Telemetry-as-a-Service: Decoupling Collection from Injection for Scalable Network Monitoring

Paper i proceeding, 2026

Network telemetry has evolved from poll-based SNMP to push-based streaming protocols such as gNMI and YANG-Push, yet current architectures tightly couple device collection with pipeline injection inside monolithic collectors. Scaling telemetry processing under this model requires replicating device credentials across every collector instance, and outsourcing processing to third parties exposes management-plane access together with internal network topology. The term “telemetry-as-a-service” (TaaS) has been introduced in the literature as an API abstraction, but its architecture still combined collection and processing while retaining full device credentials. In this paper we extend the TaaS concept by separating stateful collectors, which face devices and hold credentials, from stateless injectors, which are credential-free, topology-agnostic, and horizontally scalable. The key enabler is a push-direction inversion: collectors push telemetry to injectors rather than injectors pulling from devices, creating a natural credential boundary. We present a gRPC-based protocol that preserves gNMI semantics while enforcing this boundary, and evaluate an implementation demonstrating linear throughput scaling with minimal latency overhead. The architecture enables independent scaling of collection and injection tiers, credential isolation suitable for outsourcing, and topology privacy preservation.