Performance — Middle¶

Use CPU profiles and flame graphs for on-CPU work, allocation profiles for churn, memory profiles for retention, and tracing for waiting across components. Separate CPU-bound, I/O-bound, lock-bound, and queue-bound behavior.

Benchmark data structures and concurrency with realistic sizes. More workers can reduce latency until locks, connection pools, cache misses, or downstream limits dominate.

Test yourself¶

1. Which profile distinguishes churn from retention?
2. How does contention appear as concurrency rises?
3. Why can a faster function leave endpoint latency unchanged?
4. What is the bottleneck after your optimization?

Continue to senior.md.