pytest-dev/pytest · error · ValueError

absolute tolerance can't be negative

Error message

absolute tolerance can't be negative: {abs}

What it means

Raised by ApproxTimedelta.__init__ when abs is a timedelta but negative. A negative tolerance is meaningless (it would accept nothing or invert the comparison) and almost always indicates a sign mistake. The check runs after confirming abs is a timedelta, so the value can be safely compared to timedelta(0).

Solutions

  1. Ensure the tolerance is non-negative: use builtins.abs(...) on the computation.
  2. Pass an explicit positive timedelta literal.
  3. Review how the tolerance value is derived before passing it to approx().

Example fix

// before
approx(td, abs=expected - actual)  # may be negative
// after
approx(td, abs=abs(expected - actual))
Defensive patterns

Strategy: validation

Validate before calling

from datetime import timedelta
def nonneg_td(v):
    if v is not None and v < timedelta(0):
        raise ValueError('tolerance must be non-negative')
    return v

Type guard

from datetime import timedelta
def is_nonneg_timedelta(v) -> bool:
    return v is None or (isinstance(v, timedelta) and v >= timedelta(0))

Prevention

When it happens

Trigger: approx(td, abs=timedelta(seconds=-1)); computing abs from a subtraction that can go negative without builtins.abs().

Common situations: Tolerance computed as (expected - actual) without wrapping in abs(); off-by-sign bugs in test fixture generation.

Related errors

AI-assisted analysis of pytest-dev/pytest@0d6fbdeffa (2026-08-11). Data as JSON: /api/errors/be9be101889764ff. Report an issue: GitHub.

Appendix: source

Thrown at src/_pytest/approx.py:669

        if nan_ok:
            raise TypeError(
                "pytest.approx() does not support nan_ok for "
                "datetime/timedelta comparisons."
            )
        if abs is None and rel is None:
            raise TypeError(
                "pytest.approx() requires an explicit tolerance for "
                "datetime/timedelta comparisons: "
                "e.g. approx(expected, abs=timedelta(seconds=1)) "
                "or approx(expected, rel=0.01)"
            )
        if abs is not None and not isinstance(abs, timedelta):
            raise TypeError(
                f"absolute tolerance for datetime/timedelta must be a "
                f"timedelta, got {type(abs).__name__}"
            )
        if abs is not None and abs < timedelta(0):
            raise ValueError(f"absolute tolerance can't be negative: {abs}")
        if rel is not None:
            if not isinstance(rel, (int, float)):
                raise TypeError(
                    f"relative tolerance for timedelta must be a "
                    f"number, got {type(rel).__name__}"
                )
            if rel < 0:
                raise ValueError(f"relative tolerance can't be negative: {rel}")
            if math.isnan(rel):
                raise ValueError("relative tolerance can't be NaN.")
            if math.isinf(rel):
                raise ValueError("relative tolerance can't be infinite.")
        # Compute the effective tolerance. abs_tolerance is a timedelta, rel * expected
        # gives a timedelta (timedelta * float works in Python).
        abs_tolerance = abs
        if rel is None:
            rel_tolerance = None
        else:

View on GitHub (pinned to 0d6fbdeffa)