pyxtrace 4.0.0

Catch Python performance regressions in CI — deterministically,
so your benchmark gate stops crying wolf.

Website

Why

Wall-clock benchmarks are noisy enough to make a pull-request gate unreliable. Run the same unchanged code seven times and time it, then count the operations it performed on those same runs (python benchmarks/overhead.py):

wall-clock       : 2.06ms  2.06ms  2.20ms  2.04ms  2.06ms  2.05ms  2.17ms   → 3.1% spread
operation counts : 3196    3196    3196    3196    3196    3196    3196     → 0.0% spread

That 3.1% is on an idle local machine, which is the best case — a shared CI runner with a noisy neighbour is worse, and smaller workloads are worse still (sub-millisecond ones swing well past 15%). Meanwhile the counts are not "stable," they are identical.

So on wall-clock you either set the threshold tight and get flaky failures, or set it wide and miss real regressions. Most teams end up deleting the job.

PyxTrace counts operations — function calls and lines executed — instead of measuring time. Those counts are byte-for-byte reproducible, so a 10% threshold actually means something. Sampling profilers cannot do this: sampling is non-deterministic by construction.

Install

pip install pyxtrace

One dependency (rich). Python 3.10+.

Use

With your test suite: nothing to set up

pip install pyxtrace
pyxtrace diff main HEAD                # or `pyxtrace diff main` for uncommitted work

With no runs recorded, diff uses your pytest suite as the workload. It checks out both commits into temporary git worktrees (your checkout is never touched), runs the head's tests against both versions of the code, and counts each test on its own. Only tests that passed on both sides are compared, so a test for a feature the base does not have yet is left out rather than reported. Tests that fail on the head are named, not silently dropped; pyxTrace is not your test runner, so run pytest as its own CI job as well (or pass --fail-on-test-failure).

Each regression is then tied to the diff. A function the change edited is marked ← changed in this diff; one that only got busier names the changed functions on its call path. Changed functions the gate could not compare are listed too, split by why: no test runs them, or only tests new in this change do, so there is no baseline yet:

⚠  src/shop/orders.py::process_order  +120 calls to other functions  (same code, more work)
   ← on the call path of changed: src/shop/orders.py::handle_request
   process_order() runs 120x and makes 1 more call(s) to fetch_customer() each time

✗ FAIL — 3 function(s) regressed

1 of 2 changed function(s) were compared
⚠  1 changed function(s) no test runs, so the gate cannot see them:
     src/shop/orders.py::refund

--require-covered turns that last warning into a failure.

Only each test's own body is counted, not fixtures or imports. Each test file runs in its own pytest process, several at once, so a test's counts never depend on which other files ran before it; pytest-xdist itself is not used.

On the head, only the test files that can see the change are run: those whose tests ran a changed function on the base (or failed there). A change outside any function (an import, a constant, a class attribute), a method added to or removed from a class (an override changes what unchanged calls run), or a change to a non-Python file can reach any test, so it runs every file, and says why. --all-tests always runs every file.

A stress test can cost more traced time than the rest of the suite together (on sqlglot, one test is 82% of all counted lines). A test that makes more than 5,000,000 calls (--max-test-calls) stops being counted and runs on untraced; it is named in the report and not compared, so a regression that only it would show is missed, never silently. The largest test on jinja makes 324,000 calls, on rich 55,000.

With a workload script

Record a run per commit, then compare commits:

pyxtrace examples/orders.py            # → .pyxtrace/<commit sha>.pyxt
# ... commit your change ...
pyxtrace examples/orders.py
pyxtrace diff HEAD~1 HEAD              # any git ref, or a .pyxt path

diff exits 1 when something regressed, so it drops straight into CI.

Before committing, compare your uncommitted work against any commit. A run with uncommitted changes is filed separately, so it never overwrites the commit's clean run:

pyxtrace examples/orders.py            # → .pyxtrace/<sha>-dirty.pyxt
pyxtrace diff main                     # main vs. the checkout as it is now

What a caught regression looks like

The bundled example has a batched customer lookup. Set PYXTRACE_NPLUSONE=1 and it fetches each order's customer individually instead:

⚠  examples/orders.py::process_order  +120 calls to other functions  (same code, more work)
   process_order() runs 120x and makes 1 more call(s) to fetch_customer() each time
   Attributed to:
     examples/orders.py::fetch_customer   0 → 120 calls   new

⚠  examples/orders.py::Cursor.execute  +5,950% operations  (6 → 363)

⚠  examples/orders.py::fetch_customer  new operations  (0 → 240)
   Attributed to:
     examples/orders.py::Cursor.execute   0 → 120 calls   new

✗ FAIL — 3 function(s) regressed

process_order ranks first: it is where the extra work was introduced. Cursor.execute and fetch_customer merely run more often as a consequence.

Every line there is arithmetic over exact counts, and there is no confidence score because there is nothing to be uncertain about. Note also what it does not say. "Runs 120x and makes 1 more call each time" is a measurement; calling it an N+1 would be a conclusion, and one input size cannot support it. The same shape is produced by a function that simply does more work per call. Reading it is your job, and you have the counts to do it with.

A single run

pyxtrace examples/orders.py

                 pyxTrace – examples/orders.py
┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━┳━━━━━━━┳━━━━━━━━┓
┃ function                            ┃ calls ┃ lines ┃ own ms ┃
┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━╇━━━━━━━╇━━━━━━━━┩
│ examples/orders.py::serialize       │   120 │   480 │    0.2 │
│ examples/orders.py::process_order   │   120 │   360 │    0.2 │
│ examples/orders.py::handle_request  │     1 │   245 │    0.1 │
│ examples/orders.py::fetch_orders    │     1 │   122 │    0.0 │
└─────────────────────────────────────┴───────┴───────┴────────┘

calls and lines are deterministic. own ms is timing and varies run to run — it is shown for orientation and is never what diff gates on.

In CI

With a test suite, the gate is one step after installing your project's test dependencies:

- uses: actions/checkout@v4
  with: { fetch-depth: 0 }                  # the base commit has to be present
- run: pip install pyxtrace pytest -e .
- uses: actions/cache@v4                    # the base side, counted once per test file
  with:
    path: .pyxtrace/tests
    key: pyxtrace-${{ runner.os }}-${{ github.event.pull_request.base.sha }}
    restore-keys: pyxtrace-${{ runner.os }}-
- run: pyxtrace diff ${{ github.event.pull_request.base.sha }} HEAD --stability-runs 2

The base side's results are kept in .pyxtrace/tests/, one entry per test file, named by a hash of everything the counts depend on: the base commit, that test file, every conftest and test helper, the Python version, the installed packages and pyxTrace itself. A second diff against the same base runs only the head; a pull request that edits one test file counts only that file again on the base. An entry can never be reused for anything it does not match, so restoring an older cache is safe.

--stability-runs 2 runs the head's tests twice and leaves out any test whose counts differ between the two, so a test driven by the clock cannot fail the gate. It lists those tests, and costs one more run of the head.

With a workload script, run files are keyed by commit sha, so a PR gate records the base commit and the head and diffs them by name:

- uses: actions/checkout@v4
  with: { fetch-depth: 0 }                  # the base commit has to be present
- run: pip install pyxtrace
- run: pyxtrace benchmarks/workload.py      # HEAD → .pyxtrace/<sha>.pyxt
- run: |
    git checkout -q ${{ github.event.pull_request.base.sha }}
    pyxtrace benchmarks/workload.py         # base → .pyxtrace/<sha>.pyxt
    git checkout -q -
- run: pyxtrace diff ${{ github.event.pull_request.base.sha }} HEAD --threshold 10

Because the counts are exact, the base run can be recorded once and cached (actions/cache on .pyxtrace/) instead of re-run per PR: the same commit on the same Python version always produces the same bytes. Put the Python version in the cache key — see Limits.

Options

diff exits 0 on a pass, 1 on a regression, and 2 on a usage error or an unreadable run file. run exits with the script's own status; a script that fails records nothing, since a partial count is not a baseline.

What keeps the counts exact

Determinism is the whole product, so these are handled rather than documented away:

• String hashing. Python randomises it per process, which reorders every set and dict of strings and changes how many lines a loop over one runs. The pyxtrace command re-launches itself with PYTHONHASHSEED=0. The Python API cannot, and warns instead.
• Threads. Every thread counts into its own counters, merged at the end, so no update is lost even on free-threaded Python. Threads the script leaves running are waited for exactly as python app.py waits for them. Daemon threads still running at exit are reported, since their counts are partial.
• Randomness and garbage collection in tests. When diff runs your test suite, random is seeded the same way before every test, and garbage collection is held off while a test runs, so it cannot fire at a point that depends on earlier tests. On jinja that took the tests whose counts moved between identical runs from 5 of 911 to none. Still yours to fix: other random generators (NumPy, Faker), and work driven by the clock or by timer threads. On rich, 1 to 3 of 948 tests still move between runs by under 0.1%, because log lines carry the time and a live display refreshes on a timer. --stability-runs 2 finds such tests and leaves them out.
• Names. Functions are keyed by qualified name (Cursor.execute), so two classes' __init__ never merge into one entry. Python 3.10 has no qualified names on code objects, so pyxTrace reads them from the source instead.

Overhead

Measured on fib(22), a deliberately call-heavy worst case (python benchmarks/overhead.py):

fib(22) is a deliberate worst case: it does almost nothing between calls, so the per-event cost has nothing to amortise against. The real-code column is sqlglot parsing 838 SQL statements, which is the number to plan around for a workload script.

With your test suite, on Python 3.12+ the counting uses sys.monitoring (PEP 669), which switches code outside the root off for good instead of filtering it on every call. Each side of the diff, measured on whole suites:

The two tracers count calls differently (settrace counts a generator resuming as a call), so a run file records which one made it and runs from different tracers are never compared. PYXTRACE_TRACER=settrace forces the slower one, for example when another profiler already holds sys.monitoring.

Limits

Stated plainly, because a profiler that hides these is worse than none:

• Work inside C is invisible. PyxTrace counts Python operations. A regression that lives in a C extension, a regex, NumPy, a database driver, or any I/O wait produces no extra Python operations and will not be caught. A measured example: moving the same workload onto a backtracking regex made it 230,000x slower with byte-identical counts, and the gate passed. This is the mirror image of what sampling profilers cannot do, and it is the first thing to check before adopting PyxTrace. It is strongest on pure-Python CPU-bound code: parsers, template engines, serializers, business logic.
• Subprocesses are not traced. Threads and asyncio are: coroutines run on the event loop's thread, and worker threads are covered via threading.settrace. A thread started before tracing begins is not.
• The workload has to be deterministic itself. Unseeded random, the clock, or network responses that steer control flow will move the counts. Seed them in the benchmark script.
• Only code under one root directory is profiled — the git checkout the script is in by default (the script's own directory outside git), or --root DIR. Anything outside it is skipped, and so is any virtualenv or site-packages inside it. A run that captured nothing but the entry script says so.
• Counts are exact per Python minor version, not across them. The interpreter decides which lines emit events, and that changes between releases: 3.12 inlined comprehensions, so upgrading from 3.11 moves line counts with no code change. Record both sides of a diff on the same Python; each run file records its interpreter and diff warns when they differ.
• Operation counts are a proxy for cost, not a measure of it. An algorithmic change can cut operations and still be slower. Gate on counts, then confirm with a real timer.
• Not for production. This is a deterministic tracer for CI. To profile a live process, use py-spy.

When to use something else

Contributing

Contributions are welcome, and the most useful one right now is trying pyxtrace on a real project and reporting what happened.

git clone https://github.com/AbhineetSaha/pyxTrace.git
cd pyxTrace
python -m venv .venv && source .venv/bin/activate
pip install -e ".[dev]"

pytest -q                        # tests
python benchmarks/overhead.py    # overhead + determinism gate

CONTRIBUTING.md explains how the code is organised and the few rules that keep the counts exact. Look for good first issues to get started. This project follows the Contributor Covenant.

License

MIT — see LICENSE.