On Tue, 26 Jul 2005, Matthias Urlichs wrote:
> That's a problem with the bkcvs tree. Remember tht Bitkeeper does
> exactly the same thing -- the 1.0 version of *any* file is empty, and
> content appears only in version 1.1.
Not really. That may be how the SCCS _deltas_ end up being done
internally, but it's definitely not how BK changesets work.
For example, in BK SCCS files (if I understood it correctly, I've not
actually looked very closely) a "rename" ends up being two deltas - the
"diff" delta and the "rename" delta. That does _not_ really mean that BK
considers it two different things, it's purely a SCCS file layout thing,
and it just shows through to bkcvs.
Similarly, apparently in SCCS the "create" event is one revision, and the
"initial data" is another one, and again, that shows up in bkcvs even
though that's not really how BK works conceptually at all.
So then the cvs archives end up showing some of these things as multiple
separate patches, but that just means that cvsps doesn't understand that
they all get collapsed into one changeset in the bk model. The fact that
some changes may end up showing as multiple deltas is just a result of BK
mostly re-using an old fileformat that just doesn't know anything at all
about changesets what-so-ever.
cvsps should really have collapsed those things into _one_ changeset. I
had thought that it would do so automatically based on the date (the date
should always be exactly the same), but it turns out that since the log
messages might be different, cvsps will split _one_ ChangeSet into
multiple patches, which is _wrong_.
In fact, I guess the log messages _will_ be different, because the bkcvs
thing will always put in the BK key in the "real" message.
I _thought_ that this was exactly what the "--bkcvs" flag was going to
notice, but it seems to not be the case. Or maybe "cvsimport" just doesn't
pass that flag through.. In fact, I just checked, and "cvsps --bkcvs"
_does_ do the right thing, and collapses all of these things to one
But I think I see why "git cvsimport" does the wrong thing: since that one
commit has _all_ the deltas associated with the commit, it looks like
Date: 2002/02/05 17:40:40
and notice how "COPYING" (and all other new files) has two deltas
associated with it, the "INITIAL->1.1" and the "1.1->1.2" one.
And they are in the wrong order, so "cvsimport" ends up committing the
last one, which is the _empty_ one.
Notice? We'll end up committing "COPYING 1.1" (the empty initial create)
even though we _should_ have committed "COPYING 1.2" (the actual thing
that BK committed).
> Well, the bkcvs export preserved that ... "feature".
No, the bkcvs thing exports an atomic BK commit as several deltas (with
the same date) not because it's a "feature" of BK, but partly because you
can't express what BK does in CVS, and partly because of what appears to
be purely internal BK implementation details (ie a feature of the SCCS
BK did it right, and we just imported it wrong.
David - is there some way where cvsps could always order these things by
revision? I now realize that this is probably also what causes cvsps to
complain about things like:
PatchSet 2 has collisions
which means that cvsps actually _saw_ this, but just output the result in
the wrong order as far as "git cvsimport" was concerned.
The preferred solution would be to always just suppress the older revision
when you see multiple ones - it is by definition not interesting (you
cannot actually ever access it even in the original BK tree).
> (Side question - why aren't you doing a direct bk2git import?)
Because I don't have any BK trees left, and because I'm not going to touch
Andrews code. We'd have had a portable BK export format (in fact, I wrote
one as a proof-of-concept thing when we were tryign to convince Tridge
that he really doesn't want to muck inside BK internals), but then shit
So I'll use the CVS thing.
> > I realize you'd want to do that to
> > avoid connecting millions of times, but maybe it's better to use something
> > like cvsnup to download the whole thing, and then always use a local CVS
> > archive?
> ... I don't have a sensible RCS library for perl (the code that I could
> find is just a command line front-end). Fork+exec of some cvs checkout
> command per file is slower than just running a persistent CVS server.
Fair enough, I suspect that without a builtin RCS library it really ends
up being faster than the whole exec thing.
I thought there was a RCS library (I know I got some hits for "rcslib"
for some scripting language and was cursing the fact that it wasn't for
C), but I never really looked closer. I'm not surprised if the "library"
ends up just doing a "system()" thing.
> I've tried other ideas, but they run into problems because some
> idiots^Wpeople occasionally tag only parts of a CVS tree
Yeah, CVS really allows some very very annoying things that only work in a
To unsubscribe from this list: send the line "unsubscribe git" in
the body of a message to [EMAIL PROTECTED]
More majordomo info at http://vger.kernel.org/majordomo-info.html