# 🔗 linkedom [![Downloads](https://img.shields.io/npm/dm/linkedom.svg)](https://www.npmjs.com/package/linkedom) [![Build Status](https://travis-ci.com/WebReflection/linkedom.svg?branch=main)](https://travis-ci.com/WebReflection/linkedom) [![Coverage Status](https://coveralls.io/repos/github/WebReflection/linkedom/badge.svg?branch=main)](https://coveralls.io/github/WebReflection/linkedom?branch=main) **Social Media Photo by [JJ Ying](https://unsplash.com/@jjying) on [Unsplash](https://unsplash.com/)** ### This is not a crawler! LinkeDOM is a [triple-linked list](#data-structure) based DOM-like namespace, for DOM-less environments, with the following goals: * **avoid** maximum callstack/recursion or **crashes**, even under heaviest conditions. * guarantee **linear performance** from small to big documents. * be **close to the** current **DOM standard**, but [not too close](https://github.com/WebReflection/linkedom#faq). ```js import {DOMParser, parseHTML} from 'linkedom'; // Standard way: text/html, text/xml, image/svg+xml, etc... // const document = (new DOMParser).parseFromString(html, 'text/html'); // Simplified way for HTML const { // note, these are *not* globals window, document, customElements, HTMLElement, Event, CustomEvent // other exports .. } = parseHTML(` Hello SSR
`); // builtin extends compatible too 👍 customElements.define('custom-element', class extends HTMLElement { connectedCallback() { console.log('it works 🥳'); } }); document.body.appendChild( document.createElement('custom-element') ); document.toString(); // the SSR ready document document.querySelectorAll('form, input[name], button'); // the NodeList of elements // CSS Selector via CSSselect ``` ### What's New * in `v0.11` a new `linkedom/worker` export has been added. This works with [deno](https://deno.land/), Web, and Service Workers, and it's not strictly coupled with NodeJS. Please note, this export does not include `canvas` module, and the `performance` is retrieved from the `globalThis` context. ### Serializing as JSON *LinkeDOM* uses a blazing fast [JSDON serializer](https://github.com/WebReflection/jsdon#readme), and nodes, as well as whole documents, can be retrieved back via `parseJSON(value)`. ```js // any node can be serialized const array = toJSON(document); // somewhere else ... import {parseJSON} from 'linkedom'; const document = parseJSON(array); ``` Please note that *Custom Elements* won't be upgraded, unless the resulting nodes are imported via `document.importNode(nodeOrFragment, true)`. Alternatively, `JSDON.fromJSON(array, document)` is able to initialize right away *Custom Elements* associated with the passed `document`. ### Simulating JSDOM Bootstrap This module is based on [DOMParser](https://developer.mozilla.org/en-US/docs/Web/API/DOMParser) API, hence it creates a *new* `document` each time `new DOMParser().parseFromString(...)` is invoked. As there's *no global pollution* whatsoever, to retrieve classes and features associated to the `document` returned by `parseFromString`, you need to access its `defaultView` property, which is a special proxy that lets you get *pseudo-global-but-not-global* properties and classes. Alternatively, you can use the `parseHTML` utility which returns a pseudo *window* object with all the public references you need. ```js // facade to a generic JSDOM bootstrap import {parseHTML} from 'linkedom'; function JSDOM(html) { return parseHTML(html); } // now you can do the same as you would with JSDOM const {document, window} = new JSDOM('

Hello LinkeDOM 👋

'); ``` ## Data Structure The triple-linked list data structure is explained below in [How does it work?](#how-does-it-work), the [Deep Dive](./deep-dive.md), and the [presentation on Speakeasy JS](https://www.youtube.com/watch?v=PEESaD7Qkxs). ## F.A.Q.
Why "not too close"?
*LinkeDOM* has zero intention to: * implement all things *JSDOM* already implemented. If you need a library which goal is to be 100% standard compliant, please [use JSDOM](https://github.com/jsdom/jsdom) because *LinkeDOM* doesn't want to be neirly as bloated nor as slow as *JSDOM* is * implement features not interesting for *Server Side Rendering*. If you need to pretend your NodeJS, Worker, or any other environment, is a browser, please [use JSDOM](https://github.com/jsdom/jsdom) * other points listed, or not, in the followung *F.A.Q.s*: this project will always prefer the minimal/fast approach over 100% compliant behavior. Again, if you are looking for 100% compliant behavior and you are not willing to have any compromise in the DOM, this is **not** the project you are looking for That's it, the rule of thumb is: do I want to be able to render anything, and as fast as possible, in a DOM-less env? *LinkeDOM* is great! Do I need a 100% spec compliant env that simulate a browser? I rather use *cypress* or *JSDOM* then, as *LinkeDOM* is not meant to be a replacement for neither projects.
Are live collections supported?
The *TL;DR* answer is **no**. Live collections are considered legacy, are slower, have side effects, and it's not intention of *LinkeDOM* to support these, including: * `getElementsByTagName` does not update when nodes are added or removed * `getElementsByClassName` does not update when nodes are added or removed * `childNodes`, if trapped once, does not update when nodes are added or removed * `children`, if trapped once, does not update when nodes are added or removed * `attributes`, if trapped once, does not update when attributes are added or removed * `document.all`, if trapped once, does not update when attributes are added or removed If any code you are dealing with does something like this: ```js const {children} = element; while (children.length) target.appendChild(children[0]); ``` it will cause an infinite loop, as the `children` reference won't side-effect when nodes are moved. You can solve this in various ways though: ```js // the modern approach (suggested) target.append(...element.children); // the check for firstElement/Child approach (good enough) while (element.firstChild) target.appendChild(element.firstChild); // the convert to array approach (slow but OK) const list = [].slice.call(element.children); while (list.length) target.appendChild(list.shift()); // the zero trap approach (inefficient) while (element.childNodes.length) target.appendChild(element.childNodes[0]); ```
Are childNodes and children always same?
**Nope**, these are discovered each time, so when heavy usage of these *lists* is needed, but no mutation is meant, just trap these once and use these like a frozen array. ```js function eachChildNode({childNodes}, callback) { for (const child of childNodes) { callback(child); if (child.nodeType === child.ELEMENT_NODE) eachChildNode(child, callback); } } eachChildNode(document, console.log); ```
## How does it work? All nodes are linked on both sides, and all elements consist of 2 nodes, also linked in between. Attributes are always at the beginning of an element, while zero or more extra nodes can be found before the end. A fragment is a special element without boundaries, or parent node. ``` Node: ← node → Attr: ← attr → ↑ ownerElement? Text: ← text → ↑ parentNode? Comment: ← comment → ↑ parentNode? Element: ← start ↔ end → ↑ parentNode? Fragment: start ↔ end Element example: parentNode? (as shortcut for a linked list of previous nodes) ↑ ├────────────────────────────────────────────┐ │ ↓ node? ← start → attr* → text* → comment* → element* → end → node? ↑ │ └────────────────────────────────────────────┘ Fragment example: ┌────────────────────────────────────────────┐ │ ↓ start → attr* → text* → comment* → element* → end ↑ │ └────────────────────────────────────────────┘ ``` If this is not clear, feel free to **[read more in the deep dive page](./deep-dive.md)**. ### Why is this better? Moving *N* nodes from a container, being it either an *Element* or a *Fragment*, requires the following steps: * update the first *left* link of the moved segment * update the last *right* link of the moved segment * connect the *left* side, if any, of the moved node at the beginning of the segment, with the *right* side, if any, of the node at the end of such segment * update the *parentNode* of the segment to either *null*, or the new *parentNode* As result, there are no array operations, and no memory operations, and everything is kept in sync by updating a few properties, so that removing `3714` sparse `
` elements in a *12M* document, as example, takes as little as *3ms*, while appending a whole fragment takes close to *0ms*. Try `npm run benchmark:html` to see it yourself. This structure also allows programs to avoid issues such as "*Maximum call stack size exceeded*" (basicHTML), or "*JavaScript heap out of memory*" crashes (JSDOM), thanks to its reduced usage of memory and zero stacks involved, hence scaling better from small to very big documents. ### Are *childNodes* and *children* always computed? As everything is a `while(...)` loop away, by default this module does not cache anything, specially because caching requires state invalidation for each container, returned queries, and so on. However, you can import `linkedom/cached` instead, as long as you [understand its constraints](https://github.com/WebReflection/linkedom#cached-vs-not-cached). ## Parsing VS Node Types This module parses, and works, only with the following `nodeType`: * `ELEMENT_NODE` * `ATTRIBUTE_NODE` * `TEXT_NODE` * `COMMENT_NODE` * `DOCUMENT_NODE` * `DOCUMENT_FRAGMENT_NODE` * `DOCUMENT_TYPE_NODE` Everything else, at least for the time being, is considered *YAGNI*, and it won't likely ever land in this project, as there's no goal to replicate deprecated features of this aged Web. ## Cached VS Not Cached This module exports both `linkedom` and `linkedom/cached`, which are basically the exact same thing, except the cached version outperforms `linkedom` in these scenarios: * the document, or any of its elements, are rarely changed, as opposite of frequently mutated or manipulated * the use-case needs many repeated *CSS* selectors, over a sporadically mutated "*tree*" * the generic DOM mutation time is *not* a concern (each, removal or change requires a whole document cache invalidation) * the *RAM* is *not* a concern (all cached results are held into *NodeList* arrays until changes happen) On the other hand, the basic, *non-cached*, module, grants the following: * minimal amount of *RAM* needed, given any task to perform, as nothing is ever retained on *RAM* * linear fast performance for any *every-time-new* structure, such as those created via `importNode` or `cloneNode` (i.e. template literals based libraries) * much faster DOM manipulation, without side effect caused by cache invalidation ## Benchmarks To run the benchmark locally, please follow these commands: ```sh git clone https://github.com/WebReflection/linkedom.git cd linkedom/test npm i cd .. npm i npm run benchmark ```