CPU threshold

Learn to scale based on CPU percentage.

The CPU threshold

The CPU percentage is calculated relative to the pod's request (requests.cpu), not the node capacity. That's why CPU requests must be written in the target containers for the HPA to work; otherwise <unknown> is seen in the TARGETS column:

resources:
  requests:
    cpu: 200m
# if the pod uses 150m, utilization is 75%

$ kubectl get hpa
NAME   TARGETS              REPLICAS
api    cpu: <unknown>/70%   2         # no requests!

Choosing the threshold

If the target is too high (95%), under a sudden load the existing pods are overwhelmed until new pods are ready; if too low (30%), resources are wasted. The longer a new pod takes to start, the lower the target is chosen. 60–75% is a commonly used range.

$ kubectl describe hpa api | grep -A3 Conditions
  AbleToScale     True    ReadyForNewScale
  ScalingActive   True    ValidMetricFound
  ScalingLimited  False   DesiredWithinRange

Exercise A container with requests

Write a Deployment's container section (starting with containers:): container named api, image api:1.0; resources.requests cpu: 200m, memory: 256Mi; resources.limits cpu: "1".

Show example solution

containers:
  - name: api
    image: api:1.0
    resources:
      requests:
        cpu: 200m
        memory: 256Mi
      limits:
        cpu: "1"

Test yourself

What is the CPU utilization of a pod with requests.cpu: 500m using 400m?
- 40%
- 125%
- 20%
- 80%
Show answer

80% — 400 / 500 = 0.8.

Write a CPU request so the HPA can calculate.

resources:
  ____:
    cpu: 200m

Show answer

resources:
  requests:
    cpu: 200m

Match.

requests.cpu

Percentage base
<unknown>

No request or metric
high target

Overload on spikes
low target

Wasted resources

Next lesson: Min and max replicas

We Are Enhancing Your Experience

Cookie Preferences

Strictly Necessary Cookies

Analytical Cookies