Bundles

Status

Date:

2007-06-19

This document describes the current and future design of the bzr bundle facility.

Motivation

Bundles are intended to be a compact binary representation of the changes done within a branch for transmission between users. Bundles should be able to be used easily and seamlessly - we want to avoid having a parallel set of commands to get data from within a bundle.

A related concept is merge directives which are used to transmit bzr merge and merge-like operations from one user to another in such a way that the recipient can be sure they get the correct data the initiator desired.

Desired features

  • A bundle should be able to substitute for the entire branch in any bzr command that operates on branches in a read only fashion.

  • Bundles should be as small as possible without losing data to keep them feasible for including in emails.

Historical Design

Not formally documented, the current released implementation can be found in bzrlib.bundle.serializer. One key element is that this design included parts of the branch data as human readable diffs; which were then subject to corruption by transports such as email.

June 2007 Design

Bundle Format 4 spec

Future Plans

Bundles will be implemented as a ‘Shallow Branch’ with the branch and repository data combined into a single file. This removes the need to special case bundle handling for all command which read from branches.

Physical encoding

Bundles will be encoded using the bzr pack format. Within the pack the branch metadata will be serialised as a BzrMetaDir1 branch entry. The Repository data added by the revisions contained in the bundle will be encoded using multi parent diffs as they are the most pithy diffs we are able to create today in the presence of merges. XXX More details needed?

Code reuse

Ideally we can reuse our BzrMetaDir based branch formats directly within a Bundle by layering a Transport interface on top of the pack - or just copying the data out into a readonly memory transport when we read the pack. This suggests we will have a pack specific Control instance, replacing the usual ‘BzrDir’ instance, but use the Branch class as-is.

For the Repository access, we will create a composite Repository using the planned Repository Stacking API, and a minimal Repository implementation that can work with the multi parent diffs within the bundle.

We will need access to a branch that has the basis revision of the bundle to be able to construct revisions from within it - this is a requirement for Shallow Branches too, so hopefully we can define a single mechanism at the Branch level to gain access to that.