Re: [Gnu-arch-users] arch lkml

[Peter Conrad]

Hi,

On Wed, Dec 10, 2003 at 02:05:17PM -0500, Aaron Bentley wrote:
> 
> A format like this will not have random-access properties.  For file
> foo, if you store 12 revisions, you'll need to consult all 12 in order
> to reconstruct revision 12.  As with Arch, as the number of revisions
> increases, the work increases.  Nothing is free.

I have thought about a "blob" storage format for arch archives before,
and I think it could be very useful for many applications. Of course
I'm well aware of the disadvantages of large "blob" files as opposed to
many small changeset files, therefore I wouldn't advocate the blob format
as the (or even a) standard arch archive format.

However, it could be used very efficiently for mirrors. A well organized
blob storage format (with a library for accessing it) would be capable of

 - serving a standard changeset tar.gz files for any contained version
 - serving combined changesets as Andrea has requested ("give me a changeset
   containing all changes from patch-255 to patch-512")
 - serving changesets containing context-free diffs (useful if you want to
   update an unmodified local version from a remote mirror over a slow link)
 - serving as a revision library with access to any file in any revision
 - answer questions like "what has changed in file x between patch-x and
   patch-y?"

Dreaming one step further you could create a kernel module for serving all
of this (well, at least the archive and revision library) through a filesystem
interface...

My feeling is that such a format could not only be very space-efficient if
done properly, but also very efficient in access time even though everything
would have to be generated "on the fly".

Bye,
        Peter
-- 
Peter Conrad                        Tel: +49 6102 / 80 99 072
[ t]ivano Software GmbH             Fax: +49 6102 / 80 99 071
Bahnhofstr. 18                      http://www.tivano.de/
63263 Neu-Isenburg

Germany