The bottleneck I found: At 200 concurrent connections, pods were queuing requests — visible as a 2,100ms spike against a 45ms median. The fix: raise CPU limits and pre-scale before traffic hits, not…