A percentage needs context
High CPU is not automatically a problem. A build server doing useful work can be busy by design, while a lightly loaded interactive workstation may become unpleasant long before it is fully saturated. Memory accounting differs across operating systems, and a host-wide percentage does not identify a leaking process.
Start with a representative history window and the device's purpose. Decide what symptom you want to investigate and who will act. Do not copy one threshold across every machine solely because it is a convenient default.
Combine threshold and elapsed time
Consider an illustrative rule: CPU at or above 90% for five minutes. A short peak above 90% does not satisfy it. The condition must persist across fresh observations for the elapsed duration. A five-minute chart average is not the same as a five-minute sustained rule.
| Illustrative observations | Expected interpretation |
|---|---|
| 95% briefly, then 40% | A spike; the sustained window resets |
| Fresh readings ≥90% over five minutes | Meets the example trigger condition |
| High readings separated by a long data gap | Insufficient continuous evidence |
| Faster repeated high samples within one minute | Still not five elapsed minutes |
This table explains the rule; it is not benchmark data from a customer fleet. In IronAgents, samples more than ninety seconds apart reset a pending resource window, and server receipt time governs elapsed duration.
Give recovery its own rule
If the same 90% boundary both opens and closes an incident, readings near it can cause repeated state changes. A lower recovery threshold adds separation. For example, clear at or below 80% for two minutes.
Once an incident is open, a reading of 85% is in the middle band. It leaves the incident open. Missing data is not recovery either. A clear condition needs fresh evidence for its own duration. This behavior is often called hysteresis; the practical value is a less jumpy incident history.
Know what is being measured
IronAgents reports aggregate host CPU, memory and system-disk values. Linux memory uses total minus available memory, reflecting the kernel's estimate rather than simply subtracting every cached byte. See the Linux kernel's /proc reference.
The Mac and Windows collectors use their platform APIs and accounting. Their values are useful for their own trends but are not guaranteed to match every category in a native task manager. For process-level diagnosis, use the operating system's local tools.
Charts can average longer ranges. When an incident seems inconsistent with a smooth seven-day chart, read its triggering value and timing, then use a shorter range. Do not discard the alert merely because an averaged line hides a brief interval.
Roll out one deliberate rule
Start with an appropriate device or site, leave unrelated rules alone and review the impact preview. In IronAgents, CPU and memory rules are disabled by default. The starting values are 90% for five minutes, clearing at 80% for two, but they are configuration starting points rather than universally recommended operating targets.
Use observed workload behavior to tune the values. Record why an exception exists. Changing an effective rule resets its timing and closes an existing incident as a rule change, which is different from measured recovery.
Keep notifications separate from observations
Planned work can warrant a maintenance window while metrics continue. A missing email might reflect maintenance, stale readings or delivery state rather than a missing incident. Verify the incident and notification attempt independently.
The resource-alert documentation explains IronAgents's timing and recovery semantics. Production customer email delivery remains pending; private development-inbox testing is not an assurance of delivery to your mailbox.