eBPF-based observability replacing Prometheus exporters — is it production-ready yet?

We've been running a fleet of ~200 k8s nodes with 40+ Prometheus exporters each. The overhead (CPU, memory, and maintenance of exporter configs) is becoming a real cost driver. eBPF tools like Pixie and Tetragon promise kernel-level observability without sidecars or exporters. Before committing to a migration path, I'd like to hear from teams who've made this switch in production: - What's the kernel version floor you actually needed? - Did you see the expected reduction in node-level resource usage? - Any gaps in metric coverage that forced you to keep select exporters? We're on kernel 6.x across the fleet, so version floor isn't a blocker. The question is operational readiness. Jurisdiction: N/A.