I'm trying to understand kotlin coroutines, I'm coming from C# and there's something I'm not understanding here in kotlin. In this scenario I'm writing a webapi using Kotlin in the Quarkus framework. From what I can tell if I label a controller (or resource) function as a suspend function quarkus will automatically launch it in a coroutine.
The issue i have is i don't know what the preferred method for suspending that coroutine is. The vast majority of examples I see on kotlin coroutines use the delay() function, which internally uses suspendCancellableCoroutine() to suspend the function. That makes sense, but i don't see a lot of example calling suspendCancellableCoroutine() explicitly. I've done some reading about the underlying code that gets generated in a suspend function, and some resources lead me to believe that by virtue of calling another suspend function i'll hit a suspend point and that will suspend my coroutine. In C# i'd usually just call await() from inside my async function to execute the long running code.
In my kotlin setup i have setup an instance of jmeter and i simulate 5 threads calling my API at the same time, while limiting my program to run on a single thread in quarkus. My API then makes a call to another API (i'll call that API, data API from now on), which could be a long running operation. For the purpose of my test my data API has a 1 second sleep in it.
Essentially:
web api controller -> web api processing -> web api calls data api through client -> data API does slow operation
I've tried calling async/await on the call to the data API, which seems to work, JMeter reports that 5 requests are all completed in roughly 1 second, and the logging i have indicates that all 5 requests are handled on a single thread. This feels clunky though. I'm already in a coroutine and now my coroutine is creating a new coroutine (async is a coroutine builder) to execute the long running function.
I've also removed the async/await and updated the call to the data API to be a suspend function as well (though this is a client generated from resteasy client). This also seems to work, but resteasy reactive could be generating something that's doing the suspend for me. I need to work with a simpler example, but in the mean time...
If i'm not using the delay() function in Kotlin, and i'm executing code in a coroutine, what is the preferred method to indicate that a section of code is potentially blocking and my coroutine should be suspended? Do i launch a new coroutine? Call suspendCancellableCoroutine()? Or something else? Probably overthinking this, but i want to make sure i understand this.
The coroutines library provides several suspend functions you can use to suspend in a coroutine or in another suspend function, among them:
withContext
delay
coroutineScope
supervisorScope
suspendCoroutine
suspendCancellableCoroutine
Job.join
Deferred.await
The typical way to convert blocking (long-running synchronous) code into something you can use in a coroutine is to wrap it in withContext(Dispatchers.Default) { } or withContext(Dispatchers.IO) { }. If it's something you use repeatedly, you can write a suspend function for it:
suspend fun foo() = withContext(Dispatchers.IO) {
blockingFoo()
}
but if it's some one-off blocking chunk of code, you can use withContext directly in a coroutine.
Note, using async { }.await() is basically never done. The compiler warns you against it. You should be using withContext instead. Calling await on a Deferred is used either when one coroutine needs a result from some other coroutine that has been passed to it, or when you're working with multiple parallel children coroutines inside a coroutineScope block.
The typical way to convert asynchronous callback-based code into a suspend function so you can use it synchronously in a coroutine is to use suspendCoroutine or suspendCancellableCoroutine. You can look up how to use those. They are pretty low level. Many libraries like Retrofit and Firebase already provide suspend functions you can use instead of the callbacks.
coroutineScope and supervisorScope are for creating a scope inside your coroutine to run multiple children coroutines in parallel and wait for them all.
I have a personal project written in Kotlin, and I developed a habit of using withContext(...) very generously. I tend to use withContext(Dispatchers.IO) when calling anything that could possibly be related to I/O.
For example:
suspend fun getSomethingFromDatabase(db: AppDatabase) = withContext(Dispatchers.IO) {
return // ...
}
suspend fun doSomethingWithDatabaseItem(db: AppDatabase) {
val item = withContext(Dispatchers.IO) {
getSomethingFromDatabase(db)
}
// ...
}
You can see a redundant withContext(Dispatchers.IO) in the second function. I'm being extra cautious here, because I might not know/remember if getSomethingFromDatabase switches to an appropriate context or not. Does this impact performance? Is this bad? What's the idiomatic way of dealing with Dispatchers?
Note: I know that it's perfectly fine to switch between different contexts this way, but this question is specifically about using the same context.
You do not need withContext for anything besides calling code that demands a specific context. Therefore withContext(Dispatchers.Main) should only be used when you're working with UI functions that require the main thread. And you should only use withContext(Dispatchers.IO) when calling blocking IO related code.
A proper suspend function does not block (see Suspending convention section here), and therefore, you should never have to specify a dispatcher to call a suspend function. The exception would be if you're working with someone else's code or API and they are using suspend functions incorrectly!
I don't know what your AppDatabase class is, but if it is sensibly designed, it will expose suspend functions instead of blocking functions, so you should not need withContext to retrieve values from it. But if it does expose blocking functions for retrieving items, then the code of your first function is correct.
And your second function definitely doesn't need withContext because it's simply using it to call something that I can see is a suspend function.
As for whether it's OK to use redundant context switching...it doesn't hurt anything besides possibly wasting a tiny bit of time and memory context switching and allocating lambdas for no reason. And it makes your code less readable.
My question is rather theoretical.
I am quite new to kotlin (only passed the tutorial, didn't write any real code).
Reading through the language reference I find myself confused about the fact that "suspend" is a keyword, yet I can't find anything like "launch" in the list of keywords. That makes me think that there is some asymmetry - the "suspend" is a compiler feature, yet "launch" is a library function. Is my understanding correct? If so - wouldn't it have been better to implement both as library features or both as compiler features?
I always thought that you can always write your own standard library using the available language features, but I still can't see if this really applies to this case.
TL;DR: Can I start a coroutine using pure kotlin, without importing any libraries whatsoever (however ugly that would be)?
The suspend marker adds a hidden continuation parameter to the function signature and completely changes the implementation bytecode. Suspension points don't boil down to helper function calls, they turn your linear program code into a state machine, the state being kept in the continuation object. The resulting bytecode isn't even representable as Java program code.
As opposed to that, launch is just regular library code that builds upon the suspend/resume primitive.
#Alexey Soshin's isn't quite correct.
You can use coroutines w/o the library, and it's pretty easy. Here is a about the simplest suspending coroutine example that has 0 dependency on the coroutine library.
import kotlin.coroutines.*
fun main() {
lateinit var context: Continuation<Unit>
suspend {
val extra="extra"
println("before suspend $extra")
suspendCoroutine<Unit> { context = it }
println("after suspend $extra")
}.startCoroutine(
object : Continuation<Unit> {
override val context: CoroutineContext = EmptyCoroutineContext
// called when a coroutine ends. do nothing.
override fun resumeWith(result: Result<Unit>) {
result.onFailure { ex : Throwable -> throw ex }
}
}
)
println("kick it")
context.resume(Unit)
}
This runs fine on the play.kotlinlang.org site.
As you can see from this code, any lambda decorated with suspend has the startCourtine() on it.
In fact, I think the SequenceBuilder() from the standard collection classes uses a simple coroutine like this to generate the sequence, with no dependency on the coroutine library.
The compiler is doing the heavy lifting on the coroutines, splitting the code into different "methods" at each possible suspending point. Look at the java code for this, and you'll see it's "split" into a switch statement. one case before the suspend, and another after.
The library does a ton of nice stuff for you..... and it's likely you'll almost always use it (cuz why not?) but you don't actually need it.
Can I start a coroutine using pure kotlin, without importing any libraries whatsoever (however ugly that would be)?
No. All coroutine generators are inside kotlinx.coroutines library, so you'll need at least that. Now, very theoretically, you could reimplement this functionality yourself. But probably you shouldn't.
How this can be done is a bit too long for a StackOverflow answer, but try invoking method of this Kotlin class from Java:
class AsyncWorks {
suspend fun print() {
println("Hello")
}
}
You'll see that although Kotlin method has no arguments, in Java it requires Continuation<? super Unit>. This is what suspend keyword does. It adds Continuation<T> as the last argument of our function.
wouldn't it have been better to implement both as library features or
both as compiler features?
Ideally, you'd want everything to be a "library feature", since it's easier to evolve. Removing a keyword from a language is very hard. In theory, having suspend as a keyword could be avoided. Quasar, being a framework, uses annotations instead. Go programming language, on the other hand, assumes all functions are suspendable. All those approaches have their advantages and disadvantages.
Kotlin decided to be pragmatic, and add suspend keyword, leaving the decision on the developers. If you're interested in the topic, I highly recommend this talk by Roman Elizarov, author of Kotlin coroutines, that explains their decissions: https://www.youtube.com/watch?v=Mj5P47F6nJg
Answering my own question here.
After a year of Kotlin I tend to think that this IS indeed possible.
The suspend language feature creates an extra class and instantiates it every time your suspend function is called. This class extends ContinuationImpl and stores the progress of your coroutine - to which point it was able to execute so far.
Therefore one will need to write a custom dispatcher that would be able to manage the queue of the continuation objects to decide which one has to run now and a launch function that would take the newly created continuation object and pass it over to the dispatcher.
Now, this is still an asymmetry - the ContinuationImpl lives in kotlin.coroutines.jvm.internal so the compiler assumes this package exists. If one really wants to drop the standard library altogether - he'll need to implement that package to be able use the suspend keyword.
I'm not a kotlin expert though, so I might be wrong.
Because coroutines are valid for use cases that don't support launch. Because suspend requires some specific support from the compiler and launch doesn't if you already have suspend. Because structured concurrency is a library framework on top of the language feature, and launch is a part of that specific framework, that makes specific choices on top of what the language requires.
Starting a coroutine without any libraries can be done with startCoroutine. kotlin.coroutines is part of Kotlin, not a library.
I'm starting with coroutines in Android app. I'm rewriting callbacks to suspendCoroutine<> {} and I've got one dillema: when should I just mark the function as suspend, and when should I wrap the call in some builder (launch, async, etc.)?
Is there some best practice, rule of thumb, or something?
You should write a suspend fun for every asynchronous, callback-based API call you're currently making.
You should wrap in withContext(myThreadPool) every synchronous API call you're making.
All Android-friendly APIs that do I/O use the async approach, so for these you'll be writing suspend funs, but for CPU-intensive tasks (such as decoding images) you may need withContext.
Finally, to be able to use either kind of calls, you must create a top-level coroutine with launch(UI).
Keep in mind that the above is really just a rule of thumb. When you factor your code, many times you realize you need, for example, a suspend fun to make a CPU-intensive operation because there's a withContext somewhere on that call path.
Let me also add a warning about a very typical misuse of the coroutine APIs: you almost never need async-await. Use it only for the cases where you want to truly run it "in the background" while you continue to perform other stuff in your current context. In simpler terms, you should never write
val result = async { calculation() }.await()
Instead you should write
val result = withContext(myThreadPool) { calculation() }
Why would I want to use Kotlin's coroutines?
It seems that the RxKotlin library is much more versatile.
Kotlin's coroutines look significantly less powerful and more cumbersome to use in comparison.
I base my opinion on coroutines on this design talk by Andrey Breslav (JetBrains)
Slideshow from the talk is accessible here.
EDIT (thanks to #hotkey):
Better source on the current state of coroutines here.
Disclaimer: Parts of this answer are irrelevant since Coroutines now have the flow API, very similar to Rx one. If you want an up-to-date answer, jump to the last edit.
There is two parts in Rx; the Observable pattern, and a solid set of operators to manipulate, transform and combine them. The Observable pattern, by itself, doesn't do a lot. Same with Coroutines; it's just another paradigm to deal with asynchronism. You can compare the pro/cons of callbacks, Observable and coroutines to solve a given problem, but you can't compare a paradigm with a fully featured library. It's like comparing a language with a framework.
How Kotlin coroutines are better than RxKotlin ? Didn't used coroutines yet, but it's look similar to async/wait in C#. You just write sequential code, everything is as easy as writing synchronous code ... except it execute asynchronously. It's easier to grasp.
Why would I want to use kotlin coroutines ? I will answer for myself. Most of the time I will stick to Rx, because I favor event-driven architecture. But should arise the situation where I am writing sequential code, and I need to call an asynchronous method in the middle, I will happily leverage coroutines to keep it that way and avoiding wrapping everything in Observable.
Edit: Now that I am using coroutines it's time for an update.
RxKotlin is just syntactic sugar to use RxJava in Kotlin, so I will speak about RxJava and not RxKotlin in the following. Coroutines are a lower lever and more general concept than RxJava, they serve others use-cases. That said, there is one use-case where you could compare RxJava and coroutines (channel), it's passing around data asynchronously. Coroutines have a clear advantage over RxJava here:
Coroutines are better to deal with resources
In RxJava you can assign computations to schedulers but subscribeOn() and ObserveOn()are confusing. Every coroutine is given a thread context and return to parent context. For a channel, both side (producer, consumer) execute on his own context. Coroutines are more intuitive on thread or thread pool affectation.
Coroutines give more control on when those computation occur. You can for example pass hand (yield), prioritize (select), parallelize (multiple producer/actor on channel) or lock resource (Mutex) for a given computation. It may not matter on server (where RxJava came first) but on resources limited environment this level of control may be required.
Due to it's reactive nature, backpressure doesn't fit well in RxJava. In the other end send() to channel is a suspensive function that suspend when channel capacity is reached. It's out-of-the-box backpressure given by nature. You could also offer() to channel, in which case the call never suspend but return false in case the channel is full, effectively reproducing onBackpressureDrop() from RxJava. Or you could just write your own custom backpressure logic, which won't be difficult with coroutines, especially compared to do the same with RxJava.
There is another use-case, where coroutines shine and this will answer your second question "Why would I want to use Kotlin coroutines?". Coroutines are the perfect replacement for background threads or AsyncTask (Android). It's as easy as launch { someBlockingFunction() }. Of course you could achieve this with RxJava too, using Schedulers and Completable perhaps. You won't (or little) use the Observer pattern and the operators which are the signature of RxJava, a hint that this work is out of scope for RxJava. RxJava complexity (a useless tax here) will make your code more verbose and less clean than Coroutine's version.
Readability matters. On this regard, RxJava and coroutines approach differ a lot. Coroutines are simpler than RxJava. If you are not at ease with map(), flatmap() and functional reactive programming in general, coroutines manipulations are easier, involving basics instructions: for, if, try/catch ... But I personally find coroutine's code harder to understand for non-trivial tasks. Especially it involves more nesting and indentation whereas operator chaining in RxJava keep everything in line. Functional-style programming make processing more explicit. On top of that RxJava can solve complex transformations with a few standard operators from their rich (OK, way too rich) operator set. RxJava shine when you have complex data flows requiring a lot of combinations and transformations.
I hope those considerations will help you choose the right tool given your needs.
Edit: Coroutine now have flow, an API very, very similar to Rx. One could compare pro/cons of each, but the truth is the differences are minor.
Coroutines as it's core is a concurrency design pattern, with add-on libraries, one of those being a stream API similar to Rx. Obviously, Coroutines having a far broader scope than Rx, there is a lot of things that Coroutines can that Rx can't, and I can't list them all. But usually if I use Coroutines in one of my project it boil down to one reason:
Coroutines are better at removing callback from your code
I avoid using callback wich harm readability too much. Coroutines make asynchronous code simple and easy to write. By leveraging the suspend keyword, your code look like synchronous one.
I have seen Rx used in project mostly for the same purpose of replacing callback, but if you don't plan to modify your architecture to commit to the reactive pattern, Rx will be a burden. Consider this interface:
interface Foo {
fun bar(callback: Callback)
}
The Coroutine equivalent is more explicit, with a return type and the keyword suspend indicating it's an asynchronous operation.
interface Foo {
suspend fun bar: Result
}
But there is a problem with the Rx equivalent:
interface Foo {
fun bar: Single<Result>
}
When you call bar() in the callback or Coroutine version, you trigger the computation; with the Rx version, you get a representation of a computation that you can trigger at will. You need to call bar() then subscribing to the Single. Usually not a big deal, but it's a little confusing for beginner and can lead to subtle problem.
One exemple of such problems, suppose the callback bar function is implemented as such:
fun bar(callback: Callback) {
setCallback(callback)
refreshData()
}
If you don't port it properly, you will end with a Single that can be triggered only once because refreshData() is called in bar() function and not at subscription time. A beginner mistake, granted, but the thing is Rx is way more than a callback replacement and a lot of developers struggle to grasp Rx.
If your objective is to transform an asynchronous task from callback to a nicer paradigm, Coroutines are a perfect fit whereas Rx add some complexity.
I know RxJava very well and I've recently switched to Kotlin Coroutines and Flow.
RxKotlin is basically the same as RxJava, it just adds some syntactic sugar to make it more comfortable / idiomatic writing RxJava code in Kotlin.
A "fair" comparison between RxJava and Kotlin Coroutines should include Flow in the mix and I'm gonna try to explain why here. This is gonna be a bit long but I'll try to keep it as simple as I can with examples.
With RxJava you have different objects (since version 2):
// 0-n events without backpressure management
fun observeEventsA(): Observable<String>
// 0-n events with explicit backpressure management
fun observeEventsB(): Flowable<String>
// exactly 1 event
fun encrypt(original: String): Single<String>
// 0-1 events
fun cached(key: String): Maybe<MyData>
// just completes with no specific results
fun syncPending(): Completable
In kotlin coroutines + flow you do not need many entities cause if you do not have a stream of events you can just use simple coroutines (suspending functions):
// 0-n events, the backpressure is automatically taken care off
fun observeEvents(): Flow<String>
// exactly 1 event
suspend fun encrypt(original: String): String
// 0-1 events
suspend fun cached(key: String): MyData?
// just completes with no specific results
suspend fun syncPending()
Bonus: Kotlin Flow / Coroutines support null values (support removed with RxJava 2)
Suspending functions are what the name hint: they are function that can pause the execution of the code and resume it later when the function is completed; this allow you to write code that feels more natural.
What about the operators?
With RxJava you have so many operators (map, filter, flatMap, switchMap, ...), and for most of them there's a version for each entity type (Single.map(), Observable.map(), ...).
Kotlin Coroutines + Flow do not need that many operators, let's see why with some example on the most common operators
map()
RxJava:
fun getPerson(id: String): Single<Person>
fun observePersons(): Observable<Person>
fun getPersonName(id: String): Single<String> {
return getPerson(id)
.map { it.firstName }
}
fun observePersonsNames(): Observable<String> {
return observePersons()
.map { it.firstName }
}
Kotlin coroutines + Flow
suspend fun getPerson(id: String): Person
fun observePersons(): Flow<Person>
suspend fun getPersonName(id: String): String? {
return getPerson(id).firstName
}
fun observePersonsNames(): Flow<String> {
return observePersons()
.map { it.firstName }
}
You do not need an operator for the "single" case and it is fairly similar for the Flow case.
flatMap()
The flatMap operator and his siblings switchMap, contactMap exists to allow you to combine different RxJava object and thus execute potentially asynchronous code while mapping your events.
Say you need, for each person, to grab from a database (or remote service) it's insurance
RxJava
fun fetchInsurance(insuranceId: String): Single<Insurance>
fun getPersonInsurance(id: String): Single<Insurance> {
return getPerson(id)
.flatMap { person ->
fetchInsurance(person.insuranceId)
}
}
fun observePersonsInsurances(): Observable<Insurance> {
return observePersons()
.flatMap { person ->
fetchInsurance(person.insuranceId) // this is a Single
.toObservable() // flatMap expect an Observable
}
}
Let's see with Kotlin Coroutiens + Flow
suspend fun fetchInsurance(insuranceId: String): Insurance
suspend fun getPersonInsurance(id: String): Insurance {
val person = getPerson(id)
return fetchInsurance(person.insuranceId)
}
fun observePersonsInsurances(): Flow<Insurance> {
return observePersons()
.map { person ->
fetchInsurance(person.insuranceId)
}
}
Like before, with the simple coroutine case we do not need operators, we just write the code like we would if it wasn't async, just using suspending functions.
And with Flow that is NOT a typo, there's no need for a flatMap operator, we can just use map. And the reason is that map lambda is a suspending function! We can execute suspending code in it!!!
We don't need another operator just for that.
I cheated a bit here
Rx flatMap, switchMap and concatMap behave slightly differently.
Rx flatMap generate a new stream for each event and than merge them all together: the order of the new streams events you receive in the output is undetermined, it might not match the order or the events in input
Rx concatMap "fixes" that and guarantee you will get each new stream in the same order of your input events
Rx switchMap will instead dispose of any previously running stream when it gets a new events, only the last input received matter with this operator
So you see, it isn't true that Flow.map is the same, it is actually more similar to Rx concatMap, which is the more natural behavior you expect from a map operator.
But it is true you need less operators, inside map you can do any async operation you want and reproduce the behavior of flatMap because it is a suspendable function. The actual equivalent operator to RxJava flatMap is Flow.flatMapMerge operator.
The equivalent of the RxJava switchMap can be achieved in Flow by using the conflate() operator before the map operator.
For more complex stuff you can use the Flow transform() operator which for every event emit a Flow of your choice.
Every Flow operator accept a suspending function!
In the previous paragraph I told you I cheated. But the key get away of what I meant by Flow do not need as many operators is that most operator's callbacks are suspending function.
So say you need to filter() but your filter need to perform a network call to know if you should keep the value or not, with RxJava you need to combine multiple operators with unreadable code, with Flow you can just use filter()!
fun observePersonsWithValidInsurance(): Flow<Person> {
return observerPersons()
.filter { person ->
val insurance = fetchInsurance(person.insuranceId) // suspending call
insurance.isValid()
}
}
delay(), startWith(), concatWith(), ...
In RxJava you have many operators for applying delay or adding items before and after:
delay()
delaySubscription()
startWith(T)
startWith(Observable)
concatWith(...)
with kotlin Flow you can simply:
grabMyFlow()
.onStart {
// delay by 3 seconds before starting
delay(3000L)
// just emitting an item first
emit("First item!")
emit(cachedItem()) // call another suspending function and emit the result
}
.onEach { value ->
// insert a delay of 1 second after a value only on some condition
if (value.length() > 5) {
delay(1000L)
}
}
.onCompletion {
val endingSequence: Flow<String> = grabEndingSequence()
emitAll(endingSequence)
}
error handling
RxJava have lot of operators to handle errors:
onErrorResumeWith()
onErrorReturn()
onErrorComplete()
with Flow you don't need much more than the operator catch():
grabMyFlow()
.catch { error ->
// emit something from the flow
emit("We got an error: $error.message")
// then if we can recover from this error emit it
if (error is RecoverableError) {
// error.recover() here is supposed to return a Flow<> to recover
emitAll(error.recover())
} else {
// re-throw the error if we can't recover (aka = don't catch it)
throw error
}
}
and with suspending function you can just use try {} catch() {}.
You can achieve ALL the RxJava error operators with a single catch operator because you get a suspending function.
easy to write Flow operators
Due to the coroutines powering Flow under the hood it is way easier to write operators. If you ever checked an RxJava operator you would see how hard it is and how many things you need to learn.
Writing Kotlin Flow operators is easier, you can get an idea just by looking at the source code of the operators that are already part of Flow here. The reason is coroutines makes it easier to write async code and operators just feels more natural to use.
As a bonus, Flow operators are all kotlin Extension Functions, which means either you, or libraries, can easily add operators and they will not feel weird to use (in RxJava observable.lift() or observable.compose() are needed to combine custom operators).
Upstream thread doesn't leak downstream
What does this even mean?
This explain why in RxJava you have subscribeOn() and observeOn() while in Flow you only have flowOn().
Let's take this RxJava example:
urlsToCall()
.switchMap { url ->
if (url.scheme == "local") {
val data = grabFromMemory(url.path)
Flowable.just(data)
} else {
performNetworkCall(url)
.subscribeOn(Subscribers.io())
.toObservable()
}
}
.subscribe {
// in which thread is this call executed?
}
So where is the callback in subscribe executed?
The answer is:
depends...
if it comes from the network it's in an IO thread; if it comes from the other branch it is undefined, depends on which thread is used to send the url.
If you think about it, any code you write: you don't know in which thread it is gonna be executed: always depends on the caller. The issue here is that the Thread doesn't depends on the caller anymore, it depends on what an internal function call does.
Suppose you have this plain, standard code:
fun callUrl(url: Uri) {
val callResult = if (url.scheme == "local") {
grabFromMemory(url.path)
} else {
performNetworkCall(url)
}
return callResult
}
Imagine not having a way of knowing in which thread the line return callResult is executed in without looking inside grabFromMemory() and performNetworkCall().
Think about that for a second: having the thread change based on which function you call and what they do inside.
This happens all the time with callbacks APIs: you have no way of knowing in which thread the callback you provide will be executed unless documented.
This is the concept of "upstream thread leaking downstream".
With Flow and Coroutines this is not the case, unless you explicitly require this behavior (using Dispatchers.Unconfined).
suspend fun myFunction() {
// execute this coroutine body in the main thread
withContext(Dispatchers.Main) {
urlsToCall()
.conflate() // to achieve the effect of switchMap
.transform { url ->
if (url.scheme == "local") {
val data = grabFromMemory(url.path)
emit(data)
} else {
withContext(Dispatchers.IO) {
performNetworkCall(url)
}
}
}
.collect {
// this will always execute in the main thread
// because this is where we collect,
// inside withContext(Dispatchers.Main)
}
}
}
Coroutines code will run in the context that they have been executed into. And only the part with the network call will run on the IO thread, while everything else we see here will run on the main thread.
Well, actually, we don't know where code inside grabFromMemory() will run, but we don't care: we know that it will be called inside the Main thread, inside that suspending function we could have another Dispatcher being used, but we know when it will get back with the result val data this will be again in the main thread.
Which means, looking at a piece of code, it's easier to tell in which thread it will run, if you see an explicit Dispatcher = it's that dispatcher, if you do not see it: in whatever thread dispatcher the suspension call you are looking at is being called.
Structured Concurrency
This is not a concept invented by kotlin, but it is something they embraced more than any other language I know of.
If what i explain here is not enough for you read this article or watch this video.
So what is it?
With RxJava you subscribe to observables, and they give you a Disposable object.
You need to take care of disposing of it when it's not needed anymore. So what you usually do is keep a reference to it (or put it in a CompositeDisposable) to later call dispose() on it when it's not needed anymore. If you don't the linter will give you a warning.
RxJava is somewhat nicer than a traditional thread. When you create a new thread and execute something on it, it's a "fire and forget", you do not even get a way to cancel it: Thread.stop() is deprecated, harmful, and recent implementation actually do nothing. Thread.interrupt() makes your thread fail etc.. Any exceptions goes lost.. you get the picture.
With kotlin coroutines and flow they reverse the "Disposable" concept. You CANNOT create a coroutine without a CoroutineContext.
This context define the scope of your coroutine. Every child coroutine spawned inside that one will share the same scope.
If you subscribe to a flow you have to be inside a coroutine or provide a scope too.
You can still keep reference of the coroutines you start (Job) and cancel them. This will cancel every child of that coroutine automatically.
If you are an Android developer they give you these scopes automatically. Example: viewModelScope and you can launch coroutines inside a viewModel with that scope knowing they will automatically be cancelled when the viewmodel is cleared.
viewModelScope.launch {
// my coroutine here
}
Some scope will terminate if any children fail, some other scope will let each children leave his own lifecycle without stopping other children if one fails (SupervisedJob).
Why is this a good thing?
Let me try to explain it like Roman Elizarov did.
Some old programming language had this concept of goto which basically let you jump from one line of code to another at will.
Very powerful, but if abused you could end up with very hard to understand code, difficult to debug and reason upon.
So new programming languages eventually completely removed it from the language.
When you use if or while or when it is way easier to reason on the code: doesn't matter what happens inside those blocks, you'll eventually come out of them, it's a "context", you don't have weird jumps in and out.
Launching a thread or subscribing to an RxJava observable is similar to the goto: you are executing code which than will keep going until "elsewhere" is stopped.
With coroutines, by demanding you provide a context/scope, you know that when your scope is over everything inside that coroutines will complete when your context completes, doesn't matter if you have a single coroutines or 10 thousands.
You can still "goto" with coroutines by using GlobalScope, which you shouldn't for the same reason you shouldn't use goto in languages that provides it.
Cold vs Hot - ShareFlow and StateFlow
When we work with reactive streams we always have this concept of Cold and Hot streams. Those are concepts on both the Rx world and Kotlin Flows
Cold streams are just like a function in our code: it's there and does nothing until you call it. With a Flow that means it is defined what the stream does but it will do nothing until you start to collect on it. And, like a function, if you collect (call) it twice the stream will runs twice. (ex. a cold stream to perform an http request will execute the request twice if collected twice).
Hot streams do not work like that. When you have multiple collect call on them they all share the same Hot stream under the hood, which means your hot streams runs once and you can have multiple observers.
You can usually turn a Cold stream into an Hot streams with some operator.
On RxJava you can use this concept of Connectable Observable/Flowable.
val coldObservable: Observable<Something> = buildColdObservable()
// create an hot observable from the cold one
val connectableObservable: ConnectableObservable<Something> = coldObservable.publish()
// you can subscribe multiple times to this connectable
val subADisposable: Disposable = connectableObservable.subscribe(subscriberA)
val subBDisposable: Disposable = connectableObservable.subscribe(subscriberB)
// but nothing will be emitted there until you call
val hotDisposable: Disposable = connectableObservable.connect()
// which actually run the cold observable and share the result on bot subscriberA and subscriberB
// while it's active another one can start listening to it
val subCDisposable: Disposable = connectableObservable.subscribe(subscriberC)
You than have other helpful operators like refCount() or autoConnect() which turn back the Connectable into a standard stream and under the hood automatically .connect() when the first subscriber is attached.
buildColdObservable()
.replay(1) // when a new subscriber is attached receive the last data instantly
.autoConnect() // keep the cold observable alive while there's some subscriber
On Flow you have the shareIn() and the stateIn() operators. You can see the API design here. They are less "manual" in handling when you "connect".
buildColdFlow()
.shareIn(
// you need to specify a scope for the cold flow subscription
scope = myScope,
// when to "connect"
started = SharingStarted.WhileSubscribed(),
// how many events already emitted should be sent to new subscribers
replay = 1,
)
scope
The scope is for structured concurrency. On RxJava it's the connect() operation that actually subscribe to the cold observable, it gives you a Disposable you will have to call .dispose() on somewhere. If you use refCount() or autoConnect() it is called on the first subscriber and with refCount() is never disposed while with autoConnect() is disposed when there aren't any more subscribers.
With Flow you need to give a dedicated Scope to collect the cold stream, if you cancel that scope the cold stream will stop emitting and will not be usable anymore.
started
So this one is easy
RxJava refCount() --> Flow SharingStarted.Lazily, starts collecting on the first subscriber
RxJava autoConnect() -> Flow SharingStarted.WhileSubscribed(), starts collecting on the first subscriber and cancel it when there aren't anymore
RxJava call connect() manually before any subscription -> Flow SharingStarted.Eagerly(), starts collecting immediately
The WhileSubscribed() has useful parameters, check them out.
You can also define your own logic for SharingStarted to handle when collecting from the coldFlow.
Behavior and backpressure
When you have an hot observable you always have backpressure issues to deal with. 1 source of data being listened by many means one listener can be slower then others.
Flow .shareIn collect the cold stream in a dedicated coroutine and buffer emission by default. It means if the cold stream emit too fast it will use the buffer. You can change this behavior.
Kotlin SharedFlow also let you access the replay buffer directly to inspect previous emission if you need to.
Cancelling a subscriber will have no effect on the shared flow.
using flowOn() to change the Dispatcher on the subscriber will have no effect on the shared flow (use flowOn() before sharing if you need to run the cold stream in some specific dispatcher)
stateIn
Flow has a "special" version of ShareFlow that is called StateFlow and you can use stateIn() to create one from another stream.
A StateFlow always have 1 value, it cannot be "empty", so you need to provide the initial value when you do stateIn().
A StateFlow can never throw exceptions and can never terminate (in this way is similar to BehaviorRelay in the RxRelay library)
A StateFlow will only emit if the state change (it's like it has a build in distinctUntilChanged().
RxJava Subjects vs Mutable*Flow
A Subject in RxJava is a class that you can use to manually push your data on it while still using it as a stream.
In Flow you can use MutableSharedFlow or MutableStateFlow to achieve a similar effect.
With Kotlin coroutines you can also use Channels but they are considered somewhat a lower level API.
Any Drawback?
Flow is still in development and some features available in RxJava might be marked experimental in Kotlin Coroutines Flow or have some difference here and there.
Some niche operator or operator function might not be yet implemented and you might have to implement it yourself (at least it's easier).
But other than that there aren't any drawbacks I know of.
However there are differences to be aware of that could cause some frictions in switching from RxJava and needs you to learn new things.
Structured concurrency is a step forward, but introduces new concept you need to learn and get used to (scopes, supervisorJob): cancellation is handled completely different.
There's some gotcha to be aware of.
Gotcha: Cancellation Exception
If you cancel() job in a coroutine or throw CancellationException() the exception is propagated to parent coroutines unless you used a Supervisor scope / job.
The parent coroutine also cancel sibling coroutines of the one that got canceled if that happens.
BUT if you catch(e: Exception), even using runCatching {}, you must remember to rethrow CancellationException() otherwise you'll have unexpected results cause the coroutine has been canceled but your code is still trying to execute like it wasn't.
Gotcha: UncaughtExceptionHandler
if you do launch { ... } to create a new coroutine and that coroutine throws, by default, that will terminate the coroutine but will not crash the app and you might completely missed something went wrong.
This code will not crash your app.
launch {
throw RuntimeException()
}
In some cases it might not even print anything in the log.
If it was a cancellation exception it will definitely NOT print anything in the log.
Kotlin coroutines are different from Rx. It is hard to compare them apples-to-apples, because Kotlin coroutines are a thin language feature (with just a couple of basic concepts and a few basic functions to manipulate them), while Rx is a pretty heavy library with quite large variety of ready-to-use operators. Both are designed to address a problem of asynchronous programming, however their approach to solution is very different:
Rx comes with a particular functional style of programming that can be implemented in virtually any programming language without support from the language itself. It works well when the problem at hand easily decomposes into a sequence of standard operators and not so well otherwise.
Kotlin coroutines provide a language feature that let library writers implement various asynchronous programming styles, including, but not limited to functional reactive style (Rx). With Kotlin coroutines you can also write your asynchronous code in imperative style, in promise/futures-based style, in actor-style, etc.
It is more appropriate to compare Rx with some specific libraries that are implemented based on Kotlin coroutines.
Take kotlinx.coroutines library as one example. This library provides a set of primitives like async/await and channels that are typically baked into other programming languages. It also has support for light-weight future-less actors. You can read more in the Guide to kotlinx.coroutines by example.
Channels provided by kotlinx.coroutines can replace or augment Rx in certain use-cases. There is a separate Guide to reactive streams with coroutines that goes deeper into similarities and differences with Rx.
The talk/doc you linked does not talk about channels. Channels are what fill the gap between your current understanding of coroutines and event driven programming.
With coroutines and channels you can do event driven programming as you are probably used to do with rx, but you can do it with synchronous-looking code and without as many "custom" operators.
If you want to understand this better I suggest to look outside of kotlin, where those concepts are more mature and refined (not experimental). Look at core.async from Clojure, Rich Hickey videos, posts and related discussions.
http://discuss.purelyfunctional.tv/t/core-async-channels-vs-rx-observables/519/2
https://github.com/matthiasn/talk-transcripts/blob/master/Hickey_Rich/CoreAsync.md
Coroutines are designed to provide a lightweight asynchronous programming framework. Lightweight in terms of resources needed to start the async job. Coroutines don't enforce using an external API and are more natural for the users (programmers). In contrast, RxJava + RxKotlin has an additional data processing package that is not really needed in Kotlin which has a really rich API in the standard library for Sequences and Collections processing.
If you'd like to see more about the practical use of coroutines on Android I can recommend my article:
https://www.netguru.com/codestories/android-coroutines-%EF%B8%8Fin-2020
Coming to RxJava, RxJava streams are prone to leaks, where a stream continues to process items even when you no longer care. Kotlin coroutines use structured concurrency, which makes it much easier to manage the lifecycle of all your concurrent code.RxJava is, as it says on the tin, limited to Java. Coroutines work on any Kotlin-supported platform, so if we ever want to share asynchronous code between Android and iOS we could do it with coroutines.