Capacity curve: sampled stages

Illustrative assumptions, NOT measured | fixed capacity | load / successful throughput / p99

Capacity curve: sampled stages Illustrative assumptions, NOT measured | fixed capacity | load / successful throughput / p99 Before knee · Load 1,000/s | Done 1,000/s | p99 200ms · Architecture component Before knee Load 1,000/s | Done 1,000/s | p99 200ms Near knee · Load 2,000/s | Done 2,000/s | p99 300ms · Architecture component Near knee Load 2,000/s | Done 2,000/s | p99 300ms After knee candidate · Load 2,500/s | Done 2,200/s | p99 1,500ms · Architecture component After knee candidate Load 2,500/s | Done 2,200/s | p99 1,500ms Increase load Increase load Legend Backend

Locate the knee, then diagnose waits

  • • Throughput gain flattens while tail latency rises.
  • • Sample between stages; correlate pool, DB, I/O and queue waits.
  • • Operational capacity must meet SLO with headroom.