Non-string Dict keys - regression from JSON3.jl

[carpetdev]

Using JSON3.jl, I've been successfully roundtripping dictionaries with integer keys:

JSON3.read(JSON3.write(Dict(1=>1), Dict{Int,Int})
# Success

I'm aware that JSON object keys are strings so the produced JSON is actually "{\"1\":1}", but I've assumed the successful deserialization was by design.

Using JSON.jl, this errors:

JSON.parse(JSON.json(Dict(1=>1)), Dict{Int,Int})
# ERROR: MethodError: Cannot `convert` an object of type String to an object of type Int64

Is this to do with StructUtils.jl being less forgiving by default than StructTypes.jl?

In my specific use case I can actually replace my dictionary with a vector, but in general I might not have been able to.

[quinnj]

Thanks for reporting this! You're correct that there's a gap here - we define lowerkey to convert integer keys to strings when writing:

StructUtils.lowerkey(::JSONStyle, n::Union{Integer, Float16, Float32, Float64}) = string(n)

But we didn't add the corresponding liftkey to convert strings back to integers when parsing. This is an oversight that has now been fixed.

Fix: #423 adds the missing liftkey methods:

StructUtils.liftkey(::JSONStyle, ::Type{T}, x::AbstractString) where {T<:Integer} = Base.parse(T, x)
StructUtils.liftkey(::JSONStyle, ::Type{T}, x::AbstractString) where {T<:AbstractFloat} = Base.parse(T, x)

After this fix, round-tripping works correctly:

julia> JSON.parse(JSON.json(Dict(1=>1)), Dict{Int,Int})
Dict{Int64, Int64} with 1 entry:
  1 => 1

Workaround (until the fix is released):

You can define these methods yourself:

using JSON, StructUtils
StructUtils.liftkey(::JSON.JSONStyle, ::Type{T}, x::String) where {T<:Integer} = parse(T, x)
StructUtils.liftkey(::JSON.JSONStyle, ::Type{T}, x::String) where {T<:AbstractFloat} = parse(T, x)

[carpetdev]

Thanks for the fix and the comprehensive explanation!