After the 8:00 rush Inkwell's servers don't recover. Their queues hold 2 000 requests each, the app gives up after 300 ms and retries, and every retry lands behind requests whose callers have already left. For the rest of the morning most requests fail.

Workload

1 400 rps for most of the 15 seconds, jumping to 4 000 rps for three seconds, 90% reads / 10% writes.

Your move

You can't shorten the queues or add servers. Add a circuit breaker between the client and the load balancer, then run again.