Stream: helpdesk (published)

Topic: Confusing subtyping

Timothy (Nov 26 2021 at 07:48):

I'm a bit confused about why the expected function method isn't being dispatched in my code, and it seems it comes down to this confusing result:

julia> Dict{Char, Vector{Type}} <: Dict{Char, Vector{Type}}
true

julia> Dict{Char, Vector{Type}} <: Dict{Char, Vector{<:Type}}
false

julia> Dict{Char, Vector{DataType}} <: Dict{Char, Vector{Type}}
false

julia> Dict{Char, Vector{DataType}} <: Dict{Char, Vector{<:Type}}
false

If we could work out how to attain the result I'm expecting, that would be great.

For reference, this is the issue I'm having

MethodError: no method matching parseorg(::String, ::Dict{Char, Vector{Type}}, ::Vector{DataType}; debug=false, pointonfail=false)
Closest candidates are:
  parseorg(::AbstractString, ::Dict{Char, Vector{var"#s40"} where var"#s40"<:Type}, ::Vector{var"#s33"} where var"#s33"<:Type; debug, pointonfail)

Sukera (Nov 26 2021 at 07:54):

types in julia are invariant

Sukera (Nov 26 2021 at 07:54):

(except tuples)

Sukera (Nov 26 2021 at 07:55):

see https://docs.julialang.org/en/v1/manual/types/#man-parametric-composite-types

Sukera (Nov 26 2021 at 07:58):

I'm guessing you declared your function like parseorg(s::AbstractString, d::Dict{Char, Vector{T}}, v::Vector{U}) where {T <: Type, U <: Type}? To get the dispatch you want, you'd write that as parseorg(s::AbstractString, d::AbstractDict{Char, <:AbstractVector{T}}, v::AbstractVector{U}) where {T, U}, though that feels overly typed to me

Timothy (Nov 26 2021 at 07:59):

I've written

 parseorg(content::AbstractString, typematchers::Dict{Char, Vector{<:Type}},
                  typefallbacks::Vector{<:Type}; debug=false, pointonfail=false)

Sukera (Nov 26 2021 at 07:59):

yeah, same thing, just without the where

Timothy (Nov 26 2021 at 08:00):

So basically going from a concrete to abstract type should do the trick?

Sukera (Nov 26 2021 at 08:00):

no, acknowledging that types are invariant does the trick ;)

Sukera (Nov 26 2021 at 08:01):

what you've written is asking for a Dict taking a Char as key and pointing to _a_ vector that holds a subtype of Type though

Sukera (Nov 26 2021 at 08:01):

note the <: AbstractVector part

Timothy (Nov 26 2021 at 08:01):

Ah right, thanks for pointing out that it's that bit

Sukera (Nov 26 2021 at 08:02):

do you have to distinguish between different parseorg? if not, I wouldn't necessarily use type annotations at all here

Timothy (Nov 26 2021 at 08:02):

I have two forms, yes. I also like typing function arguments when to let myself know what I'm expecting.

Sukera (Nov 26 2021 at 08:03):

does that form have the same signature/number & order of arguments?

Timothy (Nov 26 2021 at 08:04):

The signature doesn't match. I could get away without typing any of it but then I come back to "I also like typing function arguments when to let myself know what I'm expecting."

Sukera (Nov 26 2021 at 08:04):

Yeah, for documentation purposes it's not a bad idea, though be aware that you don't necessarily want to use concrete types in the signature

Timothy (Nov 26 2021 at 08:04):

On that note, is there a particular reason why I'd want to go for AbstractDict / AbstractVec ?

Sukera (Nov 26 2021 at 08:05):

a view is an AbstractVector :)

Sukera (Nov 26 2021 at 08:05):

no need to copy arrays around when slicing after all

Sukera (Nov 26 2021 at 08:06):

As for AbstractDict- you could switch to Dictionaries.jl or an OrderedDict from DataStructures.jl if you need a different interface or some ordering in your dict

Sukera (Nov 26 2021 at 08:06):

if you have AbstractDict, you don't get MethodError all over the place, since they all satisfy that API

Timothy (Nov 26 2021 at 08:07):

Thanks for elaborating. I've got AbstractString here for exactly the reason you mention with vectors (I'm taking lots of substrings).

Sukera (Nov 26 2021 at 08:08):

but other than that - I usually don't type things and rather document in comments/docstrings what I expect to hold interface-wise, since apart from dispatch, type annotations in function signatures do not improve performance or help inference

Sukera (Nov 26 2021 at 08:09):

the only things I do type religiously is when I have a custom struct internal to my package and I want to make sure to require that struct as an input

Timothy (Nov 26 2021 at 08:09):

I do have more info in docstrings, however for er, somewhat unique reasons, I'm not putting the docstrings next to the functions here.

Timothy (Nov 26 2021 at 08:10):

( if I do I get load order errors, since the docstrings use a macro which uses the code I'm documenting )

Sukera (Nov 26 2021 at 08:12):

:thinking:

Sukera (Nov 26 2021 at 08:13):

are you using @doc?

Timothy (Nov 26 2021 at 08:14):

I am

Sukera (Nov 26 2021 at 08:14):

I'm very curious about how that breaks :o

Timothy (Nov 26 2021 at 08:16):

Well, for the sake of your curiosity, this is what I've got going on:

@doc org"""
some lovely documentation
""" parseorg

and as you may imagine, the @org_str macro definition involves calling parseorg :smile:.

Sukera (Nov 26 2021 at 08:20):

well, I'm sure you have your reasons for that :sweat_smile:

Timothy (Nov 26 2021 at 08:22):

Well, if I'm writing a package to provide nice support for a certain markup language, it would basically be heretical not to write the docs using that markup language wouldn't it? :stuck_out_tongue:

Timothy (Nov 26 2021 at 08:46):

Anyway, thanks for the help!

Last updated: Oct 02 2023 at 04:34 UTC