Adaptive container auto-scaling for fluctuating workloads in cloud

Database-as-a-Service(DBaaS) provides services for multiple tenants through resource containers, which are allowed to scale over time to fulfill the service-level agreements. Designing container auto-scaling methods for DBaaS can help reduce their expenditure. Reinforcement Learning (RL) shows powerful performance in cloud resource scaling due to its robustness in dynamic environments. However, the RL-based methods fail to maintain high performance for fluctuating workloads since their fixed-action design cannot adapt to numerous variations of the resource demand. This paper proposes an adaptive container auto-scaling method called Asner that includes an improved RL-based algorithm with a dynamic action model to solve the problem of fixed-action design. Asner consists of a resource estimation model (Estimator) and a RL-based scaling algorithm (Scaler). Estimator adopts a graph-based method to estimate the workload resource demand for container scaling. Scaler generates the container scaling strategy by employing an improved RL-based algorithm with a dynamic action model for adapting to the fluctuating workload. Our experiment results show that Estimator achieves about 93% accuracy under the TPC-DS dataset, Scale’s performance is about 30% higher than the state-of-the-art RL, and Asner improves its performance by up to 45% compared to other methods.