Datasets:
roanbrasil
/
devops-gitops-corpus

Dataset card Data Studio Files
xet
 Community
1
text
stringlengths
0
59.1k
kubectl get --raw "/apis/custom.metrics.k8s.io/v1beta1" | jq .
```
The output should be a list of available custom metrics. Look for the `pods/gpu_utilization_percent` metric, which confirms that the entire pipeline is working correctly and the metric is ready for the HPA to consume.
```json
{
"kind": "APIResourceList",
"apiVersion": "v1",
"groupVersion": "custom.metrics.k8s.io/v1beta1",
"resources": [
{
"name": "pods/gpu_utilization_percent",
"singularName": "",
"namespaced": true,
"kind": "MetricValueList",
"verbs": [
"get"
]
}
]
}
```
### 3.3. Deploy the Horizontal Pod Autoscaler (HPA)
The HPA is configured to use the final, clean metric name, `gpu_utilization_percent`, to maintain an average GPU utilization of 20%.
```bash
kubectl apply -f ./gpu-horizontal-pod-autoscaler.yaml -n vllm-example
```
Inspect the HPA's configuration to confirm it's targeting the correct metric.
```bash
kubectl describe hpa/gemma-server-gpu-hpa -n vllm-example
# Expected output should include:
# Metrics: ( current / target )
# "gpu_utilization_percent" on pods: <current value> / 20
```
---
## 4. Load Test the Autoscaling Setup
Generate a sustained load on the vLLM server to cause GPU utilization to rise.
### 4.1. Generate Inference Load
First, establish a port-forward to the vLLM service.
```bash
kubectl port-forward service/vllm-service -n vllm-example 8081:8081
```
In another terminal, execute the `request-looper.sh` script.
```bash
./request-looper.sh
```
### 4.2. Observe the HPA Scaling the Deployment
While the load script is running, monitor the HPA's behavior.
```bash
# See the HPA's metric values and scaling events
kubectl describe hpa/gemma-server-gpu-hpa -n vllm-example
# Watch the number of deployment replicas increase
kubectl get deploy/vllm-gemma-deployment -n vllm-example -w
```
As the average GPU utilization exceeds the 20% target, the HPA will scale up the deployment.
---
## 5. Cleanup
To tear down the resources from this exercise, run the following command:
```bash
kubectl delete -f . -n vllm-example
```
<|endoftext|>
# source: k8s_examples/AI/vllm-deployment/hpa/gpu-dcgm-exporter-service-gke.yaml type: yaml
# This Service provides a stable network endpoint for the NVIDIA DCGM Exporter
# pods. The Prometheus Operator's ServiceMonitor will target this Service
# to discover and scrape the GPU metrics. This is especially important
# because the exporter pods are part of a DaemonSet, and their IPs can change.
#
# NOTE: This configuration is specific to GKE, which automatically deploys the
# DCGM exporter in the 'gke-managed-system' namespace. For other cloud
# providers or on-premise clusters, you would need to deploy your own DCGM
# exporter (e.g., via a Helm chart) and update this Service's 'namespace'
# and 'labels' to match your deployment.
apiVersion: v1
kind: Service
metadata:
name: gke-managed-dcgm-exporter
# GKE-SPECIFIC: GKE deploys its managed DCGM exporter in this namespace.
# On other platforms, this would be the namespace where you deploy the exporter.
namespace: gke-managed-system
labels: