Multi-zone GKE clusters paying for cross-zone traffic, a planned check with no findings yet

Zone hops are metered traffic

VPC network pricing charges nothing for traffic inside a zone on internal addresses, but $0.01 per GiB for data sent to a different zone in the same region. The cost is billed to the project of the sending VM, and Google names GKE nodes among the VMs this applies to.

A regional or multi-zonal cluster spreads pods across zones for resilience. By default a Kubernetes Service sends traffic evenly to all endpoints, so in a three-zone cluster roughly two of every three calls between services leave the caller’s zone. Chatty microservices turn that into a steady line item.

Checking zones and service routing

gcloud container clusters list --format="table(name,location,locations.list())"
kubectl get services -A -o custom-columns=NS:.metadata.namespace,NAME:.metadata.name,TD:.spec.trafficDistribution

locations is the list of zones that hold the cluster’s nodes. An empty traffic distribution column means the Service uses the default even spread.

What qualifies a cluster

The cluster must have nodes in more than one zone. Single-zone clusters, or clusters whose zone count ZopNight could not read, are skipped. This is a planned check: today it raises no finding on any cluster, as explained below.

When no saving is shown

Only billed inter-zone transfer is a defensible base. The cluster’s compute cost says nothing about how much traffic crosses zones, so ZopNight does not apply a fraction to it. ZopNight cannot yet separate billed inter-zone transfer from other Compute Engine charges, so for now the rule stays silent on every cluster. Use the commands above and your billing export to size it by hand.

The 30% reduction factor

Once billed inter-zone transfer can be separated, the planned saving is:

saving = billed inter-zone transfer cost x 0.30

The factor is conservative on purpose: a highly available cluster still needs some traffic to cross zones.

Keeping calls inside their zone

1. On each high-traffic Service set spec.trafficDistribution: PreferSameZone, described in the Kubernetes virtual IPs reference, or the older service.kubernetes.io/topology-mode: Auto annotation.
2. Make sure each zone has enough endpoints; with none available locally, traffic falls back to the whole cluster.
3. Use pod affinity to keep tightly coupled workloads together, and topology spread constraints to keep replicas balanced across zones.
4. Measure inter-zone transfer in billing before and after the change.