Last I spent some time learning about the Google Container Engine, and was almost able to get a full continuous deployment pipeline working, but it was missing the last mile from the container building to it actually getting deployed, so I wrote up a very small proof of concept called gke_ci and put it up on Github.
It's quite, quite basic, and works by subscribing to the
cloud_builds PubSub topic, and then
calling the Kubernetes API to update the deployed version of all existing deployments that use
that container. There are so few container builds, that I discover all deployments every time
there is a successful build, so there is no "setup" process to add another service, and it'll
automatically work for new stuff.
gke_ci in a deployment with one instance, and interestingly it actually upgrades itself!
Using the same logic as for other builds: when it sees a new build for its container, it'll detect its
deployment uses that container, and trigger an upgrade. Kubernetes will then spin down the current
container and spin up a new version, with builds remaining queued in PubSub until it comes back up and
resumes its duties.
Altogether, this gives me a ~5 minute deployment from push to full rollout on my small cluster, and
doesn't require active watching from me (rolling rollouts catch vast majority of issues I've found),
although I will still need to add a build step that runs tests (should be easy in Container Builder
if I'm willing to polute the container a bit), as well as send notifications for failed builds
(easy to add to
gke_ci itself since it already listens to the
cloud_builds queue, or could add
another simple utility with a new subscription).