How to define random functions on your own types need to be better documented.

[KristofferC]

and have better error messages.

On 0.6:

julia> struct F
           v::Float64
       end

julia> rand(F)
ERROR: MethodError: no method matching rand(::MersenneTwister, ::Type{F})
Closest candidates are:
  rand(::AbstractRNG, ::Type, ::Tuple{Vararg{Int64,N}} where N) at random.jl:390
  rand(::AbstractRNG, ::Type, ::Integer, ::Integer...) at random.jl:391
  rand(::MersenneTwister, ::Type{I<:Base.Random.FloatInterval}) where I<:Base.Random.FloatInterval at random.jl:143
  ...
Stacktrace:
 [1] rand(::Type{T} where T) at ./random.jl:283

julia> Base.rand(::MersenneTwister, ::Type{F}) = F(rand())

julia> rand(F)
F(0.5219484358707998)

julia> rand(F, 2,2)
2×2 Array{F,2}:
 F(0.0131819)  F(0.0734573)
 F(0.440896)   F(0.735767)

The error message says exactly what is happening, I define the function, everything works.

On 0.7:

julia> struct F
           v::Float64
       end

julia> rand(F)
ERROR: ArgumentError: Sampler for this object is not defined
Stacktrace:
 [1] Type at ./random/random.jl:106 [inlined]
 [2] rand at ./random/random.jl:193 [inlined]
 [3] rand at ./random/random.jl:194 [inlined]
 [4] rand(::Type{F}) at ./random/random.jl:198
 [5] top-level scope

Ok... what do I do... Looking at docs for rand, says nothing about Sampler. Looking at NEWS.md, nothing about Sampler. Looking at docs, nothing about Sampler. Looking at >help? Sampler, No documentation found.. Let's try the same as in 0.6:

julia> Base.rand(::MersenneTwister, ::Type{F}) = F(rand())

julia> rand(F)
F(0.15546036169717214)

Woho, it works... or?

julia> rand(F, 2, 2)
ERROR: ArgumentError: Sampler for this object is not defined
Stacktrace:
 [1] Type at ./random/random.jl:106 [inlined]
 [2] rand at ./random/random.jl:193 [inlined]
 [3] rand! at ./random/random.jl:210 [inlined]
 [4] rand! at ./random/random.jl:206 [inlined]

Gives up.

[JeffBezanson]

This is a fine example of a general theme I've seen a few times. Sometimes people feel a MethodError is too "low level" and user-unfriendly, and that a more application-specific error message should be substituted for it. But I generally think that it's good to know how the system works, in which case a MethodError is actually quite informative and more useful than an ad-hoc message: the formatting is consistent so you get better at reading MethodErrors over time, and more effort has been put in to MethodErrors to make them include more information.

[KristofferC]

MbedTLS fails with this error message now: https://travis-ci.org/JuliaWeb/GitHub.jl/jobs/330410030#L1289.

Still no idea how to fix. Was this supposed to be just an optimization or an API change? If purely optimization, should be fixed to not break code, if API change, should have deprecation and NEWS:

[stevengj]

I agree with @KristofferC that the issue here is not MethodError at all. The issue is:

• Sampler etc should be documented if is needed to extend rand
• The problem here is that the ArgumentError (not a MethodError!) is obscure and needs to be rewritten/replaced.

[rfourquet]

I will address the issue.

[rfourquet]

The error message says exactly what is happening, I define the function, everything works.

The previous rand system could be specialized for custom types, but it was not designed for extension nor documented as such. Sure, you could get away with defining ad-hoc functions according to the MethodError messages, but you would end-up re-inventing the wheel in some cases. E.g. (julia v0.6):

julia> Base.rand(::MersenneTwister, f::F) = rand()*f.v

julia> julia> rand(MersenneTwister(0), F(1.1))
0.7166581394375923

julia> julia> rand(F(2.0), 2, 2)
ERROR: MethodError: no method matching rand(::F, ::Int64, ::Int64)
Closest candidates are:
  rand(::Integer...) at random.jl:280
  rand(::Type, ::Integer, ::Integer...) at random.jl:282
  rand(::AbstractArray, ::Integer...) at random.jl:297
  ...

and you now have to define a loop yourself for array generation.

Gives up.

Yes, I fully agree the new system needs documentation and a better error message. In the PR introducing Sampler, I had of course the intent to document this eventually (before the 0.7 release), but it was still experimental (with of course all base tests passing), and I felt this was useful to have this system used for some time in Base during the 0.7 timeframe, to catch potential design mistake and/or find improvement opportunities.

[rfourquet]

Sometimes people feel a MethodError is too "low level" and user-unfriendly, and that a more application-specific error message should be substituted for it

This is not what happened here, cf. #23964 (comment)

[KristofferC]

Bump, what is the status on this? I would like to reiterate how awesome it would be if this could be made non-breaking.

[fredrikekre]

Bump; documentation for this is very much needed now that we start upgrading to v0. 7.

[rfourquet]

Bump; documentation for this is very much needed now that we start upgrading to v0. 7.

Message received!

[rfourquet]

I would like to reiterate how awesome it would be if this could be made non-breaking.

I believe this is not possible while keeping the benefits of the new framework. I would like to reiterate how limited the previous API was; I know you are not a big fan of this change, but hope you will come to appreciate the new possibilities.

[zsunberg]

It's still kind of difficult to find and understand the documentation for this. If I google "ERROR: ArgumentError: Sampler for this object is not defined", it brings me here.

When I finally found the right part of the docs (https://docs.julialang.org/en/v1/stdlib/Random/#Generating-values-from-a-collection-1), I was very surprised to find that, instead of learning how to implement a Sampler, I needed to implement rand Random.rand(rng::AbstractRNG, st::Random.SamplerTrivial{MyDist}). I think the "Sampler for this object is not defined" error message could be greatly improved to point users more quickly toward a way to fix the problem :)