EKS: known Terraform limitations

Some things people ask the terraform-aws-modules/eks/aws module for cannot be implemented by any module, in any registry. They are limits of Terraform itself. This page collects the recurring ones for this module, says plainly what causes each, gives the native workaround in full, and - where one exists - shows the Operational Rule that removes the need for a fork.

EKS node group desired_size changes on every plan

Cluster Autoscaler and Karpenter scale node groups by writing scaling_config.desired_size. Terraform reads the scaled value as drift and plans to put the configured number back, so every plan is dirty and every apply fights the autoscaler.

Why Terraform cannot fix this

Terraform's model is that it owns every attribute it can see. When a controller, an autoscaler, a scanner, or a deployment pipeline writes to one of those attributes, Terraform reads the new value as drift and plans to put its own value back. Both systems are behaving correctly; they simply disagree about who owns the field.

The only mechanism Terraform offers for splitting ownership is ignore_changes, and that is a lifecycle argument, which brings the literal-value restriction above with it. A module cannot accept the list of attributes to ignore as an input, so a module consumer has no way to declare the split.

The native workaround

The native fix is ignore_changes = [scaling_config[0].desired_size] on the node group, which is a lifecycle argument. Fork this module and add the lifecycle block to the resource yourself. That works, and it is the honest answer - it is what module maintainers do when they need it. The cost is ongoing rather than one-off: the fork has to be re-synced with every upstream release, and each sync needs a review to confirm the block still lands on the resource it was meant for. A wrapper module does not avoid this, because the lifecycle block still has to sit inside the resource block, which is inside the module you did not write.

The other native option is to stop setting desired_size where the API allows it, and let min and max bound the autoscaler instead - that avoids the disagreement rather than suppressing it, and is worth trying first.

Managed fix

Native option only, today. compliance.tf has no rule for this yet, so the workaround above is the whole answer. The rule that would answer this is not built yet.

Using the maintained alternative

compliance.tf serves this module from a registry that applies the rules above during terraform init. The change is the source line; the inputs and outputs are the upstream module's.

module "eks" {
  source = "https://cis.compliance.tf/terraform-aws-modules/eks/aws"

  # ... the same inputs you pass today ...
}

None of the limitations above has a rule yet, so this module is served with compliance controls only. The ?add_rules= parameter takes the ids of rules that do exist - see Operational Rule Definitions.

Related

• Operational Rules - what rules are, and how they differ from compliance controls
• Operational Rule Definitions - the HCL of every selectable rule
• Stop paying the Terraform fork tax - why forking a module to add a lifecycle block costs more than it looks
• EKS module reference - inputs, outputs, and the controls applied to it