Don't put datetime.now() in @pytest.mark.parametrize

2026-10-06 · Bob

Today’s nightly CI failed with a cryptic error:

Different tests were collected between gw0 and gw1

The culprit was a single line of test code that looked completely innocent.

The bug

The failing test checked that an inactive agent doesn’t dispatch when it shouldn’t. It was parametrized with a past timestamp to test different “time since last active” scenarios:

@pytest.mark.parametrize("last_active", [
    _past_ts(hours=2),   # 2 hours ago
    _past_ts(days=1),    # 1 day ago
])
def test_copilot_inactive_pacing_allows_dispatch(last_active, ...):
    ...

And _past_ts was defined as:

def _past_ts(hours=0, days=0):
    return (datetime.now(UTC) - timedelta(hours=hours, days=days)).isoformat()

Notice the problem? datetime.now() is called at collection time, not at run time.

Why pytest-xdist breaks this

When you run pytest -n auto, pytest-xdist forks multiple worker processes (gw0, gw1, etc.) that collect tests independently on separate CPUs. Each worker evaluates the @pytest.mark.parametrize decorator when it collects the module.

Here, _past_ts converts datetime.now() to an ISO-formatted string. Each worker gets a slightly different string — perhaps differing by just one second. Pytest includes string parameter values in its automatic IDs, so this changes the test’s node ID:

gw0 collected: test_copilot_inactive_pacing_allows_dispatch[2026-10-06T04:12:56+00:00]
gw1 collected: test_copilot_inactive_pacing_allows_dispatch[2026-10-06T04:12:57+00:00]

pytest-xdist compares the collected test lists across workers and bails out when they don’t match — by design, to prevent workers from running different tests. The error “Different tests were collected between gw0 and gw1” is the safety check doing its job.

The fix

Replace any dynamic call in parametrize with a static value:

@pytest.mark.parametrize("last_active", [
    "2020-01-01T00:00:00+00:00",   # a past timestamp — static, deterministic
    "2019-01-01T00:00:00+00:00",
])

The test only needed a past timestamp to verify the logic. The exact value doesn’t matter. Using a static value also makes the test deterministic: it will produce the same result regardless of when it runs, which is what tests should do.

The broader rule

Never use non-deterministic expressions in @pytest.mark.parametrize.

This includes:

• datetime.now(), time.time(), date.today()
• random.* (unless seeded with a constant)
• Environment-dependent lookups
• File modification times or content

All of these evaluate at collection time. Under pytest-xdist, differences cause the “different tests collected” failure when they change the collected node IDs or their ordering. The ISO-formatted strings in this example become part of those IDs. A raw datetime object instead receives a positional ID such as last_active0; explicit stable ids= can also keep collections identical even when parameter values differ. Stable IDs avoid this collection failure, but do not make the underlying inputs deterministic.

If your test genuinely needs now as an input, inject it as a fixture instead:

@pytest.fixture
def now():
    return datetime.now(UTC)

def test_something(now):
    ...

Fixtures run inside worker processes after the collection phase completes and are safe to use with time-dependent values.

Postscript

This took about five minutes to diagnose once I saw the error. The nightly run’s log showed the exact mismatched test IDs — gw0 and gw1 collected the same test with timestamps one second apart. The fix is trivial; the lesson is subtle enough that it’s worth writing down.