Networking is still hard in 2026

Three things are true about a lot of EKS clusters using Karpenter (great choice!) as their scheduler:
1. One instance family, benchmarked and standardized across teams
2. Spot preferred, with On-Demand as the fallback
3. No zone-aware routing, because while all the capacity happens to live in one zone, "route anywhere" and "route locally" are the same instruction.

Then Spot capacity for that family runs out in the region. Karpenter does exactly what it was told: it falls back to On-Demand, and it places new nodes wherever it can still get instances. Within a day, a fleet that had effectively lived in one availability zone is spread across three.

Nothing about that is wrong. Spreading across zones is what you are supposed to want. Except the services still talk to each other constantly, and now roughly two thirds of every internal call crosses a zone boundary at $0.01 per GB in each direction.

The traffic pattern never changed. Only the placement did. And there is no Kubernetes signal for placement: no restarts, no failing probes, zero errors. In Cost Explorer it hides inside EC2-Other.

The full story is on the blog: how a NodePool quietly became a placement policy, why one instance family is the thing that set it up, and the four changes that stop it from happening again.