assemblage_db

Crates.ioassemblage_db
lib.rsassemblage_db
version0.1.0
sourcesrc
created_at2021-08-18 18:11:32.952129
updated_at2021-08-18 18:11:32.952129
descriptionDistributed Document/Graph DB for Connected and Overlapping Pages
homepage
repositoryhttps://github.com/fkettelhoit/assemblagedb/tree/main/assemblage_db
max_upload_size
id439178
size251,443
Frederic Kettelhoit (fkettelhoit)

documentation

README

Distributed Document/Graph DB for Connected Pages

AssemblageDB is a transactional high-level database for connected webs of pages, notes, texts and other media. Think of it like a personal web, but easily editable, with more connections and better navigation than the web. It is high-level in the sense that it defines a document model similar to HTML but vastly simpler and with graph-like 2-way links instead of tree-like 1-way jump links. The data model is both:

  • document-oriented: supports nested documents without a fixed schema
  • graph-based: documents can have multiple parents and form a directed, possibly cyclic graph

Features

  • versioned: old values remain accessible until merged
  • transactional: snapshots are isolated through MVCC
  • storage-agnostic: supports native (files) and wasm (IndexedDB) targets
  • indexed: maintains an automatic index for similarity/overlap search
  • distributed: nodes can be published/subscribed as remote broadcasts

Data Model

Nodes in an AssemblageDB can be either atomic (a line of text for example) or nested, either in a list containing multiple children or a styled node containing just a single child. List nodes have a layout, which controls how children are laid out in relation to each other, while Styled nodes have zero or more block styles or span styles, that control how their (possibly nested) child is displayed. Examples for layouts and styles are:

  • Layout::Chain: lays out children as a consecutive chain of inline spans. With 2 text children "foo" and "bar", the chain would be displayed as "foobar".
  • Layout::Page: lays out children as blocks on a page, separated vertically by a new line. With 2 text children "foo" and "bar", the page would be displayed as 2 lines, the first line containing "foo", the second line containing "bar".
  • SpanStyle::Italic: A span (inline) style that would display the child "foo" as "foo"
  • BlockStyle::Heading: A block style that would display the child "foo" in its own block with a larger font size.

A node is always either a span or a block. Text nodes are considered to be spans by default and remain spans if styled using span styles such as SpanStyle::Italic or SpanStyle::Bold. However, a single block style (such as BlockStyle::Heading) in a set of styles is always "contagious" and turns a text node "foo" styled with both SpanStyle::Italic and BlockStyle::Heading into a block. Similarly, layouts control whether a list is displayed as a span or a block: Layout::Chain turns a list into a span, while Layout::Page turns a list into a sequence of blocks.

Obligatory Warning

AssemblageDB is still in an experimental stage. It should go without saying that it is not a battle-tested production-ready database and could at any time eat all of your data. If you need to persist production data, use a real database instead.

Example

let storage = FileStorage::open("db1").await?;
let db = Db::open(storage).await?;

// Nodes support layouts and styles, for example as a page of blocks...
let page1_id = tx!(|db| {
    db.add(Node::list(
        Layout::Page,
        vec![
            Node::styled(BlockStyle::Heading, Node::text("A Heading!")),
            Node::text("This is the first paragraph."),
            Node::text("Unsurprisingly this is the second one..."),
        ],
    ))
    .await?
});

// ...or as inline spans that are chained together:
let page2_id = tx!(|db| {
    db.add(Node::list(
        Layout::Page,
        vec![Node::list(
            Layout::Chain,
            vec![
                Node::text("And this is the "),
                Node::styled(SpanStyle::Italic, Node::text("last")),
                Node::text(" paragraph..."),
            ],
        )],
    ))
    .await?
});

// Documents can form a graph, with nodes keeping track of all parents:
tx!(|db| {
    db.add(Node::list(Layout::Page, vec![page1_id, page1_id])).await?;
    assert_eq!(db.parents(page1_id).await?.len(), 2);
    assert_eq!(db.parents(page2_id).await?.len(), 0);
});

// All text is indexed, the DB supports "overlap" similarity search:
tx!(|db| {
    let paragraph1_id = db.get(page1_id).await?.unwrap().children()[1].id()?;
    let paragraph3_id = db.get(page2_id).await?.unwrap().children()[0].id()?;
    let overlaps_of_p1 = db.overlaps(paragraph1_id).await?;
    assert_eq!(overlaps_of_p1.len(), 1);
    assert_eq!(overlaps_of_p1[0].id, paragraph3_id);
    assert!(overlaps_of_p1[0].score() > 0.5);
});

// Nodes (with all their descendants) can be published globally...
let broadcast = tx!(|db| { db.publish_broadcast(page1_id).await? });

// ...and the broadcast can then be fetched in another remote DB:
let other_storage = FileStorage::open("db2").await?;
let db2 = Db::open(other_storage).await?;
tx!(|db2| {
    let paragraph_id = db2
        .add(Node::text("This is the first paragraph, right?"))
        .await?;
    db2.add(Node::list(Layout::Page, vec![paragraph_id]))
        .await?;
    // The DB is empty except for this paragraph, so no overlaps:
    assert_eq!(db2.overlaps(paragraph_id).await?.len(), 0);
    // But when the broadcast paragraph is fetched, there is an overlap:
    db2.fetch_broadcast(&broadcast.broadcast_id).await?;
    assert_eq!(db2.overlaps(paragraph_id).await?.len(), 1);
});
Commit count: 0

cargo fmt