How To

Design systems

Systems must be the only authoring point for any logic executed on entities. It's strictly forbidden to initiate any manipulation with entities outside of systems. Though, they can call into their dependencies and pieces of logic isolated in their own files and classes.

Systems can't contain any collections of components or entities that persist through multiple frames: everything should be stored in ECS directly.

Systems can contain temporary collections used for data aggregation: take a look at DeferredLoadingSystem.cs.

Systems can't contain a state. All states should be written and stored in ECS Worlds.

How to construct

Systems should have an internal constructor, it clearly indicates that we can't instantiate the system directly but are obliged to use ArchSystemsWorldBuilder.

Systems may accept shared dependencies in the .ctor such as:

• Settings that apriori exist in a single instance: Quality, Partitioning, etc.
• Pool Providers
• Utility functionality (that for some reason is not static)
• Configuration dedicated to the given system only, strategies and factories that are injected from the upper level: e.g. IConcurrentBudgetProvider

Every system should be inherited from BaseUnityLoopSystem: it provides common functionality for profiling and error reporting.

Follow Single Resposibility

Every system should execute a limited scope of responsibilities. It should be reflected in its name. There is no strict rule of how many queries it should have but if it grows beyond 200 lines of code consider splitting it into static counterparts.

Normally, every feature is represented by multiple systems that are bound by a certain execution order.

Distribute in groups

Decide in which Game Loop moment (SystemGroup) it will be executed. It purely depends on the system's designation, e.g.:

• Physics manipulation should happen in PhysicsSystemGroup
• Actions based on Transform.position or Transform.rotation - in PresentationSystemGroup as it is executed after transformation is applied in Unity

Consider creating your own group for a given feature: it will simplify defining dependencies between other groups and systems

How to write queries

There are four ways of writing queries:

1. Automatic generation is the most preferred one. It can be used in the systems only. But if you have a generic system it's impossible to use it as generic attributes are not supported in the version of C# used in Unity.
2. Iterating over chunks manually: GetChunkIterator(). The same code is generated by the source generator. You can consider this option in a generic class in some special cases.
3. World.InlineQuery can be used outside of the system itself and in generic cases. Its performance is very close to generated queries. See ReleasePoolableComponentSystem<T, TProvider> for a reference.
4. World.Query is the least preferred way of doing things as it uses delegates and can lead to closures unintentionally.

Good practices

• Filter out by DeleteEntityIntention: it's undesirable to execute logic over entities marked for destruction

Performance Implication

System's Update should be allocation-free. In order to ensure this consider profiling before sending a feature for review.

When you define a system that operates in a scene context (not a global world) there will be as many instances of this system as worlds are loaded. Thus its Update may be executed many times in the same frame. You should keep the logic as simple as possible so every step of the system takes negligible time.

Every query produces an overhead. Try avoiding introducing multiple queries in the same system with the same filter. Invoke several different methods from one handler instead:

• e.g. take a look at CalculateCharacterVelocitySystem: it uses a single entry point ResolveVelocity to calculate every kind of velocity in isolated pieces of logic: ApplyCharacterMovementVelocity, ApplyJump, ApplyGravity, ApplyAirDrag, etc.
• another example is FinalizeGltfContainerLoadingSystem: in FinalizeLoading the static method ConfigureGltfContainerColliders.SetupColliders is called to execute logic encapsulated in its own class

Throttling

In order to optimize it further there is a concept of throttling: Systems registered in a Scene World do not execute unless there is a CRDT change from the JavaScript scene. This behavior is implemented in SystemGroupsUpdateGate.cs.

Throttling must be enabled manually by annotating with ThrottlingEnabled attribute. Not every system is suitable for throttling: for example Promises resolution should happen as soon as data is ready.

Enabling Throttling will significantly relieve CPU pressure.

How to manipulate components

World.Get API and Queries provide a ref access to the component. It makes it possible to modify a value type directly without the necessity of setting it back.

⚠️ You must use ref var, otherwise the value will be copied and changes won't be reflected, e.g.: ref var meshRendererComponent = ref world.Get<PrimitiveMeshRendererComponent>(entity);

⚠️ A severe ECS pitfall you may fall into: E.g. you have a query

        protected void TestQuery(in Entity entity, ref StreamableLoadingState state)
        {
            World.Add(entity, new StreamableLoadingResult<TAsset>());
            state.Value = StreamableLoadingState.Status.Finished;
        }

state.Value = StreamableLoadingState.Status.Finished; will not apply the change to the value you expect because you make a structural change (World.Add) before that line and moving between archetypes invalidates ref StreamableLoadingState state

You should be very cautious and apply all structural changes last!

It's super hard to detect as ref StreamableLoadingState state will not throw any exception but will silently point to another cell (in fact the same cell but affected by memcpy) in the reserved array in the archetype's chunk.

So the change will apply eventually to an indefinite component: lucky you if it is just an empty reserved cell but it can be also another valid entity that will be modified accidentally!

How to test systems

UnitySystemTestBase<TSystem> provides basic functionality for world creation and disposal.

In Tests you can create systems directly by calling a constructor. Consider exposing them by [InternalsVisibleTo] to tests.

Design components

• In terms of ECS there is no difference between SDK components (from Proto) and written by us
• If you need to enrich an entity (created with an SDK component) with additional data create a separate component: by filtering you will be able to recognize which components are not processed yet
• Keep the balance between separate components and state:
  • structural changes are expensive operations, if the logic supposes frequent/uncontrolled Adding or Removing components, it's preferred to have a single component and change its state instead.
  • otherwise, it's advised to maintain a reasonable segregation and responsibilities distribution between different components
• If you need to wait for data that is retrieved asynchronously create an AssetPromise<TAsset, TLoadingIntention>, e.g.:
  • Asset Bundles
  • GLTF
  • Textures
  • Any other data from web requests
• You may have as many AssetPromises as needed and store them in a component or add them to an entity directly. Keep in mind it's a value type as well so whatever you do, ensure you operate with it by ref, otherwise the state won't be reflected.

Singletons

• Some components can be natural singletons (e.g. PhysicsTickComponent): in our case, it means they exist in a single instance per World
• They are created by systems in ctor or in Initialize
• Then they can be used in Update by other systems
• Instead of making a query every time such a component is needed consider caching it and save into a SingleInstanceEntity field