ECS is less about access patterns (relationships and queries) and more about efficient storage. Games care about page faults still so making your tight loop execute quickly on many smaller pieces of data and "do the right thing" is a win. Also, from an architectural perspective it's easier to write abstract code like `Health.value -= 10` instead of PlayerHealth.value -= 10`. By coding against a single components interface you can get consistent behavior across all entities with that component. You could also do this with inheritance but that becomes complex and (the real killer) very slow. jblow has some rants about this in the witness.
I don't think the entities/system are really that relational. Data is not denormalized in the same way and joins don't really exist. Indices are an after thought. Full sequential scans are the main access pattern.
But, it's easy to think of data for a component system as a table where each row is associated with an entity so I can see why you draw the similarity.
I don't think there is a fundamental difference. ECS is the application of relational databases in a dynamic real-time context with relatively modest amounts of data, but lots of sequential processing. Sparse sets work well in this context, but they are ultimately an implementation detail. There are other ECS implementations that don't make use of it but can still deal with large throughput.
> ECS is the application of relational databases in a dynamic real-time context with relatively modest amounts of data, but lots of sequential processing.
This is probably the best explanation that I can imagine -- that they're not fundamentally different, just that ECS tends to make a sufficiently different set of design choices that it warrants some of it's own terminology.
I can't help but wonder what a more generalized take on ECS might look like, if it were to continue drawing from relational databases. For example, support for multiple indices to assist with different query patterns. Or triggers to abstract away cascading updates. Or perhaps materialized views to keep query patterns simple.
I've never had the opportunity to use an ECS system, especially in a performance sensitive context, so I don't have a good sense of where any pain points are in practice versus what my imagination can conjure up.
I also wonder what it might look like to use SQL to generate an optimal C++ representation while keeping the data definition high level.
Just idle musings - maybe one day I'll take the time to experiment.
Triggers might be counter productive. Games usually have a fundamental concept of a game loop. Changes can be processed in the loop and side effects can be processed in the following iteration. Triggers would cause this processing to be unpredictable (or at least harder to predict). ECS provides a clean way to define the order in which systems are processed each loop. Triggers might disrupt this ordering.
Maybe that's still desired, maybe not. I just thought it would be interesting to mention.
That's a great insight. I wonder if it'd be practical to defer them until after the causal operation was complete
say translating all positions, then calling all of the triggers for the positions component.
that'd keep everything in tight single purpose loops and preserve cache lines.
fair enough that it'd probably make execution order harder to predict, but also in theory it would be in the realm of possibility to generate a plan and print out the order that things would happen in.
(I'm not positing that this is actually worth doing or that the pros out weigh the cons -- just toying with the idea)
Completely agree on the different design choices. Curious if the analogy of how different gui packages work might have some parallels. FLTK uses a "retained mode" paradigm where IMGUI uses immediate mode. More idle musings..
I'm not familiar with _this_ exact ECS library, but another header-only ECS library with comparable functionality is flecs [1] and it compiles so fast I thought I misconfigured it. Literally could not even measure the compile time increase.
Unpopular opinion, but I personally think ECS has no place in C/C++ game dev. Every supposed advantage is either curbed by limitations of the language (e.g. serialization) or can be implemented more logically in a stateless manner.
In some ways ECS could be considered stateless, as the entity holds the game state as raw data, while the system operates on that data without holding a state of its own.
How would you implement ECS, a system for tracking the state of game entities, in a stateless manner? Can you expand on that?
I'll try to explain what I mean. Suppose you want an in game character to change the displayed weapon in their inventory.
Stateful way: your Inventory Manager component iterates through its private array of item entities to mark their model components as shown or hidden, then calls the dependency injected character's avatar component to update its animation state.
Stateless way: you flip an integer in character's struct, and the rendering function does something different.
I don't think the terminology you're looking for is not stateful/stateless, it's single source of truth vs multiple.
The latter is an anti-pattern, and you can do either with ECS.
With ECS nothing prevents you from just having the "physical" game state as the single source of truth, and having the render function for the player inventories look at it.
ECS is closer to the second way that was described. Data is represented as arrays of structs (referred to as components), similar to what was described. Systems can mutate or read these structs.
A physics system might change the position components while a decoupled rendering system might read that position components and the rendering details components to add draw commands.
I can't even imagine game dev without ECS. There's more performant architectures but it's such a useful mental model for the type of rapid iteration games need that so many people are willing to put up with the drawbacks
Instead of treating each object as a collection of components, you can operate on bags of components
So say in you're modeling space invaders:
every frame:
for enemy in enemies
enemy.transform.y += 10
The CPU is hopping to a random memory address where your "transform" component is located before modifying it, over and over.
Data oriented would be something like:
each frame:
for transform in transforms
if transform is enemy
transform.y += 10
if transform is player
transform += input
So you get cache locality as you're running through your transforms.
Now imagine if instead of space invaders we're trying to model say blades of grass. Suddenly we go from a very cache unfriendly method to something that's cache friendly, branch prediction friendly, and easy to parallelize.
If you want to experiment with it Unity has some great examples under their new DOTS system
https://www.minecraft.net/en-us/attribution