Graph Databases don’t like shards.
Issues
- if using RDMS: the joins are really expensive, not efficient for relations. We often do traversal or reverse traversal in relations.
- To solve that, we use index-free adjacency
Ingredients
nodes, edges, properties, labels
Graph databases
- property graph: Neo4j
- triple stores(RDF): subject, property, object.
Abbreviation
IRI: international resource identifier
Syntax
RDF/XML
- subject
<rdf:Description rdf:about="subject-IRI"> - property: within the subject tags:
<rdf:...> <geo:property1> ...</geo:property1></rdf:...> - object: within the property tags:
<geo:property1 rdf:resource="object-IRI"/>or<geo:property2>object-value<geo:property2>JSON-LD
{"@id":subject, "rdf:type":subject-type,"property1":"object1","property2":"object2"}Turtle
@prefix sub:IRI. @prefix object:IRI. @prefix prop:IRI sub:self prop:sub-prop object:sub-obj ...
Querying
Cypher
Format: (node1)-[:edge-label1]->(node2)-[:edge-label2]<-(node3)
- anchoring a label:
(node1:label1) - filtering a property:
(node1 {prop-key: 'prop-value'} - combining:
(node1: label1 {prop-key: 'prop-value'} - variable repetition:
(node1)-[:edge-label1]->(node2)-[:edge-label2]->(node1) - variable length path:
(alpha)-[*1..4]->(beta) - MATCH clause:
MATCH (one-query) (WHERE CONDIONS) RETURN gamma - CREATE clause: use , to seperate
CREATE (),(),()
SPARQL
Architecture
No shards
Master-slave: Master has the entire graph, in the slaves, only have copies. Data replication to avoid data loss(synchronization), to improve performance of scalability(everybody gets to connect to master/slave).
Write
how to guarantee the consistency?
- write to the master
- or to a slave. It is blocked until it makes sure that the master is up-to-date.
Hardware
Fixed-size records: serialize the nodes/edges/labels/properties.
- properties storage: save key-value pairs
- relationship storage:
- double links for free iteration
- from an edge view: a pointer to the source(target) node, a pointer to the s/t-previous edge, a pointer to the s/t-next edge.
- typical size:
- node: 9 bytes
- relationship: 33 bytes
- relationship name: 5 bytes
- property: 33 bytes
Further Reading
Graph Databases, 2nd Edition Chapter 1, 2, 3, 4, 6,
