Adaptive Recollected RNN for Workload Forecasting in Database-as-a-Service

Nowadays, Database-as-a-Service (DBaaS) plays a more and more important role in the era of big data due to its convenience and manageable capacity. However, with increasing complexity of data-driven applications, the management of database systems becomes intractable. To achieve the self-management of resources, forecasting the workload turns out to be essential. In this paper, we propose a novel machine learning based model, named Adaptive Recollected Recurrent Neural Network (AR-RNN) to help DBaaS managers better capture historical information and predict future workload with a recollection mechanism based multi-encoder and an attention mechanism based decoder architecture. Experiments on two real-world datasets show that our model outperforms both traditional and other machine learning methods for workload prediction.