I run an AWS EKS cluster (v1.24) with a couple c5a.xlarge worker nodes and VPCCNI networking, with ENI trunking enabled and ENABLE_POD_ENI
set to true in the VPCCNI configuration (some of the pods need specific security groups, hence the need for assigning native ENIs).
Now, the problem is that the subnet that the nodes are in is rapidly running out of IP addresses - each node, even when it has only the aws-node
, ebs-csi-node
/efs-csi-node
and kube-proxy
pod running, has three ENIs active: the primary interface with 28 IPs, one named aws-k8s-trunk-eni
with one IP (probably the trunk controller?, and a third one named aws-K8S-i-xxxxxxxxx
with again 28 IPs.
I have not enabled ENABLE_PREFIX_DELEGATION
, and the IPs also don't look contiguous - what causes this absurd waste?