Baseline
Use timestamps, queue IDs, remote responses, source IP, VMTA, sender domain and message class to establish the failure boundary. Avoid reacting to one isolated delivery attempt.
Signals
- Connection setup latency
- Messages per connection
- 4xx rate as concurrency rises
- Remote disconnects and timeouts
- Local socket and file descriptor pressure
Method
Change one route at a time in small increments. The best setting maximizes sustained accepted throughput, not raw connection count.
Operational rule: collect evidence before changing policy, make one controlled change, verify the result, and retain a rollback path.