TypeError with pytest.approx on nested dictionaries

[frexvahi]

{'a': {'b': 0.2 + 0.1}} == pytest.approx({'a': {'b': 0.3}})

~/.conda/lib/python3.6/site-packages/_pytest/python_api.py in __eq__(self, actual)
    195         # other or not.  The abs() calls are for compatibility with complex
    196         # numbers.
--> 197         if math.isnan(abs(self.expected)):
    198             return self.nan_ok and math.isnan(abs(actual))
    199 

TypeError: bad operand type for abs(): 'dict'

Pytest 3.3.2 on python 3.6.4 (Anaconda)
Ubuntu 16.04

[The-Compiler]

That seems expected to me.

[RonnyPfannschmidt]

we should perhaps raise a better error message

im reasonably sure that assert {'a': {'b': 0.2 + 0.1}} == {'a': {'b': pytest.approx(0.3)}} should be workable (with maybe some rough edges)

does anyone want to try?

[The-Compiler]

I think it'd be quite unusual to do additional type checking here (EAFP applies 😉). You'll probably get similar error messages if you do pytest.approx({'a': None}) or whatever.

[nicoddemus]

Recursively comparing dictionaries might get tricky I think.

[kdebrab]

One can find an implementation of approx with support for nested dictionaries here: https://stackoverflow.com/a/56048692/1551810. It would be great if pytest could support this natively.

[RonnyPfannschmidt]

its important to have a definition of whats expected wrt additional/missing keys wrt nested dictionaries, plus its not hard to put approx instances actually in dictionaries as well

so we should find a clear spec of what approx should do, write it down in the docs and then go for it

[Darkdragon84]

Hi! Any progress on this? It would help tremendously and save me from writing tons and tons of ugly boilerplate code for approximate comparisons of deeply nested dataclasses (or their dictifications)

[The-Compiler]

@Darkdragon84 nothing I'm aware of - if anything is happening, it would likely be noted in this issue.

[nicoddemus]

Btw I would vote for not reusing pytest.approx to compare nested dictionaries, instead using a new API, because that can be more flexible in being configurable (for example, what do do regarding missing keys, etc).

[RonnyPfannschmidt]

@nicoddemus we might want to set up a new pytest plugin for other assert helpers (pytest-unordered, and a few other items could go there for incubation)

[Darkdragon84]

Understood. Good point @nicoddemus I will see what is out there. Thanks folks for the quick reaction 👍

[rambattu]

deepdiff module is a good alternative for nested dictionaries
{'a': {'b': 0.2 + 0.1}} == pytest.approx({'a': {'b': 0.3}})
becomes

assert not deepdiff.DeepDiff({'a': {'b': 0.2 + 0.1}}, {'a': {'b': 0.3}}, ignore_order=True, significant_digits=6)

[FrankZijlstra]

This can be fixed with a single line change:
Instead of

pytest/src/_pytest/python_api.py

Line 90 in dfc910e

a == self._approx_scalar(x) for a, x in self._yield_comparisons(actual)

a == approx(x, rel=self.rel, abs=self.abs, nan_ok=self.nan_ok) for a, x in self._yield_comparisons(actual)
will do the trick, along with removing the TypeError in ApproxMapping (

pytest/src/_pytest/python_api.py

Line 307 in dfc910e

raise TypeError(msg.format(key, value, pprint.pformat(self.expected)))

)