FAQ ALB ip vs instance mode rationale incorrect?

[anguslees]

How do I use Karpenter with the AWS load balancer controller?

Set the ALB target type to IP mode for the pods. Use IP targeting if you want the pods to receive equal weight. Instance balancing could greatly skew the traffic being sent to a node without also managing host spread of the workload.

This sounds incorrect? I think instance-mode will send traffic equally to all nodes (to a NodePort), and thence equally to all pods (that are ready and match the service selector).

Maybe this section is confusing instance-mode with internalTrafficPolicy=Local ?

[ellistarn]

Hey Gus! Feel free to cut a PR. I can check in with @chrisnegus on where this information came from.

[jicowan]

@anguslees it probably depends on how pods are distributed across the nodes in the cluster. If you don't have a pod anti-affinity policy or are not using topology spread constraints, multiple pods in a Deployment can wind up running on a small number of instances. IP targets are generally recommended because it avoids routing traffic through kube-proxy (> overhead). Unless you're using IPVS, kube-proxy will always default to RR for routing traffic so, assuming an even distribution of pods and a traffic policy of cluster, each pod should have equal weight when the target type is instance.

[ellistarn]

routing traffic through kube-proxy (> overhead)

Curious if we have benchmarks on this. IPTables should be pretty fast at most scales.

[jicowan]

Using IP targets can preserve the source IP in the HTTP header, making it easier to trace back to the source of the traffic. While I think the recommendation is sound (use IP targets) the rationale might need to be updated.

[anguslees]

it probably depends on how pods are distributed across the nodes in the cluster.

And just to be clear, I'm saying instance mode (without internalTrafficPolicy=Local) will send an equal number of queries to pods even when those pods are packed unevenly onto nodes.

(Because the pod-level assignment is actually done by kube-proxy and iptables RR as you described)

[chrisnegus]

We decided that this type of content is best suited for the AWS load balancer documentation. So I removed the content from the Karpenter FAQ (see PR #2245). I then opened an issue for the ALB docs (see Issue 2747).