Capacity Planning Tips
The “distributed” in distributed version control system should indicate that Bazaar is well suited for multi-site development situations and indeed, that is the case. The advantage comes from the ease and transparency of managing merges between branches with divergent history. Note that there are many, many different ways to manage widely-flung development setups using Bazaar and its branching and merging capabilities. These can be discovered and tested before being implemented as policy. We will describe one such possible setup here.
Consider ProjectX Corp’s international expansion with a new branch office in Darwin, Australia, in addition to the company’s headquarters in Austin, Texas, USA. One of the difficulties of a far-flung multi-site development environment such as this is that the network connection between Australia and Texas is slow and unreliable. So, each branch office would like the master branch to be local to them. (In situations with good network connectivity, a local branch bound to the remote master may be all that is needed to support multi-site development.)
Of course, with two master branches, there is always the question of which one is authoritative. Given Bazaar’s facility at managing multiple branches, we suggest that it is best not to privilege either the Texas or Australia branches, but to merge both of them into a separate master branch (which may reside at either site). For definiteness, we will locate the master branch at the Texas site. So, we will have three branches stored on two servers: trunk and texas-integration at the Texas site and australia-integration at the Darwin site. These branches are named in terms of the sites where the development takes place, but in many cases it may make more sense to name branches after the functional teams rather their geographical locations. Since we are trying illustrate the issues with multi-site development, we will persist in this naming scheme.
Using our previous setup at the Texas site, we will simply rename the old trunk branch as trunk and branch a copy as texas-integration.
$ cd /srv/bzr/projectx $ mv trunk trunk # can simply rename on the filesystem $ bzr branch trunk texas-integration # very fast in a shared repository
In Australia, we need to set up the /srv/bzr/projectx directory and get a copy of the current trunk as australia-integration:
$ mkdir -p /srv/bzr $ cd /srv/bzr $ bzr init-repo --no-trees projectx $ cd projectx $ bzr branch bzr+ssh://server.example.com/srv/bzr/trunk $ bzr branch trunk australia-integration
Merging to master
Then, each office works with their local copy of the trunk. At some point, sooner or later depending on the pace of development in the two locations, the two local trunks need to be merged. (In general, sooner beats later when merging, since there is no penalty for multiple merges.) In this example, Alice at the Texas office will do the merging on her local machine using branches on the server:
# Get a copy of the Australia branch in Texas. After the initial branch # command, use pull to keep the branch up to date. With a slow network, # this is the only slow part $ bzr branch bzr+ssh://autralia.example.com/srv/bzr/projectx/australia-integration \ bzr+ssh://server.example.com/srv/bzr/projectx/australia-integration # Check out the master branch locally for doing the merge $ cd ~/projectx $ bzr checkout bzr+ssh://server.example.com/srv/bzr/projectx/trunk $ cd trunk $ bzr merge bzr+ssh://server.example.com/srv/bzr/projectx/texas-integration # Run the test suite and resolve any conflicts $ bzr commit -m "Merge Texas branch to master" # Now, merge from Australia using the local copy of that branch $ bzr merge bzr+ssh://server.example.com/srv/bzr/projectx/australia-integration # Run the test suite and resolve any conflicts between the two offices $ bzr commit -m "Merge Australia branch to master"
Note that Bazaar does not commit even cleanly applied merges by default. This is because although a merge may apply cleanly, the merged state still needs to be checked before it is committed. (Just because there are no text conflicts does not mean that everything will work after a merge.) An alternative that can pull when possible and merge otherwise is available with bzr merge --pull.
Merging back to local trunks
Now the trunk branch is the most up-to-date version of the software and both of the local trunks need to reincorporate the changes from the master. If no new commits have been made to texas-integration, then that can happen using bzr pull:
$ cd ~/projectx $ bzr checkout bzr+ssh://server.example.com/srv/bzr/projectx/texas-integration $ cd texas-integration $ bzr pull ../trunk # Use trunk from the local disk # No need to commit
If new changes have happened on texas-integration since the integration with trunk, then the above pull will produce an error suggesting to use merge:
$ bzr merge ../trunk # Run test suite, resolve conflicts $ bzr commit -m "Merging Australian changes"
In Australia, they will need to update their local copy of trunk:
$ cd /srv/bzr/projectx/trunk $ bzr pull # parent location is used by default
Then, they need to pull or merge the changes from trunk into the local trunk. This should be done by a developer with a checkout of australia-integration so that they can run the test suite:
$ cd ~/projectx $ bzr co bzr+ssh://australia.example.com/srv/bzr/projectx/australia-integration $ cd australia-integration $ bzr merge bzr+ssh://australia.example.com/srv/bzr/projectx/trunk # Run test suite and integrate Texan changes with only recent local # development $ bzr commit -m "Integrate work from Texas"
Multi-site deployments can be complicated, due to the many possible variations of development velocity, divisions of labor, network connectivity, resources for integration, etc. The preceding description is meant to be one possible way to do fairly symmetric multi-site development. (Neither Texas or Australia is privileged in this structure.) In a situation where there is one main site and other smaller sites, one of the local trunk branches can be eliminated and trunk can be used directly for development at the main site.
It is also up to the particular situation how frequently the local trunks are integrated into the master trunk. Given resources specifically for integration, it is conceivable that a developer may be constantly responsible for integrating changes from the two teams. Alternatively, the two sites could work on well-separated, well-defined features and merge to the master trunk only when their respective features are complete. Given the difficulty of resolving conflicts in very large merges and the ease of merge handling in Bazaar, we suggest that merges be done more frequently, rather than less.