Linus Torvalds Doesn't Recommend Using ZFS On Linux

[elvis]

I've never seen the point of ZFS when we have Btrfs...

Both have merit for different reasons in different use cases.

Speaking from experience, I've run a 288 drive ZFS setup, and that was far nicer under ZFS. ZFS is also much nicer/faster when you need to export block level storage, and for VM usage than BtrFS. Swap on ZFS works wonders, and even with very recent changes is still awful under BtrFS, which means needing LVM if you want a BtrFS volume and swap on the same group of disks, and that isn't great (BtrFS and ZFS are really designed to work best with storage that's under their complete management).

For home, I run exclusively BtrFS on the file storage part of my home file server (the OS lives on separate SSDs with LVM, BtrFS and swap), because of how much better it is on low performing hardware and dealing with mismatching drives that upgrade to different sizes at different times (usually when a drive fails, and I replace it with a bigger one, and scrub+rebalance). For my non-work/personal/small-scale needs, BtrFS is the clear winner, as ZFS is too inflexible.

Horses for courses, YMMV, etc. I'm very glad both exist.

Well, a) BTRFS only supports RAID1 mirrors and tended to eat your data if running RAID5/6

RAID5/6 for data is considered stable (with the caveat below), but not metadata. BtrFS makes it easy to mix different data and metadata profiles, so that's an option.

That caveat is that the RIAD5 "write hole" problem still exists (common to almost all commercial RAID controllers too, but not in ZFS due to clever design). Even with RAID1/10 profiles on metadata, a full scrub is recommended on unclean shutdown if you run RAID5/6 on data (or invest in a good UPS with connectivity to your server, and tools to clean shutdown on a detected power failure).

On my small home file server, there's 6TB of data, and 13GB of metadata. So at around 2%, having your metadata as RAID1 instead of RAID5/6 isn't too horrible as a work around.

https://btrfs.wiki.kernel.org/index.php/Status#RAID56

https://btrfs.wiki.kernel.org/index.php/RAID56

As of a December 2019, there's also now raid1c3 and raid1c4 profiles (i.e.: 3 or 4 copies instead of just 2):

https://btrfs.wiki.kernel.org/index.php/Changelog#btrfs-progs_v5.4_.28Dec_2019.29

b) RedHat has removed it from the release branch and is no longer developing it.

RedHat hired a bunch of XFS developers who... wait for it... said that XFS was better and dumped BtrFS. Colour me surprised.

On top of that, they're now attempting a rather convoluted combination of LVM and XFS (with XFS adding features like checksums and CoW) to try and make an equivalent file system, all in the name of "getting it to market faster". I can't say I'm impressed with what's on offer yet, but I'm willing to have my mind changed if they can combine the performance, caching, block storage exporting and swap goodness of ZFS with the flexibility and on-the-fly reconfiguration of BtrFS. (I would have instead liked to see them invest in improving BtrFS, but again, let's see where they land).

https://opensource.com/article/18/4/stratis-lessons-learned

https://stratis-storage.github.io/

[boxie]

ZFS is mature, stable and doesn't eat your data.

Yeah - and Apple Macs can't get Virus's, X is secure and Santa Clause and the Easter bunny love having brunch together.

ZFS *has* eaten data and the recovery effort involved was quite amazing.

[k1e0x]

Check out one of the comments:


Anyway, regardless of that, it's fundamentally missing the point if you think that matters one way or the other. The devs who wrote this stuff don't decide if something is legal. Nobody at Sun even decides it anymore. Oracle's lawyers and Larry Ellison are the only one's opinions who matter. Given the details of their java lawsuit vs Google, you'd be insane to believe they wouldn't choose to sue over ZFS if they decide they can make money from it.

That said, they'll never sue any individuals so feel free to go nuts and use ZFS all you want personally.

No, but they do get to decide if it's moral or not.. and the creators of this technology want it in Linux by their own words.

Bryan illustrates that Sun couldn't use the GPL for Solaris because they needed closed source driver compatibility. That is the reason they chose the MPL as a model, not for GPL incompatibility. That is not revisionist it's a fact that the GPL wouldn't work for OpenSolaris.

So.
1. They can't use GPL and they chose MPL based. This is a good thing. MPL is a good license.
2. They wrote a license that express it's ability to be mixed with both other open source license and proprietary license. This is a good thing too.
3. They say they want ZFS and DTrace in Linux. This is a good thing, more technology! Yay!

Where do you get the conclusion they are evil?

I want Oracle to sue. That would be the best thing ever. I want Oracle to lose.

[jo-erlend]

Anyway, regardless of that, it's fundamentally missing the point if you think that matters one way or the other. The devs who wrote this stuff don't decide if something is legal. Nobody at Sun even decides it anymore. Oracle's lawyers and Larry Ellison are the only one's opinions who matter. Given the details of their java lawsuit vs Google, you'd be insane to believe they wouldn't choose to sue over ZFS if they decide they can make money from it.

But you're wrong. What Oracle's lawyers or Larry Ellison feels, is irrelevant. It's what the license says that's important. For the same reason, the license can be incompatible with GPL even if that wasn't the intention; it doesn't matter what the intention was or how the current owner feels about it now.

[k1e0x]

RAID5/6 for data is considered stable (with the caveat below), but not metadata. BtrFS makes it easy to mix different data and metadata profiles, so that's an option.

That caveat is that the RIAD5 "write hole" problem still exists (common to almost all commercial RAID controllers too, but not in ZFS due to clever design). Even with RAID1/10 profiles on metadata, a full scrub is recommended on unclean shutdown if you run RAID5/6 on data (or invest in a good UPS with connectivity to your server, and tools to clean shutdown on a detected power failure).

On my small home file server, there's 6TB of data, and 13GB of metadata. So at around 2%, having your metadata as RAID1 instead of RAID5/6 isn't too horrible as a work around.

So.. PSA for everyone..

You can do whatever you want at home.. but stop using Raid 6 at work. Both 5 and 6 are depreciated due to the size of the volumes and latent data errors occurring during resilver. You need triple parity at least (Raid 7.3).

https://queue.acm.org/detail.cfm?id=1670144

(Adam Leventhal is also a ZFS developer)

[k1e0x]

Also DDN has a commercial storage device based on ZFS. Where is the Oracle lawsuit? DDN isn't small.. And they are a larger storage vendor than Oracle themselves.

[Farmer]

From memory so some of this might be wrong. Most of it maybe.
Reiser and his team created their file system without using some kernel features. Getting it adopted was a long slow road as they needed to use more of the file system stuff Linux provided. Christoph Hellwig did most of the review. There were personality clashes and some competition stuff at the time as Hellwig had been on the XFS team. After Reiser was accepted he moved on to the "new big thing" and didn't want to support the older ReiserFS any longer. That "get it accepted into mainline and walk away" likely has made them gun shy. Maintenance was taken over by SUSE. Mainly Chris Mason.

Mason is behind BTRFS.

Mason works at Oracle currently if I remember.

Torvalds not wanting a file system which has been orphaned is logically due to history. Reiserfs. I'm aware that a laundry list of other systems have been mainlined, then abandoned, but with regards to file systems Reiser was a politically vicious infight.

I also get the impression that XFS was designed around RAID whereas EXT, Reiser, and perhaps BTRFS were really designed around a single physical hard drive. When flash drives entered the fray those are really internally very RAIDish. XFS came back in spades. That is when Red Hat dumped BTRFS and focused on getting XFS further along.

BTRFS was never Red Hat. It was SUSE. Chris Mason.

[smitty3268]

No, but they do get to decide if it's moral or not.. and the creators of this technology want it in Linux by their own words.

Wherever did you get the silly idea that morality came into play about legal matters?

Bryan illustrates that Sun couldn't use the GPL for Solaris because they needed closed source driver compatibility.

Sure, but they could have easily used a BSD license. I will grant you there could have been many different reasons they chose to go the MPL route, and not being in that room none of us will ever know for sure. However, I do know someone who was in that room states that being GPL-incompatible was one of the primary reasons behind it. Maybe they are lying, maybe not. In the end it really doesn't matter.

[smitty3268]

But you're wrong. What Oracle's lawyers or Larry Ellison feels, is irrelevant. It's what the license says that's important. For the same reason, the license can be incompatible with GPL even if that wasn't the intention; it doesn't matter what the intention was or how the current owner feels about it now.

Oh, if only that's the way our legal system really worked.

[k1e0x]

Sure, but they could have easily used a BSD license. I will grant you there could have been many different reasons they chose to go the MPL route, and not being in that room none of us will ever know for sure.

If you are a GPL proponent then you should applaud Sun's decisions to go copyleft over permissive. No?