Quick answer

Pick ZFS for any array with parity. Its RAIDZ levels are stable, checksum every block, self-heal from a good copy, and never suffer the RAID-5 write hole. Btrfs is the better fit for single-disk or mirror setups where you want cheap snapshots and send/receive with far less memory. Btrfs RAID5 and RAID6 remain unsafe for production data.

Contents

At a glance

Class
Homelab

Provenance

  • Nov 16, 2026Published

Two filesystems get recommended for the same job, and they are not the same tool. ZFS is a storage stack that happens to ship a filesystem inside it. Btrfs is a filesystem that learned to manage more than one device. On a NAS, a Proxmox datastore, or a backup target, the decision collapses to one question: do you need parity RAID you can trust with data you cannot rebuild?

TL;DR · ZFS vs Btrfs

  • Parity array of three or more drives: ZFS RAIDZ. Stable, checksummed, no RAID-5 write hole.
  • Two-drive mirror or single disk: Btrfs. Cheap snapshots, subvolumes, and no ARC budget.
  • Proxmox VM datastore: mirrors on either filesystem, never RAIDZ, because IOPS matter more than bandwidth.
  • Small ARM board or 4 GB NAS: Btrfs. ZFS wants 2 GiB plus 1 GiB per TiB.
  • Never: Btrfs RAID5 or RAID6 for data you care about.

What both filesystems actually guarantee

Both are copy-on-write. New data lands in free space, the checksum is written, and only then does the tree pointer move to the new block. An interrupted write leaves the old tree intact, which is why neither filesystem needs an fsck pass on every unclean shutdown.

Both checksum. ZFS states plainly that all metadata and data is checksummed, and that it automatically repairs bad data from a good copy when corruption is detected (OpenZFS zpoolconcepts). Btrfs checksums data and metadata by default and verifies those checksums after reading, using crc32c unless you switch to xxhash, sha256, or blake2 (Btrfs Checksumming).

Both snapshot, both scrub, and both have a send stream. The divergence is not in the feature list. It is in how much verification each one performs without being asked, and in which RAID profiles are actually finished.

Where ZFS wins: RAIDZ, the write hole, and repair on read

The single most important line in the OpenZFS manual is that raidz does not suffer from the RAID-5/6 write hole, the window after a power loss where data and parity can be left inconsistent (zpoolconcepts). That is the property that makes parity RAID defensible on a home array at all.

Capacity follows the usual arithmetic. A mirror of N disks each holding X bytes stores X bytes and survives N-1 failures. A raidz vdev holds roughly (N minus parity) times X, raidz1 has one parity disk, raidz2 has two, raidz3 has three, and the manual recommends keeping a raidz vdev between three and nine disks (zpoolconcepts). A four-drive raidz1 gives you three drives of usable space. The same four drives as two mirrors give you two, plus double the random read performance.

Expansion used to be the fatal flaw. It is not anymore. zpool attach can add a device to an existing raidz vdev while maintaining redundancy, and the documentation is explicit that expansion does not change the number of failures a vdev tolerates, so a RAID-Z2 stays a RAID-Z2 afterwards (zpool-attach). Two footnotes matter: expansion reserves about 3.5 MiB per disk for a boot area, and the per-vdev disk limit drops from 255 to 56 disks at ashift=16 (zpool-attach). Proxmox documents raidz expansion as available from Proxmox VE 9 on ZFS 2.3.3 (Proxmox ZFS wiki).

Self-healing is the second win, and it is subtler than the marketing suggests. ZFS repairs bad data from a good copy as corruption is detected, not only during a scrub (zpoolconcepts). The zpool scrub command exists to force a full sweep of every block (zpool-scrub), but a verified read of a damaged block already triggers repair. That difference shows up years later, when a drive starts returning bad sectors silently.

Where Btrfs wins: flexible devices, cheap snapshots, small hardware

Btrfs treats a pool as a set of devices you can add, remove, or replace on demand, and it lets you change the RAID profile of existing data with btrfs balance rather than rebuilding the filesystem (Volume-management). If you started with two drives as RAID1 and later want RAID0 for scratch space, you convert in place. ZFS has no equivalent; you plan the vdev layout up front.

Subvolumes are the other practical advantage. One pool can hold a dozen independently snapshot-able filesystems, which is how distros implement rollback-on-upgrade and how a Proxmox host can snapshot a container root without carving out a zvol. Btrfs send and receive carry those subvolumes to another host as full or incremental streams, with each command protected by CRC32C (Send-receive). ZFS does the same thing with zfs send, but the memory cost of getting there is higher (zfs-send).

Memory is where Btrfs quietly wins the small-hardware argument. ZFS runs an Adaptive Replacement Cache in RAM, and Proxmox recommends budgeting at least 2 GiB of base memory plus 1 GiB per TiB of stored data, with the ARC enforcing a hard 64 MiB minimum (Proxmox ZFS wiki). On a 4 GB ARM NAS or a mini PC you also use as a media server, that budget is the whole machine. Btrfs has no ARC and simply borrows the kernel page cache, which is why it remains the sane default for single-board boxes.

The decision table

RequirementZFSBtrfs
Parity RAID in productionRAIDZ1/2/3, stableRAID5/6 unstable, evaluation only
RAID-5/6 write holeNot present in raidzDocumented, journal not implemented
Two-drive mirror with snapshotsWorks, needs ARC budgetWorks, minimal overhead
Single disk with rollbackPossible, heavierNative subvolume snapshots
VM datastore IOPSMirrors only, raidz is one disk of IOPSMirrors only, same logic
Add a disk to existing parity setzpool attach, OpenZFS 2.3+Not applicable, no stable parity
Change RAID profile in placeNo, plan the layout firstbtrfs balance convert
Metadata checksumsAlwaysDefault, crc32c unless changed
Repair on a normal readYes, from a good copyVerifies, but needs a scrub to find all copies
Send/receive to remotezfs send/zfs receivebtrfs send/btrfs receive
Memory floor2 GiB base + 1 GiB/TiBNo dedicated cache
Per-vdev disk guidance3-9 disks for raidzProfile-dependent

Recovery and the operational caveats that bite

The Btrfs RAID56 warning is not a legacy footnote. The project’s own manual says RAID5 and RAID6 should not be used in production, only for evaluation or testing, that power-failure safety for metadata with RAID56 is not complete, and that you should not use raid5 or raid6 for metadata at all, using raid1 or raid1c3 instead (btrfs(5)). The status page still lists RAID56 as unstable (Btrfs Status). Anyone selling you a four-bay Btrfs NAS on the promise of RAID5 is selling you a test rig.

Btrfs auto-repair is real but conditional. It salvages data from another valid copy when the profile has redundancy, and the documentation notes that other damaged copies are not detected unless a scrub runs (Auto-repair). So your scrub schedule is not optional housekeeping on Btrfs; it is the only mechanism that finds the second bad copy.

Device replacement has a sharp edge on Btrfs: replace aborts if it hits an input/output error reading the source device (Status). A source disk that is dying slowly can stall the rebuild you started because it was dying. Replace early, not at the first sign of failure, and never treat a rebuild as a backup.

On the ZFS side, the caveats are about planning rather than corruption. A raidz vdev behaves approximately like a single disk for IOPS while providing a lot of bandwidth, and Proxmox is direct that IOPS is the more important metric for running virtual machines (Proxmox ZFS wiki). This is why a six-drive RAIDZ2 pool can feel sluggish hosting eight VMs while a three-mirror pool on the same disks feels instant. If your workload is databases, virtual machines, or anything with small random reads, mirrors beat parity on both filesystems.

ECC memory is worth the money on ZFS, and the reason is narrow: ZFS trusts RAM, and a bad DIMM can write a checksum that matches bad data. I cover which failure modes ECC actually prevents in what breaks without ECC on a home server, and it is a more honest treatment than the usual blanket warning.

Which one should you pick

If you are building a parity array, run ZFS. Three to nine drives in raidz1 or raidz2, checksums on everything, no write hole, and a scrub that actually finds silent corruption. Use the NAS storage calculator to size usable capacity before you buy, because raidz overhead is not intuitive once you subtract the 80 percent fill guideline you should be respecting anyway.

If you are building a two-drive mirror, a single-disk boot volume, or anything on a small ARM board, run Btrfs. You give up nothing that matters at that scale, and you keep your RAM.

If you are virtualizing, the filesystem is the second decision. The first is mirrors versus parity, and it decides everything downstream. The Proxmox capacity planner will tell you how much usable space a given layout leaves you, but the IOPS answer is already settled by the word RAIDZ.

For the wider stack decision, including whether ZFS should live on the hypervisor at all, read Proxmox vs TrueNAS vs Unraid as a storage backend and the four-way NAS OS comparison. If you have already committed to a ZFS pool, the pool setup and ARC tuning walkthrough covers the commands, and drive selection is covered in best NAS hard drives.

Frequently asked questions

Is Btrfs RAID5 or RAID6 safe to use in 2026?
No. The upstream Btrfs documentation still classifies RAID5 and RAID6 as unstable and states they should not be used in production, only for evaluation or testing. The same manual warns that power-failure safety for metadata with RAID56 is not complete and explicitly says not to use raid5 or raid6 for metadata. If you want single-disk parity on four or more drives, use ZFS RAIDZ1 instead.
Can I add a disk to an existing ZFS RAIDZ array without destroying it?
Yes, with raidz expansion. The zpool attach manual page documents adding a device to an existing raidz vdev while maintaining redundancy, and notes that a RAID-Z2 remains a RAID-Z2 after expansion, so the number of tolerated failures does not change. The feature needs OpenZFS 2.3 or newer; Proxmox VE documents it as available from Proxmox VE 9 with ZFS 2.3.3.
Does Btrfs need as much RAM as ZFS?
No. ZFS keeps its Adaptive Replacement Cache in system memory and Proxmox recommends budgeting at least 2 GiB of base memory plus 1 GiB per TiB of stored data, with a documented floor of 64 MiB for the ARC itself. Btrfs has no equivalent page cache of its own and relies on the kernel’s normal caching, which is why it runs comfortably on 4 GB ARM boards where a ZFS pool would be starved.
Do both filesystems actually repair corrupted data automatically?
Both can repair from a second valid copy when the storage profile carries redundancy, but the mechanics differ. ZFS checksums all metadata and data and repairs bad data from a good copy as corruption is detected, which includes reads. Btrfs verifies checksums after reading and can salvage from another valid copy, but its own documentation notes that other damaged copies are not detected unless a scrub runs.
Why does ZFS feel slow for virtual machines even on fast disks?
Because RAIDZ trades input/output operations per second for bandwidth. Proxmox’s ZFS documentation states that a RAIDZ of any redundancy level behaves approximately like a single disk with regard to IOPS while offering a lot of bandwidth, and that IOPS is the more important metric for running virtual machines. A pool of mirrors gives you one vdev’s worth of IOPS per mirror pair instead.
What happens if a Btrfs device fails during a replace operation?
Per the Btrfs status documentation, device replace aborts if it hits an input/output error while reading the source device. That means a degrading source disk can stop the rebuild, so replace a failing device early rather than waiting for it to drop out entirely, and keep a second copy of anything irreplaceable on separate media.
Is Btrfs send/receive a real backup?
It is a replication tool, not a backup by itself. Btrfs documents full and incremental send streams that can be piped to another host or file, with each command protected by CRC32C. A stream stored on the same pool that produced it is not a backup. Send it to a second machine, a second pool, or offline media, and verify a restore before you trust it.
Which filesystem should I format a single NVMe drive with?
Btrfs. On a single device there is no parity to protect, so the deciding factors become snapshot cost, subvolume layout, and memory overhead. Btrfs gives you cheap snapshots and per-subvolume send/receive without the ARC budget, and its DUP profile can store two copies of metadata for extra resilience on a drive that has room.

Sources and corrections

Last updated
Methodology
See our methodology for research and review standards. It draws on 11 cited sources, listed below, each checked against the original page on the date above. Filesystem behavior in this comparison was checked on November 16, 2026 against the upstream OpenZFS manual pages for zpoolconcepts, zpool attach, scrub and send, the Btrfs project’s own Status, btrfs(5), Checksumming, Auto-repair, Volume-management and Send-receive documentation, and Proxmox’s ZFS on Linux wiki. No throughput or latency figures are quoted because I did not run a benchmark for this page; every performance statement is paraphrased from the vendor documentation cited on the same sentence.
Corrections
Spotted an error? Email a correction with the affected line and what you found. The draft includes this page’s address. Verified changes are recorded here. Contact details.

About the author

Written by Lowell K. Wood IV, who builds and runs TechFuelHQ from St. Louis, Missouri.