metrics-server and kubectl top

metrics-server 6,745 (v0.9.0, installed in Autoscaling Pods and Nodes for the HPA) scrapes each kubelet every 15 seconds and serves the latest CPU and memory figures through the aggregated metrics.k8s.io API, with no history:

Current usage from the Metrics APIShell

kubectl top pods --sort-by=memory
kubectl get --raw /apis/metrics.k8s.io/v1beta1/namespaces/booknest/pods/postgres-0 \
  | jq -c '.containers[] | {name, usage}'

Output

NAME                  CPU(cores)   MEMORY(bytes)
postgres-0            95m          78Mi
load2                 156m         51Mi
...
web-ccd64df5f-jf87t   31m          6Mi
{"name":"postgres","usage":{"cpu":"94808104n","memory":"79912Ki"}}

Two test Pods (load, load2) were sending requests, and the HPA had scaled the API to six replicas. Memory is the working set, which the kubelet uses for evictions.

MENU

metrics-server and kubectl top