Skip to content

Performance — Junior

Measure representative latency and throughput with warmup, stable inputs, and repeated runs. Report distributions, not one fast run.

sequenceDiagram Benchmark->>System: representative workload System-->>Benchmark: latency distribution Profiler->>System: sample CPU and allocations System-->>Profiler: hotspot evidence

Optimize the dominant cost: algorithm, I/O, serialization, allocation, or network. Verify correctness and rerun the same benchmark. A microbenchmark cannot prove whole-system improvement.

Test yourself

  1. What makes a workload representative?
  2. Why report p95 or p99?
  3. Which profile matches allocation pressure?
  4. How do you prove the change helped users?

Continue to middle.md.