Google Container Registry lifecycle policy for images retention


Is your Google Container Registry filling up, taking up storage and becoming expensive? How to handle images retention as a service?

Amazon’s Elastic Container Registry has a feature called Lifecycle Policies to handle images retention. Google doesn’t have this feature. There is a feature request in their tracker since Aug 2018 and there is not ETA for it so far…

There is a popular bash script from Ahmet and Go in CloudRun from Seth but none of them solve the requirements I needed. What exactly do I need?

I wanna scan my whole GCR and delete the digests that are:

  • older than X days
  • not being used in my kubernetes cluster
  • not the most recent Y digests (I wanna keep, say, 20 most recent tagged digests)

When I check these requirements, I want to apply these lifecycle policies to all the images.

Say I have few images in GCR with certain prefixes: is the docker registry endpoint
my-project is ID of my GCP project
foo/bar is the prefix (“repo”)
my-service is an image name
123 is a tag

my-service:123 is an image with a tag, but wait, what is the digest?

Image vs Layers, taken from

A docker image digest is an ID (hashing algorithm used and the hash computed). The digest can look like this:


You can tag a digest with several tags, even zero tags = untagged image.

Let’s say build an image my-service and push it to docker registry. When pushing, I tag it with :123. The new produced digest has two tags, :123 and :latest.The digest that was tagged :latest before I pushed this image, got the :latest tag removed.

If I remove a tag from an image in GCR, I simply remove a tag from the digest, I don’t delete the digest though.

What I can delete, in order to save some space, is the digest, like this:

gcloud container images delete -q — force-delete-tags

Then, what do I need to do?

  • Recursively scan for all image prefixes (
  • For each prefix — list all its digests, delete the ones that don’t match my rules

How to check if they match the rules:

  • sort them, preserve the most recent Y digests
  • fetch pods and replicaSets’ image:tags (all of them, even the ones scaled to zero, we’d need these images in case of a rollback) from the k8s cluster, then go through the digests (that belong to that image name) and check if ANY of their tags contain a tag that is used in the cluster, preserve those
  • check the rest of the digests, if they are older than X days, delete them

you could use standard kubectl to fetch the data:

kubectl get rs,po --all-namespaces -o jsonpath={..image} | tr ' ' '\n'

the is exposing a docker v2 API, you can use a standard docker client or just curl using the gcloud token

ACCESS_TOKEN=$(gcloud auth print-access-token)
curl --silent --show-error -u_token:"$ACCESS_TOKEN" -X GET ""

I implemented all of this using bash/jq (yep, that wasn’t a smart idea) and published it to github:

Right now I’m running this in Gitlab-CI pipeline on a cron schedule (once a day) to evaluate it’s dry-run logs for production GCP projects.

I’m planning on rewriting this to python (py-kubeng and docker-py) if Google will not to come up with ETA for this feature :(