Question 932

Eight PM Problem

Your API averages 800 QPS across the day, but dashboards show 3,400 QPS every evening around 8pm. You provisioned enough servers for 1,000 QPS and now see timeouts nightly. What went wrong in the estimate?

A

The 25 percent headroom over average is the standard buffer for a service this size, so the fix is client retry backoff and a request queue rather than more instances

B

Capacity was sized to the daily average with a thin margin; it has to be sized to peak, which here is over 4x average, so plan for roughly 3,500 QPS or more

C

The daily average was computed over the wrong window; recompute it across the 8pm to 11pm hours alone and provision the fleet against that higher evening average

D

The 3,400 QPS figure is a dashboard sampling artifact from a short scrape interval, since sub-second spikes drain no sustained capacity from the fleet