Pausing and reverting nodes
Sometimes one node should be left alone: it is running a long batch job, or it is the node you are debugging. Two annotations on the NodeClaim control this. Only their presence matters, so any value works.
# Leave this NodeClaim exactly as it is
kubectl annotate nodeclaim default-8vbh7 bestbefore.k-belt.io/paused=true
# Undo k-belt's drift as well, then leave it alone
kubectl annotate nodeclaim default-8vbh7 bestbefore.k-belt.io/revert=true
# Hand it back to k-belt
kubectl annotate nodeclaim default-8vbh7 bestbefore.k-belt.io/revert-
paused freezes the NodeClaim in whatever state it is in. If k-belt had already marked it, it stays
marked and Karpenter will still replace it. Pause a node before it reaches maxAge and it is
never marked at all.
revert does what paused does and first puts the original hash back, so Karpenter clears the
drift. Use it when a node has been marked but you want to keep it.
Removing the annotation hands the NodeClaim back. If it is still older than maxAge, k-belt marks
it again within seconds, so remove the annotation only when you are ready for the node to go.
Reverting cannot stop a replacement Karpenter has already started. Once the node has the
karpenter.sh/disruptedtaint, it is going away. Check withkubectl get nodeclaim <name> -o jsonpath='{.status.conditions[?(@.type=="DisruptionReason")]}'.
Paused and reverted NodeClaims still count in a policy’s matched and stale totals, which is why
STALE can sit above DRIFTED indefinitely on a cluster where you have paused a few nodes.
Pausing a whole pool
There is no pause field on the policy. To stop rotations for a NodePool without deleting the policy,
set its Drifted budget to zero:
spec:
disruption:
budgets:
- nodes: "0"
reasons: [Drifted]
k-belt keeps marking nodes, Karpenter stops acting on them, and removing the budget resumes the rollout where it left off. Deleting the BestBefore instead un-marks every node it owns, which is the right move if you want to abandon the rotation rather than postpone it.