Re: [Duplicity-talk] Re: are periodic full backups necessary?

[Peter Schuller]

> I think the operations you'd need to be able to perform on the store
> are:
> 
>   1. Upload new encrypted blob plus encrypted signature
>   2. Upload encrypted directory tree (contains pointers to blobs)

This is blob id + offset in blob, right?

>   3. Download encrypted signature for blob
>   4. Download portion of encrypted directory tree
>   5. Replace subtree of encrypted directory tree (ie re-pack)
>   6. (Garbage collection) delete encrypted blob or tree

Sounds good to me, as long as you're willing to scale O(n) with the
number of files in terms of absolute overhead per backup. But this
tends to be a valid assumption under many circumstances anyway.

> The problem with literally using git for this is that (among other
> problems) it can't cope with multi-gigabyte files.  I also don't know
> if it's binary diffing facilities are as good as rdiff's (I doubt it).
> 
> But I think the general approach has merit.

I like it, since it implements something similar to the remote block
device, but without dependencies on the OS and without being massively
susceptible to latency.

The two downsides I see are:

* The need for a remote instance of the software (though this should not
  change much in terms of the interface it must implement).
* One is critically dependent on this block handling to the extent that
  we are now only dependent on the file system. The fundamental "ok, I can
  ls -l and md5sum the files and it looks good" factor is missing.

But with a sufficiently robust stand-alone, well-tested and internally
consistency-checking implementation...

-- 
/ Peter Schuller

PGP userID: 0xE9758B7D or 'Peter Schuller <address@hidden>'
Key retrieval: Send an E-Mail to address@hidden
E-Mail: address@hidden Web: http://www.scode.org

pgppBFG1ZXL7p.pgp
Description: PGP signature