Hackle's blog
between the abstractions we want and the abstractions we get.
Hackle's blog
between the abstractions we want and the abstractions we get.
In this post we look at a dialect of union type in Kotlin's sealed interface (or class). In particular, how it differs from other prominent variants (for this purpose, Haskell and TypeScript) in ergonomics.
The ergonomics of union types, or modern type systems has a bias towards correctness, which can sometimes be at odds with preference of aesthetics, accepted convention or mindset. For example, "smart" code that degrades type-checking to runtime behaviour, or results in regression of type safety should be avoided; seemingly "dumb" code with the blessing of the type system is more preferable. (An attractive deal sweetener is well-type code is easier for the modern language servers or IDEs to analyse, often seen through high-quality auto-completion). We'll see how such ergonomics pan out.
A union type is defined as a sealed interface (or sealed class but I will use interface hereafter) that is implemented by its member types (there are variations in where the member types can be placed - please see Kotlin docs). Consider the definition of Maybe<T>,
sealed interface Maybe<T>
data class Just<T>(val value: T) : Maybe<T>
data object None : Maybe<Nothing>
This looks not much more than a verbose version of the intuitive definition in Haskell,
data Maybe a = Nothing | Just a
(To disambiguate, Nothing in Haskell is a data constructor, but Kotlin's Nothing is the built-in "bottom" type.)
However, there is a notable difference: Just in Haskell is one of two data constructors of Maybe a (which itself is called a "type constructor"), but it is not a type; while Just<T> in Kotlin is a type of its own. Therefore, it's not possible to type a value as Just Int in Haskell, but completely legitimate in Kotlin as below.
fun showIncrement(mustBeJust: Just<Int>) {
println(mustBeJust.value + 1)
}
val just2: Just<Int> = Just(2)
showIncrement(just2)
This valuable feature gives Kotlin's union types an edge of expressiveness. (Or if you are cynical, yet another chance for confusion and abuse. Being cynical is no fun.)
This type-over-type hierarchy may tempt a Haskeller to comparing it to type families, such as type promotion of Just / Nothing via DataKinds; or, to question if Kotlin supports type-level programming. Unfortunately, the short answer is no - at least not along the lines of DataKinds, or that of TypeScript to effortlessly promote values to types for a flavour of dependent types. The longer answer, however, is one with slight complication.
Nothing in Kotlin's Maybe<T> is a singleton type - a type with just one value. It allows the seemingly strange expression as below,
val naught: None = None
The telling feature of a singleton type is, "if you know the type, you know the value". This innocent-looking property can be used in combination with a union type hierarchy to enforce invariants.
sealed interface SBoolean
data object STrue : SBoolean
data object SFalse : SBoolean
To the unsuspecting eyes, SBoolean ranges over two single values, it is therefore nothing more than a superfluous alternative to a plain enum. That would be a limiting view, because STrue and SFalse are not just values, but also types, and can be used individually, such as follows,
fun runRiskyOperation(mustHaveAcknowledged: STrue) {
println("By this point, you must have acknowledged.")
}
fun runAtRisk(acknowledgedOrNot: SBoolean) {
when (acknowledgedOrNot) {
STrue -> runRiskyOperation(STrue)
SFalse -> println("Cannot run operation without acknowledgement.")
}
}
In the above example, the risky operation must only be executed with acknowledgement. This is enforced by the use of singleton type STrue. The caller has no other options except passing in the only value STrue; SFalse is a type error, outright rejected by the compiler.
Had we used type mustHaveAcknowledged: Boolean, the caller could rightfully pass in false, an illegal state that requires runtime validation to avoid.
Kotlin's syntax also clearly indicates the unification of types and values - instead of is STrue for type inspection, the value STrue is pattern-matched against directly.
The same type hierarchy can also be expressed with singleton enums - you guessed it, a singleton enum is an enum with just one value.
sealed interface SEBoolean { val value: Boolean }
enum class SETrue(override val value: Boolean) : SEBoolean { TRUE(true) }
enum class SEFalse(override val value: Boolean) : SEBoolean { FALSE(false) }
The enum variation has an advantage over the singleton objects: it allows encoding extra information in the singleton values, in this case, mapping enum constants to the corresponding booleans. One use case is to use the encoded boolean value as runtime value, such as for serialisation and deserialisation (consider @JsonValue / @JsonCreator with the Jackson library). Data objects typically require some level of arm-twisting to get deserialised.
From singleton types one would be tempted to encode the likes of Peano number as well as its operations including addition, subtraction etc, but that would be pushing it too far! As seen below, it's trivial to express decrement, there is no way to express addition of two types.
sealed interface Nat
data object Zero : Nat
data class Succ<T : Nat>(val pred: T) : Nat
// easy - it's just destructuring
fun <T : Nat> decrement(n: Succ<T>): T = n.pred
// not possible
// fun <T1 : Nat, T2: Nat> add(n1: T1, n2: T2): Add<T1, T2>
The real power of union types does not lie in modelling a hierarchy of types with different shapes - the good-old non-sealed interface is more than enough for that job; instead, it's in modelling a closed type hierarchy that is known at compile time, and what follows the closedness: the ability to analyse such hierarchy exhaustively.
Consider the use of SBoolean in the slightly revised runAtRisk,
fun runAtRisk(acknowledgedOrNot: SBoolean) {
when (acknowledgedOrNot) {
STrue -> runRiskyOperation(STrue)
else -> println("Cannot run operation without acknowledgement.")
}
}
The use of else (equivalent to wildcard _ in other languages) may seem harmless, but it is a terrible idea, because it throws away the benefit of exhaustive pattern matching. Would we add a new member type SUndecided alongside STrue and SFalse, the previous implementation of runAtRisk fails to compile, rightfully and expectedly, for not handling the new scenario, whereas this version happily ignores the new addition, resulting in undesirable behaviour - treating "SUndecided" as "SFalse". It is therefore recommended to forbid the use of else with union types. With Kotlin, this is usually done with the Detekt rule ElseCaseInsteadOfExhaustiveWhen.
In practice, it's often the case a new adapter feels peeved by the size as well as the boring nature of the when expression, and tries to improve or "abstract" it, such as by introducing a map; or faced with a union type with a few more member types, attempts to find shortcuts to make the code shorter and "cleaner". Watch out - this is usually a slippery slope. Such smart workarounds usually undermines the guarantee provided by union types, leading to regression in correctness.
Yet, for the extra flexibility afforded by the union of member types (versus just different constructors as in Haskell), at times we want to differentiate a subset of the member types. Consider the example of PayResult<T>,
sealed interface PayResult<out T : Payment> {
data class Success<P : Payment>(val value: P) : PayResult<P>
data class NotEnoughBalance(
val minimum: Long
) : PayResult<Nothing> {
val errorMessage = "Not enough balance."
}
data class Unauthenticated(
val serverErrorCode: ErrorCode
) : PayResult<Nothing> {
val errorMessage = "Authentication failed."
}
}
(Note that the placement of errorMessage inside the data classes is intentional: it should not be exposed through the constructor as a free-form String, but closed off and locked down, to prevent nonsense input such as "divided by zero").
When handling an instance of PayResult, there would be need to extract any potential error message for display. One may implement extractError as following,
fun extractError(payResult: PayResult<Cash>): String? =
when (payResult) {
is PayResult.Success -> null
is PayResult.NotEnoughBalance -> payResult.errorMessage
is PayResult.Unauthenticated -> payResult.errorMessage
}
Which is not terrible, and would be expected of equivalent code in Haskell or Rust. However, if we are picky, there are 2 reasons to be critical,
payResult.errorMessage. Unlike TypeScript, Kotlin is mostly not structured typed, and does not try to figure out the presence of the common field;extractError from PayResult.Success - there is no "error" to speak of! This forces the use of String? as the return type. While one may guess that Success will result in null, from the types along, there is no stopping null for NotEnoughBalance too. In other words, the return type is polluted; the caller's hands are also forced to have to check for null.It's start with the 2nd point. Ideally, as its name states, extractError should only accept member types that encode errors. But that poses the challenge of separating Success from the rest of the "error" member types.
One may take hint from the T in PayResult<T>, and try to unify NotEnoughBalance and Unauthenticated as PayResult<Nothing>, but this only gets us half-way as follows,
fun extractErrorSafe(payResult: PayResult<Nothing>): String =
when (payResult) {
is PayResult.NotEnoughBalance -> payResult.errorMessage
is PayResult.Unauthenticated -> payResult.errorMessage
is PayResult.Success -> "Impossible"
}
Again, unlike TypeScript, Kotlin does not go the extra mile to filter out Success for PayResult<Nothing>, so although extractErrorSafe cannot be called with Success, its implementation must still check the Success case, resulting in an "impossible" branch.
Here comes the advantage of having distinct member types, not just constructors or values. We can do what's normally done to unify distinct classes: use a common interface. See HasErrorMessage as below,
interface HasErrorMessage {
val errorMessage: String
}
sealed interface PayResult<out T : Payment> {
data class Success<P : Payment>(val value: P) : PayResult<P>
data class NotEnoughBalance(
val minimum: Double
) : PayResult<Nothing>, HasErrorMessage {
override val errorMessage = "Not enough balance."
}
data class Unauthenticated(
val serverErrorCode: ErrorCode
) : PayResult<Nothing>, HasErrorMessage {
override val errorMessage = "Authentication failed."
}
}
fun extractErrorSafe(withErrorMessage: HasErrorMessage): String =
withErrorMessage.errorMessage
HasErrorMessage makes extractErrorSafe superfluous - it simply calls through to withErrorMessage.errorMessage; on the other hand, we can see how the two "error" types are unified as below, removing the duplication.
fun makeDisplayMessage(payResult: PayResult<Cash>) =
when(payResult) {
is PayResult.Success -> "All paid!"
is PayResult.NotEnoughBalance,
is PayResult.Unauthenticated,
-> payResult.errorMessage
}
One may be tempted to go even further with yet another trick - to make HasErrorMessage also sealed, so the above when block can be "simplified" as below,
fun makeDisplayMessage(payResult: PayResult<Cash>) =
when(payResult) {
is PayResult.Success -> "All paid!"
is HasErrorMessage -> payResult.errorMessage
}
However, the sharp reader would immediately find this to be a bad idea - because this "simplification" is little better than using else - if a new "error" type is added to PayResult, this version of makeDisplayMessage will happily ignore the new addition, and throw away the strong guarantee of exhaustive pattern-matching.
Unlike a normal interface, a sealed interface is closed at the time of definition, strictly speaking, together with all its member types, which are all known at compile-time. This property of closedness applies to any union types, and is what makes exhaustive analysis possible.
However, without the constraint of "tagging", or restrictive nominal typing as in Haskell and Rust, untagged unions as in TypeScript of Python can be extended to make a new union type - which are separate from the original, but with additional member types.
Surprisingly, extending a union type is also possible with Kotlin's sealed types through a form of inheritance, thanks to the use of interfaces and sub-typing.
Let's suppose we want to extend PayResult with a new member type SuspectedScam, to make a new union type CreditCardPayResult, which has dedicated use in only part of the application, while keeping the existing use of the original PayResult undisturbed. How could this be done?
A naive answer is to make CreditCardPayResult extend PayResult directly, as follows,
sealed interface CreditCardPayResult<out T : Payment> : PayResult<T> {
data class SuspectedScam(
val probability: Float
) : CreditCardPayResult<Nothing>, HasErrorMessage {
override val errorMessage = "Watch out! Likely a scam!"
}
}
// error: 'when' expression must be exhaustive.
// Add the 'is SuspectedScam' branch or an 'else' branch.
fun makeDisplayMessage(payResult: PayResult<Cash>) =
when(payResult) {
is PayResult.Success -> "All paid!"
is PayResult.NotEnoughBalance,
is PayResult.Unauthenticated,
-> payResult.errorMessage
}
Unexpectedly, this upends any existing use of PayResult. Why? Because Kotlin scans for any sub-type of PayResult to discover the closed hierarchy, a.k.a. the union; by extending PayResult, CreditCardPayResult and in turn SuspectedScam, is considered a member type of the union!
The satisfactory solution plays along with how union types are designed in Kotlin, and is a bit of a brain-twister. The sub-type relationship must be reversed, viz. by making PayResult extend CreditCardPayResult, as follows,
sealed interface PayResult<out T : Payment> : CreditCardPayResult<T> {
// left unchanged
}
sealed interface CreditCardPayResult<out T> {
data class SuspectedScam(
val probability: Float
) : CreditCardPayResult<Nothing>, HasErrorMessage {
override val errorMessage = "Watch out! Likely a scam!"
}
}
A bit strange isn't it? But this setup achieves exactly what we set out to do, as is seen below,
fun makeDisplayMessage(payResult: PayResult<Cash>) =
when(payResult) {
is PayResult.Success -> "All paid!"
is PayResult.NotEnoughBalance,
is PayResult.Unauthenticated,
-> payResult.errorMessage
}
fun makeDisplayMessage(payResult: CreditCardPayResult<Cash>) =
when(payResult) {
is PayResult.Success -> "All paid!"
is PayResult.NotEnoughBalance,
is PayResult.Unauthenticated,
is CreditCardPayResult.SuspectedScam,
-> payResult.errorMessage
}
Why does this work? It is all sub-typing. The original hierarchy of PayResult is kept undisturbed, so its number of sub-types is still exactly three. On the other hand, CreditCardPayResult has four sub-types, SuspectedScam that implements it directly, and indirectly, the 3 member types of PayResult.
In other words, extending or implementing a type results in the creation of a sub-type, which may have "more" behaviour, but also fewer values or member types, as is the case with PayResult. On the other hand, the super-type CreditCardPayResult ranges "larger" and has more values and member types. From this perspective, there is nothing surprising at all.