Izumi Framework Best Practices

General workflow

When we start working on a new component, generally we follow these steps:

Define external data model of the component, namely:
Inputs and outputs of the component methods, as case classes and sealed traits
Expected (domain) error hierarchies, as one sealed trait and final case classes inheriting it
We define abstract interface (trait) of the component,
We use an abstract bifunctor F[+_, +_] in order to encode method outputs
Methods which are not expected to fail should have Nothing in the error (left) branch of the bifunctor
It’s okay to initially have Nothing everywhere and add specific error later, once we discover them
Don’t try to encode all the possible errors, especially irrecoverable ones. E.g. it’s pointless to encode OutOfMemoryError as a possible method result. Only the expected domain errors should be encoded.
Create a stub implementation of the component. If the component is an integration point, its first implementation should be a fake
Use BIO typeclasses as F implementation, if the component is intended to be run under ZIO or another bifunctor, use Either otherwise
Try to use the minimal set of BIO typeclasses necessary for your component. It’s fine to start by using IO2 but you should remove unnecessary capabilites when possible.
Keep the stub implementation inside of the companion object of the component interface
Add distage bindings for the component and its implementation
Create a stub of an abstract test suite for the interface
Evaluate your tests in terms of Constructive Test Taxonomy
Stick to the best/cheapest test kinds (in CTT terms) which would work for your purpose
Try to stick to Blackbox-Contractual tests when you can 6 Work on the test and the implementation (the fake or the business code) simultaneously until you have them functional
It’s not very good but acceptable to use impure mutable fields in dummies instead of Refs.
If the component is an integration point, start working on a real implementation
Add integration checks and their bindings
Add docker definitions and their bindings for managed scene if that’s possible

Example

final case class User(props: Map[String, String])
final case class IdentifiedUser(id: UUID, user: User)

sealed trait DatabaseError
final case class EntityNotFoundExists(id: UUID) extends DatabaseError

trait DatabaseLayer[F[+_, +_]] {
  def store(id: UUID, user: User): F[Nothing, Unit]
  def get(id: UUID): F[EntityNotFoundExists, IdentifiedUser]
}

object DatabaseLayer {
  final class DatabaseLayerDummyImpl[F[+_, +_] : IO2]() extends DatabaseLayer[F] {
    private val content = scala.collection.mutable.HashMap.empty[UUID, IdentifiedUser]

     def store(id: UUID, user: User): F[Nothing, Unit] = F.sync(content += IdentifiedUser(id, user))
     def get(id: UUID): F[EntityNotFoundExists, IdentifiedUser] = F.fromOption(EntityNotFoundExists(id))(content.get(id))
  }
}

final class DatabaseLayerPostgresImpl[F[+_, +_] : IO2]() extends DatabaseLayer[F] {
  // ...
}

class DatabaseModule[F[+_, +_] : TagKK] extends PluginDef {
  make[DatabaseLayer[F]].tagged(Repo.Dummy).from[DatabaseLayer.DatabaseLayerDummyImpl[F]]
  make[DatabaseLayer[F]].tagged(Repo.Prod).from[DatabaseLayerPostgresImpl[F]]
}

class DatabaseLayerTest extends Spec2[zio.IO] {
   override def config: TestConfig = super.config.copy(
      pluginConfig = PluginConfig.const(new DatabaseModule[zio.IO])
   )

   "database layer" should {
      "store users" in {
         (db: DatabaseLayer[zio.IO]) =>
            for {
              // ...
            } yield {
              assert(...)
            }
      }
   }
}

BIO, TF and error encoding

Always use for-comprehension
Always use minimal possible set of BIO typeclasses
You can’t do much with irrecoverable errors. Let them be logged and fail the computation

Error propagation

If one component uses another, it either should propagate its errors or handle them. This makes things somewhat inconvenient, especially on Scala 2 which has no support for union types.

Essentially, on Scala 2 there is no perfect solution for error propagation.

The most comprehensive solution would look like:

sealed trait DatabaseError
final case class EntityNotFoundExists(id: UUID) extends DatabaseError

sealed trait BusinessError
sealed trait UserStoreMethodError {
   this: BusinessError =>
}
final case class UserInvalid(id: UUID) extends BusinessError with UserStoreMethodError
final case class InheritedDatabaseError(error: DatabaseError) extends BusinessError with UserStoreMethodError

trait BusinessLayer[F[+_, +_]] {
   def validateAndStore(id: UUID, user: User): F[UserStoreMethodError, Unit]
}

Essentially, this approach simulates the missing union types.

It kinda breaks encapsulation, because the errors specific only to one particular component implementation are being propagated into the abstract interface. Also this approach, when followed strictly, is just way too verbose.

It’s possible to modify it by just logging dependency errors and returning generic failure branch:

sealed trait BusinessError
sealed trait UserStoreMethodError {
   this: BusinessError =>
}
final case class UserInvalid(id: UUID) extends BusinessError with UserStoreMethodError
// when we return this, we would have to log the details first
final case class InheritedDatabaseError() extends BusinessError with UserStoreMethodError

In most of the cases you wouldn’t want to create separate unions for every method (like UserStoreMethodError) and would just return the base failure (BusinessError) everywhere:

sealed trait BusinessError
final case class UserInvalid(id: UUID) extends BusinessError
final case class InheritedDatabaseError() extends BusinessError

trait BusinessLayer[F[+_, +_]] {
   def validateAndStore(id: UUID, user: User): F[BusinessError, Unit]
}

It’s tempting to use exceptions everywhere either explicitly encoded or just hidden in old bad Java manner. Don’t do that. It’s always possible to have errors encoded as values and, generally, it always pays back.

Only use exceptions and try-catch blocks when you need to integrate with legacy third-party code.

Dual Test Tactic and TDD

When working on a dummy, don’t try to replicate all the specificity of a real endpoint, unless that’s really important. Dummies should be cheap. There is no point in simulating all the possible errors and conditions specific to real integration points.
Probably, you don’t want to simulate database connection errors in a database layer dummy
Though, if you work on a UDP transport layer component, you might want to simulate packet loss in its dummy
Dual test tactic always pays back, always write dual tests for your integration points
With distage it’s not expensive to have dual tests for your business logic. Do that too.
There is no reason to test all the possible combinations of production/dummy dependencies. Usually it’s enough to have an all-production and an all-dummy test configuration.

Data modeling

Denormalize inputs and outputs. E.g. it’s a common useful pattern to split “raw” “unsaved” entities and “stored” entities retrieved from the database
Make your models closed. All the classes should be final, all traits sealed.
Avoid storing untyped data (e.g. JSON) in your models at all costs

Integration points

Keep your integration points as dumb as possible, just wrap the APIs you integrate with
Generally, all the assertions, computations and decisions should happen in your business layer

Inheritance

There is very limited use of class inheritance. Usually one class shouldn’t inherit from another. Make classes final by default. If you intend to use class inheritance, mark them as open, but do not leave them unmarked.
abstract classes might be useful but, generally, you should avoid having them too

Edit this page

1.2.5