Skip to main content

The Art of Assumptions

· 5 min read

I am writing this post, because I don't want to make mistakes like this in the future.
And I think that analyzing past mistakes, and attempting to understand them, are a good way to avoid them.

For context, I have been working on UMG for about 2.5 years.
And throughout this time, over and over, I have made the same mistake of making too many assumptions about the future.

When I say "making an assumption", I mean you make a "guess" about how code/data is going to be used.
APIs can assume a lot, or can assume not much.

For example: math.distance(x, y)
This is a highly assumptionless API. We do not know (or CARE) where the x,y values are from. We just compute the euclidian distance.

Another example: uiObject:setBorderColor(color)
By contrast, this is an assumption-heavy API.
We are assuming that uiObject has a border.
We are assuming that the border has a color.
And we are also assuming that the uiObject can be rendered on the screen!

At face value, there is nothing "wrong" with this.
Assumptions work really well, IF we guess correctly. If every uiObject has a border, then this is great!
But what happens if we want a transparent image as an element?
UH OH... This isn't ideal.

It'd be much better if uiObject didn't have the :setBorderColor(color) method, and instead, we left border-handling to some other system/object.
This way, we wouldn't be left with a weird, bloaty method that is only used for some objects.

In OOP land, this is very similar to the Interface segregation principle, which attempts to isolate behaviour across interfaces as much as possible.

In a way, Interface-segregation embodies exactly what I mean by "assumptionless-programming".
By splitting things into tiny, hyper-specific interfaces; we are assuming VERY LITTLE about the final structure of our objects.
We are basically admitting to the system: "I have no idea what the final objects are going to look like! Thus, I will split the interfaces up as much as possible, so when I implement the concrete classes, the interfaces will fit nicely."


In UMG, however, we don't often do OOP; most UMG code is procedural.
So, what does procedural assumptionless-code look like?
I have some examples up my sleeve from UMG:

Some examples from the UMG ecosystem:

Perhaps the most relevant example as of (27/03/2024) is ui.
With the UI mod, I (stupidly) made an assumption that all root-UI-elements should be rendered in the main-scene.
This... was really stupid. Because I've now come across a situation where we need to render UI elements directly in the world, rather than in the upper-level main-scene.
To fix this, we need to remove the assumption that all root-elements exist in the main-scene; and instead, give the programmer the right to choose where the elements are put.
(I know; this seems obvious in hindsight!)

Another example is the inventory system.
A few years ago, I assumed that all inventories would be rendered into the world.
Think like, a chest. Or an anvil. Or a player inventory.
All of these can be rendered... no problem, right?
No!!! This is terrible!
We have created an assumption about how the inventory should be rendered. What happens if we want a minecraft-like inventory setup, with a hotbar at the bottom?
Uh oh... suddenly, we need 2 different renderers; one for the hotbar, and one for the inventory. Not possible under our current setup.




Ok, so it may be obvious by now that "highly assumptionless code" is a related concept to having low coupling.
Whilst I would agree with this, I think coupling is more often used when talking about the internals of a system; or a system as a whole. When writing an internal system, a lot of the time, you don't need to make assumptions about the future; because everything should be contained inside of said system.

Assumptions, then, are mostly needed when we are creating an external API, and we are predicting how the API is to be used.
UMG is truly a different beast, for this exact reason. With every base-mod that I write, I am FORCED to create assumptions about the future of the platform.
Which sucks, because my previous code has largely sucked at making good assumptions!!
(I will note, however; that the core UMG-API has been excellent at making no assumptions; so at least I have that haha.)

I also want to point out, a piece of software that I think has a very "lovely" assumptionless structure, is LOVE itself. With LOVE, the framework doesn't care at all how you structure your game.
There's no such thing as a gameObject. It doesn't care where you put logic.
It doesn't care how you load your files, either.
Isn't that beautiful?
Another API that is highly assumptionless is Raylib.




Avoiding bad assumptions in the future:

Obviously, I want to avoid this shit like the plague.
As such, I have created a lil checklist/procedure for myself to use.
Here it is:

Checklist for systems:

  • List (abstractly) all actions (A) that this system is doing/performing.

    • For every action A, can A be split up in any way?
      • If so, consider splitting it up.
    • Also, try make ALL As independent of one another.
  • List every dependency / bit of coupling C that this system is involved with.

    • For each C, could we remove this coupling whilst maintaining simplicity?

Checklist for components:

  • Attempt to list all the scenarios that this component could be used in.
    • If there are many scenarios; good. Keep it.
    • If the list is lacking, consider killing the component, or making it more generic.


Thanks for reading; ok, lets hope I remember to try use this thing lol.