Facilitating the Spread of Knowledge and Innovation in Professional Software Development

Write for InfoQ


Choose your language

InfoQ Homepage Presentations Learning to Love Type Systems

Learning to Love Type Systems



Lauren Tan talks about type systems in TypeScript, Flow, and GraphQL that can improve confidence and help ship less bugs to production every day. She starts with why: a practical look at what the benefits are from embracing types. Then, a gentle introduction to the ideas behind them. Finally, she explores the possibilities of a type system over the network.


Lauren Tan is an Engineering Manager at Netflix. Although formally educated in Finance, she's been involved in the crossroads of technology, design, and business for more than a decade. She's also an active contributor to open source, maintaining a number of popular JavaScript and Elixir libraries.

About the conference

Software is changing the world. QCon empowers software development by facilitating the spread of knowledge and innovation in the developer community. A practitioner-driven conference, QCon is designed for technical team leads, architects, engineering directors, and project managers who influence innovation in their teams.


Today I want to talk about type systems in JavaScript. I think you've heard a lot about it today, and you may have heard about this very surprising nugget that more than 46% of users on npm are using TypeScript. So today we're going to look at why investigating a type system, whether or not you use TypeScript or Flow, is useful, and how it can help you write better code.

So I'm really excited about this talk. I spent way too many hours down the rabbit hole researching arcane things like type theory, but I've distilled it down, and I have some pro tips to share to help you write better programs. It's a very deep topic, and I've personally learned a lot about type systems as well as functional programming, and I hope you will too. Before we begin, there is a small warning. There is quite a bit of code and math in the talk. I'll try my best to avoid using jargon unless I absolutely have to, then I'll explain it so you will know what I'm talking about.

So can I get a quick show of hands, how many of you have shipped a bug into production before? Well, I know I have, and it's okay, right? We're all human, programming is hard, and it's really easy to add bugs into software, especially in the Wild West that is JavaScript. And I don't know about you, but when bugs inevitably happen, I feel pretty terrible about it. And it's very easy to experience a real crisis of confidence, especially if you can't figure out what the cause of the bug was. But when you do figure it out, all is good in the world, and your confidence is restored.

Now, I've been programming for a while now, and I've seen my fair share of JavaScript errors, as I'm sure you have too. Here we have everyone's favorite, "undefined is not a function." Recently, I've grown to really love working with types on top of JavaScript, and primarily I use TypeScript a lot for my personal projects as well as at work, and I'm in love. Flow is also a really attractive option. It has a number of differences compared to TypeScript, but you can also choose it to statically type your code, but in this talk I will be using TypeScript examples. I also wanted to briefly talk about GraphQL, which is a really interesting way to deliver a type system over the network.

My name is Lauren and I am an engineering manager in Netflix and I ship bugs into production. Confusingly, I have many internet handles. On Twitter I am sugarpirate_, on GitHub I'm poteto, but with an “e”. So clearly I'm not very good at the internet.

How Many Ways Can This Program Fail?

So what is a type system other than something that yells at you a lot? There are many different definitions out there, but my explanation is a type system is a way to add constraints to your code. And the type system is really there to help you enforce logic that is consistent within those constraints. But why? Why are constraints useful in code? So here's my bold claim. Constraints are useful because they limit the number of bugs in your programs. And let me explain with a simple example.

Here we have a function that takes one argument and divides it by two. Can't get any simpler than that. But how many ways can this function fail? I came up with some possible inputs that I could pass into this function and here we have our usual suspects like null, undefined, numbers, strings, and so on. Now, I'm going to enumerate over all of those inputs, and after running this code, we should see an array of test results. So let's run that in our console and see what happens. However, unfortunately for us, null and undefined do not have methods on them. So my code doesn't even run.

But wouldn't it be nice if the compiler could tell you that? I don't want to have to run this code in my browser to find these bugs manually, or have to write a million unit tests, because I'm pretty lazy. And this is really only a simple function. It's just division by two. Now, in JavaScript, your first instinct might be to do something like add checks for null or undefined, or if you want to get even extreme, you can define undefined as a function, goodbye to that error. But you end up having very defensive code, and you're checking these argument types in your functions, which are only checked at run time, so you're shipping more and more code to your browser. Which is fine I guess, but I know it's frustrating not being able to catch simple bugs like these before you get into production. So if we fix the errors again and we run our program, we'll see that most of these inputs throw an error, a type error specifically, or result in some kind of nonsensical value like not a number.

So let me ask you again, how many ways can that function fail? Think about all the possible inputs that you could pass into that function, like every single one of those. There is an infinite number of permutations of garbage you could pass into this function. You could give it an array of eleven 1s, array of a million 8s, 100,000 character string, it doesn't matter, they would all cause problems.

But using TypeScript, let's just add the simplest type annotation to drastically constraining this function down into only accepting one set of types. So now, x will only be a number, at least at compile time, and the compiler will yell at me if I try to do anything else. And if you run this function with every single number that's possible in JavaScript, in theory, none of them should cause an error, because you have now typed your program correctly, and in most cases, this function is what we call total.

So my bold claim today is that types and constraints will reduce the bugs in your program. Today we're going to look at a very brief introduction to TypeScript, some background into type systems, and how it can help you write better programs. And finally, we will explore what it means to have a type system over the network.

A Gentle Introduction to Types

So we briefly saw what types in TypeScript look like. Let's learn some of the basics just to establish a foundation, so everyone's on the same footing for the next topic. And I'm not going to go through everything here, so don't worry too much about syntax. Focus on the ideas and you can always look up syntax later. So first, let's look at variables, something I hope you use all the time. In TypeScript, you specify the type after the variable name, and these two examples are actually both equivalent, since TypeScript can infer the type, even if it's not explicitly declared. For this, you have to tell TypeScript what kind of array it is. So here we have number[], which means an array of numbers.

You can also define your own types, and here I've made my own type called ServerStacks. It is a union of string literals, meaning that anything that has a type of ServerStacks can only be one of these exact strings. You can also define things like interfaces and other types like that, but here my user interface describes the shape of some arbitrary user object, and you can also use interfaces to describe functions. This is a function that takes an argument of type User. So as the early example suggests, because isAdmin is a "boolean," we know that it's not possible to assign that value, that property, to the number one, which is a number. And this is all caught for you at compile time, so if you use something like VSCode or an IDE, you'll also get a nice squiggly line and an error message while you're programming.

So those are just the basic examples, and again, don't worry about the exact syntax. Remember the idea is everything has a type. It could be a primitive type like a number or a string, it could be a function, it could be a user-defined type. It could be an interface. These are all the things that describe your application.

Now, I really want to talk about how you can make these types reusable, just like how you make your code reusable. So, bear with this contrived example. makeArray is a very simple function that takes a value X and then puts that into an array as the first element. But what we have here only covers numbers, strings, and booleans. What about objects? What about the types that I defined earlier? What about symbols? What about instances? How can you type this function so you can accept any type, without defining all the possible variations, but still have the type safety that you want in your program?

The answer is to make it reusable by making it a generic function. Now, notice that the function logic itself has not changed; only the type signature has. So let's walk through this. The first thing you'll notice here is that this single letter T in the pointy brackets. This is a type variable, and much like a regular variable, it's for types. This tells TypeScript that the function is generic and can work over more than one type. If you want to get fancy, this is also called parametric polymorphism, but most people call it generics.

Now, in your function arguments, we say that x is of type T, and this will capture the type of whatever was passed into the function, and says, "let T be the type of that argument x." So this could be a string, it could be a number, it could be a user, whatever you pass into this function. Then we can use T later in the type signature. Finally, we specify the return type of this function, which in our case is an array filled with items of type T. So now when you use your function, TypeScript will know how to guide you. If you pass in a number, it knows that it will return an array of numbers. And if you pass in a string, it will give you back an array of strings. This is a very simple example, but also very powerful, because it lets you reuse your types and helps you not to repeat yourself, because when you use a typed language, it very quickly gets pretty verbose.

Now, here's another example to drive things home. This map function is generic and takes two type arguments, A and B, which we will use later. map takes two arguments, the first is a function, and the second, a list of items. The first argument is a function which takes a single argument of type A and then returns a single value of type B. The second argument to this function is an array of items that are of type A. Now, the return value of this function is an array of items that are type B. So just based on the type signature alone, it should be fairly obvious what this function does. And as you might have guessed, this function is one of the built in methods on the array prototype, called map. And with this type signature, TypeScript will know how to type your functions accordingly based on what you pass into it. And you'll be able to tell what is returned without actually having to evaluate the function, and what this really enables you to do is to start thinking about how your function's composed, purely based on the inputs and outputs that go into them.

Why Less is More

So we look at some high level concepts and again, don't worry about the exact syntax. Focus on the ideas. If you recall earlier, we talked about how many ways that program could fail. Having types means that your functions and the classes and bits of your code can specify constraints about what you can pass into them. And there is really a ton of things that functional programming can teach us here. And if you're interested in learning more, there are a lot of videos in this playlist, which I link to at the bottom of this slide.

Now, type systems are grounded in mathematics, and so is functional programming. Interestingly, type systems can also be applied to object oriented languages, as I'm sure you're all aware. And the concepts behind this are actually pretty interesting. And I think it can help us understand why type systems help us write better code, without all the jargon that you see up here.

So let's start with some quick high level overview. Functional programming is really the practical implementation of a bunch of different mathematical theories. We have proof theory, which is about constructive logic and mathematical proofs. We have category theory, which is about algebra and how things compose, about abstractions, and if you've ever dealt with things like monoids or functors or monads, these are all examples of categories within category theory. And finally, we have type theory, which is about programs and the idea that propositions are types. So Curry, Howard and Lambek are logicians and mathematicians who independently found that these three different mathematical theories were describing the same idea, just in different ways. And the correspondence between them is really the relationship between these theories, proof theory, category theory, and type theory.

Now, I promised if this sounds arcane to you, don't worry, I'm going to explain all that crazy talk. The key takeaway here is that propositions are types. For example, my type signature here states that given two numbers x and y, a number exists, and that programs are proofs of those propositions. Because just like in mathematics or in real life, you don't just get to say that things are true. You have to prove it or you have to show your work. And in this case, x plus y is a proof that a number exists. We're going to revisit that concept again later to see why it's useful, but first let's take a look at functions not in the programming sense, but in the mathematical sense.

So this is a function as described in mathematics. In plain English, this function maps an object of type A to an object of type B. Pure functions, if you ever heard of the term, are really mappings. For a given input, you always get an output, and the same output for that input. And note that object here doesn't refer to a JavaScript object, but an ordinary object unrelated to programming. If you remember what we said earlier, types are propositions. They assert that something is true or exists, for example, this function will return a number. And our programs are proofs. A number exists because here I have produced the number one in my code. Therefore, type checking is really proof checking, and if you work backwards from that intuition, what this means is that if you have good types in your system, you can really let the type system suggest the implementation for you; basically, almost like type-driven development. And in theory, based on the kind of advancedness of the type system you're using, you could declare all of your types and function type signatures without writing a single line of actual implementation code. And if it compiles, your program should just work, trademark. Then it's really a matter of filling in the details, proving that the propositions that you've defined are true.

So let's start with a concrete example. This function takes a list of items and returns the first one. Once you've written the type signature down, the implementation can sometimes seem pretty obvious. For example, if you return the list itself, you get a TypeScript immediately yelling at you that, “Hey, this isn't the right type,” because you're returning an array instead of a single item, as your type signature suggests. So if you try returning the first item in the list, then that's really all you can do, because there's very little else that would make this proposition true. And you could actually notice that you don't have to return the first index, you could return any one, and this would still type check. And this here is really an example of a limitation of a lot of type systems, that you cannot actually check these kind of things.

And other languages like Idris do have more advanced type systems where they have concepts like dependent types, where this sort of type checking can be possible. But what I want to highlight is that TypeScript and Flow are really not silver bullets. They are a pragmatic way to balance productivity and type safety within your code. And if we look at our map function again, if you try this out iteratively and return the items, if you just return the items that you got, you'll see some compile errors, since, of course, the type signature does not match. It does not prove that our proposition or type signature is true.

If you call the callback function on a single item in the list of items, then you're getting a little bit closer, except you're not returning a list of objects of type B; you're only returning a single B. So logically, the only thing you can do here is to take that callback function and run it on every item in the list. And this seems pretty obvious once you've done it a few times.

And what about React, our favorite library? The nice thing about React is it's very elegant to mental model, that the view is a function of your state. And in React, functional components, as their name suggests, are functions, because they take props and they give you back a ReactNode. So another fairly contrived example, if you try this out with the number one, obviously it will not work, because Lauren, one is a number, not a ReactNode. If you make it a function that returns a one, of course, it doesn't compile either. So logically, again, the only thing you can do is to return a JSX element, and your types will pass. Thinking in type signatures and pure functions make it much easier to compose functions and to start thinking about creating higher order components and higher order abstractions that you can reuse throughout your code.

So let's revisit the implications of the Curry-Howard-Lambek correspondence, the combination, or the correspondence between proof theory, category theory, and type theory, and learn how they can help you write better functions in your JavaScript applications today. Here we have a graph of the function, f of x is equals to x squared, and just like in Hannah's talk, what I want to call here is that the functions that you are very familiar to using in JavaScript are really similar, if not the same, as the kind of functions that you learned in high school math, they're the same conceptually.

And some terminology before we go further, just so we're all on the same page: pure functions as I explained earlier, map an input to an output. And when you're talking about functions like these, you also talk about things like the domain and the codomain. The domain of a function is the set of all possible inputs that can go into the function, for example, in this function, it's all the numbers possible. And the codomain of this function is the set of all possible outputs, which in this case is also numbers.

Now, has anyone heard of total or partial functions? I'm not talking about partially applied functions. That's a different concept. Now, I don't think it's very commonly talked about, but functions can really be pure or impure, but that is really independent of whether they are total or partial. A partial function is a function that is not defined for all possible input values. So think about the functions in your code base that don't always return a value, or they can return undefined, or it could even throw exceptions, or worse, an infinite loop or crash your entire computer. If you take that half function again, we can see that this is a pure function. It doesn't have any side effects, and it always gives you the same output for the input that you pass into this function. However, it is a partial function, because there are inputs that could cause it to throw errors.

So if we look at all the possible domains and codomains for this function, we know that if we pass this a number, it will give us back a number. And if we pass this function a string, it could either return a number or not a number, because JavaScript. And this is actually a feature, not a bug. JavaScript was designed to work in the worst of environments and it will try its hardest to make your code at least run even if it's terrible. And this understanding of how JavaScript was designed is really the core design philosophy behind many typed languages that compile to JavaScript, like TypeScript or Flow. It's never about achieving 100% type safety in JavaScript, because it's pretty much impossible, but you can try your best.

And getting back to this example, you also have your other types like null or undefined, you have objects, you have arrays, which are really special objects, and these all return not a number, which by the way, is a number. Thanks again, JavaScript. And if you pass in a symbol, you'll get a type error, which is not actually an output or a codomain, it is an exception that's thrown by the runtime.

Now, with a type annotation, we can restrict the domain of this function to only numbers. And therefore, this function is now total, at least at compile time, because it should never be undefined or throw an exception. And what is a total function? As the name suggests, and from the previous example, a total function is a function that is defined for all possible values of its input, which means that it always terminates and it always returns a value, never undefined or an exception.

Let's take a look at another function. fetchUser takes a user name as a string. It hits some API behind the scenes and returns a promise. And this is a partial function because the fetch could fail. I'd like to introduce you to something to make this function total. The promise now resolves with a user wrapped in an Either type. And now this function can only return a promise that resolves to this Either type. Whether or not it fails, it will always return this type. And the Either type really encodes the two possible outcomes from this code. It returns a Left, which represents a failure, or a Right, which represents success. What we have here is actually a monad, but I promised no jargon, so don't worry about what that means just yet. The idea here is that regardless of what you get back as output, you want to have something with a very similar interface, so you don't have to care if it was successful or if it was a failure.

Now, here's an example of implementation or proof of this type signature, of which there are many. There are many examples of different kinds of code that you could write to fulfill this type signature. So first, I'm just going to fetch my user from my API, and if it fails, I will return a left, which again represents failure, if the fetch is not a 200 OK. And I also will pass the left function an error object, which I can use later. And note that the error here is not thrown, it's just passed into the left function as an argument.

If the fetch was successful, we want to return a right, which in this case is an object that matches the user interface that we defined earlier. When we use this function, note that irregardless of whether the request fails, the return value of that function has methods that you can call safely. So this code on screen won't cause any issues even if the fetch fails, and you'll see that we're actually handling both the success and the failure scenario in the same block of code. And when you get into functional programming, you'll learn how to compose these concepts to do even cooler things, so you can either look up category theory, or come speak to me later after my talk.

The cool thing about the Either type is that you can capture the failure and then do something with it. So if I was building a UI, maybe I could alert the user with an onscreen message telling them that something went wrong. Or if I went down the RxJS path, then maybe I could also do that as well with different observable handlers. So let's take a look at one more example. This function detects the first element in the viewport inside of a scrollable container element, because this function can return undefined, it is a partial function as well. For example, what if I gave it a selector that yielded no elements? We can reduce the codomain of this function to make this total by wrapping this in an Option type, now that every input to this function will give us an output in this codomain, and you will never get undefined being returned from this function.

The Either type and the Option type are quite similar. The main difference is that the Either type also captures the failure value, while the Option type does not. And the Option type, or the monad, represents some value or nothing. Now, this is the actual function implementation, but don't worry about the details. You can see that there are two places where this function can give you an undefined value. querySelector is part of the HTML DOM methods, and it will return undefined if it finds nothing. And Array.find will also return undefined if it does not find anything within the array. But unlike the Either type, we're not going to capture the reason for failure, we're just going to return the value none. And if we do find the first visible element, we'll return that element.

So when you use this function, what does it look like? Irregardless of whether an element was found or not, we're going to get back a consistent interface, regardless of whether the type is something or nothing. And both of these objects, some and none, have the map function defined, so my code will not throw an error either, because, oh sorry, it will not throw an error like "undefined is not a function," our favorite error.

And finally, here's an example of a react component that fetches data. Because functional components are pure functions, you can also make them total. So this component lets me define four possible states that could be present in my component. No data, the fetch is pending, the fetch is failed, and the successful case with data. Writing total functions means that you think about these scenarios up front, and you also think about the way that these abstractions can compose with each other so you can do more with less.

I mentioned that domains and codomains describe the set of all inputs and outputs into your functions. So whenever you talk about things like domains and codomains, you're also talking about set theory. And in set theory, cardinality means the number of items of the set. Let me just explain why this is a useful concept and not some arcane math thing. I did make a bold claim earlier that the more specific your types and the more constraints you have in your code, the less likely you're going to experience bugs. And this is not a silver bullet by any means, but lowering the cardinality of the code does mean constraining the number of possible inputs in the function domain, and, therefore, it should prevent bugs from happening.

So if we start from the basics, a set is informally a collection of objects. In our context, we want to think about the set of all values and the types that we're dealing with in our functions. Here, we have a type called Conferences. Note that this type is not a string, which would match every possible string, but because there are only three possible values that I passed into the type, we say that the cardinality of this type is three. But again, if it was just a string type, then there could be an infinite amount of strings that would match this type.

So what I'm trying to say is that primitive types are not precise. If you look at the cardinality of all the primitive types that you have in JavaScript, the only types with finite cardinality are booleans, nulls, and undefined, and you probably don't want to write a program just using those types, unless you are very hardcore. Even the non-primitive type object has an infinite cardinality, and any possible object that you have in your system you can think of would match this type. So defining precise types can really help reduce bugs in your code, so let's look at some examples.

Here we have a generic function, toString, which calls the toString method on the object that was passed into it. And the strange thing here is that even though null and undefined don't have the toString method defined on them, as we saw earlier, this will still be type checked, or this will still be considered type safe by TypeScript. And what this means is that you have lost type safety, because this will cause a type error if we run this code and that kind of totally defeats the purpose of using TypeScript in the first place. So this is a function with unbounded polymorphism, which basically means that T, the T type variable, can be literally any type, and you almost always want to stay away from the any type, because it's not precise.

We know in advance that null and undefined are going to give us type errors, so we want to make this a bounded generic function. To do this, we're going to add a constraint to our generic type. If you look at the code here, you'll see that NonNullable is not actually JavaScript. It is TypeScript syntax, and it is a type that will constrain down that generic type T into something that's much narrower. And what it does is it tells TypeScript to exclude null and undefined from all of the possible types that could be captured by the type variable T.

And now, assuming everything else in your code has toString defined, you have made this function total, since it will always return for every possible input. If you want to think about this visually, imagine a Venn diagram with two sets; one being the set of nullable types, null or undefined, and the set of all types, which is basically every single type in your code base. Here I've only listed the default JavaScript types, but the types that you've defined in your code could also be members of this set two. And what we really want here is the difference between these two sets. All of the types minus the nullable types, which will leave you with the non-nullable types.

The non-nullable type is implemented with conditional types, a new feature, or relatively new feature, in TypeScript 2.8. And this is the basis for many of the new helper types that they've just added. The great thing about TypeScript is that it now ships with a lot of these little helper utilities built in now. So they're very useful when you start thinking about adding constraints to your code and making it more strict. And again, there's a link at the bottom if you want to check out more details in the TypeScript source.

So we've talked a lot about functions and functional programming concepts and arcane jargon that I promised I wouldn't talk about. But at the end of the day, just keep in mind that what you really want is to be pragmatic. JavaScript is a multi-paradigm language, so you can do whatever you want. We can do object-oriented. You can do functional programming. You can do functional-reactive programming. You don't have to get too religious about it. That's what's great about JavaScript.

And TypeScript is pretty similar. It's really about pragmatism. And it wants to strike a balance between productivity and type safety, because you don't have to go all-in in order to benefit from it. And again, it's not going to solve all your problems. You're not going to become a Fortune 500 company if you use TypeScript, because, you know, TypeScript will help you catch the obvious bugs, but it can't catch every single bug for you. The nice thing about taking this approach is that you really start thinking about the different kinds of constraints that, or even invariants in your code, that always have to hold true. And you think about things like fault tolerance and making your functions total, but only where it makes sense.

Types over the Network

I want to close off by talking about GraphQL and explore the idea of an end-to-end type system. GraphQL is a query language and specification, and just like TypeScript and Flow, it's also strongly typed, and allows you to define a schema that describes your entire object graph. The benefit of this approach is that your clients ask for exactly what they want, and they get predictable results every time. GraphQL APIs are organized in terms of types and fields and not endpoints, and this great because types ensure that your client applications only ask for what's possible, and they provide you clear and helpful errors when they try something that's not implemented. They also help with avoiding writing manual parsing codes, since you have type information.

And speaking about exposing type information, this also represents a lot of opportunities for interesting tools to be built on top of this platform. For example, here we have a tool called graphical, which is a visual playground that you can mount on top of any GraphQL compliant API. And what this means is that if you're using GraphQL anywhere in your tech stack, you get documentation for almost free. And this is a really great developer experience for clients of your service.

GraphQL inverts control back to your client, and they request only the fields they want, nothing more, nothing less. And the backend developers here may be pretty alarmed by that, since the clients can just ask for anything they want, any combination of fields, so, you know, what about things like N+1's, and that's definitely a very valid concern. But GraphQL on its own is really decoupled from the way that you retrieve your data. So this is not a problem that is unique to GraphQL. And so tools already exists out there that can solve these kind of concerns. For example, you may want to look at Facebook's DataLoader module, which will coalesce requests in a batch before you execute them.

But the nice thing about GraphQL and its type schema is that you get a lot of the same developer ergonomics that you have in TypeScript. And the only difference really is that the type system is now communicating to you over the wire instead of in your editor. And this is really great when you're working in very large distributed teams. The GraphQL type system is still evolving, and there's so much potential in where it could go. But that led me to a very interesting idea. As Neha introduced me, I work at Netflix, and we do have a lot of microservices. So I was thinking to myself recently, what if we had a distributed type system? What would that look like? Not just type safety in the editor while I'm creating my client applications, but type safety across multiple teams in a very large organization.

So this is a lot of what we're thinking about in my team at the moment. We're building tools that power the Netflix studio, and we're trying to make it more efficient and to empower better decision-making. And the goal is really to help with content creation, which is a very highly complex domain, spanning multiple business functions like finance, legal, business, staffing, creative, which are also very tightly connected as a graph. And there are a lot of different core entities that we deal with everyday, such as talent, movies, vendors, deals, projects, what have you, across multiple systems.

So the challenge here is really in connecting diverse sources of enterprise data across multiple fast-moving domains. And we're investigating a couple of things here. We are investigating the idea of a GraphQL API gateway in the form of a hybrid between service orchestration and choreography. We're also exploring, as I mentioned earlier, the distributed type system, which can enable developer productivity across your organization in real time, allowing you to move quickly while building up a larger graph that describes your entire business, but allow you to stay consistent as the different data models and the different domains in your business evolve.

What's really exciting here are the possibilities at the various levels in your organization. For example, as a UI engineer, one of the benefits of this approach is that you can now think about things like autocompletion in your editor or your IDE while you're building client UIs. And you can see even new fields come online as other teams are evolving the graph, right in your editor. Another thing you may be able to do with this technology is to know things like the impact of page load time by field, because you can tell exactly how long it takes for a field to be returned by this graph.

As a backend engineer, imagine if you had pull request feedback, or even build failures, whenever a schema change in your service will break clients of that data. That's something that we encounter pretty often. And especially if you're building microservices that are heavily dependent on by different people, you do want to make sure that any interface changes are properly communicated. And also, what about visibility into the aggregated graph of all of your different data models as it evolves over time?

So to do this, we're thinking about combining gRPC, GraphQL, and TypeScript, which are all strongly type languages, combining them to create the ultimate end-to-end type system, like a kind of a strange Voltron. This is not perfect at the moment. Although Node has the best support for GraphQL tooling currently, the generated gRPC code in NodeJS isn't the greatest. And this kind of architecture still has many rough edges. For example, code generation between the three formats is not one to one, since type information from the different formats cannot be cleanly transformed from one to the other.

Now, this all sounds nice, but how do you use it? Unfortunately, a cohesive solution doesn't quite exist yet, so you do have to glue many pieces together. But I will have to say that TypeScript and Flow and type languages help make this possible. It's always tough when you're doing things like this about architecture to have multiple teams move towards a common goal. But because in JavaScript, TypeScript is so easily refactored to words in an incremental fashion, adoption hasn't been that bad.

There are also some SaaS products today that can help you get close to this pattern if you're interested. For example, this is Apollo Engine. It has a feature that lets you track the evolution of your GraphQL schema, and do things like raise alerts when breaking changes are introduced. There are also many open source tools to make translating the schemas between these different technologies and types easier. And lastly, another interesting project that you might want to look at is this unofficial Google project called Rejoiner, which allows you to stitch together gRPC microservices across multiple languages, and exposing that as a single GraphQL API.

So we've discussed a lot today. First we looked at an introduction to basic types in TypeScript. We looked at things like generics and how do we type signatures, a reminder that generics help you write reusable types, just like you do with reusable components and functions. Then we looked to functional programming to see how it can help us or inspire us to write better functions that are both pure and total. We also explored a high level introduction to the Curry-Howard-Lambek correspondence; the idea that propositions are types, programs are proofs of those propositions, and that type checking is really the same as proof checking.

And lastly, we looked briefly at GraphQL and we explored what type systems can give you, even when they're delivered over the network. I'm really excited to see how this idea evolves and enables greater productivity in large microservice-heavy organizations. We've only scratched the surface today, but I hope you learned a little bit about type systems and functional programming, and how it can help you write better JavaScript. Thank you all so much for listening, you've been great.

Questions and Answers

Participant 1: What would you recommend to help people start integrating to their existing projects? For any type system.

Tan: Any type system into an existing code base? I think TypeScript and Flow are really great tools, because they let you incrementally adopt TypeScript for example, which I have the most experience with. You can get started by simply renaming your JavaScript file to a .ts file, and you can literally do nothing else but you get some benefits already from TypeScript. And some recent news that actually makes this even easier is that now Babel has an integration with TypeScript, so Babel can actually compile your TypeScript for you without even having to do anything special.

So I think that's a really great way to just incrementally refactor your existing JavaScript into TypeScript, whereas other typed languages, like ReasonML or like PureScript, you have to kind of start, rewrite your whole system from scratch, which isn't really that incremental. So, hopefully that helps.

Participant 2: I have a question for you that's related to that, which is when you started loving type systems, was this a change that was rolled out slowly over the team as well, and how was the type system integrated into the codebase?

Tan: The question was really how did we get started adopting TypeScript in our organization? And I think it started very incrementally. One of the early ways we started was just simply by adding type definitions to some of the libraries that we share among the different teams. And if you use an editor like VSCode, you actually can benefit from those type declarations, whether or not your project was written in TypeScript, because one of the really interesting things about the TypeScript language server, from what I understand, is that it will actually attempt to install the type definition packages for you, whether or not you're using TypeScript. And that's how we can give you autocompletion in your editor for certain things like, you npm install Lodash, and suddenly you have Lodash autocompletion. How does it do that? That's actually through TypeScript. So that was one of the ways we started and it was a great way to start thinking about all the different interfaces that we were implicitly relying on and now making them more concrete.

Participant 3: So the GraphQL is not actually a JSON body that we send it. So it has its own specification. So what is the media type that we set it for sending a payload?

Tan: It is still a JSON blob, but the main body is captured as a string. So you're right in the sense that it's a little bit harder to work with; especially if you're doing service-to-service communication, you now have to implement a GraphQL parser to be able to translate that request into something else. Which is why we, for example, in the example I showed earlier, we're still using things like gRPC or Thrift for service-to-service communication. And then we expose GraphQL as a kind of product-facing API for our clients.

Participant 3: In enterprise, the teams, the enterprise has spent money on building something more like a gate pay, wherein we can define an interface and validate an interface before it goes into production. So do we have any such product for GraphQL? Open source?

Tan: In terms of validating your interfaces before you go into production?

Participant 3: Yes, something like Swagger, that validates the interface.

Tan: GraphQL is only a specification, so I don't think GraphQL itself with provide that for you, and I'm, off the top of my head, I don't know if there are packages like that. But I'm going to guess that there are. But I'm not sure.

Participant 4: Yes, I was just wondering if you have any preferred libraries for algebraic data types playing well with TypeScript, and also, TypeScript doesn't have type inference. At least it doesn't seem, it's pretty verbose, kind of like Java or something? And it's just aesthetically not pleasing to me. Do you know of any other type systems that compile to JavaScript that have type inference?

Tan: So first of all, to answer your first question, which was about what do I prefer for algebraic data types, there is a really great package called fp-ts. It's by a mathematician, his name, I can't remember his name, but his GitHub username is gcanti, and he also has a very interesting package called io-ts, which is an interesting way to actually add run time type checking. I haven't really used it myself, but it's a very interesting exploration into that. So fp-ts is Fantasy Land compliant, so we can use that, if you're familiar with that style of functional programming, and it's pretty well-written. It's a TypeScript-first project, so it has even support for higher kinded types, and he goes into quite great detail about how it supports that. So that's definitely where I would start.

And then in terms of type inference, TypeScript does have type inference, it's just not very advanced. So it has a basic understanding of some types. And here are really the main differences between TypeScript and Flow. So, the way the TypeScript compiler is created, it's based on the AST, an abstract syntax tree. So it can really only understand the types that it's aware of. And once it encounters a type that it has no idea what to do with, it just breaks. But a language like Flow is actually based on, the compiler is graph-based, so when it defines a type that it has no idea what it is, it can just wait, come back to it later, and then try to infer what that type is. So that's really one of the major differences between Flow and TypeScript. But, yes, TypeScript does have type inference, it's just not super great.


See more presentations with transcripts


Recorded at:

Jan 26, 2019