How do I profile in production safely?

Time‑box profiles (30–60s), sample not trace every call, and whitelist access to pprof endpoints via mTLS/IP allowlists. eBPF profilers (Parca, Pixie) provide low‑overhead continuous profiles with aggregation.

We’re on Kubernetes—what K8s settings matter most for perf?

Right‑size `requests/limits` to avoid CPU throttling, set `podDisruptionBudget` to keep capacity during rollouts, and use HPA on a meaningful metric (RPS or queue depth via custom metrics), not CPU% alone.

What if AI‑generated code made things worse?

Seen it. Use the same playbooks: profile, trace, and measure. Then do a vibe code cleanup: remove unnecessary layers, reduce allocations, and add tests around perf‑critical paths. GitPlumbers does code rescue and AI code refactoring with trace‑driven guidance.

Guides · Dec 9, 2025 · 10 minute read

The p95 Kill Kit: Battle‑Tested Playbooks for CPU, DB, GC, and Cache Bottlenecks

Stop vibe‑debugging. Use these focused playbooks to cut p95 latency, tame costs, and ship without fear.

Back to all posts

The p95 Kill Kit: Battle‑Tested Playbooks for CPU, DB, GC, and Cache Bottlenecks

Key takeaways

Implementation checklist