Revision as of 02:48, February 28, 2014 by 404 Error (Talk)

WARNING: This tutorial is under a heavy revision to be switched from ZFS Fuse to ZFS on Linux.


ZFS features and limitations

ZFS offers an impressive amount of features even putting aside its hybrid nature (both a filesystem and a volume manager -- zvol) covered in detail on Wikipedia. One of the most fundamental points to keep in mind about ZFS is it targets a legendary reliability in terms of preserving data integrity. ZFS uses several techniques to detect and repair (self-healing) corrupted data, simply speaking it makes an aggressive use of checksums and relies on data redundancy, the price pay is it requires a bit more CPU processing power than traditional filesystems and RAID solution. However, the Wikipedia article about ZFS also mention it is strongly discouraged to use ZFS over classic RAID arrays as it can not control the data redundancy,thus ruining most of its benefits.

In short, ZFS has the following features (not exhaustive):

  • Storage pool (if you are used to BTRFS volumes should be familiar)
  • Plenty of space:
    • 256 zettabytes per storage pool (2^64 storages pools max in a system).
    • 16 exabytes max for a single file
    • 2^48 entries max per directory
  • Virtual block-devices support support over a ZFS pool (zvol) - (extremely cool when jointly used over a RAID-Z volume)
  • Read-only Snapshot support (it is possible to get a read-write copy of them, those are named clones)
  • Encryption support (supported only at ZFS version 30 and upper, ZFS version 31 is shipped with Oracle Solaris 11 so that version is mandatory if you plan to encrypt your ZFS datasets/pools)
  • Built-in RAID-5-like-over-steroid capabilities known as RAID-Z and RAID-6-like-over-steroid capabilities known as RAID-Z2. RAID-Z3 (triple parity) also exists.
  • Copy-on-Write transactional filesystem
  • Meta-attributes support (properties) allowing you to you easily drive the show like "That directory is encrypted", "that directory is limited to 5GiB", "That directory is exported via NFS" and so on. Depending on what you define, ZFS takes the appropriates actions!
  • Dynamic striping to optimize data throughput
  • Variable block length
  • Data deduplication
  • Automatic pool re-silvering
  • Transparent data compression / encryption (later requires Solaris 11)

Most notable limitations are:

  • Lack a features ZFS developers knows as "Block Pointer rewrite functionality" (planned to be developed), without it ZFS suffers of currently not being able to:
    • Pool defragmentation (COW techniques used in ZFS mitigates the problem)
    • Pool resizing
    • Data compression (re-applying)
    • Adding an additional device in a RAID-Z/Z2/Z3 pool to increase it size (however, it is possible to replace in sequence each one of the disks composing a RAID-Z/Z2/Z3)
  • No data healing if used on a single device (corruption can still be detected), workaround if to force a data duplication on the drive
  • No support of TRIMming (SSD devices)

ZFS on well known operating systems


Despite the source code of ZFS is open, its license (Sun CDDL) is incompatible with the license governing the Linux kernel (GNU GPL v2) thus preventing its direct integration. However a couple of ports exists, but suffers of maturity and lack of features. As of writing (February 2014) two known implementations exists:

  • ZFS-fuse: a totally userland implementation relying on FUSE. This implementation can now be considered as defunct as of February 2014). The original site of ZFS FUSE seems to have disappeared nevertheless the source code is still available on http://freecode.com/projects/zfs-fuse. ZFS FUSE stalled at version 0.7.0 in 2011 and never really evolved since then.
  • ZFS on Linux: a kernel mode implementation of ZFS in kernel mode which supports a lot of NFS features. The implementation is not as complete as it is under Solaris and its siblings like OpenIndiana (e.g. SMB integration is still missing, no encryption support...) but a lot of functionality is there. This is the implementation used for this article. As ZFS on Linux is an out-of-tree Linux kernel implementation, patches must be waited after each Linux kernel release. ZfsOnLinux currently supports zpools version 28.


  • Oracle Solaris: remains the de facto reference platform for ZFS implementation: ZFS on this platform is now considered as mature and usable on production systems. Solaris 11 uses ZFS even for its "system" pool (aka rpool). A great advantage of this: it is now quite easy to revert the effect of a patch at the condition a snapshot has been taken just before applying it. In the "old good" times of Solaris 10 and before, reverting a patch was possible but could be tricky and complex when possible. ZFS is far from being new in Solaris as it takes its roots in 2005 to be, then, integrated in Solaris 10 6/06 introduced in June 2006.
  • OpenIndiana: is based on the Illuminos kernel (a derivative of the now defunct OpenSolaris) which aims to provide absolute binary compatibility with Sun/Oracle Solaris. Worth mentioning that Solaris kernel and the Illumos kernel were both sharing the same code base, however, they now follows a different path since Oracle announced the discontinuation of OpenSolaris (August 13th 2010). Like Oracle Solaris, OpenIndiana uses ZFS for its system pool. The illumos kernel ZFS support lags a bit behind Oracle: it supports zpool version 28 where as Oracle Solaris 11 has zpool version 31 support, data encryption being supported at zpool version 30.


  • FreeBSD: ZFS is present in FreeBSD since FreeBSD 7 (zpool version 6) and FreeBSD can boot on a ZFS volume (zfsboot). ZFS support has been vastly enhanced in FreeBSD 8.x (8.2 supports zpool version 15, version 8.3 supports version 28), FreeBSD 9 and FreeBSD 10 (both supports zpool version 28). ZFS in FreeBSD is now considered as fully functional and mature. FreeBSD derivatives such as the popular FreeNAS takes befenits of ZFS and integrated it in their tools. In the case of that latter, it have, for example, supports for zvol though its Web management interface (FreeNAS >= 8.0.1).
  • NetBSD: ZFS has been started to be ported as a GSoC project in 2007 and is present in the NetBSD mainstream since 2009 (zpool version 13).
  • OpenBSD: No ZFS support yet and not planned until Oracle changes some policies according to the project FAQ.

ZFS alternatives

  • WAFL seems to have severe limitation [1] (document is not dated), also an interesting article lies here
  • BTRFS is advancing every week but it still lacks such features like the capability of emulating a virtual block device over a storage pool (zvol) and built-in support for RAID-5/6 is not complete yet (cf. Btrfs mailing list). At date of writing, it is still experimental where as ZFS is used on big production servers.
  • VxFS has also been targeted by comparisons like this one (a bit controversial). VxFS has been known in the industry since 1993 and is known for its legendary flexibility. Symantec acquired VxFS and proposed a basic version (no clustering for example) of it under the same Veritas Storage Foundation Basic
  • An interesting discussion about modern filesystems can be found on OSNews.com


BTRFS and ZFS are sib in their concepts and of course have differences:

  • both are transactional filesystems (in BTRFS a transaction is a sequence of low level operations)
  • both implement for example the pool concept (called a "volume" in BTRFS)
  • both can do snapshots although in ZFS a snapshot is a read only thing and its attributes can't be modified. BTRFS on the other hand has writable snapshots (known as clones in ZFS)
  • both can organize their storage pool in several logical divisions (called datasets in ZFS and subvolumes in BTRFS).
  • As their equivalent in BTRFS (subvolumes), ZFS datasets appears as directories
  • Where as a ZFS snapshot is "hidden" in a sub-directory (named .zfs), BTRFS snapshots appears as visible directories
  • While ZFS manages rollback in a transparent manner (the filesystem knows where and how to rollback the data), rolling back data in BTRFS requires a bit more work as the system administrator must umount/remount a BTRFS subvolume.
  • ZFS has a kind of sophisticated RAID-5 called RAID-Z (and now RAID-Z2 ~ RAID-6), similar capabilities are planned for BTRFS but not yet available as of February 2014 because of some work has to be done on parity logging (see [2]).
  • Both supports the concept of sending and receiving a snapshot of dataset/subvolume, however those are not compatible (a BTRFS snapshot cannot be restored as a ZFS dataset and vice-versa)
  • Whereas ZFS makes an aggressive use of properties to govern the behaviour of the different datasets (quotas, sharing over NFS, encryption, compression and so on), BTRFS does not use this notion or in a much light manner and only through the mount command.
  • ZFS has no journal (!), this is not a design flaw but an interesting intrinsic feature :) See page 7 of "ZFS The last word on filesystems".
  • While BTRFS subvolumes data can be defragmented, ZFS, even in Solaris/OpenIndiana, has no such capability nevertheless ZFS has deduplication capabilities. Deduplication is possible with BTRFS but not with its "stock" userland software, a third party tool known as bedup is required.

ZFS resource naming restrictions

Before going further, you must be aware of restrictions concerning the names you can use on a ZFS filesystem. The general rule is: you can can use all of the alphanumeric characters plus the following specials are allowed:

  • Underscore (_)
  • Hyphen (-)
  • Colon (:)
  • Period (.)

The name used to designate a ZFS pool has no particular restriction except:

  • it can't use one the reserved words in particular:
    • mirror
    • raidz (raidz2, raidz3 and so on)
    • spare
    • cache
    • log
  • names must begin with an alphanumeric character (same for ZFS datasets).

Playing with ZFS


  • Kernel with FUSE stuff enabled
  • ZFS userland tools installed (sys-fs/zfs)
  • ZFS kernel modules built and installed (sys-fs/zfs-kmod), there is a known issue with kernel 3.13 series see this thread on Funtoo's forum
  • /etc/init.d/zfs started (automatically loads all required kernel modules then detects and mounts all of your ZFS datasets...at least all of those not having their canmount property set to false)
  • Disk size of 64 Mbytes as a bare minimum (128 Mbytes is the minimum size of a pool). Multiple disk will be simulated through the use of several raw images accessed via the Linux loopback devices.
  • At least 512 MB of RAM

Your first ZFS pool

To start with, four raw disks (2 GB each) are created:

# for i in 0 1 2 3; do dd if=/dev/zero of=/tmp/zfs-test-disk0${i}.img bs=2G count=1; done
0+1 records in
0+1 records out
2147479552 bytes (2.1 GB) copied, 40.3722 s, 53.2 MB/s

Then let's see what loopback devices are in use and which is the first free:

# losetup -a
# losetup -f

In the above example nothing is used and the first available loopback device is /dev/loop0. Now associate all of the disks with a loopback device (/tmp/zfs-test-disk00.img -> /dev/loop/0, /tmp/zfs-test-disk01.img -> /dev/loop/1 and so on):

# for i in 0 1 2 3; do losetup /dev/loop${i} /tmp/zfs-test-disk0${i}.img; done
# losetup -a
/dev/loop0: [000c]:781455 (/tmp/zfs-test-disk00.img)
/dev/loop1: [000c]:806903 (/tmp/zfs-test-disk01.img)
/dev/loop2: [000c]:807274 (/tmp/zfs-test-disk02.img)
/dev/loop3: [000c]:781298 (/tmp/zfs-test-disk03.img)

Pool creation

It is now time to create our first ZFS data pool and this is accomplished by one of the two commands you have to retain: zfspool. For now, we will ask it to do a simple job: get all of the just created devices and create an aggregated pool:

# zfs create myfirstpool /dev/loop0 /dev/loop1 /dev/loop2 /dev/loop3
# mount
kstat on /zfs-kstat type fuse (rw,nosuid,nodev,allow_other)
myfirstpool on /myfirstpool type fuse (rw,allow_other,default_permissions)

Note that the pool has also been mounted on /myfirstpool! Forget kstat for now, it is mounted automatically by zfs-fuse and countains some performance statistics. Oh by the way, we have used block devices (loopback devices are block devices) to create our ZFS pool, however ZFS can also deal directly with files and the taxonomy used in the ZFS world retains the term vdev (virtual device). Let's be curious a bit and see what df reports:

# df -h
# myfirstpool                          7.9G   21K  7.9G   1% /myfirstpool

Cool! About 8GB are reported, this is barely the sum of our four vdevs minus some metadata. What can we do with 8 GB of free storage space? Copy some files in it of course!

Some file operations

# cp -a  /usr/src/linux-3.1-rc4 /myfirstpool
# df -h
myfirstpool                          7.9G  662M  7.2G   9% /myfirstpool
# cd /myfirstpool
# ls -l /myfirstpool
total 3
drwxrwxr-x 24 root root 56 Aug 29 08:41 linux-3.1-rc4
# ls -l /myfirstpool/linux-3.1-rc4
total 29
-rw-rw-r--  1 root root    18693 Aug 29 00:16 COPYING
-rw-rw-r--  1 root root    94790 Aug 29 00:16 CREDITS
drwxrwxr-x 94 root root      222 Aug 29 00:16 Documentation
-rw-rw-r--  1 root root     2464 Aug 29 00:16 Kbuild
-rw-rw-r--  1 root root      252 Aug 29 00:16 Kconfig
-rw-rw-r--  1 root root   200918 Aug 29 00:16 MAINTAINERS
-rw-rw-r--  1 root root    53537 Aug 29 00:16 Makefile
-rw-r--r--  1 root root   364907 Aug 29 08:41 Module.symvers
-rw-rw-r--  1 root root    17459 Aug 29 00:16 README
drwxrwxr-x 22 root root       41 Aug 29 08:41 sound
drwxrwxr-x  9 root root        9 Aug 29 00:16 tools
drwxrwxr-x  2 root root       11 Aug 29 08:38 usr
drwxrwxr-x  3 root root        3 Aug 29 00:16 virt
-rwxr-xr-x  1 root root 13126551 Aug 29 08:41 vmlinux
-rw-r--r--  1 root root 14771911 Aug 29 08:41 vmlinux.o
# make clean
# df -h
Filesystem                           Size  Used Avail Use% Mounted on
myfirstpool                          7.9G  444M  7.4G   6% /myfirstpool

In fact nothing magic, a ZFS pool is acting just like any other existing filesystem :)

Unmounting/remounting the pool

If ZFS behaves just like any other filesystem, can we unmount it?

# umount /myfirstpool
# mount | grep myfirstpool

No more /myfirstpool in our light of sight. So yes, it is possible to unmount a ZFS pool just like with any other filesystem. But... How can we remount it then? Simple! First check the list of all ZFS pools known by the system:

# zpool list
myfirstpool  7.94G   444M  7.50G     5%  1.00x  ONLINE  -

Then mount it again:

# zpool list
myfirstpool  7.94G   444M  7.50G     5%  1.00x  ONLINE  -
# zfs mount myfirstpool

Oh! Did you noticed? We used the zfs command instead of the zpool command. You will understand the reason of using zfs instead of zpool a bit later, for now just remember that zfs and zpool are the only two commands used to interact with the ZFS universe. Also note that zfs mount... is the one and only way to remount a ZFS pool in the VFS arborescence so you can't be confused or do errors.


The missing leading / ahead of myfirstpool is not a typo. When a pool is created, ZFS writes in the pool metadata where it must be mounted. Unless overridden, it is assumed that the pool is to be mounted directly under the VFS root in a mountpoint which has the same name of the pool.

Let's check what happened:

# mount | grep myfirstpool
myfirstpool on /myfirstpool type fuse (rw,allow_other,default_permissions)
# ls -l /myfirstpool
total 3
drwxrwxr-x 23 root root 33 Sep  4 18:18 linux-3.1-rc4

Everything is back again!

ZFS datasets

Just like your house is a kind of big container subdivided in many others container (rooms), a ZFS pool can be divided in several logical containers known as datasets. Basically, the role of a dataset is to fullfill the so well known adage divide and conquer as they define the frontiers where all ZFS operations take place: it is only possible, for example, to take a snapshot/do a rollback of a dataset taken at whole.

Creating and destroying datasets

Creating a dataset in a pool is pretty easy to achieve: you invoke the zfs command, you give it the name of the pool to divide and the name of the dataset to create. To create three datasets named myfirstDS, mysecondDS, mythirdDS in myfirstpool(again the missing / ahead of myfirstpool is not a typo) :

# zfs create myfirstpool/myfirstDS
# zfs create myfirstpool/mysecondDS
# zfs create myfirstpool/mythirdDS
# ls -l /myfirstpool
total 7
drwxrwxr-x 23 root root 33 Sep  4 18:18 linux-3.1-rc4
drwxr-xr-x  2 root root  2 Sep  4 23:34 myfirstDS
drwxr-xr-x  2 root root  2 Sep  4 23:34 mysecondDS
drwxr-xr-x  2 root root  2 Sep  4 23:34 mythirdDS

Datasets are appearing just as if they were regular directories. Are they? Try to remove one of those:

# rmdir /myfirstpool/myfirstDS
rmdir: failed to remove `/myfirstpool/myfirstDS': Device or resource busy

This behavior is absolutely normal, datasets are special entities and must be managed via ZFS commands. Trouble: how a regular directory with files opened by a running process can be distinguished from a ZFS dataset? Both looks similar! Here again, the zfs command rescues us:

# zfs list
myfirstpool              444M  7.38G   444M  /myfirstpool
myfirstpool/myfirstDS     21K  7.38G    21K  /myfirstpool/myfirstDS
myfirstpool/mysecondDS    21K  7.38G    21K  /myfirstpool/mysecondDS
myfirstpool/mythirdDS     21K  7.38G    21K  /myfirstpool/mythirdDS

Not obvious but zfs list also reveals you a great secret: we lied you in the previous paragraphs. It it not possible to mount a ZFS pool in the VFS arborescence as only datasets can be mounted. So where is the prank? Our myfirstpool had been mounted in the VFS and you never defined any datasets in it. How is that possible? Is there some ZFS black magic lying behind? No. When you created the ZFS pool myfirstpool, a special dataset had also been created in the pool automatically for you: the root dataset. When you typed zfs mount mypool, you had in fact interact with this root dataset and not with the pool in itself. The operation was transparent for you and you never noticed its presence although using the zfs command instead of zpool could have given you a hint about what lies under the hood. You see that root dataset in the first line of what zfs list reported in the example above.

So the root dataset (myfirstpool) is mounted on /myfirstpool, myfirstDS is then mounted inside (/myfirstpool/myfirstDS) ditto for mysecondDS and mythirdDS. Mounted is the exact term because if we have a look at what the mount command reports we can see that those datasets have been effectively mounted:

# mount
rootfs on / type rootfs (rw)
myfirstpool on /myfirstpool type fuse (rw,allow_other,default_permissions)
myfirstpool/myfirstDS on /myfirstpool/myfirstDS type fuse (rw,allow_other,default_permissions)
myfirstpool/mysecondDS on /myfirstpool/mysecondDS type fuse (rw,allow_other,default_permissions)
myfirstpool/mythirdDS on /myfirstpool/mythirdDS type fuse (rw,allow_other,default_permissions)

As we did before, we can copy some files in the newly created datasets just like they were regular directories:

# cp -a /usr/portage /myfirstpool/mythirdDS
# ls -l /myfirstpool/mythirdDS/*
total 438
drwxr-xr-x   45 root root      46 Aug 31 07:37 app-accessibility
drwxr-xr-x  202 root root     203 Sep  2 07:21 app-admin
drwxr-xr-x    3 root root       4 Aug 18 18:13 app-antivirus
drwxr-xr-x   93 root root      94 Aug 18 18:13 app-arch
drwxr-xr-x   38 root root      39 Aug 18 18:13 app-backup
drwxr-xr-x   30 root root      31 Aug 18 18:13 app-benchmarks
drwxr-xr-x   66 root root      67 Aug 18 18:13 app-cdr
drwxr-xr-x   96 root root      97 Aug 18 18:13 app-crypt
drwxr-xr-x  358 root root     359 Aug 18 18:13 app-dicts
# df -h | grep DS                
myfirstpool/myfirstDS                5.6G   21K  5.6G   1% /myfirstpool/myfirstDS
myfirstpool/mysecondDS               5.6G   21K  5.6G   1% /myfirstpool/mysecondDS
myfirstpool/mythirdDS                7.4G  1.9G  5.6G  25% /myfirstpool/mythirdDS

Notice what df returns: our four datasets shares (don't forget the root dataset!) shares the same storage capacity. Logical indeed: as they are all contained in the same pool they cannot exceed its own storage capacity. Is it possible to cap the maximum capacity of a dataset? Yes, for now just retain that datasets:

  1. are logical containers where ZFS operations take place
  2. are concerned at whole by ZFS operations (again: you cannot snapshot/rollback a particular directory located in a dataset, you can only operate at the dataset level)

We have three datasets, but the third is pretty useless and contains a lot of garbage. Is it possible to remove it with a simple rm -rf? Let's try:

# rm -rf /myfirstpool/mythirdDS
rm: cannot remove `/myfirstpool/mythirdDS': Device or resource busy

This is perfectly normal, remember that datasets are special entities that requires special care and they are not deletable through regular shell commands. However it is possible to destroy them and here again, the zfs command comes at our rescue:

# zfs destroy myfirstpool/mythirdDS
# zfs list
myfirstpool              444M  7.38G   444M  /myfirstpool
myfirstpool/myfirstDS     21K  7.38G    21K  /myfirstpool/myfirstDS
myfirstpool/mysecondDS    21K  7.38G    21K  /myfirstpool/mysecondDS

Et voila! No more third dataset. :)

A bit more subtle case: let's mythirdDS and put another nested one in it then try to destroy mythirdDS again:

# zfs create myfirstpool/mythirdDS
# zfs create myfirstpool/mythirdDS/nestedSD
# zfs list
NAME                             USED  AVAIL  REFER  MOUNTPOINT
myfirstpool                      444M  7.38G   444M  /myfirstpool
myfirstpool/myfirstDS             21K  7.38G    21K  /myfirstpool/myfirstDS
myfirstpool/mysecondDS            21K  7.38G    21K  /myfirstpool/mysecondDS
myfirstpool/mythirdDS             42K  7.38G    21K  /myfirstpool/mythirdDS
myfirstpool/mythirdDS/nestedDS    21K  7.38G    21K  /myfirstpool/mythirdDS/nestedDS
# zfs destroy myfirstpool/mythirdDS
cannot destroy 'myfirstpool/mythirdDS': filesystem has children
use '-r' to destroy the following datasets:

zfs tells us it has found some others datasets located in mythirdDS and, thus, is unable to delete it without you consent to make a recursive destruction (-r parameter). Before trying to destroy the dataset again let's create some more nested datasets plus a couple of directories inside mythirdDS:

# zfs create myfirstpool/mythirdDS/nestedSD
# zfs create myfirstpool/mythirdDS/nestedSD2
# zfs create myfirstpool/mythirdDS/nestedSD3
# mkdir /myfirstpool/mythirdDS/dir1
# mkdir /myfirstpool/mythirdDS/dir2
# mkdir /myfirstpool/mythirdDS/dir3
# zfs list
NAME                                USED  AVAIL  REFER  MOUNTPOINT
myfirstpool                         444M  7.38G   444M  /myfirstpool
myfirstpool/myfirstDS                21K  7.38G    21K  /myfirstpool/myfirstDS
myfirstpool/mysecondDS               21K  7.38G    21K  /myfirstpool/mysecondDS
myfirstpool/mythirdDS                84K  7.38G    21K  /myfirstpool/mythirdDS
myfirstpool/mythirdDS/mynestedDS     21K  7.38G    21K  /myfirstpool/mythirdDS/mynestedDS
myfirstpool/mythirdDS/mynestedDS2    21K  7.38G    21K  /myfirstpool/mythirdDS/mynestedDS2
myfirstpool/mythirdDS/mynestedDS3    21K  7.38G    21K  /myfirstpool/mythirdDS/mynestedDS3
# zfs destroy -r myfirstpool/mythirdDS

Now what happens if we try to destroy mythird again this time with '-r'?

# zfs destroy -r myfirstpool/mythirdDS        
cannot destroy 'myfirstpool/mythirdDS/mynestedDS': dataset is busy

This is not as exactly normal as it should and seems to be a bug in zfs-fuse, the expected behavior is to automatically unmount any dataset contained inside mythirdDS then destroy it including mythirdDS itself. The same kind of operation on a Solaris machine with a similar dataset structure gives:

# zfs list
NAME                              USED  AVAIL  REFER  MOUNTPOINT
rpool1/swap                      4.04G  23.2G   123M  -
testpool/test                    55.4K  3.76T  55.4K  /testpool/test
testpool/test/ds1                44.9K  3.76T  44.9K  /testpool/test/ds1
testpool/test/ds2                44.9K  3.76T  44.9K  /testpool/test/ds2
testpool/test/ds3                44.9K  3.76T  44.9K  /testpool/test/ds3
testpool/test2                   44.9K  3.76T  44.9K  /testpool/test2
# mkdir /testpool/test/dir1
# mkdir /testpool/test/dir2
# mkdir /testpool/test/dir1
# zfs destroy -r testpool/test
# zfs list
NAME                              USED  AVAIL  REFER  MOUNTPOINT
rpool1/swap                      4.04G  23.2G   123M  -
testpool/test2                   44.9K  3.76T  44.9K  /testpool/test2

To go back on ZFS Fuse, just do a few attempts and mythirdDS should vanish (you may also have to do an explicit zfs destroy mythirdDS at the end).

Snapshotting and rolling back a dataset

This is, by far, one of the coolest feature of ZFS: you can litterally take a photograph of a dataset, do whatever you want with the dataset then restore it in the exact same state just as if nothing had ever happened in the middle. To start with, let's copy some files in mysecondDS:

# cp -a /usr/portage /myfirstpool/mysecondDS
# ls /myfirstpool/mysecondDS/portage
total 200
drwxr-xr-x   45 root root      46 Aug 31 07:37 app-accessibility
drwxr-xr-x  202 root root     203 Sep  2 07:21 app-admin
drwxr-xr-x    3 root root       4 Aug 18 18:13 app-antivirus
drwxr-xr-x   93 root root      94 Aug 18 18:13 app-arch
drwxr-xr-x   57 root root      58 Aug 22 08:56 x11-wm
drwxr-xr-x   16 root root      17 Aug 18 18:13 xfce-base
drwxr-xr-x   54 root root      55 Aug 18 18:13 xfce-extra

Now, let's take a snapshot of mysecondDS. Because we manipulate a dataset and not the pool, we rely on the zfs command:

# zfs snapshot myfirstpool/mysecondDS@Charlie

The syntax is always pool/dataset@snapshot-name, the name of the snapshot is left at your discretion however you must use an at sign (@) to separate the snapshot name from the rest of the path.

After running that command,

# ls -la /myfirstpool/mysecondDS
total 9
drwxr-xr-x   3 root root   3 Sep  5 16:49 .
drwxr-xr-x   6 root root   6 Sep  5 15:43 ..
drwxr-xr-x 164 root root 169 Aug 18 18:25 portage

You were not thinking you would see something like @Charlie or Charlie lying in /myfirstpool/mysecondDS were you? Of course not, this is obvious ;-) Can zfs be of any help this time? It has rescued us several times in the past:

# zfs list 
NAME                               USED  AVAIL  REFER  MOUNTPOINT
myfirstpool                       2.27G  5.54G   444M  /myfirstpool
myfirstpool/myfirstDS               21K  5.54G    21K  /myfirstpool/myfirstDS
myfirstpool/mysecondDS            1.84G  5.54G  1.84G  /myfirstpool/mysecondDS

So where the heck is Charlie? And how on earth can we use it if *nothing* is visible to us. Again the answer is: zfs! This time we invoke it with the -t parameter set to 'all' meaning "list all dataset including snapshots":

# zfs list 
NAME                               USED  AVAIL  REFER  MOUNTPOINT
myfirstpool                       2.27G  5.54G   444M  /myfirstpool
myfirstpool/myfirstDS               21K  5.54G    21K  /myfirstpool/myfirstDS
myfirstpool/mysecondDS            1.84G  5.54G  1.84G  /myfirstpool/mysecondDS
myfirstpool/mysecondDS@Charlie      37K      -  1.84G  -

Notice that Charlie is not mounted and although mysecondDS holds near 2GB of data, Charlie takes only a couple of kilobytes in the dataset. This is the consequence of ZFS being a Copy-on-write filesystem, duplicating all of the data blocks is not required. They will be duplicated only when needed: when ZFS sense a change in a data block, it will create a copy of it thus leaving intact the datablock pointed by a snapshot. At the time they are taken, snapshots occupy very little space in the datasets however as the time goes on they tend to "stick"more and more data blocks to be in use. It is wise to delete snapshots when become not needed anymore.


OpenIndiana and Oracle Solaris supports an interesting feature not available in ZFS Fuse: a kind of secret door in the form of a virtual directory named .zfs (notice the dot ahead). "secret door" because it is really secret! You cannot see it even with ls -la, however .zfs is present in just any of your datasets and holds some very interesting clues:

# zfs list -t all
testpool/test2                     205K  3.76T  70.3K  /testpool/test2
testpool/test2@snap1                  0      -  70.3K  -
# cd /testpool/test2
# ls -la
total 22
drwxr-xr-x  11 root root 11 2011-09-05 17:34 .
drwxr-xr-x   6 root root  6 2011-09-05 16:13 ..
drwxr-xr-x   2 root root  2 2011-09-05 17:34 .sometest
drwxr-xr-x   2 root root  2 2011-09-05 17:34 .xyz
drwxr-xr-x   2 root root  2 2011-09-05 16:13 dir1
drwxr-xr-x   2 root root  2 2011-09-05 16:13 dir2
# cd /testpool/test2/.zfs
# pwd
# ls -l
ls -l
total 2
dr-xr-xr-x 2 root root 2 2011-09-05 16:13 shares
dr-xr-xr-x 3 root root 3 2011-09-05 17:19 snapshot
# cd snapshot
# ls -l
total 2
drwxr-xr-x 9 root root 9 2011-09-05 17:19 snap1
# cd snap1
# ls -l
total 22
drwxr-xr-x  11 root root 11 2011-09-05 17:34 .
drwxr-xr-x   6 root root  6 2011-09-05 16:13 ..
drwxr-xr-x   2 root root  2 2011-09-05 17:34 .sometest
drwxr-xr-x   2 root root  2 2011-09-05 17:34 .xyz
drwxr-xr-x   2 root root  2 2011-09-05 16:13 dir1
drwxr-xr-x   2 root root  2 2011-09-05 16:13 dir2

Despite you cannot change the snapshot contents, you can access it without having to roll it back to examine its contents. Extremely nifty design choice from the ZFS designers!

Now we have found Charlie, let's do some changes in the mysecondDS:

# rm -rf /myfirstpool/mysecondDS/portage
# echo "Hello, world" >  /myfirstpool/mysecondDS/hello.txt
# ls -l  /myfirstpool/mysecondDS
total 1
-rw-r--r-- 1 root root 13 Sep  5 18:07 hello.txt
# cat /myfirstpool/mysecondDS/hello.txt
Hello, world

Whooops...removing portage was not the best idea to have and we do not bother about hello.txt. We will have to move back at checkpoint Charlie!

# zfs rollback myfirstpool/mysecondDS@Charlie
# ls -l /myfirstpool/mysecondDS
total 6
drwxr-xr-x 164 root root 169 Aug 18 18:25 portage

Again, ZFS handled everything for you and you now have the contents of mysecondDS exactly as it was at the time the snapshot Charlie was taken. Not more complicated than that. Hang on you hat, we have not finished.

Dealing with several snapshots (time-traveling machine)

So far we only used a single snapshot just to keep things simple. However a dataset can hold several snapshots and moreover you can do a delta between two snapshots and nothing is really much more complicated than you have seen so far.

Let's consider myfirstDS this time. This dataset should be empty as we did nothing in it so far:

# ls -la /myfirstpool/myfirstDS
total 3
drwxr-xr-x 2 root root 2 Sep  4 23:34 .
drwxr-xr-x 6 root root 6 Sep  5 15:43 ..

Now generate some contents, take a snapshot (snapshot-1), add more content, take a snapshot again (snapshot-2), do some more modifications and take a third snapshot (snapshot-3):

# echo "Hello, world" >  /myfirstpool/myfirstDS/hello.txt
# cp /usr/src/linux-3.1-rc4.tar.bz2 /myfirstpool/myfirstDS
# ls -l /myfirstpool/myfirstDS
# ls -l /myfirstpool/myfirstDS
total 75580
-rw-r--r-- 1 root root       13 Sep  5 22:38 hello.txt
-rw-r--r-- 1 root root 77220912 Sep  5 22:38 linux-3.1-rc4.tar.bz2
# zfs snapshot myfirstpool/myfirstDS@snapshot-1
# echo "Goodbye, world" >  /myfirstpool/myfirstDS/goodbye.txt
# echo "Are you there?" >> /myfirstpool/myfirstDS/hello.txt
# cp /usr/src/linux-3.0.tar.bz2 /myfirstpool/myfirstDS
# rm /myfirstpool/myfirstDS/linux-3.1-rc4.tar.bz2
# zfs snapshot myfirstpool/myfirstDS@snapshot-2
# echo "Still there?" >> /myfirstpool/myfirstDS/goodbye.txt
# rm /myfirstpool/myfirstDS/hello.txt
# cp /proc/config.gz /myfirstpool/myfirstDS
# zfs snapshot myfirstpool/myfirstDS@snapshot-3
# zfs list -t all
# zfs list -t all
NAME                               USED  AVAIL  REFER  MOUNTPOINT
myfirstpool                       2.41G  5.40G   444M  /myfirstpool
myfirstpool/myfirstDS              147M  5.40G  73.3M  /myfirstpool/myfirstDS
myfirstpool/myfirstDS@snapshot-1  73.8M      -  73.8M  -
myfirstpool/myfirstDS@snapshot-2    20K      -  73.3M  -
myfirstpool/myfirstDS@snapshot-3      0      -  73.3M  -

Wow, nice demonstration on how a Copy-on-Write filesystem like ZFS works: what do we observe? First it is quite obvious to see that snapshot-1 is quite big. Is is possible that having a so big snapshot to be the consequence of removing /myfirstDS/linux-3.1-rc4.tar.bz2? Absolutely. Remember that a snapshot is a photograph of what a dataset contains at a given time, deleted information and unmodified original information is retained by the snapshot even you delete it from the dataset or bring in some changes to it. If you look again at the command history between snapshot-2 and snapshot-3, you will notice that we removed a small file and changed another small file a bit thus having a little information delta between what the dataset content at this time and what it also actually contains leading to a very small snapshot at the end. The third dataset is the exact copy of what the current dataset contains thus its size is very close to zero (truncated to zero on what you see).

$100 question: "How can I see what changed between snapshots?". Answer: yes, you can! Nuance is: ZFS Fuse does not support it yet :( Nevertheless here is what snapshots diffing looks like on an OpenIndiana/Solaris machine:

# zfs create testpool/test2
# cd /testpool/test2
# wget http://www.kernel.org/pub/linux/kernel/v3.0/testing/patch-3.1-rc4.bz2
# echo "Hello,world" > hello.txt
# zfs snapshot testpool/test2@s1

# rm patch-3.1-rc4.bz2
# echo 'Goodbye!' > goodbye.txt
# echo 'Still there?' >> hello.txt
# zfs snapshot testpool/test2@s2

# echo 'Hello, again' >> hello.txt
# ln -s goodbye.txt goodbye2.txt
# mv hello.txt hello-new.txt
# zfs snapshot testpool/test2@s3

# zfs list -t all | grep test2
testpool/test2                    8.49M  3.76T  47.9K  /testpool/test2
testpool/test2@s1                 8.41M      -  8.42M     -
testpool/test2@s2                 29.2K      -  46.4K     -
testpool/test2@s3                     0      -  47.9K     -
# zfs diff testpool/test2@s1 testpool/test2@s2
M       /testpool/test2/
-       /testpool/test2/patch-3.1-rc4.bz2
M       /testpool/test2/hello.txt
+       /testpool/test2/goodbye.txt

# zfs diff testpool/test2@s2 testpool/test2@s3
M       /testpool/test2/
R       /testpool/test2/hello.txt -> /testpool/test2/hello-new.txt
+       /testpool/test2/goodbye2.txt 

# zfs diff testpool/test2@s1 testpool/test2@s3
M       /testpool/test2/
-       /testpool/test2/patch-3.1-rc4.bz2
R       /testpool/test2/hello.txt -> /testpool/test2/hello-new.txt
+       /testpool/test2/goodbye.txt
+       /testpool/test2/goodbye2.txt

# zfs diff testpool/test2@s3 san/test2@s1
Unable to obtain diffs: 
   Not an earlier snapshot from the same fs

Where M,R,+,- stands for:

  • M: item has been modified
  • R: item has been renamed
  • +: item has been added
  • -: item has been removed

Observe the output of each diff and draw you own conclusion on what we did at each step and what appears in the diff. It is not possible to get a detailed diff similar to what Git and others gives but you have the big picture of what changed between snapshots.

If ZFS-Fuse does not implements (yet) a snapshot diffing capability, it can deal with several snapshots and is able to jump across several steps backwards. Suppose we want myfirstDS to go back exactly is was when we took the dataset photograph named snapshot-1:

# zfs rollback myfirstpool/myfirstDS@snapshot-1 
cannot rollback to 'myfirstpool/myfirstDS@snapshot-1': more recent snapshots exist
use '-r' to force deletion of the following snapshots:

This is not a bug, this is absolutely normal. The zfs command asks you to give it the explicit permission to remove the two others snapshots as they becomes useless (restoring them would be an absolute no sense) once snapshot-1 is restored. Second attempt:

# zfs rollback myfirstpool/myfirstDS@snapshot-1
# ls -l /myfirstpool/myfirstDS
total 75580
-rw-r--r-- 1 root root       13 Sep  5 22:38 hello.txt
-rw-r--r-- 1 root root 77220912 Sep  5 22:38 linux-3.1-rc4.tar.bz2
# zfs list -t all
NAME                               USED  AVAIL  REFER  MOUNTPOINT
myfirstpool                       2.34G  5.47G   444M  /myfirstpool
myfirstpool/myfirstDS             73.8M  5.47G  73.8M  /myfirstpool/myfirstDS
myfirstpool/myfirstDS@snapshot-1      0      -  73.8M  -
myfirstpool/mysecondDS            1.84G  5.47G  1.84G  /myfirstpool/mysecondDS
myfirstpool/mysecondDS@snapshot1    37K      -  1.84G  -

myfirstDS effectively returned to its state when snapshot-2 was taken and the snapshots snapshot-2 and snapshot-3 vanished.


You can leap several steps backward at the cost of loosing your subsequent modifications forever.

Streaming datasets over the network


Nothing in an infrastructure is as much critical as having reliable backups of the data used by an organization. Whereas a server can be remounted from scratch, the data it contains is very likely to be lost forever whenever a disaster occurs. Of course, as the data is the blood of an organization business processes, its integrity and confidentiality must be preserved in all cases.

You find ZFS snaphots useful? Well, you have seen just a small part of their potential. As a snapshot is a photograph of what a dataset contains frozen in the time, snapshots can be seen as being no more than a data backup. Like any backup, they must not stay on the local machine but must be put elsewhere and the common good sense tells to keep backups in a safe place, making them travel through a secure channel. By "secure channel" we intend something like a trusted person in your organization whose job consists of bringing a box of tapes off-site in a secure location but we also intend a secure communication channel like an SSH tunnel over two hosts without any human intervention.

ZSH designers had the same vision and made possible for a dataset to be able to be sent over a network. How is that possible? Simple: the process involves two peers who can use through a communication channel like the one established by netcat (OpenSSH supports a similar functionality but with an encrypted communication channel). For the sake of the demonstration, we will use two Solaris boxes at each end-point.

How stream some ZFS bits over the network? Here again, zfs is the answer. A nifty move from the designers was to use stdin and stdout as transmission/reception channels thus allowing great a flexibility in processing the ZFS stream. You can envisage, for instance, to compress your stream then crypt it then encode it in base64 then sign it and so on. It sounds a bit overkill but it is possible and in the general case you can use any tool that swallow the data from stdin and spit it through stdout in your plumbing.


The rest of this section has been done entirely on two Solaris 11 machines.

1. Sender side:

# zfs create testpool2/zfsstreamtest
# echo 'Hello, world!' > /testpool2/zfsstreamtest/hello.txt
# echo 'Goodbye, world' > /testpool2/zfsstreamtest/goodbye.txt
# zfs snapshot zfs testpool2/zfsstreamtest@s1
# zfs list -t snapshot
NAME                               USED  AVAIL  REFER  MOUNTPOINT
testpool2/zfsstreamtest@s1            0      -    32K           -

2. Receiver side (the dataset zfs-stream-test will be created and should not be present):

# nc -l -p 7000 | zfs receive testpool/zfs-stream-test

At this point the receiver is waiting after some data.

3. Sender side:

# zfs send testpool2/zfsstreamtest@s1 | nc 192.168.aaa.bbb.ccc 7000

4. Receiver side:

# zfs list -t snapshot
NAME                          USED  AVAIL  REFER 
testpool2/zfs-stream-test@s1       0      -  46.4K  -

Note that we did not set an explicit snapshot name in the second step but it could have been possible to choose anything else but the default which is the name of the snapshot sent over the network. In that case the dataset which will contain the snapshot needs to be created first:

# nc -l -p 7000 | zfs receive testpool/zfs-stream-test@mysnapshot01

Once received you would get:

# zfs list -t snapshot
NAME                                      USED  AVAIL  REFER 
testpool2/zfs-stream-test@mysnapshot01       0      -  46.4K  -

5. Just for the sake of the curiosity let's do a rollback on the receiver side:

# zfs rollback testpool2/zfsstreamtest@s1
# ls -l /testpool2/zfs-stream-test
total 2
-rw-r--r-- 1 root root 15 2011-09-06 23:54 goodbye.txt
-rw-r--r-- 1 root root 13 2011-09-06 23:53 hello.txt
# cat /testpool2/zfs-stream-test/hello.txt
Hello, world

Because ZFS streaming operates using the starnd input and output (stdin / stdout) you can build a bit more complex pipeline like:

# zfs send testpool2/zfsstreamtest@s1 | gzip | nc 192.168.aaa.bbb.ccc 7000

The above example was using two hosts but a simpler setup is also possible: you are not required to send you data over the network with netcat, you can store it to a regular file then mail it or store it on a USB key. By the way: we have not finished! We took only a simple case here: it is absolutely possible to do the exact same operation with the difference between snapshots (incremental). Just like an incremental backup takes only what has changed, ZFS can determine the difference between two snapshots and streaming instead of streaming a snapshot taken at whole. Although ZFS can detect and act on differentials, it does not operate (yet) at the block level: if only a few bytes of a very big file have changed, the whole file will be taken into consideration (operating at data block level is possible with some tools like the well-known rsync).

Consider the following:

  • A dataset snapshot (S1) contains two files:
    • A -> 10 MB
    • B -> 4 GB
  • A bit later some files (named C, D and E) are added to the dataset and another snapshot is (S2) taken. S2 contains:
    • A -> 10 MB
    • B -> 4 GB
    • C -> 3 MB
    • D -> 500 KB
    • E -> 1GB

With a full transfer of S2 A,B,C,D and E would be streamed whereas an incremental transfert (S2-S1), zfs would only process C, D and E. The next $100 question:"How can we stream a difference of snapshot? zfs again?" Yes! This time with a subtle difference: a special option specified on the command line telling it must use a difference rather than a full snapshot. Assuming a few more files are added in testpool2/zfsstreamtest dataset and a snapshot (s2) is has been taken, the delta between s2 and s1 (s2-s1) giving s3 can be send like this (on the receiver side the same as shown above is used, nothing special is required alos notice the presence of the -i option):

  • Sender:
# zfs send -i testpool2/zfsstreamtest@s1 testpool2/zfsstreamtest@s2 | nc 192.168.aaa.bbb.ccc 7000
  • Receiver:
# nc -l -p 7000 | zfs receive testpool/zfs-stream-test
# zfs list -t snapshot
testpool/zfs-stream-test@s1       28.4K      -  46.4K  -
testpool/zfs-stream-test@s2           0      -  47.1K  -

Note that although we did not specified any snapshot name to use on the receiver side, ZFS used by default the name of the second snapshot involved in the delta (s2 here).

$200 question: suppose we delete all of the received snapshots so far on the receiver side and we try to send the difference between s2 and s1, what would happen? ZFS will protest on the receiver side although no error message will be visible on the sender side:

cannot receive incremental stream: destination testpool/zfs-stream-test has been modified
since most recent snapshot

It is even worse if we remove the dataset used to receive the data:

cannot receive incremental stream: destination 'testpool/zfs-stream-test' does not exist

ZFS streaming over a network has no underlying protocol, therefore the sender just assumes the data has been successfully received and processed. It does not care whether a processing error occurs.

Govern a dataset by attributes

So far, most of a filesystem capabilities were driven by separate and scarced command line line tools (e.g. tune2fs, edquota, rquota, quotacheck...) which all have their own ways to handle tasks and can go through tricky ways sometimes especially the quota-related management utilities. Moreover, there was no easy way to handle a limitations on a directory rather than putting it a a dedicated partition or logical volume implying downtimes when additional space was to be added. Quota management is however one of the many facets disk space management includes.

In the ZFS world, many aspects are now managed by simply setting/clearing a property attached to a ZFS dataset through the now so well-known command zfs.You can, for example:

  • put a size limit on a dataset
  • reserve a space for dataset (that space is guaranteed to be available in the future although not being allocated at the time the reservation is made)
  • control if new files are encrypted and/or compressed
  • define a quota per user or group of users
  • control checksum usage => never turn that property off unless having very good reasons you are likely to never have (no checksums = no silent data corruption detection)
  • share a dataset by NFS/CIFS
  • control automatic data deduplication

Not all of a dataset properties are settable, some of them are set and managed by the operating system in the background for you and thus cannot be modified.


Solaris/OpenIndiana users: ZFS has a tight integration with the NFS/CIFS server, thus it is possible to share a zfs dataset by setting adequate attributes. ZFS on Linux (native kernel mode port) also has a tight integration with the built-in Linux NFS server, the same for ZFS fuse although still experimental. Under FreeBSD ZFS integration has been done both with NFS and Samba (CIFS).

Like any other action concerning datasets, properties are sets and unset via the zfs command. On our Funtoo box running zfs-Fuse we can, for example, start by seeing the value of all properties for the dataset myfirstpool/myfirstDS:

# zfs get all myfirstpool/myfirstDS
 zfs get all myfirstpool/myfirstDS 
NAME                   PROPERTY              VALUE                   SOURCE
myfirstpool/myfirstDS  type                  filesystem              -
myfirstpool/myfirstDS  creation              Sun Sep  4 23:34 2011   -
myfirstpool/myfirstDS  used                  73.8M                   -
myfirstpool/myfirstDS  available             5.47G                   -
myfirstpool/myfirstDS  referenced            73.8M                   -
myfirstpool/myfirstDS  compressratio         1.00x                   -
myfirstpool/myfirstDS  mounted               yes                     -
myfirstpool/myfirstDS  quota                 none                    default
myfirstpool/myfirstDS  reservation           none                    default
myfirstpool/myfirstDS  recordsize            128K                    default
myfirstpool/myfirstDS  mountpoint            /myfirstpool/myfirstDS  default
myfirstpool/myfirstDS  sharenfs              off                     default
myfirstpool/myfirstDS  checksum              on                      default
myfirstpool/myfirstDS  compression           off                     default
myfirstpool/myfirstDS  atime                 on                      default
myfirstpool/myfirstDS  devices               on                      default
myfirstpool/myfirstDS  exec                  on                      default
myfirstpool/myfirstDS  setuid                on                      default
myfirstpool/myfirstDS  readonly              off                     default
myfirstpool/myfirstDS  zoned                 off                     default
myfirstpool/myfirstDS  snapdir               hidden                  default
myfirstpool/myfirstDS  aclmode               groupmask               default
myfirstpool/myfirstDS  aclinherit            restricted              default
myfirstpool/myfirstDS  canmount              on                      default
myfirstpool/myfirstDS  xattr                 on                      default
myfirstpool/myfirstDS  copies                1                       default
myfirstpool/myfirstDS  version               4                       -
myfirstpool/myfirstDS  utf8only              off                     -
myfirstpool/myfirstDS  normalization         none                    -
myfirstpool/myfirstDS  casesensitivity       sensitive               -
myfirstpool/myfirstDS  vscan                 off                     default
myfirstpool/myfirstDS  nbmand                off                     default
myfirstpool/myfirstDS  sharesmb              off                     default
myfirstpool/myfirstDS  refquota              none                    default
myfirstpool/myfirstDS  refreservation        none                    default
myfirstpool/myfirstDS  primarycache          all                     default
myfirstpool/myfirstDS  secondarycache        all                     default
myfirstpool/myfirstDS  usedbysnapshots       18K                     -
myfirstpool/myfirstDS  usedbydataset         73.8M                   -
myfirstpool/myfirstDS  usedbychildren        0                       -
myfirstpool/myfirstDS  usedbyrefreservation  0                       -
myfirstpool/myfirstDS  logbias               latency                 default
myfirstpool/myfirstDS  dedup                 off                     default
myfirstpool/myfirstDS  mlslabel              off                     -

How can we set a limit that prevents myfirstpool/myfirstDS to not use more than 1 GB of space in the pool? Simple, just set the quota property:

# zfs set quota=1G myfirstpool/myfirstDS
# zfs get quota myfirstpool/myfirstDS
NAME                   PROPERTY  VALUE  SOURCE
myfirstpool/myfirstDS  quota     1G     local

May be something poked your curiosity: what "SOURCE" means? "SOURCE" describes how the property has been determined for the dataset and can have several values:

  • local: the property has been explicitly set for this dataset
  • default: a default value has been assigned by the operating system if not explicitely set by the system adminsitrator (e.g SUID allowed or not in the above example).
  • dash (-): not modifiable intrinsic property (e.g. dataset creation time, whether the dataset is currently mounted or not, dataset space usage in the pool, average compression ratio...)

Before copying some files in the dataset, let's fix a binary (on/off) property:

# zfs set compression=on myfirstpool/myfirstDS

Now try to put more than 1GB of data in the dataset:

# dd if=/dev/zero of=/myfirstpool/myfirstDS/one-GB-test bs=2G count=1
dd: writing `/myfirstpool/myfirstDS/one-GB-test': Disk quota exceeded

Permission delegation

ZFS brings a feature known as delegated administration. Delegated administration enables ordinary users to handle administrative tasks on a dataset without being administrators. It is however not a sudo replacement as it covers only ZFS related tasks such as sharing/unsharing, disk quota management and so on. Permission delegation shines in flexibility because such delegation can be handled by inheritance though nested datasets. Pewrmission deleguation is handled via zfs through its allow and disallow options.

Data redundancy with ZFS

Nothing is perfect and the storage medium (even in datacenter-class equipment) is prone to failures and fails on a regular basis. Having data redundancy is mandatory to help in preventing single-points of failure (SPoF). Over the past decades, RAID technologies were powerful however their power is precisely their weakness: as operating at the block level, they do not care about what is stored on the data blocks and have no ways to interact with the filesystems stored on them to ensure data integrity is properly handled.

Some statistics

It is not a secret to tell that a general trend in the IT industry is the exponential growth of data quantities. Just thinking about the amount of data Youtube, Google or Facebook generates every day taking the case of the first some statistics gives:

  • 24 hours of video is generated every minute in March 2010 (May 2009 - 20h / October 2008 - 15h / May 2008 - 13h)
  • More than 2 billions views a day
  • More video is produced on Youtube every 60 days than 3 major US broadcasting networks did in the last 60 years

Facebook is also impressive (Facebook own stats):

  • over 900 million objects that people interact with (pages, groups, events and community pages)
  • Average user creates 90 pieces of content each month (750 millions users active)
  • More than 2.5 million websites have integrated with Facebook

What is true with Facebook and Youtube is also true with many other cases (think one minutes about the amount of data stored in iTunes) especially with the growing popularity of cloud computing infrastructures. Despite the progress of the technology a "bottleneck" still exists: the storage reliability is nearly the same over the years. If only one organization in the world generate huge quantities of data it would be the CERN (Conseil Européen pour la Recherche Nucléaire, now officially known as European Organization for Nuclear Research) as their experiments can generate spikes of many terabytes of data within a few seconds. A study done in 2007 quoted by a ZDNet article reveals that:

  • Even ECC memory cannot be always be helpful: 3 double-bit errors (uncorrectable) occurred in 3 months on 1300 nodes. Bad news: it should be zero.
  • RAID systems cannot protect in all cases: monitoring 492 RAID controller for 4 weeks showed an average error rate of 1 per ~10^14 bits, giving roughly 300 errors for every 2.4 petabytes
  • Magnetic storage is still not reliable even on high-end datacenter class drives: 500 errors found over 100 nodes while writing 2 GB file to 3000+ nodes every 2 hours then read it again and again for 5 weeks.

Overall this means: 22 corrupted files (1 in every 1500 files) for a grand total of 33700 files holding 8.7TB of data. And this study is 5 years old....

Source of silent data corruption


Not an exhaustive list but we can quote:

  • Cheap controller or buggy driver that does not reports errors/pre-failure conditions to the operating system;
  • "bit-leaking": an harddrive consists of many concentric magnetic tracks. When the hard drive magnetic head writes bits on the magnetic surface it generates a very weak magnetic field however sufficient to "leak" on the next track and change some bits. Drives can generally, compensate those situations because they also records some error correction data on the magnetic surface
  • magnetic surface defects (weak sectors)
  • Hard drives firmware bugs
  • Cosmic rays hitting your RAM chips or hard drives cache memory/electronics

Building a mirrored pool



RAID-5 is very commonly used nowadays because of its simplicity, efficiency and fault-tolerance. Although the technology did its proof over decades, it has a major drawback known as "The RAID-5 write hole". if you are familiar with RAID-5 you already know that is consists of spreading the stripes across all of the disks within the array and interleaving them with a special stripe called the parity. Several schemes of spreading stripes/parity between disks exists in the natures, each one with its own pros and cons, however the "standard" one (also known as left-asynchronous) is:

Disk_0  | Disk_1  | Disk_2  | Disk_3
[D0_S0] | [D0_S1] | [D0_S2] | [D0_P]
[D1_S0] | [D1_S1] | [D1_P]  | [D1_S2]
[D2_S0] | [D2_P]  | [D2_S1] | [D2_S2]
[D2_P]  | [D2_S0] | [D2_S1] | [D2_S2]

The parity is simply computed by XORing the stripes of the same "row", thus giving the general equation:

  • [Dn_S0] XOR [Dn_S1] XOR ... XOR [Dn_Sm] XOR [Dn_P] = 0

This equation can be rewritten in several ways:

  • [Dn_S0] XOR [Dn_S1] XOR ... XOR [Dn_Sm] = [Dn_P]
  • [Dn_S1] XOR [Dn_S2] XOR ... XOR [Dn_Sm] XOR [Dn_P] = [Dn_S0]
  • [Dn_S0] XOR [Dn_S2] XOR ... XOR [Dn_Sm] XOR [Dn_P] = [Dn_S1]
  • ...and so on!

Because the equations are a combinations of exclusive-or, it is possible to easily compute a parameter if it is missing. Let say we have 3 stripes plus one parity composed of 4 bits each but one of them is missing due to a disk failure:

  • D0_S0 = 1011
  • D0_S1 = 0010
  • D0_S2 = <missing>
  • D0_P = 0110

However we know that:

  • D0_S0 XOR D0_S1 XOR D0_S2 XOR D0_P = 0000 also rewritten as:
  • D0_S2 = D0_S1 XOR D0_S2 XOR D0_P

Applying boolean algebra it gives: D0_S2 = 1011 XOR 0010 XOR 0110 = 1111. Proof: 1011 XOR 0010 XOR 1111 = 0110 this is the same as D0_P

'So what's the deal?' Okay now the funny part, forgot the above hypothesis and imagine we have this:

  • D0_S0 = 1011
  • D0_S1 = 0010
  • D0_S2 = 1101
  • D0_P = 0110

Applying boolean algebra magics gives 1011 XOR 0010 XOR 1101 => 0100. Problem: this is different of D0_P (0110). Can you tell which one (or which ONES) of the four terms lies? If you find a mathematically acceptable solution, found your company because you have just solved a big computer science problem. If humans can't solve the question, imagine how hard it is for the poor little RAID-5 controller to determine which stripe is right and which one lies and the resulting "datageddon" (i.e. massive data corruption on the RAID-5 array) when the RAID-5 controller detect error and start to rebuild the array.

This is not science fiction, this a pure reality and the weakness stays in the RAID-5 simplicity. Here is how it can happen: an urban legend with RAID-5 arrays is that they update stripes in an atomic transaction (all of the stripes+parity are written or none of them). Too bad, this is just not true, the data is written on the fly and if for a reason or another the machine where the RAID-5 array has a power outage or crash, the RAID-5 controller will simply have no idea about what he was doing and which stripes are up to date which ones are not up to date. Of course, RAID controllers in servers do have a replaceable on-board battery and most of the time the server they reside in is connected to an auxiliary source like a battery-based UPS or a diesel/gas electricity generator. However, Murphy laws or unpredictable hazards can, sometimes, happens....

Another funny scenario: imagine a machine with a RAID-5 array (on UPS this time) but with non ECC memory. the RAID-5 controller splits the data buffer in stripes, computes a data stripe and starts to write them on the different disks of the array. But...but...but... For some odd reason, only one bit in one of the stripes flips (cosmic rays, RFI...) after the parity calculation. Too bad too sad, one of the written stripes contains corrupted data and it is silently written on the array. Datageddon in sight!

Not to make you freaking: storage units have sophisticated error correction capability (a magnetic surface or an optical recording surface is not perfect and reading/writing error occurs) masking most the cases. However, some established statistics estimates that even with error correction mechanism one bit over 10^16 bits transferred is incorrect. 10^16 is really huge but unfortunately in this beginning of the XXIst century with datacenters brewing massive amounts of data with several hundreds to not say thousands servers this this number starts to give headaches: a big datacenter can face to silent data corruption every 15 minutes (Wikepedia). No typo here, a potential disaster may silently appear 5 times an hour for every single day of the year. Detection techniques exists but traditional RAID-5 arrays in them selves can be a problem. Ironic for a so popular and widely used solution :)

If RAID-5 was an acceptable trade-off in the past decades, it simply made its time. RAID-5 is dead? *Horray!*

More advanced topics

ZFS Intention Log (ZIL)

Final words and lessons learned

ZFS surpasses by far (as of September 2011) every of the well-known filesystems around there: none of them propose such an integration of features and certainly not with this management simplicity and robustness. However in the Linux world it is definitely a no-go in the short term especially for production systems. The two known implementations are not ready for production environments and lacks some important features or behave in a clunky manner, this is absolutely correct as none of them pretend to be at this level of maturity and the licensing incompatibility between the code opened by Sun Microsystems some years ago and the GNU/GPL does not help the cause. However, both look very promising once their corners will become rounded.

For a Linux system, the nearest plan B is you seek for a BTRFS like filesystem covering some of the functionalities offered by ZFS is BTRFS (still considered as experimental, be prepared to a disaster sooner or later although BTRFS is used by some Funtoo core team members since 2 years and proved to be quite stable in practise). BTRFS however does not pushes the limits as much as ZFS does: it does not have built-in snapshot differentiation tool nor implement built-in filesystem streaming capabilities and roll-backing a BTRFS subvolume is a bit more manual than in "the ZFS way of life".

Footnotes & references

Source: solaris-zfs-administration-guide