Toward Lazy Evaluation in a Graph Database

For graph databases deployed in a server setting, high throughput is a critical design goal in performance optimization. With larger data, individual queries incur longer latency, which in turn lowers throughput. This paper explores a novel throughput optimization called in-graph batching empowered by lazy evaluation in data processing. By delaying queries and allowing them to be propagated in the graph, multiple queries can be batched together during propagation, effectively sharing the latency cost. We implement our idea in Neo4j, one of the most widely used graph databases. For graphs consisting of 100 million nodes, our preliminary implementation shows an average increase in throughput of about 282% compared to the default data processing of Neo4j. More encouragingly, the throughput improvement appears to scale with the data size.