Jiraph is an embedded graph database for Clojure. It is extremely fast and can walk 100,000 edges in about 3 seconds on my laptop. It uses Tokyo Cabinet for backend storage.

For performance and scalability, graphs in Jiraph are multi-layer graphs. Nodes exist on every layer. In this way, node data can be partitioned across all layers (similar to column families in some nosql databases). For our purposes, we'll call the node data on a particular layer a node slice. Edges, on the other hand, can only exist on a single layer. All edges on a specific layer generally correspond in some way. The layer name can be thought of as the edge type, or alternatively, multiple similar edge types can exist on one layer.

Though layers can be used to organize your data, the primary motivation for layers is performance. All data for each layer is stored in a separate data store, which partitions the graph and speeds up walks by allowing them to load only the subset of the graph data they need. This means you should strive to put all graph data needed for a particular walk in one layer. This isn't always possible, but it will improve speed because only one disk read will be required per walk step.

A Jiraph graph is just a clojure map of layer names (keywords) to datatypes that implement the jiraph.layer/Layer protocol.

Every node slice is just a clojure map of attributes. It is conventional to use keywords for the keys, but the values can be arbitrary clojure data structures. Each edge is also a map of attributes. Internally, outgoing edges are stored as a map from node-ids to attributes in the :edges attribute on the corresponding node slice. This way, a node and all its outgoing edges can be loaded with one disk read.

Nodes are not required to have a type, but it is conventional to include the node type in its id if there are multiple types of nodes (e.g. "human-144567", "robot-23131"). Here is a sample node:

{:name      "Justin"
 :nicknames ["Judd" "Huck" "Judd Huck"]
 :edges     {"human-2" {:type :spouse}
             "robot-1" {:type :friend}
             "dog-2"   {:type :pet}}
}

(use 'jiraph.graph)

(def g
  {:foo (jiraph.masai-layer/make "/tmp/foo")
   :bar (jiraph.masai-layer/make "/tmp/bar")
   :baz (jiraph.masai-layer/make "/tmp/baz")})

(with-graph g
  (add-node! :foo "human-1" {:name "Justin"  :edges {"human-2" {:type :spouse}}})
  (add-node! :foo "human-2" {:name "Heather" :edges {"human-1" {:type :spouse}}})

  (get-node :foo "human-1"))
  ;; {:name "Justin", :edges {"human-2" {:type :spouse}}}

(with-graph g
  (add-node! :foo "robot-1" {:name "Bender" :edges {"human-1" {:type :friend}}})
  (append-node! :foo "human-1" {:edges {"robot-1" {:type :friend}}})

  (get-node :foo "human-1"))
  ;; {:name "Justin", :edges {"robot-1" {:type :friend}, "human-2" {:type :spouse}}}

(with-graph g
  (assoc-node! :foo "robot-1" {:designation "Bending Unit 22"})

  (get-node :foo "robot-1"))
  ;; {:edges {"human-1" {:type :friend}}, :name "Bender", :designation "Bending Unit 22"}

(with-graph g
  (update-node! :foo "robot-1" dissoc :designation)

  (get-node :foo "robot-1"))
  ;; {:name "Bender", :edges {"human-1" {:type :friend}}}

You can use at-revision to mark changes with a given revision and rewind the state of the graph back to that revision later.

(with-graph g
  (at-revision 1
    (add-node! :foo "human-2" {:name "Ceruzzi"}))

  (at-revision 2
    (append-node! :foo "human-2" {:name "Hatcher"}))

  (:name (get-node :foo "human-2")) ;; "Hatcher"

  (at-revision 1
    (:name (get-node :foo "human-2"))) ;; "Ceruzzi"

  (:name (get-node :foo "human-2"))) ;; "Hatcher"

You can only use at-revision to rewind a layer's state if the layer was updated using only add-node! and append-node!. All other update operations are destructive, and nodes modified with update-node!, assoc-node! or delete-node! will not exist if you use at revision to go back to before they were modified. To ensure that no destructive operations are permitted on a layer, you can set :append-only in the metadata on the graph (either a set of layer names that are append-only, or true for all layers). Even if a layer is marked append-only, you can still call compact-node! to reduce the storage requirement and remove historical data.

Transactions also behave slightly different inside of at-revision. When with-transaction is complete, it sets the :rev property on current layer to the current-revision. Also only the first transaction on a layer for a given revision will be applied. Subsequent transactions are assumed to be duplicates. This permits cross-layer transactions to be performed by assigning the same revision number to all of them. Then if there is a failure in the middle of a revision, the entire revision can be reapplied and layers that have already been updated will be skipped.

For faster performance, Jiraph supports using protocol buffers for node slices and edge data.

The easiest way to use Jiraph in your project is via Cake. Add the following to the :dependencies key in your project.clj:

[jiraph "0.5.0-SNAPSHOT"]

Using Cake allows you to automatically pull in the native libraries for tokyocabinet and compile protocol buffers if you are using them. Protocol Buffers, if used, should be placed in your project's proto/ directory and can be compiled by running cake proto.