Treating constants like state objects about fusion

[filiptibell]

This is one of the first things that I made a utility function for when starting with Fusion, and would be a great addition.

Being part of the public namespace / main module exports is something I wouldn't be a fan of, however. I like how rocket.rs does it, splitting main functionality and utility into rocket and rocket_contrib, with the latter containing utilities that most if not all large scale projects would use. This would also give users the choice to implement those said utilities themselves, or simply use a different variant of them, without polluting the main modules namespace.

Another useful utility function that I've been using with Fusion, which functions similar to the proposed makeState, could be something like unwrapState:

local function unwrapState(stateOrValue, registerAsDependency)
	if stateOrValue is a state then
		return stateOrValue:get(registerAsDependency)
	end
	return stateOrValue
end

[dphfox]

I would argue that it's better to separate them for optimization sake.

Do you have benchmarks to show that a function like this poses any performance issue? Optimisation should not a primary concern until it's proven to be a problem with data.

In my mind it's much more important that developers have a good experience with Fusion than it is to always run the most optimal code path theoretically possible, though of course this is not an excuse to write horribly slow code.

This is completely unnecessary for constants, and would just result in us creating a lot of unnecessary objects. (Think creating an Observer object for every single property in the entirety of someone's UI code). For that reason I don't think this is viable long-term.

This would not be used in such a fashion. For one, this is a user-facing utility, not something for use inside Fusion itself. Secondly, it's not mandated to be used, so where performance is needed you can manually handle both cases. Thirdly, most uses for statify would be when components are processing custom props - which is largely limited to one-offs - and not used for every state object. Fourthly, the overhead is entirely at create time - constant-value state objects are completely weightless in terms of run time performance due to the push nature of the reactive graph.

[nottirb]

The optimization basis was specifically in reference to memory usage, rather than speed.

That being said, I misunderstood what you meant by this, and thought you were specifically referring to what we were doing internally. As a result all I can say is disregard what I said... totally my fault.

[dphfox]

I've so far identified three different solutions to this problem, which may not necessarily be the same as the solution proposed here.

1. a function that transforms constants into state, but leaves state alone - meaning you can operate on constants just like they're state (this is what makeState is proposed to do)
2. a function that, given a constant, returns it; but given a state, returns the value of the state
3. a function that wraps your component functions, automatically turning constants in your props into states

Some pseudocode examples:

local const = "const"
local state = Value("state")

-- option 1
local function makeState(x)
    return if x is state then x else { get = function() return x end }
end

local function process(x)
    print(x:get())
end

process(state) -- prints "state"
process(makeState(const)) -- prints "const"

-- option 2
local function get(x)
    return if x is state then x:get() else x
end

local function process(x)
    print(get(x))
end

process(state) -- prints "state"
process(const) -- prints "const"

-- option 3
local function wrap(component)
    return function(props)
        local realProps = {}
        for key, value in pairs(props) do
            realProps[key] = makeState(value)
        end
        return component(realProps)
    end
end

local component = wrap(function(props)
    print(props.value:get())
end)

-- prints "state"
component {
    value = state
}

-- prints "const"
component {
    value = const
}

I'm personally a huge fan of how option 2 doesn't require transforming variables or constructing objects, but it also means we can't use OOP notation. Perhaps we should be reconsidering the OOP syntax of state objects entirely? We're already moving away from metatable-based OOP, which would make self useless to us.

[dphfox]

This has been formalised as the unwrap() function, which is currently slated for Fusion v0.3, so deferring to the Fusion v0.3 project board as such.

[nottirb]

Still a bit confused on how we want this implemented. Do we want this to be Fusion.Statify(state | constant) or do you want to add something like what @filiptibell said and create something like fusion_contrib separate from Fusion? (and implement it there)

If we want the latter, do we want:
a) To implement it into the main fusion program (i.e. do Fusion.contrib.XYZ where "XYZ" is just some utility that most people would use.
b) To implement it into a separate organization, and users can add them manually or via wally by doing something like ``XYZ = "fusion_contrib/xyz@VERSION" Edit: this is discussed in #40

Once I get some feedback on how we want this implemented I'll work on getting it added.

Edit: After diving more into the Fusion source code, I now understand what you mean when you say that you want the ability to Statify constants. To that end, I would argue that while this might improve the programming experience a little bit (at first the separation of various things into states and constants was a bit confusing and led to a few bugs in my early contribution process), I would argue that it's better to separate them for optimization sake. Though, that's just my immediate thought.

For instance, take the New.lua file, you see something like this under "Properties bound to state":

local disconnect = Observer(value):onChange(function()
	if ref.instance == nil then
		if ENABLE_EXPERIMENTAL_GC_MODE then
			if conn.Connected then
				warn("ref is nil and instance is around!!!")
			else
				print("ref is nil, but instance was destroyed")
			end
		end
		return
	end
	Scheduler.enqueueProperty(ref.instance, key, value:get(false))
end)
table.insert(cleanupTasks, disconnect)

This is completely unnecessary for constants, and would just result in us creating a lot of unnecessary objects. (Think creating an Observer object for every single property in the entirety of someone's UI code). For that reason I don't think this is viable long-term.

If anyone thinks otherwise feel free to let me know!

[Gargafield]

So I came up with this stupid idea. (But I'm still going to say it 🙂)
What if all data has a :get function. This doesn't mean it's a state object, more like a prop object. The prop object won't update. In the new function will prop just behave like some data.

[dphfox]

In line with the motives behind #159 and #168, it might be worth considering an alternative name such as peek or once