---
title: Pyroscope
modified: 2026-03-26
tags:
  - service
  - observability
---

# Grafana Pyroscope

Continuous profiling backend for BlumeOps. Stores CPU profiles collected by Alloy's eBPF profiler on ringtail, providing function-level visibility into where compute time is spent.

## Quick Reference

| Property | Value |
|----------|-------|
| **URL** | https://pyroscope.tail8d86e.ts.net |
| **Namespace** | `pyroscope` |
| **Cluster** | ringtail (k3s) |
| **Deployment** | StatefulSet (`argocd/manifests/pyroscope/`) |
| **Image** | `grafana/pyroscope` |
| **Port** | 4040 |
| **Storage** | 10Gi PVC at `/data` |
| **Retention** | 7 days (`max_query_lookback: 168h`) |

## Architecture

Pyroscope runs on ringtail because eBPF profiling requires Linux. Grafana on indri queries it via Tailscale Ingress.

```
Alloy (pyroscope.ebpf on ringtail) → Pyroscope (ringtail) → Grafana (indri, via Tailscale)
```

## Collection

Profiles are collected by the `alloy-profiling-ringtail` DaemonSet, which runs the `pyroscope.ebpf` component in privileged mode with `hostPID: true`. It discovers Kubernetes pods automatically and excludes infrastructure namespaces (`kube-system`, `tailscale`) and Alloy pods.

The eBPF profiler works without application instrumentation — it samples CPU stack traces from the kernel, covering native code (Go, C/C++), interpreted languages (Python, Ruby, Node.js), and JIT-compiled runtimes (.NET).

**Limitations:**
- GPU workloads (e.g., Frigate inference via CUDA) are invisible to CPU profiling
- Stripped binaries (no debug symbols) produce opaque stack frames
- Python frame quality varies depending on runtime version

## Related

- [[alloy]] - Collection agent
- [[observability]] - Full observability stack overview
- [[grafana]] - Visualization
- [[tempo]] - Distributed tracing (cross-linked via traces-to-profiles)