Managing GKE Workloads

What it is

Operating GKE means knowing what is running, letting nodes pull images from Artifact Registry, managing node pools, working with Kubernetes resources (Pods, Services, StatefulSets), configuring horizontal and vertical Pod autoscaling, setting resource requests in Autopilot, and splitting traffic between versions.

How it works

Viewing cluster inventory

kubectl get nodes
kubectl get pods --all-namespaces
kubectl get services
kubectl describe pod web-7d4b9c8f6-abcde
kubectl logs web-7d4b9c8f6-abcde

The GKE pages in the console show the same: clusters, workloads, Services and Ingress, with health and events.

Accessing Artifact Registry

GKE nodes pull images using the node service account. For images in Artifact Registry in the same project, the default setup usually works; otherwise grant the node service account Artifact Registry Reader (roles/artifactregistry.reader) on the repository. Image paths look like REGION-docker.pkg.dev/PROJECT/REPOSITORY/IMAGE:TAG.

Node pools

A node pool is a group of nodes with the same configuration in a Standard cluster.

gcloud container node-pools create gpu-pool --cluster=shop-std --region=us-central1 \
  --machine-type=n1-standard-8 --accelerator=type=nvidia-tesla-t4,count=1 \
  --enable-autoscaling --min-nodes=0 --max-nodes=3

gcloud container node-pools update default-pool --cluster=shop-std --region=us-central1 \
  --enable-autoscaling --min-nodes=1 --max-nodes=5

gcloud container node-pools delete old-pool --cluster=shop-std --region=us-central1

Use separate node pools for different hardware (GPUs, high memory) or Spot VMs. The cluster autoscaler adds and removes nodes in a pool based on Pods that cannot be scheduled.

Kubernetes resources

Pod autoscaling

kubectl autoscale deployment web --cpu-percent=60 --min=2 --max=10
kubectl get hpa

Do not let HPA and VPA both act on CPU or memory for the same workload; they conflict.

Autopilot resource requests

In Autopilot you pay for the CPU, memory and ephemeral storage that Pods request. Set requests in the Pod spec; if you do not, Autopilot applies defaults, and it adjusts requests to supported values. Right-sizing requests is the main cost lever in Autopilot.

Traffic splitting on GKE

To send a share of traffic to a new version, run two Deployments (for example v1 and v2) behind routing that supports weights, such as the Gateway API with weighted backends, or use a service mesh. Simply scaling replica counts gives only rough proportions.

A worked scenario

A team's GKE app slows at lunchtime. They add an HPA on CPU, enable cluster autoscaling on the node pool, and create a separate GPU node pool scaling from zero for image processing. A Redis-like cache moves to a StatefulSet with persistent volumes. New releases go to 10 percent of users through Gateway API weighted routing before full rollout.

How the exam asks about this

• "See all Pods in every namespace." kubectl get pods --all-namespaces.
• "Add GPU nodes to an existing cluster." Create a new node pool with accelerators.
• "Scale Pods with CPU load." Horizontal Pod Autoscaler (kubectl autoscale).
• "Stable identity and storage per Pod." StatefulSet.
• "Nodes cannot pull images from Artifact Registry." Grant the node service account Artifact Registry Reader.