Plan-structured deep neural network models for query performance prediction

Abstract

Query performance prediction, the task of predicting the latency of a query, is one of the most challenging problem in database management systems. Existing approaches rely on features and performance models engineered by human experts, but often fail to capture the complex interactions between query operators and input relations, and generally do not adapt naturally to workload characteristics and patterns in query execution plans. In this paper, we argue that deep learning can be applied to the query performance prediction problem, and we introduce a novel neural network architecture for the task: a plan-structured neural network. Our approach eliminates the need for human-crafted feature selection and automatically discovers complex performance models both at the operator and query plan level. Our novel neural network architecture can match the structure of any optimizer-selected query execution plan and predict its latency with high accuracy. We also propose a number of optimizations that reduce training overhead without sacrificing effectiveness. We evaluated our techniques on various workloads and we demonstrate that our plan-structured neural network can outperform the state-of-the-art in query performance prediction.

Active Learning for ML Enhanced Database Systems

“…There are many examples of ML enhanced databases [14,37,45]. Without loss of generality, we discuss two examples that represent two distinct types of ML problems which have many applications in database systems.…”

“…Previous works have applied various ML models to these tasks, such as random forests [15], boosted regression trees [37], and deep neural networks [45]. Training data for these models typically come from the execution history of standard benchmarks or any accessible databases.…”

“…We increasingly see the promise of using machine learning (ML) techniques to enhance database systems' performance, such as in query run-time prediction [18,37], configuration tuning [51,66,77], query optimization [35,44,50], and index tuning [5,14,61]. For example, a query optimizer can use an ML model that predicts a query plan's execution cost to search for the best query plan [45]. Most of these techniques use supervised learning [54], which trains the ML model using a labeled dataset.…”

“…3 Batching: Many AL strategies assume model retraining after acquiring each label. Training ML models for databases can be expensive [14,45]. Thus, the ADCP needs to support acquiring a batch of labels before retraining the model.…”

Machine Learning for Database Management Systems

“…Seq2SQL [3] and SQLNet [4] use a large manually generated dataset of SQL and natural language query pairs. Most other systems (e.g., CardLearner [7], QPPNet [32], Naru [8], AutoAdmin [14,15], and Learned Index [26]) preform initial training from previously collected training data and then perform periodic retraining as new data becomes available or the query workload or the data in the DBMS significantly change. In some cases, collecting training data can be expensive as it requires executing a large number of queries potentially on large databases.…”