A functional Kotlin library to serialize/deserialize Json fast and safely without using reflection, annotations or code generation.

Kondor is based on the concept of a JsonConverter, which is a class that maps each domain class to its Json representation. The converter defines how the Json would look like for each of the types you want to persist, using a high level DSL.

No need of custom Data Transfer Objects, and custom serializer, no matter how complex is the Json format you need to use. You can also define more than one converter for each class if you want to have multiple formats for the same types, for example in case of versioned api or different formats for Json in HTTP and persistence.

Maven

<dependency>
   <groupId>com.ubertob.kondor</groupId>
    <artifactId>kondor-core</artifactId>
    <version>3.2.2</version>
</dependency>

Gradle

implementation 'com.ubertob.kondor:kondor-core:3.2.2'

A live code presentation to see how to use Kondor and some insights on how it works internally:

• It's not easy to insert existing Json inside a converter
• New Kondor-mongo to reuse the converters working with MongoDb
• Automatic JsonSchema generation from any converter
• Code generation of converters from data classes
• JMap works with any type as key (thanks to Alistair O'Neill)

Let's say we have a class with some fields:

FileInfo(
   name = "filename",
   date = Instant.parse("2021-07-01T10:15:30Z"),
   isDir = false,
   size = 123,
   folderPath = "/tmp"
)

And we want to render it to this Json:

{
   "creation_date": 1625134530000,
   "file_name": "filename",
   "folder_path": "/tmp",
   "is_dir": false,
   "selected": true,
   "size": 123
}

What we need to do first, is to define a converter for our class. Usually Kondor converters are Kotlin objects named with a J in front of the class name, but you can use a different convention.

With the converter JFileInfo we can parse a Json string in this way:

val fileInfo: FileInfo = JFileInfo.fromJson(jsonString).orThrow()

And we can render the object in Json in this way:

val json: String = JFileInfo.toJson(fileInfo) 

To write the converter we use a simple DSL:

object JFileInfo : JAny<FileInfo>() {
   val file_name by str(FileInfo::name)
   val creation_date by num(FileInfo::date)
   val is_dir by bool(FileInfo::isDir)
   val size by num(FileInfo::size)
   val folder_path by str(FileInfo::folderPath)

   override fun JsonNodeObject.deserializeOrThrow() =
      FileInfo(
         name = +file_name,
         date = +creation_date,
         isDir = +is_dir,
         size = +size,
         folderPath = +folder_path
      )
}

Each field (id,name) need to be associated to a decoder and a field in the mapped object. Then we need to explicitly define the function for the deserialization.

Converters themselves can be automatically generated from the domain classes, you only have to copy and paste them in your code base, and adapting them as you need. Note that there is no automatic update if you change the data class, the whole point of Kondor-Json is to have converters that maps your classes to a clearly described Json format. To use generators, you need to import kondor-tools module as test dependency.

Comparing with a solution involving writing DTOs, it's quicker to use converters (especially if you use the generator). Even without considering DTOs and the generator, the time needed to write the converters is roughly the same of annotating the classes one by one, but it's easier and more IDE friendly to write the converter. Instead of having to find the right annotation (StackOverflow anyone?), the IDE can suggest the possible converters or you can write new ones.

Another advantage of converters is that it's very easy to define different converters for same domain class in different api, for example we can define JFileInfoV2 to map the same domain class to a different Json format.

Moreover the converters have all the informations to produce very friendly and precise error messages:

error at parsing: Expected a Double at position 55 but found '"' while parsing </items/0/price>

Finally converters know how to generate a Json schema, for example:

JFileInfo.schema().render(pretty)

Will output the following Json schema:

{
   "properties": {
      "creation_date": {
         "type": "number"
      },
      "file_name": {
         "type": "string"
      },
      "folder_path": {
         "type": "string"
      },
      "is_dir": {
         "type": "boolean"
      },
      "size": {
         "type": "number"
      }
   },
   "required": [
     "file_name",
     "creation_date",
     "is_dir",
     "size",
     "folder_path"
   ],
  "type": "object"
}

I made a video explaining how to generate the converters quickly using ChatGPT:

https://youtu.be/4v0VuxZeIIY?t=500

Make sure to import kondor-tools as a dependency of the project. Then you create a new kotlin file for your converters and add a temporary main function that calls kondorGenerator, like this:

fun main() {
    println(
        kondorGenerator(MyClass::class, AnotherClass::class)
    )
}

When you run it, it will output a complete file (including imports) with all the JConverters for the specified classes.

Then you can copy-paste the result in your file (and delete the main function if you have done).

Or you can open Kotlin REPL and type it, or you can make a Kotlin script to write the file, or if you really want, you can write an IntelliJ plugin to make the generation a breeze (just let me know in case!).

Not all the types are supported by the code generator (yet) but it gives a good try.

In the module kondor-tools there are also some useful functions for testing:

expectSuccess and expectFailure simplify test expectations:

val actual = JCustomer.fromJsonNode(json).expectSuccess()

expectThat(actual).isEqualTo(expectedCustomer)

isEquivalentJson compares two Json strings and return Success if they are equivalent:

val json1 = """
    {
    "name" : "Frank",
    "age" : 32,
    "married" : true,
    "children" : ["Ann", "Bob", "Cathy"] 
    }
""".trimIndent()

val json2 = """{"age": 32,"name": "Frank", "children": [  "Ann", "Bob","Cathy"], "married": true} """

json2.isEquivalentJson(json1).expectSuccess()

To use Kondor we need to define a Converter for each type (or class of types).

Let's analyze an example in details:

data class Product(val id: Int, val shortDesc: String, val longDesc: String, val price: Double?) // 1

object JProduct : JAny<Product>() { // 2

   val id by num(Product::id) // 3
   val long_description by str(Product::longDesc) // 4
   val `short-desc` by str(Product::shortDesc) // 5
   val price by num(Product::price) // 6

   override fun JsonNodeObject.deserializeOrThrow() = // 7
      Product( // 8
         id = +id, //9 
         shortDesc = +`short-desc`,
         longDesc = +long_description,
         price = +price
      )
}

1. This is the class we want to serialize/deserialize
2. Here we define the converter, inheriting from a JAny<T> where T is our type. If we want to serialize a collection we can start from JList or JSet and so on, we can also create new abstract converters.
3. Inside the converter we need to define the fields as they will be saved in Json. For each field we need to specify the getter for the serialization, inside a function that represent the kind of Json node (boolean, number, string,array, object) and the specific converter needed for its type. If the converter or the getter is not correct it won't compile.
4. The name of the field is taken from the variable name, long_description in this case
5. Using ticks we can also use names illegal for variables in Kotlin
6. Nullable/optional fields are handled automatically.
7. We then need to define the method to create our objects from Json fields. If we are only interested in serialization we can leave the method empty.
8. Here we use the class constructor, but we could have used any function that return a Product
9. To get the value from the fields we use the unaryplus operator. It is easy to spot any mistake since we match the name of parameter with the fields.

When failing to parse a Json, Kondor is not throwing any exception, instead fromJson and fromJsonNode methods return an Outcome<T> instead of a simple T. Why is that?

Outcome is an example of the Either monad specialized for error handling patterns, if you are not familiar with it, here there are 5 ways to handle errors depending on the case:

1. orThrow()

JCustomer.parseJson(jsonString).orThrow()

this throws an exception if there is an error.

1. orNull()

JCustomer.parseJson(jsonString).orNull()
   ?.let { customer ->
      //do something only if successful
   }

this returns null if there is an error, it's not great because the error is lost but it can be convenient sometime.

1. onFailure{}

val customer = JCustomer.parseJson(jsonString)
   .onFailure { error ->
      log(error)
      return
   }

using onFailure we can return from the calling function (non-local return) in case of errors.

1. transform{} + recover{}

val htmlPage = JCustomer.parseJson(jsonString)
   .transform { customer ->
      display(customer)
   }.recover { error ->
      display(error)
   }

using transform we can convert the Outcome<Customer> to something else, for example a Outcome<HtmlPage>, then using recover we can convert the error result to the same type and remove the Outcome.

This is my favorite way to handle errors.

There are many other implementations of the Either monad in Kotlin (Result4k, Arrows, Kotlin-Result etc...), rather than importing one of these I created a new one, so you can convert Kondor Outcome to your specific result type. I choose a different name to avoid clashing with Result in the Kotlin library which works differently but it will always be imported first by the IDE. I also don't like map and flatmap be identical to the collections methods because when using a collection of results it becomes very confusing.

With Kondor is easy to solve all Json mappings, including not so trivial ones. For example:

Enums are automatically transformed in strings. For example with these types:

enum class TaxType { Domestic, Exempt, EU, US, Other }

data class Company(val name: String, val taxType: TaxType)

You can create this converter:

object JCompany : JAny<Company>() {

   private val name by str(Company::name)
   private val tax_type by str(Company::taxType)

   override fun JsonNodeObject.deserializeOrThrow() =
      Company(
         name = +name,
         taxType = +tax_type
      )
}

And it will be mapped to this Json format:

{
   "name": "Company Name",
   "tax_type": "Domestic"
}

TODO: better example with FileInfo, show types one by one and add array example JInstance on sealed example default JsonStyle

Sometimes it's convenient to declare some attributes as JsonNode, in this way we can create a structured data dynamically. Another possible case is if we want to store an external Json (for example a response from another service) as a field in our types. For this cases Kondor can serialize/deserialize directly fields of type JsonNodeObject. This can be also used in a flatten field to merge the Json attributes with the typed ones.

For example with this class and this converter:

data class DynamicAttr(
    val id: Int,
    val name: String,
    val attributes: JsonNodeObject
)

object JDynamicAttr : JAny<DynamicAttr>() {
    private val id by num(DynamicAttr::id)
    private val name by str(DynamicAttr::name)
    private val attributes by flatten(DynamicAttr::attributes)
    override fun JsonNodeObject.deserializeOrThrow() = DynamicAttr(
        id = +id,
        name = +name,
        attributes = +attributes
    )
}

We can parse and render a Json like this:

 {
   "id": 123,
   "name": "Ann",
   "aString": "String",
   "aObj": {
     "aNestedString": "NestedString",
     "aNestedNum": 123.4
   }
 }

Where id and name are typed fields of the objects and all the rest are stored in the attributes field as JsonNode.

To store in Json a sealed class, or an interface with a number of known implementations you can use the JSealed base converter.

For example assuming Customer can be either a Person or a Company:

sealed class Customer()
data class Person(val id: Int, val name: String) : Customer()
data class Company(val name: String, val taxType: TaxType) : Customer()

You just need to map each converter to a string and (optionally) specifiy the name of the discriminator field:

object JCustomer : JSealed<Customer> {

    override val discriminatorFieldName = "type"
   
    override val subtypesJObject: Map<String, JObject<out Customer>> =
        mapOf(
            "private" to JPerson,
            "company" to JCompany
        )

   override fun extractTypeName(obj: Customer): String =
      when (obj) {
         is Person -> "private"
         is Company -> "company"
      }
}

It will be mapped in a Json like this:

{
   "type": "private",
   "id": 1,
   "name": "ann"
}

Where "type" here is the discriminator field.

Let's say we have a class FileInfo that maps to this json format:

{
   "name": "filename",
   "date": 0,
   "size": 123,
   "folderPath": "/"
}

Now we need to create a type which is same as FileInfo but with a boolean field added:

data class SelectedFile(val selected: Boolean, val file: FileInfo)

Writing a converter we will get this json format:

{
   "selected": true,
   "file": {
      "name": "filename",
      "date": 0,
      "size": 123,
      "folderPath": "/"
   }
}

But instead we want something simpler:

{
   "selected": true,
   "name": "filename",
   "date": 0,
   "size": 123,
   "folderPath": "/"
}

With Kondor is easy to do this using the flatten format:

object JSelectedFile : JAny<SelectedFile>() {

   val selected by bool(SelectedFile::selected)
   val file_info by flatten(JFileInfo, SelectedFile::file)

   override fun JsonNodeObject.deserializeOrThrow() =
      SelectedFile(
         +selected,
         +file_info
      )

}

Note that it only works with non-nullable fields and it requires that there are no fields with same name on SelectedFile and FileInfo.

Let's say you have a field which is a map and you want to save it as a Json object.

For example a map of things to do, with a short key and a longer description:

data class Notes(val updated: Instant, val thingsToDo: Map<String, String>)

You just need to use the JMap converter and passing it the converter for the value type of the Map (the keys have to be String because of Json syntax):

object JNotes : JAny<Notes>() {
   private val updated by str(Notes::updated, JInstantD)
   private val things_to_do by obj(JMap(JString), Notes::thingsToDo)

   override fun JsonNodeObject.deserializeOrThrow() =
      Notes(
         updated = +updated,
         thingsToDo = +things_to_do
      )
}

The result will be a Json like this:

{
   "updated": "2021-03-26T19:19:20.093501Z",
   "things_to_do": {
      "something": "lorem ipsum",
      "something else": "Lorem ipsum dolor sit amet",
      "another thing to do": "Lorem ipsum dolor sit amet, consectetur adipiscing elit",
      "ditto": "Lorem ipsum dolor sit amet, consectetur adipiscing elit, sed do eiusmod tempor incididun"
   }
}

If you have a custom collection:

class Products : ArrayList<Product>() {
   fun total(): Double = sumOf { it.price ?: 0.0 }

   companion object {
        fun fromIterable(from: Iterable<Product>): Products =
            from.fold(Products()) { acc, p -> acc.apply { add(p) } }
    }
}

You can easily create a converter for it that will render it as a normal array:

object JProducts : JArray<Product, Products>() {
   override val helper = JProduct

   override fun convertToCollection(from: Iterable<Product>) =
      Products.fromIterable(from)

}

And it will be rendered as a standard Json array:

[
   {
      "id": 175,
      "long_description": "Lorem ipsum dolor sit amet, consectetur adipiscing elit, sed do eiusmod tempor incididun",
      "short-desc": "Good Stuff",
      "price": 223.23
   },
   {
      "id": 281,
      "long_description": "Lorem ipsum dolor sit amet, consectetur adipiscing elit",
      "short-desc": "Free Stuff"
   }
]

When it comes to convert an Xml format to json, a common practice is to map the Xml nodes to a Json array of objects with a single field that contains the type of the object as key and the object content as value.

For example an extract of the Liquibase Json format to describe a ChangeSet with two changes, one of type addColumnand the other one of type addLookupTable:

{
   "changeSet": {
      "id": "1",
      "changes": [
         {
            "addColumn": {
               "tableName": "person",
               "columns": [
                  {
                     "column": {
                        "name": "username",
                        "type": "varchar(8)"
                     }
                  }
               ]
            }
         },
         {
            "addLookupTable": {
               "existingTableName": "person",
               "existingColumnName": "state",
               "newTableName": "state",
               "newColumnName": "id",
               "newColumnDataType": "char(2)"
            }
         }
      ]
   }
}

It is quite complicated to parse this format, either using reflection, annotations or generated classes.

Once we have the types to represent that format:

 data class ChangeSet(val id: String, val author: String, val changes: List<Change>) : ChangeLogItem

sealed class Change {
   data class CreateTable(val tableName: String, val columns: List<Column>) : Change()
   data class AddColumn(val tableName: String, val columns: List<Column>) : Change()
   data class AddLookupTable(
      val existingTableName: String,
      val existingColumnName: String,
      val newTableName: String,
      val newColumnName: String,
      val newColumnDataType: String
   ) : Change()
}

With Kondor we can abstract on the protocol, and we can create a specific converter for this format, let's call it NestedPolyConverter.

We can now easily define our converters using the new format that will parse and output the Json example correctly:

object JChangeSet : JAny<ChangeSet>() {
   val id by str(ChangeSet::id)
   val author by str(ChangeSet::author)
   val changes by array(JChange, ChangeSet::changes)

   override fun JsonNodeObject.deserializeOrThrow() =
      ChangeSet("id", "author", emptyList())
}

object JChange : NestedPolyConverter<Change> {

   override fun extractTypeName(obj: Change): String =
      when (obj) {
         is Change.AddColumn -> "addColumn"
         is Change.AddLookupTable -> "addLookupTable"
         is Change.CreateTable -> "createTable"
      }

   override val subConverters = mapOf(
      "addColumn" to JAddColumn,
      "addLookupTable" to JAddLookupTable,
      "createTable" to JCreateTable
   )
}

It's very easy to create new converters to follow your team conventions.

Converters are defined using JsonNode, so you don't have to handle the parsing, and the serializing separately (which can be a source of bugs). They are easier to write than other libraries custom serialisers.

There are some converters in Kondor ready-to-use:

• Sealed classes: automatically converting your sealed classes in polymorphic json
• Maps: converting a Map<String, *> into a Json object
• Instant: both using epoch or date format
• BigDecimal: you can use numbers of any length
• String wrappers: simplify json for IDs and other types that wrap over a string

and so on...

You can choose which fields to serialize or even use functions, and for deserialization you don't have to use the constructor.

TODO: add example of class with private constructor and custom serializer/deserializer

If you need to parse or produce some arbitrary Json that you don't need to map you any of your domain object, you can use Kondor JsonNode DSL

How to "prettify" any json on the fly:

parseJsonNode("my json here").expectSuccess().render(pretty)

TODO: add example of writing Json

TODO: add example of parsing Json

Using Kondor it's easy to integrate with Http4k since they use the same functional approach.

For example you can create a BodyLens directly from a JConverter using an ext function like this:

fun <T : Any> JConverter<T>.toBodyLens(vararg metas: Meta): BiDiBodyLens<T> =
   httpBodyRoot(metas.toList(), APPLICATION_JSON, ContentNegotiation.None)
      .map({ fromJson(it.payload.asString()).orThrow() }, { Body(toJson(it)) })
      .toLens()

It's easy to write generic converters for special types, much easier and safer than defining custom serializer.

Faster than reflection based parsers.

Immutable node objects are convenient if you want to manipulate Json trees.

Easy DSL to parse arbitrary Json or to produce arbitrary Json.

No external dependencies, not even one apart from Kotlin stdlib, and it will stay so.

Doesn't throw any Exception.

I wrote this library to solve a specific problem of our team. It was useful for us, so there could be other people that could find this beneficial.

To describe the problem, let's say you need to map a Json like this:

 {
   "id": "1001",
   "vat-to-pay": true,
   "customer": {
      "id": 1,
      "name": "ann"
   },
   "items": [
      {
         "id": 1001,
         "short_desc": "toothpaste",
         "long_description": "toothpaste \"whiter than white\"",
         "price": 125
      },
      {
         "id": 10001,
         "short_desc": "special offer",
         "long_description": "bla bla"
      }
   ],
   "total": 123.45
}

To your own domain objects:

data class Person(val id: Int, val name: String)

data class Product(val id: Int, val shortDesc: String, val longDesc: String, val price: Double?)

data class InvoiceId(override val raw: String) : StringWrapper

data class Invoice(
   val id: InvoiceId,
   val vat: Boolean,
   val customer: Person,
   val items: List<Product>,
   val total: Double
)

The Json format is quite similar to the domain objects but there are some differences:

• InvoiceId is a custom type wrapping a string, but in the Json there is only the string.
• field names follow a different conventions (snake instead of camel) or totally different (vat-to-pay).
• nullable fields are optional in Json.

We also used the same domain classes inside other Json format, with slightly different field mappings, moreover we have to handle different versions of the Json format.

Another big requirement for us was not having reflection on our domain classes, we all got bad experiences with refactors that broke Json api.

Finally, we prefer to avoid annotating domain classes with serialization details.

The possible solutions we examined were:

• Libraries based on reflection like Jackson or Gson: to meet our requirements we would have to create DTO for all our types with fields heavily annotated.

• KotlinSerializer: even if it's based on compile-time reflection, it has the same problems of the other libraries based on reflection: any changes on the domain classes would be reflected in a Json format change. Moreover KotlinSerializer is also sensible to changes to class packages for sealed classes and it doesn't support Java classes without custom serializers.

So we did several progressive improvements over the idea of defining bidirectional converter explicitly using a simple DSL.

Technically each converter is a "Profunctor", which is a special kind of bifunctor where one of the functors is covariant and the other is contravariant. So instead of trying to explain to the Json mapper how to serialize/deserialize our class using annotations we define the converter (technically a profunctor) for each class. Thanks to Kotlin DSL capabilities, it doesn't require much code.

And after a few iterations, this is how Kondor-Json was born.

• Jackson
• KotlinSerializer
• Moshi

TODO: examples of annotations vs Kondor converters

TODO: comparison of handling errors

TODO: comparison of performance

• A DSL for Java

• Generating random values from the converters

• Add a converter that use Jackson to simplify the migration/adoption

• Add integration with Snodge for fuzzy testing

I've got the inspiration for Kondor while studying Adjoint functors. But further studies showed me that Kondor two functors don't form an adjunction but instead they form a profunctor.

Profunctor and Adjunctions are a fascinating part of Category Theory, you can find some more materials about them here:

https://en.wikipedia.org/wiki/Profunctor

https://typeclasses.com/profunctors