FluidStream: Decentralized Tuple Routing for Low-Latency Stream Processing in the Cloud Continuum

Scalable stream processing across the cloud continuum, from the core cloud to the network edge, must effectively deal with two sources of dynamism: unpredictable workload variations combined with extreme infrastructure heterogeneity, network volatility, and resource constraints. While input rates and query loads fluctuate, the continuum environment introduces profound resource disparities, frequent partial failures, and elastic scaling requirements, which collectively complicate the provision of consistent end-to-end latency guarantees. Existing approaches, designed for centralized clouds, struggle due to static placement, centralized control, or inefficient state migration. In this paper, we propose FluidStream, an adaptive query-processing framework designed for the cloud continuum that employs a decentralized tuple-forwarding scheme to dynamically route stream tuples across heterogeneous nodes. We formulate the end-to-end latency minimization problem using stochastic optimization and derive a practical routing policy that balances processing load while considering operator state migration costs across potentially high-latency links. Our approach enables any operator to execute on any node in the continuum, with forwarding decisions made locally based on current queue lengths, network conditions, and state transfer overhead. Through extensive simulation experiments, we demonstrate that FluidStream achieves significantly lower query response times compared to static, random, and YARN-based approaches across scenarios involving flashcrowds, straggler nodes, and the resource heterogeneity inherent to the continuum. Finally, we describe a Kubernetes-based implementation, validating the practical feasibility of our architecture using cloud-native primitives for container orchestration across distributed environments.