Julia: make order irrelevant to ambiguous method warnings

[JeffBezanson]

It would be nice to be smarter about giving these warnings, so that if the needed method is defined, say, before the end of a file, no warning is printed.

[vtjnash]

would it be reasonable to make the warnings print when a module is finished loading (with a special case for Main that continues with the current behavior)? Anyways, I don't see this as release blocking.

[StefanKarpinski]

Another option is making it parse-unit-based. So warn at the end of file for file input or per input expression in the repl.

[timholy]

Now I'm beginning to wonder whether these warnings are the right way to go. I just loaded two packages, Images and Calendar, and was encouraged to define methods for doing arithmetic on Images of dates:

Warning: New definition 
    +(AbstractArray{T1<:Union(AbstractCalendarDuration,CalendarTime),N},AbstractArray{T2<:Union(AbstractCalendarDuration,CalendarTime),N}) at operators.jl:228
is ambiguous with 
    +(AbstractImageDirect{T,N},AbstractArray{T,N}) at /home/tim/.julia/Images/src/algorithms.jl:102.
Make sure 
    +(AbstractImageDirect{T1<:Union(AbstractCalendarDuration,CalendarTime),N},AbstractArray{T2<:Union(AbstractCalendarDuration,CalendarTime),N})
is defined first.

and so on.

While it's interesting to consider, I don't currently intend to define an image where each pixel is a CalendarTime. So in this case, this warning isn't terribly helpful.

I guess the problem is that, as the number of packages grows, it gets to be hard to avoid these warnings because one can't anticipate the various combinations of packages that will be used.

[diegozea]

Related: https://github.com/JuliaLang/julia/issues/3178

[JeffBezanson]

The warnings are always telling you something useful, even if adding crazy new methods is not the best fix. In fact the warning should also say "If the suggested new method seems crazy, then one of the other methods should probably be deleted."

Packages should really not define things like +(Array{Foo}, Array{Bar}), because + for arrays is already defined generically. And when others (like yourself) define other kinds of arrays, conflicts are then inevitable.

[vtjnash]

My initial thought was that perhaps a module should have a keyword to turn off all ambiguity warnings within that module (and make them happen at runtime when the user attempts to do something like images with CalendarTime pixels).

However, in this case, it's not immediately clear to me why an AbstractImage is a type of AbstractArray. While an image can be represented as an array and in many cases converted to an array, I think that is a rather different thing from saying that an image is a type of array.

[JeffBezanson]

Maybe, but I say overriding +(AbstractArray, AbstractArray) for a certain element type is the worse infraction here. There is something intrinsically sketchy about that, while there is nothing intrinsically sketchy about defining a new type of AbstractArray.

[JeffBezanson]

Also, I suspect our vectorize macros should only generate definitions for Array (since they only construct Arrays) and not AbstractArray.

[timholy]

Wow, this is the 11th-oldest open issue. Fun!

[StefanKarpinski]

Thanks, @timholy – I actually strongly believe that this will improve the "feel of maturity" of Julia considerably. All those warnings when loading packages, even if they were harmless did not look good or give a warm and fuzzy feeling.

[timholy]

julia-0.5 just feels SO solid. Aside from having some amazing new features and fewer performance gotchas than ever, I am grateful on a nearly daily basis for the work ferreting out GC/dispatch/etc problems and things like the work done improving the line numbers in backtraces. It really adds up to a great experience.