Changed doesn't explain any potential drawback
#3082
Comments
djeedai
added
C-Docs
|
Every time you mutate a component, the "tick" at which it changed is stored. |
|
Unless there's another alternative I'm not seeing, you can't know if a system mutated a component unless either 1) you assume that any |
|
You're missing option 3: utilize rust's DerefMut implementation to automatically overwrite the tick whenever a component or resource is accessed mutably. This means just querying for these things isn't enough to trigger a mutation. You must actually mutate (according to the rust compiler). Of course, you could mutate a component by setting to the exact same value, but I think this is the correct tradeoff relative to actually diffing components. We need recording a change to be very cheap (because mutations can happen frequently), so its literally just setting the tick value. We don't collect a list of mutated entities because allocating on mutation to populate a list (and de-duplicating the list) would be prohibitively expensive. We also don't do actual-value comparisons for the same reason. The tradeoff we landed on is that iterating Changed components checks every component's tick for the given type: we iterate every component of the given type ... not just the changed ones. This happens internally and users just receive actual Changed components. |
|
Ah the doc in question is wrong though:
I think something was lost in translation here by the doc author: Changed can be used as a way to optimize expensive operations by only doing them for changed components (ex: only updating the GlobalTransform when the Transform changes). The act of iterating changed components can actually be slightly more expensive than iterating all components of a given type because we need to check the tick in addition to retrieving the component for changed entities (although it might be faster in some cases because we dont actually need to retrieve unchanged components). |
|
Ok thanks for the detailed explanation. That's 1b) really, under the general umbrella of "assume modified on mutating getter". Don't get me wrong, that's absolutely the right thing to do for sure, but the details should probably be explained in the docs to call out the potential false positive (which may matter I'm some cases) and the extra work you mention. I'll try to make a proposal PR. |
|
Sounds good to me! |
|
@cart the current docs says:
Is that really the case?! I'd assume if system A mutates a component, system B filtering with |
|
A system doesn't have to run every frame. The nice thing about bevy's change detection is that skipping a system won't break change detection. It will report it the next time the system runs. That is why every system has it's own "tick" indicating when it ran the last time that is compared with the component "tick" indicating when it was changed the last time. |
Fix the documentation of the `Changed` filter to detail its mutating detection functioning, and explain the advantages and drawbacks of using it. Bug: bevyengine#3082
|
Thanks, but that's not exactly the question. Sure, it's nice that if a system doesn't tick a frame, |
|
Opened PR #3084, we can iterate on proper wording there. |
I think @bjorn3 did pretty much answer your question by saying that each system has its own tick, which is compared against the component tick. Change detection is per system. If a change (which happened anywhere / in any system) occurred since the last time a system ran, the change will show up in that system's Change queries. |
|
Removed from the milestone because this isn't really needed for the release and we're nearing the cutoff point. If it makes it in, thats cool though :) |
## Objective - ~~Make absurdly long-lived changes stay detectable for even longer (without leveling up to `u64`).~~ - Give all changes a consistent maximum lifespan. - Improve code clarity. ## Solution - ~~Increase the frequency of `check_tick` scans to increase the oldest reliably-detectable change.~~ (Deferred until we can benchmark the cost of a scan.) - Ignore changes older than the maximum reliably-detectable age. - General refactoring—name the constants, use them everywhere, and update the docs. - Update test cases to check for the specified behavior. ## Related This PR addresses (at least partially) the concerns raised in: - #3071 - #3082 (and associated PR #3084) ## Background - #1471 Given the minimum interval between `check_ticks` scans, `N`, the oldest reliably-detectable change is `u32::MAX - (2 * N - 1)` (or `MAX_CHANGE_AGE`). Reducing `N` from ~530 million (current value) to something like ~2 million would extend the lifetime of changes by a billion. | minimum `check_ticks` interval | oldest reliably-detectable change | usable % of `u32::MAX` | | --- | --- | --- | | `u32::MAX / 8` (536,870,911) | `(u32::MAX / 4) * 3` | 75.0% | | `2_000_000` | `u32::MAX - 3_999_999` | 99.9% | Similarly, changes are still allowed to be between `MAX_CHANGE_AGE`-old and `u32::MAX`-old in the interim between `check_tick` scans. While we prevent their age from overflowing, the test to detect changes still compares raw values. This makes failure ultimately unreliable, since when ancient changes stop being detected varies depending on when the next scan occurs. ## Open Question Currently, systems and system states are incorrectly initialized with their `last_change_tick` set to `0`, which doesn't handle wraparound correctly. For consistent behavior, they should either be initialized to the world's `last_change_tick` (and detect no changes) or to `MAX_CHANGE_AGE` behind the world's current `change_tick` (and detect everything as a change). I've currently gone with the latter since that was closer to the existing behavior. ## Follow-up Work (Edited: entire section) We haven't actually profiled how long a `check_ticks` scan takes on a "large" `World` , so we don't know if it's safe to increase their frequency. However, we are currently relying on play sessions not lasting long enough to trigger a scan and apps not having enough entities/archetypes for it to be "expensive" (our assumption). That isn't a real solution. (Either scanning never costs enough to impact frame times or we provide an option to use `u64` change ticks. Nobody will accept random hiccups.) To further extend the lifetime of changes, we actually only need to increment the world tick if a system has `Fetch: !ReadOnlySystemParamFetch`. The behavior will be identical because all writes are sequenced, but I'm not sure how to implement that in a way that the compiler can optimize the branch out. Also, since having no false positives depends on a `check_ticks` scan running at least every `2 * N - 1` ticks, a `last_check_tick` should also be stored in the `World` so that any lull in system execution (like a command flush) could trigger a scan if needed. To be completely robust, all the systems initialized on the world should be scanned, not just those in the current stage.
## Objective - ~~Make absurdly long-lived changes stay detectable for even longer (without leveling up to `u64`).~~ - Give all changes a consistent maximum lifespan. - Improve code clarity. ## Solution - ~~Increase the frequency of `check_tick` scans to increase the oldest reliably-detectable change.~~ (Deferred until we can benchmark the cost of a scan.) - Ignore changes older than the maximum reliably-detectable age. - General refactoring—name the constants, use them everywhere, and update the docs. - Update test cases to check for the specified behavior. ## Related This PR addresses (at least partially) the concerns raised in: - #3071 - #3082 (and associated PR #3084) ## Background - #1471 Given the minimum interval between `check_ticks` scans, `N`, the oldest reliably-detectable change is `u32::MAX - (2 * N - 1)` (or `MAX_CHANGE_AGE`). Reducing `N` from ~530 million (current value) to something like ~2 million would extend the lifetime of changes by a billion. | minimum `check_ticks` interval | oldest reliably-detectable change | usable % of `u32::MAX` | | --- | --- | --- | | `u32::MAX / 8` (536,870,911) | `(u32::MAX / 4) * 3` | 75.0% | | `2_000_000` | `u32::MAX - 3_999_999` | 99.9% | Similarly, changes are still allowed to be between `MAX_CHANGE_AGE`-old and `u32::MAX`-old in the interim between `check_tick` scans. While we prevent their age from overflowing, the test to detect changes still compares raw values. This makes failure ultimately unreliable, since when ancient changes stop being detected varies depending on when the next scan occurs. ## Open Question Currently, systems and system states are incorrectly initialized with their `last_change_tick` set to `0`, which doesn't handle wraparound correctly. For consistent behavior, they should either be initialized to the world's `last_change_tick` (and detect no changes) or to `MAX_CHANGE_AGE` behind the world's current `change_tick` (and detect everything as a change). I've currently gone with the latter since that was closer to the existing behavior. ## Follow-up Work (Edited: entire section) We haven't actually profiled how long a `check_ticks` scan takes on a "large" `World` , so we don't know if it's safe to increase their frequency. However, we are currently relying on play sessions not lasting long enough to trigger a scan and apps not having enough entities/archetypes for it to be "expensive" (our assumption). That isn't a real solution. (Either scanning never costs enough to impact frame times or we provide an option to use `u64` change ticks. Nobody will accept random hiccups.) To further extend the lifetime of changes, we actually only need to increment the world tick if a system has `Fetch: !ReadOnlySystemParamFetch`. The behavior will be identical because all writes are sequenced, but I'm not sure how to implement that in a way that the compiler can optimize the branch out. Also, since having no false positives depends on a `check_ticks` scan running at least every `2 * N - 1` ticks, a `last_check_tick` should also be stored in the `World` so that any lull in system execution (like a command flush) could trigger a scan if needed. To be completely robust, all the systems initialized on the world should be scanned, not just those in the current stage.
## Objective - ~~Make absurdly long-lived changes stay detectable for even longer (without leveling up to `u64`).~~ - Give all changes a consistent maximum lifespan. - Improve code clarity. ## Solution - ~~Increase the frequency of `check_tick` scans to increase the oldest reliably-detectable change.~~ (Deferred until we can benchmark the cost of a scan.) - Ignore changes older than the maximum reliably-detectable age. - General refactoring—name the constants, use them everywhere, and update the docs. - Update test cases to check for the specified behavior. ## Related This PR addresses (at least partially) the concerns raised in: - bevyengine#3071 - bevyengine#3082 (and associated PR bevyengine#3084) ## Background - bevyengine#1471 Given the minimum interval between `check_ticks` scans, `N`, the oldest reliably-detectable change is `u32::MAX - (2 * N - 1)` (or `MAX_CHANGE_AGE`). Reducing `N` from ~530 million (current value) to something like ~2 million would extend the lifetime of changes by a billion. | minimum `check_ticks` interval | oldest reliably-detectable change | usable % of `u32::MAX` | | --- | --- | --- | | `u32::MAX / 8` (536,870,911) | `(u32::MAX / 4) * 3` | 75.0% | | `2_000_000` | `u32::MAX - 3_999_999` | 99.9% | Similarly, changes are still allowed to be between `MAX_CHANGE_AGE`-old and `u32::MAX`-old in the interim between `check_tick` scans. While we prevent their age from overflowing, the test to detect changes still compares raw values. This makes failure ultimately unreliable, since when ancient changes stop being detected varies depending on when the next scan occurs. ## Open Question Currently, systems and system states are incorrectly initialized with their `last_change_tick` set to `0`, which doesn't handle wraparound correctly. For consistent behavior, they should either be initialized to the world's `last_change_tick` (and detect no changes) or to `MAX_CHANGE_AGE` behind the world's current `change_tick` (and detect everything as a change). I've currently gone with the latter since that was closer to the existing behavior. ## Follow-up Work (Edited: entire section) We haven't actually profiled how long a `check_ticks` scan takes on a "large" `World` , so we don't know if it's safe to increase their frequency. However, we are currently relying on play sessions not lasting long enough to trigger a scan and apps not having enough entities/archetypes for it to be "expensive" (our assumption). That isn't a real solution. (Either scanning never costs enough to impact frame times or we provide an option to use `u64` change ticks. Nobody will accept random hiccups.) To further extend the lifetime of changes, we actually only need to increment the world tick if a system has `Fetch: !ReadOnlySystemParamFetch`. The behavior will be identical because all writes are sequenced, but I'm not sure how to implement that in a way that the compiler can optimize the branch out. Also, since having no false positives depends on a `check_ticks` scan running at least every `2 * N - 1` ticks, a `last_check_tick` should also be stored in the `World` so that any lull in system execution (like a command flush) could trigger a scan if needed. To be completely robust, all the systems initialized on the world should be scanned, not just those in the current stage.
## Objective - ~~Make absurdly long-lived changes stay detectable for even longer (without leveling up to `u64`).~~ - Give all changes a consistent maximum lifespan. - Improve code clarity. ## Solution - ~~Increase the frequency of `check_tick` scans to increase the oldest reliably-detectable change.~~ (Deferred until we can benchmark the cost of a scan.) - Ignore changes older than the maximum reliably-detectable age. - General refactoring—name the constants, use them everywhere, and update the docs. - Update test cases to check for the specified behavior. ## Related This PR addresses (at least partially) the concerns raised in: - bevyengine#3071 - bevyengine#3082 (and associated PR bevyengine#3084) ## Background - bevyengine#1471 Given the minimum interval between `check_ticks` scans, `N`, the oldest reliably-detectable change is `u32::MAX - (2 * N - 1)` (or `MAX_CHANGE_AGE`). Reducing `N` from ~530 million (current value) to something like ~2 million would extend the lifetime of changes by a billion. | minimum `check_ticks` interval | oldest reliably-detectable change | usable % of `u32::MAX` | | --- | --- | --- | | `u32::MAX / 8` (536,870,911) | `(u32::MAX / 4) * 3` | 75.0% | | `2_000_000` | `u32::MAX - 3_999_999` | 99.9% | Similarly, changes are still allowed to be between `MAX_CHANGE_AGE`-old and `u32::MAX`-old in the interim between `check_tick` scans. While we prevent their age from overflowing, the test to detect changes still compares raw values. This makes failure ultimately unreliable, since when ancient changes stop being detected varies depending on when the next scan occurs. ## Open Question Currently, systems and system states are incorrectly initialized with their `last_change_tick` set to `0`, which doesn't handle wraparound correctly. For consistent behavior, they should either be initialized to the world's `last_change_tick` (and detect no changes) or to `MAX_CHANGE_AGE` behind the world's current `change_tick` (and detect everything as a change). I've currently gone with the latter since that was closer to the existing behavior. ## Follow-up Work (Edited: entire section) We haven't actually profiled how long a `check_ticks` scan takes on a "large" `World` , so we don't know if it's safe to increase their frequency. However, we are currently relying on play sessions not lasting long enough to trigger a scan and apps not having enough entities/archetypes for it to be "expensive" (our assumption). That isn't a real solution. (Either scanning never costs enough to impact frame times or we provide an option to use `u64` change ticks. Nobody will accept random hiccups.) To further extend the lifetime of changes, we actually only need to increment the world tick if a system has `Fetch: !ReadOnlySystemParamFetch`. The behavior will be identical because all writes are sequenced, but I'm not sure how to implement that in a way that the compiler can optimize the branch out. Also, since having no false positives depends on a `check_ticks` scan running at least every `2 * N - 1` ticks, a `last_check_tick` should also be stored in the `World` so that any lull in system execution (like a command flush) could trigger a scan if needed. To be completely robust, all the systems initialized on the world should be scanned, not just those in the current stage.
## Objective - ~~Make absurdly long-lived changes stay detectable for even longer (without leveling up to `u64`).~~ - Give all changes a consistent maximum lifespan. - Improve code clarity. ## Solution - ~~Increase the frequency of `check_tick` scans to increase the oldest reliably-detectable change.~~ (Deferred until we can benchmark the cost of a scan.) - Ignore changes older than the maximum reliably-detectable age. - General refactoring—name the constants, use them everywhere, and update the docs. - Update test cases to check for the specified behavior. ## Related This PR addresses (at least partially) the concerns raised in: - bevyengine#3071 - bevyengine#3082 (and associated PR bevyengine#3084) ## Background - bevyengine#1471 Given the minimum interval between `check_ticks` scans, `N`, the oldest reliably-detectable change is `u32::MAX - (2 * N - 1)` (or `MAX_CHANGE_AGE`). Reducing `N` from ~530 million (current value) to something like ~2 million would extend the lifetime of changes by a billion. | minimum `check_ticks` interval | oldest reliably-detectable change | usable % of `u32::MAX` | | --- | --- | --- | | `u32::MAX / 8` (536,870,911) | `(u32::MAX / 4) * 3` | 75.0% | | `2_000_000` | `u32::MAX - 3_999_999` | 99.9% | Similarly, changes are still allowed to be between `MAX_CHANGE_AGE`-old and `u32::MAX`-old in the interim between `check_tick` scans. While we prevent their age from overflowing, the test to detect changes still compares raw values. This makes failure ultimately unreliable, since when ancient changes stop being detected varies depending on when the next scan occurs. ## Open Question Currently, systems and system states are incorrectly initialized with their `last_change_tick` set to `0`, which doesn't handle wraparound correctly. For consistent behavior, they should either be initialized to the world's `last_change_tick` (and detect no changes) or to `MAX_CHANGE_AGE` behind the world's current `change_tick` (and detect everything as a change). I've currently gone with the latter since that was closer to the existing behavior. ## Follow-up Work (Edited: entire section) We haven't actually profiled how long a `check_ticks` scan takes on a "large" `World` , so we don't know if it's safe to increase their frequency. However, we are currently relying on play sessions not lasting long enough to trigger a scan and apps not having enough entities/archetypes for it to be "expensive" (our assumption). That isn't a real solution. (Either scanning never costs enough to impact frame times or we provide an option to use `u64` change ticks. Nobody will accept random hiccups.) To further extend the lifetime of changes, we actually only need to increment the world tick if a system has `Fetch: !ReadOnlySystemParamFetch`. The behavior will be identical because all writes are sequenced, but I'm not sure how to implement that in a way that the compiler can optimize the branch out. Also, since having no false positives depends on a `check_ticks` scan running at least every `2 * N - 1` ticks, a `last_check_tick` should also be stored in the `World` so that any lull in system execution (like a command flush) could trigger a scan if needed. To be completely robust, all the systems initialized on the world should be scanned, not just those in the current stage.
|
I don't think this is particularly relevant any more. Closing. |
How can Bevy's documentation be improved?
https://docs.rs/bevy/0.5.0/bevy/ecs/query/struct.Changed.html
The docs mention this is a performance optimization to iterate over changed entities/components only. There's no mention of what the cost of maintaining such information is. Surely this doesn't come "for free"; there must be some caching somewhere to compare with the previous frame's version and be able to tell. What's the cost? Can this be used on many different single-components (like, should I use
Changed<MyType>for most of my own types, even if used once or twice only)? Or is it more beneficial on a single component with many instances (like theChanged<Transform>example? Is there a minimum order of magnitude number below whichChangedis prohibitive compared to just iterating over all existing items without any caching?I checked the docs on
maintoo, and it's no better. It just mentionsChangeTrackersin addition; I'm not even understanding the difference betweenChangeTrackersandChanged.The text was updated successfully, but these errors were encountered: