pytest-speedguard

A pytest plugin that passively tracks every test's wall-clock duration across runs, keeps a noise-tolerant rolling baseline, and flags tests that regress (get meaningfully slower) or newly appear as slow. Cross-platform, with zero runtime dependencies beyond pytest — the statistics are implemented in pure Python.

Badges reflect live state: CI shows the real GitHub Actions result and the PyPI badge shows the real published version (unpublished = "not found").

The problem: test suites rot silently

A suite creeps from 2 minutes to 12 minutes over a year and nobody notices the single test that quietly got 5× slower. Each individual run looks "basically the same as last time," so the slowdown hides in the noise until CI is painfully slow and the culprit commit is long gone.

Why pytest-benchmark doesn't cover this

pytest-benchmark is excellent, but it solves a different problem: deliberate microbenchmarks you write by hand for a handful of hot functions (benchmark(func, ...)). It won't tell you that test_import_users in your integration suite doubled last Tuesday, because you never wrote a benchmark for it.

pytest-speedguard is passive and whole-suite: it watches every test you already have, needs no code changes, and answers a different question — "did anything in my existing suite get slower?"

Install

pip install pytest-speedguard

The plugin auto-registers via pytest's pytest11 entry point — nothing else to configure.

Quickstart

# 1. Run once to seed a baseline (every test is "new" — nothing to compare yet).
pytest

# 2. ... time passes, code changes ...

# 3. Run again. Tests that got meaningfully slower are flagged.
pytest

Example output when something regressed:

=============================== speedguard ================================
Top regressions:
  test                          baseline  current    delta  n
  ----------------------------  --------  -------  -------  -
  tests/test_api.py::test_sync   208.0ms  504.5ms  +142.5%  8
1 regression, 0 new slow tests (threshold 50%, floor 50ms)

By default this is informational (it never fails your build). Add --speedguard-fail to turn it into a CI gate.

How the statistical model works (and why)

For each test, the plugin keeps a rolling window of recent durations. A test is flagged as a regression only if all four guards agree:

1. Noise floor — current >= min-duration (default 50 ms). Sub-50ms timings are dominated by scheduler jitter; a test going from 1ms to 3ms is noise, not a regression. Kills tiny-fast-test false positives.
2. Warm-up guard — there must be at least min-samples (default 3) baseline samples. With fewer, the test is still "warming up" and can't be judged. Stops a brand-new or barely-seen test from firing on its second run.
3. Relative guard — current > median × (1 + threshold) (default +50%). The robust median of the window is the reference, not the mean, so one slow outlier in the history doesn't move the bar.
4. Absolute statistical guard — current > median + noise_factor × scaled_mad (default noise_factor 3.0). scaled_mad = 1.4826 × MAD (median absolute deviation) is a robust, outlier-resistant estimate of the standard deviation. This demands the jump be larger than the test's own normal run-to-run jitter. Kills CI-runner-noise false positives on tests that are naturally variable. When dispersion is zero (few or identical samples) this guard is skipped and the decision rests on the relative guard, so we never divide by — or flag on — zero dispersion.

Using median + MAD instead of mean + standard deviation is deliberate: test timings on shared CI runners are heavy-tailed, and robust statistics ignore the occasional 10× outlier from a noisy neighbour instead of letting it poison the baseline.

New slow tests (no baseline history yet) whose duration is at or above the floor are listed separately so you notice expensive newcomers, but they are never counted as regressions.

Options

Every option is available on the command line and as an ini key (in pytest.ini, tox.ini, or pyproject.toml's [tool.pytest.ini_options]). The command line wins over ini.

CLI flag	ini key	Type	Default	Description
--speedguard / --no-speedguard	speedguard	bool	true	Master switch.
--speedguard-threshold	speedguard_threshold	float	0.5	Relative regression threshold (0.5 = +50%).
--speedguard-min-duration	speedguard_min_duration	float (s)	0.05	Ignore tests faster than this (noise floor).
--speedguard-noise-factor	speedguard_noise_factor	float	3.0	Require current > median + factor × scaled_mad.
--speedguard-window	speedguard_window	int	20	Rolling window size (samples kept per test).
--speedguard-min-samples	speedguard_min_samples	int	3	Minimum baseline samples before flagging (warm-up).
--speedguard-phase	speedguard_phase	call | total	call	Time the test body (call) or setup+call+teardown (total).
--speedguard-update	speedguard_update	auto | always | never	auto	Baseline update policy (see below).
--speedguard-accept	—	flag	off	Re-baseline: record all current durations, suppress reporting, don't fail.
--speedguard-fail	speedguard_fail	bool	false	Fail the session (non-zero exit) on any regression — the CI gate.
--speedguard-baseline	speedguard_baseline	path	.pytest_speedguard/baseline.json	Baseline file location.
--speedguard-top	—	int	20	Max rows per report section.

Update modes and --accept

Controls how the baseline window is updated after a run:

• auto (default) — append the new sample to the rolling window only if the test did not regress. A genuine regression therefore keeps firing on every run until it's fixed or explicitly accepted, instead of being silently absorbed into the baseline.
• always — always append (a pure rolling window). Simpler, but it only ever catches sudden spikes, because gradual slowdowns are continuously absorbed.
• never — compare only, never write. The baseline is frozen. Recommended for CI against a committed baseline file (see below).

--speedguard-accept is the deliberate way to move the baseline after an intended slowdown: it records all current durations as the new baseline, suppresses regression reporting for that run, and won't fail on regressions.

# You intentionally made things slower and that's fine — accept the new normal:
pytest --speedguard-accept

Known limitation: slow-creep under the per-run threshold

A rolling baseline (auto/always) compares each run against recent history. If a test creeps up by, say, 3% every week — always staying under the per-run threshold — each run looks fine while the baseline slowly drifts up with it. Over a year that's a large regression that was never flagged, because no single run ever crossed the line.

Mitigation — freeze the reference in CI:

1. Seed and commit a baseline file to your repo:

   pytest --speedguard-update=always   # build up samples over a few runs
   git add .pytest_speedguard/baseline.json
   git commit -m "Freeze test duration baseline"

2. In CI, compare against that fixed reference and never rewrite it:

   pytest --speedguard-update=never --speedguard-fail

Now drift is measured against a fixed point in time, so slow-creep accumulates against a stable reference and is caught. Re-baseline deliberately (--speedguard-accept, then commit the file) whenever you accept a new normal.

Parallel runs (pytest-xdist)

pytest-speedguard is xdist-aware in v1 via controller-side aggregation. Workers just run tests and stream their reports back to the controller; all comparison, reporting, and baseline writing happen only on the controller (guarded by hasattr(config, "workerinput")), so there are no racing writers and the baseline stays consistent. Just run pytest as usual:

pytest -n auto

Baseline JSON schema

The baseline is a small, human-readable JSON file written atomically (temp file + os.replace) so an interrupted run can never corrupt it. A missing or corrupt file is treated as an empty baseline (it never crashes your run).

{
  "schema": 1,
  "phase": "call",
  "tests": {
    "tests/test_api.py::test_sync": { "samples": [0.201, 0.199, 0.204] },
    "tests/test_api.py::test_list[page-2]": { "samples": [0.512] }
  }
}

• schema — on-disk format version.
• phase — which timing phase the samples represent (call or total). Loading a baseline recorded under a different phase resets it, since the numbers aren't comparable.
• tests — a map of pytest node id → rolling window of recent durations (seconds). Parametrized tests have distinct node ids (e.g. test_x[a-1]) and are tracked separately.

Development

pip install -e .[dev]
pytest -q

License

MIT — see LICENSE.