-
-
Notifications
You must be signed in to change notification settings - Fork 3.7k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
[Merged by Bors] - System sets and run criteria v2 #1675
Conversation
`RunCriteriaDescriptor` prototype; mass system description superpowers for `SystemSet`; removed `VirtualSystemSet`; individual systems can now have run criteria; centralized all run criteria of `SystemStage`; "criteria" -> "criterion" where singular.
# Conflicts: # crates/bevy_ecs/src/lib.rs # crates/bevy_ecs/src/schedule/system_set.rs # examples/game/alien_cake_addict.rs
Relevant to this question: #1295. I think we leave this for another PR and keep this scope small so we can try to get this out ASAP to fix bugs. |
Personally I think this is very important to reinforce the idea that run criteria and systems are separate things. Also, this trivially solves the second problem.
As a first pass, I'd like to propose |
|
In my view, I would expect most run criteria to be reused, especially as games grow. So I think my ideal API would look like: app
.add_run_criteria(FixedTimeStep{ 2.0 }, MyRunCriteria::EveryTwo)
.add_system(blinken_lights.system()).with_run_criteria(MyRunCriteria::EveryTwo)
.add_system_set([
forces.system(),
physics.system(),
]).with_run_criteria(MyRunCriteria::EveryTwo); With distinct Edit: Ideally, I'd love to be able to attach run criteria to system labels as well, allowing you to tie run criteria behavior to a label directly. I would prefer this to being able to attach run criteria to system sets at all, provided we can label each system in a system set (I haven't poked at the new labelling API deeply). That might be out of scope though and I could live without it. |
Yes. I think we can do this and use familiar APIs.
I'm not sure I see how?..
Maybe; I'll see how that feels in tests. What's in now already returns a type that can have special APIs on it: the
Yes, something close to that should be possible. However, arrays of systems coercing into system sets is not trivial, and not because const generics:
"Define stuff by doing things to a label" is on the docket as "future API R&D" - after stageless, most likely. |
While arrays are definitely problematic, we should be able to make something that works using tuples. With a trait for these systems, and tuples of these systems, we could have far more elegant description syntax. |
More "I wish we had variadic generics" macros ;) I like this idea though: tuples are better than nothing. |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
This generally looks good so far. I don't think we should merge it before 0.5 unless it:
- passes existing state tests (as written)
- solves State-dependent system runs twice on wrong
add_state
call order #1672 (i don't think Hard lock using State across stages #1671 needs solving before 0.5 and i dont really see how this pr could help there). It looks like the State-dependent system runs twice on wrongadd_state
call order #1672 solution would be to:- Add a new
StateDriver<T>
run criteria label, which is automatically added to the driver criteria - Make
State::<T>::on_enter()
return aRunCriteriaDescriptor
(StateDriverLabel<T>::new().chain(State::<T>::current_on_enter_impl()
)
- Add a new
Its a little odd that we're making criteria chaining kind of like after()
and kind of like chain()
. It feels like consistency would be a good thing there.
If criteria can be ordered (and use labels), imo we should use before()
and after()
for those cases and leave chain()
for system chaining.
Yes, it is a bit odd, but it's doing something that we don't have a precedent for: "chaining" systems without welding them together into one system. Without a syntax for this the alternative is to use the regular decoupling techniques, like events or channels or any other kind of buffer resource, all of which are more cumbersome. It bears mentioning that this does not actually replace regular chaining, so I think a more distinct name would be better. I already have |
I definitely dislike using the chain name here when the behaviour is so different. It's also not super clear what "chain" might do at all to a new user. If we need the behaviour, what about |
Discussion on Discord so far seems to converge towards |
# Conflicts: # crates/bevy_ecs/src/schedule/mod.rs # crates/bevy_ecs/src/schedule/stage.rs
I don't love this behavior :/ It's very implicit and hard to wrap your head around. Which run criteria label is the "first" one in the case of duplicates?
Why do we have duplicate run criteria labels at all? This being silent seems very dangerous and hard to debug: I'd prefer a panic over it working (unless there's a good reason the duplicates can't be avoided). Can we avoid using labels at all here? E.g. from this snippet: .with_run_criteria(RunCriteria::pipe("every other time", eot_piped.system())), Why not just use .with_run_criteria(eot.system().pipe(eot_piped.system()))), and duplicate the run criteria? |
The silentness allows for the auto-labelling of state run criteria, which fixes ordering bugs and deduplicates unnecessary run criteria. For "why labels", these run criteria aren't guaranteed to be side effect free and the state driver is an example of a run criteria with side effects. |
These are good answers; hopefully we can revisit this with a more elegant fix later. |
I don't see how: "criteria labels must be unique" is as simple a rule as it gets. It should be expanded with "if you define a criteria with the same label as some other criteria it will be discarded" to fully document the behavior, though. Looking at it now, I think this could be split to be explicit: panic on duplicate labels by default, discard criteria with duplicate labels as an opt-in.
I've reiterated this several times already and I was certain we were all on the same page with regards to what the piping operation does, exactly. I don't know how else I could explain why it works via a label, but I'll try again anyway: A run criteria is defined somewhere. We want to use its result somewhere else. We attach a label to that first definition, then refer to the criteria by that label in the somewhere else. If we didn't use a label but the criteria itself again, we would have two copies of the same function running every tick. With piping, that "somewhere else" is another criteria. If we want that behavior but are okay with having duplicate code, we use a chain instead. |
let run_criteria_label = descriptor | ||
.run_criteria | ||
.take() | ||
.map(|criteria| self.process_run_criteria(criteria)) |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I expect this behavior to be a bit surprising for users. People using a SystemSet with RunCriteria probably did that with the intent to apply the criteria to all systems (ex: run all of these systems with a fixed timestep).
As a user, I would expect adding criteria to a system inside one of these sets to apply that criteria "in addition to" the system set criteria (ex: every_other_time
criteria on a system should run every other time a fixed timestep occurs).
As a Bevy internals developer, I understand the run criteria implementation and why it can't do that, but that doesn't change the fact that users will be surprised when the "every_other_run" criteria overrides the fixed timestep criteria.
After this pr I would like us to explore the idea of multiple run criteria (ShouldRun::Yes + ShouldRun::YesAndLoop == ShouldRunYesAndLoop, ShouldRun::Yes + ShouldRun::No == ShouldRun::No, etc), which would hopefully solve this problem.
In the interim, I almost want to make this panic because it goes against user expectations, which we ourselves have already set due to how things like "SystemSet labels" now work (system set labels and system labels are combined).
Users hitting the panic could just move the affected system to a separate system set (or add it without a set)
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
As a user, I would expect adding criteria to a system inside one of these sets to apply that criteria "in addition to" the system set criteria (ex: every_other_time criteria on a system should run every other time a fixed timestep occurs).
I definitely agree here.
After this pr I would like us to explore the idea of multiple run criteria (ShouldRun::Yes + ShouldRun::YesAndLoop == ShouldRunYesAndLoop, ShouldRun::Yes + ShouldRun::No == ShouldRun::No, etc), which would hopefully solve this problem.
@TheRawMeatball, @Ratysz and I talked about this at length during #1144 on Discord. Our conclusion was that no obvious and clearly correct logic table exists for this quaternary logic :(
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
The rule encoded here is "systems' own descriptors are applied on top of the descriptor of their set" - which means overwriting any non-additive field; run criteria is just the only one such field we have right now. SystemSet
s are not intended to be any sort of a hierarchy block or a container: they are now used only for mass description. Could probably do with a rename to avoid this confusion.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I totally understand the rule (i fully groked the "system sets are weapons of mass system description). Thats not why I left the comment. The behavior in the pr is "consistent", I just think it produces surprising behaviors. I think the FixedUpdate and State lifecycle criterias will be the two most common cases for SystemSets. People using sets under those circumstances are using them to group systems that should run under a certain condition. With that mindset, adding another criteria inside the set will likely be motivated by needing "additional" criteria (otherwise the user wouldn't be putting the system in a "fixed update" system set). Replacing the criteria defeats the purpose the set was created for.
I would rather fail in those cases to tell users they did something wrong than produce surprising results. I would also prefer to not fail (but I understand that this particular case is challenging so I'm cool with tabling the multiple run criteria conversation).
Renaming won't do anything to solve the issue above, or the fact that labels do work as "expected" by "adding" but run criteria "replace".
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
My previous comment had one more paragraph that got cut off 🤦 Network's been crapping out all morning. The gist was that I share your concerns but I don't have a good way of addressing them yet, hence this behavior - there can't be just no behavior. Attempt to reproduce it:
I had thought about forcing the criteria defined on systems that are inside sets to be "pipable", but that seemed more arbitrary than overwriting. It can't be strait-up forbidden because having to move a system outside of a set just to be able to set its run criteria is unergonomic - you'd end up with another copy of the descriptor, all because of the very thing that was supposed to prevent having to copy the descriptor. It also can't be "do it, but warn in console": every time we do that we have to think about how to silence the warning for legit cases.
In the short-term, until we figure it out (probably via a dedicated system set API improvement PR - system sets v3?..), I'm more okay with this overwriting behavior than I would be with forbidding this outright.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I had thought about forcing the criteria defined on systems that are inside sets to be "pipable", but that seemed more arbitrary than overwriting.
Yeah I agree. Seems too complicated / arbitrary.
It also can't be "do it, but warn in console": every time we do that we have to think about how to silence the warning for legit cases.
Agreed
It can't be strait-up forbidden because having to move a system outside of a set just to be able to set its run criteria is unergonomic - you'd end up with another copy of the descriptor, all because of the very thing that was supposed to prevent having to copy the descriptor.
I don't think this transform is particularly unergonomic. And it significantly increases clarity:
// before
app
.add_system_set(SystemSet::parallel()
.with_run_criteria(FixedTimestep::step(0.4))
.with_system(a.system())
.with_system(b.system())
.with_system(c.system().with_run_criteria(State::on_enter(Foo::Bar))
)
// after
app
.add_system_set(SystemSet::parallel()
.with_run_criteria(FixedTimestep::step(0.4))
.with_system(a.system())
.with_system(b.system())
)
.add_system(c.system().with_run_criteria(State::on_enter(Foo::Bar)))
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
For a case where you only have run criteria, sure. But consider this:
// before
app
.add_system_set(SystemSet::parallel()
.with_run_criteria(FixedTimestep::step(0.4))
.label(MyLabel::Logic)
.after(MyLabel::Physics)
.before(MyLabel::Rendering)
.with_system(a.system())
.with_system(b.system())
.with_system(c.system().with_run_criteria(State::on_enter(Foo::Bar))
)
// after
app
.add_system_set(SystemSet::parallel()
.with_run_criteria(FixedTimestep::step(0.4))
.label(MyLabel::Logic)
.after(MyLabel::Physics)
.before(MyLabel::Rendering)
.with_system(a.system())
.with_system(b.system())
)
.add_system(c.system()
.with_run_criteria(State::on_enter(Foo::Bar))
.label(MyLabel::Logic)
.after(MyLabel::Physics)
.before(MyLabel::Rendering)
)
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Yeah thats a fair point. I'd argue that c.system()
is in a "similar" set, but not an identical set, due to the differing run criteria. Regardless it sounds like we've agreed on the implementation (at least as a short term solution while we work the details of future run criteria improvements).
Just pushed a "proposal" that removes RunCriteriaIndex labels, which were both unnecessary/confusing (#1675 (comment)) and broken (#1675 (comment)). The new impl relies on "fixing up" stored run criteria indices like we do for systems. I'm not a huge fan of this machinery, but I think its preferable to adding a bunch of extra (broken/meaningless) labels. We could remove the index-fixup machinery by making runcriteria indices stable upon insertion (ex: the run_criteria list only allows push ops), but that also requires keeping more things in memory and more indirection when iterating/evaluating run criteria (ex: use the "run criteria index" to look up the "packed run criteria index"). |
Forgot to mention that I also included my preferred solution to #1675 (comment). I'm not putting my foot down there / its easily revertible if we opt for a different approach. |
Good catch on the bug, and thanks for fixing it! I like the other changes too - I've warmed up to the idea of forbidding for now individual criteria on systems that are inserted as part of a set: a clear behavior with a meh API is better than surprising behavior with good API. |
Awesome I think this is good to go! |
bors r+ |
I'm opening this prematurely; consider this an RFC that predates RFCs and therefore not super-RFC-like. This PR does two "big" things: decouple run criteria from system sets, reimagine system sets as weapons of mass system description. ### What it lets us do: * Reuse run criteria within a stage. * Pipe output of one run criteria as input to another. * Assign labels, dependencies, run criteria, and ambiguity sets to many systems at the same time. ### Things already done: * Decoupled run criteria from system sets. * Mass system description superpowers to `SystemSet`. * Implemented `RunCriteriaDescriptor`. * Removed `VirtualSystemSet`. * Centralized all run criteria of `SystemStage`. * Extended system descriptors with per-system run criteria. * `.before()` and `.after()` for run criteria. * Explicit order between state driver and related run criteria. Fixes #1672. * Opt-in run criteria deduplication; default behavior is to panic. * Labels (not exposed) for state run criteria; state run criteria are deduplicated. ### API issues that need discussion: * [`FixedTimestep::step(1.0).label("my label")`](https://github.com/Ratysz/bevy/blob/eaccf857cdaeb5a5632b6e75feab5c1ad6267d1d/crates/bevy_ecs/src/schedule/run_criteria.rs#L120-L122) and [`FixedTimestep::step(1.0).with_label("my label")`](https://github.com/Ratysz/bevy/blob/eaccf857cdaeb5a5632b6e75feab5c1ad6267d1d/crates/bevy_core/src/time/fixed_timestep.rs#L86-L89) are both valid but do very different things. --- I will try to maintain this post up-to-date as things change. Do check the diffs in "edited" thingy from time to time. Co-authored-by: Carter Anderson <mcanders1@gmail.com>
Pull request successfully merged into main. Build succeeded: |
Fixes #1753. The problem was introduced while reworking the logic around stages' own criteria. Before #1675 they used to be stored and processed inline with the systems' criteria, and systems without criteria used that of their stage. After, criteria-less systems think they should run, always. This PR more or less restores previous behavior; a less cludge solution can wait until after 0.5 - ideally, until stageless. Co-authored-by: Carter Anderson <mcanders1@gmail.com>
I'm opening this prematurely; consider this an RFC that predates RFCs and therefore not super-RFC-like.
This PR does two "big" things: decouple run criteria from system sets, reimagine system sets as weapons of mass system description.
What it lets us do:
Things already done:
SystemSet
.RunCriteriaDescriptor
.VirtualSystemSet
.SystemStage
..before()
and.after()
for run criteria.add_state
call order #1672.API issues that need discussion:
FixedTimestep::step(1.0).label("my label")
andFixedTimestep::step(1.0).with_label("my label")
are both valid but do very different things.I will try to maintain this post up-to-date as things change. Do check the diffs in "edited" thingy from time to time.