How to Sync CI (Hudson) Activity into an existing automated Build Process (phing, svn)?

Question

OUR CURRENT BUILD PROCESS

We're a small team of developers (2 to 4 people depending on project) who currently use Phing to deploy code to a staging environment, before going live. We keep our code in a SVN repo, where the trunk holds current active development and, at certain times, we do make branches that we test and then (if successful), tag and export to the staging env. If everything goes well there too, we finally deploy'em in production servers. Actions are highly automated, but always triggered by human intervention.

---

THE DOUBT

We'd now like to introduce Continuous Integration (with Hudson) in the process; unfortunately we have a few doubts about activity syncing, since we're afraid that CI could somewhat interfere with our build process and cause certain problems.

Considering that an automated CI cycle has a certain frequency of automatically executed actions, we see 2 possible cases for "integration", each with its own problems:

1. Case A: each CI cycle produces a new branch with its own name; we do use such a name to manually (through phing as it happens now) export the code from the SVN to the staging env. The problem I see here is that (unless specific countermeasures are taken - IE deletion) the number of branches we have can easily grow out of control (let's suppose we commit often, so that we have a fresh new build/branch every N minutes).

2. Case B: each CI cycle creates a new branch named 'current', which is then tagged with a unique name only when we manually decide to export it to staging; the current branch, at any case is then deleted, as soon as the next CI cycle starts up. The problem we see here is that a new cycle could kick in while someone is tagging/exporting the 'current' branch to staging thus creating an inconsistent build (but maybe here I'm just too pessimist, since I confess I don't know whether SVN offers some built-in protection against this).

---

With all this being said, I was wondering if anyone with similar experiences could be so kind to give us some hints on the subject, since none of the approaches depicted above looks completely satisfing to us.

Is there something important we just completely left off in the overall picture? Thanks for your attention & (in advance) for your help!

Answer 1

An approach we used in our project was to run CI builds only when there was a code change. This can be configured on your SVN as a post commit hook. You can then remotely trigger builds in HUDSON via an authenticated URL. Problem I see though is that since jobs have to be created, unless your build system supports it, there is no way for hudson to figure out there is a new branch on the repo and create a job for that.

Answer 2

In both options you start with "each CI cycle produces a new branch". Don't do that. You want to keep your number of branches to the minimum and always under control (manually created) to avoid that your project becomes a mess. The decision of whether the development in the mainline is ready and you can produce a release candidate (branch from the trunk) is not trivial.

CI cycles are triggered by changes in the code to ensure that the integration of those changes doesn't break the application. Therefore, you'd rather set up a project in Hudson for each active stream of development, this is, one for the mainline, one for the branch that represents the production version (for bug-fixing) and eventually one for the RC.

Martin Fowler's article about Continuous Integration is an excellent guide to the whys and hows of CI implementation.