A CRDT framework with a powerful abstraction of shared data
Yjs is a CRDT implementation that exposes its internal
data structure as shared types. Shared types are common data types like Map
or Array
with superpowers: changes are automatically distributed to other
peers and merged without merge conflicts.
Yjs is network agnostic (p2p!), supports many existing rich text editors, offline editing, version snapshots, undo/redo and shared cursors. It scales well with an unlimited number of users and is well suited for even large documents.
👷♀️ If you are looking for professional (paid) support to build collaborative or distributed applications ping us at [email protected]. Otherwise you can find help on our discussion board.
I'm currently looking for sponsors that allow me to be less dependent on contracting work. These awesome backers already fund further development of Yjs:
Sponsorship also comes with special perks!
This repository contains a collection of shared types that can be observed for changes and manipulated concurrently. Network functionality and two-way-bindings are implemented in separate modules.
Name | Cursors | Binding | Demo |
---|---|---|---|
ProseMirror | ✔ | y-prosemirror | demo |
Quill | ✔ | y-quill | demo |
CodeMirror | ✔ | y-codemirror | demo |
Monaco | ✔ | y-monaco | demo |
Slate | ✔ | slate-yjs | demo |
Setting up the communication between clients, managing awareness information, and storing shared data for offline usage is quite a hassle. Providers manage all that for you and are the perfect starting point for your collaborative app.
Install Yjs and a provider with your favorite package manager:
npm i yjs y-websocket
Start the y-websocket server:
PORT=1234 node ./node_modules/y-websocket/bin/server.js
const yarray = doc.getArray('my-array')
yarray.observe(event => {
console.log('yarray was modified')
})
// every time a local or remote client modifies yarray, the observer is called
yarray.insert(0, ['val']) // => "yarray was modified"
Remember, shared types are just plain old data types. The only limitation is that a shared type must exist only once in the shared document.
const ymap = doc.getMap('map')
const foodArray = new Y.Array()
foodArray.insert(0, ['apple', 'banana'])
ymap.set('food', foodArray)
ymap.get('food') === foodArray // => true
ymap.set('fruit', foodArray) // => Error! foodArray is already defined
Now you understand how types are defined on a shared document. Next you can jump to the demo repository or continue reading the API docs.
Any of the Yjs providers can be combined with each other. So you can sync data over different network technologies.
In most cases you want to use a network provider (like y-websocket or y-webrtc) in combination with a persistence provider (y-indexeddb in the browser). Persistence allows you to load the document faster and to persist data that is created while offline.
For the sake of this demo we combine two different network providers with a persistence provider.
import * as Y from 'yjs'
import { WebrtcProvider } from 'y-webrtc'
import { WebsocketProvider } from 'y-websocket'
import { IndexeddbPersistence } from 'y-indexeddb'
const ydoc = new Y.Doc()
// this allows you to instantly get the (cached) documents data
const indexeddbProvider = new IndexeddbPersistence('count-demo', ydoc)
indexeddbProvider.whenSynced.then(() => {
console.log('loaded data from indexed db')
})
// Sync clients with the y-webrtc provider.
const webrtcProvider = new WebrtcProvider('count-demo', ydoc)
// Sync clients with the y-websocket provider
const websocketProvider = new WebsocketProvider(
'wss://demos.yjs.dev', 'count-demo', ydoc
)
// array of numbers which produce a sum
const yarray = ydoc.getArray('count')
// observe changes of the sum
yarray.observe(event => {
// print updates when the data changes
console.log('new sum: ' + yarray.toArray().reduce((a,b) => a + b))
})
// add 1 to the sum
yarray.push([1]) // => "new sum: 1"
import * as Y from 'yjs'
A shareable Array-like type that supports efficient insert/delete of elements at any position. Internally it uses a linked list of Arrays that is split when necessary.
const yarray = new Y.Array()
parent:Y.AbstractType|null
insert(index:number, content:Array<object|boolean|Array|string|number|Uint8Array|Y.Type>)
array.insert(0, [1])
splices the list and inserts 1 at
position 0.
push(Array<Object|boolean|Array|string|number|Uint8Array|Y.Type>)
unshift(Array<Object|boolean|Array|string|number|Uint8Array|Y.Type>)
delete(index:number, length:number)
get(index:number)
slice(start:number, end:number):Array<Object|boolean|Array|string|number|Uint8Array|Y.Type>
length:number
forEach(function(value:object|boolean|Array|string|number|Uint8Array|Y.Type,
index:number, array: Y.Array))
map(function(T, number, YArray):M):Array<M>
toArray():Array<object|boolean|Array|string|number|Uint8Array|Y.Type>
toJSON():Array<Object|boolean|Array|string|number>
toJSON
method.
[Symbol.Iterator]
for (let value of yarray) { .. }
observe(function(YArrayEvent, Transaction):void)
unobserve(function(YArrayEvent, Transaction):void)
observe
event listener from this type.
observeDeep(function(Array<YEvent>, Transaction):void)
unobserveDeep(function(Array<YEvent>, Transaction):void)
observeDeep
event listener from this type.
A shareable Map type.
const ymap = new Y.Map()
parent:Y.AbstractType|null
get(key:string):object|boolean|string|number|Uint8Array|Y.Type
set(key:string, value:object|boolean|string|number|Uint8Array|Y.Type)
delete(key:string)
has(key:string):boolean
get(index:number)
clone():Y.Map
toJSON():Object<string, Object|boolean|Array|string|number|Uint8Array>
[key,value]
pairs of this YMap to a new Object.It
transforms all child types to JSON using their toJSON
method.
forEach(function(value:object|boolean|Array|string|number|Uint8Array|Y.Type,
key:string, map: Y.Map))
[Symbol.Iterator]
[key, value]
pairs.
for (let [key, value] of ymap) { .. }
entries()
[key, value]
pairs.
values()
keys()
observe(function(YMapEvent, Transaction):void)
unobserve(function(YMapEvent, Transaction):void)
observe
event listener from this type.
observeDeep(function(Array<YEvent>, Transaction):void)
unobserveDeep(function(Array<YEvent>, Transaction):void)
observeDeep
event listener from this type.
A shareable type that is optimized for shared editing on text. It allows to assign properties to ranges in the text. This makes it possible to implement rich-text bindings to this type.
This type can also be transformed to the delta format. Similarly the YTextEvents compute changes as deltas.
const ytext = new Y.Text()
parent:Y.AbstractType|null
insert(index:number, content:string, [formattingAttributes:Object<string,string>])
ytext.insert(0, 'bold text', { bold: true })
delete(index:number, length:number)
format(index:number, length:number, formattingAttributes:Object<string,string>)
applyDelta(delta, opts:Object<string,any>)
ytext.applyDelta(delta, { sanitize: false })
length:number
toString():string
toJSON():string
toString
toDelta():Delta
observe(function(YTextEvent, Transaction):void)
unobserve(function(YTextEvent, Transaction):void)
observe
event listener from this type.
observeDeep(function(Array<YEvent>, Transaction):void)
unobserveDeep(function(Array<YEvent>, Transaction):void)
observeDeep
event listener from this type.
A container that holds an Array of Y.XmlElements.
const yxml = new Y.XmlFragment()
parent:Y.AbstractType|null
firstChild:Y.XmlElement|Y.XmlText|null
insert(index:number, content:Array<Y.XmlElement|Y.XmlText>)
delete(index:number, length:number)
get(index:number)
slice(start:number, end:number):Array<Y.XmlElement|Y.XmlText>
length:number
clone():Y.XmlFragment
toArray():Array<Y.XmlElement|Y.XmlText>
toDOM():DocumentFragment
toString():string
toJSON():string
toString
.observe(function(YXmlEvent, Transaction):void)
unobserve(function(YXmlEvent, Transaction):void)
observe
event listener from this type.
observeDeep(function(Array<YEvent>, Transaction):void)
unobserveDeep(function(Array<YEvent>, Transaction):void)
observeDeep
event listener from this type.
A shareable type that represents an XML Element. It has a nodeName
,
attributes, and a list of children. But it makes no effort to validate its
content and be actually XML compliant.
const yxml = new Y.XmlElement()
parent:Y.AbstractType|null
firstChild:Y.XmlElement|Y.XmlText|null
nextSibling:Y.XmlElement|Y.XmlText|null
prevSibling:Y.XmlElement|Y.XmlText|null
insert(index:number, content:Array<Y.XmlElement|Y.XmlText>)
delete(index:number, length:number)
get(index:number)
length:number
setAttribute(attributeName:string, attributeValue:string)
removeAttribute(attributeName:string)
getAttribute(attributeName:string):string
getAttributes(attributeName:string):Object<string,string>
get(i:number):Y.XmlElement|Y.XmlText
slice(start:number, end:number):Array<Y.XmlElement|Y.XmlText>
clone():Y.XmlElement
toArray():Array<Y.XmlElement|Y.XmlText>
toDOM():Element
toString():string
toJSON():string
toString
.observe(function(YXmlEvent, Transaction):void)
unobserve(function(YXmlEvent, Transaction):void)
observe
event listener from this type.
observeDeep(function(Array<YEvent>, Transaction):void)
unobserveDeep(function(Array<YEvent>, Transaction):void)
observeDeep
event listener from this type.
const doc = new Y.Doc()
clientID
gc
transact(function(Transaction):void [, origin:any])
update
event are called after each transaction. You should
bundle changes into a single transaction to reduce the amount of event
calls. I.e. doc.transact(() => { yarray.insert(..); ymap.set(..) })
triggers a single change event. origin
parameter that is stored on transaction.origin
and
on('update', (update, origin) => ..)
.
toJSON():any
get(string, Y.[TypeClass]):[Type]
getArray(string):Y.Array
y.get(string, Y.Array)
.getMap(string):Y.Map
y.get(string, Y.Map)
.getXmlFragment(string):Y.XmlFragment
y.get(string, Y.XmlFragment)
.on(string, function)
off(string, function)
on('update', function(updateMessage:Uint8Array, origin:any, Y.Doc):void)
on('beforeTransaction', function(Y.Transaction, Y.Doc):void)
on('afterTransaction', function(Y.Transaction, Y.Doc):void)
on('beforeAllTransactions', function(Y.Doc):void)
on('afterAllTransactions', function(Y.Doc, Array<Y.Transaction>):void)
Changes on the shared document are encoded into document updates. Document updates are commutative and idempotent. This means that they can be applied in any order and multiple times.
const doc1 = new Y.Doc()
const doc2 = new Y.Doc()
doc1.on('update', update => {
Y.applyUpdate(doc2, update)
})
doc2.on('update', update => {
Y.applyUpdate(doc1, update)
})
// All changes are also applied to the other document
doc1.getArray('myarray').insert(0, ['Hello doc2, you got this?'])
doc2.getArray('myarray').get(0) // => 'Hello doc2, you got this?'
Yjs internally maintains a state vector that denotes the next expected clock from each client. In a different interpretation it holds the number of structs created by each client. When two clients sync, you can either exchange the complete document structure or only the differences by sending the state vector to compute the differences.
const state1 = Y.encodeStateAsUpdate(ydoc1)
const state2 = Y.encodeStateAsUpdate(ydoc2)
Y.applyUpdate(ydoc1, state2)
Y.applyUpdate(ydoc2, state1)
This example shows how to sync two clients with the minimal amount of exchanged data by computing only the differences using the state vector of the remote client. Syncing clients using the state vector requires another roundtrip, but can safe a lot of bandwidth.
const stateVector1 = Y.encodeStateVector(ydoc1)
const stateVector2 = Y.encodeStateVector(ydoc2)
const diff1 = Y.encodeStateAsUpdate(ydoc1, stateVector2)
const diff2 = Y.encodeStateAsUpdate(ydoc2, stateVector1)
Y.applyUpdate(ydoc1, diff2)
Y.applyUpdate(ydoc2, diff1)
It is possible to sync clients and compute delta updates without loading the Yjs document to memory. Yjs exposes an API to compute the differences directly on the binary document updates.
// encode the current state as a binary buffer
let currentState1 = Y.encodeStateAsUpdate(ydoc1)
let currentState2 = Y.encodeStateAsUpdate(ydoc2)
// now we can continue syncing clients using state vectors without using the Y.Doc
ydoc1.destroy()
ydoc2.destroy()
const stateVector1 = Y.encodeStateVectorFromUpdate(currentState1)
const stateVector2 = Y.encodeStateVectorFromUpdate(currentState2)
const diff1 = Y.diffUpdate(currentState1, stateVector2)
const diff2 = Y.diffUpdate(currentState2, stateVector1)
// sync clients
currentState1 = Y.mergeUpdates([currentState1, diff2])
currentState1 = Y.mergeUpdates([currentState1, diff1])
Y.applyUpdate(Y.Doc, update:Uint8Array, [transactionOrigin:any])
transactionOrigin
that will be stored on
transaction.origin
and ydoc.on('update', (update, origin) => ..)
.
Y.encodeStateAsUpdate(Y.Doc, [encodedTargetStateVector:Uint8Array]):Uint8Array
Y.encodeStateVector(Y.Doc):Uint8Array
Y.mergeUpdates(Array<Uint8Array>)
Y.encodeStateVectorFromUpdate(Uint8Array): Uint8Array
Y.diffUpdate(update: Uint8Array, stateVector: Uint8Array): Uint8Array
Y.encodeStateAsUpdate(ydoc, stateVector)
but works
on updates instead.
const relPos = Y.createRelativePositionFromTypeIndex(ytext, 2)
const pos = Y.createAbsolutePositionFromRelativePosition(relPos, doc)
pos.type === ytext // => true
pos.index === 2 // => true
const relPos = Y.createRelativePositionFromTypeIndex(ytext, 2)
const encodedRelPos = JSON.stringify(relPos)
// send encodedRelPos to remote client..
const parsedRelPos = JSON.parse(encodedRelPos)
const pos = Y.createAbsolutePositionFromRelativePosition(parsedRelPos, remoteDoc)
pos.type === remoteytext // => true
pos.index === 2 // => true
const relPos = Y.createRelativePositionFromTypeIndex(ytext, 2)
const encodedRelPos = Y.encodeRelativePosition(relPos)
// send encodedRelPos to remote client..
const parsedRelPos = Y.decodeRelativePosition(encodedRelPos)
const pos = Y.createAbsolutePositionFromRelativePosition(parsedRelPos, remoteDoc)
pos.type === remoteytext // => true
pos.index === 2 // => true
Y.createRelativePositionFromTypeIndex(Uint8Array|Y.Type, number)
Y.createAbsolutePositionFromRelativePosition(RelativePosition, Y.Doc)
Y.encodeRelativePosition(RelativePosition):Uint8Array
Y.decodeRelativePosition(Uint8Array):RelativePosition
Yjs ships with an Undo/Redo manager for selective undo/redo of of changes on a Yjs type. The changes can be optionally scoped to transaction origins.
const ytext = doc.getText('text')
const undoManager = new Y.UndoManager(ytext)
ytext.insert(0, 'abc')
undoManager.undo()
ytext.toString() // => ''
undoManager.redo()
ytext.toString() // => 'abc'
constructor(scope:Y.AbstractType|Array<Y.AbstractType>
[, {captureTimeout:number,trackedOrigins:Set<any>,deleteFilter:function(item):boolean}])
undo()
redo()
stopCapturing()
on('stack-item-added', { stackItem: { meta: Map<any,any> }, type: 'undo'
| 'redo' })
StackItem
is added to the
undo- or the redo-stack.
on('stack-item-popped', { stackItem: { meta: Map<any,any> }, type: 'undo'
| 'redo' })
StackItem
is popped from
the undo- or the redo-stack.
UndoManager merges Undo-StackItems if they are created within time-gap
smaller than options.captureTimeout
. Call um.stopCapturing()
so that the next
StackItem won't be merged.
// without stopCapturing
ytext.insert(0, 'a')
ytext.insert(1, 'b')
undoManager.undo()
ytext.toString() // => '' (note that 'ab' was removed)
// with stopCapturing
ytext.insert(0, 'a')
undoManager.stopCapturing()
ytext.insert(0, 'b')
undoManager.undo()
ytext.toString() // => 'a' (note that only 'b' was removed)
Every change on the shared document has an origin. If no origin was specified,
it defaults to null
. By specifying trackedOrigins
you can
selectively specify which changes should be tracked by UndoManager
. The
UndoManager instance is always added to trackedOrigins
.
class CustomBinding {}
const ytext = doc.getText('text')
const undoManager = new Y.UndoManager(ytext, {
trackedOrigins: new Set([42, CustomBinding])
})
ytext.insert(0, 'abc')
undoManager.undo()
ytext.toString() // => 'abc' (does not track because origin `null` and not part
// of `trackedTransactionOrigins`)
ytext.delete(0, 3) // revert change
doc.transact(() => {
ytext.insert(0, 'abc')
}, 42)
undoManager.undo()
ytext.toString() // => '' (tracked because origin is an instance of `trackedTransactionorigins`)
doc.transact(() => {
ytext.insert(0, 'abc')
}, 41)
undoManager.undo()
ytext.toString() // => '' (not tracked because 41 is not an instance of
// `trackedTransactionorigins`)
ytext.delete(0, 3) // revert change
doc.transact(() => {
ytext.insert(0, 'abc')
}, new CustomBinding())
undoManager.undo()
ytext.toString() // => '' (tracked because origin is a `CustomBinding` and
// `CustomBinding` is in `trackedTransactionorigins`)
When undoing or redoing a previous action, it is often expected to restore additional meta information like the cursor location or the view on the document. You can assign meta-information to Undo-/Redo-StackItems.
const ytext = doc.getText('text')
const undoManager = new Y.UndoManager(ytext, {
trackedOrigins: new Set([42, CustomBinding])
})
undoManager.on('stack-item-added', event => {
// save the current cursor location on the stack-item
event.stackItem.meta.set('cursor-location', getRelativeCursorLocation())
})
undoManager.on('stack-item-popped', event => {
// restore the current cursor location on the stack-item
restoreCursorLocation(event.stackItem.meta.get('cursor-location'))
})
Conflict-free replicated data types (CRDT) for collaborative editing are an alternative approach to operational transformation (OT). A very simple differenciation between the two approaches is that OT attempts to transform index positions to ensure convergence (all clients end up with the same content), while CRDTs use mathematical models that usually do not involve index transformations, like linked lists. OT is currently the de-facto standard for shared editing on text. OT approaches that support shared editing without a central source of truth (a central server) require too much bookkeeping to be viable in practice. CRDTs are better suited for distributed systems, provide additional guarantees that the document can be synced with remote clients, and do not require a central source of truth.
Yjs implements a modified version of the algorithm described in this paper. This article explains a simple optimization on the CRDT model and gives more insight about the performance characteristics in Yjs. More information about the specific implementation is available in INTERNALS.md and in this walkthrough of the Yjs codebase.
CRDTs that suitable for shared text editing suffer from the fact that they only grow in size. There are CRDTs that do not grow in size, but they do not have the characteristics that are benificial for shared text editing (like intention preservation). Yjs implements many improvements to the original algorithm that diminish the trade-off that the document only grows in size. We can't garbage collect deleted structs (tombstones) while ensuring a unique order of the structs. But we can 1. merge preceeding structs into a single struct to reduce the amount of meta information, 2. we can delete content from the struct if it is deleted, and 3. we can garbage collect tombstones if we don't care about the order of the structs anymore (e.g. if the parent was deleted).
Examples:
array.insert(0, ['a']), array.insert(0, ['b']);
is
first represented as two structs ([{id: {client, clock: 0}, content: 'a'}, {id: {client, clock: 1}, content: 'b'}
) and then merged into a single
struct: [{id: {client, clock: 0}, content: 'ab'}]
.ItemString
) is deleted, the
struct will be replaced with an ItemDeleted
that does not contain content
anymore.GC
structs. A
GC
struct only denotes the existence of a struct and that it is deleted.
GC
structs can always be merged with other GC
structs if the id's are
adjacent.Especially when working on structured content (e.g. shared editing on ProseMirror), these improvements yield very good results when benchmarking random document edits. In practice they show even better results, because users usually edit text in sequence, resulting in structs that can easily be merged. The benchmarks show that even in the worst case scenario that a user edits text from right to left, Yjs achieves good performance even for huge documents.
Yjs has the ability to exchange only the differences when syncing two clients.
We use lamport timestamps to identify structs and to track in which order a
client created them. Each struct has an struct.id = { client: number, clock: number}
that uniquely identifies a struct. We define the next expected clock
by each client as the state vector. This data structure is similar to the
version vectors data structure.
But we use state vectors only to describe the state of the local document, so we
can compute the missing struct of the remote client. We do not use it to track
causality.
Yjs and all related projects are MIT licensed.
Yjs is based on my research as a student at the RWTH i5. Now I am working on Yjs in my spare time.
Fund this project by donating on GitHub Sponsors or hiring me as a contractor for your collaborative app.