approx: Decimal and numpy cleanup

[RonnyPfannschmidt]

This was explored and written by an AI agent (Claude Opus 5 via Claude Code) that I prompted. The diagnosis, the patches, the tests and this description are its work; I read it, reproduced it and am posting it.

Closes #15005.

A cleanup pass over pytest.approx, starting from #15005 and following the same mistake through the module. Five commits, each standalone with its own changelog fragment — reviewable and revertable one at a time.

Decimal and float meet in _pytest/approx.py in two ways, and only one is lossy:

• float(Decimal) saturates. float(Decimal("1e400")) is inf and float(Decimal("1e-400")) is 0.0, silently. Every math.* call on a Decimal takes this route.
• Comparing a Decimal to a float is exact. CPython compares true values (Decimal("0.1") < 0.1 is correctly True). Safe arithmetically; it only signals decimal.FloatOperation when that trap is set.

1. nan/inf checks — closes #15005

math.isnan/math.isinf take the lossy route, so approx() treated a finite Decimal("1e400") as infinite and short-circuited before the exact Decimal tolerance was used:

>>> Decimal("1.0000001e400") == pytest.approx(Decimal("1e400"), rel=Decimal("1e-6"))
False   # difference 1E+393, tolerance 1E+394

Five sites, not just the infinity guard the issue names: the same guard in __repr__, the NaN checks in __eq__ (which also raised ValueError on Decimal("sNaN")), and both NaN validations in tolerance. Fixed with _is_nan/_is_inf helpers using Decimal.is_nan()/Decimal.is_infinite(). No change for float/complex; Decimal("Infinity") still compares only to itself.

2. The displayed tolerance

#13543 fixed the FloatOperation crash of #13530 by overriding __repr__ in ApproxDecimal so self.tolerance was never touched. The override copied ApproxScalar's 1e-3 <= self.tolerance < 1e3 — where that bound only picks between :n and :.1e formatting — and reinterpreted it as a gate on whether to show rel at all. Its test asserted only that == stopped crashing, so no repr was pinned and the regression shipped.

	8.3.5	9.1.1	here
approx(Decimal("2.60"))	2.60 ± 2.6e-6	2.60 ± ???	2.60 ± 2.6e-6
rel=Decimal("1e-6")	2.60 ± 2.6e-6	2.60 ± ???	2.60 ± 2.6e-6
rel=Decimal("0.01")	2.60 ± 2.6e-2	2.60 ± 1.0e-2	2.60 ± 2.6e-2
approx(Decimal("Infinity"))	Infinity	Infinity ± ???	Infinity
approx(Decimal("NaN"))	raises InvalidOperation	NaN ± ???	NaN ± ???

The ± N is the absolute band that decides pass/fail; rel is a dimensionless ratio. Printing rel there is a unit error — 2.60 ± 1.0e-2 reads as "within 0.01 of 2.60" when the band is 0.026.

The override is dropped and ApproxScalar.__repr__ made type-correct: for a Decimal tolerance the float-literal bound is skipped and :.1e used unconditionally. That is also the only readable choice, since ApproxDecimal.__init__ converts float tolerances with Decimal.from_float — exact, and so 55 digits for abs=0.01, which :n would print in full.

Removing the override re-exposed a bug it had masked: for a NaN Decimal the negativity check in tolerance runs before the NaN check and raises decimal.InvalidOperation instead of the ValueError that __repr__ catches. NaN is now checked first, matching the float path — which is why the last row improves on 8.3.5 rather than merely restoring it.

3. Float sentinels in failure messages

All three _repr_compare implementations seeded their running maxima with -math.inf and tested == 0.0, so building a message for Decimal collections compared a float against a Decimal. Under the trap the two container variants diverged:

• ApproxSequenceLike swallowed it. decimal.FloatOperation is a TypeError subclass, so the except TypeError added for Fix 13010 - Ensure pytest.approx in sequence/collections only process instances of Number #13012 to skip non-numbers caught it, and the message reported Max absolute difference: -inf.
• ApproxMapping did not, so it escaped out of the assertion formatting as a bare decimal.FloatOperation.

Now accumulated from None, with -math.inf substituted only at the end where it still means "no difference could be computed" as #13012 needs. Divide-by-zero is a flag rather than a mid-loop math.inf assignment, which would reintroduce the mixing on the next element. The sequence variant re-raises decimal.FloatOperation explicitly, so a future leak fails loudly instead of reporting -inf.

4. numpy object dtype

>>> np.array([1.0, 2.0], dtype=object) == pytest.approx(np.array([1.0, 2.0], dtype=object))
AttributeError: 'float' object has no attribute 'item'

_yield_comparisons calls .item() to unbox numpy scalars; indexing an object array yields the stored Python object, which has none. Not Decimal-specific — object arrays of float and Fraction fail identically — but Decimal is how you get there without asking, since np.asarray([Decimal(1)]) picks object dtype on its own. It fires whenever the expected side is an object array.

Not a regression: .item() replaced np.asscalar() in 42bb0b390 when numpy 1.16 deprecated it, and np.asscalar() required an ndarray, so object dtype never worked. Unbox only when the element offers item().

I could not find an existing issue for this (searched asscalar, object dtype, dtype=object, and the error text); the nearest is #12114, a different attribute in a different method.

5. Warning on inexact float tolerances

A float tolerance with a Decimal expected value is converted exactly, so it is wider than what was written, and that can change the result:

>>> Decimal("1.10000000000000000001") == approx(Decimal("1"), rel=0.1)
True     # float band is 0.1000000000000000055511151231
>>> Decimal("1.10000000000000000001") == approx(Decimal("1"), rel=Decimal("0.1"))
False    # Decimal band is exactly 0.1

Of the tolerances anyone writes, only dyadic rationals (0.5, 0.25) survive unchanged; 1e-6, 1e-3, 0.01, 0.1, 5e-6, 1e-12 are all inexact. New pytest.PytestApproxDecimalToleranceWarning fires on those and stays quiet for exact floats, ints, Decimals, and non-Decimal comparisons.

The check lives in approx(), not ApproxDecimal.__init__. Containers build one ApproxDecimal per element from inside generator expressions — measured at 100 for a single repr() of a 100-element list — so warning there reports pytest's own source as the location, several times over. In approx() it warns once, at the line the user wrote, for scalars and containers alike.

This is also the tolerance half of what @RonnyPfannschmidt asked for on #8495 in 2021 ("we should definitively warn when we approx compare floats vs Decimals") — the half that currently succeeds silently.

Validation

Beyond the suite, the displayed tolerance was checked to be the actual pass/fail boundary across 65 combinations — magnitudes from 1e-400 to 1e100000, spanning both float underflow and overflow, crossed with default/rel/abs/both/rel=0 — with decimal.FloatOperation trapped throughout, so any remaining leak raises rather than passing quietly. The full approx suite also passes with that trap set globally.

Test gaps closed

• test_decimal_approx_repr (the pytest.approx(Decimal(...)) Causes decimal.FloatOperation in __repr__ (pytest 8.4.0+) #13530 regression test) asserted only that == worked, never a repr — which is exactly how the repr regression shipped. It now pins the repr under the trap.
• test_decimal_approx_float_rel pinned the regressed string 1.0e-2; it was added after the regression, so it encoded it.
• New coverage for: effective-tolerance reprs cross-checked against the real comparison boundary; reprs beyond the float range in both directions; Infinity/NaN reprs; NaN tolerance raising ValueError not InvalidOperation; Decimal failure messages under the trap for sequences, mappings and both divide-by-zero paths; numpy object arrays of Decimal, float and Fraction; and the warning firing once at the caller's line while staying silent for exact tolerances.

6. Coverage

src/_pytest/approx.py goes from 99% (4 missed statements and a partial branch) to 100% statement and branch coverage. Two of the three gaps predate this branch, and all three turned out to be reachable rather than dead:

• a ragged nested sequence is what makes np.asarray() fail, reaching cannot compare ... to numpy.ndarray;
• ApproxMapping guards its relative difference with expected == 0, but timedelta(0) does not compare equal to 0, so the division still raises ZeroDivisionError;
• the running maximum compares the current best against the next candidate, so a sequence mixing Decimal and float elements compares the two there — which is exactly the decimal.FloatOperation that commit 3 stops swallowing. Before that commit it reported 7 where the real maximum was 8.

Not addressed

#8495 (float vs Decimal values in __eq__) still fails identically on this branch — it needs the coercion decision from that thread, which is a larger question than this cleanup. #3247 was checked and no longer reproduces.

🤖 Generated with Claude Code

[Pierre-Sassoulas]

LGTM !