Re: snapshotting - what data gets shared?

[Sean Bartell <wingedtachikoma@gmail.com>]

On Wed, Jul 14, 2010 at 11:27:39PM +0200, Mathijs Kwik wrote:
> Hi all,
> 
> I'm used to snapshots with LVM and I would like to compare them to btrfs.
> 
> The case I want to compare is the following:
> At the moment a snapshot is created, no extra space is needed (maybe
> some metadata overhead) and all data is shared between the original
> and the snapshot.
> In LVM, snapshots work at the block-level, so any changes done to the
> original volume trigger a COW to the snapshot.
> If LVM is configured to use 4Mb blocks (default), this means that
> overwriting a 100k file, will lead to 4Mb "snapshot data" to be backed
> up.
> A 800Mb file will take around 800Mb.
> So, for small files (that are not on the same extent/block) this can
> waste quite some space, while for bigger files, or lots of files
> "close" to each other, it doesn't matter much.
> 
> How is this for btrfs snapshots?
> Do they work at the file-level? or also at blocks/extents?
> 
> I mean, does changing a 100k file lead to 100k being snapshotted?

Btrfs CoWs file extents, and files can use only the parts of an extent
they need, so a 1-byte change would only require one additional 4K data
block. Of course, metadata also needs to be updated, and will require
a number of additional blocks.

> What would happen if I have a 20G file (for example a disk image for kvm)?
> Would minor changes in that file lead to the entire 20G to be COWed/"backed up"?

No, only the relevant portion.

> Is there a distinction between data and metadata?
> Or does touching (ctime/mtime) or visiting (atime) a file cause it to be COWed?

Metadata is CoWed separately, so there will still only be one copy of
the data.

> Thanks for any info on this.
> Mathijs