Evaluate every rotation condition so a skipped one cannot fire spuriously

[januththedev]

Rotation: every condition is now evaluated, so a skipped one cannot fire spuriously

Disclosure, as required by the Automated Contributions Policy in CONTRIBUTING.rst: this bug was found, proved, and fixed with the assistance of an LLM. I have reviewed the change line by line and verified the tests myself. There is no LLM co-author on the commit.

Description

A list of rotation conditions is combined with any():

def __call__(self, message, file) -> bool:
    return any(rotation(message, file) for rotation in self._rotations)

any() over a generator short-circuits. That is fine for the return value, but each RotationTime condition is stateful: it lazily computes self._limit on first call and then advances it as a side effect of being invoked (loguru/_file_sink.py:154-157, the while self._limit <= record_time: self._step_forward(...) catch-up loop).

When the first condition returns True, the remaining conditions are never invoked, so their _limit is never advanced past the current time. A skipped condition is left holding a limit that is now in the past, and it therefore fires on the very next logged message — producing a second, immediate, spurious rotation.

Reproduction

logger.add(path, rotation=["13:00", "12:00"]), file created at 06:00, process idle until 13:00:01, then two messages:

before — 3 files (BUG):
  file.2020-01-01_06-00-00_000000.log     'a\n'
  file.2020-01-01_13-00-01_000000.log     'b\n'    <-- spurious extra rotation
  file.log                                'c\n'

after — 2 files (correct):
  file.2020-01-01_06-00-00_000000.log     'a\n'
  file.log                                'b\nc\n'

Message "b" legitimately rotates at 13:00:01; "c" should not. The "12:00" condition's stale 2020-01-01 12:00 limit is what caused the second rotation.

The change

# Note that the conditions must all be evaluated, even after one of them returned
# True, because "RotationTime" instances keep their limit up-to-date as a side-effect
# of being called. Skipping them would leave a stale limit in the past, causing a
# spurious extra rotation on the next logged message.
should_rotate = False
for rotation in self._rotations:
    if rotation(message, file):
        should_rotate = True
return should_rotate

This keeps the documented "any of which can trigger the rotation" contract for the return value while restoring the side effect the conditions rely on.

Tests

test_multiple_rotation_conditions_all_evaluated is added to tests/test_filesink_rotation.py.

• Fails before the change with assert 3 == 2; passes after.
• pytest tests/test_filesink_rotation.py → 147 passed, 4 skipped (146 before).
• Filesink suites → 251 passed, 6 skipped.
• Full suite → 1646 passed, 70 skipped, 0 failed. The baseline on unmodified HEAD was 1645 passed, 0 failed, so the delta is exactly this one test.
• ruff check clean.

One note: ruff format --check reports one pre-existing trailing-comma difference at _file_sink.py:189 (**kwargs vs **kwargs,) because my ruff differs from the pinned v0.15.7. I left it untouched.

The list-of-conditions feature itself is unreleased (still in Unreleased from #1174), so the changelog entry is filed under that issue rather than needing a new one.