Public pages
Your pod’s CPU usage looks green. Requests and limits are set correctly, and kubectl top shows nothing alarming. Yet application latency is spiking. The hidden cause? A noisy neighbour sharing the same node is winning scheduling, and standard Kubernetes metrics are completely blind to it.
Standard metrics track CPU utilisation and throttling, but they cannot measure CPU contention: the delay an application experiences waiting in the kernel run-queue to get CPU time. Building on Netflix’s eBPF research into measuring kernel run-queue latency, this talk bridges the gap between raw eBPF code and a practical Kubernetes pattern, complete with monitoring dashboards.
The talk walks through deploying an eBPF collector via a Kubernetes DaemonSet that aggregates scheduling latency directly inside the kernel with minimal overhead, then exports these metrics into Prometheus and explores them on live dashboards. Attendees leave with an open-source end-to-end pattern they can deploy in their own clusters, and an understanding of how eBPF opens up visibility into other invisible node-level bottlenecks, from disk I/O queuing to memory contention.
Event: CloudCon 2026 (CloudCon Sydney)
Date: 9-10 November 2026 (exact session time TBC)
Location: ICC Sydney
Format: Full length talk (20 minutes)
Slides will be available after the talk.