Performance Prediction for Graph Queries

Abstract

Query performance prediction has shown benefits to query optimization and resource allocation for relational databases. Emerging applications are leading to search scenarios where workloads with heterogeneous, structure-less analytical queries are processed over large-scale graph and network data. This calls for effective models to predict the performance of graph analytical queries, which are often more involved than their relational counterparts. In this paper, we study and evaluate predictive techniques for graph query performance prediction. We make several contributions. (1) We propose a general learning framework that makes use of practical and computationally efficient statistics from query scenarios and employs regression models. (2) We instantiate the framework with two routinely issued query classes, namely, reachability and graph pattern matching, that exhibit different query complexity. We develop modeling and learning algorithms for both query classes. (3) We show that our prediction models readily apply to resource-bounded querying, by providing a learningbased workload optimization strategy. Given a query workload and a time bound, the models select queries to be processed with a maximized query profit and a total cost within the bound. Using real-world graphs, we experimentally demonstrate the efficacy of our framework in terms of accuracy and the effectiveness of workload optimization. Example 1: Consider a query posed on a knowledge graph DBpedia that finds the artists who work with "J.Lo" in a Band [19]. This query can be represented by a graph pattern Q that carries (ambiguous) keywords, with a corresponding approximate match as illustrated in Fig. 1. Each pattern node in Q may have a large number of candidate matches. The quality of the answer is usually determined only at run-time via similarity functions [29]. For example, "Canela Cox" is a best answer when "Jennifer Lopez" is matched to the ambiguous keyword "J.Lo" in the query. Conventional QPP that exploits relational algebra may not be applicable for graph pattern Q, as the syntax and declarative operators are hard to be derived from the "structureless" Q. Moreover, deriving statistics from the graph data alone is already expensive, due to the sheer size of data, and the fact that the underlying graph may change from time to time.

Execution Time Prediction for Cypher Queries in the Neo4j Database Using a Learning Approach

“…Figure 5. The mean relative error of the four benchmarks on the three datasets in five types of models: LR [22], SVR [22], RT [22], FCNN [44], and RBF [38].…”

“…The results for the FIFA2021 dataset and CORD-19 dataset are omitted for space considerations. ) with five types of models: LR [22], SVR [22], RT [22], FCNN [44], and RBF [38].…”

“…In [33], the authors provided a benchmarking framework for the systematic investigation of query evaluation performance. In [22], the authors developed modeling and learning algorithms for readability and graph pattern matching queries. The former provided three features for graph queries and predicted the execution time using the statistical learning models with LR, RT, RF, and SVM.…”

“…Traditional machine learning: We implemented the traditional machine-learning approaches of line regression (LR), regression trees (RT), and support vector regression (SVR) that were proposed in [22]. The authors proposed five types of statistical machinelearning models (linear regression [39], regression tree [40], random forest [41], KNN [42], and SVM [43]) to predict graph query performance.…”

“…Second, traditional graph query performance prediction methods are mostly based on a query optimizer [20,21] which requires tuning by the database manager. Third, for graph query performance prediction problems, existing research work commonly uses the linear regression approach, which cannot obtain accurate prediction results for complex queries [22,23]. Moreover, graph query performance prediction features selection is also a challenge; the quality of the prediction models depends on the selected query features, and irrelevant features will increase noise during the training process.…”