[ruby-core:22791] Re: suggestions for float

On Sat, Mar 7, 2009 at 13:43, Charles Oliver Nutter
<[e-mail]>wrote:

> While I appreciate that returning -0.9 for the above "looks" nicer, it also
> helps reinforce the myth that floating-point results are 100% accurate. The
> bugs JRuby had (and in some cases still has) representing floating point
> math results the same way as MRI have often led to broken code as a
> result...such as a certain piece of code in Rails that multiplied a float
> and an integer and compared it with another float with no margin for error.
> Pretending Floats are accurate leads to bugs.


I agree.

I suspect that in the vast majority of real-world cases of use of
non-integer numbers, the *correct* thing is to use decimal arithmetic. Think
of all the code which does computation of currency amounts, durations of
time, physical quantities, and so on. In all of those cases, binary rounding
errors are unexpected, and people often need to be educated as to why they
are getting them and what they need to do to deal with them.

The number of cases where you actually *want* binary float behavior is
incredibly small. In fact, I can only think of one case in the last decade
or so where I personally wanted binary floats, and that was purely for
performance, not because of their peculiar properties.

So I'd favor making decimal arithmetic be the default in Ruby, and having a
FastFloat or IEEEFloat or some other special class for the few special
situations where you actually want fast but inaccurate arithmetic.


mathew
[I feel that real numbers are a mathematical concept that doesn't exist in
reality. In the physical world, you can go down as far as measuring things
in Planck lengths, and that's about it--there's no such thing as an
infinitely divisible real number in the physical universe. Fixed point
decimal numbers, on the other hand--those actually exist, at least in my
wallet.]