eBPF Autoscaler
Runqueue-aware autoscaling sandbox for Kubernetes on Kind.
An eBPF collector reads Linux scheduler tracepoints to export runqueue latency histograms. Prometheus scrapes the metrics and Grafana visualizes them for scaling experiments.
The problem
CPU utilization alone does not describe how long a task waits to run. This sandbox investigates runqueue latency as an additional signal for Kubernetes scaling experiments.
The approach
An eBPF collector attaches to Linux scheduler tracepoints and exports latency histograms through a Go exporter. A Kind cluster runs the collector; Prometheus scrapes its metrics and Grafana displays the distributions.
What it demonstrates
Kernel instrumentation, metric collection and a Kubernetes observability pipeline. The repository documents environment limits and the remaining Prometheus Adapter integration, rather than presenting the sandbox as a completed production autoscaler.
- Linux scheduler
- eBPF + Go
- Prometheus
- Grafana
An experimental sandbox; the repository documents measurement limitations and remaining scaling integration work.