GPU profiling frequently asked questions for Azure Kubernetes Service (AKS) (Preview)

Get answers to frequently asked questions about monitoring integrations and profile storage for GPU profiling on Azure Kubernetes Service (AKS). Before you begin, set up GPU profiling.

Monitoring integrations

How do I connect my own Prometheus instance to Inspektor Gadget?

If you run your own Prometheus instance instead of Azure Monitor managed service for Prometheus, configure it to scrape metrics from the Inspektor Gadget pods.

Inspektor Gadget exposes its metrics at the following endpoint:

Setting Value
Namespace gadget
Port 2224
Path /metrics

Add a scrape job to your Prometheus configuration to discover and scrape the pods:

Tip

For Prometheus Operator deployments, create a PodMonitor with the equivalent namespace, label selector, port, and metrics path.

scrape_configs:
  - job_name: "inspektor-gadget"
    kubernetes_sd_configs:
      - role: pod
        namespaces:
          names:
            - gadget
    relabel_configs:
      # Only keep pods with the k8s-app=gadget label.
      - source_labels: [__meta_kubernetes_pod_label_k8s_app]
        action: keep
        regex: gadget

      # Set the metrics path.
      - target_label: __metrics_path__
        replacement: /metrics

      # Set the metrics port.
      - source_labels: [__address__]
        action: replace
        regex: ([^:]+)(?::\d+)?
        replacement: $1:2224
        target_label: __address__

Verify the connection in your Prometheus UI under Status > Targets. The inspektor-gadget job should show its target as UP.

How do I visualize Inspektor Gadget metrics in my own Grafana instance?

Make sure your Prometheus instance is connected to Grafana as a data source. Then, import the Inspektor Gadget dashboard from the following JSON definition:

https://raw.githubusercontent.com/inspektor-gadget/grafana-dashboards/refs/heads/main/dashboards/gpu-observability/AdvancedGPUObservability.json

Profile storage

How do I use Azure Blob Storage as the Pyroscope backend?

By default, the Pyroscope deployment in the GPU profiling setup article stores profiles on the local filesystem. Profiling data is lost if the pod is recreated. To persist profiles, configure Pyroscope to use Azure Blob Storage with Microsoft Entra Workload ID.

Export the required variables:

export RESOURCE_GROUP="<your-resource-group>"
export AKS_CLUSTER="<your-aks-cluster-name>"
export LOCATION="<your-aks-cluster-location>"
export BLOB_STORAGE_NAME="<your-storage-account-name>"

export BLOB_STORAGE_CONTAINER="pyroscope-blob-storage"
export PYROSCOPE_IDENTITY="pyroscope-id"
export PYROSCOPE_SERVICE_ACCOUNT="pyroscope-sa"
export PYROSCOPE_FEDERATED_CREDS="pyroscope-federated-creds"

Create the storage account and container

az storage account create \
  --name "$BLOB_STORAGE_NAME" \
  --resource-group "$RESOURCE_GROUP" \
  --location "$LOCATION" \
  --sku Standard_ZRS \
  --kind StorageV2 \
  --min-tls-version TLS1_2 \
  --allow-blob-public-access false \
  --output none

az storage container create \
  --name "$BLOB_STORAGE_CONTAINER" \
  --account-name "$BLOB_STORAGE_NAME" \
  --auth-mode login \
  --output none

Create a managed identity and grant access to the storage account

az identity create \
  --name "$PYROSCOPE_IDENTITY" \
  --resource-group "$RESOURCE_GROUP" \
  --location "$LOCATION" \
  --output none

IDENTITY_CLIENT_ID=$(az identity show \
  --name "$PYROSCOPE_IDENTITY" \
  --resource-group "$RESOURCE_GROUP" \
  --query 'clientId' \
  --output tsv)

SUBSCRIPTION_ID=$(az account show --query id --output tsv)

# Restrict the role assignment scope to the storage account.
STORAGE_SCOPE="/subscriptions/${SUBSCRIPTION_ID}/resourceGroups/${RESOURCE_GROUP}/providers/Microsoft.Storage/storageAccounts/${BLOB_STORAGE_NAME}"

az role assignment create \
  --assignee "$IDENTITY_CLIENT_ID" \
  --role "Storage Blob Data Contributor" \
  --scope "$STORAGE_SCOPE" \
  --output none

Enable workload identity and create federated credentials

az aks update \
  --name "$AKS_CLUSTER" \
  --resource-group "$RESOURCE_GROUP" \
  --enable-oidc-issuer \
  --enable-workload-identity \
  --output none

OIDC_ISSUER=$(az aks show \
  --name "$AKS_CLUSTER" \
  --resource-group "$RESOURCE_GROUP" \
  --query 'oidcIssuerProfile.issuerUrl' \
  --output tsv)

az identity federated-credential create \
  --name "$PYROSCOPE_FEDERATED_CREDS" \
  --identity-name "$PYROSCOPE_IDENTITY" \
  --resource-group "$RESOURCE_GROUP" \
  --issuer "$OIDC_ISSUER" \
  --subject "system:serviceaccount:gadget:${PYROSCOPE_SERVICE_ACCOUNT}" \
  --audiences "api://AzureADTokenExchange" \
  --output none

Important

The service account name in the federated credential subject must match the service account name used by the Pyroscope Helm release. Otherwise, Pyroscope can't authenticate to Azure Blob Storage.

Deploy Pyroscope with the Azure Blob Storage backend

Instead of the Helm command in Enable profile visualization with Pyroscope, run the following command. This configuration uses Azure Blob Storage, creates the service account, and enables workload identity on the Pyroscope pod.

helm install pyroscope -n gadget \
  oci://ghcr.io/grafana/helm-charts/pyroscope \
  --version 1.15.0 \
  --set pyroscope.image.repository=grafana/pyroscope \
  --set-string pyroscope.image.tag=1.15.0 \
  --set pyroscope.replicaCount=1 \
  --set pyroscope.structuredConfig.self_profiling.disable_push=true \
  --set pyroscope.structuredConfig.storage.backend=azure \
  --set-string pyroscope.structuredConfig.storage.azure.account_name="$BLOB_STORAGE_NAME" \
  --set-string pyroscope.structuredConfig.storage.azure.container_name="$BLOB_STORAGE_CONTAINER" \
  --set pyroscope.serviceAccount.create=true \
  --set-string pyroscope.serviceAccount.name="$PYROSCOPE_SERVICE_ACCOUNT" \
  --set-string pyroscope.serviceAccount.annotations."azure\.workload\.identity/client-id"="$IDENTITY_CLIENT_ID" \
  --set-string pyroscope.podLabels."azure\.workload\.identity/use"=true \
  --set pyroscope.service.type=LoadBalancer \
  --set pyroscope.service.port=4040 \
  --set-string pyroscope.service.annotations."service\.beta\.kubernetes\.io/azure-load-balancer-internal"=true \
  --set-string pyroscope.service.annotations."service\.beta\.kubernetes\.io/azure-pls-create"=true \
  --set-string pyroscope.service.annotations."service\.beta\.kubernetes\.io/azure-pls-name"=pyroscope-pls \
  --set-string pyroscope.service.annotations."service\.beta\.kubernetes\.io/azure-pls-proxy-protocol"=false \
  --set-string pyroscope.service.annotations."service\.beta\.kubernetes\.io/azure-pls-visibility"='*' \
  --set alloy.enabled=false \
  --set minio.enabled=false

Verify that the pod is running:

kubectl get pods -n gadget pyroscope-0

Continue with Connect Pyroscope to Azure Managed Grafana.

Note

Helm value keys can change between chart versions. To confirm the available keys for your version, run helm show values oci://ghcr.io/grafana/helm-charts/pyroscope --version 1.15.0.