What causes "relative tolerance can't be NaN." in pytest-dev/pytest?

To treat NaN as equal to NaN, use `approx(value, nan_ok=True)` and avoid rel on NaN expected.

pytest-dev/pytest · error · ValueError

relative tolerance can't be NaN.

Error message

relative tolerance can't be NaN.

What it means

In ApproxScalar.tolerance, if the computed relative tolerance (rel * abs(expected)) is NaN, pytest raises ValueError. This happens when rel is NaN or when expected is NaN (since abs(NaN) is NaN, and rel*NaN is NaN). Comparisons against NaN always fail, so pytest surfaces the bad input explicitly instead of producing silently-failing assertions.

Solutions

To treat NaN as equal to NaN, use `approx(value, nan_ok=True)` and avoid rel on NaN expected.
Sanitize the expected value: `assert not math.isnan(v)` before constructing approx.
Fix the upstream NaN source (zero-denominator division, missing data).

Example fix

# before
approx(float('nan'), rel=0.01)
# after
approx(float('nan'), nan_ok=True)

Defensive patterns

Strategy: validation

Validate before calling

import math
def is_safe_for_rel(expected, rel) -> bool:
    if rel is None:
        return True
    if isinstance(expected, (int, float)) and math.isnan(expected):
        return False
    return isinstance(rel, (int, float)) and not math.isnan(rel)
# usage
if not is_safe_for_rel(expected, rel):
    # use nan_ok instead of rel for NaN expected
    assert actual == approx(expected, nan_ok=True)
else:
    assert actual == approx(expected, rel=rel)

Type guard

import math
def is_finite_rel_source(expected, rel) -> bool:
    return (
        isinstance(rel, (int, float)) and not math.isnan(rel)
        and not (isinstance(expected, float) and math.isnan(expected))
    )

Prevention

Use nan_ok=True instead of rel when the expected value may be NaN.
Sanitize expected values: assert not math.isnan(v).
Fix upstream computations that produce NaN.

When it happens

Trigger: Calling approx(float('nan'), rel=0.1) (expected is NaN), approx(1.0, rel=float('nan')), or approx(value, rel=nan_from_config). Triggered at ApproxScalar.tolerance (src/_pytest/approx.py:588-589).

Common situations: Comparing NaN expected values with a relative tolerance (use nan_ok=True for NaN-equality instead); tolerance parsed from data with missing/non-numeric fields; upstream computation producing NaN.

Related errors

AI-assisted analysis of pytest-dev/pytest@0d6fbdeffa (2026-08-11). Data as JSON: /api/errors/e10b63e3e0116544. Report an issue: GitHub.

Appendix: source

Thrown at src/_pytest/approx.py:589

                return absolute_tolerance

        # Figure out what the relative tolerance should be.  ``self.rel`` is
        # either None or a value specified by the user.  This is done after
        # we've made sure the user didn't ask for an absolute tolerance only,
        # because we don't want to raise errors about the relative tolerance if
        # we aren't even going to use it.
        rel = self.rel if self.rel is not None else self.DEFAULT_RELATIVE_TOLERANCE
        # expected is SupportAbs, checked in __init__.
        # The typing here is not exact...
        abs_expected: ExpectedT = abs(self.expected)  # type: ignore[arg-type]
        relative_tolerance: float | Decimal = rel * abs_expected  # type: ignore[operator]

        if relative_tolerance < 0:
            raise ValueError(
                f"relative tolerance can't be negative: {relative_tolerance}"
            )
        if math.isnan(relative_tolerance):
            raise ValueError("relative tolerance can't be NaN.")

        # Return the larger of the relative and absolute tolerances.
        return max(relative_tolerance, absolute_tolerance)


class ApproxDecimal(ApproxScalar[Decimal]):
    """Perform approximate comparisons where the expected value is a Decimal."""

    DEFAULT_ABSOLUTE_TOLERANCE = Decimal("1e-12")
    DEFAULT_RELATIVE_TOLERANCE = Decimal("1e-6")
    rel: Decimal | None
    abs: Decimal | None

    def __init__(
        self,
        expected: Decimal,
        rel: float | Decimal | timedelta | None,
        abs: float | Decimal | timedelta | None,

View on GitHub (pinned to 0d6fbdeffa)