I've had incidents where CPU sat at 80% for hours and fixing it meant deploying experimental changes and hoping. Metrics told us which services, traces showed request flow, but we still didn't know which function was actually hot. We added Parca for continuous profiling. It uses eBPF to sample stack traces in production without touching application code. Flamegraphs show exactly where CPU goes. Found things like JSON serialization and regex loops consuming 30-40% of resources in services we thought were optimized. Small fixes, big impact. The ROI was real. We dropped CPU enough to downsize node pools. The post covers the setup, integration with existing observability stacks, when to adopt, and the actual ROI we saw: [eBPF Observability and Continuous Profiling with Parca](https://fatihkoc.net/posts/ebpf-parca-observability/) What's your approach to performance optimization? Are you profiling in prod or still relying on metrics and intuition?