kubernetes-yaml/learn/learn-kubernetes-master/kiamol/ch19/lab/README.md

73 lines
1.7 KiB
Markdown
Raw Normal View History

# ch19 lab
## Setup
Deploy metrics-server if you need it (check with `kubectl top nodes` - no stats means you need it):
```
kubectl apply -f metrics-server/
```
Run the app:
```
kubectl apply -f lab/pi/
```
Confirm the metrics are coming through:
```
kubectl top pods -l app=pi-web-lab
```
> Browse to the app and check the CPU spikes - e.g. http://localhost:8032/?dp=100000
## Sample Solution
You need to label your node to indicate it's in the EU region - you can use any key and value for this, but you'll need to use the same in your affinity rules:
```
kubectl label node --all kiamol.net/region=eu
```
### Pod with affinity rules
The updated deployment in [solution/pi.yaml](./solution/pi.yaml) adds these settings:
- **node affinity** - require to run on nodes with region=eu
- **pod anti-affinity** - prefer to run on nodes without any other Pi pods
- **resources** - add memory request for the HPA to use
- **replicas** - start with 2 as that's the desired minimum
```
kubectl apply -f lab/solution/pi.yaml
```
> You'll have two Pods running; browse to the app in a few tabs and both will spike CPU
### HPA for scaling on CPU
The HPA spec in [solution/hpa-cpu.yaml](./solution/hpa-cpu.yaml) scales from 2 to 5 Pods based on target CPU utilization of 50%.
```
kubectl apply -f lab/solution/hpa-cpu.yaml
```
> Make lots of browser requests in different tabs (or adapt the `ch19/loadpi` script in) and you'll see the Pods scale up to a maximum of five replicas:
![Horizontal pod autoscaling in action](./solution/hpa.png)
## Teardown
Delete all the resources:
```
kubectl delete all,hpa -l kiamol=ch19-lab
```
And metrics-server if you deployed it:
```
kubectl delete -f metrics-server/
```