Peterbe.com

Next.js is a full front-end web framework. Vite is a build tool so they don't easily compare. But if you're building a single-page app ("SPA"), the difference isn't that big, especially if you bolt on a routing library which is something that Next.js has built in.

My SPA is a relatively straight forward one. It's a React app that uses wonderful Mantine UI framework. The app is CRM for real-estate agents that I've been hacking on with my wife. SEO is not a concern because you can't do anything until you've signed in. So server-side rendering is not a requirement. In that sense, it's like loading Gmail. Yes, users might want a speedy first load when they open it in a fresh new browser tab, but the static assets are most likely going to be heavily (browser) cached by the few users it has.

With that out of the way, let's skim through some of the differences.

Build times

Immediately, this is a tricky one to compare because Next.js has the ability to cache. You get that .next/cache/ directory which is black magic to me, but it clearly speeds things up. And it's incremental so the caching can help partially when only some of the code has changed.

Running, npm run build && npm run export a couple of times yields:

Next.js

Without no .next/cache/ directory

Total time to run npm run build && npm run export: 52 seconds

With the .next/cache/ left before each build

Total time to run npm run build && npm run export: 30 seconds

Vite

Total time to run npm run build: 12 seconds

A curious thing about Vite here is that its output contains a measurement of the time it took. But I ignored that and used /usr/bin/time -h ... instead. This gives me the total time.
I.e. the output of npm run build will say:

✓ built in 7.67s

...but it actually took 12.2 seconds with /usr/bin/time.

Build artifacts

Perhaps not very important because Next.js automatically code splits in its wonderfully clever way.

Next.js

❯ du -sh out
1.8M    out

❯ tree out | rg '\.js|\.css' | wc -l
      52

Vite

❯ du -sh dist
960K    dist

and

❯ tree dist/assets
dist/assets
├── index-1636ae43.css
└── index-d568dfbf.js

Again, it's probably unfair to compare at this point. Most of the weight of these static assets (particularly the .js files) is due to Mantine components being so heavy.

Routing

This isn't really a judgment in any way. More of a record how it differs in functionality.

Next.js

In my app, that I'm switching from Next.js to Vite + wouter, I use the old way of using Next.js which is to use a src/pages/* directory. For example, to make a route to the /account/settings page I first create:

// src/pages/account/settings.tsx

import { Settings } from "../../components/account/settings"

const Page = () => {
  return <Settings />
}
export default Page

I'm glad I built it this way in the first place. When I now port to Vite + wouter, I don't really have to touch that src/components/account/settings.tsx code because that component kinda assumes it's been invoked by some routing.

Vite + wouter

First I installed the router in the src/App.tsx. Abbreviated code:

// src/App.tsx

import { Routes } from "./routes"

export default function App() {
  const { myTheme, colorScheme, toggleColorScheme } = useMyTheme()
  return (
    <ColorSchemeProvider
      colorScheme={colorScheme}
      toggleColorScheme={toggleColorScheme}
    >
      <MantineProvider withGlobalStyles withNormalizeCSS theme={myTheme}>
        <Routes />
      </MantineProvider>
    </ColorSchemeProvider>
  )
}

By the way, the code for Next.js looks very similar in its src/pages/_app.tsx with all those contexts that Mantine make you wrap things in.

And here's the magic routing:

// src/routes.tsx

import { Router, Switch, Route } from "outer"

import { Home } from "./components/home"
import { Authenticate } from "./components/authenticate"
import { Settings } from "./components/account/settings"
import { Custom404 } from "./components/404"

export function Routes() {
  return (
    <Router>
      <Switch>
        <Route path="/signin" component={Authenticate} />
        <Route path="/account/settings" component={Settings} />
        {/* many more lines like this ... */}

        <Route path="/" component={Home} />

        <Route>
          <Custom404 />
        </Route>
      </Switch>
    </Router>
  )
}

Redirecting with router

This is a made-up example, but it demonstrates the pattern with wouter compared to Next.js

Next.js

const { push } = useRouter()

useEffect(() => {
  if (user) {
    push('/signedin')
  }
}, [user])

wouter

const [, setLocation] = useLocation()

useEffect(() => {
  if (user) {
    setLocation('/signedin')
  }
}, [user])

Linking

Next.js

import Link from 'next/link'

// ...

<Link href="/settings" passHref>
  <Anchor>Settings</Anchor>
</Link>

wouter

import { Link } from "wouter"

// ...

<Link href="/settings">
  <Anchor>Settings</Anchor>
</Link>

Getting a query string value

Next.js

import { useRouter } from "next/router"

// ...

const { query } = useRouter()

if (query.name) {
  const name = Array.isArray(query.name) ? query.name[0] : query.name
  // ...
}

wouter

import { useSearch } from "wouter/use-location"

// ...

const search = useSearch()
const searchParams = new URLSearchParams(search)

if (searchParams.get('name')) {
  const name = searchParams.get('name')
  // ...
}

Conclusion

The best thing about Next.js is its momentum. It gets lots of eyes on it. Lots of support opportunities and great chance of its libraries being maintained well into the future. Vite also has great momentum and adaptation. But wouter is less "common".

Comparing apples and oranges is often counter-productive if you don't take all constraints and angles into account and those are usually quite specific. In my case, I just want to build a single-page app. I don't want a Node server. In fact, my particular app is a Python backend that does all the API responses from a fetch in the JavaScript app. That Python app also serves the built static files, including the dist/index.html file. That's how my app can serve the app straight away if the current URL is something like /account/settings. A piece of Python code (more or less the only code that doesn't serve /api/* URLs) collapses all initial serving URLs to serve the dist/index.html file. It's a classic pattern and honestly feels a bit dated in 2023. But it works. And what's so great about all of this is that I have a multi-stage Dockerfile that first does the npm run build (and some COPY --from=frontend /home/node/app/dist ./server/out) and now I can "lump" together the API backend and the front-end code in just 1 server (which I host on Digital Ocean).

If you had to write a SPA in 2023 what would you use? In particular, if it has to be React. Remix is all about server-side rendering. Create-react-app is completely unsupported. Building it from scratch yourself rolling your own TypeScript + Eslint + Rollup/esbuild/Parcel/Webpack does not feel productive unless you have enough time and energy to really get it all right.

In terms of comparing the performance between Next.js and Vite + wouter, the time it takes to build the whole app is actually not that big a deal. It's a rare thing to do. It's something I do after a long coding/debugging session. What's more pressing is how npm run dev works.
With Vite, I type npm run dev and hit Enter. Faster than I can almost notice, after hitting Enter I see...

VITE v4.4.6  ready in 240 ms

  ➜  Local:   http://localhost:3000/
  ➜  Network: use --host to expose
  ➜  press h to show help

and I'm ready to open http://localhost:3000/ to play. With Next.js, after having typed npm run dev and Enter, there's this slight but annoying delay before it's ready.

tl;dr; fvh is marginally faster than unified and unified is a bit faster than plain.

When you send a full-text search query to Elasticsearch, you can specify (if and) how it should highlight, with HTML tags, highlights. E.g.

The correct way to index data into <mark>Elasticsearch</mark> with (Python) <mark>elasticsearch</mark>-dsl

Among other configuration options, you can pick one of 3 different highlighter algorithms:

1. unified (default)
2. plain
3. fvh

The last one, fvh, requires that you index more at index-time (in particular to add term_vector="with_positions_offsets" to the mapping). In a previous benchmark I did, the total document size on disk, as described by http://localhost:9200/_cat/indices?v grew by 38%.

I bombarded my local Elasticsearch 7.7 instance with thousands of queries collected from logs. Some single-word, some multi-word. The fields it highlights are things like title (~5-50 words) and body (~100-2,000 words).
Basically, I edited the search query by testing one at a time. For example:

search_query = search_query.highlight(
-   "title", fragment_size=120, number_of_fragments=1, type="unified"
+   "title", fragment_size=120, number_of_fragments=1, type="plain"
)

...etc.

After doing 1,000 searches 3 different times per each highlighter type option, and recording the times it took I recorded the following:

(milliseconds per query, lower is better)

UNIFIED:
  MEAN  18.1ms
  MEDIAN 19.0ms

PLAIN:
  MEAN  24.5ms
  MEDIAN 27.5ms

FVH:
  MEAN  16.1ms
  MEDIAN 17.6ms

Thin marginal win for fvh over unified.

Conclusion

Conclusion? Or should I say "Caveats" instead? There's a lot more to it than raw performance speed. In this benchmark, it takes ~20 milliseconds to search on 2 different indexes, each with a scoring function and indexes containing between 1,000 and 5,000 documents with hundreds of thousands of words. So it's pretty minor.

Each highlighter performs slightly differently too, so you'd have to study the outcome a bit more carefully to get a better feel for if it works the way you and your team prefer it to work.

If there's any conclusion, other than the boring usual "it depends on your setup and preferences", the performance difference is noticeable but not blowing you away. It makes sense that fvh is a bit faster because you've paid for it by indexing more upfront (the offsets) at the expense of memory.

I recently switch from Nextjs to Remix for my personal website. One thing I struggled with was to have it merge individual .css files into one. So I solved it with the Parcel CLI. This blog post demonstrates how.

The problem

Note, first of all, this talks about the global CSS. You can and should still employ CSS Modules or something equivalent for CSS that is tied directly to a React component.

But global CSS has its place and purpose. The problem is that there's no convenient way to bundle multiple little .css files into one which you can then nest into routes in Remix.

The way you inject CSS into a Remix page is like this:

import highlight from "~/styles/highlight.css";
import blogpost from "~/styles/blogpost.css";

...


export function links() {
  return [
    { rel: "stylesheet", href: highlight },
    { rel: "stylesheet", href: blogpost },
}

And for the record, suppose you have a nested route that needs those, and another one you do:

import banner from "~/styles/banner.css";
import { links as rootLinks } from "./_index";

...


export function links() {
  return [
    ...rootLinks().filter((x) => !x.extra),
    { rel: "stylesheet", href: banner },
  ];
}

This will nicely pick up those source .css files, minify them and produce in the final HTML SSR output:

<link rel="stylesheet" href="/build/_assets/highlight-KI4AX52K.css"/>
<link rel="stylesheet" href="/build/_assets/blogpost-75V4EYTP.css"/>

Nice. Http2 is famously good at parallel downloads. But even that has its physical limits. Especially if you have many little .css files that make up all the CSS you need. Now you have multiple files that can get stuck on the network. Yes, you might be able to update 1 and keep caching the others if their fingerprint don't change, but this is likely to be rare.

Parcel to the rescue

I solved it by using the Parcel CLI. In package.json I have:

"parcel:build": "parcel build --dist-dir app/styles/build app/*.css",

And in app/global.css I have this:

/* This is app/global.css */

@import "../node_modules/@picocss/pico/css/pico.css";
@import "./styles/globals.css";
@import "./styles/message.css";
@import "./styles/nav.css";
@import "./styles/comments.css";
@import "./styles/carbonads.css";
@import "./styles/carbonads-outer.css";
@import "./styles/modal-search.css";

That means, that Parcel will bundle all of these app/*.css files into 1 app/styles/build/global.css
Now, I can refer to that built on in the Remix app:

import global from "~/styles/build/global.css";

...

export function links() {
  return [
    { rel: "stylesheet", href: global },
  ]
}

Build vs. dev

Ok, so that explains how to bundle individual CSS files before you actually use the bundled CSS files. Remix doesn't care (a good thing).
At this point, we've modularized the problem. Now Parcel can do what it does best (CSS bundling (among other things it can do)) and Remix can do what it does (serving the .css files into the HTML).

But just like it's ergonomically pleasant to bundle CSS files like this, we still want it so that you don't have to manually run a separate step to build the bundle every time you edit an individual source .css file (e.g. app/styles/nav.css)

Here's how I solved that split up by Dev and Build

Build

  "scripts": {
-   "build": "remix build",
+   "build": "npm run parcel:build && remix build",
+   "parcel:build": "parcel build --dist-dir app/styles/build app/*.css",

Now, npm run build will do both things.

Dev

  "scripts": {
    "dev": "npm-run-all build --parallel \"dev:*\"",
    "dev:node": "cross-env NODE_ENV=development esrun --watch ./server.ts ",
    "dev:remix": "remix watch",
+   "dev:parcel": "parcel watch --dist-dir app/styles/build app/global.css",

In conclusion

I admit, I'm a CSS Modules fan-boy and it saddens me how much global CSS I have. One thing at a time, I guess. They both have their powers; global and modular CSS, but I'll admit that my own personal site still relies a bit too much on global CSS. At least, little goes to waste because Remix makes it relatively easy to pick exactly which files you need for individual routes.

tl;dr; Always pass timeZone:"UTC" when calling Date.toLocaleDateString

The surprise

In my browser's web console:

>>> new Date('2014-11-27T02:50:49Z').toLocaleDateString("en-us", {day: "numeric"})
"26"

On my server located in the same time zone:

Welcome to Node.js v16.13.0.
Type ".help" for more information.
> process.env.TZ
undefined
> new Date('2014-11-27T02:50:49Z').toLocaleDateString("en-us", {day: "numeric"})
'26'

Here on my laptop:

Welcome to Node.js v16.20.0.
Type ".help" for more information.
> process.env.TZ
undefined
> new Date('2014-11-27T02:50:49Z').toLocaleDateString("en-us", {day: "numeric"})
'27'

What! Despite $TZ not being set, it formats according to something else.

02:50 Zulu means, to me, in the US Eastern time zone, the day before.

Why this matters

I kept getting this production error from React that the SSR-rendered HTML differed from the client-side rendered HTML. Strangely, I could never reproduce this locally and the error doesn't say what's different. All the Stack Overflow suggestions and Google results speak of the most basic easy things to check. It's not unusual that this happens when dealing with dates because even though the database (PostgreSQL) stores the dates in full UTC, sometimes when data travels via app servers through JSON pipelines, date formatting can drop important bits.
But here, '2014-11-27T02:50:49Z' is specific.

What made this so incredibly hard to debug was that it worked on one page but not on the other even though the two had the same exact component code. I broke it apart thinking there was something nasty in the content of the Markdown-rendered HTML. No. The reason it only happened on some pages was that I had a function that looked like this:

export function formatDateBasic(date: string) {
  return new Date(date).toLocaleDateString("en-us", {
    year: "numeric",
    month: "long",
    day: "numeric",
  });
}

And, different pages listed, almost non-deterministic, with different dates for related content which was referred to along with their dates. So on one page, there might be a single date that formats differently in EDT (Eastern daylight-saving time) compared to UTC. For example, Apr 1 at 18:00 Zulu, is still Apr 1 in EDT.

The explanation

I'm sorry that I don't understand this better, but Node's implementation of Date.toLocaleDateString does more than depend on process.env.TZ. I think $TZ is just a way to gain control.

For example, start the node REPL like this:

On my Ubuntu 20.04 server:

$ TZ=utc node
Welcome to Node.js v16.20.0.
Type ".help" for more information.
> new Date('2014-11-27T02:50:49Z').toLocaleDateString("en-us", {day: "numeric"})
'27'

On my MacBook:

❯ TZ=utc node
Welcome to Node.js v16.13.0.
Type ".help" for more information.
> new Date('2014-11-27T02:50:49Z').toLocaleDateString("en-us", {day: "numeric"})
'27'

To find out what timezone your computer has:

On Ubuntu:

$ timedatectl
               Local time: Mon 2023-05-08 12:42:03 UTC
           Universal time: Mon 2023-05-08 12:42:03 UTC
                 RTC time: Mon 2023-05-08 12:42:04
                Time zone: Etc/UTC (UTC, +0000)
System clock synchronized: yes
              NTP service: active
          RTC in local TZ: no

On macOS:

❯ sudo systemsetup -gettimezone
Password:
Time Zone: America/New_York

The solution

Setting TZ is probably a good thing. That can get a bit tricky though. Your code needs to run consistently on your laptop, in GitHub Actions, on a VPS server, in an Edge cloud function, etc.

A better way is to force Date.toLocaleString to be fed a timezone. Now it's controlled at the highest level:

export function formatDateBasic(date: string) {
  return new Date(date).toLocaleDateString("en-us", {
    year: "numeric",
    month: "long",
    day: "numeric",
+   timeZone: "UTC"
  });
}

Now, it no longer depends on the OS it runs on.

On my Ubuntu server:

Welcome to Node.js v16.20.0.
Type ".help" for more information.
> new Date('2014-11-27T02:50:49Z').toLocaleDateString("en-us", {day: "numeric", timeZone: "UTC"})
'27'

On my macOS:

Welcome to Node.js v16.13.0.
Type ".help" for more information.
> new Date('2014-11-27T02:50:49Z').toLocaleDateString("en-us", {day: "numeric", timeZone: "UTC"})
'27'

Fun fact

I once made it unnecessarily weird for me in the debugging session, when I figured out about the timeZone option. What I ran was this:

Welcome to Node.js v16.13.0.
Type ".help" for more information.
> new Date('2014-11-27T02:50:49Z').toLocaleDateString("en-us", {day: "numeric", zimeZone: "UTC"})
'26'

I expected it to be '27' now but why did it revert?? Notice the typo? And Date.toLocaleDateString won't throw an error for passing in options it doesn't expect.

Suppose you have a GitHub Action workflow that does some computation and a possible outcome is that file comes into existence. How do you run a follow-up step based on whether a file was created?

tl;dr

- name: Is file created?
  if: ${{ hashFiles('test.txt') != '' }}
  run: echo "File exists"

The "wrong" way

Technically, there's no wrong way, but an alternative might be to rely on exit codes. This would work.

- name: Check if file was created
  run: |
      if [ -f test.txt ]; then
          echo "File exists"
          exit 1
      else
          echo "File does not exist"
      fi
- name: Did the last step fail?
  if: ${{ failure() }}
  run: echo "Last step failed, so file must have maybe been created"

The problem with this is that not only leaves a red ❌ in the workflow logs, but it could also lead to false positives. For example, if the step that might create a file is non-trivial, you don't want to lump the creation of the file with a possible bug in your code.

A use case

What I needed this for was a complex script that was executed to find broken links in a web app. If there were broken links, only then do I want to file a new issue about that. If the script failed for some reason, you want to know that and work on fixing whatever its bug might be. It looked like this:

  - name: Run broken link check
    run: |
      script/check-links.js broken_links.md

  - name: Create issue from file
    if: ${{ hashFiles('broken_links.md') != '' }}
    uses: peter-evans/create-issue-from-file@433e51abf769039ee20ba1293a088ca19d573b7f
    with:
      token: ${{ env.GITHUB_TOKEN }}
      title: More than one zero broken links found
      content-filepath: ./broken_links.md
      repository: ${{ env.REPORT_REPOSITORY }}
      labels: ${{ env.REPORT_LABEL }}

That script/check-links.js script is given an argument which is the name of the file to write to if it did indeed find any broken links. If there were any, it generates a snippet of Markdown about them which is the body of filed new issue.

Demo

To be confident this works, I created a dummy workflow in a test repo to test. It looks like this: .github/workflows/maybe-fail.yml

I implemented this at work recently and although it felt like a hack, I've come to like it and it's been very helpful to our many contributors.
As (Node) engineers, we know that you should keep your node_modules up-to-date by running npm install periodically or every time you git pull from the upstream. It could be that some package got upgraded last night since you git pulled last time.
But not everyone remembers to run npm install often enough. They might do git pull origin main && npm start and now the code that starts up depends on some latest version that was upgraded in package.json and package-lock.json.

How we solved it was that we added this script:

node script/cmp-files.js package-lock.json .installed.package-lock.json || npm install && cp package-lock.json .installed.package-lock.json

And it's hooked up as a script in package.json called prestart:

"scripts": {
  ...
  "prestart": "node script/cmp-files.js ...",
  ...
}

Now, every time you run npm start to start up the local development server, it will run that piece of bash. No more having to remember to run npm install after every git pull.

A note on performance

The npm install command is fast when all packages are already updated. You can see it with:

# First time
$ npm install

# Second time when nothing should happen
$ time npm install
...
2.53s user 0.37s system 134% cpu 2.166 total

So it only takes 2 seconds. Not bad.

$ time node script/cmp-files.js package-lock.json .installed.package-lock.json
...
0.08s user 0.03s system 100% cpu 0.110 total

But 0.08 seconds is better :)

The comparison script

The cmp-files.js script looks like this:

#!/usr/bin/env node

// Given N files. Exit 0 if they all exist and are identical in content.

import fs from 'fs'

import { program } from 'commander'

program.description('Compare N files').arguments('[files...]', '').parse(process.argv)

main(program.args)

function main(files) {
  if (files.length < 2) throw new Error('Must be at least 2 files')
  try {
    const contents = files.map((file) => fs.readFileSync(file, 'utf-8'))
    if (new Set(contents).size > 1) {
      process.exit(1)
    }
  } catch (error) {
    if (error.code === 'ENOENT') {
      process.exit(1)
    } else {
      throw error
    }
  }
}

The file .installed.package-lock.json file is added to the repo's .gitignore

Note; given how well this works for running before npm start we can probably add this to a post-checkout git hook too.