Suspending functions need to pass continuations to one another. They do not have any trouble calling normal functions, but normal functions cannot call suspending functions.
Every suspending function needs to be called by another suspending function, which is called by another suspending function, and so on. This all needs to start somewhere. It starts with a coroutine builder, a bridge from the normal to the suspending world1.
We are going to explore the three essential coroutine builders provided by the kotlinx.coroutines library:
Each has its own use cases. Let's explore them.
launch works is conceptually similar to starting a new thread (
thread function). We just start a coroutine, and it will run independently, like a firework that is launched into the air. This is how we use
launch - to start a process.
launch function is an extension function on the
CoroutineScope interface. This is part of an important mechanism called structured concurrency, whose purpose is to build a relationship between the parent coroutine and a child coroutine. Later in this chapter, we will learn about structured concurrency, but for now we will avoid this topic by calling
launch (and later
async) on the
GlobalScope object. This is not a standard practice though as we should rarely use
GlobalScope in real-life projects.
Another thing you might have noticed is that at the end of the
main function we need to call
Thread.sleep. Without doing this, this function would end immediately after launching the coroutines, so they wouldn't have a chance to do their job. This is because
delay does not block the thread: it suspends a coroutine. You might remember from the How does suspension work? chapter that
delay just sets a timer to resume after a set amount of time, and suspends a coroutine until then. If the thread is not blocked, nothing is busy, so nothing stops the program from finishing (later we will see that if we use structured concurrency,
Thread.sleep is not needed).
To some degree, how
launch works is similar to a daemon thread2 but much cheaper. This metaphor is useful initially but it becomes problematic later. Maintaining a blocked thread is always costly, while maintaining a suspended coroutine is almost free (as explained in the Coroutines under the hood chapter). They both start some independent processes and need something that will prevent the program ending before they are done (in the example below, this is
The general rule is that coroutines should never block threads, only suspend them. On the other hand, there are cases in which blocking is necessary. Like in the main function, we need to block the thread, otherwise our program will end too early. For such cases, we might use
runBlocking is a very atypical builder. It blocks the thread it has been started on whenever its coroutine is suspended3 (similar to suspending main). This means that
runBlocking will behave like
There are actually a couple of specific use cases in which
runBlocking is used. The first one is the main function, where we need to block the thread, because otherwise the program will end. Another common use case is unit tests, where we need to block the thread for the same reason.
We might use
runBlocking in our example to replace
delay(2000). Later we will see that it is even more useful once we introduce structured concurrency.
runBlocking used to be an important builder, but in modern programming it is used rather rarely. In unit tests, we often use its successor
runTest instead, which makes coroutines operate in virtual time (a very useful feature for testing which we will describe in the Testing coroutines chapter). For the main function, we often make it suspending.
main is convenient, but for now we will keep using
async coroutine builder is similar to
launch, but it is designed to produce a value. This value needs to be returned by the lambda expression5. The
async function returns an object of type
T is the type of the produced value.
Deferred has a suspending method
await, which returns this value once it is ready. In the example below, the produced value is
42, and its type is
Deferred<Int> is returned, and
42 of type
Just like the
async starts a coroutine immediately when it is called. So, it is a way to start a few processes at once and then await all their results. The returned
Deferred stores a value inside itself once it is produced, so once it is ready it will be immediately returned from
await. However, if we call
await before the value is produced, we are suspended until the value is ready.
async builder works is very similar to
launch, but it has additional support for returning a value. If all
launch functions were replaced with
async, the code would still work fine. But don't do that!
async is about producing a value, so if we don't need a value, we should use
async builder is often used to parallelize two processes, such as obtaining data from two different places, to combine them together.
If a coroutine is started on
GlobalScope, the program will not wait for it. As previously mentioned, coroutines do not block any threads, and nothing prevents the program from ending. This is why, in the below example, an additional
delay at the end of
runBlocking needs to be called if we want to see "World!" printed.
Why do we need this
GlobalScope in the first place? It is because
async are extension functions on the
CoroutineScope. However, if you take a look at the definitions of these and of
runBlocking, you will see that the
block parameter is a function type whose receiver type is also
This means that we can get rid of the
launch can be called on the receiver provided by
runBlocking, so with
this.launch or simply
launch. As a result,
launch becomes a child of
runBlocking. As parents might recognize, a parental responsibility is to wait for all their children, so
runBlocking will suspend until all its children are finished.
A parent provides a scope for its children, and they are called in this scope. This builds a relationship that is called a structured concurrency. Here are the most important effects of the parent-child relationship:
- children inherit context from their parent (but they can also overwrite it, as will be explained in the Coroutine context chapter);
- a parent suspends until all the children are finished (this will be explained in the Job and children awaiting chapter);
- when the parent is cancelled, its child coroutines are cancelled too (this will be explained in the Cancellation chapter);
- when a child raises an error, it destroys the parent as well (this will be explained in the Exception handling chapter).
Notice that, unlike other coroutine builders,
runBlocking is not an extension function on
CoroutineScope. This means that it cannot be a child: it can only be used as a root coroutine (the parent of all the children in a hierarchy). This means that
runBlocking will be used in different cases than other coroutines. As we mentioned before, this is very different from other builders.
The bigger picture
Suspending functions need to be called from other suspending functions. This all needs to start with a coroutine builder. Except for
runBlocking, builders need to be started on
CoroutineScope. In our simple examples, the scope is provided by
runBlocking, but in bigger applications it is either constructed by us (we will explain how to do this in the Constructing coroutine scope chapter) or it is provided by the framework we use (for instance, Ktor on a backend or Android KTX on Android). Once the first builder is started on a scope, other builders can be started on the scope of the first builder, and so on. This is in essence how our applications are structured.
Here are a few examples of how coroutines are used in real-life projects. The first two are typical for both backend and Android.
MainPresenter represents a case typical for Android.
UserController represents a case typical for backend applications.
There is one problem though: what about suspending functions? We can suspend there, but we do not have any scope. Passing scope as an argument is not a good solution (as we will see in the Scoping functions chapter). Instead, we should use the
coroutineScope function, which is a suspending function that creates a scope for builders.
Imagine that in some repository function you need to asynchronously load two resources, for example user data and a list of articles. In this case, you want to return only those articles that should be seen by the user. To call
async, we need a scope, but we don't want to pass it to a function6. To create a scope out of a suspending function, we use the
coroutineScope is just a suspending function that creates a scope for its lambda expression. The function returns whatever is returned by the lambda expression (like
runBlocking). So, in the above example, it returns
List<ArticleJson> because this is what is returned from the lambda expression.
coroutineScope is a standard function we use when we need a scope inside a suspending function. It is really important. The way it is designed is perfect for this use case, but to analyze it we first need to learn a bit about context, cancelling, and exception handling. This is why the function will be explained in detail later in a dedicated chapter (Coroutine scope functions).
We can also start using the suspending main function together with
coroutineScope, which is a modern alternative to using the main function with
runBlocking. In these, we can call other builders or suspending functions. We cannot run builders on suspending functions, so we use coroutine scope functions (like
This knowledge is enough for most Kotlin coroutine usages. In most cases, we just have suspending functions calling other suspending or normal functions. If we need to introduce concurrent processing, we wrap a function with a
coroutineScope and use builders on its scope. Everything needs to start with some builders called on some scope. We will learn in later parts how to construct such a scope, but for most projects it needs to be defined once and is touched only rarely.
Even though we have learned the essentials, there is still a lot to learn about. In the next chapters, we will dive deeper into coroutines. We will learn to use different contexts, how to tame cancellations, exception handling, how to test coroutines, etc. There are still a lot of great features to discover.
It can also start from a suspending main, but even though we use it in a lot in examples, it is not helpful when we write Android or backend applications. It is also good to know that a suspending main is also just started by a builder, but we don't see it because Kotlin does that for us.
A daemon thread is a low-priority thread that runs in the background. Here I compare
launch to a daemon thread because both do not stop the program from ending.
To be more precise, it runs a new coroutine and blocks the current thread interruptibly until its completion.
The reason is that
runBlocking creates a scope, while suspending main does not unless we use the
coroutineScope function, which we will introduce later.
To be strict with wording: by the argument of a functional type placed on the last position.
In the Coroutine scope functions chapter, we will explain in detail why.
Using a dispatcher, we can make
runBlocking run on a different thread. But still, the thread on which this builder has been started will be blocked until the coroutine is done.