I currently have a secondary pool (with raidz2) that I was originally going to use for my important documents, such as storage for Paperless-ngx, as raidz offers corruption detection and repair. The pool is encrypted.
However, I’m concerned about rebuild times (it’s a pool of 4 22TB drives). Is btrfs a better choice for this use case, or should I just go with raidz like I originally planned?
Edit: I should have mentioned that I already have 4-3-2 backups configured - I’m primarily interested in the “self-healing” aspect of ZFS so that I don’t have to recover from backups unless necessary, and to resolve corruption on the fly without me having to notice that a file is corrupt.
You must log in or register to comment.
Don’t use any btrfs raid levels besides mirroring. There are long-standing raid bugs that they wontfix in 5 and 6 that have led to data loss.
I’ve been using a raid1 btrfs pool to store offline backups for around 10 years. It’s 4 rotating drives (2x4TB+2x12TB). I replaced / rebalanced 3 disks with larger / newer ones already (went fine). I identified a bad usb/sata controller, and lots of bitrots on one old disk (scrub was able to correct a few thousands errors).
I’m getting around 80MB/s read/write throughput (not great but OK for offline backup). I’m able to mount it on low-powered / low-memory devices (not the case for ZFS). Scrub takes around 2 days IIRC (for around 10TB of actual data), so I run it once a year.
I keep it simple and thus am not using advanced features (dedup / encryption / snapshots / subvolumes / raid5/6/10). So far its a good match for my needs.
If rebuild time is truly your main concern, you could configure the 4 drives as raid 1+0. On truenas I think it would be called a pool of 2 mirror vdevs. It would give you much faster rebuild time and still have 1-2 drive failure tolerance.
RAID is not a backup.
RAID is not for data safety.
RAID is for:
- Ensuring availability of data in the face of hardware failure. That means your files don’t disappear when a drive dies and you have some time to swap out for functional hardware and restore redundancy.
- Presenting multiple drives as one larger unit. This is what striping does, and to a lesser extent the parity-mode levels.
- Improving performance (sometimes). A RAID mirror is generally much faster to read from than any individual drive because reads can be interleaved across drive members. A stripe can be much faster because writes are distributed across drive members. This is less of a bonus today with solid state/nvme drives, but it’s still applicable to spinning rust.
If your concern is protecting your data, set up a 3-2-1 backup strategy.
I’m very aware and have full 4-3-2 backups already. I’m also not interested in a standard raid. Thanks though! It’s always good to mention that raid is not a backup. I simply want to add more protection from disk corruption (not necessarily full failure) so that I don’t need to recover from backups unless I absolutely must. A benefit would also be resolving corruption before I even notice it.
I was originally going to use for my important documents
Not quite what you’re asking, but if your concern is avoiding data loss, if you haven’t already, I’d set up a backup before I started setting up a RAID or similar setup.
Filesystem doesn’t really matter once you have a reliable, redundant off-site backup and recovery plan set up and tested.
Really, use what fs feels best for you. And do your backups.
Did I mention backups are important?
I should have mentioned that I was (and still am) using XFS for my primary pool. XFS has been great, I was just hoping for a solution which would make it so I don’t have to hit my backups unless absolutely necessary (such as full array loss, for example). ZFS’ data corruption protection is very intriguing, it’s mostly the rebuild time and write speed that is making me think about doing something else.
What about 2 mirrored pools of 2 drives each, then back up the main pool to the other with either ZFS snapshots or a tool like Restic.
Ideally you also need an offsite backup of important files too, but that gets you part way to a robust system that can handle corruption or accidental deletions.
Would this just be to help with the rebuild time? Raid10 in ZFS is an interesting idea, which would also require two mirrors and striping.
Yes mirrors are the fastest to rebuild I believe, it’s also to give you a backup, as any kind of raid or mirror is not a functional backup, it only provides redundancy.
I would not do raid 10 for the same reason of no backup that way.
While redundancy in a drive setup helps, it’s not really a backup and thus not a “safe” way to store important information on it’s own.
That said, selecting the way you setup a raid system is based on risk and utility. I have a raid1 with a hotspare for important files. And use raid5 with 3&4 drives for less important stuff. You can also optimise for reading speed for example (as the same file can be drawn from multiple drives)
Like you said: RAID is not a backup. If it’s import follow at least the 3-2-1 rule. 3 copies on at least 2 different media, 1 of them off site.
Absolutely, 4-3-2 is what I use now! MDisc backups have been great.
I’d stay on zfs, I simply don’t trust btrfs’ raid implementation. For very important documents I also set copies=2 (or 3) on that dataset, just in case.
And as others already said: 3-2-1 backups ;)
Yeah, I’m leaning towards ZFS for sure after reading about btrfs’ past…
If you’re storing, then offline, read only, checksummed and preferably encrypted. And above all, several tested backups. I’m partial to making erofs images from dirs, but any archive that fits the content type works.
For the random access cache on top, whatever.
For your situation I would be more likely to go with a single drive with btrfs and dup for metadata redundancy. Regular snapshots and scrubs.
Use a second drive in the same system with btrfs to store snapshots at wider scheduled intervals. These will be bigger since no CoW on the separate file system. Scheduled scrub here too.
Use a third drive with ext4 as a backup target using a separate backup mechanism.
Use the fourth drive as a spare, or in a separate location as a target to send the backups if you don’t already have an off-site solution.
Interesting idea.
Acronyms, initialisms, abbreviations, contractions, and other phrases which expand to something larger, that I’ve seen in this thread:
Fewer Letters More Letters HTTP Hypertext Transfer Protocol, the Web IP Internet Protocol RAID Redundant Array of Independent Disks for mass storage SSL Secure Sockets Layer, for transparent encryption TLS Transport Layer Security, supersedes SSL VPN Virtual Private Network ZFS Solaris/Linux filesystem focusing on data integrity nginx Popular HTTP server
6 acronyms in this thread; the most compressed thread commented on today has 15 acronyms.
[Thread #314 for this comm, first seen 25th May 2026, 11:10] [FAQ] [Full list] [Contact] [Source code]
Why was this upvoted? It’s AI slop giving definitions for acronyms that aren’t in this thread and not even related to backups.
At least this time it has a few terms that people might not know. Usually it just spasms obvious trivialities.
It isn’t AI, you can take a look at the source code for it from the url it provides. Obviously the detection needs some tweaking, but extra acronyms in the list doesn’t really hurt anything when the other half are relevant.
Detection is completely broken because it finds terms that aren’t anywhere in the thread, even as substrings.
AI isn’t just LLM.
It wasnt even LLMs until the public took the term and changed it lol. Unless you are calling every algorithim ever made AI these days, this isnt AI.
Chess programs were AI. Expert systems which were regular logic were AI. Lisp was an AI language. Chat bots were AI.
This is a bot which makes it a type of AI and it’s really inaccurate.
uhm, no? Literally none of that was considered AI. Even chatbots, people weren’t calling them AI until LLMs came around and were stuck in them. Lisp is a language USED for AI research, that doesn’t make it AI itself.
This bot is most definitely not even close to what people consider AI
@Zeoic You are both correct. Chess was chosen as an early problem domain for work on first-order logic-based programming. That certainly was considered AI. People really interested in chess later abandoned logic programming in favor of brute-force, highly parallel special purpose hardware. That was not AI.
“Expert systems” (I hate that term) are application area of “Pattern-Directed Inference Systems” (PDIS). Rule-based systems are just one type of PDIS. For example, “Constraint Satisfaction” is another powerful AI technique often used in resource optimization and scheduling systems.
https://www.chessprogramming.org/Artificial_Intelligence
" the term ‘artificial intelligence’ was coined by John McCarthy in the proposal for the 1956 Dartmouth Conference [4] . In its beginning, Computer Chess was called the Drosophila of Artificial Intelligence. "
Expert Systems:
https://en.wikipedia.org/wiki/Expert_system "In artificial intelligence (AI), an expert system is a computer system emulating the decision-making ability of a human expert.[1] "
Chatbots in AI:
https://liacademy.co.uk/the-story-of-eliza-the-ai-that-fooled-the-world/
https://en.wikipedia.org/wiki/Eugene_Goostman
Lisp is a language USED for AI research, that doesn’t make it AI itself. “Lisp was an AI language.”
I didn’t say Lisp was AI. I said it was a language used for AI.
Maybe you could switch to a raid10 (mirrored striped vdevs) for faster rebuild time.
BTRFS is relatively similar to ZFS when it comes to their raid implementation, though using raid5 or raid6 comes with some caveats.
Raid 10 is and interesting idea for sure. It would certainly help with write speeds, although if I want to utilize the “self-healing” of ZFS, I’d need to do two (separately striped) vdevs that mirror each other to get the equivalent to RAID10, right?
You’d create two mirror vdevs in the same zpool to get the raid 10 equivalent
I would absolutely not trust BTRFS’s implementation. Maybe things are better now but it earned the backronym Bro The RAID Fuckin Sucks for a reason.
