The extensive use of APIs as the entry point to many Cloud-based applications has created challenging problems, especially concerning API quality properties such as performance and reliability. API best practices and patterns, such as bundling requests, rate limiting, or load balancing, have been proposed to solve these challenges. Unfortunately, no study investigating the impact of existing API practices and patterns on such quality properties exists beyond informal recommendations. In this paper, we fill this gap by proposing a pattern-based, automated recommendation approach to improve the performance and reliability of API operations. We provide a benchmark suite based on a realistic open-source microservice application to enable the automatic generation of comprehensive decision tree models. These models are then processed to generate API design recommendation algorithms to improve API operations regarding performance and reliability stored in catalogs for reuse. We validate our algorithms using extensive data sets generated by running the benchmark on a private cloud and AWS. For both environments, based on the decision tree models automatically generated from the measured data, API design recommendation algorithms have been calculated using our approach.

There are many challenges in maintaining the desired quality of service levels in modern microservice and cloud applications. Numerous techniques and patterns, such as API Rate Limit, Load Balancing, and Request Bundle, have been suggested for API services and clients to improve quality properties related to performance and reliability. However, no study has measured the impact of these techniques and their combinations in a specific configuration, especially using a large distributed system workload setting. This paper experimentally studies the effects of combining the API Rate Limit, Load Balancing, and Request Bundle patterns based on a realistic, third-party microservice-based application deployed in a private cloud and on the Amazon Web Services cloud (AWS) using 130 different configurations. We have run each configuration 500 times in the private cloud, totaling more than 4500 hours of runtime, and 200 times on AWS, totaling more than 3900 hours of runtime. We developed regression models from the collected data to predict the performance and reliability impacts of combining such techniques and patterns. We found that the models provide acceptable prediction errors below 30% on the private cloud and AWS. Further, we found that the models work best in highly reliable environments like AWS. In addition to the concrete analyses provided in our work, we propose a general and largely automated method that can be followed iteratively to evaluate similar techniques and patterns for their quality properties.