What the heck is data-oriented programming?

Yehonathan Sharvit explains the why data-oriented programming is better than object-oriented programming.


We also talk about:
  • the basics of Data-Oriented Programming,
  • how Data-Oriented Programming compares to object-oriented programming and functional programming,
  • and how we can apply it in our codebases
Picture of Yehonathan Sharvit
About Yehonathan Sharvit
Yehonathan Sharvit, is author of the book Data-oriented programming. Yehonathan is a developer, author, and speaker. He has tons of experience in full-stack development using various languages such as Java, Javascript, and Ruby. But his favorite language is Clojure. He bundled all this experience and knowledge into his book which should be finished this summer.
This episode is sponsored by Tonic.ai – where your data is modeled from your production data to help you tell an identical story in your testing environments.


Read the whole episode "what the heck is data-oriented programming?" (Transcript)

[Make the transcript better, and improve the podcast’s accessibility via Github..]

Dr. McKayla 00:04 Hello and welcome to the Software Engineering Unlocked podcast. I'm your host, Dr. McKayla and today I have the pleasure to talk to Yehonathan Sharvit. But before I start, let me introduce you to an amazing start-up: tonic.ai - the fake data company. So what does tonic.ai do? I'm sure you know how complex and cumbersome it is to create quality test! It’s a never-ending chore that eats into valuable engineering resources. Random data doesn’t do it — and production data is not safe (or legal) for developers to use. What if you could mimic your entire production database to create a realistic dataset with zero sensitive data? That sounds amazing, right Tonic.ai does exactly that. With Tonic, you can generate fake data that looks, acts, and behaves like production because it’s made from production. Yet, tonic.ai guarantees privacy so your Datasets are safe to share with developers, QA, data scientists—heck, even distributed teams around the world. Visit Tonic.ai, to sign-up today or click the link in the shownotes to get a free, 2-week Sandbox. But now back to Yehonathan. Yehonathan is a software developer, author, and speaker. He has tons of experience in full-stack development using various languages such as Java, JavaScript, Ruby, but his favorite language is Clojure. And he bundled and packaged all of his experience and knowledge into his new book Data-Oriented Programming. This is also the reason why I invited him today. And I'm super thrilled that I will dive with him into how to design and build software that deals with information. So Yehonathan, thank you so much for being on my show. Welcome here.

Yehonathan Sharvit 01:19 Hello, McKayla, I'm really glad to be here with you.

Dr. McKayla 01:22 Yeah, I'm also really thrilled. And I heard that you are giving away one of your books, digital copy of your book to one of my listeners. So what do you have to do to win this book? Well, retweet today's episode and like it, and then you're in the pot to get one copy of Data-Oriented Programming from Yehonathan. So, but, Yehonathan, what is this book actually about?

Yehonathan Sharvit 01:46 Okay, so this book is about a simple way to reduce complexity of information systems. And by information systems, I mean, a program that manipulates data that doesn't belong to the program., data that comes from the outside world. For example, a front-end application that receives data from a back end, or a back-end application that fetches data from the database or an API. It could be also a web worker that reads data from RabbitMQ or Kafka needs to process it and passes it forward. So all those systems have in common that they manipulate data. But the data does not belong to it, the data or the information existed before the program, and will continue to exist after the program dies. And those kinds of systems, which is basically what we do on our day-to-day basis as full-stack developers need, I think, a different and a simpler approach to how we represent data inside our programs.

Dr. McKayla 02:50 So I realized that when I looked at the first topic, or the first subtitle, Unlearning Objects, it was very, very clear to me that it somehow has something to do with object-oriented programming, and that you want actually a paradigm shift here. And you think now about data-oriented programming, more or less, right? And so, that this can somehow improve our applications, and probably our maintainability. So what is that shift from object-oriented programming to data-oriented programming? And why should we unlearn objects?

Yehonathan Sharvit 03:24 Okay, that's a great question. And let me tell you how I see it. I practice meditation. I've been practicing meditation for around 10 years now. And to me, there is quite a similarity between what meditation guides us to do and objects. So in meditation, the basic assumption of the basic principle is that the main cause of our suffering doesn't have to do with the reality itself, but it has to do with the way we perceive reality. So our mind projects, something, an object on the reality. Object causes us suffering. But the reality on itself does not cause any suffering. And meditation guides us to remove the glasses that our mind puts on reality and to look at the reality as it is. And if you are able for a moment to look at the flower as an idea, if you are lucky enough to explain that it's a joyful experience. And even pain, if something is painful, if you're able just to feel the feeling of the pain with no interpretation of the mind, with no meaning about what does it mean about me but just experiencing the pain as a physical trigger, then the pain is not as painful as it seems to be. So that's what' my take on meditation is. And in a sense, object-oriented programming is causing suffering to the developers because instead of looking at reality, at data, at information as it is, we took our mental model, and then we stepped into infinite complexity system. But that's not that the data on itself or the information on itself is complex, it's the tool. The object that we use to grasp data is causing the complexity. So in a nutshell, I would say that most of the complexity that we have in our program is accidental complexity. It's complexity that we have created because of our mental model. It's not inherent complexity that has to do with the business domain.

Dr. McKayla 05:25 Can you give a concrete example of how object and object-oriented programming increase the complexity of our system?

Yehonathan Sharvit 05:33 Yes, so let's take... In the book, I take a simple example of building a library management system. And let's say in the library, you have books, and you can do many kinds of operations with books and authors. And in object-oriented, what you do is that you create an object that represents an author, and you have methods inside the objects to manipulate the author object. But the fact that you bundle together code and data or behavior and data creates complex hierarchy diagram. Suddenly, many classes need to input the author definition, just in order to access a simple field from the author. Or if you want to create a new author or an author with different fields, you have to create all the method of the author. Or if you want to test how the function of something behaves on an author, you have to create this complex object just to check for little things. Sometimes we say that you want the banana but instead of having the banana, you get the banana and the gorilla and the jungle, while you only want the banana, only one, the name of the author.

Dr. McKayla 06:36 And so how would I design that system? Let's say with this library, and the books, how would I design that in data-oriented programming? How would that look differently?

Yehonathan Sharvit 06:46 So it will be very simple. First of all, you separate data from code. And now you ask yourself, how should I represent data. So you have stateless objects or modules for the behavior. And you have data with no methods, data objects, or data classes, or data maps for the data. And what is data? Most of it is just record. So its maps are HashMaps, with fields and values. And that's it. So you represent your whole domain as a nested combination of maps and arrays, and integers and strings and kind of JSON, what JSON has brought to the world since it has been adopted, and not only JavaScript, it’s simple tools to represent data. What do we have in JSON? You have strings, you have integers, numbers, Boolean, node, to represent information that belongs to the real world.

Dr. McKayla 07:37 Okay, and so how would I structure the code in terms of files, for example? Where do I put all that data information? And the functional information will have separate files, two files, three files, 10 files? How does it work? One file?

Yehonathan Sharvit 07:52 Yeah, yeah, the thing that I didn't mention is that in data-oriented programming, we separate between the data representation and data validation. So let me give you an example. When you create a book with the title, and the publication year, and the list of authors, you just create a map, you don't create a book object. So you don't need any class to instantiate a new book, it's just a HashMap. If you're in a language like JavaScript, Ruby, or Python, you have dataliterate for that, if you are in a language like Java, or C # or statically typed language, you have a HashMap constructor. So that's it, you don't need any type definition for the book. And you organize your functions, you would in a regular object-oriented program, but the functions are much simpler, because they are stateless, they always receive as an explicit argument the data that they manipulate, so there is no implicit leads or set or reference to the object because there is no object.

Dr. McKayla 08:47 Okay, and so you were already talking a little bit about verification. And so for me, the first question that now comes to my mind is that actually with the typing, right, and that I can even create custom types, and then, you know, enforce them in the system. This gives me somehow also security that a book actually is how I imagined a book to be, how do we have that in data-oriented programming? How do we verify that it looks like this? And you know, like, because JSON again, right, we have we have some simple types. But you know, you could present me any data. And I'm not sure if this is...

Yehonathan Sharvit 09:22 Yes, exactly. So you could, for example, receive a map that you think is a is a book and a book, you're expected to have a field name title. And suddenly you get a map. And the field is named D title, for example. So how are you going to deal with that? And that's a great question. And the thing is that you deal with that, like when you encounter surprises in the real world. So surprises happen in the real world. If you, over the wire, let's say you access an API and you ask for a book. And for some reason, there is a bug in the API, the server that serves the API, and you receive a book with a room field. You need to deal with that. And the types won't be helpful there. The types will just fail, try to pass the JSON into a book with a field title, and there is no field title. And you will have either an exception or a nice error message. So this is applicable exactly the same way. In data-oriented programming, you define your schema. In a schema language like JSON schema, you receive your data presented as a hash map. And there are libraries in all languages alone to validate a piece of data against a schema, you validate, if it's valid, to move forward, if it's invalid, you deal with that by sending an error message to the user, by whatever, how you deal with that. But the important thing is that you embrace surprises, you don't assume that everything is going to work as expected. Like in real life, you know, you learn at university, you read books, you consult professionals, if then you expect that life will behave exactly as you have been told, you're going to be quite upset. So the approach is that you embrace changes and surprises. So you have modeled, you have expectations, but you have little maps or glasses, but you don't confuse... There is a great quote that I remembered today or yesterday about the map is not the territory. So objects are like map, schemas are like map. Quite often in object-oriented programming, we tend to think that the map is the territory. There are no objects in the real world, the information about the book is not an object. It's a piece of information, you might decide to represent it as an object in your program. So that's your map. But that's not the territory. While in when you represent your program information about the book, string map, just a map with keys and values, this is much, much closer to what it looks like for real, there is less of an impedance mismatch between how the information exists in the world and how you present in your program. So it's a move of humility in the sense.

Dr. McKayla 09:38 And so I think I get a good understanding of how it's different from object-oriented programming. But what's the difference to functional programming now? So where do they overlap? Or where do they differ, those two different ways?

Yehonathan Sharvit 12:10 Yeah, that's a great question. So first of all, they overlap, functional programming and data-oriented programming, by both approach in that we need to separate between code and data. So that's exactly the same. But when it comes to how should we represent data, in most functional programming languages, like Haskell, OCaml, etc, you'd use custom types to represent data. While in data-oriented programming, you don't use custom types, you use generic data structures. So that's the main difference. And there is another area where they overlap, the data is immutable, in both in functional programming, and in that alternate programming, we never ever mutate, we always create a new version of data to manage changes.

Dr. McKayla 12:54 Okay, so one of the big benefits from data-oriented programming that you see is that we don't have the custom types. I'm a big custom-type person, right? I'm always happy if I can create the type. And I have, you know, some enforcement around that, you know, you already described it to me, but still, what are some of the really cool things? Or why does it help me not to have the custom type and not such a strong typing system around what I'm doing?

Yehonathan Sharvit 13:23 I'm going to send you back the question, what is it that you liked so much about type?

Dr. McKayla 13:27 Well, so we were thinking about complexity, and I think I've never, not at least intentionally wrote a data-oriented program, right? So I can just go from my experience that I have so far. And I would say that strong typing helps me to avoid a lot of complexity around ensuring that, you know, the thing that I'm getting is actually what I'm getting, because it's already, you know, somehow, even before it runs, right, so it's definitely, I can statically assure that this works. Now, if I remove that layer, I feel like very insecure, I feel like oh my god, now I have to write all this logic that you describe of like, I'm dealing with this uncertain world. And I have to check this and that. And I totally understand that this is somehow reality. But on the other hand, why do I have to create my reality more complex, in my mind, right? More challenging than it has to be? Because if I have the type system behind it, I can say, well, I know that this is a book and this is how it looks like, and otherwise I will not receive it here.

Yehonathan Sharvit 14:27 Perfect. Great question. Before I answer, let me ask you another question. Do you also enjoy the kind of help that you ID gives you when you have type, autocomplete... that you can never miss type a field?

Dr. McKayla 14:40 Yes, I actually like that. Yeah. And I like that. I have a tooltip that I know Oh, these are the fields that I need. You know, like, I love all of that very much.

Yehonathan Sharvit 14:49 Okay, perfect. What you said totally resonate to me. Now, are there anything that you dislike this type, or for you it's the best that could be in the world?

Dr. McKayla 14:59 There are probably tons of things that I disliked. But now that I have to think about it, but you probably can tell me...

Yehonathan Sharvit 15:06 Okay, let me challenge you, when you have, let's say a book record. That is, how easy it is to serialize it to JSON?

Dr. McKayla 15:13 It depends. If I need a higher library, sometimes it's easy. Sometimes it's hard.

Yehonathan Sharvit 15:18 For my experience, it's always had with static typing. There is no native way to take record a custom type and to serialize it to JSON. You need always to hack something to use reflection, or some kind of hack. You agree.

Dr. McKayla 15:35 It's not straightforward. It could be, it could be more straightforward. Yes, I agree.

Yehonathan Sharvit 15:39 if if you program let's say you write a game, for playing with the library, then you create the books and the books stay in the program. So you never need to externalize data, that's fine. I have nothing to say against types, because it's a closed system. But if the moment you need to externalize your information to other, you need to go out of your types anyway.

Dr. McKayla 16:04 I never saw that as a real problem. Because it's just another step of things that we have to do, right.

Yehonathan Sharvit 16:11 I'm curious, how do you serialize? Let's say you have a nested record that describes the whole library with books and authors. How do you make from it? No, I'm ot saying that... Yeah, it definitely takes time, right? It takes time and effort. And you have to write it, right? I probably write the record, but I'm not sure if this is. Or I asked the other person... Yeah, you write to custom logic for any piece? Okay. So one challenge, another challenge. Let's say you have in the book, you have a field called title and field, publication year and a field author. And let's say you want to rename one of the fields, you want to call it, author, because that's how you need to send it over the wire for some reason, what are you going to do? Create a new type, exactly like the first type. So it's going to be called Book Two, which is exactly a book, but just the field also is called the book author. Usually there is a profusion, a profusion of types. And yeah, each little module that needs to have a different look at the same data or very similar data will create its own type. That's the kind of complexity I'm referring to. There are types that are intertwined with different things, different glasses to look at the same reality. Another example, let's say you have two modules that have the same type with the exact same field, but one type belongs to one module, and one type belongs to another module. Now you have one by module, a one by module b with the exact same field. If you compare them, the language will tell you it's not the same, two different type instance. But the fields are the same. So it's two glasses, that look like the exact same reality. So instead of comparing reality, you compare your glasses.

Dr. McKayla 17:53 Yeah, I think I probably would need a very concrete example to see the different instances of how or when I encountered those problems. But I see where you are heading towards. So what are the programs that you wrote, or the products that they created, where all of those challenges that you describe have been so predominant, that data oriented-programming really made sense? And how did you switch from, you know, like, probably had like already a codebase? So how do we go from that codebase that we have to a data-oriented code base? What steps do we need here?

Yehonathan Sharvit 18:20 Yeah, if you take a look at the book, you will see that basically, data-oriented programming is made of four core principles. You don't have to apply all of them, you can apply principle number one of data-oriented programming, and change your code base according only to this principle. So for example, you will take a class that combines code and data and split it into two classes: one for the code, one for the data, and it will be already... It will have beneficial impact, if will lower the complexity of your class diagram. After that, you can say okay, this piece of data, maybe I can represent it as a HashMap. Instead of creating my custom type, and see if it makes sense, maybe something that you need to send over the wire as JSON, instead of creating a custom type and then a custom JSON serializer take a look and say, Okay, this, it will make sense to leave it as a HashMap. And then you get the serialization JSON for free. Then the third principle, you can play with schema languages like JSON schema, and see how it looks like to define the schema of your API, defining the rules inside your classes. And then for example, you have a REST endpoint that expects a payload with a specific shape. You define the payload of the request... JSON schema, and you validate that the requests that you receive correspond to the schema. And finally, the last principle, you can avoid mutation and instead of mutating in place you try to create when it makes sense a new version of data and see if it helps in terms of state management.

Dr. McKayla 20:03 But it also means that it plays nicely together we can have an object-oriented or functional programming environment or system and then add...

Yehonathan Sharvit 20:12 Yes.

Dr. McKayla 20:13 Parts that are benefiting from the data-oriented paradigm and transfer that into that?

Yehonathan Sharvit 20:20 You could have an object-oriented programming style or language, embracing the data-oriented programming paradigm.

Dr. McKayla 20:28 Okay, but it can still live together, right? So I can have like a large part of my codebase has objects and works in an object-oriented way. And as you said, Maybe I'm taking just parts of some of those objects and separate data from logic and so on. But the rest will still be very OO, right? Object-oriented.

Yehonathan Sharvit 20:48 Yes. And that will be the advice I would give to someone, not to adopt the will directly but to try when it makes sense. And if you... it's a place where you're really scared, let's say if you don't have your types, it's fine. Don't start from this place.

Dr. McKayla 21:01 Okay. Another question that I had for you is that I wanted to say data-intensive application. And in our pre-recording session, you said no, no, don't say data-intensive application, say, information systems. Why? What's the difference for you here? I think I grasp it now a little bit. But I think probably it has to do that it doesn't have to be data-intensive. You can just have normal data usage. But why do you prefer information systems and not data-intensive applications?

Yehonathan Sharvit 21:32 Yeah, that's a great question. When I hear the word information, it's clear to me that it's something that exists outside my program. information could be in a database information could be in a program, it could be in a file, it could be in the real world. And when you say data-intensive application, to me, it sounds like something at scale with lots of data, lots of traffic, that has nothing to say about how to manage data at scale, it has to say how to represent data in a simple way.

Dr. McKayla 22:03 Okay. Yeah. I mean, a lot of systems have databases at least, right? Like they are probably applications that don't have a lot of information. They don't need a database, but I don't know many...

Yehonathan Sharvit 22:15 I mean, they need a database. But what I mean, is that the scale doesn't have to be... the throughput doesn't have to be big.

Dr. McKayla 22:22 I understand. Yeah. So actually, every application that has or deals with data and information that stores it, somehow retrieves it somehow, transforms it, it would be interested to look at data-oriented programming, in your mind, or are there some specific kinds of applications that you would say benefit more?

Yehonathan Sharvit 22:40 Any application that is in the stack of what we call full-stack development, front-end, back-end worker will be... it will benefit. I think where it doesn't make any difference with.... it won't benefit. If you want to write a compiler, or to write a game engine, or something that lives in a closed ecosystem, where you have no surprises. If your data never surprises you. And you don't need to send it to other, you don't need data-oriented programming.

Dr. McKayla 23:09 Yeah, I like that. That's very clear.

Yehonathan Sharvit 23:11 If you communicate with data to different systems that use different programming languages, then it makes sense.

Dr. McKayla 23:18 Okay, so the last question I have for you is about the style of the book, actually. So your book is actually written a little bit differently. It's a story, right? It's a story and the whole information, the learning experience is actually guided through personas, characters that are doing something, why did you choose to write the book that way? Which is very different from other technical books?

Yehonathan Sharvit 23:43 Yeah. So there are several reasons. One of them is that I like story. And I need to admit, it's difficult for me to read technical books. It's very difficult, I get bored, or I fall asleep. And when I read the story, I'm entertained, I'm energized. So I wanted to invent a story. But it's tricky to make a story in the context of technical material, but I wanted to. I had this idea in the back of my mind. And then when I wrote the first chapter, as a regular chapter, I struggled with finding the proper tone. I was either too enthusiastic and too selling or too boring. I didn't know how to deal with the objections that I know the readers will have. So I played a little game between me and myself, part of me that played the data with the programming mentor, and other parts of myself that would play the developer. I made little discussions between them. So the mentor would say, types are problematic. And the developer say, but, why? I like types. It brings me safety. Uh huh. And then he challenges m,e then I need him to answer his question.

Dr. McKayla 24:54 Oh, that's what you did with me today.

Yehonathan Sharvit 24:56 Yes, yes, exactly. I was already prepared. And sometimes, I felt like kind of schizophrenic. Like I have two personalities. But for me, it was the best way I found to deal with questions. And many readers told me that quite often when they read the book, in their mind, they see the character as the question just the moment the question arises in the mind. So that's how I decided to make dialogues between the character and others. Okay, now I have the dialogue with, they're going to talk forever with no context? So sometimes I put them in the coffee shop, and sometimes in the university and sometime in the office at some time in the park, and sometimes in the countryside. And then I made a little story around the character and their journey toward enlightenment.

Dr. McKayla 25:45Yeah. And so Manning, right, this is a Manning book, right? Manning has this pre-access, early access thing. Yeah. And you have people that actually read already your book, and you're getting feedback. Do they like this kind of new style? Are they surprised?

Yehonathan Sharvit 26:01 I think they're surprised and they like it, and they comment on it. And they provide very useful feedback on it allows me to improve the book, while I'm writing, I also have a reader from... As you may have guessed, I'm not a native American speaker. I was born in France. And I have a friend in Boston, I think, that liked the book and offered me to review the whole book, and reformulate it in proper English.

Dr. McKayla 26:26 Oh, nice. That's very good.

Yehonathan Sharvit 26:27 We had two weeks of fun together, we sent him the chapters, and it will send me reviews and comments. And it move dbeyond only correcting the English. He suggested little changes to the story, to the setup, to the characters. That was one of my best experience, my best interaction with readers, that they actually participated in the book, they contributed in the book. Is that open source, it's like an open-source book, in a sense. A collaborative book.

Dr. McKayla 26:54 Very nice, yeah. And so this actually brings me to the end of our episode, and I want to remind our listeners that they can win a book, you are giving away one copy. They have to retweet and like this episode to have this chance. And then in a week, 10 days later, I will, you know, raffle and pick one lucky person that can read the whole story and find out where the characters are around and how they're discussing the pros and cons of data-oriented programming. So Yehonathan, thank you so much for being at my show. It was really a pleasure to talk to you. And yeah, thank you so much.

Yehonathan Sharvit 27:32 Thank you, McKayla.

Dr. McKayla 27:33 This was another episode of the Software Engineering Unlocked podcast. If you enjoyed the episode, please help me spread the word about the podcast, send the episode to a friend via email, Twitter, LinkedIn, well, whatever messaging system you use. Or give it a positive review on your favorite podcasting platforms such as Spotify or iTunes. This would mean really a lot to me. So thank you for listening. Don't forget to subscribe and I will talk to you in two weeks. Bye.

Copyright 2022 Doctor McKayla