
EKS can roll back an upgrade now.
On July 1, 2026 EKS gained Version Rollback: revert a control plane from N to N-1 within 7 days. It's narrower than it sounds. It reverts the control plane, not your nodes, and the kubelet skew rule closes the window the moment you recycle them. Here's the upgrade shape that keeps the rollback available, plus the Terraform timeout that quietly breaks it.
Series: EKS (3 articles)
- 2EKS can roll back an upgrade now. This article
The problem
For years, an EKS control-plane upgrade was a one-way door. Push the cluster from 1.32 to 1.33, discover on day two that a controller or an admission webhook doesn't like the new API server, and your recovery options were both bad: fix forward under pressure, or stand up a whole blue/green cluster and cut traffic across at roughly double the running cost until you were done.
On July 1, 2026 AWS shipped Version Rollback. You can now revert an EKS control plane from N to N-1, in place, within 7 days of the upgrade, at no extra charge. That is a real improvement. It is also narrower than the word "rollback" makes it sound, and if your upgrade process doesn't already have the right shape, the safety net won't be there when you reach for it.
What actually rolls back
The word implies the whole cluster steps back a version. It doesn't. Two planes, two rules:
- The control plane reverts on every cluster, N to N-1, and AWS operates that revert for you.
- The data plane auto-reverts only on EKS Auto Mode. On managed node groups and self-managed nodes, you roll the nodes back yourself with
UpdateNodegroupVersionor your own process.
That split is the whole story, because of one Kubernetes rule most upgrade runbooks never had to think about.
Step 1: Upgrade the control plane on its own
The way to keep the rollback available is to stop upgrading the control plane and the data plane in the same maintenance window. Upgrade the control plane first, and only the control plane:
aws eks update-cluster-version \ --name prod \ --kubernetes-version 1.33
That returns an update object you can poll. The
status moves through InProgress to Successful :aws eks describe-update \ --name prod \ --update-id b5f0ba18-9a87-4450-b5a0-825e6e84496f \ --query 'update.{status:status,type:type}'
{ "status": "Successful", "type": "VersionUpdate"}
At this point the control plane is on 1.33 and every node is still on 1.32. The cluster is fully reversible.
Step 2: Bake, and watch the rollback window
Let the new control plane run against real traffic for a few days. This is the part teams compress away to save a window, and it's exactly the part that makes the rollback worth having. While you bake, the revert path is open and nothing on the data plane has moved.
Before it lets you revert, EKS runs rollback readiness checks and surfaces them as insights. They cover API compatibility, API field changes, cluster health, kubelet version skew, kube-proxy compatibility, and add-on version compatibility. If a workload uses an API version that was removed between N-1 and N, the readiness check flags it as an error rather than letting you roll into a broken state.
If something does go wrong in the bake window, the rollback is the same shape of call as the upgrade:
aws eks update-cluster-version \ --name prod \ --kubernetes-version 1.32
Tip
There's a
--force flag that bypasses the readiness insights if you want to override a warning you understand. Know its limit: it skips the advisory checks, but it does not override PodDisruptionBudgets or pod-level disruption controls. Force is for "I've read the warning and accept it," not for "make the impossible happen."The kubelet skew rule that closes the door
Here's the constraint that reshapes everything. Kubernetes enforces a version skew between the control plane and the kubelet on each node. A kubelet may run several minor versions behind the API server. On EKS 1.28 and above that tolerance is three minors behind. It may never run ahead.
Now trace a rushed upgrade against that rule:
- You upgrade the control plane to N+1. Reversible.
- You upgrade the node group in the same window. Nodes recycle and the kubelet comes up on N+1.
- You try a control-plane-only rollback to N. The kubelet is now newer than the control plane you're asking for.
- The skew check refuses. To roll the control plane back, you'd have to roll the nodes back too.
The feature is still technically present. You've just made yourself ineligible for it by moving both planes at once.
Step 3: Recycle the data plane, on purpose
Only once you're confident in the new control plane do you upgrade the nodes, and at that point you're committing deliberately:
aws eks update-nodegroup-version \ --cluster-name prod \ --nodegroup-name workers \ --kubernetes-version 1.33
On EKS Auto Mode the ordering is handled for you. When you initiate a control-plane rollback there, EKS rolls the worker nodes back first, then the control plane, and it respects your NodePool disruption budgets, PodDisruptionBudgets, and the
karpenter.sh/do-not-disrupt annotation while doing it. Auto Mode also lets you cancel an in-progress rollback:aws eks cancel-update \ --name prod \ --update-id <update-id> \ --region <aws-region>
Standard clusters don't support cancellation, because their control-plane rollback completes quickly.
The Terraform timeout that marks a healthy rollback as failed
This one has nothing to do with Kubernetes and will catch IaC users specifically. By default, EKS gives a rollback operation up to 7 days to finish before marking it failed. That's sane for a human in the console. It's wrong for a Terraform apply, which times out at 24 hours, or CloudFormation, which gives up at 36. Your IaC tool declares the rollback failed and starts unwinding while EKS is still happily rolling back.
The fix is one field. Cap the rollback's own timeout under your IaC tool's:
aws eks update-cluster-version \ --name prod \ --kubernetes-version 1.33 \ --rollback-config timeoutMinutes=1440
1440 is 24 hours, which sits inside a default Terraform apply. Pick whatever fits the timeout your pipeline actually runs with, and the two systems stop fighting each other.Warning
Rollback supports one minor version, N to N-1 only. Multi-version rollbacks are not supported, both the rollback and its readiness insights expire 7 days after the upgrade, and you can only roll back to a version EKS still supports. Upgrade off an expiring release and the door is already closed.
Honest trade-offs
This isn't a free undo, and it's worth being clear about what it doesn't buy you.
It rewards a habit, not a click. If your upgrades already look like "bump the control plane and the node group in one Terragrunt run on a Saturday," this feature does nothing for you until you split them. The rollback only exists in the gap between the two steps, and the phased upgrade costs you a longer overall change window.
The 7-day clock is short. For a cluster you touch rarely, a regression can surface after the window has closed, and then you're back to blue/green anyway. The rollback is insurance against fast-surfacing problems, not slow ones.
And the data-plane story on managed node groups is still manual. If you're not on Auto Mode, "roll back the cluster" is really "roll back the control plane, then separately reconcile the nodes yourself." Plan the node rollback as its own step, not an afterthought.
Wrapping up
On the clusters I run, this release changes one thing concretely: the control-plane upgrade becomes its own change, with a bake period measured in days on real traffic, disruption budgets kept tight so a data-plane recycle can't stampede, and a
--rollback-config timeout that matches the pipeline instead of the 7-day default. The phased upgrade isn't new. What's new is that being careful now comes with a parachute, and rushing the upgrade is the one reliable way to leave it on the ground.Series: EKS (3 articles)
- 2EKS can roll back an upgrade now. This article
Enjoyed reading this content? Let the author know!
Your likes, comments, shares, and saves help creators reach more builders.
Loading recommendations
Loading article