Should I use ZFS or hardware RAID on a rented storage server?

Why does OpenZFS advise against hardware RAID?

Because a RAID card hides the disks from ZFS. ZFS then sees one virtual disk and cannot use its own redundancy. The OpenZFS hardware guide lists the problems:

• Weaker self-healing. When a checksum fails, ZFS often cannot fetch a good copy from another disk, since the card owns the redundancy.
• Wrong sector sizes. The card can misreport them, which causes read-modify-write overhead.
• Locked-in replacements. A failed card may need the same model to read the array again.
• Write cache risk. Cached writes can vanish on a power loss or a dead battery.
• Unknown rebuild behavior. The guide cites reports of RAID 5 and 6 arrays lost during rebuilds after silent corruption.
• Closed firmware and shifting device names. You cannot inspect the code, and one failed array can rename the others.

Its advice is to use a host bus adapter (HBA) instead of a RAID controller, for performance and reliability.

ZFS can only heal what it can see: give it the drives, not a virtual disk.

What does ZFS do that a RAID card cannot?

ZFS checks every block it reads against a checksum. When a block is wrong and the pool has redundancy, ZFS reads a good copy and repairs the bad one. A scrub does this for all data, to catch silent errors from hardware faults or disk failure.

The filesystem and the array are one layer, so ZFS knows which blocks hold data. A resilver only examines out-of-date data, not empty space. That shortens rebuilds on a pool that is not full; the hours are worked out on the RAID rebuild time page.

It also brings snapshots, compression and native encryption per dataset. RAIDZ2 has no RAID 5/6 write hole, where data and parity disagree after a power loss.

When is hardware RAID or mdadm the better choice?

When you want a plain filesystem such as ext4 or XFS. Then something below it must provide the redundancy: a RAID card or Linux mdadm.

mdadm needs no card. Its arrays move to any Linux machine, and LUKS encryption sits cleanly under it. Its limit is the one ZFS removes: md does not know which blocks hold data, so it rebuilds whole drives.

If a server only comes with a RAID card, set the card to pass drives through before you build ZFS. That keeps ZFS in charge of the disks.

Of 23 storage-server sellers Basalthold checked on 2 November 2026, 4 stated a hardware RAID controller or option on a storage plan. Basalthold promises none: the data drives come blank, for your own software array.

Which one fits a backup server?

ZFS, in most cases. Backups are written once and read rarely, so silent corruption can sit unseen for months. Checksums and scrubs find it while a good copy still exists.

Proxmox VE and Proxmox Backup Server include ZFS, so the same pool serves the datastore; see Proxmox Backup Server offsite. ZFS wants memory for its cache; the sizing for 100 TB is on ZFS RAM requirements. The layout choice is on RAID 6 or RAID 10.

How do I check that ZFS sees the raw drives?

List the drives before you build anything. Each data drive should appear on its own, with its real model and serial number. Then create the pool from the stable names in /dev/disk/by-id, so a renamed device cannot confuse it.

Both data sheets read for this page describe 4,096-byte physical sectors. Seagate's Exos X20 ships in 512e format, and Western Digital calls the HC560 an Advanced Format drive. Set ashift=12 to match.

# every data drive on its own line, with model and serial
lsblk -d -o NAME,SIZE,ROTA,MODEL,SERIAL
smartctl -i /dev/sdb
ls -l /dev/disk/by-id/ | grep -v part

# RAIDZ2 over the data drives, by stable ID, 4K sectors
zpool create -o ashift=12 tank raidz2 \
  /dev/disk/by-id/ata-MODEL_SERIAL01 /dev/disk/by-id/ata-MODEL_SERIAL02 \
  /dev/disk/by-id/ata-MODEL_SERIAL03 /dev/disk/by-id/ata-MODEL_SERIAL04
zpool status tank

# scrub once a month to find silent errors
zpool scrub tank