Apple kicks ZFS in the butt
blogs.zdnet.com
blogs.zdnet.com
* maybe with ZFS, there were problems like slow speed, unreliability, poor resource use, etc... more speculation of Apple's testing benchmarks...
* maybe without ZFS, no problems were found in testing. If this was the case, and remember I am speculating, I would ship without ZFS, easy.
What I would really like to see, though, is some modern filesystem with a license that is neither offensive to a commercial vendor like Apple or Microsoft, nor offensive to the GNU crowd, or the BSD crowd. I'm not sure that's even possible, but it would be nice to know that I could format a flash drive in a universal way that isn't FAT32.
If your talking about writing something and being sure to be able to read it at some future time UFS is read-compatible pretty much everywhere.
Microsoft will never, ever support a technology they don't control. FAT32 is just fine by them and you can't count them out when you say "universal".
They could, of course, use BSD licensed stuff, but they would not be able to put pressure and extract licensing fees like they do with FAT.
ZFS isn't quite what's billed to be, but it's (mostly) better than every other current choice. You can read the positives everywhere else, but let me point out some of the negatives.
* Space wastage: block sizes for files grow up to 128K, but they never shrink, and the last block in a file is currently always the same as the overall file block size. This means that files e.g. 129K in size will have 127K of wasted space on disk, and files that were once 128K and truncated back down to 1 byte will still have 128K block size. Depending on the kinds of files, you can see space wastage rates in the region of 25-45%. This isn't theoretical: I had to reduce max block size to 8K on one filesystem to avoid huge wastage.
* Streaming latency: as a media server, I often play movies from data stored on this server, but there's a problem somewhere in the stack (network, OS file cache, filesystem implementation). If I'm playing a movie or TV episode from a ripped DVD (i.e. VIDEO_TS directory) I can guarantee that there'll be at least one glitch in every 40 minutes of playback, where network transfer has dropped to zero for 3 or 4 seconds.
* Device removal: you can add more storage to a ZFS pool, and one can replace any given device with a drive equal or larger in size, but it's not possible to remove a device. If you start out with 8 disks, you're stuck with 8 disks unless you want to backup and restore the entire FS - but at least FS streaming is easy with e.g. zfs send <snapshot> | ssh <remote> zfs receive <filesystem>.
* Fragmentation: there is no defragmentation solution currently. Sure, people say it's "not an issue", that "ZFS is not pathalogical", but the same was said about NTFS, and NTFS fragments like crazy when small incremental writes are used. ZFS performance is known to fall off a cliff (e.g. less than 1% of prior perf) as you approach capacity (symptomatic of fragmented free space) and, more importantly, stay abysmal until you back off quite a bit and preferably drop a whole file system or two. Recommendation is to stay under 80%, which isn't very full.
* Performance: performance is a bit of a red herring; you can make ZFS as fast as you want by adding mirroring and striping. The more interesting case is when you don't have mirrors and stripes out the wazoo, and are relying on RAIDZ or RAIDZ2, and here the performance depends on access patterns. The way the parity and checksumming works, you need to burn a lot of cache to get decent small random read/write performance, because entire blocks need to get read and written even for small updates. So performance looks great for a while, then falls off.
I still think ZFS is one of the best choices for this kind of application - the software RAID, multiple-volume handling is terrific - but it's far from the last word, and not 100% baked (needs defragmentation and data recovery tool ecosystem). BTRFS looks like it could be a very worthy contender, but from what I've read, I haven't seen that it values a storage-unit/pool/filesystem stack, where storage-unit can be a disk, a file or a partition, and still seems stuck on the old device/filesystem approach.
http://www.cluboverclocker.com/reviews/cases/gigabyte/aurora...
HDs are both in the internal HD bays, which have a dedicated fan, and in 3.5/5.25 adapter in the 5.25 bays up top, with active cooling, it's something like this:
That's all that really matters to me.
My media system is using ZFS (currently 98.7% full), and I keep a small zpool in my backpack with a bunch of data on it I don't want wasting space on my SSD.
There's nothing else I trust my data on at this point.
I have a FreeBSD machine with a raidz for all my other data (databases, mail, a copy of my music, my photo album, etc...) It was RAID5 before that and it was pretty terrible.
My current plan is to stick with ZFS until something better comes along, and there isn't all that much better showing up in this neighborhood. That does mean my media system can't run Snow Leopard, though. :(
It's more than likely that ZFS meant biting off more than they could chew, so they deferred it for another release.
1. Sun's shared source license sucks and this is why ZFS isn't natively supported in Linux. 2. btrfs will soon have better features/performance/stability than ZFS, and a better architecture to boot: http://lwn.net/Articles/342892/
I've tried it out on OpenSolaris a few times in an attempt to learn why it's so hyped up. The setup process is a bit cryptic and confusing. Even after the initial setup I was really confused about what commands were destructive and which ones were not. Based on that alone I would not consider using ZFS yet. The flexibility of ZFS was a bit lost on me. It seems like there's a ton of limitations and caveats to consider. I can't really comment on performance or reliability since this was a short lab test but in the last 5 years or so I've never lost a HFS+, EXT3 or NTFS file system so I'm not sure just how much more reliable ZFS can be. Is all this hype just acronym lust? I feel like a good RAID card combined with a semi-modern FS is still a better solution. It's certainly easier to setup IMO.
On top of that, ZFS stores all data along with checksums, so when you read a file it will pick up errors and correct them. How do you know the photos you haven't accessed recently are OK? That they will not be corrupt? You don't - you assume that if the disk shows up and the directory list shows up then they will be there. Wih ZFS, you set up scheduled scrubs and it verifies all the data is present, readable and correct - and because of this regular check, it picks up hardware failures more quickly.
Also, it has instant (well, small constant time) snapshots, which doubles as a time-machine/filesystem versioning system so you can see files as they were, and you can clone snapshots to take backups / send them over a network to keep a clone in sync.
Also, it was going to be Zettabyte File System (rumour) because it was designed to take filesystem size limits out of consideration.
Did I mention it writes data, then updates a pointer to the data in one operation so as long as your disks /controller do not cheap out and misrepresent what they are doing, data is always in the old or the new state - never half written.
Also, because of the lack of RAID controller, you can move the disks to another system and attach the zpool with leas requirements in an emergency.
That's pretty awesome - not just acronyms and buzzwords.
I'm amazed you found set-up cryptic. Once I understood the concepts of ZFS pools and filesystems, I found the toolset among the most elegant and well-designed out there.
I think it might actually be a little daunting in it's simplicity, maybe. ("that's it?!")
If you have never lost a file system, I'm guessing you haven't had very many.
Also, ZFS gracefully handles cheap disk. RAID cards don't handle the case where you are using consumer grade disk and it retries for hours rather than failing. (I recently tested this on a rather nice 3ware 9550SX. It handled the problem no better than linux md. Performance was worse on the 3ware, too. I'm sending it back.)
I am personally switching to 'enterprise' (read, 20% more cost for 30% less capacity) sata just so I get drives that actually fail rather than retrying when they are failed.
And this is ignoring the 'silent corruption' issue the CERN guys were on about. Block level checksumming fixes that too.
Also, cheap snapshots are awesome. LVM snapshots work, and they can save your ass in a pinch, but they are expensive performance-wise. In my case, where I'm pushing my disk to it's performance limits already, LVM snapshots are nearly unusable. ZFS snapshots, I understand, have almost zero performance overhead. (space is cheap. Performance, not so much.)
Possibly there are other good reasons to buy the enterprise models, but using the cheaper drives is working out for me. I'm running 6 WD15EADS drives with TLER disabled in a software RAID6 configuration on a GNU/Linux machine, and except for the fact that two of the drives failed within hours of first use (the replacement drives work fine), I've had no problems running the array 24/7 for over a month now.
As for the relation to ZFS, it is supposed to timeout ill-behaving consumer drives like this. In fact, from what I understand, you can configure ZFS's behavior, for example, to fail the first drive that times out, and then hang on the second drive (which would perhaps be a reasonable configuration for a ZRAID volume)
I believe Time Machine is still built on top of ZFS snapshot though. So obviously some ZFS code remains.
Edit: Apparently Time Machine is not based on ZFS. So it seems ZFS is completely off the map after all.
P.S. Woops, iPhone + airport == spotty internet.
ZFS contains a lot of features that sound like they overlap with Time Machine, but in practice they really don't in a way that would make Time Machine a "free" feature.
Steps: First run:
* Rsync data from source drive to backup drive (can be done over network, ssh, whatever). We'll call this backup "0"
Second run:
* On the backup drive, copy the last backup folder (backup "0") to a new backup folder (backup "1"), using hardlinks. Hardlink copies don't take any additional space, since they're pointing to the original resource on the disk (folders do add some negligible space).
* Rsync data from source drive to backup drive (backup "1"). Rsync, of course, only updates what's new/changed.
Repeat, saving as far back as you'd like, and deleting old backups if the drive is full, after N backups, or so many days. Now you've got an easy incremental backup bash script.
Hardlinks are beautiful, because they don't add extra space, and once all the hardlinks to a resource are deleted the resource becomes free space. This isn't an enterprise-level backup solution, but it's a great quick-and-dirty way to do incremental backups, and is exactly how Time Machine works.
But, yeah, same basic concept for sure.
Obviously the rest of Time Machine would not fit in one line, but allow me the right to exaggerate a bit.
I bet the timeline feature in Nautilus on OpenSolaris, while not a one-liner, must be quite simple an add-on.
Now, local snapshots would be useful for sure (and could be integrated into the TM system in various ways), but they're totally orthogonal to separate-disk backup/versioning systems like TM.
With ZFS, Time Machine would not require an extra disk and could maintain nearly continuous history of the whole file system even in single disk configurations. The minimum time between snapshots could be easily adjusted so that not much disk space is consumed in the process.
To say it would be "much better" ignores that one of the primary usages of Time Machine is that it provides easy whole-disk backup in case of catastrophe. In this scenario (the primary reason why many people use it), ZFS is completely irrelevant.
Now, perhaps one could imagine a future version of Time Machine that would read from source file system snapshots while making backups to the external target in order to capture versions that were made while the user was unplugged from the external disk; and perhaps ZFS would be appropriate for Time Machine target disks in order to save space via block copy-on-write. I'd say I expect both of these features at SOME point in the future ("expect" inasmuch as they're obvious wins and one has to think that the future direction of file systems is to include constant time low cost snapshots via copy-on-write), though I don't know if the file system(s) will actually be ZFS-derived or not.
But this is a far cry from saying that Time Machine functionality would be free on top of ZFS. Any future version of Time Machine will include replication to an external disk (or, at least, "the Cloud") and that's NOT a feature of ZFS as it stands right now.