raquo / airstream   16.0.0

MIT License GitHub

State propagation and event streams with mandatory ownership and no glitches

Scala versions: 3.x 2.13 2.12 2.11
Scala.js versions: 1.x 0.6

Airstream

Build status Chat on https://discord.gg/JTrUxhq7sj Maven Central

Airstream is a small state propagation and streaming library for Scala.js. Primary differences from other solutions:

  • Mandatory ownership of leaky resources – it is impossible to create a subscription without specifying when it shall be destroyed. This helps prevent memory leaks and unexpected behaviour.

  • No FRP glitches – neither observables themselves nor their observers will ever see inconsistent state within a transaction, at no runtime cost.

  • One integrated system for two core types of observables

    • EventStream for events (lazy, no current value)
    • Signal for state (lazy, has current value, only state-safe operators)
    • Seamless interop between the two types
  • Small size, simple implementation – easy to understand, easy to create custom observables. Does not bloat your Scala.js bundle size.

Airstream has a very generic design, but is primarily intended to serve as a reactive layer for unidirectional dataflow architecture in UI components. As such, it is not burdened by features that cause more problems than they solve in frontend development, such as backpressure and typed effects.

I created Airstream because I found existing solutions were not suitable for building reactive UI components. My original need for Airstream was to replace the previous reactive layer of Laminar, but I'll be happy to see it used by other reactive UI libraries as well. Another piece of Laminar you can reuse is Scala DOM Types.

"com.raquo" %%% "airstream" % "<version>"  // Requires Scala.js 1.13.0+

Table of Contents

Community

Documentation

API doc

The documentation provided below is a high level overview that occasionally dives into gritty details for things that are hard to figure out on your own. It is pretty comprehensive, but is not a full replacement to discovering every last available method by reading the code (which is quite simple, and has comments) or simply with an IDE's autocomplete functionality.

This documentation explains how Airstream works. Although I try to explain all required concepts and the rationale for doing things a certain way, if you need a primer on reactive programming using streams, consider this guide by André Staltz or its video adaptation.

This documentation is intended to be read top to bottom, sections further down the line assume knowledge of concepts and behaviours introduced in earlier sections.

For examples of Airstream usage, see laminar-examples, Laminar source code, as well as Airstream tests.

EventStream

EventStream is a reactive variable that represents a stream of discrete events.

EventStream has no concept of "current value". It is a stream of discrete events, and there is no such thing as a "current event".

EventStream is a lazy observable. That means that it will not receive or process events unless it has at least one Observer listening to it (more on this below).

Generally, when you add an Observer to a stream, it starts to send events to the observer from that point on.

The result of calling observable.addObserver(observer)(owner) or observable.foreach(onNext)(owner) is a Subscription. To remove the observer manually, you can call subscription.kill(), but usually it's the owner's job to do that. Hold that thought for now, read about owners later in the Ownership section.

Laziness

Before exploring Signals, the other kind of Observable, let's outline how exactly laziness works in Airstream. All Airstream Observables are lazy, but we will use EventStream-s here to make our explanation less abstract.

Every Observable has zero or more observers – both "external" observers that you add manually using addObserver or foreach methods, and InternalObserver-s representing dependant observables. More on those soon.

Starting Observables

When a stream acquires its first observer (does not matter if external or internal), it is said to be started. So when you call addObserver on a stream for the first time, you start the stream. Airstream will then call the stream.onStart method, which must ensure that this stream wakes up and starts working. Someone started observing (caring about the output of) this stream – and so the stream must ensure that the events start coming in.

Usually the stream accomplishes that by adding an InternalObserver to the parent (upstream) stream – the stream on which this one depends. For example, let's consider this scenario:

val foo: EventStream[Foo] = ???
val bar: EventStream[Bar] = foo.map(fooToBar)
val baz: EventStream[Baz] = bar.map(barToBaz)
val qux: EventStream[Qux] = baz.map(bazToQux)
val rap: EventStream[Rap] = qux.map(quxToRap)

baz.addObserver(bazObserver)

Until baz.addObserver(bazObserver) was called, these streams would not be receiving or emitting any events because they have no observers, internal or external. After baz.addObserver is called, an external observer bazObserver is added to baz, starting it. Then, baz.onStart is called, adding baz as an InternalObserver to bar. baz will now receive and process any events emitted by bar.

But this means that bar just got its first observer – even if an internal one, it still matters – someone started caring, even if indirectly. So its onStart method is called, and it adds bar as the first InternalObserver to foo. Now, foo is started as well, and its onStart method does something to ensure that foo will now start sending out events. We don't actually know what foo's onStart method does because we didn't define foo's implementation. For example, it could be adding a DOM listener to a DOM element.

Now we see how adding an observer resulted in a chain of activations of all upstream streams that were required, directly or indirectly, to get the events out of the stream we actually wanted to observe. The onStart method ensured – recursively – that the observed stream is now running.

Adding another observer to the now already running streams – foo or bar or baz – would not need to cause such a chain reaction because the stream to which it is being added already has observers (internal or external).

Lastly, notice that the qux and rap streams are untouched by all this. No one cares for their output yet, so those streams will not receive any events, and neither bazToQux nor quxToRap will ever run (well, not until we add observers that need them to run, directly or indirectly).

On a lower level, how exactly is it that qux will not run? Put simply, it needs to be getting events from baz to process them and produce its own events, but it's getting nothing from baz simply because at this point baz does not know that qux exists. baz sends out its events to all of its observers, but so far nothing added qux as an observer to baz.

For extra clarity, while the stream rap does depend on qux, rap itself has no observers, so it is stopped. Nothing started it yet, and so nothing triggered its onStart method which would have added rap as an InternalObserver to qux, starting qux recursively as described above.

Stopping Observables

Just like Observers can be added to streams, they can also be removed, e.g. with subscription.kill(). When you remove the last observer (internal or external) from a stream, the stream is said to be stopped. The same domino effect as when starting streams applies, except the onStop method recursively undoes everything that was done by onStart – instead of adding an InternalObserver to parent stream, we remove it, and if that causes the grand-parent stream to be stopped, we call its onStop method, and the chain continues upstream.

When the dust settles, streams that are now without observers (internal or external) will be stopped, and those that still have observers will otherwise be untouched, except they will stop referencing the now-stopped observables in their lists of internal observers.

Very often, an observable is started, used for a while, then stopped, and is discarded afterwards, never to be used again. However, Airstream does support restarting previously stopped observables, and Laminar has good use cases for that. Restarting Observables is an advanced topic that you can read about after you get a good understanding of other Airstream concepts like transactions.

Memory Management Implications

Every observable that depends on another – parent, or upstream observable, – always has a reference to that parent, regardless of whether it's started or stopped.

However, the parent/upstream observable has no references to its child/downstream observable(s) until the child observable is started. Only then does the parent obtain a reference to the child, adding it to the list of its internal observers.

This has straightforward memory management implications: nothing in Airstream is keeping references to stopped observables. So, if you don't have any of your own references to a stopped Observable, it will be garbage collected, as expected.

However, a started observable has additional references to it from:

  1. The parent/upstream observable on which this observable depends (via the parent's list of internal observers)
  2. The Subscription objects created by addObserver or foreach calls on this observable, if this observable has external observers. Those subscriptions are in turn referenced by their Owner-s (more on those later)

Remember that if a given observable is started, its parent is also guaranteed to be started, and so on. This creates a potentially long chain of observables that typically terminate with external observers on the downstream end, and some kind of event producer on the upstream end. All of these reference each other, directly or indirectly, and so will not be garbage collected unless there are no more references in your program to any observable or observer in this whole graph.

Now imagine that in the chain of activated observables mentioned above the most downstream observable is related to a UI component that has since then been destroyed. You would want that now-irrelevant observable to be stopped in order for it to be garbage collected, since it's not needed anymore, but it will continue to run for as long as it has its observer. And if you forgot to remove that observer when you destroyed the UI component it related to, you got yourself a memory leak.

This is a common memory management pattern for almost all streaming libraries out there, so this should come as no surprise to anyone familiar with event streams.

Some reactive UI libraries such as Outwatch give you a way to bind the lifecycle of subscriptions to the lifecycle of corresponding UI components, and that automatically kills the subscription (removes the observer) when the UI component it relates to is destroyed. However, the underlying streaming libraries that such UI libraries use have no concept of such binding, and so in those libraries you can manually call stream.addObserver and create a subscription that will not be automatically killed when the UI component that it conceptually relates to is unmounted.

What makes Airstream special is a built-in concept of ownership. When creating a leaky resource, e.g. when calling addObserver, you have to also provide a reference to an Owner who will eventually kill the subscription. For example, that owner could be a UI component to which the subscription relates, and it could automatically kill all subscriptions that it owns when it is destroyed, allowing the now-irrelevant observables to be stopped and garbage collected. This is essentially how Laminar's ReactiveElement works. For more details, see the Ownership section.

Signal

Signal is a reactive variable that represents a time-varying value, or an accumulated value. In other words, "state".

Similar to EventStream, Signal is lazy, so everything in the Laziness section applies to Signals as well.

Unlike EventStream, a Signal always has a current value. For instance, you could create a Signal by calling val signal = eventStream.startWith(initialValue). In that example, signal's current value would first equal to initialValue, and then any time eventStream emits a value, signal's current value would be updated to the emitted value, and then signal would emit this new current value.

However, all of that would only happen if signal had one or more observers (because of Laziness). If signal had no observers, its current value would be stuck at the last current value it saved while it had observers, or at initialValue if it never had observers.

When adding an Observer to a Signal, the observer will immediately receive the signal's current value, as well as any future values. If you don't want the observer to receive the current value, observe the stream signal.changes instead.

Note: Signal's initial value is evaluated lazily. For example:

val fooStream: EventStream[Foo] = ???
val fooSignal: Signal[Foo] = fooStream.startWith(myFoo)
val barSignal: Signal[Bar] = fooSignal.map(fooToBar)

In this example, barSignal's initial value would be equal to fooToBar(myFoo), but that expression will not be evaluated until it is needed (i.e. until barSignal acquires an observer). And once evaluated, it will not be re-evaluated again.

Similarly, myFoo expression will not be evaluated immediately as it is passed by name. It will only be evaluated if and when it is needed (e.g. to pass it down to an observer of barSignal).


Note: before Airstream 15, Signal only fired an event when its next value was different from its current value. The comparison was made using Scala's == operator. If you see references to "signals' == checks" in past issues / discussions, this is what they're talking about. In v15.0.0, this built-in auto-distinction filter is eliminated (see blog post), and you need to explicitly use one of the distinction operators to achieve such behaviour.

Getting Signal's current value

See relevant RFC: signal.peekNow()

Signal's laziness means that its current value might get stale / inconsistent in the absence of observers. Airstream therefore does not allow you to access a Signal's current value without proving that it has observers.

You can use stream.withCurrentValueOf(signal).mapN((lastStreamEvent, signalCurrentValue) => ???) to access signal's current value. The resulting stream will still be lazy, but this way the processing of currentValue is just as lazy as currentValue itself, so there is no risk of looking at a stale currentValue.

If you don't need lastStreamEvent, use stream.sample(signal).map(signalCurrentValue => ???) instead. Note: both of these output streams will emit only when stream emits, as documented in the code. If you want updates from signal to also trigger an event, look into the combineWith operator.

Note: withCurrentValueOf and sample operators are also available on signals, not just streams.

If you want to get a Signal's current value without the complications of sampling, or even if you just want to make sure that a Signal is started, just call observe on it. That will add a noop observer to the signal, and return an OwnedSignal instance which being a StrictSignal, does expose now() and tryNow() methods that safely provide you with its current value. However, you will need to provide an Owner to do that. More on those later.

Relationship between EventStream and Signal

Signals and EventStreams are distinct concepts with different use cases as described above, but both are Observables.

You can scanLeft(initialValue)(fn) an EventStream into a Signal, or make a Signal directly with stream.startWith(initialValue), or stream.startWithNone (which creates a "weak" signal, one that initially starts out with None, and has events wrapped in Some).

You can get an EventStream of changes from a Signal – signal.changes – this stream will re-emit whatever the parent signal emits (subject to laziness of the stream), minus the Signal's initial value.

If you have an observable, you can refine it to a Signal with Observable#toWeakSignal or Observable#toSignalIfStream(ifStream = streamToSignal), and to a Stream with Observable#toStreamOrSignal(ifSignal = signalToStream). For example, if you want to convert Observable[String] into Signal[String] with empty string as initial value in case this Observable is a stream, use observable.toSignalIfStream(_.startWith("")).

See also: Sources & Sinks

Observer

Observer[A] is a modest wrapper around an onNext: A => Unit callback that represents an external observer (see sections above for the distinction with InternalObserver-s). Observers have no knowledge of which observables, if any, they're observing, they have no power to choose whether they want to observe a given observable, etc.

Observers are intended to contain side effects, and to trigger evaluation of observables by their presence (remember, all Observables are lazy).

You usually create observers with Observer.apply or myObservable.foreach. There are a few more methods on Observer that support error handling.

Observers have a few convenience methods:

def contramap[B](project: B => A): Observer[B] – This is useful for separation of concerns. For example your Ajax service might expose an Observer[Request], but you don't want a simple UserProfile component to know about your Ajax implementation details (Request), so you can instead provide it with requestObserver.contramap(makeUpdateRequest) which is a Observer[User].

def filter(passes: A => Boolean): Observer[A] – useful if you have an Observable that you need to observe while filtering out some events (there is no Observable.filter method, only EventStream.filter).

def contramapSome is just an easy way to get Observer[A] from Observer[Option[A]]

def contracollect[B](pf: PartialFunction[B, A]): Observer[B] – when you want to both contramap and filter at once.

def contracollectOpt[B](project: B => Option[A]): Observer[B] – like contracollect but designed for APIs that return Options, such as NonEmptyList.fromList.

delay(ms: Int) – creates an observer that calls the original observer after the specified delay (for both events and errors)

Observer.combine[A](observers: Observer[A]) creates an observer that triggers all of the observers provided to it. Unlike Observer[A](nextValue => observers.foreach(_.onNext(nextValue))), the combined observer will also trigger its child observers in case of .onError (more about that in Error Handling).

Ownership

Alright, this is it. By now you've read enough to have many questions about how ownership works. This assumes you've read all the docs above, but to recap the core problem that ownership solves:

  • Adding an Observer to the lazily evaluated Observable is a leaky operation. That is, these resources will not be garbage collected even if the observable and the observer are both unreachable to user code. This is because the observable's parent observables will keep an internal reference to it for as long as it has observers.
  • Therefore, without Ownership you would have needed to remember to remove the observers that you added when the observers are no longer needed.
  • But doing that manually is insane, you will eventually forget and cause memory leaks and undesired behaviour. You should not need to take out your own garbage in a garbage collected language.

If any of the above does not make sense, the rest of this section might be confusing. Make sure you at least understand the entirety of the Laziness section before proceeding.

Without further ado:

Subscription is a resource that must be killed in order to release memory or prevent some other leak. You can get it by calling observable.addObserver(observer), writeBus.addSource(stream), or other similar methods that all take an implicit owner param.

Every Subscription has an Owner. An Owner is an object that keeps track of its subscriptions and knows when to kill them, and kills them when it's time (determined at its sole discretion). Airstream does not offer any concrete Owner classes (aside from the very basic ManualOwner), just the base trait. Unless you use Dynamic Ownership, you need to instantiate (and thus implement) your own Owner-s.

For example, until v0.8, my reactive UI library Laminar's ReactiveElement (a wrapper class for managing a JS DOM Element) used to implement Owner. When a ReactiveElement was discarded (unmounted from the DOM), it would kill all of its subscriptions, i.e. all the Subscriptions that were bound to its lifetime. That would remove the observers that those subscriptions installed on the observables, stopping them if they have no other observers. Note: Laminar switched to Dynamic Ownership in v0.8 (more on that later).

When creating a Subscription, you can perform whatever leaky operations you wanted, and just provide the cleanup method to perform any required cleanup.

Subscriptions are bound to a specific Owner upon the creation of the Subscription, and this link stays unchanged for the lifetime of the Subscription.

Subscriptions are normally killed by their Owner, but you can also .kill() the subscription manually. The Owner will be notified about this via owner.onKilledExternally(subscription) so that it can drop the reference to the killed subscription from its list.

Killing the same Subscription more than once throws an exception, don't do it.

Built-in Owner carefully tracks a list of its subscriptions, making sure to call the right hooks, and create and dispose the right references for memory management. If you extend Owner and change that logic, memory management of that owner and its subscriptions is on you. Generally you shouldn't need to mess with any of that logic though, just make sure to call killSubscriptions when it's time.

Ownership & Memory Management

In broad terms, ownership solves memory leaks by tying the lifecycle of Subscriptions which would be otherwise hard to track manually to the lifecycle of an Owner which is expected to be tracked automatically by a UI library like Laminar.

In practice, Airstream's memory management has no magic to it. It uses Javascript's standard garbage collection, same as the rest of your Scala.js code. You just need to understand what references what, and the documentation here explains it.

For example, a Subscription created by observable.addObserver method keeps references to both the Observable and the Observer (via the function passed as its cleanup param). That means that if you're keeping a reference to a Subscription, you're also keeping those references. Given that the Subscription has a kill method that lets you remove the observer from the observable, the presence of these references should be obvious. So like I said – no magic, you just need to internalize the basic ideas of lazy observables, just like you've already internalized the basic ideas of classes and functions.

OneTimeOwner

The basic Owner trait provides a high degree of flexibility, and therefore lacks some behaviour that you might expect in Owners.

For example, you can kill an Owner multiple times. Every time you do, its subscriptions will be killed, and the list of subscriptions cleared, but the Owner will remain usable after that, letting you add more subscriptions and kill them again later.

If you want an Owner that can only be killed once, and does not let you add subscriptions to it after it was killed, use the OneTimeOwner class instead. DynamicOwner below uses OneTimeOwner, and that is how Laminar provides element Owners that can not be used after the element is unmounted and its owner is killed.

If you try to create a subscription that uses a OneTimeOwner, the subscription will be killed immediately, and OneTimeOwner's onAccessAfterKilled callback will be fired. You can throw in that callback, then subscription initialization will throw too.

Note that the subscription itself does not contain any activation logic (i.e. what needs to happen when subscription is activated), that user-provided logic is external to subscription initialization, and is typically run before the subscription is initialized, so before OneTimeOwner can prevent that from happening. So when try to use a dead OneTimeOwner, instead of completely ignoring the effective payload of the subscription, unless you take special measures, it will still execute, but the subscription will be cancelled and cleaned up immediately. But if the subscription's payload was to e.g. make a network request, you can't put that back in the bottle.

Bottom line, you should not be deliberately sending events to dead OneTimeOwner-s. They just fix what otherwise could be a memory leak, not completely prevent your code from running. They report the error so that you can fix your code that's doing this.

ManualOwner

The basic Owner trait also doesn't allow external code to kill it, because some owners are supposed to manage themselves. All you need to overcome that is expose the killSubscriptions method to the public, or just use the ManualOwner class that does this.

Dynamic Ownership

Dynamic Ownership is not a replacement for standard Ownership described above. Rather, it is a self-contained feature built on top of regular Ownership. No APIs in Airstream itself require Dynamic Ownership, it is intended to be consumed by the user or by other libraries depending on Airstream.

The premise of Dynamic Ownership is similar to that of regular ownership: you can create DynamicSubscription-s owned by DynamicOwner-s. Here is what's different:

Regular Subscription-s can never recover from being kill()-ed, whereas DynamicSubscription-s can be activated and deactivated, and then activated again, and so on, as many times as their DynamicOwner wants. For example:

val stream: EventStream[Int] = ???
val observer: Observer[Int] = ???
 
val dynOwner = new DynamicOwner
 
val dynSub = DynamicSubscription.unsafe(
  dynOwner,
  activate = (owner: Owner) => stream.addObserver(observer)(owner)
)
 
// Run dynSub's activate method and save the resulting non-dynamic Subscription
dynOwner.activate()
 
// Kill the regular Subscription that we saved 
dynOwner.deactivate()
 
// Run dynSub's activate method again, obtaining a new Subscription
dynOwner.activate()
 
// Kill the new Subscription that we saved 
dynOwner.deactivate()

Every time a DynamicOwner is activate()-d, it creates a new OneTimeOwner, and uses it to activate every DynamicSubscription that it owns. It saves the resulting non-dynamic Subscription, which the DynamicOwner later kills() when it's deactivate()-d.

Now you can see how this integrates with regular ownership. Anything that requires a non-dynamic Owner produces a Subscription. So to create a DynamicSubscription you need to provide an activate method that does this. That could be a call to addObserver, addSource, etc.

As a result, we have a dynamic owner that can add or remove observer from stream at any time. In Laminar starting with v0.8 every ReactiveElement has a DynamicOwner. When the element is mounted, that owner is activated, activating all dynamic subscriptions using a newly created non-dynamic owner. Then when the element is later unmounted, those subscriptions are deactivated, removing observers from observables, sources from event buses, etc.

Previously in Laminar v0.7 every ReactiveElement used to extend the non-dynamic Owner, so once it was unmounted, all the subscriptions were killed forever, so if the user re-mounted that element, its subscriptions would not have come back to life. But now that Laminar uses Dynamic Ownership, you can re-mount previously unmounted elements, and their dynamic subscriptions will spring back to life.

Note that a DynamicSubscription is not automatically activated upon creation. Its DynamicOwner controls its activation and deactivation. You can still permanently kill() a DynamicSubscription manually – it will be deactivated if it's currently active, and removed from its DynamicOwner.

Dynamic Ownership & Memory Management

I created Dynamic Ownership specifically to solve this long-standing Laminar memory management issue: if a non-dynamic Subscription is created when ReactiveElement is initialized, and is killed when that element is unmounted, what happens to elements that get initialized but are never mounted into the DOM? That's right, their subscriptions are never killed (because they are technically never unmounted) and so they are essentially never garbage collected.

Laminar v0.8 had to fix this by creating Subscription-s every time the element is mounted, and killing them when the element was unmounted. Long story short, Dynamic Ownership is exactly this, slightly generalized for wider use.

There is really nothing special in Dynamic Ownership memory management. It's just a helper to create and destroy subscriptions repeatedly. In practice DynamicSubscription's activate method generally contains the same references that Subscription's cleanup method would, so it's all the same considerations as before.

Transferable Subscription

What, a helper for subscription helpers? Yes, indeed. This one behaves like a DynamicSubscription that lets you transfer it from one active DynamicOwner to another active DynamicOwner without deactivating and re-activating the subscription.

The API is simple:

class TransferableSubscription(
  activate: () => Unit,
  deactivate: () => Unit
) { 
  def setOwner(nextOwner: DynamicOwner): Unit
  def clearOwner(): Unit
}

Note that you don't get access to Owner in activate. This is the tradeoff required to achieve this flexibility safely. TransferableSubscription is useful in very specific cases when you only care about continuity of active ownership, such as when moving an element from one mounted parent to another mounted parent in Laminar (you wouldn't expect Unmount / Mount events to fire in this case).

Sources of Events

We now understand how events propagate through streams and signals, but the events in Airstream have to originate somewhere, right?

Creating Observables from Futures

EventStream.fromFuture[A] creates a stream that emits the value that the future completes with, when that happens.

  • The event is emitted asynchronously relative to the future's completion
  • Creating a stream from an already completed future results in a stream that emits the future's value when it starts.

Signal.fromFuture[A] creates a Signal of Option[A] that emits the value that the future completes with, wrapped in Some().

  • The initial value of this signal is always equal to None – even if the future has already completed when the initial value was evaluated. In that case, the initial None will be quickly (but asynchronously) followed by Some(resolvedValue).
  • If the Signal was created from a not yet completed future, the completion event is emitted asynchronously relative to when the future completes, because that is how future.onComplete works.
  • Being a StrictSignal, this signal exposes now and tryNow methods that provide its current value. However, note that there is a short asynchronous delay between the completion of the Future and this signal's current value updating, as explained above.

Signal.fromFuture(future, initialValue) is a variation of this method that returns a Signal[A] instead of Signal[Option[A]]. Otherwise, it behaves just as described above, with the initial None replaced by initialValue.

Note that all observables created from futures fire their events in a new transaction because they don't have a parent observable to be synchronous with.

If you have an Observable[Future[A]], you can flatten it into Observable[A] in a few ways, see Flattening Observables.

A failed future results in an error (see Error Handling).

Creating Observables from other Streaming APIs

EventStream.fromPublisher[A] creates a stream that subscribes to a Flow.Publisher, and emits the values that it produces.

Flow.Publisher is a Java Reactive Streams interface that is useful for interoperating between streaming APIs. For example, you can transform an FS2 Stream[IO, A] into an Airstream EventStream[A].

The resulting EventStream creates a new Flow.Subscriber and subscribes it to the publisher every time the EventStream is started, and cancels the subscription when the stream is stopped.

import cats.effect.unsafe.implicits._ // imports implicit IORuntime
EventStream.fromPublisher(fs2Stream.unsafeToPublisher())

EventStream.fromJsPromise and Signal.fromJsPromise

Behave the same as fromFuture above, but accept js.Promise instead. Useful for integration with JS libraries and APIs.

EventStream.fromSeq

object EventStream {
  def fromSeq[A](events: Seq[A]): EventStream[A] = ...
  ...
}

This method creates an event stream that synchronously emits events from the provided sequence to any newly added observer.

Each event is emitted in a separate transaction, meaning that the propagation of the previous event will fully complete before the propagation of the new event starts.

Note: you should avoid using this factory, at least with multiple events. You generally shouldn't need to emit more than one event at the same time like this stream does. If you do, I think your model is likely abusing the concept of "event". This method is provided as a kludge until I can make a more confident determination.

EventStream.fromValue

Like EventStream.fromSeq (see right above), but only allows for a single event.

EventStream.fromTry

Like EventStream.fromValue (see right above), but also allows an error.

EventStream.delay(ms)

Fires a Unit, or another value, if provided, ms milliseconds after the stream is started.

EventStream.periodic

An event stream that emits events at an interval. EventStream.periodic emits the index of the event, starting with 0 for the initial event that's emitted without delay. If you want to skip the initial event, use .drop(1). The resetOnStop option (false by default) determines whether the index will be reset to 0 when the stream is stopped due to lack of observers. You can also reset the stream to any index manually by calling resetTo(value) on it. This will immediately emit this new index.

The underlying PeriodicStream class offers more functionality, including the ability to emit values other than index, set a custom interval for every subsequent event, and stop the stream while it still has observers.

EventStream.empty

A stream that never emits any events.

EventStream.withCallback and withObserver

EventStream.withCallback[A] creates and returns a tuple of a stream and an A => Unit callback that, when called, passes the callback's parameter to that stream. Of course, as streams are lazy, the stream will only emit if it has observers.

val (stream, callback) = EventStream.withCallback[Int]
callback(1) // nothing happens because stream has no observers
stream.foreach(println)
callback(2) // `2` will be printed  

EventStream.withJsCallback[A] works similarly except it returns a js.Function for easier integration with Javascript libraries.

EventStream.withUnitCallback works similarly except it provides a callback that accepts no arguments, and a stream that emits Unit.

EventStream.withObserver[A] works similarly but creates an observer, which among other conveniences passes the errors that it receives into the stream.

EventBus

new EventBus[MyEvent] is a more powerful way to create a stream on which you can manually trigger events. The resulting EventBus exposes two properties:

events is the stream of events emitted by the EventBus.

writer is a WriteBus object that lets you trigger EventBus events in a few ways.

WriteBus extends Observer, so you can call onNext(newEventValue) on it, or pass it as an observer to another stream's addObserver method. This will cause the event bus to emit newEventValue in a new transaction.

Or you can just call eventBus.emit(newEvent) for the same effect.

What sets EventBus apart from e.g. EventStream.withObserver is that you can also call eventBus.addSource(otherStream)(owner), and the event bus will re-emit every event emitted by that stream. This is somewhat similar to adding writer as an observer to otherStream, except this will not cause otherStream to be started unless/until the EventBus's own stream is started (see Laziness).

You've probably noticed that addSource takes owner as an implicit param – this is for memory management purposes. You would typically pass a WriteBus to a child component if you want the child to send any events to the parent. Thus, we want addSource to be automatically undone when said child is discarded (see Ownership), even if writer.stream is still being observed.

An EventBus can have multiple sources simultaneously. In that case it will emit events from all of those sources in the order in which they come in. EventBus always emits every event in a new Transaction. Note that EventBus lets you create loops of Observables. It is up to you to make sure that a propagation of an event through such loops eventually terminates (via a proper .filter(passes) gate for example, or the implicit == equality filter in Signal).

You can manually remove a previously added source stream by calling kill() on the Subscription object returned by the addSource call.

EventBus is particularly useful to get a single stream of events from a dynamic list of child components. You basically pass down the writer to every child component, and inside the child component you can add a source stream to it, or add the writer as an observer to some stream. Then when any given child component is discarded (i.e. its owner kills its subscriptions), its connection to the event bus will also be severed.

Typically you don't pass EventBus itself down to child components as it provides both read and write access. Instead, you pass down either the writer or the event stream, depending on what is needed. This separation of concerns is the reason why EventBus doesn't just extend WriteBus and EventStream, by the way.

WriteBus comes with a few ways to create new writers. Consider this:

val eventBus = new EventBus[Foo]
val barWriter: WriteBus[Bar] = eventBus.writer
  .filterWriter(isGoodFoo)
  .contramapWriter(barToFoo)

Now you can send Bar events to barWriter, and they will appear in eventBus processed with barToFoo then and filtered by isGoodFoo. This is useful when you want to get events from a child component, but the child component does not or should not know what Foo is. Generally if you don't need such separation of concerns, you can just map/filter the stream that's feeding the EventBus instead.

WriteBus also offers a powerful contracomposeWriter method, which is like contramapWriter but with compose rather than map as the underlying transformation.

Batch EventBus Updates

EventBus emits every event in a new transaction. However, similar to Var batch updates you can call EventBus.emit or EventBus.emitTry to send values into several EventBus-es simultaneously, within the same transaction, to avoid glitches downstream.

val valuesEventBus = new EventBus[Int]
val labelsEventBus = new EventBus[String]
 
EventBus.emit(
  valuesEventBus -> 100,
  labelsEventBus -> "users"
)

Similar to Vars, you can't emit more than one event into the same EventBus in the same transaction. Airstream will throw if you attempt to do this, so you can't have duplicate inputs like EventBus.emit(bus1 -> ev1, bus1 -> ev2, bus2 -> ev3). If you need to emit more than one event into the same EventBus, just call the method twice, and they will be sent in separate transactions.

Var

Var is a reactive variable that you can update manually, and that exposes its current value at all times, as well as a .signal of its current value.

Creating a Var is straightforward: Var(initialValue), Var.fromTry(tryValue).

Simple Updates

You can update a Var using one of its methods: set(value), setTry(Try(value)), update(currentValue => nextValue), tryUpdate(currentValueTry => Try(nextValue)). Note that update will send a VarError into unhandled errors if the Var's current value is an error. Use set* or tryUpdate methods to update failed Vars.

Observers Feeding into Var

Every Var provides a writer which is an Observer that writes the values it receives into the Var.

In addition to writer, Var also offers updaters, making it easy to create an Observer that updates the Var based on both the Observer's input value and the Var's current value:

val v = Var(List(1, 2, 3))
val adder = v.updater[Int]((currValue, nextInput) => currValue :+ nextInput)

adder.onNext(4)
v.now() // List(1, 2, 3, 4)

val inputStream: EventStream[Int] = ???

inputStream.foreach(adder)
div(inputStream --> adder) // Laminar syntax

updater will send a VarError into unhandled errors if you ask it to update a Var that is in a failed state. In such cases, use writer or tryUpdater instead.

Vars of Options, i.e. Var[Option[A]], also offer someWriter: Observer[A] for convenience.

Reading Values from a Var

You can get the Var's current value using now() and tryNow(). now throws if the current value is an error. Var also exposes a signal of its values.

SourceVar, i.e. any Var that you create with Var(...), follows strict (not lazy) execution – it will update its current value as instructed even if its signal has no observers. Unlike most other signals, the Var's signal is also strict – its current value matches the Var's current value at all times regardless of whether it has observers. Of course, any downstream observables that depend on the Var's signal are still lazy as usual.

Being a StrictSignal, the Var's signal also exposes now and tryNow methods, so if you need to provide your code with read-only access to a Var, sharing only its signal is the way to go.

Var Transaction Delay

Var emits every event in a new transaction. This has important ramifications when writing to and reading from a Var. Consider the following code:

val myVar = Var(0)

println("Start")

myVar.set(1)
println(s"After set: ${myVar.now()}")

myVar.update(_ + 1)
println(s"After update: ${myVar.now()}")

Transaction { _ =>
  println(s"After trx: ${myVar.now()}")
}

println("Done")

If you put this code in your app's main method or inside a setTimeout callback, it will print:

Start
After set: 1
After update: 2
After trx: 2
Done

But if you try to run this same code while another transaction is being executed, for example inside one of your observers, in response to an incoming stream event, this is what will be printed:

Start
Done
After set: 0
After update: 0
After trx: 2

Why the difference? Var's current value exposed by now() only updates when the Var emits the updated value, and as we now know, this always happens in a new transaction. But we ran our code inside an observer, that is, while another transaction was running. And the new transaction will only run after the current transaction has finished propagating, so the Var' current value will not update until then.

This is why reading myVar.now() after calling myVar.set(1) gives you a stale value in this case. Var tries very hard to do the right thing though. While you can't expect to see the new value in now(), the update method does provide the updated value, 1, as the input to its callback. This is because the update callback is also scheduled for a new transaction, and so it is executed after the transaction in which the Var's value was set to 1 has finished propagating.

Finally, in both cases the code prints "After trx: 2". This is because that println is only executed in a new transaction. Similar to the update callback, this only gets run when the previously scheduled transactions have finished propagating, so it will always see the final Var value.

So there you have it, you have two ways to read the Var's new current value: either call now() inside a new transaction, or use update. And of course you can also listen to the Var's signal.

Keep in mind that transaction scheduling is fully synchronous, we do not introduce an asynchronous delay anywhere, we merely order the execution chunks to make the maximum amount of sense possible. Read more about transaction scheduling in the Transactions section.

Derived Vars

If you have a Var[A], you can get zoomed / derived Var[B] by providing A => B, (A, B) => A, and an owner. The result is a DerivedVar, essentially a combination of var.signal.map and writer.contramap packaged in a Var, and so to simulate the strictness of Var, creating DerivedVar requires an owner.

When you have a derived var, updates to it are propagated to the source var and vice versa, as long as derived var's signal has listeners. Don't try to set/update a derived Var's value when it has no listeners. The value will not be updated, and Airstream will emit an unhandled error.

case class FormData(num: Int, str: String)
val owner: Owner = ???
val source = Var(FormData(0, "a"))
val derived = source.zoom(_.num)((f, n) => f.copy(num = n))(owner)

source.update(_.copy(num = 2))
// source.now() == Form(2, "a")
// derived.now() == 2

derived.set(3)
// source.now() == Form(3, "a")
// derived.now() == 3

owner.killSubscriptions()

source.update(_.copy(num = 3))
// derived var did not update:
// derived.now() == 3

derived.set(5)
// neither var updated:
// source.now() == Form(4, "a")
// derived.now() == 3

Note: DerivedVar starts out with a subscription owned by owner, that counts as a listener of course. However, just like OwnedSignal in general, if it obtains any other listeners, it will continue running even if the original owner kills its subscription.

Batch Updates

Similar to EventBus, Var emits each event in a new transaction. And, similar to EventBus.emit, you can put values into multiple Vars "at the same time", in the same transaction, to avoid glitches downstream. To do that, use the set / setTry / update / tryUpdate methods on the Var companion object. For example:

val value = Var(1)
val isEven = Var(false)

val sumSignal = x.signal.combineWith(y.signal)

// batch updates!
Var.set(x -> 2, y -> true)

With such a batched update, sumSignal will only emit (1, false) and (2, true). It will not emit an inconsistent value like (1, true) or (2, false).

Batch updates are also atomic in the following ways:

  • update and tryUpdate will only execute the provided mods when the transaction is actually executed, not immediately as it's scheduled. This ensures that the mods operate on the latest available Var state.
  • Similar to Var#update, Var.update sends an error into unhandled if you try to apply mod to a failed Var. In the batch case none of the input Vars will be updated, although some of the mod functions will be executed. For this reason, mod functions should be pure of side effects. Use tryUpdate when you need more control over error handling.
  • Similar to Var#tryUpdate, Var.tryUpdate sends an error into unhandled if any of the provided mods throw. None of the Vars will update in this case. You should return Failure() from your mod instead of throwing if this is not what you want.

Also, since an Airstream observable can't emit more than once per transaction, the inputs to batch Var methods must have no duplicate vars. For example, you can't do this: Var.set(var1 -> 1, var1 -> 2, var2 -> 3). Airstream will detect that you're attempting to put two events into var1 in the same transaction, and will send an error into unhandled. Use two separate calls if you want to send two updates into the same Var.

Keep in mind that derived vars count as the underlying source vars for duplicate detection purposes, so you can't update vars var1 and var1.zoom(fa)(fb) in the same transaction.

Those are the only ways in which setting / updating a Var can trigger an error. If any of those happen when batch-updating Var values, Airstream will none of the involved Vars will fail to update, keeping their current value.

Remember that this atomicity guarantee only applies to failures which would have caused an individual update / tryUpdate call to throw. For example, if the mod function provided to update throws, update will not throw, it will instead successfully set that Var Failure(err).

For extra clarity, note that "sending error into unhandled" simply reports the error and cancels the update of the Var, it does not stop the execution of the program like a real throw could.

Val

Val(value) / Val.fromTry(tryValue) is a Signal "constant" – a Signal that never changes its value. Unlike other Signals, its value is evaluated immediately upon creation, and is exposed in public now() and tryNow() methods.

Val is useful when a component wants to accept either a Signal or a constant value as input. You can just wrap your constant in a Val, and make the component accept a Signal (or a StrictSignal) instead.

FetchStream

Airstream has a convenient interface to make network requests using the modern Fetch browser API:

FetchStream.get(
  url,
  _.redirect(_.follow),
  _.referrerPolicy(_.`no-referrer`),
  _.abortStream(...)
) // EventStream[String] of response body

You can also get a stream of raw dom.Response-s, or use a custom codec for requests and responses, all with the same API:

FetchStream.raw.get(url) // EventStream[dom.Response]
val Fetch = FetchStream.withCodec(encodeRequest, decodeResponse)

Fetch.post(url, _.body(myRequest)) // EventStream[MyResponse]

Ajax

Ajax (XMLHttpRequest) is a legacy web technology that was largely replaced by the Fetch API (see above). Nevertheless, Airstream has a built-in way to perform Ajax requests:

AjaxStream
  .get("/api/kittens") // EventStream[dom.XMLHttpRequest]
  .map(req => req.responseText) // EventStream[String]

Methods for POST, PUT, PATCH, and DELETE are also available.

The request is made every time the stream is started. If the stream is stopped while the request is pending, the old request will not be cancelled, but its result will be discarded.

If the request times out, is aborted, returns an HTTP status code that isn't 2xx or 304, or fails in any other way, the stream will emit an AjaxStreamError.

If you want a stream that never fails, a stream that emits an event regardless of all those errors, call .completeEvents on your ajax stream.

You can listen for progress or readyStateChange events by passing in the corresponding observers to AjaxEventStream.get et al, for example:

val (progressObserver, progressS) = EventStream.withObserver[(dom.XMLHttpRequest, dom.ProgressEvent)]

val requestS = AjaxEventStream.get(
  url = "/api/kittens",
  progressObserver = progressObserver
)

val bytesLoadedS = progressS.mapN((xhr, ev) => ev.loaded)

In a similar manner, you can pass a requestObserver that will be called with the newly created dom.XMLHttpRequest just before the request is sent. This way you can save the pending request into a Var and e.g. abort() it if needed.

Warning: dom.XmlHttpRequest is an ugly, imperative JS construct. We set event callbacks for onload, onerror, onabort, ontimeout, and if requested, also for onprogress and onreadystatechange. Make sure you don't override Airstream's listeners in your own code, or this stream will not work properly.

Websockets

Airstream has no official websockets integration yet.

For several users' implementations, search the old Laminar gitter room, and the issues in this repo.

DOM Events

val element: dom.Element = ???
DomEventStream[dom.MouseEvent](element, "click") // EventStream[dom.MouseEvent]

This stream, when started, registers a click event listener on element, and emits all events the listener receives until the stream is stopped, at which point the listener is removed.

Airstream does not know the names & types of DOM events, so you need to manually specify both. You can get those manually from MDN or programmatically from event props such as onClick available in Laminar.

DomEventStream works not just on elements but on any dom.EventTarget. However, make sure to check browser compatibility for weird EventTarget-s such as XMLHttpRequest.

Custom Event Sources

If simpler event sources (see above) do not suit your needs, consider using CustomSource. This mechanism lets you create a custom stream or signal as long as it does not depend on other Airstream observables. So, it's good for bringing third party sources of events into Airstream.

You can create custom event sources using EventStream.fromCustomSource and Signal.fromCustomSource, which are convenience wrappers over the underlying CustomEventSource and CustomSignalSource classes. This section will explain how to use those underlying classes, and after that the understanding of fromCustomSource methods should come naturally.

Airstream's DomEventStream.apply creates a stream of events by wrapping the DOM API into CustomStreamSource. Let's see how it works:

def apply[Ev <: dom.Event](
  eventTarget: dom.EventTarget,
  eventKey: String,
  useCapture: Boolean = false
): EventStream[Ev] = {

  CustomStreamSource[Ev]( (fireValue, fireError, getStartIndex, getIsStarted) => {

    val eventHandler: js.Function1[Ev, Unit] = fireValue

    CustomSource.Config(
      onStart = () => {
        eventTarget.addEventListener(eventKey, eventHandler, useCapture)
      },
      onStop = () => {
        eventTarget.removeEventListener(eventKey, eventHandler, useCapture)
      }
    )
  })
}

When we create a CustomStreamSource, we need to provide a callback that accepts some useful arguments and returns an instance of CustomSource.Config, which is essentially a bundle of two callbacks: onStart which fires when your stream is started, and onStop which fires when your stream is stopped (see Laziness).

Here we see that DomEventStream registers fireValue as an event listener on the DOM element when the stream starts, and unregisters that listener when the stream stops. This way the resulting stream will properly clean up its resources.

Side note: val eventHandler is cached to avoid implicitly creating a new instance of js.Function1. We need to keep this exact reference to be able to unregister the listener. Just a bit of Scala-vs-js friction here.

Let's look at the methods that CustomStreamSource makes available to us:

  • fireValue - call this with a value to make the custom stream emit that value in a new transaction
  • fireError - call this with a Throwable to make the custom stream emit an error (see Error Handling)
  • getStartIndex – call this to check how many times the custom stream has been started. Airstream uses this for the emitOnce param in streams like EventStream.fromSeq.
  • getIsStarted – call this to check if the custom stream is currently started

CustomSource.Config instances have a when(passes: () => Boolean) method that returns a config that, when the predicate does not pass, will not call your onStart callback when the stream starts, and will not call your onStop callback when the stream is subsequently stopped (we assume that your onStop code cleans up after your onStart code). To clarify, the predicate is evaluated when the custom stream is about to start. And the stream *will actually start – you can't break this part of Airstream contract – the predicate only controls whether your callbacks defined in the config will be run. You can see this predicate being useful to implement the emitOnce param in streams like EventStream.fromSeq.

CustomSignalSource is the Signal version of CustomStreamSource, and works similarly, just with a slightly different set of params:

class CustomSignalSource[A] (
  getInitialValue: => Try[A],
  makeConfig: (SetCurrentValue[A], GetCurrentValue[A], GetStartIndex, GetIsStarted) => CustomSource.Config
)

fireValue and fireError are merged into one setCurrentValue callback that expects a Try[A], and this being a Signal, we also provide a getCurrentValue param to check the custom signal's current value.

Generally signals need to be started in order for their current value to update. Stopped signals generally don't update without listeners, unless they are a StrictSignal like Var#signal. CustomSignalSource is not a StrictSignal so there is no expectation for it to keep updating its value when it's stopped. Users should keep listening to signals that they care about.

Extending Observables

If you need a custom observable that depends on another Airstream observable, you can subclass WritableEventStream or WritableSignal. See existing classes for inspiration, such as MapSignal and MapEventStream.

You will likely want to mix in either SingleParentSignal with InternalTryObserver or SingleParentStream with InternalNextErrorObserver. Then you will just need to implement onTry (for signals) or onNext / onError (for streams) methods, which will be triggered when the parent observable emits. In turn, those methods should call fireValue, fireError or fireTry to make your custom observable emit its own value. Also, for signals you will need to implement initialValue which you should derive from the parent observable's current value (NOT from the parent observable's initialValue).

If you want to put asynchronous logic in your observable, make sure to have a good understanding of Airstream transactions and topoRank, and consult with other asynchronous observables implementations such as DelayEventStream.

If your custom observable does not depend on any Airstream observables, e.g. if you're writing a compatibility layer for a third party library, you generally should be able to use the simpler Custom Sources API.

Accessing Protected Members

Some values and methods that you might want to access on observables are protected. That means that the compiler will only let you access those values and methods on the same instance. So, you can read this.topoRank, but you can't read parentObservable.topoRank. To get around this, use the Protected object: Protected.topoRank(parentObservable).

Aside from topoRank, you will need to access tryNow() and now() this way, e.g. when implementing a custom signal's initialValue. These methods require an implicit evidence of type Protected, which is automatically in scope if you're calling these methods from inside your custom observable. You're not supposed to access a signal's current value from the outside, without proving that the signal is running (e.g. by subscribing to it), otherwise you might get a stale value.

Honestly all this "protected" business smells funny to me, but I couldn't figure out a better way to allow third party extensions without making these protected members public.

Sources & Sinks

A Source[A] in Airstream is something that exposes a toObservable method, something that can be (explicitly, not implicitly) converted into an Observable[A]. For example, the observables themselves are Sources, but so are EventBus-es (def toObservable = this.events) and Var-s (def toObservable = this.signal).

Source is further subtyped into – EventSource (EventStream, EventBus) and SignalSource (Signal, Var). Predictably, eventSource.toObservable returns an EventStream, whereas signalSource.toObservable returns a Signal.

These types are useful when you want to create a method that can accept "anything that you can get a stream from". For example, it's used in Laminar:

val textBus = new EventBus[String]
div(value <-- textBus.events)
div(value <-- textBus) // Also works because this <-- accepts Source[String]

The counterparty to Source in Airstream is Sink. Sink[A] is something that exposes a toObserver method that can be explicitly (not implicitly) convert a Sink to an Observer. So Observers are sinks, as are EventBus-es and Var-s, and even js.Function1[A, Unit] has an implicit conversion to Sink[A].

However, there is no implicit conversion from A => Unit to Sink because unfortunately Scala requires a lambda's type param to have a type ascription to implicitly convert it into a Sink[A], so syntax like div(value <-- (_ => println("x")) would not be possible with such an implicit defined. In Laminar we get around this by overloading the <-- method to accept either a Sink[A] or A => Unit. If you need this conversion, just wrap your function in Observer(). You'll still need to ascribe the types though.

Speaking of implicits, why don't we have EventBus extend both EventStream[A] and Observer[A] instead of having separate Source and Sink types? On a technical level simply because Observable and Observer have overlapping methods defined, such as filter and delay, but more importantly, it would just be confusing. The whole point of Source and Sink is to not expose any methods other than toObservable, so that these types are only used as input types to methods that the developer wants to be flexible.

FRP Glitches

Other Libraries

A glitch in Functional Reactive Programming is a situation where inconsistent state is allowed to exist and exposed to either an observable or an observer. For example, consider the typical diamond case:

val numbers: EventStream[Int] = ???
val isPositive: EventStream[Boolean] = numbers.map(_ > 0)
val doubledNumbers: EventStream[Int] = numbers.map(_ * 2)
val combinedStream: EventStream[(Int, Boolean)] = doubledNumbers.combineWith(isPositive)
combinedStream.addObserver(combinedStreamObserver)(owner)

Now, without thinking too hard, what do you think combinedStream will emit when numbers emits 1, assuming -1 was previously emitted? You might expect that isPositive would emit true, doubledNumbers would emit 2, and then combinedStream would emit a tuple (2, true). That would make sense, and this is how Airstream works at no cost to you, and yet this is not how most streaming and state propagation libraries behave.

Most streaming libraries will introduce a glitch in this scenario, as they are implemented with unconditional depth-first propagation. So in other libraries when the event from numbers (1) propagates, it goes to isPositive (true), then to combinedStream ((-1, true)). And that's a glitch. (-1, true) is not a valid state, as -1 is not a positive number. Immediately afterwards, doubledNumbers will emit 2, and finally combinedStream would emit (2, true), the correct event.

Such behaviour is problematic in a few ways – first, you are now propagating two events on equal standing. Any observables (and in most other libraries, even observers!) downstream of combinedStream will see two events come in, the first one carrying invalid/incorrect state, and they will probably perform incorrect calculations or side effects because of that.

In general, glitches happen when you have an observable that synchronously depends on multiple observables that synchronously depend on a common ancestor or one of themselves. I'm using the term synchronously depends to describe a situation where emitting an event to a parent observable might result in the child observable also emitting it – synchronously. So map and filter would fall into this category, but delay wouldn't.

Topological Rank

In the diamond-combine case described above Airstream avoids a glitch because CombineObservable-s (those created using the combineWith method) do not propagate downstream immediately. Instead, they are put into a pendingObservables queue in the current Transaction (we'll get to those soon). When the rest of the propagation within a transaction finishes, the propagation of the first pending observable is resumed. When that is finished, we propagate the first remaining pending observable, and so on.

So in our example, what happens in Airstream: after isPositive emits true, combinedStream is notified that one of its parents emitted a new event. Instead of emitting its own event, it adds itself to the list of pending observables. Then, as the isPositive branch finished propagating (for now), doubledNumbers emits 2, and then again notifies combinedStream about this. combinedStream is already pending, so it just grabs and remembers the new value from this parent. At this point the propagation of numbers is complete (assuming no other branches exist), and Airstream checks pendingObservableson the current transaction. It finds only one – combinedStream, and re-starts the propagation from there. The only thing left to do in our example is to send the new event – (2, true) to combinedStreamObserver.

Now, only this simple example could work with such logic. The important bit that makes this work for complex observable graphs is topological rank. Topological rank in Airstream is defined as follows: if observable A synchronously depends (see definition above) on observable B, its topological rank will be greater than that of B. In practical terms, doubledNumbers.topoRank = numbers.topoRank + 1 and combinedStream.topoRank == max(isPositive.topoRank, doubledNumbers.topoRank) + 1.

In case of combineWith, Airstream uses topological rank for one thing – do determine which of the pending observables to resolve first. So when I said that Airstream continues the propagation of the "first" pending observable, I meant the one with the lowest topoRank among pending observables. This ensures that if you have more than one combined observable pending, that the one that doesn't depend on the other one will be propagated first.

So this is how Airstream avoids the glitch in the diamond-combine case.

Transactions

Before we dive into other kinds of glitches (ha! you thought that was it!?), we need to know what a Transaction is.

Philosophically, a Transaction in Airstream encapsulates a part of the propagation that 1) happens synchronously, and 2) contains no loops of observables. Within the confines of a single Transaction Airstream guarantees a) no glitches, and b) that no observable will emit more than once.

Async streams such as stream.delay(500) emit their events in a new transaction because Airstream executes transactions sequentially – and there is no sense in keeping other transactions blocked until some Promise or Future decides to resolve itself.

Events that come from outside of Airstream – see Sources of Events – each come in a new Transaction, and those source observables have a topoRank of 1. I guess it makes sense why EventStream.periodic would behave that way, but why wouldn't EventBus reuse the transaction of whatever event came in from one of its source streams?

And the answer is the limitation of our topological ranking approach: it does not work for loops of observables. A topoRank is a property of an observable, not of the event coming in. And an observable's topoRank is static, determined at its creation. EventBus on its creation has no sources, and allows you to fire events into it manually, so its stream needs to emit all those events in a new Transaction because there is no way to guarantee correct topological ranking to avoid glitches.

That said, in practice this is not a big deal because the events that an EventBus receives from different sources should be usually independent of each other because they are coming from different child components or from different browser events.

Apart from EventBus there is another way to create a loop – the eventStream.flatten method. And that one too, produces an event stream that emits all events in a new transaction, for all the same reasons.

Loops and potentially-loopy constructs necessarily require a new transaction as a tradeoff. Some other libraries do some kinds of dynamic topological sorting which is less predictable and whose performance worsens as your observables graph gets more complicated, but with Airstream there are no such costs. The only – and tiny – cost is when Airstream inserts a CombineObservable into the list of pending observables – that list is sorted by a static topoRank field, so it takes O(n) where n is the number of currently pending observables, which is usually zero or not much more than that.

Lastly, keep in mind that emitting events inside Observer-s will necessarily happen in a new transaction as you will need to use EventBus / Var APIs that create new transactions. Observers are generally intended for side effects. Those effects might be emitting other events, but in that case we consider them independent events, not a continuation of the current transaction. Philosophically, Observers should not know what they're observing (and they can observe multiple things at a time).

Avoiding Glitches When Merging

Consider this:

val numbers: EventStream[Int] = ???
val tens: EventStream[Int] = numbers.map(_ * 10)
val hundreds: EventStream[Int] = tens.map(_ * 10)
val multiples: EventStream[Int] = EventStream.merge(hundreds, tens)
multiples.addObserver(multiplesObserver)(owner)

What do you expect multiples to emit when numbers emits 1? I expect it to emit 10, and then 100. Two important considerations here:

  1. On a high level, the order of output events is determined by the order of input events: hundreds emits 100 after tens emits 10, so the merged stream does the same. On a technical level, the order of events emitted in the same transaction is determined by the parent observables' topological rank.

  2. The merged stream can, by design, emit multiple events per one origination event (the 1 event), as shown in our example above. This means that it can't always emit all of the events in the same incoming transaction, because any observable can only emit one event per transaction. At the same time, in cases when the merge stream depends only on mutually unrelated observables (that never emit in the same transaction), we don't want to force the merge stream to fire all of its events in a new transaction, as this could cause FRP glitches down the road. And so, the merge stream takes a compromise: in every transaction, it emits the first parent observable's event as-is, but if any other parent observable also emits in the same transaction (like tens and hundreds do in our example), the merge stream re-emits that event in a new transaction. So, in our example, it would emit 10 in the same transaction as the original event, and then emit 100 in a new transaction.

Such handling of transactions might seem arbitrary, but it actually matches the semantics of merge streams. As a result, such a mechanism produces desired behaviour. Even though we're emitting some events in new transactions, which would normally increase the chance of FRP glitches downstream, we only do it when it's necessary (when the merge stream emits more than one event per one originating event), and so in practice we don't see glitches. In fact, any other behaviour is guaranteed to cause glitches (unexpected behaviour). This might sound handwavy, but there's actually a lot of real life experience and unit tests behind that principle. See Operators vs Transactions for more on that.

Scheduling of Transactions

When you call methods like Var#set, EventBus.emit, etc. we create a new transaction. If another transaction is currently executing, which is often the case (e.g. if you're doing this inside a stream.foreach callback), this transaction will not be executed immediately, but will be scheduled to be executed later, because to avoid glitches, the current transaction needs to finish first before any other transaction can put more events onto the observable graph.

So if you set a Var's value, you will not be able to read it in the same transaction, because this instruction will only be executed after the current transaction finishes:

val logVar: Var[List[Event]] = ???
stream.foreach { ev =>
  logVar.set(logVar.now() :+ ev)
  logVar.set(logVar.now() :+ ev)
  println(logVar.now())
  // NONE of the logVar.now() calls here will contain any `ev`
  // because they are all executed before the .set transaction executes.
  // Because of this, after all of the transactions are executed,
  // logVar will only contain one instance of `ev`, not two.
}

If you need to read of Var after writing to it, you can use Var#update, which will evaluate its mod only when its transaction runs, so it will always look at the freshest state of the Var:

val logVar = Var(List[Event]())
stream.foreach { ev =>
  logVar.update(_ :+ ev)
  logVar.update(_ :+ ev)
  // After both transactions execute, logVar will have two `ev`-s in it
}

Let's expand our example above:

val bus = new EventBus[Event]
val logVar = Var(List[Event]())
val countVar = Var(0)
bus.events.foreach { ev =>
  logVar.update(_ :+ ev)
  logVar.update(_ :+ ev)
  // After both transactions execute, logVar will have two `ev`-s in it
}
logVar.signal.foreach { log =>
  sideEffect(log.size)
  countVar.update(_ += 1)
}

Let's say you fires an event into bus, and its transaction A started executing. The callback provided to bus.events.foreach will schedule two transactions to update logVar, B and C. After that, transaction A will finish as there are no other listeners.

Transaction B will immediately start executing. ev will be appended to logVar state, then this new state will be propagated to logVar.signal. sideEffect(1) will be called, and another transaction D to update countVar will be scheduled. After that, transaction B will finish as there are no other listeners.

Now, which transaction will execute next, C (the second update to logVar), or D (update to countVar)? Since Airstream v0.11.0, D will execute next, because its considered to be a child of the transaction B that just finished, because it was scheduled while transaction B was running. After a transaction finishes, Airstream first executes any pending transactions that were scheduled while it was running, in the order in which they were scheduled. This is recursive, so effectively we iterate over an hierarchy of transactions in a depth-first search.

In practice, this makes sense: in the code, the first logVar.update(_ :+ ev) is seen before the second logVar.update(_ :+ ev), so the first transaction will completely finish, including any descendant transactions it creates, before we hand over control to its sibling transaction.

Remember that all of this happens synchronously. There can be no async boundaries within a transaction. Any event fires after an async delay is necessarily fired in a new transaction that is initialized / scheduled after the async delay, so it's not part of the pending transaction queue until the async delay resolves, and when it does, it's guaranteed that there are no pending transactions in the queue as Javascript is single-threaded.

Operators

Airstream offers standard observables operators like map / filter / collect / compose / combineWith etc. You will need to read the API doc or the actual code or use IDE autocompletion to discover those that aren't documented here or in other section of the documentation. In the code, see BaseObservable, Observable, EventStream, and Signal traits and their companion objects.

Some of the more interesting / non-standard operators are documented below:

Distinction Operators

Both streams and signals have various distinct* operators to filter updates using == or other comparisons. These can be used to make your signals behave like they did prior to v15.0.0 (see blog post), or to achieve different, custom logic:

signal.distinct // performs `==` checks, similar to pre-15.0.0 behaviour
signal.distinctBy(_.id) // performs `==` checks on a certain key
signal.distinctByRef // performs reference equality checks
signal.distinctByFn((prevValue, nextValue) => isSame) // custom checks
signal.distinctErrors((prevErr, nextErr) => isSame) // filter errors in the error channel
signal.distinctTry((prevTryValue, nextTryValue) => isSame) // one comparator for both event and error channels

The same operators are available on streams too.

N-arity Operators

Airstream offers several methods and operators that work on up to 9 observables or tuples up to Tuple9:

mapN((a, b, ...) => ???)

Available on observables of (A, B, ...) tuples

filterN((a, b, ...) => ???)

Available on observables of (A, B, ...) tuples

observableA.combineWith(observableB, observableC, ...)

There is a bit of magic to this method for convenience. streamOfA.combineWith(streamOfB) returns a stream of (A, B) tuples only if neither A nor B are tuple types. Otherwise, combineWith flattens the tuple types, so for example both streamOfA.combineWith(streamOfB).combineWith(streamOfC) and streamOfA.combineWith(streamOfB, streamOfC) return a stream of (A, B, C), not ((A, B), C). We achieve this using implicit Composition instances provided by the tuplez library.

observableA.combineWithFn(observableB, ...)((a, b, ...) => ???)

Similar to combineWith, but you get to provide the combinator instead of relying on tuples. For example: streamOfX.combineWithFn(streamOfY)(Point) where Point is case class Point(x: Int, y: Int).

EventStream.combine(streamA, streamB, ...) et al.

N-arity combine and combineWithFn methods are also available on EventStream and Signal companion objects.

observableA.withCurrentValueOf(signalB, signalC, ...)

Same auto-flattening of tuples as combineWith.

observable.sample(signalA, signalB, ...)

Returns an observable of (A, B, ...) tuples

Compose Changes

Some operators are available only on Event Streams, not Signals. This is by design. For example, filter is not applicable to Signals because a Signal can't exist without a current value, so signal.filter(_ => false) would not make any sense. Similarly, you can't delay(ms) a signal because you can't delay its initial value.

However, you can still use those operators with Signals, you just need to be explicit that you're applying them only to the Signal's changes, not to the initial value of the Signal:

val signal: Signal[Int] = ???
val delayedSignal = signal.composeChanges(changes => changes.delay(1000)) // all updates delayed by one second
val filteredSignal = signal.composeChanges(_.filter(_ % 2 == 0)) // only allows changes with even numbers (initial value can still be odd)

For more advanced transformations, composeAll operator lets you transform the Signal's initial value as well.

Sync Delay

Suppose you have two streams that emit in the same Transaction. Generally you don't know in which order they will emit, unless one of them depends on the other.

If this order matters to you, you can use delaySync operator to establish the desired order:

val stream1: EventStream[Int] = ???
val stream2: EventStream[Int] = ???
 
val stream1synced = stream1.delaySync(after = stream2)

stream1synced synchronously re-emits all values that stream1 feeds into it. Its only guarantee is that if stream1 and stream2 emit in the same transaction, stream1synced will emit AFTER stream2 (assuming it has observers of course, or it won't emit at all, as usual). Otherwise, stream2 does not affect stream1synced in any way. Don't confuse this with the sample operator.

Note: delaySync is better than a simple delay because it does not introduce an asynchronous boundary. delaySync does not use a setTimeout under the hood. In Airstream terms, stream1synced synchronously depends on stream1, so all events in stream1synced fire in the same transaction as stream1, which is not the case with stream1.delay(1000) – those events would fire in a separate Transaction, and at an async delay.

Under the hood delaySync uses the same pendingObservables machinery as combinedWith operator – see Topological Rank docs for an explanation.

Splitting Observables

Airstream offers powerful split and splitToSignals operators that split an observable of M[Input] into an observable of M[Output] based on Input => Key. The functionality of these operators is very generic, so we will explore its properties by diving into concrete examples.

Note: These operators are available on qualifying streams and signals by means of SplittableSignal and SplittableEventStream value classes.

Example 0: Tests

This operator is particularly hard to put into words, at least on my first try. You might want to read the split signal into signals test in SplitEventStreamSpec.scala

And hey, don't be a stranger, remember we have Discord for chat.

Example 1: Latest Version of Foo by Id

If you are familiar with Laminar, consider skipping to the second example

Suppose you have an Signal[List[Foo]], and you want to get Signal[Map[String, Signal[Foo]]] where the keys of the map are Foo ids, and the values of the map are signals of the latest version of a Foo with that id.

The important part here is the desire to obtain individual signals of Foo by id, not to transform a List into a Map. Here is how we could do this:

case class Foo(id: String, version: Int)
val inputSignal: Signal[List[Foo]] = ???
 
val outputSignal: Signal[List[(String, Signal[Foo])]] = inputSignal.split(
  key = _.id
)(
  project = (key, initialFoo, thisFooSignal) => (key, thisFooSignal)
)
 
val resultSignal: Signal[Map[String, Signal[Foo]]] = outputSignal.map(list => Map(list: _*))

Let's unpack all this.

In this example our input is a signal of a list of Foo-s, and we split it into a signal of a list of (fooId, fooSignal) pairs. In each of those pairs, fooSignal is a signal that emits a new Foo whenever inputSignal emits a value that contains a Foo such that foo.id == fooId.

So essentially each of the pairs in outputSignal contains a a foo id and a signal of the latest version of a Foo for this id, as found in inputSignal.

Finally, in resultSignal we trivially transform outputSignal to convert a list to a map.

Example 2: Dynamic Lists of Laminar Elements

Suppose you want to render a list of Foo-s into a list of elements. You know how to render an individual Foo from its signal, but the list of Foo-s changes over time, and you want to avoid unnecessarily re-creating DOM elements.

This is what you can do:

case class Foo(id: String, version: Int)
   
def renderFoo(fooId: String, initialFoo: Foo, fooSignal: Signal[Foo]): HtmlElement = {
  div(
    "foo id: " + fooId,
    "first seen foo with this id: " + initialFoo.toString,
    "last seen foo with this id: ",
    child <-- fooSignal.map(_.toString)
  )
}
 
val inputSignal: Signal[List[Foo]] = ???
val outputSignal: Signal[List[HtmlElement]] = inputSignal.split(
  key = _.id
)(
  project = renderFoo
)

This works somewhat similarly to React.js keys, if you're familiar with that API:

  • As soon as a Foo with id="123" appears in inputSignal, we call renderFoo to render it. This gives us a Div element that we will reuse for all future versions of this foo.

  • We remember this Div element. Whenever inputSignal updates with a new version of the id="123" foo, the fooSignal in renderFoo is updated with this new version.

  • Similarly, when other foo-s are updated in inputSignal their updates are scoped to their own invocations of renderFoo. The grouping happens by key, which in our case is the id of Foo.

  • When the list emitted by inputSignal no longer contains a Foo with id="123", we remove its Div from the output and forget that we ever made it.

  • Thus, the output signal contains a list of Div elements matching one-to-one to the Foo-s in the input signal list.

So in essence, our outputSignal is broadly equivalent to inputSignal.map(_.map(renderFoo)), except that renderFoo requires memoization and data that simple mapping can't provide, thus the need for split.

To drive the point home, let's see how we would likely do this without split:

case class Foo(id: String, version: Int)
   
def renderFoo(foo: Foo): Div = {
  div(
    "foo id: " + foo.id,
    "last seen foo with this id: " + foo.toString
  )
}
 
val inputSignal: Signal[List[Foo]] = ???
val outputSignal: Signal[List[Div]] = inputSignal.map(_.map(renderFoo))

Same input and output types, but the behaviour is very different.

  • First, renderFoo is now called every time inputSignal updates, for every Foo. This means that we are recreating the entire list of DOM nodes from scratch on every update. This is very inefficient.

    • In contrast, with split, we would only call renderFoo whenever we would see a new Foo with a previously unseen id in inputSignal, and the child <-- ... modifier would take care of updating existing elements as new versions foo came in.
  • Second, renderFoo no longer has access to initialFoo and fooSignal. It does not know anymore if the foo it's rendering has changed over time, it can't listen for those changes, etc.

distinctCompose parameter

The split operator internally uses == checks to determine whether each record in the collection has "changed" or not. If not for these == checks, split would trigger a useless update for every record on every incoming event, instead of triggering only on the record that was actually affected by the event.

To allow customization, the split operator has a second parameter called distinctCompose which indicates how exactly the values are to be distinct-ed, and defaults to _.distinct. You can override it to provide a custom distinctor function if desired:

children <-- nodesStream.split(_.id, _.distinctByFn(customComparator))(/*...*/)
splitByIndex

The split operator requires a unique key for each item, such as _.id. But what if you don't have such a key?

One option is to change your model to create an ephemeral key, that is generated on the frontend and never sent to the backend. Airstream does not need the key to be meaningful or consistent across user sessions – the key simply needs to identify a certain item in the dynamic list for as long as Laminar is rendering that list.

Another, simpler solution, is often quite workable – with splitByIndex you can use the index of the item in the list as its key. This works as well as any other key if you only ever append items to the list, and don't insert items in the middle, remove items from the middle, or reorder the items.

splitOption

This variation of the split operator is available for observables of Option[Foo], and it uses _.isDefined as a key, thus, it calls the render function when switching from None to Some(foo), and updates the signal provided to the render function when switching from Some(foo1) to Some(foo2).

splitOne

Now that you know how the split operator works, it's only a small leap to understand its special-cased cousin splitOne. Where split works on observables of List[Foo], Option[Foo] etc., splitOne works on observables of Foo itself, that is, on any observable:

case class Editor(text: Boolean, isMultiLine: Boolean)
   
def renderEditor(
  isMultiLine: Boolean,
  initialEditor: Editor,
  editorSignal: Signal[Editor]
): HtmlElement = {
  val tag = if (isMultiLine) textArea else input
  tag(value <-- editorSignal.map(_.text))
}
 
val inputSignal: Signal[Editor] = ???
 
val outputSignal: Signal[HtmlElement] = 
  inputSignal.split(key = _.isMultiLine)(project = renderEditor)

The example is a bit contrived to demonstrate that key does not need to be a record ID but could be any property. In this case, renderEditor will be called only when the next emitted word's isMultiline value is different from that of the last emitted editor, because that is when we need to change the tag we use – it's impossible to update the tag name of an existing element.

Another use case for this is when you want to reset a complex component's state, for example:

case class Document(id: String, content: DocumentContent, ...)
   
def editor(
  documentId: Boolean,
  initialDocument: Document,
  documentSignal: Signal[Document]
): HtmlElement = {
  val documentState = Var(initialDocument)
  ... // More complex setup here, with some internal state specific to a particular document
  div(
    someInput --> documentState,
    documentSignal.map(...) --> documentState,
    documentState --> ...
  )
}
 
val inputSignal: Signal[Document] = ???
 
val outputSignal: Signal[HtmlElement] = 
  inputSignal.split(key = _.id)(project = editor)

In this case we don't have a strict technical constraint like changing the element type that we need to work around. In principle, we don't absolutely need splitOne. However, imagine that this component is rendering an editor for a complex document, similar to e.g. Google Docs. This kind of component has a lot of internal state which is all tied to a specific document, to a specific document ID. When inputSignal emits an update to the current document (documentId is the same), we want documentSignal to update as usual, but when inputSignal emits a new document ID... we don't want to manually clear / reset all that complex documentId-specific state inside editor – with all the redundant state / caches / etc. in it, this could easily become prone to bugs. It would be much easier to simply discard the old editor component and create a new one for the new document. And this is exactly what the code above does. Less code – fewer bugs.

In other words, our splitOne code above guarantees that every instance of documentSignal will always have unchanging document.id matching documentId, and that the editor method will be called whenever we switch to rendering a new documentId.

This all might seem similar to what a distinct operator would do, and indeed there is some conceptual overlap, because distinct is a fundamental part of split semantics, but you will be hard pressed to implement this pattern with distinct instead of splitOne, at least if you have a single observable like Signal[Document] as your input.

Duplicate Key Warnings

The split operator does not tolerate items with non-unique keys – this is simply invalid input for it, and it will crash and burn if provided such bad data.

Therefore, Airstream offers duplicate key warnings by default. Your code will still break if the split operator encounters duplicate keys, but Airstream will first print a warning in the browser console listing the duplicate keys at fault.

Thus, these new warnings do not affect the execution of your code, and can be safely turned on for debugging or turned off for performance. You can adjust this setting both for your entire application, and for individual usages of split:

// Disable duplicate key warnings by default
DuplicateKeysConfig.setDefault(DuplicateKeysConfig.noWarnings)

// Disable warnings for just one split observable
stream.split(_.id, duplicateKeys = DuplicateKeysConfig.noWarnings)(...)

Flattening Observables

Warning! Avoid using the flatMap / flatten methods when they are not absolutely needed. The canonical way to combine two observables is combineWith, not flatMap, so don't use Scala's for-comprehensions just to get the value of several observables into one scope. flatMap should only be used when you need to switch your stream to mirroring a new observable in response to an incoming event. Using it where a simpler combineWith would suffice will degrade performance and cause FRP glitches. /Warning

Flattening generally refers to reducing the number of nested container layers. In Airstream the precise type definition can be found in the FlattenStrategy trait.

Aside from the def flatMap[...](implicit strategy: FlattenStrategy[...]) method on the Observable itself, a similar flatten method is available on all observables by means of MetaObservable implicit value class. To use these methods, you just need to provide an instance of FlattenStrategy that works for your specific observable's type. While you can of course implement your own flattening strategy, we have a few predefined in Airstream:

SwitchStreamStrategy flattens an Observable[EventStream[A]] into an EventStream[A]. The resulting stream will emit events from the latest stream emitted by the parent observable. So, as the parent observable emits a new stream, the resulting flattened stream switches to imitating this last emitted stream.

  • This strategy is the default for the parent observable type that it supports. So if you want to flatten an Observable[EventStream[A]] using this strategy, you don't need to pass it to flatten explicitly, it is provided implicitly.

SwitchSignalStrategy flattens an Signal[Signal[A]] into a Signal[A]. Works similar to SwitchStreamStrategy but with Signal mechanics, tracking the last emitted value from the last signal emitted by the parent signal. Note:

  • This strategy is the default for flattening Signal[Signal[A]].
  • When switching to a new signal, the flattened signal emits the current value of the new signal (unless it's the same as the flattened signal's previous current value, as usual).
  • The flattened signal follows standard signal expectations – it's lazy. If you stop observing it, its current value might get stale even if the signal that it's tracking has other observers.

SwitchSignalStreamStrategy and SwitchSignalObservableStrategy behave similarly, but work on different combinations of streams and signals. All of these are implicitly available, so you shouldn't need to choose manually.

To summarize, the above strategies result in an observable that imitates the latest stream / signal emitted by the parent observable. So as soon as the parent observable emits a new stream / signal, it stops listening for values produced by previously emitted futures / signals / streams.

ConcurrentStreamStrategy is essentially a dynamic version of EventStream.merge.

  • The resulting stream re-emits all the events emitted by all of the streams emitted by the input observable.
  • If you stop observing the resulting stream, it will forget all of the streams it previously listened to. When you start it up again, it will start listening to the input observable from scratch, as if it's the first time you started it.
  • To use this strategy, you need to explicitly pass it to flatMap.

All built-in strategies result in observables that emit each event in a new transaction, because the toporank of the resulting observable is not knowable in advance, so Airstream can't prove that you aren't creating a loop of observables by using flatMap, so it has to play it safe by emitting events in a new transaction.

Other Notable Operators

Once again, this is not a full list of Airstream operators, just some of the interesting / non-standard ones.

take and drop Operators

The stream.take(numEvents) operator returns a stream that re-emits the first numEvents events emitted by the parent stream, and then stops emitting. stream.drop(numEvents) does the opposite, skipping the first numEvents events and then starting to re-emit everything that the parent stream emits.

These operators are available with several signatures:

stream.take(numEvents = 5)
stream.takeWhile(ev => passes(ev)) // stop taking when `passes(ev)` returns `false`
stream.takeUntil(ev => passes(ev)) // stop taking when `passes(ev)` returns `true`
stream.drop(numEvents = 5)
stream.dropWhile(ev => passes(ev)) // stop skipping when `passes(ev)` returns `false`
stream.dropUntil(ev => passes(ev)) // stop skipping when `passes(ev)` returns `true`

Like some other operators, these have an optional resetOnStop argument. Defaults to false, but if set to true, they "forget" all past events, and are reset to their original state when the parent stream is stopped and then started again.

filterWith

stream.filterWith(signalOfBooleans) emits events from stream, but only when the given signal's (or Var's) current value is true.

Operators vs Transactions

This section is intended to help understand the practical aspects of transactions and FRP glitches. It assumes familiarity with those documentation sections, and seeks to improve their understanding.

We say that some Airstream operators like flatMap fire events in a new transaction. This can cause FRP glitches in cases when, at a high level, using flatMap or similar transaction-creating methods was unnecessary. In contrast, when the usage of flatMap is truly necessary to achieve your desired behaviour, you will not see FRP glitches. Let's try to categorize the methods and operators that Airstream offers, to get a more intuitive understanding of why that is.

Flowy Operators

First of all, remember that Airstream guarantees no FRP glitches within a Transaction, and places the following limits on what observables can do in a transaction:

  • No observable can emit more than once in a transaction.
  • An observable can only emit in a transaction synchronously, in response to its parent observable emitting in the same transaction

Basically, observables need to be expressible as map and/or filter to avoid the need for emitting events in a new transaction. We call such operators "flowy".

With such requirements, we can essentially forget about the concept of time within a transaction, and live in a beautiful world where all the observables are connected in a directed acyclic graph, and their updates propagate from the root(s) to the leaves of that graph.

This is a very desirable feel, because it comes closest to imperative programming with plain values. Normally when working with observables we have to worry about glitches, delays, side effects, recursive updates, etc., but within a transaction, we have none of that.

And so, we want to confine the propagation of each event to a single transaction, as much as possible. We do this by using "flowy" operators – these operators keep the flow going smoothly, and emit events in the same transaction as their input events, because they satisfy Airstream transaction guarantees:

  • They emit events only in response to parent observable emitting events
  • They emit at most one event per one input event

The basic operators are all flowy: map, filter, collect, take, drop, etc.

Even operators that convert streams to signals and back like startWith and changes are flowy, because aside from adding or removing the initial value, they are a mere map(identity) operator.

In addition, in Airstream the operators that combine multiple observables are also flowy – combineWith / withCurrentValueOf / sample all have map + filter semantics, albeit relating to several input observables, and are implemented in a special way to remain flowy and avoid FRP glitches. Explanation of their exact mechanics is here.

If you think real hard, you can imagine the split and distinct operators as (rather complicated) versions of map + filter, and so they too are flowy operators.

The same goes for the delaySync operator, because it reorders / prioritizes events within a transaction, without introducing an async delay, and it only emits as many events as it receives.

As you see, you can get pretty far with just flowy operators – you can map, filter, combine, and split observables in all sorts of ways. All other operators emit their events in a new transaction, and now we will see why, for each category.

Async Operators

Flowy operators are by definition synchronous, because there can be no async delay within a transaction. Thus, all operators that emit after an async delay are not flowy, and must emit their events in a new transaction.

This includes EventStream.periodic, delay, throttle, debounce, fromFuture, fromJsPromise.

Can these operators cause FRP glitches? The answer is the same – did you need the delay, either for business logic reasons, or for technical reasons (e.g. making a network request)? If yes, then you will see no glitches.

However, if you add .delay(0) to doubledNumbers in the diamond glitch example for no good reason, you would get glitch-like behaviour. I say "no good reason" because if you actually had a reason to do that, then the "glitch" wouldn't be a glitch, but would be your desired behaviour. I know this seems a bit hand-wavy, but when you actually run into such cases it's extremely obvious.

When we combine two observables into one (e.g. using combineWith), and we expect a given event in both parent observables to happen "at the same time", we also expect the combined observable to emit only one combined event as a result. When something else happens instead (e.g. the observable emits two events), we call it a glitch. But, if one of your parent observables is asynchronous, e.g. it's making a network request for each of its input events, you don't expect its output to happen at the same time as the other parent, and so even though you technically get the exact same behaviour with the combined observable emitting two events, it's a not a glitch, that's now your expected outcome, due to the asynchrony involved.

So, FRP glitches are expectation-vs-reality mismatches in behaviour. When you use async operators, you don't have expectations of events happening "at the same time", so the concept of glitches that we're used to in synchronous operators is not applicable.

Loopy Operators

When we earlier called the flatMap operator "unnecessary", we implied that it is more powerful, or perhaps somehow more expensive, than "flowy" operators that don't need to create a new transaction for every event they emit.

And this is indeed the case. flatMap allows you to trivially violate the constraints of a transaction, for example you can emit more than one event per incoming event:

val intStream: EventStream[Int] = ???
val flatIntStream: EventStream[Int] = intStream.flatMap { ev =>
  EventStream.fromSeq(List(ev + 1, ev + 2))
}

Or create a loop in the observable graph, i.e. a stream that depends on itself:

val intStream: EventStream[Int] = ???
val flatIntStream: EventStream[Int] = intStream.flatMap { ev =>
  flatIntStream.map(_ + 1)
}

As we mentioned before, Airstream can only propagate events within a single transaction as long as we're walking down a direct acyclic graph of observables. "acyclic" means that such a graph must have no cycles in it, that is, no observable can depend on itself, whether directly or indirectly. Thus, flatMap can not possibly be a part of an acyclic graph, as it can depend on itself.

Or, to be more precise, flatMap can be used without creating cycles / loops in the observable graph, but whether it is or isn't used that way in your program is not knowable in advance. Airstream calculates a static topological rank for every observable at its creation time, and uses that to make combineWith and similar operators glitch-free. But the topological rank of a flatMap observable is not knowable at its creation time, because the observable it depends on can change dynamically based on what its parent observable emits.

And so, every method and operator in Airstream that lets you feed events from an observable back into itself, has to emit its events in a new transaction. This includes the flatMap and flatten operators, but also every method used to update a Var or send an event into an EventBus, such as Var.set, EventBus.emit, etc.

Bottom line: Avoid using loopy operators where they are not needed. Don't use flatMap if combineWith works just as well. Don't put state in a Var if you can just map over some signal to compute the same state.

Merge Streams Special Case

stream1.mergeWith(stream2) can behave either as a flowy operator or a loopy operator depending on which streams are being merged.

If both streams never emit in the same transaction, the merged stream behaves as a flowy operator, avoiding FRP glitches as flowy operators do.

However, if one of the parent streams synchronously depends on the other, e.g. if you do stream1.mergeWith(stream1.map(_ * 10)), you must expect one incoming event to produce two independent events with separate propagations because... how else could it possibly work in this case?

And so, in such cases, when both of merged stream's parents emit in the same transaction, the merged stream emits the first event in the same transaction, but it emits the second event in a separate transaction, allowing it to propagate separately. This actually eliminates FRP glitches that could otherwise happen when using the combineWith operator on the output of combineWith.

See also: Avoiding Glitches When Merging

Restarting Observables

This documentation section assumes a good understanding of all sections above, especially Signals, Laziness and Transactions.

When an Airstream observable loses its last observer, it is stopped, and must detach itself from the rest of the observable graph, so that it can be garbage collected. Specifically, it must remove itself from the list of observers in its parent observable.

Because of that, parent observables stopped observables are unable to receive updates from their parent observables, for example if numSignal continues to update while fooSignal is stopped (we assume that numSignal has other observers and is not stopped itself).

val numSignal: Signal[Int] = ???
val fooSignal: Signal[Foo] = numSignal.map(Foo(_))

Signals Re-Syncing on Restart

Now let's suppose that fooSignal is started again, after being stopped for a while: we add an observer to it, and fooSignal adds itself to numSignal's observers, thus starting to receive its updates again. However, those updates have not come in yet (perhaps they only happen in response to user clicks), but as we already started the signal, we need to provide fooSignal's current value to this new observer that we added, per Signal's contract. Unfortunately, fooSignal's current value is actually stale, because it missed all of the updates in numSignal while fooSignal was stopped. To mitigate this inconsistency, when restarting, signals attempt to sync their value with the parent signal.

In our case, when restarting, fooSignal will check whether its parent numSignal has emitted any updates while fooSignal was stopped, and if that is the case, then fooSignal will apply the Foo(_) function to update its current value to match numSignal's latest value. All signals that have one parent signal do it this way (see trait SingleParentSignal).

Different kinds of signals have slightly different re-syncing logic, but at the high level it is all the same – the signals recompute their own current value if the parent signal(s) have emitted any updates. This latter condition is very important, because you don't want to run the signal's computation more than once per incoming event. Not so much because of performance, but because Airstream guarantees shared execution, i.e. that an observable will only process any incoming event only once, regardless of whether it has 1 or 100 subscribers, and regardless of other factors. That way, users' computations and side effects get executed a predictable number of times, even if the users are not disciplined about separating side effects from pure computations. Forcing purity on users in a language that does not enforce purity goes against Airstream design goals.

Restarting Streams

Unlike signals, streams do not have a "current value", so we generally can not sync a stream with its parent stream. For example, if barStream was stopped for a while, and then is started again, it will not emit any new events until boolStream emits a new event.

This might sound like an unfortunate limitation, but it actually works just fine when you're using event streams for their intended purpose – to represent events. For example, if you're re-mounting a Laminar component, you probably don't want to suddenly receive a click event that happened a while ago. On the other hand, if you have compiled some state based on those click events, that state should already live in a Signal, and your component should be able to sync with it. For example:

val parentClickEvents: EventStream[dom.MouseEvent] = ???
val state: Signal[State] = parentClickEvents.scanLeft((initialState, ev) => newState)

Now, if the state signal lives in your parent Laminar component, and is never stopped, any signals in your child component can re-sync with it when they are stopped and then restarted (when the child component is unmounted and then re-mounted).

Restarting Streams That Depend on Signals (signal.changes)

Airstream offers both state-like and stream-like observables under one roof, and integrating them smoothly in one system is profoundly hard. Consider the case of signal.changes – this is a stream that emits all events that the signal emits, except for its initial value.

What happens when signal.changes is stopped and then restarted, but signal itself is never stopped, continuing to emit new updates? On the surface, the answer is simple – signal keeps track of its current value, so when restarting the signal.changes stream, we just do the same thing as we do when restarting signals that depend on other signals – check if the parent signal has emitted while signal.changes was stopped, and if so, update the current value of signal.changes. That would be the desirable behaviour.

However... signal.changes is a stream, not a signal, so it does not have a "current value" that can be updated, nor a mechanism by which a new observer could pull its updated value. As an event stream, the only thing signal.changes can do is emit the signal's updated value as a new event, and under normal circumstances that's exactly what it would do, but restarting of a stream is not exactly a "normal circumstance", it is a special moment in its lifetime.

Specifically, when a stream is getting started, we don't really have access to the current / ongoing transaction, and we don't know how far the transaction has already propagated, so I don't think signal.changes can safely emit the signal's updated value in the current transaction when signal.changes is restarting. That does not seem safe, although I haven't quite proved it definitively to myself if I'm being honest. There could be room for improvement here.

The naive alternative then would be for signal.changes to emit the signal's updated value in a new transaction when signal.changes is being restarted. However, that is problematic: what if you add not one, but two observers to signal.changes at the same time? The first observer would trigger the restart of this stream, and would cause signal.changes to emit the new event in a new transaction, but then the second new observer would not receive that event, because it already missed it. This difference in behaviour is obviously undesirable, but that's not even the end of it.

If instead of adding two observers to signal.changes, we add one observer to signal.changes and another one to signal.map(foo).changes at the same time (while they are both stopped), then those streams would emit the new event different transactions, which would never happen under normal circumstances. This could have caused an FRP glitch when restarting these streams like that if we combineWith them somewhere downstream.

To avoid all this, we have a special mechanism that lets us batch simultaneous events in a new transaction when restarting observables. Currently, it's only used to restart signal.changes. Basically, to avoid this restarting glitch, you want to wrap all your simultaneous observer additions into Transaction.onStart.shared { /* code here */ }, so any signal.changes you restart within that block will all emit in the same transaction. This is also not a perfect match for "normal" Airstream behaviour, and could potentially cause the other kind of FRP glitch where an intermediary event that you do expect to happen is swallowed by the system instead. However, of the two evils I think this is a lesser one, because it's much less common for that to be a problem. Ideally I would like to find a more robust mechanism for this edge case, but currently I lack the time required to do the research.

Airstream's DynamicOwner uses this onStart.shared mechanism when activating all of its subscriptions, and all Laminar's methods like amend do the same when applying multiple modifiers at a time, so you really shouldn't ever need to use onStart.shared manually, unless you're an advanced user creating your own custom modifiers or ownership primitives.

Restarting Signals That Depend on Streams

As event streams have no concept of "current value" and don't remember their "last emitted event", a signal that depends on a stream can not sync its value with the parent stream when it's being restarted, it simply starts listening to the parent stream again, and the signal's current value remains (potentially) stale until the parent stream emits a new event.

Of course, if you want signals like stream.scanLeft(intial)((acc, ev) => ???) to keep receiving events from stream, you need to prevent them from getting stopped. In Laminar, this is usually accomplished by either choosing a better owner for the signal (e.g. move it to the parent component that does not get unmounted), or by hiding the owner element with CSS (display: none) instead of unmounting it.

Stopping is Actually Pausing

Starting with Airstream 15, Airstream's general paradigm is to "pause" the observables when they are stopped, and seamlessly "resume" them when they are restarted, instead of tearing them down on stop, and restarting them from scratch. That is, when an observable is stopped and then re-started, it now tries to retain its internal state. This applies to streams too. Even though they generally do not represent "state" (that's what signals are for), streams still have internal state that they manage. For example:

  • stream.take(numEvents = 2) internally remembers how many events it took from stream, so if stream emits 3 events, and then is stopped and re-started, the take stream will not emit any more events if stream continues to emit, because it already took 2 events. This can be overriden by passing resetOnStop = true to the take operator.

  • stream1.combineWith(stream2) internally remembers the last event(s) that it saw stream1 and stream2 emit even after it was stopped. That way, when the combined stream is restarted, it behaves as if it was never stopped (sort of), instead of behaving as if we were starting it for the first time (i.e. waiting for both stream1 and stream2 to emit).

As we've established before, signals even sync their value with their parent signal when they're being restarted, following the same paradigm of pausing and unpausing.

Signals That Keep Updating When Stopped

Generally, signals do not update their current value while they are stopped because they get disconnected from the source of updates, but there are several exceptions to this:

  • Signals that have never been started – their initial value might not be determined yet, for example if you have stream.toSignal(initial = foo), the foo initial value will not be evaluated until the signal is started. Until then, the signal's current value is unknown, and Airstream will not evaluate it until and unless the signal is started.

  • Future / Promise signals – the current value of Signal.fromFuture(future) and Signal.fromJsPromise(promise) mirrors the current value of the future / promise regardless of whether the signal is started.

  • Var signals – the current value of Var.signal mirrors the current value of the future / promise regardless of whether the signal is started.

  • Custom signals – signals made using CustomSource API or by extending the WritableSignal trait might behave in whatever way makes sense for them. Remember that if a signal is stopped, it has no observers, so instead of spending resources on keeping the signal's value up to date while it is stopped, you might want to update its value once when the signal is restarting, in its onWillStart method.

Debugging

Debugging Observables

val stream: EventStream[Int] = ???
val useJsLogger: Boolean = false

val debugStream = stream
  .debugWithName("MyStream") // optional: use this prefix when logging below
  .debugLogEvents(when = _ < 0, useJsLogger) // optional: only log negative numbers
  .debugSpyStarts(topoRank => ???)
  .debugBreakErrors()

// Before:
stream.addObserver(obs)

// After: when debugging, replace with:
debugStream.addObserver(obs)

Airstream offers many debugging operators for observables, letting you run a callbacks (debugSpy*), log (debugLog*), or set a JS breakpoint (debugBreak*) at the most important times in the observable's lifecycle, including when emitting events or errors, starting or stopping, and evaluating the initial value (for signals). Those methods are available implicitly on every observable via DebuggableObservable and DebuggableSignal.

To debug an observable, call one of the available debug* methods to produce a new observable that listens to the original one, and re-emits all its events and errors, and also performs the specified debug action.

For example, stream.debugLog() will create a new observable that will simply print every event and error that it emits, that stream feeds to it, and stream.debugLog().debugBreakErrors() will do the same plus also set a JS breakpoint on errors in stream.

Very importantly, debugLog does not monkey-patch the original observable to add debugging functionality to it. We create a new observable that depends on the original, and debug that. This is true for every debug operator: in stream.debugLog().debugBreakErrors(), debugLog() creates an observable based on stream, and debugBreakErrors() creates an observable based on stream.debugLog().

To make it crystal clear: stream.debugLog() will only log the events that stream.debugLog() emits, so you need to make sure to listen to it instead of listening to stream. Easiest is to just use stream.debugLog() in place of the original stream in your code.

Another important consideration is the order of debug procedures. It follows the propagation order. For example, in stream.debugLog().debugSpy(fn) the observable stream.debugLog() will obviously emit before the observable stream.debugLog().debugSpy(fn) emits, and so you will see the event printed first, and only after that will fn be called with that event. So if you're ever confused about why your debugger prints stuff in a weird order (e.g. event fired before stream is started), make sure your debug operators are in the expected order.

You can also use debugWith(debugger) method instead of debug* methods to provide an ObservableDebugger with all of the behaviour that you want, instead of adding it piece by piece.

Also regarding timing, the per-event debuggers like (debugSpy() / debugLog() / debugBreak()) do their thing right before the event is fired (by the debugged observable, not the original), and the start/stop debuggers do their thing right after the observable is started or stopped.

Note for signals: Any debug* method that triggers when the signal emits an event or an error will also trigger when the signal is started. This ensures that you won't miss the signal's initial value when debugging, even though the Signal's initial value technically isn't ever "emitted", in Airstream terms.

Logging debug operators generally offer an optional when filter to reduce noise, and a useJsLogger option that you should enable when you're logging native JS values. Logging plain JS objects with println will often result in printing [object Object], but JS dom.console.log can print them nicely.

Also, your logs will be prefixed with sourceName which defaults to stream.toString but you can provide a prettier name as a param to the .debug() method.

We try to minimize the impact of debugging on the execution of your observable graph. To that end, any errors you throw in the callbacks you provide to debugSpy* methods will be reported as unhandled (wrapped in DebuggerError), and will not affect the propagation of events.

displayName

When debugging, it's very useful to name your observables. There are two ways to do this: stream.setDisplayName("MyStream") patches stream in place to set its displayName. It returns the same stream, it does not create a new observable.

This should be the preferred method for adding permanent names to important observables. For example, if you have a global val AuthSignal: Signal[AuthContext], you might want to add .setDisplayName("AuthSignal") to its definition.

displayName, if explicitly set, is returned by the observable's toString method. If not explicitly set, it defaults to defaultDisplayName, which in turn defaults to java.Object's default type@hashcode toString implementation. If subclassing Observable, you can't override its toString method directly, but you can override defaultDisplayName.

displayName is of course useful to give human-readable identifiers to observables – seeing AuthSignal when print-debugging is much more useful than MapSignal@f161.

displayName is also prepended to logs produced by debugLog* methods. However, given stream.setDisplayName("MyStream"), while the displayName of stream is "MyStream", the displayName of stream.setDisplayName("MyStream").debugLog() is "MyStream|Debug", to differentiate it from stream.

You can of course also setDisplayName on the debugged stream directly: stream.debugLog().setDisplayName("MyDebuggedStream"), but if you have a chain of debug streams that you want to apply the same name to, you can use the withDisplayName method: stream.withDisplayName("MyDisplayName").debugLogEvents().debugSpyErrors().debugLogStarts() – in this case all three debug* streams will have their displayName set to "MyDisplayName", but stream will not. This is because unlike setDisplayName, withDisplayName does not patch the original observable, it creates a new debug observable and patches that instead. withDisplayName works with debug chains because unlike regular observables, debug observables inherit their parent debug observable's displayName by default.

Airstream does not require displayName to be unique, although if you want to keep your sanity, it should be descriptive enough to clearly tell you which instance it refers to.

debugTopoRank

Topological Ranks of observables determine the order of Airstream observable graph propagation. The new observables created using debug* operators will necessarily have different topoRank-s from the original observables. Due to (mostly) depth-first propagation in Airstream, debugging observables generally don't affect your graph's propagation, but if you're specifically debugging an issue related to topoRanks, you might want to avoid adding temporary observables to your observable graph.

You can check the topoRank of an observable using observable.debugTopoRank. Unlike other debug* operators, this one does not affect the observable or create new observables, it just returns the observable's topoRank.

Debugging Observers

val stream: EventStream[Int] = ???
val obs: Observer[Int] = ???

val debuggedObs = obs
  .debugWithName("MyObserver")
  .debugLog()
  .debugSpyErrors(err => ???)

// Before:
stream.addObserver(obs)

// After: when debugging, replace with:
stream.addObserver(debuggedObs)

Observers have debugging functionality very similar to observables, but it works slightly differently. Whereas adding debuggers to observables is similar to mapping them, adding debuggers to observers is similar to contramapping them.

Just as in typical contramapping, the order of execution is reversed, so in the example above debugSpyErrors() callback will run before the error is logged by log(), and all the debugging happens before the original observer runs.

Similarly to debugged observables, your debuggers throwing do not affect the execution of the observable graph, instead they result in an unhandled DebuggerError being reported.

Just like observables, observers can be named using setDisplayName and debugWithName, with similar semantics. Note that even though observers are executed in "reverse" order (contramap semantics), debugged observers inherit displayName from their parent, so the displayName of all debugged observers in obs.debugWithName("MyObserver").debugLog().debugSpy(???) is "MyObserver".

Error Handling

Airstream error handling is very different from conventional streaming libraries. Before diving into implementation details we first need to understand these differences and the reasons why such a departure from the norm is beneficial.

Scala Exception Handling

First, to clarify what kind of errors we're talking about here: this whole section is concerned with exceptions thrown by user-provided code inside Airstream observables. For example, consider the case of the project function in stream.map(project) throwing. Without special error handling capabilities in Airstream, such an exception would terminate the propagation of this stream and bubble up the call stack.

However, this behaviour is vastly undesirable in FRP context because the caller (which would be the source of events) is not normally in a position to handle failures of child observables. For example, a DOM event listener that publishes DOM events onto a stream can't do much about some other component failing to process some of those events. So we need a different strategy to deal with errors in observables.

Conventional Streaming Libraries

Conventional streaming libraries basically don't propagate errors in observables. If your stream fails, it gets completed, meaning that it informs all of its dependant observables and observers that it errored, and will no longer produce events, and is now shutting down forever.

If you don't want this outcome, you need to recover from such an error by creating a stream that takes two inputs: this original stream, and a stream factory that returns a new stream that you will switch to if the original stream errors.

This model does not make any sense to me. Consider this chain of observables:

object OtherModule {
  def doubled(parentStream: EventStream[Double]): EventStream[Double] = {
    parentStream.map(num => num * 2)
  }
}
 
// ----
 
import OtherModule.doubled
 
val stream1: EventStream[Double] = ???
val invertedStream: EventStream[Double] = stream1.map(num => 1 / num)
 
doubled(invertedStream).foreach(dom.console.log(_))

Inside doubled, there is nothing you can do to recover from an error in parentStream. You don't know what that stream does, so once it's broke, it's broke. You can't replace it with anything meaningful, just maybe some sentinel value to indicate failure.

So essentially, this requires you to either manually guard every single user-provided input to every stream, or lose your sanity to crazy amounts of coupling. In this scenario the more likely outcome is that you'll just ignore error handling, and your program will stop working on an error that would have been recoverable if only it wasn't in your streaming code.

Fundamentally, the problem here is conceptual: conventional streaming libraries see any exception in a stream as a terminal diagnosis for it.

Yes, such an error could have made parts of your application state inconsistent, that is a legitimate problem. However, unconditionally killing parts of your program that depend on a stream is not a way to mitigate inconsistent state for the simple reason that the streaming library has no idea which, if any, state became inconsistent, therefore it has no idea whether allowing the error to propagate will mitigate state breakage or make it even worse.

Airstream's Approach

Airstream aspires to replicate the feel of native exception handling in FRP. However, whereas in imperative code we typically want to propagate exceptions up the call stack, in Airstream we instead want to propagate errors in observables to their observers and dependent observables.

Think about it this way: in imperative code, you call a function which can throw, and you can either handle the error or decide to let your caller handle it, and so on, recursively. In Airstream, you make an observable that depends on another observable which can "throw". So you can either handle the error, or let any downstream observables or observers handle it.

This FRP adaptation of classic exception handling is somewhat similar to the approach of Scala.rx, however since Airstream offers a unified reactive system, we have to adjust this basic idea to support event streams as well.

To contrast our approach with conventional streaming libraries: where they see a failed stream, we only see a failed value, generally expecting the next emitted value to work fine. This is similar to plain Scala exceptions: if a function throws an exception, it does not suddenly become broken forever. You can call it again with perhaps a different value, and it will perhaps not fail that time. Yes, if such an exception produces invalid state, you as a programmer need to address it. Same in Airstream. We give you the tools (more on this below), you do the work, because you're the only one who knows how.

We elaborate on the call stack analogy in the subsection Errors Multiply below.

Error Propagation

Airstream does not complete or terminate observables on error. Instead, we essentially propagate error values the same way we propagate normal values. Each Observer and InternalObserver has onNext(A) and onError(Throwable) methods defined, so both observables and observers are capable of accepting error values as inputs.

If you do not take special action, Airstream observables will generally (but not always, we'll get to that) pass through the errors to their observers untouched. Eventually the error will reach one or more external Observers.

When an error reaches an external observer, its onError method will handle it. If you didn't specify onError (e.g. if you used the Observer(onNext) factory to create an observer) or if your onError partial function is not defined for a given error, Airstream will report this error as unhandled.

You can get notified about unhandled errors by registering a callback using AirstreamError.registerUnhandledErrorCallback(fn: Throwable => Unit). Similarly, you can unregister it using AirstreamError.unregisterUnhandledErrorCallback(fn: Throwable => Unit).

By default Airstream registers AirstreamErrors.consoleErrorCallback to log all errors to the console because there is nothing worse than silently swallowed unhandled errors. You can unregister it and/or register more callbacks, e.g. to terminate your program, or to report the error to a service like Sentry.

Now let's dive deeper into each step of this process.

User Input is Generally Guarded Against Exceptions

Airstream guards from exceptions almost all user (developer) inputs that are used to build observables. For example, if the project function that you provided to stream.map(project) throws an exception, the resulting stream will emit an error, kick-starting this whole error handling process.

However, user inputs that are supposed to return a Try are assumed to not throw. Various public class constructors that require the new keyword are generally not intended to be used directly, these can also have params that are not guarded against exceptions.

For clarity, every Airstream method and constructor available to you has a Scaladoc comment reaffirming whether its parameters are guarded or not.

Errors Do Not Affect Propagation

When an error happens in an observable (assuming that it was guarded against, as it should always be, see above), the observable emits this error to all of its observers – internal and external – in the same manner that it would emit a value. Think of it as propagating a Try[A] value instead of just A.

Errors in an Observer Do Not Affect Other Observers

Similarly, and unlike conventional streaming libraries, if an error happened in an observer's onNext or onError callback, this error will be reported as unhandled, but it will not prevent other observers or the code immediately following this observer's definition from running.

If you're concerned that the contents of your onNext / onError can fail and make your app state inconsistent, you do the same thing you've always done in this situation – try some code and catch the error, just keep it all inside the callback in question:

stream.addObserver(Observer(onNext = v =>
  try { riskyFoo(v) }
  catch { case err => recoverFromFooFail() }
))

Remember, this is just for observers. We have a better way for observables, read on.

Unhandled Errors Do Not Terminate The Program

In Airstream, unhandled errors do not result in the program terminating. By default they are reported to the console. You can specify different or additional handlers such as AirstreamError.debuggerErrorCallback or even a custom handler that effectively terminates your program.

Regardless of this seeming leniency, you should still handle all of your errors at some point before they become unhandled. In a good Airstream codebase every unhandled error must be treated as a bug.

Error Timing is Consistent

Observables generally emit errors at the same time as they would have emitted a non-error value. For example, a CombinedObservable like stream.combineWith(stream.map(foo)) will only emit a single value if stream emits a value (yes, Airstream deals with FRP Glitches for you). Similarly, it will only emit a single error if stream emits an error. However, that error will be wrapped in CombinedError because it needs to support the case when its parent observables simultaneously emit a different error.

On the other hand, MergeStream does no such error combination, it emits the errors similarly to how it emits non-error events – as they come, putting all but the first seen one in a new transaction.

DelayEventStream re-emits incoming errors with the same delay that it uses for normal values.

Event Streams Generally Forget Errors

Similarly to how event streams generally do not keep track of their last emitted value, they also forget their last emitted error once they finish propagating it to their observers.

Signals Remember Errors

If a Signal[A] observable runs into an error that it doesn't handle itself, this error becomes its current state. Signals' current state is actually Try[A], not just A.

StrictSignal[A] has a public now(): A method that returns its current value. Calling it is equivalent to tryNow().get– it will throw if current state is a Failure.

Errors Can Become Wrapped

Errors originating in observables error handling code (e.g. the pf in stream.recover(pf)) are wrapped in ErrorHandlingError.

In Observers created using Airstream-provided constructors:

  • If the user-provided observer callback (e.g. the handleError in Observer.withRecover[Int](cb, handleError)) throws while processing an incoming error, the error is wrapped into ObserverErrorHandlingError and sent off as unhandled error.

  • If handleObserverErrors constructor param is true (that's the default), and the user-provided observer callback throws while processing an incoming event (as opposed to an incoming error), the error is wrapped in ObserverError and sent to this same observer's onError method. So, such an observer has a chance to process its own failure. The usefulness of that lies mostly in being able to automatically pass this error up the chain of observers.

    For example:

    val sourceObs = Observer.fromTry[Int](println)
    val derivedObs = sourceObs.contramap[Int] { value =>
      if (value >= 0) 1 else throw new Exception("negative!")
    }
    derivedObs.onNext(1) // prints "Success(1)"
    derivedObs.onNext(-1) // prints "Failure(ObserverError: java.lang.Exception: negative!)"
    // no errors are sent into unhandled because `println` is a total function, it handles them all.

You can always access the original errors on wrapped errors, and it will always provide you with a stack trace that includes the line where your code failed. Make sure to configure error reporting as well as Scala.js source maps to make use of this in fullOpt / production.

Errors Multiply

The same error might need to be handled multiple times to avoid becoming unhandled.

It is perhaps counter-intuitive, but it's obvious in retrospect:

val upStream = ???
val fooStream = upStream.map(foo)
val barStream = upStream.map(bar)
fooStream.foreach(dom.console.log("foo"))
barStream.foreach(dom.console.log("bar"))

If upStream emits an error, both fooStream and barStream will emit an error – the same instance of it, actually. Then it will end up reported as unhandled – twice!

If we try to lean on our call stack analogy, it kinda breaks this time. Call stack is a stack, a linear data structure much like a list. An exception can only ever bubble to just one caller before bubbling up to its one and only caller. But in our case it feels like the bubbles are splitting into multiple parallel "streams".

Instead, think of fooStream and barStream as independent function calls that both require the value of upStream, except by magic of FRP upStream is only executed once and its value (well, its exception in this case) is shared with both fooStream and barStream calls. If you call a broken function twice, you get two exceptions (identical ones, thanks to FRP) and thus two call stacks to propagate through.

To be extra clear, if we add .recoverIgnoreErrors to the fooStream definition above, its observer will not receive the error coming from upStream, and will not report that error as unhandled. We did, after all, handle it, so that's fair. However, barStream's observer would still receive that same error, and would still report it as unhandled. This is why handling errors in the right place is very important, just like handling exceptions is in plain Scala code.

Recovering from Errors

As we mentioned, generally observables propagate the errors they receive with no change. However, we can recover from errors, like this:

val signal0: Signal[A] = ???
val signal1 = signal0.map(whatever)
val signal2 = signal1.recover {
  case MyException(foo) if foo.isGood => Some(foo) // emit foo
  case MyException(foo) if !foo.isGood => None // skip this, emit nothing
  case o: MyOtherException => throw new Exception("lolwat") // emit new error
}

Importantly, this recover method does not affect signal0, or even signal1. Like any other operator it's just creating a new observable (signal2), except this one has a defined error handling strategy, so it will process any errors that come through it.

When signal0 emits an error, signal2 will feed it to the partial function we provided to it. That function should normally emit either Some(value) to make signal2 emit some value, or None to just skip this event altogether, as if it never happened. Yes, this latter case means that signal2's current state would remain at whatever it was before this error came in, meanwhile signal0's current state would be updated to the error (same for signal1).

Any errors that this partial function is not defined for will pass through without being handled. If the partial function throws an error, it will be passed down wrapped in ErrorHandlingError.

In addition to recover, Observables have a couple shorthand operators:

recoverIgnoreErrors just skips all errors, emitting only good values. This is an FRP equivalent of an empty catch block.

Note that you can't ignore an error in a Signal's initial value, as it needs a Try[A] value to be its state whenever it's started. Therefore, if the initial value is an error, and recover returns None while handling it (that's what recoverIgnoreErrors does), the initial value is set to the original error.

recoverToTry transforms an Observable[A] into an Observable[Try[A]] that never emits an error (it emits Failure(err) as a value instead).

You can use it to get a stream of errors as plain values, for example:

stream.recoverToTry.collect { case Failure(err) => err } // EventStream[Throwable]

throwFailure lets you "un-recover" errors, i.e. it converts an observable of Try[A] into an observable of A, throwing the failure case into the error channel.

Handling Errors Using Observers

Observer.fromTry(nextTry => ...) and Observer.withRecover(onNext, onError: PartialFunction[Throwable, Unit]) let you handle all or some of the errors coming from upstream observables. Errors for which onError is not defined get reported as unhandled.

Observer.ignoreErrors(onNext) is similar to recoverIgnoreErrors on observables – it simply silences any error it receives, so that it does not get reported as unhandled.

Observers that are derived from other observers, e.g. observer.contramap[Int](...), pass the error to the original observer, and so maintain the original observer's error handling behaviour.

Airstream-provided Observer constructors that let you specify an error handling callback (e.g. Observer.fromTry and Observer.withRecover) also have a handleObserverErrors param. When this param is true (that's the default), if the observer's callback throws while processing an incoming event, the same observer's error callback (the onError method) gets called with the error wrapped in ObserverError. This lets you automatically propagate unexpected exceptions up the chain of observers instead of sending the error into unhandled right away. For more details, see Errors Can Become Wrapped above.

Other Error Handling Considerations

  • scanLeft operator is unable to proceed when encountering an error, so such an observable will enter a permanent error state if it encounters an error. You can not use the standard recover method to recover from this. You need to use scanLeftRecover instead of scanLeft to supply your error handling logic.

  • filter operator can't filter if its passes function fails, so it will pass through all errors that it receives, unfiltered. You can filter errors using recover, by returning None.

  • Remember that Signal's initial value is not evaluated until and unless it is needed. That is true even if the initial value would have been an error because obviously you can't know what it is without evaluating it. And if an error is not evaluated, then it can't possibly be reported anywhere because, well, it didn't actually happen. In practice this means that the initial value of a Signal whose only consumer is its .changes stream is completely ignored (because no one cares about it). @TODO[API] Should we reconsider this particular aspect of laziness? Either way, we should document the rationale for that some more.

Limitations

  • Airstream only runs on Scala.js because its primary intended use case is unidirectional dataflow architecture on the frontend. I have no plans to make it run on the JVM. It would require too much of my time and too much compromise, complicating the API to support a completely different environment and use cases.
  • Airstream has no concept of observables "completing". Personally I don't think this is much of a limitation, but I can see it being viewed as such. See Issue #23.

My Related Projects

  • Laminar – Efficient reactive UI library for Scala.js that uses Airstream
  • Waypoint – Efficient router for Laminar made with Airstream
  • XStream.scala – streaming library used by Laminar before Airstream

Other building blocks of Laminar:

  • Scala DOM Types – Type definitions that we use for all the HTML tags, attributes, properties, and styles
  • Scala DOM TestUtils – Test that your Javascript DOM nodes match your expectations

Author

Nikita Gazarov – @raquo

Please support Airstream development if you use it commercially.

License

Airstream is provided under the MIT license.