Two Failure Modes of Polling
The conventional approach has an agent read the SMART log on a timer. Long intervals miss transient events — a thermal spike or a single media error can come and go between samples. Short intervals impose real management-path overhead across a thousand-drive fleet, and every Get Log Page consumes a queue slot that competes with production I/O. Neither end of the dial is satisfactory.
Retrieve-Then-Re-Arm Semantics
AEN works by having the host post several Asynchronous Event Request commands to the drive in advance. They do not return; they sit pending. When an event occurs the drive completes one of them, returning the event type and the identifier of the associated log page. The host must then read that log page to clear the condition and post a fresh AER to re-arm the channel. The consequence is that a driver which forgets to re-post silently loses every subsequent event. This is the most common integration defect, and it manifests as fewer alerts rather than errors, which is easily mistaken for a healthy fleet.
Which Events Deserve the Channel
Events fall into three groups: error status such as media and internal errors; health status changes including reliability degradation, temperature threshold crossings, spare capacity below threshold, and read-only transitions; and notices such as namespace attribute changes, firmware activation, and Telemetry log updates. An enterprise drive should support all three and let the host enable them bit by bit through Set Features, so machines in different roles subscribe only to what they act on and real alerts are not buried in noise.
Division of Labour with SMART, Telemetry and NVMe-MI
These four complement rather than replace one another. AEN says something happened, quickly, carrying very little payload. The SMART log carries quantitative health counters for trend analysis. Telemetry captures an internal state snapshot at the moment of failure for offline analysis. NVMe-MI runs out of band, still readable when the host OS is down or the drive will not answer in-band commands. AEN without Telemetry tells operators that something broke but not why; SMART without AEN is always one polling interval behind.
Suppressing Chatter and False Positives
Temperature oscillating around a threshold produces alert storms. Firmware needs hysteresis windows and a minimum reporting interval, coalescing same-class events within a window and using different thresholds for rising and falling crossings. J750/J770 are still in product definition/EVT, and these parameters depend on the customer's actual chassis thermal profile, so they are calibrated during platform bring-up. The event coverage list and suppression thresholds are being validated item by item on the internal JH-E4U-C engineering platform and will ship with the firmware release notes, giving host-side subscription configuration a definite reference to work against.