backing up zfs dataset

[Andrew Reilly]

On Fri, Nov 11, 2011 at 04:36:29PM +0000, Johannes Totz wrote:
> To back up a zfs dataset there are a few possibilities:
> 1) rsync file data to another machine
> 2) zfs-send to another machine, into a zfs dataset
> 3) zfs-send to another machine, dumping the stream to a file

Backing up to another machine clearly has advantages, but
it's not the only way.  I use ZFS send/receive to backup to a
removable drive on the same machine, and it seems to be working
quite nicely.

> The first one works alright but you loose admin info, properties set on 
> the dataset, etc
> The second is prefered but requires another machine which runs zfs.

Method 2 does require the receiver to be running zfs, but
it seems to be the only good way.  Well, if you want to
remote-backup to a machine not running ZFS then rsync from a
zfs snapshot should get most of the way there.  Modern rsync is
pretty good about metadata and sparse files, I think.  (Having
said that, I haven't used that method for a long time.)

> The third is bad.

Tell me about it!  I was saving zfs send streams to a UFS disk
as a backkup strategy for ages.  When I needed it, I discovered
that there is no recovery or redundancy in zfs receive: it just
stopped, unable to read the filesystem.  Toast.

> So far I have been doing (3), for daily short-term backups, works, 
> tested, everything is peachy. However, I dont like it anymore for 
> obvious reasons. Ideally, I would like to go with (2). But I dont have 
> another zfs-capable machine, or the machine that I would like to backup 
> onto will not ever run zfs.

FWIW, my backup script is now based on:

    zfs snapshot tank/${1}@0 || error "problem making new snapshot on $1"
    zfs send -i tank/${1}@1 tank/${1}@0 | zfs recv -vF bkp2pool/$1 || error "problem sending incremental fs packet on $1"

with appropriate snapshot rotation and checking scripts before
and after.  Seems to be working OK, and I really like the
confidence that being able to wander around inside the readable
backups provides.

I do think that backup is something of a weakness for ZFS at
the moment.  Sure, live filesystems and snapshots are clearly
cool, and the modern way and all, but there is an awful lot of
flexibility and ease of undersanding in the model of a "backup
file on a tape."  Doesn't have to be on a tape, but the moral
equivalent to dump/restore would (in my book) be a wonderful
addition to ZFS, if anyone felt inclined.  Just padding the
send/receive serialisation format with enough checksum and
restart information to allow detection and graceful recovery
from read errors in the backup medium would do the job.

Cheers,

-- 
Andrew