I've been going through what changes for workloads on the 1.34 to 1.35 hop. EKS 1.34 leaves standard support on December 2, 2026, so a lot of clusters will make that hop soon. If you use a private registry, this is the change I'd check first.
What changes
Up to 1.34, an image on a node is shared. If one pod pulls a private image with its imagePullSecrets, any other pod scheduled to that node can start from the cached copy with imagePullPolicy: IfNotPresent, even if it has no secret at all.
In 1.35, KubeletEnsureSecretPulledImages is Beta and enabled by default. A pod now reuses a cached image only if its own credentials pulled that image successfully before. A pod with different credentials, or none, has to pull again. If it has no secret for a private registry, that pull fails.
So the pods at risk are the ones that have been working by accident. They use a private image, they carry no pull secret, and some other pod has always pulled the image for them.
One command
This lists pods with no imagePullSecrets that share an image with a pod that has one. It only reads.
kubectl get pods -A -o json | jq -r '
[.items[] | {img: .spec.containers[].image,
pod: (.metadata.namespace + "/" + .metadata.name),
secret: ((.spec.imagePullSecrets // []) | length > 0)}]
| group_by(.img)[]
| select(any(.secret) and any(.secret | not))
| .[] | select(.secret | not) | .img + "\t" + .pod'
I ran it on sample pod JSON with a billing API that has a pull secret and a nightly job that doesn't. It prints:
ghcr.io/acme/base-worker:4.2 reports/nightly-28861
That job is the one to fix.
What to do
Give every pod that uses a private image its own pull secret. If you need more time, the kubelet setting imagePullCredentialsVerificationPolicy: NeverVerify keeps the old behaviour.
Two things the command can't tell you. Images that were already on a node when the kubelet first turned the feature on are treated as pre-pulled, and the default policy (NeverVerifyPreloadedImages) doesn't check those. So this tends to show up on fresh nodes. And I haven't confirmed whether every managed provider keeps the upstream default, so read the kubelet config on a 1.35 node before you rely on it.
The Kubernetes docs cover it under "Ensure image pull credential verification" on the Images page.
Top comments (0)