DOT

The DOT project is a blend of operational transformation, CmRDT, persistent/immutable datastructures and reactive stream processing.

This is an implementation of distributed data synchronization of rich custom data structures with conflict-free merging.

Status

This is very close to v1 release. The ES6 version interoperates well right now but outstanding short-term issues have more to do with consistency of the API surface than features:

• The ES6 version has a simpler polling-based Network API that seems worth adopting here. ** Adopted **
• The ES6 branch/undo integration also feels a lot simpler. ** Adopted **
• The ES6 version prefers replace() instead of update().
• Nullable value types (i.e typed Nil values vs change.Nil vs nil) seems confusing.

Features

1. Small, well tested mutations and immutable persistent values
2. Support for rich user-defined types, not just collaborative text
3. Streams and Git-like branching, merging support
4. Simple network support (Gob serialization) and storage support
5. Strong references support that are automatically updated with changes
6. Rich builtin undo support for any type and mutation
7. Folding (committed changes on top of uncommitted changes)
8. Support for CmRDT types (see crdt)

An interoperable ES6 version is available on dotchain/dotjs with a TODO MVC demo of it here

Contents

1. Status
2. Features
3. CRDTs
4. TODO Example
   1. Server
   2. Types
   3. Type registration
   4. Code generation
   5. Toggling Complete
   6. Changing description
   7. Adding Todos
   8. Client connection
   9. Running the demo
   10. In browser demo
5. How it all works
   1. Applying changes
   2. Applying changes with streams
   3. Composition of changes
   4. Convergence
   5. Convergence using streams
   6. Revert and undo
   7. Folding
   8. Branching of streams
   9. References
   10. Network synchronization and server
6. Broad Issues
7. Contributing

CRDTs

Much of the framework can support operation-based CRDT changes which simply appear as commutative operations (and so the merge operation is trivial). A set of types built this way is available in the crdt folder.

TODO Example

The standard TODO-MVC example demonstrates the features of collaborative (eventually consistent) distributed data structures.

Server

The DOT backend is essentially a simple log store. All mutations to the application state are represented as a sequence of operations and written in append-only fashion onto the log. The following snippet shows how to start a web server (though it does not include authentication or CORs for example).

func Server() {
	// import net/http
	// import github.com/dotchain/dot

        // uses a local-file backed bolt DB backend
	http.Handle("/dot/", dot.BoltServer("file.bolt"))
        http.ListenAndServe(":8080", nil)
}

The example above uses the Bolt backend for the actual storage of the operations. There is also a Postgres backend available.

Note that the server above has no real reference to any application logic: it simply accepts operations and writes them out in a guaranteed order broadcasting these to all the clients.

Types

A TODO MVC app consists of only two core types: Todo and TodoList:

// Todo tracks a single todo item
type Todo struct {
	Complete bool
        Description string
}

// TodoList tracks a collection of todo items
type TodoList []Todo

Type registration

To use the types across the network, they have to be registered with the codec (which will be sjson in this example)

// import github.com/dotchain/dot/ops/nw

func init() {
	nw.Register(Todo{})
        nw.Register(TodoList{})
}

Code generation

For use with DOT, these types need to be augmented with standard methods of the Value interface (or in the case of lists like TodoList, also implement the Collection interface).

These interfaces are essentially the ability to take changes of the form replace a sub field or replace items in the array and calculate the result of applying them. They are mostly boilerplate and so can be autogenerated easily via the dotc package. See code generation for augmenting the above type information.

The code generation not only implements these two interfaces, it also produces a new Stream type for Todo and TodoList. A stream type is like a linked list with the Value field being the underlying value and Next() returning the next entry in the stream (in case the value was modified). And Latest returns the last entry in the stream at that point. Also, each stream type implements mutation methods to easily modify the value associated with a stream.

What makes the streams interesting is that two different modifications from the same state cause the Latest of both to be the same with the effect of both merged. (This is done using the magic of operational transformations)

Toggling Complete

The code to toggle the Complete field of a particular todo item looks like the following:

func Toggle(t *TodoListStream, index int) {
	// TodoListStream.Item() is generated code. It returns
        // a stream of the n'th element of the slice so that
        // particular stream can be modified. When that stream is
        // modified, the effect is automatically merged into the
        // parent (and available via .Next of the parent stream)
	todoStream := t.Item(index) 

	// TodoStream.Complete is generated code. It returns a stream
        // for the Todo.Complete field so that it can be modified. As
        // with slices above, mutations on the field's stream are
        // reflected on the struct stream (via .Next or .Latest())
        completeStream := todoStream.Complete()

	// completeStream is of type streams.Bool. All streams
        // implement the simple Update(newValue) method that replaces
        // the current value with a new value.
        completeStream.Update(!completeStream.Value)
}

Note that the function does not return any value here but the updates can be fetched by calling .Latest() on any of the corresponding streams. If a single stream instance has multiple edits, the Latest() value is the merged value of all those edits.

Changing description

The code for changing the Description field is similar. The string Description field in Todo maps to a streams.S16 stream. This implements an Update() method like all streams.

But to make things interesting, lets look at splicing rather than replacing the whole string. Splicing is taking a subsequence of the string at a particular position and replacing it with the provided value. It captures insert, delete and replace in one operation.

This probably better mimics what text editors do and a benefit of such high granularity edits is that when two users edit the same text, the edits will merge quite cleanly so long as they don't directly touch the same characters.

func SpliceDescription(t *TodoListStream, index, offset, count int, replacement string) {
	// TodoListStream.Item() is generated code. It returns
        // a stream of the n'th element of the slice so that
        // particular stream can be modified. When that stream is
        // modified, the effect is automatically merged into the
        // parent (and available via .Next of the parent stream)
	todoStream := t.Item(index) 

	// TodoStream.Description is generated code. It returns a
        // stream for the Todo.Description field so that it can be
        // modified. As with slices above, mutations on the field's
        // stream are reflected on the struct stream (via .Next or
        // .Latest()) 
	// TodoStream.Description() returns streams.S16 type
        descStream := todoStream.Description()

	// streams.S16 implements Splice(offset, removeCount, replacement)
        descStream.Splice(offset, count, replacement)
}

Adding Todos

Adding a Todo is relatively simple as well:

func AddTodo(t *TodoListStream, todo Todo) {
	// All slice streams implement Splice(offset, removeCount, replacement)
	t.Splice(len(t.Value), 0, todo)
}

The use of Splice in this example should hint that (just like strings) collections support insertion/deletion at arbitrary points within via the Splice method. In addition to supporting this, collections also support the Move(offset, count, distance) method to move some items around within the collection

Client connection

Setting up the client requires connecting to the URL where the server is hosted. In addition, the code below illustrates how sessions could be saved and restarted if needed.

// import time
// import sync
// import github.com/dotchain/dot

var Lock sync.Mutex
func Client(stop chan struct{}, render func(*TodoListStream)) {
	url := "http://localhost:8080/dot/"
        session, todos := SavedSession()
	s, store := session.NonBlockingStream(url, nil)
        defer store.Close()

	todosStream := &TodoListStream{Stream: s, Value: todos}

        ticker := time.NewTicker(500*time.Millisecond)
        changed := true
	for {
        	if changed {
			render(todosStream)
                }
        	select {
                case <- stop:
                	return
                case <- ticker.C:
                }

                Lock.Lock()
		s.Push()
                s.Pull()
                next := todosStream.Latest()
                changed = next != todosStream
                todosStream, s = next, next.Stream
                Lock.Unlock()
        }

       	SaveSession(session, todosStream.Value)
}


func SaveSession(s *dot.Session, todos TodoList) {
	// this is not yet implemented. if it were, then
        // this value should be persisted locally and returned
        // by the call to savedSession
}

func SavedSession() (s *dot.Session, todos TodoList) {
	// this is not yet implemented. return default values
        return dot.NewSession(), nil
}

Running the demo

The TODO MVC demo is in the example folder.

The snippets in this markdown file can be used to generate the todo.go file and then auto-generate the "generated.go" file:

$ go get github.com/tvastar/test/cmd/testmd
$ testmd -pkg example -o examples/todo.go README.md
$ testmd -pkg main codegen.md > examples/generated.go

The server can then be started by:

$ go run server.go

The client can then be started by:

$ go run client.go

The provide client.go stub file simply appends a task every 10 seconds.

In browser demo

The fuss project has demos of a TODO-MVC app built on top of this framework using gopherjs. In particular, the collab folder illustrates how simple the code is to make something work collaboratively (the rest of the code base is not even aware of whether things are collaborative).

How it all works

There are values, changes and streams.

1. Values implement the Value interface. If the value represents a collection, it also implements the Collection interface.
2. Changes represent mutations to values that can be merged. If two independent changes are made to the same value, they can be merged so that the A + merged(B) = B + merged(A). This is represented by the Change interface. The changes package implements the core changes with composition that allow richer changes to be implemented.
3. Streams represent a sequence of changes to a value, except it is convergent -- if multiple writers modify a value, they each get a separate stream instance that only reflects their local change but following the Next chain will guarantee that all versions end up with the same final value.

Applying changes

The following example illustrates how to edit a string with values and changes

	// import fmt
        // import github.com/dotchain/dot/changes
        // import github.com/dotchain/dot/changes/types

	// S8 is DOT-compatible string type with UTF8 string indices
	initial := types.S8("hello")

        append := changes.Splice{
        	Offset: len("hello"), // end of "hello"
                Before: types.S8(""), // nothing to remove
                After: types.S8(" world"), // insert " world"
        }

        // apply the change
        updated := initial.Apply(nil, append)

	fmt.Println(updated)
        // Output: hello world

Applying changes with streams

A less verbose stream based version (preferred) would look like so:

	// import fmt
        // import github.com/dotchain/dot/streams

        initial := &streams.S8{Stream: streams.New(), Value: "hello"}
        updated := initial.Splice(5, 0, " world")

	fmt.Println(updated.Value)
        // Output: hello world

The changes package implements the core changes: Splice, Move and Replace. The logical model for these changes is to treat all values as either being like arrays or like maps. The actual underlying datatype can be different as long as the array/map semantics is implemented.

Composition of changes

Changes can be composed together. A simple form of composition is just a set of changes:

	// import fmt
        // import github.com/dotchain/dot/changes
        // import github.com/dotchain/dot/changes/types

	initial := types.S8("hello")

        // append " world" => "hello world"
        append1 := changes.Splice{
        	Offset: len("hello"),
                Before: types.S8(""),
                After: types.S8(" world"),
        }

        // append "." => "hello world."
        append2 := changes.Splice{
        	Offset: len("hello world"),
                Before: types.S8(""),
                After: types.S8("."),
        }
        
        // now combine the two appends and apply
        both := changes.ChangeSet{append1, append2}
        updated := initial.Apply(nil, both)
        fmt.Println(updated)

	// Output: hello world.

Another form of composition is modifying a sub-element such as an array element or a dictionary path:

	// import fmt
        // import github.com/dotchain/dot/changes
        // import github.com/dotchain/dot/changes/types

        // types.A is a generic array type and types.M is a map type
        initial := types.A{types.M{"hello": types.S8("world")}}

        // replace "world" with "world!"
        replace := changes.Replace{Before: types.S8("world"), After: types.S8("world!")}

        // replace "world" with "world!" of initial[0]["hello"]
        path := []interface{}{0, "hello"}
        c := changes.PathChange{Path: path, Change: replace}
        updated := initial.Apply(nil, c)
        fmt.Println(updated)

	// Output: [map[hello:world!]]        

Convergence

The core property of all changes is the ability to guarantee convergence when two mutations are attempted on the same state:

	// import fmt
        // import github.com/dotchain/dot/changes
        // import github.com/dotchain/dot/changes/types

	initial := types.S8("hello")

	// two changes: append " world" and delete "lo"
	insert := changes.Splice{Offset: 5, Before: types.S8(""), After: types.S8(" world")}
	remove := changes.Splice{Offset: 3, Before: types.S8("lo"), After: types.S8("")}

	// two versions derived from initial
        inserted := initial.Apply(nil, insert)
        removed := initial.Apply(nil, remove)

        // merge the changes
        removex, insertx := insert.Merge(remove)

        // converge by applying the above
        final1 := inserted.Apply(nil, removex)
        final2 := removed.Apply(nil, insertx)

        fmt.Println(final1, final1 == final2)
        // Output: hel world true

Convergence using streams

The same convergence example is a lot easier to read with streams:

	// import fmt
        // import github.com/dotchain/dot/streams

	initial := streams.S8{Stream:  streams.New(), Value: "hello"}

	// two changes: append " world" and delete "lo"
        s1 := initial.Splice(5, 0, " world")
	s2 := initial.Splice(3, len("lo"), "")

	// streams automatically merge because they are both
        // based on initial
        s1 = s1.Latest()
        s2 = s2.Latest()

        fmt.Println(s1.Value, s1.Value == s2.Value)
        // Output: hel world true

The ability to merge two independent changes done to the same initial state is the basis for the eventual convergence of the data structures. The changes package has fairly intensive tests to cover the change types defined there, both individually and in composition.

Revert and undo

All the predefined types of changes in DOT (see changes) are carefully designed so that every change can be inverted easily without reference to the underlying value. For example, changes.Replace has both the Before and After fields instead of just keeping the After. This allows the reverse to be computed quite easily by swapping the two fields. This does generally incur additional storage expenses but the tradeoff is that code gets much simpler to work with.

In particular, it is possible to build generic undo support quite easily and naturally. The following example shows both Undo and Redo being invoked from an undo stack.

	// import fmt
        // import github.com/dotchain/dot/streams
        // import github.com/dotchain/dot/changes
        // import github.com/dotchain/dot/changes/types
        // import github.com/dotchain/dot/streams/undo

	// create master, undoable child and the undo stack itself
	master := &streams.S16{Stream: streams.New(), Value: "hello"}
        s := undo.New(master.Stream)
        undoableChild := &streams.S16{Stream: s, Value: master.Value}

	// change hello => Hello
	undoableChild = undoableChild.Splice(0, len("h"), "H")
	fmt.Println(undoableChild.Value)

	// for kicks, update master hello => hello$ as if it came
        // from the server
        master.Splice(len("hello"), 0, "$")

	// now undo this via the stack
        s.Undo()

	// now undoableChild should be hello$
        undoableChild = undoableChild.Latest()
        fmt.Println(undoableChild.Value)

	// now redo the last operation to get Hello$
        s.Redo()
        undoableChild = undoableChild.Latest()
        fmt.Println(undoableChild.Value)
        
	// Output:
        // Hello
        // hello$
        // Hello$

Folding

In the case of editors, folding refers to a piece of text that has been hidden away. The difficulty with implementing this in a collaborative setting is that as external edits come in, the fold has to be maintained.

The design of DOT allows for an elegant way to achieve this: consider the "folding" as a local change (replacing the folded region with say "..."). This local change is never meant to be sent out. All changes to the unfolded and folded versions can be proxied quite nicely without much app involvement:

	// import fmt
        // import github.com/dotchain/dot/streams
        // import github.com/dotchain/dot/changes
        // import github.com/dotchain/dot/changes/types
        // import github.com/dotchain/dot/x/fold

	// create master, folded child and the folding itself
	master := &streams.S16{Stream: streams.New(), Value: "hello world!"}
        foldChange := changes.Splice{
        	Offset: len("hello"),
                Before: types.S16(" world"),
                After: types.S16("..."),
        }
        foldedStream := fold.New(foldChange, master.Stream)
        folded := &streams.S16{Stream: foldedStream, Value :"hello...!"}

        // folded:  hello...! => Hello...!!!
	folded = folded.Splice(0, len("h"), "H")
        folded = folded.Splice(len("Hello...!"), 0, "!!")
        fmt.Println(folded.Value)

	// master: hello world => hullo world
	master = master.Splice(len("h"), len("e"), "u")
        fmt.Println(master.Value)

        // now folded = Hullo...!!!
        fmt.Println(folded.Latest().Value)

        // master = Hullo world!!!
        fmt.Println(master.Latest().Value)

	// Output:
        // Hello...!!!
        // hullo world!
        // Hullo...!!!
        // Hullo world!!!

Branching of streams

Streams in DOT can also be branched a la Git. Changes made in branches do not affect the master or vice-versa -- until one of Pull or Push are called.

	// import fmt
        // import github.com/dotchain/dot/streams
        // import github.com/dotchain/dot/changes
        // import github.com/dotchain/dot/changes/types        
        
        // local is a branch of master
        master := &streams.S16{Stream: streams.New(), Value: "hello"}
        local := &streams.S16{Stream: streams.Branch(master.Stream), Value: master.Value}

	// edit locally: hello => hallo
	local.Splice(len("h"), len("e"), "a")

	// changes will not be reflected on master yet
        fmt.Println(master.Latest().Value)

	// push local changes up to master now
        local.Stream.Push()

	// now master = hallo
	fmt.Println(master.Latest().Value)

        // Output:
        // hello
        // hallo

There are other neat benefits to the branching model: it provides a fine grained control for pulling changes from the network on demand and suspending it as well as providing a way for making local changes.

References

There are two broad cases where a JSON-like structure is not quite enough.

1. Editors often need to track the cursor or selection which can be thought of as offsets in the editor text. When changes happen to the text, for example, the offset would need to be updated.
2. Objects often need to refer to other parts of the JSON-tree. For example, one can represent a graph using the array, map primitives with the addition of references. When changes happen, these too would need to be updated.

The refs package implements a set of types that help work with these. In particular, it defines a Container value that allows elements within to refer to other elements.

Network synchronization and server

DOT uses a fairly simple backend Store interface: an append-only dumb log. The Bolt and Postgres implementations are quite simple and other data backends can be easily added.

See Server and Client connection for sample server and client applications. Note that the journal approach used implies that the journal size only increases and so clients will eventually take a while to rebuild their state from the journal. The client API allows snapshotting state to make the rebuilds faster. There is no server support for snapshots though it is possible to build one rather easily

Broad Issues

1. changes.Context/changes.Meta are not fully integrated
2. gob-encoding makes it harder to deal with other languages but JSON encodindg wont work with interfaces.
   • Added sjson encoding as a portable (if verbose) format.
   • The ES6 dotjs package uses this as the native format.
3. Cross-object merging and persisted branches need more platform support
   • Snapshots are somewhat related to this as well.
4. Full rich-text support with collaborative cursors still needs work with references and reference containers.
5. Code generation can infer types from regular go declarations
6. Snapshots and transient states need some sugar.

Contributing

Please see CONTRIBUTING.md.