The constraint

The previous peak season had degraded for ninety minutes during the highest-revenue window of the year. The post-mortem identified the symptom; nobody had the confidence to say it would not recur.

What we changed

We rebuilt caching and queueing, then load-tested against a replay of the actual incident traffic rather than a synthetic curve.

  • Edge caching moved in front of the catalogue path
  • Write queue decoupled from the checkout response
  • Load tests replay real incident traffic, run before every peak

Where it landed

120,000 requests per second sustained through peak with no incidents, and a repeatable test that makes next year an evidence question rather than a hope.