ADT (Algebraic Data Types)

ADTs encoding and decoding

The most straightforward way to encode / decode ADTs is by using generic derivation for the case classes but explicitly defined instances for the ADT type.

Consider the following ADT:

sealed trait Event

case class Foo(i: Int) extends Event
case class Bar(s: String) extends Event
case class Baz(c: Char) extends Event
case class Qux(values: List[String]) extends Event

And the encoder / decoder instances:

import cats.syntax.functor._
import io.circe.{ Decoder, Encoder }, io.circe.generic.auto._
import io.circe.syntax._

object GenericDerivation {
  implicit val encodeEvent: Encoder[Event] = Encoder.instance {
    case foo: Foo => foo.asJson
    case bar: Bar => bar.asJson
    case baz: Baz => baz.asJson
    case qux: Qux => qux.asJson
  }

  implicit val decodeEvent: Decoder[Event] =
    List[Decoder[Event]](
      Decoder[Foo].widen,
      Decoder[Bar].widen,
      Decoder[Baz].widen,
      Decoder[Qux].widen
    ).reduceLeft(_ or _)
}

Note that we have to call widen (which is provided by Cats's Functor syntax, which we bring into scope with the first import) on the decoders because the Decoder type class is not covariant. The invariance of circe's type classes is a matter of some controversy (Argonaut for example has gone from invariant to covariant and back), but it has enough benefits that it's unlikely to change, which means we need workarounds like this occasionally.

It's also worth noting that our explicit Encoder and Decoder instances will take precedence over the generically-derived instances we would otherwise get from the io.circe.generic.auto._ import (see slides from Travis Brown's talk here for some discussion of how this prioritization works).

We can use these instances like this:

import GenericDerivation._
import io.circe.parser.decode

decode[Event]("""{ "i": 1000 }""")
// res0: Either[io.circe.Error, Event] = Right(value = Foo(i = 1000))

(Foo(100): Event).asJson.noSpaces
// res1: String = "{\"i\":100}"

This works, and if you need to be able to specify the order that the ADT constructors are tried, it's currently the best solution. Having to enumerate the constructors like this is obviously not ideal, though, even if we get the case class instances for free.

A more generic solution

We can avoid the fuss of writing out all the cases by using the circe-shapes module:

// To suppress previously imported implicit codecs.
import GenericDerivation.{ decodeEvent => _, encodeEvent => _ }

object ShapesDerivation {
  import io.circe.shapes
  import shapeless.{ Coproduct, Generic }

  implicit def encodeAdtNoDiscr[A, Repr <: Coproduct](implicit
    gen: Generic.Aux[A, Repr],
    encodeRepr: Encoder[Repr]
  ): Encoder[A] = encodeRepr.contramap(gen.to)

  implicit def decodeAdtNoDiscr[A, Repr <: Coproduct](implicit
    gen: Generic.Aux[A, Repr],
    decodeRepr: Decoder[Repr]
  ): Decoder[A] = decodeRepr.map(gen.from)

}

And then:

import ShapesDerivation._
import io.circe.parser.decode, io.circe.syntax._

decode[Event]("""{ "i": 1000 }""")
// res2: Either[io.circe.Error, Event] = Right(value = Foo(i = 1000))

(Foo(100): Event).asJson.noSpaces
// res3: String = "{\"i\":100}"

This will work for any ADT anywhere that encodeAdtNoDiscr and decodeAdtNoDiscr are in scope. If we wanted it to be more limited, we could replace the generic A with our ADT types in those definitions, or we could make the definitions non-implicit and define implicit instances explicitly for the ADTs we want encoded this way.

The main drawback of this approach (apart from the extra circe-shapes dependency) is that the constructors will be tried in alphabetical order, which may not be what we want if we have ambiguous case classes (where the member names and types are the same).

The future

The generic-extras module provides a little more configurability in this respect. We can write the following, for example:

import io.circe.generic.extras.auto._
import io.circe.generic.extras.Configuration

implicit val genDevConfig: Configuration =
  Configuration.default.withDiscriminator("what_am_i")

And then:

import io.circe.parser.decode, io.circe.syntax._

(Foo(100): Event).asJson.noSpaces
// res4: String = "{\"i\":100}"

decode[Event]("""{ "i": 1000, "what_am_i": "Foo" }""")
// res5: Either[io.circe.Error, Event] = Right(value = Foo(i = 1000))

Instead of a wrapper object in the JSON we have an extra field that indicates the constructor. This isn't the default behavior since it has some weird corner cases (e.g. if one of our case classes had a member named what_am_i), but in many cases it's reasonable and it's been supported in generic-extras since that module was introduced.

Notes

This still doesn't get us exactly what we want, but it's closer than the default behavior. It's also been considered to change withDiscriminator to take an Option[String] instead of a String, with None indicating that we don't want an extra field indicating the constructor, giving us the same behavior as our circe-shapes instances in the previous section, but haven't been implemented so far.