The box underneath it all.
Sometimes the application is fine and the machine is not. A single agent per host reports what the box is actually doing, on the same timeline as the sessions and traces, so you can rule the infrastructure in or out instead of arguing about it.
The box, on the same timeline as the visit.
8 stops, each one a real screen from the console.
A page of hosts, and the one that isn't fine is first
An average across a fleet hides the one box about to hurt you. The card names it instead, and the list is ranked by health rather than by hostname, so the row you need is at the top whether you run three machines or three hundred.
Six hours of the box, on one click
The row opens where it sits, and the second chart is the one that matters: CPU split into I/O wait and steal. That split is the difference between a machine that is busy and a machine that is blocked — the first needs nothing from you.
The numbers that explain the headline ones
CPU percentage tells you a box is busy. It never tells you a box is stalled — waiting on memory it hasn't got, or on a disk that can't keep up. Pressure stall does, and it is on the same six hours as everything else.
One verdict, and how long the disk has left
One verdict for the host, with its reasoning under it — and a forecast that answers the question you actually have, which is how long you have got. Seventeen days on this disk, at the rate its daily peaks are climbing.
Is nginx up, and turn on tracing from here
Watch the units you care about, then instrument the app running on them without opening an SSH session. Enable attaches the OpenTelemetry agent from here, health-gated, and rolls itself back if the service comes up unhappy.
Which service is eating the host
“The box is at 90%” is not an answer. “Postgres is” is. cgroup v2 makes every service and container a leaf, so the host's CPU and memory land on whatever actually spent them — and on a web host, so does its own access log.
db-01 by workload: postgresql, pgbouncer, the backup container, the agent itself.
What was running when it spiked, and what the kernel killed
A spike you can still investigate after it is over. The agent takes a top-style snapshot while it happens and the console writes the verdict over it: which process, what it was running, and whether anything else was starved.
Rules on the host, the fleet, or any host at all
A rule sits on one host, on a fleet, or on every host at once — and it closes itself when the metric recovers, so nobody has to remember to clear an alert that stopped being true an hour ago.
Everything under infrastructure
One lightweight agent per host reporting CPU, memory, disk and network — so “the site feels slow” can be answered with the machine it runs on.
Host metrics every ten seconds
CPU including I/O wait, memory and swap, load, pressure stall (PSI), disk usage and I/O, network throughput and packets — from one agent.
Fleet at a glance
Hosts reporting, fleet health with the worst host named, fleet network in and out, total cores and RAM.
One row per host
CPU, memory, disk, load per core and network in/out, coloured against alert thresholds — the noisy box surfaces itself.
Host detail in place
A row opens on the page it lives on: CPU and memory, load, network throughput, TCP, disk and pressure over 1 h, 6 h, 24 h or 7 d.
Health, services, workloads, diagnostics
Beyond metrics: what is running, the services the agent found, and a process snapshot taken when CPU spikes.
Lifecycle, not zombies
Active, inactive after an hour of silence, retired on a clean shutdown. A host that goes away stops alarming.
Fleets
Group hosts with --group web. The group is shown on the host and is what an alert rule scopes to — including fleet size, so an autoscaler that drops below two members pages you.
Enable APM and logs from here
Once a host reports, turn on tracing or log shipping for a service on it in one click. No second agent.
Agent upgrades
Hosts on an old agent are flagged; a blessed version can be rolled out from the console, or the command copied.
Why it matters.
One agent, whole fleet
Install once per host. CPU, memory, disk, network and process-level detail report in without per-service configuration.
The worst box first
Rows are ranked by health, not alphabetically, so the machine that is drifting is at the top of a fleet of any size instead of hiding in an average.
Alerts on what matters
Thresholds held over a window — memory, CPU, I/O wait, disk, fleet size, a host gone quiet — delivered through the same inbox and email as everything else.
Minutes, not quarters.
Install the agent
One command per host with an ingest key. Give it a --group and it joins a fleet.
Metrics roll up
Samples are aggregated server-side into the same time series everything else uses.
Answer the question
When a session was slow, check whether the box was busy at that exact minute — then enable APM on the service that was.