The Computing Continuum (CC) integrates different layers of processing infrastructure, from Edge to Cloud, to optimize service quality through ubiquitous and reliable computation. Compared to central architectures, however, heterogeneous and dynamic infrastructure increases the complexity for service orchestration. To guide research, this article first summarizes structural problems of the CC, and then, envisions an ideal solution for autonomous service orchestration across the CC. As one instantiation, we show how Active Inference, a concept from neuroscience, can support self-organizing services in continuously interpreting their environment to optimize service quality. Still, we conclude that no existing solution achieves our vision, but that research on service orchestration faces several structural challenges. Most notably: provide standardized simulation and evaluation environments for comparing the performance of orchestration mechanisms. Together, the challenges outline a research roadmap toward resilient and scalable service orchestration in the CC.
Sensory data from Internet of Things (IoT) devices form the backbone of pervasive applications. As running example, consider a digital twin of an entirely smart city that allows clients to interact with content through Augmented Reality (AR) across surfaces, like handhelds or windows. To enter and navigate such environments in real-time, computation is shifted from Cloud centers towards Edge devices; thus, it is possible to process and render content on nearby devices. While this reduces latency, Edge devices offer limited and less predictable resources. During environmental changes, e.g., when facing higher load at rush hours, Edge devices require fallback mechanisms that ensure service quality. To combine the strengths of both worlds, Cloud and Edge layers are integrated in one composite architecture-the Computing Continuum (CC). Thus, applications can place latency-aware services on the Edge, and use the abundance of Cloud XXXX-XXX © 2026 IEEE Digital Object Identifier 10.1109/XXX.0000.0000000 resources for hosting the remaining services.
While the CC aims to improve Quality of Experience (QoE), it introduces numerous orchestration challenges [1]. Most notably, by distributing services from one application across different physical devices with heterogeneous characteristics, it gets complex to predict application behavior (e.g., after scaling up one service). This is further complicated by deploying applications across multiple vendors and jurisdictions, as providers may be unwilling to share their full system state. This demands solutions that allow parties with partial observability to collaborate towards common optima, while accounting for their individual behavior.
Driven by the numbers of IoT devices and the need to process data nearby, the last decade has produced a large amount of research to cope with the heterogeneity and distribution of processing systems. Common topics revolve around latency-aware service placement [2], or forming agreements between service and device providers [3]. These mechanisms are often based on static architectures and require a priori understanding of services and devices, e.g., benchmarking the service quality across device types. However, individual devices in the CC will (dis)appear dynamically, which can include new device types that are not benchmarked yet. Also, clients can always redefine the desired service operation-usually specified through Service Level Objectives (SLOs)-Thus putting the service in an unknown context. To summarize these problems, this article provides a structured overview of hypotheses that describe the CC, its applications, and challenges for service orchestration.
To guide research on service orchestration-also inspired by the well-known vision of autonomic computing [4]-this article outlines how continuous service interpretation and adaptation can help optimize SLO fulfillment. To instantiate this design, we present a preliminary implementation using Active Inference (AIF) [5]-a concept from neuroscience that aims to create self-organizing components that maintain internal requirements fulfilled: First, we model the interactions between a service and its environment through a behavioral Markov blanket (MB)-a description of how a service interprets its current state and which corrective action to take [6]. Thus, services can decide how to react depending on the context, e.g., by shifting computation or services accordingly. Second, whenever the context changes, e.g., after dynamically reconfiguring SLOs, these behavioral models must be updated. Therefore, we wrap each component in a continuous action-perception cycle and adjust its MBs according to environmental feedback [7]. Third, we analyze interactions between services, and additionally, their hosting devices, by composing their MBs [8]. This enhances collaboration within the CC because services can estimate how local actions impact dependent services and the corresponding resource demand. This implementation, using AIF, allows collaborative agents to continuously model and understand the environment-a first step to address fundamental problems of dynamic CC systems. However, when mapping the implementation to the structural hypotheses and problems, we still identify multiple challenges that impede research on service orchestration. Most notably, we see a clear gap for: (1) large-scale, standardized simulation environments that simplify testing hypotheses and comparing solutions, (2) continuous and context-aware mechanisms for accurate ML inference, and (3) seamless infrastructure composition that allows clients to use infrastructure from multiple vendors or individuals. To achieve structural improvement for service computing, we summarize these challenges in more detail under three key challenge areas.
Research on CC systems is in an early stage, resulting in numerous beliefs and views about their inherent challenges, like in [3], [7], [9]. In this section, we provide a structura
This content is AI-processed based on open access ArXiv data.