The first part describes query answering over data subject to RDFS constraints, stored in relational data management systems. The implicit information resulting from RDF reasoning is required to correctly answer such queries. We introduce the database fragment of RDF, going beyond the expressive power of previously studied fragments. We devise novel techniques for answering Basic Graph Pattern queries within this fragment, exploring the two established approaches for handling RDF semantics, namely graph saturation and query reformulation. In particular, we consider graph updates within each approach and propose a method for incrementally maintaining the saturation. We experimentally study the performance trade-offs of our techniques, which can be deployed on top of any relational data management engine.

The second part considers the new requirements for data analytics tools and methods emerging from the development of the Semantic Web. We fully redesign, from the bottom up, core data analytics concepts and tools in the context of RDF data. We propose the first complete formal framework for warehouse-style RDF analytics. Notably, we define analytical schemas tailored to heterogeneous, semantic-rich RDF graphs, analytical queries which (beyond relational cubes) allow flexible querying of the data and the schema as well as powerful aggregation and OLAP-style operations. Experiments on a fully-implemented platform demonstrate the practical interest of our approach.

There has recently been an increase in the number of RDF knowledge bases published on the Internet. These rich RDF data sets can be useful in answering many queries, but much more interesting queries can be answered by integrating information from different data sets. This has given rise to research on automatically linking different RDF data sets representing different knowledge bases. This is challenging due to their scale and semantic heterogeneity. Various approaches have been proposed, but there is room for improving the quality of the generated links.

In this paper, we present ALEX, a system that aims at improving the quality of links between RDF data sets by using feedback provided by users on the answers to linked data queries. ALEX starts with a set of candidate links obtained using any automatic linking algorithm. ALEX utilizes user feedback to discover new links that did not exist in the set of candidate links while preserving link precision. ALEX discovers these new links by finding links that are similar to a link approved by the user through feedback on queries. ALEX uses a Monte-Carlo reinforcement learning method to learn how to explore in the space of possible links around a given link. Our experiments on real-world data sets show that ALEX is efficient and significantly improves the quality of links.

Global Web applications face the problem of high network latency due to their need to communicate with distant data centers. Many applications use edge networks for caching images, CSS, javascript, and other static content in order to avoid some of this network latency. However, for updates and for anything other than static content, communication with the data center is still required, and can dominate application request latencies. One way to address this problem is to push more of the web application, as well the database on which it depends, from the remote data center towards the edge of the network. In this paper, we present preliminary work in this direction. Specifically, we present an edge-aware dynamic data replication architecture for relational database systems supporting web applications. Our objective is to allow dynamic content to be served from the edge of the network, with low latency.

This will be a short practice talk.

The considerable interest in distributed systems that can execute algorithms to process large graphs has led to the creation of many graph processing systems. However, existing systems suffer from two major issues: (1) poor performance due to frequent global synchronization barriers and limited scalability; and (2) lack of support for graph algorithms that require serializability, the guarantee that parallel executions of an algorithm produce the same results as some serial execution of that algorithm.

Many graph processing systems use the bulk synchronous parallel (BSP) model, which allows graph algorithms to be easily implemented and reasoned about. However, BSP suffers from poor performance due to stale messages and frequent global synchronization barriers. While asynchronous models have been proposed to alleviate these overheads, existing systems that implement such models have limited scalability or retain frequent global barriers and do not always support graph mutations or algorithms with multiple computation phases. We propose barrierless asynchronous parallel (BAP), a new computation model that overcomes the limitations of existing asynchronous models by reducing both message staleness and global synchronization while retaining support for graph mutations and algorithms with multiple computation phases. We present GiraphUC, which implements our BAP model in the open source distributed graph processing system Giraph, and evaluate it at scale to demonstrate that BAP provides efficient and transparent asynchronous execution of algorithms that are programmed synchronously.

Secondly, very few systems provide serializability, despite the fact that many graph algorithms require it for accuracy, correctness, or termination. To address this deficiency, we provide a complete solution that can be implemented on top of existing graph processing systems to provide serializability. Our solution formalizes the notion of serializability and the conditions under which it can be provided for graph processing systems. We propose a partition-based synchronization technique that enforces these conditions efficiently to provide serializability. We implement this technique into Giraph and GiraphUC to demonstrate that it is configurable, transparent to algorithm developers, and more performant than existing techniques.

This will be a short practice talk.

This is joint work with Carlos Teixeira, Alexandre Fonseca, Marco Serafini, Georgos Siganos, and Mohammed Zaki. It appears in SOSP 2015.