analyze-perfetto-trace

Use the perfetto Python library's TraceProcessor to load the .perfetto-trace file, then run SQL queries against the slice table. Query name and dur columns ordered by duration to find the largest spans, and divide dur by 1e6 to convert nanoseconds to milliseconds.

Start with a top-level slice duration query ordered by dur descending, then drill into the largest phase by listing its direct children. Compare the sum of children to the parent duration; a large gap indicates untraced work that needs new trace instrumentation.

Run ./metrow trace with a target project directory and Gradle compile task, which publishes Metro locally, enables tracing, and copies the trace to tmp/traces. For benchmark iteration, use benchmark/trace_compile.sh against the generated benchmark project.

The perfetto package installs against whichever Python pip points to, which may differ from the interpreter you invoke. On this setup it is installed for Python 3.13, so run queries with the python3.13 binary and verify with an import check first.

No, FIR and IR traces are separate files with independent timelines, so durations cannot be directly compared across them. Group files by their shared timestamp prefix and analyze the fir-* or ir-* file matching the phase you are investigating.