Two Autoscaling Approaches on Kubernetes Clusters Against Data Streaming Applications

This paper is aimed to study the performance comparison of two autoscaling approaches against data streaming and visualization applications under Kubernetes clusters. In particular, vertical and horizontal scaling methods are focused. The experimental setup has been done extensively based on the architecture stack which includes Apache Kafka, Apache Spark, Filebeat, Logstash, Elasticsearch, Kibana, Prometheus, and Grafana. The performance and resource utilization of the data flow display system in Kubernetes clusters are evaluated using Apache JMeter. Three metrics: CPU utilization, response time, throughput, and error rate are used. The experimental results show that on average, the CPU utilization of the horizontally scaled system is higher by 18.48% compared to the non-scaling system, while the vertically scaled system uses 49.45% using less CPU resources than the non-scaling system. The average response time of the horizontally scaled system is 65.60% lower than the non-scaling system, while the vertically scaled system is 1.19% lower than the non-scaling system. The error rate of the horizontally scaled system is 0.04%, but the error rate of the vertically scaled system is as high as 15.16%.