[Kde-scm-interest] Have we arrived to a dead end?

[Oswald Buddenhagen]

On Sat, Feb 13, 2010 at 01:25:10PM +0100, Thiago Macieira wrote:
> Em Sábado 13. Fevereiro 2010, às 12.51.42, Oswald Buddenhagen escreveu:
> > and if one wanted to go really overboard, one could implement a proper
> > transactional system in which the repos coordinate with each other
> > directly. i mean, distributed databases isn't really a completely new
> > research topic.
> 
> I hope you were not thinking of having the repositories coordinate the 
> locking.
>
in the above paragraph, i did.

> This has to be client-driven: the client locks all the repository 
> branches it wants to push to, then it pushes, then unlocks.
> 
that's the simple variant. i don't think any amount of handshaking can
guarantee real transactions in such a setup, but then, what do i know
about distributed systems ...

> Otherwise you'd end up requiring that you can only push to one server.
> 
that's a completely acceptable requirement, wouldn't you think?

> But with a client-driven locking, the servers might want to time out
> the locking, which means you'll be limited in time (and thus amount of
> data) that you can "atomically" push.
>
an a-priori check whether the updates would have any chance to succeed
would be certainly sensible, but all that really needs locking are the
ref updates themselves. the rest would be just some more garbage to
collect in case some ref update fails.