gRPC Streaming Telemetry

Uses Protocol Buffers (protobuf) as the interface definition language and wire format. This provides a strongly-typed, compact binary encoding that is 3-10x smaller than JSON and significantly faster to serialize/deserialize. The schema is defined in .proto files, which are compiled to generate client and server code in multiple languages. This strict contract eliminates ambiguity in data models and ensures backward-compatible evolution of telemetry paths.

Operates on a dial-out paradigm where the collector subscribes to specific telemetry paths on the network device. The device then pushes data continuously over a persistent HTTP/2 connection. Subscription types include:

• ONCE: A single snapshot of current state
• POLL: On-demand retrieval
• STREAM: Continuous, low-latency streaming of state changes This eliminates the overhead of periodic SNMP GET requests and enables sub-second cadence for high-resolution data.

Leverages HTTP/2 as the transport layer, providing native multiplexing of multiple concurrent streams over a single TCP connection. This eliminates head-of-line blocking, reduces connection overhead, and enables bidirectional streaming. The transport also supports TLS 1.3 encryption by default, ensuring all telemetry data is authenticated and encrypted in transit without the performance penalty of separate VPN tunnels.

Telemetry paths are defined using YANG (RFC 7950) data models, providing a vendor-neutral, structured representation of device configuration and operational state. OpenConfig and IETF YANG models define standardized paths for common network functions. This allows a single collector to subscribe to telemetry from multi-vendor environments without custom parsing logic. The YANG model defines the hierarchy, types, and constraints that protobuf serializes.

gRPC Network Management Interface (gNMI) is the standardized service definition for streaming telemetry. It defines four RPCs:

• Capabilities: Discover supported models and encodings
• Get: Retrieve a snapshot of state
• Set: Modify device configuration
• Subscribe: Stream telemetry updates The Subscribe RPC supports sample (periodic) and on-change (event-driven) modes, with the latter only transmitting data when a value changes, dramatically reducing bandwidth for stable metrics.

Supports two fundamental streaming modes that optimize for different use cases:

• On-Change (Event-Driven): The device transmits only when a monitored value changes. Ideal for interface state transitions, BGP session flaps, or alarm conditions. Reduces bandwidth by 90%+ compared to fixed-interval polling.
• Periodic (Sample): The device transmits at a configured interval (e.g., every 10 seconds). Essential for trend analysis, capacity planning, and metrics that change continuously like CPU utilization or throughput counters. Collectors can mix both modes within a single subscription.

Two fundamental subscription initiation modes:

• Dial-Out: The network device initiates the gRPC connection to a pre-configured collector. This is firewall-friendly and ideal for devices behind NAT.
• Dial-In: The collector initiates the connection to the device. This centralizes connection management but requires the device to be network-reachable. Both modes support dynamic subscriptions without configuration changes.

The traditional Simple Network Management Protocol that gRPC streaming telemetry replaces. SNMP relies on a central manager periodically 'walking' Management Information Bases (MIBs) via UDP, which is connectionless and stateless. This pull-based model struggles with scale, high-frequency data, and CPU overhead on devices, making it unsuitable for the sub-second granularity required by modern AI-driven RAN optimization.

A deployment pattern where a third-party analytics application runs directly on a network device inside a container. The application subscribes to gRPC telemetry locally, processes data at the source, and streams only actionable insights or aggregated metrics externally. This architecture minimizes backhaul bandwidth and enables ultra-low-latency closed-loop control for functions like predictive load balancing.

A common architectural pattern where a gRPC telemetry collector acts as a bridge, decoding Protobuf streams and publishing them into a Kafka cluster. This decouples data production from consumption, allowing multiple downstream systems—such as a time-series database for anomaly detection and a data lake for offline model training—to consume the same telemetry stream reliably and at their own pace.