Could You Debug This Without AI?
The LLM is unavailable. You have a production problem, the dashboards, and what you know. Decide what to check and why before you look; commit to a root cause; then see where AI would genuinely have helped once the diagnosis existed — and where it would have sent you the wrong way.
ArchitectureIntermediate
API latency went from 100 ms to 4 seconds
Page: `GET /orders/:id` p50 latency 108 ms → 4.1 s, starting 14:20 — ten minutes after release 2024.31 reached 100% of pods. Gateway 504s went from 0.2% to 3.1% of all requests, not just order lookups.
DatabaseIntermediate
Database CPU is at 100%
Page: Postgres primary CPU pinned at 100% since 09:40. p95 for `GET /accounts/:id` went from 8 ms to 1.9 s; the connection pool is reporting "waiting for connection" on every service that touches `accounts`. No deploy is on today’s release calendar.
DatabaseIntermediate
Requests occasionally return stale data
Ticket: since Tuesday, about 2% of profile saves "don’t stick" — the user changes their display name, the page reloads and shows the old name, and a second reload shows the new one. `PATCH /me` returns 200 every time; nothing is in the error logs.
Operating SystemsAdvanced
Memory usage keeps increasing
Page: `catalog-api` pods are OOM-killed every 5–7 hours. Working-set memory climbs in a straight line from 380 MiB after start to the 2 GiB limit, then the pod restarts and the line starts again. Traffic has been flat for a month.
NetworkingAdvanced
Under load, some requests hang for exactly 60 seconds
Page: during the 12:00 peak, ~4% of calls from `checkout` to `pricing` hang for exactly 60 s and fail with `ETIMEDOUT`. The other 96% complete in 12 ms. Off-peak the failure rate is zero and `pricing`’s own p99 is 14 ms all day.