Skip to main content


· 8 min read

I've been using ZIO at work for about a year now, and thought I would share some of my learnings. On a couple of occasions, I've helped bring people up to speed on using ZIO in our code bases, so this could be thought of as a getting-started highlight for Scala developers who are familiar with the language, but not necessarily functional effect-based system - in this case, ZIO.

Anatomy of a ZIO Application

The main components to start discussing are ZIO[R, E, A] (the computational effects you want to run) , ZLayer[R, E, A] (the dependencies you need to run your effects), and the Runtime[R] (ZIO - the platform / effect system).

ZIO[R, E, A]

If I were to try to explain what a ZIO/effect is, in as few words as possible, I would say

A ZIO[R, E, A] will compute a result of type A, and will need resources of type R to do it. If it recoverably fails, it will fail with an exception of type E.

Let's dig into that.


There type aliases and companion objects that simplify common cases:

  • Task[A] == ZIO[Any, Throwable, A] - Doesn't need any dependencies to compute A, and can recover from a a failure that is Throwable.
  • UIO[A] == ZIO[Any, Nothing, A - Doesn't need any dependencies to compute A, and won't fail with something you could recover from.
  • IO[E, A] == ZIO[Any, E, A] - Doesn't need any dependencies to compute A, and can recover from a failure that is E.
  • RIO[R,A] == ZIO[R, Throwble, A] - Requires R to compute A, and can recover from a a failure that is Throwable.
  • URIO[R, A] == ZIO[R, Nothing, A]- Requires R to compute A, and won't fail with something you could recover from.

The abbreviations may seem daunting at first, but if you feel like they're too much at start - just don't use them! They're just aliases, and ZIO[Any, Throwable, A] is just as valid as Task[A]. You'll get used them pretty quick though, and if you use IntelliJ Idea + the ZIO Plugin, it'll likely even suggest the shorter version for you to help out.

The E is Not Silent

The U in UIO, or URIO above is sometimes said to be for "Un-failing" (i.e. it can't fail), but it's important to state right off that your FP/Effect system application can still absolutely crash! Errors are not magically handled. This is more of a conceptual thing, and realizing that the E error channel is about exceptions you can and want to recover from. If, for example, you were reading numbers from a database to perform math on and your error channel was an ArithmeticException (e.g. ZIO[Any, ArithmeticException, Int]), you could still crash from an un-checked SQLException - because you said "I'm only concerned with recovering from ArithmeticExceptions". This also isn't Akka, so don't "let it crash" - you still need to catch your exceptions!

For example: This is still going to crash your application if you pass it zero:

def danger(denom: Int): ZIO[Console, Throwable, Int] = for {
result <- ZIO.attempt(42 / denom)
_ <- printLine(s"Computed $result")
} yield result

so, you should be sure to handle the exceptions you want to recover from, e.g.:

def lessDanger(denom: Int) = danger(denom).catchSome {
case _: ArithmeticException => ZIO.succeed(0)

The R: ZLayer[R, E, A]

A lot of people seem to struggle up with ZLayers at first, but I think they aren't that complicated once you get used to them. A ZLayer provides the R resources for a ZIO. Sometimes those resources need dependencies themselves, so just like with a ZIO, a ZLayer[R, E, A] will give you a dependency resource A you can inject into your application, and will need dependencies of type R to do it. If it recoverably fails, it will fail with an exception of type E. Also, like with the ZIO, there are corresponding type aliases which match above.

The tricky part is combining all the layers for you application. For example, if you have a ZLayer[A, Throwable, B] and a ZLayer[B, Throwable, C], depending on how you combine them, you can get a ZLayer[A, Throwable, C] , ZLayer[A with B, Throwable, B with C], or even a ZLayer[A, Throwable, B with C]. This is due to the fact that you can horizontally, and vertically combine layers.

For example, let's look at some type signatures:

val l1: ZLayer[Console, Throwable, Random] = ???
val l2: ZLayer[Random, Throwable, Clock] = ???
val l3: ZLayer[Console, Throwable, Clock] = l1 >>> l2 // Vertically
val l4: ZLayer[Console with Random, Throwable, Random with Clock] = l1 ++ l2 // Horizontally
val l5: ZLayer[Console, Throwable, Random with Clock] = l1 >+> l2 // A bit of both

For l3, we have combined the layers vertically. This means we used the output of l1 and fed it intol2 - which now means in this example we now have a layer where "If you give me a Console I will produce a Clock for you".

For l4, we have combined them horizontally, which mainly just means we stack the Rs and the As - here, you end up with a layer that when given a Console and a Random, it will produce a Random and a Clock.

In the case of l5, it's a bit of both. With >+> it just stacks the As - so we end up with a layer that says "Give me a Console, and I'll give you a Random and a Clock".

So which of these you need, really just depends on if you are going to use the resulting layer to build any other layers - and if you wanted/needed to easily re-use the dependencies in the R channel. A really nice thing is that your overall program is a collection of ZIOs, and as you combine them all, all of their resources stack up - so you know exactly what dependencies your program needs to run, and then you can build a layer to provide them all! For example:

// Get the current time
def currentTime: URIO[Clock, OffsetDateTime] = Clock.currentDateTime
//Log something
def log(msg: String): ZIO[Console, IOException, Unit] = printLine(msg)
// Log the current time
def logTime: ZIO[Console with Clock, IOException, Unit] = for {
time <- currentTime
_ <- log(s"The current time is ${time}")
} yield ()

We can see that if I want to run logTime, I need to provide Console with Clock, which is the combined set of dependencies of the individual ZIOs used to build that method.

The Runtime[Env]

The awesome follow up to the concepts of ZLayers, and knowing what resources your applications needs to run - is that they're just there. By that, I mean the Runtime which is running our application has to know about all the resources needed. For logTime above, that means I have at least a Runtime[Clock with Console]. Whatever the ultimate layer provided to the application (call it AppEnv where type AppEnv = This with That with Other...), you have a Runtime[AppEnv] - and that means you can access any of those dependencies! For example, logTime could be written as

val fromEnv: ZIO[Console with Clock, IOException, Unit] = for {
clock <- ZIO.service[Clock]
time <- clock.currentDateTime
console <- ZIO.service[Console]
_ <- console.printLine(s"The current time is $time")
} yield ()

Looking at clock <- ZIO.service[Clock] - that's basically saying "from the runtime environment, grab a Clock for me to use". So anywhere in your program's logic, if you're writing a line in a ZIO for-comprehension, and you know there's a service of type S provided, you could quickly grab a reference to it with s <- ZIO.service[S] - even if a companion/helper object hasn't been set up to provide it "nicely" via something like Clock.currentDateTime.

Why use an effect system?

Ok, cool. You can do dependency injection and exception handling without an effect system - so what? Well, in addition to the powerful, tightly integrated ergonomics above - this is all run on a performant fiber based system, which means that it takes near zero effort to take any of your code and add retry logic, scheduling, and async operations. What if I wanted to print the current time, 30 seconds in the future? logTime.delay(30 second). Do that 5 times? logTime.delay(30 second).repeatN(5). Log forever in the background while moving ahead in the application? logTime.repeat(Schedule.spaced(1.second)).forkDaemon

What if you're asking for user input, and you want to retry some number of times in case of mistyping?

val fromuser = (for {
_ <- Console.printLine("Enter a number")
input <- Console.readLine
number <- ZIO.attempt(input.toInt) // This could blow up!
_ <- Console.printLine(s"You entered number $number")
} yield ()).retryN(5)

These are of course silly example, but in a real-world application what if you're making a REST call, get an error code with a Retry-After header set you can recursively call yourself with the appropriate timeout with ease!

Wrapping up

I hope that helped hit some highlights of ZIO, and perhaps make it less scary to jump into!