D31: CODECO Federated Cluster Operation Architectural Design

31 provides the final architectural specification of the CODECO Federated Cluster frame-
work. It supersedes D10 and reflects the complete design achieved across the full CODECO
project lifecycle (M1–M39). D31 is an additional deliverable proposed by the consortium to the
European Commission in support of the open-source publication and long-term reuse of the
CODECO architectural design. D31 establishes the three design principles that jointly distin-
guish the CODECO Federated Cluster architecture from existing approaches. Data-compute-
network co-orchestration treats networking, data observability, and computation as per-
spectives that must be jointly considered in every placement and adaptation decision, rather
than optimised in isolation. A hybrid governance model separates governance centralised
at a Hub cluster via Open Cluster Management (OCM) from coordination and execution, which
are progressively decentralised toward Managed Clusters, enabling global policy consistency
while supporting autonomous far Edge operation during Hub disconnections. Application
neighbourhoods define bounded, semantically selected subsets of Managed Clusters rele-
vant to a specific application, limiting the scope of monitoring, learning, and scheduling ex-
changes to provide federation-scale coordination without prohibitive communication overhead.
It also analyses federated operation priority areas, namely, control-plane decentralisation, pri-
vacy-preserving decentralised AI, semantic abstraction and adaptive networking, and feder-
ated scheduling, providing for each a structured analysis of existing approaches, a set of open
questions, and the architectural directions adopted in CODECO.
The deliverable then presents the overall CODECO federated architecture, including its three-
layer governance model, the OCM-based federation control structure, the application neigh-
bourhood concept, and the AI-driven decision flow. Each of its components is presented, and
the single and federated operation is explained: the CODECO Neighbourhood Composer
(CNC) plugin computes and maintains application neighbourhood definitions, determining the
bounded cluster scope for each application's scheduling and monitoring exchanges. ACM-FC
is the sole user-facing entry point, responsible for application lifecycle management via the
CODECO Application Model (CAM), cluster configuration, and coordination of all other com-
ponents. PDLC-FC provides AI-assisted placement recommendations through neighbour-
hood-scoped federated learning and multi-agent reinforcement learning, keeping raw teleme-
try local to each Managed Cluster and exchanging only encrypted model artefacts. MDM-FC
maintains a semantic metadata graph capturing data locality, freshness, compliance, and port-
ability across the federation, enabling data-aware placement decisions. NetMA-FC delivers
cross-layer network observability through ALTO-based cost maps and passive eBPF-instru-
mented monitoring and manages secure inter-cluster connectivity through a Layer 2 mesh
approach integrated with a hybrid SDN control plane. SWM-FC implements a two-layer feder-
ated scheduler supporting stateful migration. The most consequential directions for future
work, and the Eclipse KUDECO project as the long-term vehicle for CODECO's community
development are debated in the final part of the deliverable.