Evaluating Query Languages and Systems for High-Energy Physics Data

Query languages in general and SQL in particular are arguably one of the most successful programming interfaces. Yet, in the domain of high-energy physics (HEP), they have found limited acceptance. This is surprising since data analysis in HEP matches the SQL model well: it is fully structured data queried using combinations of selections, projections, joins, and reductions. To gain insights on why this is the case, in this paper we perform an exhaustive performance and functionality analysis of several data processing platforms (Amazon Athena, Google Big Query, Presto, Rumble) and compare them to the new RDataFrame interface of the ROOT framework, the most commonly used system by particle physicists today. The goal of the analysis is to identify the potential advantages and shortcomings of each system considering not only performance but also cost for cloud deployments, suitability of the query dialect, and resulting query complexity. The analysis is done using a HEP workload: the Analysis Description Languages (ADL) benchmark, created by physicists to capture representative aspects of their data processing tasks. The evaluation of these systems results in an interesting and rather complex picture of existing solutions: those offering the best possibilities in terms of expressiveness, conciseness, and usability turn out to be the slowest and most expensive; the fastest ones are not the most cost-efficient and involve complex queries; RDataFrame, the baseline we use as a reference, is often faster and cheaper but is currently facing scalability issues with large multi-core machines. In the paper, we analyze all the aspects that lead to such results and discuss how systems should evolve to better support HEP workloads. In the process, we identify several weaknesses of existing systems that should be relevant to a wide range of use cases beyond particle physics.