Track spawn Tick of entities, offer methods, query data SpawnDetails and query filter Spawned

[urben1680]

Objective

In my own project I was encountering the issue to find out which entities were spawned after applying commands. I began maintaining a vector of all entities with generational information before and after applying the command and diffing it. This was awfully complicated though and has no constant complexity but grows with the number of entities.

Solution

Looking at EntyMeta it seemed obvious to me that struct can track the tick just as it does with MaybeLocation, updated from the same call. After that it became almost a given to also introduce query data SpawnDetails which offers methods to get the spawn tick and location, and query filter Spawned that filters entities out that were not spawned since the last run.

Testing

I expanded a few tests and added new ones, though maybe I forgot a group of tests that should be extended too. I basically searched bevy_ecs for mentions of Changed and Added to see where the tests and docs are.

Benchmarks of spawn/despawn can be found here.

---

Showcase

From the added docs, systems with equal complexity since the filter is not archetypal:

fn system1(q: Query<Entity, Spawned>) {
    for entity in &q { /* entity spawned */ }
}

fn system2(query: Query<(Entity, SpawnDetails)>) {
    for (entity, spawned) in &query {
        if spawned.is_spawned() { /* entity spawned */ }
    }
}

SpawnedDetails has a few more methods:

fn print_spawn_details(query: Query<(Entity, SpawnDetails)>) {
    for (entity, spawn_details) in &query {
        if spawn_details.is_spawned() {
            print!("new ");
        }
        println!(
            "entity {:?} spawned at {:?} by {:?}",
            entity,
            spawn_details.spawned_at(),
            spawn_details.spawned_by()
        );        
    }
}

Changes

No public api was changed, I only added to it. That is why I added no migration guide.

• query data SpawnDetails
• query filter Spawned
• method Entities::entity_get_spawned_or_despawned_at
• method EntityRef::spawned_at
• method EntityMut::spawned_at
• method EntityWorldMut::spawned_at
• method UnsafeEntityCell::spawned_at
• method FilteredEntityRef::spawned_at
• method FilteredEntityMut::spawned_at
• method EntityRefExcept::spawned_at
• method EntityMutExcept::spawned_at

[alice-i-cecile]

I really like this idea: it's useful and consistent. That said, we should see how this affects spawning performance, both in the worst case and on more realistic cases.

[ElliottjPierce]

Over all, looks good. Just some minor questions, etc.

One thing we should consider is benching this. I have a feeling that having a separate vec for SpawnedOrDespawned, separate from EntityMeta would be faster. Almost every iteration cares about where the entity is. Asking when it was spawned and by who is more like a map thing IIUC. But this might be better as a separate PR.

Another option is making an observer and event for Spawned. I feel like that's going to be the primary use of this, and exposing an observer might be better anyway for this use case. Then again, it makes sense to ask when an entity was spawned. Ex: This particle entity has been around for 100 ticks; despawn it. (Not that each particle should be an entity.) In other words, this might be a cheaper alternative to a timer component in some cases.

Anyway, that's my thoughts. I'm generally on board with this, provided the separate vecs and observer APIs are investigated (maybe in future prs).

[urben1680]

Regarding introducing an event for observers, I considered it but I think the PR is big enough as is. I also imagine spawns triggering observers will have a more noticeable performance impact on all spawns than what I did here.

[urben1680]

I did another change.

Previously, Entities::set_spawned_or_despawned_by was called at the end of spawns/despawns, which was fine for logging something you usually deactivate during releases.

However, since with this change the method is no longer noop even when you deactivate tracking locations, this always added to the number of accesses on Entities.

So I removed that method entirely and added Entities::set_spawn_despawn that is used instead of Entities::set in the spawning/despawning context.

Entities::set is still used at

• BundleInserter::insert
• EntityWorldMut::move_entity_from_remove (which makes me think Add BundleRemover #18521 does not need to adjust here)

Entities::set_spawn_despawn is now used instead at

• BundleSpawner::spawn_non_existent
• EntityWorldMut::despawn_with_caller (I don't know why this method has two set calls, please tell me)
• World::spawn_at_empty_internal

When doing spawn benchmarks, this change greatly reduced the regressions from up to +6% down to +/-1%.
Despawn benchmarks are still a bit wildly going in +/- 5%.

I will do a separate post with benchmarks once I am sure I do them correctly.

[urben1680]

I wrote release notes. I wanted to do it detailed, I hope not too much.

rendered (up to date to following commits)

I was also close to give a short recap that Ticks becoming too old is something to keep in mind but I was not sure if this belongs there.

[NthTensor]

Oh these are perfect! Thank you!!!

[alice-i-cecile]

Good stuff! Thanks for the release note; it'll need an editing pass but capturing the motivation and usage is the key thing at this stage.

[alice-i-cecile]

Merge conflicts, sorry!

[urben1680]

@alice-i-cecile resolved them. 👍

[akimakinai]

caller is no longer set for entities not swap_removed? #21293

_{(sorry, I commented this on merge commit initially)}