Observability Fabric
A high-performance network observability platform in Go, using eBPF for kernel tracing and DPDK for line-rate packet processing — 150+ metrics, ML anomaly detection, zero application-code changes.
The problem
Diagnosing issues in a large, high-throughput environment meant stitching together siloed data by hand. Investigations were slow, and problems often hit service before they were understood.
What I built
A centralized observability platform in Go that captures telemetry at the kernel and packet level and turns it into real-time signal. eBPF traces kernel and network events with negligible overhead; DPDK enables packet inspection at line rate; a metrics pipeline feeds Prometheus and Grafana.
How it works
- eBPF probes capture TCP behavior, packet drops, latency and syscalls in-kernel with minimal overhead.
- DPDK enables userspace packet processing fast enough for 10 Gbps interfaces.
- Concurrent Go pipelines aggregate 50,000+ samples/second into Prometheus metrics.
- An ML layer scores telemetry to flag anomalies and predict failures before service impact.
- Grafana dashboards unify network, kernel and system health in real time.
Related: carrier-grade GCP core
I also led the first-time GCP deployment of a core network function and contributed to a world-first GCP voice-core rollout for a major European carrier — GKE, VPC networking, ingress, DNS/TLS, IAM and Workload Identity, with end-to-end traffic troubleshooting.
Impact
Investigation time dropped ~40% and anomaly detection went from ~15 minutes to under 5 — reactive firefighting became proactive prevention.