Sorry, just saw this.
I'll have to look at this while less tired, but going through rationals (I realize I talked about this first) might not be the correct approach, as that rational can be smaller than the equivalent decimal representation.

Sorry, just saw this.
I'll have to look at this while less tired, but going through rationals (I realize I talked about this first) might not be the correct approach, as that rational can be smaller than the equivalent decimal representation.

I'm open to other approaches. This isn't my area of expertise, I was just trying to get these bugs fixed based on patches already provided.

Do you have a case that still fails with the proposed approach (proving this approach doesn't go far enough)? Do you have a case that fails with the proposed approach but works with the current approach (proving the proposed approach introduces a regression)? If this causes regressions, we definitely shouldn't do it. If it fixes some cases but not others, but doesn't break any existing cases, it may still be worth applying. If you have an intuition that this approach isn't going to work for all cases, but cannot come up with a failing case, I'm willing to accept that and won't press for this to be merged, but it would be nicer to have proof that this approach has problems.

I didn't try your patch, but IIUC, it uses Rational(self).round(n).to_f? If so, I suspect that (2.914e-17).round(20) would return 2.913e-17 instead of 2.914e-17, right?

I think the correct version would be Rational(self.to_s).round(n).to_f. I believe that this definition would also fix the issue for 291.4 in https://bugs.ruby-lang.org/issues/18018 (but more slowly).

Basically, for this rounding algorithm of floats, the convertion to a rational should be using the decimal notation instead of the base-2 representation. (e.g.: 291.4 => Rational(2914, 10) instead of 2563181506671411/8796093022208)

I checked out this branch and ran p (2.914e-17).round(20) and it printed 2.914e-17. I'm not opposed to switching to the to_s route, but we should at least confirm it is necessary.

Oh, my bad, I confused round with floor 🤦‍♂️.

Thank you @jeremyevans <3