4 / 4 Capacity Planning & Chaos Engineering

Wrong Number to Watch

Your capacity model sizes the fleet on CPU: instances autoscale at 60% CPU and load tests confirm headroom to 3x traffic. In a real 3x event the service failed at 40% CPU with connection acquisition timeouts, because 40 instances x 20 pooled connections exceeded what the database pooler would accept. What is the core lesson for the capacity model?

A

Read replicas should absorb the additional query volume so that the primary stops being the limiting resource during a traffic peak like this

B

The model must use the binding constraint, here total database connections, since scaling on a non-binding metric hides the real ceiling

C

The per-instance pool size should be raised above 20 so that every instance can acquire the connections it needs without waiting during a surge

D

The autoscaler's CPU target should be lowered from 60% so that the fleet grows earlier and never reaches the point where it failed before