Converting a svn repository with layout changes

[Jelmer Vernooij]

On Thu, 2010-04-15 at 00:44 +0300, Marius Gedminas wrote:
> On Wed, Apr 14, 2010 at 06:33:36PM +0200, Jelmer Vernooij wrote:
> > On Wed, 2010-04-14 at 19:27 +0300, Marius Gedminas wrote:
> > > On Mon, Apr 12, 2010 at 06:53:08PM +0300, Marius Gedminas wrote:
> > > > On Mon, Apr 12, 2010 at 06:34:51PM +0300, Marius Gedminas wrote:
> > > > > On Sat, Apr 10, 2010 at 06:37:10PM +0200, Jelmer Vernooij wrote:
> > > > > > On Sat, 2010-04-10 at 19:01 +0300, Marius Gedminas wrote:
> > > > > > > http://mg.pov.lt/eazysvn/svn is an Subversion repository that used the
> > > > > > > trivial "everything in root" layout for 49 revisions, then switched to
> > > > > > > trunk/ + branches/ + tags/ for the rest.
> ...
> > > Now
> > > 
> > > $ bzr branch http://mg.pov.lt/eazysvn/svn eazysvn
> > > 
> > > gives me a repository containing 66 revisions, half of which contain
> > > directories named trunk/, branches/ and tags/
> > > 
> > > $ bzr branch http://mg.pov.lt/eazysvn/svn/trunk eazysvn
> > > 
> > > gives me a repository containing 15 revisions, losing old history.
> >
> > This is correct - see the output of:
> > 
> > svn log http://mg.pov.lt/eazysvn/svn/trunk
> > 
> > which also reports just 15 revisions.
> 
> You asked svn log to filter by pathname and it does so, but the full history
> itself is still there.  Compare
> 
>   svn log http://mg.pov.lt/eazysvn/svn/trunk/README.txt
>   (10 changesets)
The history of the trunk directory is what matters, that is the branch.
Bazaar does not support tracking cherry-picks and that's what is
happening here with README.txt. The branch "trunk" was created in r50
and /trunk/README.txt was cherry-picked into it from the / branch
(/trunk is not an ancestor of /trunk). Since Bazaar cannot represent the
cherrypick bzr-svn discards that information.

> I think my approaches are:
> 
>   * bzr branch http://mg.pov.lt/eazysvn/svn eazysvn, followed by
>     bzr rm tags branches; bzr mv trunk/* ./; bzr rm trunk; bzr commit
>       - upside: simple
>       - downside: ugly history
> 
>   * svnadmin dump, tweak the history using svndumpfilter, import into a
>     new svn revision, then convert it with bzr.
>       - upside: clean history
>       - downside: medium amount of effort
>       - upside: I've used svndumpfilter before; can't say I've loved it
> 
>   * bzr branch http://mg.pov.lt/eazysvn/svn eazysvn, then dump the bzr
>     branch into some format (fast-export?), filter it (with sed? perl?),
>     import into a new bzr branch
>       - upside: clean history
>       - downside: medium amount of effort
>       - upside: unfamiliar tools
> 
>   * write a custom conversion script/fork bzr svn-import.
>       - upside: clean history
>       - downside: a lot of effort
This would be tricky to get right, see earlier comments. You might have
more luck patching svn2bzr for something like this.

>   * implement and submit a patch for bzr svn-import to handle
>     mixed-style layouts (I can't be the only one on the planet who
>     tended to start with simple small Svn repositories and later
>     migrated to trunk+tags+branches!)
>       - upside: giving back etc.
>       - downside: even more effort
It's not the mixed layout that is giving problems. It's the fact that
your trunk branch didn't exist before r50. bzr-svn handles layout
changes just fine as long as the branch was copied.

To give an example: It's perfectly possible to have a branch at /foo
that was later renamed to /trunk. bzr-svn will happily import the
history beyond the rename.

>   * convince someone else to implement support of mixed-style svn
>     repositories
>       - downside: I'd have to be very persuasive
See comment above, mixed-style layouts are already supported.

>   * try Tom Widmer's Mercurial recipe, then see about converting to bzr
>     (or stay with hg, it can't hurt to be fluent in more than one DVCS)
>       - upside: simple
>       - downside: not Bazaar
>       - upside: fluency in another DVCS
I doubt Mercurial handles this better. If they do, I'd be interested to
hear in how they represent it as they do not support tracking
cherrypicks either AFAIK.

Cheers,

Jelmer