Validated
Imagine you are filling out a web form to signup for an account. You input your username and password and submit. Response comes back saying your username can't have dashes in it, so you make some changes and resubmit. Can't have special characters either. Change, resubmit. Passwords need to have at least one capital letter. Change, resubmit. Password needs to have at least one number.
Or perhaps you're reading from a configuration file. One could imagine the configuration library you're using returns
a scala.util.Try, or maybe a scala.util.Either (or cats.data.Xor). Your parsing may look something like:
for {
url <- config[String]("url")
port <- config[Int]("port")
} yield ConnectionParams(url, port)
You run your program and it says key "url" not found, turns out the key was "endpoint". So you change your code and re-run. Now it says the "port" key was not a well-formed integer.
It would be nice to have all of these errors be reported simultaneously. That the username can't have dashes can be validated separately from it not having special characters, as well as from the password needing to have certain requirements. A misspelled (or missing) field in a config can be validated separately from another field not being well-formed.
Enter Validated.
Parallel validation
Our goal is to report any and all errors across independent bits of data. For instance, when we ask for several pieces of configuration, each configuration field can be validated separately from one another. How then do we enforce that the data we are working with is independent? We ask for both of them up front.
As our running example, we will look at config parsing. Our config will be represented by a
Map[String, String]. Parsing will be handled by a Read type class - we provide instances
just for String and Int for brevity.
trait Read[A] {
def read(s: String): Option[A]
}
object Read {
def apply[A](implicit A: Read[A]): Read[A] = A
implicit val stringRead: Read[String] =
new Read[String] { def read(s: String): Option[String] = Some(s) }
implicit val intRead: Read[Int] =
new Read[Int] {
def read(s: String): Option[Int] =
if (s.matches("-?[0-9]+")) Some(s.toInt)
else None
}
}
Then we enumerate our errors - when asking for a config value, one of two things can go wrong: the field is missing, or it is not well-formed with regards to the expected type.
sealed abstract class ConfigError
final case class MissingConfig(field: String) extends ConfigError
final case class ParseError(field: String) extends ConfigError
We need a data type that can represent either a successful value (a parsed configuration), or an error.
sealed abstract class Validated[+E, +A]
object Validated {
final case class Valid[+A](a: A) extends Validated[Nothing, A]
final case class Invalid[+E](e: E) extends Validated[E, Nothing]
}
Now we are ready to write our parser.
import cats.data.Validated
import cats.data.Validated.{Invalid, Valid}
case class Config(map: Map[String, String]) {
def parse[A : Read](key: String): Validated[ConfigError, A] =
map.get(key) match {
case None => Invalid(MissingConfig(key))
case Some(value) =>
Read[A].read(value) match {
case None => Invalid(ParseError(key))
case Some(a) => Valid(a)
}
}
}
Everything is in place to write the parallel validator. Recall that we can only do parallel validation if each piece is independent. How do we enforce the data is independent? By asking for all of it up front. Let's start with two pieces of data.
def parallelValidate[E, A, B, C](v1: Validated[E, A], v2: Validated[E, B])(f: (A, B) => C): Validated[E, C] =
(v1, v2) match {
case (Valid(a), Valid(b)) => Valid(f(a, b))
case (Valid(_), i@Invalid(_)) => i
case (i@Invalid(_), Valid(_)) => i
case (Invalid(e1), Invalid(e2)) => ???
}
We've run into a problem. In the case where both have errors, we want to report both. But we have
no way of combining the two errors into one error! Perhaps we can put both errors into a List,
but that seems needlessly specific - clients may want to define their own way of combining errors.
How then do we abstract over a binary operation? The Semigroup type class captures this idea.
import cats.Semigroup
def parallelValidate[E : Semigroup, A, B, C](v1: Validated[E, A], v2: Validated[E, B])(f: (A, B) => C): Validated[E, C] =
(v1, v2) match {
case (Valid(a), Valid(b)) => Valid(f(a, b))
case (Valid(_), i@Invalid(_)) => i
case (i@Invalid(_), Valid(_)) => i
case (Invalid(e1), Invalid(e2)) => Invalid(Semigroup[E].combine(e1, e2))
}
Perfect! But.. going back to our example, we don't have a way to combine ConfigErrors. But as clients,
we can change our Validated values where the error can be combined, say, a List[ConfigError]. It is
more common however to use a NonEmptyList[ConfigError] - the NonEmptyList statically guarantees we
have at least one value, which aligns with the fact that if we have an Invalid, then we most
certainly have at least one error. This technique is so common there is a convenient method on Validated
called toValidatedNel that turns any Validated[E, A] value to a Validated[NonEmptyList[E], A].
Additionally, the type alias ValidatedNel[E, A] is provided.
Time to parse.
import cats.SemigroupK
import cats.data.NonEmptyList
import cats.implicits._
case class ConnectionParams(url: String, port: Int)
val config = Config(Map(("endpoint", "127.0.0.1"), ("port", "not an int")))
implicit val nelSemigroup: Semigroup[NonEmptyList[ConfigError]] =
SemigroupK[NonEmptyList].algebra[ConfigError]
implicit val readString: Read[String] = Read.stringRead
implicit val readInt: Read[Int] = Read.intRead
Any and all errors are reported!
val v1 = parallelValidate(config.parse[String]("url").toValidatedNel,
config.parse[Int]("port").toValidatedNel)(ConnectionParams.apply)
// v1: cats.data.Validated[cats.data.NonEmptyList[ConfigError],ConnectionParams] = Invalid(OneAnd(MissingConfig(url),List(ParseError(port))))
val v2 = parallelValidate(config.parse[String]("endpoint").toValidatedNel,
config.parse[Int]("port").toValidatedNel)(ConnectionParams.apply)
// v2: cats.data.Validated[cats.data.NonEmptyList[ConfigError],ConnectionParams] = Invalid(OneAnd(ParseError(port),List()))
val config = Config(Map(("endpoint", "127.0.0.1"), ("port", "1234")))
// config: Config = Config(Map(endpoint -> 127.0.0.1, port -> 1234))
val v3 = parallelValidate(config.parse[String]("endpoint").toValidatedNel,
config.parse[Int]("port").toValidatedNel)(ConnectionParams.apply)
// v3: cats.data.Validated[cats.data.NonEmptyList[ConfigError],ConnectionParams] = Valid(ConnectionParams(127.0.0.1,1234))
Apply
Our parallelValidate function looks awfully like the Apply#map2 function.
def map2[F[_], A, B, C](fa: F[A], fb: F[B])(f: (A, B) => C): F[C]
Which can be defined in terms of Apply#ap and Apply#map, the very functions needed to create an Apply instance.
Can we perhaps define an Apply instance for Validated? Better yet, can we define an Applicative instance?
Note: the example below assumes usage of the kind-projector compiler plugin and will not compile if it is not being used in a project.
import cats.Applicative
implicit def validatedApplicative[E : Semigroup]: Applicative[Validated[E, ?]] =
new Applicative[Validated[E, ?]] {
def ap[A, B](f: Validated[E, A => B])(fa: Validated[E, A]): Validated[E, B] =
(fa, f) match {
case (Valid(a), Valid(fab)) => Valid(fab(a))
case (i@Invalid(_), Valid(_)) => i
case (Valid(_), i@Invalid(_)) => i
case (Invalid(e1), Invalid(e2)) => Invalid(Semigroup[E].combine(e1, e2))
}
def pure[A](x: A): Validated[E, A] = Validated.valid(x)
}
Awesome! And now we also get access to all the goodness of Applicative, which includes map{2-22}, as well as the
Cartesian syntax |@|.
We can now easily ask for several bits of configuration and get any and all errors returned back.
import cats.Apply
import cats.data.ValidatedNel
implicit val nelSemigroup: Semigroup[NonEmptyList[ConfigError]] =
SemigroupK[NonEmptyList].algebra[ConfigError]
val config = Config(Map(("name", "cat"), ("age", "not a number"), ("houseNumber", "1234"), ("lane", "feline street")))
case class Address(houseNumber: Int, street: String)
case class Person(name: String, age: Int, address: Address)
Thus.
val personFromConfig: ValidatedNel[ConfigError, Person] =
Apply[ValidatedNel[ConfigError, ?]].map4(config.parse[String]("name").toValidatedNel,
config.parse[Int]("age").toValidatedNel,
config.parse[Int]("house_number").toValidatedNel,
config.parse[String]("street").toValidatedNel) {
case (name, age, houseNumber, street) => Person(name, age, Address(houseNumber, street))
}
// personFromConfig: cats.data.ValidatedNel[ConfigError,Person] = Invalid(OneAnd(MissingConfig(street),List(MissingConfig(house_number), ParseError(age))))
Of flatMaps and Xors
Option has flatMap, Xor has flatMap, where's Validated's? Let's try to implement it - better yet,
let's implement the Monad type class.
import cats.Monad
implicit def validatedMonad[E]: Monad[Validated[E, ?]] =
new Monad[Validated[E, ?]] {
def flatMap[A, B](fa: Validated[E, A])(f: A => Validated[E, B]): Validated[E, B] =
fa match {
case Valid(a) => f(a)
case i@Invalid(_) => i
}
def pure[A](x: A): Validated[E, A] = Valid(x)
}
Note that all Monad instances are also Applicative instances, where ap is defined as
trait Monad[F[_]] {
def flatMap[A, B](fa: F[A])(f: A => F[B]): F[B]
def pure[A](x: A): F[A]
def map[A, B](fa: F[A])(f: A => B): F[B] =
flatMap(fa)(f.andThen(pure))
def ap[A, B](fa: F[A])(f: F[A => B]): F[B] =
flatMap(fa)(a => map(f)(fab => fab(a)))
}
However, the ap behavior defined in terms of flatMap does not behave the same as that of
our ap defined above. Observe:
val v = validatedMonad.tuple2(Validated.invalidNel[String, Int]("oops"), Validated.invalidNel[String, Double]("uh oh"))
// v: cats.data.Validated[cats.data.NonEmptyList[String],(Int, Double)] = Invalid(OneAnd(oops,List()))
This one short circuits! Therefore, if we were to define a Monad (or FlatMap) instance for Validated we would
have to override ap to get the behavior we want. But then the behavior of flatMap would be inconsistent with
that of ap, not good. Therefore, Validated has only an Applicative instance.
Validated vs Xor
We've established that an error-accumulating data type such as Validated can't have a valid Monad instance. Sometimes the task at hand requires error-accumulation. However, sometimes we want a monadic structure that we can use for sequential validation (such as in a for-comprehension). This leaves us in a bit of a conundrum.
Cats has decided to solve this problem by using separate data structures for error-accumulation (Validated) and short-circuiting monadic behavior (Xor).
If you are trying to decide whether you want to use Validated or Xor, a simple heuristic is to use Validated if you want error-accumulation and to otherwise use Xor.
Sequential Validation
If you do want error accumulation but occasionally run into places where you sequential validation is needed, then Validated provides a couple methods that may be helpful.
andThen
The andThen method is similar to flatMap (such as Xor.flatMap). In the cause of success, it passes the valid value into a function that returns a new Validated instance.
val houseNumber = config.parse[Int]("house_number").andThen{ n =>
if (n >= 0) Validated.valid(n)
else Validated.invalid(ParseError("house_number"))
}
// houseNumber: cats.data.Validated[ConfigError,Int] = Invalid(MissingConfig(house_number))
withXor
The withXor method allows you to temporarily turn a Validated instance into an Xor instance and apply it to a function.
import cats.data.Xor
def positive(field: String, i: Int): ConfigError Xor Int = {
if (i >= 0) Xor.right(i)
else Xor.left(ParseError(field))
}
Thus.
val houseNumber = config.parse[Int]("house_number").withXor{ xor: ConfigError Xor Int =>
xor.flatMap{ i =>
positive("house_number", i)
}
}
// houseNumber: cats.data.Validated[ConfigError,Int] = Invalid(MissingConfig(house_number))