Explicit polymorphism discriminator

[Tmpod]

What is your use-case and why do you need this feature?

I've been experimenting with MongoDB through KMongo and KtMongo, specifically around polymorphic documents. Both kbson (used by KMongo) and the official Kotlin Mongo driver (used by KtMongo) support polymorphic serialization, but always through a virtual field (whose name is dependent on the format).

This isn't ideal. You have to resort to clunky workarounds to use the type-safe query API, you have use SerialName everywhere and in general it isn't intuitive at all. Implicit fields make code harder to reason about, and often require diving into library sources to really understand what's happening.

Here's an example with current implicit fields

enum class FooType { A, B }

@Serializable
sealed interface Foo {
    val common: Int
    val type: FooType
}

@Serializable
@SerialName("A")
class FooA(
    override val common: Int,
    val bar: Long,
) : Foo {
    // Can't be concrete otherwise you get weird behaviour between
    // the type field generated by kx.ser (with value from @SerialName)
    // and the actual concrete field
    override val type get() = FooType.A
}

@Serializable
@SerialName("B")
class FooB(
    override val common: Int,
    val baz: String,
) : Foo {
    override val type get() = FooType.B
}

suspend fun someQueries(db: Database) {
    db.getCollection<FooA>().insertOne(FooA(common = 1, bar = 2))
    db.getCollection<FooB>().insertOne(FooB(common = 1, baz = "3"))
    
    // Only works if the format-specific discriminator is called "type".
    // If it is something else you either have to change the property name
    // or use @SerialName
    val fooA = db.getCollection<FooA>().findOne { Foo::type eq FooType.A }
    val fooB = db.getCollection<FooB>().findOne { Foo::type eq FooType.B }
    val foo = db.getCollection<Foo>().findOne()
    println(fooA.bar)
    println(fooB.baz)
    println(foo.type)
}

Describe the solution you'd like

Let's consider a new @PolymorphicDiscriminator annotation that tells kx.ser and implementing formats which field should be used for polymorphic (de)serialization. One could write the following instead:

The same example but using this suggested annotation

enum class FooType { A, B }

@Serializable
sealed class Foo(
    @PolymorphicDiscriminator
    // Could also use @SerialName here if wanted
    val type: FooType,
    val common: Int,
)

@Serializable
class FooA(
    common: Int,
    val bar: Long,
) : Foo(FooType.A, common)
// EDIT: I'm aware these constructors are invalid with @Serializable,
// you'd have to override instead, like in the previous example.

@Serializable
class FooB(
    common: Int,
    val baz: String,
) : Foo(FooType.B, common)

suspend fun someQueries(db: Database) {
    db.getCollection<FooA>().insertOne(FooA(common = 1, bar = 2))
    db.getCollection<FooB>().insertOne(FooB(common = 1, baz = "3"))

    // We are now sure the type-safe API will resolve to the correct discriminator name,
    // because that's defined by the user, not the format developer
    val fooA = db.getCollection<FooA>().findOne { Foo::type eq FooType.A }
    val fooB = db.getCollection<FooB>().findOne { Foo::type eq FooType.B }
    val foo = db.getCollection<Foo>().findOne()
    println(fooA.bar)
    println(fooB.baz)
    println(foo.type)
}

I'm not well versed in compiler plugins or annotation processors, so I'm not entirely sure on the implementation details, but all the semantic information required for such an annotation exist. When you're processing a class that is part of a sealed hierarchy, you check if the value it defines is always known at compile-time and if it is unique (e.g. you can't have two implementors use the same discriminator value). I'd even say the value could be anything directly matchable in when expressions.
If the values are all consistent, you can generate the polymorphic deserialization function which selects the correct implementor serializer based on the decoded discriminator, akin to what already happens on the current implementation based on an implicit field.

---

Please correct me if anything I've said is wrong or if my suggestion is unattainable at the moment. Thank you!

[CLOVIS-AI]

Hey, KtMongo author here. I'll give a bit more context.

KtMongo relies on KotlinX.Serialization to convert Kotlin types to and from MongoDB BSON. Through this, we support sealed class and sealed interface simply because KotlinX.Serialization supports them, using an imaginary field type which contains a discriminant.

KtMongo provides an extensive DSL over MongoDB operators. This DSL is implemented through extension functions and generics, allowing us to refer to fields when writing requests:

@Serializable
class User(
    val name: String,
    val age: Int,
)

users.find {
    User::age gt 18
}

So far, so good. The problem comes because of the imaginary field KotlinX.Serialization creates on sealed types:

@Serializable
sealed class Foo

@Serializable
class FooA(val a: Int) : Foo()

@Serializable
class FooB(val b: String) : Foo()

A instance of FooB("foo") will be serialized as:

{
    "type": "FooB",
    "b": "foo"
}

We would like to be able to write a query for specifically instances of FooB, but this requires referring to the imaginary "field" type:

users.find {
    User::foo / Foo::type eq "foo"    // ⚠ Foo::type does not exist
}

This could be worked around by creating an extension function and hard-coding the name type, but KotlinX.Serialization will actually name the discriminant something else if there is already a field named type, so there is no good typesafe way of knowing how the field is called.

By using explicit enums, this issue removes the magic around the discriminant, so it can be accessed through regular Kotlin code.

[pdvrieze]

@Tmpod The information you are looking for is where @SerialName is applied to the type of the serialized object (defaulting to the FQN of the type). You can give it any desired value (but it should be unique within the context of a single format instance).

@CLOVIS-AI You can get this information as text from the descriptor: serializer<Foo>.descriptor.serialName

[Tmpod]

@CLOVIS-AI Thank you for adding more context!

This could be worked around by creating an extension function and hard-coding the name type, but KotlinX.Serialization will actually name the discriminant something else if there is already a field named type, so there is no good typesafe way of knowing how the field is called.

What I'm proposing would avoid all that hardcoding. By explicitly marking a field as the discriminant, you remove the unintuitive behaviour (imo) of virtual fields. The serialization library knows which field it should encode/decode first and how to map values to serializers. I just don't know how feasible it is to get the discriminant value from each implementor in the context of an annotation processor or compiler plugin. To me, it should be possible, the value is static, but I really am not familiar with the capabilities exposed to processors/plugins (though I've already noticed limitations, such as annotations being incapable of taking values I'd consider compile-time constants, such as enum variants).

The information you are looking for is where @SerialName is applied to the type of the serialized object (defaulting to the FQN of the type). You can give it any desired value (but it should be unique within the context of a single format instance).

I'm not sure I follow. I used @SerialName in my first example, but the point is that, because it is useful to access the type of documents both for querying Mongo and other situations (without having to resort to when matches), kx.ser should allow explicit marking of discriminants in closed polymorphic hierarchies. It would also have the added benefit of making the code clearer to developers, since there's less "magic" involved — you know exactly which field is used to decide how to deserialize.

[sandwwraith]

I guess what you want here is #1664, so it's a duplicate.

I also would like to point out that kotlinx.serialization by default always performs polymorphic serialization in array format: [type, object]. Insertion of type keyword is a Json-specific code path.

[pdvrieze]

@CLOVIS-AI Thank you for adding more context!
What I'm proposing would avoid all that hardcoding. By explicitly marking a field as the discriminant, you remove the unintuitive behaviour (imo) of virtual fields.
This is inconsistent with the design of the library (and thus would be very hard to introduce). The way the type is exposed is in many ways format specific, and doesn't need to take the shape of a synthetic attribute. In all cases it must be constant, so having it as an attribute is poor design. Note that the structure exposed to formats is (in pseudo-json): "myPolyValue"= [ "ConcreteTypeSerialName", { /* some struct value */ }]. Many formats (including json) collapse this into a single struct (or for json as an array), but that is a format specific transformation.

The serialization library knows which field it should encode/decode first and how to map values to serializers. I just don't know how feasible it is to get the discriminant value from each implementor in the context of an annotation processor or compiler plugin. To me, it should be possible, the value is static, but I really am not familiar with the capabilities exposed to processors/plugins (though I've already noticed limitations, such as annotations being incapable of taking values I'd consider compile-time constants, such as enum variants).

I'm not sure I follow. I used @SerialName in my first example, but the point is that, because it is useful to access the type of documents both for querying Mongo and other situations (without having to resort to when matches), kx.ser should allow explicit marking of discriminants in closed polymorphic hierarchies. It would also have the added benefit of making the code clearer to developers, since there's less "magic" involved — you know exactly which field is used to decide how to deserialize.

Do you mean you want to read the value of the type without using runtime type information (using the fact that the instance is of the specific type). As to what "field" is used to decide how to deserialize, the name is always the same (except when explicitly overridden). In the specific case of Json this can be provided either globally in the format configuration, or per type (not attribute) using @JsonClassDiscriminator, this annotation is then inherited.

As to "not having to resort to when matches" you can use a calculated property:

@Serializable
abstract class Base {
    abstract val type: String
}

@Serializable
class Child(val myProp:String) {
    override val type: String get() = Child.serializer().descriptor.serialName // or a hardcoded/cached version
}

[Tmpod]

I guess what you want here is #1664, so it's a duplicate.

I suppose it is related, yes, but what I was suggesting was more versatile, as I see it. The referenced issue is also in part about nested hierarchies and

I also would like to point out that kotlinx.serialization by default always performs polymorphic serialization in array format: [type, object]. Insertion of type keyword is a Json-specific code path.

I had read about that, but I'm not grasping the reasoning behind why kx.ser doesn't support field-based discriminants at the core library level. Why does it need to be a format-specific thing? Either way, an annotation that could tell the format implementors which field should be used for discriminating would be nice, instead of having each format re-implement the same annotations (kbson has a similar BsonClassDiscriminator) annotation.

[CLOVIS-AI]

As to "not having to resort to when matches" you can use a calculated property:

@serializable
abstract class Base {
abstract val type: String
}

@serializable
class Child(val myProp:String) {
override val type: String get() = Child.serializer().descriptor.serialName // or a hardcoded/cached version
}

Can you clarify what the value will return? "type" or "Child"? KtMongo needs to know both.

Also, that is a bit verbose to add to each class. Could there be a way to automatically generate this somehow?

[Tmpod]

Do you mean you want to read the value of the type without using runtime type information (using the fact that the instance is of the specific type).

Yeah. Since that field exists in the serialized document, why not have it in the deserialized object too? Makes some things easier, and in the case of KMongo/KtMongo's and similar DSLs, it's important to have it.

As to what "field" is used to decide how to deserialize, the name is always the same (except when explicitly overridden). In the specific case of Json this can be provided either globally in the format configuration, or per type (not attribute) using @JsonClassDiscriminator, this annotation is then inherited.

I am aware of that annotation, however it's format specific (and string-based, but that's a minor nitpick and a limitation of annotations). I think an annotation on the field itself would be clearer to developers — this field will be used to resolve polymorphic serializers; no virtual fields at play, no using @SerialName in classes (unintuitive, at least to me). It would also work across formats, even if they can only implement array-based polymorphism.

Again, not entirely sure if this is truly implementable, but perhaps this could work: generate and expose a getter for the discriminant field (its SerialDescriptor, maybe?), that the implementing formats and serializers could use to know which field to encode first and how to match the first decoded value to select the deserializer. It sounds to me like it could even be implemented a separate thing, though it would be more useful to have that built into the core library itself.

If this sounds plausible, I could try implement it to help illustrate what I mean by all this 😅

[pdvrieze]

As to "not having to resort to when matches" you can use a calculated property:

[@serializable](https://github.com/serializable)
abstract class Base {
abstract val type: String
}
[@serializable](https://github.com/serializable)
class Child(val myProp:String) {
override val type: String get() = Child.serializer().descriptor.serialName // or a hardcoded/cached version
}

Can you clarify what the value will return? "type" or "Child"? KtMongo needs to know both.

This returns com.example.Child (the FQN of Child or the value of the @SerialName annotation on the class.

Also, that is a bit verbose to add to each class. Could there be a way to automatically generate this somehow?

You can always use kapt

I had a bit more of a look. Basically you want to do typesafe queries. That makes sense. However, rather than:

    User::foo / Foo::type eq "foo"

Why not have something like:

    User::foo / isInstance<Foo>()

Given that KtMongo wants to be typesafe it should use these type annotations transparently and have a primacy of types. Of course it is also valid to also allow a string parameter.

As a bit of format based perspective, if you look at PolymorphicSerializer it has a descriptor with two values (and could be serialized as array or struct). It also has a kind that marks is as polymorphic. The names of the attributes of the polymorphic serializer are type(String) and value(CONTEXTUAL). But nothing stops the format to treat it specially (this is indeed why there can be a virtual attribute used by Json, or why XML has different modes (as struct, using the tagname to distinguish, or using xsi:type).

Given the application area, it may actually make a lot of sense to have a custom format for MongoKt. Using such it would be possible to do "shape matching", work directly in Bson, and have your type attributes (formats on decoding can provide virtual attributes that don't exist in the serial form, as long as a valid value can be produced). Note that formats are ultimately in control over what happens

[CLOVIS-AI]

Why not have something like:

User::foo.isInstance<Foo>()

In theory, that would be the best option, yes. But it cannot be implemented due to KotlinX.Serialization being magic:

• We can't know whether a type is polymorphically serializable using generics, because they don't have a specific supertype for that. If a user used this operator on a non-polymorphically serializable type, this filter operator would always evaluate to false, and thus this query would never return anything, without any errors or warnings. This would be a very easy footgun and particularly hard to debug.
• This operator actually cannot be implemented, because it would need to know what the name of the magically generated field is, and KotlinX.Serialization names it differently based on the case. I have at least seen it named type and __type. Worst, the name can change when other fields are added, which breaks database compatibility.

The third blocking issue was finding out the class's serial name, but your previous comment solves that one.

Given the application area, it may actually make a lot of sense to have a custom format for MongoKt.

There is one, but it's maintained by MongoDB Inc and I cannot easily extend it. Aside the governance issue, changing the format now would break all existing production deployments, which is not acceptable :/

[pdvrieze]

• We can't know whether a type is polymorphically serializable using generics, because they don't have a specific supertype for that. If a user used this operator on a non-polymorphically serializable type, this filter operator would always evaluate to false, and thus this query would never return anything, without any errors or warnings. This would be a very easy footgun and particularly hard to debug.

Actually, if you have the type of User::foo which is the base type. You can then get the serializer for that. It has a serial descriptor with SerialKind. If it is PolymorphicKind then it is polymorphic, otherwise not. For sealed polymorphic values you can determine all valid children from the base serializer, for a open polymorphic value you need to get the same information from the serializersModule (you can get the list of all registered serializers for a base class by dumping the module - I know a good api would be nice)

If you want to know the name of the type parameter, this is actually part of the serial descriptor of the polymorphic serializer (with 2 parts: type, name) and may carry annotations as well. For element annotations you need to use the descriptor of the container object (for example if that overrides the default name in the case of Json). You can also use that annotation to get the static element type (base class).

A key challenge will be translating your type to the expected value of the type annotation. We can know whether it is polymorphic. We can extract the type tree from the serializer/serial descriptor of the container, the defaults can be retrieved from the format (or your own configuration that creates the format) and thus retrieve both annotations and serial names. Note that the json format doesn't actually fallback to alternative type discriminator names.

What I suspect is that you have existing data where the discriminator is somewhat unpredictable, and potentially requiring different configurations for the format as well. You will have to somehow collect that information, be it in code, or annotations. This can then be picked up my the type safe accessors.

[CLOVIS-AI]

Sorry, this explanation is quite abstract and I'm not sure I'm following everything. To make it a bit more concrete, how would you implement the following?

@Serializable
class User(
    val name: String,
    val role: UserRole,
)

@Serializable
sealed class UserRole {
    @Serializable
    @SerialName("Guest")
    data object Guest : UserRole()

    @Serializable
    @SerialName("Employee")
    data class Employee(val id: String) : UserRole()
}

data class TypeInfo(
    val discriminatorFieldName: String,
    val discriminatorValue: String,
)

inline fun <reified T, reified F, reified F2 : F> KProperty1<T, F>.typeInfo(): TypeInfo {
    TODO("what should be here?")
}

// Expected usage:
User::role.typeInfo<_, _, UserRole.Employee>() // TypeInfo(…fieldName: "type", …value: "Employee")