Skip to main content

Hosts

The fleet view​

Hosts shows every server as a card (or a compact list, switchable). Each card carries a status dot, the key meters, a 24-hour sparkline and the host's role at a glance. The dot folds in open incidents: a host that is up but has a firing incident is not green. The amber mark a meter earns on the card (CPU, memory or disk running high) holds until that meter clearly recovers, per meter, so a card does not blink between green and amber on a value dancing around the line. Needs attention filters to hosts that actually need a human. Acknowledged incidents do not count as "needs attention": an ack means someone took it.

The host page​

Click a host. The page is built from blocks, each refreshing on its own. Click any chart (or a day in the availability strip) to pin the entire page to that moment; "back to live" resumes.

  • Meters and charts: CPU, memory, disk, network, load, with per-mount disk detail.
  • Disk I/O and pressure: read and write throughput across the host's disks, how busy the busiest disk was, and the two pressure figures, the share of time tasks waited for a CPU and for disk I/O. Pressure is the saturation signal a busy-percent hides: a box at 40% CPU whose tasks still queue is a box with a problem.
  • Availability: a year strip of up/down/degraded days, with short outages that never spanned a whole minute still shown as seconds of downtime on their day. Days beyond your plan's retention render locked. Exclude a period opens a dialog with two date pickers and the list of real heartbeat gaps of the last 30 days that still count against this host, so the window you exclude (a migration, a planned power-down) matches what actually happened; excluded minutes leave the availability math, they are neither uptime nor downtime.
  • Services: what is monitored right now (live checks), plus the full catalog of what was discovered, each with a switch. Turning a service off is complete: probes and log shipping stop, its open incidents close at once, and its entries leave the inventory. Off means "stop asking me about it", not "keep alerting quietly".
  • Sites: every website served by this host with certificate, DNS and domain state, sorted worst-first, and what stands behind each: the PHP version of its pool, or the process and unit that serve it (a site proxied to a Node app reads "node (app.service)"), with a badge when that backend is not running or the site's document root is gone. Sites that point at a CDN are marked "moved" and excluded from the certificate verdicts; a site whose certificate does not cover one of its names says which names.
  • PHP-FPM pools: per-pool workers, limits, and the listen queue. A pool with a queue is saturated; a pool at its limit with no queue has no headroom and is marked "at max".
  • Node services: per-service up/down, crash restarts, event loop and heap once instrumented.
  • Top processes: the host's top consumers of CPU, memory and disk I/O (process names only), on Linux and Windows.
  • Storage and cluster: hosts running replicated storage get a card with DRBD and LINSTOR state (replication links of this node, storage pools, network failsafe, error reports); cluster members get a cluster card with quorum, per-node liveness and open storage incidents across the cluster. Migrations and backups appear as events in the story with the node that ran them.
  • Action required: staged changes waiting for a human, each with a ready-to-paste command that includes its own backup and rollback. This is where APM activations land; see APM.
  • Story and logs: the host's event timeline (deploys, discovery changes, incidents, version hops) and its recent log lines.
  • Security: the host's patch level, open doors, and integrity events, linking into the Security module.

Manage: grants and switches​

The Manage card holds the host-level controls:

  • eBPF and CPU profiler: deep-capability grants. Switching one off reaches the agent within about a minute, stops the machinery with no reload, and withdraws pending staged changes.
  • Deep trace (WordPress hook-level spans) and Node profiling: opt-in grants, off by default.
  • Log shipping: the host-wide off switch for log collection.
  • Baseline: restricts the host to metrics, discovery and heartbeat only, denying every privileged capability at once.
  • Agent key: shows how long ago the per-server key last rotated. Rotate now issues a new key at once, for example after someone had access to the box. The agent switches without dropping its connection, and the old key keeps working until the agent next reconnects. The button is disabled while the host is offline, because the new key can only reach a connected agent, and a second rotation within 10 minutes is refused. A rotation is also refused while the agent has not yet confirmed the previous one by reconnecting with the new key (the button says so); a host with a long-lived connection confirms at its next reconnect, for example after an agent update, and keeps a valid key meanwhile.

Per-host alert rules​

Above Manage, Alert rules for this host lets a host live by its own thresholds: a database server where 90% memory is normal, a batch box where CPU pegging is routine. Switch host rules on and the host follows the catalog defaults plus your host-specific rows, instead of the workspace layer. Only rules whose telemetry actually flows from this host are offered. See Alerting.

Host settings​

Rename the host, pin the agent version, or remove the host (typed confirmation). See Adding a host for pinning and removal semantics.