Encoder-decoder-based workload forecasting framework for database-as-a-service

Yunlong Cheng, Xiuqi Huang, Xiaofeng Gao, Guihai Chen

Abstract: With the development of cloud computing, Database-as-a-Service (DBaaS) plays an increasingly important role due to its convenience and manageability. However, as the demand for large-scale databases serving heterogeneous applications increases, DBaaS management becomes a progressively complex challenge. Workload forecasting is crucial for DBaaS automation, allowing it to uphold quality of service (QoS) efficiently with reduced resources. In this paper, we propose an encoder–decoder-based workload forecasting framework, to help DBaaS managers better investigate historical data and accurately predict future workload information. Specifically, for data preprocessing, we first parse the query log and performance log of the workloads to extract their corresponding time series and store them in a k-d Tree structure. For workload forecasting, we propose an adaptive recollected recurrent neural network (AR-RNN). In AR-RNN, the recollection mechanism-based encoder first encodes workload with most similar historical workloads to capture similar patterns. Then, the attention mechanism-based decoder adaptively selects useful information and outputs the predictions. Experiments on two real-world datasets demonstrate the superiority of our proposed method over state-of-the-art workload forecasting methods.