Ubuntu 26.04 Server: Guided ZFS Root + Extra Datasets via Autoinstall

• CIS compliant disk layout — separate mount points for /var, /tmp, /var/log, /var/tmp, and /var/log/audit, matching the CIS Benchmark's partitioning recommendations
• CIS security settings are NOT included — no nodev/nosuid/noexec mount-option hardening, no benchmark packages, no other CIS controls. This gives you the disk layout only; hardening is left for you to add separately

This documents an unattended (autoinstall) install of Ubuntu Server 26.04.1 with:

• ZFS root, using Subiquity's guided ZFS layout (the only ZFS layout exposed to autoinstall today; it is not offered in the interactive Server installer UI, only via autoinstall or the Desktop installer)
• Two extra datasets the guided layout doesn't create on its own (var/tmp, var/log/audit), plus a tmp dataset, added via late-commands
• Quotas on var (8G) and tmp (6G)
• A couple of quality-of-life tweaks (zsh as the login shell, vim/zsh packages, a pinned 2G swap size)

It was built and tested in a Proxmox VM. Nothing here is Proxmox-specific except the "build the VM" section — the user-data file itself works on any hypervisor or bare metal that boots the Ubuntu Server 26.04.1 ISO.

Result

zfs list / zpool list on a freshly installed system, showing the guided layout plus the var/tmp, var/log/audit, and tmp datasets added by late-commands:

Why autoinstall instead of the interactive installer

The interactive Server installer (Subiquity's text UI) only offers LVM or a custom manual partition layout — there's no ZFS option in that menu, even though the Desktop installer has one. Subiquity's own maintainers have confirmed this directly: ZFS root on Server requires either a Desktop install, autoinstall, or a fully manual process. Autoinstall is the only one of those that stays server-oriented and scriptable.

Requirements

• Ubuntu Server 26.04.1 ISO
• A way to build an ISO with genisoimage (or equivalent) and attach it as a second virtual CD-ROM alongside the install media
• Basic familiarity with zfs/zpool

1. Build the VM (Proxmox specifics)

• Machine type: q35, BIOS: OVMF (UEFI), with an EFI disk added
• Leave Secure Boot off for simplicity (untested with Secure Boot + ZFS DKMS)
• Disk: SCSI, VirtIO SCSI single controller, Discard enabled, cache = "No cache"
• At least 4GB RAM (8GB+ more comfortable), 2 vCPUs is plenty for a test box
• 20-32GB disk is comfortable; ZFS datasets share pool space dynamically, so you don't need to size /var, /tmp, etc. individually the way you would with fixed partitions
• Attach the Ubuntu Server ISO as one CD-ROM, and the autoinstall seed ISO (built below) as a second CD-ROM

None of this is required on other hypervisors — just make sure UEFI boot is available and you can attach a second CD-ROM/virtual media device.

2. The autoinstall files

Two files are needed: user-data (the actual autoinstall config) and an empty meta-data (required by the NoCloud datasource format, even with minimal content).

meta-data

instance-id: ubuntu-zfs-test-01
local-hostname: ubuntu-zfs-test

user-data

#cloud-config
autoinstall:
  version: 1
  locale: en_US.UTF-8
  keyboard:
    layout: us

  network:
    network:
      version: 2
      ethernets:
        alleths:
          match:
            name: "en*"
          dhcp4: true

  # Guided ZFS root layout: creates bpool (/boot), rpool/ROOT/ubuntu_<id> (/),
  # rpool/home (/home), and a swap partition. This is the only ZFS layout
  # exposed to autoinstall today; extra datasets are added below.
  storage:
    layout:
      name: zfs
    swap:
      size: 2G

  identity:
    hostname: ubuntu-zfs-test
    username: youruser
    # Generate with: openssl passwd -6
    password: "$6$REPLACE_WITH_YOUR_HASH"

  ssh:
    install-server: true
    allow-pw: true

  packages:
    - vim
    - zsh

  user-data:
    chpasswd:
      expire: false

  late-commands:
    # Set zsh as the login shell -- update the username to match identity.username above
    - curtin in-target -- chsh -s /usr/bin/zsh youruser

    # Ubuntu's guided ZFS layout already creates var and its usual zsys-style
    # children (var/log, var/lib/apt, etc.) automatically. It does NOT create
    # var/tmp, var/log/audit, or tmp -- those we create ourselves. Then apply
    # quotas to var and tmp either way.
    - |
      curtin in-target -- bash -c '
      set -euo pipefail
      ROOTDS=$(zfs list -H -o name -t filesystem | grep -E "^rpool/ROOT/[^/]+$")

      for sub in var/tmp var/log/audit tmp; do
        ds="$ROOTDS/$sub"
        if zfs list -H "$ds" >/dev/null 2>&1; then
          continue
        fi
        path="/$sub"
        stage=/mnt/zfs-stage
        mkdir -p "$stage"
        zfs create -o mountpoint="$stage" "$ds"
        if [ -d "$path" ] && [ -n "$(ls -A "$path" 2>/dev/null)" ]; then
          cp -a "$path"/. "$stage"/
        fi
        rm -rf "${path:?}"/*
        zfs umount "$ds"
        zfs set mountpoint="$path" "$ds"
      done

      zfs set quota=8G "$ROOTDS/var"
      zfs set quota=6G "$ROOTDS/tmp"
      '

Before building the ISO, edit two things:

1. identity.username and the matching username in the chsh late-command
2. identity.password — generate a real hash with openssl passwd -6 and paste it in (keep the quotes)

3. Build the seed ISO

genisoimage -output seed.iso -volid cidata -joliet -rock user-data meta-data

The volume label must be cidata — that's what the NoCloud datasource looks for.

4. Boot and install

1. Attach both the Ubuntu Server ISO and seed.iso to the VM, boot from the Server ISO.
2. Recent Subiquity builds auto-detect the cidata volume and prompt to continue with autoinstall. If it doesn't, hit e at the GRUB menu and append autoinstall ds=nocloud; to the linux line.
3. The install runs unattended and reboots when done.

5. Verify

zfs list          # confirm var, var/tmp, var/log, var/log/audit, tmp, home all exist
zfs get quota rpool/ROOT/ubuntu_xxxxx/var rpool/ROOT/ubuntu_xxxxx/tmp
zpool status      # pool health
cat /etc/cron.d/zfsutils-linux   # confirm the built-in monthly scrub/TRIM cron job

zfsutils-linux (pulled in automatically as a dependency of ZFS root) ships a cron job that TRIMs on the first Sunday of the month and scrubs on the second Sunday, for every imported pool, with no extra configuration needed.

Lessons learned / troubleshooting notes

These are the non-obvious failures hit while building this, in case they save someone else the same round trips:

• The Server installer UI has no ZFS option at all, not even hidden in "Custom storage layout" — only Desktop, autoinstall, or a fully manual process support ZFS root on Server.
• Interface names aren't eth0 on modern kernels/VirtIO NICs — use a match: name: "en*" pattern in the network config instead of a hardcoded name, or DHCP silently never applies and anything requiring network access during install (e.g. a package not in the ISO's offline pool) will fail.
• Deleting/recreating a virtual disk doesn't guarantee a blank device on LVM-backed storage (including LVM over iSCSI) — LVM by default only zeroes the front of a new logical volume, not the whole thing. Leftover ZFS pool labels from a previous attempt can survive wipefs -a and sgdisk --zap-all (both only touch the partition table, not partition contents) and cause zpool create to fail with is part of potentially active pool. The reliable fix is a full-disk zero: dd if=/dev/zero of=/dev/sda bs=1M status=progress.
• zpool get bootfs rpool is not a reliable way to find the root dataset. Ubuntu's guided ZFS layout doesn't set the bootfs property, so this returns -. Use zfs list -H -o name -t filesystem | grep -E "^rpool/ROOT/[^/]+$" instead.
• Watch your quoting if you nest a script inside curtin in-target -- bash -c '...' inside a late-commands entry — single quotes can't nest inside single quotes. Use double quotes for anything inside the outer single-quoted block (e.g. a grep -E pattern).
• The guided ZFS layout already creates more separate datasets than you'd expect: var and its classic zsys-style children (var/lib, var/lib/apt, var/lib/dpkg, var/log, var/mail, var/snap, var/spool, var/www, var/games), plus srv, usr/local, home, and root. It does not create var/tmp, var/log/audit, or tmp — those need to be added manually if you want them.
• zfs set mountpoint=<path> <dataset> already mounts the dataset as a side effect of the property change. A following explicit zfs mount <dataset> fails with "already mounted" — harmless in itself, but fatal under set -e. Don't call zfs mount right after zfs set mountpoint=.
• If an install fails, the installer drops you to a shell with the message "An error occurred. Press enter to start a shell." From there, curtin itself isn't on PATH (it lives inside subiquity's snap) — to poke around inside the partially-installed target, bind-mount and chroot manually:

  mount --bind /proc /target/proc
  mount --bind /sys /target/sys
  mount --bind /dev /target/dev
  chroot /target /bin/bash

Customizing

• Different quotas or paths: edit the QUOTAS-equivalent lines and the for sub in ... list in the late-commands script.
• Different packages: edit the packages: list.
• CIS mount-option hardening (nodev/nosuid/noexec) is intentionally not included here — this only reproduces the CIS-style dataset/partition layout, not the security settings. Add ZFS dataset properties like setuid=off/exec=off/devices=off per-dataset if you want that too.