I'm currently transitioning from EKS (non-Auto) with VPC CNI and prefix delegation enabled to EKS Auto Mode. Right now, I have 3 t3a.medium nodes and I can run around 110 pods each, thanks to prefix delegation. My goal is to take advantage of EKS Auto Mode for easier management and scaling, but I've hit the wall with the 15-pod limit per node since prefix delegation isn't supported in Auto Mode. I want to know if there are any hidden tricks, workarounds, or suggestions for managing this situation without having to double my node count. Has anyone found success in using Auto Mode while achieving higher pod density?
4 Answers
Why not give Cluster Autoscaler a shot? It could help with your situation. Although I see your point—if IP addresses are your main limitation, scaling up nodes won’t address that cost-effectively. So keep an eye on your IP usage.
It’s a bit puzzling why you want to mix prefix delegation with Auto Mode. If you primarily need managed scaling and operations, installing Karpenter could help while maintaining your prefix delegation setup without switching to Auto Mode. Auto Mode mainly reduces your need for configuration and shifts the responsibility to AWS.
While prefix delegation isn’t supported in Auto Mode, going for a hack might not be reliable in the long run. If you have a Technical Account Manager (TAM), ask for insight from the EKS team about this limitation. They might have some valuable updates or suggestions for you. If you don’t have enterprise support, you could check the containers roadmap on GitHub and maybe open an issue for clarification.
Auto Mode does default to prefix delegation but switches to using secondary IPs when it can’t find a contiguous block for the prefix. If your VPC is a bit fragmented, that’s likely causing the issue. You can learn more about this detail in the EKS documentation for auto networking.
Yeah, I can see that. It seems like the IP addresses become the bottleneck here, not the resources themselves. So even if you scale up nodes, in the long run, it might not fit your needs.