Closed bacongobbler closed 7 years ago
@mboersma mentioned to me during manual testing that this is a one-off container. That is, it runs to completion then dies. It looks like the container is unhealthy in a CrashLoopBackoff but in reality it's just running a few times during the e2e job and that's all we really need.
Open to other suggestions to improve upon this, but I think this is good enough to ship so we stop breaking GKE :)
I made a trivial fork of spotify/docker-gc that keeps the container from exiting by running the script every two hours via crond
. You can test it using quay.io/mboersma/spotify-gc:0.1.2
if you're curious. Maybe something like that could be included with Deis Workflow itself, although it might be better to start from deis/base
.
As seen in our GKE clusters, the docker images for the applications spawned by e2e are never reaped. This causes thousands of images to persist, all identical to each other on a single node and can trigger etcd issues.
See https://github.com/deis/controller/issues/1082#issuecomment-248079239 for some background context.