TLDR: if any of the values inside the PhasePoint are NaN's, we want Turing's sampling to error immediately instead of the current behaviour which is to enter an infinite loop.

Not sure we want this?

I guess there are two things here:

1. Is it sensible to implement isnan for PhasePoint? Probably 🤷
2. Is it sensible to error in Turing.jl when PhasePoint is NaN rather than just rejecting? Probably not 😕

Though I understand (2) is maybe a bit counter-intuitive, the reality is that with these random programs, you might encounter NaNs even though the code is all correct (just because we encounter particularly badly behaved parts of the space). In these cases, we want to reject the points, as they are just 0-probability choices, rather than error and break stop the sample call completely. In particular during warmup/adaptation phases where we can be quite aggressive in tuning of sampler parameters, leading to numerical issues, etc.

I agree with Tor here, we probably don't want to error directly. Introducing some kind of guard with number of attempts should be good.

If we do decide not to error, would isnans here be still useful?

Introducing some kind of guard with number of attempts should be good.

There's already a guard that warns the user, buuuuut I think in practice this isn't as useful as I was hoping when I first introduced it because IO isn't given priority and the user never ends up seeing the warning until they interrupt the computation manually 😬

I did actually see the warning when I ran it!

Yes, if you'd rather error after N attempts (for some large N) then this PR isn't strictly needed. Unless – maybe we could say that after N attempts we can error, and if isnan(z) then add more description to the error (just to help with troubleshooting)? In that case then this would still be useful.

I suppose the question is whether seeing ±Inf means something different from seeing NaN.

My assumption is that NaN indicates that something is wrong and the user should look at it. But if that's not true and you can get NaN's even when doing everything right, then there's no difference (and we shouldn't include it in an error message either).

My assumption is that NaN indicates that something is wrong and the user should look at it.

This is unfortunately not the case 😕 For example

julia> Inf / Inf
NaN

(got it from here: https://docs.julialang.org/en/v1/manual/integers-and-floating-point-numbers/#Special-floating-point-values)

Closing in favour of TuringLang/Turing.jl#2392