// the find
cloudflare/ebpf_exporter
Prometheus exporter for custom eBPF metrics
A Prometheus exporter that lets you write custom eBPF programs and expose their output as metrics. Built by Cloudflare and backed by libbpf with CO-RE support, so compiled programs run across kernel versions without recompilation. The target audience is platform engineers who want kernel-level observability (block IO latency, TCP retransmits, scheduler events) without instrumenting application code.
CO-RE via libbpf means you compile once and the binary works on any BTF-enabled kernel — no per-host recompilation. The decoder pipeline (ksym, cgroup path, inet_ip, regexp filtering, static maps) handles the tedious kernel-to-human translation that you'd otherwise write yourself every time. Capability dropping after probe attachment is a real security win for production deployments. The 40+ included examples covering bio, TCP, cgroups, scheduler, and XDP are genuinely useful as starting points rather than toy demos.
You must pre-compile the BPF object files before deploying — there's no JIT-from-config path, which means your CI/CD needs a clang toolchain and a matching vmlinux.h for each target architecture. The histogram bucketing configuration (exp2, exp2zero, linear, fixed) requires you to match your kernel-side map layout precisely; a mismatch silently produces wrong data rather than an error. OpenTelemetry tracing support is present but clearly secondary — it's documented in a subdirectory and the examples are sparse. The third-party Helm chart is explicitly disclaimed by the maintainers, so Kubernetes users are on their own for production packaging.