#sync

melda

Melda is a Delta-State JSON CRDT

12 releases

0.2.3 Aug 11, 2023
0.2.2 Aug 11, 2023
0.2.1 Mar 13, 2023
0.1.15 Jun 21, 2022

#62 in #sync


Used in libmelda-tools

GPL-3.0 license

260KB
4K SLoC

What is Melda?

Melda is a Delta-State JSON CRDT. CRDTs, which stand for Conflict-free Replicated Data Types, are data structures which can be replicated (copied) across multiple computers in a network. Each replica can be individually and concurrently updated without the need for central coordination or synchronization. Updates made on each replica can be merged at any time.

There exist different types of CRDTs: operation-based CRDTs (which generate and exchange update operations between replicas), state-based CRDTS (which exchange and merge the full state of each replica) and delta-state CRDT, such as Melda, (which exchange only the differences between versions, or states, of the data type).

Melda natively supports the JSON data format and provides a way to synchronize changes made to arbitrary JSON documents. You can work with Melda CRDTs either using this Rust library or using a command line tool. In the Kibi w/Melda repository you can find a fork of the original Kibi text editor with collaboration features implemented using Melda.

How do I use Melda?

The following guide is also availbe in examples. You can run it with cargo run --examples simple.

First of all, in Cargo.toml add the following dependency

melda = { git = "https://github.com/slashdotted/libmelda" }

or

melda = "0.2.3"

If using the crate from crates.io adapt the version string as needed. Then import the required modules. For this example you will need:

use melda::{filesystemadapter::FilesystemAdapter, melda::Melda};
use serde_json::json;
use std::sync::{Arc, RwLock};

To understand how to use Melda we consider the following situation, where a shared JSON document used by a fictitious activity planning software (i.e. a todo management software) is concurrently updated by multiple parties. The provided JSON is generated by the application (by serializing its data model). We assume that user Alice creates the first version of the shared JSON document, which will be named v1_alice.json. This first version contains the following data:

{
	"software" : "MeldaDo",
	"version" : "1.0.0",
	"items♭" : []
}

The root object contains three fields: a software field which defines the name of the application, a version field, which sets the version of the software, and an items♭ field, which maps to an array of JSON objects (one for each todo). Since this is the first version, the array of items is empty. The suffix is used to ask Melda to flatten the contents of the array, by extracting the contained JSON objects in order to keep track of their changes individually.

To better understand the purpose of the flattening procedure, consider how Melda processes the following two JSON files. The first one, named v2_alice_noflat.json contains:

{
	"software" : "MeldaDo",
	"version" : "1.0.0",
	"items" : [
	   {"_id" : "alice_todo_01", "title" : "Buy milk", "description" : "Go to the grocery store"}
	]
}

In this case, Melda will keep the root object as is, and the changes made to the items array by one user will not merge with changes made by other users. So, for example, if two users add an element to the array on their replica and later merge those replicas, only one of the elements will be visible. On the contrary, consider now another version of the document, named v2_alice.json, which contains:

{
	"software" : "MeldaDo",
	"version" : "1.0.0",
	"items♭" : [
	   {"_id" : "alice_todo_01", "title" : "Buy milk", "description" : "Go to the grocery store"}
	]
}

In this case the object within the items♭ array will be extracted and tracked individually. In particular, two JSON objects results from the above document:

{
	"_id" : "",
	"software" : "MeldaDo",
	"version" : "1.0.0",
	"items♭" : [
	  "alice_todo_01"
	]
}

And the todo item itself:

{
	"_id" : "alice_todo_01",
	"title" : "Buy milk",
	"description" : "Go to the grocery store"
}

Please notice that each object has its own unique identifier stored in the _id field. If an identifier is not provided by the client application, Melda will auto-generate one. The root object is always identified by (this identifier cannot be changed by the client application). Since each object of the items♭ array is tracked individually, if an user adds an element to the array and later merges his/her replica with another user all changes will be preserved.

If the collection of items becomes too large we can ask Melda to only store difference arrays between the newest revision of the document and the previous one. For that we simply need to prefix the key of the items field with the Δ character (greek capital letter delta). Version delta_alice.json might therefore become:

{
	"software" : "MeldaDo",
	"version" : "1.0.0",
	"Δitems♭" : [
	   {"_id" : "alice_todo_01", "title" : "Buy milk", "description" : "Go to the grocery store"}
	]
}

To keep things simple, in the following we will not use difference arrays. Let's go back to our example situation... Up until this point we only considered some JSON data, but we have yet to see how we can interact with Melda in order to update the data structure.

Adapters

Melda implements a modular design where the logic of the CRDT is separated from the data storage. Storing the data (in our case, delta states) is achieved by means of Adapters. Melda already provides different types of adapters, supporting in-memory storage (MemoryAdapter), a folder in the filesystem (FilesystemAdapter), a SQLite database (SQLiteAdapter), and a Solid Pod (SolidAdapter). Furthermore, it is possible to use a meta-adapter to compress data using the Flate2 algorithm (Flate2Adapter): other adapters can be composed with the Flate2Adapter to store compressed data on the chosen backend.

We can initialize an adapter that will store data on the filesystem (in the todolist directory) as follows (FilesystemAdapter):

let adapter = Box::new(FilesystemAdapter::new("todolist").expect("Cannot initialize adapter"));

If we want to used compression we would add the Flate2Adapter as follows:

let adapter = Box::new(Flate2Adapter::new(Arc::new(RwLock::new(Box::new(
            FilesystemAdapter::new("todolist").expect("Cannot initialize adapter"))))));

Alternatively you can use the get_adapter function to initialize an adapter from an Url:

let adapter = get_adapter("file+flate://todolist").unwrap();

Valid schemes for the get_adapter function are:

Storage type Example path Description
In memory (memory://) memory://
In memory w/Deflate compression (memory+flate://) memory+flate://
In memory w/Brotli compression (memory+brotli://) memory+brotli://
Folder (file://) file://mycrdtdocument The absolute path of a folder (can be on a network share)
Folder w/Deflate compression (file+flate://) file+flate://mycrdtdocument The absolute path of a folder (can be on a network share)
Folder w/Brotli compression (file+brotli://) file+brotli://mycrdtdocument The absolute path of a folder (can be on a network share)
Solid Pod (solid://) solid://anuser.solidcommunity.net/mycrdtdocument The URL of a Solid Pod
Solid Pod w/Deflate compression (solid+flate://) solid+flate://anuser.solidcommunity.net/mycrdtdocument The URL of a Solid Pod
Solid Pod w/Brotli compression (solid+brotli://) solid+brotli://anuser.solidcommunity.net/mycrdtdocument The URL of a Solid Pod
SQLite (sqlite://) sqlite://mycrdtdocument The name of the database is required (use :memory: for in-memory storage)
SQLite w/Deflate compression (sqlite+flate://) sqlite+flate://mycrdtdocument The name of the database is required (use :memory: for in-memory storage)
SQLite w/Brotli compression (sqlite+brotli://) sqlite+brotli://mycrdtdocument The name of the database is required (use :memory: for in-memory storage)

For Solid Pod's access, a username and a password are required.

Initializing Melda

To initialize Melda we use the new method, passing the chosen adapter:

let mut m = Melda::new(Arc::new(RwLock::new(adapter))).expect("Failed to inizialize Melda");

or you can use an Url

let mut m = Melda::new_from_url("file+flate://todolist").expect("Failed to inizialize Melda");

Please note that we can remove the mut modifier if we only intend to read the CRDT.

Updating the CRDT

In order to update the state of the CRDT we use the update method. First we need to parse the JSON data into a JSON value: since we use serde_json we call serde_json::from_str or the json! macro. Subsequently we call the update method on the resulting object:

let v = json!({ "software" : "MeldaDo", "version" : "1.0.0", "items♭" : []})
        .as_object()
        .expect("Not an object")
        .clone();
m.update(v).expect("Failed to update");

Updates made to the CRDT are now staged. In order to persist them we need to commit to the data storage backend. We commit using the commit method: we can pass an optional JSON object containing some additional information that will be stored along with the updates. Please note that it is possible to perform as many updates as needed before commiting, however it is not possible to commit if no updates have been made to the CRDT.

let info = json!({ "author" : "Alice", "description" : "First commit" })
	.as_object()
	.expect("Not an object")
	.clone();
let commit_result = m.commit(Some(info));

The result of the commit is either an error, None if there were no changes to be committed or Some(String) if changes were committed: the string is the commit identifier. Upon success, on disk (in the todolist directory) the following content should have been created:

todolist/
├── 49
│   └── 49ccea4d5797250208edf9bc5d0b89edf23c30a61f5cb3fafb87069f07276a62.delta
└── b4
    └── b4e50e445542c4737f4cfd7a9193ffd3be3794049d361d114a44f36434257cb3.pack

The .delta file is called delta block, and contains the versioning information of each object in the CRDT, whereas the .pack file is the data pack which stores the actual JSON content of each object. Each commit produces a new delta block (with a different name, which corresponds to the hash digest of its content) and possibly a data pack (if new JSON values are produced). The directory structure of the todolist directory organizes files into sub-directories according to their prefix.

We can perform another update using (again) the update method and commit the resulting changes:

let v = json!({ "software" : "MeldaDo", "version" : "1.0.0", "items♭" : [
       {"_id" : "alice_todo_01", "title" : "Buy milk", "description" : "Go to the grocery store"}
    ]
    })
    .as_object()
    .expect("Not an object")
    .clone();
m.update(v).expect("Failed to update");
let info = json!({ "author" : "Alice", "description" : "Add buy milk" })
        .as_object()
        .expect("Not an object")
        .clone();
let commit_result = m.commit(Some(info));

The changes will reflect on disk (with new packs and blocks created in the corresponding directories):

todolist/
├── 2b
│   └── 2b0a463fcba92d5cf7dae531a5c40b67aaa0f45ab351c15613534fb5bba28564.pack
├── 49
│   └── 49ccea4d5797250208edf9bc5d0b89edf23c30a61f5cb3fafb87069f07276a62.delta
├── b4
│   └── b4e50e445542c4737f4cfd7a9193ffd3be3794049d361d114a44f36434257cb3.pack
└── b6
    └── b6297035f06f13186160577099759dea843addcd1fbd05d24da87d9ac071da3b.delta

Reading the data

At any time it is possible to read the state of the CRDT back into a JSON document using the read method:

let data = m.read().expect("Failed to read");
let content = serde_json::to_string(&data).unwrap();
println!("{}", content);

This additional code will print the following on the terminal:

{"_id":"","items♭":
    [
        {"_id":"alice_todo_01","description":"Go to the grocery store","title":"Buy milk"}
    ],
    "software":"MeldaDo","version":"1.0.0"
}

Each object managed by Melda will contain the _id field with the corresponding unique identifier.

Sharing data

We now suppose that Alice shares the current state of the todolist directory with Bob (she can simply zip the contents and send the compressed file by e-mail to Bob). We assume that Bob saves the contents in the todolist_bob directory. Bob initializes Melda and can perform some updates:

let adapter_bob =
        Box::new(FilesystemAdapter::new("todolist_bob").expect("Cannot initialize adapter"));
    let mut m_bob =
        Melda::new(Arc::new(RwLock::new(adapter_bob))).expect("Failed to inizialize Melda");
let v = json!({ "software" : "MeldaDo", "version" : "1.0.0", "items♭" : [
       {"_id" : "alice_todo_01", "title" : "Buy milk", "description" : "Go to the grocery store"},
       {"_id" : "bob_todo_01", "title" : "Pay bills", "description" : "Withdraw 500 to pay bill"},
       {"_id" : "bob_todo_02", "title" : "Call mom", "description" : "Call mom to schedule dinner"},
    ]
    })
    .as_object()
    .expect("Not an object")
    .clone();
m_bob.update(v).expect("Failed to update");
let info = json!({ "author" : "Bob", "description" : "Add some todos" })
        .as_object()
        .expect("Not an object")
        .clone();
let commit_result = m_bob.commit(Some(info));

As you might notice, two new items have been added by Bob. In the meantime, Alice continues to work on her replica, by removing one item (alice_todo_01) and adding a new item (alice_todo_02):

let v = json!({ "software" : "MeldaDo", "version" : "1.0.0", "items♭" : [
        {"_id" : "alice_todo_02", "title" : "Take picture of our dog", "description" : "It must be a nice one"}
     ]
     })
     .as_object()
     .expect("Not an object")
     .clone();
    m.update(v).expect("Failed to update");
let info = json!({ "author" : "Alice", "description" : "Some more stuff to do" })
        .as_object()
        .expect("Not an object")
        .clone();
let commit_result = m.commit(Some(info));

Finally, Bob shares his own copy with Alice: now Alice simply needs to merge the content of the directory (as received from Bob) with the local directory. Alternatively, suppose that the data modified by Bob is in the todolist_bob directory on Alice's computer. To merge changes back into the todolist directory, Alice can use the meld method:

let adapter_bob = Box::new(FilesystemAdapter::new("todolist_bob").expect("Cannot initialize adapter"));
let m_bob = Melda::new(Arc::new(RwLock::new(adapter_bob))).expect("Failed to inizialize Melda");
m.meld(&m_bob).expect("Failed to meld");
m.refresh();

The refresh method is used to load updates from the storage backend after the meld operation. Alice can then read the new state of the CRDT with:

let data = m.read().expect("Failed to read");
let content = serde_json::to_string(&data).unwrap();
println!("{}", content);

The result, printed on the terminal should look like:

{
  "_id": "",
  "items♭": [
    {
      "_id": "bob_todo_01",
      "description": "Withdraw 500 to pay bill",
      "title": "Pay bills"
    },
    {
      "_id": "bob_todo_02",
      "description": "Call mom to schedule dinner",
      "title": "Call mom"
    },
    {
      "_id": "alice_todo_02",
      "description": "It must be a nice one",
      "title": "Take picture of our dog"
    }
  ],
  "software": "MeldaDo",
  "version": "1.0.0"
}

As you can see, there is only one todo from Alice, as well as the two todos added by Bob.

Getting the commit history

When we commit to the CRDT a new delta block is created. Each block is linked to some parent block, so as to create a chain of blocks. Upon commit Melda looks for anchor blocks, which are the ones that are currently not referenced as parent by any other block. We can get the set of current anchors using get_anchors, so if Alice wants to get the anchors for her CRDT she can use:

let anchors = m.get_anchors();

When we commit, the anchor list will contain just the identifier of the committed block, whereas when we meld from another replica we could end up with multiple anchors (which correspond to different branches in the commit tree). We can obtain additional information about each block using the get_block method. If Alice wants to fetch additional information about the anchors, she can run the following code:

let anchors = m.get_anchors();
for block_id in anchors {
	if let Some(block) = m.get_block(&block_id).expect("Failed to get block") {
	    let parents = block.parents;
	    let info = block.info;
	    let packs = block.packs;
	    println!("Block {}", block_id);
	    println!("\t Information: {:?}", info);
	    println!("\t Parents: {:?}", parents);
	    println!("\t Packs: {:?}", packs);
	}
}

The parents field contains the identifiers of parent blocks (it is None if there are no parents, i.e. we are at an origin block), whereas the optional info field corresponds to the commit information. Finally, the packs field contains the identifiers of the data packs generated during the commit (it is set to None if no new data was produced for that commit).

Going back in time

It is possible to navigate through commits by means of the reload_until method. As an example, suppose that Alice wants to go back to the origin:

 let mut anchors: Vec<String> = m.get_anchors().into_iter().collect();
    let mut block_id = anchors.get(0).cloned();
    while block_id.is_some() {
        let block = m
            .get_block(block_id.as_ref().unwrap())
            .expect("Failed to get block")
            .unwrap();
        if let Some(parents) = block.parents {
            let parents: Vec<String> = parents.into_iter().collect();
            block_id = parents.get(0).cloned();
        } else {
            // We reached the origin
            break;
        }
    }
    m.reload_until(block_id.as_ref().unwrap())
        .expect("Failed to reload until origin");

As expected, reading the JSON document now shows the first version:

{"_id":"","items♭":[],"software":"MeldaDo","version":"1.0.0"}

Alice can go back to the last version by reloading the CRDT:

m.reload();

Conflicts and resolution

When we meld two replicas that had some concurrent updates it is likely that conflicts arise. In our scenario, the root document (with identifier √) has a conflict since both Alice and Bob modified the todo items on their own replica (which was later merged). It is possible to check for conflicts by means of the in_conflict method, which returns a set of all objects with conflicts. If an object has a conflict we can retrieve conflicting revisions with the get_conflicting method. Melda chooses a winning revision using a deterministic algorithm. We can get the currently winning revision using the get_winner method. Finally we can view the value associated with a winner or a conflicting revision using the get_value method. Accordingly, Alice can get a summary of conflicting objects and the corresponding values with something like:

for uuid in  m.in_conflict() {
	let winner = m.get_winner(&uuid).unwrap();
	let conflicting = m.get_conflicting(&uuid).unwrap();
	println!("Winner: {:?} -> {:?}", winner, m.get_value(&uuid, &winner));
	for c in conflicting {
	    println!("Conflict {:?}", m.get_value(&uuid, &c));
	}
}

We can resolve a conflict using the resolve_as method. For example, if Alice wants to accept the current winner and resolve all conflicts, she can use the following code:

    for uuid in  m.in_conflict() {
        let winner = m.get_winner(&uuid).unwrap();
        m.resolve_as(&uuid, &winner).expect("Failed to resolve");
    }
    assert!(m.in_conflict().is_empty());

Benchmarks

In the libmelda-benchmarks repository you will find a benchmark comparing Melda to Automerge

Example integration

In the Kibi repository you will find an example of integration of Melda into a text-editor. There is also another project libmelda-tools which implements a simple command line tool to update, read, and meld Melda structures.

Publications

2022

Amos Brocco "Melda: A General Purpose Delta State JSON CRDT". 9th Workshop on Principles and Practice of Consistency for Distributed Data (PaPoC'22). April 2022. PDF BibTex

2021

Amos Brocco "Delta-State JSON CRDT: Putting Collaboration on Solid Ground". (Brief announcement). 23rd International Symposium on Stabilization, Safety, and Security of Distributed Systems (SSS 2021). November 2021. PDF BibTex

Contact

amos dot brocco at supsi dot ch

License

(c)2021-2022 Amos Brocco, GPL v3 (for now... but I will evaluate a change of license - to something like BSD3/MIT/... in the near future)

Acknowledgements

Many thanks to @ngortheone (documentation updates, examples)

Dependencies

~9–48MB
~849K SLoC