Metrics

What the platform measures, how often, and how long it keeps it.

Every host reports metrics to the control plane continuously. They drive the dashboard charts, the host detail views, and the alert rules.

Collection and retention

   
Metric interval 10 seconds
Host heartbeat 30 seconds
Alert evaluation Every 60 seconds
Retention 168 hours (7 days) by default

Retention is configurable. Ten-second samples for a week is a lot of rows on a large fleet, so raising it has a real cost in database size — and lowering it below a week means a Monday investigation can no longer see the previous weekend.

What is collected

Per host: CPU utilization, memory, load, disk usage and I/O, network throughput, and hardware sensors where the server exposes them — temperatures in particular.

Per VM: CPU, memory, disk I/O and network throughput.

Cluster-wide: aggregates across hosts, plus vCPU and memory allocation against physical capacity.

Allocation is not utilization

The dashboard shows both, and the difference matters.

Allocation is what has been promised to VMs. Utilization is what is being used right now.

vCPU allocation above 100% is normal and expected — oversubscribing CPU is most of the point of virtualization. Memory allocation above 100% is a different matter: a host that fails leaves VMs with nowhere to restart that has the memory they were promised.

Watch allocation for capacity planning and utilization for performance. Neither on its own tells you much: high allocation with low utilization is healthy consolidation; high allocation with high utilization means no headroom.

Where to see them

View Shows
Dashboard Fleet-wide capacity, live load, per-host bars
Host detail CPU/RAM and network I/O over the last six hours, plus sensors
VM detail That VM’s own usage

Live values arrive over a WebSocket, so charts update without reloading the page.

Sensors

Where the hardware reports them, temperatures are shown per host, with thresholds: amber from 68 °C, red from 78 °C.

A host climbing into amber under normal load is usually a cooling problem — a failed fan, a blocked intake, a rack that is warmer than it should be. It is worth an alert rule, because the failure it precedes is abrupt.

Utilization thresholds

CPU, memory and disk bars change colour at 75% (amber) and 90% (red) throughout the console. These are display thresholds — they do not raise alerts by themselves. Create rules for that.

Host health

The heartbeat is what decides whether a host is online. A host that stops heartbeating is eventually declared OFFLINE, which is what starts the failover pipeline.

There is exactly one health check making that decision. A second detector with its own timeout would eventually disagree with the first, and two mechanisms disagreeing about whether a host is dead is precisely the condition that corrupts disks.

External monitoring

The platform exposes metrics in a standard format for scraping, so a deployment can feed an existing monitoring system rather than being watched only from its own console.

That is worth doing. The console’s metrics are stored on the cluster; if the cluster is the thing having a problem, an external system is what still has the history.

Reading a problem

Symptom Look at
A guest is slow That host’s CPU and disk I/O, and its neighbours
Everything on one host is slow Host CPU, memory, and whether storage is degraded
Storage is slow fleet-wide Storage cluster health, and whether a rebuild is running
Network throughput capped Uplink state, bond members, MTU probe result
A host climbing in temperature Sensors, then the hardware

A rebuild in progress is the most common explanation for storage that is suddenly slower without anything else changing. See Replicated storage.