Intro to Clean Architectures

Introduction

Once your mod reaches a respectable size, you’ll find it harder and harder to keep the code clean and free of bugs. This is an especially big problem when using a dynamically typed language like Lua, given that the compiler gives you very little compiler-time help when it comes to things like making sure that types are used correctly.

This chapter covers important concepts needed to keep your code clean, and common design patterns to achieve that. Please note that this chapter isn’t meant to be prescriptive, but to instead give you an idea of the possibilities. There is no one good way of designing a mod, and good mod design is very subjective.

Cohesion, Coupling, and Separation of Concerns
Observer
Model-View-Controller
- API-View
Conclusion

Cohesion, Coupling, and Separation of Concerns

Without any planning, a programming project will tend to gradually descend into spaghetti code. Spaghetti code is characterised by a lack of structure - all the code is thrown in together with no clear boundaries. This ultimately makes a project completely unmaintainable, ending in its abandonment.

The opposite of this is to design your project as a collection of interacting smaller programs or areas of code.

Inside every large program, there is a small program trying to get out.

–C.A.R. Hoare

This should be done in such a way that you achieve Separation of Concerns - each area should be distinct and address a separate need or concern.

These programs/areas should have the following two properties:

High Cohesion - the area should be closely/tightly related.
Low Coupling - keep dependencies between areas as low as possible, and avoid relying on internal implementations. It’s a very good idea to make sure you have a low amount of coupling, as this means that changing the APIs of certain areas will be more feasible.

Note that these apply both when thinking about the relationship between mods, and the relationship between areas inside a mod.

Observer

A simple way to separate different areas of code is to use the Observer pattern.

Let’s take the example of unlocking an achievement when a player first kills a rare animal. The naïve approach would be to have achievement code in the mob kill function, checking the mob name and unlocking the award if it matches. This is a bad idea, however, as it makes the mobs mod coupled to the achievements code. If you kept on doing this - for example, adding XP to the mob death code - you could end up with a lot of messy dependencies.

Enter the Observer pattern. Instead of the mymobs mod caring about awards, the mymobs mod exposes a way for other areas of code to register their interest in an event and receive data about the event.

mymobs.registered_on_death = {}
function mymobs.register_on_death(func)
    table.insert(mymobs.registered_on_death, func)
end

-- in mob death code
for i=1, #mymobs.registered_on_death do
    mymobs.registered_on_death[i](entity, reason)
end

Then the other code registers its interest:

mymobs.register_on_death(function(mob, reason)
    if reason.type == "punch" and reason.object and
            reason.object:is_player() then
        awards.notify_mob_kill(reason.object, mob.name)
    end
end)

You may be thinking - wait a second, this looks awfully familiar. And you’re right! The Minetest API is heavily Observer-based to stop the engine having to care about what is listening to something.

Model-View-Controller

In the next chapter, we will discuss how to automatically test your code and one of the problems we will have is how to separate your logic (calculations, what should be done) from API calls (core.*, other mods) as much as possible.

One way to do this is to think about:

What data you have.
What actions you can take with this data.
How events (ie: formspec, punches, etc) trigger these actions, and how these actions cause things to happen in the engine.

Let’s take an example of a land protection mod. The data you have is the areas and any associated metadata. Actions you can take are create, edit, or delete. The events that trigger these actions are chat commands and formspec receive fields. These are 3 areas that can usually be separated pretty well.

In your tests, you will be able to make sure that an action when triggered does the right thing to the data. You won’t need to test that an event calls an action (as this would require using the Minetest API, and this area of code should be made as small as possible anyway.)

You should write your data representation using Pure Lua. “Pure” in this context means that the functions could run outside of Minetest - none of the engine’s functions are called.

-- Data
function land.create(name, area_name)
    land.lands[area_name] = {
        name  = area_name,
        owner = name,
        -- more stuff
    }
end

function land.get_by_name(area_name)
    return land.lands[area_name]
end

Your actions should also be pure, but calling other functions is more acceptable than in the above.

-- Controller
function land.handle_create_submit(name, area_name)
    -- process stuff
    -- (ie: check for overlaps, check quotas, check permissions)

    land.create(name, area_name)
end

function land.handle_creation_request(name)
    -- This is a bad example, as explained later
    land.show_create_formspec(name)
end

Your event handlers will have to interact with the Minetest API. You should keep the number of calculations to a minimum, as you won’t be able to test this area very easily.

-- View
function land.show_create_formspec(name)
    -- Note how there's no complex calculations here!
    return [[
        size[4,3]
        label[1,0;This is an example]
        field[0,1;3,1;area_name;]
        button_exit[0,2;1,1;exit;Exit]
    ]]
end

core.register_chatcommand("/land", {
    privs = { land = true },
    func = function(name)
        land.handle_creation_request(name)
    end,
})

core.register_on_player_receive_fields(function(player,
            formname, fields)
    land.handle_create_submit(player:get_player_name(),
            fields.area_name)
end)

The above is the Model-View-Controller pattern. The model is a collection of data with minimal functions. The view is a collection of functions which listen to events and pass it to the controller, and also receives calls from the controller to do something with the Minetest API. The controller is where the decisions and most of the calculations are made.

The controller should have no knowledge about the Minetest API - notice how there are no Minetest calls or any view functions that resemble them. You should NOT have a function like view.hud_add(player, def). Instead, the view defines some actions that the controller can tell the view to do, like view.add_hud(info) where info is a value or table which doesn’t relate to the Minetest API at all.

It is important that each area only communicates with its direct neighbours, as shown above, in order to reduce how much you need to change if you modify an area’s internals or externals. For example, to change the formspec you would only need to edit the view. To change the view API, you would only need to change the view and the controller, but not the model at all.

In practice, this design is rarely used because of the increased complexity and because it doesn’t give many benefits for most types of mods. Instead, you will commonly see a less formal and strict kind of design - variants of the API-View.

API-View

In an ideal world, you’d have the above 3 areas perfectly separated with all events going into the controller before going back to the normal view. But this isn’t the real world. A good compromise is to reduce the mod into two parts:

API - This was the model and controller above. There should be no uses of core. here.
View - This was also the view above. It’s a good idea to structure this into separate files for each type of event.

rubenwardy’s crafting mod roughly follows this design. api.lua is almost all pure Lua functions handling the data storage and controller-style calculations. gui.lua is the view for formspecs and formspec submission, and async_crafter.lua is the view and controller for a node formspec and node timers.

Separating the mod like this means that you can very easily test the API part, as it doesn’t use any Minetest APIs - as shown in the next chapter and seen in the crafting mod.

Conclusion

Good code design is subjective, and highly depends on the project you’re making. As a general rule, try to keep cohesion high and coupling low. Phrased differently, keep related code together and unrelated code apart, and keep dependencies simple.

I highly recommend reading the Game Programming Patterns book. It’s freely available to read online and goes into much more detail on common programming patterns relevant to games.