swift-composable-architecture-mirror/Sources/ComposableArchitecture/Documentation.docc/Articles/Performance.md

Performance

Learn how to improve the performance of features built in the Composable Architecture.

As your features and application grow you may run into performance problems, such as reducers becoming slow to execute, SwiftUI view bodies executing more often than expected, and more. This article outlines a few common pitfalls when developing features in the library, and how to fix them.

• Sharing logic with actions
• CPU-intensive calculations
• High-frequency actions
• Store scoping

Sharing logic with actions

There is a common pattern of using actions to share logic across multiple parts of a reducer. This is an inefficient way to share logic. Sending actions is not as lightweight of an operation as, say, calling a method on a class. Actions travel through multiple layers of an application, and at each layer a reducer can intercept and reinterpret the action.

It is far better to share logic via simple methods on your Reducer conformance. The helper methods can take inout State as an argument if it needs to make mutations, and it can return an Effect<Action>. This allows you to share logic without incurring the cost of sending needless actions.

For example, suppose that there are 3 UI components in your feature such that when any is changed you want to update the corresponding field of state, but then you also want to make some mutations and execute an effect. That common mutation and effect could be put into its own action and then each user action can return an effect that immediately emits that shared action:

@Reducer
struct Feature {
  @ObservableState
  struct State { /* ... */ }
  enum Action { /* ... */ }

  var body: some Reducer<State, Action> {
    Reduce { state, action in
      switch action {
      case .buttonTapped:
        state.count += 1
        return .send(.sharedComputation)

      case .toggleChanged:
        state.isEnabled.toggle()
        return .send(.sharedComputation)

      case let .textFieldChanged(text):
        state.description = text
        return .send(.sharedComputation)

      case .sharedComputation:
        // Some shared work to compute something.
        return .run { send in
          // A shared effect to compute something
        }
      }
    }
  }
}

This is one way of sharing the logic and effect, but we are now incurring the cost of two actions even though the user performed a single action. That is not going to be as efficient as it would be if only a single action was sent.

Besides just performance concerns, there are two other reasons why you should not follow this pattern. First, this style of sharing logic is not very flexible. Because the shared logic is relegated to a separate action it must always be run after the initial logic. But what if instead you need to run some shared logic before the core logic? This style cannot accommodate that.

Second, this style of sharing logic also muddies tests. When you send a user action you have to further assert on receiving the shared action and assert on how state changed. This bloats tests with unnecessary internal details, and the test no longer reads as a script from top-to-bottom of actions the user is taking in the feature:

let store = TestStore(initialState: Feature.State()) {
  Feature()
}

store.send(.buttonTapped) {
  $0.count = 1
}
store.receive(\.sharedComputation) {
  // Assert on shared logic
}
store.send(.toggleChanged) {
  $0.isEnabled = true
}
store.receive(\.sharedComputation) {
  // Assert on shared logic
}
store.send(.textFieldChanged("Hello")) {
  $0.description = "Hello"
}
store.receive(\.sharedComputation) {
  // Assert on shared logic
}

So, we do not recommend sharing logic in a reducer by having dedicated actions for the logic and executing synchronous effects.

Instead, we recommend sharing logic with methods defined in your feature's reducer. The method has full access to all dependencies, it can take an inout State if it needs to make mutations to state, and it can return an Effect<Action> if it needs to execute effects.

The above example can be refactored like so:

@Reducer
struct Feature {
  @ObservableState
  struct State { /* ... */ }
  enum Action { /* ... */ }

  var body: some Reducer<State, Action> {
    Reduce { state, action in
      switch action {
      case .buttonTapped:
        state.count += 1
        return self.sharedComputation(state: &state)

      case .toggleChanged:
        state.isEnabled.toggle()
        return self.sharedComputation(state: &state)

      case let .textFieldChanged(text):
        state.description = text
        return self.sharedComputation(state: &state)
      }
    }
  }

  func sharedComputation(state: inout State) -> Effect<Action> {
    // Some shared work to compute something.
    return .run { send in
      // A shared effect to compute something
    }
  }
}

This effectively works the same as before, but now when a user action is sent all logic is executed at once without sending an additional action. This also fixes the other problems we mentioned above.

For example, if you need to execute the shared logic before the core logic, you can do so easily:

case .buttonTapped:
  let sharedEffect = self.sharedComputation(state: &state)
  state.count += 1
  return sharedEffect

You have complete flexibility to decide how, when and where you want to execute the shared logic.

Further, tests become more streamlined since you do not have to assert on internal details of shared actions being sent around. The test reads like a user script of what the user is doing in the feature:

let store = TestStore(initialState: Feature.State()) {
  Feature()
}

store.send(.buttonTapped) {
  $0.count = 1
  // Assert on shared logic
}
store.send(.toggleChanged) {
  $0.isEnabled = true
  // Assert on shared logic
}
store.send(.textFieldChanged("Hello") {
  $0.description = "Hello"
  // Assert on shared logic
}

Sharing logic in child features

There is another common scenario for sharing logic in features where the parent feature wants to invoke logic in a child feature. One can technically do this by sending actions from the parent to the child, but we do not recommend it (see above in doc:Performance#Sharing-logic-with-actions to learn why):

// Handling action from parent feature:
case .buttonTapped:
  // Send action to child to perform logic:
  return .send(.child(.refresh))

Instead, we recommend invoking the child reducer directly:

case .buttonTapped:
  return Child().reduce(into: &state.child, action: .refresh)
    .map(Action.child)

CPU intensive calculations

Reducers are run on the main thread and so they are not appropriate for performing intense CPU work. If you need to perform lots of CPU-bound work, then it is more appropriate to use an Effect, which will operate in the cooperative thread pool, and then send actions back into the system. You should also make sure to perform your CPU intensive work in a cooperative manner by periodically suspending with Task.yield() so that you do not block a thread in the cooperative pool for too long.

So, instead of performing intense work like this in your reducer:

case .buttonTapped:
  var result = // ...
  for value in someLargeCollection {
    // Some intense computation with value
  }
  state.result = result

...you should return an effect to perform that work, sprinkling in some yields every once in awhile, and then delivering the result in an action:

case .buttonTapped:
  return .run { send in
    var result = // ...
    for (index, value) in someLargeCollection.enumerated() {
      // Some intense computation with value

      // Yield every once in awhile to cooperate in the thread pool.
      if index.isMultiple(of: 1_000) {
        await Task.yield()
      }
    }
    await send(.computationResponse(result))
  }

case let .computationResponse(result):
  state.result = result

This will keep CPU intense work from being performed in the reducer, and hence not on the main thread.

High-frequency actions

Sending actions in a Composable Architecture application should not be thought as simple method calls that one does with classes, such as ObservableObject conformances. When an action is sent into the system there are multiple layers of features that can intercept and interpret it, and the resulting state changes can reverberate throughout the entire application.

Because of this, sending actions does come with a cost. You should aim to only send "significant" actions into the system, that is, actions that cause the execution of important logic and effects for your application. High-frequency actions, such as sending dozens of actions per second, should be avoided unless your application truly needs that volume of actions in order to implement its logic.

However, there are often times that actions are sent at a high frequency but the reducer doesn't actually need that volume of information. For example, say you were constructing an effect that wanted to report its progress back to the system for each step of its work. You could choose to send the progress for literally every step:

case .startButtonTapped:
  return .run { send in
    var count = 0
    let max = await self.eventsClient.count()

    for await event in self.eventsClient.events() {
      defer { count += 1 }
      await send(.progress(Double(count) / Double(max)))
    }
  }
}

However, what if the effect required 10,000 steps to finish? Or 100,000? Or more? It would be immensely wasteful to send 100,000 actions into the system to report a progress value that is only going to vary from 0.0 to 1.0.

Instead, you can choose to report the progress every once in awhile. You can even do the math to make it so that you report the progress at most 100 times:

case .startButtonTapped:
  return .run { send in
    var count = 0
    let max = await self.eventsClient.count()
    let interval = max / 100

    for await event in self.eventsClient.events() {
      defer { count += 1 }
      if count.isMultiple(of: interval) {
        await send(.progress(Double(count) / Double(max)))
      }
    }
  }
}

This greatly reduces the bandwidth of actions being sent into the system so that you are not incurring unnecessary costs for sending actions.

Another example that comes up often is sliders. If done in the most direct way, by deriving a binding from the store to hand to a Slider:

Slider(value: store.$opacity, in: 0...1)

This will send an action into the system for every little change to the slider, which can be dozens or hundreds of actions as the user is dragging the slider. If this turns out to be problematic then you can consider alternatives.

For example, you can hold onto some local @State in the view for using with the Slider, and then you can use the trailing onEditingChanged closure to send an action to the store:

Slider(value: self.$opacity, in: 0...1) {
  self.store.send(.setOpacity(self.opacity))
}

This way an action is only sent once the user stops moving the slider.

Store scoping

In the 1.5.6 release of the library a change was made to Store/scope(state:action:)-90255 that made it more sensitive to performance considerations.

The most common form of scoping, that of scoping directly along boundaries of child features, is the most performant form of scoping and is the intended use of scoping. The library is slowly evolving to a state where that is the only kind of scoping one can do on a store.

The simplest example of this directly scoping to some child state and actions for handing to a child view:

ChildView(
  store: store.scope(state: \.child, action: \.child)
)

Furthermore, scoping to a child domain to be used with one of the libraries navigation view modifiers, such as SwiftUI/View/sheet(store:onDismiss:content:), also falls under the intended use of scope:

.sheet(store: store.scope(state: \.child, action: \.child)) { store in
  ChildView(store: store)
}

All of these examples are how Store/scope(state:action:)-90255 is intended to be used, and you can continue using it in this way with no performance concerns.

Where performance can become a concern is when using scope on computed properties rather than simple stored fields. For example, say you had a computed property in the parent feature's state for deriving the child state:

extension ParentFeature.State {
  var computedChild: ChildFeature.State {
    ChildFeature.State(
      // Heavy computation here...
    )
  }
}

And then in the view, say you scoped along that computed property:

ChildView(
  store: store.scope(state: \.computedChild, action: \.child)
)

If the computation in that property is heavy, it is going to become exacerbated by the changes made in 1.5, and the problem worsens the closer the scoping is to the root of the application.

The problem is that in version 1.5 scoped stores stopped directly holding onto their local state, and instead hold onto a reference to the store at the root of the application. And when you access state from the scoped store, it transforms the root state to the child state on the fly.

This transformation will include the heavy computed property, and potentially compute it many times if you need to access multiple pieces of state from the store. If you are noticing a performance problem while depending on 1.5+ of the library, look through your code base for any place you are using computed properties in scopes. You can even put a print statement in the computed property so that you can see first hand just how many times it is being invoked while running your application.

To fix the problem we recommend using Store/scope(state:action:)-90255 only along stored properties of child features. Such key paths are simple getters, and so not have a problem with performance. If you are using a computed property in a scope, then reconsider if that could instead be done along a plain, stored property and moving the computed logic into the child view. The further you push the computation towards the leaf nodes of your application, the less performance problems you will see.