Use component lifecycle hooks to make Parent + Children hierarchy code simpler and more robust
#12235
Labels
A-ECS
Entities, components, systems, and events
C-Code-Quality
A section of code that is hard to understand or change
D-Complex
Quite challenging from either a design or technical perspective. Ask for help!
S-Ready-For-Implementation
This issue is ready for an implementation PR. Go for it!
Milestone
Comments
alice-i-cecile
added
A-ECS
|
Is this just using hooks for cleanup? Like removing dangling, despawn recursive, etc? |
|
Yeah, I think that's about all we'll be able to do currently. |
|
If we don't allow users to mutably access the components directly we can enforce an always up to date hierarchy by only allowing commands to modify it. Although this would be an access pattern that I don't believe bevy currently uses anywhere. |
|
That's what I do for aery. Private |
|
I did most of the work for this a while ago as I was testing the hooks PR. |
|
Awesome, thank you @irate-devil :) And yeah, ironing out the kinks of the component hooks API is a big part of why I'm keen to do this. |
|
Doing this is going to require making |
alice-i-cecile
moved this to Active: engine observers and hooks
in Alice's Work Planning
|
EDIT: Tabbed over to the wrong issue! |
|
What is the point of using hooks if we restrict access to commands? Can't we just do the cleanup in the commands? |
BenjaminBrienen
added
the
S-Ready-For-Implementation
|
The fix for this has been marked for 0.16, probably should do the same here. |
|
There is one more advantage of using hooks for hierarchy. |
github-merge-queue bot
pushed a commit
that referenced
this issue
Jan 18, 2025
This adds support for one-to-many non-fragmenting relationships (with planned paths for fragmenting and non-fragmenting many-to-many relationships). "Non-fragmenting" means that entities with the same relationship type, but different relationship targets, are not forced into separate tables (which would cause "table fragmentation"). Functionally, this fills a similar niche as the current Parent/Children system. The biggest differences are: 1. Relationships have simpler internals and significantly improved performance and UX. Commands and specialized APIs are no longer necessary to keep everything in sync. Just spawn entities with the relationship components you want and everything "just works". 2. Relationships are generalized. Bevy can provide additional built in relationships, and users can define their own. **REQUEST TO REVIEWERS**: _please don't leave top level comments and instead comment on specific lines of code. That way we can take advantage of threaded discussions. Also dont leave comments simply pointing out CI failures as I can read those just fine._ ## Built on top of what we have Relationships are implemented on top of the Bevy ECS features we already have: components, immutability, and hooks. This makes them immediately compatible with all of our existing (and future) APIs for querying, spawning, removing, scenes, reflection, etc. The fewer specialized APIs we need to build, maintain, and teach, the better. ## Why focus on one-to-many non-fragmenting first? 1. This allows us to improve Parent/Children relationships immediately, in a way that is reasonably uncontroversial. Switching our hierarchy to fragmenting relationships would have significant performance implications. ~~Flecs is heavily considering a switch to non-fragmenting relations after careful considerations of the performance tradeoffs.~~ _(Correction from @SanderMertens: Flecs is implementing non-fragmenting storage specialized for asset hierarchies, where asset hierarchies are many instances of small trees that have a well defined structure)_ 2. Adding generalized one-to-many relationships is currently a priority for the [Next Generation Scene / UI effort](#14437). Specifically, we're interested in building reactions and observers on top. ## The changes This PR does the following: 1. Adds a generic one-to-many Relationship system 3. Ports the existing Parent/Children system to Relationships, which now lives in `bevy_ecs::hierarchy`. The old `bevy_hierarchy` crate has been removed. 4. Adds on_despawn component hooks 5. Relationships can opt-in to "despawn descendants" behavior, meaning that the entire relationship hierarchy is despawned when `entity.despawn()` is called. The built in Parent/Children hierarchies enable this behavior, and `entity.despawn_recursive()` has been removed. 6. `world.spawn` now applies commands after spawning. This ensures that relationship bookkeeping happens immediately and removes the need to manually flush. This is in line with the equivalent behaviors recently added to the other APIs (ex: insert). 7. Removes the ValidParentCheckPlugin (system-driven / poll based) in favor of a `validate_parent_has_component` hook. ## Using Relationships The `Relationship` trait looks like this: ```rust pub trait Relationship: Component + Sized { type RelationshipSources: RelationshipSources<Relationship = Self>; fn get(&self) -> Entity; fn from(entity: Entity) -> Self; } ``` A relationship is a component that: 1. Is a simple wrapper over a "target" Entity. 2. Has a corresponding `RelationshipSources` component, which is a simple wrapper over a collection of entities. Every "target entity" targeted by a "source entity" with a `Relationship` has a `RelationshipSources` component, which contains every "source entity" that targets it. For example, the `Parent` component (as it currently exists in Bevy) is the `Relationship` component and the entity containing the Parent is the "source entity". The entity _inside_ the `Parent(Entity)` component is the "target entity". And that target entity has a `Children` component (which implements `RelationshipSources`). In practice, the Parent/Children relationship looks like this: ```rust #[derive(Relationship)] #[relationship(relationship_sources = Children)] pub struct Parent(pub Entity); #[derive(RelationshipSources)] #[relationship_sources(relationship = Parent)] pub struct Children(Vec<Entity>); ``` The Relationship and RelationshipSources derives automatically implement Component with the relevant configuration (namely, the hooks necessary to keep everything in sync). The most direct way to add relationships is to spawn entities with relationship components: ```rust let a = world.spawn_empty().id(); let b = world.spawn(Parent(a)).id(); assert_eq!(world.entity(a).get::<Children>().unwrap(), &[b]); ``` There are also convenience APIs for spawning more than one entity with the same relationship: ```rust world.spawn_empty().with_related::<Children>(|s| { s.spawn_empty(); s.spawn_empty(); }) ``` The existing `with_children` API is now a simpler wrapper over `with_related`. This makes this change largely non-breaking for existing spawn patterns. ```rust world.spawn_empty().with_children(|s| { s.spawn_empty(); s.spawn_empty(); }) ``` There are also other relationship APIs, such as `add_related` and `despawn_related`. ## Automatic recursive despawn via the new on_despawn hook `RelationshipSources` can opt-in to "despawn descendants" behavior, which will despawn all related entities in the relationship hierarchy: ```rust #[derive(RelationshipSources)] #[relationship_sources(relationship = Parent, despawn_descendants)] pub struct Children(Vec<Entity>); ``` This means that `entity.despawn_recursive()` is no longer required. Instead, just use `entity.despawn()` and the relevant related entities will also be despawned. To despawn an entity _without_ despawning its parent/child descendants, you should remove the `Children` component first, which will also remove the related `Parent` components: ```rust entity .remove::<Children>() .despawn() ``` This builds on the on_despawn hook introduced in this PR, which is fired when an entity is despawned (before other hooks). ## Relationships are the source of truth `Relationship` is the _single_ source of truth component. `RelationshipSources` is merely a reflection of what all the `Relationship` components say. By embracing this, we are able to significantly improve the performance of the system as a whole. We can rely on component lifecycles to protect us against duplicates, rather than needing to scan at runtime to ensure entities don't already exist (which results in quadratic runtime). A single source of truth gives us constant-time inserts. This does mean that we cannot directly spawn populated `Children` components (or directly add or remove entities from those components). I personally think this is a worthwhile tradeoff, both because it makes the performance much better _and_ because it means theres exactly one way to do things (which is a philosophy we try to employ for Bevy APIs). As an aside: treating both sides of the relationship as "equivalent source of truth relations" does enable building simple and flexible many-to-many relationships. But this introduces an _inherent_ need to scan (or hash) to protect against duplicates. [`evergreen_relations`](https://github.com/EvergreenNest/evergreen_relations) has a very nice implementation of the "symmetrical many-to-many" approach. Unfortunately I think the performance issues inherent to that approach make it a poor choice for Bevy's default relationship system. ## Followup Work * Discuss renaming `Parent` to `ChildOf`. I refrained from doing that in this PR to keep the diff reasonable, but I'm personally biased toward this change (and using that naming pattern generally for relationships). * [Improved spawning ergonomics](#16920) * Consider adding relationship observers/triggers for "relationship targets" whenever a source is added or removed. This would replace the current "hierarchy events" system, which is unused upstream but may have existing users downstream. I think triggers are the better fit for this than a buffered event queue, and would prefer not to add that back. * Fragmenting relations: My current idea hinges on the introduction of "value components" (aka: components whose type _and_ value determines their ComponentId, via something like Hashing / PartialEq). By labeling a Relationship component such as `ChildOf(Entity)` as a "value component", `ChildOf(e1)` and `ChildOf(e2)` would be considered "different components". This makes the transition between fragmenting and non-fragmenting a single flag, and everything else continues to work as expected. * Many-to-many support * Non-fragmenting: We can expand Relationship to be a list of entities instead of a single entity. I have largely already written the code for this. * Fragmenting: With the "value component" impl mentioned above, we get many-to-many support "for free", as it would allow inserting multiple copies of a Relationship component with different target entities. Fixes #3742 (If this PR is merged, I think we should open more targeted followup issues for the work above, with a fresh tracking issue free of the large amount of less-directed historical context) Fixes #17301 Fixes #12235 Fixes #15299 Fixes #15308 ## Migration Guide * Replace `ChildBuilder` with `ChildSpawnerCommands`. * Replace calls to `.set_parent(parent_id)` with `.insert(Parent(parent_id))`. * Replace calls to `.replace_children()` with `.remove::<Children>()` followed by `.add_children()`. Note that you'll need to manually despawn any children that are not carried over. * Replace calls to `.despawn_recursive()` with `.despawn()`. * Replace calls to `.despawn_descendants()` with `.despawn_related::<Children>()`. * If you have any calls to `.despawn()` which depend on the children being preserved, you'll need to remove the `Children` component first. --------- Co-authored-by: Alice Cecile <[email protected]>
mrchantey
pushed a commit
to mrchantey/bevy
that referenced
this issue
Feb 4, 2025
This adds support for one-to-many non-fragmenting relationships (with planned paths for fragmenting and non-fragmenting many-to-many relationships). "Non-fragmenting" means that entities with the same relationship type, but different relationship targets, are not forced into separate tables (which would cause "table fragmentation"). Functionally, this fills a similar niche as the current Parent/Children system. The biggest differences are: 1. Relationships have simpler internals and significantly improved performance and UX. Commands and specialized APIs are no longer necessary to keep everything in sync. Just spawn entities with the relationship components you want and everything "just works". 2. Relationships are generalized. Bevy can provide additional built in relationships, and users can define their own. **REQUEST TO REVIEWERS**: _please don't leave top level comments and instead comment on specific lines of code. That way we can take advantage of threaded discussions. Also dont leave comments simply pointing out CI failures as I can read those just fine._ ## Built on top of what we have Relationships are implemented on top of the Bevy ECS features we already have: components, immutability, and hooks. This makes them immediately compatible with all of our existing (and future) APIs for querying, spawning, removing, scenes, reflection, etc. The fewer specialized APIs we need to build, maintain, and teach, the better. ## Why focus on one-to-many non-fragmenting first? 1. This allows us to improve Parent/Children relationships immediately, in a way that is reasonably uncontroversial. Switching our hierarchy to fragmenting relationships would have significant performance implications. ~~Flecs is heavily considering a switch to non-fragmenting relations after careful considerations of the performance tradeoffs.~~ _(Correction from @SanderMertens: Flecs is implementing non-fragmenting storage specialized for asset hierarchies, where asset hierarchies are many instances of small trees that have a well defined structure)_ 2. Adding generalized one-to-many relationships is currently a priority for the [Next Generation Scene / UI effort](bevyengine#14437). Specifically, we're interested in building reactions and observers on top. ## The changes This PR does the following: 1. Adds a generic one-to-many Relationship system 3. Ports the existing Parent/Children system to Relationships, which now lives in `bevy_ecs::hierarchy`. The old `bevy_hierarchy` crate has been removed. 4. Adds on_despawn component hooks 5. Relationships can opt-in to "despawn descendants" behavior, meaning that the entire relationship hierarchy is despawned when `entity.despawn()` is called. The built in Parent/Children hierarchies enable this behavior, and `entity.despawn_recursive()` has been removed. 6. `world.spawn` now applies commands after spawning. This ensures that relationship bookkeeping happens immediately and removes the need to manually flush. This is in line with the equivalent behaviors recently added to the other APIs (ex: insert). 7. Removes the ValidParentCheckPlugin (system-driven / poll based) in favor of a `validate_parent_has_component` hook. ## Using Relationships The `Relationship` trait looks like this: ```rust pub trait Relationship: Component + Sized { type RelationshipSources: RelationshipSources<Relationship = Self>; fn get(&self) -> Entity; fn from(entity: Entity) -> Self; } ``` A relationship is a component that: 1. Is a simple wrapper over a "target" Entity. 2. Has a corresponding `RelationshipSources` component, which is a simple wrapper over a collection of entities. Every "target entity" targeted by a "source entity" with a `Relationship` has a `RelationshipSources` component, which contains every "source entity" that targets it. For example, the `Parent` component (as it currently exists in Bevy) is the `Relationship` component and the entity containing the Parent is the "source entity". The entity _inside_ the `Parent(Entity)` component is the "target entity". And that target entity has a `Children` component (which implements `RelationshipSources`). In practice, the Parent/Children relationship looks like this: ```rust #[derive(Relationship)] #[relationship(relationship_sources = Children)] pub struct Parent(pub Entity); #[derive(RelationshipSources)] #[relationship_sources(relationship = Parent)] pub struct Children(Vec<Entity>); ``` The Relationship and RelationshipSources derives automatically implement Component with the relevant configuration (namely, the hooks necessary to keep everything in sync). The most direct way to add relationships is to spawn entities with relationship components: ```rust let a = world.spawn_empty().id(); let b = world.spawn(Parent(a)).id(); assert_eq!(world.entity(a).get::<Children>().unwrap(), &[b]); ``` There are also convenience APIs for spawning more than one entity with the same relationship: ```rust world.spawn_empty().with_related::<Children>(|s| { s.spawn_empty(); s.spawn_empty(); }) ``` The existing `with_children` API is now a simpler wrapper over `with_related`. This makes this change largely non-breaking for existing spawn patterns. ```rust world.spawn_empty().with_children(|s| { s.spawn_empty(); s.spawn_empty(); }) ``` There are also other relationship APIs, such as `add_related` and `despawn_related`. ## Automatic recursive despawn via the new on_despawn hook `RelationshipSources` can opt-in to "despawn descendants" behavior, which will despawn all related entities in the relationship hierarchy: ```rust #[derive(RelationshipSources)] #[relationship_sources(relationship = Parent, despawn_descendants)] pub struct Children(Vec<Entity>); ``` This means that `entity.despawn_recursive()` is no longer required. Instead, just use `entity.despawn()` and the relevant related entities will also be despawned. To despawn an entity _without_ despawning its parent/child descendants, you should remove the `Children` component first, which will also remove the related `Parent` components: ```rust entity .remove::<Children>() .despawn() ``` This builds on the on_despawn hook introduced in this PR, which is fired when an entity is despawned (before other hooks). ## Relationships are the source of truth `Relationship` is the _single_ source of truth component. `RelationshipSources` is merely a reflection of what all the `Relationship` components say. By embracing this, we are able to significantly improve the performance of the system as a whole. We can rely on component lifecycles to protect us against duplicates, rather than needing to scan at runtime to ensure entities don't already exist (which results in quadratic runtime). A single source of truth gives us constant-time inserts. This does mean that we cannot directly spawn populated `Children` components (or directly add or remove entities from those components). I personally think this is a worthwhile tradeoff, both because it makes the performance much better _and_ because it means theres exactly one way to do things (which is a philosophy we try to employ for Bevy APIs). As an aside: treating both sides of the relationship as "equivalent source of truth relations" does enable building simple and flexible many-to-many relationships. But this introduces an _inherent_ need to scan (or hash) to protect against duplicates. [`evergreen_relations`](https://github.com/EvergreenNest/evergreen_relations) has a very nice implementation of the "symmetrical many-to-many" approach. Unfortunately I think the performance issues inherent to that approach make it a poor choice for Bevy's default relationship system. ## Followup Work * Discuss renaming `Parent` to `ChildOf`. I refrained from doing that in this PR to keep the diff reasonable, but I'm personally biased toward this change (and using that naming pattern generally for relationships). * [Improved spawning ergonomics](bevyengine#16920) * Consider adding relationship observers/triggers for "relationship targets" whenever a source is added or removed. This would replace the current "hierarchy events" system, which is unused upstream but may have existing users downstream. I think triggers are the better fit for this than a buffered event queue, and would prefer not to add that back. * Fragmenting relations: My current idea hinges on the introduction of "value components" (aka: components whose type _and_ value determines their ComponentId, via something like Hashing / PartialEq). By labeling a Relationship component such as `ChildOf(Entity)` as a "value component", `ChildOf(e1)` and `ChildOf(e2)` would be considered "different components". This makes the transition between fragmenting and non-fragmenting a single flag, and everything else continues to work as expected. * Many-to-many support * Non-fragmenting: We can expand Relationship to be a list of entities instead of a single entity. I have largely already written the code for this. * Fragmenting: With the "value component" impl mentioned above, we get many-to-many support "for free", as it would allow inserting multiple copies of a Relationship component with different target entities. Fixes bevyengine#3742 (If this PR is merged, I think we should open more targeted followup issues for the work above, with a fresh tracking issue free of the large amount of less-directed historical context) Fixes bevyengine#17301 Fixes bevyengine#12235 Fixes bevyengine#15299 Fixes bevyengine#15308 ## Migration Guide * Replace `ChildBuilder` with `ChildSpawnerCommands`. * Replace calls to `.set_parent(parent_id)` with `.insert(Parent(parent_id))`. * Replace calls to `.replace_children()` with `.remove::<Children>()` followed by `.add_children()`. Note that you'll need to manually despawn any children that are not carried over. * Replace calls to `.despawn_recursive()` with `.despawn()`. * Replace calls to `.despawn_descendants()` with `.despawn_related::<Children>()`. * If you have any calls to `.despawn()` which depend on the children being preserved, you'll need to remove the `Children` component first. --------- Co-authored-by: Alice Cecile <[email protected]>
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
What problem does this solve or what need does it fill?
bevy_hierarchy's internals are complex, performance-critical, and have panicking unsafe code in them.The core reason for this is that we cannot let the graph become invalid: this leads to wasted work, subtle bugs and huge amounts of user frustration. And because we use
unsafecode internally that relies on the assumption of validity for soundness, this can become a soundness problem.They're painful to maintain, and nearly impossible to replicate externally for other hierarchy-like code that wants to uphold these same variants.
What solution would you like?
#10756 adds component lifecycle hooks: operations that take effect when components are added or removed.
We should replace the existing complex code in
bevy_hierarchywith this idiom, being sure to benchmark and optimize performance.What alternative(s) have you considered?
We could wait until the fabled day that relations (#3742) finally gets shipped.
Additional context
This has been widely anticipated as one of the key benefits of that PR: I'm recording it here to make sure we don't forget (and can coordinate work).
While the user-facing API is also pretty gnarly and inconsistent, I'd like to leave that for future work to make sure that the correctness-sensitive internals get the review attention it needs.
The text was updated successfully, but these errors were encountered: