antimatter: an algorithm that prunes CRDT/OT history

Antimatter is the world's first peer-to-peer synchronization algorithm that can prune its history, while peers disconnect and reconnect and need to merge their old changes. Antimatter supports arbitrary simultaneous edits, from arbitrary peers, under arbitrary network delays and partitions, and guarantees full CRDT/OT consistency, while pruning unnecessary history within each partitioned subnet, and across subnets once they reconnect. In steady state, it prunes down to zero overhead. This lets you put synchronizing data structures in more parts of your software, without worrying about memory overhead.

This package implements an antimatter peer composed of three objects:

var {sequence, json, antimatter} = require('@braid.org/antimatter')

• sequence is a pruneable sequence CRDT — sequence meaning it represents a javascript string or array, CRDT meaning this structure can be merged with other ones, and pruneable meaning that it supports an operation to remove meta-data when it is no longer needed (whereas CRDT's often keep track of this meta-data forever).
• json is a pruneable JSON CRDT — JSON meaning it represents an arbitrary JSON datstructure, and CRDT and pruneable having the same meaning as for sequence above. json makes recursive use of sequence structures to represent arbitrary JSON (for instance, a map is represented with a sequence structure for each value, where the first element in the sequence is the value).
• antimatter subclasses the json CRDT object, and adds antimatter algorithm methods to it so that it can communicate with other peers to learn which history can be pruned, and tells the json object to prune it.

click here to see this README side-by-side with the source code.

API

antimatter.create(send[, init])

creates and returns a new antimatter object (or adds antimatter methods and properties to init)

• send: a callback function to be called whenever this antimatter wants to send a message to a peer antimatter. the function takes two parameters: peer, and message, where peer is a string id of the peer to send to, and message is a javascript object to send to them, where ultimately we want to call receive on the target peer, and pass it message as a parameter.
• init: (optional) an antimatter object to start with, which we'll add any properties to that it doesn't have, and we'll add all the antimatter methods to it. this option exists so you can serialize an antimatter instance as JSON, and then restore it later.

var antimatter_instance = antimatter.create((peer, msg) => {
    websockets[peer].send(JSON.stringify(msg))
}, JSON.parse(fs.readFileSync('./antimatter.backup')))

antimatter_instance.receive(message)

let this antimatter object "receive" a message from another antimatter object, presumably from its send callback

websocket.on('message', data => {
    antimatter_instance.receive(JSON.parse(data))
});

you generally do not need to mess with a message object directly, but here are the various message objects you might see:

message get or get_back

get is the first message sent to a newly connected peer, and it will respond with get_back

{cmd: 'get', peer: 'PEER_ID', conn: 'CONN_ID'}

message forget

used to disconnect without creating a fissure

{cmd: 'forget', peer: 'PEER_ID', conn: 'CONN_ID'}

message forget_ack

sent in response to forget

{cmd: 'forget_ack', peer: 'PEER_ID', conn: 'CONN_ID'}

message disconnect

issued locally when we detect that a peer has disconnected, in which case we'll set fissure to true; or when we are forgetting a peer (causing a disconnection), in which case we'll set fissure to false.

{cmd: 'disconnect', peer: 'PEER_ID', fissure: true or false}

message fissure

sent to alert peers about a fissure. the fissure object contains information about the two peers involved in the fissure, the specific connection id that broke, the versions that need to be protected, and the time of the fissure (in case we want to ignore it after some time).

{
    cmd: 'fissure',
    fissure: {
        a: 'PEER_A_ID',
        b: 'PEER_B_ID',
        conn: 'CONN_ID',
        versions: {'VERSION_ID': true, ...},
        time: Date.now()
    }
}

message set

sent to alert peers about a change in the document. the change is represented as a version, with a unique id, a set of parent versions (the most recent versions known before adding this version), and an array of patches, where the offsets in the patches do not take into account the application of other patches in the same array.

{
    cmd: 'set',
    version: 'VERSION_ID',
    parents: {'PARENT_VERSION_ID': true, ...},
    patches: [
        {range: '.json.path.a.b', content: 42}, ...
    ]
}

message local ack

sent in response to set, but not right away; a peer will first send the set to all its other peers, and only after they have all responded with a local ack will the peer send a local ack to the originating peer

{cmd: 'ack', seen: 'local', version: 'VERSION_ID', peer: 'PEER_ID', conn: 'CONN_ID'}

message global ack

sent after an originating peer has received a local ack from all its peers, or by any peer who receives it to all its peers, so that everyone may come to know that this version has been seen by everyone in this peer group.

{cmd: 'ack', seen: 'global', version: 'VERSION_ID', peer: 'PEER_ID', conn: 'CONN_ID'}

message welcome

sent in response to a get, basically contains the initial state of the document; welcome messages are also propogated to our peers, with the inclusion of two extra fields: unack_boundary and min_leaves (these are meant to deal with the issue of a peer disconnecting during the connection process itself, in which case we'll want to include these "unack" versions in the resulting fissure)

{
    cmd: 'welcome',
    versions: [each version looks like a set message...],
    fissures: [each fissure looks like the fissure property in a fissure message...],
    parents: {'PARENT_VERSION_ID': true,
        ...versions you must have before consuming these new versions},

    unack_boundary: {'VERSION_ID': true,
        ...mark these and their ancestors as not-globally-acknowledged,
        even if they were marked as such...},
    min_leaves: {'VERSION_ID': true,
        ...protect these and their ancestors from the unack_boundary's unacknowledging...},

    peer: 'PEER_ID', conn: 'CONN_ID'
}

antimatter_instance.get(peer) or connect(peer)

connect to the given peer — triggers this antimatter object to send a get message to the given peer

alice_antimatter_instance.get('bob')

antimatter_instance.forget(peer)

disconnect from the given peer without creating a fissure — we don't need to reconnect with them.. it seems.. if we do, then we need to call disconnect instead, which will create a fissure allowing us to reconnect.

alice_antimatter_instance.forget('bob')

antimatter_instance.disconnect(peer)

if we detect that a peer has disconnected, let the antimatter object know by calling this method with the given peer — this will create a fissure so we can reconnect with this peer if they come back

alice_antimatter_instance.disconnect('bob')

antimatter_instance.set(...patches)

modify this antimatter object by applying the given patches. each patch looks like {range: '.life.meaning', content: 42}. calling this method will trigger calling the send callback to let our peers know about this change.

antimatter_instance.set({range: '.life.meaning', content: 42})

json.create([init])

create a new json crdt object (or start with init, and add stuff to that).

var json_instance = json.create()

json_instance.read()

returns an instance of the json object represented by this json data-structure

console.log(json_instance.read())

json_instance.generate_braid(versions)

returns an array of set messages that each look like this: {version, parents, patches, sort_keys}, such that if we pass all these messages to antimatter.receive(), we'll reconstruct the data in this json datastructure, assuming the recipient already has the given versions (which is represented as an object where each key is a version, and each value is true).

json_instance.generate_braid({alice2: true, bob3: true})

json_instance.apply_bubbles(to_bubble)

this method helps prune away meta data and compress stuff when we have determined that certain versions can be renamed to other versions — these renamings are expressed in to_bubble, where keys are versions and values are "bubbles", each bubble represented with an array of two elements, the first element is the "bottom" of the bubble, and the second element is the "top" of the bubble; "bottom" and "top" make sense when viewing versions in a directed graph with the oldest version(s) at the top, and each version pointing up to it's parents. a bubble is then a set of versions where the only arrows leaving the bubble upward are from the "top" version, and the only arrows leaving the bubble downward are from the "bottom" version. this method effectively combines all the versions in a bubble into a single version, and may allow the data structure to be compressed, since now we don't need to distinguish between certain versions that we used to need to.

json_instance.apply_bubbles({alice4: ['bob5', 'alice4'], bob5: ['bob5', 'alice4']})

json_instance.add_version(version, parents, patches[, sort_keys])

the main method for modifying a json data structure.

• version: unique string associated with this edit.
• parents: a set of versions that this version is aware of, represented as a map with versions as keys, and values of true.
• patches: an array of patches, where each patch is an object like this {range: '.life.meaning', content: 42}
• sort_keys: (optional) an object where each key is an index, and the value is a sort_key to use with the patch at the given index in the patches array — a sort_key overrides the version for a patch for the purposes of sorting.. this can be useful after doing some pruning.

json_instance.add_version('alice6',
    {alice5: true, bob7: true},
    [{range: '.a.b', content: 'c'}])

json_instance.ancestors(versions, ignore_nonexistent=false)

gather versions and all their ancestors into a set. versions is a set of versions, i.e. a map with version-keys and values of true — we'll basically return a larger set. if ignore_nonexistent is true, then we won't throw an exception if we encounter a version that we don't have in our datastructure.

json_instance.ancestors({alice12: true, bob10: true})

json_instance.get_leaves(versions)

returns a set of versions from versions which don't also have a child in versions. versions is itself a set of versions, represented as an object with version keys and true values, and the return value is represented the same way.

json_instance.parse_patch(patch)

takes a patch in the form {range, content}, and returns an object of the form {path: [...], [slice: [...]], [delete: true], content}; basically calling parse_json_path on patch.range, and adding patch.content along for the ride.

json_instance.parse_json_path(json_path)

parses the string json_path into an object like: {path: [...], [slice: [...]], [delete: true]}.

• a.b[3] --> {path: ['a', 'b', 3]}
• a.b[3:5] --> {path: ['a', 'b'], slice: [3, 5]}
• delete a.b --> {path: ['a', 'b'], delete: true}

console.log(json_instance.parse_json_path('a.b.c'))

sequence.create_node(version, elems, [end_cap, sort_key])

creates a node for a sequence sequence CRDT with the given properties. the resulting node will look like this:

{
    version, // globally unique string
    elems, // a string or array representing actual data elements of the underlying sequence
    end_cap, // this is useful for dealing with replace operations
    sort_key, // version to pretend this is for the purposes of sorting
    deleted_by : {}, // if this node gets deleted, we'll mark it here
    nexts : [], // array of nodes following this one
    next : null // final node following this one (after all the nexts)
}

var sequence_node = sequence.create_node('alice1', 'hello')

sequence.generate_braid(root_node, version, is_anc)

reconstructs an array of splice-information which can be passed to sequence.add_version in order to add version to another sequence instance — the returned array looks like: [[insert_pos, delete_count, insert_elems, sort_key], ...]. is_anc is a function which accepts a version string and returns true if and only if the given version is an ancestor of version (i.e. a version which the author of version knew about when they created that version).

var root_node = sequence.create_node('alice1', 'hello')
console.log(sequence.generate_braid(root_node, 'alice1', x => false)) // outputs [0, 0, "hello"]

sequence.apply_bubbles(root_node, to_bubble)

this method helps prune away meta data and compress stuff when we have determined that certain versions can be renamed to other versions — these renamings are expressed in to_bubble, where keys are versions and values are "bubbles", each bubble represented with an array of two elements, the first element is the "bottom" of the bubble, and the second element is the "top" of the bubble. we will rename the given version to the "bottom" of the bubble. "bottom" and "top" make sense when viewing versions in a directed graph with the oldest version(s) at the top, and each version pointing up to it's parents. a bubble is then a set of versions where the only arrows leaving the bubble upward are from the "top" version, and the only arrows leaving the bubble downward are from the "bottom" version. this method effectively combines all the versions in a bubble into a single version, and may allow the data structure to be compressed, since now we don't need to distinguish between certain versions that we used to need to.

sequence.apply_bubbles(root_node, {alice4: ['bob5', 'alice4'], bob5: ['bob5', 'alice4']})

sequence.get(root_node, i, is_anc)

returns the element at the ith position (0-based) in the sequence rooted at root_node, when only considering versions which result in true when passed to is_anc.

var x = sequence.get(root_node, 2, {alice1: true})

sequence.set(root_node, i, v, is_anc)

sets the element at the ith position (0-based) in the sequence rooted at root_node to the value v, when only considering versions which result in true when passed to is_anc.

sequence.set(root_node, 2, 'x', {alice1: true})

sequence.length(root_node, is_anc)

returns the length of the sequence rooted at root_node, when only considering versions which result in true when passed to is_anc.

console.log(sequence.length(root_node, {alice1: true}))

sequence.break_node(node, break_position, end_cap, new_next)

this methods breaks apart a sequence node into two nodes, each representing a subsequence of the sequence represented by the original node; the node parameter is modified into the first node, and the second node is returned. the first node represents the elements of the sequence before break_position, and the second node represents the rest of the elements. if end_cap is truthy, then the first node will have end_cap set — this is generally done if the elements in the second node are being replaced. this method will add new_next to the first node's nexts array.

var node = sequence.create_node('alice1', 'hello')
// node node.elems == 'hello'

var second = sequence.break_node(node, 2)
// now node.elems   == 'he',
// and second.elems == 'llo'

sequence.add_version(root_node, version, splices, [is_anc])

this is the main method of sequence, used to modify the sequence. the modification must be given a unique version string, and the modification itself is represented as an array of splices, where each splice looks like this: [position, num_elements_to_delete, elements_to_insert, optional_sort_key]. note that all positions are relative to the original sequence, before any splices have been applied. positions are counted by only considering nodes with versions which result in true when passed to is_anc (and are not deleted_by any versions which return true when passed to is_anc).

var node = sequence.create_node('alice1', 'hello')
sequence.add_version(node, 'alice2', [[5, 0, ' world']], null, v => v == 'alice1')

sequence.traverse(root_node, is_anc, callback, [view_deleted, tail_callback])

traverses the subset of nodes in the tree rooted at root_node whos versions return true when passed to is_anc. for each node, callback is called with these parameters: node, offset, has_nexts, prev, version, deleted, where node is the current node being traversed; offset says how many elements we have passed so far getting here; has_nexts is true if some of this node's nexts will be traversed according to is_anc; prev is a pointer to the node whos next points to this one, or null if this is the root node; version is the version of this node, or this node's prev if our version is null, or that node's prev if it is also null, etc; deleted is true if this node is deleted according to is_anc (usually we skip deleted nodes when traversing, but we'll include them if view_deleted is true). tail_callback is an optional callback that will get called with a single parameter node after all of that node's children nexts and next have been traversed.

sequence.traverse(node, () => true, node => process.stdout.write(node.elems))