DEV Community

Cover image for The Junior Developer's Complete Guide to SSR, SSG and SPA
JulesWritesCode
JulesWritesCode

Posted on

The Junior Developer's Complete Guide to SSR, SSG and SPA

Every dev tools company and -team seems to assume that Junior Devs are familiar with these terms.

When I started to code, I saw them everywhere: Nuxt is an SSR framework, you can use Gatsby for SSG, and you can enable SPA mode if you set this or that flag in your next.config.js.

What the hell?

As a first step, here's a glossary – though it won't help you to understand the details:

  • CSR = Client-Side Rendering
  • SPA = Single Page Application
  • SSR = Server-Side Rendering
  • SSG = Static Site Generation

Next, let's shed some light into the dark.

Static Web Servers

Initially, a website was an HTML file you requested from a server.

Your browser would ask the server, "Hey, can you hand me that /about page?" and the server would respond with an about.html file. Your browser knew how to parse said file and rendered a beautiful website such as this one.

We call such a server a static web server. A developer wrote HTML and CSS (and maybe a bit of JS) by hand, saved it as a file, placed it into a folder, and the server delivered it upon request. There was no user-specific content, only general, static (unchanging) content accessible to everybody.

app.get('/about', async (_, res) => {
  const file = fs.readFileSync('./about.html').toString();
  res.set('Content-Type', 'text/html');
  res.status(200).send(file);
})
Enter fullscreen mode Exit fullscreen mode

Interactive Web Apps & Request-Specific Content

Static websites are, however, boring.

It's much more fun for a user if she can interact with the website. So developers made it possible: With a touch of JS, she could click on buttons, expand navigation bars, or filter her search results. The web became interactive.

This also meant that the /search-results.html page would contain different elements depending on what the user sent as search parameters.

So, the user would type into the search bar, hit Enter, and send a request with his search parameters to the server. Next, the server would grab the search results from a database, convert them into valid HTML, and create a complete /search-results.html file. The user received the resulting file as a response.

(To simplify creating request-specific HTML, developers invented HTML templating languages, such as Handlebars.)

app.get('/search-results', async (req, res) => {
  const searchParams = req.query.q;
  const results = await search(searchParams);

  let htmlList = '<ul>';
  for (const result of results) {
    htmlList += `<li>${result.title}</li>`;
  }
  htmlList += '</ul>';

  const template = fs.readFileSync('./search-results.html').toString();
  const fullPage = embedIntoTemplate(htmlList, template);

  res.set('Content-Type', 'text/html');
  res.status(200).send(fullPage);
});
Enter fullscreen mode Exit fullscreen mode

A short detour about "rendering"

For the longest time, I found the term rendering highly confusing.

In its original meaning, rendering describes the computer creating a human-processable image. In video games, for example, rendering refers to the process of creating, say, 60 images per second, which the user could consume as an engaging 3D-experience. I wondered, already having heard about Server Side Rendering, how that could work — how could the server render images for the user to see?

But it turned out, and I realized this a bit too late, that "rendering" in the context of Server- or Client-Side Rendering means a different thing.

In the context of the browser, "rendering" keeps its original meaning. The browser does render an image for the user to see (the website). To do so, it needs a blueprint of what the final result should look like. This blueprint comes in the form of HTML and CSS files. The browser will interpret those files and derive from it a model representation, the Document Object Model (DOM), which it can then render and manipulate.

Let's map this to buildings and architecture so we can understand it a bit better: There's a blueprint of a house (HTML & CSS), the architect turns it into a small-scale physical model on his desk (the DOM) so that he can manipulate it, and when everybody agrees on the result, construction workers look at the model and "render" it into an actual building (the image the user sees).

When we talk about "rendering" in the context of the Server, however, we talk about creating, as opposed to parsing, HTML and CSS files. This is done first so the browser can receive the files to interpret.

Moving on to Client-Side Rendering, when we talk about "rendering", we mean manipulating the DOM (the model that the browser creates by interpreting the HTML & CSS files). The browser then converts the DOM into a human-visible image.

Client-Side Rendering & Single Page Applications (SPAs)

With the rise of platforms like Facebook, developers needed more and faster interactivity.

Processing a button-click in an interactive web app took time — the HTML file had to be created, it had to be sent over the network, and the user's browser had to render it.

All that hassle while the browser could already manipulate the website without requesting anything from the server. It just needed the proper instructions — in the form of JavaScript.

So that's where devs placed their chips.

Large JavaScript files were written and sent to the users. If the user clicked on a button, the browser would insert an HTML component; if the user clicked a "show more" button below a post, the text would be expanded — without fetching anything.

<!DOCTYPE html>
<html lang="en">
  <head>
    <meta charset="UTF-8" />
    <meta name="viewport" content="width=device-width, initial-scale=1.0" />
    <title>Document</title>
  </head>
  <body>
    <div id="root"></div>
    <script>
      document.addEventListener('DOMContentLoaded', () => {
        const root = document.getElementById('root');

        root.innerHTML = `
            <h1>Home</h1>
            <button>About</button>
        `;

        const btn = document.querySelector('button');
        btn.addEventListener('click', () => {
          root.innerHTML = `
            <h1>About</h1>
        `;
        });
      });
    </script>
  </body>
</html>

Enter fullscreen mode Exit fullscreen mode

Though the code snippet suggests the opposite, developers didn't write vanilla JavaScript.

Ginormous web apps like Facebook had so much interactivity and duplicate components (such as the infamous Like-button) that writing plain JS became cumbersome. Developers needed tools that made it simpler to deal with all the rendering, so around 2010, frameworks like Ember.js, Backbone.js, and Angular.js were born.

Of them, Angular.js was the one that brought Single Page Applications (SPAs) into the mainstream.

An SPA is the same as Client-Side Rendering, but it is taken a step further. The conventional page navigation, where a click on a link would fetch and render another HTML document, was taken over by JavaScript. A click on a link would now fire a JS function that replaced the page's contents with other, already preloaded content.

For this to work properly, devs needed to bypass existing browser mechanisms.

For example, if you click on a <button type="submit"> that's wrapped in a <form> element, the browser normally sends a POST request to the server and renders the HTML file of the response (an agreed upon standard all browsers conform to). But, since the server had already sent all the instructions, and the browser had now taken control over rendering pages, that behavior needed to be prevented.

Devs invented all kinds of hacks to bypass this and other mechanisms, but discussing those hacks is outside the scope of this post.

Server-Side Rendering

So what were the issues with that approach?

SEO and Performance.

First, if you look closely at the above HTML file, you'll barely see any content in the <body> tags (except for the script). The content was stored in JS and only rendered once the browser executed the <script>. Hence, Google's robots had a hard time guessing what the page's content was about — in fact, they couldn't guess anything.

The site couldn't be indexed and thus wouldn't rank highly on Google.

Second, since the browser would only send a single request to the server and then continue to render the SPA on its own, all content that could ever be rendered had to be delivered with the initial request. With large web apps, this could easily surpass a couple megabytes, which slowed down the page load significantly. Amazon conducted a study that concluded businesses would lose 1% of revenue with every 100ms in added page load time, so that was a huge no-no for most companies.

In short, developers needed to create HTML files on the server again.

But they couldn't circle back to templating-languages and request-specific content — by now, everybody was writing React, and they loved the component-driven approach. Wasn't there a way to write React and render it on the server? Of course! With the advent of Node.js, developers were already writing JS on the backend, and the road to full-stack frameworks such as Next.js or Nuxt was paved.

Those frameworks render React (or Vue, Svelte...) on the server. Basically, React turned into a templating language such as Handlebars.

// List.tsx
import React from "react"

export const List = (props: {
  results: { title: string }[]
}) => (
  <ul>
    {props.results.map((r) => (
      <ListElement title={r.title} />
    ))}
  </ul>
)

export const ListElement = (props: { title: string }) => (
  <li>{props.title}</li>
)

// server.ts
app.get("/search-results", async (req, res) => {
  const searchParams = req.query.q
  const results = await search(searchParams)

  const fullPage = renderToString(
    React.createElement(List, { results })
  )

  res.set("Content-Type", "text/html")
  res.status(200).send(fullPage)
})


Enter fullscreen mode Exit fullscreen mode

(Note: the only difference to the "Request-Specific Content" approach is that we're now using modern frontend frameworks to write our HTML.)

Static Site Generation

Great, so our pages were indexable and fast again.

However, some sites weren't as fast as they could be.

If the content didn't change (like the page of a blog post), why should the server newly fetch and build the HTML for every user visiting the page? That seemed wasteful. Wouldn't it be enough to build the HTML once and reuse it whenever somebody requested it?

That approach is called static site generation.

In contrast to the original static web server approach, where developers manually wrote and stored HTML files on the server, here, the HTML files were generated:

Non-technical people wrote content and uploaded images in a CMS (Content Management System, such as WordPress or Sanity). Developers wrote React components, used the CMS's API, and executed a build script that fetched the data for every page and built the HTML file according to the outlined blueprint. The finished file was then stored on the server, ready to be delivered upon a user's request.

Developers could then re-trigger the build script to create new files when new content became available.

The new kid on the block: Server Components

As of August 2024, React Server Components (RSCs) – though still marked as experimental – are all the rage.

The basic idea is this: Before RSCs, a React Component first had to be delivered to the client. There, it would render and, if it needed to fetch data, send another request to the server; it had to wait for the response and re-render. This is wasteful. RSCs make it possible to both fetch the data and render the component on the server. Only the finished, isolated HTML component is sent to the client, and there merged with the existing HTML – the rest of the page is left as is.

The result is a much better performance: less back-and-forth between the server, fewer kilobytes over the network, and less re-rendering.

Plus, you can await your data right in your React Code, making it much simpler to write.

But, as usual when new features are added to existing frameworks, RSCs make React much more complex. You must constantly consider whether you need a component on the server or the client; caching layers are involved; interactivity is impossible with RSCs – tl;dr, the waters are still muddy.

As a counter-movement, developers who are fed up with the front-end complexity have started to write and favor simple frameworks, such as HTMX.

Top comments (0)