Monitoring without a monitoring stack
The reflex on "I want to monitor my server" is Prometheus, Node Exporter, Alertmanager and Grafana. Four components, a few hundred megabytes of memory, a configuration language for alert rules, and a dashboard nobody opens after the first week.
On my machine, 6.0 of 7.7 GB of memory are in use. A monitoring stack would have needed more memory than the automation service I ended up monitoring with. That is no sensible ratio for the question I actually want answered: is everything running?
What I do instead
A scheduled job in n8n, a command, a comparison, a message. That is all. Three lines of shell that establish whether a service is running, how much disk is left, when the certificate expires – and a Telegram message when an answer is not what it should be.
The important part is that last clause: a message when something is off. No dashboard you have to actively open. No report that arrives every morning unread and gets filtered away after two weeks. Silence means everything is fine.
For me that covers: are the services running, is there enough disk, are the certificates fresh, did the backup complete, are updates pending. Five questions – and honestly, the five that are the problem 95 percent of the time.
What it cannot do
So this does not read like a silver bullet, here is what I am missing:
- No history. I can see that the disk is 80 percent full now. I cannot see that it was 40 percent last week. That trend is the interesting information.
- No correlation. If three things fail at once I get three messages, not a cause.
- No nuance. "Service is running" is not the same as "service responds within 200 milliseconds".
- And the big one: it does not monitor itself. If the machine goes down, the watchman goes down with it. No error arrives – nothing arrives, and in my setup silence means "all good".
That last point needs something external, without exception. A third-party ping service that raises the alarm when the machine stops reporting in is the one addition you should not skip. It costs nothing and closes the only gap that really hurts.
When the stack earns its keep
I am not against Prometheus. I am against starting with the tool instead of the question. I would switch as soon as any of these applies:
- More than a handful of machines
- Availability committed to customers
- A team that needs to look at the same state together
- Questions about trends and capacity rather than "is it up"
Until then, a Telegram message on deviation is not a compromise, it is simply the proportionate answer. The best monitoring setup is the one you actually run – and a neglected Grafana stack with three muted alert rules watches less, in the end, than three lines of shell that text your phone.