Fabric Grooming

Overview

Fabric grooming is the operational practice of proactively and reactively identifying suspect or degraded optical links in an AI cluster fabric and removing them from service before or after they cause a link flap event. The goal is to sacrifice capacity in order to preserve fabric stability — a trade-off supported by operational evidence that a stable lower-capacity fabric outperforms an unstable full-capacity one for AI training workloads.

OCI's fabric grooming process uses a multi-layer toolset: trend analysis on optical power levels, predictive metrics across multiple telemetry dimensions, FEC histogram pattern matching, automated fast-flap detection and link-down triggers, link debounce, and Oracle Cloud Agent (OCA) host-side telemetry. Root cause analysis (RCA) during grooming is difficult and extends deployment timelines — grooming an unstable fabric while simultaneously trying to deliver it to customers is a compounding operational burden.