Re: Pushing a bzr branch with rsync

This article responds to some of the points in Andrew’s post about Pushing a bzr branch with rsync.

bzr rspush and shared repositories

First of all, to understand why bzr rspush refuses to operate on a non-standalone branch, it is worth looking at what it does:

  1. Download the revision history of the remote branch, and check to see that the remote head revision is an ancestor of the local head revision. If it is not, error out.
  2. If it is an ancestor, use rsync to copy the local branch and repository information to the remote location.

Now if you bring shared repositories into the mix, and there is a different set of branches in the local and remote repositories, then step (2) is liable to delete revision information needed by those branches that don’t exist locally. This is not a theoretical concern if you do development from multiple machines (e.g. a desktop and a laptop) and publish to the same repository.

Storage Formats and Hard linking

The data storage format used by Bazaar was designed to be cross platform and compact. The compactness is important for the dumb/passive server mode, since the on-disk representation has a large impact on how much data needs to be transferred to pull or update a branch.

The representation chosen effectively has one “knit” file per file in the repository, which is only ever appended to (with deltas to the previous revision, and occasional full texts), plus a “knit index” file per knit that describes the data stored inside the knit. Knit index files are much smaller than their corresponding knit files.

When pushing changes, it is a simple matter of downloading the knit index, working out which revisions are missing, append those to the knit and update the index. When pulling changes, the knit index is downloaded and the required sections of the knit file are downloaded (e.g. via an HTTP range request).

The fact that the knit files get appended to is what causes problems with hard linked trees. Unfortunately the SFTP protocol doesn’t really provide a way to tell whether a file has multiple links or a way to do server side file copies, so while it would be possible to break the links locally, it would not be possible when updating a remote branch.

Furthermore, relying on hard links for compact local storage of related branches introduces platform compatibility problems. Win32 does not support hard links (update: apparently they are supported, but hidden in the UI), and while MacOS X does support them its HFS+ file system has a very inefficient implementation (see this article for a description).

Rsync vs. The Bazaar smart server

As described above, Bazaar is already sending deltas across the wire. However, it is slower than rsync due due to it waiting on more round trips. The smart server is intended to eventually resolve this discrepancy. It is a fairly recent development though, so hasn’t achieved its full potential (the development plan has been to get Bazaar to talk to the smart server first, and then start accelerating certain operations).

When it is more mature, a push operation would effectively work out which revisions the remote machine is missing, and then send a bundle of just that revision data in one go, which is about the same amount of round trips as you’d get with rsync.

This has the potential to be faster than an equivalent rsync:

  • Usually each revision only modifies a subset of the files in the tree. By checking which files have been changed in the revisions to be transferred, Bazaar will only need to open those knit files. In contrast, rsync will check every file in the repository.
  • In Andrew’s rsync script, the entire repository plus a single branch are transferred to the server. While only one branch is transferred, the revision information for all branches will be transferred. It is not too difficult to reconstruct the branches from that data (depending on what else is in the repository, this could be a problem). In contrast, Bazaar only transfers the revisions that are part of the branch being transferred.

So it is worth watching the development of the smart server over the next few months: it is only going to get faster.