First scheduled pod starts using the wrong CNI

[2ZZ]

Hi,

The first pods scheduled on a node can sometimes get an IP from the non-default CNI.
I think this is because CNI-Genie daemonset pod can start too late maybe due to delays in pulling the image.
I added the system-node-critical priorityClass to the CNI-Genie daemonset but it has not helped.

Example timeline:

• Nginx app pod triggers cluster scale up
• AWS-CNI, Calico and CNI-Genie daemonsets are scheduled on the new node
• Nginx pod starts up before GNI-Genie pod has finished starting so the config is not in /etc/cni/net.d at this point
• Nginx pod gets an IP from AWS-CNI instead of the default set in CNI-Genie
• Future pods on that new node are given correct IPs once CNI-Genie has started

Setup:
Cluster: Amazon EKS 1.18
Calico version: 3.16.3
CNI-Genie version: latest

[shinebayar-g]

That sounds possible scenario. We're using CNI-Genie for running Cilium + AWS VPC.

Fortunately we didn't observe this issue when upgrading worker versions.

[2ZZ]

Hi, I modified AWS CNI to wait for CNI-Genie to be present in /etc/cni and haven't seen this issue since.