An Adaptive Scaling Method For Microservice Instances In Cloud Environment

In recent years,microservice architecture has become an important foundation for cloud native,and containerization technology is used as the basic unit of application deployment and resource allocation in cloud platforms due to its flexibility.In a cloud-native system,the automatic scaling of containers is one of the important technologies for resource management.Microservice instances are adaptively and automatically scaled according to real-time changing traffic and service requests to meet consumption and meet the service provider’s SLA requirements while reducing resource consumption.Therefore,it’s a challenging task to determine the number of service instances required for each microservice based on the monitoring metrics.Due to its adaptability,Reinforcement Learning has a wide range of applications in the scene of automatic scaling of microservice instances.Usually the convergence of general reinforcement learning requires a large amount of training data,and the results of the scaling strategy on one microservice cannot be well generalized and applied to other services.In order to solve the above problems,the specific work of this paper is as follows: 1.In view of the large state space of the auto scaling scene,the delayed cluster state observation after scaling makes it difficult to learn the strategy,and the slow iteration,this paper proposes a method to build a virtual environment based on offline pressure testing for meta learning,and models the auto scaling scene.2.By analyzing the limitations of the existing Kubernetes HPA,this paper proposes a meta automatic scaling framework Meta Autoscaler,which is based on meta reinforcement learning Meta-Q-Learning.Through offline training in the virtual environment of multiple services,the strategy under the automatic scaling scenario is obtained,which can enable the automatic scaler to achieve good results when facing the decision of new services.Secondly,a meta decision algorithm WEDC-evo based on the evolution of decision window is proposed to prevent the policy decision results from fluctuating and respond to emergency situations.3.Design a meta automatic scaling system,which can create a scaler through configuration,obtain cluster information,make decisions,and put decisions into effect in the Kubernetes cluster.Through the simulation experiment of real NASA access traffic and comparison with Kubernetes native HPA,the SLA violation rate of Meta-Autoscaler decreased by31.60%,and the resource consumption decreased by 1.45%.