; its ownership moved from your live files to the snapshots. Deleting things in the present cannot un-keep the past. (Operational source: the practitioner's own hourly-snapshot dataset doing exactly this.)

The diagnostic, in order:

  1. zfs list -o space tank/...—read USEDDS (the live dataset) vs USEDSNAP (held by snapshots). A big USEDSNAP is this exact situation.
  2. zfs list -t snapshot -o name,used,creation tank/...—see who's holding. Remember each snapshot's used shows only unique space, so a chain can collectively hold far more than the column suggests.
  3. To reclaim: destroy the snapshots that span the deletion. Dry-run first—zfs destroy -nv tank/ds@old1%old8 (a range, with -n showing what would free without doing it).

⚠ Destructive—zfs destroy of snapshots is permanent, and under a policy tool it's also usually wrong: sanoid will keep pruning on schedule anyway. Adjust the retention policy, not individual snapshots—hand-culling what a policy engine manages is how you end up with neither the space nor the history.

The two intuitions this scenario finally reconciles: "snapshots are free" (true—at creation) and "deleting frees space" (true—absent references). Snapshot cost is divergence: churn-heavy datasets (databases, build dirs, VM disks) under hourly retention accumulate held space fast; static media barely accrues any. Which means retention policy is a per-dataset decision—the databases might warrant aggressive hourlies pruned in days, the media library weekly snapshots kept for months. One-size retention is how the pool fills.

The graduating insight: USEDSNAP isn't waste—it's the storage price of your history, itemized. The question is never "how do I make it zero" but "is this history worth this rent," dataset by dataset.