| name | perfetto-infra-getting-trace-processor |
| description | Use when the user does not yet have `trace_processor` installed, or has not set up the Python `perfetto` client used to drive the long-running RPC mode. Covers fetching the prebuilt binary and an opinionated venv-based install for the Python client. This is the OSS / public path; teams in restricted environments may have their own version of this skill that overrides it. |
Getting trace_processor and the Python client
This skill is one answer to the question "where does trace_processor
come from?". It is intentionally orthogonal to the
querying-traces skill: that one teaches what to do once you have a
working trace_processor; this one teaches how to obtain it.
There are many possible answers depending on the environment — Google
internal users, OEM build environments, CI images, and other teams
typically have their own preferred or mandatory path. If the user is
in such an environment, prefer their team-specific version of this
skill. Use this one when nothing more specific is available.
Part 1 — The trace_processor binary
Fetch the official prebuilt binary. This is what 99% of users should
do — a build from source is only needed when developing Perfetto
itself.
curl -LO https://get.perfetto.dev/trace_processor
chmod +x trace_processor
./trace_processor --version
get.perfetto.dev/trace_processor is a small wrapper script that picks
the right prebuilt binary for the host platform (Linux x86_64 / arm64,
macOS, Windows) and caches it. Re-running the curl -LO (or just
re-invoking ./trace_processor) is the recommended way to stay
current — the script is hash-idempotent, so it only re-downloads when
the upstream binary actually changes. Don't move it onto PATH as
a one-shot install: doing so encourages users to keep running a stale
binary indefinitely.
If a Perfetto source checkout is available, tools/trace_processor
inside that checkout does the same fetch under the hood and
additionally supports --build for a from-source build.
Part 2 — The Python client (for long-running RPC mode)
The querying-traces skill recommends starting trace_processor server http TRACE_FILE once and connecting from Python so iteration doesn't
re-parse the trace on every query. That requires the perfetto Python
package (and protobuf).
Default: an isolated venv
Modern macOS / Debian / Ubuntu Python installs (PEP 668) refuse global
pip install by default, and even where they don't, mixing the
Perfetto client into the system interpreter is a recipe for version
conflicts. The opinionated path is a dedicated venv:
python3 -m venv ~/.venv/perfetto
~/.venv/perfetto/bin/pip install -U perfetto protobuf
~/.venv/perfetto/bin/pip install -U pandas
~/.venv/perfetto/bin/python -c \
"from perfetto.trace_processor import TraceProcessor; print('ok')"
Then invoke Python from the querying skill as
~/.venv/perfetto/bin/python, or source ~/.venv/perfetto/bin/activate
once per shell and use plain python.
When to override the default
- The user already has an environment manager (
uv, pipx,
conda, poetry, an existing project venv): install perfetto and
protobuf into that environment instead of creating a fresh
~/.venv/perfetto.
- The user explicitly wants a global install:
pip install --break-system-packages perfetto protobuf is the escape hatch on
PEP-668 distros. Don't reach for it on the user's behalf — only
when they ask.
- The environment has its own packaging story (Google internal,
CI images with pre-baked deps, locked-down corporate Python): defer
to that. This skill is the OSS default, not a mandate.
Verifying the full setup
A one-shot end-to-end check that the binary and the client agree:
PORT=$((9100 + RANDOM % 900))
trace_processor server http --port $PORT /path/to/some_trace.pftrace &
SERVER_PID=$!
sleep 1
~/.venv/perfetto/bin/python - "$PORT" <<'PY'
import sys
from perfetto.trace_processor import TraceProcessor
tp = TraceProcessor(addr=f'127.0.0.1:{sys.argv[1]}')
print(next(iter(tp.query('SELECT count(*) AS c FROM slice'))).c)
tp.close()
PY
kill $SERVER_PID
If this prints a non-zero count and exits cleanly, the user is ready to
follow the querying-traces skill.
