Self-Optimizing Cloud to Edge: Machine Learning for Serverless, Drones, and Distributed Databases

Advances in cloud and edge computing have enabled scalable, on-demand resource provisioning across heterogeneous platforms. However, these environments pose critical challenges: ensuring performance reliability under variable or spiky workloads, optimizing resource allocation, and balancing cost-efficiency. This thesis bridges these gaps through three adaptive, machine learning-driven contributions.

First, we develop a serverless cloud computing optimization solution for reducing cold starts—a latency caused by initializing new function instances after periods of inactivity—by using an enhanced Genetic Algorithm that dynamically balances cold-start frequency, execution time, and energy overhead. Second, we design novel UAV mobility algorithms—DroneZoom, DroneCycle, and a hybrid approach—to optimize drone-based autonomous surveillance. By dynamically adapting flight trajectories and altitudes based on real-time event rates, these methods achieve efficient large-area coverage while preserving high detection accuracy under strict latency constraints. Finally, we present OptimusCassandra and OptimusRedis for fine-grained configuration tuning in cloud-hosted NoSQL databases for two distinct types of NoSQL databases. Our algorithm relies on workload prediction and partial reconfiguration to create heterogeneous clusters that adapt to changing demands. Collectively, these works illustrate how predictive modeling and intelligent control can significantly improve system performance, from faster serverless response times to robust aerial surveillance and cost-effective data management. Our results not only demonstrate immediate gains in latency reduction, resource usage, and throughput but also highlight the potential for broader applications in future adaptive, self-managing cloud and edge infrastructures.