Kotlin

Asynchronous
Programming
in Kotlin

@kotlin | Developed by JetBrains

What weʼll cover

● Parallel and asynchronous programming

● The history of coroutines

● Kotlin coroutines

● Inside CoroutineScope

● Channels

● More
Parallel programming

New Runnable scheduler Running Terminated

start done
interrupted

blocking
call

Blocked
Parallel programming

● In reality, programs (threads) spend a lot of time waiting for data to be fetched from disk,
network, etc.

● The number of threads that can be launched is limited by the underlying operating system (each
takes some number of MBs).

● Threads aren’t cheap, as they require context switches which are costly.

● Threads aren’t always available. Some platforms, such as JavaScript, do not even support them.

● Working with threads is hard. Bugs in threads (which are extremely difficult to debug), race
conditions, and deadlocks are common problems we suffer from in multi-threaded
programming.

● Threads terminating due to exceptions is a problem that deserves to be a separate point.

An example

fun postItem(item: Item) {

val token = preparePost()
val post = submitPost(token, item)
processPost(post)
}

fun preparePost(): Token { ./ requestToken

./ makes a request and consequently blocks the execution thread
return token
}
An example

How this code gets executed on a single thread

Thread
An example

What we want

submitPost

Thread

processPost processPost
An example

What we get

blocked submitPost blocked

Thread

processPost processPost
An example

What happens when we go multi-threaded

put results into shared memory

T#2 blocked sync

aware of T#1 and T#2 results

Thread work sync work

put results into shared memory

T#3 blocked sync throws

terminated due to exception

Asynchronous
Programming
Continuation passing style
fun preparePostAsync(callback: (Token) .> Unit) {
./ make request and return immediately
./ arrange callback to be invoked later
}

With callbacks, the idea is to pass one function as a parameter to another function and have this one
invoked once the process has completed.

fun postItem(item: Item) {

preparePostAsync { token .>
submitPostAsync(token, item) { post .>
processPost(post)
}
}
}
Continuation passing style
fun postItem(item: Item) {
preparePostAsync { token .>
submitPostAsync(token, item) {
post .> processPostAsync(post) {
…
}
}
}
}

● The } ladder is the Stairway to Heaven Highway to “Callback Hell”.

● Where is the error handling?

● Callbacks are not asynchronous “by nature”.

Futures, promises, and other approaches

Promise<T> encapsulates the callback.

fun preparePostAsync(): Promise<Token> {

./ makes request and returns a promise that is completed later
return promise
}
Futures, promises, and other approaches

Promise<T> encapsulates the callback.

fun postItem(item: Item) {

preparePostAsync()
.thenCompose { token .> submitPostAsync(token, item) }
.thenAccept { post .> processPost(post) }
…
}

● This model differs from the typical top-down imperative approach.

● There are different APIs, which vary across libraries, frameworks, and platforms.

● It employs the Promise<T> return type instead of the actual we need.

● Each thenCompute/Accept/Handle creates a new object.

● Error handling can be complicated.

Kotlin coroutines
suspend — a keyword in Kotlin marking suspendable function.
suspend fun submitPost(token: Token, item: Item): Post {
...
}

suspend fun postItem(item: Item) {

val token = preparePost()
val post = submitPost(token, item)
processPost(post)
}

This looks and feels sequential, allowing you to focus on the logic of your code.
marks suspension points in IntelliJ IDEA.
The history of coroutines
History and definition

● Melvin Conway coined the term “coroutine” in 1958 for his assembly program.

● Coroutines were first introduced as a language feature in Simula’67 with the detach and
resume commands.

● A coroutine can be thought of as an instance of a suspendable computation, i.e. one that can
suspend at some point and later resume execution, possibly even on another thread.

● Coroutines calling each other (and passing data back and forth) can form the machinery for
cooperative multitasking.

● Go’09, C#’12, Kotlin’17, C++’20, OpenJDK, Project Loom.

Kotlin

Coroutines came to Kotlin in version 1.1, and they became stable in version 1.3.

● suspend – A keyword for marking suspendable functions.

● kotlin.coroutines – A tiny part of the standard library.
● kotlinx.coroutines – A library with all the necessary functionality. It is not a part of the
standard library, meaning there are no additional requirements for the host platform,
facilitating multiplatform development.

A coroutine is an instance of suspendable computation. It is conceptually similar to a thread in the

sense that it takes a block of code to run and has a similar life-cycle. It is created and started, but it is
not bound to any particular thread. It may suspend its execution in one thread and resume in
another one. Moreover, like a future or a promise, it can complete with some result (which is either a
value or an exception).
Kotlin coroutines
Under the hood

The compiler turns your suspend function:

suspend fun submitPost(token: Token, item: Item): Post {...}

Into:
fun submitPost(token: Token, item: Item, cont: Continuation<Post>) {...}

Where:
public interface Continuation<in T> {
public val context: CoroutineContext
public fun resumeWith(result: Result<T>)
}

Continuation<in T> ∼ Generic callback

Under the hood

Code with suspending calls:

./ code inside postItem

./ suspend call 0
val token = preparePost()
./ suspend call 1
val post = submitPost(token, item)
./ suspend call 2
processPost(post)

Is compiled into (simplified version)

→
Under the hood
./ some code here, continuation is created

Code with suspending calls: when(continuation.label) {

0 .> { ./ suspend call 0
cont.label = 1;
./ code inside postItem
./ suspend call 0 preparePost(cont);
val token = preparePost() }
./ suspend call 1 1 .> { ./ suspend call 1
val post = submitPost(token, item) val token = prevResult;
./ suspend call 2 cont.label = 2;
processPost(post)
submitPost(token, item, cont);
}
Is compiled into ( simplified version ) 2 .> { ./ suspend call 2
→ val post = prevResult;
processPost(post, cont);
A large switch??? A state machine! }
}
Each label marks a suspension point.
./ more code here
State of a coroutine

label = 0 label = 1 label = 2

processPost

token submitPost

preparePost post

SUSPENDED
Practice

Now we can finally post items without blocking the execution thread!

fun nonBlockingItemPosting(...) {
...
postItem(item)
}
Practice

Now we can finally post items without blocking the execution thread!

fun nonBlockingItemPosting(...) {
...
postItem(item)
}

The suspending function postItem should be called only from a coroutine or another suspending
function.

One cannot just walk into a suspending function.

Inside CoroutineScope
Practice
suspend functions can be called from other suspend functions or within CoroutineScope.

fun main() = runBlocking { ./ this: CoroutineScope

launch { ./ launch a new coroutine and continue
delay(1000L) ./ non-blocking delay for 1 second (default time unit is ms)
println("World!") ./ print after delay
}
print("Hello ") ./ main coroutine continues while the previous one is delayed
}

HOFs like launch are called coroutine builders.

Sophisticated practice
val jobs: List<Job> = List(1_000_000) {
launch(Dispatchers.Default + CoroutineName("#$it")
+ CoroutineExceptionHandler { context, error .>
println("${context[CoroutineName]..name}: $error")
}, ./ CoroutineContext
CoroutineStart.LAZY ./ do not start instantly
) {
delay(Random.nextLong(1000))
if (it % 10 .= 0) { throw Exception("No comments") }
println("Hello from coroutine $it!")
}
}

jobs.forEach { it.start() }

Now we are going to cover all of this step by step.

Scope and context
public interface CoroutineScope {
public val coroutineContext: CoroutineContext
}

Easy, isn’t it?

Scope and context
public interface CoroutineScope {
public val coroutineContext: CoroutineContext
}

public interface CoroutineContext {

public operator fun <E : Element> get(key: Key<E>): E?
…

public interface Element : CoroutineContext {

public val key: Key..>
...
}
}

You can think of context like Map<Key<Element>, Element>

Inside CoroutineScope
Job
Job
public interface Job : CoroutineContext.Element {
public companion object Key : CoroutineContext.Key<Job>
public fun start(): Boolean
public fun cancel(cause: CancellationException? = null) public val children: Sequence<Job>
...
}

● A Job is work that is executed in the background.

● It is a cancellable work item with a life-cycle that culminates in its completion.

● Jobs can be arranged into parent-child hierarchies.

● A child’s failure immediately cancels its parent along with all its other children. This behavior can be customized using
SupervisorJob.
Job States

start complete finish

New Active Completing Completed

wait for
children

cancel/fail

finish
Cancelling Cancelled
Job states

state isAlive isCompleted isCancelled

New false false false

Active true false false

Completing true false false

Cancelling false false true

Cancelled false true true

Completed false true false

Inside CoroutineScope
Dispatchers
Dispatchers

public abstract class CoroutineDispatcher : ... {

...
public abstract fun dispatch(context: CoroutineContext, block: Runnable)
}

● Dispatchers.Default – A shared pool of background threads, at least 2, depending on the

default number of CPU cores. It is an appropriate choice for compute-intensive coroutines.

● Dispatchers.IO – A shared pool of on-demand created threads and is designed for offloading
IO-intensive blocking operations (such as file/socket IO).
Dispatchers

● Dispatchers.Main – A dispatcher that is confined to the Main thread operating with UI

objects. Usually single-threaded, it is not present in core, but is instead provided by packages
like android, swing, etc.

● Dispatchers.Unconfined – The unconfined dispatcher should not normally be used in code.

● Private thread pools can be created with newSingleThreadContext and

newFixedThreadPoolContext. (Both are @ExperimentalCoroutinesApi.)

● A view of a dispatcher with the guarantee that no more than parallelism coroutines are
executed at the same time can be created via:

./ method of public abstract class CoroutineDispatcher

@ExperimentalCoroutinesApi
public open fun limitedParallelism(parallelism: Int): CoroutineDispatcher
{ ... }
Dispatchers

● An arbitrary ExecutorService can be converted into a dispatcher with the

asCoroutineDispatcher extension function.

interface ExecutorService : Executor {

fun execute(command: Runnable) ./ Executor is a SAM with this method
...
}

val myExecutorService: ExecutorService = ...

val myDispatcher = myExecutorService.asCoroutineDispatcher()
A peek under the hood

internal class GlobalQueue : LockFreeTaskQueue<Task>(singleConsumer = false)

internal class CoroutineScheduler(

@JvmField val corePoolSize: Int, @JvmField val maxPoolSize: Int,
@JvmField val idleWorkerKeepAliveNs: Long = ...,
@JvmField val schedulerName: String = ...
) : Executor, Closeable {
...
val globalCpuQueue = GlobalQueue()

val globalBlockingQueue = GlobalQueue()

...
}
A peek under the hood

internal class CoroutineScheduler(...) : Executor, Closeable {

...
val workers = AtomicReferenceArray<Worker?>(maxPoolSize + 1)

fun dispatch(
block: Runnable,
taskContext: TaskContext = NonBlockingContext,
tailDispatch: Boolean = false
) {
...
}
}
A peek under the hood

internal inner class Worker private constructor() : Thread() {

...
val localQueue: WorkQueue = WorkQueue()
var state = WorkerState.DORMANT
fun findTask(scanLocalQueue: Boolean): Task? {
./ localQueue .> globalBlockingQueue
return task .: trySteal(blockingOnly = true)
}
}
Inside CoroutineScope
Coroutines vs threads
Adding contexts
val jobs: List<Job> = List(1_000_000) {
launch(
BaseContext
+ SupervisorJob()
+ CoroutineName("#$it")
+ CoroutineExceptionHandler { context, error .>
println("${context[CoroutineName]..name}: $error")
}, ./ launch’s first argument is CoroutineContext, which is a sum here
...
) { ... }

● Contexts can be added together. In this case, the rightmost value for a Key is taken as the
resulting context.

● Since each Element implements CoroutineContext, this looks like a sum of elements.
Context switching

suspend fun preparePost(): Token = withContext(Dispatchers.IO) { ... }

./ submitPost also withContext(Dispatchers.IO)

suspend fun processPost(post: Post) =

withContext(Dispatchers.Default) { ... }

suspend fun postItem(item: Item) {

val token = preparePost()
val post = submitPost(token, item)
processPost(post)
}

./ somewhere in our application's code there is a View and a CoroutineScope related to it

viewScope.launch {
postItem(someItem)
./ show the result in the UI somehow
}
How is this actually better than threads?
UI

Main

IO

Default

time
How is this actually better than threads?

Main post

IO

Default

time
How is this actually better than threads?

Main post
withContext(Dispatchers.IO)

IO fetch

Default

time
How is this actually better than threads?

Main post

IO fetch blocked

Default

time
How is this actually better than threads?

Main post

IO fetch blocked done

Default

time
How is this actually better than threads?

Main post

IO fetch blocked done

withContext(Dispatchers.Defalt)

Default process

time
How is this actually better than threads?

Main post show

Back to the original scope/dispatcher

IO fetch blocked done

Default process

time
How is this actually better than threads?

Main post Not blocked show

IO fetch blocked done

Default process

time
How is this actually better than threads?

Main post like sub dwnld post calc post upd

IO#1 fetch blocked done fetch blocked

IO#2 blocked done fetch blocked done fetch blocked

IO#3 fetch blocked done error fetch blocked done

Def#1 process analyze compute intense compute

Def#2 harder error better faster error stronger

time
Coroutines - fibers - threads

fun main(): Unit = runBlocking {

repeat(1_000_000) { ./ it: Int
delay(Random.nextLong(1000))
println("Hello from coroutine $it!")
}
}
Coroutines - fibers - threads

fun main(): Unit = runBlocking {

repeat(1_000_000) { ./ it: Int
delay(Random.nextLong(1000))
println("Hello from coroutine $it!")
}
}

WRONG!
The default behavior is sequential, you have to ask for concurrency.
Coroutines - fibers - threads

fun main(): Unit = runBlocking {

repeat(1_000_000) { ./ it: Int
launch { ./ new asynchronous activity
delay(1000L)
println("Hello from coroutine $it!")
}
}
}

Coroutines are like light-weight threads.

Coroutines - fibers - threads

fun main(): Unit {

repeat(1_000_000) { ./ it: Int
thread { ./ new thread
sleep(1000L)
println("Hello from thread $it!")
}
}
}
Coroutines - fibers - threads

fun main(): Unit {

repeat(1_000_000) { ./ it: Int
thread { ./ new thread
sleep(1000L)
println("Hello from thread $it!")
}
}
}

Exception in thread "main" java.lang.OutOfMemoryError: unable to create native thread: possibly out
of memory or process/resource limits reached.
Coroutines - fibers - threads

fun main(): Unit = runBlocking {

repeat(1_000_000) { ./ it: Int
launch { ./ new asynchronous activity
delay(1000L)
println("Hello from coroutine $it!")
}
}
}

Coroutines are like light-weight threads.

Inside Coroutine Scope
Thread switching problem
An important non-guarantee

It is not guaranteed that the coroutine is going to be resumed on the same thread, so be very careful about calling suspending
function while holding any monitor.

val lock = ReentrantLock()

suspend fun russianRoulette() {

lock.lock()
pullTheTrigger()
lock.unlock()
}

Unlock might happen on another thread.

Murphy’s law: “Anything that can go wrong will go wrong.”
Then unlock will throw IllegalMonitorStateException.
Mutual Exclusion

Mutual Exclusion ..> Mutex.

val mutex = Mutex() ./ .lock() suspends, .tryLock() does not suspend

var counter = 0

suspend fun withMutex() {

repeat(1_000) {
launch {
./ protect each increment with lock
mutex.withLock { counter.+ }
}
}
println("Counter = $counter") ./ Guaranteed `1000`
}
Inside CoroutineScope
Exceptions
Exception handling

public interface CoroutineExceptionHandler : CoroutineContext.Element {

public companion object Key : CoroutineContext.Key<...>

public fun handleException(context: CoroutineContext, exception: Throwable)

}

● Children coroutines delegate handling to their parents.

● Coroutines running with SupervisorJob do not propagate exceptions to their parents.

● CancellationExceptions are ignored.

● If there is a Job in the context, then Job.cancel is invoked.

● All instances of CoroutineExceptionHandler found via ServiceLoader are invoked.

● The current thread’s Thread.uncaughtExceptionHandler is invoked.

Exception propagation

Job

launch SupervisorJob

launch launch launch

Exception propagation

Job

launch SupervisorJob

launch launch launch

exception
Exception propagation

Job

launch SupervisorJob

Check this out

launch launch launch

exception
Exception propagation

Job

launch SupervisorJob

No, thank you

launch launch launch

exception
Exception propagation

Job

launch SupervisorJob
handler

launch launch launch

Exception propagation

Job

launch SupervisorJob
handler

launch launch launch

exception
Exception propagation

Job

launch SupervisorJob
handler
Check this out

launch launch launch

exception
Exception propagation

Job
Check this out

launch SupervisorJob
handler

launch launch launch

exception
Exception propagation

Job Cancel

launch SupervisorJob
handler

launch launch launch

Exception propagation

Job

launch SupervisorJob
Cancel
handler

launch launch launch

Exception propagation

Job

launch SupervisorJob
handler

launch launch launch

Exception Propagation

Job

launch SupervisorJob
handler

launch launch launch

Now you see it
val jobs: List<Job> = List(1_000_000) {
launch(Dispatchers.Default + CoroutineName("#$it")
+ CoroutineExceptionHandler { context, error .>
println("${context[CoroutineName]..name}: $error")
},
CoroutineStart.LAZY
) {
delay(Random.nextLong(1000))
if (it % 10 .= 0) { throw Exception("No comments") }
println("Hello from coroutine $it!")
}
}

jobs.forEach { it.start() }

This exception handler is useless if this code is not inside a SupervisorJob.

Error handling

fun main() = runBlocking { ./ root coroutine

val job1 = launch {
delay(500)
throw Exception("Some jobs just want to watch the world burn")
}
val job2 = launch {
println("Going to do something extremely useful")
delay(10000)
println("I've done something extremely useful")
}
}

Exception in job1 .> propagate to parent .> job2 gets cancelled

Error handling

fun main() {
val scope = CoroutineScope(Dispatchers.Default + SupervisorJob())
with(scope) {
val job1 = launch {
throw Exception("Some jobs just want to watch the world burn")
}
val job2 = launch {
delay(3000)
println("I've done something extremely useful")
}
}
scope.coroutineContext[Job]..let { job .>
runBlocking { job.children.forEach { it.join() } }
} ./ `job1.join()` will throw, so `it.join()` should actually be in a `try/catch` block
}
Error handling
fun main() {
./ `someScope: CoroutineScope` already exists
someScope.launch { ./ this coroutine is a child of someScope
supervisorScope { ./ SupervisorJob inside
val job1 = launch {
throw Exception("Some jobs just want to watch the world burn")
}
val job2 = launch {
println("Going to do something extremely useful")
delay(3000)
println("I've done something extremely useful")
}
}
}
...
}
Error handling

fun main() {
val scopeWithHandler = CoroutineScope(CoroutineExceptionHandler {
context, error .> println("root handler called")
})
scopeWithHandler.launch {
supervisorScope {
launch { throw Exception() }
launch(CoroutineExceptionHandler { context, error .>
println("personal handler called")
}) { throw Exception() }
}
}
...
}

Exceptions are not propagated to parents meaning you can override the handler.
Inside CoroutineScope
Structured concurrency
Error handling (revisited)

fun processReferences(refs: List<Reference>) {

for (ref in refs) {
val location = ref.resolveLocation()
GlobalScope.launch {
val content = downloadContent(location)
processContent(content)
}
}
}

● Downloads are launched in the background.

● GlobalScope – This is a delicate API and its use requires care. Make sure you fully read and understand the documentation of
any declaration that is marked as a delicate API. The delicate part is that no Job is attached to the GlobalScope, making
its use dangerous and inconvenient.

● Any downloadContent or processContent crash results in a coroutines leak.

Structured concurrency

suspend fun processReferences(refs: List<Reference>) {

coroutineScope { ./ new scope with outer context, but a new Job
for (ref in refs) {
val location = ref.resolveLocation()
launch { ./ child of the coroutineScope above
val content = downloadContent(location)
processContent(content)
}
}
}
}

In the event of a downloadContent or processContent crash, the exception goes to the coroutineScope, which stores links
to all child coroutines and will cancel them. This is an example of structured concurrency, a concept that is not present in
threads.
A helpful convention
This function takes a long time and waits for something:
suspend fun work(...) { ... }

This function launches more background work and quickly returns:

fun CoroutineScope.backgroundWork(...) {
launch { ... }
}

or:
fun CoroutineScope.moreWork(...): Job = launch { ... }

Not:
suspend fun CoroutineScope.dontDoThisPlease()
A helpful convention

fun CoroutineScope.processReferences(refs: List<Reference>) {

for (ref in refs) {
val location = ref.resolveLocation()
launch { ./ child of coroutineScope
val content = downloadContent(location)
processContent(content)
}
}
}
Inside CoroutineScope
Coroutine cancellation
Cancelling coroutines
val job = launch(Dispatchers.Default) {
repeat(5) {
println("job: I'm sleeping $it...")
Thread.sleep(500) ./ simulate blocking work
}
}
yield() ./ lets the childJob work
println("main: I'm tired of waiting!")
job.cancel() ./ cancels the `job`
job.join() ./ waits for `job`'s completion
println("main: Now I can quit.")

● The coroutine (job) does not know that somebody is trying to cancel it.

● Cancellation is cooperative.
Cancelling coroutines
val job = launch(Dispatchers.Default) {
repeat(5) {
try {
println("job: I'm sleeping $it...")
delay(500)
} catch (e: CancellationException) {
println("job: I won't give up $it")
}
}
}
yield()
println("main: I'm tired of waiting!")
job.cancelAndJoin() ./ cancel + join
println("main: Now I can quit.")
Cancelling coroutines

val job = launch(Dispatchers.Default) {

var i = 0
while (isActive .& i < 5) { ./ check Job status
println("job: I'm sleeping ${i.+}...")
Thread.sleep(500)
}
}
delay(1300L)
println("main: I'm tired of waiting!")
job.cancelAndJoin()
println("main: Now I can quit.")
Cancelling coroutines

val job = launch {

try {
repeat(1_000) {
println("job: I'm sleeping $it...")
delay(500L)
}
} finally {
withContext(NonCancellable) {
println("job: I'm running finally")
delay(1000L)
println("job: Delayed for 1 sec thanks to NonCancellable")
}
}
}
...
job.cancelAndJoin()
Channels
Communicating sequential processes

Channel is like BlockingQueue, but with suspending calls instead of blocking ones.

● Blocking put → suspending send

● Blocking take → suspending receive

● No shared mutable state!

● Channels are still experimental

public interface Channel<E> : SendChannel<in E>, ReceiveChannel<out E> {

... suspend fun send(element: E)
... suspend fun recieve(): E
...
}

There are also trySend and alike that do not wait.

Practice

fun main() = runBlocking {

val channel = Channel<Int>()
launch {
for (x in 1..5)
channel.send(x * x)
}
repeat(5) {
println(channel.receive())
}
println("Done!")
}
Prime Numbers

fun CoroutineScope.numbersFrom(start: Int) = produce<Int> {

var x = start
while (true) send(x.+) ./ infinite stream of integers from start
}

fun CoroutineScope.filter(numbers: ReceiveChannel<Int>, prime: Int) =

produce<Int> { for (x in numbers) if (x % prime .= 0) send(x) }

fun main() = runBlocking {

var cur = numbersFrom(2)
repeat(10) {
println(cur.receive())
cur = filter(cur, prime)
}
coroutineContext.cancelChildren()
}
Fan-in and fan-out

fun <T> CoroutineScope.production(ch: SendChannel<T>, msg: T) =

launch { while (true) { delay(Random.nextLong(23)); ch.send(msg) } }

fun <T> CoroutineScope.processing(ch: ReceiveChannel<T>, name: String) =

launch { for (msg in ch) { println("$name: received $msg") } }

fun main() = runBlocking {

val channel = Channel<String>()
listOf("foo", "bar", "baz").forEach { production(channel, it) }
repeat(8) { processing(channel, "worker #$it") }
delay(700)
coroutineContext.cancelChildren(CancellationException("Enough!"))
}
Details

● Channels are still experimental.

● Channels are fair, meaning that send and receive calls are served in a first-in first-out order.

● By default, channels have RENDEZVOUS capacity: no buffer at all. This behavior can be tweaked:
The user can specify buffer capacity, what to do when buffer overflows, and what to do with
undelivered items.
Select (experimental!

suspend fun selector(

channel1: ReceiveChannel<String>,
channel2: ReceiveChannel<String>
): String = select<String> {
./ onReceive clause in select fails when the channel is closed
channel1.onReceive { it: String .> "b .> '$it'" }
channel2.onReceiveCatching { it: ChannelResult<String> .>
val value = it.getOrNull()
if (value .= null) {
"a .> '$value'"
} else {
"Channel 'a' is closed" ./ Select does not stop!
}
}
}
Miscellaneous
 Miscellaneous
Beyond asynchronous programming
Sequences
val fibonacci = sequence { ./ A coroutine builder!
var cur = 1
var next = 1
while (true) {
yield(cur) ./ A suspending call!
cur += next
next = cur - next
}
}

val iter = fibonacci.iterator() ./ nothing happens yet

println(iter.next()) ./ process up to the first yield .> 1
println(iter.next()) ./ wake up and continue .> 1
println(iter.next()) ./ 2 and then to infinity and beyond
Miscellaneous
Under the hood: advanced
Under the hood

Remember this code?

suspend fun postItem(item: Item) {

val token = preparePost()
val post = submitPost(token, item)
processPost(post)
}

Now that we know much more, let’s get a better approximation of what’s going on under the hood.
Under the hood

fun postItem(item: Item, completion: Continuation<Any?>) {

class PostItemStateMachine(
completion: Continuation<Any?>?,
context: CoroutineContext?
): ContinuationImpl(completion) {
var result: Result<Any?> = Result(null)
var label: Int = 0

var token: Token? = null

var post: Post? = null
...
}
}
Under the hood

fun postItem(item: Item, completion: Continuation<Any?>) {

class PostItemStateMachine(...): ... {

...
override fun invokeSuspend(result: Result<Any?>) {
this.result = result
postItem(item, this)
}
}

val continuation = completion as? PostItemStateMachine .: PostItemStateMachine(completion)

...
}
Under the hood
...
when(continuation.label) {
0 .> { ... }
1 .> {
continuation.token = continuation.result.getOrThrow() as Token
continuation.label = 2
submitPost(continuation.token.!, continuation.item.!, continuation)
}
2 .> { ... }
3 .> {
continuation.finalResult = continuation.result.getOrThrow() as FinalResult
continuation.completion.resume(continuation.finalResult.!)
}
else .> throw IllegalStateException(...)
}
...
 More
Continuation as generic callback
Continuation
Here’s a refresher on what Continuation looks like:
public interface Continuation<in T> {
public val context: CoroutineContext
public fun resumeWith(result: Result<T>)
}

We are given:
suspend fun suspendAnswer() = 42
suspend fun suspendSqr(x: Int) = x * x

How can we run suspendSqr(suspendAnswer) without kotlinx.coroutines?

Continuation
Continuation is a generic callback, so we can go back to the continuation passing style:

fun main() {
.:suspendAnswer.startCoroutine(object : Continuation<Int> {
override val context: CoroutineContext
get() = CoroutineName("Empty Context Simulation")

override fun resumeWith(result: Result<Int>) {

val prevResult = result.getOrThrow()
.:suspendSqr.startCoroutine(
prevResult,
Continuation(CoroutineName("Only name Context")) {
it: Result<Int> .> println(it.getOrNull())
}
)
} ./ Oh no!
}) ./ Closing brackets are coming!
} ./ Please help! I am being dragged into Callback Hell!!!
 Miscellaneous
To wrap existing async code or to implement your
own?
To wrap existing async code or to
implement your own?
suspend fun AsynchronousFileChannel.aRead(b: ByteBuffer, p: Int = 0) =
./ Scheme: call-with-current-continuation; call/cc
suspendCoroutine { cont .>
./ CompletionHandler ~ Continuation
read(b, p.toLong(), Unit, object : CompletionHandler<Int, Unit> {
override fun completed(bytesRead: Int, attachment: Unit) {
cont.resume(bytesRead)
}

override fun failed(exception: Throwable, attachment: Unit) {

cont.resumeWithException(exception)
}
})
}
To wrap existing async code or to
implement your own?
fun main() = runBlocking {
val readJob = launch(Dispatchers.IO) {
val fileName = ...
val channel = AsynchronousFileChannel.open(Paths.get(fileName))
val buf = ByteBuffer.allocate(...)
channel.use { ./ syntactic sugar for `try { ... } finally { channel.close() }`
while (isActive) {
... = it.aRead(buf)
...
}
}
}
...
}
To wrap existing async code or to
implement your own?
suspend fun cancellable(…) =
suspendCancellableCoroutine { cancellableCont .>
cancellableCont.invokeOnCancellation { throwable: Throwable? .>
./ release resources, etc.
...
}

...

cancellableCont.cancel(…)
}
Miscellaneous
async / await
async / await in Kotlin
async Task PostItem(Item item) {
Task<Token> tokenTask = PreparePost();
Post post = await SubmitPost(tokenTask.await(), item);
ProcessPost();
}

● async and await are keywords in C#.

● Awaiting does not block heavy OS thread.

● await is an explicit suspension point.

● await is a single function, but depending on its environment it can result in 2 different behaviours.

● The C# approach was a great inspiration for the Kotlin team when they were designing coroutines, as it was for Dart, TS,
JS, Python, Rust, C++...
async / await in Kotlin
fun CoroutineScope.preparePostAsync(): Deferred<Token> = async<Token> { ... }

suspend fun postItem(item: Item) {

coroutineScope {
val token = preparePost().await()
val post = submitPost(token, item).await()
processPost(post)
}
}

Deferred<T> : Job is a Job that we can get some result from. async is just another coroutine builder. You can write exactly
the same in Kotlin!

But why would you do that? This is not idiomatic Kotlin.

async / await in Kotlin

suspend fun postItemAsyncAwait(item: Item) {

coroutineScope {
val deferredToken = async { preparePost() }
./ some work
val token = deferredToken.await()
val deferredPost = async { submitPost(token, item) }
./ more work
val post = deferredPost.await()
processPost(post)
}
}
Miscellaneous
Coroutine builders
A zoo of them!

public fun CoroutineScope.launch(

context: CoroutineContext,
start: CoroutineStart,
block: suspend CoroutineScope.() .> Unit ./ suspend lambda
): Job
public fun <T> future(...): CompletableFuture<T> ./ jdk8/experimental
public fun <T> CoroutineScope.async(...): Deferred<T>
public fun <T> runBlocking(...): T ./ Avoid using it
public fun <E> CoroutineScope.produce(
context: CoroutineContext,
capacity: Int,
@BuilderInference block: suspend ProducerScope<E>.() .> Unit
): ReceiveChannel<E>

And many more! Like actor.

Actor

Actor ∼ coroutine + channel

./ Message types for counterActor – Command pattern

sealed class CounterMsg
./ one-way message to increment
object IncCounter : CounterMsg() counter
./ a request with a reply
class GetCounter(val response: CompletableDeferred<Int>) : CounterMsg()
Actor

./ This function launches a new counter actor

fun CoroutineScope.counterActor() = actor<CounterMsg> {
var counter = 0 ./ actor state
for (msg in channel) { ./ iterate over incoming messages
when (msg) {
is IncCounter .> counter.+
is GetCounter .> msg.response.complete(counter)
}
}
}

Frequently encapsulated into a separate class.

Miscellaneous
Android
Android

Check out developer.android.com to learn how coroutines are used (extensively) in modern
Android development.

class MyViewModel: ViewModel() {

init {
viewModelScope.launch { ... }
}
}

● A ViewModelScope is defined for each ViewModel in your app.

● A LifecycleScope is defined for each Lifecycle object.

● Flow, which is not covered in this lecture, is common in Android.

Further Reading

● github.com/Kotlin/KEEP/ – Kotlin design proposals, including coroutines

● kotlinlang.org – “Coroutines overview” and “Official libraries/kotlinx.coroutines”

● github.com/Kotlin/kotlinx.coroutines – A nicely documented resource

● Roman Elizarov’s talks on YouTube and posts on medium

● Flow<T> – Asynchronous Flow

● Kotlin sources at github.com/JetBrains/kotlin/

● All of the code from this presentation can be found in the corountines folder at
github.com/bochkarevko/kotlin-things/
Thanks!

@kotlin | Developed by JetBrains