Metric events
The Recent events panel is a compact list of discrete things that happened in the selected range: a kill, a disk fault, a NIC that flapped, a GPU Xid. These are not time-series. A chart can look calm in the same window an event fired.
The panel is omitted entirely when nothing fell in range — not an empty-state card. Each row shows kind, severity (info / warning / critical), entity, source, and time.
Events are sampled and disposable like the rest of metrics. They are history, not an alert product.
Two classes
| Class | Gated by | Examples |
|---|---|---|
| Hardware health | hardwareHealthEventsEnabled on the capability plan (on by default for physical hosts) | SMART/NVMe/RAID, NIC link, fans, thermal, PSU/voltage, memory ECC, GPU faults |
| Operational | Always collected | OOM, hung task, conntrack, filesystem state, TurboFabric mesh, clock sync, topology/boot generation |
Disabling hardware-health events does not drop OOM, filesystem, fabric, or clock events.
Kernel and OS
Out of memory kill
The kernel invoked the OOM killer. A process was sacrificed so the rest of the system could continue. Pair with memory PSI and memory used in the same window.
Hung task
The kernel reported a task stuck in uninterruptible sleep longer than the hung-task timeout (often disk or a blocked driver). Pair with I/O PSI and blocked processes.
Conntrack exhaustion
The connection-tracking table hit its ceiling. New tracked connections fail. See Kernel resource usage.
Filesystems
Read-only
A mounted filesystem flipped to read-only — often the kernel's reaction to I/O errors. Writes start failing while reads may still work.
Disappeared
A mount the daemon had been watching is gone (unmounted, failed, or the device vanished).
Remount
A filesystem was remounted (options or read-write state changed). Info-level context next to the two events above.
Disks and RAID
SMART critical
The drive's SMART attributes crossed a vendor critical threshold (reallocations, pending sectors, and similar). Replace-the-disk territory; pair with block temperature and latency.
NVMe critical
Controller-level NVMe critical warning (spare capacity, reliability, read-only, or similar bits).
NVMe media error
The NVMe device reported media errors. Data at risk on that namespace.
RAID degraded
An array is running without a full complement of members.
RAID rebuild started
A rebuild began. Rebuilds cost I/O — expect latency and PSI to rise until completed or failed.
RAID rebuild completed
The array finished rebuilding. I/O pressure should fall back toward the host's baseline.
RAID rebuild failed
The rebuild did not finish. The array stays degraded — see RAID degraded.
Network and TurboFabric
NIC link down
Carrier lost or returned on a physical interface. A single down/up pair is a blip. Many pairs are flapping.
NIC link up
Carrier returned. Pair with the preceding link down to see how long the interface was dark.
NIC flapping
Link state changed too often in a short window — bad cable, bad NIC, or a switch port in trouble.
TurboFabric peer change
The mesh peer set changed (a peer appeared or left). Operational, not hardware-health.
TurboFabric unavailable
The mesh overlay was not usable. Workloads that need the mesh fail until it recovers.
TurboFabric recovered
The mesh overlay is usable again after unavailable.
Thermal, power, and memory ECC
Fan fault
A fan stopped reporting a working speed, or the hardware marked it failed.
Fan alarm
A fan is still present but outside its expected range (too slow, too fast, or the controller raised an alarm). Temperatures on physical signals and GPU usually follow.
Temperature alarm
A probe crossed its warning threshold. The matching gauge is on physical signals or GPU.
Temperature critical
A probe crossed its critical threshold — the vendor's limit, not a TurboPanel policy.
PSU fault
A power supply reported a fault (failed, missing, or out of spec).
Voltage alarm
A rail left its allowed band.
ECC corrected
A memory error the hardware fixed. Occasional corrected errors happen; a rising rate is a DIMM that will get worse.
ECC uncorrected
A memory error that was not corrected — processes can crash or the host can panic.
GPU
Xid
NVIDIA Xid (or vendor equivalent) — the GPU driver reported a fault. Some Xids are application errors; others are the card.
ECC
GPU memory ECC error. Same seriousness ladder as host ECC: corrected vs uncorrected lives in the event payload when the vendor provides it.
Row remap
The GPU remapped a bad memory row. The card is still running; this is wear.
Retirement
The driver retired a memory page. Accumulating retirements is a card approaching failure.
Fallen off bus
The GPU disappeared from PCIe. Often power, seating, or a crashed device.
Thermal critical
GPU temperature crossed the vendor critical limit. See throttle.
Disappeared
The GPU is gone from inventory (not just off the bus for a moment) — uninstalled, failed, or a VM lost its vGPU.
Clock and topology
Clock sync lost
NTP (or the configured time source) lost sync. Metrics timestamps can skew while sync is lost; compare hosts carefully in that window.
Clock sync restored
Time sync is healthy again after clock sync lost.
Topology generation changed
The daemon's enumerated set of NICs, disks, GPUs, filesystems, or sensors changed shape. Charts draw a vertical divider here so two different devices never look like one line.
Boot generation changed
The daemon restarted (or the host rebooted in a way that reset that counter). Rate metrics re-baseline after this.
Related
Last updated on