required components with constructors like FromWorld

[tomaspecl]

What problem does this solve or what need does it fill?

Problem is when the required component should be initialized using the data from the World.

What alternative(s) have you considered?

You can have the required component first be created with a placeholder value and then use a system or on_add component hook to initialize it correctly.

Additional context

Currently I am writing a rollback plugin, I want to use the disabling components to "virtualy" despawn an entity so that it can be put back when time is rolled back to a point when the entity was supposed to still exist. Currently I have this:

/// put this marker disabling component on a rollback entity to virtualy despawn it
#[derive(Component)]
#[require(DespawnedSince(Frame(u64::MAX)))]    //the u64::MAX is placeholder value until the correct value is loaded from World
pub struct Despawn;

/// when the entity is virtualy despawed for too long (until snapshot storage length)
/// it will despawn properly as the other data needed to restore it is not stored anymore
/// and its assumed that rolling back too far will not happen/will not be needed
#[derive(Component)]
pub struct DespawnedSince(pub Frame);

pub struct RollbackDespawnPlugin;

impl Plugin for RollbackDespawnPlugin {
    fn build(&self, app: &mut App) {
        // currently I have to modify the component after its created by required component
        app.world_mut().register_component_hooks::<DespawnedSince>()
        .on_add(|mut world, context| {
            let Some(&frame) = world.get_resource::<Frame>() else{
                warn!("can not initialize DespawnedSince component, the Frame resource does not exist");
                return
            };
            world.entity_mut(context.entity).get_mut::<DespawnedSince>().unwrap().0 = frame;
        });
        ....
    }

What solution would you like?

the world.register_required_components_with(constructor_function) should take a function that takes &World (or maybe DeferredWorld) so that the constructor can initialize the component using world data.

world.register_required_components could instead of taking a component that implements Default, take component that implements FromWorld, as FromWorld is automatically implemented for any type implementing Default, it should work nice. And then it would internally call the world.register_required_components_with using that FromWorld::from_world() implementation instead of Default::default().

The try_ versions should be also changed the same way.

Then I could specify that the DespawnedSince required component is initialized correctly from the begining and I would not need to first have it be made with a placeholder and then overwritten, it seems logical to me that it should be possible when a costructor function is passed in already, it should allow taking arguments for initialization.

[alice-i-cecile]

The other wrinkle to this is that I would really like it to be easy to take &self in required components, and construct values that are derived from the initial component value.

[SkiFire13]

The other wrinkle to this is that I would really like it to be easy to take &self in required components, and construct values that are derived from the initial component value.

On the other hand, if a component "requires" another component I would the expect the required component to be constructed first (including the hooks, cc #20001), which is the opposite order to what's needed for your proposal.

[tomaspecl]

The other wrinkle to this is that I would really like it to be easy to take &self in required components, and construct values that are derived from the initial component value.

The other wrinkle to this is that I would really like it to be easy to take &self in required components, and construct values that are derived from the initial component value.

On the other hand, if a component "requires" another component I would the expect the required component to be constructed first (including the hooks, cc #20001), which is the opposite order to what's needed for your proposal.

how about when you insert a component (call it struct A) into an entity that component is first stored somewhere (the place where it should be placed, I dont know, some table or whatever), then the required component (struct B) constructor is run with the &World or DeferredWorld, that way it can access what it needs including the A, the hooks for B are run after that and then hooks for A are run.

that way you can use values from A in the construction of required component B and also finish init of A using values from B

[tomaspecl]

but I am not sure what happens when A requires B and B requires C. Maybe its just inconsistent idea.

[tomaspecl]

or @alice-i-cecile please describe more precisely what you meant by

take &self in required components

I assume &self was meant the component inserted that triggered the insertion of the required component

[alice-i-cecile]

I assume &self was meant the component inserted that triggered the insertion of the required component

This was my intent. I don't know if it can literally be &self.

On the other hand, if a component "requires" another component I would the expect the required component to be constructed first (including the hooks, cc #20001), which is the opposite order to what's needed for your proposal.

That makes sense to me, but that feels like a naming issue with "required components". I tend to think about required components as flowing down from the initially inserted value, where it makes sense that the original value is available to other components spawned in follow-up.

[tomaspecl]

@alice-i-cecile so what about when having A -> B -> C (A->B means A requires B), and A is inserted onto an entity, first A gets placed there, then B constructor runs and places B, then C constructor runs and places C, then hooks run in reverse so C hook, B hook, A hook.

But I dont know how the ECS works under the hood while the A,B,C are being initialized, I would assume you would have some optimization such that it knows it should allocate entity that will have space for A, B, C and they are initialized one after another but when A gets initialized B and C are still Uninit or something so their state is Uninit, so the entire entity is also Uninit, therefore you cant read the entity data before it gets completed fully (fully initialized). And that would make it impossible to run the B and C constructor with World access as it would be unsound, it would be UB or something

[alice-i-cecile]

I'm not sure! I haven't given a ton of thought with respect to the optimal hook / observer ordering. @SanderMertens may have insight or opinions here based on flecs though.