BETTING THE COMPANYBETTING THE COMPANY
(LITERALLY) ON A
GRAPH DATABASEGRAPH DATABASE
TIPS, TRICKS, AND LESSONS LEARNED
Aseem Kishore
Jan 2013

START user=node(1), other=node(2)
MATCH (user) -[r1:has|wants]-> (thing) <-[r2:has|wants]- (other)
WHERE TYPE(r1) <> TYPE(r2)
RETURN TYPE(r1), TYPE(r2), thing

Daniel Gasienica
@gasi

SO…
JUST WHAT IS A
GRAPH DATABASEGRAPH DATABASE?

# adjacency list:
nodes = List<Node>
neighbors = Map<Node, List<Node>>
neighbors[node1].add(node2)
# adjacency matrix:
nodes = List<Node>
connections = Map<Node, Map<Node, bool>>
connections[node1][node2] = true

“ By definition, a graph database is any storage
system that provides index-free adjacency. ”
“ This means that every element contains a
direct pointer to its adjacent element and no
index lookups are necessary. ”

QUERYING
1. Start somewhere
2. Traverse elsewhere

QUERYING IN NEO4J
1. Start somewhere
Root node
ID directly (file offset)
Lucene index
2. Traverse elsewhere
Traversal APIs
Cypher patterns
Built-in graph algos (Djikstra, A*, etc.)

NEO4J USAGE
Embedded mode (Java API)
Server mode (REST API)
Cypher query language (both)

OUR USAGE
NODE.JS
+
REST API
+
CYPHER

NEO4J EDITIONS
Community edition
Single instance
Offline backup
Advanced edition
Meh
Enterprise edition
Multi-instance cluster!
Online backup!

NEO4J SCALING
Master-slave replication
Cache-based sharding
Feature-based polyglot'ing
64B limit on nodes, rels, props
But can be easily upped; just flipping some bits
100 props/node (high) ⇒ 640M nodes

OKAY...
LET'S TALK ABOUT
WHAT WE LEARNEDWHAT WE LEARNED

WHAT WE LEARNED
Unique, expressive relationship types

WHAT WE LEARNED
Unique, expressive relationship types
Cache stats where possible

WHAT WE LEARNED
Unique, expressive relationship types
Cache stats where possible
Capture history through event nodes

WHAT WE LEARNED
Unique, expressive relationship types
Cache stats where possible
Capture history through event nodes
First-class objects ⇒ nodes, not rels

WHAT WE LEARNED
Unique, expressive relationship types
Cache stats where possible
First-class objects ⇒ nodes, not rels
Capture history through event nodes
Connected data ⇒ nodes, not props

WHAT WE LEARNED
Unique, expressive relationship types
Cache stats where possible
First-class objects ⇒ nodes, not rels
Capture history through event nodes
Connected data ⇒ nodes, not props
Maintain linked lists for O(1) queries

NEO4J ROADMAP
Overhaul of indexing API
Relationship type grouping
Socket and/or binary protocol
Automatic sharding?

THANKS!
TWITTER: @ASEEMK
GITHUB: @ASEEMK
EMAIL: ASEEM.KISHORE@GMAIL.COM
Questions?