Coroutines and channels − tutorial

Before you start

1. Download and install the latest version of IntelliJ IDEA.

2. Clone the project template by choosing Get from VCS on the Welcome screen or selecting File | New | Project from Version Control.

   You can also clone it from the command line:

   git clone https://github.com/kotlin-hands-on/intro-coroutines

Blocking requests

You will use the Retrofit library to perform HTTP requests to GitHub. It allows requesting the list of repositories under the given organization and the list of contributors for each repository:

This API is used by the loadContributorsBlocking() function to fetch the list of contributors for the given organization.

1. Open src/tasks/Request1Blocking.kt to see its implementation:

   fun loadContributorsBlocking( service: GitHubService, req: RequestData ): List<User> { val repos = service .getOrgReposCall(req.org) // #1 .execute() // #2 .also { logRepos(req, it) } // #3 .body() ?: emptyList() // #4 return repos.flatMap { repo -> service .getRepoContributorsCall(req.org, repo.name) // #1 .execute() // #2 .also { logUsers(repo, it) } // #3 .bodyList() // #4 }.aggregate() }

   • At first, you get a list of the repositories under the given organization and store it in the repos list. Then for each repository, the list of contributors is requested, and all of the lists are merged into one final list of contributors.

   • getOrgReposCall() and getRepoContributorsCall() both return an instance of the *Call class (#1). At this point, no request is sent.

   • *Call.execute() is then invoked to perform the request (#2). execute() is a synchronous call that blocks the underlying thread.

   • When you get the response, the result is logged by calling the specific logRepos() and logUsers() functions (#3). If the HTTP response contains an error, this error will be logged here.

   • Finally, get the response's body, which contains the data you need. For this tutorial, you'll use an empty list as a result in case there is an error, and you'll log the corresponding error (#4).

2. To avoid repeating .body() ?: emptyList(), an extension function bodyList() is declared:

   fun <T> Response<List<T>>.bodyList(): List<T> { return body() ?: emptyList() }

3. Run the program again and take a look at the system output in IntelliJ IDEA. It should have something like this:

   1770 [AWT-EventQueue-0] INFO Contributors - kotlin: loaded 40 repos 2025 [AWT-EventQueue-0] INFO Contributors - kotlin-examples: loaded 23 contributors 2229 [AWT-EventQueue-0] INFO Contributors - kotlin-koans: loaded 45 contributors ...

   • The first item on each line is the number of milliseconds that have passed since the program started, then the thread name in square brackets. You can see from which thread the loading request is called.

   • The final item on each line is the actual message: how many repositories or contributors were loaded.

   This log output demonstrates that all of the results were logged from the main thread. When you run the code with a BLOCKING option, the window freezes and doesn't react to input until the loading is finished. All of the requests are executed from the same thread as the one called loadContributorsBlocking() is from, which is the main UI thread (in Swing, it's an AWT event dispatching thread). This main thread becomes blocked, and that's why the UI is frozen:

   

   After the list of contributors has loaded, the result is updated.

4. In src/contributors/Contributors.kt, find the loadContributors() function responsible for choosing how the contributors are loaded and look at how loadContributorsBlocking() is called:

   when (getSelectedVariant()) { BLOCKING -> { // Blocking UI thread val users = loadContributorsBlocking(service, req) updateResults(users, startTime) } }

   • The updateResults() call goes right after the loadContributorsBlocking() call.

   • updateResults() updates the UI, so it must always be called from the UI thread.

   • Since loadContributorsBlocking() is also called from the UI thread, the UI thread becomes blocked and the UI is frozen.

Task 1

The first task helps you familiarize yourself with the task domain. Currently, each contributor's name is repeated several times, once for every project they have taken part in. Implement the aggregate() function combining the users so that each contributor is added only once. The User.contributions property should contain the total number of contributions of the given user to all the projects. The resulting list should be sorted in descending order according to the number of contributions.

Open src/tasks/Aggregation.kt and implement the List<User>.aggregate() function. Users should be sorted by the total number of their contributions.

The corresponding test file test/tasks/AggregationKtTest.kt shows an example of the expected result.

After implementing this task, the resulting list for the "kotlin" organization should be similar to the following:

Solution for task 1

1. To group users by login, use groupBy(), which returns a map from a login to all occurrences of the user with this login in different repositories.

2. For each map entry, count the total number of contributions for each user and create a new instance of the User class by the given name and total of contributions.

3. Sort the resulting list in descending order:

   fun List<User>.aggregate(): List<User> = groupBy { it.login } .map { (login, group) -> User(login, group.sumOf { it.contributions }) } .sortedByDescending { it.contributions }

An alternative solution is to use the groupingBy() function instead of groupBy().

Callbacks

The previous solution works, but it blocks the thread and therefore freezes the UI. A traditional approach that avoids this is to use callbacks.

Instead of calling the code that should be invoked right after the operation is completed, you can extract it into a separate callback, often a lambda, and pass that lambda to the caller in order for it to be called later.

To make the UI responsive, you can either move the whole computation to a separate thread or switch to the Retrofit API which uses callbacks instead of blocking calls.

Use the Retrofit callback API

In the previous solution, the whole loading logic is moved to the background thread, but that still isn't the best use of resources. All of the loading requests go sequentially and the thread is blocked while waiting for the loading result, while it could have been occupied by other tasks. Specifically, the thread could start loading another request to receive the entire result earlier.

Handling the data for each repository should then be divided into two parts: loading and processing the resulting response. The second processing part should be extracted into a callback.

The loading for each repository can then be started before the result for the previous repository is received (and the corresponding callback is called):

The Retrofit callback API can help achieve this. The Call.enqueue() function starts an HTTP request and takes a callback as an argument. In this callback, you need to specify what needs to be done after each request.

Open src/tasks/Request3Callbacks.kt and see the implementation of loadContributorsCallbacks() that uses this API:

fun loadContributorsCallbacks( service: GitHubService, req: RequestData, updateResults: (List<User>) -> Unit ) { service.getOrgReposCall(req.org).onResponse { responseRepos -> // #1 logRepos(req, responseRepos) val repos = responseRepos.bodyList() val allUsers = mutableListOf<User>() for (repo in repos) { service.getRepoContributorsCall(req.org, repo.name) .onResponse { responseUsers -> // #2 logUsers(repo, responseUsers) val users = responseUsers.bodyList() allUsers += users } } } // TODO: Why doesn't this code work? How to fix that? updateResults(allUsers.aggregate()) }

• For convenience, this code fragment uses the onResponse() extension function declared in the same file. It takes a lambda as an argument rather than an object expression.

• The logic for handling the responses is extracted into callbacks: the corresponding lambdas start at lines #1 and #2.

However, the provided solution doesn't work. If you run the program and load contributors by choosing the CALLBACKS option, you'll see that nothing is shown. However, the test from Request3CallbacksKtTest immediately returns the result that it successfully passed.

Think about why the given code doesn't work as expected and try to fix it, or see the solutions below.

Suspending functions

You can implement the same logic using suspending functions. Instead of returning Call<List<Repo>>, define the API call as a suspending function as follows:

• getOrgRepos() is defined as a suspend function. When you use a suspending function to perform a request, the underlying thread isn't blocked. More details about how this works will come in later sections.

• getOrgRepos() returns the result directly instead of returning a Call. If the result is unsuccessful, an exception is thrown.

Alternatively, Retrofit allows returning the result wrapped in Response. In this case, the result body is provided, and it is possible to check for errors manually. This tutorial uses the versions that return Response.

In src/contributors/GitHubService.kt, add the following declarations to the GitHubService interface:

Coroutines

The code with suspending functions looks similar to the "blocking" version. The major difference from the blocking version is that instead of blocking the thread, the coroutine is suspended:

Starting a new coroutine

If you look at how loadContributorsSuspend() is used in src/contributors/Contributors.kt, you can see that it's called inside launch. launch is a library function that takes a lambda as an argument:

launch { val users = loadContributorsSuspend(req) updateResults(users, startTime) }

Here launch starts a new computation that is responsible for loading the data and showing the results. The computation is suspendable – when performing network requests, it is suspended and releases the underlying thread. When the network request returns the result, the computation is resumed.

Such a suspendable computation is called a coroutine. So, in this case, launch starts a new coroutine responsible for loading data and showing the results.

Coroutines run on top of threads and can be suspended. When a coroutine is suspended, the corresponding computation is paused, removed from the thread, and stored in memory. Meanwhile, the thread is free to be occupied by other tasks:

When the computation is ready to be continued, it is returned to a thread (not necessarily the same one).

In the loadContributorsSuspend() example, each "contributors" request now waits for the result using the suspension mechanism. First, the new request is sent. Then, while waiting for the response, the whole "load contributors" coroutine that was started by the launch function is suspended.

The coroutine resumes only after the corresponding response is received:

While the response is waiting to be received, the thread is free to be occupied by other tasks. The UI stays responsive, despite all the requests taking place on the main UI thread:

1. Run the program using the SUSPEND option. The log confirms that all of the requests are sent to the main UI thread:

   2538 [AWT-EventQueue-0 @coroutine#1] INFO Contributors - kotlin: loaded 30 repos 2729 [AWT-EventQueue-0 @coroutine#1] INFO Contributors - ts2kt: loaded 11 contributors 3029 [AWT-EventQueue-0 @coroutine#1] INFO Contributors - kotlin-koans: loaded 45 contributors ... 11252 [AWT-EventQueue-0 @coroutine#1] INFO Contributors - kotlin-coroutines-workshop: loaded 1 contributors

2. The log can show you which coroutine the corresponding code is running on. To enable it, open Run | Edit configurations and add the -Dkotlinx.coroutines.debug VM option:

   

   The coroutine name will be attached to the thread name while main() is run with this option. You can also modify the template for running all of the Kotlin files and enable this option by default.

Now all of the code runs on one coroutine, the "load contributors" coroutine mentioned above, denoted as @coroutine#1. While waiting for the result, you shouldn't reuse the thread for sending other requests because the code is written sequentially. The new request is sent only when the previous result is received.

Suspending functions treat the thread fairly and don't block it for "waiting". However, this doesn't yet bring any concurrency into the picture.

Concurrency

Kotlin coroutines are much less resource-intensive than threads. Each time you want to start a new computation asynchronously, you can create a new coroutine instead.

To start a new coroutine, use one of the main coroutine builders: launch, async, or runBlocking. Different libraries can define additional coroutine builders.

async starts a new coroutine and returns a Deferred object. Deferred represents a concept known by other names such as Future or Promise. It stores a computation, but it defers the moment you get the final result; it promises the result sometime in the future.

The main difference between async and launch is that launch is used to start a computation that isn't expected to return a specific result. launch returns a Job that represents the coroutine. It is possible to wait until it completes by calling Job.join().

Deferred is a generic type that extends Job. An async call can return a Deferred<Int> or a Deferred<CustomType>, depending on what the lambda returns (the last expression inside the lambda is the result).

To get the result of a coroutine, you can call await() on the Deferred instance. While waiting for the result, the coroutine that this await() is called from is suspended:

runBlocking is used as a bridge between regular and suspending functions, or between the blocking and non-blocking worlds. It works as an adaptor for starting the top-level main coroutine. It is intended primarily to be used in main() functions and tests.

If there is a list of deferred objects, you can call awaitAll() to await the results of all of them:

When each "contributors" request is started in a new coroutine, all of the requests are started asynchronously. A new request can be sent before the result for the previous one is received:

The total loading time is approximately the same as in the CALLBACKS version, but it doesn't need any callbacks. What's more, async explicitly emphasizes which parts run concurrently in the code.

Structured concurrency

• The coroutine scope is responsible for the structure and parent-child relationships between different coroutines. New coroutines usually need to be started inside a scope.

• The coroutine context stores additional technical information used to run a given coroutine, like the coroutine custom name, or the dispatcher specifying the threads the coroutine should be scheduled on.

When launch, async, or runBlocking are used to start a new coroutine, they automatically create the corresponding scope. All of these functions take a lambda with a receiver as an argument, and CoroutineScope is the implicit receiver type:

• New coroutines can only be started inside a scope.

• launch and async are declared as extensions to CoroutineScope, so an implicit or explicit receiver must always be passed when you call them.

• The coroutine started by runBlocking is the only exception because runBlocking is defined as a top-level function. But because it blocks the current thread, it's intended primarily to be used in main() functions and tests as a bridge function.

A new coroutine inside runBlocking, launch, or async is started automatically inside the scope:

When you call launch inside runBlocking, it's called as an extension to the implicit receiver of the CoroutineScope type. Alternatively, you could explicitly write this.launch.

The nested coroutine (started by launch in this example) can be considered as a child of the outer coroutine (started by runBlocking). This "parent-child" relationship works through scopes; the child coroutine is started from the scope corresponding to the parent coroutine.

It's possible to create a new scope without starting a new coroutine, by using the coroutineScope function. To start new coroutines in a structured way inside a suspend function without access to the outer scope, you can create a new coroutine scope that automatically becomes a child of the outer scope that this suspend function is called from. loadContributorsConcurrent()is a good example.

You can also start a new coroutine from the global scope using GlobalScope.async or GlobalScope.launch. This will create a top-level "independent" coroutine.

The mechanism behind the structure of the coroutines is called structured concurrency. It provides the following benefits over global scopes:

• The scope is generally responsible for child coroutines, whose lifetime is attached to the lifetime of the scope.

• The scope can automatically cancel child coroutines if something goes wrong or a user changes their mind and decides to revoke the operation.

• The scope automatically waits for the completion of all child coroutines. Therefore, if the scope corresponds to a coroutine, the parent coroutine does not complete until all the coroutines launched in its scope have completed.

When using GlobalScope.async, there is no structure that binds several coroutines to a smaller scope. Coroutines started from the global scope are all independent – their lifetime is limited only by the lifetime of the whole application. It's possible to store a reference to the coroutine started from the global scope and wait for its completion or cancel it explicitly, but that won't happen automatically as it would with structured concurrency.

Showing progress

Despite the information for some repositories being loaded rather quickly, the user only sees the resulting list after all of the data has been loaded. Until then, the loader icon runs showing the progress, but there's no information about the current state or what contributors are already loaded.

You can show the intermediate results earlier and display all of the contributors after loading the data for each of the repositories:

To implement this functionality, in the src/tasks/Request6Progress.kt, you'll need to pass the logic updating the UI as a callback, so that it's called on each intermediate state:

On the call site in Contributors.kt, the callback is passed to update the results from the Main thread for the PROGRESS option:

• The updateResults() parameter is declared as suspend in loadContributorsProgress(). It's necessary to call withContext, which is a suspend function inside the corresponding lambda argument.

• updateResults() callback takes an additional Boolean parameter as an argument specifying whether the loading has completed and the results are final.

Task 6

In the Request6Progress.kt file, implement the loadContributorsProgress() function that shows the intermediate progress. Base it on the loadContributorsSuspend() function from Request4Suspend.kt.

• Use a simple version without concurrency; you'll add it later in the next section.

• The intermediate list of contributors should be shown in an "aggregated" state, not just the list of users loaded for each repository.

• The total number of contributions for each user should be increased when the data for each new repository is loaded.

Solution for task 6

To store the intermediate list of loaded contributors in the "aggregated" state, define an allUsers variable which stores the list of users, and then update it after contributors for each new repository are loaded:

suspend fun loadContributorsProgress( service: GitHubService, req: RequestData, updateResults: suspend (List<User>, completed: Boolean) -> Unit ) { val repos = service .getOrgRepos(req.org) .also { logRepos(req, it) } .bodyList() var allUsers = emptyList<User>() for ((index, repo) in repos.withIndex()) { val users = service.getRepoContributors(req.org, repo.name) .also { logUsers(repo, it) } .bodyList() allUsers = (allUsers + users).aggregate() updateResults(allUsers, index == repos.lastIndex) } }

Consecutive vs concurrent

An updateResults() callback is called after each request is completed:

This code doesn't include concurrency. It's sequential, so you don't need synchronization.

The best option would be to send requests concurrently and update the intermediate results after getting the response for each repository:

To add concurrency, use channels.

Channels

Writing code with a shared mutable state is quite difficult and error-prone (like in the solution using callbacks). A simpler way is to share information by communication rather than by using a common mutable state. Coroutines can communicate with each other through channels.

Channels are communication primitives that allow data to be passed between coroutines. One coroutine can send some information to a channel, while another can receive that information from it:

A coroutine that sends (produces) information is often called a producer, and a coroutine that receives (consumes) information is called a consumer. One or multiple coroutines can send information to the same channel, and one or multiple coroutines can receive data from it:

When many coroutines receive information from the same channel, each element is handled only once by one of the consumers. Once an element is handled, it is immediately removed from the channel.

You can think of a channel as similar to a collection of elements, or more precisely, a queue, in which elements are added to one end and received from the other. However, there's an important difference: unlike collections, even in their synchronized versions, a channel can suspend send()and receive() operations. This happens when the channel is empty or full. The channel can be full if the channel size has an upper bound.

Channel is represented by three different interfaces: SendChannel, ReceiveChannel, and Channel, with the latter extending the first two. You usually create a channel and give it to producers as a SendChannel instance so that only they can send information to the channel. You give a channel to consumers as a ReceiveChannel instance so that only they can receive from it. Both send and receive methods are declared as suspend:

The producer can close a channel to indicate that no more elements are coming.

Several types of channels are defined in the library. They differ in how many elements they can internally store and whether the send() call can be suspended or not. For all of the channel types, the receive() call behaves similarly: it receives an element if the channel is not empty; otherwise, it is suspended.

Unlimited channel

An unlimited channel is the closest analog to a queue: producers can send elements to this channel and it will keep growing indefinitely. The send() call will never be suspended. If the program runs out of memory, you'll get an OutOfMemoryException. The difference between an unlimited channel and a queue is that when a consumer tries to receive from an empty channel, it becomes suspended until some new elements are sent.

Buffered channel

The size of a buffered channel is constrained by the specified number. Producers can send elements to this channel until the size limit is reached. All of the elements are internally stored. When the channel is full, the next `send` call on it is suspended until more free space becomes available.

Rendezvous channel

The "Rendezvous" channel is a channel without a buffer, the same as a buffered channel with zero size. One of the functions (send() or receive()) is always suspended until the other is called.

If the send() function is called and there's no suspended receive() call ready to process the element, then send() is suspended. Similarly, if the receive() function is called and the channel is empty or, in other words, there's no suspended send() call ready to send the element, the receive() call is suspended.

The "rendezvous" name ("a meeting at an agreed time and place") refers to the fact that send() and receive() should "meet on time".

Conflated channel

A new element sent to the conflated channel will overwrite the previously sent element, so the receiver will always get only the latest element. The send() call is never suspended.

When you create a channel, specify its type or the buffer size (if you need a buffered one):

By default, a "Rendezvous" channel is created.

In the following task, you'll create a "Rendezvous" channel, two producer coroutines, and a consumer coroutine:

Testing coroutines

Let's now test all solutions to check that the solution with concurrent coroutines is faster than the solution with the suspend functions, and check that the solution with channels is faster than the simple "progress" one.

In the following task, you'll compare the total running time of the solutions. You'll mock a GitHub service and make this service return results after the given timeouts:

The sequential solution with the suspend functions should take around 4000 ms (4000 = 1000 + (1000 + 1200 + 800)). The concurrent solution should take around 2200 ms (2200 = 1000 + max(1000, 1200, 800)).

For the solutions that show progress, you can also check the intermediate results with timestamps.

The corresponding test data is defined in test/contributors/testData.kt, and the files Request4SuspendKtTest, Request7ChannelsKtTest, and so on contain the straightforward tests that use mock service calls.

However, there are two problems here:

• These tests take too long to run. Each test takes around 2 to 4 seconds, and you need to wait for the results each time. It's not very efficient.

• You can't rely on the exact time the solution runs because it still takes additional time to prepare and run the code. You could add a constant, but then the time would differ from machine to machine. The mock service delays should be higher than this constant so you can see a difference. If the constant is 0.5 sec, making the delays 0.1 sec won't be enough.

A better way would be to use special frameworks to test the timing while running the same code several times (which increases the total time even more), but that is complicated to learn and set up.

To solve these problems and make sure that solutions with provided test delays behave as expected, one faster than the other, use virtual time with a special test dispatcher. This dispatcher keeps track of the virtual time passed from the start and runs everything immediately in real time. When you run coroutines on this dispatcher, the delay will return immediately and advance the virtual time.

Tests that use this mechanism run fast, but you can still check what happens at different moments in virtual time. The total running time drastically decreases:

To use virtual time, replace the runBlocking invocation with a runTest. runTest takes an extension lambda to TestScope as an argument. When you call delay in a suspend function inside this special scope, delay will increase the virtual time instead of delaying in real time:

You can check the current virtual time using the currentTime property of TestScope.

The actual running time in this example is several milliseconds, whereas virtual time equals the delay argument, which is 1000 milliseconds.

To get the full effect of "virtual" delay in child coroutines, start all of the child coroutines with TestDispatcher. Otherwise, it won't work. This dispatcher is automatically inherited from the other TestScope, unless you provide a different dispatcher:

If launch is called with the context of Dispatchers.Default in the example above, the test will fail. You'll get an exception saying that the job has not been completed yet.

You can test the loadContributorsConcurrent() function this way only if it starts the child coroutines with the inherited context, without modifying it using the Dispatchers.Default dispatcher.

You can specify the context elements like the dispatcher when calling a function rather than when defining it, which allows for more flexibility and easier testing.

By default, the compiler shows warnings if you use the experimental testing API. To suppress these warnings, annotate the test function or the whole class containing the tests with @OptIn(ExperimentalCoroutinesApi::class). Add the compiler argument instructing the compiler that you're using the experimental API:

In the project corresponding to this tutorial, the compiler argument has already been added to the Gradle script.

What's next

• Check out the Asynchronous Programming with Kotlin workshop at KotlinConf.

• Find out more about using virtual time and the experimental testing package.