Question

A worker fleet's CPU jumps to 95% across all instances at 18:00 and stays there even after the original problem clears. Useful work output (jobs completed/sec) collapses while CPU stays pinned. A profile shows most CPU in a tight retry loop: serialize request, call a dependency, catch the error, immediately retry with no backoff. The trigger was a brief 90-second blip in that dependency at 17:58; the dependency is healthy again, but CPU never recovered. Job arrival rate is flat. Triage and mitigate.