APJSON

Aron Kovacs

4.67/5 (5 votes)

28 Aug 2013CPOL13 min read

42.4K

1.2K

This is an alternative for fastJSON

Introduction

This is an alternative version of Gholam's great JSON library which brings a new deserializer, code and workflow optimizations and some new features at acceptable performance cost (if).

Key Features (v1.0)

Event faster deserialization of a json string into generic values (IJsonValue)
Type extension support (custom type names)
Custom type support (custom (de-)serializers)
Silverlight 5 support
Optionally checks for DataMember attribute.
Handles DataMember.Name representing the json field name
Handles IgnoreDataMember by ignoring properties decorated with it.
Handles date time kind through the JsonDateTimeOptions attribute.
Handles custom date formats through the JsonDateTimeOptions attribute.
[New] Supports HashSet<T> (with some performance penalty)
[New] Json.Current.BuildUp supports now JsonArray into collection types.
(Most) built-in value types are supported.
Non-abstract reference types are supported (including inheritance)
Custom type support.
Enhanced debugging capabilities purposes see the GetSerializationMembers method
Committed to quality over performance.

==> See changelog at the bottom of this article.

==> See fastJSON 2.0.9 feature list for more complete list of features.

Using the code

Since ApJson is built on top of fastJSON, therefore it's API has some overlap:

Serialize an object

Apolyton.FastJson.Json.Current.ToJson(c);

Note that in all cases in which no explicit parameter object is provided, the default parameters are used.

Deserialize an object

Option 1: Straight forward

In order to deserialize an object directly (with custom type support), your call should like the following:

string jsonText =  "{...}"; // your json string 
var myObject = (MyClass)Apolyton.FastJson.Json.Current.ReadObject(jsonText);

Note 1: type extensions must be enabled and present in the jsin string for this feature to work (otherwise the deserializer cannot determine, which kind of object to create.
Note 2: there is also a generic version of that method

Option 2: Sniff'n go

However, A-FastJson comes with another deserializer, the JsonValueDeserializer which returns a value store based on IJsonValue. This operation is a lot faster than the ReadObject method shown above, but is not that strongly typed. Reading into a json value can be accomplished by

string jsonText =  "{...}"; // your json string 
JsonObject myObject = Apolyton.FastJson.Json.Current.ReadJsonValue(jsonText);

The JsonObject class is essentially a dictionary which allows you to sniff into its values before continuing the deserialization process. This is for instance useful for protocol validation since it allows you to trash the json request before full deserialization has happened which saves time and increases potentially your i/o (ReadJsonValue is almost twice as fast than ReadObject, see below).

An instance of JsonObject can be used to populate an existing instance of your CLR objects. For that, use the BuilUp method on the Json singleton:

string jsonText =  "{...}"; // your json string 
var deserializedStore = (JsonObject)Apolyton.FastJson.Json.Current.ReadJsonValue(jsonText);
var target = new MyClass();

Apolyton.FastJson.Json.Current.BuildUp(target, deserializedStore);

For optimal performance, you should pool your target objects; The BuildUp method is supposed to ensure that you can recycle your instances.

Configuration

By default, the Json class instance uses the default parameters for serialization and deserialization. Due to internal mechanisms, these configuration objects are expensive in contrast to those in fastJSON. This is because all serialization meta-information is attached to it.

The parameter names should be self explicit and are explained briefly in code. Since v0.93, there is a clear separation of parameters which refer to serialization and deserialization. Properties on the JsonParameters class refer to both operations. UseGlobalTypes is obsolete.

Inspecting the configuration

For debugging or automated protocol checks, one may want to see the list of properties which are visible to the JSON (de-)serializer. For that, you can use the GetSerializationMembers method.

// For the default parameter
Apolyton.FastJson.Json.Current.GetSerializationMembers(typeof(MyClass)); 
or  
// For any parameter object
new JsomParameter().GetSerializationMembers(typeof(MyClass);

Custom Type Support

The custom type support of FastJSON has been extended and reviewed in order to be platform independent. In essence, a parameter object can register a serialization and deserialization handler which are plain delegates. This can happen for instance through:

// Preamble for illustration purposes only
JsonParameters parameters = CreateTestParameters();
JsonSerializer serializer = new JsonSerializer(parameters);
CustomTypeClass customTypeClass = new CustomTypeClass() { Custom = new CustomTypeClass.CustomType() };

// Register two delegates, we serialize to 'yes', we deserialize to 'null' 
parameters.RegisterCustomType(typeof(CustomTypeClass.CustomType), 
       (obj) => { return "yes"; }, 
       (obj) => { return null;  });

String jsonString = serializer.Serialize(customTypeClass);

That will render to a JSON string like:

{"Custom":"yes"}

The submitted inline-delegate is called for each instance of the given type. Note that you cannot register a custom type handler for internal types. If you try so, an ArgumentException is thrown by the register method.

For more maintainable code, one can also implement the ICustomTypeSerializer interface and register that instance as a custom type handler:

There is already a base class, namely CustomTypeSerializer, which pre-implements that interface and lets us focus on the important part of our code (and not the formal one):

internal class ObjectIdSerializer : CustomTypeSerializer
{
    /// <summary>
    /// Gets the type for which this serializer is responsible for.
    /// </summary>
    public override Type Type
    {
        get { return typeof(ObjectId); }
    }

    /// <summary>
    /// Returns true: yes, this class can deserialize ObjectId.
    /// </summary>
    public override bool CanDeserialize
    {
        get{return true; }
    }

    /// <summary>
    /// Returns true: yes, this class can serialize ObjectId.
    /// </summary>
    public override bool CanSerialize
    {
        get { return true; }
    }

    /// <summary>
    /// Returns the string of the object id.
    /// </summary>
    /// <param name="data"></param>
    /// <returns></returns>
    public override string Serialize(object data)
    {
        return data.ToString();
    }

    /// <summary>
    /// Returns the object id representing the string.
    /// </summary>
    public override object Deserialize(string jsonString)
    {
        if (!String.IsNullOrEmpty(jsonString))
        {
            return new ObjectId(jsonString);
        }
        else
        {
            return ObjectId.Empty;
        }
    }
}

That example serializes Mongo ObjectId structs into strings and the way back. Note that we didn't implement TypeName which can define the name of the type (see Type Descriptors and Polymorphism); returning null leads to default behavior (recommended).

Type Descriptors and Polymorphism

This chapter is mostly relevant for advanced deserialization scenarios. Serializing objects works normally without need of interaction or configuration. The limits of configuration-free deserialization is reached, when polymorphic objects are used.

Understanding the problem

A simple scenario involving polymorph objects is a list of animals in which each item can be some concrete sort of animal, like dog or cat. Without type descriptors, the deserializer would deserialize each item into an animal. So given a list of 3 animals, 2 dogs and one cat, the serializer would do its job properly by writing a json array with 3 items and their properties within:

[
  { "name"="Brutus", barkLevel="high", "power"="medium" }, 
  { "name"="Bronto", barkLevel="high", "power"="high"},
  { "name"="Silvester", "intelligence"="low", "creativity"="high", "luck"="not-existent" }
]

However, the deserializer just sees a list of 3 animals. Given the text information above, the type of line is lost. Therefore, the deserliazed list would only contain a list with 3 Animals ignoring all additional information in the stream. A solution to this problem is to serialize the type information as well, normally into the special property named $type:

[
  { "$type"="dog", "name"="Brutus", barkLevel="high", "power"="medium" }, 
  { "$type"="dog", "name"="Bronto", barkLevel="high", "power"="high"},
  { "$type"="cat", "name"="Silvester", "intelligence"="low", "creativity"="high", "luck"="not-existent" }
]

The value of the type field can be controlled through type descriptors:

Type Descriptors

The key responsibility of type descriptors is to describe a type by a string. By default, FastJson and ApJson describe a given type by the assembly qualified name which is an awkward long string:

For System.Object:
"System.Object, mscorlib, Version=4.0.0.0, Culture=neutral, PublicKeyToken=b77a5c561934e089"

The good about this string is, that within the .NET world it should work nicely, the bad thing about these type names are their portability and length. Obviously, the question 'What is a good type name for a user type?' cannot be answered by this toolkit. Although the recommendation is to use the DataContractTypeDescriptor, the type description can be custom by implementing an own sub-class of JsonTypeDescriptor and register it to the parameters:

As mentioned above, the JsonTypeDescriptor generates assembly qualified type names. It is easy to imagine (and implement) type descriptors which use the FullName of a type or just the Name. However, the recommended way is to use the DataContractTypeDescriptor:

As its name indicates, the descriptor is using the DataContractAttribute to determine the name of the type. As a fallback mechanism, the FullName of the type is used (if the given type does not have a DataContract attribute on it or it has no Name assigned). Assigning a name to one of our classes is straight forward:

[System.Runtime.Serialization.DataContract(Name="dog")]
public class Dog
{
  ...
}

Optionally, we also might want to provide a namespace, but the example above will already lead to a readable output which is illustrated above.

Points of Interest

Always reuse configuration (JsonParameter) objects. Performance drops dramatically, if you don't.
The IJsonValue concept is inspired from the JSON API in Silverlight done by Microsoft. I found it very useful to sniff into incoming requests and trash them when they are not following basic expectations (mandatory fields missing). Running a full deserialization in order to do the samewas appearing to me as a waste of resources.
The code is covered by unit tests (around 160+) which should ensure high quality of each release.
Byte enumeration is not supported. Use byte[] instead.
DataMember.OrderNumber is ignored
JsonDateTimeOptions.Format follows the specification of DateTime.ParseExact.
Avoid using internal properties or members in (de-) serialization. The framework reacts partially quite strangely when those properties are accessed through reflection -and in most cases, there is a negative performance impact (that's a .NET thing). Silverlight will for instance throw access violation or method not found exceptions for internal types.

Benchmarks

Before crunching some numbers, some points should be enlightened since they are easily forgotten:

All benchmarks published by framework developers, like this or fastJSON are optimized results. Real life results can look very differently. This is not bad intention, but perfectly normal. Therefore:
All results published here just performance indicators. You should compare the framework's performance in your end-to-end scenario while considering that:
Performance varies not only from run-to-run, but also from class-to-class and the data in it.
All benchmarks are run against fastJSON 2.0.13 which was the reference at publish time.
When reading benchmark results ensure that input and output are the same (*1)

(*1) For instance, the fastJSON benchmark compares with BinaryFormatter which takes a stream as input. fastJSON cannot process streams. This aspect is ignored, but it leads to a potentially wrong assumption that fastJSON is faster than BinaryFormatter..

Scenario 1: A-FastJSON vs fastJSON x86 (custom types)

(A)-FastJSON serialization is usually 20% faster than fastJSON.
(A1)-FastJSON deserialization to IJsonValue is faster than serialization and ~300% faster fastJSON.
(A2)-FastJSON deserialization to IJsonValue, then build up of into given class is almost ~20% faster than fastJSON (+10% compared to v0.91)
(A3)-FastJSON deserializatio9n to IJsonValue, then build up of into given class with type extensions, custom type name (Data Contract support) is ~9% faster than FastJson.
(A4)-FastJSON deserialization into object is ~10% faster or equal to fastJSON
(A5)-FastJSON deserialization into a known object is ~20% faster than fastJSON (ReadObject<T>)

Scenario 2: A-FastJSON vs fastJSON x86 (custom types & exotic types)

(A)-FastJSON serialization is usually 15% faster than fastJSON.
(A1)-FastJSON deserialization to IJsonValue is faster than serialization and ~300% faster fastJSON.
(A2)-FastJSON deserialization to IJsonValue, then build up of into given class is almost ~20% faster than fastJSON (+10% compared to v0.91)
(A3)-FastJSON deserializatio9n to IJsonValue, then build up of into given class with type extensions, custom type name (Data Contract support) is slower than FastJson.
(A4)-FastJSON deserialization into object is ~19% faster or equal to fastJSON
(A5)-FastJSON deserialization into a known object is ~15% faster than fastJSON (ReadObject<T>)

Note that 64 Bit scenarios are not listed anymore since they don't reveal surprising measures (and are therefore not meaningful).

Conclusions

Both, ApJson and fastJSON are reasonably fast.
In 64 bit scenarios the advantage of ApJson is lower.
If one needs built in data set/ data table support, fastJSON is likely the better option.
If one needs exotic type support (dictionaries, hash-sets etc), ApJson is likely the better option.
The IJsonValue conversion is surprisingly fast and seems to be a very good option -especially, if one considers that the second step, converting the generic dictionary into a given type, is optional.
Due to code optimizations, ApJson is meanwhile faster than FastJson in most of the scenarios it defines.

Known Issues and Limitations

DataSet and DataTable are not supported by the BuildUp method (*2). Anyway, the value to convert from one generic type into another one should be quite low. Given enough interest, extension methods might be added on IJsonValue.
HashSet in deserialization is not supported (lack of interface to populate the hash set collection).
DataMember with Name = "$type" and comparable is still allowed and leads potentially to wrong behavior.
InvalidProgramException is thrown when property indexer (aka this[]) is defined on class to serialize
Deserialization into non-public types fails with TypeAccessException.
UseGlobalTypes doesn't work with JsonValueDeserializer.

(*2) DataTable and DataSet can be deserialized by the ReadDataTable/ ReadDataSet methods, however this part of the code is currently not tested.

Release Notes:

v1.0:

Small performance improvements using JsonValue deserialization.
[NEW] Can now deserialize into a HashSet<>
[NEW] Allowing to buildup json arrays.
[NEW] Deserialization into enumeration types is now trying to set an array of the items (instead of exception)
[FIX] Registry did not always respect property read/ write specifiers
[FIX] Crash on deserializing into non-generic Array.
[FIX] JsonPropertyInfo.Copy copied too not relevant values for Silverlight.
[FIX] Silverlight property/ field getter for certain types resulted in NullReferenceExceptions
[CHANGE] JsonParameter.UseGlobalTypes has been removed.
Other fixes..

v.93 Release Candidate

Small performance improvement 5%-10%.
[Cancelled] (Cannot declare explicit cast operators on interfaces) IJsonValue.
[New] Explicit cast operators on JsonPrimitive class
[New] JsonParameter is reviewed. The old version misses clear separation of what is serialization option, what is deserialization and what is common.
[New] Custom date format support through 'JsonDateTimeOptions.Format' format string (note that automatic conversions are supported).
[Fix] JsonPrimitive decimal was using integer parsing.
[Fix] JsonPrimitive ToChar was not throwing an exception when local value had more than one character
[Fix] IJsonValue implementations now throw the promised NotSupportedException instead of InvalidOperationException.
[Fix] BuildUp built Dictionary<,> incorrectly.
[Change] Setting default UseExtensions to false since it is not required in most of the cases.
[Change] JsonValue builder will become the default deserializer except for data table and data set objects.
[Change] SerializationPolicy is obsolete and to be replaced with MemberStrategy.
[Change] Visibility scope of JsonObject, JsonValue, JsonArray corrected down to internal (they are read-only objects).
[Change] DateTimeKind specifications (local, Utc, etc) have only an implicit effect on deserialized values, if the string values is zulu time. Otherwise, the kind is set according to the option, but the value is not modified.
[Change] Type Extensions are now disabled by default as this is an advanced case and has a significant performance impact.

v0.92

First intermediate release before API stability is guaranteed (in v1.0)

Removing support of xmlignore attribute, it is replaced by DataMember, IgnoreDataMember
New: JsonPrimitive custom type support.
New: Silverlight 5 support.
New: A SerializationException is thrown when an attempt is done to serialize a class without (visible) properties. Previously, this was rendering to '{}' which will lead to problems on receivers side.
New: Json.ToJsonBytes returns a byte array representing the json string bytes in the configured encoding.
New: A date time property can be decorated with the DateTimeOptions attribute which defines the expected kind of date time (utc or not). The deserializer will automatically convert, if appropriate.
Change: Thread static attribute has been removed from JSON class. It was causing irritations, since each thread had its own default parameters. Any change applied to the parameters needed to be re-set for each worker thread.
Fix: Can register custom type handler for non default JSON parameters.
Fix: DateTime deserialization was always converting to local time. Now the field can declare the JsonDateTimeOptions attribute which allows to specify the desired value.
Fix: If last property is null or empty and SerializeNullValues is true, a final ',' was rendered anyway (ie. { i:1,} where 'j' is nullable and null).
Increasing unit test depth on JsonRegistry and other related classes.
Improved error reporting when duplicate data member is detected.

v0.9 Fork from fastJSON 2.0.9

Serialization

DateTime to UTC is now respecting the kind of the date time (JsonSerializer_DateTimeUtc)
Byte array was not proper when a List of bytes was serialized (see JsonSerializer_ByteEnumeration)
Serialization of custom types implementing IList was not considered
TimeSpan was wrongly serialized.

Deserialization

Deserialization of a number of bytes was failing
Wrong date time was returned when given string was in UTC and parameter was set to avoid UTC dates

JsonParameter

Changing serialization/ deserialization values at runtime could lead to unexpected output.
Code was changing some properties on its own behalf.

MyPropInfo

CanWrite was always false for fields
Fill was never false and unused. Removed.
GenericTypes was always null except for dictionaries with the name 'Dictionary'.

Benchmark

Fixing time measure flaw in benchmark tool.

History

28th August 2013: Release of v1.0 (after long testing period).

30th January 2013: Release of v0.93
4th January 2013: Release of v.92
26th November 2012: Release of v.90
25th October 2012: Fork of FastJSON. Development started.

License

This article, along with any associated source code and files, is licensed under The Code Project Open License (CPOL)