2. Scope

2.1 Scope and Compatibility

Supported environments

Environment	Status
EKS Auto Mode	Primary target
Self-managed Karpenter v1+ on EKS	Supported
Karpenter on other CNCF (AKS NAP, etc.)	Best-effort

• EKS Auto Mode: the 21-day hard cap is the strongest motivating constraint.
• Other CNCF distributions: the CRD API is the same, but underlying operator behavior may differ.

Karpenter compatibility policy

karpenter.sh/v1 is required; earlier versions (v1beta1, v1alpha5) are not supported.

The compatibility contract is the stable karpenter.sh/v1 CRD surface — not a specific Karpenter controller minor. This matters for EKS Auto Mode, which does not expose the exact managed Karpenter minor to users.

• Runtime target: any cluster serving a compatible karpenter.sh/v1 NodePool/NodeClaim API
• Build/test baseline: the bundled sigs.k8s.io/karpenter Go module version in go.mod (currently v1.13.0). This pins the typed Go API, not a runtime requirement
• Interaction boundary: solely through Kubernetes API objects (NodeClaim/NodePool CRDs, plus core Node/Pod). No Karpenter internals, no cloud-provider API
• Runtime enforcement: a startup preflight (§5.1) fails fast when the cluster does not serve karpenter.sh/v1 or RBAC cannot read it

Required compatibility surface

The controller depends only on the following public karpenter.sh/v1 fields:

Kind / field / key	Used for
NodeClaim, NodePool	Rotation unit and owning pool
NodeClaim.spec.expireAfter	Per-node deadline (§3.2)
NodeClaim.spec.terminationGracePeriod	Drain bound for t_rot (§3.2)
NodeClaim.spec.requirements	Placeholder requirement fallback (§3.3)
NodeClaim.status.nodeName	Claim → Node mapping (§3.3, §5.2)
NodeClaim.status.conditions[Ready]	Selection eligibility (§3.2)
NodePool.spec.template.spec.expireAfter	Representative E for validation
NodePool.spec.template.spec.terminationGracePeriod	Representative tGP
NodePool.spec.template.spec.requirements	Placeholder requirement replication (§3.3)
NodePool.spec.template.spec.taints	Placeholder tolerations (§3.3)
NodePool.spec.limits	Surge headroom check (§3.2, §5.2)
NodePool.status.resources	Provisioned footprint for headroom
label karpenter.sh/nodepool	Pairing nodes/placeholder to pool
annotation karpenter.sh/do-not-disrupt	Freeze surge pair (§3.3)

Anything outside this set — including all Karpenter controller internals — is irrelevant to compatibility.

2.2 Composition with Existing Mechanisms

Mechanism	Relationship
Consolidation / Drift	Coexists
NodePool expireAfter	Coexists as backstop
terminationGracePeriod	Depended on
PodDisruptionBudget	Depended on
topologySpreadConstraints	Depended on

• Consolidation / Drift: the controller takes over only the Expiration path. Voluntary disruption from Consolidation/Drift still flows through Karpenter.
• expireAfter: the derived ageThreshold sits below expireAfter by construction (A = E − (K·P + t_rot), §3.2). Validation fails (fatal) when the schedule cannot guarantee the configured rotation chances — the gap is not hand-tuned.
• terminationGracePeriod: after the controller deletes an old NodeClaim, Karpenter's termination controller honors PDBs during drain, bounded by terminationGracePeriod.
• PodDisruptionBudget: drain follows the voluntary path, so PDBs are strictly respected.
• topologySpreadConstraints: even with surge, all pods on a node disappear together when that node is drained. Spreading remains essential.