Grafana Pyroscope — Continuous Profiling Backend

Introduction

Pyroscope is a continuous profiling backend that records CPU, memory, goroutine, and lock profiles from production services and lets you slice them by label like a time-series. After joining the Grafana Labs family, it was merged with Phlare into a single Grafana-integrated backend that supports both pull (pprof-scrape) and push (agent SDK) ingestion.

What Pyroscope Does

• Collects pprof-compatible profiles from Go, Java, Python, Ruby, .NET, Node.js, Rust, and eBPF.
• Provides a flamegraph-first UI with "diff" mode for comparing time ranges or deployments.
• Exposes profiles under labels (service, region, deployment) like Prometheus time series.
• Integrates natively with Grafana for dashboards combining metrics, logs, traces, and profiles.
• Supports "span profiles" — traces linked to the specific flamegraph frames that ran in them.

Architecture Overview

Pyroscope v1 is a horizontally-scalable backend derived from Grafana Phlare. Distributors accept push ingestion; Ingesters build time-partitioned profile blocks in object storage (S3/GCS/Azure); Queriers fan out reads across block metadata, with Compactors merging blocks for long-term retention. A single-binary mode collapses all roles for dev and small clusters. On the agent side, Grafana Alloy (or the standalone Pyroscope agent) handles scraping pprof endpoints and running eBPF profilers node-wide.

Self-Hosting & Configuration

• Helm chart: helm install pyroscope grafana/pyroscope (single-binary by default).
• Configure storage in pyroscope.yaml — set storage.backend=s3 and an S3 bucket for prod.
• Push SDKs auto-profile at 100 Hz by default; tune SampleRate for heavy workloads.
• eBPF profiler DaemonSet runs CPU profiles across a node for uninstrumented languages.
• Grafana plugin ships built-in — add Pyroscope as a data source to get flamegraphs.

Key Features

• pprof-compatible format — dump, diff, and inspect with standard tooling.
• Label-based profile queries with PromQL-like selectors.
• eBPF whole-node profiling without touching application code.
• Trace-to-profile linking: click a trace span to see the exact flamegraph for that span.
• Scales from single binary to multi-service HA on object storage.

Comparison with Similar Tools

• Parca — open-source eBPF continuous profiling; Pyroscope is broader in language support.
• Datadog Continuous Profiler — SaaS, similar capabilities; Pyroscope is self-hostable.
• Polar Signals Cloud — managed Parca; same idea, hosted.
• elastic/otel-profiling-agent — eBPF profiler feeding Elastic; Pyroscope supports pprof and OTel profiles.
• Classic pprof — CLI only; Pyroscope makes it a continuous time-series system.

FAQ

Q: Do I have to instrument code to use it? A: Not for CPU: the eBPF profiler covers Go, Rust, C/C++, Python, Java, .NET without code changes.

Q: How much storage does it use? A: Profiles are small and compress well; typical services cost far less than their metrics stream.

Q: Can I diff two deployments? A: Yes, the UI supports flamegraph diff across arbitrary label sets.

Q: Is it part of Grafana Cloud? A: Yes; the open-source code here is the same engine that powers the Grafana Cloud Profiles product.

Sources

• https://github.com/grafana/pyroscope
• https://grafana.com/docs/pyroscope/latest/