Datasets:
roanbrasil
/
devops-gitops-corpus

Dataset card Data Studio Files
xet
 Community
1
text
stringlengths
0
59.1k
<|endoftext|>
# source: k8s_examples/AI/vllm-deployment/hpa/horizontal-pod-autoscaler.yaml type: yaml
# This HorizontalPodAutoscaler (HPA) targets the vLLM deployment and scales
# it based on the average number of concurrent requests across all pods.
# It uses the custom metric 'vllm_num_requests_running', which is provided
# by the Prometheus Adapter.
apiVersion: autoscaling/v2
kind: HorizontalPodAutoscaler
metadata:
name: gemma-server-hpa
spec:
# scaleTargetRef points the HPA to the deployment it needs to scale.
scaleTargetRef:
apiVersion: apps/v1
kind: Deployment
name: vllm-gemma-deployment
minReplicas: 1
maxReplicas: 5
metrics:
- type: Pods
pods:
metric:
# This is the custom metric that the HPA will query.
# IMPORTANT: This name ('vllm_num_requests_running') is not the raw metric
# from the vLLM server. It is the clean, renamed metric that is
# exposed by the Prometheus Adapter. The names must match exactly.
name: vllm_num_requests_running
target:
type: AverageValue
# This is the target value for the metric. The HPA will add or remove
# pods to keep the average number of running requests per pod at 4.
averageValue: 4
behavior:
# The scaling behavior can be customized to control how quickly the
# deployment scales up or down.
scaleDown:
# The stabilizationWindowSeconds is set to 30 to prevent the HPA from
# scaling down too aggressively. This means the controller will wait for
# 30 seconds after a scale-down event before considering another one.
# This helps to smooth out the scaling behavior and prevent "flapping"
# (rapidly scaling up and down). A larger value will make the scaling
# more conservative, which can be useful for workloads with fluctuating
# metrics, but it may also result in higher costs if the resources are
# not released quickly after a load decrease.
stabilizationWindowSeconds: 30
policies:
- type: Percent
value: 100
periodSeconds: 15
<|endoftext|>
# source: k8s_examples/AI/vllm-deployment/hpa/gpu-service-monitor-gke.yaml type: yaml
# This ServiceMonitor tells the Prometheus Operator how to discover and scrape
# metrics from the NVIDIA DCGM Exporter. It is designed to find the
# 'gke-managed-dcgm-exporter' Service in the 'gke-managed-system' namespace
# and scrape its '/metrics' endpoint.
apiVersion: monitoring.coreos.com/v1
kind: ServiceMonitor
metadata:
name: nvidia-dcgm-exporter-servicemonitor
namespace: monitoring
labels:
# This label is used by the Prometheus Operator to discover this
# ServiceMonitor. It must match the 'serviceMonitorSelector' configured
# in the Prometheus custom resource.
release: prometheus
spec:
# This selector identifies the specific Service to scrape. It must match
# the labels on the 'gke-managed-dcgm-exporter' Service.
selector:
matchLabels:
# GKE-SPECIFIC: This label matches the Service for GKE's managed DCGM
# exporter. If you are using a different DCGM daemon, you must
# update this label to match the label of the corresponding Service.
app.kubernetes.io/name: gke-managed-dcgm-exporter
# This selector specifies which namespace to search for the target Service.
# For GKE, the DCGM service is in 'gke-managed-system'.
namespaceSelector:
matchNames:
# GKE-SPECIFIC: This is the namespace for GKE's managed DCGM exporter.
# For other environments, this should be the namespace where you have
# deployed the DCGM exporter Service.
- gke-managed-system
endpoints:
- port: metrics
interval: 15s
<|endoftext|>
# source: k8s_examples/AI/vllm-deployment/hpa/vllm-hpa.md type: docs
# Autoscaling an AI Inference Server with HPA using vLLM Server Metrics
This guide provides a comprehensive walkthrough for configuring a Kubernetes Horizontal Pod Autoscaler (HPA) to dynamically scale a vLLM AI inference server. The autoscaling logic is driven by a custom metric, `vllm:num_requests_running`, which is exposed directly by the vLLM server. This approach allows the system to ...
This guide assumes you have already deployed the vLLM inference server from the [parent directory's exercise](../README.md) into the `vllm-example` namespace.
---
## 1. Verify vLLM Server Metrics