Need for Big Data Technologies: A Review

Teseo and the analysis of structural dynamic graphs

“…In our experiments, we explicitly issued a new level every 10 seconds, a value suggested in the original paper [46]. • GraphOne [35,37]: a library also inspired by a delta design with multiple levels. The read store is logically arranged as an adjacency list, with the change sets maintained at the node granularity.…”

“…Upon inspection, these approaches have been shown to come short in terms of performance [48,50], compared to systems for static graphs, while offering a somewhat more restricted abstraction model. Nowadays, single machines can process relatively large graphs [51], and, recently, for this architecture, several libraries to tackle dynamic graphs have been published [20,35,37,46,63]. These systems, regardless of their physical implementation, all expose a logical view based on adjacency lists, one of the most conventional abstractions presented in algorithmic textbooks [14].…”

“…Regrettably, most of these systems forgo transactional support [20,35,43,46] or incur overhead [63]. Analysis of dynamic graphs is increasingly relevant in sectors such as Finance, Insurance, Logistics, Media and Infrastructure and include applications like Fraud Detection [4], Threat Detection [34], Information Diffusion [24], Risk Analysis, Compliance as well as Supply Chain Optimization, and more [55].…”

“…Through these techniques, Teseo aims to exploit the access patterns of graph analysis, while retaining sustainable throughput in updates. Designs based on trees, and 𝐵 + -trees in particular, are predominant in RDBMSes [56], but they are a novelty with respect to the high-throughput and low-latency systems for dynamic graphs we examined [20,35,37,46,63]. These systems share a common scheme, outlined in Figure 1, ultimately anchored to a vertex table, a container where all vertices are stored together and whose entries refer to lists of their adjacent edges.…”

“…Our results complement and, at first glance, can conflict with the experimental results reported in previous work. As we detail in Section 8, in 3 /4 of our competitors [20,35,46] updates have only been evaluated as bulk loads, while, generally, different, but, occasionally incompatible implementations of the same graph algorithms have been compared. Our experiments, instead, target fine-grained updates, whereas all systems are evaluated in the LDBC Graphalytics standard benchmark [29], under the same implementation for the graph algorithms.…”