Deserialization depends on lexicographical ordering of types

[simsurace]

This is unexpected and may be a bug.

using JSON3, StructTypes

struct Type1
    name::String
    n::Int
end

struct Type2
    name::String
end

struct Type3
    name::String
    n::Int
end

json_string = """
    [
        {
            "name": "test",
            "n": 1
        },
        {
            "name": "test"
        }
    ]
    """

JSON3.read(json_string, Vector{Union{Type1,Type2}}) # [Type1("test", 1), Type2("test")]
JSON3.read(json_string, Vector{Union{Type3,Type2}}) # [Type2("test"), Type2("test")]

The parsing as Type2 in the second example drops one field of the json data.

[simsurace]

Ok, I digged into it a bit. This seems to come down to the comment here:

JSON3.jl/src/structs.jl

Lines 70 to 89 in 02e2ecf

	@generated function read(::Struct, buf, pos, len, b, T::Union; kw...)
	U = first(T.parameters) # Extract Union from Type
	# Julia implementation detail: Unions are sorted :)
	# This lets us avoid the below try-catch when U <: Union{Missing,T}
	if U.a === Nothing || U.a === Missing
	:(if buf[pos] == UInt8('n')
	return read(StructType($(U.a)), buf, pos, len, b, $(U.a))
	else
	return read(StructType($(U.b)), buf, pos, len, b, $(U.b); kw...)
	end)
	else
	return :(
	try
	return read(StructType($(U.a)), buf, pos, len, b, $(U.a); kw...)
	catch e
	return read(StructType($(U.b)), buf, pos, len, b, $(U.b); kw...)
	end
	)
	end
	end

And JSON3.jl not failing when there are extra fields. So JSON3.read("{\"name\":\"test\",\"n\":1}", Type2) does not fail.
Might we add a keyword argument ignore_extra_fields::Bool=false to JSON3.read to opt into throwing an error when an extra field is encountered?