omar@omr.sh:~/notes$

What Should I Use to Monitor My Infrastructure?

I went looking for a monitoring tool and quickly ran into a familiar problem: almost every recommendation answers a slightly different question. Some tools check whether a service responds. Others measure what happens inside a server. Others make the results look good on a dashboard.

So I started looking at the candidates one at a time.

A dark server room with a central rack; on the right, dashboards show CPU, memory and health metrics; on the left, a small external server sits apart, connected by a dashed pulse line — the outside check that would notice if the rack itself went dark.
fig. 01 — The dashboards live next to the rack. The check that matters most lives outside it.

Uptime Kuma: is it reachable?

Uptime Kuma is the easiest one to understand. Give it an HTTP(S) endpoint, a TCP port, a ping target or a DNS record, and it checks whether it responds — with intervals as short as twenty seconds and notifications through dozens of channels.

That makes it appealing for websites, exposed endpoints and services I expect to be available. A reachable port or a successful ping does not by itself prove that a WireGuard tunnel is actually passing traffic correctly, though. And an HTTPS check cannot tell me why a machine is struggling: a server can answer requests while its disk is nearly full or an array is degraded. Kuma would answer an important question, just not every question.

Beszel: what is happening inside the machine?

Beszel takes the next step. An agent on each host reports resource usage to a central hub: CPU, memory, disks, network and Docker containers. It keeps history and can raise alerts. The setup looks approachable for a small collection of servers, which explains much of its appeal in homelab discussions.

Would it be enough on its own? Closer than I expected, but the details matter. Beszel documents two different things here: per-disk S.M.A.R.T. data, which comes from smartctl and needs real permissions on the agent — raw disk access for SATA, admin capabilities for NVMe, smartctl 7.0 or newer — and, separately on Linux, the health of an mdraid array. Titan has two mdraid arrays, md0 and md1, both healthy today. Whether Beszel actually reports a real degradation there, and whether the alert reaches me when it happens, is something I still need to set up and test, not something I’ve tried yet. At least one open report also describes a healthy array getting flagged as WARNING, which is reason enough to watch it carefully rather than trust a green status on day one. I would also still want an independent check for public services — a healthy server does not guarantee that its website is reachable.

Zabbix: should I use one broader system?

Zabbix is tempting for a different reason. It’s built to watch hosts, network equipment, services and applications together in one place, with templates, automatic discovery and event-driven alerting covering most of it. If I wanted one central monitoring platform and expected the infrastructure to grow, it would deserve a serious look.

The question is whether I want to learn and maintain that platform now. It is not something you point at a server and walk away from — hosts, items, triggers and templates take real setup time, which is exactly the complaint I keep seeing from people running small labs. For a small lab, a broad feature set can be an advantage or simply more configuration than I need. I would choose Zabbix if consolidating checks became more valuable than getting a few essential alerts running quickly.

Prometheus and Grafana: how deep do I want to go?

Prometheus collects metrics over time; Grafana lets me query and display them in highly customizable dashboards. This is the most interesting route if I want to correlate resource usage, network behaviour and application metrics during an incident.

It also gives me more pieces to configure: exporters, targets, retention, queries, dashboards and alert rules. None of those alerts would come from Claude Code — Prometheus itself evaluates its rules continuously, the same way a Kuma or Beszel check fires on its own, at 3am, without anyone asking it to. Actually getting a notification out of a firing rule needs Alertmanager, or alerts configured directly in Grafana — Prometheus on its own only knows that a condition is true. What Claude Code could actually do is help me keep those rule and dashboard definitions in Git, reviewed and versioned like any other config — it would not remove the need to decide which metrics matter. I would pick this route when I have questions that simpler tools cannot answer, rather than start by collecting everything.

Homepage: do I need a dashboard or a monitor?

Homepage is the odd one out. Its job is to assemble links and widgets into a useful entry point. It does have its own siteMonitor feature — a periodic HTTP request, HEAD falling back to GET, that shows a service as up or down with a response time — and it can also embed the status already computed by Uptime Kuma, Healthchecks or Gatus through their widgets. That’s a real check, not nothing. But it stops at the current state: no history of past outages, no escalation, no notification channel of its own. For that I would still want a dedicated monitor doing the actual watching, with Homepage as the entry point in front of it rather than the thing recording what happened at 3am.

That distinction matters for Titan. I can imagine Homepage as its visual control panel, Beszel for host health, and a small set of uptime checks for services such as Pulsar. Those are candidates, not a deployed stack. Whichever tools I choose, the small VPS outside Titan must check whether Titan itself is reachable. A monitor running only inside the host cannot report that the host has disappeared.

For now, the useful decision is not which dashboard looks best. It is which failures I need to detect first — and which tool can tell me about each one without becoming another infrastructure project.


← ~/notes