Home Android & Kotlin Books Kotlin Coroutines by Tutorials

What Is Asynchronous Programming? Written by Filip Babić

The UI (user interface) is a fundamental part of almost every application. It’s what users see and interact with in order to do their tasks. More often than not, applications do complex work, such as talking to external services or processing data from a database. Then, when the work is done, they show a result, mostly in some form of a message.

The UI must be responsive. If the work at hand takes a lot of time to complete, it’s necessary to provide feedback to the user so that they don’t feel like the application has frozen, that they didn’t click a button properly — or perhaps that a feature doesn’t work at all.

In this chapter, you’ll learn how to provide useful information to users about what’s happening in the application and what different mechanisms exist for working with multiple tasks. You’ll see what problems arise while trying to do complex and long-running synchronous operations and how asynchronous programming comes to the rescue.

You’ll start off by analyzing the flow of a function that deals with data processing and provides feedback to the user.

Providing feedback

Suppose you have an application that needs to upload content to a network. When the user selects the Upload button, loading bars or spinners appear to indicate that something is ongoing and the application hasn’t stopped working. This information is crucial for a good user experience since no one likes unresponsive applications. But what does providing feedback look like in code?

Consider the following task wherein you want to upload an image but must wait for the application to complete the upload:

fun uploadImage(image: Image) {
  // Do some work
  // Work’s done, hide the spinner

At first glance, the code gives you an idea of what’s happening:

  • You start by showing a spinner.
  • You then upload an image.
  • When complete, you hide the spinner.

Unfortunately, it’s not exactly that simple because the spinner contains an animation, and there must be code responsible for that. showLoadingSpinner() must then contain code such as this:

fun showLoadingSpinner() {
  while(running) {

showSpinnerView() displays the actual View component, and the following cycle manages the image rotation. But when does this function actually return?

In uploadImage(), you assumed that the spinner animation was running even after the completion of showLoadingSpinner(), so that the uploading of the image could start. Looking at the previous code, this is not possible. If the spinner is animating, it means that showLoadingSpinner() has not completed. If showLoadingSpinner() has completed, then the upload has started. This means that the spinner is not animating anymore. This is happening because when you invoke showLoadingSpinner() you’re making a blocking call.

Blocking calls

A blocking call is essentially a function that only returns when it has completed. In the example above, showLoadingSpinner() prevents the upload of an image because it keeps the main thread of execution busy until it returns. But when it returns (because running becomes false), the spinner stops rotating.

So how can you solve this problem and animate the spinner even while the upload function is executing?

Simply put, you need additional threads on which to execute your long-running tasks.

The main thread is also known as the UI thread, because it’s responsible for rendering everything on the screen, and this should be the only thing it does. This means that it should manage the rotation of the spinner but not the upload of the image — that has nothing to do with the UI. But if the main thread cannot do this because that isn’t its job, what can execute the upload task? Well, quite simply, you need a new thread on which to execute your long-running tasks!

Computers nowadays are far more advanced than they were 10 or 15 years ago. Back in the day computers could only have one thread of execution making them freeze up if you tried to do multiple things at once. But because of technological advancements, your applications support a mechanism known as multi-threading. It’s the art of having multiple threads, where each can process a piece of work, collectively finishing the needed tasks.

Why multithreading?

There’s always been a hardware limit on how fast computers could be — that’s not really about to change. Moreover, the number of operations a single processor in a computer can complete is reaching the law of diminishing returns.

Because of that, technology has steered in the direction of increasing the number of cores each processor has, and the number of threads each core can have running concurrently. This way, you could logically divide any number of tasks between different threads, and the cores could prioritize their work by organizing them. And, by doing so, multithreading has drastically improved how computer systems optimize work and the speed of execution.

You can apply the same idea to modern applications. For example, rather than spending large amounts of money on servers with better hardware, you can speed up the entire system using multithreading and the smart application of concurrency.

Comparing the main and worker threads

The main thread, or the UI thread, is the thread responsible for managing the UI. Every application can only have one main thread in order to avoid a classical problem called deadlock. This can happen when many threads access the same resources — in this case, UI components — in a different order. The other threads, which are not responsible for rendering the UI, are called worker threads or background threads. The ability to allow the execution of multiple threads of control is called multithreading, and the set of techniques used to control their collaboration and synchronization, is called concurrency.

Given this, you can rethink how uploadImage() should work. showLoadingSpinner() starts a new thread that is responsible for the rotation of the spinner image, which interacts with the main thread just to notify a refresh in the UI. Starting a new thread, the function is now a non-blocking call and can return immediately, allowing the image upload to start its own worker thread. When completed, this background thread will notify the main thread to hide the spinner.

Once the program launches a background thread, it can either forget about it or expect some result. You will see how background threads process the result, and communicate with the main thread, in the following section.

Interacting with the UI thread from the background

The upload image example demonstrates how important managing threads is. The thread responsible for rotating the spinner image needs to communicate with the main thread in order to refresh the UI at each frame. The worker thread is responsible for the actual upload and needs to communicate with the UI thread which handles the animation when it completes in order to stop it, and to hide the spinner. All of this must happen without any type of blocks. Knowing how threads communicate is key to achieving the full potential of concurrency.

Sharing data

In order to communicate, different threads need to share data. For instance, the thread responsible for the rotation of the spinner image needs to notify the main thread that a new image is ready to be displayed. Sharing data is not simple, and it needs some sort of synchronization, which is one of the main benefits of well written concurrency code.

What happens, for instance, if the main thread receives a notification that a new image is available and, before displaying it, the image is replaced? In this case, the application would skip a frame and a race condition would happen. You then need some sort of a thread safe data structure. This means that the data structure should work correctly even if accessed by multiple threads at the same time.

Accessing the same data from multiple threads, maintaining the correct behavior and good performance, is the real challenge of concurrent programming.

There are special cases, however. What if the data is only accessed and never updated? In this case, multiple threads can read the same data without any race condition, and your data structure is referred to as immutable. Immutable objects are always thread safe.

As a practical example, take a coffee machine in an office. If two people shared it, and it wasn’t thread safe, they could easily make bad coffee or spill it and make a mess. As one person started making a mocha latte and another wanted a black coffee, they would ultimately ruin the machine — or worse, the coffee.

What are the data structures that you can use in order to safely share data in a thread? The most important data structures are queues and, as a special case, pipelines.


Threads usually communicate using queues, and they can act on them as producers or consumers. A producer is a thread that puts information into the queue, and the consumer is the one that reads and uses them. You can think of a queue as a list in which producers append data to the end, and then consumers read data from the top, following a logic called FIFO (First In First Out). Threads usually put data into the queue as objects called messages, which encapsulate the information to share.

A queue is not just a container, but it also provides synchronization in order to allow a thread to consume a message only if it is available. Otherwise, it waits if the message is not available. If the queue is a blocking queue, the consumer can block and wait for a new message — or just retry later.

The same can happen for the producer if the queue is full. Queues are thread safe, so it is possible to have multiple producers and multiple consumers.

A great real-life example of queues are fast food lines.

Imagine having three lines at a fast food restaurant. The first line has no customers, so the person working the line is blocked until someone arrives. The second has customers, so the line is slowly getting smaller as the worker serves customers. However, the last line is full of customers, but there’s no one to serve them; this, in turn, blocks the line until help arrives.

In this example customers form a queue waiting to consume what the fast food workers are preparing for them. When the food is available, the customer consumes it and leaves the queue. You could also look at the customers as produced work, which the workers need to consume and serve, but the idea stays the same.


If you think about pipes or faucets and how they work, it’s a fairly simple concept. When you release the pressure by turning the valve, you’re actually requesting water. On the other side of that request, there’s a system that regulates the flow of water. As soon as you make a request, it is blocked until the water comes running — just like a blocking call.

The same process is used for pipelines or pipes in programming. There’s a pipe that allows streams of data to flow, and there are listeners. The data is usually a stream of bytes, which the listeners parse into something more meaningful.

As an example, you can also think about factory lines. Just like in a factory line, if there’s too much product, the line has to stop until you process everything. That is, if there’s too much data that you haven’t yet processed, the pipeline is blocked until you consume some of the data and make room for more to flow. And, alternatively, if there’s not enough product, the person processing it sits and waits until something comes up.

In other words, if there’s not enough data to flow — the pipe is empty — you’re blocked until some data emerges. Because you’re either trying to send data to an overflowed stream, or trying to get data from an empty stream, the mechanism doesn’t know how to react but to block until the conditions are met.

You can think of pipes as blocking queues wherein you don’t have messages, but chunks of bytes.

Handling work completion using callbacks

Out of all the asynchronous programming mechanisms, callbacks are the most often used. This consists of the creation of objects that encapsulate code that somebody else can execute later, like when a specific task completes . This approach can also be used in real life when you ask somebody to push a button when they have completed some task you have assigned to them. When using callbacks, the button is analogous to code for them to execute; the person executing the task is a non-blocking function.

How can you put some code into an object to pass around? One way is by using interfaces. You can create the interface in this way:

interface OnUploadCallback {

  fun onUploadCompleted()

With this, you are passing an implementation of the interface to the function that is executing the long-running task. At completion, this function will invoke onUploadCompleted() on the object. The function doesn’t know what that implementations does, and it’s not supposed to know.

In modern programming languages like Kotlin, which support functional programming features, you can do the same with a lambda expression. In the previous example, you could pass the lambda to the upload function as a callback. The lambda would then contain the code to execute when the upload task completes:

fun uploadImage(image: Image) {

  uploadService.upload(image) { hideLoadingSpinner() }

Looking back at the very first snippet, not much has changed. You still show a loading spinner, call upload() and hide the spinner when the upload is done. The core difference, though, is that you’re not calling hideLoadingSpinner() right after the upload. That function is now part of the lambda block, passed as a parameter to upload(), which will be executed at completion. Doing so, you can call the wrapped function anytime you’re done with the connected task. And the lambda block can do pretty much anything, not just hide a loading spinner.

In case some value is returned, it is passed down into the lambda block, so that you can use it from within. Of course, the inner implementation of the uploadService depends on the service and the library that you’re using. Generally, each library has its own types of callbacks. However, even though callbacks are one of the most popular ways to deal with asynchronicity, they have become notorious over the years. You’ll see how in the next section.

Indentation hell

Callbacks are simpler than building your own mechanisms for thread communication. Their syntax is also fairly readable, when it comes to simple functions. However, it’s often the case that you have multiple function calls, which need to be connected or combined somehow, mapping the results into more complex objects.

In these cases, the code becomes extremely difficult to write, maintain and reason about. Since you can’t return a value from a callback, but have to pass it down the lambda block itself, you have to nest callbacks. It’s similar to nesting forEach or map statements on collections, where each operation has its own ambda parameter.

When nesting callbacks, or lambdas, you get a large number of braces ’{}’, each forming a local scope. This, in turn, creates a structure called indentation hell — or callback hell (when it’s specific to callbacks). A good example would be the fetching, resizing and uploading images:

fun uploadImage(imagePath: String) {

  loadImage(imagePath) { image ->
    resizeImage(image) { resizedImage ->
      uploadImage(resizedImage) {

You show the upload spinner before the upload itself, as before. But, after you load the image from a file, you proceed to resize it. Next, when you’ve resized the image successfully, you start uploading it. Finally, once you manage to upload it, you hide the loading spinner.

The first thing you notice is the amount of braces and indentation that form a stair-like code structure. This makes the code very hard to read, and it’s not even a complex operation. When building services on the web, nesting can easily reach 10 levels, if not more. Not only is the code hard to read, but it’s also extremely hard to maintain such code. Because of the structure, you suffer from cognitive load, making it harder to reason about the functionality and flow. Trying to add a step in between, or change the lambda-result types, will break all the subsequent levels.

Additionally, some people find callbacks really hard to grasp at first. Their steep learning curve, combined with the cognitive load and the lack of extensibility, make people look elsewhere for a solution to asynchronous programming. This is where reactive extensions come to life. You’ll see how they solve the nesting problem in the next section.

Using reactive extensions for background work

The most significant issue of a callback-based approach is passing the data from one function to another. This results in nested callbacks, which are tough to read and maintain.

If you think about the queues and pipes, they operate with streams of data, wherein you can listen to the data as long as you need. Reactive extensions, or Rx, are built upon the idea of having asynchronous operations wrapped up in streams of events.

Rx incorporates the observer pattern into helpful constructs. Furthermore, there are a large number of operators that extend the behavior of observable streams, allowing for clean and expressive data processing. You can subscribe to a stream of events, map, filter, reduce and combine the events in numerous ways, as well as handle errors in the entire chain of operations, using a single lambda function.

The previous example of loading, uploading and resizing an image, using Rx, can be represented as:

fun uploadImage(imagePath: String) {
    .subscribe(::hideLoadingSpinner, ::handleError)

At first, this code might look weird. In reality, it’s a stream of data modified by using a bunch of operators. It begins with the flatMap operator, which takes some data — the image from loadImage() — and passes it to another function, creating a new stream. Then, the new stream sends events in the form of resizedImage, which gets passed to uploadImage(), using flatMapCompletable(), and operator chaining.

Finally, the uploadImage stream doesn’t pass data but, rather, completion events, which tell you to hide the loading spinner when the upload has finished.

These streams of data and operations don’t actually get executed until someone subscribes to them, using subscribe(onComplete, onError).

Additionally, doOnSubscribe() takes an action that the stream executes whenever you subscribe to it. There are also functions like doOnSuccess and doOnError, which propagate their respective events.

Further, it’s important to know that, if any error or exception occurs in any of the operations in a chain, it’s not thrown, and the application doesn’t crash. Instead, the stream passes it down the chain, finally reaching the onError lambda. Callbacks do not have this behavior; they just throw the exception and you have to handle it yourself, using try/catch blocks.

Reactive extensions are cleaner than callbacks when it comes to asynchronous programming, but they also have a steeper learning curve.

With dozens of operators, different types of streams and a lot of edge cases with switching between threads, it takes a large amount of time to fully understand them.

The learning curve, and a few other issues, will be discussed in the next section.

Diving deeper into the complexity of Rx

Since this book isn’t about Rx, you’ll only have a narrow overview of its positive and negative features. As seen before, Rx makes asynchronous programming clean and readable. Further, in addition to the operators that allow for data processing, Rx is a powerful mechanism. Moreover, the error handling concept of streams adds extra safety to applications.

But Rx is not perfect. It has problems like any other framework, or paradigm, some of which are showing up in the programming community lately.

To start, there is the learning curve. When you start learning Rx, you have to learn a number of additional concepts, such as the observer pattern and streams. You will also find that Rx is not just a framework; it brings a completely new paradigm called reactive programming. Because of this, it’s very hard to start working with Rx. But it’s even harder to grasp the finesse of using its operators. The amount of operators, types of thread scheduling, and the combinations between the two, creates so many options that it’s nearly impossible to know the full extent of Rx.

Another problematic issue with using Rx is the hype. Over the years, people have moved towards Rx as a silver bullet for asynchronous operations.

This eventually led to such programming being Rx-driven, introducing even more complexity to existing applications. Finding workarounds and using numerous design patterns, just to make Rx work, introduced new layers of unwanted complexity. Because of this, in Android, the Rx community has been debating if programmers should represent things like network requests as streams of data versus just a single event that they could handle using callbacks or something even simpler.

The same debate transitions to navigation events, as an example. Should programmers represent clicks as streams of events, too? The community opinion is very divided on this topic.

So, with all this in mind, is there a better or simpler way to deal with asynchronicity? Oddly enough, there’s a concept dating back decades, which has recently become a hot topic.

A blast from the past

This is a book about coroutines. They’re a mechanism dating back to the 1960’s, depicting a unique way of handling asynchronous programming. The concept revolves around the use of suspension points, suspendable functions and continuations as first-class citizens in a language.

They’re a bit abstract, so it’s better to show an example:

fun fetchUser(userId: String) {
  val user = userService.getUser(userId) // 1
  print("Fetching user") // 2
  print(user.name) // 3
  print("Fetched user") // 4

Using the above code snippet, and revisiting what you learned about blocking calls, you’d say that the execution order was 1, 2, 3 and 4. If you carefully look at the code, you realize that this is not the only possible logical sequence. For instance, the order between 1 and 2 is not important, nor is the order between 3 and 4. What is important is that the user data is fetched before it is displayed; 1 must happen before 3. You can also delay the fetching of the user data to a convenient time before the user data is actually displayed. Managing these issues in a transparent way is the black magic of coroutines!

They’re a part-thread, part-callback mechanism, which use the system’s power of scheduling and suspending work. This way, you can immediately return a result from a call without using callbacks, threads or streams. Think of it this way, once you start a coroutine, or call a suspendable function, it gets nicely wrapped up and prepared like a taco. But, until you want to eat the taco, the code inside might not get executed.

Explaining coroutines: The inner works

It’s not really black magic — only a smart way of using low-level processing. getUser() is marked as a suspendable function, meaning the system prepares the call in the background, and you get an unfinished, wrapped taco. But it might not execute the function yet. The system moves it to a thread pool, where it waits for further commands. Once you’re ready to eat the taco and you request the result, the program can block until you get a ready-to-go snack, or suspend and wait for it within the coroutine.

Knowing this, the program can skip over the rest of the function code, until it reaches the first line of code on which it uses the user. This is called awaiting the result. At that point, it executes getUser() and if it hasn’t already, suspends the program.

This means you can do as much processing as you want, in between the call itself and using its result. Because the compiler knows suspension points and suspendable functions are asynchronous and treats their execution sequentially, you can write understandable and clean code. This makes your code very extensible and easy to maintain.

Since writing asynchronous code is so simple with coroutines, you can easily combine multiple requests or transformations of data. No more staircases, strange stream mapping to pass the data around, or complex operators to combine or transform the result. All you need to do is mark functions as suspendable, and call them in a coroutine block.

Another, extremely important thing to note about coroutines is that they’re not threads. They are a low-level mechanism that utilizes thread pools to shuffle work between multiple, existing threads. This allows you to create millions of coroutines, without overflowing memory. A million threads would take so much memory, even today’s state-of-the-art computers would crash.

Although many languages support coroutines, each has a different implementation.

Variations through history

As mentioned, coroutines are a dated but powerful concept. Throughout the years, several programming languages have evolved their versions of the implementation. For example, in languages like Python and Smalltalk, coroutines are first-class citizens, and can be used without an external library.

A generator in Python would look like this:

def coroutine():
    while True:
        value = yield
        print(’Received a value:’, value)

This code defines a function, which loops forever, listening and printing any arguments you send to it. The concept of an infinite loop, which listens for data is called a generator. The keyword yield is what triggers the generator, receiving the value. As you can see, there’s a while True statement in the function. In regular code, this would create a standard infinite loop, effectively blocking the program, since there’s no exit condition. But this is a coroutine-powered call, so it waits in the background until you send some value to the function, which is why it doesn’t block.

Another language with first-class coroutines is C#. In C#, there’s support for the yield statement, like in Python, but also for async and await calls, like this:

MyResult result = await AsyncMethodThatReturnsAResult();

await AsyncMethodWithoutAResult();

By adding the await keyword, you can return an asynchronous result, using normal, sequential code. It’s pretty much what you saw in the example above, where you first learned about coroutines.

Both Python and C# have first-class support for coroutines. By including them in the language itself, it allows you to make asynchronous calls without including a third-party framework. Many other programming languages utilize external libraries in order to support programming with coroutines. Kotlin also has coroutine support in its standard library. Additionally, the way Kotlin coroutines are built using global and extension functions with receivers, makes them very extensible. You can also create your own APIs by building on top of the existing functions.

You’ll see how to do this in the next chapters of the book.

Key points

  • Multithreading allows you to run multiple tasks in parallel.
  • Asynchronous programming is a common pattern for thread communication.
  • There are different mechanisms for sharing data between threads, some of which are queues and pipelines.
  • Most mechanisms rely on a push-pull tactic, blocking threads when there is too much, or not enough data, to process
  • Callbacks are a complex, hard-to-maintain and cognitive-load-heavy mechanism.
  • It’s easy to reach callback hell when doing complex operations using callbacks.
  • Reactive extensions provide clean solutions for data transformation, combination and error handling.
  • Rx can be too complex, and doesn’t fit all applications.
  • Coroutines are an established, and reliable concept, based on low-level scheduling.
  • Too many threads can take up a lot of memory, ultimately crashing your program or computer.
  • Coroutines don’t always create new threads, they can reuse existing ones from thread pools.
  • It’s possible to have asynchronous code, written in a clean, sequential style, using coroutines.

Where to go from here?

Well that was a really brief overview of the history and theory behind asynchronous programming and coroutines.

If you’re excited about seeing some code and Kotlin’s coroutines, in the next section of the book you’ll learn about suspendable functions and suspension points. Moreover, you’ll see how coroutines are created in Kotlin, using coroutine builders. Next, you’ll build asynchronous calls, which return some data with the async function, and see how you await the result. And, finally, you’ll learn about jobs and their children, in coroutines.

You’ll cover the entire base API for Kotlin Coroutines, learn how to wrap asynchronous calls into async blocks, how to combine multiple operations and how to build Jobs which have multiple layers of coroutines.

But before that, you have to set up your build environment, so let’s get going!

Have a technical question? Want to report a bug? You can ask questions and report bugs to the book authors in our official book forum here.

Have feedback to share about the online reading experience? If you have feedback about the UI, UX, highlighting, or other features of our online readers, you can send them to the design team with the form below:

© 2021 Razeware LLC