Template engine with streaming capability - part 1/2

First publication date:

Browsers can render HTML on the fly as chunks of text arrive. This is called HTML streaming, and it is not new. I recently read Chris Haynes' article explaining how you can stream HTML out of order (OOO) without JavaScript. This opens up a lot of possibilities and new patterns at low cost. However, I have not found much in the way of simple yet efficient template engine libraries with streaming support. Let's build one!

Streaming HTML

Conceptually, streaming HTML is simple: it consists in building a sequence of text strings. While some parts of the sequences are static, others are dynamic, derived from contextual data. There is a helpful Javascript built-in construct for implementing such sequences: tagged templates

Consider the following example:

const Greet = ({ name }) => html`<p>hello ${name}</p>`;

Greet({name: 'Lorenzofx'});

We can build up the following sequence:

['<p>hello ', 'Lorenzofox', '</p>'];

The first and the last item are static whereas the second is a dynamic one.

Tagged templates

To implement the html tagged template (or any tagged template), you have to define a function whose signature is:

function html(templateParts, ...values) {
    // code

templatesParts is an array of the static strings and values is the list of variables to interpolate. templateParts size is always equal to the size of values list, plus one. What html does or returns is then left to your imagination.

html function


We want our html function to generate a sequence with the following rules for the interpolated values:

The actual library has actually more extensive specifications, but they are not relevant to this article.


To build a sequence we will use a generator function(you probably know that I am a big fan of generators). We will do something close to TDD(Test Driven Development) to see the evolution of our implementation.

Interpolate literals

import {test} from 'zora';

const stringify = iterable => [...iterable].join('');

test('literals are intepolated as strings', ({eq}) => {
    eq(stringify(html`<p>${'hello'}</p>`), '<p>hello</p>');
    eq(stringify(html`<p>${42}</p>`), '<p>42</p>');
    eq(stringify(html`<p aria-hidden="${true}">blah</p>`), '<p aria-hidden="true">blah</p>');
    eq(stringify(html`<p>${undefined}</p>`), '<p>undefined</p>');

function *html(templateParts, ...values){
    const [first, ...rest] = templateParts;
    yield first;
    for (const [templatePart, value] of zip(rest, values)){
        yield String(value);
        yield templatePart;

First, we take out the first template part, so that we know that both arrays rest and values have exactly the same number of items. We can then zip them together to form pairs.

zip is a common function whose implementation looks like:

const zip = (a, b) => a.map((item, i) => [item, b[i]]);

We can then iterate over the pairs, yield each element one by one and ensure that the interpolated value is a string.

Escape literals

test('Strings are HTML escpaed when interpolated', ({eq}) => {
        html`<p attr="${"/><script></script>"}"></p>`), 
        '<p attr="/&gt;&lt;script&gt;&lt;/script&gt;"></p>');

const escapeMap = {
    '&': '&amp;',
    '<': '&lt;',
    '>': '&gt;',
    '"': '&quot;',
    "'": '&#39;',

const htmlEntities = /[&<>"']/g;
const escape = (value) => {
    if (/[&<>"']/.test(value)) {
        return value.replace(htmlEntities, (char) => escapeMap[char]);

    return value;

function *html(templateParts, ...values){
    const [first, ...rest] = templateParts;
    yield first;
    for (const [templatePart, value] of zip(rest, values)){
        yield escape(String(value));
        yield templatePart;

The escape function encodes a set of specific characters into their entity equivalent to make sure that a malicious string cannot be injected.

Compose templates

test(`templates can be composed together`, ({ eq }) => {
        html`<p>foo ${html`<span>${42}</span>`}</p>`),
        '<p>foo <span>42</span></p>',

function* html(templateParts, ...values) {
    const [first, ...rest] = templateParts;
    yield first;
    for (const [templatePart, value] of zip(rest, values)) {
        if (value?.[Symbol.iterator] && typeof value !== 'string') {
            yield* value;
        } else {
            yield escape(String(value));
        yield templatePart;

The main difference is that we now check the type of the interpolated value. Generators are iterables, so that they implement [Symbol.iterator]. This means that a template will check this condition. Strings are also iterables, but we don’t want to iterate over every character of the string: hence the second check. If both checks pass, we simply delegate the control to the iterable value using yield* operator.

This has the side effect of implementing the rules on arrays as well. At this point, an array can contain nested arrays, templates, or types that we can’t technically support. But we won’t go any deeper, because recursivity can be a bit tricky to implement with tagged templates. We leave that job to the upcoming render function

Thenable (Promises)

test(`html yield Promise like as they are`, ({ eq }) => {
    const promise = Promise.resolve(42);
    const thenable = {
        then() {
            return 42;
        ['<div>', promise, '</div><p>', thenable, '</p>'],

function* html(templateParts, ...values) {
    const [first, ...rest] = templateParts;
    yield first;
    for (const [templatePart, value] of zip(rest, values)) {
        if (value?.[Symbol.iterator] && typeof value !== 'string') {
            yield* value;
        } else if (value?.then) {
            yield value;
        } else {
            yield escape(String(value));
        yield templatePart;

Anything that can be awaited implement then. If this check passes we just yield the “Promise like” value to the render function: there is nothing we can do at this point, and the strategy for handling asynchronous structures may vary form renderer to renderer. One renderer could choose to pause the stream and wait for the Promise to resolve. Other could insert a placeholder and, later when the Promise has resolved, append some code to replace that placeholder with the actual value (this technique can be used for out of order streaming).

render function

Now that we have a sequence that has already been partially converted into bits of strings, we can convert that sequence into a proper stream. This will also be the time to handle late-arriving values (such as what promises resolve to), and eventually reject unsupported chunks.

There is a one-to-one relationship between a stream and an async generator. Again, we will use the latter to take advantage of recursivity, delegation, etc.

async function* _render(template) {
    for (const chunk of template) {
        if (typeof chunk === 'string') {
            yield chunk;
        } else if (chunk?.then) {
            yield* _render(await chunk);
        } else if (chunk?.[Symbol.iterator]) {
            yield* _render(chunk);
        }   else {
            throw new Error('Unsupported chunk');

export function render(template) {
    return ReadableStream.from(_render(template));

export async function renderAsString(template) {
    const buffer = [];
    for await (const chunk of render(template)){
    return buffer.join('');

Let’s focus on the _render generator (the other functions will just convert it into other data structures). It takes an iterable (like a template!) as a parameter and iterates over its sequence. By this point, all the literals should already have been converted to strings, and that is our first check.

If it hits a Promise, we are now in an asynchronous context and can wait for it to resolve. We can then recursively delegate to the generator.

In the third check, we convert nested arrays or templates(and arrays) that could have been resolved by a Promise. Finally, we throw an error for any other chunk type: this prevents, for example, to have arrays with number elements or any other invalid literal elements.

All in all, the code is concise and fairly easy to follow, but…

Performance evaluation

To evaluate the performance of our library, we will use a real-world application: we will build a blog like the one you are reading at the moment, using fastify for the backend, with different template engines. pug, the library I would have used by default; and ejs, which seems to be quite popular with 13M download every week, will be our baselines.

screen shot of the blog application
screenshot of the test application

Code base with our library

The interesting thing about our library is that we can combine templates together using all the scripting abilities of Javascript.

// layout.js
import {html} from 'tpl-stream'
import {Navigation} from './navigation.js'

export const Page = ({ title, content }) => html`
  <!doctype html>
      <meta charset="utf-8" />
      <meta name="viewport" content="width=device-width,initial-scale=1" />
      <link href="/public/main.css" rel="stylesheet" type="text/css" />
      <link rel="icon" href="./public/favicon.ico" />
      <header id="main-header">
        <img id="logo" alt="blog logo" src="/public/logo.webp" />
        ${Navigation({ currentPage: '/blog' })}
          Hi! I am Laurent and this is my dev blog. This is where I collect what
          I learn, what I experiment and what I find interesting.
      <footer><p>© Laurent RENARD. All Rights Reserved.</p></footer>

where Navigation is

// navigation.js
import {html} from 'tpl-stream'

const links = [
    { href: '/', name: 'Home' },
    { href: '/blog', name: 'Blog' },
    { href: '/about', name: 'About' },

const NavigationLink = ({ name, href, ...rest }) =>
    html`<li><a href="${href}" ${rest}}>${name}</a></li>`;

const Navigation = ({ currentPage = '/' }) =>
      ${links.map((linkDef) =>
            ['aria-current']: currentPage === linkDef.href ? 'page' : false,

You will have noticed in NavigationLink that this version of html can interpolate objects: each key-value pair becomes an attribute if the value is not equal to false (we left this feature out of this article to keep it more concise).

The content that will be rendered in the layout comes from:

import { html } from 'tpl-stream';

const BlogPost = ({
}) => html`
  <article class="post-preview">
    <p class="meta">
      Published by ${author} on <time>${formatDate(publicationDate)}</time>
    <a rel="bookmark" href="${permalink}">Read full article</a>

export const Blog = ({ posts }) =>
    <h2>Latest articles</h2>

So that the request handler looks like:

import {render} from 'tpl-stream';

return render(
        title: 'Blog',
        content: getPosts().then((posts) => Blog({ posts })),


The posts are loaded with the code below, to fake a database latency and make it more realistic:

const LATENCY = env.DB_LATENCY || 10;

export async function getPosts() {
  const latency = Math.round(Math.random() * LATENCY); // between 0 and 10ms of latency
  await setTimeout(latency);
  return [...postList];

Load test

We use autocannon to perform the load test and see how many requests per second the server can handle. We get the following results (median requests per second):

tpl-stream pug ejs
238 1632 670

This is pretty bad and gets worst the more complicated the page is. Nothing to worry about, we will fix the performance issues in the next article.


In this article we have learned what HTML streaming is. We have built a templating engine that supports streaming, with a straightforward implementation (less than 100 SLOC). The API gives us a lot of flexibility, and it is fairly easy and pleasant to work with. Unfortunately, our naive implementation performs poorly compared to the baselines made with popular libraries from the ecosystem. In the next post we will fix these performance issues, and hopefully learn a few things along the way.