The growing complexity of modern IT systems has motivated the development of new paradigms (objects, components, services,...) to better cope with the critical size of their functionalities. Such systems are then built as a modular and dynamically adaptable compositions, allowing them to minimise their down-times while performing evolutions or fixes. In order to ensure non-functional properties (i.e. request latency) such systems are distributed across different computation nodes. Besides the added value in term of computational power (cloud), this distribution can also target nodes with dedicated properties such as mobile nodes and sensors (internet of things), physically close to users for interactions. Adapting a system requires knowledge about its current state in order to adapt its architecture to its evolving needs. A new state must be then disseminated to other nodes to synchronise them. Maintaining its consistency and sharing this state is a difficult task especially in case of sporadic connexions which lead to divergent state between sub-systems. To tackle these scientific problems, this thesis proposes an abstraction to design and deploy distributed adaptive systems following the Model@Runtime paradigm. From this abstraction, the proposed approach allows defining a distributed reflexive layer to manipulate heterogeneous distributed nodes. In particular, this contribution introduces variable consistencies in model definition and divergence in system conception. This reflexive layer, eventually consistent allows the construction of distributed adapted systems even on mobile nodes with intermittent connectivity. This work has been realized in an open source project named Kevoree, and validated on various distributed systems ranging from sensor networks to “cloud” computing.
