Aerospike Inc., maker of a highly scalable NoSQL database, today is entering the graph database market with an offering that it claims can outperform and outscale offerings from market leaders such as Neo4j Inc. and TigerGraph Inc.

Based on the Apache TinkerPop open-source graph framework that is also the basis of Amazon Web Services Inc.’s Neptune graph database, Aerospike’s offering supports the Gremlin query language rather than Neo4J’s Cypher or TigerGraph’s Graph SQL. Support for the open-source implementation of Cypher is planned for the first quarter of next year. It also uses Aerospike’s key-value storage underlying a graph data model rather than native graph storage.

Gremlin is prized for its flexibility, portability and scalability, although it has a steep learning curve and a limited ecosystem of third-party tools. Aerospike said the underlying architecture of its new graph offering permits it to surpass its competitors in both performance and scale, delivering millisecond multihop graph queries across trillions of vertices and edges. The company says benchmarks indicate a throughput of more than 100,000 queries per second with a latency of less than five milliseconds running on as little as one-10th as much hardware infrastructure as competitors.

Hybrid memory architecture

Aerospike’s claim to fame is what it calls a “hybrid memory architecture” that enables each node or server to treat a bank of solid-state storage devices as a type of extended memory. Frequently accessed data is stored in memory movement between memory and disk is done automatically without accessing the underlying file or operating systems.

The approach doesn’t use caching but finely tuned clustering algorithms and networking paths. “We auto-partition the whole data set and can go from the client to the data in a single hop because the clients are first-class participants in a roster-based clustering algorithm as opposed to a quorum-based clustering algorithm,” said Chief Product Officer Lenley Hensarling.

Roster-based clustering organizes data points into clusters based on the concept of rosters and is widely used in data mining, pattern recognition and machine learning to identify natural groupings or clusters within a dataset.

Horizontal scale

Aerospike is also designed to scale horizontally across multiple nodes with the capability to manage data distributed across a cluster of machines for high throughput and low latency. It also provides configurable durability options that let users to control the level of data persistence and replication.

“Neo4j is a great product with great tooling but it doesn’t scale out,” Hensarling said. “We can support graphs with billions of vertices, each with thousands of edges and do two- to four-hop queries on the order of eight milliseconds per query.”

The company’s core markets are advertising technology and financial services and it expects to initially focus on those areas of strength.

“We have seen a lot of demand from the digital marketing and the ad tech world driven primarily by streaming video,” Hensarling said. For example, graph technology can be used to identify multiple people who are watching a streaming video in the same room by analyzing the device usage patterns of each.

“With graph technology, you can understand who’s in the proximity, what people are watching, what they’re all doing on their phones, and then use the graph to pull that together to determine what ads they should be shown,” he said.

Aerospike said its graph product can independently scale up computing and storage, enabling users to pay only for the infrastructure they need. The company didn’t announce pricing, but Hensarling said it will be based on a combination of storage volume and the number of virtual CPUs used. It will also offer pricing that varies by month to accommodate customers and businesses such as retail and media that experience spikes in demand.

Image: Wikipedia