ZFS, Actually Storage you can finally reason about.

ZFS Under Proxmox

scenario
depth 3/3 − deeper + surface

Proxmox is a popular system for running virtual machines—computers simulated in software—on one physical box, and it offers ZFS during installation. This is how most home servers end up running ZFS for years without their owners ever meeting it: click yes, machines appear, storage works. The natural assumption—each virtual machine's disk is a big file sitting in a folder somewhere—sends you hunting for files that don't exist. Actually, Proxmox carves each virtual machine's disk straight from the pool (a zvol—a slice of the pool pretending to be a real disk), and the mysterious rpool is just the pool the installer made for the operating system itself. (Operational source: the practitioner's own Proxmox hosts run exactly this stack.)

Decoding a typical host:

  • zpool status—rpool (the installer's, often a mirror of the boot disks) plus any pool you added. All the vdev/scrub/status reasoning from this site applies verbatim.
  • zfs list—rpool/ROOT/... (the OS), rpool/data/vm-103-disk-0 and friends (zvols, one per VM disk), and on a services host, datasets like tank/services/<name> per deployed service—filesystem face and block-device face of the same pool, side by side.
  • The RAM ledger—the ARC and your VMs share physical RAM, and the give-it-back dance is imperfect on a VM host; this is the one place capping zfs_arc_max is orthodox rather than superstition (see The ARC). Budget it deliberately: RAM = VMs + ARC cap + host slack.
  • Proxmox's own VM snapshots ride the same machinery—but sanoid policies on the zvols and syncoid replication offsite are what turn "snapshots exist" into a strategy (see Sanoid and Syncoid).

The two models reconciled—the point of this node: "datasets for services" and "zvols for VMs" felt like different systems when learned separately. They're one pool wearing two faces per consumer type: containers and services want filesystems (datasets—snapshot boundaries per service); VMs want disks (zvols—block semantics per guest). Policy, retention, replication, scrubs: identical underneath.

Two Proxmox-specific edges worth respecting: double-CoW—a guest running ZFS/btrfs (or qcow2-on-ZFS) on top of a zvol amplifies every write; give guests ext4/XFS unless you have a reason. And thin zvols overcommit the pool—five VMs each promised 100GB on a 300GB pool works until they all mean it; the free-space discipline gets less forgiving when the claimants are machines (see The Free-Space Discipline).