Why Titan Outperforms Traditional RDBMS for Complex Graph Queries

Background

Relational databases model data with tables, rows, and foreign keys, but they become inefficient when handling highly complex many‑to‑many relationships, often requiring costly joins and producing massive intermediate results.

Graph Database Comparison

Graph databases store data as vertices and edges, giving them a natural advantage for knowledge graphs and social network analysis. Three popular options are compared:

OrientDB – easy to use, active community.

Neo4j – simple but not truly distributed; struggles with data beyond a single machine.

Titan – supports HBase or Cassandra as storage, integrates with Hadoop, and can run OLTP and Spark‑based OLAP workloads.

Based on these criteria, the team selected Titan.

Titan Architecture

Titan’s architecture consists of modular, open‑source components:

Pluggable storage (HBase/Cassandra) for virtually unlimited capacity.

External indexing plugins (Elasticsearch, Solr, Lucene) for non‑equality queries.

Management API for schema and instance management.

TinkerPop API for graph operations.

Internal layers that translate high‑level graph commands into storage‑specific actions (e.g., HBase put/get/scan).

GraphComputer for Spark or MapReduce‑based OLAP, such as PageRank calculations.

Data Model

Titan represents a graph with three element types:

Vertex : an entity identified by a unique Vertex ID and distinguished by labels (e.g., Titan, Location).

Edge : a directed relationship between an out‑vertex and an in‑vertex, also labeled (e.g., father, lives) and uniquely identified by an Edge ID.

Property : key‑value pairs attached to vertices or edges; each property has a key, optional cardinality, and a unique Property ID. Indexes can be defined on properties to speed up lookups.

Gremlin Query Examples

Titan exposes the Gremlin language via a Gremlin Server. Example commands:

# Create a cluster

gremlin> graph = TitanFactory.open('conf/titan-hbase.properties')

# Find a vertex named 'saturn'

gremlin> saturn = g.V().has('name', 'saturn').next()

# Retrieve its properties

gremlin> g.V(saturn).valueMap()

# Traverse to the grandfather's name

gremlin> g.V(saturn).in('father').in('father').values('name')

Storage Layout

Titan stores vertices and edges using an adjacency‑list model that follows the BigTable data model. Each row key is a Vertex ID; cells contain either vertex properties or edge information. Edge cells encode direction, ordering attributes, adjacent vertex IDs, and edge IDs, while property cells store type IDs and values.

Paipaidai Use Case

At Paipaidai, a heterogeneous network of over 1 billion vertices and 500 billion edges was built in Titan on HBase, representing users, devices, and social relationships. This graph supports fraud detection and user‑association analysis, enabling queries such as:

Find all users linked to a given user.

Identify features of users associated with a target.

Trace the connection path between two users.

The system can explore up to six degrees of separation, reducing investigation effort and improving overall efficiency by more than 25%.