Blog post · November 19, 2026 · aviznetworks.com
Aviz Networks outlines two AI fabric monitoring data paths
- About Aviz Networks
- Also names ECN
- Internet Engineering Task Force (IETF)
- It
- Network
- Read the original
Summary
Aviz Networks argues AI fabrics need separate streaming telemetry and fast notification data paths. Brief congestion events can stall training jobs and waste GPU time before dashboards register them, making the distinction relevant to enterprise network and IT leaders.
Role of Streaming Telemetry
Streaming telemetry, delivered through mechanisms such as gNMI Subscribe, supplies counters, queue statistics, optics health, routing state and historical baselines. The post describes it as a step up from polling, because systems subscribe to state and receive pushed updates instead of querying devices on an interval. It is characterized as the operational record of the network, but it is optimized for breadth and history rather than speed on urgent events.
Why AI Workloads Are Sensitive to Network Events
The article states that distributed AI training and inference depend on the network as part of the application's performance envelope, since GPUs coordinate over tight collective-communication windows and a disruption in one path can stall an entire job. Event types cited include queue buildup during collective communication, packet drops, ECN and PFC-related congestion, link flaps, optical pre-failure signals, mid-job path changes, telemetry gaps and NIC issues.
The post references the IETF Fast Network Notifications (FANN) effort, which remains in draft. According to the article, modern forwarding hardware can often detect a problem in microseconds to low milliseconds, but delivering that information to a system that can act on it typically takes far longer.
Fast Notification as a Separate Lane
The article characterizes fast notification as a complement to streaming telemetry rather than a replacement. A useful notification, according to the post, includes event type and location, severity and affected scope, timestamp and source, a recommended consumer with required urgency, and a correlation key linking it to downstream telemetry. Consumers may include a network controller that reroutes traffic, a workload scheduler that avoids impaired segments, or a remediation system that opens an incident.
Correlating Both Paths During an Incident
The post uses an example in which a queue congestion notification on port 12 of leaf switch L14 is correlated with streaming telemetry and topology. The correlated view shows the queue had been rising for six minutes, that L14 connects to eight GPU nodes running two training jobs, and that a similar symptom appeared on a neighboring leaf twice in the past hour. Policy allowed an automatic reroute for this severity but required approval for a scheduler pause, and telemetry confirmed queue depth returned to baseline within 90 seconds after the reroute.
Architecture Requirements for AI Fabric Monitoring
The article recommends separate pipelines rather than a single one, including high-volume streaming collectors, a low-latency ingestion path for urgent events, and normalized schemas that allow correlation. It also lists topology and workload correlation, time-windowed incident objects, policy-based remediation tiers of auto, approve and block, and verification with audit history after each action.
Aviz ONES Capabilities
According to the post, Aviz ONES currently correlates streaming telemetry collected via gNMI and related sources with topology and device state to provide real-time and historical views. The company says it is building toward adding a fast-notification lane to the same correlation layer, with longer-term workflows covering detection, correlation, recommendation, approval, execution, verification and rollback.
This Blog Summary is a fact-based summary of the vendor blog. For enterprise decision-makers running AI training and inference fabrics, the post argues that streaming telemetry and fast notification serve distinct purposes and that correlating both with topology and policy is needed to produce actionable, verifiable incidents. The planned capabilities described are the vendor's stated direction rather than released functionality.
Our summary. The original keeps its own byline.