RAIDZ2 or RAIDZ3: how much parity does a large ZFS pool need?

What does each parity level survive?

The OpenZFS manual sets the rule: a raidz group of N disks with P parity disks holds about (N − P) times the disk size and survives P failed devices. RAIDZ2 survives two, RAIDZ3 three.

The same page recommends 3 to 9 devices per group, "to help increase performance". A ten-drive group sits one above that range. The range is about performance, not safety: parity covers P failures at any width.

Each extra parity drive buys one more failure, and costs one drive of space. On Hold 160 that is 16 TB, or 12.5% of the RAIDZ2 capacity.

How long is the pool exposed during a resilver?

As long as the new drive takes to fill. ZFS resilvers only the data the pool holds, so a half-full pool rebuilds about half the bytes.

• Best case, full 16 TB drive: 16 TB at the 285 MB/s maximum of a Seagate Exos X20 takes about 15.6 hours.
• Busy pool, example at 50 MB/s: the same drive takes about 89 hours, or 3.7 days.
• Before the resilver starts: a failed drive in a rented server waits for the datacenter to swap it. When a drive fails, open a ticket: the team has it replaced at no cost, and rebuilding the array on the new drive is up to your setup.

During that window, RAIDZ2 has one parity left and RAIDZ3 has two. The rebuild arithmetic for 10 to 20 TB drives is on how long a RAID 6 or RAIDZ2 rebuild takes.

What do unreadable sectors change?

They make the third parity drive useful even without a third failure. The Exos X20 data sheet rates non-recoverable read errors at 1 sector per 10^15 bits read. That is one sector per 125 TB read, at the data-sheet ceiling.

A resilver reads the surviving drives. On a full ten-drive pool of 16 TB drives, nine survivors hold 144 TB, so the ceiling allows about one unreadable sector. Real drives usually do better than their rating; the rating is the worst case the maker accepts.

• RAIDZ2, one drive down: ZFS rebuilds an unreadable sector from the remaining parity. Nothing is lost.
• RAIDZ2, two drives down: no parity is left. The block is lost, and zpool status -v lists the files containing data errors.
• RAIDZ3, two drives down: one parity is left, so the sector is rebuilt.

RAIDZ3 is insurance for the second failure, not the first.

How likely is a second failure during the window?

Small, on paper. Backblaze reported a 1.41% lifetime annualized failure rate across its drives in its Q2 2026 report. Over an 89-hour resilver, one drive fails with a probability of about 1.41% × 89 ÷ 8,760, or 0.014%. With nine survivors, the chance of one more failure is about 0.13%.

Those sums treat failures as independent. Drives bought together share a model, a batch and an age, and a resilver works every survivor hard. The figures are a floor, not a forecast.

The market is moving to that size. Of the 23 storage-server sellers Basalthold checked on 2 November 2026, 8 listed a server with 100 TB or more of hard drives.

Can I switch from RAIDZ2 to RAIDZ3 later?

No. OpenZFS 2.3 added RAID-Z expansion: you can attach a new disk to an existing group, with the raidz_expansion pool feature. The manual is explicit that "a RAID-Z2 is still a RAID-Z2 even after expansion". Old blocks also keep their old data-to-parity ratio.

Changing the parity level means copying the data out, destroying the pool and creating it again. On a rented server, the right moment is delivery, when the data drives are still blank. The operating system sits on its own SSD (two mirrored SSDs on the larger holds); the data drives arrive blank and unpartitioned, so you can build ZFS, an mdadm array or a set of independent disks. If you want the team to prepare a RAID 6 or RAID 10 array before delivery, say so to support before the server is handed over.

# your ten data drives by stable ID; the system SSDs stay out of the pool
ls -l /dev/disk/by-id/ | grep -v part

# three drives' worth of parity (replace DRIVE01 to DRIVE10 with the real IDs)
zpool create -o ashift=12 -O compression=lz4 tank raidz3 /dev/disk/by-id/ata-DRIVE{01..10}

# capacity, health, and the files hit by any data error
zpool list tank
zpool status -v tank

Which should I choose on a rented storage server?

Choose by what the pool is for and how long a failure can wait.

• RAIDZ2: a backup target whose originals live elsewhere. A lost file can be copied again.
• RAIDZ3: an archive that is the only copy for a while, a pool you cannot watch every day, or drives so large that a resilver under load runs for days.
• Two or four drives: a mirror on Hold 24; on Hold 40, RAIDZ2 and striped mirrors both keep 20 TB: RAIDZ2 survives any two failures, mirrors give faster random I/O.

Hold 100 and Hold 160 carry ten data drives each; their terms are on the Basalthold pricing page.

Parity is not a backup. A storage server is not a backup of itself: we keep no copy of your data anywhere else, so a second copy in another place stays your job. The usual second copy for a NAS is set out on what the best offsite backup for a NAS is. The same trade-off without ZFS is on RAID 6 or RAID 10, and why ZFS wants the bare drives is on ZFS or hardware RAID.