Let’s talk about C#’s finalizers (also called destructors in C#) and how a common mistake when using them might lead to unwanted behaviour, especially in applications made with the Unity engine.
A finalizer is a method that is called whenever an instance of a class is being garbage-collected. It is used for cleanup, commonly to release resources. The code below contains an example class for a music player, where its destructor closes an open file. Note that finalizers always start with a tilde and can’t have any parameters.
public class MusicPlayer
private FileStream _file;
public void Play(string filePath)
_file = File.Open(filePath, FileMode.Open);
if (_file != null)
Disclaimer: even though destructors and finalizers are two different things, C#’s spec treats them as the same. If you’re acquainted with both definitions, the C#’s mechanism discussed in this article is actually a finalizer. Also, if you come from a C++ background, even though C#’s syntax for finalizers resembles C++’s syntax, keep in mind that they’re not the same. C++ destructors are called explicitly (and deterministically) by the user whereas C#’s finalizers are called implicitly (and nondeterministically) by the garbage collector.
A quick intro to Unity’s MonoBehaviour
The Unity engine provides a handy base class (MonoBehaviour) that contains common behaviour often needed for game development. Among other things, it contain event methods that are automatically called by the engine under given scenarios. These event methods are guaranteed to be called under the right circumstances and therefore developers can rely on them. Some of these methods are called when an object is created and when it’s destroyed, analogous to constructors and finalizers (which should not exist for MonoBehaviours). For example, the Awake methods is called when an object is created and the OnDestroy method is called when it’s destroyed.
It is common practice to use Awake and OnDestroy to subscribe and unsubscribe to events, respectively. The code below shows an example.
The ExampleButton behaviour contains only an event, for example purposes:
public class ExampleButton : MonoBehaviour
public event Action OnClick;
The MyBehaviour behaviour subscribes to _button‘s OnClick event when it’s created and unsubscribes to the same event when it’s destroyed.
The code above works as expected and both initialization and cleanup execute as expected.
The naive thought
At some point during development, we introduce a regular, non-MonoBehaviour class called ExampleClass. Unlike MonoBehaviours, it can’t rely on event methods like Awake for initialization. As an alternative, we normally use the class’ constructor. Analogously, ExampleClass can’t rely on event methods like OnDestroy for cleanup. As an alternative, we use its finalizer.
Let’s define ExampleBehaviour, a MonoBehaviour that contains an ExampleButton and an ExampleClass. On its Awake method, the _object field (of type ExampleClass) is initialized.
public class ExampleBehaviour : MonoBehaviour
[SerializeField] private ExampleButton _button;
private ExampleClass _object;
private void Awake()
_object = new ExampleClass("Foo", _button);
A new Unity scene is created and only 2 objects are added to it: one containing an ExampleButton (called MyButtonGameObject) and one containing an ExampleBehaviour (called MyBehaviourGameObject). Whenever the scene is played, MyBehaviourGameObject’s Awake method is invoked and the _object variable is assigned, as expected. Inside ExampleClass constructor, the event subscription is executed, as expected. Nothing unusual so far.
Then, because of some design decision, the MyBehaviourGameObject object is destroyed along the application lifetime. We realize that we might have to do some cleanup because _object should unsubscribe from _button‘s OnClick event. But then we come to the conclusion that it actually should be alright and that no additional cleanup should be necessary. Whenever MyBehaviourGameObject gets destroyed, the garbage collector will collect its ExampleBehaviour script. Since _object only belongs to MyBehaviourGameObject, it should be collected as well, which should trigger its finalizer and unsubscribe from the events.
Later during development, we notice some weird behaviour whenever the user clicks on MyButtonGameObject: there’s ExampleClass behaviour still being executed. But at that point of the execution, there should be no activeExampleClass in the scene because MyBehaviourGameObject was destroyed! After checking the call stack, we find out that Bar is being called by ExampleButton‘s OnClick event. But wait a second, something is wrong. ExampleClass‘s finalizer was responsible for event unsubscription. What happened? After some more debugging, you finally find out that the object’s finalizer doesn’t ever get called. But why isn’t the finalizer being called at all?
Maybe the garbage collector is not running, for whatever reason. Let’s try to force collection using a simple script.
public class ManualGarbageCollector : MonoBehaviour
private void Update()
For testing purposes, I manually delete MyBehaviourGameObject and press the C key. The finalizer is still not executing. What is going on? Is the garbage collector broken? MyBehaviourGameObject was destroyed and _object should be collected, which should trigger the event unsubscription.
To answer those question, we need to understand how C#’s garbage collector works. First, no, it’s not broken at all. In fact, it’s doing exactly what it’s told to. It’s our minds that forgot what we were doing. An object will only get collected whenever there are no references to it. None at all. Zero. If there’s one, even if really hidden, forgotten reference to an object, it will not get collected by the garbage collector. It’s as simple as that. But it still doesn’t make sense. The only reference to _object was inside ExampleBehaviour, right?
We forgot about the one reference we were trying to get rid of: the one inside ExampleButton‘s OnClick event. But wait a second, we didn’t store a reference to _object in that event, we just stored a reference to a method, right?
Again, wrong. Although it looks like we’re subscribing to a method, we need to keep in mind that it’s an instance method. It belongs to an instance of a class, a.k.a. an object. Under the hood, that method’s reference consists – among other things – of a reference to the method in memory and a reference to the instance the methods should be called on. If you ever programmed in Python, think of how the first argument of an instance method is self. As a consequence, an event subscription to an instance method will keep a reference to the instance itself. Therefore, it will stop the object of being collected by the garbage collector.
We can show show that _object‘s reference is, in fact, being kept by MyButtonGameObject by destroying the latter. Once destroyed, the garbage collector will collect ExampleButton‘s memory and later, _object.
If Bar was a static method, this wouldn’t happen because references to static methods don’t include a reference to an object. Consequently, _object‘s reference count would drop to 0 and it would be eventually collected by the garbage collector.
Thankfully, there is an easy solution for that: create a cleanup method and explicitly call it whenever necessary. On this article’s example, the perfect candidate would be ExampleBehaviour‘s OnDestroy method.
public void Cleanup()
_button.OnClick -= Bar;
private void OnDestroy()
This fix not only will guarantee that events get unsubscribed but it will also – ironically – remove the last references to _object, which allows its (now obsolete) finalizer to be called.
The bonus trap
There’s another trap regarding finalizers that is not related to the one described above. Finalizers are called by the garbage collector, which runs on a separate thread than Unity’s main thread. As a consequence, two problems might come up.
First – as usual – Unity engine code can not be called from a separate thread. Calling something as simple as _foo.gameObject will throw a UnityEngine.UnityException with the message “get_gameObject can only be called from the main thread“.
Second, Unity will not catch and log exceptions running on separate threads. Therefore, any exceptions thrown inside finalizers (like the one described above) or in any separate thread might fly under the radar and never get acknowledged by the developers. There are two possible fixes for this problem. One is the obvious: catch the exceptions inside the thread itself. Another one can be used for user-defined threads (and thus is not applicable to GC threads): useTask instead of Thread to start a new thread with exception handling support.
Whenever using C# finalizers, keep in mind that they will only get called when no references to their respective objects are left, including references to instance methods. Therefore, using finalizers to unsubscribe from events and to remove delegate references might lead to unwanted behaviour. As an alternative, create a cleanup method and explicitly call it whenever necessary.
As a good practice, try to use finalizers for what they are good for: freeing resources. For other usages, don’t rely on them and explicitly invoke cleanup methods. Some even say that pure finalizers should avoided and the disposable pattern should be used instead.
That’s all for today. As usual, feel free to leave a comment with corrections, questions, criticism or compliments. See you next time!
In this article I analyze the development Voedingscentrum VR, an educational, kiosk-style and Virtual Reality game I helped to develop at Fantazm from October 2018 until April 2019. After a quick introduction about the project, the biggest problems that surfaced during development are explained along with the solutions we found to solve them. At the end, a quick conclusion wraps the article up.
The project was commissioned by Voedingscentrum (The Netherlands Nutrition Center), an independent organization funded by the Dutch government. Their aim is to use relevant research regarding sustainable nutrition to promote balanced, sustainable food consumption and production.
The game was planned as a fun, educacional VR experience targeted at children from 9 to 12 years old. Players are placed into a cartoonish world where they need to feed Smikkel – an elephant-like creature – for 3 to 4 meals: breakfast, lunch, afternoon snack and dinner.
A robot waiter teaches the players about the game mechanics and its main loop: for each meal the player chooses, Smikkel comes to the dinner table and you feed it by choosing, grabbing and placing food items (served by the waiter) on a plate. Once all the food items of a specific meal are chosen (between 2 and 6 items), Smikkel eats the food and its mood, energy levels and appearance change according to the food items you fed him. Following, playtime starts and the player can interact with Smikkel by petting it and throwing balls so it can fetch them. During playtime, it should be evident that the food Smikkel ate affected its behavior. Additionally, the waiter displays a report showing the selected food items and gives the user feedback on the chosen food.
Once playtime finishes, the feeding loop starts again. Once all the meal loops are over, the waiter gives a final feedback, announces that the experience is over and instructs the player to remove their VR headset. During the entire experience, spectators can watch the gameplay through a TV.
The educational value is clear: children learn that what they ingest might affect their mood, stamina and health. The fun aspect is present: immersing into a cartoonish VR world and interacting with fictional characters is exciting, even for adults.
Voedingscentrum was responsible for nutrition expertise and game design insights. The development team consisted of one game designer/project manager, two game developers (I inherited the project from another developer), one 2D artist, two 3D artists, one external 3D artist/animator and an external voice actor agency.
Prototyping started as early as May of 2018 and the last version of the game was deployed in April 2019. Although this period might seem long, development didn’t take place continuously because the development team didn’t work on this project exclusively. Therefore, it’s hard to estimate how many hours were put into development.
The VR experience was meant to be installed in Dutch museums. Besides the VR headset and controllers, walls with game art, a grass carpet and a TV (for spectators) would be installed to add a fun, immersive and inviting environment. By design, the experience should work unsupervised and require low maintenance. Windows Mixed Reality (hereafter referred as WMR) was chosen as the target platform because it offers the features we sought (VR headset with controllers and inside-out tracking) at an affordable price.
The development environment
The project was developed using the Unity engine (by the end of the project, 2018.3.11f) and C# as a programming language. On initial development stages, Unity Collaborate was used for version controlling, but later we switched to a Git-based solution hosted by BitBucket.
Three versions (1.0, 1.1 and 1.2) of the game were shipped. The first one was installed at Boerhaave Museum (Leiden, the Netherlands) in December of 2018 and it showed us right away that there were some challenges we had to overcome before the game was ready for the world. The exhibition was taken down and we came back to the drawing board. After a quick break, we started the development of version 1.1, on which we tackled most of the problems the previous version unearthed. Although most of the design problems found in version 1.0 were fixed in version 1.1, technical problems (mostly regarding Windows Mixed Reality) could not be overcome. The application was migrated from WMR to HTC VIVE + SteamVR on version 1.2.
As expected, some challenges surfaced during the development process and after we installed the experience for the first time. In this section we discuss how we overcame them – whenever possible. Keep in mind that this article was published months after development ceased. Therefore, some of these problems might not exist anymore. If that’s the case and you are aware of a possible solution, please leave a comment on the comments section so this post can be updated. Adding new solutions to this postmorten can help many developers that might run into the same problems we did.
Arguably the biggest challenge we faced during the game development were the WMR controllers. It’s nothing about their build quality or responsiveness (which are both great, by the way), but we ran into a few different aspects that didn’t match our project. We did not find solutions for some of these challenges, and there’s a reason behind that. At the end, the accumulation of problems regarding the controllers was so overwhelming that we decided to ditch the controllers altogether. We used a Leap Motion to track hand movement so we could use the player’s hands as interaction interfaces instead. Nevertheless, the problems we ran into before ditching the controllers are listed below.
Although the WMR controller has many buttons (touchpad, thumbstick, menu, windows, trigger, grab), the proposed gameplay was so simple that it only required one button. This sounds like the opposite of a problem and we didn’t expect it to be one at all, but something surfaced during playtests with children. Unless they were instructed beforehand, players had a hard time figuring out how to interact with the elements in the VR world, usually because they didn’t know which button to press. Changing the button mapping didn’t seem to solve the problem because different children were attracted to different buttons. Our proposed solution was to create either a sleeve or a case around the controller to hide unused buttons and to guide the children towards the desired one. We considered modifying the controller, removing the thumbstick altogether and even placing a banner with gameplay instructions next to the headset. None of these sounded like good solutions, but we were open to trying them out in order to allow a smooth gameplay.
The game was supposed to run unsupervised in a museum. Therefore, the entire experience should always be ready for a new game session, requiring no setup from the player whatsoever. This requirement was put into check by the controllers because when left idle, they turn off automatically and – unlike the headset – there is no way (via software or hardware) to setup the idle timer duration. Therefore, if there was enough time between game sessions to bring the controllers to an idle state, the next player would have to turn them on in order to play. We were left with no other choice but to add an instruction banner explaining how to turn the controllers on, which again, was far from ideal.
WMR controllers are turned on by pressing their Windows key for a few seconds. Since it had already been stablished that the users might have to turn the controllers on before playing, that key must have been easily accessible. That introduced an unexpected problem that didn’t surface until we installed the first version of the game for public access: the same key serves as a “home” key for the entire WMR ecosystem. When pressed, the current application is suspended and the user is brought to the WMR home, an environment that serves as a start screen for MR applications. From there, the user can have access to other applications (e.g. web browser, other games, settings, YouTube…) without any constraints at all. Unfortunately, this feature can not be disabled via software and physically disabling the key wasn’t viable because it’s the same key that turns the controllers on. Hence, we were stuck in a pit: disabling or hiding the key incapacitates the controller, and leaving it accessible gives the user total freedom over the computer. We could not solve this problem and it was one of the key factors that lead us to the Leap Motion switch.
Finally, the controllers are wireless and no cords keep someone from removing them from the experience. Attaching them to a piece of hardware would add more wires to the existing headset cords and could hurt gameplay and overall enjoyment. We didn’t expect anyone to bring the controllers home as a memento and decided to risk losing them, which didn’t happen during the short period of time the installation was displayed.
Facing all these challenges, we started to question our platform choice and to search for alternatives to WMR. But other platforms either didn’t solve our existing problems or introduced new ones. Our team had some experience with the Leap Motion and we decided it was worth a shot. We took a couple of days to implement the Leap Motion and its library into the game and by the end of it, we were quite surprised. Not only using the hands to interact with the VR elements solved all of the problems we had with the WMR controllers, but it also felt way more natural than pressing buttons. We decided to keep the WMR headset but we ditched the controllers for the Leap Motion. After some tests with children, we identified some problems regarding the Leap Motion (discussed on the next section), but they appeared to be way more manageable than the challenges we had with the WMR controllers.
The Leap Motion was meant to replace the WMR controllers as interaction devices in the VR world. Although the switch eliminated the challenges we faced with WMR, other characteristic issues surfaced once we started to use the Leap Motion.
Whilst migrating to the Leap Motion solution, it was clear the one of the game’s key interactions was far from ideal: throwing balls. Previously, the user could grab balls using the WMR controller’s trigger button and throw them by releasing it, using a bezier curve as an indicator of where the ball would land. Naturally, we thought that when using your hands, the user could simply grab the ball and instinctively throw it. It sounded like the obvious design choice. Although, during internal playtests, an instinctive throw was never successful and the ball would always drop during the movement. The reason became evident: a natural ball throw usually requires a backwards arm movement, bringing the player’s hands outside of the Leap Motion tracking area. In order to execute a successful throw, the user had to keep their hands in the tracking area during the entire movement, which did not look natural whatsoever. We overcame this obstacle by changing the ball throwing design and adding a ball cannon which throws balls inserted into its feeder. After some external playtests, we added some visual clues to the ball cannon (on the right) and it was clear that the ball throwing problem was solved. Children are naturally attracted to the cannon and the visual cues make its purpose obvious.
All playtests with the Leap Motion happened during winter, which introduced an unexpected and somewhat funny new challenge. Long sleeve shirts (wore due to low temperatures) often covered part of the children’s hands, occluding them from the Leap Motion. As a consequence, tracking was regularly compromised. It’d never happen during development because the developer (a.k.a. me) comes from a warm part of Brazil and isn’t comfortable wearing long sleeve shirts. Therefore, clothing occlusion was never present during internal playtests. This is yet another – funny – example of how even the smallest cultural differences can affect game development. We didn’t focus too much on this problem because of the following reasons. First, it didn’t seem to harm gameplay substantially because the players were still able to interact with the VR elements, despite he occasional loss of tracking. Second, a small percentage of children experienced this clothing occlusion. Third, the problem was seasonal and wasn’t present for a good part of the year. Lastly, instructions to tuck away long sleeves were added to the game instructions.
In order to integrate the Leap Motion into the VR setup, the sensor had to be fixed to the headset. During development, we examined the VR headset (Lenovo Explorer), trying to find the best place to attach the Leap Motion sensor. The gameplay focus mostly on hand movement at and above eye level. Therefore, the Leap Motion sensor should be angled on a way that is optimal for the proposed gameplay. Additionally, the sensor’s location could not block the headset sensors. Once we found the perfect spot, we attached the sensor to the Lenovo headset using regular office tape. Before shipping the game, we had to come up with a more permanent, safe solution. We ended up using glue to attach the Leap Motion to the headset. The Leap Motion never fell off the VR headset during the couple of months the installation was publicly available.
During the installation of the experience, one problem became evident: the headset the client chose (Samsung Odyssey) was different from the one used during development and it had a different sensor placement. Finding an optimal spot to attach the Leap Motion became a challenge because there’s less free room on the Samsung headset. After a few tests, it became obvious that there was no spot that would not occlude (even if just a bit) the headset sensors without compromising the hand tracking. In the end, we glued the Leap Motion in such a way that it would slightly obstruct the headset sensor, but not enough to compromise the WMR tracking. As a lesson, make sure you use the target hardware during development to avoid problems on installation/deployment day.
Since the game is intended to be displayed on public spaces, the environmental noise could muffle the game voices and sounds. Thus, a setup with headphones was desirable. Luckily, there was a WMR headset with built-in headphones: the Samsung Odyssey. During playtests, a new game requirement was added: the external TV used to display the gameplay to spectators should also play the game sounds. By the time the game’s first version was installed, there was a Windows Mixed Reality setting to enable or disable audio redirection to the headset, but no audio mirroring option was available. The audio could go either to the headset or to the TV, but never to both. After some research, we found some third-party Windows applications that could simulate audio cards and that could potentially solve the problem, but it introduced some delay between the headset and the TV audio, which was undesirable. By the time we started working on the game’s second version, the WMR settings were updated and a mirroring option was added, but it presented the same delay problem as the third-party mirroring solution. Currently, the delay is still present and we have not found a solution for it yet.
Windows Mixed Reality
Even though the WMR platform sounds really promising and the quality of the compatible hardware is quite impressive for a first generation of devices, the development platform has a lot of room for improvement.
Sadly, the WMR settings panel is quite basic and lacks many features that are present on other VR platforms (idling timeout, stop rendering when idling, “locked” mode, etc). We needed some of these features for our game and it was sad to see that they were just not there. When searching for solutions, we found out that it was possible to develop SteamVR applications for WMR compatible headsets and controllers, so we decided to give it a try. Since we were using VRTK – which supports both WMR and SteamVR – the switch to SteamVR took less than a day and it was successful. We we able not only to use the WMR hardware with SteamVR, but we also had access to SteamVR’s control panel. Both versions of the game are actually SteamVR applications.
Additionally, there’s not much control over WMR’s play area. First, you can’t customize the play area walls (e.g. color, shader). Second, you can’t choose which way is forward while setting up the area. Apparently, the WMR platform totally ignores which way you start setting up the area and chooses whichever play area side is the longest one. As a consequence, which way the player faces when the game starts might differ based on the WMR room setup process you perform. Being able to select which way is forward was crucial on this project because the physical, printed walls (seen at the picture in the intro) should match the virtual environment. As a workaround, scripts to rotate and save the player’s rotation using keyboard shortcuts were added to the game. After doing the WMR setup, we do our own VR play area rotation adjustments using the keyboard shortcuts. This problem could be avoided if the user had more control over the play area settings during room setup. Additionally, we often experience wrong floor height values and have to use the WMR standard “Floor Check” app to fix it. I believe that a simple and easy floor height check should be included during room setup, like the one from the Oculus Quest.
Overall, I found the WMR platform to be too closed to developers. Sometimes it felt like I was fighting against the platform in order to accomplish my goals. Many settings that should be available (either on the WMR control panel or via an API) can’t be changed whatsoever. That being said, I did notice that the team responsible for the WMR platform is listening to developers and is slowly implementing requested features. Maybe the platform is just not mature enough and all it needs is some time to evolve.
The game had specific requirements that made it a kiosk application: the computer should be easy to turn on and off (ideally just one physical button), the application should open automatically when the system finishes booting up, nobody should assist or instruct players, the application should run as long as the computer was on and the game should reset every time someone takes the headset off. At the time, unlike with the HoloLens, WMR didn’t have a kiosk mode, so we had to implement the requirements by ourselves.
We cleaned the target computer from anything that could interrupt gameplay: notifications and Windows Update were disabled, anti-virus and performance improvement programs were uninstalled, Wi-Fi was turn off and unnecessary startup items were removed. Then, we made sure the application would never be interrupted by energy saving features: we disabled Window’s energy saving options and the headset sleep timer [1, 2]. To prevent screen burn-in on the headset display, we added a “headset fade” component that fades the display to black when nobody is wearing the headset.
The WMR Control Panel automatically shows up once a WMR application is open. This behavior may seem logical, but it might get on the way of some kiosk applications. Two applications are being loaded and are competing for windows focus (i.e. stay on top): the Control Panel and the game. Even though the VR application is set to run on full screen mode, it might lose focus for the WMR Control Panel, which is undesired, given that spectators would not be able to watch the gameplay with the Control Panel on top of the game window. Luckily, we found a workaround on Unity’s forums that worked perfectly. Once implemented, this solution enables to developer to steal windows focus anytime. The game steals focus once its setup process (waiting for the headset to return to an idle state) is over and the WMR Control Panel never occludes the game screen.
The game should reset every time the headset is taken off so the next player gets to experience a fresh environment. We used SteamVR’s library to check for user presence via the built-in headset’s proximity sensor instead of Unity’s user presence check because the latter only changed state once the headset idled. As a consequence, if a player handed the headset to another person, the headset wouldn’t idle, the user presence variable wouldn’t change and the game wouldn’t reset. Using the proximity sensor enabled a more controlled reset process which always reset the game when players were switched.
Once the above features were implemented, the game behaved like a kiosk application that could be turn on and off using a single button (the computer’s power button) and that would work uninterruptedly for hours on end.
Before starting external playtests, we would introduce the game and instruct the children on how to play it. In retrospect, we shouldn’t done that. Since the game would be used as an unsupervised installation, nobody would explain to children how to play the game (in the end we added some small icons with basic guidelines on the spectator screen, but it still wasn’t as informative as the instructions we gave them). Therefore, we should’ve replicated the same kind of environment during external playtests. As a consequence, we didn’t catch usability and UX problems as soon as we should have. In fact, some of the problems were only evident after the first day the game was available to the public. On future projects, we will try to mimic the real world environment on external playtests in order to catch usability problems as early as possible.
Attention: this subsection is targeted at programmers.
I tried two new features on this project: async/await and C# 7. The first one was used as a replacement for coroutines because it delivers the same features but enables the usage of return values and error handling. You can read more about the motivation behind the switch and how to use async/await in Unity here. The new version of C# brings features as nested functions, pattern matching, tuples, more expression-bodied members and out variables. These features allow for more succinct but easily understandable code. You can read more about them here and here. Both async/await and C# 7 additions proved to be worth the time investment and I’m glad I took the time to learn about them.
Looking back, I wish we’d created more tools to automate repetitive tasks we had to perform during development. An example is taking screenshots of food items within the environment for client approval. The first time I had to take screenshots, I didn’t think that we would end up having so many approval iterations, so I didn’t implement a tool to automate it and took them manually. And every time I had to take the screenshots again, I’d think “I’m sure this is the last time. It’s not worth implementing a tool just for this”. And as one can imagine, I was wrong. Many times.
VoedingscentrumVR was a challenging project, mostly from usability and UX perspectives. We learned valuable lessons on how to design unsupervised VR applications, how to deal with WMR limitations, how to plan better playtests and how to design better interactions using the Leap Motion. It was a project that seemed an easy task at the beginning, but that showed us we had more to learn than expected. But at the end, we accomplished our task and developed a fun, educational VR game that children enjoy playing.
This was my first project after going back to game development. It was also my first VR project since 2016 and my first WMR project ever. Besides learning more about VR development, this project taught me a lot about project and time management and client communication. Although there were times I wish this project was over already (as we do in every project), I’m now grateful I was part of it.
As usual, if you have any questions, comments or corrections (specially corrections!), please leave a comment on the comments section. I hope to see you all on the next blog post!
While learning Functional Programming, there were two concepts that for me, were really hard set apart: currying and partial application. One day, I finally decided I would dig deeper and learn what each one represents and more importantly, why I mixed them up. In this article we’ll see what currying and partial application are, distinguish them, see how they’re used in functional languages, discuss why these concepts get mixed up so easily and present why they are so important.
One tiny thing before we start. Even though the concepts we focus are widely used in functional languages, we won’t use a purely functional language to explain them – Python. You have to trust me on this.
Currying is the process of transforming a function that takes many (N) arguments into a series of (N) functions, each one taking exactly one argument. Well, maybe doesn’t tell us much, but let’s take a look at the example code below and learn from example. The Python code above shows the definition of the mult_pow function, which takes 3 numbers, raises the first one to the power of the second, and multiplies the result by the third one. Pretty straight forward. The only way to evaluate this function is by calling it and providing the 3 arguments.
return x ** y * z
Now let’s try to curry this function. As described above, we will transform this function into 3 new functions, each one taking exactly one argument. We call mult_pow_c the curried version of mult_pow.
return lambda y: lambda z: x ** y * z
Let’s analyze this pice of code for a minute. The function mult_pow_c(x) takes one argument and will always return a lambda, which is… a function! The returned function takes exactly one argument – z – and returns another lambda. This last lambda, again, takes one argument and returns the result. Note that the outer variables x and y are part of the inner function’s closure and can be accessed by it. As you noticed, we did exactly what the definition of currying said two paragraphs ago: we turned a function that takes many arguments into a series of nested functions, each one taking exactly one argument. To put it in a more formal way, currying transforms a function of N-arity into N functions of 1-arity.
Let’s now try to use these two functions and see how their application differ. The first function call uses the uncurried version and the second one the curried version.
The first call (lines 1-2) demonstrates function application on the uncurried version and behaves as expected. But when we try to run the second call (lines 4-5), invoking mult_pow_c instead of mult_pow and passing three arguments to it, we get a type error telling us that we called that function with the wrong number of arguments. And indeed we did. Remember that mult_pow_c is the curried version of mult_pow, hence it only takes one argument. When we give it only one argument (lines 7-8), it returns a function, exactly what we expected from a curried version of mult_pow. In this context, if we have a function, what’s the most reasonable thing to do with it? Apply it! We do so (lines 10-11) and we still got a function in return, which makes sense given that mult_pow originally is a function that takes 3 arguments (3-arity). Only when we apply the functions three times (lines 13-14), we get the result we expected.
Well, well, well. It looks like we just created a super complicated way of defining functions. We need lambdas, unnecessary parenthesis when calling a function and there is a big knot in my head when I think of it. At first sight, it might look like a cumbersome way of doing simple things, but bare with me along the next section so we can understand why this is a powerful tool.
Let’s again use a Python example to illustrate the concept of partial application. The following code defines a function to multiply two numbers – mult – and a function to double (multiply by two) a number – double. Easy peasy.
Notice that the double is defined in terms of mult. This is a pattern that we’ve seen before and some call it “wrapper functions”. Another good example are the functions pow (power) and square (^2). We use previously implemented functionality and we fix one of the parameters to create a new function. This is also known as partially applying a function, simple like that. Given the original function, we limit its expressiveness by tying some of its parameters. For example, take the * operator (multiplication): initially, its image (the set of possible outcomes) includes all the integers, but when we partially apply it, by tying the first argument to 2 (*2), we reduce the function’s expressiveness by limiting its image to even numbers.
So that’s partial application. On a formal way, partially applying a function to X arguments is the process of transforming a N-arity function into a (N-X)-arity one by means of tying X arguments to values.
Currying vs. Partial Application
Even though the general idea – and even the semi-formal definitions – might seem closely related, keep in mind that unlike currying, partial application does change the function expressiveness and its meaning. Another way of seeing this is to look at both as a function, where currying would be Curry(f) and partial application would be PA(f, x…xn). Currying only takes a function as input and returns another function, whilst partial application takes a function and N (N > 0) parameters.
Also, remember the definitions previously introduced: currying is the process of transforming a function of N-arity into N functions of 1-arity. Partial application is the process of taking a N-arity function, X parameters and transforming it into a (N-X)-arity function. Now, let’s take a look on how these concepts are introduced in functional languages and why we mixed them up so easily.
Currying and Partial Application in Functional Languages
Even though we can manually curry a Python function (just like we did earlier), Python doesn’t support currying by default, which means that the following code doesn’t run, where mult_pow is the function we defined on the first section.
That might seem logical for a Python (Or C, C++, Java…) programmer, but it makes currying – something that as we’re going to see, is extremely powerful – a manual, programmer-dependent task.
The twist is that – unlike in imperative languages – in most functional languages, functions are curried by default. This might go by as a simple detail, but it has big consequences, as we’re going to see later. In Haskell or Clean, every function is curried by default. There’s no way of escaping out of it. As a consequence, every function that takes N arguments is actually a function that takes one argument and returns a function that takes N-1 arguments, just like we did on our first Python currying example (mult_pow_c). In addition, we get that for free: there is no need for any special implementation, syntax or annotation. That, aligned with higher-order functions gives us a great toolset to play with functions.
Partial application in such languages becomes so easy because of default currying. As we saw in the Python examples, partially applying a function isn’t as trivial as in a functional language.
Take the following Haskell example in consideration (if you’re not used to Haskell, don’t stop here, I won’t use it too much). We define the function multPow that takes 3 integers as arguments and returns an integer. Remember that since Haskell functions are curried by default, multPow actually is a function that takes an integer as parameter and returns another function that takes 2 integers as parameter.
To prove that, we use GHCi (GHC’s interactive environment), to partially apply multPow to the parameter 2 by doing pMultPow = multPow 2. Due to currying and partial application, pMultPow is a function that takes 2 integers as parameter and returns another one. You don’t believe me? Ask GHCi to give you some information about pMultPow by typing “:i pMultPow” and it will tell you exactly what we expected.
multPow :: Int -> Int -> Int -> Int
multPow x y z = x ^ y * z
-- In GHCi:
> pMultPow = multPow 2
> pMultPow 3 4
> :i pMultPow
pMultPow :: Int -> Int -> Int
One interesting fact is that you can notice that from the function’s type annotation itself. We’re used to read Haskell function type signatures in a non-curried manner. For example, we’d read multPow’s type signature as “it takes 3 integers and return an integer”, but we might as well read it as “it takes one integer and returns a function that takes 2 integers and returns an integer”. Both are equivalent in a language that is curried by default.
As we saw, partial application really shines in languages that have functions curried by default because there’s no extra step necessary to define curried function as in most imperative languages. That’s exactly why we easily mix up those concepts: we don’t see the currying, it’s done automatically. So every time someone tries to give an example of currying in a functional language, they fail to point exactly what it represents without using partial application, and the confusion is born. We can avoid that by explaining both concepts using a non-functional language and then bring the discussion back to the shiny, fancy functional world. It’s also why I chose to start this article with Python instead of Clean or Haskell.
Why is it so important?
The triad of higher-order functions, currying and partial application provides powerful benefits including expressiveness, higher level of abstraction and scalability, but there are two aspects that I want to focus: reusability and modularity.
With higher order functions, functions can be returned and passed as parameters, so you don’t need to overload a method, creating 10 different instances of it that are almost equal, except for one function call. Instead, make only one instance of that (previously overloaded) function that takes a function as parameter and every time you call it, pass the desired function.
With default currying, partially applying a function is trivial, which stops you from creating many wrapper functions just to fix some of the arguments. Let’s take the famous map as an example of how these concepts increase modularity and reusability. In Clean, map can be defined as in the code below. Start is Clean’s equivalent to C’s main and function type signatures use the arrow only before the output type (not in between inputs as in Haskell).
map :: (a -> b) [a] -> [b]
map _  = 
map f [x:xs] = [f x: map f xs]
double :: Int -> Int
double x = x * 2
Start = map double [1,2,3,4,5]
Start = map ((*)2) [1,2,3,4,5]
As we can see – by the (a -> b) – map is a higher-order function: it takes one function as input. It’s also polymorphic, given that it operates on lists of any type (a and b here are type variables). So, as expected, it maps a function over a list, element by element. In addition, we define the double function which… well, doubles a number. After mapping double over a list, we get a list in which all of its elements are the double of the original elements. Notice that we don’t limit the types a and b by any means. Given that, every time I need to apply a function over a list, I use map. Due to its high-order nature, I don’t need to worry about the list type as long as I provide a fit function as argument. That means that I don’t need to write 10 versions of overloaded maps, one for each kind I need, and even for different functions to map over! Imagine creating versions of map in C for ints and doubles, for doubling, squaring, dividing by 2 and many other possible operations. In a functional language, that’s not necessary.
Finally, let’s take a look on the second map example. Instead of defining a new function just to multiply an element by two, I can take advantage of Clean’s default currying and partially apply – on the spot – the * operator on the argument 2, which would give me a new function that takes one integer as input and return an integer – exactly what I need to use map, and exactly what double is.
This is a small and naive example to illustrate one of the biggest advantages of pure functional programming. Since function outputs only depend on its inputs (and nothing else!), and since we have higher-order functions with default currying and partial application, functions are little building blocks. You can just plug them, pass them around, partially apply them, run them over a list and all. That allows a level of reusability and modularity that simply isn’t possible (and won’t ever be) in imperative languages. Functions like map and foldr are reused on a daily bases by functional programmers, along with many functions they’re written themselves. Reusing functions is the bread and butter of a functional programmer. Every time you see yourself writing code that somehow resembles some code you’ve written, you think “how can I generalize this and then specialize it when I need to so I don’t need to write many slightly different versions of the same code?”
Even though currying and partial application are concepts that are related, they are not the same thing. Currying is a process of transforming a function into an equivalent one and partial application involves fixing a function to one of its arguments. Partial application really shines on functional languages where functions are curried by default, and that might be the reason why functional programmers get those concepts mixed up so easily. Allied with higher-order functions, these concepts increase modularity and reusability in an extraordinary way, and that’s one of the aspects where functional programming shines the most.
If you have any comments, suggestions, corrections (especially corrections), please fell free to leave a comments below.