JSON.jl

This package provides for parsing and printing JSON in pure Julia. See also alternative packages JSON3.jl, and Serde.jl (based on the Serde Rust library that handles more formats than just JSON).

Type ] add JSON and then hit ⏎ Return at the REPL. You should see pkg> add JSON.

import JSON

# JSON.parse - string or stream to Julia data structures
s = "{\"a_number\" : 5.0, \"an_array\" : [\"string\", 9]}"
j = JSON.parse(s)
#  Dict{AbstractString,Any} with 2 entries:
#    "an_array" => {"string",9}
#    "a_number" => 5.0

# JSON.json - Julia data structures to a string
JSON.json([2,3])
#  "[2,3]"
JSON.json(j)
#  "{\"an_array\":[\"string\",9],\"a_number\":5.0}"

JSON.print(io::IO, s::AbstractString)
JSON.print(io::IO, s::Union{Integer, AbstractFloat})
JSON.print(io::IO, n::Nothing)
JSON.print(io::IO, b::Bool)
JSON.print(io::IO, a::AbstractDict)
JSON.print(io::IO, v::AbstractVector)
JSON.print(io::IO, v::Array)

Writes a compact (no extra whitespace or indentation) JSON representation to the supplied IO.

JSON.print(a::AbstractDict, indent)
JSON.print(io::IO, a::AbstractDict, indent)

Writes a JSON representation with newlines, and indentation if specified. Non-zero indent will be applied recursively to nested elements.

json(a::Any)

Returns a compact JSON representation as an AbstractString.

JSON.parse(s::AbstractString; dicttype=Dict, inttype=Int64)
JSON.parse(io::IO; dicttype=Dict, inttype=Int64)
JSON.parsefile(filename::AbstractString; dicttype=Dict, inttype=Int64, use_mmap=true)

Parses a JSON AbstractString or IO stream into a nested Array or Dict.

The dicttype indicates the dictionary type (<: Associative), or a function that returns an instance of a dictionary type, that JSON objects are parsed to. It defaults to Dict (the built-in Julia dictionary), but a different type can be passed for additional functionality. For example, if you import DataStructures (assuming the DataStructures package is installed)

• you can pass dicttype=DataStructures.OrderedDict to maintain the insertion order of the items in the object;
• or you can pass ()->DefaultDict{String,Any}(Missing) to having any non-found keys return missing when you index the result.

The inttype argument controls how integers are parsed. If a number in a JSON file is recognized to be an integer, it is parsed as one; otherwise it is parsed as a Float64. The inttype defaults to Int64, but, for example, if you know that your integer numbers are all small and want to save space, you can pass inttype=Int32. Alternatively, if your JSON input has integers which are too large for Int64, you can pass inttype=Int128 or inttype=BigInt. inttype can be any subtype of Real.

JSONText(s::AbstractString)

A wrapper around a Julia string representing JSON-formatted text, which is inserted as-is in the JSON output of JSON.print and JSON.json.

JSON.lower(p::Point2D) = [p.x, p.y]

Define a custom serialization rule for a particular data type. Must return a value that can be directly serialized; see help for more details.

Users may find the default behaviour of JSON inappropriate for their use case. In such cases, JSON provides two mechanisms for users to customize serialization. The first method, JSON.Writer.StructuralContext, is used to customize the cosmetic properties of the serialized JSON. (For example, the default pretty printing vs. compact printing is supported by provided two different StructuralContexts.) Examples of applications for which StructuralContext is appropriate include: particular formatting demands for JSON (maybe not in compliance with the JSON standard) or JSON-like formats with different syntax.

The second method, JSON.Serializations.Serialization, is used to control the translation of Julia objects into JSON serialization instructions. In most cases, writing a method for JSON.lower (as mentioned above) is sufficient to define JSON serializations for user-defined objects. However, this is not appropriate for overriding or deleting predefined serializations (since that would globally affect users of the JSON module and is an instance of dangerous type piracy). For these use-cases, users should define a custom instance of Serialization. An example of an application for this use case includes: a commonly requested extension to JSON which serializes float NaN and infinite values as NaN or Inf, in contravention of the JSON standard.

Both methods are controlled by the JSON.show_json function, which has the following signature:

JSON.show_json(io::StructuralContext, serialization::Serialization, object)

which is expected to write to io in a way appropriate based on the rules of Serialization, but here io is usually (but not required to be) handled in a higher-level manner than a raw IO object would ordinarily be.

To define a new StructuralContext, the following boilerplate is recommended:

import JSON.Writer.StructuralContext
[mutable] struct MyContext <: StructuralContext
    io::IO
    [ ... additional state / settings for context goes here ... ]
end

If your structural context is going to be very similar to the existing JSON contexts, it is also possible to instead subtype the abstract subtype JSONContext of StructuralContext. If this is the case, an io::IO field (as above) is preferred, although the default implementation will only use this for write, so replacing that method is enough to avoid this requirement.

The following methods should be defined for your context, regardless of whether it subtypes JSONContext or StructuralContext directly. If some of these methods are omitted, then CommonSerialization cannot be generally used with this context.

# called when the next object in a vector or next pair of a dict is to be written
# (requiring a newline and indent for some contexts)
# can do nothing if the context need not support indenting
JSON.Writer.indent(io::MyContext)

# called for vectors/dicts to separate items, usually writes ","
# unless this is the first element in a JSON array
# (default implementation for JSONContext exists, but requires a mutable bool
#  `first` field, and this is an implementation detail not to be relied on; 
#  to define own or delegate explicitly)
JSON.Writer.delimit(io::MyContext)

# called for dicts to separate key and value, usually writes ": "
JSON.Writer.separate(io::MyContext)

# called to indicate start and end of a vector
JSON.Writer.begin_array(io::MyContext)
JSON.Writer.end_array(io::MyContext)

# called to indicate start and end of a dict
JSON.Writer.begin_object(io::MyContext)
JSON.Writer.end_object(io::MyContext)

For the following methods, JSONContext provides a default implementation, but it can be specialized. For StructuralContexts which are not JSONContexts, the JSONContext defaults are not appropriate and so are not available.

# directly write a specific byte (if supported)
# default implementation writes to underlying `.io` field
# note that this enables JSONContext to act as any `io::IO`,
# i.e. one can use `print`, `show`, etc.
Base.write(io::MyContext, byte::UInt8)

# write "null"
# default implementation writes to underlying `.io` field
JSON.Writer.show_null(io::MyContext)

# write an object or string in a manner safe for JSON string
# default implementation calls `print` but escapes each byte as appropriate
# and adds double quotes around the content of `elt`
JSON.Writer.show_string(io::MyContext, elt)

# write a new element of JSON array
# default implementation calls delimit, then indent, then show_json
JSON.Writer.show_element(io::MyContext, elt)

# write a key for a JSON object
# default implementation calls delimit, then indent, then show_string,
# then separate
JSON.Writer.show_key(io::MyContext, elt)

# write a key-value pair for a JSON object
# default implementation calls show key, then show_json
JSON.Writer.show_pair(io::MyContext, s::Serialization, key, value)

What follows is an example of a JSONContext subtype which is very similar to the default context, but which uses None instead of null for JSON nulls, which is then generally compatible with Python object literal notation (PYON). It wraps a default JSONContext to delegate all the required methods to. Since the wrapped context already has a .io, this object does not need to include an .io field, and so the write method must also be delegated, since the default is not appropriate. The only other specialization needed is show_null.

import JSON.Writer
import JSON.Writer.JSONContext
mutable struct PYONContext <: JSONContext
    underlying::JSONContext
end

for delegate in [:indent,
                 :delimit,
                 :separate,
                 :begin_array,
                 :end_array,
                 :begin_object,
                 :end_object]
    @eval JSON.Writer.$delegate(io::PYONContext) = JSON.Writer.$delegate(io.underlying)
end
Base.write(io::PYONContext, byte::UInt8) = write(io.underlying, byte)

JSON.Writer.show_null(io::PYONContext) = print(io, "None")
pyonprint(io::IO, obj) = let io = PYONContext(JSON.Writer.PrettyContext(io, 4))
    JSON.print(io, obj)
    return
end

The usage of this pyonprint function is as any other print function, e.g.

julia> pyonprint(stdout, [1, 2, nothing])
[
    1,
    2,
    None
]

julia> sprint(pyonprint, missing)
"None"

For cases where the JSON cosmetics are unimportant, but how objects are converted into their JSON equivalents (arrays, objects, numbers, etc.) need to be changed, the appropriate abstraction is Serialization. A Serialization instance is used as the second argument in show_json. Thus, specializing show_json for custom Serialization instances enables either creating more restrictive or different ways to convert objects into JSON.

The default serialization is called JSON.Serializations.StandardSerialization, which is a subtype of CommonSerialization. Methods of show_json are not added to StandardSerialization, but rather to CommonSerialization, by both JSON and by other packages for their own types. The lower functionality is also specific to CommonSerialization. Therefore, to create a serialization instance that inherits from and may extend or override parts of the standard serialization, it suffices to define a new serialization subtyping CommonSerialization. In the example below, the new serialization is the same as StandardSerialization except that numbers are serialized with an additional type tag.

import JSON.Serializations: CommonSerialization, StandardSerialization
import JSON.Writer: StructuralContext, show_json
struct TaggedNumberSerialization <: CommonSerialization end

tag(f::Real) = Dict(:type => string(typeof(f)), :value => f)
show_json(io::StructuralContext,
            ::TaggedNumberSerialization,
           f::Union{Integer, AbstractFloat}) =
    show_json(io, StandardSerialization(), tag(f))

Note that the recursive call constructs a StandardSerialization(), as otherwise this would result in a stack overflow, and serializes a Dict using that. In this toy example, this is fine (with only the overhead of constructing a Dict), but this is not always possible. (For instance, if the constructed Dict could have other numbers within its values that need to be tagged.)

To deal with these more complex cases, or simply to eliminate the overhead of constructing the intermediate Dict, the show_json method can be implemented more carefully by explicitly calling the context’s begin_object, show_pair, and end_object methods, as documented above, and use the StandardSerialization() only for the show_pair call for f.

# More careful implementation
# No difference in this case, but could be needed if recursive data structures are to be
# serialized in more complex cases.
import JSON.Writer: begin_object, show_pair, end_object
function show_json(io::StructuralContext,
                    s::TaggedNumberSerialization,
                    f::Union{Integer, AbstractFloat})
    begin_object(io)
    show_pair(io, s, :tag => string(typeof(f)))
    show_pair(io, StandardSerialization(), :value => f)
    end_object(io)
end

To use the custom serialization, sprint can be used (and this can be encapsulated by a convenient user-defined interface):

julia> JSON.parse(sprint(show_json, TaggedNumberSerialization(), Any[1, 2.0, "hi"]))
3-element Array{Any,1}:
 Dict{String,Any}("value" => 1,"type" => "Int64")
 Dict{String,Any}("value" => 2.0,"type" => "Float64")
 "hi"

If it is not desired to inherit all the functionality of StandardSerialization, users may also choose to start from scratch by directly subtyping JSON.Serializations.Serialization. This is useful if the user wishes to enforce a strict JSON which throws errors when attempting to serialize objects that aren’t explicitly supported. Note that this means you will need to define a method to support serializing any kind of object, including the standard JSON objects like booleans, integers, strings, etc.!