Scalable software services are expected to be always available. Hence, service software must be updated on the fly, while the service continues to function. The central challenge for on-the-fly update is to ensure the absence of observable, destructive interference between current- and new-version components, whose inconsistencies may lead to unexpected outcomes and catastrophic service failures.
This work proposes new theoretical foundations for modular, consistent, on-the-fly updates for networks of processes. In a process network, updates to one component may cause destructive interference at a distant, non-updating component. We introduce coherence, a causality-based formulation of update consistency that guarantees a clear separation between old- and new-version transactions. We establish through an impossibility result that myopic update strategies, which observe and control only the updating components, are insufficient for consistency. This result leads to an isolation principle that cleanly separates a core, controlled update strategy from surrounding uncontrolled network activity, which enables modular reasoning about sequential and even concurrent updates. Applying this principle to microservice networks, we show how consistent update strategies for the full network can be constructed from strategies for updating isolated microservices based on compatibility and commutativity properties of the microservice actions.