Keyboard shortcuts

Press ← or → to navigate between chapters

Press S or / to search in the book

Press ? to show this help

Press Esc to hide this help

Profiling

Cadmus supports continuous profiling with Pyroscope when the profiling feature is enabled.

What it does

  • Uses jemalloc heap profiling for allocation data
  • Uses pprof for CPU profiling data
  • Pushes both profile types to Pyroscope

Feature flags

  • profiling enables profiling support in the codebase
  • telemetry enables profiling together with tracing

Runtime behavior

Profiling is initialized during app startup and shut down during app exit. If a Pyroscope endpoint is configured, Cadmus starts collecting profiles right away so early startup work is included.

The app configures jemalloc as the global allocator when profiling is enabled and enables heap profiling at process startup. CPU profiling runs alongside it through the Pyroscope agent.

Configuration

See the settings reference for the full logging and profiling configuration.

  • logging.pyroscope-endpoint
  • PYROSCOPE_SERVER_URL

PYROSCOPE_SERVER_URL overrides logging.pyroscope-endpoint when both are set.

To build Cadmus with profiling support:

cargo build --features profiling

To build Cadmus with tracing and profiling together:

cargo build --features telemetry

Profile types

Cadmus currently exports:

  • heap allocation profiles via jemalloc
  • CPU profiles via pprof-rs

Both profile streams are pushed to the configured Pyroscope server.

Local development

The cadmus-dev-otel command starts the emulator with profiling enabled and the local Pyroscope service available at http://localhost:4040.

With the full devenv stack running, traces go to Tempo, logs go to Loki, and profiles go to Pyroscope. That makes it possible to correlate a single run across all three observability backends.