bzr and centralised repositories

[Martin Pool]

On Tue, 2005-04-19 at 14:45 -0500, Nicholas Nethercote wrote:
> Hi,
> 
> I'm following the Bazaar-NG development with interest.  I've read through 
> the design documents, but I'm a bit unclear about how it would support 
> centralised repositories (I have no experience with decentralised VCS).
> Basically I'm interested in knowing how I could use Bazaar-NG for the 
> projects I work on now, which is entirely selfish, but perhaps what 
> follows might also provide a useful use-case for the Bazaar-NG developers.
> 
> The main project I work on is Valgrind.  We just switched from CVS to 
> Subversion, which wasn't a very big change.  We have a centralised 
> repository, of course.  There are about 7 developers who have commit 
> access, 3-4 of whom do most of the development.  Most development is done 
> on the HEAD.  We only use branches for maintaining stable releases, and 
> only bug fixes go into those branches.  (And so such bug fixes have to get 
> merged into both the stable branch and the HEAD, which is a chore.)  We 
> can get away with this because there are few enough developers.  One 
> consequence is that we have stable releases and development releases, and 
> the development releases often get way ahead of the stable releases 
> (because the development pace is quite fast and so large architectural 
> changes still occur quite often), which is not ideal.
> 
> So that's how we work as a group. My personal way of working with SVN is 
> that I usually have several versions of the HEAD checked out into several 
> workspaces, which I call head1, head2, head3, etc.  I have multiple 
> workspaces because I often work on more than one change at a time.  One 
> nice thing is that I can just use "svn update" to keep these workspaces 
> up-to-date with changes that are committed into the head.  This way I can 
> run the regression tests just before committing and be fairly confident 
> that my change hasn't broken anything.  An important thing is that when I 
> run the regression tests in any of these directories, I'm testing the HEAD 
> plus a single changeset.  I like doing that before committing, rather than 
> testing and committing multiple changesets at once (which I think 
> decentralised VCS systems might encourage, but I'm not sure).

You can think of these checkouts as being a kind of degenerate private
branch: you can work in parallel in each of them until your changes are
ready to merge back to the rest of the world.  But they're degenerate
branches because you can't version your own changes until you're ready
to publish them.  In bzr you have the option of doing versioning in your
own private branch before publishing your work.

> This suits me fine because I almost always have network access.  Another 
> Valgrind developer finds Subversion more frustrating because he likes to 
> work on his laptop on the train, so he wants to be able to do offline 
> development.  Because of this, he's a fan of BitKeeper.
> 
> I saw what Greg Hudson said at 
> http://web.mit.edu/ghudson/thoughts/bitkeeper.whynot.  A lot of what he 
> said sounded right to me, given our model of working.  It's unclear to me 
> how my way of working would fit with Bazaar-NG, or how it could be 
> adapted.  Would each workspace be a branch?  How would the branches be 
> committed to the central repository? 
> http://www.bazaar-ng.org/doc/shared-branches.html kind of covers this 
> ground, but I still don't get the big picture.

Workspaces in bzr can be either of two modes: "bound" or "independent".
(I'm still thinking of good names for these.)

A bound branch works like a CVS or Svn checkout: when you commit, the
changes are pushed into possibly-remote possibly-shared storage.  You
need to come up to date with other changes in that branch before you
commit.

An independent branch lets you record changes within itself, and later
push them elsewhere.  

The two modes can interoperate and you can flip a branch from one mode
to the other.  So if, for example, your central server has a hardware
failure everybody might switch to independent mode and then reintegrate
their work later.

> What would be the benefits of Bazaar-NG for Valgrind over Subversion? 
> Maybe there would be no great benefits.  Or maybe I can't see them because 
> I'm so used to centralised VCS that I don't understand how decentralised 
> VCS could help us -- ie. I don't realise what shortcomings of centralised 
> VCS we're working around.  For example, maybe our very limited use of 
> branching is because the branching and merging in centralised VCS systems 
> is not very good.

One advantage is that you can accommodate the guy on the train while
everyone else works in the way they're used to.

Another is that the merge facilities for maintaining branches with
different goals for stability or features should work better.  

Of course at this point svn is much more mature.

(Sorry, out of time now, will write more or you can ask more questions.)
-- 
Martin

-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 189 bytes
Desc: This is a digitally signed message part
Url : https://lists.ubuntu.com/archives/bazaar/attachments/20050420/7960ba61/attachment.pgp