push from lightweight checkout to location with bzrdir creates BranchReference

[James Henstridge]

I don't have a reliable way of reproducing this problem, but thought I'd report it anyway in case anyone else has an insight.

As part of Launchpad development, I'll often push branches to our development server, roughly following this workflow:
 1. create a lightweight checkout of the local copy of the branch
 2. make my changes in the lightweight checkout and commit
 3. run "bzr push" in the lightweight checkout, which will create a new branch and push it via bzr+ssh to the development server.

The destination directory on the development server is part of a shared repository.

Usually this works fine, but once or twice I ended up with a BranchReference object on the server containing the path to the local copy of the branch (which of course does not exist remotely).

It only happens occasionally, but it has also happened to other members of the team so I don't think it is purely related to my local environment.

[Christian Reis (kiko)]

This happens to me consistently. Here are some details:

kiko@whatever:~$ bzr --version
Bazaar (bzr) 0.90.0

I use a script to create my branches:

# Create the branch in the repository
cd ~/repo/
bzr branch ~/devel/archives/launchpad $1

# Make a lightweight checkout of the new branch
cd ~/devel
bzr checkout --lightweight ~/repo/$1 $1

bzr push --remember \
    bzr+ssh://devpad.canonical.com/code/kiko/launchpad/$1 &

I then use the checkout, commit, commit and then push, and it says no new revisions. I can reproduce this pretty reliably using whatever.async.com.br and devpad.canonical.com.

[Martin Albisetti (beuno)]

Is this still happening with 1.3?

[James Henstridge (jamesh)]

I haven't seen reports of problems recently, but that might just be because we've been lucky (I've never been able to find a way of reliably reproducing it).

[Aaron Bentley (abentley)]

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

James Henstridge wrote:
> I haven't seen reports of problems recently, but that might just be
> because we've been lucky (I've never been able to find a way of reliably
> reproducing it).

It seemed to be repeatable if you created a .bzrdir and then pushed.

Aaron
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.4.6 (GNU/Linux)
Comment: Using GnuPG with Mozilla - http://enigmail.mozdev.org

iD8DBQFH94oP0F+nu1YWqI0RAhK8AJ9tV67ERP8iJizUr8BESNd6b4x00wCfdXCN
fbz7bgiWbgSRgncg/h7gRJ0=
=bw0k
-----END PGP SIGNATURE-----

[Martin Pool (mbp)]

I could not reproduce this using kiko's recipe in bzr 1.9. If you're still hitting this, please help us work out how to reproduce it.

[James Henstridge (jamesh)]

Attached is a script that replicates the problem for me with bzr-1.9. This is using Aaron's idea of pushing to a location with an existing .bzr dir (in this case, a directory initialised as a repository).

On my system, the script prints the following at the end:

  ./branch format: Bazaar Branch Format 6 (bzr 0.15)
  ./checkout format: Bazaar-NG Branch Reference Format 1
  ./target format: Bazaar-NG Branch Reference Format 1

I'd expect ./target to be a real branch rather than a reference, given that I get a real branch when pushing to a non-existent location.

[Jelmer Vernooij (jelmer)]

Reproduced here as well. Thanks for the script.

[Jelmer Vernooij (jelmer)]

related to bug 462513?