Svelte and Reactivity with Rich Harris

Hosted byCharles Lowell and Taras Mankovski

September 4th, 2019.

Rich Harris talks about Svelte and Reactivity.

Rich Harris: Graphics Editor on The New York Times investigations team.


Please join us in these conversations! If you or someone you know would be a perfect guest, please get in touch with us at Our goal is to get people thinking on the platform level which includes tooling, internalization, state management, routing, upgrade, and the data layer.

This show was produced by Mandy Moore, aka @therubyrep of DevReps, LLC.


CHARLES: Hello and welcome to The Frontside Podcast, a place where we talk about user interfaces and everything that you need to know to build them right.

TARAS: It's actually a really nice, Rich and I'm really, really happy to have a chance to actually chat with you about this because Svelte is a really fun piece technology. In many ways, it's interesting to see our technology evolve and our industry evolve through innovation, real innovation. I think Svelte 3 has really been kind of that next thought provoking technology that kind of makes you think about different ways that we can approach problems in our space. So, really excited to chat with you about this stuff.

RICH: Well, thank you. Excited to be here.

TARAS: I think quite a lot of people know, Rich, about your history, like how you got into what you're doing now. But I'm not sure if Charles is aware, so if you could kind of give us a little bit of a lowdown on where you kind of come from in terms of your technical background and such.

RICH: Sure. I'll give you the 30-second life history. I started out as a reporter at a financial news organization. I had a Philosophy Degree and didn't know what else to do with it. So, I went into journalism. This was around the time of the great recession. And within a few weeks of me joining this company, I watched half of my colleagues get laid off and it's like, "Shit, I need to make myself more employable." And so gradually, sort of took on more and more technical responsibilities until I was writing JavaScript as part of my day job. Then from there, all these opportunities kind of opened up. And the big thing that I had in mind was building interactive pieces of journalism, data-driven, personalized, all of that sort of thing, which were being built at places like the New York Times, and The Guardian, and the BBC. That was the reason that I really wanted to get into JavaScript. And that's guided my career path ever since.

CHARLES: It's interesting that this D3 and all that did come out of journalism.

RICH: It's not a coincidence because when you're working under extreme time pressure and you're not building things with a view to maintain them over a long period of time, you just need to build something and get it shipped immediately. But it needs to be built in a way that is going to work across a whole range of devices. We've got native apps, we've got [inaudible], we've got our own website. And in order to do all that, you need to have tools that really guide you into the pit of success. And D3 is a perfect example of that. And a lot of people have come into JavaScript through D3.

CHARLES: And so, are you still working for the same company?

RICH: No. That's ancient history at this point.

CHARLES: Because I'm wondering, are you actually getting to use these tools that you've been building to actually do the types of visualizations and stuff that we've been talking about?

RICH: Very much so. I moved to The Guardian some years ago. And then from there, moved to Guardian US, which has an office in New York. And it was there that I started working on Svelte. I then moved to the New York Times and I'm still working on Svelte. I've used it a number of times to build things at the New York Times and the people have built things with it too. And so, yeah, it's very much informed by the demands of building high performance interactive applications on a very tight deadline.

CHARLES: Okay, cool. So I've probably used, I mean, I'm an avid reader of both Guardian and the New York Times, so I've probably used a bunch of these visualizations. I had no idea what was driving them. I just assumed it was all D3.

RICH: There is a lot of D3. Mike Bostock, the creator of D3, he was a linchpin at the graphics department for many years. Unfortunately we didn't overlap. He left the Times before I joined the Times, but his presence is still very much felt in the department. And a lot of people who are entering the industry, they're still becoming database practitioners by learning from D3 examples. It's been a hugely influential thing in our industry.

TARAS: How long is a typical project? How long would it take to put together a visualization for an article that we typically see?

RICH: It varies wildly. The graphics desk is about 50 strong and they will turn around things within a day. Like when the Notre Dame burnt down a couple of months ago, my colleagues turned around this interactive scroll driven webGL 3D reconstruction of how the fire spreads through the cathedral in less than 24 hours, which was absolutely mind blowing. But at the same time, there are projects that will take months. I work on the investigations team at the Times. And so, I'm working with people who are investigating stories for the best part of the year or sometimes more. And I'm building graphics for those. And so that, it's two very different timescales, but you need to be able to accommodate all of those different possibilities.

CHARLES: So, what does the software development practice look like? I mean, because it sounds like some of this stuff, are you just throwing it together? I guess what I mean by that is, I guess the projects that we typically work on, three months is kind of a minimum that you would expect. So, you go into it, we need to make sure we've got good collaboration practices around source control and continuous integration and testing and all this stuff. But I mean, you're talking about compressing that entire process into a matter of hours. So what, do you just throw right out the window? What do you say? "We're just doing a live version of this."

RICH: Our collaboration processes consist of sitting near each other. And when the time calls for it, getting in the same room as each other and just hammering stuff out on the laptop together. There's no time for messing around with continuous integration and writing tests. No one writes tests in the news graphics, it's just not a thing.

CHARLES: Right. But then for those projects that stretch into like three months, I imagine there are some. Do you run into like quality concerns or things like that where you do have to take into account some of those practices? I'm just so curious because it sounds like there's actually, the difference between two hours and two months is, that's several orders of magnitude and complexity of what you're developing.

RICH: It is. Although I haven't worked on a news project yet that has involved tests. And I know that's a shocking admission to a lot of people who have a development background, but it's just not part of the culture. And I guess the main difference between the codebase for a two-hour project and a two-month project is that the two-month project will strive to have some reasonable components. And that's, I think, the main thing that I've been able to get out of working on the kinds of projects that I do is instead of just throwing code at the page until it works, we actually have a bit of time to extract out common functionality and make components that can be used in subsequent interactives. So, things like scroll driven storytelling, that's much easier for me now than it was when I first built a scroll driven storytelling component like a couple of years ago.

CHARLES: Yeah. That was actually literally my next question is how do you bridge that, given that you've got kind of this frothy experimentation, but you are being, sounds like, very deliberate about extracting those tools and extracting those common components? And how do you find the time to even do that?

RICH: Well, this is where the component driven mindset comes in really handy, I think. I think that five or 10 years ago when people thought in terms of libraries and scripts, there wasn't like that good unit of reusability that wasn't the sort of all encompassing, like a component is just the right level of atomicity or whatever the word is. It makes sense to have things that are reusable but also very easy to tweak and manipulate and adapt to your current situation. And so, I think that the advent of component oriented development is actually quite big for those of us working in this space. And it hasn't really caught on yet to a huge degree because like I say, a lot of people are still coming with this kind of D3 script based mindset because the news industry, for some interesting and historical reasons, is slightly out of step with mainstream mode development in some ways. We don't use things like Babel a lot, for example.

CHARLES: That makes sense, right? I mean, the online print is not like it's a React application or it's not like the application is all encompassing, so you really need to have a light footprint, I would imagine, because it really is a script. What you're doing is scripting in the truest sense of the word where you essentially have a whole bunch of content and then you just need to kind of --

RICH: Yeah. And the light footprint that you mentioned is key because like most new sites, we have analytics on the page and we have ads and we have comments and all of these things that involve JavaScript. And by the time our code loads, all of this other stuff is already fighting for the main thread. And so, we need to get in there as fast as we can and do our work with a minimum fuss. We don't have the capacity to be loading big frameworks and messing about on the page. So that again is one of these sort of downward pressures that kind of enforces a certain type of tool to come out of the news business.

TARAS: A lot of the tooling that's available, especially on like the really fatter, bigger frameworks, the tools that you get with those frameworks, they benefit over long term. So if you have like a long running project, the weight of the abstractions, you've experienced that benefit over time and it adds up significantly. But if you're working to ship something in a day, you want something that is just like a chisel. It does exactly what you want it to do. You want to apply it in exactly the right place and you want to get it done exactly, like you want the outcome to be precise.

RICH: That's true. And I think a lot of people who have built large React apps, for example, or large Ember apps, they sort of look at Svelte and think, "Well, maybe this isn't going to be applicable to my situation," because it has this bias towards being able to very quickly produce something. And I'm not convinced that that's true. I think that if you make something easier to get started with, then you're just making it easier. If you build something that is simple for beginners to use, then you're also building something simple for experts to use. And so, I don't necessarily see it as a tradeoff, I don't think we're trading long-term maintainability for short term production. But it is certainly a suspicion that I've encountered from people.

TARAS: This is something that we've also encountered recently. It's been kind of a brewing discussion inside a front side about the fact that it seems to be that certain problems are actually better to rewrite than they are to maintain or refactor towards an end goal. And we found this, especially as the tools that we create have gotten more precise and more refined and simplified and lighter, it is actually easier to rewrite those things five times than it is to refactor it one time to a particular place that we want it to be. And it's interesting, like I find this to be very recent, this idea is blossoming in my mind very recently. I didn't observe this in the past.

CHARLES: Do you mean in the sense that like if a tool is focused enough and a tool is simple enough, then refactoring is tantamount to a rewrite if you're talking about 200 or 300 lines of code? Is that what you mean?

TARAS: Yeah. If you're sitting down to make a change or you have something in mind, it is actually easy to say, "Let's just start from scratch and then we're going to get exactly the same place in the same amount of time." But this kind of mantra of not rewriting makes me think about that, makes me question whether that's actually something that is always the right answer.

RICH: I definitely question that conventional wisdom at all levels, as well. I started a bundler called Rollup as well as Svelte more recently. And Rollup was the second JavaScript bundler that I wrote, because the first one that I wrote wasn't quite capable of doing the things that I wanted. And it was easier to just start from scratch than to try and shift the existing user base of its predecessor over to this new way of doing things. Svelte 3 is a more or less complete rewrite. Svelte has had multiple, more or less, complete rewrite. Some of them weren't breaking changes. But Svelte itself was a rewrite of an earlier project that I'd started in 2013. And so in my career, I've benefited massively from learning from having built something. But then when the time comes and you realize that you can't change it in the ways that you need to change it, just rewrite it.

And I think that at the other end of the spectrum, the recent debate about micro frontend has largely missed this point. People think that the benefit of the micro frontend is that people don't need to talk to each other, which is absolute nonsense. I think the benefit of this way of thinking about building applications is that it optimizes for this fact of life that we all agree is inevitable, which is that at some point, you're going to have to rewrite your code. And we spend so much energy trying to optimize for the stability of a code base over the long term. And in the process, lock ourselves into architectural and technical decisions that don't necessarily make sense three or four years down the line. And I think as an industry, would be a lot better placed if we all started thinking about how to optimize for rewrites.

CHARLES: So for those of us who aren't familiar, what is the debate surrounding micro frontends? This is actually something I've heard a lot about, but I've actually never heard what micro frontends actually are.

RICH: Yeah. I mean, to be clear, I don't really have a dog in this fight because I'm not building products, but the nub of it is that typically if you're building a website that maybe has like an admin page, maybe it has a a settings page, maybe it has product pages, whatever. Traditionally, these would all be parts of a single monolithic application. The micro frontend approach is to say, "Well, this team is going to own the settings page. This team is going to own the product page." And they can use whatever technologies they want to bring that about. And the detractors sort of attack a straw man version of this, "You're going to have different styles in every page. You're going to have to load Vue on one page. You're going to have to load React on the other page. It's going to be a terrible user experience," when actually its proponents aren't suggesting that at all. They're suggesting that people from these different teams coordinate a lot more that are free to deviate from some kind of grand master architectural plan when it's not suitable for a given task. And darn right. I think it means that you have a lot more agility as an engineering organization than you would if you're building this monolithic app where someone can't say, "Oh, we should use this new tool for this thing. We should use microstates when the rest of the organization is using Google docs." It's not possible. And so, you get locked into the decisions of a previous generation.

CHARLES: Right. No, it makes sense. It's funny because my first reaction is like, "Oh my goodness, that's a potential for disaster." The klaxon's going to go off in your head, but then you think, really then the work is how do you actually manage it so it doesn't become a disaster. And if you can figure that out, then yeah, there is a lot of potential.

RICH: Yeah. People always try and solve social problems with technology. You solve social problems with social solutions.

CHARLES: Right. And you have to imagine it too, it depends on the application, right? I think Amazon, the Amazon website is developed that way where they have different teams that are responsible even down to little content boxes that are up on the toolbar. And the site doesn't really, it shows, right? Like it shows like this is kind of like slapped together, but that's not what they need. They don't need it to not look like there's slight variation with the different ways that things behave. They need to be showing for their business to work. They need to be showing the right thing at the right time. And that's the overriding concern. So having it look very beautiful and very coherent isn't necessarily a thing. Same thing in Spotify, used as another example of this. I didn't know if it was called micro frontends, but I know that they've got a similar type thing, but they are clearly the experience and having it look coherent is more important. And so, they make it work somehow. And then like you're saying, it probably involves groups of people talking to other groups of people about the priorities.

So yeah, it doesn't sound to me like just like you're going to adopt micro frontends guarantees one particular set of outcomes. It really is context dependent on what you make of it.

RICH: Totally.

TARAS: I'm curious though, so with Svelte, essentially for your reactivity engine, you have to compile to get that reactive behavior.

RICH: Yeah.

TARAS: How does that play with other tools like when you actually integrate it together? I've never worked with Svelte on a large project, so I can't imagine what it looks like at scale. I was wondering if you've seen those kind of use cases and what that ends up, if there's any kind of side effects from that.

RICH: As you say, the reactivity within a component is only in the local state within that component or to state that is patched in as a prop from a parent component. But we also have this concept called a store. And a store is just a project that represents a specific value and you import it from svelte/store. And there are three types of store that you get out of the box. A writable, a readable and a derived. And a writeable is just, var count = writable (0) and then you can update that and you can set it using methods on that store. Inside your marker, you can reference or in fact inside the script block in the component, you can reference the value of that store just by prefacing it with a dollar sign. And the compiler sees that and says, "Okay, we need to subscribe to this store as value and then assign it and apply the reactivity." And that is the primary way of having state that exists outside the component hierarchy. Now, I mentioned the writable, readable, and derived are the built in stores that you get, but you can actually implement your own stores. You just need to implement this very simple contract. And so,, it's entirely possible to use that API to wrap any state management solution you have. So you can wrap redux, you can wrap microstates, you can wrap state, you can wrap whatever it is, whatever your preferred state management solution is, you can adapt it to use with Svelte. And it's very sort of idiomatic and streamlined. Like it takes care of unsubscriptions when the component is unmounted. All of that stuff is just done for you.

CHARLES: Digging a little bit deeper into the question of integration, how difficult would it be to take wholesale components that were implemented in Svelte and kind of integrate them with some other component framework like React?

RICH: If the component is a leaf node, then it's fairly straightforward. There is a project called react-svelte which is, I say project, it's like 20 lines of code and I don't think it's [inaudible] they did for Svelte 3, which I should probably do. But that allows you to use a Svelte component in the context of React application, just using the component API the same way that you would [inaudible] or whatever. You can do that inside a React component. Or you could compile the Svelte component to a web component. And this is one of the great benefits of being a compiler is that you can target different things. You can generate a regular JavaScript class and you've got an interactive application. Or you can target a server side rendering component which will just generate some html for some given state which can then later be hydrated on the client. Or you can target a web component which you can use like any other element in the context of any framework at all. And because it's a compiler, because it's discarding all of the bits of the framework that you're not using, it's not like you're bundling an entire framework to go along with your component. And I should mention while I'm talking about being able to target different outputs, we can also, as a NativeScript project, you can target iOS and Android that same way. Where it gets a little bit more complicated is if it's not a leaf node. If you want to have a React app that contains a Svelte component that has React [inaudible], then things start to get a little bit more unwieldy, I think. It's probably technically possible, but I don't know that I would recommend it. But the point is that it is definitely possible to incrementally adopt Svelte inside an existing application, should that be what you need to do.

CHARLES: You said there's a NativeScript project, but it sounds to me like you shouldn't necessarily need NativeScript, right? If you're a compiler, you can actually target Android and you could target iOS directly instead of having NativeScript as an intermediary, right?

RICH: Yes. If, if we had the time to do the work, then yes. I think the big thing there would be getting styles to work because Svelte components have styles. And a regular style tag just to CSS and you can't just throw CSS in a native app.

CHARLES: Right. Sometimes, I feel like it'd be a lot cooler if you could.


RICH: NativeScript really is doing a lot of heavy lifting. Basically what it's doing is it's providing a fake dom. And so, what the NativeScript does is it targets that dom instead of the real dom and then NativeScript turns that into the native instructions.

CHARLES: Okay. And you can do that because you're a compiler.

TARAS: Compilers has been on our radar for some time, but I'm curious like what is your process for figuring out what it should compile to? Like how do you arrive at the final compile output? Manually, have you written that code and then, "I'm going to now change this to be dynamically generated." Or like how do you figure out what the output should be?

RICH: That's pretty much it. Certainly, when the project started, it was a case of, I'm going to think like a compiler, I'm going to hand convert this declarative component code into some framework plus JavaScript. And then once that's done, sort of work backwards and figure out how a compiler would generate that code. And then the process, you do learn certain things about what the points of reusability are, which things should be abstracted out into a shared internal helper library and what things should be generated in line. The whole process is designed to produce output that is easy for a human to understand and reason about. It's not like what you would imagine compile [inaudible] to be like, it's not completely inscrutable. It's designed to be, even to that level of being well formatted, it's designed to be something that someone can look at and understand what the compiler was thinking at that moment. And there's definitely ways that we could change and improve it. There are some places where there's more duplication than we need to have. There are some places where we should be using classes instead of closures for performance and memory benefits. But these are all things that once you've got that base, having gone through that process, that you can begin to iterate on.

CHARLES: It's always curious to me about when is the proper time to move to a compiler, because when you're doing everything at runtime, there's more flexibility there. But at what point do you decide, "You know what? I know that these pathways are so well worn that I'm going to lay down pavement. And I'm going to write a compiler." What was the decision process in your mind about, "Okay, now it's time." Because I think that that's maybe not a thought that occurs to most of us. It's like, "I had to write a compiler for this." Is this something that people should do more often?

RICH: The [inaudible] of 'this should be a compiler' is one that is worth sort of having at the back of your head. I think there are a lot of opportunities not just in DUI framework space but in general, like is there some way that we can take this work that is currently happening at runtime and shift it into a step that only happens once. That obviously benefits users. And very often we find that benefits developers as well. I don't think there was a point at which I said, "Oh, this stuff that's happening at runtime should be happening at compile time." It was more, I mean, the actual origin has felt that it was a brain worm that someone else infected me with. Judgment is a very well known figure in the JavaScript world. He had been working on this exact idea but hadn't taken it to the point where he was ready to open source it. But he had shared like his findings and the general idea and I was just immediately smitten with this concept of getting rid of the framework runtime. At the time, the big conversation happening in the JavaScript community was about the fact that we're shipping too much JavaScript and it's affecting startup performance time. And so the initial thought was, "Well, maybe we can solve that problem by just not having the runtime." And so, that was the starting point with Svelte. Over time, I've come to realize that that is maybe not the main benefit. That is just one of the benefits that you get from this approach. You also get much faster update performance because you don't have to do this fairly expensive virtual dom different process. Lately, I've come to think that the biggest win from it is that you can write a lot less code. If you're a compiler, then you're not kind of hemmed in by the constraints of the language, so you can almost invent your own language. And if you can do that, then you can do the same things that you have been doing with an API in the language itself. And that's the basis of our system of reactivity, for example. We can build these apps that are smaller and by extension, less bug prone and more maintainable.

I just wanted to quickly address the point you made about flexibility. This is a theoretical downside of being a compiler. We're throwing away the constraints about the code needing to be something that runs in the browser, but we're adding a constraint, which is that the code needs to be statically analyzable. And in theory, that results in a loss of flexibility. In practice, we haven't found that to affect the things that we can build. And I think that a lot of times when people have this conversation, they're focusing on the sort of academic concepts of flexibility. But what matters is what can you build? How easy is it to build a certain thing? And so if empirically you find that you're not restricted in the things that you can build and you can build the same things much faster, then that academic notion of flexibility doesn't, to my mind, have any real value.

CHARLES: Hearing you talk reminded me of kind of a quote that I heard that always stuck with me back from early in my career. I came into programming through Perl. Perl was my first language and Perl is a very weird language. But among other things, you can actually just change the way that Perl parses code. You can write Perl that makes Perl not throw, if that makes any sense. And when asked about this feature, the guy, Larry Wall, who came up with Perl, he's like, "You program Perl, but really what you're doing is you're programming Perl with a set of semantics that you've negotiated with the compiler." And that was kind of a funny way of saying like, "You get to extend the compiler yourself." Here's like the default set of things that you can do with our compiler, but if you want to tweak it or add or modify, you can do that. And so, you can utilize the same functionality that makes it powerful in the first place. You can kind of inject that whole mode of operation into the entire workflow. Does that make sense? That's like a long way of saying, have you thought about, and is it possible to kind of extend the Svelte compiler as part of a customization or as part of the Svelte programming experience?

RICH: We have a very rudimentary version of that, which is pre-processing. There's an API that comes with Svelte called preprocess. And the idea there is that you can pass in some code and it will do some very basic, like it will extract your styles, it will extract your script and it will extract your markup. And then it will give you the opportunity to replace those things with something else. So for example, you could write some futuristic JavaScript and then compile it with Babel before it gets passed to the Svelte compiler, which uses acorn and therefore needs to be able to have managed other scripts so that it can construct an abstract syntax tree. A more extreme version of that, people can use [inaudible] to write their markup instead of html. You can use Sass and Less and things like that. Generally, I don't recommend that people do because it adds these moving parts and it makes like a lot of bug reports of people just trying to figure out how to get these different moving parts to operate together. I don't know, it means that your editor plugins can't understand what's inside your style tag all of a sudden and stuff like that. So, it definitely adds some complexity, but it is possible.

At the other end, at a slightly more extreme level, we have talked about making the cogeneration part plugable so that for example, the default renderer and the SSR renderer are just two examples of something that plugs into the compiler that says, "Here is the component, here's the abstract syntax tree, here's some metadata about which values are in scope," all of this stuff and then go away and generate some code from this. We haven't done that so far, partly because there hasn't been a great demand for it, but also because it's really complicated. As soon as you turn something into a plugin platform, you just magnify the number of connection points and the number of ways that things could go wrong by an order of magnitude. And so, we've been a little bit wary of doing that, but it is something that we've talked about primarily in the context of being able to do new and interesting things like target webGL directly or target the command line. There are renders for React that let you build command line apps using React components. And like we've talked about, maybe we should be able to do that. Native is another example. The NativeScript integration as you say, it could be replaced with the compiler doing that work directly, but for that to work presently, that would mean that all of that logic would need to sit in core. And it would be nice if that could be just another extension to the compiler. We're talking about a lot of engineering effort and there's higher priority items on our to do list at the moment. So, it's filed under one day.

CHARLES: Right. What are those high priority items?

RICH: The biggest thing I think at the moment is TypeScript integration. Surprisingly, this is probably like the number one feature request I think is that people want to be able to write Typescript inside the Svelte components and they want to be able to get TypeScript when they import the Svelte component into something else. They want to be able to get completion [inaudible] and type checking and all the rest of it. A couple of years ago, that would've been more or less than thinkable but now it's like table stakes is that you have to have first-class TypeScript support.

CHARLES: Yeah, TypeScript is as popular as Babel these days, right?

RICH: Yeah, I think so. I don't need to be sold on the benefits. I've been using TypeScript a lot myself. Svelte is written in TypeScript, but actually being able to write it inside your components is something that would involve as hacking around in the TypeScript compiler API in a way that, I don't know if anyone actually or any of us on the team actually knows how to do. So, we just need to spend some time and do that. But obviously when you've got an open source project, you need to deal with the bugs that arise and stuff first. So, it's difficult to find time to do a big project like that.

CHARLES: So, devil's advocate here is if the compiler was open for extension, couldn't a TypeScript support be just another plugin?

RICH: It could, but then you could end up with a situation where there's multiple competing TypeScript plugins and no one's sure which ones are used and they all have slightly different characteristics. I always think it's better if these things that are common feature requests that a lot of people would benefit from, if they're built into the project themselves. I go really light in the batteries included way of developing and I think this is something that we've sort of drifted away from in the frontend world over the last few years, we've drifted away from batteries included towards do it yourself.

CHARLES: Assemble the entire thing. Step one, open the box and pour the thousand Lego pieces onto the floor.

RICH: Yeah, but it's worse than that because at least, with a Lego set, you get the Lego pieces. It's like if you had the Lego manual showing you how to build something, but you were then responsible for going out and getting the Lego pieces, that's frontend development and I don't like it.

CHARLES: Right. Yeah. I don't like that either. But still, there's a lot of people advocating directly. You really ought to be doing everything completely and totally yourself.

RICH: Yes.

CHARLES: And a lot of software development shops still operate that way.

RICH: Yeah. I find that the people advocating for that position the most loudly, they tend to be the maintainers of the projects in question. The whole small modules philosophy, they exist for the benefit primarily of library authors and framework authors, not for the benefit of developers, much less users. And the fact that the people who are building libraries and frameworks tend to have the loudest megaphones means that that mindset, that philosophy is taken as a best practice for the industry as a whole. And I think it's a mistake to think that way.

TARAS: There is also, I think, a degree of a sliding scale where you start off with like as the more experience you get, because there is more experience you get closer, you get to that kind of wanting granular control and then they kind of slides down towards granular control and then slice back up to, once you've got a lot of experience, you're like, "Okay, I don't want this control anymore." And then you kind of cast that and you get into like, "I'm now responsible for tools that my team uses," and now you're back to wanting that control because you want things to be able to click together. It's kind of like a way that your interest in that might change over time depending on your experience level and your position in the organization. So yeah, there's definitely different motivating factors. Like one of the things that we've been thinking a lot about is designing tools that are composable and granular at individual module level, but combined together into a system for consumption by regular people. So like finding those primitives that will just click together when you know how to click them together. But when you're consuming them, just feel like a holistic whole, but at the same time not being monolithic. That's a lot of things to figure out and it's a lot of things to manage over time, but that's solely the kind of things we've been thinking about a lot.

RICH: I think that's what distinguishes the good projects that are going to have a long lifespan from the projects that are maybe interesting but don't have a long shelf life is whether they're designed in such a way that permits that kind of cohesion and innovation tradeoff, if you think of it as a trade off. Anyone can build the fastest thing or the smallest thing or the whatever it is thing. But building these things in a way that feels like it was designed holistically but is also flexible enough to be used with everything else that you use, that's the real design challenge.

CHARLES: It's hard to know where to draw that line. Maybe one good example of this and, these are actually two projects that I'm not particularly a fan of, but I think they do a good job of operating this way. So, I guess in that sense, it means I can even be more honest about it. I don't particularly care for Redux or like observables, but we ended up using, in one of our last React projects, we had to choose between using Redux-Saga and Redux-Observable. The Redux-Observable worked very well for us. And I think one of the reasons is because they both had to kind of exist. They had to kind of co-exist is their own projects. Like Redux exists as its own entity and Observables exist as their own kind of whole ecosystem. And so, they put a lot of thought in like what is the natural way in which these two primitives compose together? As opposed to the Saga, which I don't want to disparage the project because I think it actually is a really good project. There's a lot of really good ideas there but because it's more like just bolted on to Redux and it doesn't exist outside of the ecosystem of Redux and the ideas can't flourish outside and figure out how it interfaces with other things. Like the true primitive is still unrevealed there. And so, whereas I feel like with Redux you actually have to really, really true primitives. Now, they're not necessarily my favorite primitives, but they are very refined and very like these do exactly what they are meant to do. And so when you find how they connect together, that experience is also really good. And the primitive that arises there I think ends up being better. Is that an example of what you guys are talking about?

RICH: Maybe. [Laughs]

TARAS: No, I think so. I mean, it's distilling to the essence, the core of what you're trying to do and then be able to combine it together. I mean, that's been kind of the thing that we've been working on at the Frontside. But also within this context, it makes me think of how does a compiler fit into that? How does that work with the compiler? It's just like when you add the compiler element, it just makes it like my mind just goes poof!


CHARLES: Yeah, exactly. That's why I keep coming back to like, how do you, and maybe I haven't, you just have to kind of go through the experience, but it feels like maybe there's this cycle of like you build up the framework and then once it's well understood, you throw the framework away in favor of like just wiring it straight in there with the compiler and then you iterate on that process. Is that fair to say?

RICH: Kind of, yeah. At the moment, I'm working on this project, so I referred a moment ago to being able to target webGL directly. At the moment, the approach that I'm taking to building webGL apps is to have webGL components inside Svelte in this project called SvelteGL. And we've used it a couple of times at the Times. It's not really production ready yet, but I think it has some promise. But it's also slightly inefficient, like it needs to have all of the shade of code available for whichever path you're going to take, whatever characteristics your materials have, you need to have all of the shade of code. And if we're smart about it, then the compiler could know ahead of time which bits of shade of code it needed to include. At the moment, it just doesn't have a way of figuring that out. And so that would be an example of paving those cow paths. Like if you do try and do everything within the compiler universe, it does restrict your freedom of movement. It's true. And to qualify my earlier statements about how the small modules philosophy is to the benefit of authors over developers, it has actually enabled this huge flourishing of innovation, particularly in the React world. We've got this plethora of different state management solutions and CSS and JS solutions. And while I, as a developer, probably don't want to deal with that, I just want there to be a single correct answer. It's definitely been to the advantage of the ecosystem as a whole to have all of this experimentation. Then in the wild, there are projects like Svelte they can then take advantage of. We can say, "Oh well, having observed all of this, this is the right way to solve this problem." And so, we can kind of bake in that and take advantage of the research that other people have done. And I think we have made contributions of our own but there is a lot of stuff in Svelte like the fact that data generally flows one way instead of having [inaudible] everywhere. Things like that are the results of having seen everyone make mistakes in the past and learning from them. So, there are tradeoffs all around.

TARAS: One thing on topic of data flow here and there, one thing that I've been kind of struggling to compute is the impact of that as opposed to something where you have like one directional data flow because it seems like conceptually it's really simple. You set a property like in two way balance system, like you just propagate through stuff but we don't really have a way, you don't have any way of assessing what is the true impact of that computation. Like what is the cost of that propagation where I think it's almost easier to see the cost of that computation if you have like one directional data flow because you know that essentially everything between the moment that you invoke transition to computing the next state, that is the cost of your computation where you don't have that way of computing the result in a two way balance system. Something like Ember Run Loop or mobx or zones, Vues, reactive system. All these systems make it really difficult to understand what is the real cost of setting state. And that's something that I personally find difficult because this clarity that you have about the one directional data flow and what it takes to compute the next state, it's almost like because that cost is tangible where you're thinking about like mutation of objects and tracking their change like that cost is almost immeasurable. It just seems like a blob of changes that they have to propagate. I don't know. That's just something that I've been thinking a lot because especially with the work that we'll be doing with microstates because as you're figuring out what the next state is, you know exactly what operations are performed in a process where that might not be the case with the system that tracks changes like where you'd have with zones or with Ember Run Loop, or Vue.

RICH: I would agree with that. The times that I found it to be beneficial to deviate from the top-down ideology is when you have things like form elements and you want to bind to the values of those form elements. You want to use them in some other computation. And when you do all that by having props going in and then events going out and then you intercept the event and then you set the prop, you're basically articulating what the compiler can articulate for you more effectively anyway. And so conceptually, we have two way bindings within Svelte, but mechanically everything is top down, if that makes sense.

CHARLES: Is it because you can analyze the tree of top down and basically understanding when you can cheat. This might be really over-simplistic, but if you're kind of with the event, you're collecting the water and then you have to put it way up on top of the thing and it flows down. But if you can see the entire apparatus, you can say, "Actually, I've got this water and it's going to end up here, so I'm just going to cheat and put it over right there." Is that the type of thing that you're talking about where you're effectively getting a two way binding, but you're skipping the ceremony.

RICH: It's kind of writing the exact same code that you would write if you were doing it using events. But if you're writing it yourself, then maybe you would do something in a slightly inefficient way perhaps. For example, with some kinds of bindings, you have to be careful to avoid an infinite loop. If you have an event that triggers a state change, the state change could trigger the event again and you get this infinite loop. A compiler can guard against that. It can say this is a binding that could have that problem, so we're going to just keep track of whether the state changes as a result of the binding. And so, the compiler can sort of solve all of these really hairy problems that you had faced as a developer while also giving you the benefit in terms of being able to write much less code and write code that expresses the relationship between these two things in a more semantic and declarative way without the danger.

TARAS: This is one of the reasons why I was so excited to talk to you about this stuff, Rich, because this stuff is really interesting. I mentioned that we might, so we have a little bit more time. So I just want to mention, because I think that you might find this interesting, the [inaudible], the stuff that we were talking about that I mentioned to you before. So, I want to let Charles talk about it briefly because it's interesting, because it essentially comes down to managing asynchrony as it ties to life cycle of objects. Life cycle of objects and components are something we deal with on a regular basis. So, it's been an interesting exercise and experimenting with that. Charles, do you want to give kind of a low down?

CHARLES: Sure. It's definitely something that I'm very excited about. So, Taras gets to hear like an earful pretty much every day. But the idea behind structure concurrency, I don't know if you're familiar with it. It's something that I read a fantastic -- so people have been using this for a while in the Ember community. So Alex Matchneer, who's a friend and often time guest on the podcast created a library called ember-concurrency where he brought these ideas of structure concurrency to the ember world. But it's actually very prevalent. There's C libraries and Python libraries. There's not a generic one for JavaScript yet, but the idea is just really taking the same concepts of scope that you have with variables and with components, whether they be ember components, Svelte components, React components or whatever there is, you have a tree of components or you have a of parents and children and modeling every single asynchronous process as a tree rather than what we have now, which is kind of parallel linear stacks. You call some tick happens in the event loop and you drill down and you either edit an exception or you go straight back up. The next tick of the event loop comes, you drill down to some stack and then you go back up. A promise resolves, you do that stack. And so with structure concurrency, essentially every stack can have multiple children. And so, you can fork off multiple children. But if you have an error in any of these children, it's going to propagate up the entire tree. And so, it's essentially the same idea as components except to apply to concurrent processes. And you can do some just really, really amazing things because you don't ever have to worry about some process going rogue and you don't have to worry about coordinating all these different event loops. And one of the things that I'm discovering is that I don't need like event loops. I don't really use promises anymore. Like actually, I was watching, I think it was why I was watching your talk when you're talking about Svelte 3, when you're like -- or maybe did you write a blog post about we've got to stop saying that virtual doms are fast?

RICH: Yes, I did.

CHARLES: So I think it was that one. I was reading that one and it jived with me because it's just like, why can't we just go and do the work? We've got the event, we can just do the work. And one of the things that I'm discovering is with using the construction concurrency with generators, I'm experiencing a very similar phenomenon where these stack traces, like if there's an error, the stack traces like three lines long because you're basically doing the work and you're executing all these stacks and you're pausing them with a generator. And then when an event happens, you just resume right where you left off. There's no like, we've got this event, let's push it into this event queue that's waiting behind these three event loops. And then we're draining these queues one at a time. It's like, nope, the event happens. You can just resume right where you were. You're in the middle of a function call, in the middle of like [inaudible] block. You just go without any ceremony, without any fuss. You just go straight to where you were, and the stack and the context and all the variables and everything is there preserved exactly where you left it. So, it's really like you're just taking the book right off the shelf and going right to your bookmark and continuing along. Rather than when you've got things like the run loop in ember or the zones in angular where you have all these mechanics to reconstruct the context of where you were to make sure that you don't have some event listener. An event listeners created inside of a context and making sure that that context is either reconstructed or the event listener doesn't fire. All these problems just cease to exist when you take this approach. And so, if it's pertinent to this conversation, that was a surprising result for me was that if you're using essentially code routines to manage your concurrency, you don't need event loops, you don't need buffers, you don't need any of this other stuff. You just use the JavaScript call stack. And that's enough.

RICH: I'm not going to pretend to have fully understood everything you just said but it does sound interesting. It does have something not that dissimilar to ember's run loop because if you have two state changes right next to each other, X+=1, Y+=1, you want to have a single update resulting from those. So instead of instruments in the code such that your components are updated immediately after X+=1, it waits until the end of the event loop and then it will flush all of the pending changes simultaneously. So, what you're describing sounds quite wonderful and I hope to understand that better. You have also reminded me that Alex Matchneer implemented this idea in Svelte, it's called svelte-concurrency. And when he sent it to me, I was out in the woods somewhere and I couldn't take a look at it and it went on my mental to do list and you just brought it to the top of that to do list. So yeah, we have some common ground here, I think.

CHARLES: All right.

TARAS: This is a really, really fascinating conversation. Thank you, Rich, so much for joining us.

CHARLES: Thank you for listening. If you or someone you know has something to say about building user interfaces that simply must be heard, please get in touch with us. We can be found on Twitter at @thefrontside or over just plain old email at Thanks and see you next time.

Listen to our podcast:

Listen on Apple Podcasts