Profiling Artifacts

Profiling Artifacts

This page is a stable reference for what profiling produces and how to interpret it.

Directory structure

Under profiling/results/:

• latest.json, latest.csv: aggregate summary of the most recent profiling run
• <suite>_<run_id>.json, <suite>_<run_id>.csv: per-suite summaries
• raw/
  • <run_id>_<scenario>_<iter>.prof: cProfile dumps
  • <run_id>/<scenario>/<iteration>/: computed outputs (e.g. CSV/Parquet) for e2e/write suites

Record fields (what you’ll see in JSON/CSV)

Each scenario produces a record with:

• run_id, timestamp
• suite, scenario, mode, engine
• status (completed / skipped / failed)
• elapsed_seconds
• phase_seconds (where applicable): source, calculation, write
• prof_path (path to the .prof file)

Cached vs live

• cached uses a synthetic fixture frame (deterministic, no network)
• live calls source_data(...) and depends on upstream data and dependencies

GPU scenarios

GPU scenarios are auto-skipped if Polars GPU collect is unavailable; the skip reason is recorded.

Related pages:

• Profiling Runner
• Profiling