trace0

A low-overhead tracing profiler for Python, written in Rust.

trace0 hooks PEP 669 sys.monitoring, packs each event into 8 bytes in a thread-local buffer, and hands whole batches to a background thread over lock-free SPSC rings. It emits Perfetto-compatible traces — open them at https://ui.perfetto.dev.

How much does it cost?

Across 14 benchmarks from pyperformance, the suite CPython itself is judged on, run unmodified on an M4 Mac:

Median slowdown 1.19×, ranging 1.00× to 1.48×.

	slowdown
`telco`, `fannkuch`, `nbody`, `scimark`	1.00–1.04×	arithmetic in tight loops
`json_dumps`, `pyflate`, `chaos`, `float`, `go`, `spectral_norm`	1.12–1.25×
`raytrace`, `hexiom`, `deltablue`, `richards`	1.36–1.48×	call-heavy

The spread is the story: a tracer charges per call, so what a program pays depends entirely on how many calls it makes per unit of work, not on how long it runs. Reproduce with scripts/benchmark.py suite.

Against the alternatives, on a call-dense workload — transpiling SQL with sqlglot, 10.8M events per second of work. Reproduce with scripts/benchmark.py alternatives:

	wall time	vs vanilla
vanilla	0.97 s	—
trace0 (protobuf)	1.16 s	1.20×
trace0 (json)	1.35 s	1.40×
`cProfile`	2.36 s	2.45×
`profile`	19.11 s	19.8×

trace0 adds about a seventh of cProfile's overhead and keeps every event, where cProfile only keeps per-function aggregates. That protobuf trace was 197 MB — a full timeline is not free, it is just cheaper in time than in disk. The same run as JSON is an order of magnitude larger.

For the callback in isolation, scripts/benchmark.py overhead reports the difference against the same workload untraced: ~4 ns per event at 8 threads, ~13 ns on a single thread. An event is one PY_START or PY_RETURN, so a function call is two.

Try it

No install — uvx fetches the wheel, traces your script, and leaves nothing behind:

uvx trace0 run --output trace.pb your_script.py

Drop trace.pb onto https://ui.perfetto.dev and you have a flame chart.

The traced script runs inside the tracer's environment, so name what it imports with --with, and the interpreter with --python:

uvx --with httpx --python 3.13t trace0 run --output trace.pb your_script.py

Free-threaded builds (3.13t, 3.14t) are a first-class target — every thread is traced, not just the calling one.

-m runs a library module, exactly as python -m does; everything after the module name belongs to it:

uvx --with uvicorn trace0 run --output trace.pb -m uvicorn app:app --port 8000

A server traced from launch spends most of its trace importing. To trace only the part you care about, wrap it with the Tracer API below instead.

Install

Install it for the Tracer API, which also puts the same CLI on your project's path. Requires Python 3.13+.

uv add trace0

from trace0 import Tracer

with Tracer("trace.pb"):
    your_workload()

Output is Perfetto protobuf by default. --format json, or Tracer(..., format="json"), writes Chrome Trace Event instead — larger and slower to write, but human-readable and diffable.

Child processes

Children are traced too, whether they arrive by fork or by exec, so multiprocessing, subprocess and worker-based servers all show up. They write into the same file as they run, so there is nothing to merge afterwards:

trace0 run --output trace.pb -m uvicorn app:app --workers 4

leaves one trace.pb holding every worker, on a track of its own. Each process buffers whole packets and commits them under a lock, so two processes writing at once never splice into each other, and each namespaces its packet sequences by pid so Perfetto keeps them apart.

JSON traces use the Chrome JSON Array Format for the same reason — entries follow one another with a trailing comma and no closing bracket. That is what lets several processes append to one array, and it means a process killed mid-run still leaves everything up to its last whole entry.

Status

Early. The tracer works end to end and is covered by tests, but the API is not yet stable and wheels are not yet built for every platform.

Developing

The benchmarks behind every number above live in one script:

uv run scripts/benchmark.py --help

prek runs the CI checks before a commit lands — formatting and clippy on commit, both test suites on push:

uv tool install prek && prek install --hook-type pre-commit --hook-type pre-push

Layout

Rust lives under crates/, Python under python/.

crate	role
`trace0-core`	clock, event model, event queue, exporter contract
`trace0-json`	Chrome Trace Event output
`trace0-proto`	Perfetto protobuf output
`trace0-py`	the `sys.monitoring` callbacks and the `_core` extension

Only trace0-py links against CPython, and it is the one crate with no tests: a cdylib built with extension-module leaves the interpreter's symbols undefined, so it cannot link outside maturin. Everything worth testing lives in the other three, which cargo test runs directly.

License

MIT

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

trace0-0.1.10.tar.gz (45.3 kB view details)

Uploaded Aug 2, 2026 Source

Built Distributions

If you're not sure about the file name format, learn more about wheel file names.

The dropdown lists show the available interpreters, ABIs, and platforms. Enable javascript to be able to filter the list of wheel files.

trace0-0.1.10-cp314-cp314t-manylinux_2_17_x86_64.manylinux2014_x86_64.whl (645.8 kB view details)

Uploaded Aug 2, 2026 CPython 3.14tmanylinux: glibc 2.17+ x86-64

trace0-0.1.10-cp314-cp314t-manylinux_2_17_aarch64.manylinux2014_aarch64.whl (630.7 kB view details)

Uploaded Aug 2, 2026 CPython 3.14tmanylinux: glibc 2.17+ ARM64

trace0-0.1.10-cp314-cp314t-macosx_11_0_arm64.whl (578.8 kB view details)

Uploaded Aug 2, 2026 CPython 3.14tmacOS 11.0+ ARM64

trace0-0.1.10-cp314-cp314t-macosx_10_12_x86_64.whl (592.2 kB view details)

Uploaded Aug 2, 2026 CPython 3.14tmacOS 10.12+ x86-64

trace0-0.1.10-cp314-cp314-manylinux_2_17_x86_64.manylinux2014_x86_64.whl (646.4 kB view details)

Uploaded Aug 2, 2026 CPython 3.14manylinux: glibc 2.17+ x86-64

trace0-0.1.10-cp314-cp314-manylinux_2_17_aarch64.manylinux2014_aarch64.whl (633.1 kB view details)

Uploaded Aug 2, 2026 CPython 3.14manylinux: glibc 2.17+ ARM64

trace0-0.1.10-cp314-cp314-macosx_11_0_arm64.whl (579.5 kB view details)

Uploaded Aug 2, 2026 CPython 3.14macOS 11.0+ ARM64

trace0-0.1.10-cp314-cp314-macosx_10_12_x86_64.whl (593.4 kB view details)

Uploaded Aug 2, 2026 CPython 3.14macOS 10.12+ x86-64

trace0-0.1.10-cp313-cp313t-macosx_11_0_arm64.whl (577.1 kB view details)

Uploaded Aug 2, 2026 CPython 3.13tmacOS 11.0+ ARM64

trace0-0.1.10-cp313-cp313t-macosx_10_12_x86_64.whl (592.1 kB view details)

Uploaded Aug 2, 2026 CPython 3.13tmacOS 10.12+ x86-64

trace0-0.1.10-cp313-cp313-manylinux_2_17_x86_64.manylinux2014_x86_64.whl (646.5 kB view details)

Uploaded Aug 2, 2026 CPython 3.13manylinux: glibc 2.17+ x86-64

trace0-0.1.10-cp313-cp313-manylinux_2_17_aarch64.manylinux2014_aarch64.whl (632.8 kB view details)

Uploaded Aug 2, 2026 CPython 3.13manylinux: glibc 2.17+ ARM64

trace0-0.1.10-cp313-cp313-macosx_11_0_arm64.whl (579.1 kB view details)

Uploaded Aug 2, 2026 CPython 3.13macOS 11.0+ ARM64

trace0-0.1.10-cp313-cp313-macosx_10_12_x86_64.whl (592.8 kB view details)

Uploaded Aug 2, 2026 CPython 3.13macOS 10.12+ x86-64

File details

Details for the file trace0-0.1.10.tar.gz.

File metadata

Download URL: trace0-0.1.10.tar.gz
Upload date: Aug 2, 2026
Size: 45.3 kB
Tags: Source
Uploaded using Trusted Publishing? Yes
Uploaded via: twine/7.0.0 CPython/3.13.14

File hashes

Hashes for trace0-0.1.10.tar.gz
Algorithm	Hash digest
SHA256	`839c543da2b650ba3bd071584ab556b5b4bc7a3ee1ae11693e913fcf3d728d07`
MD5	`d8be08fde6d9bf5da152eb018f733ba0`
BLAKE2b-256	`4649263112b6ef8729b32471cfad06ca04c0c0375d0fc5a9820d013f03ecf241`

Algorithm	Hash digest
SHA256	`5403f4b6ddf4fea4dbde06196af5bd93b5abf5fb2ed0c0a3925842f88b5fb304`
MD5	`6729015b17e872e56f301c09e91e366c`
BLAKE2b-256	`c637cc6f14cdb339bb3e6666722a1162048d8498efcae78784d167084b667ed3`

Algorithm	Hash digest
SHA256	`1861f583766d93e6392b686f9a1551ae987a0e7ae29806274b20218f9c03018f`
MD5	`422e23192277ae866abda31b5f30f6a3`
BLAKE2b-256	`122b12f11395794ebf78523dd3d35774b52aa3b078e910ca63e3562ac08c197d`

Algorithm	Hash digest
SHA256	`c23e520b2c664c20c95bd0fceeb4062e6f14f2f7a555cd1f83b8c3c4e8e726cb`
MD5	`bd8af338de25ad1ce09a55d1f7a70a91`
BLAKE2b-256	`b7a0bf52b3227457c232148cd9f49792466b1add7db166277db13f69be9a1448`

Algorithm	Hash digest
SHA256	`6abe6300b610918c34ef224b1e3cc14758c0c64cf9f3cc779a567d7b519e998b`
MD5	`021d5748f097a620b97ff8955ab4f54a`
BLAKE2b-256	`00db1c41fd81bbcd3f162d3a2bc1dfc05c6d42e91e7fc666fcf105267fa1086a`

Algorithm	Hash digest
SHA256	`3e3edd74e68906e43cee612c64d3d00ce16e903d25759f7a243231c872dc10cf`
MD5	`ffdc05ecb446bb5d063a3e13861fc0a0`
BLAKE2b-256	`8578c1fb260d3cf629c6dcc0c98814a710ac6a2c5be809ef042d0b989fc2b183`

Algorithm	Hash digest
SHA256	`27d4b741eb8ced5d849051d19bb70a87cbc4e2fc609fb97a43ff746fd4754720`
MD5	`1f1c0de35818061fb4f8086248907dc7`
BLAKE2b-256	`4e371da68fdb460f5b6820be40b35cbbd61eaa6639d8672cc146975e3dd9e569`

Algorithm	Hash digest
SHA256	`26997e11e02ecf32052b4f633eaae8b9dceecc6dd45dfd3a4efbd30c9f5ca4c7`
MD5	`b53f6d1a3e741fc59e6d97eab10476d6`
BLAKE2b-256	`8aeeeaa426f57f91aac098b38baa7ab715b296897032664bf27e63d1d77d4353`

Algorithm	Hash digest
SHA256	`405d4323f035de8c3767666adef292d4859a54b1604e1df39ad82985dd93e7a1`
MD5	`1686849de0c721c1641dcea3cabbc823`
BLAKE2b-256	`a834067c7658d31d4e03a267bd5f81244c544e1e816cbed950467320b2d449b7`

Algorithm	Hash digest
SHA256	`4f35c4faf70bedf86beabe7e8148971466c74446088b789f0a90102f43f22307`
MD5	`a0540dd4d0bfee30c1ec81689f1e17f1`
BLAKE2b-256	`6712672aa8cbe3c9a9b05aaf14c92f613ea08a6fb6aeb93ddf8503c47f7754cf`

Algorithm	Hash digest
SHA256	`4edf21602af7d5cd2215e380f995f2a711cdce5b5dd6c7443b6ebdad44101ab2`
MD5	`d9910ad4603bdd2e911ad0d6f67db0af`
BLAKE2b-256	`55b6dfad76d39dc54debbc27445baadbe57254099830badf5d1e9f16f00da53e`

Algorithm	Hash digest
SHA256	`25b6bffae448ef2095dae340538a31e514105d2def3f731db130c13f0354da32`
MD5	`3b0b6bcafc5a0a587fc7614edb5f8955`
BLAKE2b-256	`7f7c892bf8e577a234886185dfa6c22726cbe6a064acf7bc1563157ae98727e7`

Algorithm	Hash digest
SHA256	`4622efc6bd8fc1fa035ed3646ebba9dabc93124f4812b787c4df9184d9c059d3`
MD5	`f59ea464cfebace1b354d74a8e292092`
BLAKE2b-256	`8b106402a17a7147c67204b9fb5d9c24dd4bf451b51b4ce3e9df07cb939a8c41`

Algorithm	Hash digest
SHA256	`c365397bd817306b8c30bfcee491413cd846fa7f5082179c33c6ffdaebd23c08`
MD5	`ebe5739dc7f4ab325616f2b1becf28a0`
BLAKE2b-256	`b1eea20deb6db566406335b1a01fa9620049321104d6cf76ce35cb102f3a8b71`

Algorithm	Hash digest
SHA256	`8484563130b275396613cfb509b355b8f38eb39e68550fbfe6e1f90f3f9344a9`
MD5	`a118eadecd035c886c86c8cac248c98c`
BLAKE2b-256	`bf4b990f47a9d0320282e92957183f8ca92f8e9680dfea7dcda3348381227d72`

trace0 0.1.10

trace0

How much does it cost?

Try it

Install

Child processes

Status

Developing

Layout

License

Project links

Key dates

1 maintainer

Credits

GitHub Statistics

License expression

Requires

Download files

Source Distribution

Built Distributions

File details

File metadata

File hashes

Provenance

File details

File metadata

File hashes

Provenance

File details

File metadata

File hashes

Provenance

File details

File metadata

File hashes

Provenance

File details

File metadata

File hashes

Provenance

File details

File metadata

File hashes

Provenance

File details

File metadata

File hashes

Provenance

File details

File metadata

File hashes

Provenance

File details

File metadata

File hashes

Provenance

File details

File metadata

File hashes

Provenance

File details

File metadata

File hashes

Provenance

File details

File metadata

File hashes

Provenance

File details

File metadata

File hashes

Provenance

File details

File metadata

File hashes

Provenance

File details

File metadata

File hashes

Provenance