eBPF-based observability replacing sidecars — real production experience?

Looking at Cilium Tetragon and Pixie for replacing our sidecar-based observability stack. Sidecars add 30-40ms latency per hop and consume ~200MB RAM per pod. eBPF promises zero-overhead kernel-level tracing. Has anyone actually run this in prod with 50+ services? Concerns: kernel version lock-in, debugging when eBPF programs crash, and whether the tooling is mature enough for on-call.