Sun, 24 Apr 2016

Automating Deployments: Version Recycling Considered Harmful

Permanent link

In the previous installment we saw a GoCD configuration that automatically built a Debian package from a git repository whenever somebody pushes a new commit to the git repo.

The version of the generated Debian package comes from the debian/changelog file of the git repository. Which means that whenever somebody pushes code or doc changes without a new changelog entry, the resulting Debian package has the same version number as the previous one.

The problem with this version recycling is that most Debian tooling assumes that the tuple of package name, version and architecture uniquely identifies a revision of a package. So stuffing a new version of a package with an old version number into a repository is bound to cause trouble; most repository management software simply refuses to accept that. On the target machine, upgrade the package won't do anything if the version number stays the same.

So, its a good idea to put a bit more thought into the version string of the automatically built Debian package.

Constructing Unique Version Numbers

There are several source that you can tap to generate unique version numbers:

  • Randomness (for example in the form of UUIDs)
  • The current date and time
  • The git repository itself
  • GoCD exposes several environment variables that can be of use

The latter is quite promising: GO_PIPELINE_COUNTER is a monotonic counter that increases each time GoCD runs the pipeline, so a good source for a version number. GoCD allows manual re-running of stages, so it's best to combine it with GO_STAGE_COUNTER. In terms of shell scripting, using $GO_PIPELINE_COUNTER.$GO_STAGE_COUNTER as a version string sounds like a decent approach.

But, there's more. GoCD allows you to trigger a pipeline with a specific version of a material, so you can have a new pipeline run to build an old version of the software. If you do that, using GO_PIPELINE_COUNTER as the first part of the version string doesn't reflect the use of an old code base.

To construct a version string that primarily reflects the version of the git repository, and only secondarily the build iteration, the first part of the version string has to come from git. As a distributed version control system, git doesn't supply a single, numeric version counter. But if you limit yourself to a single repository and branch, you can simply count commits.

git describe is an established way to count commits. By default it prints the last tag in the repo, and if HEAD does not resolve to the same commit as the tag, it adds the number of commits since that tag, and the abbreviated sha1 hash prefixed by g, so for example 2016.04-32-g4232204 for the commit 4232204, which is 32 commits after the tag 2016.04. The option --long forces it to always print the number of commits and the hash, even when HEAD points to a tag.

We don't need the commit hash for the version number, so a shell script to construct a good version number looks like this:


set -e
set -o pipefail
version=$(git describe --long |sed 's/-g[A-Fa-f0-9]*$//')

Bash's ${VARIABLE:-default} syntax is a good way to make the script work outside a GoCD agent environment.

This script requires a tag to be set in the git repository. If there is none, it fails with this message from git describe:

fatal: No names found, cannot describe anything.

Other Bits and Pieces Around the Build

Now that we have a version string, we need to instruct the build system to use this version string. This works by writing a new entry in debian/changelog with the desired version number. The debchange tool automates this for us. A few options are necessary to make it work reliably:

export DEBFULLNAME='Go Debian Build Agent'
export DEBEMAIL=''
debchange --newversion=$version  --force-distribution -b  \
    --distribution="${DISTRIBUTION:-jessie}" 'New Version'

When we want to reference this version number in later stages in the pipeline (yes, there will be more), it's handy to have it available in a file. It is also handy to have it in the output, so two more lines to the script:

echo $version
echo $version > ../version

And of course, trigger the actual build:

debuild -b -us -uc

Plugging It Into GoCD

To make the script accessible to GoCD, and also have it under version control, I put it into a git repository under the name debian-autobuild and added the repo as a material to the pipeline:

<pipeline name="package-info">
    <git url="" dest="package-info" />
    <git url="" dest="deployment-utils" materialName="deployment-utils" />
  <stage name="build" cleanWorkingDir="true">
      <job name="build-deb" timeout="5">
          <exec command="../deployment-utils/debian-autobuild" workingdir="#{package}" />
          <artifact src="version" />
          <artifact src="package-info*_*" dest="package-info/" />

Now GoCD automatically builds Debian packages on each commit to the git repository, and gives each a distinct version string.

The next step is to add it to a repository, so that it can be installed on a target machine with a simple apt-get command.

I'm writing a book on automating deployments. If this topic interests you, please sign up for the Automating Deployments newsletter. It will keep you informed about automating and continuous deployments. It also helps me to gauge interest in this project, and your feedback can shape the course it takes.

Subscribe to the Automating Deployments mailing list

* indicates required

[/automating-deployments] Permanent link