Profile-guided optimization (PGO) in the Go toolchain uses CPU pprof profiles as the PGO profile format. Though pprof is a widely-used format across many tools, Go's PGO imposes specific requirements on the contents of profiles, which many tools across the ecosystem may not be compatible with.

This (non-exhaustive) page lists tools for collecting and working with profiles that are known to be compatible with PGO.

Collecting profiles

  • runtime/pprof, net/http/pprof: The Go standard library profiling functionality always provides PGO-compatible profiles.
  • Parca Agent produces PGO-compatible profiles when paired with a symbolizer that produces metadata that includes function start lines, such as Polar Signals Cloud or Parca starting at version v0.19.0.

Working with profiles

  • go tool pprof/standalone pprof CLI, github.com/google/pprof/profile: The official pprof CLI and Go packages can perform various operations on profiles (filtering, merging multiple profiles, etc). These tools/packages generally leave metadata (e.g., symbolization, function start lines) intact across operations. Thus given a PGO-compatible input, they should produce a PGO-compatible output.
  • Parca and Polar Signals Cloud provide various mechanisms to query and filter profiling data and download any query as a pprof file that will contain function start line metadata to be PGO-compatible.