Advanced EKS Topics
A small EKS cluster mostly works with defaults. The topics here are what you run into as it grows: address space runs out, scaling is too slow or too jumpy, and the bill needs explaining. Each is a system of interacting settings rather than a single feature, which is why they get their own notes.
Read the fundamentals first: networking, compute and observability.
Notes
| Note | Start here when |
|---|---|
| Advanced Networking | Pods cannot get IPs, the cluster must be fully private, CIDRs overlap, or egress must be controlled |
| Advanced Autoscaling | Scale-up is too slow for your traffic, or the pod and node autoscalers fight each other |
| Cost Optimization | You need to know what the cluster costs per team, and how to reduce it without hurting reliability |
How the three interact
They are not independent, and optimising one in isolation tends to damage another:
- Autoscaling and cost. Aggressive consolidation saves money and increases pod disruption. Generous headroom makes scaling fast and costs idle capacity.
- Networking and autoscaling. A scale-out burst needs IP addresses and ENI attachments immediately; a subnet that is nearly full, or warm-pool settings tuned too tightly, turn a scaling event into
Pendingpods. - Cost and networking. Cross-zone traffic, NAT gateway data processing and VPC endpoints are frequently a larger line item than people expect, and are decided by topology, not by instance choice.
A sensible order: make scaling correct (right-sized requests, PodDisruptionBudgets, limits), then make it fast enough, and only then make it cheap — with Spot, Graviton and consolidation, once you can see the effect of each change.
Prerequisites that pay off everywhere
| Have this in place | Because |
|---|---|
| Resource requests on every container | Scheduling, autoscaling and cost allocation all depend on them |
| PodDisruptionBudgets and more than one replica | Every optimisation here moves pods around |
| Metrics for IP usage, pending pods, node utilisation and cost per namespace | You cannot tune what you cannot see |
| Infrastructure and add-ons defined as code | These settings need to be reproducible and reviewable |
Related
- EKS implementation track
- Karpenter and Spot
- VPC CNI
- Non-functional guides — the provider-neutral versions of these topics
- AWS cost management