22 short tests of Combine – Part 1: Protocols

I wrote some experiments around Combine, Apple’s reactive programming framework, to gain insight into how Combine handles edge cases that have caused problems for me in other reactive programming frameworks. How do subscriptions work? How do I cache computations? When are publishers and subscribers released? Under what circumstances is Combine thread-safe? Is re-entrancy possible? Does Combine guarantee delivery-order? How does Combine’s performance compare to pre-existing reactive frameworks?

Looking at everything in one article got much too long so I broke it into three parts:

  1. re-implementing the core protocols of Combine
  2. a trio of topics: shared computation, shared reference lifetimes and sharing subscribers
  3. asynchrony, threading and performance

This article will be the first third of my investigation, covering an effort to re-implement the three key protocols of Combine: Publisher, Subscriber and Subscription.

Download: The code for this series, CombineExploration, is available on github.

Warning: This is not a tutorial for Combine. I won’t be using Combine in anything resembling a conventional manner. This is going to be a look at some edge cases in Combine, testing behaviors that aren’t really documented and may therefore change in future.

Publisher and Subscriber

Apple’s Combine is built around two key protocols, Publisher and Subscriber.

The naïve intepretation of Combine is that a Publisher emits a sequence of values.

This common interpretation is not accurate but the distinction between this “naïve” interpretation and an “accurate” interpretation is rare enough that we often ignore the difference.

Publisher is defined as:

protocol Publisher {
   associatedtype Output
   associatedtype Failure : Error
   func receive<S>(subscriber: S) where S : Subscriber, Self.Failure == S.Failure, Self.Output == S.Input
}

According to the protocol, a Publisher does not emit values but receives Subscribers. Of course, depending on what a Subscriber is, the Publisher might still directly emit values into these Subscribers that it receives.

So let’s look at the Subscriber protocol for a clearer picture:

protocol Subscriber : CustomCombineIdentifierConvertible {
   associatedtype Input
   associatedtype Failure : Error
   func receive(_ input: Self.Input) -> Subscribers.Demand
   func receive(completion: Subscribers.Completion<Self.Failure>)
   func receive(subscription: Subscription)
}

Ignoring that last function for now, the other functions here on Subscriber appear to support the “naïve” interpretation: the Publisher receives Subscribers and can send Input values or Completion events directly to all known Subscribers.

Let’s establish a “control case” to which we can compare other tests, starting with a fairly standard test of the Publisher and Subscriber model where a Subscribers.Sink (a typical Subscriber) is subscribed to a PassthroughSubject (not exactly a typical Publisher but helpful in tests since it lets us inject values easily from outside) and we record the values that pass from the subject to the sink.

func testSubjectSink() {
   let subject = PassthroughSubject<Int, Never>()
   var received = [Subscribers.Event<Int, Never>]()
   let sink = Subscribers.Sink<Int, Never>(
      receiveCompletion: { received.append(.complete($0)) },
      receiveValue: { received.append(.value($0)) }
   )
   
   subject.subscribe(sink)
   subject.send(sequence: 1...3, completion: .finished)
      
   XCTAssertEqual(received, (1...3).asEvents(completion: .finished))
}

This test includes a few of my own additions to make the tests easier:

  • Subscribers.Event is just an “either” over the Value and Completion types of a Combine sequence
  • send(sequence:completion:) sends all values in the sequence and the completion
  • asEvents creates an array of Subscribers.Event from an array of Value and a Completion.

This test conforms to the “naïve” interpretation: values are sent to the subject are received by the closure we passed to the sink.

Graph mutations over time

Imagine a basic subject, A, that generates values over time (e.g. a network connection), followed by a stateful transforming node B (e.g. a scan or similar streaming processor), followed by an observer C (e.g. a Sink):

func testScan() {
   let subjectA = PassthroughSubject<Int, Never>()
   let scanB = Publishers.Scan(upstream: subjectA, initialResult: 10) { state, next in state + next }
   var receivedC = [Subscribers.Event<Int, Never>]()
   let sinkC = Subscribers.Sink<Int, Never>(
      receiveCompletion: { receivedC.append(.complete($0)) },
      receiveValue: { receivedC.append(.value($0)) }
   )

   scanB.subscribe(sinkC)
   subjectA.send(sequence: 1...4, completion: .finished)
   
   XCTAssertEqual(receivedC, [11, 13, 16, 20].asEvents(completion: .finished))
}

There’s an extra transformation line (the scanB line) but relative to the original control case it’s not much different.

Now, what happens when, halfway through A streaming its data, a new observer D subscribes to B, totally unaware that B is already in the middle of its output?

Should the new listener D get half the data it expected, even though it doesn’t know about C and the fact that the connection is already started?

The answer is complicated. Depending on your program’s logic, you may want any of the following options:

  1. multicastD receives the second half of the values that C receives
  2. caching – the first half is buffered and D immediately receives the first half of the message upon joining and new values like multicast
  3. latest valueD receives the last emitted value immediately and new values like multicast
  4. custom cachingD receives only as much as needed (e.g. since the last keyframe or resume point) and new values like multicast
  5. resubscribeD should trigger all upstream nodes to restart their work, go all the way back to the network and re-request all data, performing all calculations, again

In this article, I will focus only on the last of these options since it is, arguably, the default behavior in Combine. In the next article, I’ll look at the other approaches.

For now though, here’s an example of resubscribe behavior:

func testSequenceABCD() {
   let sequenceA = Publishers.Sequence<ClosedRange<Int>, Never>(sequence: 1...4)
   let scanB = Publishers.Scan(upstream: sequenceA, initialResult: 10) { state, next in state + next }
   var receivedC = [Subscribers.Event<Int, Never>]()
   let sinkC = Subscribers.Sink<Int, Never>(
      receiveCompletion: { receivedC.append(.complete($0)) },
      receiveValue: { receivedC.append(.value($0)) }
   )
   var receivedD = [Subscribers.Event<Int, Never>]()
   let sinkD = Subscribers.Sink<Int, Never>(
      receiveCompletion: { receivedD.append(.complete($0)) },
      receiveValue: { receivedD.append(.value($0)) }
   )

   scanB.subscribe(sinkC)
   scanB.subscribe(sinkD)
   
   XCTAssertEqual(receivedC, [11, 13, 16, 20].asEvents(completion: .finished))
   XCTAssertEqual(receivedD, [11, 13, 16, 20].asEvents(completion: .finished))
}

None of the nodes here are recreated and most importantly, the B node – the stateful scan processor – is shared between subscriptions, yet each of C and D receive an independent version of the values.

In case you think something weird is happening because the sequences don’t actually overlap in time, here’s an equivalent test where the sequences values are manually delivered in an overlapping fashion:

func testOverlappingABCD() {
   var subjects = [PassthroughSubject<Int, Never>]()
   let deferred = Deferred { () -> PassthroughSubject<Int, Never> in
      let request = PassthroughSubject<Int, Never>()
      subjects.append(request)
      return request
   }
   let scanB = Publishers.Scan(upstream: deferred, initialResult: 10) { state, next in state + next }
   var receivedC = [Subscribers.Event<Int, Never>]()
   let sinkC = Subscribers.Sink<Int, Never>(
      receiveCompletion: { receivedC.append(.complete($0)) },
      receiveValue: { receivedC.append(.value($0)) }
   )
   var receivedD = [Subscribers.Event<Int, Never>]()
   let sinkD = Subscribers.Sink<Int, Never>(
      receiveCompletion: { receivedD.append(.complete($0)) },
      receiveValue: { receivedD.append(.value($0)) }
   )

   scanB.subscribe(sinkC)
   subjects[0].send(sequence: 1...2, completion: nil)
   scanB.subscribe(sinkD)
   subjects[0].send(sequence: 3...4, completion: .finished)
   subjects[1].send(sequence: 1...4, completion: .finished)
   
   XCTAssertEqual(receivedC, [11, 13, 16, 20].asEvents(completion: .finished))
   XCTAssertEqual(receivedD, [11, 13, 16, 20].asEvents(completion: .finished))
}

This test case shows that the “naïve” interpretation of Combine cannot properly describe how Combine works in all cases. While there are two PassthroughSubjects, two Subscriber.Sinks, there is only one scanB node in the Publisher graph, yet it behaves like two completely different nodes – one for the sinkC and one for the sinkD.

Subscription, the mostly-hidden type

How does this work?

Despite the programmer creating a single graph of Publishers, there is a shadow graph of other instances that really performs the value processing and sending. We can see this shadow graph in the last function in Subscriber protocol.

func receive(subscription: Subscription)

Every Publisher in your graph is shadowed by one instance of Subscription per active Subscriber.

We didn’t see the effects of this shadow Subscription graph in the first testScan example because the shared PassthroughSubject tied all the subscriptions together but when we moved to using Deferred, the graphs become untied and independent and we could see the effects of multiple Subscriptions at the scan node.

The accurate interpretation of Combine is that values are sent and processed through a graph of Subscription instances, lazily constructed by Publisher instances on a per-subscribe basis.

We don’t usually interact with Subscription instances. Subscription instances are created automatically by Publishers when a Subscriber subscribes. The graph of Subscription instances mirrors the graph of Publishers.

You can see why the distinction between the Publisher graph and the Subscriber graph (the distinction between the “naïve” and “accurate” interpretations) can be confusing. Further adding to the confusion is that there are no usable public implementations of Subscription (I’m ignoring Subscriptions.empty which is a placeholder that ignores everything).

The subscription concept was introduced by the Reactive Extensions for .NET, attempting to make each mutation of the graph behave like a completely separate, unrelated graph – as it might appear in a strict functional programming language. However, strict functional programming languages cache function results, so redundant recalculation of upstream values is avoided. In Swift, if we don’t cache it ourselves, everything is repeated.

If I wanted to repeat all the processing, I would have recreated the publisher graph.

When I wrote my own reactive programming framework, CwlSignal, the main Signal instances were the delivery graph – the “naïve” interpretation was the same as the “accurate” interpretation. I handled the problem with multiple subscribers a different way: Signal nodes allowed only a single child to observe. For those specific cases where you need multiple listeners, CwlSignal offered a special SignalMulti node that encoded choices like “multicast”, “continuous” (cache latest), “playback” (cache all). But a re-subscribe option deliberately wasn’t offered.

In any case, let’s look under the hood at the definition of Subscription:

public protocol Subscription : Cancellable, CustomCombineIdentifierConvertible {
    func request(_ demand: Subscribers.Demand)
}

It’s pretty terse. If this is the entire definition of the shadow value sending graph, it’s not revealing much.

Custom implementations

Fortunately, Subscription isn’t impossible to understand. It generally just performs all the roles of the Publisher and Subscriber in the “naïve” interpretation: it receives values, processes them and passes them down the line.

A Subscription should replicate everything important from its associated Publisher, copying any closures and state from the initial values stored in the Publisher. In this way, the Subscription is independent and has everything needed to handle the processing, without further assistance from the Publisher.

The trickiest part is working out when to create a Subscriber from a Publisher and getting everything to piece together. I arrived at the following steps, centered on Publisher.receive, after a little experimentation:

NOTE: the words Subscriber and Subscription are very similar. I’m sure this is going to get confusing (it was confusing to write).

  1. You invoke Combine’s subscribe function on your Publisher, passing your Subscriber.
  2. This will call through to your Publisher’s receive function passing the Subscriber you provided to the subscribe function
  3. In the receive function Publisher creates a custom Subscription instance, which should also conform to Subscriber and should hold a reference to the downstream Subscriber.
  4. Your Publisher calls subscribe on its upstream Publisher (if any) passing the custom Subscription (this is why it should conform to Subscriber).
  5. The upstream Publisher calls receive on your custom Subscription, passing its own subscription instance.
  6. Your Subscriber should call receive on its downstream Subscriber
  7. The downstream Subscriber will invoke request on your Subscription and your Subscription should invoke request on its upstream Subscription.

The exact steps tend to vary based on whether your Publisher has an upstream Publisher or is a Subject.

Let’s focus on a transforming Publisher with an upstream Publisher, since that’s the canonical case. Such a Publisher would have a receive function that looks like this:

public func receive<S>(subscriber: S) where S: Subscriber, Failure == S.Failure, Output == S.Input {
   let downstream = AnySubscriber(subscriber)
   let behavior = CustomTransformer.Behavior(
      downstream: downstream,
      processor: processor,
      state: initialState)
   let subscription = CustomSubscription(behavior: behavior)
   upstreamPublisher.subscribe(subscription)
}

There’s a very important trick here: even though we are a downstream node (and could implement Subscriber for ourselves), we don’t pass ourselves to the upstreamPublisher. Instead we pass the newly constructed Subscription instance instead. This is why Subscription implementations are often Subscriber implementations too. The Subscription instances are their own, independent delivery graph, connected only to other Subscription instances.

I chose to design my custom subscription in two parts: a wrapper (to apply mutex behaviors) and a behavior protocol (used to apply Publisher specific behaviors inside the mutex). The mutex wrapper is therefore implemented just once and the behavior content is simpler.

Here’s the wrapper interface:

public struct CustomSubscription<Content: SubscriptionBehavior>: Subscriber, Subscription {
   public typealias Input = Content.Input
   public typealias Failure = Content.Failure
   
   public var combineIdentifier: CombineIdentifier { return content.combineIdentifier }
   let recursiveMutex = NSRecursiveLock()
   let content: Content
}

and the SubscriptionBehavior inside it looks like this:

public protocol SubscriptionBehavior: class, Cancellable, CustomCombineIdentifierConvertible {
   associatedtype Input
   associatedtype Failure: Error
   associatedtype Output
   associatedtype OutputFailure: Error

   var demand: Subscribers.Demand { get set }
   var upstream: Subscription? { get set }
   var downstream: AnySubscriber<Output, OutputFailure> { get }
   
   func request(_ d: Subscribers.Demand)
   func receive(_ input: Input) -> Subscribers.Demand
   func receive(completion: Subscribers.Completion<Failure>)
}

The implementations are then straightforward: values arrive via the receive functions and are processed as appropriate for the publisher that created the instance and emitted to the downstream AnySubscriber.

You can see the full implementation of CustomSubject, CustomScan, CustomSubscription and CustomSink in the CombineExploration repository.

Is this how the implemention of Subscription looks in Combine? Almost certainly not. As far as I can tell, Combine uses a type called Conduit which applies its mutex once at the start, rather than once for every Publisher stage in the pipeline. Conduit does use a recursive mutex implementation (more on that in part 3 of this series) but it appears to be implemented on top of os_unfair_lock (which is usually a non-recursive mutex).

However, these implementations do appear to behave correctly and interoperate correctly with the official Combine implementations.

Here’s the previous testOverlappingABCD rewritten with these implementations, showing that they function as drop-in replacements for the default implementations:

func testCustomABCD() {
   var subjects = [CustomSubject<Int, Never>]()
   let deferred = Deferred { () -> CustomSubject<Int, Never> in
      let request = CustomSubject<Int, Never>()
      subjects.append(request)
      return request
   }
   let scanB = CustomScan(upstream: deferred, initialResult: 10) { state, next in state + next }
   var receivedC = [Subscribers.Event<Int, Never>]()
   let sinkC = CustomSink<Int, Never>(
      receiveCompletion: { receivedC.append(.complete($0)) },
      receiveValue: { receivedC.append(.value($0)) }
   )
   var receivedD = [Subscribers.Event<Int, Never>]()
   let sinkD = CustomSink<Int, Never>(
      receiveCompletion: { receivedD.append(.complete($0)) },
      receiveValue: { receivedD.append(.value($0)) }
   )

   scanB.subscribe(sinkC)
   subjects[0].send(sequence: 1...2, completion: nil)
   scanB.subscribe(sinkD)
   subjects[0].send(sequence: 3...4, completion: .finished)
   subjects[1].send(sequence: 1...4, completion: .finished)
   
   XCTAssertEqual(receivedC, [11, 13, 16, 20].asEvents(completion: .finished))
   XCTAssertEqual(receivedD, [11, 13, 16, 20].asEvents(completion: .finished))
}

Conclusion

Download: The code for this series, CombineExploration, is available on github.

We frequently talk about our Publisher graphs as though they perform a calculation and emit values but this isn’t really true. Values in Combine are sent by the Subscription graph and the calculation is repeated for each Subscription graph.

The distinction between Publisher and Subscription graphs exists to prevent separate subscribers from interferring with each other. For this to work, all stream processing state you set up in a custom Publisher must be copied into a Subscription and mutated there, exclusively.

Looking forward…

In most cases, we don’t want redundant calculations. Where possible, we want values calculated once per Publisher graph and we want the latest value shared between all subscribers.

How do we avoid “resubscription” in Combine? How do we get multicast or cached results? Will we need to use connect or hold redundant subscribe cancellables as we do in RxSwift? For that matter, what is needed, in general, to keep Combine subscriptions alive? What are the rules by which Combine keeps anything (Publishers, Subscribers or Subscriptions) alive?

This is what I’ll look at in the next article: sharing.