Stay up to date
Subscribe to the newsletter to stay up to date with articles, courses and much more!
I respect your privacy. Unsubscribe at any time.
The web is made up of technologies that got their start over 25 years ago. HTTP, HTML, CSS, and JS were all first standardized in the mid-nineties (when I was 8 years old). Since then, the web evolved into a ubiquitous application platform. As the web has evolved, so too has the architecture for the development of these applications. There are many core architectures for building applications for the web these days. The most popular architecture employed by web developers today is the Single Page App (SPA), but we are transitioning to a new and improved architecture for building web applications.
<form> elements have been around from the very beginning. Links for a browser to get things from a server, and forms for a browser to send things to a server (and get things in return). With this two-way communication established as a part of the specification from the start, it has been possible to create powerful applications on the web forever.
Here are the major architectures (in chronological order of popular use):
Each architecture of web development has benefits and pain points. Eventually, the pain points became enough of a problem to motivate the move to the next architecture which came with its own trade-offs.
No matter how we build our applications, we’re almost always going to need code running on a server (notable exceptions includes games like Wordle which [used to] store game state in local storage). One of the things that distinguishes these architectures is where the code lives. Let’s explore each of these in turn and watch how the location of code changed over time. As we cover each architecture, we’ll consider specifically the following use cases of code:
There are, naturally, more parts of a web application than these bits, but these are the bits that move around the most and where we spend the bulk of our time as web devs. Depending on project scale and team structure we may work in all of these categories of code or we may work on only a part of one.
In the early days, this is the only architecture that worked at all on the web based on the capabilities of web browsers at the time.
With Multi-Page Apps, all of the code we write lives on the server. The UI Feedback code on the client is handled by the user’s browser.
Document Request: When the user enters a URL in the address bar, the browser sends a request to our server. Our routing logic will call a function to fetch data which communicates with the persistence code to retrieve the data. This data then gets used by the rendering logic to determine the HTML which will be sent as a response to the client. All the while, the browser is giving the user feedback with some kind of pending state (normally in the favicon position).
Mutation Request: When the user submits a form, the browser serializes the form into a request sent to our server. Our routing logic will call a function to mutate the data which communicates with the persistence code to make database updates. Then it will respond with a redirect so the browser triggers a GET request to get fresh UI (which will trigger the same thing that happened when the user entered the URL to begin with). Again, the browser will give the user feedback with pending UI.
Note: It’s important that successful mutations send a redirect response rather than just the new HTML. Otherwise you’ll have the POST request in your history stack and hitting the back button will trigger the POST request again (ever wonder why apps sometimes say “DON’T HIT THE BACK BUTTON!!” Yeah, that’s why. They should’ve responded with a redirect).
The mental model of MPAs is simple. We didn’t appreciate it back then. While there was some state and complicated flows handled primarily by cookies in the requests, for the most part everything happened within the time of a request/response cycle.
Where this architecture falls short:
It’s notable that the web platform is constantly improving with the upcoming page transitions API which makes MPAs a more viable option for more use cases. But for the majority of web applications, that’s still not enough. In any case, at the time this problem was far from the minds of standards committees and our users wanted more now!
Progressive Enhancement is the idea that our web applications should be functional and accessible to all web browsers and then leverage whatever extra capabilities the browser has to enhance the experience. The term was coined in 2003 by Nick Finck and Steve Champeon. Speaking of the capabilities of the browser…
XMLHttpRequest was initially developed by Microsoft’s Outlook Web Access team in 1998 but it wasn’t standardized until 2016 (can you believe that!?). Of course that never stopped browser vendors and web devs before. AJAX was popularized as a term in 2005 and a lot of people started making HTTP requests in the browser. Businesses were built on the idea that we don’t have to go back to the server for any more than little bits of data to update the UI in place. With that, we could build Progressively Enhanced Multi-Page Apps:
“Whoa!” you might be thinking, “wait a minute… where’d all this code come from?” So now we’ve not only taken responsibility of UI feedback from the browser, we also have Routing, Data fetching, Data mutation, and Rendering logic to the client in addition to what we already had on the server. “What gives?”
That said, depending on the level of enhancement we’re talking about, we may indeed have to write code in almost all of our categories, persistence being the exception (unless we want offline mode support which is really neat, but not an industry standard practice, so it’s not included in the chart).
In addition, we even had to add more code to the backend to support AJAX requests our client would make. So more on both sides of the network.
This is the era of jQuery, MooTools, etc.
<script> tags which will be used for the enhancement capabilities.
Client-side Navigation: When the user clicks an anchor element with an
We definitely solved the problems with MPAs by bringing along client-side code and taking the UI Feedback responsibility onto ourselves. We have much more control and can give users a more custom app-like feel.
Unfortunately, to give the users the best experience they’re looking for, we have to be responsible for routing, data fetching, mutations, and rendering logic. There are a few problems with this:
On a personal note, this is around the time I entered the web development world. I recall this time with a mix of longing nostalgia and shivery fright 🍝.
It didn’t take long before we realized we could remove the duplication problems if we just deleted the UI code from the backend. So that’s what we did:
You’ll notice this graphic is almost identical to the PEMPA one. The only difference is the Rendering logic is gone. Some of the routing code is gone as well because we no longer need to have routes for UI. All we’re left with is API routes. This is the era of Backbone, Knockout, Angular, Ember, React, Vue, Svelte, etc. This is the strategy used by most of the industry today.
Because the backend no longer has rendering logic, all document requests (the first request a user makes when they enter our URL) are served by a static file server (normally a CDN). In the early days of SPAs, that HTML document was almost always an effectively empty HTML file with a
<div id="root"></div> in the
<body> which would be used to “mount” the application. These days, however, frameworks allow us to pre-render as much of the page as we know at build-time using a technique known as “Static Site Generation” (SSG).
The other behaviors in this strategy are the same as they are with PEMPAs. Only now we mostly use
fetch instead of
What’s interesting is the only difference from PEMPAs in the architectural behaviors above is that the document request is worse! So why did we do this!?
By far the biggest pro here is the developer experience. That was the original driving force for the transition from PEMPAs to SPAs in the first place. Not having code duplication was an enormous benefit. We justified this change via various means (DX is an input to UX after all). Unfortunately improving DX is all SPAs really did for us.
I remember personally being convinced that SPA architecture helped with perceived performance because a CDN could respond with an HTML document faster than a server could generate one, but in real world scenarios that never seemed to make a difference (and this is even less true thanks to modern infrastructure). The sad reality is that SPAs still have all the same other issues as PEMPAs, albeit with more modern tools that make things much easier to deal with.
To make matters worse, SPAs also introduced several new problems:
image.png. This is not great and ultimately results in a much worse user experience. For static content we can avoid much of this, but there are a whole host of issues and limitations with that which the purveyors of SSG strategies are working on and are happy to sell us their vendor-specific solutions to.
Libraries have been created to help wrangle these issues and reduce their impact. This has been incredibly helpful, but some would call the churn fatiguing. This has become the de-facto standard way to build web apps since the mid-2010s. We’re well into the 2020s and there are some new ideas on the horizon.
MPAs have a simple mental model. SPAs have more powerful capabilities. Folks who have been through the MPA stage and are working in SPAs truly lament the simplicity we’ve lost in the last decade. This is particularly interesting if you consider the fact that the motivation behind the SPA architecture was primarily for improving the developer experience over PEMPAs. If we could somehow merge SPAs and MPAs into a single architecture to get the best of both, then hopefully we’ll have something that is both simple and more capable. That’s what Progressively Enhanced Single Page Apps are.
Remember we had one significant problem with PEMPAs as well: code duplication. PESPAs solve this problem by making the backend UI code and frontend UI code the exact same. By using a UI library capable of both rendering on the server and becoming interactive/handling updates on the client, then we don’t have code duplication issues.
You’ll notice there are small boxes for data fetching, mutation, and rendering. These bits are for enhancement. For example, pending states, optimistic UI, etc. don’t really have a place on the server, so we’re going to have some code that’s run only on the client. But even with that, with modern UI libraries the colocation we get makes it tenable.
Client-side Navigation: When the user clicks a link, we’ll prevent the default behavior. Our router will determine the data and UI needed for the new route and trigger data fetching for whatever data the next route needs and render the UI that’s rendered for that route.
Mutation Requests: Did you notice those two charts are the same? Yeah! That's not an accident! Mutations with PESPAs are done via form submissions. No more of this
fetch nonsense (however, imperative mutations are fine for progressive enhancement like redirecting to the login screen when the user’s session times out). When the user submits a form, we’ll prevent the default behavior. Our mutation code serializes the form and sends it as a request to the route associated to the
action of the form (which defaults to the current URL). The routing logic on the backend calls the action code which communicates with the persistence code to perform the update and sends back a successful response (for example: a tweet like) or redirect (for example: creating a new GitHub repo). If it’s a redirect, the router loads code/data/assets for that route (in parallel) and then triggers the rendering logic. If it’s not a redirect, the router revalidates the data for the current UI and triggers the rendering logic to update the UI. Interestingly though, regardless of whether it’s an inline mutation or a redirect, the router is involved, giving us the same mental model for both types of mutations.
PESPAs eliminate a ton of problems from previous architectures. Let’s look at them one by one:
What distinguishes a PESPA:
As for the cons. We’re still discovering what those are. But here are some thoughts and initial reactions:
Many who are used to SPAs and SSG will lament that we now have server-side code running our app. However for any real-world app we can’t avoid server-side code. There are certainly some use cases where we can build the entire site once and stick it on a CDN, but most apps we work on for our day jobs don’t fit into this category.
Related to this is people are concerned about server cost. The idea is that SSG allows us to build our app once and then serve it via a CDN to an almost infinite number of users at very low cost. There are two flaws with this criticism. 1) We’re probably hitting APIs in our app, so those users will still be triggering plenty of our most expensive server-side code on their visits anyway. 2) CDNs support HTTP caching mechanisms, so if we’re really able to use SSG, then we can definitely make use of that to give both fast responses and limit the amount of work our rendering server is dealing with.
Another common issue people have with leaving SPAs is that now we have to deal with the challenges of rendering on the server. This is definitely a different model for folks used to running their code only on the client, but if we’re using tools that have taken this into consideration, it’s hardly a challenge. If we’re not, then it can definitely be a challenge, but there are reasonable workarounds to force certain code to only run client-side while we migrate.
As I said, we’re still discovering the cons of Progressively Enhanced Single Page Apps, but I think the benefits are worth the trade-offs we can perceive so far.
I should also mention that even though we've had the capabilities of a PESPA architecture for quite some time with existing tools, the focus on Progressive Enhancement while also sharing rendering logic code is new. This post is primarily interested in demonstrating the de-facto standard architectures, not just the capabilities of the platform.
Leading the charge for PESPAs is Remix, a web framework with a laser focus on web fundamentals and modern user experience. Remix is the first web framework to out of the box offer everything that I described a PESPA offering. Other frameworks can and are adapting to follow Remix’s lead on this. I’m specifically aware of both SvelteKit and SolidStart working in PESPA principles into their implementations. I imagine more will follow (again, meta-frameworks have been capable of PESPA architecture for quite some time, however Remix has put this architecture in the forefront and others are following suite). Here’s how things look when we’ve got a web framework for our PESPA:
In this case, Remix acts as a bridge across the network. Without Remix, we’d have to implement this ourselves to have a complete PESPA. Remix also handles our routing via a combination of convention-based and config based routing. Remix will also help with the progressively enhanced bits of our data fetching and mutations (like the twitter like button) and the UI feedback for implementing things like pending states and optimistic UI.
Thanks to the nested routing built into Remix, we get better code organization as well (something Next.js is pursuing as well). While nested routing isn’t required for the PESPA architecture specifically, route-based code splitting is an important part. Also, we get much more granular code splitting with nested routing so it is an important aspect.
Remix is demonstrating that we can have more fun building better experiences faster with the PESPA architecture. And we end up with situations like this one:
A perfect perf lighthouse score without trying? Sign me up!
Personally, I’m super here for this transition. Getting a better UX and DX at the same time is a solid win. I think it’s an important one and I’m excited about what the future holds for us. As a reward to you for finishing this blog post, I’ve made a repository that demonstrates all this code moving around through the ages using a TodoMVC app! Find it here: kentcdodds/the-webs-next-transformation. Hopefully it helps make some of the ideas more concrete.
And this is what I'm excited to teach you here on EpicWeb.dev. If you'd like to keep up with my progress here, stick your email in the form below. Let's make the web better 🎉
Stay up to date
Subscribe to the newsletter to stay up to date with articles, courses and much more!
I respect your privacy. Unsubscribe at any time.