Idea for definition of conversion constants

[giordano]

Currently, the package isn't very arbitrary-precision-friendly, because conversion constants are often of Float64 type. My idea is to define them as Irrational:

julia> using Unitful

julia> Base.@irrational μ 1.256637061435917295385e-06 (4 * big(π) * (1//10) ^ 7)

julia> μ0 = μ * Unitful.H / Unitful.m 
μ = 1.2566370614359... H m^-1

julia> float(μ)
1.2566370614359173e-6

julia> float(Unitful.μ0.val)
1.2566370614359173e-6

julia> big(μ)
1.256637061435917295385057353311801153678867759750042328389977836923126562514468e-06

julia> big(float(μ))
1.256637061435917288506887261323452520400678622536361217498779296875000000000000e-06

julia> big(Unitful.μ0.val)
1.256637061435917288506887261323452520400678622536361217498779296875000000000000e-06

The last three lines show that in this way the conversion factor has full precision as a BigFloat.

Edit: I initially proposed to define a type that was a simplification of Irrational, but in the end there wasn't much difference between my custom type and Irrational.

[ajkeller34]

Sorry for not getting back to you earlier. I like this idea a lot. I'm a little worried that we'd get collisions because someone else might define Irrational{:μ}, otherwise it seems great.

In practice, I'd imagine we could avoid this by using the irrational macro with very unique names like _Unitful_μ. Unique Irrational{:_Unitful_μ} methods would result, as well as a _Unitful_μ symbol. Then we could just say μ = _Unitful_μ so that the symbol is cleaner but we still avoid type collisions.

Also, while I defined float for quantities, I neglected to define big apparently. That'd need to be done too.

I'll get to this eventually but would also welcome a PR if you're feeling up for it.

[giordano]

Thank you! I like your approach to avoid type collision.

I might have some time to bake a PR, but not anytime soon.

[sostock]

With #754, floating-point conversion factors are now <: AbstractIrrational. However, this is only the first step towards arbitrary precision, since these factors are currently still calculated with Float64 precision.