`atol` used by `isapprox` for affine units

[sostock]

As noted in https://github.com/PainterQubits/Unitful.jl/pull/719#discussion_r1600230695, the specified atol is stripped of its units by calling ustrip(unit(y), atol):

https://github.com/PainterQubits/Unitful.jl/blob/b80ab85b5e1bb59c376945e54e723a07b2330cc7/src/quantities.jl#L254

This means that the atol that is actually used depends on whether the units of x and y are affine or not:

julia> isapprox(274.15K, 275.15K, atol=10K) # passing everything in Kelvin gives the expected result
true

julia> isapprox(1°C, 2°C, atol=10K) # this should be the same …
false

[sostock]

The case where the atol itself is specified in °C might be more controversial:

julia> isapprox(1K, 2K, atol=0.1°C)
true

julia> isapprox(1°C, 2°C, atol=0.1°C)
false

IMO, these should both return true, since atol == 273.25K and 1K < 273.25K. But I can see that someone might disagree and think of isapprox(1°C, 2°C, atol=0.1°C) as equivalent to isapprox(1, 2, atol=0.1). Especially since logarithmic quantities already handle it like this:

julia> isapprox(20dBm, 21dBm, atol=10dBm) # equivalent to isapprox(20, 21, atol=10)
true

julia> isapprox(mW(20dBm), mW(21dBm), atol=mW(10dBm)) # equivalent to isapprox(100, 125.89…, atol=10)
false

[cstjean]

FWIW, over here we adhere strictly to "Temperature measurements can be either Kelvin or Celsius, temperature deltas must be in Kelvin", which Unitful nicely checks for us thanks to 0.1°C + 0.1°C being an AffineError. In that perspective, atol=0.1°C would be an error (too ambiguous to be safe).

(I'm aware this isn't necessarily a very popular perspective...)

EDIT: Ah, I guess it's https://github.com/PainterQubits/Unitful.jl/issues/521

[sostock]

I also think that making atol=0.1°C error is the best solution. And since the current behavior is broken (in my opinion), I don’t consider it a breaking change.

(Making all comparisons between affine and linear quantities error, as proposed in #521, is something I consider breaking, since the current behavior is well-defined and consistent, although unexpected to some and therefore error-prone).

[cstjean]

Fair enough!