The document presents a performance comparison of native multi-model databases: ArangoDB, MongoDB, and Neo4j, using a dataset from a Slovakian social network. It outlines the test setup, including data characteristics, use cases like single reads/writes and aggregations, and benchmark criteria, emphasizing that ArangoDB serves as the baseline for performance metrics. It encourages readers to conduct their own benchmarks rather than solely trusting the presented results.

1. Native multi-model
is competitive
Performance comparison between
ArangoDB, MongoDB and Neo4j
Claudius Weinberger (ArangoDB)
June, 2015

2. Native Multi-Model DB
“A native multi-model database is from my perspective a document
store (JSON documents), a key/value store and a graph database, all
in one database engine and with a unifying query language and API
that covers all three data models and even allows to mix them in a
single query.” — Claudius Weinberger

3. Data-Set
Our test data are a snapshot from a social
network in Slovakia and provided by the
Stanford University SNAP collection. It
contains 1 632 803 vertices describing
people and 30 622 564 edges describing
their friendship relation. We use the vertex
data for the document tests and the combined
vertex and edge data for the graph tests. We
wanted to study the standard client/server
setup rather than embedding the database
into the application, because this model is
used more widely and therefore more relevant
in practice. We used JavaScript/node.js as
client language and environment

4. Use-Cases
• single read: single document reads of profiles (100 000 documents)
• single write: single document writes of profiles (100 000 documents)
• aggregation: aggregation over a single collection (1 632 803 documents)
Here, we compute statistics about the age distribution for everyone in the network,
simply counting which age occurs how often.
• neighbors: finding direct neighbors plus the neighbors of the neighbors (for 500
vertices)
• shortest path: finding 19 shortest paths (in a highly connected social graph)
This answers the question how close to each other two people are in the social
network.

5. The throughput measurements on the test machine for ArangoDB define the baseline (100%)
for the comparisons. Lower percentages point to higher throughput - so less is better.

6. Test Setup
• Hardware:
• virtual machine of type n1-standard-16 in Google Compute Engine with 16 virtual cores (on these, a virtual
core is implemented as a single hardware hyper-thread on a 2.3 GHz Intel Xeon E5 v3) and altogether 60
GB of RAM.
• the client was an n1-standard-8 (8 vCPU, 30 GB RAM) in the same network.
• Software:
• ArangoDB V2.6.0 alpha3 (pre-release) for x86_64
driver: arangojs in version 3.8.0
• MongoDB V3.0.3 for x86_64, using the WiredTiger storage engine
driver: mongodb in v2.0.33, which builds on top of mongodb-core in v1.1.32
• Neo4j Enterprise 2.3 SNAPSHOT running on JDK 1.7.0_79
driver: node-neo4j v2.0.0 RC1
more details in the blog post

7. feedback welcome
Don’t trust benchmarks. Make your own.
github.com/weinberger/nosql-test