The host collector reads Linux kernel counters from /proc and hardware sensors from /sys and turns them into Prometheus-style metrics with familiar node_* names — CPU per mode, load average, memory, network, disk. It is off by default and enabled per host.enabled in the config.

host:
  enabled: true
  proc_path: "/host/proc"
  sys_path: "/host/sys"
  filesystems_path: "/host/fs"
  interval: 5s

The bundled compose file bind-mounts /proc:/host/proc:ro and /sys:/host/sys:ro so the container sees the real kernel's counters. The bundled Host dashboard plots them.

CPU by mode

percent
—

Docker Desktop VM caveat

On Linux hosts, the container shares the host kernel directly. Bind-mounting /proc gives the collector an honest view of the real host's CPU, memory, load, disk and network.

On macOS and Windows, Docker Desktop runs a hidden Linux VM. The "host" the container sees is that VM, not the underlying OS. The collector will surface the VM's metrics — useful for confirming the loop works end-to-end, but not the OS-level metrics you would see on a production Linux deployment. There is no clean fix from owl's side; it is a property of Docker Desktop's architecture.

In practice: develop and smoke-test on Docker Desktop, then trust real numbers only when reading from a Linux production host.

What is collected

  • CPU — node_cpu_seconds_total per cpu × mode (user, system, iowait, idle, …). Use rate(...[1m]) to plot percentage usage.
  • Load average — node_load1, node_load5, node_load15.
  • Memory — node_memory_MemTotal_bytes, node_memory_MemAvailable_bytes, and the cached / buffered counterparts used by the Host dashboard.
  • Network — node_network_receive_bytes_total / _transmit_bytes_total per device.
  • Disk IO — node_disk_*_bytes_total and _completed_total per device.
  • Disk space — node_filesystem_{size,free,avail}_bytes, node_filesystem_files{,_free} and node_filesystem_readonly per filesystem, labelled with the host's real device, fstype and mountpoint (see below).
  • Temperatures — node_hwmon_temp_celsius (plus node_hwmon_temp_max_celsius and node_hwmon_temp_crit_celsius), node_hwmon_fan_rpm and node_thermal_zone_temp.

The collector is intentionally lean: no per-process scrape, no /proc/pressure parsing, no GPU, no SMART. owl does not ship gopsutil and hand-rolls the few files it needs to keep the binary small.

Disk space

Disk space panels on the Host dashboard

The disk Docker stores containers on, normally the host's system disk, is measured with no extra configuration: a container's / is an overlay whose free space is that disk's free space, and owl resolves which host filesystem it is from /host/proc/1/mountinfo. When owl runs as a plain binary, / is the real root filesystem. If the host filesystem cannot be resolved, the sample falls back to mountpoint="/" (or the container path), with device and fstype taken from the container's own mount table; they are unknown only if that lookup fails too.

For each additional disk, bind-mount an empty, dedicated folder of it under filesystems_path (default /host/fs):

volumes:
  - /srv/disk1/.owl:/host/fs/disk1:ro
  - /srv/disk2/.owl:/host/fs/disk2:ro

statfs on any folder returns the whole filesystem's figures, so owl reports the disk's size and free space without seeing a single file on it. Do not mount a folder with content: owl would be able to read it. The folder name becomes a name label; mountpoint is the host's real one (/srv/disk1). Entries whose name starts with . are skipped. Two folders on the same filesystem (same device major:minor) are reported once, and the overlay root is never labelled with its own merged mount. As a plain binary, put symlinks to the mountpoints in filesystems_path instead.

node_filesystem_readonly comes from the host's mount options, so a :ro bind mount of an otherwise writable disk does not report as read-only. Only when the host mount cannot be resolved does owl fall back to the statfs read-only flag.

Every probe (stat and statfs) runs concurrently with a 5 s timeout. A path that hangs, such as a dead network mount, is skipped until its call returns; owl logs a warning once when a path starts timing out and an info line when it recovers.

Caveats: for alerts use avail (what non-root users can write), not free, because ext4 reserves about 5 % for root. btrfs reports approximate figures, and ZFS datasets share their pool's free space, so summing them overcounts. On Docker Desktop the "host" is its Linux VM.

Temperatures

Read from /sys/class/hwmon and /sys/class/thermal; this needs only the existing /sys mount. Sensors are read from hwmonN/ and, for legacy drivers that expose them there, from hwmonN/device/. Chip naming matches node_exporter. Virtual machines and most VPSs expose no sensors, in which case the Host dashboard's Temperatures panel shows "No data".

Labels: chip and sensor as node_exporter, plus chip_name (the driver, e.g. k10temp, coretemp) and label (e.g. Tctl, Package id 0; it falls back to the sensor name when the chip gives no label). Joins on node_hwmon_sensor_label or node_hwmon_chip_names are not supported: those series do not exist, use chip_name and label directly.

SATA disk temperatures: load the kernel's drivetemp module on the host (modprobe drivetemp, Linux 5.6 or later) and the disks appear as hwmon chips. Check that periodic reads do not keep your disks from spinning down before relying on it. On some older AMD CPUs k10temp refuses to load (erratum 319) or needs force=1.