On Implementing Autonomic Systems with a Serverless Computing Approach: The Case of Self-Partitioning Cloud Caches

The research community has made significant advances towards realizing self-tuning cloud caches; notwithstanding, existing products still require manual expert tuning to maximize performance. Cloud (software) caches are built to swiftly serve requests; thus, avoiding costly functionality additions not directly related to the request-serving control path is critical. We show that serverless computing cloud services can be leveraged to solve the complex optimization problems that arise during self-tuning loops and can be used to optimize cloud caches for free. To illustrate that our approach is feasible and useful, we implement SPREDS (Self-Partitioning REDiS), a modified version of Redis that optimizes memory management in the multi-instance Redis scenario. A cost analysis shows that the serverless computing approach can lead to significant cost savings: The cost of running the controller as a serverless microservice is 0.85% of the cost of the always-on alternative. Through this case study, we make a strong case for implementing the controller of autonomic systems using a serverless computing approach.