The four numbers that matter
CPU, physical memory, entry processes and I/O. Hitting a limit produces different symptoms: CPU throttling makes everything slow, memory limits produce 500 errors, entry process limits make the site unreachable under concurrency.
Faults are the signal
A fault means a limit was hit and something was refused. Occasional faults during a backup are normal; sustained faults during business hours mean the plan or the code needs attention.
Correlate before upgrading
Match spikes against what happened — a crawler, a scheduled import, a campaign. Upgrading to absorb a badly behaved bot is the most expensive way to solve a robots.txt problem.
