What’s the fastest way to tell if it’s the DB or the app?

Check RED: if latency spikes align with `db_cpu` and `active_connections`, and traces show long DB spans, it’s the DB. Otherwise, capture a flamegraph; if most time is in wait/lock or syscalls, look at thread pools and external calls.

Should we scale up hardware first?

Only as a stopgap. Without fixing the bottleneck, you’ll pay more for the same problem at a higher RPS. Apply the matching playbook, then right-size infra.

How do we keep AI-generated code from regressing performance?

Add performance linters and tests in CI. Flag dangerous SQL patterns, enforce query timeouts, and run a light load test gate. Schedule periodic vibe code cleanup and code rescue reviews.

We use a vendor APM. Do we still need OpenTelemetry?

Yes. OTel gives you portable context propagation and avoids lock-in. Most APMs ingest OTel now; you can switch vendors without ripping out instrumentation.

Guides · Dec 12, 2025 · 9 minute read

The P99 Killers: Playbooks We Actually Use to Fix DB Hotspots, Thread Pools, and K8s Throttle Loops

If you keep firefighting the same latency spikes, it’s time to standardize the fix. Here are the performance playbooks we install at clients so teams can diagnose and resolve bottlenecks in hours, not quarters.

Back to all posts

The P99 Killers: Playbooks We Actually Use to Fix DB Hotspots, Thread Pools, and K8s Throttle Loops

Key takeaways

Implementation checklist