OpenTelemetry eBPF Profiler with Pyroscope

This example demonstrates system-wide continuous profiling using the OpenTelemetry eBPF profiler, with profiles exported to Pyroscope and visualized in Grafana.

Linux only (amd64/arm64). The eBPF profiler requires the Linux kernel and privileged access to system resources (/proc, /sys/kernel, cgroups).

Architecture

flowchart LR
    A["OTel eBPF Profiler
<i>collects CPU profiles
system-wide via eBPF</i>"] -- "OTLP gRPC" --> B["Pyroscope
<i>stores & aggregates
profiling data</i>"]
    B -- query --> C["Grafana
<i>visualize as
flamegraphs</i>"]

The profiler runs with host PID namespace access and collects CPU profiles at 97 samples/second from all processes on the host. Profiles are sent to Pyroscope via OTLP gRPC, where process.executable.name is relabeled to service_name.

Docker Compose

Prerequisites

• Docker with Compose v2
• Linux host (amd64 or arm64)

Run

cd docker
docker compose up

Services:

• Grafana: http://localhost:3000 (anonymous auth, no login required)
• Pyroscope: http://localhost:4040

To stop:

docker compose down

Kubernetes

Prerequisites

• A Kubernetes cluster (e.g., minikube, kind)
• kubectl with kustomize support

Deploy

kubectl apply -k kubernetes/

This deploys:

• otel-ebpf-profiler DaemonSet -- runs the profiler on every node (privileged, hostPID)
• pyroscope Deployment + Service -- stores profiling data
• grafana Deployment + Service -- visualizes profiles (pre-configured with Pyroscope datasource)
• cpu-stress Deployment -- sample workload to generate visible profiles
• RBAC resources for the profiler's Kubernetes metadata enrichment

Access

kubectl port-forward svc/grafana 3000:3000

Then open http://localhost:3000.

Clean up

kubectl delete -k kubernetes/

Configuration

Example output